{"doc_id": "007ff2ca5f297b04636699ce4d01ca6d6f21dc77", "text": ["document", ":", "Attention", "Boosted", "Sequential", "Inference", "Model", "Attention", "mechanism", "has", "been", "proven", "effective", "on", "natural", "language", "processing", ".", "This", "paper", "proposes", "an", "attention", "boosted", "natural", "language", "inference", "model", "named", "aESIM", "by", "adding", "word", "attention", "and", "adaptive", "direction", "-", "oriented", "attention", "mechanisms", "to", "the", "traditional", "Bi", "-", "LSTM", "layer", "of", "natural", "language", "inference", "models", ",", "e.g.", "ESIM", ".", "This", "makes", "the", "inference", "model", "aESIM", "has", "the", "ability", "to", "effectively", "learn", "the", "representation", "of", "words", "and", "model", "the", "local", "subsentential", "inference", "between", "pairs", "of", "premise", "and", "hypothesis", ".", "The", "empirical", "studies", "on", "the", "SNLI", ",", "MultiNLI", "and", "Quora", "benchmarks", "manifest", "that", "aESIM", "is", "superior", "to", "the", "original", "ESIM", "model", ".", "acmlicensed", "[", "DAPA’19", "]", "DAPA2019WSDMWorkshoponDeepmatchinginPracticalApplicationsFebruary15th", ",", "2019Melbourne", ",", "Australia", "2019", "2019", "printacmref", "=", "false", "1234", "-", "5678", "-", "9012", "1234", "-", "5678", "-", "9012", "1234", "-", "5678", "-", "9012", "section", ":", "Introduction", "Natural", "language", "inference", "(", "NLI", ")", "is", "an", "important", "and", "significant", "task", "in", "natural", "language", "processing", "(", "NLP", ")", ".", "It", "concerns", "whether", "a", "hypothesis", "can", "be", "inferred", "from", "a", "premise", ",", "requiring", "understanding", "of", "the", "semantic", "similarity", "between", "the", "hypothesis", "and", "the", "premise", "to", "discriminate", "their", "relation", ".", "Table", "[", "reference", "]", "shows", "several", "samples", "of", "natural", "language", "inference", "from", "SNLI", "(", "Stanford", "Natural", "Language", "Inference", ")", "corpus", ".", "In", "the", "literature", ",", "the", "task", "of", "NLI", "is", "usually", "viewed", "as", "a", "relation", "classification", ".", "It", "learns", "the", "relation", "between", "a", "premise", "and", "a", "hypothesis", "in", "a", "large", "training", "set", ",", "then", "predicts", "the", "relation", "between", "a", "new", "pair", "of", "premise", "and", "hypothesis", ".", "The", "existing", "methods", "of", "NLI", "can", "be", "roughly", "partitioned", "into", "two", "categories", ":", "feature", "-", "based", "models", "and", "neural", "network", "-", "based", "models", ".", "Feature", "-", "based", "models", "represent", "a", "premise", "and", "a", "hypothesis", "by", "their", "unlexicalized", "and", "lexicalized", "features", ",", "such", "as", "-", "gram", "length", "and", "the", "real", "-", "valued", "feature", "of", "length", "difference", ",", "then", "train", "a", "classifier", "to", "perform", "relation", "classification", ".", "Recently", ",", "end", "-", "to", "-", "end", "neural", "network", "-", "based", "models", "have", "drawn", "worldwide", "attention", "since", "they", "have", "demonstrated", "excellent", "performance", "on", "quite", "a", "few", "NLP", "tasks", "including", "machine", "translation", ",", "natural", "language", "inference", ",", "etc", ".", "On", "the", "basis", "of", "their", "model", "structures", ",", "we", "can", "divide", "neural", "network", "-", "based", "models", "for", "NLI", "into", "two", "classes", ",", "sentence", "encoding", "models", "and", "sentence", "interaction", "-", "aggregation", "models", ".", "The", "architectures", "of", "the", "two", "types", "of", "models", "are", "shown", "in", "Figure", "[", "reference", "]", ".", "Sentence", "encoding", "models", "(", "their", "main", "architecture", "is", "shown", "in", "Figure", "[", "reference", "]", ".a", ")", "independently", "encode", "a", "pair", "of", "sentences", ",", "a", "premise", "and", "a", "hypothesis", "using", "pre", "-", "trained", "word", "embedding", "vectors", ",", "then", "learn", "semantic", "relation", "between", "two", "sentences", "with", "a", "multi", "-", "layer", "perceptron", "(", "MLP", ")", ".", "In", "these", "models", ",", "LSTM", "(", "Long", "Short", "-", "Term", "Memory", "networks", ")", ",", "its", "variants", "GRU", "(", "Gated", "Recurrent", "Units", ")", "and", "Bi", "-", "LSTM", ",", "are", "usually", "utilized", "to", "encode", "the", "sentences", "since", "they", "were", "capable", "of", "learning", "long", "-", "term", "dependencies", "inside", "sentences", ".", "For", "example", ",", "Conneau", "et", "al", ".", "proposed", "a", "generic", "NLI", "training", "scheme", "and", "compared", "several", "sentence", "encoding", "architectures", ":", "LSTM", "or", "GRU", ",", "Bi", "-", "LSTM", "with", "mean", "/", "max", "pooling", ",", "self", "-", "attention", "network", "and", "hierarchical", "convolutional", "networks", ".", "The", "experimental", "results", "demonstrated", "that", "the", "Bi", "-", "LSTM", "with", "max", "pooling", "achieved", "the", "best", "performance", ".", "Talman", "et", "al", ".", "designed", "a", "hierarchical", "Bi", "-", "LSTM", "max", "pooling", "(", "HBMP", ")", "model", "to", "encode", "sentences", ".", "This", "model", "applied", "parameters", "of", "one", "Bi", "-", "LSTM", "to", "initialize", "the", "next", "Bi", "-", "LSTM", "to", "convey", "information", ",", "which", "shown", "better", "results", "than", "the", "model", "with", "a", "single", "Bi", "-", "LSTM", ".", "Besides", "LSTM", ",", "attention", "mechanisms", "could", "also", "be", "used", "to", "boost", "the", "effectiveness", "of", "sentence", "encoding", ".", "The", "model", "developed", "by", "Ghaeini", "et", "al", ".", "added", "self", "-", "attention", "to", "LSTM", "model", ",", "and", "achieved", "better", "performance", ".", "Sentence", "interaction", "-", "aggregation", "models", "(", "their", "main", "architecture", "is", "shown", "in", "Figure", "[", "reference", "]", ".b", ")", "learn", "vector", "representations", "of", "pairs", "of", "sentences", "in", "the", "way", "similar", "to", "sentence", "encoding", "models", "and", "calculate", "pairwise", "word", "interaction", "matrix", "between", "two", "sentences", "using", "the", "newly", "updated", "word", "vectors", ",", "and", "then", "the", "matching", "results", "are", "aggregated", "into", "a", "vector", "to", "make", "the", "final", "decision", ".", "Compared", "with", "sentence", "encoding", "model", ",", "sentence", "interaction", "-", "aggregation", "models", "aggregate", "word", "similarities", "between", "a", "pair", "of", "sentences", ",", "are", "capable", "of", "capturing", "the", "relevant", "information", "between", "two", "sentences", ",", "a", "premise", "and", "a", "hypothesis", ".", "Bahdanau", "et", "al", ".", "translated", "and", "aligned", "text", "simultaneously", "in", "machine", "translation", "task", ",", "innovatively", "introducing", "attention", "mechanism", "to", "natural", "language", "process", "(", "NLP", ")", ".", "He", "et", "al", ".", "designed", "a", "pairwise", "word", "interaction", "model", "(", "PWIM", ")", ",", "which", "made", "full", "use", "of", "word", "-", "level", "fine", "-", "grained", "information", ".", "Wang", "et", "al", ".", "put", "forward", "a", "bilateral", "multi", "-", "perspective", "matching", "(", "BiMPM", ")", "model", ",", "focusing", "on", "various", "matching", "strategies", "that", "could", "be", "seen", "as", "different", "types", "of", "attention", ".", "The", "empirical", "studies", "of", "Lan", "et", "al", ".", "and", "Chen", "et", "al", ".", "concluded", "that", "sentence", "interation", "-", "aggregation", "models", ",", "especially", "ESIM", "(", "Enhanced", "Sequential", "Inference", "Model", ")", ",", "a", "carefully", "designed", "sequential", "inference", "model", "based", "on", "chain", "LSTMs", ",", "outperformed", "all", "previous", "sentence", "encoding", "models", ".", "Although", "ESIM", "has", "achieved", "excellent", "achievements", ",", "this", "model", "does", "n’t", "consider", "the", "attention", "along", "the", "words", "in", "a", "sentence", "in", "its", "Bi", "-", "LSTM", "layer", ".", "Word", "attention", "can", "characterize", "the", "different", "contribution", "of", "each", "word", ".", "Therefore", ",", "it", "will", "be", "beneficial", "to", "put", "word", "attention", "into", "the", "Bi", "-", "LTSM", "layer", ".", "Moreover", ",", "the", "orientation", "of", "the", "words", "represents", "the", "direction", "of", "the", "information", "flow", ",", "either", "forward", "or", "backward", ",", "should", "not", "be", "ignored", ".", "In", "traditional", "Bi", "-", "LSTM", "model", ",", "the", "forward", "and", "the", "backward", "vectors", "learnt", "by", "Bi", "-", "LSTM", "are", "simply", "jointed", ".", "It", "’s", "necessary", "to", "consider", "whether", "each", "orientation", "(", "forward", "or", "backward", ")", "has", "different", "importance", "on", "word", "encoding", ",", "thus", "adaptively", "joint", "the", "two", "orientation", "vectors", "together", "with", "different", "weights", ".", "Therefore", ",", "in", "this", "study", ",", "using", "ESIM", "model", "as", "the", "baseline", ",", "we", "add", "an", "attention", "layer", "behind", "each", "Bi", "-", "LSTM", "layer", ",", "then", "use", "an", "adaptive", "orientation", "embedding", "layer", "to", "jointly", "represent", "the", "forward", "and", "backward", "vectors", ".", "We", "name", "this", "attention", "boosted", "Bi", "-", "LSTM", "as", "Bi", "-", "aLSTM", ",", "and", "denote", "the", "modified", "ESIM", "as", "aESIM", ".", "Experimental", "results", "on", "SNLI", ",", "MultiNLI", "and", "Quora", "benchmarks", "have", "demonstrated", "better", "performance", "of", "aESIM", "model", "than", "that", "of", "the", "baseline", "ESIM", "and", "the", "other", "state", "-", "of", "-", "the", "-", "art", "models", ".", "We", "believe", "that", "the", "architecture", "of", "Bi", "-", "aLSTM", "has", "potentially", "to", "be", "used", "in", "other", "NLP", "tasks", "such", "as", "text", "classification", ",", "machine", "translation", "and", "so", "on", ".", "This", "paper", "is", "organized", "as", "follows", ".", "We", "introduce", "the", "general", "frameworks", "of", "ESIM", "and", "aESIM", "in", "Section", "2", ".", "We", "describe", "the", "datasets", "and", "the", "experiment", "settings", ",", "and", "analyze", "our", "experimental", "results", "in", "Section", "3", ".", "We", "then", "draw", "conclusions", "in", "Section", "4", ".", "section", ":", "Attention", "Boosted", "Sequential", "Inference", "Model", "Supposed", "that", "we", "have", "two", "sentences", "and", ",", "where", "represents", "premise", "and", "represents", "hypothesis", ".", "The", "goal", "is", "to", "predict", "the", "label", "meaning", "for", "their", "relation", ".", "subsection", ":", "ESIM", "model", "Enhanced", "Sequential", "Inference", "Model", "(", "ESIM", ")", "is", "composed", "of", "four", "main", "components", ":", "input", "encoding", "layer", ",", "local", "inference", "modeling", "layer", ",", "inference", "composition", "layer", "and", "classification", "layer", ".", "In", "the", "input", "encoding", "layer", ",", "ESIM", "first", "uses", "Bi", "-", "LSTM", "layer", "to", "encode", "input", "sentence", "pairs", "(", "Equations", "1", "-", "2", ")", ",", "which", "can", "be", "initialized", "using", "pre", "-", "trained", "word", "embeddings", "(", "e.g.", "Glove", "840B", "vectors", ")", ",", "where", "is", "the", "word", "embedding", "vector", "of", "the", "-", "th", "word", "in", ",", "is", "that", "of", "word", "in", ".", "Secondly", ",", "ESIM", "implements", "the", "local", "inference", "layer", "for", "enhancing", "the", "sentence", "information", ".", "First", "it", "calculates", "a", "similarity", "matrix", "based", "on", "and", ".", "It", "then", "gets", "the", "new", "expression", "for", "and", "with", "the", "equation", "below", ":", "where", "and", "represent", "the", "weighted", "summation", "of", "and", ".", "It", "further", "enhances", "the", "local", "inference", "information", "collected", "as", "below", ".", "After", "the", "enhancement", "of", "local", "inference", ",", "another", "Bi", "-", "LSTM", "layer", "is", "used", "to", "capture", "local", "inference", "information", "and", "their", "context", "for", "inference", "composition", ".", "Instead", "of", "summation", "adopted", "by", "Parikh", "et", "al", ".", ",", "ESIM", "proposes", "to", "compute", "both", "max", "and", "average", "pooling", "and", "feeds", "the", "concatenate", "fixed", "length", "vector", "to", "the", "final", "classifier", ":", "a", "fully", "connected", "multi", "-", "layer", "perceptron", ".", "Figure", "[", "reference", "]", "shows", "a", "high", "-", "level", "view", "of", "the", "ESIM", "architecture", ",", "where", "the", "bottom", "LSTM1", "layer", "of", "Figure", "[", "reference", "]", "is", "the", "input", "encoding", "layer", ",", "the", "middle", "part", "with", "LSTM2", "layer", "is", "the", "local", "inference", "layer", ",", "the", "upper", "part", "is", "the", "inference", "composition", "layer", ".", "subsection", ":", "aESIM", "model", "The", "overall", "architecture", "of", "our", "newly", "proposed", "attention", "boosted", "sequential", "inference", "model", "(", "named", "aESIM", ")", "based", "on", "ESIM", "is", "similar", "to", "ESIM", ".", "In", "detail", ",", "aESIM", "also", "consists", "of", "four", "main", "parts", ":", "encoding", "layer", ",", "local", "inference", "modeling", "layer", ",", "decoding", "layer", "and", "classification", "layer", ".", "The", "only", "difference", "between", "ESIM", "and", "aESIM", "is", "that", "we", "substitute", "the", "two", "Bi", "-", "LSTM", "layers", "(", "LSTM1", "and", "LSTM2", ")", "in", "ESIM", "with", "two", "Bi", "-", "aLSTM", "layers", "in", "aESIM", ".", "Therefore", ",", "as", "illustrated", "in", "Figure", "[", "reference", "]", ",", "the", "layers", "with", "red", "-", "dotted", "circles", "in", "ESIM", "will", "be", "replaced", "by", "the", "Bi", "-", "aLSTM", "layers", "shown", "in", "the", "right", "upper", "corner", "of", "the", "Figure", "[", "reference", "]", "and", "the", "details", "of", "Bi", "-", "aLSTM", "can", "be", "found", "in", "Figure", "[", "reference", "]", ".", "Given", "the", "word", "vector", "of", "the", "-", "th", "word", "in", "sentence", ",", "which", "can", "be", "obtained", "by", "pre", "-", "trained", "word", "embeddings", "such", "as", "Glove", "840B", "vectors", "in", "the", "first", "Bi", "-", "aLSTM", "layer", "or", "obtained", "from", "the", "local", "inference", "modeling", "layer", "in", "the", "second", "Bi", "-", "aLSTM", "layer", ".", "We", "utilize", "a", "forward", "LSTM", "layer", "and", "a", "backward", "LSTM", "layer", "to", "collect", "both", "direction", "information", "and", ".", "As", "described", "in", "introduction", "section", ",", "in", "the", "following", "newly", "proposed", "Bi", "-", "aLSTM", ",", "we", "add", "word", "attention", "and", "additive", "operation", "on", "both", "orientations", "of", "traditional", "Bi", "-", "LSTM", "layer", ".", "Word", "attention", "layer", "It", "’s", "obvious", "that", "not", "all", "words", "contribute", "equally", "to", "the", "representation", "of", "a", "sentence", ".", "Attention", "mechanism", ",", "which", "is", "introduced", "in", ",", "is", "extremely", "effective", "to", "extract", "vital", "words", "from", "the", "whole", "sentence", ",", "and", "is", "particularly", "beneficial", "to", "generate", "the", "sentence", "vector", ".", "Therefore", ",", "we", "use", "the", "following", "attention", "mechanism", "after", "we", "get", "and", ".", "Suppose", ",", "we", "then", "have", "where", "is", "obtained", "after", "one", "-", "layer", "MLP", "for", "the", "input", ",", "is", "the", "importance", "of", "word", ",", "is", "calculated", "by", "the", "SoftMax", "unit", "on", "the", "context", "vector", "of", "the", "sentence", "which", "is", "randomly", "initialized", "and", "modified", "during", "the", "training", ",", "is", "the", "attention", "enhanced", "vector", "through", "multiplying", "the", "weight", "and", "original", "vector", ",", "where", "correspond", "to", "the", "forward", "vector", "and", "the", "backward", "vector", ",", "respectively", ".", "Adaptive", "word", "direction", "layer", "In", "traditional", "Bi", "-", "LSTM", "model", ",", "the", "forward", "and", "the", "backward", "vectors", "of", "a", "word", "are", "considered", "to", "have", "equal", "importance", "on", "the", "word", "representation", ".", "The", "model", "simply", "connects", "the", "forward", "and", "backward", "vectors", "head", "and", "tail", "without", "weighing", "their", "importance", ".", "For", "a", "word", "in", "different", "direction", "or", "orientation", ",", "the", "former", "and", "the", "latter", "words", "are", "reversed", ".", "Thus", ",", "different", "direction", "vectors", "of", "a", "word", "make", "different", "contribution", "to", "the", "representation", ",", "especially", "the", "words", "in", "a", "long", "sentence", ".", "Therefore", ",", "we", "propose", "a", "new", "adaptive", "direction", "layer", "to", "learn", "the", "contribution", "of", "different", "directions", "for", "a", "single", "word", ".", "Formally", ",", "given", "two", "direction", "word", "vectors", "and", ",", "the", "whole", "word", "vector", "can", "be", "expressed", "as", ":", "where", ",", "and", "denote", "weight", "matrix", "and", "the", "bias", ",", "denotes", "the", "nonlinear", "function", ",", "denotes", "the", "concentration", ".", "All", "the", "parameters", "can", "be", "learned", "during", "training", ".", "Then", "we", "can", "get", "the", "whole", "sentence", "vector", "as", "below", ":", "This", "word", "and", "orientation", "enhanced", "Bi", "-", "LSTM", "is", "called", "Bi", "-", "aLSTM", ".", "Its", "whole", "architecture", "is", "shown", "in", "the", "Figure", "[", "reference", "]", ",", "is", "applied", "in", "ESIM", "model", "to", "replace", "the", "two", "Bi", "-", "LSTM", "layers", "for", "the", "task", "of", "natural", "language", "inference", ".", "Besides", ",", "this", "Bi", "-", "aLSTM", "can", "be", "used", "to", "other", "natural", "language", "processing", "tasks", "and", "our", "preliminary", "experiments", "have", "demonstrated", "that", "Bi", "-", "aLSTM", "is", "capable", "of", "improving", "the", "performance", "of", "Bi", "-", "LSTM", "models", "on", "sentimental", "classification", "task", "(", "for", "space", "limitation", ",", "this", "results", "will", "not", "be", "shown", "in", "the", "paper", ")", ".", "section", ":", "Experiment", "Setup", "subsection", ":", "Datasets", "We", "evaluated", "our", "model", "on", "three", "datasets", ":", "the", "Stanford", "Natural", "Language", "Inference", "(", "SNLI", ")", "corpus", ",", "the", "Multi", "-", "Genre", "Natural", "Language", "Inference", "(", "MultiNLI", ")", "corpus", ",", "and", "Quora", "duplicate", "question", "dataset", ".", "We", "selected", "these", "three", "relatively", "large", "corpora", "out", "of", "eight", "corpora", "in", "since", "deep", "learning", "models", "usually", "show", "better", "generalization", "ability", "on", "large", "training", "sets", "and", "produce", "more", "convincing", "results", "than", "on", "small", "training", "sets", ".", "SNLI", "The", "Stanford", "Natural", "Language", "Inference", "(", "SNLI", ")", "corpus", "contains", "570", ",", "152", "sentence", "pairs", ",", "including", "549", "K", "training", "pairs", ",", "10", "K", "validation", "pairs", "and", "10", "K", "testing", "pairs", ".", "Each", "pair", "has", "one", "of", "relation", "classes", "(", "entailment", ",", "neutral", ",", "contradiction", "and", "‘", "-", "’", ")", ".", "The", "‘", "-", "’", "class", "indicates", "there", "is", "no", "conclusion", "between", "the", "two", "sentences", ".", "Consequently", ",", "we", "remove", "all", "pairs", "with", "relation", "’", "-", "’", "during", "training", ",", "validating", "and", "testing", "processes", ".", "MultiNLI", "This", "corpus", "is", "a", "crowd", "-", "sourced", "collection", "of", "433", "K", "sentence", "pairs", "annotated", "with", "textual", "entailment", "information", ".", "The", "corpus", "is", "modeled", "on", "the", "SNLI", "corpus", ",", "but", "differs", "in", "that", "covers", "a", "range", "of", "genres", "of", "spoken", "and", "written", "text", ",", "and", "supports", "a", "distinctive", "cross", "-", "genre", "generation", "evaluation", ".", "Quora", "The", "Quora", "dataset", "contains", "400", ",", "000", "question", "pairs", ".", "The", "task", "of", "this", "corpus", "is", "to", "judge", "whether", "the", "two", "sentences", "means", "the", "same", "affair", ".", "subsection", ":", "Setting", "We", "use", "the", "validation", "set", "to", "select", "models", "for", "testing", ".", "The", "hyper", "-", "parameters", "of", "aESIM", "model", "are", "listed", "as", "follows", ".", "We", "use", "the", "Adam", "method", "for", "optimization", ".", "The", "first", "momentum", "is", "set", "to", "be", "0.9", "and", "the", "second", "0.999", ".", "The", "initial", "learning", "rate", "is", "set", "to", "0.0005", ",", "and", "the", "batch", "size", "is", "128", ".", "The", "dimensions", "of", "all", "hidden", "states", "of", "Bi", "-", "aLSTM", "and", "word", "embedding", "are", "300", ".", "We", "employ", "non", "-", "linearity", "function", "replacing", "rectified", "linear", "unit", "on", "account", "of", "its", "faster", "convergence", "rate", ".", "Dropout", "rate", "is", "set", "to", "0.2", "during", "training", ".", "We", "use", "pre", "-", "trained", "300", "-", "D", "Glove", "840B", "vectors", "to", "initialize", "word", "embeddings", ".", "Out", "-", "of", "-", "vocabulary", "(", "OOV", ")", "words", "are", "initialized", "randomly", "with", "Gaussian", "samples", ".", "All", "vectors", "are", "updated", "during", "training", ".", "subsection", ":", "Experiment", "results", "Except", "for", "comparing", "our", "method", "aESIM", "with", "ESIM", ",", "we", "listed", "the", "experimental", "results", "of", "methods", "with", "their", "references", "in", "Table", "[", "reference", "]", "on", "SNIL", ".", "In", "Table", "[", "reference", "]", ",", "the", "method", "in", "the", "first", "block", "is", "a", "traditional", "feature", "engineering", "method", ",", "those", "in", "the", "second", "are", "the", "sentence", "vector", "-", "based", "models", ",", "those", "in", "the", "third", "are", "attention", "-", "based", "models", ",", "and", "ESIM", "and", "our", "aESIM", "are", "shown", "in", "the", "fourth", "block", ".", "Where", "the", "results", "of", "ESIM", "and", "aESIM", "are", "implemented", "by", "ourselves", "on", "Keras", ",", "the", "results", "of", "the", "others", "are", "taken", "from", "their", "original", "publications", ".", "We", "then", "compare", "the", "baseline", "models", ",", "CBOW", ",", "Bi", "-", "LSTM", "with", "ESIM", "and", "our", "aESIM", "on", "MultiNLI", "corpus", "shown", "In", "Table", "[", "reference", "]", ",", "where", "the", "results", "of", "the", "baselines", "are", "taken", "from", ".", "Finally", ",", "we", "compare", "several", "types", "of", "CNN", "and", "RNN", "models", "on", "Quroa", "corpus", "shown", "in", "Table", "[", "reference", "]", ",", "the", "results", "of", "theses", "CNN", "and", "RNN", "models", "are", "taken", "from", ".", "The", "accuracy", "(", "ACC", ")", "of", "each", "method", "is", "measured", "by", "the", "commonly", "used", "precision", "score", ",", "and", "the", "methods", "with", "the", "best", "accuracy", "are", "marked", "in", "bold", ".", "According", "to", "the", "results", "in", "Tables", "2", "-", "4", ",", "aESIM", "model", "achieved", "88.1", "%", "on", "SNLI", "corpus", ",", "elevating", "0.8", "percent", "higher", "than", "ESIM", "model", ".", "It", "promoted", "almost", "0.5", "percent", "accuracy", "and", "outperformed", "the", "baselines", "on", "MultiNLI", ".", "It", "also", "achieved", "88.01", "%", "on", "Quora", ".", "Therefore", ",", "we", "concluded", "that", "aESIM", "with", "further", "word", "attention", "and", "word", "orientation", "operation", "was", "superior", "to", "ESIM", "model", ".", "subsection", ":", "Attention", "visualization", "We", "selected", "three", "types", "of", "sentence", "pairs", "from", "a", "premise", "and", "its", "three", "hypothesis", "sentences", "in", "the", "test", "set", "of", "SNLI", "corpus", "as", "shown", "in", "Figure", "[", "reference", "]", ",", "where", "the", "premise", "sentence", "is", "‘", "A", "woman", "with", "a", "green", "headscarf", ",", "blue", "shirt", "and", "a", "very", "big", "grin", "’", ",", "and", "three", "hypothesis", "sentences", "are", "‘", "the", "woman", "has", "been", "shot", "’", ",", "‘", "the", "woman", "is", "very", "happy", "’", "and", "‘", "the", "woman", "is", "young", "’", "with", "relation", "labels", "‘", "contradiction", "’", ",", "‘", "entailment", "’", ",", "and", "‘", "neutral", "’", ",", "respectively", ".", "Each", "pair", "of", "sentences", "has", "their", "key", "word", "pairs", ":", "grin", "-", "shot", ",", "grin", "-", "happy", "and", "grin", "-", "young", ",", "which", "determines", "whether", "the", "premise", "can", "entail", "the", "hypothesis", ".", "Figures", "4.a", "-", "4.c", "are", "the", "visualization", "of", "the", "attention", "layer", "between", "sentence", "pairs", "after", "the", "Bi", "-", "LSTM", "layer", "in", "ESIM", "model", "and", "that", "after", "Bi", "-", "aLSTM", "layer", "in", "aESIM", "model", "for", "contrasting", "ESIM", "and", "aESIM", ".", "By", "doing", "so", ",", "we", "could", "understand", "how", "the", "models", "judge", "the", "relation", "between", "two", "sentences", ".", "In", "each", "Figure", ",", "the", "brighter", "the", "color", ",", "the", "higher", "the", "weight", "is", ".", "We", "could", "conclude", "that", "our", "aESIM", "model", "had", "the", "higher", "weight", "than", "ESIM", "model", "on", "each", "key", "word", "pair", ",", "especially", "in", "Figure", "[", "reference", "]", ".b", ",", "where", "the", "similarity", "of", "‘", "happy", "’", "and", "‘", "grin", "’", "in", "aESIM", "model", "is", "much", "higher", "than", "that", "in", "ESIM", "model", ".", "Therefore", ",", "our", "aESIM", "model", "was", "able", "to", "capture", "the", "most", "important", "word", "pair", "in", "each", "pair", "of", "sentences", ".", "section", ":", "Conclusion", "In", "this", "study", ",", "we", "propose", "an", "improved", "version", "of", "ESIM", "named", "aESIM", "for", "NLI", ".", "It", "modifies", "the", "Bi", "-", "LSTM", "layer", "to", "collect", "more", "information", ".", "We", "evaluate", "our", "aESIM", "model", "on", "three", "NLI", "corpora", ".", "Experimental", "results", "show", "that", "aESIM", "model", "achieves", "better", "performance", "than", "ESIM", "model", ".", "In", "the", "future", ",", "we", "will", "evaluate", "how", "attention", "mechanisms", "can", "be", "applied", "on", "other", "tasks", "and", "explore", "a", "way", "to", "use", "less", "time", "and", "space", "with", "guaranteed", "accuracy", ".", "section", ":", "Acknowledgement", "This", "work", "is", "supported", "in", "part", "by", "the", "National", "Nature", "Science", "Foundation", "of", "China", "(", "No", ".", "61876016", "and", "No", ".", "61632004", ")", ",", "the", "Fundamental", "Research", "Funds", "for", "the", "Central", "Universities", "(", "No", ".", "2018JBZ006", ")", ".", "bibliography", ":", "References"], "entity": [[2, 7, ["Attention", "Boosted", "Sequential", "Inference", "Model"], "Method"], [7, 9, ["Attention", "mechanism"], "Method"], [14, 17, ["natural", "language", "processing"], "Task"], [22, 28, ["attention", "boosted", "natural", "language", "inference", "model"], "Method"], [29, 30, ["aESIM"], "Method"], [32, 34, ["word", "attention"], "Method"], [35, 41, ["adaptive", "direction", "-", "oriented", "attention", "mechanisms"], "Method"], [44, 53, ["Bi", "-", "LSTM", "layer", "of", "natural", "language", "inference", "models"], "Method"], [55, 56, ["ESIM"], "Method"], [60, 63, ["inference", "model", "aESIM"], "Method"], [76, 79, ["local", "subsentential", "inference"], "Task"], [91, 92, ["SNLI"], "Material"], [93, 94, ["MultiNLI"], "Material"], [95, 96, ["Quora"], "Material"], [99, 100, ["aESIM"], "Method"], [105, 107, ["ESIM", "model"], "Method"], [140, 143, ["Natural", "language", "inference"], "Task"], [144, 146, ["NLI", ")"], "Task"], [153, 159, ["natural", "language", "processing", "(", "NLP", ")"], "Task"], [197, 200, ["natural", "language", "inference"], "Task"], [217, 218, ["NLI"], "Task"], [223, 225, ["relation", "classification"], "Task"], [259, 260, ["NLI"], "Task"], [268, 272, ["feature", "-", "based", "models"], "Method"], [273, 278, ["neural", "network", "-", "based", "models"], "Method"], [279, 283, ["Feature", "-", "based", "models"], "Method"], [314, 315, ["classifier"], "Method"], [317, 319, ["relation", "classification"], "Task"], [322, 332, ["end", "-", "to", "-", "end", "neural", "network", "-", "based", "models"], "Method"], [346, 348, ["NLP", "tasks"], "Task"], [349, 351, ["machine", "translation"], "Task"], [352, 355, ["natural", "language", "inference"], "Task"], [369, 374, ["neural", "network", "-", "based", "models"], "Method"], [375, 376, ["NLI"], "Task"], [380, 383, ["sentence", "encoding", "models"], "Method"], [384, 389, ["sentence", "interaction", "-", "aggregation", "models"], "Method"], [406, 409, ["Sentence", "encoding", "models"], "Method"], [451, 457, ["multi", "-", "layer", "perceptron", "(", "MLP"], "Method"], [463, 471, ["LSTM", "(", "Long", "Short", "-", "Term", "Memory", "networks"], "Method"], [475, 480, ["GRU", "(", "Gated", "Recurrent", "Units"], "Method"], [482, 485, ["Bi", "-", "LSTM"], "Method"], [516, 519, ["NLI", "training", "scheme"], "Task"], [522, 525, ["sentence", "encoding", "architectures"], "Method"], [526, 529, ["LSTM", "or", "GRU"], "Method"], [530, 533, ["Bi", "-", "LSTM"], "Method"], [534, 538, ["mean", "/", "max", "pooling"], "Method"], [539, 543, ["self", "-", "attention", "network"], "Method"], [544, 547, ["hierarchical", "convolutional", "networks"], "Method"], [554, 557, ["Bi", "-", "LSTM"], "Method"], [558, 560, ["max", "pooling"], "Method"], [571, 581, ["hierarchical", "Bi", "-", "LSTM", "max", "pooling", "(", "HBMP", ")", "model"], "Method"], [591, 594, ["Bi", "-", "LSTM"], "Method"], [598, 601, ["Bi", "-", "LSTM"], "Method"], [615, 618, ["Bi", "-", "LSTM"], "Method"], [620, 621, ["LSTM"], "Method"], [622, 624, ["attention", "mechanisms"], "Method"], [633, 635, ["sentence", "encoding"], "Task"], [649, 651, ["LSTM", "model"], "Method"], [657, 662, ["Sentence", "interaction", "-", "aggregation", "models"], "Method"], [676, 678, ["vector", "representations"], "Method"], [687, 690, ["sentence", "encoding", "models"], "Method"], [724, 727, ["sentence", "encoding", "model"], "Method"], [728, 733, ["sentence", "interaction", "-", "aggregation", "models"], "Method"], [769, 772, ["machine", "translation", "task"], "Task"], [775, 777, ["attention", "mechanism"], "Method"], [778, 781, ["natural", "language", "process"], "Task"], [782, 783, ["NLP"], "Task"], [791, 795, ["pairwise", "word", "interaction", "model"], "Method"], [796, 798, ["PWIM", ")"], "Method"], [819, 828, ["bilateral", "multi", "-", "perspective", "matching", "(", "BiMPM", ")", "model"], "Method"], [832, 834, ["matching", "strategies"], "Method"], [859, 864, ["sentence", "interation", "-", "aggregation", "models"], "Method"], [866, 867, ["ESIM"], "Method"], [868, 872, ["Enhanced", "Sequential", "Inference", "Model"], "Method"], [877, 880, ["sequential", "inference", "model"], "Method"], [882, 884, ["chain", "LSTMs"], "Method"], [888, 891, ["sentence", "encoding", "models"], "Method"], [893, 894, ["ESIM"], "Method"], [914, 918, ["Bi", "-", "LSTM", "layer"], "Method"], [919, 921, ["Word", "attention"], "Task"], [942, 946, ["Bi", "-", "LTSM", "layer"], "Method"], [974, 978, ["Bi", "-", "LSTM", "model"], "Method"], [987, 990, ["Bi", "-", "LSTM"], "Method"], [1011, 1013, ["word", "encoding"], "Task"], [1033, 1035, ["ESIM", "model"], "Method"], [1042, 1044, ["attention", "layer"], "Method"], [1046, 1050, ["Bi", "-", "LSTM", "layer"], "Method"], [1054, 1058, ["adaptive", "orientation", "embedding", "layer"], "Method"], [1070, 1075, ["attention", "boosted", "Bi", "-", "LSTM"], "Method"], [1076, 1079, ["Bi", "-", "aLSTM"], "Method"], [1084, 1085, ["ESIM"], "Method"], [1086, 1087, ["aESIM"], "Method"], [1091, 1092, ["SNLI"], "Material"], [1093, 1094, ["MultiNLI"], "Material"], [1095, 1096, ["Quora"], "Material"], [1102, 1104, ["aESIM", "model"], "Method"], [1109, 1110, ["ESIM"], "Method"], [1128, 1131, ["Bi", "-", "aLSTM"], "Method"], [1138, 1140, ["NLP", "tasks"], "Task"], [1142, 1144, ["text", "classification"], "Task"], [1145, 1147, ["machine", "translation"], "Task"], [1164, 1165, ["ESIM"], "Method"], [1166, 1167, ["aESIM"], "Method"], [1199, 1204, ["Attention", "Boosted", "Sequential", "Inference", "Model"], "Method"], [1233, 1235, ["ESIM", "model"], "Method"], [1235, 1239, ["Enhanced", "Sequential", "Inference", "Model"], "Method"], [1240, 1241, ["ESIM"], "Method"], [1249, 1252, ["input", "encoding", "layer"], "Method"], [1253, 1257, ["local", "inference", "modeling", "layer"], "Method"], [1258, 1261, ["inference", "composition", "layer"], "Method"], [1262, 1264, ["classification", "layer"], "Method"], [1267, 1270, ["input", "encoding", "layer"], "Method"], [1271, 1272, ["ESIM"], "Method"], [1274, 1278, ["Bi", "-", "LSTM", "layer"], "Method"], [1298, 1300, ["word", "embeddings"], "Method"], [1328, 1329, ["ESIM"], "Method"], [1331, 1334, ["local", "inference", "layer"], "Method"], [1387, 1389, ["local", "inference"], "Task"], [1391, 1395, ["Bi", "-", "LSTM", "layer"], "Method"], [1406, 1408, ["inference", "composition"], "Task"], [1411, 1412, ["summation"], "Method"], [1419, 1420, ["ESIM"], "Method"], [1424, 1428, ["max", "and", "average", "pooling"], "Method"], [1438, 1439, ["classifier"], "Method"], [1441, 1447, ["fully", "connected", "multi", "-", "layer", "perceptron"], "Method"], [1466, 1468, ["LSTM1", "layer"], "Method"], [1476, 1478, ["encoding", "layer"], "Method"], [1483, 1485, ["LSTM2", "layer"], "Method"], [1487, 1490, ["local", "inference", "layer"], "Method"], [1496, 1499, ["inference", "composition", "layer"], "Method"], [1502, 1504, ["aESIM", "model"], "Method"], [1511, 1516, ["attention", "boosted", "sequential", "inference", "model"], "Method"], [1518, 1520, ["aESIM", ")"], "Method"], [1522, 1523, ["ESIM"], "Method"], [1526, 1527, ["ESIM"], "Method"], [1531, 1532, ["aESIM"], "Method"], [1539, 1541, ["encoding", "layer"], "Method"], [1542, 1546, ["local", "inference", "modeling", "layer"], "Method"], [1547, 1549, ["decoding", "layer"], "Method"], [1550, 1552, ["classification", "layer"], "Method"], [1557, 1558, ["ESIM"], "Method"], [1559, 1560, ["aESIM"], "Method"], [1566, 1570, ["Bi", "-", "LSTM", "layers"], "Method"], [1571, 1572, ["LSTM1"], "Method"], [1573, 1574, ["LSTM2"], "Method"], [1576, 1577, ["ESIM"], "Method"], [1579, 1583, ["Bi", "-", "aLSTM", "layers"], "Method"], [1584, 1585, ["aESIM"], "Method"], [1604, 1605, ["ESIM"], "Method"], [1610, 1614, ["Bi", "-", "aLSTM", "layers"], "Method"], [1630, 1633, ["Bi", "-", "aLSTM"], "Method"], [1662, 1664, ["word", "embeddings"], "Method"], [1672, 1676, ["Bi", "-", "aLSTM", "layer"], "Method"], [1680, 1684, ["local", "inference", "modeling", "layer"], "Method"], [1687, 1691, ["Bi", "-", "aLSTM", "layer"], "Method"], [1695, 1698, ["forward", "LSTM", "layer"], "Method"], [1700, 1703, ["backward", "LSTM", "layer"], "Method"], [1721, 1724, ["Bi", "-", "aLSTM"], "Method"], [1730, 1732, ["additive", "operation"], "Method"], [1737, 1741, ["Bi", "-", "LSTM", "layer"], "Method"], [1742, 1745, ["Word", "attention", "layer"], "Method"], [1761, 1763, ["Attention", "mechanism"], "Method"], [1797, 1799, ["attention", "mechanism"], "Method"], [1813, 1817, ["one", "-", "layer", "MLP"], "Method"], [1831, 1833, ["SoftMax", "unit"], "Method"], [1876, 1880, ["Adaptive", "word", "direction", "layer"], "Method"], [1882, 1886, ["Bi", "-", "LSTM", "model"], "Method"], [1904, 1906, ["word", "representation"], "Method"], [1971, 1974, ["adaptive", "direction", "layer"], "Method"], [2044, 2051, ["word", "and", "orientation", "enhanced", "Bi", "-", "LSTM"], "Method"], [2053, 2056, ["Bi", "-", "aLSTM"], "Method"], [2072, 2074, ["ESIM", "model"], "Method"], [2078, 2082, ["Bi", "-", "LSTM", "layers"], "Method"], [2086, 2089, ["natural", "language", "inference"], "Task"], [2093, 2096, ["Bi", "-", "aLSTM"], "Method"], [2112, 2115, ["Bi", "-", "aLSTM"], "Method"], [2122, 2126, ["Bi", "-", "LSTM", "models"], "Method"], [2127, 2130, ["sentimental", "classification", "task"], "Task"], [2202, 2205, ["deep", "learning", "models"], "Method"], [2208, 2210, ["generalization", "ability"], "Metric"], [2225, 2226, ["SNLI"], "Material"], [2311, 2312, ["MultiNLI"], "Material"], [2359, 2364, ["cross", "-", "genre", "generation", "evaluation"], "Task"], [2365, 2366, ["Quora"], "Material"], [2412, 2414, ["aESIM", "model"], "Method"], [2422, 2424, ["Adam", "method"], "Method"], [2425, 2426, ["optimization"], "Task"], [2442, 2444, ["learning", "rate"], "Metric"], [2463, 2466, ["Bi", "-", "aLSTM"], "Method"], [2467, 2469, ["word", "embedding"], "Method"], [2479, 2482, ["rectified", "linear", "unit"], "Method"], [2487, 2489, ["convergence", "rate"], "Metric"], [2490, 2492, ["Dropout", "rate"], "Metric"], [2515, 2524, ["Out", "-", "of", "-", "vocabulary", "(", "OOV", ")", "words"], "Task"], [2528, 2530, ["Gaussian", "samples"], "Method"], [2547, 2548, ["aESIM"], "Method"], [2549, 2550, ["ESIM"], "Method"], [2567, 2568, ["SNIL"], "Material"], [2584, 2587, ["feature", "engineering", "method"], "Method"], [2594, 2599, ["sentence", "vector", "-", "based", "models"], "Method"], [2605, 2609, ["attention", "-", "based", "models"], "Method"], [2611, 2612, ["ESIM"], "Method"], [2614, 2615, ["aESIM"], "Method"], [2626, 2627, ["ESIM"], "Method"], [2628, 2629, ["aESIM"], "Method"], [2634, 2635, ["Keras"], "Method"], [2655, 2656, ["CBOW"], "Method"], [2657, 2660, ["Bi", "-", "LSTM"], "Method"], [2661, 2662, ["ESIM"], "Method"], [2664, 2665, ["aESIM"], "Method"], [2697, 2699, ["Quroa", "corpus"], "Material"], [2710, 2711, ["CNN"], "Method"], [2712, 2714, ["RNN", "models"], "Method"], [2719, 2720, ["accuracy"], "Metric"], [2721, 2722, ["ACC"], "Metric"], [2732, 2734, ["precision", "score"], "Metric"], [2741, 2742, ["accuracy"], "Metric"], [2757, 2759, ["aESIM", "model"], "Method"], [2771, 2773, ["ESIM", "model"], "Method"], [2779, 2780, ["accuracy"], "Metric"], [2785, 2786, ["MultiNLI"], "Material"], [2793, 2794, ["Quora"], "Material"], [2800, 2801, ["aESIM"], "Method"], [2812, 2814, ["ESIM", "model"], "Method"], [2817, 2819, ["Attention", "visualization"], "Task"], [2839, 2840, ["SNLI"], "Material"], [2964, 2968, ["Bi", "-", "LSTM", "layer"], "Method"], [2969, 2971, ["ESIM", "model"], "Method"], [2974, 2978, ["Bi", "-", "aLSTM", "layer"], "Method"], [2979, 2981, ["aESIM", "model"], "Method"], [2983, 2984, ["ESIM"], "Method"], [2985, 2986, ["aESIM"], "Method"], [3024, 3026, ["aESIM", "model"], "Method"], [3031, 3033, ["ESIM", "model"], "Method"], [3059, 3061, ["aESIM", "model"], "Method"], [3067, 3069, ["ESIM", "model"], "Method"], [3073, 3075, ["aESIM", "model"], "Method"], [3103, 3104, ["ESIM"], "Method"], [3105, 3106, ["aESIM"], "Method"], [3107, 3108, ["NLI"], "Task"], [3112, 3116, ["Bi", "-", "LSTM", "layer"], "Method"], [3124, 3126, ["aESIM", "model"], "Method"], [3135, 3137, ["aESIM", "model"], "Method"], [3141, 3143, ["ESIM", "model"], "Method"], [3152, 3154, ["attention", "mechanisms"], "Method"], [201, 202, ["SNLI"], "Material"], [1460, 1461, ["ESIM"], "Method"], [2101, 2104, ["natural", "language", "processing"], "Task"], [2167, 2168, ["SNLI"], "Material"], [2179, 2180, ["MultiNLI"], "Material"], [2184, 2185, ["Quora"], "Material"], [2232, 2233, ["SNLI"], "Material"], [2337, 2338, ["SNLI"], "Material"], [2367, 2368, ["Quora"], "Material"], [2666, 2667, ["MultiNLI"], "Material"], [2692, 2693, ["CNN"], "Method"], [2694, 2696, ["RNN", "models"], "Method"], [2763, 2764, ["SNLI"], "Material"], [3172, 3173, ["accuracy"], "Metric"]]} {"doc_id": "027d6c52be01b583b9a0d9eb8c9364c6b701c656", "text": ["document", ":", "All", "you", "need", "is", "a", "good", "init", "Layer", "-", "sequential", "unit", "-", "variance", "(", "LSUV", ")", "initialization", "–", "a", "simple", "method", "for", "weight", "initialization", "for", "deep", "net", "learning", "–", "is", "proposed", ".", "The", "method", "consists", "of", "the", "two", "steps", ".", "First", ",", "pre", "-", "initialize", "weights", "of", "each", "convolution", "or", "inner", "-", "product", "layer", "with", "orthonormal", "matrices", ".", "Second", ",", "proceed", "from", "the", "first", "to", "the", "final", "layer", ",", "normalizing", "the", "variance", "of", "the", "output", "of", "each", "layer", "to", "be", "equal", "to", "one", ".", "Experiment", "with", "different", "activation", "functions", "(", "maxout", ",", "ReLU", "-", "family", ",", "tanh", ")", "show", "that", "the", "proposed", "initialization", "leads", "to", "learning", "of", "very", "deep", "nets", "that", "(", "i", ")", "produces", "networks", "with", "test", "accuracy", "better", "or", "equal", "to", "standard", "methods", "and", "(", "ii", ")", "is", "at", "least", "as", "fast", "as", "the", "complex", "schemes", "proposed", "specifically", "for", "very", "deep", "nets", "such", "as", "FitNets", "(", ")", "and", "Highway", "(", ")", ".", "Performance", "is", "evaluated", "on", "GoogLeNet", ",", "CaffeNet", ",", "FitNets", "and", "Residual", "nets", "and", "the", "state", "-", "of", "-", "the", "-", "art", ",", "or", "very", "close", "to", "it", ",", "is", "achieved", "on", "the", "MNIST", ",", "CIFAR", "-", "10", "/", "100", "and", "ImageNet", "datasets", ".", "update", "section", ":", "Introduction", "Deep", "nets", "have", "demonstrated", "impressive", "results", "on", "a", "number", "of", "computer", "vision", "and", "natural", "language", "processing", "problems", ".", "At", "present", ",", "state", "-", "of", "-", "the", "-", "art", "results", "in", "image", "classification", "(", ")", "and", "speech", "recognition", "(", ")", ",", "etc", ".", ",", "have", "been", "achieved", "with", "very", "deep", "(", "layer", ")", "CNNs", ".", "Thin", "deep", "nets", "are", "of", "particular", "interest", ",", "since", "they", "are", "accurate", "and", "at", "the", "same", "inference", "-", "time", "efficient", "(", ")", ".", "One", "of", "the", "main", "obstacles", "preventing", "the", "wide", "adoption", "of", "very", "deep", "nets", "is", "the", "absence", "of", "a", "general", ",", "repeatable", "and", "efficient", "procedure", "for", "their", "end", "-", "to", "-", "end", "training", ".", "For", "example", ",", "VGGNet", "(", ")", "was", "optimized", "by", "a", "four", "stage", "procedure", "that", "started", "by", "training", "a", "network", "with", "moderate", "depth", ",", "adding", "progressively", "more", "layers", ".", "stated", "that", "deep", "and", "thin", "networks", "are", "very", "hard", "to", "train", "by", "backpropagation", "if", "deeper", "than", "five", "layers", ",", "especially", "with", "uniform", "initialization", ".", "On", "the", "other", "hand", ",", "showed", "that", "it", "is", "possible", "to", "train", "the", "VGGNet", "in", "a", "single", "optimization", "run", "if", "the", "network", "weights", "are", "initialized", "with", "a", "specific", "ReLU", "-", "aware", "initialization", ".", "The", "procedure", "generalizes", "to", "the", "ReLU", "non", "-", "linearity", "the", "idea", "of", "filter", "-", "size", "dependent", "initialization", ",", "introduced", "for", "the", "linear", "case", "by", "(", ")", ".", "Batch", "normalization", "(", ")", ",", "a", "technique", "that", "inserts", "layers", "into", "the", "the", "deep", "net", "that", "transform", "the", "output", "for", "the", "batch", "to", "be", "zero", "mean", "unit", "variance", ",", "has", "successfully", "facilitated", "training", "of", "the", "twenty", "-", "two", "layer", "GoogLeNet", "(", ")", ".", "However", ",", "batch", "normalization", "adds", "a", "30", "%", "computational", "overhead", "to", "each", "iteration", ".", "The", "main", "contribution", "of", "the", "paper", "is", "a", "proposal", "of", "a", "simple", "initialization", "procedure", "that", ",", "in", "connection", "with", "standard", "stochastic", "gradient", "descent", "(", "SGD", ")", ",", "leads", "to", "state", "-", "of", "-", "the", "-", "art", "thin", "and", "very", "deep", "neural", "nets", ".", "The", "result", "highlights", "the", "importance", "of", "initialization", "in", "very", "deep", "nets", ".", "We", "review", "the", "history", "of", "CNN", "initialization", "in", "Section", "[", "reference", "]", ",", "which", "is", "followed", "by", "a", "detailed", "description", "of", "the", "novel", "initialization", "method", "in", "Section", "[", "reference", "]", ".", "The", "method", "is", "experimentally", "validated", "in", "Section", "[", "reference", "]", ".", "section", ":", "Initialization", "in", "neural", "networks", "After", "the", "success", "of", "CNNs", "in", "IVSRC", "2012", "(", ")", ",", "initialization", "with", "Gaussian", "noise", "with", "mean", "equal", "to", "zero", "and", "standard", "deviation", "set", "to", "0.01", "and", "adding", "bias", "equal", "to", "one", "for", "some", "layers", "become", "very", "popular", ".", "But", ",", "as", "mentioned", "before", ",", "it", "is", "not", "possible", "to", "train", "very", "deep", "network", "from", "scratch", "with", "it", "(", ")", ".", "The", "problem", "is", "caused", "by", "the", "activation", "(", "and", "/", "or", ")", "gradient", "magnitude", "in", "final", "layers", "(", ")", ".", "If", "each", "layer", ",", "not", "properly", "initialized", ",", "scales", "input", "by", ",", "the", "final", "scale", "would", "be", ",", "where", "is", "a", "number", "of", "layers", ".", "Values", "of", "lead", "to", "extremely", "large", "values", "of", "output", "layers", ",", "leads", "to", "a", "diminishing", "signal", "and", "gradient", ".", "proposed", "a", "formula", "for", "estimating", "the", "standard", "deviation", "on", "the", "basis", "of", "the", "number", "of", "input", "and", "output", "channels", "of", "the", "layers", "under", "assumption", "of", "no", "non", "-", "linearity", "between", "layers", ".", "Despite", "invalidity", "of", "the", "assumption", ",", "Glorot", "initialization", "works", "well", "in", "many", "applications", ".", "extended", "this", "formula", "to", "the", "ReLU", "(", ")", "non", "-", "linearity", "and", "showed", "its", "superior", "performance", "for", "ReLU", "-", "based", "nets", ".", "Figure", "[", "reference", "]", "shows", "why", "scaling", "is", "important", ".", "Large", "weights", "lead", "to", "divergence", "via", "updates", "larger", "than", "the", "initial", "values", ",", "small", "initial", "weights", "do", "not", "allow", "the", "network", "to", "learn", "since", "the", "updates", "are", "of", "the", "order", "of", "0.0001", "per", "iteration", ".", "The", "optimal", "scaling", "for", "ReLU", "-", "net", "is", "around", "1.4", ",", "which", "is", "in", "line", "with", "the", "theoretically", "derived", "by", ".", "proposed", "the", "so", "called", "Random", "walk", "initialization", ",", "RWI", ",", "which", "keeps", "constant", "the", "log", "of", "the", "norms", "of", "the", "backpropagated", "errors", ".", "In", "our", "experiments", ",", "we", "have", "not", "been", "able", "to", "obtain", "good", "results", "with", "our", "implementation", "of", "RWI", ",", "that", "is", "why", "this", "method", "is", "not", "evaluated", "in", "experimental", "section", ".", "and", "take", "another", "approach", "to", "initialization", "and", "formulate", "training", "as", "mimicking", "teacher", "network", "predictions", "(", "so", "called", "knowledge", "distillation", ")", "and", "internal", "representations", "(", "so", "called", "Hints", "initialization", ")", "rather", "than", "minimizing", "the", "softmax", "loss", ".", "proposed", "a", "LSTM", "-", "inspired", "gating", "scheme", "to", "control", "information", "and", "gradient", "flow", "through", "the", "network", ".", "They", "trained", "a", "1000", "-", "layers", "MLP", "network", "on", "MNIST", ".", "Basically", ",", "this", "kind", "of", "networks", "implicitly", "learns", "the", "depth", "needed", "for", "the", "given", "task", ".", "Independently", ",", "showed", "that", "orthonormal", "matrix", "initialization", "works", "much", "better", "for", "linear", "networks", "than", "Gaussian", "noise", ",", "which", "is", "only", "approximate", "orthogonal", ".", "It", "also", "work", "for", "networks", "with", "non", "-", "linearities", ".", "The", "approach", "of", "layer", "-", "wise", "pre", "-", "training", "(", ")", "which", "is", "still", "useful", "for", "multi", "-", "layer", "-", "perceptron", ",", "is", "not", "popular", "for", "training", "discriminative", "convolution", "networks", ".", "section", ":", "Layer", "-", "sequential", "unit", "-", "variance", "initialization", "To", "the", "best", "of", "our", "knowledge", ",", "there", "have", "been", "no", "attempts", "to", "generalize", "formulas", "to", "non", "-", "linearities", "other", "than", "ReLU", ",", "such", "as", "tanh", ",", "maxout", ",", "etc", ".", "Also", ",", "the", "formula", "does", "not", "cover", "max", "-", "pooling", ",", "local", "normalization", "layers", "and", "other", "types", "of", "layers", "which", "influences", "activations", "variance", ".", "Instead", "of", "theoretical", "derivation", "for", "all", "possible", "layer", "types", ",", "or", "doing", "extensive", "parameters", "search", "as", "in", "Figure", "[", "reference", "]", ",", "we", "propose", "a", "data", "-", "driven", "weights", "initialization", ".", "We", "thus", "extend", "the", "orthonormal", "initialization", "to", "an", "iterative", "procedure", ",", "described", "in", "Algorithm", "[", "reference", "]", ".", "could", "be", "implemented", "in", "two", "steps", ".", "First", ",", "fill", "the", "weights", "with", "Gaussian", "noise", "with", "unit", "variance", ".", "Second", ",", "decompose", "them", "to", "orthonormal", "basis", "with", "QR", "or", "SVD", "-", "decomposition", "and", "replace", "weights", "with", "one", "of", "the", "components", ".", "The", "LSUV", "process", "then", "estimates", "output", "variance", "of", "each", "convolution", "and", "inner", "product", "layer", "and", "scales", "the", "weight", "to", "make", "variance", "equal", "to", "one", ".", "The", "influence", "of", "selected", "mini", "-", "batch", "size", "on", "estimated", "variance", "is", "negligible", "in", "wide", "margins", ",", "see", "Appendix", ".", "The", "proposed", "scheme", "can", "be", "viewed", "as", "an", "orthonormal", "initialization", "combined", "with", "batch", "normalization", "performed", "only", "on", "the", "first", "mini", "-", "batch", ".", "The", "similarity", "to", "batch", "normalization", "is", "the", "unit", "variance", "normalization", "procedure", ",", "while", "initial", "ortho", "-", "normalization", "of", "weights", "matrices", "efficiently", "de", "-", "correlates", "layer", "activations", ",", "which", "is", "not", "done", "in", ".", "Experiments", "show", "that", "such", "normalization", "is", "sufficient", "and", "computationally", "highly", "efficient", "in", "comparison", "with", "full", "batch", "normalization", ".", "[", "b", "]", "Layer", "-", "sequential", "unit", "-", "variance", "orthogonal", "initialization", ".", "L", "–", "convolution", "or", "full", "-", "connected", "layer", ",", "WL", "-", "its", "weights", ",", "BL", "-", "its", "output", "blob", ".", ",", "Tolvar", "-", "variance", "tolerance", ",", "Ti", "–", "current", "trial", ",", "Tmax", "–", "max", "number", "of", "trials", ".", "Pre", "-", "initialize", "network", "with", "orthonormal", "matrices", "as", "in", "[", "]", "layer", "L", "|Var", "(", "BL", ")-", "1.0|≥Tolvar", "and", "(", "Ti", "<", "Tmax", ")", "Forward", "pass", "with", "a", "mini", "-", "batch", "⁢Var", "(", "BL", ")", "=", "WL", "/", "⁢Var", "(", "BL", ")", "The", "LSUV", "algorithm", "is", "summarized", "in", "Algorithm", "[", "reference", "]", ".", "The", "single", "parameter", "influences", "convergence", "of", "the", "initialization", "procedure", ",", "not", "the", "properties", "of", "the", "trained", "network", ".", "Its", "value", "does", "not", "noticeably", "influence", "the", "performance", "in", "a", "broad", "range", "of", "0.01", "to", "0.1", ".", "Because", "of", "data", "variations", ",", "it", "is", "often", "not", "possible", "to", "normalize", "variance", "with", "the", "desired", "precision", ".", "To", "eliminate", "the", "possibility", "of", "an", "infinite", "loop", ",", "we", "restricted", "number", "of", "trials", "to", ".", "However", ",", "in", "experiments", "described", "in", "paper", ",", "the", "was", "never", "reached", ".", "The", "desired", "variance", "was", "achieved", "in", "1", "-", "5", "iterations", ".", "We", "tested", "a", "variant", "LSUV", "initialization", "which", "was", "normalizing", "input", "activations", "of", "the", "each", "layer", "instead", "of", "output", "ones", ".", "Normalizing", "the", "input", "or", "output", "is", "identical", "for", "standard", "feed", "-", "forward", "nets", ",", "but", "normalizing", "input", "is", "much", "more", "complicated", "for", "networks", "with", "maxout", "(", ")", "or", "for", "networks", "like", "GoogLeNet", "(", ")", "which", "use", "the", "output", "of", "multiple", "layers", "as", "input", ".", "Input", "normalization", "brought", "no", "improvement", "of", "results", "when", "tested", "against", "the", "LSUV", "Algorithm", "[", "reference", "]", ",", "LSUV", "was", "also", "tested", "with", "pre", "-", "initialization", "of", "weights", "with", "Gaussian", "noise", "instead", "of", "orthonormal", "matrices", ".", "The", "Gaussian", "initialization", "led", "to", "small", ",", "but", "consistent", ",", "decrease", "in", "performance", ".", "section", ":", "Experimental", "validation", "Here", "we", "show", "that", "very", "deep", "and", "thin", "nets", "could", "be", "trained", "in", "a", "single", "stage", ".", "Network", "architectures", "are", "exactly", "as", "proposed", "by", ".", "The", "architectures", "are", "presented", "in", "Table", "[", "reference", "]", ".", "subsection", ":", "MNIST", "First", ",", "as", "a", "”", "sanity", "check", "”", ",", "we", "performed", "an", "experiment", "on", "the", "MNIST", "dataset", "(", ")", ".", "It", "consists", "of", "60", ",", "000", "28x28", "grayscale", "images", "of", "handwritten", "digits", "0", "to", "9", ".", "We", "selected", "the", "FitNet", "-", "MNIST", "architecture", "(", "see", "Table", "[", "reference", "]", ")", "of", "and", "trained", "it", "with", "the", "proposed", "initialization", "strategy", ",", "without", "data", "augmentation", ".", "Recognition", "results", "are", "shown", "in", "Table", "[", "reference", "]", ",", "right", "block", ".", "LSUV", "outperforms", "orthonormal", "initialization", "and", "both", "LSUV", "and", "orthonormal", "outperform", "Hints", "initialization", ".", "The", "error", "rates", "of", "the", "Deeply", "-", "Supervised", "Nets", "(", "DSN", ",", ")", "and", "maxout", "networks", ",", "the", "current", "state", "-", "of", "-", "art", ",", "are", "provided", "for", "reference", ".", "Since", "the", "widely", "cited", "DSN", "error", "rate", "of", "0.39", "%", ",", "the", "state", "-", "of", "-", "the", "-", "art", "(", "until", "recently", ")", "was", "obtained", "after", "replacing", "the", "softmax", "classifier", "with", "SVM", ",", "we", "do", "the", "same", "and", "also", "observe", "improved", "results", "(", "line", "FitNet", "-", "LSUV", "-", "SVM", "in", "Table", "[", "reference", "]", ")", ".", "subsection", ":", "CIFAR", "-", "10", "/", "100", "We", "validated", "the", "proposed", "initialization", "LSUV", "strategy", "on", "the", "CIFAR", "-", "10", "/", "100", "(", ")", "dataset", ".", "It", "contains", "60", ",", "000", "32x32", "RGB", "images", ",", "which", "are", "divided", "into", "10", "and", "100", "classes", ",", "respectively", ".", "The", "FitNets", "are", "trained", "with", "the", "stochastic", "gradient", "descent", "with", "momentum", "set", "to", "0.9", ",", "the", "initial", "learning", "rate", "set", "to", "0.01", "and", "reduced", "by", "a", "factor", "of", "10", "after", "the", "100th", ",", "150th", "and", "200th", "epoch", ",", "finishing", "at", "230th", "epoch", ".", "and", "trained", "their", "networks", "for", "500", "epochs", ".", "Of", "course", ",", "training", "time", "is", "a", "trade", "-", "off", "dependent", "on", "the", "desired", "accuracy", ";", "one", "could", "train", "a", "slightly", "less", "accurate", "network", "much", "faster", ".", "Like", "in", "the", "MNIST", "experiment", ",", "LSUV", "and", "orthonormal", "initialized", "nets", "outperformed", "Hints", "-", "trained", "Fitnets", ",", "leading", "to", "the", "new", "state", "-", "of", "-", "art", "when", "using", "commonly", "used", "augmentation", "–", "mirroring", "and", "random", "shifts", ".", "The", "gain", "on", "the", "fine", "-", "grained", "CIFAR", "-", "100", "is", "much", "larger", "than", "on", "CIFAR", "-", "10", ".", "Also", ",", "note", "that", "FitNets", "with", "LSUV", "initialization", "outperform", "even", "much", "larger", "networks", "like", "Large", "-", "All", "-", "CNN", "and", "Fractional", "Max", "-", "pooling", "trained", "with", "affine", "and", "color", "dataset", "augmentation", "on", "CIFAR", "-", "100", ".", "The", "results", "of", "LSUV", "are", "virtually", "identical", "to", "the", "orthonormal", "initialization", ".", "section", ":", "Analysis", "of", "empirical", "results", "subsection", ":", "Initialization", "strategies", "and", "non", "-", "linearities", "For", "the", "FitNet", "-", "1", "architecture", ",", "we", "have", "not", "experienced", "any", "difficulties", "training", "the", "network", "with", "any", "of", "the", "activation", "functions", "(", "ReLU", ",", "maxout", ",", "tanh", ")", ",", "optimizers", "(", "SGD", ",", "RMSProp", ")", "or", "initialization", "(", "Xavier", ",", "MSRA", ",", "Ortho", ",", "LSUV", ")", ",", "unlike", "the", "uniform", "initialization", "used", "in", ".", "The", "most", "probable", "cause", "is", "that", "CNNs", "tolerate", "a", "wide", "variety", "of", "mediocre", "initialization", ",", "only", "the", "learning", "time", "increases", ".", "The", "differences", "in", "the", "final", "accuracy", "between", "the", "different", "initialization", "methods", "for", "the", "FitNet", "-", "1", "architecture", "is", "rather", "small", "and", "are", "therefore", "not", "presented", "here", ".", "The", "FitNet", "-", "4", "architecture", "is", "much", "more", "difficult", "to", "optimize", "and", "thus", "we", "focus", "on", "it", "in", "the", "experiments", "presented", "in", "this", "section", ".", "We", "have", "explored", "the", "initializations", "with", "different", "activation", "functions", "in", "very", "deep", "networks", ".", "More", "specifically", ",", "ReLU", ",", "hyperbolic", "tangent", ",", "sigmoid", ",", "maxout", "and", "the", "VLReLU", "–", "very", "leaky", "ReLU", "(", ")", "–", "a", "variant", "of", "leaky", "ReLU", "(", ",", "with", "a", "large", "value", "of", "the", "negative", "slope", "0.333", ",", "instead", "of", "the", "originally", "proposed", "0.01", ")", "which", "is", "popular", "in", "Kaggle", "competitions", ",", ")", ".", "Testing", "was", "performed", "on", "CIFAR", "-", "10", "and", "results", "are", "in", "Table", "[", "reference", "]", "and", "Figure", "[", "reference", "]", ".", "Performance", "of", "orthonormal", "-", "based", "methods", "is", "superior", "to", "the", "scaled", "Gaussian", "-", "noise", "approaches", "for", "all", "tested", "types", "of", "activation", "functions", ",", "except", "tanh", ".", "Proposed", "LSUV", "strategy", "outperforms", "orthonormal", "initialization", "by", "smaller", "margin", ",", "but", "still", "consistently", "(", "see", "Table", "[", "reference", "]", ")", ".", "All", "the", "methods", "failed", "to", "train", "sigmoid", "-", "based", "very", "deep", "network", ".", "Figure", "[", "reference", "]", "shows", "that", "LSUV", "method", "not", "only", "leads", "to", "better", "generalization", "error", ",", "but", "also", "converges", "faster", "for", "all", "tested", "activation", "functions", ",", "except", "tanh", ".", "We", "have", "also", "tested", "how", "the", "different", "initializations", "work", "”", "out", "-", "of", "-", "the", "-", "box", "”", "with", "the", "Residual", "net", "training", ";", "a", "residual", "net", "won", "the", "ILSVRC", "-", "2015", "challenge", ".", "The", "original", "paper", "proposed", "different", "implementations", "of", "residual", "learning", ".", "We", "adopted", "the", "simplest", "one", ",", "showed", "in", "Table", "[", "reference", "]", ",", "FitResNet", "-", "4", ".", "The", "output", "of", "each", "even", "convolutional", "layer", "is", "summed", "with", "the", "output", "of", "the", "previous", "non", "-", "linearity", "layer", "and", "then", "fed", "into", "the", "next", "non", "-", "linearity", ".", "Results", "are", "shown", "in", "Table", "[", "reference", "]", ".", "LSUV", "is", "the", "only", "initialization", "algorithm", "which", "leads", "nets", "to", "convergence", "with", "all", "tested", "non", "-", "linearities", "without", "any", "additional", "tuning", ",", "except", ",", "again", ",", "sigmoid", ".", "It", "is", "worth", "nothing", "that", "the", "residual", "training", "improves", "results", "for", "ReLU", "and", "maxout", ",", "but", "does", "not", "help", "tanh", "-", "based", "network", ".", "subsection", ":", "Comparison", "to", "batch", "normalization", "(", "BN", ")", "LSUV", "procedure", "could", "be", "viewed", "as", "batch", "normalization", "of", "layer", "output", "done", "only", "before", "the", "start", "of", "training", ".", "Therefore", ",", "it", "is", "natural", "to", "compare", "LSUV", "against", "a", "batch", "-", "normalized", "network", ",", "initialized", "with", "the", "standard", "method", ".", "subsubsection", ":", "Where", "to", "put", "BN", "–", "before", "or", "after", "non", "-", "linearity", "?", "It", "is", "not", "clear", "from", "the", "paper", "where", "to", "put", "the", "batch", "-", "normalization", "layer", "–", "before", "input", "of", "each", "layer", "as", "stated", "in", "Section", "3.1", ",", "or", "before", "non", "-", "linearity", ",", "as", "stated", "in", "section", "3.2", ",", "so", "we", "have", "conducted", "an", "experiment", "with", "FitNet4", "on", "CIFAR", "-", "10", "to", "clarify", "this", ".", "Results", "are", "shown", "in", "Table", "[", "reference", "]", ".", "Exact", "numbers", "vary", "from", "run", "to", "run", ",", "but", "in", "the", "most", "cases", ",", "batch", "normalization", "put", "after", "non", "-", "linearity", "performs", "better", ".", "In", "the", "next", "experiment", "we", "compare", "BN", "-", "FitNet4", ",", "initialized", "with", "Xavier", "and", "LSUV", "-", "initialized", "FitNet4", ".", "Batch", "-", "normalization", "reduces", "training", "time", "in", "terms", "of", "needed", "number", "of", "iterations", ",", "but", "each", "iteration", "becomes", "slower", "because", "of", "extra", "computations", ".", "The", "accuracy", "versus", "wall", "-", "clock", "-", "time", "graphs", "are", "shown", "in", "Figure", "[", "reference", "]", ".", "LSUV", "-", "initialized", "network", "is", "as", "good", "as", "batch", "-", "normalized", "one", ".", "However", ",", "we", "are", "not", "claiming", "that", "batch", "normalization", "can", "always", "be", "replaced", "by", "proper", "initialization", ",", "especially", "in", "large", "datasets", "like", "ImageNet", ".", "subsection", ":", "Imagenet", "training", "We", "trained", "CaffeNet", "(", ")", "and", "GoogLeNet", "(", ")", "on", "the", "ImageNet", "-", "1000", "dataset", "(", ")", "with", "the", "original", "initialization", "and", "LSUV", ".", "CaffeNet", "is", "a", "variant", "of", "AlexNet", "with", "the", "nearly", "identical", "performance", ",", "where", "the", "order", "of", "pooling", "and", "normalization", "layers", "is", "switched", "to", "reduce", "the", "memory", "footprint", ".", "LSUV", "initialization", "reduces", "the", "starting", "flat", "-", "loss", "time", "from", "0.5", "epochs", "to", "0.05", "for", "CaffeNet", ",", "and", "starts", "to", "converge", "faster", ",", "but", "it", "is", "overtaken", "by", "a", "standard", "CaffeNet", "at", "the", "30", "-", "th", "epoch", "(", "see", "Figure", "[", "reference", "]", ")", "and", "its", "final", "precision", "is", "1.3", "%", "lower", ".", "We", "have", "no", "explanation", "for", "this", "empirical", "phenomenon", ".", "On", "the", "contrary", ",", "the", "LSUV", "-", "initialized", "GoogLeNet", "learns", "faster", "than", "hen", "then", "original", "one", "and", "shows", "better", "test", "accuracy", "all", "the", "time", "–", "see", "Figure", "[", "reference", "]", ".", "The", "final", "accuracy", "is", "0.680", "vs.", "0.672", "respectively", ".", "subsection", ":", "Timings", "A", "significant", "part", "of", "LSUV", "initialization", "is", "SVD", "-", "decomposition", "of", "the", "weight", "matrices", ",", "e.g.", "for", "the", "fc6", "layer", "of", "CaffeNet", ",", "an", "SVD", "of", "a", "9216x4096", "matrix", "is", "required", ".", "The", "computational", "overhead", "on", "top", "of", "generating", "almost", "instantly", "the", "scaled", "random", "Gaussian", "samples", "is", "shown", "in", "Table", "[", "reference", "]", ".", "In", "the", "slowest", "case", "–", "CaffeNet", "–", "LSUV", "initialization", "takes", "3.5", "minutes", ",", "which", "is", "negligible", "in", "comparison", "the", "training", "time", ".", "section", ":", "Conclusions", "LSUV", ",", "layer", "sequential", "uniform", "variance", ",", "a", "simple", "strategy", "for", "weight", "initialization", "for", "deep", "net", "learning", ",", "is", "proposed", ".", "We", "have", "showed", "that", "the", "LSUV", "initialization", ",", "described", "fully", "in", "six", "lines", "of", "pseudocode", ",", "is", "as", "good", "as", "complex", "learning", "schemes", "which", "need", ",", "for", "instance", ",", "auxiliary", "nets", ".", "The", "LSUV", "initialization", "allows", "learning", "of", "very", "deep", "nets", "via", "standard", "SGD", ",", "is", "fast", ",", "and", "leads", "to", "(", "near", ")", "state", "-", "of", "-", "the", "-", "art", "results", "on", "MNIST", ",", "CIFAR", ",", "ImageNet", "datasets", ",", "outperforming", "the", "sophisticated", "systems", "designed", "specifically", "for", "very", "deep", "nets", "such", "as", "FitNets", "(", ")", "and", "Highway", "(", ")", ".", "The", "proposed", "initialization", "works", "well", "with", "different", "activation", "functions", ".", "Our", "experiments", "confirm", "the", "finding", "of", "that", "very", "thin", ",", "thus", "fast", "and", "low", "in", "parameters", ",", "but", "deep", "networks", "obtain", "comparable", "or", "even", "better", "performance", "than", "wider", ",", "but", "shallower", "ones", ".", "subsubsection", ":", "Acknowledgments", "The", "authors", "were", "supported", "by", "The", "Czech", "Science", "Foundation", "Project", "GACR", "P103", "/", "12", "/", "G084", "and", "CTU", "student", "grant", "SGS15", "/", "155", "/", "OHK3", "/", "2T", "/", "13", ".", "bibliography", ":", "References", "appendix", ":", "Technical", "details", "subsection", ":", "Influence", "of", "mini", "-", "batch", "size", "to", "LSUV", "initialization", "We", "have", "selected", "tanh", "activation", "as", "one", ",", "where", "LSUV", "initialization", "shows", "the", "worst", "performance", "and", "tested", "the", "influence", "of", "mini", "-", "batch", "size", "to", "training", "process", ".", "Note", ",", "that", "training", "mini", "-", "batch", "is", "the", "same", "for", "all", "initializations", ",", "the", "only", "difference", "is", "mini", "-", "batch", "used", "for", "variance", "estimation", ".", "One", "can", "see", "from", "Table", "[", "reference", "]", "that", "there", "is", "no", "difference", "between", "small", "or", "large", "mini", "-", "batch", ",", "except", "extreme", "cases", ",", "where", "only", "two", "sample", "are", "used", ".", "subsection", ":", "LSUV", "weight", "standard", "deviations", "in", "different", "networks", "Tables", "[", "reference", "]", "and", "[", "reference", "]", "show", "the", "standard", "deviations", "of", "the", "filter", "weights", ",", "found", "by", "the", "LSUV", "procedure", "and", "by", "other", "initialization", "schemes", ".", "subsection", ":", "Gradients", "To", "check", "how", "the", "activation", "variance", "normalization", "influences", "the", "variance", "of", "the", "gradient", ",", "we", "measure", "the", "average", "variance", "of", "the", "gradient", "at", "all", "layers", "after", "10", "mini", "-", "batches", ".", "The", "variance", "is", "close", "to", "for", "all", "convolutional", "layers", ".", "It", "is", "much", "more", "stable", "than", "for", "the", "reference", "methods", ",", "except", "MSRA", ";", "see", "Table", "[", "reference", "]", "."], "entity": [[24, 26, ["weight", "initialization"], "Task"], [27, 30, ["deep", "net", "learning"], "Task"], [50, 56, ["convolution", "or", "inner", "-", "product", "layer"], "Method"], [57, 59, ["orthonormal", "matrices"], "Method"], [94, 97, ["ReLU", "-", "family"], "Method"], [98, 99, ["tanh"], "Method"], [107, 112, ["learning", "of", "very", "deep", "nets"], "Task"], [119, 121, ["test", "accuracy"], "Metric"], [143, 146, ["very", "deep", "nets"], "Task"], [148, 149, ["FitNets"], "Method"], [152, 153, ["Highway"], "Method"], [160, 161, ["GoogLeNet"], "Method"], [162, 163, ["CaffeNet"], "Method"], [164, 165, ["FitNets"], "Method"], [166, 168, ["Residual", "nets"], "Method"], [188, 189, ["MNIST"], "Material"], [190, 193, ["CIFAR", "-", "10"], "Material"], [194, 195, ["100"], "Material"], [203, 205, ["Deep", "nets"], "Method"], [213, 220, ["computer", "vision", "and", "natural", "language", "processing", "problems"], "Task"], [233, 235, ["image", "classification"], "Task"], [238, 240, ["speech", "recognition"], "Task"], [251, 256, ["deep", "(", "layer", ")", "CNNs"], "Method"], [257, 260, ["Thin", "deep", "nets"], "Method"], [291, 293, ["deep", "nets"], "Method"], [306, 312, ["end", "-", "to", "-", "end", "training"], "Task"], [316, 319, ["VGGNet", "(", ")"], "Method"], [324, 326, ["stage", "procedure"], "Method"], [343, 347, ["deep", "and", "thin", "networks"], "Method"], [353, 354, ["backpropagation"], "Method"], [378, 379, ["VGGNet"], "Method"], [393, 397, ["ReLU", "-", "aware", "initialization"], "Method"], [410, 415, ["filter", "-", "size", "dependent", "initialization"], "Method"], [419, 421, ["linear", "case"], "Task"], [425, 427, ["Batch", "normalization"], "Method"], [438, 440, ["deep", "net"], "Method"], [462, 465, ["two", "layer", "GoogLeNet"], "Method"], [470, 472, ["batch", "normalization"], "Method"], [476, 478, ["computational", "overhead"], "Metric"], [494, 496, ["initialization", "procedure"], "Method"], [502, 505, ["stochastic", "gradient", "descent"], "Method"], [506, 508, ["SGD", ")"], "Method"], [521, 524, ["deep", "neural", "nets"], "Method"], [531, 532, ["initialization"], "Task"], [533, 536, ["very", "deep", "nets"], "Task"], [542, 544, ["CNN", "initialization"], "Task"], [560, 562, ["initialization", "method"], "Method"], [581, 582, ["Initialization"], "Task"], [583, 585, ["neural", "networks"], "Method"], [589, 590, ["CNNs"], "Method"], [596, 597, ["initialization"], "Method"], [637, 639, ["deep", "network"], "Method"], [716, 718, ["standard", "deviation"], "Metric"], [748, 750, ["Glorot", "initialization"], "Method"], [773, 777, ["ReLU", "-", "based", "nets"], "Method"], [825, 826, ["scaling"], "Method"], [827, 830, ["ReLU", "-", "net"], "Task"], [848, 851, ["Random", "walk", "initialization"], "Method"], [852, 853, ["RWI"], "Method"], [884, 885, ["RWI"], "Method"], [903, 904, ["initialization"], "Task"], [906, 907, ["training"], "Task"], [915, 917, ["knowledge", "distillation"], "Task"], [924, 926, ["Hints", "initialization"], "Method"], [936, 941, ["LSTM", "-", "inspired", "gating", "scheme"], "Method"], [954, 959, ["1000", "-", "layers", "MLP", "network"], "Method"], [960, 961, ["MNIST"], "Material"], [982, 985, ["orthonormal", "matrix", "initialization"], "Method"], [989, 991, ["linear", "networks"], "Task"], [1007, 1010, ["non", "-", "linearities"], "Task"], [1014, 1022, ["layer", "-", "wise", "pre", "-", "training", "(", ")"], "Method"], [1027, 1032, ["multi", "-", "layer", "-", "perceptron"], "Task"], [1038, 1041, ["discriminative", "convolution", "networks"], "Method"], [1044, 1051, ["Layer", "-", "sequential", "unit", "-", "variance", "initialization"], "Method"], [1089, 1092, ["max", "-", "pooling"], "Method"], [1131, 1136, ["data", "-", "driven", "weights", "initialization"], "Method"], [1141, 1143, ["orthonormal", "initialization"], "Method"], [1145, 1147, ["iterative", "procedure"], "Method"], [1182, 1187, ["QR", "or", "SVD", "-", "decomposition"], "Method"], [1205, 1210, ["convolution", "and", "inner", "product", "layer"], "Method"], [1230, 1232, ["estimated", "variance"], "Metric"], [1249, 1251, ["orthonormal", "initialization"], "Method"], [1253, 1255, ["batch", "normalization"], "Method"], [1267, 1269, ["batch", "normalization"], "Method"], [1271, 1275, ["unit", "variance", "normalization", "procedure"], "Method"], [1278, 1284, ["ortho", "-", "normalization", "of", "weights", "matrices"], "Method"], [1301, 1302, ["normalization"], "Method"], [1311, 1314, ["full", "batch", "normalization"], "Method"], [1320, 1326, ["sequential", "unit", "-", "variance", "orthogonal", "initialization"], "Method"], [1329, 1335, ["convolution", "or", "full", "-", "connected", "layer"], "Method"], [1389, 1391, ["Forward", "pass"], "Method"], [1425, 1427, ["initialization", "procedure"], "Method"], [1433, 1435, ["trained", "network"], "Method"], [1469, 1470, ["precision"], "Metric"], [1540, 1544, ["feed", "-", "forward", "nets"], "Method"], [1555, 1557, ["maxout", "("], "Method"], [1562, 1564, ["GoogLeNet", "("], "Method"], [1575, 1577, ["Input", "normalization"], "Task"], [1586, 1588, ["LSUV", "Algorithm"], "Method"], [1592, 1593, ["LSUV"], "Method"], [1611, 1613, ["Gaussian", "initialization"], "Method"], [1633, 1637, ["deep", "and", "thin", "nets"], "Method"], [1645, 1647, ["Network", "architectures"], "Method"], [1665, 1666, ["MNIST"], "Material"], [1723, 1725, ["initialization", "strategy"], "Method"], [1727, 1729, ["data", "augmentation"], "Method"], [1730, 1731, ["Recognition"], "Task"], [1743, 1744, ["LSUV"], "Method"], [1745, 1747, ["orthonormal", "initialization"], "Method"], [1749, 1750, ["LSUV"], "Method"], [1751, 1752, ["orthonormal"], "Method"], [1753, 1755, ["Hints", "initialization"], "Method"], [1757, 1759, ["error", "rates"], "Metric"], [1761, 1769, ["Deeply", "-", "Supervised", "Nets", "(", "DSN", ",", ")"], "Method"], [1770, 1772, ["maxout", "networks"], "Method"], [1790, 1793, ["DSN", "error", "rate"], "Metric"], [1814, 1816, ["softmax", "classifier"], "Method"], [1817, 1818, ["SVM"], "Method"], [1844, 1847, ["CIFAR", "-", "10"], "Material"], [1848, 1849, ["100"], "Material"], [1858, 1861, ["CIFAR", "-", "10"], "Material"], [1862, 1863, ["100"], "Material"], [1888, 1889, ["FitNets"], "Method"], [1893, 1896, ["stochastic", "gradient", "descent"], "Method"], [1904, 1906, ["learning", "rate"], "Metric"], [1941, 1943, ["training", "time"], "Metric"], [1952, 1953, ["accuracy"], "Metric"], [1980, 1981, ["Fitnets"], "Method"], [1995, 1996, ["augmentation"], "Method"], [1997, 1998, ["mirroring"], "Method"], [1999, 2001, ["random", "shifts"], "Method"], [2006, 2012, ["fine", "-", "grained", "CIFAR", "-", "100"], "Material"], [2017, 2020, ["CIFAR", "-", "10"], "Material"], [2025, 2026, ["FitNets"], "Method"], [2035, 2040, ["Large", "-", "All", "-", "CNN"], "Method"], [2041, 2045, ["Fractional", "Max", "-", "pooling"], "Method"], [2053, 2056, ["CIFAR", "-", "100"], "Material"], [2060, 2061, ["LSUV"], "Method"], [2066, 2068, ["orthonormal", "initialization"], "Method"], [2077, 2079, ["Initialization", "strategies"], "Method"], [2085, 2089, ["FitNet", "-", "1", "architecture"], "Method"], [2113, 2114, ["optimizers"], "Method"], [2115, 2116, ["SGD"], "Method"], [2117, 2118, ["RMSProp"], "Method"], [2120, 2121, ["initialization"], "Method"], [2128, 2129, ["LSUV"], "Method"], [2133, 2135, ["uniform", "initialization"], "Method"], [2144, 2145, ["CNNs"], "Method"], [2155, 2157, ["learning", "time"], "Metric"], [2164, 2165, ["accuracy"], "Metric"], [2168, 2170, ["initialization", "methods"], "Method"], [2172, 2176, ["FitNet", "-", "1", "architecture"], "Method"], [2187, 2191, ["FitNet", "-", "4", "architecture"], "Method"], [2222, 2224, ["deep", "networks"], "Method"], [2228, 2229, ["ReLU"], "Method"], [2233, 2234, ["sigmoid"], "Method"], [2235, 2236, ["maxout"], "Method"], [2238, 2239, ["VLReLU"], "Method"], [2249, 2251, ["leaky", "ReLU"], "Method"], [2283, 2286, ["CIFAR", "-", "10"], "Material"], [2302, 2306, ["orthonormal", "-", "based", "methods"], "Method"], [2310, 2315, ["scaled", "Gaussian", "-", "noise", "approaches"], "Method"], [2324, 2325, ["tanh"], "Method"], [2330, 2332, ["orthonormal", "initialization"], "Method"], [2353, 2359, ["sigmoid", "-", "based", "very", "deep", "network"], "Method"], [2373, 2375, ["generalization", "error"], "Metric"], [2387, 2388, ["tanh"], "Method"], [2409, 2412, ["Residual", "net", "training"], "Method"], [2414, 2416, ["residual", "net"], "Method"], [2430, 2432, ["residual", "learning"], "Method"], [2446, 2449, ["FitResNet", "-", "4"], "Method"], [2454, 2457, ["even", "convolutional", "layer"], "Method"], [2465, 2469, ["non", "-", "linearity", "layer"], "Method"], [2475, 2478, ["non", "-", "linearity"], "Method"], [2488, 2489, ["LSUV"], "Method"], [2492, 2494, ["initialization", "algorithm"], "Method"], [2522, 2524, ["residual", "training"], "Method"], [2527, 2528, ["ReLU"], "Task"], [2529, 2530, ["maxout"], "Task"], [2535, 2539, ["tanh", "-", "based", "network"], "Method"], [2544, 2546, ["batch", "normalization"], "Method"], [2547, 2548, ["BN"], "Method"], [2575, 2576, ["LSUV"], "Method"], [2578, 2582, ["batch", "-", "normalized", "network"], "Method"], [2614, 2618, ["batch", "-", "normalization", "layer"], "Method"], [2649, 2650, ["FitNet4"], "Method"], [2651, 2654, ["CIFAR", "-", "10"], "Material"], [2681, 2683, ["batch", "normalization"], "Method"], [2685, 2688, ["non", "-", "linearity"], "Method"], [2703, 2704, ["Xavier"], "Method"], [2705, 2708, ["LSUV", "-", "initialized"], "Method"], [2708, 2709, ["FitNet4"], "Method"], [2710, 2713, ["Batch", "-", "normalization"], "Method"], [2714, 2716, ["training", "time"], "Metric"], [2735, 2736, ["accuracy"], "Metric"], [2737, 2743, ["wall", "-", "clock", "-", "time", "graphs"], "Metric"], [2759, 2763, ["batch", "-", "normalized", "one"], "Method"], [2771, 2773, ["batch", "normalization"], "Method"], [2779, 2780, ["initialization"], "Method"], [2790, 2792, ["Imagenet", "training"], "Task"], [2794, 2795, ["CaffeNet"], "Method"], [2798, 2800, ["GoogLeNet", "("], "Method"], [2812, 2813, ["initialization"], "Method"], [2814, 2815, ["LSUV"], "Method"], [2816, 2817, ["CaffeNet"], "Method"], [2821, 2822, ["AlexNet"], "Method"], [2849, 2853, ["flat", "-", "loss", "time"], "Metric"], [2859, 2860, ["CaffeNet"], "Method"], [2874, 2875, ["CaffeNet"], "Method"], [2891, 2892, ["precision"], "Metric"], [2926, 2927, ["accuracy"], "Metric"], [2939, 2940, ["accuracy"], "Metric"], [2956, 2963, ["SVD", "-", "decomposition", "of", "the", "weight", "matrices"], "Method"], [2967, 2969, ["fc6", "layer"], "Method"], [2970, 2971, ["CaffeNet"], "Method"], [2973, 2974, ["SVD"], "Method"], [2982, 2984, ["computational", "overhead"], "Metric"], [3022, 3024, ["training", "time"], "Metric"], [3028, 3029, ["LSUV"], "Method"], [3030, 3034, ["layer", "sequential", "uniform", "variance"], "Method"], [3039, 3041, ["weight", "initialization"], "Task"], [3042, 3045, ["deep", "net", "learning"], "Task"], [3070, 3072, ["learning", "schemes"], "Method"], [3085, 3090, ["learning", "of", "very", "deep", "nets"], "Task"], [3092, 3093, ["SGD"], "Method"], [3112, 3113, ["MNIST"], "Material"], [3114, 3115, ["CIFAR"], "Material"], [3126, 3129, ["very", "deep", "nets"], "Task"], [3131, 3132, ["FitNets"], "Method"], [3135, 3136, ["Highway"], "Method"], [3167, 3169, ["deep", "networks"], "Method"], [3236, 3238, ["tanh", "activation"], "Method"], [3284, 3286, ["variance", "estimation"], "Task"], [3353, 3355, ["initialization", "schemes"], "Method"], [3363, 3366, ["activation", "variance", "normalization"], "Method"], [3376, 3378, ["average", "variance"], "Metric"], [3397, 3399, ["convolutional", "layers"], "Method"], [3412, 3413, ["MSRA"], "Method"], [16, 17, ["LSUV"], "Method"], [1197, 1198, ["LSUV"], "Method"], [1408, 1409, ["LSUV"], "Method"], [1515, 1516, ["LSUV"], "Method"], [1681, 1682, ["MNIST"], "Material"], [1707, 1708, ["MNIST"], "Material"], [1832, 1833, ["LSUV"], "Method"], [1834, 1835, ["SVM"], "Method"], [1854, 1855, ["LSUV"], "Method"], [1882, 1883, ["100"], "Material"], [1968, 1969, ["MNIST"], "Material"], [1971, 1972, ["LSUV"], "Method"], [2027, 2028, ["LSUV"], "Method"], [2122, 2123, ["Xavier"], "Method"], [2327, 2328, ["LSUV"], "Method"], [2366, 2367, ["LSUV"], "Method"], [2549, 2550, ["LSUV"], "Method"], [2699, 2700, ["FitNet4"], "Method"], [2751, 2752, ["LSUV"], "Method"], [2844, 2845, ["LSUV"], "Method"], [2911, 2912, ["LSUV"], "Method"], [2953, 2954, ["LSUV"], "Method"], [3010, 3011, ["LSUV"], "Method"], [3054, 3055, ["LSUV"], "Method"], [3082, 3083, ["LSUV"], "Method"], [3231, 3232, ["LSUV"], "Method"], [3242, 3243, ["LSUV"], "Method"], [3321, 3322, ["LSUV"], "Method"], [3348, 3349, ["LSUV"], "Method"]]} {"doc_id": "0496691c6d54fd746c111d1004349acf7654145e", "text": ["CALIBRATING", "ENERGY", "-", "BASED", "GENERATIVE", "ADVER", "-", "SARIAL", "NETWORKS", "section", ":", "ABSTRACT", "In", "this", "paper", "we", "propose", "equipping", "Generative", "Adversarial", "Networks", "with", "the", "ability", "to", "produce", "direct", "energy", "estimates", "for", "samples", ".", "Specifically", ",", "we", "develop", "a", "flexible", "adversarial", "training", "framework", ",", "and", "prove", "this", "framework", "not", "only", "ensures", "the", "generator", "converges", "to", "the", "true", "data", "distribution", ",", "but", "also", "enables", "the", "discriminator", "to", "retain", "the", "density", "information", "at", "the", "global", "optimum", ".", "We", "derive", "the", "analytic", "form", "of", "the", "induced", "solution", ",", "and", "analyze", "its", "properties", ".", "In", "order", "to", "make", "the", "proposed", "framework", "trainable", "in", "practice", ",", "we", "introduce", "two", "effective", "approximation", "techniques", ".", "Empirically", ",", "the", "experiment", "results", "closely", "match", "our", "theoretical", "analysis", ",", "verifying", "that", "the", "discriminator", "is", "able", "to", "recover", "the", "energy", "of", "data", "distribution", ".", "section", ":", "INTRODUCTION", "Generative", "Adversarial", "Networks", "(", "GANs", ")", "[", "reference", "]", "represent", "an", "important", "milestone", "on", "the", "path", "towards", "more", "effective", "generative", "models", ".", "GANs", "cast", "generative", "model", "training", "as", "a", "minimax", "game", "between", "a", "generative", "network", "(", "generator", ")", ",", "which", "maps", "a", "random", "vector", "into", "the", "data", "space", ",", "and", "a", "discriminative", "network", "(", "discriminator", ")", ",", "whose", "objective", "is", "to", "distinguish", "generated", "samples", "from", "real", "samples", ".", "Multiple", "researchers", "[", "reference", "]", ";", "[", "reference", "]", ";", "[", "reference", "]", "have", "shown", "that", "the", "adversarial", "interaction", "with", "the", "discriminator", "can", "result", "in", "a", "generator", "that", "produces", "compelling", "samples", ".", "The", "empirical", "successes", "of", "the", "GAN", "framework", "were", "also", "supported", "by", "the", "theoretical", "analysis", "of", "[", "reference", "]", "showed", "that", ",", "under", "certain", "conditions", ",", "the", "distribution", "produced", "by", "the", "generator", "converges", "to", "the", "true", "data", "distribution", ",", "while", "the", "discriminator", "converges", "to", "a", "degenerate", "uniform", "solution", ".", "While", "GANs", "have", "excelled", "as", "compelling", "sample", "generators", ",", "their", "use", "as", "general", "purpose", "probabilistic", "generative", "models", "has", "been", "limited", "by", "the", "difficulty", "in", "using", "them", "to", "provide", "density", "estimates", "or", "even", "unnormalized", "energy", "values", "for", "sample", "evaluation", ".", "It", "is", "tempting", "to", "consider", "the", "GAN", "discriminator", "as", "a", "candidate", "for", "providing", "this", "sort", "of", "scoring", "function", ".", "Conceptually", ",", "it", "is", "a", "trainable", "sample", "evaluation", "mechanism", "that", "-", "owing", "to", "GAN", "training", "paradigm", "-", "could", "be", "closely", "calibrated", "to", "the", "distribution", "modeled", "by", "the", "generator", ".", "If", "the", "discriminator", "could", "retain", "fine", "-", "grained", "information", "of", "the", "relative", "quality", "of", "samples", ",", "measured", "for", "instance", "by", "probability", "density", "or", "unnormalized", "energy", ",", "it", "could", "be", "used", "as", "an", "evaluation", "metric", ".", "Such", "data", "-", "driven", "evaluators", "would", "be", "highly", "desirable", "for", "problems", "where", "it", "is", "difficult", "to", "define", "evaluation", "criteria", "that", "correlate", "well", "with", "human", "judgment", ".", "Indeed", ",", "the", "real", "-", "valued", "discriminator", "of", "the", "recently", "introduced", "energy", "-", "based", "GANs", "[", "reference", "]", "might", "seem", "like", "an", "ideal", "candidate", "energy", "function", ".", "Unfortunately", ",", "as", "we", "will", "show", ",", "the", "degenerate", "fate", "of", "the", "GAN", "discriminator", "at", "the", "optimum", "equally", "afflicts", "the", "energy", "-", "based", "GAN", "of", "[", "reference", "]", "In", "this", "paper", "we", "consider", "the", "questions", ":", "(", "i", ")", "does", "there", "exists", "an", "adversarial", "framework", "that", "induces", "a", "non", "-", "degenerate", "discriminator", ",", "and", "(", "ii", ")", "if", "so", ",", "what", "form", "will", "the", "resulting", "discriminator", "take", "?", "We", "introduce", "a", "novel", "adversarial", "learning", "formulation", ",", "which", "leads", "to", "a", "non", "-", "degenerate", "discriminator", "while", "ensuring", "the", "generator", "distribution", "matches", "the", "data", "distribution", "at", "the", "global", "optimum", ".", "We", "derive", "a", "general", "analytic", "form", "of", "the", "optimal", "discriminator", ",", "and", "discuss", "its", "properties", "and", "their", "Following", "the", "original", "work", "on", "GANs", "[", "reference", "]", ",", "our", "analysis", "focuses", "on", "the", "nonparametric", "case", ",", "where", "all", "models", "are", "assumed", "to", "have", "infinite", "capacities", ".", "While", "many", "of", "the", "nonparametric", "intuitions", "can", "directly", "transfer", "to", "the", "parametric", "case", ",", "we", "will", "point", "out", "cases", "where", "this", "transfer", "fails", ".", "We", "assume", "a", "finite", "data", "space", "throughout", "the", "analysis", ",", "to", "avoid", "technical", "machinery", "out", "of", "the", "scope", "of", "this", "paper", ".", "Our", "results", ",", "however", ",", "can", "be", "extended", "to", "continuous", "data", "spaces", ",", "and", "our", "experiments", "are", "indeed", "performed", "on", "continuous", "data", ".", "Let", "X", "be", "the", "data", "space", "under", "consideration", ",", "and", "P", "=", "{", "p", "|", "p", "(", "x", ")", "≥", "0", ",", "∀x", "∈", "X", ",", "x∈X", "p", "(", "x", ")", "=", "1", "}", "be", "the", "set", "of", "all", "proper", "distributions", "defined", "on", "X", ".", "Then", ",", "p", "data", "∈", "P", ":", "X", "→", "R", "and", "p", "gen", "∈", "P", ":", "X", "→", "R", "will", "denote", "the", "true", "data", "distribution", "and", "the", "generator", "distribution", ".", "E", "x∼p", "f", "(", "x", ")", "denotes", "the", "expectation", "of", "the", "quantity", "f", "(", "x", ")", "w.r.t", ".", "x", "drawn", "from", "p.", "Finally", ",", "the", "term", "\"", "discriminator", "\"", "will", "refer", "to", "any", "structure", "that", "provides", "training", "signals", "to", "the", "generator", "based", "on", "some", "measure", "of", "difference", "between", "the", "generator", "distribution", "and", "the", "real", "data", "distribution", ",", "which", "which", "includes", "but", "is", "not", "limited", "to", "f", "-", "divergence", ".", "section", ":", "PROPOSED", "FORMULATION", "In", "order", "to", "understand", "the", "motivation", "of", "the", "proposed", "approach", ",", "it", "is", "helpful", "to", "analyze", "the", "optimization", "dynamics", "near", "convergence", "in", "GANs", "first", ".", "When", "the", "generator", "distribution", "matches", "the", "data", "distribution", ",", "the", "training", "signal", "(", "gradient", ")", "w.r.t", ".", "the", "discriminator", "vanishes", ".", "At", "this", "point", ",", "assume", "the", "discriminator", "still", "retains", "density", "information", ",", "and", "views", "some", "samples", "as", "more", "real", "and", "others", "as", "less", ".", "This", "discriminator", "will", "produce", "a", "training", "signal", "(", "gradient", ")", "w.r.t", ".", "the", "generator", ",", "pushing", "the", "generator", "to", "generate", "samples", "that", "appear", "more", "real", "to", "the", "discriminator", ".", "Critically", ",", "this", "training", "signal", "is", "the", "sole", "driver", "of", "the", "generator", "'s", "training", ".", "Hence", ",", "the", "generator", "distribution", "will", "diverge", "from", "the", "data", "distribution", ".", "In", "other", "words", ",", "as", "long", "as", "the", "discriminator", "retains", "relative", "density", "information", ",", "the", "generator", "distribution", "can", "not", "stably", "match", "the", "data", "distribution", ".", "Thus", ",", "in", "order", "to", "keep", "the", "generator", "stationary", "as", "the", "data", "distribution", ",", "the", "discriminator", "must", "assign", "flat", "(", "exactly", "the", "same", ")", "density", "to", "all", "samples", "at", "the", "optimal", ".", "From", "the", "analysis", "above", ",", "the", "fundamental", "difficulty", "is", "that", "the", "generator", "only", "receives", "a", "single", "training", "signal", "(", "gradient", ")", "from", "the", "discriminator", ",", "which", "it", "has", "to", "follow", ".", "To", "keep", "the", "generator", "stationary", ",", "this", "single", "training", "signal", "(", "gradient", ")", "must", "vanish", ",", "which", "requires", "a", "degenerate", "discriminator", ".", "In", "this", "work", ",", "we", "propose", "to", "tackle", "this", "single", "training", "signal", "constraint", "directly", ".", "Specifically", ",", "we", "introduce", "a", "novel", "adversarial", "learning", "formulation", "which", "incorporates", "an", "additional", "training", "signal", "to", "the", "generator", ",", "such", "that", "this", "additional", "signal", "can", "•", "balance", "(", "cancel", "out", ")", "the", "discriminator", "signal", "at", "the", "optimum", ",", "so", "that", "the", "generator", "can", "stay", "stationary", "even", "if", "the", "discriminator", "assigns", "non", "-", "flat", "density", "to", "samples", "•", "cooperate", "with", "the", "discriminator", "signal", "to", "make", "sure", "the", "generator", "converges", "to", "the", "data", "distribution", ",", "and", "the", "discriminator", "retains", "the", "correct", "relative", "density", "information", "The", "proposed", "formulation", "can", "be", "written", "as", "the", "following", "minimax", "training", "objective", ",", "where", "c", "(", "x", ")", ":", "X", "→", "R", "is", "the", "discriminator", "that", "assigns", "each", "data", "point", "an", "unbounded", "scalar", "cost", ",", "and", "K", "(", "p", "gen", ")", ":", "P", "→", "R", "is", "some", "(", "functionally", ")", "differentiable", ",", "convex", "function", "of", "p", "gen", ".", "Compared", "to", "the", "original", "GAN", ",", "despite", "the", "similar", "minimax", "surface", "form", ",", "the", "proposed", "fomulation", "has", "two", "crucial", "distinctions", ".", "Firstly", ",", "while", "the", "GAN", "discriminator", "tries", "to", "distinguish", "\"", "fake", "\"", "samples", "from", "real", "ones", "using", "binary", "classification", ",", "the", "proposed", "discriminator", "achieves", "that", "by", "assigning", "lower", "cost", "to", "real", "samples", "and", "higher", "cost", "to", "\"", "fake", "\"", "one", ".", "This", "distinction", "can", "be", "seen", "from", "the", "first", "two", "terms", "of", "Eqn", ".", "(", "1", ")", ",", "where", "the", "discriminator", "c", "(", "x", ")", "is", "trained", "to", "widen", "the", "expected", "cost", "gap", "between", "\"", "fake", "\"", "and", "real", "samples", ",", "while", "the", "generator", "is", "adversarially", "trained", "to", "minimize", "it", ".", "In", "addition", "to", "the", "different", "adversarial", "mechanism", ",", "a", "calibrating", "term", "K", "(", "p", "gen", ")", "is", "introduced", "to", "provide", "a", "countervailing", "source", "of", "training", "signal", "for", "p", "gen", "as", "we", "motivated", "above", ".", "For", "now", ",", "the", "form", "of", "K", "(", "p", "gen", ")", "has", "not", "been", "specified", ".", "But", "as", "we", "will", "see", "later", ",", "its", "choice", "will", "directly", "decide", "the", "form", "of", "the", "optimal", "discriminator", "c", "*", "(", "x", ")", ".", "With", "the", "specific", "optimization", "objective", ",", "we", "next", "provide", "theoretical", "characterization", "of", "both", "the", "generator", "and", "the", "discriminator", "at", "the", "global", "optimum", ".", "is", "the", "Lagrange", "dual", "function", "of", "the", "following", "optimization", "problem", "where", "c", "(", "x", ")", ",", "∀x", "appears", "in", "L", "(", "p", "gen", ",", "c", ")", "as", "the", "dual", "variables", "introduced", "for", "the", "equality", "constraints", ".", "This", "duality", "relationship", "has", "been", "observed", "previously", "in", "[", "reference", "]", ")", "under", "the", "adversarial", "imitation", "learning", "setting", ".", "However", ",", "in", "their", "case", ",", "the", "focus", "was", "fully", "on", "the", "generator", "side", "(", "induced", "policy", ")", ",", "and", "no", "analysis", "was", "provided", "for", "the", "discriminator", "(", "reward", "function", ")", ".", "In", "order", "to", "characterize", "c", "*", ",", "we", "first", "expand", "the", "set", "constraint", "on", "p", "gen", "into", "explicit", "equality", "and", "inequality", "constraints", ":", "min", "Notice", "that", "K", "(", "p", "gen", ")", "is", "a", "convex", "function", "of", "p", "gen", "(", "x", ")", "by", "definition", ",", "and", "both", "the", "equality", "and", "inequality", "constraints", "are", "affine", "functions", "of", "p", "gen", "(", "x", ")", ".", "Thus", ",", "problem", "(", "2", ")", "is", "a", "convex", "optimization", "problem", ".", "What", "'s", "more", ",", "since", "(", "i", ")", "dom", "K", "is", "open", ",", "and", "(", "ii", ")", "there", "exists", "a", "feasible", "solution", "p", "gen", "=", "p", "data", "to", "(", "3", ")", ",", "by", "the", "refined", "Slater", "'s", "condition", "[", "reference", "]", ",", "page", "226", ")", ",", "we", "can", "further", "verify", "that", "strong", "duality", "holds", "for", "(", "3", ")", ".", "With", "strong", "duality", ",", "a", "typical", "approach", "to", "characterizing", "the", "optimal", "solution", "is", "to", "apply", "the", "Karush", "-", "Kuhn", "-", "Tucker", "(", "KKT", ")", "conditions", ",", "which", "gives", "rise", "to", "this", "theorem", ":", "Proposition", "3.1", ".", "By", "the", "KKT", "conditions", "of", "the", "convex", "problem", "(", "3", ")", ",", "at", "the", "global", "optimum", ",", "the", "optimal", "generator", "distribution", "p", "*", "gen", "matches", "the", "true", "data", "distribution", "p", "data", ",", "and", "the", "optimal", "discriminator", "c", "*", "(", "x", ")", "has", "the", "following", "form", ":", "is", "an", "under", "-", "determined", "real", "number", "independent", "of", "x", ",", "u", "x", "∈", "R", "+", ",", "is", "an", "under", "-", "determined", "non", "-", "negative", "real", "number", ".", "The", "detailed", "proof", "of", "proposition", "3.1", "is", "provided", "in", "appendix", "A.1", ".", "From", "(", "4", ")", ",", "we", "can", "see", "the", "exact", "form", "of", "the", "optimal", "discriminator", "depends", "on", "the", "term", "K", "(", "p", "gen", ")", ",", "or", "more", "specifically", "its", "gradient", ".", "But", ",", "before", "we", "instantiate", "K", "(", "p", "gen", ")", "with", "specific", "choices", "and", "show", "the", "corresponding", "forms", "of", "c", "*", "(", "x", ")", ",", "we", "first", "discuss", "some", "general", "properties", "of", "c", "*", "(", "x", ")", "that", "do", "not", "depend", "on", "the", "choice", "of", "K.", "Weak", "Support", "Discriminator", ".", "As", "part", "of", "the", "optimal", "discriminator", "function", ",", "the", "term", "µ", "*", "(", "x", ")", "plays", "the", "role", "of", "support", "discriminator", ".", "That", "is", ",", "it", "tries", "to", "distinguish", "the", "support", "of", "the", "data", "distribution", ",", "i.e.", "SUPP", "(", "p", "data", ")", "=", "{", "x", "∈", "X", "|", "p", "data", "(", "x", ")", ">", "0", "}", ",", "from", "its", "complement", "set", "with", "zeroprobability", ",", "i.e.", "SUPP", "(", "p", "data", ")", "=", "{", "x", "∈", "X", "|", "p", "data", "(", "x", ")", "=", "0}.", "Specifically", ",", "for", "any", "x", "∈", "SUPP", "(", "p", "data", ")", "and", "x", "∈", "SUPP", "(", "p", "data", ")", ",", "it", "is", "guaranteed", "that", "µ", "*", "(", "x", ")", "≤", "µ", "*", "(", "x", ")", ".", "However", ",", "because", "µ", "*", "(", "·", ")", "is", "underdetermined", ",", "there", "is", "nothing", "preventing", "the", "inequality", "from", "degenerating", "into", "an", "equality", ".", "Therefore", ",", "we", "name", "it", "the", "weak", "support", "discriminator", ".", "But", ",", "in", "all", "cases", ",", "µ", "*", "(", "·", ")", "assigns", "zero", "cost", "to", "all", "data", "points", "within", "the", "support", ".", "As", "a", "result", ",", "it", "does", "not", "possess", "any", "fine", "-", "grained", "density", "information", "inside", "of", "the", "data", "support", ".", "It", "is", "worth", "pointing", "out", "that", ",", "in", "the", "parametric", "case", ",", "because", "of", "the", "smoothness", "and", "the", "generalization", "properties", "of", "the", "parametric", "model", ",", "the", "learned", "discriminator", "may", "generalize", "beyond", "the", "data", "support", ".", "Global", "Bias", ".", "In", "(", "4", ")", ",", "the", "term", "λ", "*", "is", "a", "scalar", "value", "shared", "for", "all", "x.", "As", "a", "result", ",", "it", "does", "not", "affect", "the", "relative", "cost", "among", "data", "points", ",", "and", "only", "serves", "as", "a", "global", "bias", "for", "the", "discriminator", "function", ".", "Having", "discussed", "general", "properties", ",", "we", "now", "consider", "some", "specific", "cases", "of", "the", "convex", "function", "K", ",", "and", "analyze", "the", "resulting", "optimal", "discriminator", "c", "*", "(", "x", ")", "in", "detail", ".", "1", ".", "First", ",", "let", "us", "consider", "the", "case", "where", "K", "is", "the", "negative", "entropy", "of", "the", "generator", "distribution", ",", "i.e.", "K", "(", "p", "gen", ")", "=", "−H", "(", "p", "gen", ")", ".", "Taking", "the", "derivative", "of", "the", "negative", "entropy", "w.r.t", ".", "p", "gen", "(", "x", ")", ",", "we", "have", "where", "µ", "*", "(", "x", ")", "and", "λ", "*", "have", "the", "same", "definitions", "as", "in", "(", "4", ")", ".", "Up", "to", "a", "constant", ",", "this", "form", "of", "c", "*", "ent", "(", "x", ")", "is", "exactly", "the", "energy", "function", "of", "the", "data", "distribution", "p", "data", "(", "x", ")", ".", "This", "elegant", "result", "has", "deep", "connections", "to", "several", "existing", "formulations", ",", "which", "include", "max", "-", "entropy", "imitation", "learning", "[", "reference", "]", "and", "the", "directed", "-", "generator", "-", "trained", "energybased", "model", "[", "reference", "]", ".", "The", "core", "difference", "is", "that", "these", "previous", "formulations", "are", "originally", "derived", "from", "maximum", "-", "likelihood", "estimation", ",", "and", "thus", "the", "minimax", "optimization", "is", "only", "implicit", ".", "In", "contrast", ",", "with", "an", "explicit", "minimax", "formulation", "we", "can", "develop", "a", "better", "understanding", "of", "the", "induced", "solution", ".", "For", "example", ",", "the", "global", "bias", "λ", "*", "suggests", "that", "there", "exists", "more", "than", "one", "stable", "equilibrium", "the", "optimal", "discriminator", "can", "actually", "reach", ".", "Further", ",", "µ", "*", "(", "x", ")", "can", "be", "understood", "as", "a", "support", "discriminator", "that", "poses", "extra", "cost", "on", "generator", "samples", "which", "fall", "in", "zero", "-", "probability", "regions", "of", "data", "space", ".", ",", "which", "can", "be", "understood", "as", "posing", "2", "regularization", "on", "p", "gen", ",", "we", "have", "with", "µ", "*", "(", "x", ")", ",", "λ", "*", "similarly", "defined", "as", "in", "(", "4", ")", ".", "Surprisingly", ",", "the", "result", "suggests", "that", "the", "optimal", "discriminator", "c", "*", "2", "(", "x", ")", "directly", "recovers", "the", "negative", "probability", "−p", "data", "(", "x", ")", ",", "shifted", "by", "a", "constant", ".", "Thus", ",", "similar", "to", "the", "entropy", "solution", "(", "5", ")", ",", "it", "fully", "retains", "the", "relative", "density", "information", "of", "data", "points", "within", "the", "support", ".", "However", ",", "because", "of", "the", "under", "-", "determined", "term", "µ", "*", "(", "x", ")", ",", "we", "can", "not", "recover", "the", "distribution", "density", "p", "data", "exactly", "from", "either", "c", "*", "2", "or", "c", "*", "ent", "if", "the", "data", "support", "is", "finite", ".", "Whether", "this", "ambiguity", "can", "be", "resolved", "is", "beyond", "the", "scope", "of", "this", "paper", ",", "but", "poses", "an", "interesting", "research", "problem", ".", "3", ".", "Finally", ",", "let", "'s", "consider", "consider", "a", "degenerate", "case", ",", "where", "K", "(", "p", "gen", ")", "is", "a", "constant", ".", "That", "is", ",", "we", "do", "nt", "provide", "any", "additional", "training", "signal", "for", "pgen", "at", "all", ".", "With", "K", "(", "p", "gen", ")", "=", "const", ",", "we", "simply", "have", "whose", "discriminative", "power", "is", "fully", "controlled", "by", "the", "weak", "support", "discriminator", "µ", "*", "(", "x", ")", ".", "Thus", ",", "it", "follows", "that", "c", "*", "cst", "(", "x", ")", "wo", "n't", "be", "able", "to", "discriminate", "data", "points", "within", "the", "support", "of", "p", "data", ",", "and", "its", "power", "to", "distinguish", "data", "from", "SUPP", "(", "p", "data", ")", "and", "SUPP", "(", "p", "data", ")", "is", "weak", ".", "This", "closely", "matches", "the", "intuitive", "argument", "in", "the", "beginning", "of", "this", "section", ".", "Note", "that", "when", "K", "(", "p", "gen", ")", "is", "a", "constant", ",", "the", "objective", "function", "(", "1", ")", "simplifies", "to", ":", "which", "is", "very", "similar", "to", "the", "EBGAN", "objective", "[", "reference", "]", "(", "2", ")", "and", "(", "4", ")", ")", ".", "As", "we", "show", "in", "appendix", "A.2", ",", "compared", "to", "the", "objective", "in", "(", "8", ")", ",", "the", "EBGAN", "objective", "puts", "extra", "constraints", "on", "the", "allowed", "discriminator", "function", ".", "In", "spite", "of", "that", ",", "the", "EBGAN", "objective", "suffers", "from", "the", "single", "-", "training", "-", "signal", "problem", "and", "does", "not", "guarantee", "that", "the", "discriminator", "will", "recover", "the", "real", "energy", "function", "(", "see", "appendix", "A.2", "for", "detailed", "analysis", ")", ".", "As", "we", "finish", "the", "theoretical", "analysis", "of", "the", "proposed", "formulation", ",", "we", "want", "to", "point", "out", "that", "simply", "adding", "the", "same", "term", "K", "(", "p", "gen", ")", "to", "the", "original", "GAN", "formulation", "will", "not", "lead", "to", "both", "a", "generator", "that", "matches", "the", "data", "distribution", ",", "and", "a", "discriminator", "that", "retains", "the", "density", "information", "(", "see", "appendix", "A.3", "for", "detailed", "analysis", ")", ".", "section", ":", "PARAMETRIC", "INSTANTIATION", "WITH", "ENTROPY", "APPROXIMATION", "While", "the", "discussion", "in", "previous", "sections", "focused", "on", "the", "non", "-", "parametric", "case", ",", "in", "practice", "we", "are", "limited", "to", "a", "finite", "amount", "of", "data", ",", "and", "the", "actual", "problem", "involves", "high", "dimensional", "continuous", "spaces", ".", "Thus", ",", "we", "resort", "to", "parametric", "representations", "for", "both", "the", "generator", "and", "the", "discriminator", ".", "In", "order", "to", "train", "the", "generator", "using", "standard", "back", "-", "propagation", ",", "we", "do", "not", "parametrize", "the", "generator", "distribution", "directly", ".", "Instead", ",", "we", "parametrize", "a", "directed", "generator", "network", "that", "transforms", "random", "noise", "z", "∼", "p", "z", "(", "z", ")", "to", "samples", "from", "a", "continuous", "data", "space", "R", "n", ".", "Consequently", ",", "we", "do", "n't", "have", "analytical", "access", "to", "the", "generator", "distribution", ",", "which", "is", "defined", "implicitly", "by", "the", "generator", "network", "'s", "noise→data", "mapping", ".", "However", ",", "the", "regularization", "term", "K", "(", "p", "gen", ")", "in", "the", "training", "objective", "(", "1", ")", "requires", "the", "generator", "distribution", ".", "Faced", "with", "this", "problem", ",", "we", "focus", "on", "the", "max", "-", "entropy", "formulation", ",", "and", "exploit", "two", "different", "approximations", "of", "the", "regularization", "term", "K", "(", "p", "gen", ")", "=", "−H", "(", "p", "gen", ")", ".", "section", ":", "NEAREST", "-", "NEIGHBOR", "ENTROPY", "GRADIENT", "APPROXIMATION", "The", "first", "proposed", "solution", "is", "built", "upon", "an", "intuitive", "interpretation", "of", "the", "entropy", "gradient", ".", "Firstly", ",", "since", "we", "construct", "p", "gen", "by", "applying", "a", "deterministic", ",", "differentiable", "transform", "g", "θ", "to", "samples", "z", "from", "a", "fixed", "distribution", "p", "z", ",", "we", "can", "write", "the", "gradient", "of", "H", "(", "p", "gen", ")", "with", "respect", "to", "the", "generator", "parameters", "θ", "as", "follows", ":", "where", "the", "first", "equality", "relies", "on", "the", "\"", "reparametrization", "trick", "\"", ".", "Equation", "9", "implies", "that", ",", "if", "we", "can", "compute", "the", "gradient", "of", "the", "generator", "log", "-", "density", "log", "p", "gen", "(", "x", ")", "w.r.t", ".", "any", "x", "=", "g", "θ", "(", "z", ")", ",", "then", "we", "can", "directly", "construct", "the", "Monte", "-", "Carlo", "estimation", "of", "the", "entropy", "gradient", "∇", "θ", "H", "(", "p", "gen", ")", "using", "samples", "from", "the", "generator", ".", "Intuitively", ",", "for", "any", "generated", "data", "x", "=", "g", "θ", "(", "z", ")", ",", "the", "term", "essentially", "describes", "the", "direction", "of", "local", "change", "in", "the", "sample", "space", "that", "will", "increase", "the", "log", "-", "density", ".", "Motivated", "by", "this", "intuition", ",", "we", "propose", "to", "form", "a", "local", "Gaussian", "approximation", "p", "i", "gen", "of", "p", "gen", "around", "each", "point", "x", "i", "in", "a", "batch", "of", "samples", "{", "x", "1", ",", "...", ",", "x", "n", "}", "from", "the", "generator", ",", "and", "then", "compute", "the", "gradient", "∂", "log", "pgen", "(", "xi", ")", "∂xi", "based", "on", "the", "Gaussian", "approximation", ".", "Specifically", ",", "each", "local", "Gaussian", "approximation", "p", "i", "gen", "is", "formed", "by", "finding", "the", "k", "nearest", "neighbors", "of", "x", "i", "in", "the", "batch", "{", "x", "1", ",", "...", ",", "x", "n", "}", ",", "and", "then", "placing", "an", "isotropic", "Gaussian", "distribution", "at", "their", "mean", "(", "i.e.", "maximimum", "likelihood", ")", ".", "Based", "on", "the", "isotropic", "Gaussian", "approximation", ",", "the", "resulting", "gradient", "has", "the", "following", "form", "x", "is", "the", "mean", "of", "the", "Gaussian", "(", "10", ")", "Finally", ",", "note", "the", "scale", "of", "this", "gradient", "approximation", "may", "not", "be", "reliable", ".", "To", "fix", "this", "problem", ",", "we", "normalize", "the", "approximated", "gradient", "into", "unit", "norm", ",", "and", "use", "a", "single", "hyper", "-", "parameter", "to", "model", "the", "scale", "for", "all", "x", ",", "leading", "to", "the", "following", "entropy", "gradient", "approximation", "where", "α", "is", "the", "hyper", "-", "parameter", "and", "µ", "i", "is", "defined", "as", "in", "equation", "(", "10", ")", ".", "An", "obvious", "weakness", "of", "this", "approximation", "is", "that", "it", "relies", "on", "Euclidean", "distance", "to", "find", "the", "k", "nearest", "neighbors", ".", "However", ",", "Euclidean", "distance", "is", "usually", "not", "the", "proper", "metric", "to", "use", "when", "the", "effective", "dimension", "is", "very", "high", ".", "As", "the", "problem", "is", "highly", "challenging", ",", "we", "leave", "it", "for", "future", "work", ".", "section", ":", "VARIATIONAL", "LOWER", "BOUND", "ON", "THE", "ENTROPY", "Another", "approach", "we", "consider", "relies", "on", "defining", "and", "maximizing", "a", "variational", "lower", "bound", "on", "the", "entropy", "H", "(", "p", "gen", "(", "x", ")", ")", "of", "the", "generator", "distribution", ".", "We", "can", "define", "the", "joint", "distribution", "over", "observed", "data", "and", "the", "noise", "variables", "as", "is", "a", "fixed", "prior", ".", "Using", "the", "joint", ",", "we", "can", "also", "define", "the", "marginal", "p", "gen", "(", "x", ")", "and", "the", "posterior", "p", "gen", "(", "z", "|", "x", ")", ".", "We", "can", "also", "write", "the", "mutual", "information", "between", "the", "observed", "data", "and", "noise", "variables", "as", ":", "where", "H", "(", "p", "gen", "(", ".", "|", ".", ")", ")", "denotes", "the", "conditional", "entropy", ".", "By", "reorganizing", "terms", "in", "this", "definition", ",", "we", "can", "write", "the", "entropy", "H", "(", "p", "gen", "(", "x", ")", ")", "as", ":", "We", "can", "think", "of", "p", "gen", "(", "x", "|", "z", ")", "as", "a", "peaked", "Gaussian", "with", "a", "fixed", ",", "diagonal", "covariance", ",", "and", "hence", "its", "conditional", "entropy", "is", "constant", "and", "can", "be", "dropped", ".", "Furthermore", ",", "H", "(", "p", "gen", "(", "z", ")", ")", "is", "also", "assumed", "to", "be", "fixed", "a", "priori", ".", "Hence", ",", "we", "can", "maximize", "H", "(", "p", "gen", "(", "x", ")", ")", "by", "minimizing", "the", "conditional", "entropy", ":", "Optimizing", "this", "term", "is", "still", "problematic", ",", "because", "(", "i", ")", "we", "do", "not", "have", "access", "to", "the", "posterior", "p", "gen", "(", "z", "|", "x", ")", ",", "and", "(", "ii", ")", "we", "can", "not", "sample", "from", "it", ".", "Therefore", ",", "we", "instead", "minimize", "a", "variational", "upper", "bound", "defined", "by", "an", "approximate", "posterior", "q", "gen", "(", "z", "|", "x", ")", ":", "We", "can", "also", "rewrite", "the", "variational", "upper", "bound", "as", ":", "which", "can", "be", "optimized", "efficiently", "with", "standard", "back", "-", "propagation", "and", "Monte", "Carlo", "integration", "of", "the", "relevant", "expectations", "based", "on", "independent", "samples", "drawn", "from", "the", "joint", "p", "gen", "(", "x", ",", "z", ")", ".", "By", "minimizing", "this", "upper", "bound", "on", "the", "conditional", "entropy", "H", "(", "p", "gen", "(", "z", "|", "x", ")", ")", ",", "we", "are", "effectively", "maximizing", "a", "variational", "lower", "bound", "on", "the", "entropy", "H", "(", "p", "gen", "(", "x", ")", ")", ".", "section", ":", "EXPERIMENTS", "In", "this", "section", ",", "we", "verify", "our", "theoretical", "results", "empirically", "on", "several", "synthetic", "and", "real", "datasets", ".", "In", "particular", ",", "we", "evaluate", "whether", "the", "discriminator", "obtained", "from", "the", "entropy", "-", "regularized", "adversarial", "training", "can", "capture", "the", "density", "information", "(", "in", "the", "form", "of", "energy", ")", ",", "while", "making", "sure", "the", "generator", "distribution", "matches", "the", "data", "distribution", ".", "For", "convenience", ",", "we", "refer", "to", "the", "obtained", "models", "as", "EGAN", "-", "Ent", ".", "Our", "experimental", "setting", "follows", "closely", "recommendations", "from", "[", "reference", "]", ",", "except", "in", "Sec", ".", "5.1", "where", "we", "use", "fully", "-", "connected", "models", "(", "see", "appendix", "B.1", "for", "details", ")", ".", "Gaussians", "arranged", "as", "two", "spirals", "(", "100", "components", "each", "spiral", ")", ",", "and", "(", "iii", ")", "Mixture", "of", "2", "Gaussians", "with", "highly", "biased", "mixture", "weights", ",", "P", "(", "c", "1", ")", "=", "0.9", ",", "P", "(", "c", "2", ")", "=", "0.1", ".", "We", "visualize", "the", "ground", "-", "truth", "energy", "of", "these", "distributions", "along", "with", "100", "K", "training", "samples", "in", "Figure", "1", ".", "Since", "the", "data", "lies", "in", "2", "-", "dimensional", "space", ",", "we", "can", "easily", "visualize", "both", "the", "learned", "generator", "(", "by", "drawing", "samples", ")", "and", "the", "discriminator", "for", "direct", "comparison", "and", "evaluation", ".", "We", "evaluate", "here", "our", "EGAN", "-", "Ent", "model", "using", "both", "approximations", ":", "the", "nearest", "-", "neighbor", "based", "approximation", "(", "EGAN", "-", "Ent", "-", "NN", ")", "and", "the", "variational", "-", "inference", "based", "approximation", "(", "EGAN", "-", "Ent", "-", "VI", ")", ",", "and", "compare", "them", "with", "two", "baselines", ":", "the", "original", "GAN", "and", "the", "energy", "based", "GAN", "with", "no", "regularization", "(", "EGAN", "-", "Const", ")", ".", "Experiment", "results", "are", "summarized", "in", "Figure", "2", "for", "baseline", "models", ",", "and", "Figure", "3", "for", "the", "proposed", "models", ".", "As", "we", "can", "see", ",", "all", "four", "models", "can", "generate", "perfect", "samples", ".", "However", ",", "for", "the", "discriminator", ",", "both", "GAN", "and", "EGAN", "-", "Const", "lead", "to", "degenerate", "solution", ",", "assigning", "flat", "energy", "inside", "the", "empirical", "data", "support", ".", "In", "comparison", ",", "EGAN", "-", "Ent", "-", "VI", "and", "EGAN", "-", "Ent", "-", "NN", "clearly", "capture", "the", "density", "information", ",", "though", "to", "different", "degrees", ".", "Specifically", ",", "on", "the", "equally", "weighted", "Gaussian", "mixture", "and", "the", "two", "-", "spiral", "mixture", "datasets", ",", "EGAN", "-", "Ent", "-", "NN", "tends", "to", "give", "more", "accurate", "and", "fine", "-", "grained", "solutions", "compared", "to", "EGAN", "-", "Ent", "-", "VI", ".", "However", ",", "on", "the", "biased", "weighted", "Gaussian", "mixture", "dataset", ",", "EGAN", "-", "Ent", "-", "VI", "actually", "fails", "to", "captures", "the", "correct", "mixture", "weights", "of", "the", "two", "modes", ",", "incorrectly", "assigning", "lower", "energy", "to", "the", "mode", "with", "lower", "probability", "(", "smaller", "weight", ")", ".", "In", "contrast", ",", "EGANEnt", "-", "NN", "perfectly", "captures", "the", "bias", "in", "mixture", "weight", ",", "and", "obtains", "a", "contour", "very", "close", "to", "the", "ground", "truth", ".", "To", "better", "quantify", "these", "differences", ",", "we", "present", "detailed", "comparison", "based", "on", "KL", "divergence", "in", "appendix", "B.2", ".", "What", "'s", "more", ",", "the", "performance", "difference", "between", "EGAN", "-", "Ent", "-", "VI", "and", "EGAN", "-", "Ent", "-", "NN", "on", "biased", "Gaussian", "mixture", "reveals", "the", "limitations", "of", "the", "variational", "inference", "based", "approximation", ",", "i.e.", "providing", "inaccurate", "gradients", ".", "Due", "to", "space", "consideratiosn", ",", "we", "refer", "interested", "readers", "to", "the", "appendix", "B.3", "for", "a", "detailed", "discussion", ".", "section", ":", "RANKING", "NIST", "DIGITS", "In", "this", "experiment", ",", "we", "verify", "that", "the", "results", "in", "synthetic", "datasets", "can", "translate", "into", "data", "with", "higher", "dimensions", ".", "While", "visualizing", "the", "learned", "energy", "function", "is", "not", "feasible", "in", "high", "-", "dimensional", "space", ",", "we", "can", "verify", "whether", "the", "learned", "energy", "function", "learns", "relative", "densities", "by", "inspecting", "the", "ranking", "of", "samples", "according", "to", "their", "assigned", "energies", ".", "We", "train", "on", "28", "×", "28", "images", "of", "a", "single", "handwritten", "Figure", "2", ":", "Learned", "energies", "and", "samples", "from", "baseline", "models", "whose", "discriminator", "can", "not", "retain", "density", "information", "at", "the", "optimal", ".", "In", "the", "sample", "plots", ",", "blue", "dots", "indicate", "generated", "samples", ",", "and", "red", "dots", "indicate", "real", "ones", ".", "section", ":", "(", "a", ")", "Entropy", "regularized", "Energy", "GAN", "with", "variational", "inference", "approximation", "(", "EGAN", "-", "Ent", "-", "VI", ")", "(", "b", ")", "Entropy", "regularized", "Energy", "GAN", "with", "nearest", "neighbor", "approximation", "(", "EGAN", "-", "Ent", "-", "NN", ")", "Figure", "3", ":", "Learned", "energies", "and", "samples", "from", "proposed", "models", "whose", "discriminator", "can", "retain", "density", "information", "at", "the", "optimal", ".", "Blue", "dots", "are", "generated", "samples", ",", "and", "red", "dots", "are", "real", "ones", ".", "digit", "from", "the", "NIST", "dataset", ".", "[", "reference", "]", "We", "compare", "the", "ability", "of", "EGAN", "-", "Ent", "-", "NN", "with", "both", "EGAN", "-", "Const", "and", "GAN", "on", "ranking", "a", "set", "of", "1", ",", "000", "images", ",", "half", "of", "which", "are", "generated", "samples", "and", "the", "rest", "are", "real", "test", "images", ".", "Figures", "4", "and", "5", "show", "the", "top", "-", "100", "and", "bottom", "-", "100", "ranked", "images", "respectively", "for", "each", "model", ",", "after", "training", "them", "on", "digit", "1", ".", "We", "also", "show", "in", "Figure", "7", "the", "mean", "of", "all", "training", "samples", ",", "so", "we", "can", "get", "a", "sense", "of", "what", "is", "the", "most", "common", "style", "(", "highest", "density", ")", "of", "digit", "1", "in", "NIST", ".", "We", "can", "notice", "that", "all", "of", "the", "top", "-", "ranked", "images", "by", "EGAN", "-", "Ent", "-", "NN", "look", "similar", "to", "the", "mean", "sample", ".", "In", "addition", ",", "the", "lowest", "-", "ranked", "images", "are", "clearly", "different", "from", "the", "mean", "image", ",", "with", "either", "high", "(", "clockwise", "or", "counter", "-", "clockwise", ")", "rotation", "degrees", "from", "the", "mean", ",", "or", "an", "extreme", "thickness", "level", ".", "We", "do", "not", "see", "such", "clear", "distinction", "in", "other", "models", ".", "We", "provide", "in", "the", "appendix", "B.4", "the", "ranking", "of", "the", "full", "set", "of", "images", ".", "section", ":", "SAMPLE", "QUALITY", "ON", "NATURAL", "IMAGE", "DATASETS", "In", "this", "last", "set", "of", "experiments", ",", "we", "evaluate", "the", "visual", "quality", "of", "samples", "generated", "by", "our", "model", "in", "two", "datasets", "of", "natural", "images", ",", "namely", "CIFAR", "-", "10", "and", "CelebA.", "We", "employ", "here", "the", "variationalbased", "approximation", "for", "entropy", "regularization", ",", "which", "can", "scale", "well", "to", "high", "-", "dimensional", "data", ".", "Figure", "6", "shows", "samples", "generated", "by", "EGAN", "-", "Ent", "-", "VI", ".", "We", "can", "see", "that", "despite", "the", "noisy", "gradients", "provided", "by", "the", "variational", "approximation", ",", "our", "model", "is", "able", "to", "generate", "high", "-", "quality", "samples", ".", "We", "futher", "validate", "the", "quality", "of", "our", "model", "'s", "samples", "on", "CIFAR", "-", "10", "using", "the", "Inception", "score", "proposed", "by", "[", "reference", "]", "3", ".", "Table", "1", "shows", "the", "scores", "of", "our", "EGAN", "-", "Ent", "-", "VI", ",", "the", "best", "GAN", "model", "from", "[", "reference", "]", "which", "uses", "only", "unlabeled", "data", ",", "and", "an", "EGAN", "-", "Const", "model", "which", "has", "the", "same", "architecture", "as", "our", "model", ".", "We", "notice", "that", "even", "without", "employing", "suggested", "techniques", "in", "[", "reference", "]", ",", "energy", "-", "based", "models", "perform", "quite", "similarly", "to", "the", "GAN", "model", ".", "Furthermore", ",", "the", "fact", "that", "our", "model", "scores", "higher", "than", "EGAN", "-", "Const", "highlights", "the", "importance", "of", "entropy", "regularization", "in", "obtaining", "good", "quality", "samples", ".", "section", ":", "CONCLUSION", "In", "this", "paper", "we", "have", "addressed", "a", "fundamental", "limitation", "in", "adversarial", "learning", "approaches", ",", "which", "is", "their", "inability", "of", "providing", "sensible", "energy", "estimates", "for", "samples", ".", "We", "proposed", "a", "novel", "adversarial", "learning", "formulation", "which", "results", "in", "a", "discriminator", "function", "that", "recovers", "the", "true", "data", "energy", ".", "We", "provided", "a", "rigorous", "characterization", "of", "the", "learned", "discriminator", "in", "the", "non", "-", "parametric", "setting", ",", "and", "proposed", "two", "methods", "for", "instantiating", "it", "in", "the", "typical", "parametric", "setting", ".", "Our", "experimental", "results", "verify", "our", "theoretical", "analysis", "about", "the", "discriminator", "properties", ",", "and", "show", "that", "we", "can", "also", "obtain", "samples", "of", "state", "-", "of", "-", "the", "-", "art", "quality", ".", "Figure", "6", ":", "Samples", "generated", "from", "our", "model", ".", "section", ":", "Model", "Our", "model", "Improved", "GAN", "†", "EGAN", "-", "Const", "Score", "±", "std", ".", "7.07", "±", ".10", "6.86", "±", ".06", "6.7447", "±", "0.09", "Table", "1", ":", "Inception", "scores", "on", "CIFAR", "-", "10", ".", "†", "As", "reported", "in", "Salimans", "et", "al", ".", "(", "2016", ")", "without", "using", "labeled", "data", ".", "(", "17", ")", "where", "c", "(", "x", ")", "∈", "R", ",", "∀x", ",", "µ", "(", "x", ")", "∈", "R", "+", ",", "∀x", ",", "and", "λ", "∈", "R", "are", "the", "dual", "variables", ".", "The", "KKT", "conditions", "for", "the", "optimal", "primal", "and", "dual", "variables", "are", "as", "follows", "Rearranging", "the", "conditions", "above", ",", "we", "get", "p", "*", "gen", "(", "x", ")", "=", "p", "data", "(", "x", ")", ",", "∀x", "∈", "X", "as", "well", "as", "equation", "(", "4", ")", ",", "which", "concludes", "the", "proof", ".", "section", ":", "A.2", "OPTIMAL", "CONDITIONS", "OF", "EBGAN", "In", "[", "reference", "]", ",", "the", "training", "objectives", "of", "the", "generator", "and", "the", "discriminator", "can", "not", "be", "written", "as", "a", "single", "minimax", "optimization", "problem", "since", "the", "margin", "structure", "is", "only", "applied", "to", "the", "objective", "of", "the", "discriminator", ".", "In", "addition", ",", "the", "discriminator", "is", "designed", "to", "produce", "the", "mean", "squared", "reconstruction", "error", "of", "an", "auto", "-", "encoder", "structure", ".", "This", "restricted", "the", "range", "of", "the", "discriminator", "output", "to", "be", "non", "-", "negative", ",", "which", "is", "equivalent", "to", "posing", "a", "set", "constraint", "on", "the", "discriminator", "under", "the", "non", "-", "parametric", "setting", ".", "Thus", ",", "to", "characterize", "the", "optimal", "generator", "and", "discriminator", ",", "we", "adapt", "the", "same", "analyzing", "logic", "used", "in", "the", "proof", "sketch", "of", "the", "original", "GAN", "[", "reference", "]", ".", "Specifically", ",", "given", "a", "specific", "generator", "distribution", "p", "gen", ",", "the", "optimal", "discriminator", "function", "given", "the", "generator", "distribution", "c", "*", "(", "x", ";", "p", "gen", ")", "can", "be", "derived", "by", "examining", "the", "objective", "of", "the", "discriminator", ".", "Then", ",", "the", "conditional", "optimal", "discriminator", "function", "is", "substituted", "into", "the", "training", "objective", "of", "p", "gen", ",", "simplifying", "the", "\"", "adversarial", "\"", "training", "as", "a", "minimizing", "problem", "only", "w.r.t", ".", "p", "gen", ",", "which", "can", "be", "well", "analyzed", ".", "Firstly", ",", "given", "any", "generator", "distribution", "p", "gen", ",", "the", "EBGAN", "training", "objective", "for", "the", "discriminator", "can", "be", "written", "as", "the", "following", "form", "where", "C", "=", "{", "c", ":", "c", "(", "x", ")", "≥", "0", ",", "∀x", "∈", "X", "}", "is", "the", "set", "of", "allowed", "non", "-", "negative", "discriminator", "functions", ".", "Note", "this", "set", "constraint", "comes", "from", "the", "fact", "the", "mean", "squared", "reconstruction", "error", "as", "discussed", "above", ".", "Since", "the", "problem", "(", "19", ")", "is", "independent", "w.r.t", ".", "each", "x", ",", "the", "optimal", "solution", "can", "be", "easily", "derived", "as", "where", "α", "x", "∈", "[", "0", ",", "m", "]", "is", "an", "under", "-", "determined", "number", ",", "a", "β", "x", "∈", "[", "0", ",", "∞", ")", "is", "another", "under", "-", "determined", "nonnegative", "real", "number", ",", "and", "the", "subscripts", "in", "m", ",", "α", "x", ",", "β", "x", "reflect", "that", "fact", "that", "these", "under", "-", "determined", "values", "can", "be", "distinct", "for", "different", "x.", "This", "way", ",", "the", "overall", "training", "objective", "can", "be", "cast", "into", "a", "minimization", "problem", "w.r.t", ".", "p", "gen", ",", "where", "the", "second", "term", "of", "the", "first", "line", "is", "implicitly", "defined", "as", "the", "problem", "is", "an", "adversarial", "game", "between", "p", "gen", "and", "c.", "Proposition", "A.1", ".", "The", "global", "optimal", "of", "the", "EBGAN", "training", "objective", "is", "achieved", "if", "and", "only", "if", "p", "gen", "=", "p", "data", ".", "At", "that", "point", ",", "c", "*", "(", "x", ")", "is", "fully", "under", "-", "determined", ".", "Proof", ".", "The", "proof", "is", "established", "by", "showing", "contradiction", ".", "Firstly", ",", "assume", "the", "optimal", "p", "*", "gen", "=", "p", "data", ".", "Thus", ",", "there", "must", "exist", "a", "non", "-", "equal", "set", "X", "=", "=", "{", "x", "|", "p", "data", "(", "x", ")", "=", "p", "*", "gen", "(", "x", ")", "}", ",", "which", "can", "be", "further", "splitted", "into", "two", "subsets", ",", "the", "greater", "-", "than", "set", "X", ">", "=", "{", "x", "|", "p", "*", "gen", "(", "x", ")", ">", "p", "data", "(", "x", ")", "}", ",", "and", "the", "less", "-", "than", "set", "X", "<", "=", "{", "x", "|", "p", "*", "gen", "(", "x", ")", "<", "p", "data", "(", "x", ")", "}.", "Similarly", ",", "we", "define", "the", "equal", "set", ",", "substituting", "the", "results", "from", "equation", "(", "20", ")", "into", "(", "21", ")", ",", "the", "L", "(", "p", "gen", ")", "*", "can", "be", "written", "as", "However", ",", "when", "p", "gen", "=", "p", "data", ",", "we", "have", "which", "contradicts", "the", "optimal", "(", "miminum", ")", "assumption", "of", "p", "*", "gen", ".", "Hence", ",", "the", "contradiction", "concludes", "that", "at", "the", "global", "optimal", ",", "p", "*", "gen", "=", "p", "data", ".", "By", "equation", "(", "20", ")", ",", "it", "directly", "follows", "that", "c", "*", "(", "x", ";", "p", "*", "gen", ")", "=", "α", "x", ",", "which", "completes", "the", "proof", ".", "section", ":", "A.3", "ANALYSIS", "OF", "ADDING", "ADDITIONAL", "TRAINING", "SIGNAL", "TO", "GAN", "FORMULATION", "To", "show", "that", "simply", "adding", "the", "same", "training", "signal", "to", "GAN", "will", "not", "lead", "to", "the", "same", "result", ",", "it", "is", "more", "convenient", "to", "directly", "work", "with", "the", "formulation", "of", "f", "-", "GAN", "[", "reference", "]", ")", "family", ",", "which", "include", "the", "original", "GAN", "formulation", "as", "a", "special", "case", ".", "Specifically", ",", "the", "general", "f", "-", "GAN", "formulation", "takes", "the", "following", "form", "where", "the", "f", "(", "·", ")", "denotes", "the", "convex", "conjugate", "[", "reference", "]", "of", "the", "f", "-", "divergence", "function", ".", "The", "optimal", "condition", "of", "the", "discriminator", "can", "be", "found", "by", "taking", "the", "variation", "w.r.t", ".", "c", ",", "which", "gives", "the", "optimal", "discriminator", "where", "f", "(", "·", ")", "is", "the", "first", "-", "order", "derivative", "of", "f", "(", "·", ")", ".", "Note", "that", ",", "even", "when", "we", "add", "an", "extra", "term", "L", "(", "p", "gen", ")", "to", "equation", "(", "24", ")", ",", "since", "the", "term", "K", "(", "p", "gen", ")", "is", "a", "constant", "w.r.t", ".", "the", "discriminator", ",", "it", "does", "not", "change", "the", "result", "given", "by", "equation", "(", "25", ")", "about", "the", "optimal", "discriminator", ".", "As", "a", "consequence", ",", "for", "the", "optimal", "discriminator", "to", "retain", "the", "density", "information", ",", "it", "effectively", "means", "p", "gen", "=", "p", "data", ".", "Hence", ",", "there", "will", "be", "a", "contradiction", "if", "both", "c", "*", "(", "x", ")", "retains", "the", "density", "information", ",", "and", "the", "generator", "matches", "the", "data", "distribution", ".", "Intuitively", ",", "this", "problem", "roots", "in", "the", "fact", "that", "f", "-", "divergence", "is", "quite", "\"", "rigid", "\"", "in", "the", "sense", "that", "given", "the", "p", "gen", "(", "x", ")", "it", "only", "allows", "one", "fixed", "point", "for", "the", "discriminator", ".", "In", "comparison", ",", "the", "divergence", "used", "in", "our", "proposed", "formulation", ",", "which", "is", "the", "expected", "cost", "gap", ",", "is", "much", "more", "flexible", ".", "By", "the", "expected", "cost", "gap", "itself", ",", "i.e.", "without", "the", "K", "(", "p", "gen", ")", "term", ",", "the", "optimal", "discriminator", "is", "actually", "under", "-", "determined", ".", "section", ":", "B", "SUPPLEMENTARY", "MATERIALS", "FOR", "SECTION", "5", "B.1", "EXPERIMENT", "SETTING", "Here", ",", "we", "specify", "the", "neural", "architectures", "used", "for", "experiements", "presented", "in", "Section", "5", ".", "Firstly", ",", "for", "the", "Egan", "-", "Ent", "-", "VI", "model", ",", "we", "parameterize", "the", "approximate", "posterior", "distribution", "q", "gen", "(", "z", "|", "x", ")", "with", "a", "diagonal", "Gaussian", "distribution", ",", "whose", "mean", "and", "covariance", "matrix", "are", "the", "output", "of", "a", "trainable", "inference", "network", ",", "i.e.", "where", "f", "infer", "denotes", "the", "inference", "network", ",", "and", "I", "is", "the", "identity", "matrix", ".", "Note", "that", "the", "Inference", "Network", "only", "appears", "in", "the", "Egan", "-", "Ent", "-", "VI", "model", ".", "For", "experiments", "with", "the", "synthetic", "datasets", ",", "the", "following", "fully", "-", "connected", "feed", "forward", "neural", "networks", "are", "employed", "where", "FC", "and", "BN", "denote", "fully", "-", "connected", "layer", "and", "batch", "normalization", "layer", "respectively", ".", "Note", "that", "since", "the", "input", "noise", "to", "the", "generator", "has", "dimension", "4", ",", "the", "Inference", "Net", "output", "has", "dimension", "4", "*", "2", ",", "where", "the", "first", "4", "elements", "correspond", "the", "inferred", "mean", ",", "and", "the", "last", "4", "elements", "correspond", "to", "the", "inferred", "diagonal", "covariance", "matrix", "in", "log", "scale", ".", "For", "the", "handwritten", "digit", "experiment", ",", "we", "closely", "follow", "the", "DCGAN", "[", "reference", "]", "architecture", "with", "the", "following", "configuration", "Here", ",", "LRec", "is", "the", "leaky", "rectified", "non", "-", "linearity", "recommended", "by", "[", "reference", "]", ".", "In", "addition", ",", "CV", "(", "128", ",", "256", ",", "4c2s", ")", "denotes", "a", "convolutional", "layer", "with", "128", "input", "channels", ",", "256", "output", "channels", ",", "and", "kernel", "size", "4", "with", "stride", "2", ".", "Similarly", ",", "DC", "(", "256", ",", "128", ",", "4c2s", ")", "denotes", "a", "corresponding", "transposed", "convolutional", "operation", ".", "Compared", "to", "the", "original", "DCGAN", "architecture", ",", "the", "discriminator", "under", "our", "formulation", "does", "not", "have", "the", "last", "sigmoid", "layer", "which", "squashes", "a", "scalar", "value", "into", "a", "probability", "in", "[", "reference", "][", "reference", "]", ".", "For", "celebA", "experiment", "with", "64", "×", "64", "color", "images", ",", "we", "use", "the", "following", "architecture", "Given", "the", "chosen", "architectures", ",", "we", "follow", "[", "reference", "]", "and", "use", "Adam", "as", "the", "optimization", "algorithm", ".", "For", "more", "detailed", "hyper", "-", "parameters", ",", "please", "refer", "to", "the", "code", ".", "In", "order", "to", "quantify", "the", "quality", "of", "recovered", "distributions", ",", "we", "compute", "the", "pairwise", "KL", "divergence", "of", "the", "following", "four", "distributions", ":", "section", ":", "B.2", "QUANTITATIVE", "COMPARISON", "OF", "DIFFERENT", "MODELS", "•", "The", "real", "data", "distribution", "with", "analytic", "form", ",", "denoted", "as", "p", "data", "•", "The", "empirical", "data", "distribution", "approximated", "from", "the", "100", "K", "training", "data", ",", "denoted", "as", "p", "emp", "•", "The", "generator", "distribution", "approximated", "from", "100", "K", "generated", "data", ",", "denoted", "as", "p", "gen", "•", "The", "discriminator", "distribution", "re", "-", "normalized", "from", "the", "learned", "energy", ",", "denoted", "as", "p", "disc", "Since", "the", "synthetic", "datasets", "are", "two", "dimensional", ",", "we", "approximate", "both", "the", "empirical", "data", "distribution", "and", "the", "generator", "distribution", "using", "the", "simple", "histogram", "estimation", ".", "Specifically", ",", "we", "divide", "the", "canvas", "into", "a", "100", "-", "by", "-", "100", "grid", ",", "and", "assign", "each", "sample", "into", "its", "nearest", "grid", "cell", "based", "on", "euclidean", "distance", ".", "Then", ",", "we", "normalize", "the", "number", "of", "samples", "in", "each", "cell", "into", "a", "proper", "distribution", ".", "When", "recovering", "the", "discriminator", "distribution", "from", "the", "learned", "energy", ",", "we", "assume", "that", "µ", "*", "(", "x", ")", "=", "0", "(", "i.e.", "infinite", "data", "support", ")", ",", "and", "discretize", "the", "distribution", "into", "the", "same", "grid", "cells", ",", "∀x", "∈", "Grid", "Based", "on", "these", "approximation", ",", "Table", "2", "summarizes", "the", "results", ".", "For", "all", "measures", "related", "to", "the", "discriminator", "distribution", ",", "EGAN", "-", "Ent", "-", "VI", "and", "EGAN", "-", "Ent", "-", "NN", "significantly", "outperform", "the", "other", "two", "baseline", "models", ",", "which", "matches", "our", "visual", "assessment", "in", "Figure", "2", "and", "3", ".", "Meanwhile", ",", "the", "generator", "distributions", "learned", "from", "our", "proposed", "framework", "also", "achieve", "relatively", "lower", "divergence", "to", "both", "the", "empirical", "data", "distribution", "and", "the", "true", "data", "distribution", ".", "section", ":", "B.3", "COMPARISON", "OF", "THE", "ENTROPY", "(", "GRADIENT", ")", "APPROXIMATION", "METHODS", "In", "order", "to", "understand", "the", "performance", "difference", "between", "EGAN", "-", "Ent", "-", "VI", "and", "EGAN", "-", "Ent", "-", "NN", ",", "we", "analyze", "the", "quality", "of", "the", "entropy", "gradient", "approximation", "during", "training", ".", "To", "do", "that", ",", "we", "visualize", "some", "detailed", "training", "information", "in", "Figure", "8", ".", "Fig", ".", "(", "1", ",", "2", ")", ":", "frequency", "map", "of", "generated", "samples", "in", "the", "current", "batch", ".", "Fig", ".", "(", "1", ",", "3", ")", ":", "frequency", "map", "of", "real", "samples", "in", "the", "current", "batch", ".", "Fig", "-(", "1", ",", "4", ")", ":", "frequency", "difference", "between", "real", "and", "generated", "samples", ".", "Fig", ".", "(", "2", ",", "1", ")", "comparison", "between", "more", "generated", "from", "current", "model", "and", "real", "sample", ".", "Fig", ".", "(", "2", ",", "2", ")", ":", "the", "discriminator", "gradient", "w.r.t", ".", "each", "training", "sample", ".", "Fig", ".", "(", "2", ",", "3", ")", ":", "the", "entropy", "gradient", "w.r.t", ".", "each", "training", "samples", ".", "Fig", ".", "(", "2", ",", "4", ")", ":", "all", "gradient", "(", "discriminator", "+", "entropy", ")", "w.r.t", ".", "each", "training", "sample", ".", "As", "we", "can", "see", "in", "figure", "8a", ",", "the", "viarational", "entropy", "gradient", "approximation", "w.r.t", ".", "samples", "is", "not", "accurate", ":", "•", "It", "is", "inaccurate", "in", "terms", "of", "gradient", "direction", ".", "Ideally", ",", "the", "direction", "of", "the", "entropy", "gradient", "should", "be", "pointing", "from", "the", "center", "of", "its", "closest", "mode", "towards", "the", "surroundings", ",", "with", "the", "direction", "orthogonal", "to", "the", "implicit", "contour", "in", "Fig", ".", "(", "1", ",", "2", ")", ".", "However", ",", "the", "direction", "of", "gradients", "in", "the", "Fig", ".", "(", "2", ",", "3", ")", "does", "not", "match", "this", ".", "•", "It", "is", "inaccurate", "in", "magnitude", ".", "As", "we", "can", "see", ",", "the", "entropy", "approximation", "gradient", "(", "Fig", ".", "(", "2", ",", "3", ")", ")", "has", "much", "larger", "norm", "than", "the", "discriminator", "gradient", "(", "Fig", ".", "(", "2", ",", "2", ")", ")", ".", "As", "a", "result", ",", "the", "total", "gradient", "(", "Fig", ".", "(", "2", ",", "4", ")", ")", "is", "fully", "dominated", "by", "the", "entropy", "approximation", "gradient", ".", "Thus", ",", "it", "usually", "takes", "much", "longer", "for", "the", "generator", "to", "learn", "to", "generate", "rare", "samples", ",", "and", "the", "training", "also", "proceeds", "much", "slower", "compared", "to", "the", "nearest", "neighbor", "based", "approximation", ".", "In", "comparison", ",", "the", "nearest", "neighbor", "based", "gradient", "approximation", "is", "much", "more", "accurate", "as", "shown", "in", "8b", ".", "As", "a", "result", ",", "it", "leads", "to", "more", "accurate", "energy", "contour", ",", "as", "well", "as", "faster", "training", ".", "What", "'s", "more", ",", "from", "Figure", "8b", "Fig", ".", "(", "2", ",", "4", ")", ",", "we", "can", "see", "the", "entropy", "gradient", "does", "have", "the", "cancel", "-", "out", "effect", "on", "the", "discriminator", "gradient", ",", "which", "again", "matches", "our", "theory", ".", "B.4", "RANKING", "NIST", "DIGITS", "Figure", "9", "shows", "the", "ranking", "of", "all", "1000", "generated", "and", "real", "images", "(", "from", "the", "test", "set", ")", "for", "three", "models", ":", "EGAN", "-", "Ent", "-", "NN", ",", "EGAN", "-", "Const", ",", "and", "GAN", ".", "We", "can", "clearly", "notice", "that", "in", "EGAN", "-", "Ent", "-", "NN", "the", "topranked", "digits", "look", "very", "similar", "to", "the", "mean", "digit", ".", "From", "the", "upper", "-", "left", "corner", "to", "the", "lower", "-", "right", "corner", ",", "the", "transition", "trend", "is", ":", "the", "rotation", "degree", "increases", ",", "and", "the", "digits", "become", "increasingly", "thick", "or", "thin", "compared", "to", "the", "mean", ".", "In", "addition", ",", "samples", "in", "the", "last", "few", "rows", "do", "diverge", "away", "from", "the", "mean", "image", ":", "either", "highly", "diagonal", "to", "the", "right", "or", "left", ",", "or", "have", "different", "shape", ":", "very", "thin", "or", "thick", ",", "or", "typewriter", "script", ".", "Other", "models", "are", "not", "able", "to", "achieve", "a", "similar", "clear", "distinction", "for", "high", "versus", "low", "probability", "images", ".", "Finally", ",", "we", "consistently", "observe", "the", "same", "trend", "in", "modeling", "other", "digits", ",", "which", "are", "not", "shown", "in", "this", "paper", "due", "to", "space", "constraint", ".", "section", ":", "B.5", "CLASSIFIER", "PERFORMANCE", "AS", "A", "PROXY", "MEASURE", "As", "mentioned", "in", "Section", "5", ",", "evaluating", "the", "proposed", "formulation", "quantitatively", "on", "high", "-", "dimensional", "data", "is", "extremely", "challenging", ".", "Here", ",", "in", "order", "to", "provide", "more", "quantitative", "intuitions", "on", "the", "learned", "discriminator", "at", "convergence", ",", "we", "adopt", "a", "proxy", "measure", ".", "Specifically", ",", "we", "take", "the", "last", "-", "layer", "activation", "of", "the", "converged", "discriminator", "network", "as", "fixed", "pretrained", "feature", ",", "and", "build", "a", "linear", "classifier", "upon", "it", ".", "Hypothetically", ",", "if", "the", "discriminator", "does", "not", "degenerate", ",", "the", "extracted", "last", "-", "layer", "feature", "should", "maintain", "more", "information", "about", "the", "data", "points", ",", "especially", "compared", "to", "features", "from", "degenerated", "discriminators", ".", "Following", "this", "idea", ",", "we", "first", "train", "EGAN", "-", "Ent", "-", "NN", ",", "EGAN", "-", "Const", ",", "and", "GAN", "on", "the", "MNIST", "till", "convergence", ",", "and", "then", "extract", "the", "last", "-", "layer", "activation", "from", "their", "discriminator", "networks", "as", "fixed", "feature", "input", ".", "Based", "on", "fixed", "feature", ",", "a", "randomly", "initialized", "linear", "classifier", "is", "trained", "to", "do", "classification", "on", "MNIST", ".", "Based", "on", "10", "runs", "(", "with", "different", "initialization", ")", "of", "each", "of", "the", "three", "models", ",", "the", "test", "classification", "performance", "is", "summarized", "in", "Table", "3", ".", "For", "comparison", "purpose", ",", "we", "also", "include", "a", "baseline", "where", "the", "input", "features", "are", "extracted", "from", "a", "discriminator", "network", "with", "random", "weights", ".", "Table", "3", ":", "Test", "performance", "of", "linear", "classifiers", "based", "on", "last", "-", "layer", "discriminator", "features", ".", "Based", "on", "the", "proxy", "measure", ",", "EGAN", "-", "Ent", "-", "NN", "seems", "to", "maintain", "more", "information", "of", "data", ",", "which", "suggests", "that", "the", "discriminator", "from", "our", "proposed", "formulation", "is", "more", "informative", ".", "Despite", "the", "positive", "result", ",", "it", "is", "important", "to", "point", "out", "that", "maintaining", "information", "about", "categories", "does", "not", "necessarily", "mean", "maintaining", "information", "about", "the", "energy", "(", "density", ")", ".", "Thus", ",", "this", "proxy", "measure", "should", "be", "understood", "cautiously", ".", "section", ":", "section", ":", "ACKNOWLEDGEMENTS", "We", "would", "like", "to", "thank", "the", "developers", "of", "Theano", "(", "Theano", "Development", "Team", ",", "2016", ")", "for", "developing", "such", "a", "powerful", "tool", "for", "scientific", "computing", ".", "Amjad", "Almahairi", "was", "supported", "by", "funding", "from", "Maluuba", "Research", ".", "section", ":"], "entity": [[0, 9, ["CALIBRATING", "ENERGY", "-", "BASED", "GENERATIVE", "ADVER", "-", "SARIAL", "NETWORKS"], "Method"], [18, 21, ["Generative", "Adversarial", "Networks"], "Method"], [38, 41, ["adversarial", "training", "framework"], "Method"], [50, 51, ["generator"], "Method"], [62, 63, ["discriminator"], "Method"], [80, 82, ["induced", "solution"], "Method"], [103, 105, ["approximation", "techniques"], "Method"], [120, 121, ["discriminator"], "Method"], [134, 137, ["Generative", "Adversarial", "Networks"], "Method"], [138, 139, ["GANs"], "Method"], [153, 155, ["generative", "models"], "Method"], [156, 157, ["GANs"], "Method"], [158, 161, ["generative", "model", "training"], "Method"], [163, 165, ["minimax", "game"], "Method"], [167, 169, ["generative", "network"], "Method"], [170, 171, ["generator"], "Method"], [185, 187, ["discriminative", "network"], "Method"], [188, 189, ["discriminator"], "Method"], [223, 224, ["discriminator"], "Method"], [228, 229, ["generator"], "Method"], [264, 265, ["generator"], "Method"], [274, 275, ["discriminator"], "Method"], [278, 281, ["degenerate", "uniform", "solution"], "Method"], [283, 284, ["GANs"], "Method"], [288, 290, ["sample", "generators"], "Method"], [294, 299, ["general", "purpose", "probabilistic", "generative", "models"], "Method"], [310, 312, ["density", "estimates"], "Method"], [318, 320, ["sample", "evaluation"], "Task"], [345, 349, ["trainable", "sample", "evaluation", "mechanism"], "Method"], [367, 368, ["generator"], "Method"], [401, 403, ["evaluation", "metric"], "Metric"], [405, 409, ["data", "-", "driven", "evaluators"], "Method"], [421, 423, ["evaluation", "criteria"], "Metric"], [433, 437, ["real", "-", "valued", "discriminator"], "Method"], [441, 445, ["energy", "-", "based", "GANs"], "Method"], [500, 502, ["adversarial", "framework"], "Method"], [507, 509, ["degenerate", "discriminator"], "Method"], [522, 523, ["discriminator"], "Method"], [529, 532, ["adversarial", "learning", "formulation"], "Method"], [537, 541, ["non", "-", "degenerate", "discriminator"], "Method"], [559, 561, ["analytic", "form"], "Method"], [563, 565, ["optimal", "discriminator"], "Method"], [577, 578, ["GANs"], "Method"], [587, 589, ["nonparametric", "case"], "Task"], [611, 613, ["parametric", "case"], "Task"], [741, 743, ["generator", "distribution"], "Method"], [771, 772, ["discriminator"], "Method"], [793, 795, ["generator", "distribution"], "Method"], [834, 838, ["optimization", "dynamics", "near", "convergence"], "Method"], [839, 840, ["GANs"], "Method"], [844, 846, ["generator", "distribution"], "Method"], [860, 861, ["discriminator"], "Method"], [927, 930, ["generator", "'s", "training"], "Method"], [934, 936, ["generator", "distribution"], "Method"], [940, 942, ["data", "distribution"], "Method"], [958, 960, ["generator", "distribution"], "Method"], [983, 984, ["discriminator"], "Method"], [1011, 1012, ["generator"], "Method"], [1023, 1024, ["discriminator"], "Method"], [1050, 1052, ["degenerate", "discriminator"], "Method"], [1074, 1077, ["adversarial", "learning", "formulation"], "Method"], [1134, 1135, ["generator"], "Method"], [1159, 1162, ["minimax", "training", "objective"], "Task"], [1174, 1175, ["discriminator"], "Method"], [1186, 1190, ["K", "(", "p", "gen"], "Method"], [1212, 1213, ["GAN"], "Method"], [1223, 1224, ["fomulation"], "Method"], [1246, 1248, ["binary", "classification"], "Method"], [1289, 1293, ["discriminator", "c", "(", "x"], "Method"], [1299, 1302, ["expected", "cost", "gap"], "Metric"], [1325, 1327, ["adversarial", "mechanism"], "Method"], [1329, 1335, ["calibrating", "term", "K", "(", "p", "gen"], "Method"], [1386, 1388, ["optimal", "discriminator"], "Method"], [1397, 1399, ["optimization", "objective"], "Metric"], [1408, 1409, ["generator"], "Method"], [1411, 1412, ["discriminator"], "Method"], [1425, 1427, ["optimization", "problem"], "Task"], [1467, 1471, ["adversarial", "imitation", "learning", "setting"], "Task"], [1487, 1489, ["induced", "policy"], "Method"], [1498, 1499, ["discriminator"], "Method"], [1573, 1576, ["convex", "optimization", "problem"], "Task"], [1637, 1639, ["strong", "duality"], "Method"], [1646, 1648, ["optimal", "solution"], "Task"], [1652, 1657, ["Karush", "-", "Kuhn", "-", "Tucker"], "Method"], [1678, 1680, ["convex", "problem"], "Task"], [1691, 1693, ["generator", "distribution"], "Method"], [1707, 1708, ["discriminator"], "Method"], [1771, 1773, ["optimal", "discriminator"], "Method"], [1835, 1838, ["Weak", "Support", "Discriminator"], "Method"], [1843, 1846, ["optimal", "discriminator", "function"], "Method"], [1987, 1990, ["weak", "support", "discriminator"], "Method"], [2042, 2044, ["parametric", "case"], "Task"], [2055, 2057, ["parametric", "model"], "Method"], [2060, 2061, ["discriminator"], "Method"], [2068, 2070, ["Global", "Bias"], "Metric"], [2112, 2114, ["discriminator", "function"], "Method"], [2136, 2138, ["optimal", "discriminator"], "Method"], [2163, 2165, ["generator", "distribution"], "Method"], [2236, 2238, ["data", "distribution"], "Method"], [2257, 2262, ["max", "-", "entropy", "imitation", "learning"], "Method"], [2267, 2274, ["directed", "-", "generator", "-", "trained", "energybased", "model"], "Method"], [2290, 2294, ["maximum", "-", "likelihood", "estimation"], "Method"], [2298, 2300, ["minimax", "optimization"], "Task"], [2310, 2312, ["minimax", "formulation"], "Method"], [2320, 2322, ["induced", "solution"], "Task"], [2342, 2343, ["discriminator"], "Method"], [2359, 2361, ["support", "discriminator"], "Method"], [2447, 2449, ["entropy", "solution"], "Method"], [2563, 2564, ["pgen"], "Method"], [2580, 2582, ["discriminative", "power"], "Metric"], [2587, 2590, ["weak", "support", "discriminator"], "Method"], [2731, 2733, ["EBGAN", "objective"], "Method"], [2736, 2742, ["single", "-", "training", "-", "signal", "problem"], "Task"], [2748, 2749, ["discriminator"], "Method"], [2802, 2803, ["generator"], "Method"], [2811, 2812, ["discriminator"], "Method"], [2828, 2830, ["PARAMETRIC", "INSTANTIATION"], "Method"], [2831, 2833, ["ENTROPY", "APPROXIMATION"], "Method"], [2842, 2846, ["non", "-", "parametric", "case"], "Task"], [2874, 2876, ["parametric", "representations"], "Method"], [2879, 2880, ["generator"], "Method"], [2882, 2883, ["discriminator"], "Method"], [2889, 2890, ["generator"], "Method"], [2892, 2895, ["back", "-", "propagation"], "Method"], [2901, 2903, ["generator", "distribution"], "Method"], [2910, 2913, ["directed", "generator", "network"], "Method"], [2944, 2946, ["generator", "distribution"], "Method"], [2953, 2958, ["generator", "network", "'s", "noise→data", "mapping"], "Method"], [2971, 2973, ["training", "objective"], "Task"], [2990, 2994, ["max", "-", "entropy", "formulation"], "Method"], [3002, 3008, ["regularization", "term", "K", "(", "p", "gen"], "Method"], [3018, 3024, ["NEAREST", "-", "NEIGHBOR", "ENTROPY", "GRADIENT", "APPROXIMATION"], "Method"], [3036, 3038, ["entropy", "gradient"], "Method"], [3044, 3046, ["p", "gen"], "Method"], [3049, 3055, ["deterministic", ",", "differentiable", "transform", "g", "θ"], "Method"], [3069, 3075, ["gradient", "of", "H", "(", "p", "gen"], "Method"], [3111, 3117, ["generator", "log", "-", "density", "log", "p"], "Method"], [3138, 3146, ["Monte", "-", "Carlo", "estimation", "of", "the", "entropy", "gradient"], "Method"], [3147, 3152, ["θ", "H", "(", "p", "gen"], "Method"], [3157, 3158, ["generator"], "Method"], [3204, 3207, ["local", "Gaussian", "approximation"], "Method"], [3234, 3235, ["generator"], "Method"], [3251, 3253, ["Gaussian", "approximation"], "Method"], [3257, 3260, ["local", "Gaussian", "approximation"], "Method"], [3291, 3294, ["isotropic", "Gaussian", "distribution"], "Method"], [3306, 3309, ["isotropic", "Gaussian", "approximation"], "Method"], [3334, 3336, ["gradient", "approximation"], "Method"], [3374, 3377, ["entropy", "gradient", "approximation"], "Method"], [3430, 3432, ["effective", "dimension"], "Metric"], [3452, 3455, ["VARIATIONAL", "LOWER", "BOUND"], "Metric"], [3468, 3471, ["variational", "lower", "bound"], "Metric"], [3473, 3478, ["entropy", "H", "(", "p", "gen"], "Metric"], [3484, 3486, ["generator", "distribution"], "Method"], [3599, 3601, ["peaked", "Gaussian"], "Method"], [3611, 3613, ["conditional", "entropy"], "Metric"], [3655, 3657, ["conditional", "entropy"], "Metric"], [3702, 3705, ["variational", "upper", "bound"], "Method"], [3708, 3712, ["approximate", "posterior", "q", "gen"], "Method"], [3735, 3738, ["back", "-", "propagation"], "Method"], [3739, 3742, ["Monte", "Carlo", "integration"], "Method"], [3787, 3790, ["variational", "lower", "bound"], "Metric"], [3829, 3830, ["discriminator"], "Method"], [3833, 3838, ["entropy", "-", "regularized", "adversarial", "training"], "Method"], [3895, 3899, ["fully", "-", "connected", "models"], "Method"], [3907, 3908, ["Gaussians"], "Method"], [3923, 3927, ["Mixture", "of", "2", "Gaussians"], "Method"], [3985, 3987, ["learned", "generator"], "Method"], [3994, 3995, ["discriminator"], "Method"], [3999, 4000, ["evaluation"], "Task"], [4005, 4009, ["EGAN", "-", "Ent", "model"], "Method"], [4014, 4019, ["nearest", "-", "neighbor", "based", "approximation"], "Method"], [4020, 4025, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [4028, 4033, ["variational", "-", "inference", "based", "approximation"], "Method"], [4034, 4039, ["EGAN", "-", "Ent", "-", "VI"], "Method"], [4050, 4051, ["GAN"], "Method"], [4057, 4059, ["no", "regularization"], "Method"], [4060, 4063, ["EGAN", "-", "Const"], "Method"], [4101, 4102, ["discriminator"], "Method"], [4126, 4131, ["EGAN", "-", "Ent", "-", "VI"], "Method"], [4132, 4137, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [4152, 4156, ["equally", "weighted", "Gaussian", "mixture"], "Method"], [4164, 4169, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [4181, 4186, ["EGAN", "-", "Ent", "-", "VI"], "Method"], [4197, 4202, ["EGAN", "-", "Ent", "-", "VI"], "Method"], [4233, 4236, ["EGANEnt", "-", "NN"], "Method"], [4281, 4286, ["EGAN", "-", "Ent", "-", "VI"], "Method"], [4287, 4292, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [4293, 4296, ["biased", "Gaussian", "mixture"], "Method"], [4301, 4305, ["variational", "inference", "based", "approximation"], "Method"], [4331, 4334, ["RANKING", "NIST", "DIGITS"], "Task"], [4411, 4413, ["baseline", "models"], "Method"], [4414, 4415, ["discriminator"], "Method"], [4452, 4455, ["variational", "inference", "approximation"], "Method"], [4456, 4461, ["EGAN", "-", "Ent", "-", "VI"], "Method"], [4470, 4473, ["nearest", "neighbor", "approximation"], "Method"], [4474, 4479, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [4491, 4492, ["discriminator"], "Method"], [4527, 4532, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [4534, 4537, ["EGAN", "-", "Const"], "Method"], [4538, 4539, ["GAN"], "Method"], [4540, 4541, ["ranking"], "Task"], [4638, 4643, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [4706, 4707, ["ranking"], "Task"], [4716, 4718, ["SAMPLE", "QUALITY"], "Metric"], [4732, 4736, ["visual", "quality", "of", "samples"], "Metric"], [4748, 4751, ["CIFAR", "-", "10"], "Material"], [4757, 4759, ["variationalbased", "approximation"], "Method"], [4760, 4762, ["entropy", "regularization"], "Task"], [4779, 4784, ["EGAN", "-", "Ent", "-", "VI"], "Method"], [4796, 4798, ["variational", "approximation"], "Method"], [4821, 4824, ["CIFAR", "-", "10"], "Material"], [4826, 4828, ["Inception", "score"], "Metric"], [4842, 4847, ["EGAN", "-", "Ent", "-", "VI"], "Method"], [4864, 4868, ["EGAN", "-", "Const", "model"], "Method"], [4890, 4894, ["energy", "-", "based", "models"], "Method"], [4912, 4915, ["EGAN", "-", "Const"], "Method"], [4919, 4921, ["entropy", "regularization"], "Method"], [4940, 4943, ["adversarial", "learning", "approaches"], "Method"], [4960, 4963, ["adversarial", "learning", "formulation"], "Method"], [4967, 4969, ["discriminator", "function"], "Method"], [4983, 4985, ["learned", "discriminator"], "Method"], [4987, 4991, ["non", "-", "parametric", "setting"], "Task"], [5002, 5004, ["parametric", "setting"], "Task"], [5071, 5073, ["Inception", "scores"], "Metric"], [5074, 5077, ["CIFAR", "-", "10"], "Material"], [5181, 5182, ["EBGAN"], "Method"], [5192, 5193, ["generator"], "Method"], [5195, 5196, ["discriminator"], "Method"], [5203, 5206, ["minimax", "optimization", "problem"], "Task"], [5224, 5225, ["discriminator"], "Method"], [5230, 5234, ["mean", "squared", "reconstruction", "error"], "Metric"], [5236, 5240, ["auto", "-", "encoder", "structure"], "Method"], [5278, 5282, ["optimal", "generator", "and", "discriminator"], "Method"], [5297, 5298, ["GAN"], "Method"], [5307, 5309, ["generator", "distribution"], "Method"], [5314, 5316, ["discriminator", "function"], "Method"], [5318, 5320, ["generator", "distribution"], "Method"], [5342, 5346, ["conditional", "optimal", "discriminator", "function"], "Method"], [5350, 5352, ["training", "objective"], "Metric"], [5353, 5355, ["p", "gen"], "Method"], [5359, 5362, ["adversarial", "\"", "training"], "Task"], [5364, 5366, ["minimizing", "problem"], "Task"], [5382, 5386, ["generator", "distribution", "p", "gen"], "Method"], [5388, 5391, ["EBGAN", "training", "objective"], "Metric"], [5393, 5394, ["discriminator"], "Method"], [5438, 5442, ["mean", "squared", "reconstruction", "error"], "Metric"], [5532, 5534, ["training", "objective"], "Metric"], [5539, 5541, ["minimization", "problem"], "Task"], [5562, 5564, ["adversarial", "game"], "Method"], [5577, 5580, ["EBGAN", "training", "objective"], "Task"], [5825, 5829, ["ADDING", "ADDITIONAL", "TRAINING", "SIGNAL"], "Task"], [5842, 5843, ["GAN"], "Method"], [5988, 5989, ["discriminator"], "Method"], [6004, 6006, ["optimal", "discriminator"], "Method"], [6014, 6015, ["discriminator"], "Method"], [6051, 6052, ["generator"], "Method"], [6066, 6069, ["f", "-", "divergence"], "Method"], [6093, 6094, ["discriminator"], "Method"], [6120, 6123, ["expected", "cost", "gap"], "Metric"], [6137, 6138, ["discriminator"], "Method"], [6160, 6162, ["neural", "architectures"], "Method"], [6174, 6180, ["Egan", "-", "Ent", "-", "VI", "model"], "Method"], [6184, 6189, ["approximate", "posterior", "distribution", "q", "gen"], "Method"], [6196, 6199, ["diagonal", "Gaussian", "distribution"], "Method"], [6210, 6213, ["trainable", "inference", "network"], "Method"], [6220, 6222, ["inference", "network"], "Method"], [6233, 6235, ["Inference", "Network"], "Method"], [6239, 6245, ["Egan", "-", "Ent", "-", "VI", "model"], "Method"], [6255, 6262, ["fully", "-", "connected", "feed", "forward", "neural", "networks"], "Method"], [6265, 6266, ["FC"], "Method"], [6267, 6268, ["BN"], "Method"], [6269, 6273, ["fully", "-", "connected", "layer"], "Method"], [6274, 6277, ["batch", "normalization", "layer"], "Method"], [6287, 6288, ["generator"], "Method"], [6330, 6333, ["handwritten", "digit", "experiment"], "Task"], [6338, 6339, ["DCGAN"], "Method"], [6349, 6350, ["LRec"], "Method"], [6366, 6374, ["CV", "(", "128", ",", "256", ",", "4c2s", ")"], "Method"], [6376, 6378, ["convolutional", "layer"], "Method"], [6397, 6405, ["DC", "(", "256", ",", "128", ",", "4c2s", ")"], "Method"], [6408, 6411, ["transposed", "convolutional", "operation"], "Method"], [6416, 6418, ["DCGAN", "architecture"], "Method"], [6420, 6421, ["discriminator"], "Method"], [6429, 6431, ["sigmoid", "layer"], "Method"], [6447, 6448, ["celebA"], "Task"], [6473, 6474, ["Adam"], "Method"], [6476, 6478, ["optimization", "algorithm"], "Method"], [6505, 6508, ["pairwise", "KL", "divergence"], "Metric"], [6554, 6556, ["generator", "distribution"], "Method"], [6595, 6598, ["empirical", "data", "distribution"], "Method"], [6600, 6602, ["generator", "distribution"], "Method"], [6605, 6607, ["histogram", "estimation"], "Method"], [6713, 6718, ["EGAN", "-", "Ent", "-", "VI"], "Method"], [6719, 6724, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [6746, 6748, ["generator", "distributions"], "Method"], [6757, 6758, ["divergence"], "Metric"], [6776, 6782, ["ENTROPY", "(", "GRADIENT", ")", "APPROXIMATION", "METHODS"], "Method"], [6790, 6795, ["EGAN", "-", "Ent", "-", "VI"], "Method"], [6796, 6801, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [6808, 6811, ["entropy", "gradient", "approximation"], "Method"], [6942, 6945, ["discriminator", "+", "entropy"], "Method"], [6961, 6966, ["viarational", "entropy", "gradient", "approximation", "w.r.t"], "Method"], [7118, 7119, ["generator"], "Method"], [7136, 7140, ["nearest", "neighbor", "based", "approximation"], "Method"], [7145, 7150, ["nearest", "neighbor", "based", "gradient", "approximation"], "Method"], [7175, 7176, ["training"], "Task"], [7217, 7220, ["RANKING", "NIST", "DIGITS"], "Metric"], [7242, 7247, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [7248, 7251, ["EGAN", "-", "Const"], "Method"], [7253, 7254, ["GAN"], "Method"], [7261, 7266, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [7399, 7401, ["CLASSIFIER", "PERFORMANCE"], "Metric"], [7403, 7405, ["PROXY", "MEASURE"], "Metric"], [7437, 7438, ["discriminator"], "Method"], [7444, 7446, ["proxy", "measure"], "Metric"], [7458, 7461, ["converged", "discriminator", "network"], "Method"], [7469, 7471, ["linear", "classifier"], "Method"], [7478, 7479, ["discriminator"], "Method"], [7503, 7505, ["degenerated", "discriminators"], "Method"], [7513, 7518, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [7519, 7522, ["EGAN", "-", "Const"], "Method"], [7524, 7525, ["GAN"], "Method"], [7527, 7528, ["MNIST"], "Method"], [7541, 7543, ["discriminator", "networks"], "Method"], [7554, 7558, ["randomly", "initialized", "linear", "classifier"], "Method"], [7562, 7563, ["classification"], "Task"], [7564, 7565, ["MNIST"], "Task"], [7583, 7586, ["test", "classification", "performance"], "Metric"], [7609, 7611, ["discriminator", "network"], "Method"], [7621, 7623, ["linear", "classifiers"], "Method"], [7625, 7630, ["last", "-", "layer", "discriminator", "features"], "Method"], [7634, 7636, ["proxy", "measure"], "Metric"], [7637, 7642, ["EGAN", "-", "Ent", "-", "NN"], "Method"], [7695, 7697, ["proxy", "measure"], "Metric"], [7730, 7732, ["scientific", "computing"], "Task"], [239, 240, ["GAN"], "Method"], [327, 328, ["GAN"], "Method"], [353, 354, ["GAN"], "Method"], [469, 470, ["GAN"], "Method"], [480, 481, ["GAN"], "Method"], [1233, 1234, ["GAN"], "Method"], [2794, 2795, ["GAN"], "Method"], [4055, 4056, ["GAN"], "Method"], [4104, 4105, ["GAN"], "Method"], [4450, 4451, ["GAN"], "Method"], [4468, 4469, ["GAN"], "Method"], [4850, 4851, ["GAN"], "Method"], [4899, 4900, ["GAN"], "Method"], [5050, 5051, ["GAN"], "Method"], [5830, 5831, ["GAN"], "Method"], [5864, 5865, ["GAN"], "Method"], [5875, 5876, ["GAN"], "Method"], [5888, 5889, ["GAN"], "Method"]]} {"doc_id": "0680f04750b1e257ffdd161e85382031dc73ea7f", "text": ["document", ":", "End", "-", "to", "-", "End", "Answer", "Chunk", "Extraction", "and", "Ranking", "for", "Reading", "Comprehension", "This", "paper", "proposes", "dynamic", "chunk", "reader", "(", "DCR", ")", ",", "an", "end", "-", "to", "-", "end", "neural", "reading", "comprehension", "(", "RC", ")", "model", "that", "is", "able", "to", "extract", "and", "rank", "a", "set", "of", "answer", "candidates", "from", "a", "given", "document", "to", "answer", "questions", ".", "DCR", "is", "able", "to", "predict", "answers", "of", "variable", "lengths", ",", "whereas", "previous", "neural", "RC", "models", "primarily", "focused", "on", "predicting", "single", "tokens", "or", "entities", ".", "DCR", "encodes", "a", "document", "and", "an", "input", "question", "with", "recurrent", "neural", "networks", ",", "and", "then", "applies", "a", "word", "-", "by", "-", "word", "attention", "mechanism", "to", "acquire", "question", "-", "aware", "representations", "for", "the", "document", ",", "followed", "by", "the", "generation", "of", "chunk", "representations", "and", "a", "ranking", "module", "to", "propose", "the", "top", "-", "ranked", "chunk", "as", "the", "answer", ".", "Experimental", "results", "show", "that", "DCR", "achieves", "state", "-", "of", "-", "the", "-", "art", "exact", "match", "and", "F1", "scores", "on", "the", "SQuAD", "dataset", ".", "section", ":", "Introduction", "Reading", "comprehension", "-", "based", "question", "answering", "(", "RCQA", ")", "is", "the", "task", "of", "answering", "a", "question", "with", "a", "chunk", "of", "text", "taken", "from", "related", "document", "(", "s", ")", ".", "A", "variety", "of", "neural", "models", "have", "been", "proposed", "recently", "either", "for", "extracting", "a", "single", "entity", "or", "a", "single", "token", "as", "an", "answer", "from", "a", "given", "text", ";", "or", "for", "selecting", "the", "correct", "answer", "by", "ranking", "a", "small", "set", "of", "human", "-", "provided", "candidates", ".", "In", "both", "cases", ",", "an", "answer", "boundary", "is", "either", "easy", "to", "determine", "or", "already", "given", ".", "Different", "from", "the", "above", "two", "assumptions", "for", "RCQA", ",", "in", "the", "real", "-", "world", "QA", "scenario", ",", "people", "may", "ask", "questions", "about", "both", "entities", "(", "factoid", ")", "and", "non", "-", "entities", "such", "as", "explanations", "and", "reasons", "(", "non", "-", "factoid", ")", "(", "see", "Table", "[", "reference", "]", "for", "examples", ")", ".", "In", "this", "regard", ",", "RCQA", "has", "the", "potential", "to", "complement", "other", "QA", "approaches", "that", "leverage", "structured", "data", "(", "e.g.", ",", "knowledge", "bases", ")", "for", "both", "the", "above", "question", "types", ".", "This", "is", "because", "RCQA", "can", "exploit", "the", "textual", "evidences", "to", "ensure", "increased", "answer", "coverage", ",", "which", "is", "particularly", "helpful", "for", "non", "-", "factoid", "answers", ".", "However", ",", "it", "is", "also", "challenging", "for", "RCQA", "to", "identify", "answer", "in", "arbitrary", "position", "in", "the", "passage", "with", "arbitrary", "length", ",", "especially", "for", "non", "-", "factoid", "answers", "which", "might", "be", "clauses", "or", "sentences", ".", "As", "a", "result", ",", "apart", "from", "a", "few", "exceptions", ",", "this", "research", "direction", "has", "not", "been", "fully", "explored", "yet", ".", "Compared", "to", "the", "relatively", "easier", "RC", "task", "of", "predicting", "single", "tokens", "/", "entities", ",", "predicting", "answers", "of", "arbitrary", "lengths", "and", "positions", "significantly", "increase", "the", "search", "space", "complexity", ":", "the", "number", "of", "possible", "candidates", "to", "consider", "is", "in", "the", "order", "of", ",", "where", "is", "the", "number", "of", "passage", "words", ".", "In", "contrast", ",", "for", "previous", "works", "in", "which", "answers", "are", "single", "tokens", "/", "entities", "or", "from", "candidate", "lists", ",", "the", "complexity", "is", "in", "or", "the", "size", "of", "candidate", "lists", "(", "usually", "5", ")", ",", "respectively", ".", "To", "address", "the", "above", "complexity", ",", "Rajpurkar", "et", "al", ".", "rajpurkar2016squad", "used", "a", "two", "-", "step", "chunk", "-", "and", "-", "rank", "approach", "that", "employs", "a", "rule", "-", "based", "algorithm", "to", "extract", "answer", "candidates", "from", "a", "passage", ",", "followed", "by", "a", "ranking", "approach", "with", "hand", "-", "crafted", "features", "to", "select", "the", "best", "answer", ".", "The", "rule", "-", "based", "chunking", "approach", "suffered", "from", "low", "coverage", "(", "70", "%", "recall", "of", "answer", "chunks", ")", "that", "can", "not", "be", "improved", "during", "training", ";", "and", "candidate", "ranking", "performance", "depends", "greatly", "on", "the", "quality", "of", "the", "hand", "-", "crafted", "features", ".", "More", "recently", ",", "Wang", "and", "Jiang", "wang2016machine", "proposed", "two", "end", "-", "to", "-", "end", "neural", "network", "models", ",", "one", "of", "which", "chunks", "a", "candidate", "answer", "by", "predicting", "the", "answer", "’s", "two", "boundary", "indices", "and", "the", "other", "classifies", "each", "passage", "word", "into", "answer", "/", "not", "-", "answer", ".", "Both", "models", "improved", "significantly", "over", "the", "method", "proposed", "by", "Rajpurkar", "et", "al", ".", "rajpurkar2016squad", ".", "Our", "proposed", "model", ",", "called", "dynamic", "chunk", "reader", "(", "DCR", ")", ",", "not", "only", "significantly", "differs", "from", "both", "the", "above", "systems", "in", "the", "way", "that", "answer", "candidates", "are", "generated", "and", "ranked", ",", "but", "also", "shares", "merits", "with", "both", "works", ".", "First", ",", "our", "model", "uses", "deep", "networks", "to", "learn", "better", "representations", "for", "candidate", "answer", "chunks", ",", "instead", "of", "using", "fixed", "feature", "representations", "as", "in", ".", "Second", ",", "it", "represents", "answer", "candidates", "as", "chunks", ",", "as", "in", ",", "instead", "of", "word", "-", "level", "representations", ",", "to", "make", "the", "model", "aware", "of", "the", "subtle", "differences", "among", "candidates", "(", "importantly", ",", "overlapping", "candidates", ")", ".", "The", "contributions", "of", "this", "paper", "are", "three", "-", "fold", ".", "(", "1", ")", "We", "propose", "a", "novel", "neural", "network", "model", "for", "joint", "candidate", "answer", "chunking", "and", "ranking", ",", "where", "the", "candidate", "answer", "chunks", "are", "dynamically", "constructed", "and", "ranked", "in", "an", "end", "-", "to", "-", "end", "manner", ".", "(", "2", ")", "we", "propose", "a", "new", "question", "-", "attention", "mechanism", "to", "enhance", "passage", "word", "representation", ",", "which", "is", "subsequently", "used", "to", "construct", "chunk", "representations", ".", "(", "3", ")", "We", "also", "propose", "several", "simple", "but", "effective", "features", "to", "strengthen", "the", "attention", "mechanism", ",", "which", "fundamentally", "improves", "candidate", "ranking", ",", "with", "the", "by", "-", "product", "of", "higher", "exact", "boundary", "match", "accuracy", ".", "The", "experiments", "on", "the", "Stanford", "Question", "Answering", "Dataset", "(", "SQuAD", ")", ",", "which", "contains", "a", "variety", "of", "human", "-", "generated", "factoid", "and", "non", "-", "factoid", "questions", ",", "have", "shown", "the", "effectiveness", "of", "above", "three", "contributions", ".", "Our", "paper", "is", "organized", "as", "follows", ".", "We", "formally", "define", "the", "RCQA", "problem", "first", ".", "Next", ",", "we", "describe", "our", "baseline", "with", "a", "neural", "network", "component", ".", "We", "present", "the", "end", "-", "to", "-", "end", "dynamic", "chunk", "reader", "model", "next", ".", "Finally", ",", "we", "analyze", "our", "experimental", "results", "and", "discuss", "the", "related", "work", ".", "section", ":", "Problem", "Definition", "Table", "[", "reference", "]", "shows", "an", "example", "of", "our", "RC", "setting", "where", "the", "goal", "is", "to", "answer", "a", "question", ",", "factoid", "(", "Q1", ")", "or", "non", "-", "factoid", "(", "Q2", "and", "Q3", ")", ",", "based", "on", "a", "supporting", "passage", ",", "by", "selecting", "a", "continuous", "sequence", "of", "text", "as", "answer", ".", ",", ",", "and", "are", "all", "word", "sequences", ",", "where", "each", "word", "is", "drawn", "from", "a", "vocabulary", ",", ".", "The", "-", "th", "instance", "in", "the", "training", "set", "is", "a", "triple", "in", "the", "form", "of", ",", "where", ",", ",", "and", "(", ")", ".", "Owing", "to", "the", "disagreement", "among", "annotators", ",", "there", "could", "be", "more", "than", "one", "correct", "answer", "for", "the", "same", "question", ";", "and", "the", "-", "th", "answer", "to", "is", "denoted", "by", ".", "An", "answer", "candidate", "for", "the", "-", "th", "training", "example", "is", "defined", "as", ",", "a", "sub", "-", "sequence", "in", ",", "that", "spans", "from", "position", "to", "(", ")", ".", "The", "ground", "truth", "answer", "could", "be", "included", "in", "the", "set", "of", "all", "candidates", ",", "where", "is", "the", "constraint", "put", "on", "the", "candidate", "chunk", "for", ",", "such", "as", ",", "“", "can", "have", "at", "most", "10", "tokens", "”", ",", "or", "“", "must", "have", "a", "pre", "-", "defined", "POS", "pattern", "”", ".", "To", "evaluate", "a", "system", "’s", "performance", ",", "its", "top", "answer", "to", "a", "question", "is", "matched", "against", "the", "corresponding", "gold", "standard", "answer", "(", "s", ")", ".", "paragraph", ":", "Remark", ":", "Categories", "of", "RC", "Tasks", "Other", "simpler", "variants", "of", "the", "aforementioned", "RC", "task", "were", "explored", "in", "the", "past", ".", "For", "example", ",", "quiz", "-", "style", "datasets", "(", "e.g.", ",", "MCTest", ",", "MovieQA", ")", "have", "multiple", "-", "choice", "questions", "with", "answer", "options", ".", "Cloze", "-", "style", "datesets", ",", "usually", "automatically", "generated", ",", "have", "factoid", "“", "question”s", "created", "by", "replacing", "the", "answer", "in", "a", "sentence", "from", "the", "text", "with", "blank", ".", "For", "the", "answer", "selection", "task", "this", "paper", "focuses", "on", ",", "several", "datasets", "exist", ",", "e.g.", "TREC", "-", "QA", "for", "factoid", "answer", "extraction", "from", "multiple", "given", "passages", ",", "bAbI", "designed", "for", "inference", "purpose", ",", "and", "the", "SQuAD", "dataset", "used", "in", "this", "paper", ".", "To", "the", "best", "of", "our", "knowledge", ",", "the", "SQuAD", "dataset", "is", "the", "only", "one", "for", "both", "factoid", "and", "non", "-", "factoid", "answer", "extraction", "with", "a", "question", "distribution", "more", "close", "to", "real", "-", "world", "applications", ".", "section", ":", "Baseline", ":", "Chunk", "-", "and", "-", "Rank", "Pipeline", "with", "Neural", "RC", "In", "this", "section", "we", "modified", "a", "state", "-", "of", "-", "the", "-", "art", "RC", "system", "for", "cloze", "-", "style", "tasks", "for", "our", "answer", "extraction", "purpose", ",", "to", "see", "how", "much", "gap", "we", "have", "for", "the", "two", "type", "of", "tasks", ",", "and", "to", "inspire", "our", "end", "-", "to", "-", "end", "system", "in", "the", "next", "section", ".", "In", "order", "to", "make", "the", "cloze", "-", "style", "RC", "system", "to", "make", "chunk", "-", "level", "decision", ",", "we", "use", "the", "RC", "model", "to", "generate", "features", "for", "chunks", ",", "which", "are", "further", "used", "in", "a", "feature", "-", "based", "ranker", "like", "in", ".", "As", "a", "result", ",", "this", "baseline", "can", "be", "viewed", "as", "a", "deep", "learning", "based", "counterpart", "of", "the", "system", "in", ".", "It", "has", "two", "main", "components", ":", "1", ")", "a", "stand", "-", "alone", "answer", "chunker", ",", "which", "is", "trained", "to", "produce", "overlapping", "candidate", "chunks", ",", "and", "2", ")", "a", "neural", "RC", "model", ",", "which", "is", "used", "to", "score", "each", "word", "in", "a", "given", "passage", "to", "be", "used", "thereafter", "for", "generating", "chunk", "scores", ".", "Answer", "Chunking", "To", "reduce", "the", "errors", "generated", "by", "the", "rule", "-", "based", "chunker", "in", ",", "first", ",", "we", "capture", "the", "part", "-", "of", "-", "speech", "(", "POS", ")", "pattern", "of", "all", "answer", "sub", "-", "sequences", "in", "the", "training", "dataset", "to", "form", "a", "POS", "pattern", "trie", "tree", ",", "and", "then", "apply", "the", "answer", "POS", "patterns", "to", "passage", "to", "acquire", "a", "collection", "of", "all", "subsequences", "(", "chunk", "candidates", ")", "whose", "POS", "patterns", "can", "be", "matched", "to", "the", "POS", "pattern", "trie", ".", "This", "is", "equivalent", "to", "putting", "an", "constraint", "to", "candidate", "answer", "chunk", "generation", "process", "that", "only", "choose", "the", "chunk", "with", "a", "POS", "pattern", "seen", "for", "answers", "in", "the", "training", "data", ".", "Then", "the", "sub", "-", "sequences", "are", "used", "as", "answer", "candidates", "for", ".", "Note", "that", "overlapping", "chunks", "could", "be", "generated", "for", "a", "passage", ",", "and", "we", "rely", "on", "the", "ranker", "to", "choose", "the", "best", "candidate", "based", "on", "features", "from", "the", "cloze", "-", "style", "RC", "system", ".", "Experiments", "showed", "that", "for", "of", "the", "questions", "on", "the", "development", "set", ",", "the", "ground", "truth", "answer", "is", "included", "in", "the", "candidate", "set", "constructed", "in", "such", "manner", ".", "Feature", "Extraction", "and", "Ranking", "For", "chunk", "ranking", ",", "we", "(", "1", ")", "use", "neural", "RCQA", "model", "to", "annotate", "each", "in", "passage", "to", "get", "score", ",", "then", "(", "2", ")", "for", "every", "chunk", "in", "passage", ",", "collect", "scores", "for", "all", "the", "contained", "within", ",", "and", "(", "3", ")", "extract", "features", "on", "the", "sequence", "of", "scores", "to", "characterize", "its", "scale", "and", "distribution", "information", ",", "which", "serves", "as", "the", "feature", "representation", "of", ".", "In", "step", "(", "1", ")", "to", "acquire", "we", "train", "and", "apply", "a", "word", "-", "level", "single", "-", "layer", "Gated", "Attention", "Reader", ",", "which", "has", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "CNN", "/", "DailyMail", "cloze", "-", "style", "RC", "task", ".", "In", "step", "(", "3", ")", "for", "chunk", ",", "we", "designed", "5", "features", ",", "including", "4", "statistics", "on", ":", "maximum", ",", "minimum", ",", "average", "and", "sum", ";", "as", "well", "as", "the", "count", "of", "matched", "POS", "pattern", "within", "the", "chunk", ",", "which", "serves", "as", "an", "answer", "prior", ".", "We", "use", "these", "5", "features", "in", "a", "state", "-", "of", "-", "the", "-", "art", "ranker", ".", "section", ":", "Dynamic", "Chunk", "Reader", "The", "dynamic", "chunk", "reader", "(", "DCR", ")", "model", "is", "presented", "in", "Figure", "[", "reference", "]", ".", "Inspired", "by", "the", "baseline", "we", "built", ",", "DCR", "is", "deemed", "to", "be", "superior", "to", "the", "baseline", "for", "3", "reasons", ".", "First", ",", "each", "chunk", "has", "a", "representation", "constructed", "dynamically", ",", "instead", "of", "having", "a", "set", "of", "pre", "-", "defined", "feature", "values", ".", "Second", ",", "each", "passage", "word", "’s", "representation", "is", "enhanced", "by", "word", "-", "by", "-", "word", "attention", "that", "evaluates", "the", "relevance", "of", "the", "passage", "word", "to", "the", "question", ".", "Third", ",", "these", "components", "are", "all", "within", "a", "single", ",", "end", "-", "to", "-", "end", "model", "that", "can", "be", "trained", "in", "a", "joint", "manner", ".", "DCR", "works", "in", "four", "steps", ".", "First", ",", "the", "encoder", "layer", "encodes", "passage", "and", "question", "separately", ",", "by", "using", "bidirectional", "recurrent", "neural", "networks", "(", "RNN", ")", ".", "Second", ",", "the", "attention", "layer", "calculates", "the", "relevance", "of", "each", "passage", "word", "to", "the", "question", ".", "Third", ",", "the", "chunk", "representation", "layer", "dynamically", "extracts", "the", "candidate", "chunks", "from", "the", "given", "passage", ",", "and", "create", "chunk", "representation", "that", "encodes", "the", "contextual", "information", "of", "each", "chunk", ".", "Fourth", ",", "the", "ranker", "layer", "scores", "the", "relevance", "between", "the", "representations", "of", "a", "chunk", "and", "the", "given", "question", ",", "and", "ranks", "all", "candidate", "chunks", "using", "a", "softmax", "layer", ".", "We", "describe", "each", "step", "in", "details", "below", ".", "Encoder", "Layer", "We", "use", "bi", "-", "directional", "RNN", "encoder", "to", "encode", "and", "of", "example", ",", "and", "get", "hidden", "state", "for", "each", "word", "position", "and", ".", "As", "RNN", "input", ",", "a", "word", "is", "represented", "by", "a", "row", "vector", ".", "can", "be", "the", "concatenation", "of", "word", "embedding", "and", "word", "features", "(", "see", "Fig", ".", "[", "reference", "]", ")", ".", "The", "word", "vector", "for", "the", "-", "th", "word", "is", ".", "A", "word", "sequence", "is", "processed", "using", "an", "RNN", "encoder", "with", "gated", "recurrent", "units", "(", "GRU", ")", ",", "which", "was", "proved", "to", "be", "effective", "in", "RC", "and", "neural", "machine", "translation", "tasks", ".", "For", "each", "position", ",", "GRU", "computes", "with", "input", "and", "previous", "state", ",", "as", ":", "where", ",", ",", "and", "are", "d", "-", "dimensional", "hidden", "state", ",", "reset", "gate", ",", "and", "update", "gate", ",", "respectively", ";", ",", "and", ",", "are", "the", "parameters", "of", "the", "GRU", ";", "is", "the", "sigmoid", "function", ",", "and", "denotes", "element", "-", "wise", "production", ".", "For", "a", "word", "at", ",", "we", "use", "the", "hidden", "state", "from", "the", "forward", "RNN", "as", "a", "representation", "of", "the", "preceding", "context", ",", "and", "the", "from", "a", "backward", "RNN", "that", "encodes", "text", "reversely", ",", "to", "incorporate", "the", "context", "after", ".", "Next", ",", ",", "the", "bi", "-", "directional", "contextual", "encoding", "of", ",", "is", "formed", ".", "is", "the", "concatenation", "operator", ".", "To", "distinguish", "hidden", "states", "from", "different", "sources", ",", "we", "denote", "the", "of", "-", "th", "word", "in", "and", "the", "of", "-", "th", "word", "in", "as", "and", "respectively", ".", "Attention", "Layer", "Attention", "mechanism", "in", "previous", "RC", "tasks", "enables", "question", "-", "aware", "passage", "representations", ".", "We", "propose", "a", "novel", "attention", "mechanism", "inspired", "by", "word", "-", "by", "-", "word", "style", "attention", "methods", ".", "For", "each", ",", "a", "question", "-", "attended", "representation", "is", "computed", "as", "follows", "(", "example", "index", "is", "omitted", "for", "simplicity", ")", ":", "where", "and", "are", "hidden", "states", "from", "the", "bi", "-", "directional", "RNN", "encoders", "(", "see", "Figure", "[", "reference", "]", ")", ".", "An", "inner", "product", ",", ",", "is", "calculated", "between", "and", "every", "question", "word", ".", "It", "indicates", "how", "well", "the", "passage", "word", "matches", "with", "every", "question", "word", ".", "is", "a", "weighted", "pooling", "of", "question", "hidden", "states", ",", "which", "serves", "as", "a", "-", "aware", "question", "representation", ".", "The", "concatenation", "of", "and", "leads", "to", "a", "passage", "-", "question", "joint", "representation", ",", ".", "Next", ",", "we", "apply", "a", "second", "bi", "-", "GRU", "layer", "taking", "the", "s", "as", "inputs", ",", "and", "obtain", "forward", "and", "backward", "representations", "and", ",", "and", "in", "turn", "their", "concatenation", ",", ".", "Chunk", "Representation", "Layer", "A", "candidate", "answer", "chunk", "representation", "is", "dynamically", "created", "given", "attention", "layer", "output", ".", "We", "first", "decide", "the", "text", "boundary", "for", "the", "candidate", "chunk", ",", "and", "then", "form", "a", "chunk", "representation", "using", "all", "or", "part", "of", "those", "outputs", "inside", "the", "chunk", ".", "To", "decide", "a", "candidate", "chunk", "(", "boundary", ")", ":", "we", "tried", "two", "ways", ":", "(", "1", ")", "adopt", "the", "POS", "trie", "-", "based", "approach", "used", "in", "our", "baseline", ",", "and", "(", "2", ")", "enumerate", "all", "possible", "chunks", "up", "to", "a", "maximum", "number", "of", "tokens", ".", "For", "(", "2", ")", ",", "we", "create", "up", "to", "(", "max", "chunk", "length", ")", "chunks", "starting", "from", "any", "position", "in", ".", "Approach", "(", "1", ")", "can", "generate", "candidates", "with", "arbitrary", "lengths", ",", "but", "fails", "to", "recall", "candidates", "whose", "POS", "pattern", "is", "unseen", "in", "training", "set", ";", "whereas", "approach", "(", "2", ")", "considers", "all", "possible", "candidates", "within", "a", "window", "and", "is", "more", "flexible", ",", "but", "over", "-", "generates", "invalid", "candidates", ".", "For", "a", "candidate", "answer", "chunk", "spanning", "from", "position", "to", "inclusively", ",", "we", "construct", "chunk", "representation", "using", "every", "within", "range", ",", "with", "a", "function", ".", "Formally", ",", "We", "experimented", "with", "several", "pooling", "functions", "(", "e.g.", ",", "max", ",", "average", ")", "for", ",", "and", "found", "out", "that", ",", "instead", "of", "pooling", ",", "the", "best", "function", "is", "to", "concatenate", "the", "hidden", "state", "of", "the", "first", "word", "in", "a", "chunk", "in", "forward", "RNN", "and", "that", "of", "the", "last", "word", "in", "backward", "RNN", ".", "Formally", ",", "We", "hypothesize", "that", "the", "hidden", "states", "at", "that", "two", "ends", "can", "better", "represent", "the", "chunk", "’s", "contexts", ",", "which", "is", "critical", "for", "this", "task", ",", "than", "the", "states", "within", "the", "chunk", ".", "This", "observation", "also", "agrees", "with", ".", "Ranker", "Layer", "Each", "chunk", "is", "evaluated", "on", "its", "context", "similarity", "to", "the", "question", ",", "by", "taking", "the", "cosine", "similarity", "between", "the", "chunk", "context", "representation", "acquired", "from", "chunk", "representation", "layer", ",", "and", "the", "question", "representation", "which", "is", "the", "concatenation", "of", "the", "last", "hidden", "state", "in", "forward", "RNN", "and", "the", "first", "hidden", "state", "in", "backward", "RNN", ".", "Thus", ",", "for", "training", "example", ",", "we", "have", "the", "probability", "of", "the", "chunk", "as", "where", "denotes", "representation", "of", "the", "chunk", ",", "or", "is", "the", "-", "th", "hidden", "state", "output", "from", "question", "’s", "forward", "and", "backward", "RNN", "encoder", ",", "respectively", ".", "In", "runtime", ",", "the", "chunk", "with", "the", "highest", "probability", "is", "taken", "as", "the", "answer", ".", "In", "training", ",", "the", "following", "negative", "log", "likelihood", "is", "minimized", ":", "Note", "that", "the", "-", "th", "training", "instance", "is", "only", "used", "when", "is", "included", "in", "the", "corresponding", "candidate", "chunk", "set", ",", "i.e.", ".", "The", "softmax", "in", "the", "final", "layer", "serves", "as", "the", "list", "-", "wise", "ranking", "module", "similar", "in", "spirit", "to", ".", "section", ":", "Experiments", "paragraph", ":", "Dataset", "We", "used", "the", "Stanford", "Question", "Answering", "Dataset", "(", "SQuAD", ")", "for", "the", "experiment", ".", "SQuAD", "came", "into", "our", "sight", "because", "it", "is", "a", "mix", "of", "factoid", "and", "non", "-", "factoid", "questions", ",", "a", "real", "-", "world", "data", "(", "crowd", "-", "sourced", ")", ",", "and", "of", "large", "scale", "(", "over", "100", "K", "question", "-", "answer", "pairs", "collected", "from", "536", "Wikipedia", "articles", ")", ".", "Answers", "range", "from", "single", "words", "to", "long", ",", "variable", "-", "length", "phrase", "/", "clauses", ".", "It", "is", "a", "relaxation", "of", "assumptions", "by", "the", "cloze", "-", "style", "and", "quiz", "-", "style", "RC", "datasets", "in", "the", "Problem", "Definition", "section", ".", "Features", "The", "input", "vector", "representation", "of", "each", "word", "to", "encoder", "RNNs", "has", "six", "parts", "including", "a", "pre", "-", "trained", "300", "-", "dimensional", "GloVe", "embedding", "and", "five", "features", "(", "see", "Figure", "[", "reference", "]", ")", ":", "(", "1", ")", "a", "one", "-", "hot", "encoding", "(", "46", "dimensions", ")", "for", "the", "part", "-", "of", "-", "speech", "(", "POS", ")", "tag", "of", ";", "(", "2", ")", "a", "one", "-", "hot", "encoding", "(", "14", "dimensions", ")", "for", "named", "entity", "(", "NE", ")", "tag", "of", ";", "(", "3", ")", "a", "binary", "value", "indicating", "whether", "’s", "surface", "form", "is", "the", "same", "to", "any", "word", "in", "the", "quesiton", ";", "(", "4", ")", "if", "the", "lemma", "form", "of", "is", "the", "same", "to", "any", "word", "in", "the", "question", ";", "and", "(", "5", ")", "if", "is", "caplitalized", ".", "Feature", "(", "3", ")", "and", "(", "4", ")", "are", "designed", "to", "help", "the", "model", "align", "the", "passage", "text", "with", "question", ".", "Note", "that", "some", "types", "of", "questions", "(", "e.g.", ",", "“", "who", "”", ",", "“", "when", "”", "questions", ")", "have", "answers", "that", "have", "a", "specific", "POS", "/", "NE", "tag", "pattern", ".", "For", "instance", ",", "“", "who", "”", "questions", "mostly", "have", "proper", "nouns", "/", "persons", "as", "answers", "and", "“", "when", "”", "questions", "may", "frequently", "have", "numbers", "/", "dates", "(", "e.g.", ",", "a", "year", ")", "as", "answers", ".", "Thus", ",", "we", "believe", "that", "the", "model", "could", "exploit", "the", "co", "-", "relation", "between", "question", "types", "and", "answer", "POS", "/", "NE", "patterns", "easier", "with", "POS", "and", "NE", "tag", "features", ".", "Implementation", "Details", "We", "pre", "-", "processed", "the", "SQuAD", "dataset", "using", "Stanford", "CoreNLP", "tool", "with", "its", "default", "setting", "to", "tokenize", "the", "text", "and", "obtain", "the", "POS", "and", "NE", "annotations", ".", "To", "train", "our", "model", ",", "we", "used", "stochastic", "gradient", "descent", "with", "the", "ADAM", "optimizer", ",", "with", "an", "initial", "learning", "rate", "of", "0.001", ".", "All", "GRU", "weights", "were", "initialized", "from", "a", "uniform", "distribution", "between", "(", "-", "0.01", ",", "0.01", ")", ".", "The", "hidden", "state", "size", ",", ",", "was", "set", "to", "300", "for", "all", "GRUs", ".", "The", "question", "bi", "-", "GRU", "shared", "parameters", "with", "the", "passage", "bi", "-", "GRU", ",", "while", "the", "attention", "-", "based", "passage", "bi", "-", "GRU", "had", "its", "own", "parameters", ".", "We", "shuffled", "all", "training", "examples", "at", "the", "beginning", "of", "each", "epoch", "and", "adopted", "a", "curriculum", "learning", "approach", ",", "by", "sorting", "training", "instances", "by", "length", "in", "every", "10", "batches", ",", "to", "enable", "the", "model", "start", "learning", "from", "relatively", "easier", "instances", "and", "to", "harder", "ones", ".", "We", "also", "applied", "dropout", "of", "rate", "0.2", "to", "the", "embedding", "layer", "of", "input", "bi", "-", "GRU", "encoder", ",", "and", "gradient", "clipping", "when", "the", "norm", "of", "gradients", "exceeded", "10", ".", "We", "trained", "in", "mini", "-", "batch", "style", "(", "mini", "-", "batch", "size", "is", "180", ")", "and", "applied", "zero", "-", "padding", "to", "the", "passage", "and", "question", "inputs", "in", "each", "batch", ".", "We", "also", "set", "the", "maximum", "passage", "length", "to", "be", "300", "tokens", ",", "and", "pruned", "all", "the", "tokens", "after", "the", "300", "-", "th", "token", "in", "the", "training", "set", "to", "save", "memory", "and", "speed", "up", "the", "training", "process", ".", "This", "step", "reduced", "the", "training", "set", "size", "by", "about", "1.6", "%", ".", "During", "test", ",", "we", "test", "on", "the", "full", "length", "of", "passage", ",", "so", "that", "we", "do", "n’t", "prune", "out", "the", "potential", "candidates", ".", "We", "trained", "the", "model", "for", "at", "most", "30", "epochs", ",", "and", "in", "case", "the", "accuracy", "did", "not", "improve", "for", "10", "epochs", ",", "we", "stopped", "training", ".", "For", "the", "feature", "ranking", "-", "based", "system", ",", "we", "used", "jforest", "ranker", "with", "LambdaMART", "-", "RegressionTree", "algorithm", "and", "the", "ranking", "metric", "was", "NDCG@10", ".", "For", "the", "Gated", "Attention", "Reader", "in", "baseline", "system", ",", "we", "replicated", "the", "method", "and", "use", "the", "same", "configurations", "as", "in", ".", "Results", "Table", "[", "reference", "]", "shows", "our", "main", "results", "on", "the", "SQuAD", "dataset", ".", "Compared", "to", "the", "scores", "reported", "in", ",", "our", "exact", "match", "(", "EM", ")", "and", "F1", "on", "the", "development", "set", "and", "EM", "score", "on", "the", "test", "set", "are", "better", ",", "and", "F1", "on", "the", "test", "set", "is", "comparable", ".", "We", "also", "studied", "how", "each", "component", "in", "our", "model", "contributes", "to", "the", "overall", "performance", ".", "Table", "[", "reference", "]", "shows", "the", "details", "as", "well", "as", "the", "results", "of", "the", "baseline", "ranker", ".", "As", "the", "first", "row", "of", "Table", "[", "reference", "]", "shows", ",", "our", "baseline", "system", "improves", "10", "%", "(", "EM", ")", "over", "Rajpurkar", "et", "al", ".", "rajpurkar2016squad", "(", "Table", "[", "reference", "]", ",", "row", "1", ")", ",", "the", "feature", "-", "based", "ranking", "system", ".", "However", "when", "compared", "to", "our", "DCR", "model", "(", "Table", "[", "reference", "]", ",", "row", "2", ")", ",", "the", "baseline", "(", "row", "1", ")", "is", "more", "than", "12", "%", "(", "EM", ")", "behind", "even", "though", "it", "is", "based", "on", "the", "state", "-", "of", "-", "the", "-", "art", "model", "for", "cloze", "-", "style", "RC", "tasks", ".", "This", "can", "be", "attributed", "to", "the", "advanced", "model", "structure", "and", "end", "-", "to", "-", "end", "manner", "of", "DCR", ".", "We", "also", "did", "ablation", "tests", "on", "our", "DCR", "model", ".", "First", ",", "replacing", "the", "word", "-", "by", "-", "word", "attention", "with", "Attentive", "Reader", "style", "attention", "decreases", "the", "EM", "score", "by", "about", "4.5", "%", ",", "showing", "the", "strength", "of", "our", "proposed", "attention", "mechanism", ".", "Second", ",", "we", "remove", "the", "features", "in", "input", "to", "see", "the", "contribution", "of", "each", "feature", ".", "The", "result", "shows", "that", "POS", "feature", "(", "1", ")", "and", "question", "-", "word", "feature", "(", "3", ")", "are", "the", "two", "most", "important", "features", ".", "Finally", ",", "combining", "the", "DCR", "model", "with", "the", "proposed", "POS", "-", "trie", "constraints", "yields", "a", "score", "similar", "to", "the", "one", "obtained", "using", "the", "DCR", "model", "with", "all", "possible", "-", "gram", "chunks", ".", "The", "result", "shows", "that", "(", "1", ")", "our", "chunk", "representations", "are", "powerful", "enough", "to", "differentiate", "even", "a", "huge", "amount", "of", "chunks", "when", "no", "constraints", "are", "applied", ";", "and", "(", "2", ")", "the", "proposed", "POS", "-", "trie", "reduces", "the", "search", "space", "at", "the", "cost", "of", "a", "small", "drop", "in", "performance", ".", "Analysis", "To", "better", "understand", "our", "system", ",", "we", "calculated", "the", "accuracy", "of", "the", "attention", "mechanism", "of", "the", "gated", "attention", "reader", "used", "in", "our", "deep", "learning", "-", "based", "baseline", ".", "We", "found", "that", "it", "is", "72", "%", "accurate", "i.e.", ",", "72", "%", "of", "the", "times", "a", "word", "with", "the", "highest", "attention", "score", "is", "inside", "the", "correct", "answer", "span", ".", "This", "means", "that", ",", "if", "we", "could", "accurately", "detect", "the", "boundary", "around", "the", "word", "with", "the", "highest", "attention", "score", "to", "form", "the", "answer", "span", ",", "we", "could", "achieve", "an", "accuracy", "close", "to", "72", "%", ".", "In", "addition", ",", "we", "checked", "the", "answer", "recall", "of", "our", "candidate", "chunking", "approach", ".", "When", "we", "use", "a", "window", "size", "of", "10", ",", "92", "%", "of", "the", "time", ",", "the", "ground", "truth", "answer", "will", "be", "included", "in", "the", "extracted", "Candidate", "chunk", "set", ".", "Thus", "the", "upper", "bound", "of", "the", "exact", "match", "score", "of", "our", "baseline", "system", "is", "around", "66", "%", "(", "92", "%", "(", "the", "answer", "recall", ")", "72", "%", ")", ".", "From", "the", "results", ",", "we", "see", "our", "DCR", "system", "’s", "exact", "match", "score", "is", "at", "62", "%", ".", "This", "shows", "that", "DCR", "is", "proficient", "at", "differentiating", "answer", "spans", "dynamically", ".", "[", "]", "[", "]", "To", "further", "analyze", "the", "system", "’s", "performance", "while", "predicting", "answers", "of", "different", "lengths", ",", "we", "show", "the", "exact", "match", "(", "EM", ")", "and", "F1", "scores", "for", "answers", "with", "lengths", "up", "to", "10", "tokens", "in", "Figure", "2", "(", "a", ")", ".", "From", "the", "graph", ",", "we", "can", "see", "that", ",", "with", "the", "increase", "of", "answer", "length", ",", "both", "EM", "and", "F1", "drops", ",", "but", "in", "different", "speed", ".", "The", "gap", "between", "F1", "and", "exact", "match", "also", "widens", "as", "answer", "length", "increases", ".", "However", ",", "the", "model", "still", "yields", "a", "decent", "accuracy", "when", "the", "answer", "is", "longer", "than", "a", "single", "word", ".", "Additionally", ",", "Figure", "2", "(", "b", ")", "shows", "that", "the", "system", "is", "better", "at", "“", "when", "”", "and", "“", "who", "”", "questions", ",", "but", "performs", "poorly", "on", "“", "why", "”", "questions", ".", "The", "large", "gap", "between", "exact", "match", "and", "F1", "on", "“", "why", "”", "questions", "means", "that", "perfectly", "identifying", "the", "span", "is", "harder", "than", "locating", "the", "core", "of", "the", "answer", "span", ".", "Since", "“", "what", "”", ",", "“", "which", "”", ",", "and", "“", "how", "”", "questions", "contain", "a", "broad", "range", "of", "question", "types", ",", "we", "split", "them", "further", "based", "on", "the", "bigram", "a", "question", "starts", "with", ",", "and", "Figure", "[", "reference", "]", "shows", "the", "breakdown", "for", "“", "what", "”", "questions", ".", "We", "can", "see", "that", "“", "what", "”", "questions", "asking", "for", "explanations", "such", "as", "“", "what", "happens", "”", "and", "“", "what", "happened", "”", "have", "lower", "EM", "and", "F1", "scores", ".", "In", "contrast", ",", "“", "what", "”", "questions", "asking", "for", "year", "and", "numbers", "have", "much", "higher", "scores", "and", ",", "for", "these", "questions", ",", "exact", "match", "scores", "are", "close", "to", "F1", "scores", ",", "which", "means", "chunking", "for", "these", "questions", "are", "easier", "for", "DCR", ".", "section", ":", "Related", "Work", "Attentive", "Reader", "was", "the", "first", "neural", "model", "for", "factoid", "RCQA", ".", "It", "uses", "Bidirectional", "RNN", "(", "Cho", "et", "al", ".", ",", "2014", ";", "Chung", "et", "al.", ",", "2014", ")", "to", "encode", "document", "and", "query", "respectively", ",", "and", "use", "query", "representation", "to", "match", "with", "every", "token", "from", "the", "document", ".", "Attention", "Sum", "Reader", "simplifies", "the", "model", "to", "just", "predicting", "positions", "of", "correct", "answer", "in", "the", "document", "and", "the", "training", "speed", "and", "test", "accuracy", "are", "both", "greatly", "improved", "on", "the", "CNN", "/", "Daily", "Mail", "dataset", ".", "also", "simplified", "Attentive", "Reader", "and", "reported", "higher", "accuracy", ".", "Window", "-", "based", "Memory", "Networks", "(", "MemN2N", ")", "is", "introduced", "along", "with", "the", "CBT", "dataset", ",", "which", "does", "not", "use", "RNN", "encoders", ",", "but", "embeds", "contexts", "as", "memory", "and", "matches", "questions", "with", "embedded", "contexts", ".", "Those", "models", "’", "mechanism", "is", "to", "learn", "the", "match", "between", "answer", "context", "with", "question", "/", "query", "representation", ".", "In", "contrast", ",", "memory", "enhanced", "neural", "networks", "like", "Neural", "Turing", "Machines", "and", "its", "variants", "were", "also", "potential", "candidates", "for", "the", "task", ",", "and", "Gulcehre", "et", "al", ".", "gulcehre2016dynamic", "reported", "results", "on", "the", "bAbI", "task", ",", "which", "is", "worse", "than", "memory", "networks", ".", "Similarly", ",", "sequence", "-", "to", "-", "sequence", "models", "were", "also", "used", ",", "but", "they", "did", "not", "yield", "better", "results", "either", ".", "Recently", ",", "several", "models", "have", "been", "proposed", "to", "enable", "more", "complex", "inference", "for", "RC", "task", ".", "For", "instance", ",", "gated", "attention", "model", "employs", "a", "multi", "-", "layer", "architecture", ",", "where", "each", "layer", "encodes", "the", "same", "document", ",", "but", "the", "attention", "is", "updated", "from", "layer", "to", "layer", ".", "EpiReader", "adopted", "a", "joint", "training", "model", "for", "answer", "extractor", "and", "reasoner", ",", "where", "the", "extractor", "proposes", "top", "candidates", ",", "and", "the", "reasoner", "weighs", "each", "candidate", "by", "examining", "entailment", "relationship", "between", "question", "-", "answer", "representation", "and", "the", "document", ".", "An", "iterative", "alternating", "attention", "mechanism", "and", "gating", "strategies", "were", "proposed", "in", "to", "optimize", "the", "attention", "through", "several", "hops", ".", "In", "contrast", ",", "Cui", "et", "al", ".", "cui2016aoa", ",", "cui2016consensus", "introduced", "fine", "-", "grained", "document", "attention", "from", "each", "question", "word", "and", "then", "aggregated", "those", "attentions", "from", "each", "question", "token", "by", "summation", "with", "or", "without", "weights", ".", "This", "system", "achieved", "the", "state", "-", "of", "-", "the", "-", "art", "score", "on", "the", "CNN", "dataset", ".", "Those", "different", "variations", "all", "result", "in", "roughly", "3", "-", "5", "%", "improvement", "over", "attention", "sum", "reader", ",", "but", "none", "of", "those", "could", "achieve", "higher", "than", "that", ".", "Other", "methods", "include", "using", "dynamic", "entity", "representation", "with", "max", "-", "pooling", "that", "aims", "to", "change", "entity", "representation", "with", "context", ",", "and", "Weissenborn", "’s", "weissenborn2016separating", "system", ",", "which", "tries", "to", "separate", "entity", "from", "the", "context", "and", "then", "matches", "the", "question", "to", "context", ",", "scoring", "an", "accuracy", "around", "70", "%", "on", "the", "CNN", "dataset", ".", "However", ",", "all", "of", "those", "models", "assume", "that", "the", "answers", "are", "single", "tokens", ".", "This", "limits", "the", "type", "of", "questions", "the", "models", "can", "answer", ".", "Wang", "and", "Jiang", "wang2016machine", "proposed", "a", "match", "-", "lstm", "and", "achieved", "good", "results", "on", "SQuAD", ".", "However", ",", "this", "approach", "predicts", "a", "chunk", "boundary", "or", "whether", "a", "word", "is", "part", "of", "a", "chunk", "or", "not", ".", "In", "contrast", ",", "our", "approach", "explicitly", "constructs", "the", "chunk", "representations", "and", "similar", "chunks", "are", "compared", "directly", "to", "determine", "correct", "answer", "boundaries", ".", "section", ":", "Conclusion", "In", "this", "paper", "we", "proposed", "a", "novel", "neural", "reading", "comprehension", "model", "for", "question", "answering", ".", "Different", "from", "the", "previously", "proposed", "models", "for", "factoid", "RCQA", ",", "the", "proposed", "model", ",", "dynamic", "chunk", "reader", ",", "is", "not", "restricted", "to", "predicting", "a", "single", "named", "entity", "as", "an", "answer", "or", "selecting", "an", "answer", "from", "a", "small", ",", "pre", "-", "defined", "candidate", "list", ".", "Instead", ",", "it", "is", "capable", "of", "answering", "both", "factoid", "and", "non", "-", "factoid", "questions", "as", "it", "learns", "to", "select", "answer", "chunks", "that", "are", "suitable", "for", "an", "input", "question", ".", "DCR", "achieves", "this", "goal", "with", "a", "joint", "deep", "learning", "model", "enhanced", "with", "a", "novel", "attention", "mechanism", "and", "five", "simple", "yet", "effective", "features", ".", "Error", "analysis", "shows", "that", "the", "DCR", "model", "achieves", "good", "performance", ",", "but", "still", "needs", "to", "improve", "on", "predicting", "longer", "answers", ",", "which", "are", "usually", "non", "-", "factoid", "in", "nature", ".", "bibliography", ":", "References"], "entity": [[7, 10, ["Answer", "Chunk", "Extraction"], "Method"], [11, 12, ["Ranking"], "Method"], [13, 15, ["Reading", "Comprehension"], "Task"], [18, 21, ["dynamic", "chunk", "reader"], "Method"], [22, 23, ["DCR"], "Method"], [26, 38, ["end", "-", "to", "-", "end", "neural", "reading", "comprehension", "(", "RC", ")", "model"], "Method"], [58, 59, ["DCR"], "Method"], [70, 73, ["neural", "RC", "models"], "Method"], [82, 83, ["DCR"], "Method"], [91, 94, ["recurrent", "neural", "networks"], "Method"], [99, 106, ["word", "-", "by", "-", "word", "attention", "mechanism"], "Method"], [108, 112, ["question", "-", "aware", "representations"], "Method"], [121, 123, ["chunk", "representations"], "Method"], [125, 127, ["ranking", "module"], "Method"], [142, 143, ["DCR"], "Method"], [151, 153, ["exact", "match"], "Metric"], [154, 156, ["F1", "scores"], "Metric"], [158, 160, ["SQuAD", "dataset"], "Material"], [164, 170, ["Reading", "comprehension", "-", "based", "question", "answering"], "Task"], [171, 172, ["RCQA"], "Task"], [196, 198, ["neural", "models"], "Method"], [260, 261, ["RCQA"], "Task"], [264, 269, ["real", "-", "world", "QA", "scenario"], "Task"], [308, 309, ["RCQA"], "Task"], [315, 317, ["QA", "approaches"], "Method"], [337, 338, ["RCQA"], "Task"], [366, 367, ["RCQA"], "Task"], [418, 420, ["RC", "task"], "Task"], [421, 426, ["predicting", "single", "tokens", "/", "entities"], "Task"], [437, 440, ["search", "space", "complexity"], "Metric"], [482, 483, ["complexity"], "Metric"], [514, 520, ["chunk", "-", "and", "-", "rank", "approach"], "Method"], [523, 527, ["rule", "-", "based", "algorithm"], "Method"], [538, 540, ["ranking", "approach"], "Method"], [552, 557, ["rule", "-", "based", "chunking", "approach"], "Method"], [559, 561, ["low", "coverage"], "Metric"], [564, 565, ["recall"], "Metric"], [578, 580, ["candidate", "ranking"], "Task"], [602, 610, ["end", "-", "to", "-", "end", "neural", "network", "models"], "Method"], [660, 663, ["dynamic", "chunk", "reader"], "Method"], [664, 665, ["DCR"], "Method"], [700, 702, ["deep", "networks"], "Method"], [714, 717, ["fixed", "feature", "representations"], "Method"], [734, 738, ["word", "-", "level", "representations"], "Method"], [774, 777, ["neural", "network", "model"], "Method"], [778, 784, ["joint", "candidate", "answer", "chunking", "and", "ranking"], "Task"], [811, 815, ["question", "-", "attention", "mechanism"], "Method"], [817, 820, ["passage", "word", "representation"], "Task"], [827, 829, ["chunk", "representations"], "Method"], [844, 846, ["attention", "mechanism"], "Method"], [850, 852, ["candidate", "ranking"], "Task"], [860, 864, ["exact", "boundary", "match", "accuracy"], "Metric"], [869, 873, ["Stanford", "Question", "Answering", "Dataset"], "Material"], [874, 876, ["SQuAD", ")"], "Material"], [912, 914, ["RCQA", "problem"], "Task"], [924, 927, ["neural", "network", "component"], "Method"], [1187, 1189, ["RC", "Tasks"], "Task"], [1195, 1197, ["RC", "task"], "Task"], [1255, 1258, ["answer", "selection", "task"], "Task"], [1272, 1275, ["factoid", "answer", "extraction"], "Task"], [1280, 1281, ["bAbI"], "Method"], [1283, 1285, ["inference", "purpose"], "Task"], [1288, 1290, ["SQuAD", "dataset"], "Material"], [1303, 1305, ["SQuAD", "dataset"], "Material"], [1311, 1318, ["factoid", "and", "non", "-", "factoid", "answer", "extraction"], "Task"], [1334, 1340, ["Chunk", "-", "and", "-", "Rank", "Pipeline"], "Task"], [1341, 1343, ["Neural", "RC"], "Method"], [1356, 1358, ["RC", "system"], "Method"], [1359, 1363, ["cloze", "-", "style", "tasks"], "Task"], [1365, 1368, ["answer", "extraction", "purpose"], "Task"], [1403, 1408, ["cloze", "-", "style", "RC", "system"], "Method"], [1410, 1414, ["chunk", "-", "level", "decision"], "Method"], [1418, 1420, ["RC", "model"], "Method"], [1432, 1436, ["feature", "-", "based", "ranker"], "Method"], [1450, 1454, ["deep", "learning", "based", "counterpart"], "Method"], [1468, 1473, ["stand", "-", "alone", "answer", "chunker"], "Method"], [1487, 1490, ["neural", "RC", "model"], "Method"], [1511, 1513, ["Answer", "Chunking"], "Method"], [1520, 1524, ["rule", "-", "based", "chunker"], "Method"], [1586, 1589, ["POS", "pattern", "trie"], "Method"], [1599, 1603, ["answer", "chunk", "generation", "process"], "Task"], [1659, 1664, ["cloze", "-", "style", "RC", "system"], "Method"], [1692, 1694, ["Feature", "Extraction"], "Task"], [1695, 1696, ["Ranking"], "Task"], [1697, 1699, ["chunk", "ranking"], "Task"], [1758, 1760, ["feature", "representation"], "Method"], [1774, 1783, ["word", "-", "level", "single", "-", "layer", "Gated", "Attention", "Reader"], "Method"], [1795, 1803, ["CNN", "/", "DailyMail", "cloze", "-", "style", "RC", "task"], "Task"], [1810, 1811, ["chunk"], "Task"], [1864, 1865, ["ranker"], "Method"], [1868, 1871, ["Dynamic", "Chunk", "Reader"], "Method"], [1872, 1875, ["dynamic", "chunk", "reader"], "Method"], [1876, 1877, ["DCR"], "Method"], [1894, 1895, ["DCR"], "Method"], [1967, 1973, ["end", "-", "to", "-", "end", "model"], "Method"], [1982, 1983, ["DCR"], "Method"], [1991, 1993, ["encoder", "layer"], "Method"], [2001, 2005, ["bidirectional", "recurrent", "neural", "networks"], "Method"], [2006, 2007, ["RNN"], "Method"], [2012, 2014, ["attention", "layer"], "Method"], [2028, 2031, ["chunk", "representation", "layer"], "Method"], [2043, 2045, ["chunk", "representation"], "Method"], [2057, 2059, ["ranker", "layer"], "Method"], [2080, 2082, ["softmax", "layer"], "Method"], [2091, 2093, ["Encoder", "Layer"], "Method"], [2095, 2100, ["bi", "-", "directional", "RNN", "encoder"], "Method"], [2165, 2167, ["RNN", "encoder"], "Method"], [2168, 2171, ["gated", "recurrent", "units"], "Method"], [2172, 2174, ["GRU", ")"], "Method"], [2182, 2188, ["RC", "and", "neural", "machine", "translation", "tasks"], "Task"], [2193, 2194, ["GRU"], "Method"], [2231, 2232, ["GRU"], "Method"], [2240, 2244, ["element", "-", "wise", "production"], "Method"], [2257, 2259, ["forward", "RNN"], "Method"], [2271, 2273, ["backward", "RNN"], "Method"], [2330, 2334, ["Attention", "Layer", "Attention", "mechanism"], "Method"], [2336, 2338, ["RC", "tasks"], "Task"], [2339, 2344, ["question", "-", "aware", "passage", "representations"], "Task"], [2349, 2351, ["attention", "mechanism"], "Method"], [2353, 2361, ["word", "-", "by", "-", "word", "style", "attention", "methods"], "Method"], [2366, 2370, ["question", "-", "attended", "representation"], "Method"], [2390, 2395, ["bi", "-", "directional", "RNN", "encoders"], "Method"], [2431, 2437, ["weighted", "pooling", "of", "question", "hidden", "states"], "Method"], [2442, 2446, ["-", "aware", "question", "representation"], "Method"], [2454, 2459, ["passage", "-", "question", "joint", "representation"], "Method"], [2467, 2471, ["bi", "-", "GRU", "layer"], "Method"], [2492, 2495, ["Chunk", "Representation", "Layer"], "Method"], [2497, 2500, ["answer", "chunk", "representation"], "Method"], [2523, 2525, ["chunk", "representation"], "Method"], [2555, 2560, ["POS", "trie", "-", "based", "approach"], "Method"], [2664, 2666, ["chunk", "representation"], "Method"], [2681, 2683, ["pooling", "functions"], "Method"], [2699, 2700, ["pooling"], "Method"], [2718, 2720, ["forward", "RNN"], "Method"], [2727, 2729, ["backward", "RNN"], "Method"], [2770, 2772, ["Ranker", "Layer"], "Method"], [2791, 2794, ["chunk", "context", "representation"], "Method"], [2796, 2799, ["chunk", "representation", "layer"], "Method"], [2802, 2804, ["question", "representation"], "Method"], [2814, 2816, ["forward", "RNN"], "Method"], [2822, 2824, ["backward", "RNN"], "Method"], [2859, 2862, ["backward", "RNN", "encoder"], "Method"], [2881, 2882, ["training"], "Task"], [2922, 2927, ["list", "-", "wise", "ranking", "module"], "Method"], [2941, 2945, ["Stanford", "Question", "Answering", "Dataset"], "Material"], [2946, 2947, ["SQuAD"], "Material"], [2952, 2953, ["SQuAD"], "Material"], [3023, 3032, ["cloze", "-", "style", "and", "quiz", "-", "style", "RC", "datasets"], "Method"], [3041, 3043, ["vector", "representation"], "Method"], [3047, 3049, ["encoder", "RNNs"], "Method"], [3057, 3062, ["300", "-", "dimensional", "GloVe", "embedding"], "Method"], [3077, 3081, ["one", "-", "hot", "encoding"], "Method"], [3102, 3106, ["one", "-", "hot", "encoding"], "Method"], [3289, 3291, ["SQuAD", "dataset"], "Material"], [3292, 3295, ["Stanford", "CoreNLP", "tool"], "Method"], [3318, 3321, ["stochastic", "gradient", "descent"], "Method"], [3323, 3325, ["ADAM", "optimizer"], "Method"], [3329, 3331, ["learning", "rate"], "Metric"], [3366, 3370, ["question", "bi", "-", "GRU"], "Method"], [3374, 3378, ["passage", "bi", "-", "GRU"], "Method"], [3381, 3388, ["attention", "-", "based", "passage", "bi", "-", "GRU"], "Method"], [3407, 3410, ["curriculum", "learning", "approach"], "Method"], [3446, 3454, ["embedding", "layer", "of", "input", "bi", "-", "GRU", "encoder"], "Method"], [3456, 3458, ["gradient", "clipping"], "Method"], [3469, 3473, ["mini", "-", "batch", "style"], "Method"], [3483, 3486, ["zero", "-", "padding"], "Method"], [3582, 3583, ["accuracy"], "Metric"], [3596, 3601, ["feature", "ranking", "-", "based", "system"], "Method"], [3604, 3606, ["jforest", "ranker"], "Method"], [3607, 3611, ["LambdaMART", "-", "RegressionTree", "algorithm"], "Method"], [3613, 3615, ["ranking", "metric"], "Metric"], [3616, 3617, ["NDCG@10"], "Method"], [3620, 3623, ["Gated", "Attention", "Reader"], "Task"], [3650, 3652, ["SQuAD", "dataset"], "Material"], [3661, 3663, ["exact", "match"], "Metric"], [3664, 3665, ["EM"], "Metric"], [3667, 3668, ["F1"], "Metric"], [3683, 3684, ["F1"], "Metric"], [3741, 3742, ["EM"], "Metric"], [3760, 3765, ["feature", "-", "based", "ranking", "system"], "Method"], [3771, 3773, ["DCR", "model"], "Method"], [3795, 3796, ["EM"], "Metric"], [3814, 3819, ["cloze", "-", "style", "RC", "tasks"], "Task"], [3837, 3838, ["DCR"], "Method"], [3846, 3848, ["DCR", "model"], "Method"], [3879, 3881, ["attention", "mechanism"], "Method"], [3926, 3928, ["DCR", "model"], "Method"], [3945, 3947, ["DCR", "model"], "Method"], [3962, 3964, ["chunk", "representations"], "Method"], [3987, 3990, ["POS", "-", "trie"], "Method"], [4014, 4015, ["accuracy"], "Metric"], [4017, 4019, ["attention", "mechanism"], "Method"], [4021, 4024, ["gated", "attention", "reader"], "Method"], [4027, 4032, ["deep", "learning", "-", "based", "baseline"], "Method"], [4091, 4092, ["accuracy"], "Metric"], [4103, 4105, ["answer", "recall"], "Metric"], [4107, 4110, ["candidate", "chunking", "approach"], "Method"], [4162, 4164, ["answer", "recall"], "Metric"], [4176, 4178, ["DCR", "system"], "Method"], [4190, 4191, ["DCR"], "Method"], [4226, 4228, ["F1", "scores"], "Metric"], [4260, 4261, ["EM"], "Metric"], [4262, 4263, ["F1"], "Metric"], [4273, 4274, ["F1"], "Metric"], [4292, 4293, ["accuracy"], "Metric"], [4342, 4343, ["F1"], "Metric"], [4438, 4439, ["EM"], "Metric"], [4440, 4441, ["F1"], "Metric"], [4471, 4473, ["F1", "scores"], "Metric"], [4483, 4484, ["DCR"], "Method"], [4489, 4491, ["Attentive", "Reader"], "Method"], [4494, 4496, ["neural", "model"], "Method"], [4497, 4499, ["factoid", "RCQA"], "Task"], [4502, 4504, ["Bidirectional", "RNN"], "Method"], [4527, 4529, ["query", "representation"], "Method"], [4556, 4558, ["training", "speed"], "Metric"], [4559, 4561, ["test", "accuracy"], "Metric"], [4575, 4577, ["Attentive", "Reader"], "Method"], [4580, 4581, ["accuracy"], "Metric"], [4582, 4587, ["Window", "-", "based", "Memory", "Networks"], "Method"], [4588, 4590, ["MemN2N", ")"], "Method"], [4602, 4604, ["RNN", "encoders"], "Method"], [4630, 4634, ["question", "/", "query", "representation"], "Method"], [4638, 4642, ["memory", "enhanced", "neural", "networks"], "Method"], [4643, 4646, ["Neural", "Turing", "Machines"], "Method"], [4662, 4663, ["gulcehre2016dynamic"], "Method"], [4667, 4669, ["bAbI", "task"], "Task"], [4674, 4676, ["memory", "networks"], "Method"], [4679, 4685, ["sequence", "-", "to", "-", "sequence", "models"], "Method"], [4709, 4710, ["inference"], "Method"], [4711, 4713, ["RC", "task"], "Task"], [4717, 4720, ["gated", "attention", "model"], "Method"], [4722, 4726, ["multi", "-", "layer", "architecture"], "Method"], [4745, 4746, ["EpiReader"], "Method"], [4748, 4751, ["joint", "training", "model"], "Method"], [4752, 4754, ["answer", "extractor"], "Task"], [4755, 4756, ["reasoner"], "Task"], [4766, 4767, ["reasoner"], "Method"], [4775, 4779, ["question", "-", "answer", "representation"], "Method"], [4784, 4788, ["iterative", "alternating", "attention", "mechanism"], "Method"], [4789, 4791, ["gating", "strategies"], "Method"], [4868, 4871, ["attention", "sum", "reader"], "Metric"], [4886, 4889, ["dynamic", "entity", "representation"], "Method"], [4890, 4893, ["max", "-", "pooling"], "Method"], [4897, 4899, ["entity", "representation"], "Method"], [4903, 4907, ["Weissenborn", "’s", "weissenborn2016separating", "system"], "Method"], [4926, 4927, ["accuracy"], "Metric"], [4966, 4969, ["match", "-", "lstm"], "Method"], [4974, 4975, ["SQuAD"], "Material"], [5004, 5006, ["chunk", "representations"], "Method"], [5028, 5032, ["neural", "reading", "comprehension", "model"], "Method"], [5033, 5035, ["question", "answering"], "Task"], [5043, 5045, ["factoid", "RCQA"], "Task"], [5050, 5053, ["dynamic", "chunk", "reader"], "Method"], [5088, 5094, ["factoid", "and", "non", "-", "factoid", "questions"], "Task"], [5109, 5110, ["DCR"], "Method"], [5115, 5119, ["joint", "deep", "learning", "model"], "Method"], [5123, 5125, ["attention", "mechanism"], "Method"], [5132, 5134, ["Error", "analysis"], "Task"], [5137, 5139, ["DCR", "model"], "Method"], [936, 939, ["dynamic", "chunk", "reader"], "Method"], [1706, 1707, ["RCQA"], "Task"], [3673, 3674, ["EM"], "Metric"], [3866, 3867, ["EM"], "Metric"], [4146, 4148, ["exact", "match"], "Metric"], [4179, 4181, ["exact", "match"], "Metric"], [4220, 4222, ["exact", "match"], "Metric"], [4223, 4224, ["EM"], "Metric"], [4275, 4277, ["exact", "match"], "Metric"], [4339, 4341, ["exact", "match"], "Metric"], [4465, 4467, ["exact", "match"], "Metric"]]} {"doc_id": "0690ba31424310a90028533218d0afd25a829c8d", "text": ["document", ":", "Semantic", "Image", "Segmentation", "with", "Deep", "Convolutional", "Nets", "and", "Fully", "Connected", "CRFs", "Deep", "Convolutional", "Neural", "Networks", "(", "DCNNs", ")", "have", "recently", "shown", "state", "of", "the", "art", "performance", "in", "high", "level", "vision", "tasks", ",", "such", "as", "image", "classification", "and", "object", "detection", ".", "This", "work", "brings", "together", "methods", "from", "DCNNs", "and", "probabilistic", "graphical", "models", "for", "addressing", "the", "task", "of", "pixel", "-", "level", "classification", "(", "also", "called", "”", "semantic", "image", "segmentation", "”", ")", ".", "We", "show", "that", "responses", "at", "the", "final", "layer", "of", "DCNNs", "are", "not", "sufficiently", "localized", "for", "accurate", "object", "segmentation", ".", "This", "is", "due", "to", "the", "very", "invariance", "properties", "that", "make", "DCNNs", "good", "for", "high", "level", "tasks", ".", "We", "overcome", "this", "poor", "localization", "property", "of", "deep", "networks", "by", "combining", "the", "responses", "at", "the", "final", "DCNN", "layer", "with", "a", "fully", "connected", "Conditional", "Random", "Field", "(", "CRF", ")", ".", "Qualitatively", ",", "our", "“", "DeepLab", "”", "system", "is", "able", "to", "localize", "segment", "boundaries", "at", "a", "level", "of", "accuracy", "which", "is", "beyond", "previous", "methods", ".", "Quantitatively", ",", "our", "method", "sets", "the", "new", "state", "-", "of", "-", "art", "at", "the", "PASCAL", "VOC", "-", "2012", "semantic", "image", "segmentation", "task", ",", "reaching", "71.6", "%", "IOU", "accuracy", "in", "the", "test", "set", ".", "We", "show", "how", "these", "results", "can", "be", "obtained", "efficiently", ":", "Careful", "network", "re", "-", "purposing", "and", "a", "novel", "application", "of", "the", "’", "hole", "’", "algorithm", "from", "the", "wavelet", "community", "allow", "dense", "computation", "of", "neural", "net", "responses", "at", "8", "frames", "per", "second", "on", "a", "modern", "GPU", ".", "section", ":", "Introduction", "Deep", "Convolutional", "Neural", "Networks", "(", "DCNNs", ")", "had", "been", "the", "method", "of", "choice", "for", "document", "recognition", "since", "LeCun1998", ",", "but", "have", "only", "recently", "become", "the", "mainstream", "of", "high", "-", "level", "vision", "research", ".", "Over", "the", "past", "two", "years", "DCNNs", "have", "pushed", "the", "performance", "of", "computer", "vision", "systems", "to", "soaring", "heights", "on", "a", "broad", "array", "of", "high", "-", "level", "problems", ",", "including", "image", "classification", "KrizhevskyNIPS2013", ",", "sermanet2013overfeat", ",", "simonyan2014very", ",", "szegedy2014going", ",", "papandreou2014untangling", ",", "object", "detection", "girshick2014rcnn", ",", "fine", "-", "grained", "categorization", "zhang2014part", ",", "among", "others", ".", "A", "common", "theme", "in", "these", "works", "is", "that", "DCNNs", "trained", "in", "an", "end", "-", "to", "-", "end", "manner", "deliver", "strikingly", "better", "results", "than", "systems", "relying", "on", "carefully", "engineered", "representations", ",", "such", "as", "SIFT", "or", "HOG", "features", ".", "This", "success", "can", "be", "partially", "attributed", "to", "the", "built", "-", "in", "invariance", "of", "DCNNs", "to", "local", "image", "transformations", ",", "which", "underpins", "their", "ability", "to", "learn", "hierarchical", "abstractions", "of", "data", "zeiler2014visualizing", ".", "While", "this", "invariance", "is", "clearly", "desirable", "for", "high", "-", "level", "vision", "tasks", ",", "it", "can", "hamper", "low", "-", "level", "tasks", ",", "such", "as", "pose", "estimation", "chen2014articulated", ",", "tompson2014joint", "and", "semantic", "segmentation", "-", "where", "we", "want", "precise", "localization", ",", "rather", "than", "abstraction", "of", "spatial", "details", ".", "There", "are", "two", "technical", "hurdles", "in", "the", "application", "of", "DCNNs", "to", "image", "labeling", "tasks", ":", "signal", "downsampling", ",", "and", "spatial", "‘", "insensitivity", "’", "(", "invariance", ")", ".", "The", "first", "problem", "relates", "to", "the", "reduction", "of", "signal", "resolution", "incurred", "by", "the", "repeated", "combination", "of", "max", "-", "pooling", "and", "downsampling", "(", "‘", "striding", "’", ")", "performed", "at", "every", "layer", "of", "standard", "DCNNs", "KrizhevskyNIPS2013", ",", "simonyan2014very", ",", "szegedy2014going", ".", "Instead", ",", "as", "in", "papandreou2014untangling", ",", "we", "employ", "the", "‘", "atrous", "’", "(", "with", "holes", ")", "algorithm", "originally", "developed", "for", "efficiently", "computing", "the", "undecimated", "discrete", "wavelet", "transform", "Mall99", ".", "This", "allows", "efficient", "dense", "computation", "of", "DCNN", "responses", "in", "a", "scheme", "substantially", "simpler", "than", "earlier", "solutions", "to", "this", "problem", "GCMG", "+", "13", ",", "sermanet2013overfeat", ".", "The", "second", "problem", "relates", "to", "the", "fact", "that", "obtaining", "object", "-", "centric", "decisions", "from", "a", "classifier", "requires", "invariance", "to", "spatial", "transformations", ",", "inherently", "limiting", "the", "spatial", "accuracy", "of", "the", "DCNN", "model", ".", "We", "boost", "our", "model", "’s", "ability", "to", "capture", "fine", "details", "by", "employing", "a", "fully", "-", "connected", "Conditional", "Random", "Field", "(", "CRF", ")", ".", "Conditional", "Random", "Fields", "have", "been", "broadly", "used", "in", "semantic", "segmentation", "to", "combine", "class", "scores", "computed", "by", "multi", "-", "way", "classifiers", "with", "the", "low", "-", "level", "information", "captured", "by", "the", "local", "interactions", "of", "pixels", "and", "edges", "rother2004grabcut", ",", "shotton2009textonboost", "or", "superpixels", "lucchi2011spatial", ".", "Even", "though", "works", "of", "increased", "sophistication", "have", "been", "proposed", "to", "model", "the", "hierarchical", "dependency", "he2004multiscale", ",", "ladicky2009associative", ",", "lempitsky2011pylon", "and", "/", "or", "high", "-", "order", "dependencies", "of", "segments", "delong2012fast", ",", "gonfaus2010harmony", ",", "kohli2009robust", ",", "CPY13", ",", "Wang15", ",", "we", "use", "the", "fully", "connected", "pairwise", "CRF", "proposed", "by", "krahenbuhl2011efficient", "for", "its", "efficient", "computation", ",", "and", "ability", "to", "capture", "fine", "edge", "details", "while", "also", "catering", "for", "long", "range", "dependencies", ".", "That", "model", "was", "shown", "in", "krahenbuhl2011efficient", "to", "largely", "improve", "the", "performance", "of", "a", "boosting", "-", "based", "pixel", "-", "level", "classifier", ",", "and", "in", "our", "work", "we", "demonstrate", "that", "it", "leads", "to", "state", "-", "of", "-", "the", "-", "art", "results", "when", "coupled", "with", "a", "DCNN", "-", "based", "pixel", "-", "level", "classifier", ".", "The", "three", "main", "advantages", "of", "our", "“", "DeepLab", "”", "system", "are", "(", "i", ")", "speed", ":", "by", "virtue", "of", "the", "‘", "atrous", "’", "algorithm", ",", "our", "dense", "DCNN", "operates", "at", "8", "fps", ",", "while", "Mean", "Field", "Inference", "for", "the", "fully", "-", "connected", "CRF", "requires", "0.5", "second", ",", "(", "ii", ")", "accuracy", ":", "we", "obtain", "state", "-", "of", "-", "the", "-", "art", "results", "on", "the", "PASCAL", "semantic", "segmentation", "challenge", ",", "outperforming", "the", "second", "-", "best", "approach", "of", "mostajabi2014feedforward", "by", "a", "margin", "of", "7.2", "and", "(", "iii", ")", "simplicity", ":", "our", "system", "is", "composed", "of", "a", "cascade", "of", "two", "fairly", "well", "-", "established", "modules", ",", "DCNNs", "and", "CRFs", ".", "section", ":", "Related", "Work", "Our", "system", "works", "directly", "on", "the", "pixel", "representation", ",", "similarly", "to", "long2014fully", ".", "This", "is", "in", "contrast", "to", "the", "two", "-", "stage", "approaches", "that", "are", "now", "most", "common", "in", "semantic", "segmentation", "with", "DCNNs", ":", "such", "techniques", "typically", "use", "a", "cascade", "of", "bottom", "-", "up", "image", "segmentation", "and", "DCNN", "-", "based", "region", "classification", ",", "which", "makes", "the", "system", "commit", "to", "potential", "errors", "of", "the", "front", "-", "end", "segmentation", "system", ".", "For", "instance", ",", "the", "bounding", "box", "proposals", "and", "masked", "regions", "delivered", "by", "arbelaez2014multiscale", ",", "Uijlings13", "are", "used", "in", "girshick2014rcnn", "and", "hariharan2014simultaneous", "as", "inputs", "to", "a", "DCNN", "to", "introduce", "shape", "information", "into", "the", "classification", "process", ".", "Similarly", ",", "the", "authors", "of", "mostajabi2014feedforward", "rely", "on", "a", "superpixel", "representation", ".", "A", "celebrated", "non", "-", "DCNN", "precursor", "to", "these", "works", "is", "the", "second", "order", "pooling", "method", "of", "carreira2012semantic", "which", "also", "assigns", "labels", "to", "the", "regions", "proposals", "delivered", "by", "carreira2012cpmc", ".", "Understanding", "the", "perils", "of", "committing", "to", "a", "single", "segmentation", ",", "the", "authors", "of", "cogswell2014combining", "build", "on", "yadollahpour2013discriminative", "to", "explore", "a", "diverse", "set", "of", "CRF", "-", "based", "segmentation", "proposals", ",", "computed", "also", "by", "carreira2012cpmc", ".", "These", "segmentation", "proposals", "are", "then", "re", "-", "ranked", "according", "to", "a", "DCNN", "trained", "in", "particular", "for", "this", "reranking", "task", ".", "Even", "though", "this", "approach", "explicitly", "tries", "to", "handle", "the", "temperamental", "nature", "of", "a", "front", "-", "end", "segmentation", "algorithm", ",", "there", "is", "still", "no", "explicit", "exploitation", "of", "the", "DCNN", "scores", "in", "the", "CRF", "-", "based", "segmentation", "algorithm", ":", "the", "DCNN", "is", "only", "applied", "post", "-", "hoc", ",", "while", "it", "would", "make", "sense", "to", "directly", "try", "to", "use", "its", "results", "during", "segmentation", ".", "Moving", "towards", "works", "that", "lie", "closer", "to", "our", "approach", ",", "several", "other", "researchers", "have", "considered", "the", "use", "of", "convolutionally", "computed", "DCNN", "features", "for", "dense", "image", "labeling", ".", "Among", "the", "first", "have", "been", "farabet2013learning", "who", "apply", "DCNNs", "at", "multiple", "image", "resolutions", "and", "then", "employ", "a", "segmentation", "tree", "to", "smooth", "the", "prediction", "results", ";", "more", "recently", ",", "hariharan2014hypercolumns", "propose", "to", "concatenate", "the", "computed", "inter", "-", "mediate", "feature", "maps", "within", "the", "DCNNs", "for", "pixel", "classification", ",", "and", "dai2014convolutional", "propose", "to", "pool", "the", "inter", "-", "mediate", "feature", "maps", "by", "region", "proposals", ".", "Even", "though", "these", "works", "still", "employ", "segmentation", "algorithms", "that", "are", "decoupled", "from", "the", "DCNN", "classifier", "’s", "results", ",", "we", "believe", "it", "is", "advantageous", "that", "segmentation", "is", "only", "used", "at", "a", "later", "stage", ",", "avoiding", "the", "commitment", "to", "premature", "decisions", ".", "More", "recently", ",", "the", "segmentation", "-", "free", "techniques", "of", "long2014fully", ",", "eigen2014predicting", "directly", "apply", "DCNNs", "to", "the", "whole", "image", "in", "a", "sliding", "window", "fashion", ",", "replacing", "the", "last", "fully", "connected", "layers", "of", "a", "DCNN", "by", "convolutional", "layers", ".", "In", "order", "to", "deal", "with", "the", "spatial", "localization", "issues", "outlined", "in", "the", "beginning", "of", "the", "introduction", ",", "long2014fully", "upsample", "and", "concatenate", "the", "scores", "from", "inter", "-", "mediate", "feature", "maps", ",", "while", "eigen2014predicting", "refine", "the", "prediction", "result", "from", "coarse", "to", "fine", "by", "propagating", "the", "coarse", "results", "to", "another", "DCNN", ".", "The", "main", "difference", "between", "our", "model", "and", "other", "state", "-", "of", "-", "the", "-", "art", "models", "is", "the", "combination", "of", "pixel", "-", "level", "CRFs", "and", "DCNN", "-", "based", "‘", "unary", "terms", "’", ".", "Focusing", "on", "the", "closest", "works", "in", "this", "direction", ",", "cogswell2014combining", "use", "CRFs", "as", "a", "proposal", "mechanism", "for", "a", "DCNN", "-", "based", "reranking", "system", ",", "while", "farabet2013learning", "treat", "superpixels", "as", "nodes", "for", "a", "local", "pairwise", "CRF", "and", "use", "graph", "-", "cuts", "for", "discrete", "inference", ";", "as", "such", "their", "results", "can", "be", "limited", "by", "errors", "in", "superpixel", "computations", ",", "while", "ignoring", "long", "-", "range", "superpixel", "dependencies", ".", "Our", "approach", "instead", "treats", "every", "pixel", "as", "a", "CRF", "node", ",", "exploits", "long", "-", "range", "dependencies", ",", "and", "uses", "CRF", "inference", "to", "directly", "optimize", "a", "DCNN", "-", "driven", "cost", "function", ".", "We", "note", "that", "mean", "field", "had", "been", "extensively", "studied", "for", "traditional", "image", "segmentation", "/", "edge", "detection", "tasks", ",", "e.g", ".", ",", "geiger1991parallel", ",", "geiger1991common", ",", "kokkinos2008computational", ",", "but", "recently", "krahenbuhl2011efficient", "showed", "that", "the", "inference", "can", "be", "very", "efficient", "for", "fully", "connected", "CRF", "and", "particularly", "effective", "in", "the", "context", "of", "semantic", "segmentation", ".", "After", "the", "first", "version", "of", "our", "manuscript", "was", "made", "publicly", "available", ",", "it", "came", "to", "our", "attention", "that", "two", "other", "groups", "have", "independently", "and", "concurrently", "pursued", "a", "very", "similar", "direction", ",", "combining", "DCNNs", "and", "densely", "connected", "CRFs", "bell2014material", ",", "zheng2015crfrnn", ".", "There", "are", "several", "differences", "in", "technical", "aspects", "of", "the", "respective", "models", ".", "bell2014material", "focus", "on", "the", "problem", "of", "material", "classification", ",", "while", "zheng2015crfrnn", "unroll", "the", "CRF", "mean", "-", "field", "inference", "steps", "to", "convert", "the", "whole", "system", "into", "an", "end", "-", "to", "-", "end", "trainable", "feed", "-", "forward", "network", ".", "We", "have", "updated", "our", "proposed", "“", "DeepLab", "”", "system", "with", "much", "improved", "methods", "and", "results", "in", "our", "latest", "work", "chen2016deeplab", ".", "We", "refer", "the", "interested", "reader", "to", "the", "paper", "for", "details", ".", "section", ":", "Convolutional", "Neural", "Networks", "for", "Dense", "Image", "Labeling", "Herein", "we", "describe", "how", "we", "have", "re", "-", "purposed", "and", "finetuned", "the", "publicly", "available", "Imagenet", "-", "pretrained", "state", "-", "of", "-", "art", "16", "-", "layer", "classification", "network", "of", "simonyan2014very", "(", "VGG", "-", "16", ")", "into", "an", "efficient", "and", "effective", "dense", "feature", "extractor", "for", "our", "dense", "semantic", "image", "segmentation", "system", ".", "subsection", ":", "Efficient", "Dense", "Sliding", "Window", "Feature", "Extraction", "with", "the", "Hole", "Algorithm", "Dense", "spatial", "score", "evaluation", "is", "instrumental", "in", "the", "success", "of", "our", "dense", "CNN", "feature", "extractor", ".", "As", "a", "first", "step", "to", "implement", "this", ",", "we", "convert", "the", "fully", "-", "connected", "layers", "of", "VGG", "-", "16", "into", "convolutional", "ones", "and", "run", "the", "network", "in", "a", "convolutional", "fashion", "on", "the", "image", "at", "its", "original", "resolution", ".", "However", "this", "is", "not", "enough", "as", "it", "yields", "very", "sparsely", "computed", "detection", "scores", "(", "with", "a", "stride", "of", "32", "pixels", ")", ".", "To", "compute", "scores", "more", "densely", "at", "our", "target", "stride", "of", "8", "pixels", ",", "we", "develop", "a", "variation", "of", "the", "method", "previously", "employed", "by", "GCMG", "+", "13", ",", "sermanet2013overfeat", ".", "We", "skip", "subsampling", "after", "the", "last", "two", "max", "-", "pooling", "layers", "in", "the", "network", "of", "simonyan2014very", "and", "modify", "the", "convolutional", "filters", "in", "the", "layers", "that", "follow", "them", "by", "introducing", "zeros", "to", "increase", "their", "length", "(", "in", "the", "last", "three", "convolutional", "layers", "and", "in", "the", "first", "fully", "connected", "layer", ")", ".", "We", "can", "implement", "this", "more", "efficiently", "by", "keeping", "the", "filters", "intact", "and", "instead", "sparsely", "sample", "the", "feature", "maps", "on", "which", "they", "are", "applied", "on", "using", "an", "input", "stride", "of", "2", "or", "4", "pixels", ",", "respectively", ".", "This", "approach", ",", "illustrated", "in", "Fig", ".", "[", "reference", "]", "is", "known", "as", "the", "‘", "hole", "algorithm", "’", "(", "‘", "atrous", "algorithm", "’", ")", "and", "has", "been", "developed", "before", "for", "efficient", "computation", "of", "the", "undecimated", "wavelet", "transform", "Mall99", ".", "We", "have", "implemented", "this", "within", "the", "Caffe", "framework", "jia2014caffe", "by", "adding", "to", "the", "im2col", "function", "(", "it", "converts", "multi", "-", "channel", "feature", "maps", "to", "vectorized", "patches", ")", "the", "option", "to", "sparsely", "sample", "the", "underlying", "feature", "map", ".", "This", "approach", "is", "generally", "applicable", "and", "allows", "us", "to", "efficiently", "compute", "dense", "CNN", "feature", "maps", "at", "any", "target", "subsampling", "rate", "without", "introducing", "any", "approximations", ".", "We", "finetune", "the", "model", "weights", "of", "the", "Imagenet", "-", "pretrained", "VGG", "-", "16", "network", "to", "adapt", "it", "to", "the", "image", "classification", "task", "in", "a", "straightforward", "fashion", ",", "following", "the", "procedure", "of", "long2014fully", ".", "We", "replace", "the", "1000", "-", "way", "Imagenet", "classifier", "in", "the", "last", "layer", "of", "VGG", "-", "16", "with", "a", "21", "-", "way", "one", ".", "Our", "loss", "function", "is", "the", "sum", "of", "cross", "-", "entropy", "terms", "for", "each", "spatial", "position", "in", "the", "CNN", "output", "map", "(", "subsampled", "by", "8", "compared", "to", "the", "original", "image", ")", ".", "All", "positions", "and", "labels", "are", "equally", "weighted", "in", "the", "overall", "loss", "function", ".", "Our", "targets", "are", "the", "ground", "truth", "labels", "(", "subsampled", "by", "8", ")", ".", "We", "optimize", "the", "objective", "function", "with", "respect", "to", "the", "weights", "at", "all", "network", "layers", "by", "the", "standard", "SGD", "procedure", "of", "KrizhevskyNIPS2013", ".", "During", "testing", ",", "we", "need", "class", "score", "maps", "at", "the", "original", "image", "resolution", ".", "As", "illustrated", "in", "Figure", "[", "reference", "]", "and", "further", "elaborated", "in", "Section", "[", "reference", "]", ",", "the", "class", "score", "maps", "(", "corresponding", "to", "log", "-", "probabilities", ")", "are", "quite", "smooth", ",", "which", "allows", "us", "to", "use", "simple", "bilinear", "interpolation", "to", "increase", "their", "resolution", "by", "a", "factor", "of", "8", "at", "a", "negligible", "computational", "cost", ".", "Note", "that", "the", "method", "of", "long2014fully", "does", "not", "use", "the", "hole", "algorithm", "and", "produces", "very", "coarse", "scores", "(", "subsampled", "by", "a", "factor", "of", "32", ")", "at", "the", "CNN", "output", ".", "This", "forced", "them", "to", "use", "learned", "upsampling", "layers", ",", "significantly", "increasing", "the", "complexity", "and", "training", "time", "of", "their", "system", ":", "Fine", "-", "tuning", "our", "network", "on", "PASCAL", "VOC", "2012", "takes", "about", "10", "hours", ",", "while", "they", "report", "a", "training", "time", "of", "several", "days", "(", "both", "timings", "on", "a", "modern", "GPU", ")", ".", "subsection", ":", "Controlling", "the", "Receptive", "Field", "Size", "and", "Accelerating", "Dense", "Computation", "with", "Convolutional", "Nets", "Another", "key", "ingredient", "in", "re", "-", "purposing", "our", "network", "for", "dense", "score", "computation", "is", "explicitly", "controlling", "the", "network", "’s", "receptive", "field", "size", ".", "Most", "recent", "DCNN", "-", "based", "image", "recognition", "methods", "rely", "on", "networks", "pre", "-", "trained", "on", "the", "Imagenet", "large", "-", "scale", "classification", "task", ".", "These", "networks", "typically", "have", "large", "receptive", "field", "size", ":", "in", "the", "case", "of", "the", "VGG", "-", "16", "net", "we", "consider", ",", "its", "receptive", "field", "is", "(", "with", "zero", "-", "padding", ")", "and", "pixels", "if", "the", "net", "is", "applied", "convolutionally", ".", "After", "converting", "the", "network", "to", "a", "fully", "convolutional", "one", ",", "the", "first", "fully", "connected", "layer", "has", "4", ",", "096", "filters", "of", "large", "spatial", "size", "and", "becomes", "the", "computational", "bottleneck", "in", "our", "dense", "score", "map", "computation", ".", "We", "have", "addressed", "this", "practical", "problem", "by", "spatially", "subsampling", "(", "by", "simple", "decimation", ")", "the", "first", "FC", "layer", "to", "(", "or", ")", "spatial", "size", ".", "This", "has", "reduced", "the", "receptive", "field", "of", "the", "network", "down", "to", "(", "with", "zero", "-", "padding", ")", "or", "(", "in", "convolutional", "mode", ")", "and", "has", "reduced", "computation", "time", "for", "the", "first", "FC", "layer", "by", "times", ".", "Using", "our", "Caffe", "-", "based", "implementation", "and", "a", "Titan", "GPU", ",", "the", "resulting", "VGG", "-", "derived", "network", "is", "very", "efficient", ":", "Given", "a", "input", "image", ",", "it", "produces", "dense", "raw", "feature", "scores", "at", "the", "top", "of", "the", "network", "at", "a", "rate", "of", "about", "8", "frames", "/", "sec", "during", "testing", ".", "The", "speed", "during", "training", "is", "3", "frames", "/", "sec", ".", "We", "have", "also", "successfully", "experimented", "with", "reducing", "the", "number", "of", "channels", "at", "the", "fully", "connected", "layers", "from", "4", ",", "096", "down", "to", "1", ",", "024", ",", "considerably", "further", "decreasing", "computation", "time", "and", "memory", "footprint", "without", "sacrificing", "performance", ",", "as", "detailed", "in", "Section", "[", "reference", "]", ".", "Using", "smaller", "networks", "such", "as", "KrizhevskyNIPS2013", "could", "allow", "video", "-", "rate", "test", "-", "time", "dense", "feature", "computation", "even", "on", "light", "-", "weight", "GPUs", ".", "section", ":", "Detailed", "Boundary", "Recovery", ":", "Fully", "-", "Connected", "Conditional", "Random", "Fields", "and", "Multi", "-", "scale", "Prediction", "subsection", ":", "Deep", "Convolutional", "Networks", "and", "the", "Localization", "Challenge", "As", "illustrated", "in", "Figure", "[", "reference", "]", ",", "DCNN", "score", "maps", "can", "reliably", "predict", "the", "presence", "and", "rough", "position", "of", "objects", "in", "an", "image", "but", "are", "less", "well", "suited", "for", "pin", "-", "pointing", "their", "exact", "outline", ".", "There", "is", "a", "natural", "trade", "-", "off", "between", "classification", "accuracy", "and", "localization", "accuracy", "with", "convolutional", "networks", ":", "Deeper", "models", "with", "multiple", "max", "-", "pooling", "layers", "have", "proven", "most", "successful", "in", "classification", "tasks", ",", "however", "their", "increased", "invariance", "and", "large", "receptive", "fields", "make", "the", "problem", "of", "inferring", "position", "from", "the", "scores", "at", "their", "top", "output", "levels", "more", "challenging", ".", "Recent", "work", "has", "pursued", "two", "directions", "to", "address", "this", "localization", "challenge", ".", "The", "first", "approach", "is", "to", "harness", "information", "from", "multiple", "layers", "in", "the", "convolutional", "network", "in", "order", "to", "better", "estimate", "the", "object", "boundaries", "long2014fully", ",", "eigen2014predicting", ".", "The", "second", "approach", "is", "to", "employ", "a", "super", "-", "pixel", "representation", ",", "essentially", "delegating", "the", "localization", "task", "to", "a", "low", "-", "level", "segmentation", "method", ".", "This", "route", "is", "followed", "by", "the", "very", "successful", "recent", "method", "of", "mostajabi2014feedforward", ".", "In", "Section", "[", "reference", "]", ",", "we", "pursue", "a", "novel", "alternative", "direction", "based", "on", "coupling", "the", "recognition", "capacity", "of", "DCNNs", "and", "the", "fine", "-", "grained", "localization", "accuracy", "of", "fully", "connected", "CRFs", "and", "show", "that", "it", "is", "remarkably", "successful", "in", "addressing", "the", "localization", "challenge", ",", "producing", "accurate", "semantic", "segmentation", "results", "and", "recovering", "object", "boundaries", "at", "a", "level", "of", "detail", "that", "is", "well", "beyond", "the", "reach", "of", "existing", "methods", ".", "subsection", ":", "Fully", "-", "Connected", "Conditional", "Random", "Fields", "for", "Accurate", "Localization", "Traditionally", ",", "conditional", "random", "fields", "(", "CRFs", ")", "have", "been", "employed", "to", "smooth", "noisy", "segmentation", "maps", "rother2004grabcut", ",", "kohli2009robust", ".", "Typically", "these", "models", "contain", "energy", "terms", "that", "couple", "neighboring", "nodes", ",", "favoring", "same", "-", "label", "assignments", "to", "spatially", "proximal", "pixels", ".", "Qualitatively", ",", "the", "primary", "function", "of", "these", "short", "-", "range", "CRFs", "has", "been", "to", "clean", "up", "the", "spurious", "predictions", "of", "weak", "classifiers", "built", "on", "top", "of", "local", "hand", "-", "engineered", "features", ".", "Compared", "to", "these", "weaker", "classifiers", ",", "modern", "DCNN", "architectures", "such", "as", "the", "one", "we", "use", "in", "this", "work", "produce", "score", "maps", "and", "semantic", "label", "predictions", "which", "are", "qualitatively", "different", ".", "As", "illustrated", "in", "Figure", "[", "reference", "]", ",", "the", "score", "maps", "are", "typically", "quite", "smooth", "and", "produce", "homogeneous", "classification", "results", ".", "In", "this", "regime", ",", "using", "short", "-", "range", "CRFs", "can", "be", "detrimental", ",", "as", "our", "goal", "should", "be", "to", "recover", "detailed", "local", "structure", "rather", "than", "further", "smooth", "it", ".", "Using", "contrast", "-", "sensitive", "potentials", "rother2004grabcut", "in", "conjunction", "to", "local", "-", "range", "CRFs", "can", "potentially", "improve", "localization", "but", "still", "miss", "thin", "-", "structures", "and", "typically", "requires", "solving", "an", "expensive", "discrete", "optimization", "problem", ".", "To", "overcome", "these", "limitations", "of", "short", "-", "range", "CRFs", ",", "we", "integrate", "into", "our", "system", "the", "fully", "connected", "CRF", "model", "of", "krahenbuhl2011efficient", ".", "The", "model", "employs", "the", "energy", "function", "where", "is", "the", "label", "assignment", "for", "pixels", ".", "We", "use", "as", "unary", "potential", ",", "where", "is", "the", "label", "assignment", "probability", "at", "pixel", "as", "computed", "by", "DCNN", ".", "The", "pairwise", "potential", "is", ",", "where", ",", "and", "zero", "otherwise", "(", "i.e", ".", ",", "Potts", "Model", ")", ".", "There", "is", "one", "pairwise", "term", "for", "each", "pair", "of", "pixels", "and", "in", "the", "image", "no", "matter", "how", "far", "from", "each", "other", "they", "lie", ",", "i.e", ".", "the", "model", "’s", "factor", "graph", "is", "fully", "connected", ".", "Each", "is", "the", "Gaussian", "kernel", "depends", "on", "features", "(", "denoted", "as", ")", "extracted", "for", "pixel", "and", "and", "is", "weighted", "by", "parameter", ".", "We", "adopt", "bilateral", "position", "and", "color", "terms", ",", "specifically", ",", "the", "kernels", "are", "where", "the", "first", "kernel", "depends", "on", "both", "pixel", "positions", "(", "denoted", "as", ")", "and", "pixel", "color", "intensities", "(", "denoted", "as", ")", ",", "and", "the", "second", "kernel", "only", "depends", "on", "pixel", "positions", ".", "The", "hyper", "parameters", ",", "and", "control", "the", "“", "scale", "”", "of", "the", "Gaussian", "kernels", ".", "Crucially", ",", "this", "model", "is", "amenable", "to", "efficient", "approximate", "probabilistic", "inference", "krahenbuhl2011efficient", ".", "The", "message", "passing", "updates", "under", "a", "fully", "decomposable", "mean", "field", "approximation", "can", "be", "expressed", "as", "convolutions", "with", "a", "Gaussian", "kernel", "in", "feature", "space", ".", "High", "-", "dimensional", "filtering", "algorithms", "adams2010fast", "significantly", "speed", "-", "up", "this", "computation", "resulting", "in", "an", "algorithm", "that", "is", "very", "fast", "in", "practice", ",", "less", "that", "0.5", "sec", "on", "average", "for", "Pascal", "VOC", "images", "using", "the", "publicly", "available", "implementation", "of", "krahenbuhl2011efficient", ".", "subsection", ":", "Multi", "-", "Scale", "Prediction", "Following", "the", "promising", "recent", "results", "of", "hariharan2014hypercolumns", ",", "long2014fully", "we", "have", "also", "explored", "a", "multi", "-", "scale", "prediction", "method", "to", "increase", "the", "boundary", "localization", "accuracy", ".", "Specifically", ",", "we", "attach", "to", "the", "input", "image", "and", "the", "output", "of", "each", "of", "the", "first", "four", "max", "pooling", "layers", "a", "two", "-", "layer", "MLP", "(", "first", "layer", ":", "128", "3x3", "convolutional", "filters", ",", "second", "layer", ":", "128", "1x1", "convolutional", "filters", ")", "whose", "feature", "map", "is", "concatenated", "to", "the", "main", "network", "’s", "last", "layer", "feature", "map", ".", "The", "aggregate", "feature", "map", "fed", "into", "the", "softmax", "layer", "is", "thus", "enhanced", "by", "5", "*", "128", "=", "640", "channels", ".", "We", "only", "adjust", "the", "newly", "added", "weights", ",", "keeping", "the", "other", "network", "parameters", "to", "the", "values", "learned", "by", "the", "method", "of", "Section", "[", "reference", "]", ".", "As", "discussed", "in", "the", "experimental", "section", ",", "introducing", "these", "extra", "direct", "connections", "from", "fine", "-", "resolution", "layers", "improves", "localization", "performance", ",", "yet", "the", "effect", "is", "not", "as", "dramatic", "as", "the", "one", "obtained", "with", "the", "fully", "-", "connected", "CRF", ".", "section", ":", "Experimental", "Evaluation", "paragraph", ":", "Dataset", "We", "test", "our", "DeepLab", "model", "on", "the", "PASCAL", "VOC", "2012", "segmentation", "benchmark", "everingham2014pascal", ",", "consisting", "of", "20", "foreground", "object", "classes", "and", "one", "background", "class", ".", "The", "original", "dataset", "contains", ",", ",", "and", "images", "for", "training", ",", "validation", ",", "and", "testing", ",", "respectively", ".", "The", "dataset", "is", "augmented", "by", "the", "extra", "annotations", "provided", "by", "hariharan2011semantic", ",", "resulting", "in", "training", "images", ".", "The", "performance", "is", "measured", "in", "terms", "of", "pixel", "intersection", "-", "over", "-", "union", "(", "IOU", ")", "averaged", "across", "the", "21", "classes", ".", "paragraph", ":", "Training", "We", "adopt", "the", "simplest", "form", "of", "piecewise", "training", ",", "decoupling", "the", "DCNN", "and", "CRF", "training", "stages", ",", "assuming", "the", "unary", "terms", "provided", "by", "the", "DCNN", "are", "fixed", "during", "CRF", "training", ".", "For", "DCNN", "training", "we", "employ", "the", "VGG", "-", "16", "network", "which", "has", "been", "pre", "-", "trained", "on", "ImageNet", ".", "We", "fine", "-", "tuned", "the", "VGG", "-", "16", "network", "on", "the", "VOC", "21", "-", "way", "pixel", "-", "classification", "task", "by", "stochastic", "gradient", "descent", "on", "the", "cross", "-", "entropy", "loss", "function", ",", "as", "described", "in", "Section", "[", "reference", "]", ".", "We", "use", "a", "mini", "-", "batch", "of", "20", "images", "and", "initial", "learning", "rate", "of", "(", "for", "the", "final", "classifier", "layer", ")", ",", "multiplying", "the", "learning", "rate", "by", "0.1", "at", "every", "2000", "iterations", ".", "We", "use", "momentum", "of", "and", "a", "weight", "decay", "of", ".", "After", "the", "DCNN", "has", "been", "fine", "-", "tuned", ",", "we", "cross", "-", "validate", "the", "parameters", "of", "the", "fully", "connected", "CRF", "model", "in", "Eq", ".", "(", "[", "reference", "]", ")", "along", "the", "lines", "of", "krahenbuhl2011efficient", ".", "We", "use", "the", "default", "values", "of", "and", "and", "we", "search", "for", "the", "best", "values", "of", ",", ",", "and", "by", "cross", "-", "validation", "on", "a", "small", "subset", "of", "the", "validation", "set", "(", "we", "use", "100", "images", ")", ".", "We", "employ", "coarse", "-", "to", "-", "fine", "search", "scheme", ".", "Specifically", ",", "the", "initial", "search", "range", "of", "the", "parameters", "are", ",", "and", "(", "MATLAB", "notation", ")", ",", "and", "then", "we", "refine", "the", "search", "step", "sizes", "around", "the", "first", "round", "’s", "best", "values", ".", "We", "fix", "the", "number", "of", "mean", "field", "iterations", "to", "10", "for", "all", "reported", "experiments", ".", "paragraph", ":", "Evaluation", "on", "Validation", "set", "We", "conduct", "the", "majority", "of", "our", "evaluations", "on", "the", "PASCAL", "‘", "val", "’", "set", ",", "training", "our", "model", "on", "the", "augmented", "PASCAL", "‘", "train", "’", "set", ".", "As", "shown", "in", "Tab", ".", "[", "reference", "]", "(", "a", ")", ",", "incorporating", "the", "fully", "connected", "CRF", "to", "our", "model", "(", "denoted", "by", "DeepLab", "-", "CRF", ")", "yields", "a", "substantial", "performance", "boost", ",", "about", "4", "%", "improvement", "over", "DeepLab", ".", "We", "note", "that", "the", "work", "of", "krahenbuhl2011efficient", "improved", "the", "result", "of", "TextonBoost", "shotton2009textonboost", "to", ",", "which", "makes", "the", "improvement", "we", "report", "here", "(", "from", "to", ")", "all", "the", "more", "impressive", ".", "Turning", "to", "qualitative", "results", ",", "we", "provide", "visual", "comparisons", "between", "DeepLab", "and", "DeepLab", "-", "CRF", "in", "Fig", ".", "[", "reference", "]", ".", "Employing", "a", "fully", "connected", "CRF", "significantly", "improves", "the", "results", ",", "allowing", "the", "model", "to", "accurately", "capture", "intricate", "object", "boundaries", ".", "paragraph", ":", "Multi", "-", "Scale", "features", "We", "also", "exploit", "the", "features", "from", "the", "intermediate", "layers", ",", "similar", "to", "hariharan2014hypercolumns", ",", "long2014fully", ".", "As", "shown", "in", "Tab", ".", "[", "reference", "]", "(", "a", ")", ",", "adding", "the", "multi", "-", "scale", "features", "to", "our", "DeepLab", "model", "(", "denoted", "as", "DeepLab", "-", "MSc", ")", "improves", "about", "performance", ",", "and", "further", "incorporating", "the", "fully", "connected", "CRF", "(", "denoted", "as", "DeepLab", "-", "MSc", "-", "CRF", ")", "yields", "about", "4", "%", "improvement", ".", "The", "qualitative", "comparisons", "between", "DeepLab", "and", "DeepLab", "-", "MSc", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "Leveraging", "the", "multi", "-", "scale", "features", "can", "slightly", "refine", "the", "object", "boundaries", ".", "paragraph", ":", "Field", "of", "View", "The", "‘", "atrous", "algorithm", "’", "we", "employed", "allows", "us", "to", "arbitrarily", "control", "the", "Field", "-", "of", "-", "View", "(", "FOV", ")", "of", "the", "models", "by", "adjusting", "the", "input", "stride", ",", "as", "illustrated", "in", "Fig", ".", "[", "reference", "]", ".", "In", "Tab", ".", "[", "reference", "]", ",", "we", "experiment", "with", "several", "kernel", "sizes", "and", "input", "strides", "at", "the", "first", "fully", "connected", "layer", ".", "The", "method", ",", "DeepLab", "-", "CRF", "-", "7x7", ",", "is", "the", "direct", "modification", "from", "VGG", "-", "16", "net", ",", "where", "the", "kernel", "size", "=", "and", "input", "stride", "=", "4", ".", "This", "model", "yields", "performance", "of", "on", "the", "‘", "val", "’", "set", ",", "but", "it", "is", "relatively", "slow", "(", "images", "per", "second", "during", "training", ")", ".", "We", "have", "improved", "model", "speed", "to", "images", "per", "second", "by", "reducing", "the", "kernel", "size", "to", ".", "We", "have", "experimented", "with", "two", "such", "network", "variants", "with", "different", "FOV", "sizes", ",", "DeepLab", "-", "CRF", "and", "DeepLab", "-", "CRF", "-", "4x4", ";", "the", "latter", "has", "large", "FOV", "(", "i.e", ".", ",", "large", "input", "stride", ")", "and", "attains", "better", "performance", ".", "Finally", ",", "we", "employ", "kernel", "size", "and", "input", "stride", "=", "12", ",", "and", "further", "change", "the", "filter", "sizes", "from", "4096", "to", "1024", "for", "the", "last", "two", "layers", ".", "Interestingly", ",", "the", "resulting", "model", ",", "DeepLab", "-", "CRF", "-", "LargeFOV", ",", "matches", "the", "performance", "of", "the", "expensive", "DeepLab", "-", "CRF", "-", "7x7", ".", "At", "the", "same", "time", ",", "it", "is", "times", "faster", "to", "run", "and", "has", "significantly", "fewer", "parameters", "(", "20.5", "M", "instead", "of", "134.3", "M", ")", ".", "The", "performance", "of", "several", "model", "variants", "is", "summarized", "in", "Tab", ".", "[", "reference", "]", ",", "showing", "the", "benefit", "of", "exploiting", "multi", "-", "scale", "features", "and", "large", "FOV", ".", "paragraph", ":", "Mean", "Pixel", "IOU", "along", "Object", "Boundaries", "To", "quantify", "the", "accuracy", "of", "the", "proposed", "model", "near", "object", "boundaries", ",", "we", "evaluate", "the", "segmentation", "accuracy", "with", "an", "experiment", "similar", "to", "kohli2009robust", ",", "krahenbuhl2011efficient", ".", "Specifically", ",", "we", "use", "the", "‘", "void", "’", "label", "annotated", "in", "val", "set", ",", "which", "usually", "occurs", "around", "object", "boundaries", ".", "We", "compute", "the", "mean", "IOU", "for", "those", "pixels", "that", "are", "located", "within", "a", "narrow", "band", "(", "called", "trimap", ")", "of", "‘", "void", "’", "labels", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "exploiting", "the", "multi", "-", "scale", "features", "from", "the", "intermediate", "layers", "and", "refining", "the", "segmentation", "results", "by", "a", "fully", "connected", "CRF", "significantly", "improve", "the", "results", "around", "object", "boundaries", ".", "paragraph", ":", "Comparison", "with", "State", "-", "of", "-", "art", "In", "Fig", ".", "[", "reference", "]", ",", "we", "qualitatively", "compare", "our", "proposed", "model", ",", "DeepLab", "-", "CRF", ",", "with", "two", "state", "-", "of", "-", "art", "models", ":", "FCN", "-", "8s", "long2014fully", "and", "TTI", "-", "Zoomout", "-", "16", "mostajabi2014feedforward", "on", "the", "‘", "val", "’", "set", "(", "the", "results", "are", "extracted", "from", "their", "papers", ")", ".", "Our", "model", "is", "able", "to", "capture", "the", "intricate", "object", "boundaries", ".", "paragraph", ":", "Reproducibility", "We", "have", "implemented", "the", "proposed", "methods", "by", "extending", "the", "excellent", "Caffe", "framework", "jia2014caffe", ".", "We", "share", "our", "source", "code", ",", "configuration", "files", ",", "and", "trained", "models", "that", "allow", "reproducing", "the", "results", "in", "this", "paper", "at", "a", "companion", "web", "site", ".", "paragraph", ":", "Test", "set", "results", "Having", "set", "our", "model", "choices", "on", "the", "validation", "set", ",", "we", "evaluate", "our", "model", "variants", "on", "the", "PASCAL", "VOC", "2012", "official", "‘", "test", "’", "set", ".", "As", "shown", "in", "Tab", ".", "[", "reference", "]", ",", "our", "DeepLab", "-", "CRF", "and", "DeepLab", "-", "MSc", "-", "CRF", "models", "achieve", "performance", "of", "and", "mean", "IOU", ",", "respectively", ".", "Our", "models", "outperform", "all", "the", "other", "state", "-", "of", "-", "the", "-", "art", "models", "(", "specifically", ",", "TTI", "-", "Zoomout", "-", "16", "mostajabi2014feedforward", ",", "FCN", "-", "8s", "long2014fully", ",", "and", "MSRA", "-", "CFM", "dai2014convolutional", ")", ".", "When", "we", "increase", "the", "FOV", "of", "the", "models", ",", "DeepLab", "-", "CRF", "-", "LargeFOV", "yields", "performance", "of", ",", "the", "same", "as", "DeepLab", "-", "CRF", "-", "7x7", ",", "while", "its", "training", "speed", "is", "faster", ".", "Furthermore", ",", "our", "best", "model", ",", "DeepLab", "-", "MSc", "-", "CRF", "-", "LargeFOV", ",", "attains", "the", "best", "performance", "of", "by", "employing", "both", "multi", "-", "scale", "features", "and", "large", "FOV", ".", "section", ":", "Discussion", "Our", "work", "combines", "ideas", "from", "deep", "convolutional", "neural", "networks", "and", "fully", "-", "connected", "conditional", "random", "fields", ",", "yielding", "a", "novel", "method", "able", "to", "produce", "semantically", "accurate", "predictions", "and", "detailed", "segmentation", "maps", ",", "while", "being", "computationally", "efficient", ".", "Our", "experimental", "results", "show", "that", "the", "proposed", "method", "significantly", "advances", "the", "state", "-", "of", "-", "art", "in", "the", "challenging", "PASCAL", "VOC", "2012", "semantic", "image", "segmentation", "task", ".", "There", "are", "multiple", "aspects", "in", "our", "model", "that", "we", "intend", "to", "refine", ",", "such", "as", "fully", "integrating", "its", "two", "main", "components", "(", "CNN", "and", "CRF", ")", "and", "train", "the", "whole", "system", "in", "an", "end", "-", "to", "-", "end", "fashion", ",", "similar", "to", "Koltun13", ",", "chen2014learning", ",", "zheng2015crfrnn", ".", "We", "also", "plan", "to", "experiment", "with", "more", "datasets", "and", "apply", "our", "method", "to", "other", "sources", "of", "data", "such", "as", "depth", "maps", "or", "videos", ".", "Recently", ",", "we", "have", "pursued", "model", "training", "with", "weakly", "supervised", "annotations", ",", "in", "the", "form", "of", "bounding", "boxes", "or", "image", "-", "level", "labels", "papandreou15weak", ".", "At", "a", "higher", "level", ",", "our", "work", "lies", "in", "the", "intersection", "of", "convolutional", "neural", "networks", "and", "probabilistic", "graphical", "models", ".", "We", "plan", "to", "further", "investigate", "the", "interplay", "of", "these", "two", "powerful", "classes", "of", "methods", "and", "explore", "their", "synergistic", "potential", "for", "solving", "challenging", "computer", "vision", "tasks", ".", "subsection", ":", "Acknowledgments", "This", "work", "was", "partly", "supported", "by", "ARO", "62250", "-", "CS", ",", "NIH", "Grant", "5R01EY022247", "-", "03", ",", "EU", "Project", "RECONFIG", "FP7", "-", "ICT", "-", "600825", "and", "EU", "Project", "MOBOT", "FP7", "-", "ICT", "-", "2011", "-", "600796", ".", "We", "also", "gratefully", "acknowledge", "the", "support", "of", "NVIDIA", "Corporation", "with", "the", "donation", "of", "GPUs", "used", "for", "this", "research", ".", "We", "would", "like", "to", "thank", "the", "anonymous", "reviewers", "for", "their", "detailed", "comments", "and", "constructive", "feedback", ".", "subsection", ":", "Paper", "Revisions", "Here", "we", "present", "the", "list", "of", "major", "paper", "revisions", "for", "the", "convenience", "of", "the", "readers", ".", "paragraph", ":", "v1", "Submission", "to", "ICLR", "2015", ".", "Introduces", "the", "model", "DeepLab", "-", "CRF", ",", "which", "attains", "the", "performance", "of", "on", "PASCAL", "VOC", "2012", "test", "set", ".", "paragraph", ":", "v2", "Rebuttal", "for", "ICLR", "2015", ".", "Adds", "the", "model", "DeepLab", "-", "MSc", "-", "CRF", ",", "which", "incorporates", "multi", "-", "scale", "features", "from", "the", "intermediate", "layers", ".", "DeepLab", "-", "MSc", "-", "CRF", "yields", "the", "performance", "of", "on", "PASCAL", "VOC", "2012", "test", "set", ".", "paragraph", ":", "v3", "Camera", "-", "ready", "for", "ICLR", "2015", ".", "Experiments", "with", "large", "Field", "-", "Of", "-", "View", ".", "On", "PASCAL", "VOC", "2012", "test", "set", ",", "DeepLab", "-", "CRF", "-", "LargeFOV", "achieves", "the", "performance", "of", ".", "When", "exploiting", "both", "mutli", "-", "scale", "features", "and", "large", "FOV", ",", "DeepLab", "-", "MSc", "-", "CRF", "-", "LargeFOV", "attains", "the", "performance", "of", ".", "paragraph", ":", "v4", "Reference", "to", "our", "updated", "“", "DeepLab", "”", "system", "chen2016deeplab", "with", "much", "improved", "results", ".", "bibliography", ":", "References"], "entity": [[2, 5, ["Semantic", "Image", "Segmentation"], "Task"], [6, 9, ["Deep", "Convolutional", "Nets"], "Method"], [10, 13, ["Fully", "Connected", "CRFs"], "Method"], [13, 17, ["Deep", "Convolutional", "Neural", "Networks"], "Method"], [18, 19, ["DCNNs"], "Method"], [29, 33, ["high", "level", "vision", "tasks"], "Task"], [36, 38, ["image", "classification"], "Task"], [39, 41, ["object", "detection"], "Task"], [48, 49, ["DCNNs"], "Method"], [50, 53, ["probabilistic", "graphical", "models"], "Method"], [58, 62, ["pixel", "-", "level", "classification"], "Task"], [81, 82, ["DCNNs"], "Method"], [87, 90, ["accurate", "object", "segmentation"], "Task"], [101, 102, ["DCNNs"], "Method"], [104, 107, ["high", "level", "tasks"], "Task"], [112, 114, ["localization", "property"], "Metric"], [115, 117, ["deep", "networks"], "Method"], [124, 126, ["DCNN", "layer"], "Method"], [128, 133, ["fully", "connected", "Conditional", "Random", "Field"], "Method"], [134, 135, ["CRF"], "Method"], [141, 144, ["DeepLab", "”", "system"], "Method"], [154, 155, ["accuracy"], "Metric"], [204, 209, ["Careful", "network", "re", "-", "purposing"], "Method"], [215, 219, ["’", "hole", "’", "algorithm"], "Method"], [221, 223, ["wavelet", "community"], "Method"], [232, 235, ["frames", "per", "second"], "Metric"], [238, 239, ["GPU"], "Method"], [243, 247, ["Deep", "Convolutional", "Neural", "Networks"], "Method"], [248, 249, ["DCNNs"], "Method"], [257, 259, ["document", "recognition"], "Task"], [270, 275, ["high", "-", "level", "vision", "research"], "Task"], [281, 282, ["DCNNs"], "Method"], [287, 290, ["computer", "vision", "systems"], "Method"], [304, 306, ["image", "classification"], "Task"], [316, 318, ["object", "detection"], "Task"], [320, 324, ["fine", "-", "grained", "categorization"], "Task"], [337, 338, ["DCNNs"], "Method"], [356, 358, ["engineered", "representations"], "Method"], [379, 380, ["DCNNs"], "Method"], [404, 409, ["high", "-", "level", "vision", "tasks"], "Task"], [413, 417, ["low", "-", "level", "tasks"], "Task"], [420, 422, ["pose", "estimation"], "Task"], [426, 428, ["semantic", "segmentation"], "Task"], [432, 434, ["precise", "localization"], "Task"], [451, 452, ["DCNNs"], "Method"], [453, 456, ["image", "labeling", "tasks"], "Task"], [457, 459, ["signal", "downsampling"], "Task"], [475, 479, ["reduction", "of", "signal", "resolution"], "Task"], [485, 488, ["max", "-", "pooling"], "Method"], [489, 490, ["downsampling"], "Method"], [501, 502, ["DCNNs"], "Method"], [518, 525, ["atrous", "’", "(", "with", "holes", ")", "algorithm"], "Method"], [531, 535, ["undecimated", "discrete", "wavelet", "transform"], "Method"], [540, 545, ["dense", "computation", "of", "DCNN", "responses"], "Task"], [556, 559, ["GCMG", "+", "13"], "Method"], [577, 578, ["classifier"], "Method"], [587, 589, ["spatial", "accuracy"], "Metric"], [591, 593, ["DCNN", "model"], "Method"], [607, 613, ["fully", "-", "connected", "Conditional", "Random", "Field"], "Method"], [614, 615, ["CRF"], "Method"], [617, 620, ["Conditional", "Random", "Fields"], "Method"], [625, 627, ["semantic", "segmentation"], "Task"], [633, 637, ["multi", "-", "way", "classifiers"], "Method"], [693, 694, ["CPY13"], "Method"], [740, 747, ["boosting", "-", "based", "pixel", "-", "level", "classifier"], "Method"], [770, 777, ["DCNN", "-", "based", "pixel", "-", "level", "classifier"], "Method"], [785, 788, ["DeepLab", "”", "system"], "Method"], [792, 793, ["speed"], "Metric"], [799, 802, ["atrous", "’", "algorithm"], "Method"], [804, 806, ["dense", "DCNN"], "Method"], [812, 815, ["Mean", "Field", "Inference"], "Method"], [828, 829, ["accuracy"], "Metric"], [842, 846, ["PASCAL", "semantic", "segmentation", "challenge"], "Task"], [864, 865, ["simplicity"], "Metric"], [872, 873, ["cascade"], "Method"], [881, 882, ["DCNNs"], "Method"], [883, 884, ["CRFs"], "Method"], [895, 897, ["pixel", "representation"], "Method"], [908, 912, ["two", "-", "stage", "approaches"], "Method"], [918, 920, ["semantic", "segmentation"], "Task"], [921, 922, ["DCNNs"], "Method"], [936, 941, ["DCNN", "-", "based", "region", "classification"], "Method"], [952, 957, ["front", "-", "end", "segmentation", "system"], "Method"], [983, 984, ["DCNN"], "Method"], [990, 992, ["classification", "process"], "Method"], [1002, 1004, ["superpixel", "representation"], "Method"], [1007, 1011, ["non", "-", "DCNN", "precursor"], "Method"], [1016, 1020, ["second", "order", "pooling", "method"], "Method"], [1079, 1080, ["DCNN"], "Method"], [1085, 1087, ["reranking", "task"], "Task"], [1101, 1106, ["front", "-", "end", "segmentation", "algorithm"], "Method"], [1115, 1117, ["DCNN", "scores"], "Method"], [1126, 1127, ["DCNN"], "Method"], [1147, 1148, ["segmentation"], "Task"], [1167, 1171, ["convolutionally", "computed", "DCNN", "features"], "Method"], [1172, 1175, ["dense", "image", "labeling"], "Task"], [1184, 1185, ["DCNNs"], "Method"], [1193, 1195, ["segmentation", "tree"], "Method"], [1217, 1218, ["DCNNs"], "Method"], [1219, 1221, ["pixel", "classification"], "Task"], [1234, 1236, ["region", "proposals"], "Method"], [1243, 1245, ["segmentation", "algorithms"], "Method"], [1250, 1252, ["DCNN", "classifier"], "Method"], [1261, 1262, ["segmentation"], "Task"], [1281, 1285, ["segmentation", "-", "free", "techniques"], "Method"], [1291, 1292, ["DCNNs"], "Method"], [1305, 1308, ["fully", "connected", "layers"], "Method"], [1310, 1311, ["DCNN"], "Method"], [1312, 1314, ["convolutional", "layers"], "Method"], [1321, 1324, ["spatial", "localization", "issues"], "Task"], [1362, 1363, ["DCNN"], "Method"], [1372, 1380, ["state", "-", "of", "-", "the", "-", "art", "models"], "Method"], [1384, 1388, ["pixel", "-", "level", "CRFs"], "Method"], [1389, 1391, ["DCNN", "-"], "Method"], [1408, 1409, ["CRFs"], "Method"], [1411, 1413, ["proposal", "mechanism"], "Method"], [1415, 1420, ["DCNN", "-", "based", "reranking", "system"], "Method"], [1434, 1437, ["graph", "-", "cuts"], "Method"], [1438, 1440, ["discrete", "inference"], "Task"], [1487, 1492, ["DCNN", "-", "driven", "cost", "function"], "Method"], [1496, 1498, ["mean", "field"], "Method"], [1504, 1506, ["image", "segmentation"], "Task"], [1507, 1510, ["edge", "detection", "tasks"], "Task"], [1526, 1527, ["inference"], "Method"], [1542, 1544, ["semantic", "segmentation"], "Task"], [1577, 1578, ["DCNNs"], "Method"], [1579, 1582, ["densely", "connected", "CRFs"], "Method"], [1604, 1606, ["material", "classification"], "Task"], [1629, 1634, ["trainable", "feed", "-", "forward", "network"], "Method"], [1641, 1644, ["DeepLab", "”", "system"], "Method"], [1669, 1672, ["Convolutional", "Neural", "Networks"], "Method"], [1673, 1676, ["Dense", "Image", "Labeling"], "Task"], [1698, 1703, ["16", "-", "layer", "classification", "network"], "Method"], [1706, 1709, ["VGG", "-", "16"], "Method"], [1715, 1718, ["dense", "feature", "extractor"], "Method"], [1729, 1734, ["Dense", "Sliding", "Window", "Feature", "Extraction"], "Method"], [1736, 1738, ["Hole", "Algorithm"], "Method"], [1738, 1742, ["Dense", "spatial", "score", "evaluation"], "Method"], [1749, 1753, ["dense", "CNN", "feature", "extractor"], "Method"], [1765, 1769, ["fully", "-", "connected", "layers"], "Method"], [1770, 1773, ["VGG", "-", "16"], "Method"], [1774, 1776, ["convolutional", "ones"], "Method"], [1782, 1784, ["convolutional", "fashion"], "Method"], [1837, 1840, ["GCMG", "+", "13"], "Method"], [1850, 1854, ["max", "-", "pooling", "layers"], "Method"], [1862, 1864, ["convolutional", "filters"], "Method"], [1882, 1884, ["convolutional", "layers"], "Method"], [1888, 1891, ["fully", "connected", "layer"], "Method"], [1944, 1946, ["hole", "algorithm"], "Method"], [1949, 1951, ["atrous", "algorithm"], "Method"], [1963, 1966, ["undecimated", "wavelet", "transform"], "Method"], [1974, 1976, ["Caffe", "framework"], "Method"], [1981, 1983, ["im2col", "function"], "Method"], [2023, 2025, ["subsampling", "rate"], "Metric"], [2037, 2044, ["Imagenet", "-", "pretrained", "VGG", "-", "16", "network"], "Method"], [2049, 2052, ["image", "classification", "task"], "Task"], [2066, 2071, ["1000", "-", "way", "Imagenet", "classifier"], "Method"], [2076, 2079, ["VGG", "-", "16"], "Method"], [2081, 2085, ["21", "-", "way", "one"], "Method"], [2087, 2089, ["loss", "function"], "Method"], [2127, 2129, ["loss", "function"], "Metric"], [2146, 2148, ["objective", "function"], "Metric"], [2160, 2162, ["SGD", "procedure"], "Method"], [2216, 2218, ["bilinear", "interpolation"], "Method"], [2230, 2232, ["computational", "cost"], "Metric"], [2243, 2245, ["hole", "algorithm"], "Method"], [2269, 2271, ["upsampling", "layers"], "Method"], [2275, 2276, ["complexity"], "Metric"], [2277, 2279, ["training", "time"], "Metric"], [2289, 2292, ["PASCAL", "VOC", "2012"], "Material"], [2301, 2303, ["training", "time"], "Metric"], [2312, 2313, ["GPU"], "Method"], [2324, 2326, ["Dense", "Computation"], "Task"], [2327, 2329, ["Convolutional", "Nets"], "Method"], [2339, 2342, ["dense", "score", "computation"], "Task"], [2354, 2360, ["DCNN", "-", "based", "image", "recognition", "methods"], "Method"], [2362, 2363, ["networks"], "Method"], [2368, 2374, ["Imagenet", "large", "-", "scale", "classification", "task"], "Task"], [2389, 2393, ["VGG", "-", "16", "net"], "Method"], [2413, 2414, ["convolutionally"], "Method"], [2421, 2424, ["fully", "convolutional", "one"], "Method"], [2427, 2430, ["fully", "connected", "layer"], "Method"], [2446, 2450, ["dense", "score", "map", "computation"], "Task"], [2458, 2460, ["spatially", "subsampling"], "Method"], [2463, 2464, ["decimation"], "Method"], [2466, 2469, ["first", "FC", "layer"], "Method"], [2502, 2504, ["computation", "time"], "Metric"], [2507, 2509, ["FC", "layer"], "Method"], [2514, 2518, ["Caffe", "-", "based", "implementation"], "Method"], [2520, 2522, ["Titan", "GPU"], "Method"], [2525, 2529, ["VGG", "-", "derived", "network"], "Method"], [2556, 2559, ["frames", "/", "sec"], "Metric"], [2565, 2566, ["training"], "Task"], [2568, 2571, ["frames", "/", "sec"], "Metric"], [2601, 2603, ["computation", "time"], "Metric"], [2604, 2606, ["memory", "footprint"], "Metric"], [2619, 2621, ["smaller", "networks"], "Method"], [2632, 2635, ["dense", "feature", "computation"], "Task"], [2644, 2647, ["Detailed", "Boundary", "Recovery"], "Task"], [2648, 2654, ["Fully", "-", "Connected", "Conditional", "Random", "Fields"], "Task"], [2655, 2659, ["Multi", "-", "scale", "Prediction"], "Task"], [2661, 2664, ["Deep", "Convolutional", "Networks"], "Method"], [2666, 2668, ["Localization", "Challenge"], "Task"], [2676, 2679, ["DCNN", "score", "maps"], "Method"], [2713, 2715, ["classification", "accuracy"], "Metric"], [2716, 2718, ["localization", "accuracy"], "Metric"], [2719, 2721, ["convolutional", "networks"], "Method"], [2722, 2724, ["Deeper", "models"], "Method"], [2726, 2730, ["max", "-", "pooling", "layers"], "Method"], [2735, 2737, ["classification", "tasks"], "Task"], [2772, 2774, ["localization", "challenge"], "Task"], [2787, 2789, ["convolutional", "network"], "Method"], [2808, 2812, ["super", "-", "pixel", "representation"], "Method"], [2816, 2818, ["localization", "task"], "Task"], [2820, 2825, ["low", "-", "level", "segmentation", "method"], "Method"], [2855, 2857, ["recognition", "capacity"], "Metric"], [2858, 2859, ["DCNNs"], "Method"], [2861, 2866, ["fine", "-", "grained", "localization", "accuracy"], "Metric"], [2867, 2870, ["fully", "connected", "CRFs"], "Method"], [2880, 2882, ["localization", "challenge"], "Task"], [2885, 2887, ["semantic", "segmentation"], "Task"], [2909, 2915, ["Fully", "-", "Connected", "Conditional", "Random", "Fields"], "Method"], [2916, 2918, ["Accurate", "Localization"], "Task"], [2920, 2923, ["conditional", "random", "fields"], "Method"], [2924, 2925, ["CRFs"], "Method"], [2966, 2970, ["short", "-", "range", "CRFs"], "Method"], [2979, 2981, ["weak", "classifiers"], "Method"], [2995, 2996, ["classifiers"], "Method"], [2998, 3000, ["DCNN", "architectures"], "Method"], [3010, 3012, ["score", "maps"], "Metric"], [3013, 3016, ["semantic", "label", "predictions"], "Task"], [3030, 3032, ["score", "maps"], "Metric"], [3039, 3040, ["classification"], "Task"], [3047, 3051, ["short", "-", "range", "CRFs"], "Method"], [3080, 3084, ["local", "-", "range", "CRFs"], "Method"], [3087, 3088, ["localization"], "Task"], [3100, 3103, ["discrete", "optimization", "problem"], "Task"], [3109, 3113, ["short", "-", "range", "CRFs"], "Method"], [3158, 3159, ["DCNN"], "Method"], [3174, 3176, ["Potts", "Model"], "Method"], [3207, 3209, ["factor", "graph"], "Method"], [3216, 3218, ["Gaussian", "kernel"], "Method"], [3292, 3294, ["Gaussian", "kernels"], "Method"], [3303, 3306, ["approximate", "probabilistic", "inference"], "Task"], [3309, 3312, ["message", "passing", "updates"], "Method"], [3314, 3319, ["fully", "decomposable", "mean", "field", "approximation"], "Method"], [3323, 3324, ["convolutions"], "Method"], [3326, 3328, ["Gaussian", "kernel"], "Method"], [3332, 3337, ["High", "-", "dimensional", "filtering", "algorithms"], "Method"], [3362, 3365, ["Pascal", "VOC", "images"], "Material"], [3375, 3379, ["Multi", "-", "Scale", "Prediction"], "Task"], [3393, 3398, ["multi", "-", "scale", "prediction", "method"], "Method"], [3401, 3404, ["boundary", "localization", "accuracy"], "Metric"], [3422, 3425, ["max", "pooling", "layers"], "Method"], [3428, 3430, ["layer", "MLP"], "Method"], [3435, 3438, ["3x3", "convolutional", "filters"], "Method"], [3444, 3446, ["convolutional", "filters"], "Method"], [3454, 3456, ["main", "network"], "Method"], [3457, 3461, ["last", "layer", "feature", "map"], "Method"], [3469, 3471, ["softmax", "layer"], "Method"], [3526, 3527, ["localization"], "Task"], [3557, 3559, ["DeepLab", "model"], "Method"], [3561, 3564, ["PASCAL", "VOC", "2012"], "Material"], [3590, 3591, ["validation"], "Task"], [3593, 3594, ["testing"], "Task"], [3621, 3627, ["pixel", "intersection", "-", "over", "-", "union"], "Metric"], [3628, 3629, ["IOU"], "Metric"], [3645, 3647, ["piecewise", "training"], "Method"], [3663, 3664, ["DCNN"], "Method"], [3671, 3673, ["DCNN", "training"], "Task"], [3676, 3680, ["VGG", "-", "16", "network"], "Method"], [3694, 3698, ["VGG", "-", "16", "network"], "Method"], [3700, 3708, ["VOC", "21", "-", "way", "pixel", "-", "classification", "task"], "Task"], [3709, 3712, ["stochastic", "gradient", "descent"], "Method"], [3714, 3719, ["cross", "-", "entropy", "loss", "function"], "Metric"], [3739, 3741, ["learning", "rate"], "Metric"], [3746, 3748, ["classifier", "layer"], "Method"], [3752, 3754, ["learning", "rate"], "Metric"], [3773, 3774, ["DCNN"], "Method"], [3825, 3828, ["cross", "-", "validation"], "Method"], [3845, 3852, ["coarse", "-", "to", "-", "fine", "search", "scheme"], "Method"], [3905, 3907, ["Validation", "set"], "Metric"], [3916, 3917, ["PASCAL"], "Material"], [3927, 3933, ["augmented", "PASCAL", "‘", "train", "’", "set"], "Material"], [3957, 3960, ["DeepLab", "-", "CRF"], "Method"], [3972, 3973, ["DeepLab"], "Method"], [4015, 4016, ["DeepLab"], "Method"], [4017, 4020, ["DeepLab", "-", "CRF"], "Method"], [4089, 4091, ["DeepLab", "model"], "Method"], [4094, 4097, ["DeepLab", "-", "MSc"], "Method"], [4112, 4117, ["DeepLab", "-", "MSc", "-", "CRF"], "Method"], [4128, 4129, ["DeepLab"], "Method"], [4130, 4133, ["DeepLab", "-", "MSc"], "Method"], [4162, 4164, ["atrous", "algorithm"], "Method"], [4218, 4221, ["fully", "connected", "layer"], "Method"], [4225, 4230, ["DeepLab", "-", "CRF", "-", "7x7"], "Method"], [4236, 4240, ["VGG", "-", "16", "net"], "Method"], [4280, 4282, ["model", "speed"], "Metric"], [4299, 4301, ["network", "variants"], "Method"], [4306, 4309, ["DeepLab", "-", "CRF"], "Method"], [4310, 4315, ["DeepLab", "-", "CRF", "-", "4x4"], "Method"], [4368, 4373, ["DeepLab", "-", "CRF", "-", "LargeFOV"], "Method"], [4380, 4385, ["DeepLab", "-", "CRF", "-", "7x7"], "Method"], [4441, 4444, ["Mean", "Pixel", "IOU"], "Metric"], [4450, 4451, ["accuracy"], "Metric"], [4462, 4464, ["segmentation", "accuracy"], "Metric"], [4497, 4499, ["mean", "IOU"], "Metric"], [4541, 4542, ["segmentation"], "Task"], [4579, 4582, ["DeepLab", "-", "CRF"], "Method"], [4592, 4595, ["FCN", "-", "8s"], "Method"], [4642, 4645, ["excellent", "Caffe", "framework"], "Method"], [4695, 4696, ["PASCAL"], "Material"], [4696, 4698, ["VOC", "2012"], "Material"], [4714, 4717, ["DeepLab", "-", "CRF"], "Method"], [4718, 4723, ["DeepLab", "-", "MSc", "-", "CRF"], "Method"], [4728, 4730, ["mean", "IOU"], "Metric"], [4757, 4760, ["FCN", "-", "8s"], "Method"], [4778, 4783, ["DeepLab", "-", "CRF", "-", "LargeFOV"], "Method"], [4790, 4795, ["DeepLab", "-", "CRF", "-", "7x7"], "Method"], [4798, 4800, ["training", "speed"], "Metric"], [4809, 4816, ["DeepLab", "-", "MSc", "-", "CRF", "-", "LargeFOV"], "Method"], [4841, 4845, ["deep", "convolutional", "neural", "networks"], "Method"], [4846, 4852, ["fully", "-", "connected", "conditional", "random", "fields"], "Method"], [4860, 4863, ["semantically", "accurate", "predictions"], "Task"], [4864, 4867, ["detailed", "segmentation", "maps"], "Task"], [4922, 4923, ["CNN"], "Method"], [4924, 4925, ["CRF"], "Method"], [4977, 4979, ["model", "training"], "Task"], [5009, 5012, ["convolutional", "neural", "networks"], "Method"], [5013, 5016, ["probabilistic", "graphical", "models"], "Method"], [5039, 5042, ["computer", "vision", "tasks"], "Task"], [5148, 5152, ["model", "DeepLab", "-", "CRF"], "Method"], [5175, 5181, ["model", "DeepLab", "-", "MSc", "-", "CRF"], "Method"], [5193, 5198, ["DeepLab", "-", "MSc", "-", "CRF"], "Method"], [5235, 5240, ["DeepLab", "-", "CRF", "-", "LargeFOV"], "Method"], [5276, 5279, ["DeepLab", "”", "system"], "Method"], [67, 69, ["image", "segmentation"], "Task"], [180, 182, ["image", "segmentation"], "Task"], [187, 188, ["IOU"], "Metric"], [703, 704, ["CRF"], "Method"], [820, 821, ["CRF"], "Method"], [933, 935, ["image", "segmentation"], "Task"], [1057, 1058, ["CRF"], "Method"], [1119, 1120, ["CRF"], "Method"], [1431, 1432, ["CRF"], "Method"], [1470, 1471, ["CRF"], "Method"], [1481, 1482, ["CRF"], "Method"], [1534, 1535, ["CRF"], "Method"], [1611, 1612, ["CRF"], "Method"], [1722, 1724, ["image", "segmentation"], "Task"], [3122, 3123, ["CRF"], "Method"], [3545, 3546, ["CRF"], "Method"], [3652, 3653, ["CRF"], "Method"], [3667, 3668, ["CRF"], "Method"], [3790, 3791, ["CRF"], "Method"], [3950, 3951, ["CRF"], "Method"], [4031, 4032, ["CRF"], "Method"], [4108, 4109, ["CRF"], "Method"], [4547, 4548, ["CRF"], "Method"], [4892, 4895, ["PASCAL", "VOC", "2012"], "Material"], [4896, 4898, ["image", "segmentation"], "Task"], [5159, 5162, ["PASCAL", "VOC", "2012"], "Material"], [5203, 5206, ["PASCAL", "VOC", "2012"], "Material"], [5229, 5232, ["PASCAL", "VOC", "2012"], "Material"], [5256, 5259, ["DeepLab", "-", "MSc"], "Method"], [5260, 5261, ["CRF"], "Method"]]} {"doc_id": "07a9478e87a8304fc3267fa16e83e9f3bbd98b27", "text": ["document", ":", "A", "Decomposable", "Attention", "Model", "for", "Natural", "Language", "Inference", "We", "propose", "a", "simple", "neural", "architecture", "for", "natural", "language", "inference", ".", "Our", "approach", "uses", "attention", "to", "decompose", "the", "problem", "into", "subproblems", "that", "can", "be", "solved", "separately", ",", "thus", "making", "it", "trivially", "parallelizable", ".", "On", "the", "Stanford", "Natural", "Language", "Inference", "(", "SNLI", ")", "dataset", ",", "we", "obtain", "state", "-", "of", "-", "the", "-", "art", "results", "with", "almost", "an", "order", "of", "magnitude", "fewer", "parameters", "than", "previous", "work", "and", "without", "relying", "on", "any", "word", "-", "order", "information", ".", "Adding", "intra", "-", "sentence", "attention", "that", "takes", "a", "minimum", "amount", "of", "order", "into", "account", "yields", "further", "improvements", ".", "section", ":", "Introduction", "Natural", "language", "inference", "(", "NLI", ")", "refers", "to", "the", "problem", "of", "determining", "entailment", "and", "contradiction", "relationships", "between", "a", "premise", "and", "a", "hypothesis", ".", "NLI", "is", "a", "central", "problem", "in", "language", "understanding", "and", "recently", "the", "large", "SNLI", "corpus", "of", "570", "K", "sentence", "pairs", "was", "created", "for", "this", "task", ".", "We", "present", "a", "new", "model", "for", "NLI", "and", "leverage", "this", "corpus", "for", "comparison", "with", "prior", "work", ".", "A", "large", "body", "of", "work", "based", "on", "neural", "networks", "for", "text", "similarity", "tasks", "including", "NLI", "has", "been", "published", "in", "recent", "years", ".", "The", "dominating", "trend", "in", "these", "models", "is", "to", "build", "complex", ",", "deep", "text", "representation", "models", ",", "for", "example", ",", "with", "convolutional", "networks", "or", "long", "short", "-", "term", "memory", "networks", "with", "the", "goal", "of", "deeper", "sentence", "comprehension", ".", "While", "these", "approaches", "have", "yielded", "impressive", "results", ",", "they", "are", "often", "computationally", "very", "expensive", ",", "and", "result", "in", "models", "having", "millions", "of", "parameters", "(", "excluding", "embeddings", ")", ".", "Here", ",", "we", "take", "a", "different", "approach", ",", "arguing", "that", "for", "natural", "language", "inference", "it", "can", "often", "suffice", "to", "simply", "align", "bits", "of", "local", "text", "substructure", "and", "then", "aggregate", "this", "information", ".", "For", "example", ",", "consider", "the", "following", "sentences", ":", "Bob", "is", "in", "his", "room", ",", "but", "because", "of", "the", "thunder", "and", "lightning", "outside", ",", "he", "can", "not", "sleep", ".", "Bob", "is", "awake", ".", "It", "is", "sunny", "outside", ".", "The", "first", "sentence", "is", "complex", "in", "structure", "and", "it", "is", "challenging", "to", "construct", "a", "compact", "representation", "that", "expresses", "its", "entire", "meaning", ".", "However", ",", "it", "is", "fairly", "easy", "to", "conclude", "that", "the", "second", "sentence", "follows", "from", "the", "first", "one", ",", "by", "simply", "aligning", "Bob", "with", "Bob", "and", "can", "not", "sleep", "with", "awake", "and", "recognizing", "that", "these", "are", "synonyms", ".", "Similarly", ",", "one", "can", "conclude", "that", "It", "is", "sunny", "outside", "contradicts", "the", "first", "sentence", ",", "by", "aligning", "thunder", "and", "lightning", "with", "sunny", "and", "recognizing", "that", "these", "are", "most", "likely", "incompatible", ".", "We", "leverage", "this", "intuition", "to", "build", "a", "simpler", "and", "more", "lightweight", "approach", "to", "NLI", "within", "a", "neural", "framework", ";", "with", "considerably", "fewer", "parameters", ",", "our", "model", "outperforms", "more", "complex", "existing", "neural", "architectures", ".", "In", "contrast", "to", "existing", "approaches", ",", "our", "approach", "only", "relies", "on", "alignment", "and", "is", "fully", "computationally", "decomposable", "with", "respect", "to", "the", "input", "text", ".", "An", "overview", "of", "our", "approach", "is", "given", "in", "Figure", "[", "reference", "]", ".", "Given", "two", "sentences", ",", "where", "each", "word", "is", "represented", "by", "an", "embedding", "vector", ",", "we", "first", "create", "a", "soft", "alignment", "matrix", "using", "neural", "attention", ".", "We", "then", "use", "the", "(", "soft", ")", "alignment", "to", "decompose", "the", "task", "into", "subproblems", "that", "are", "solved", "separately", ".", "Finally", ",", "the", "results", "of", "these", "subproblems", "are", "merged", "to", "produce", "the", "final", "classification", ".", "In", "addition", ",", "we", "optionally", "apply", "intra", "-", "sentence", "attention", "to", "endow", "the", "model", "with", "a", "richer", "encoding", "of", "substructures", "prior", "to", "the", "alignment", "step", ".", "Asymptotically", "our", "approach", "does", "the", "same", "total", "work", "as", "a", "vanilla", "LSTM", "encoder", ",", "while", "being", "trivially", "parallelizable", "across", "sentence", "length", ",", "which", "can", "allow", "for", "considerable", "speedups", "in", "low", "-", "latency", "settings", ".", "Empirical", "results", "on", "the", "SNLI", "corpus", "show", "that", "our", "approach", "achieves", "state", "-", "of", "-", "the", "-", "art", "results", ",", "while", "using", "almost", "an", "order", "of", "magnitude", "fewer", "parameters", "compared", "to", "complex", "LSTM", "-", "based", "approaches", ".", "section", ":", "Related", "Work", "Our", "method", "is", "motivated", "by", "the", "central", "role", "played", "by", "alignment", "in", "machine", "translation", "and", "previous", "approaches", "to", "sentence", "similarity", "modeling", ",", "natural", "language", "inference", ",", "and", "semantic", "parsing", ".", "The", "neural", "counterpart", "to", "alignment", ",", "attention", ",", "which", "is", "a", "key", "part", "of", "our", "approach", ",", "was", "originally", "proposed", "and", "has", "been", "predominantly", "used", "in", "conjunction", "with", "LSTMs", "and", "to", "a", "lesser", "extent", "with", "CNNs", ".", "In", "contrast", ",", "our", "use", "of", "attention", "is", "purely", "based", "on", "word", "embeddings", "and", "our", "method", "essentially", "consists", "of", "feed", "-", "forward", "networks", "that", "operate", "largely", "independently", "of", "word", "order", ".", "section", ":", "Approach", "Let", "and", "be", "the", "two", "input", "sentences", "of", "length", "and", ",", "respectively", ".", "We", "assume", "that", "each", ",", "is", "a", "word", "embedding", "vector", "of", "dimension", "and", "that", "each", "sentence", "is", "prepended", "with", "a", "“", "NULL", "”", "token", ".", "Our", "training", "data", "comes", "in", "the", "form", "of", "labeled", "pairs", ",", "where", "is", "an", "indicator", "vector", "encoding", "the", "label", "and", "is", "the", "number", "of", "output", "classes", ".", "At", "test", "time", ",", "we", "receive", "a", "pair", "of", "sentences", "and", "our", "goal", "is", "to", "predict", "the", "correct", "label", ".", "paragraph", ":", "Input", "representation", ".", "Let", "and", "denote", "the", "input", "representation", "of", "each", "fragment", "that", "is", "fed", "to", "subsequent", "steps", "of", "the", "algorithm", ".", "The", "vanilla", "version", "of", "our", "model", "simply", "defines", "and", ".", "With", "this", "input", "representation", ",", "our", "model", "does", "not", "make", "use", "of", "word", "order", ".", "However", ",", "we", "discuss", "an", "extension", "using", "intra", "-", "sentence", "attention", "in", "Section", "[", "reference", "]", "that", "uses", "a", "minimal", "amount", "of", "sequence", "information", ".", "The", "core", "model", "consists", "of", "the", "following", "three", "components", "(", "see", "Figure", "[", "reference", "]", ")", ",", "which", "are", "trained", "jointly", ":", "paragraph", ":", "Attend", ".", "First", ",", "soft", "-", "align", "the", "elements", "of", "and", "using", "a", "variant", "of", "neural", "attention", "and", "decompose", "the", "problem", "into", "the", "comparison", "of", "aligned", "subphrases", ".", "paragraph", ":", "Compare", ".", "Second", ",", "separately", "compare", "each", "aligned", "subphrase", "to", "produce", "a", "set", "of", "vectors", "for", "and", "for", ".", "Each", "is", "a", "nonlinear", "combination", "of", "and", "its", "(", "softly", ")", "aligned", "subphrase", "in", "(", "and", "analogously", "for", ")", ".", "paragraph", ":", "Aggregate", ".", "Finally", ",", "aggregate", "the", "sets", "and", "from", "the", "previous", "step", "and", "use", "the", "result", "to", "predict", "the", "label", ".", "subsection", ":", "Attend", "We", "first", "obtain", "unnormalized", "attention", "weights", ",", "computed", "by", "a", "function", ",", "which", "decomposes", "as", ":", "This", "decomposition", "avoids", "the", "quadratic", "complexity", "that", "would", "be", "associated", "with", "separately", "applying", "times", ".", "Instead", ",", "only", "applications", "of", "are", "needed", ".", "We", "take", "to", "be", "a", "feed", "-", "forward", "neural", "network", "with", "ReLU", "activations", ".", "These", "attention", "weights", "are", "normalized", "as", "follows", ":", "Here", "is", "the", "subphrase", "in", "that", "is", "(", "softly", ")", "aligned", "to", "and", "vice", "versa", "for", ".", "subsection", ":", "Compare", "Next", ",", "we", "separately", "compare", "the", "aligned", "phrases", "and", "using", "a", "function", ",", "which", "in", "this", "work", "is", "again", "a", "feed", "-", "forward", "network", ":", "where", "the", "brackets", "denote", "concatenation", ".", "Note", "that", "since", "there", "are", "only", "a", "linear", "number", "of", "terms", "in", "this", "case", ",", "we", "do", "not", "need", "to", "apply", "a", "decomposition", "as", "was", "done", "in", "the", "previous", "step", ".", "Thus", "can", "jointly", "take", "into", "account", "both", "and", ".", "subsection", ":", "Aggregate", "We", "now", "have", "two", "sets", "of", "comparison", "vectors", "and", ".", "We", "first", "aggregate", "over", "each", "set", "by", "summation", ":", "and", "feed", "the", "result", "through", "a", "final", "classifier", ",", "that", "is", "a", "feed", "forward", "network", "followed", "by", "a", "linear", "layer", ":", "where", "represents", "the", "predicted", "(", "unnormalized", ")", "scores", "for", "each", "class", "and", "consequently", "the", "predicted", "class", "is", "given", "by", ".", "For", "training", ",", "we", "use", "multi", "-", "class", "cross", "-", "entropy", "loss", "with", "dropout", "regularization", ":", "Here", "denote", "the", "learnable", "parameters", "of", "the", "functions", "F", ",", "G", "and", "H", ",", "respectively", ".", "subsection", ":", "Intra", "-", "Sentence", "Attention", "(", "Optional", ")", "In", "the", "above", "model", ",", "the", "input", "representations", "are", "simple", "word", "embeddings", ".", "However", ",", "we", "can", "augment", "this", "input", "representation", "with", "intra", "-", "sentence", "attention", "to", "encode", "compositional", "relationships", "between", "words", "within", "each", "sentence", ",", "as", "proposed", "by", "cheng2016long", ".", "Similar", "to", "Eqs", ".", "[", "reference", "]", "and", "[", "reference", "]", ",", "we", "define", "where", "is", "a", "feed", "-", "forward", "network", ".", "We", "then", "create", "the", "self", "-", "aligned", "phrases", "The", "distance", "-", "sensitive", "bias", "terms", "provides", "the", "model", "with", "a", "minimal", "amount", "of", "sequence", "information", ",", "while", "remaining", "parallelizable", ".", "These", "terms", "are", "bucketed", "such", "that", "all", "distances", "greater", "than", "10", "words", "share", "the", "same", "bias", ".", "The", "input", "representation", "for", "subsequent", "steps", "is", "then", "defined", "as", "and", "analogously", ".", "section", ":", "Computational", "Complexity", "We", "now", "discuss", "the", "asymptotic", "complexity", "of", "our", "approach", "and", "how", "it", "offers", "a", "higher", "degree", "of", "parallelism", "than", "LSTM", "-", "based", "approaches", ".", "Recall", "that", "denotes", "embedding", "dimension", "and", "means", "sentence", "length", ".", "For", "simplicity", "we", "assume", "that", "all", "hidden", "dimensions", "are", "and", "that", "the", "complexity", "of", "matrix", "(", ")", "-", "vector", "(", ")", "multiplication", "is", ".", "A", "key", "assumption", "of", "our", "analysis", "is", "that", ",", "which", "we", "believe", "is", "reasonable", "and", "is", "true", "of", "the", "SNLI", "dataset", "where", ",", "whereas", "recent", "LSTM", "-", "based", "approaches", "have", "used", ".", "This", "assumption", "allows", "us", "to", "bound", "the", "complexity", "of", "computing", "the", "attention", "weights", ".", "paragraph", ":", "Complexity", "of", "LSTMs", ".", "The", "complexity", "of", "an", "LSTM", "cell", "is", ",", "resulting", "in", "a", "complexity", "of", "to", "encode", "the", "sentence", ".", "Adding", "attention", "as", "in", "rocktaschel2015reasoning", "increases", "this", "complexity", "to", ".", "paragraph", ":", "Complexity", "of", "our", "Approach", ".", "Application", "of", "a", "feed", "-", "forward", "network", "requires", "steps", ".", "Thus", ",", "the", "Compare", "and", "Aggregate", "steps", "have", "complexity", "and", "respectively", ".", "For", "the", "Attend", "step", ",", "is", "evaluated", "times", ",", "giving", "a", "complexity", "of", ".", "Each", "attention", "weight", "requires", "one", "dot", "product", ",", "resulting", "in", "a", "complexity", "of", ".", "Thus", "the", "total", "complexity", "of", "the", "model", "is", ",", "which", "is", "equal", "to", "that", "of", "an", "LSTM", "with", "attention", ".", "However", ",", "note", "that", "with", "the", "assumption", "that", ",", "this", "becomes", "which", "is", "the", "same", "complexity", "as", "a", "regular", "LSTM", ".", "Moreover", ",", "unlike", "the", "LSTM", ",", "our", "approach", "has", "the", "advantage", "of", "being", "parallelizable", "over", ",", "which", "can", "be", "useful", "at", "test", "time", ".", "section", ":", "Experiments", "We", "evaluate", "our", "approach", "on", "the", "Stanford", "Natural", "Language", "Inference", "(", "SNLI", ")", "dataset", ".", "Given", "a", "sentences", "pair", ",", "the", "task", "is", "to", "predict", "whether", "is", "entailed", "by", ",", "contradicts", ",", "or", "whether", "their", "relationship", "is", "neutral", ".", "subsection", ":", "Implementation", "Details", "The", "method", "was", "implemented", "in", "TensorFlow", ".", "Data", "preprocessing", ":", "Following", "bowman2015large", ",", "we", "remove", "examples", "labeled", "“", "–", "”", "(", "no", "gold", "label", ")", "from", "the", "dataset", ",", "which", "leaves", "549", ",", "367", "pairs", "for", "training", ",", "9", ",", "842", "for", "development", ",", "and", "9", ",", "824", "for", "testing", ".", "We", "use", "the", "tokenized", "sentences", "from", "the", "non", "-", "binary", "parse", "provided", "in", "the", "dataset", "and", "prepend", "each", "sentence", "with", "a", "“", "NULL", "”", "token", ".", "During", "training", ",", "each", "sentence", "was", "padded", "up", "to", "the", "maximum", "length", "of", "the", "batch", "for", "efficient", "training", "(", "the", "padding", "was", "explicitly", "masked", "out", "so", "as", "not", "to", "affect", "the", "objective", "/", "gradients", ")", ".", "For", "efficient", "batching", "in", "TensorFlow", ",", "we", "semi", "-", "sorted", "the", "training", "data", "to", "first", "contain", "examples", "where", "both", "sentences", "had", "length", "less", "than", "20", ",", "followed", "by", "those", "with", "length", "less", "than", "50", ",", "and", "then", "the", "rest", ".", "This", "ensured", "that", "most", "training", "batches", "contained", "examples", "of", "similar", "length", ".", "Embeddings", ":", "We", "use", "300", "dimensional", "GloVe", "embeddings", "to", "represent", "words", ".", "Each", "embedding", "vector", "was", "normalized", "to", "have", "norm", "of", "1", "and", "projected", "down", "to", "200", "dimensions", ",", "a", "number", "determined", "via", "hyperparameter", "tuning", ".", "Out", "-", "of", "-", "vocabulary", "(", "OOV", ")", "words", "are", "hashed", "to", "one", "of", "100", "random", "embeddings", "each", "initialized", "to", "mean", "0", "and", "standard", "deviation", "1", ".", "All", "embeddings", "remain", "fixed", "during", "training", ",", "but", "the", "projection", "matrix", "is", "trained", ".", "All", "other", "parameter", "weights", "(", "hidden", "layers", "etc", ".", ")", "were", "initialized", "from", "random", "Gaussians", "with", "mean", "0", "and", "standard", "deviation", "0.01", ".", "Each", "hyperparameter", "setting", "was", "run", "on", "a", "single", "machine", "with", "10", "asynchronous", "gradient", "-", "update", "threads", ",", "using", "Adagrad", "for", "optimization", "with", "the", "default", "initial", "accumulator", "value", "of", "0.1", ".", "Dropout", "regularization", "was", "used", "for", "all", "ReLU", "layers", ",", "but", "not", "for", "the", "final", "linear", "layer", ".", "We", "additionally", "tuned", "the", "following", "hyperparameters", "and", "present", "their", "chosen", "values", "in", "parentheses", ":", "network", "size", "(", "2", "-", "layers", ",", "each", "with", "200", "neurons", ")", ",", "batch", "size", "(", "4", ")", ",", "dropout", "ratio", "(", "0.2", ")", "and", "learning", "rate", "(", "0.05–vanilla", ",", "0.025–intra", "-", "attention", ")", ".", "All", "settings", "were", "run", "for", "50", "million", "steps", "(", "each", "step", "indicates", "one", "batch", ")", "but", "model", "parameters", "were", "saved", "frequently", "as", "training", "progressed", "and", "we", "chose", "the", "model", "that", "did", "best", "on", "the", "development", "set", ".", "subsection", ":", "Results", "Results", "in", "terms", "of", "3", "-", "class", "accuracy", "are", "shown", "in", "Table", "[", "reference", "]", ".", "Our", "vanilla", "approach", "achieves", "state", "-", "of", "-", "the", "-", "art", "results", "with", "almost", "an", "order", "of", "magnitude", "fewer", "parameters", "than", "the", "LSTMN", "of", "cheng2016long", ".", "Adding", "intra", "-", "sentence", "attention", "gives", "a", "considerable", "improvement", "of", "0.5", "percentage", "points", "over", "the", "existing", "state", "of", "the", "art", ".", "Table", "[", "reference", "]", "gives", "a", "breakdown", "of", "accuracy", "on", "the", "development", "set", "showing", "that", "most", "of", "our", "gains", "stem", "from", "neutral", ",", "while", "most", "losses", "come", "from", "contradiction", "pairs", ".", "Table", "[", "reference", "]", "shows", "some", "wins", "and", "losses", ".", "Examples", "A", "-", "C", "are", "cases", "where", "both", "variants", "of", "our", "approach", "are", "correct", "while", "both", "SPINN", "-", "PI", "and", "the", "mLSTM", "are", "incorrect", ".", "In", "the", "first", "two", "cases", ",", "both", "sentences", "contain", "phrases", "that", "are", "either", "identical", "or", "highly", "lexically", "related", "(", "e.g.", "“", "Two", "kids", "”", "and", "“", "ocean", "/", "beach", "”", ")", "and", "our", "approach", "correctly", "favors", "neutral", "in", "these", "cases", ".", "In", "Example", "C", ",", "it", "is", "possible", "that", "relying", "on", "word", "-", "order", "may", "confuse", "SPINN", "-", "PI", "and", "the", "mLSTM", "due", "to", "how", "“", "fountain", "”", "is", "the", "object", "of", "a", "preposition", "in", "the", "first", "sentence", "but", "the", "subject", "of", "the", "second", ".", "The", "second", "set", "of", "examples", "(", "D", "-", "F", ")", "are", "cases", "where", "our", "vanilla", "approach", "is", "incorrect", "but", "mLSTM", "and", "SPINN", "-", "PI", "are", "correct", ".", "Example", "F", "requires", "sequential", "information", "and", "neither", "variant", "of", "our", "approach", "can", "predict", "the", "correct", "class", ".", "Examples", "D", "-", "E", "are", "interesting", "however", ",", "since", "they", "do", "n’t", "require", "word", "order", "information", ",", "yet", "intra", "-", "attention", "seems", "to", "help", ".", "We", "suspect", "this", "may", "be", "because", "the", "word", "embeddings", "are", "not", "fine", "-", "grained", "enough", "for", "the", "algorithm", "to", "conclude", "that", "“", "play", "/", "watch", "”", "is", "a", "contradiction", ",", "but", "intra", "-", "attention", ",", "by", "adding", "an", "extra", "layer", "of", "composition", "/", "nonlinearity", "to", "incorporate", "context", ",", "compensates", "for", "this", ".", "Finally", ",", "Examples", "G", "-", "I", "are", "cases", "that", "all", "methods", "get", "wrong", ".", "The", "first", "is", "actually", "representative", "of", "many", "examples", "in", "this", "category", "where", "there", "is", "one", "critical", "word", "that", "separates", "the", "two", "sentences", "(", "close", "vs", "open", "in", "this", "case", ")", "and", "goes", "unnoticed", "by", "the", "algorithms", ".", "Examples", "H", "requires", "inference", "about", "numbers", "and", "Example", "I", "needs", "sequence", "information", ".", "section", ":", "Conclusion", "We", "presented", "a", "simple", "attention", "-", "based", "approach", "to", "natural", "language", "inference", "that", "is", "trivially", "parallelizable", ".", "The", "approach", "outperforms", "considerably", "more", "complex", "neural", "methods", "aiming", "for", "text", "understanding", ".", "Our", "results", "suggest", "that", ",", "at", "least", "for", "this", "task", ",", "pairwise", "comparisons", "are", "relatively", "more", "important", "than", "global", "sentence", "-", "level", "representations", ".", "section", ":", "Acknowledgements", "We", "thank", "Slav", "Petrov", ",", "Tom", "Kwiatkowski", ",", "Yoon", "Kim", ",", "Erick", "Fonseca", ",", "Mark", "Neumann", "for", "useful", "discussion", "and", "Sam", "Bowman", "and", "Shuohang", "Wang", "for", "providing", "us", "their", "model", "outputs", "for", "error", "analysis", ".", "bibliography", ":", "References"], "entity": [[3, 6, ["Decomposable", "Attention", "Model"], "Method"], [7, 10, ["Natural", "Language", "Inference"], "Task"], [14, 16, ["neural", "architecture"], "Method"], [17, 20, ["natural", "language", "inference"], "Task"], [45, 49, ["Stanford", "Natural", "Language", "Inference"], "Material"], [50, 53, ["SNLI", ")", "dataset"], "Material"], [106, 109, ["Natural", "language", "inference"], "Task"], [110, 111, ["NLI"], "Task"], [117, 122, ["determining", "entailment", "and", "contradiction", "relationships"], "Task"], [129, 130, ["NLI"], "Task"], [135, 137, ["language", "understanding"], "Task"], [141, 143, ["SNLI", "corpus"], "Material"], [160, 161, ["NLI"], "Task"], [178, 180, ["neural", "networks"], "Method"], [181, 184, ["text", "similarity", "tasks"], "Task"], [185, 186, ["NLI"], "Task"], [204, 208, ["deep", "text", "representation", "models"], "Method"], [213, 215, ["convolutional", "networks"], "Method"], [216, 222, ["long", "short", "-", "term", "memory", "networks"], "Method"], [226, 229, ["deeper", "sentence", "comprehension"], "Task"], [269, 272, ["natural", "language", "inference"], "Task"], [341, 343, ["compact", "representation"], "Method"], [430, 431, ["NLI"], "Task"], [433, 435, ["neural", "framework"], "Method"], [447, 449, ["neural", "architectures"], "Method"], [461, 462, ["alignment"], "Method"], [509, 511, ["neural", "attention"], "Method"], [544, 545, ["classification"], "Task"], [569, 571, ["alignment", "step"], "Task"], [582, 585, ["vanilla", "LSTM", "encoder"], "Method"], [601, 605, ["low", "-", "latency", "settings"], "Task"], [610, 612, ["SNLI", "corpus"], "Material"], [638, 642, ["LSTM", "-", "based", "approaches"], "Method"], [657, 658, ["alignment"], "Task"], [659, 661, ["machine", "translation"], "Task"], [665, 668, ["sentence", "similarity", "modeling"], "Task"], [669, 672, ["natural", "language", "inference"], "Task"], [674, 676, ["semantic", "parsing"], "Task"], [678, 680, ["neural", "counterpart"], "Method"], [681, 682, ["alignment"], "Task"], [683, 684, ["attention"], "Task"], [705, 706, ["LSTMs"], "Method"], [712, 713, ["CNNs"], "Method"], [720, 721, ["attention"], "Method"], [725, 727, ["word", "embeddings"], "Method"], [733, 737, ["feed", "-", "forward", "networks"], "Method"], [835, 837, ["Input", "representation"], "Method"], [889, 893, ["intra", "-", "sentence", "attention"], "Task"], [908, 910, ["core", "model"], "Method"], [946, 948, ["neural", "attention"], "Method"], [954, 958, ["comparison", "of", "aligned", "subphrases"], "Task"], [1046, 1048, ["quadratic", "complexity"], "Metric"], [1070, 1075, ["feed", "-", "forward", "neural", "network"], "Method"], [1076, 1078, ["ReLU", "activations"], "Method"], [1127, 1131, ["feed", "-", "forward", "network"], "Method"], [1207, 1208, ["classifier"], "Method"], [1212, 1215, ["feed", "forward", "network"], "Method"], [1218, 1220, ["linear", "layer"], "Method"], [1242, 1243, ["training"], "Task"], [1246, 1253, ["multi", "-", "class", "cross", "-", "entropy", "loss"], "Method"], [1254, 1256, ["dropout", "regularization"], "Method"], [1275, 1279, ["Intra", "-", "Sentence", "Attention"], "Task"], [1292, 1294, ["word", "embeddings"], "Method"], [1340, 1344, ["feed", "-", "forward", "network"], "Method"], [1406, 1408, ["Computational", "Complexity"], "Metric"], [1412, 1414, ["asymptotic", "complexity"], "Metric"], [1427, 1431, ["LSTM", "-", "based", "approaches"], "Method"], [1456, 1464, ["matrix", "(", ")", "-", "vector", "(", ")", "multiplication"], "Method"], [1485, 1487, ["SNLI", "dataset"], "Material"], [1491, 1495, ["LSTM", "-", "based", "approaches"], "Method"], [1514, 1515, ["Complexity"], "Metric"], [1516, 1517, ["LSTMs"], "Method"], [1519, 1520, ["complexity"], "Metric"], [1522, 1524, ["LSTM", "cell"], "Method"], [1556, 1560, ["feed", "-", "forward", "network"], "Method"], [1566, 1570, ["Compare", "and", "Aggregate", "steps"], "Metric"], [1586, 1587, ["complexity"], "Metric"], [1600, 1601, ["complexity"], "Metric"], [1606, 1607, ["complexity"], "Metric"], [1619, 1622, ["LSTM", "with", "attention"], "Method"], [1638, 1639, ["complexity"], "Metric"], [1641, 1643, ["regular", "LSTM"], "Method"], [1648, 1649, ["LSTM"], "Method"], [1677, 1681, ["Stanford", "Natural", "Language", "Inference"], "Material"], [1682, 1685, ["SNLI", ")", "dataset"], "Material"], [1719, 1720, ["TensorFlow"], "Method"], [1721, 1723, ["Data", "preprocessing"], "Task"], [1756, 1757, ["development"], "Task"], [1808, 1809, ["training"], "Task"], [1829, 1830, ["batching"], "Task"], [1831, 1832, ["TensorFlow"], "Task"], [1879, 1880, ["Embeddings"], "Task"], [1883, 1887, ["300", "dimensional", "GloVe", "embeddings"], "Method"], [1912, 1914, ["hyperparameter", "tuning"], "Method"], [1930, 1932, ["random", "embeddings"], "Method"], [1990, 1995, ["asynchronous", "gradient", "-", "update", "threads"], "Method"], [1997, 1998, ["Adagrad"], "Method"], [1999, 2000, ["optimization"], "Task"], [2009, 2011, ["Dropout", "regularization"], "Method"], [2015, 2017, ["ReLU", "layers"], "Method"], [2023, 2025, ["linear", "layer"], "Method"], [2040, 2042, ["network", "size"], "Metric"], [2053, 2055, ["batch", "size"], "Metric"], [2059, 2061, ["dropout", "ratio"], "Material"], [2065, 2067, ["learning", "rate"], "Metric"], [2132, 2134, ["vanilla", "approach"], "Method"], [2153, 2154, ["LSTMN"], "Metric"], [2186, 2187, ["accuracy"], "Metric"], [2235, 2238, ["SPINN", "-", "PI"], "Method"], [2240, 2241, ["mLSTM"], "Method"], [2300, 2303, ["SPINN", "-", "PI"], "Method"], [2305, 2306, ["mLSTM"], "Method"], [2343, 2345, ["vanilla", "approach"], "Method"], [2348, 2349, ["mLSTM"], "Method"], [2350, 2353, ["SPINN", "-", "PI"], "Method"], [2521, 2525, ["attention", "-", "based", "approach"], "Method"], [2526, 2529, ["natural", "language", "inference"], "Task"], [2540, 2542, ["neural", "methods"], "Method"], [2544, 2546, ["text", "understanding"], "Task"], [2558, 2560, ["pairwise", "comparisons"], "Task"], [2565, 2570, ["global", "sentence", "-", "level", "representations"], "Method"], [2606, 2608, ["error", "analysis"], "Task"], [2122, 2123, ["accuracy"], "Metric"]]} {"doc_id": "087337fdad69caaab8ebd8ae68a731c5bf2e8b14", "text": ["document", ":", "Fully", "Convolutional", "Networks", "for", "Semantic", "Segmentation", "Convolutional", "networks", "are", "powerful", "visual", "models", "that", "yield", "hierarchies", "of", "features", ".", "We", "show", "that", "convolutional", "networks", "by", "themselves", ",", "trained", "end", "-", "to", "-", "end", ",", "pixels", "-", "to", "-", "pixels", ",", "improve", "on", "the", "previous", "best", "result", "in", "semantic", "segmentation", ".", "Our", "key", "insight", "is", "to", "build", "“", "fully", "convolutional", "”", "networks", "that", "take", "input", "of", "arbitrary", "size", "and", "produce", "correspondingly", "-", "sized", "output", "with", "efficient", "inference", "and", "learning", ".", "We", "define", "and", "detail", "the", "space", "of", "fully", "convolutional", "networks", ",", "explain", "their", "application", "to", "spatially", "dense", "prediction", "tasks", ",", "and", "draw", "connections", "to", "prior", "models", ".", "We", "adapt", "contemporary", "classification", "networks", "(", "AlexNet", ",", "the", "VGG", "net", ",", "and", "GoogLeNet", ")", "into", "fully", "convolutional", "networks", "and", "transfer", "their", "learned", "representations", "by", "fine", "-", "tuning", "to", "the", "segmentation", "task", ".", "We", "then", "define", "a", "skip", "architecture", "that", "combines", "semantic", "information", "from", "a", "deep", ",", "coarse", "layer", "with", "appearance", "information", "from", "a", "shallow", ",", "fine", "layer", "to", "produce", "accurate", "and", "detailed", "segmentations", ".", "Our", "fully", "convolutional", "network", "achieves", "improved", "segmentation", "of", "PASCAL", "VOC", "(", "30", "%", "relative", "improvement", "to", "67.2", "%", "mean", "IU", "on", "2012", ")", ",", "NYUDv2", ",", "SIFT", "Flow", ",", "and", "PASCAL", "-", "Context", ",", "while", "inference", "takes", "one", "tenth", "of", "a", "second", "for", "a", "typical", "image", ".", "Semantic", "Segmentation", ",", "Convolutional", "Networks", ",", "Deep", "Learning", ",", "Transfer", "Learning", "section", ":", "Introduction", "Convolutional", "networks", "are", "driving", "advances", "in", "recognition", ".", "Convnets", "are", "not", "only", "improving", "for", "whole", "-", "image", "classification", ",", "but", "also", "making", "progress", "on", "local", "tasks", "with", "structured", "output", ".", "These", "include", "advances", "in", "bounding", "box", "object", "detection", ",", "part", "and", "keypoint", "prediction", ",", "and", "local", "correspondence", ".", "The", "natural", "next", "step", "in", "the", "progression", "from", "coarse", "to", "fine", "inference", "is", "to", "make", "a", "prediction", "at", "every", "pixel", ".", "Prior", "approaches", "have", "used", "convnets", "for", "semantic", "segmentation", ",", "in", "which", "each", "pixel", "is", "labeled", "with", "the", "class", "of", "its", "enclosing", "object", "or", "region", ",", "but", "with", "shortcomings", "that", "this", "work", "addresses", ".", "We", "show", "that", "fully", "convolutional", "networks", "(", "FCNs", ")", "trained", "end", "-", "to", "-", "end", ",", "pixels", "-", "to", "-", "pixels", "on", "semantic", "segmentation", "exceed", "the", "previous", "best", "results", "without", "further", "machinery", ".", "To", "our", "knowledge", ",", "this", "is", "the", "first", "work", "to", "train", "FCNs", "end", "-", "to", "-", "end", "(", "1", ")", "for", "pixelwise", "prediction", "and", "(", "2", ")", "from", "supervised", "pre", "-", "training", ".", "Fully", "convolutional", "versions", "of", "existing", "networks", "predict", "dense", "outputs", "from", "arbitrary", "-", "sized", "inputs", ".", "Both", "learning", "and", "inference", "are", "performed", "whole", "-", "image", "-", "at", "-", "a", "-", "time", "by", "dense", "feedforward", "computation", "and", "backpropagation", ".", "In", "-", "network", "upsampling", "layers", "enable", "pixelwise", "prediction", "and", "learning", "in", "nets", "with", "subsampling", ".", "This", "method", "is", "efficient", ",", "both", "asymptotically", "and", "absolutely", ",", "and", "precludes", "the", "need", "for", "the", "complications", "in", "other", "works", ".", "Patchwise", "training", "is", "common", ",", "but", "lacks", "the", "efficiency", "of", "fully", "convolutional", "training", ".", "Our", "approach", "does", "not", "make", "use", "of", "pre", "-", "and", "post", "-", "processing", "complications", ",", "including", "superpixels", ",", "proposals", ",", "or", "post", "-", "hoc", "refinement", "by", "random", "fields", "or", "local", "classifiers", ".", "Our", "model", "transfers", "recent", "success", "in", "classification", "to", "dense", "prediction", "by", "reinterpreting", "classification", "nets", "as", "fully", "convolutional", "and", "fine", "-", "tuning", "from", "their", "learned", "representations", ".", "In", "contrast", ",", "previous", "works", "have", "applied", "small", "convnets", "without", "supervised", "pre", "-", "training", ".", "Semantic", "segmentation", "faces", "an", "inherent", "tension", "between", "semantics", "and", "location", ":", "global", "information", "resolves", "what", "while", "local", "information", "resolves", "where", ".", "What", "can", "be", "done", "to", "navigate", "this", "spectrum", "from", "location", "to", "semantics", "?", "How", "can", "local", "decisions", "respect", "global", "structure", "?", "It", "is", "not", "immediately", "clear", "that", "deep", "networks", "for", "image", "classification", "yield", "representations", "sufficient", "for", "accurate", ",", "pixelwise", "recognition", ".", "In", "the", "conference", "version", "of", "this", "paper", ",", "we", "cast", "pre", "-", "trained", "networks", "into", "fully", "convolutional", "form", ",", "and", "augment", "them", "with", "a", "skip", "architecture", "that", "takes", "advantage", "of", "the", "full", "feature", "spectrum", ".", "The", "skip", "architecture", "fuses", "the", "feature", "hierarchy", "to", "combine", "deep", ",", "coarse", ",", "semantic", "information", "and", "shallow", ",", "fine", ",", "appearance", "information", "(", "see", "Section", "[", "reference", "]", "and", "Figure", "[", "reference", "]", ")", ".", "In", "this", "light", ",", "deep", "feature", "hierarchies", "encode", "location", "and", "semantics", "in", "a", "nonlinear", "local", "-", "to", "-", "global", "pyramid", ".", "This", "journal", "paper", "extends", "our", "earlier", "work", "through", "further", "tuning", ",", "analysis", ",", "and", "more", "results", ".", "Alternative", "choices", ",", "ablations", ",", "and", "implementation", "details", "better", "cover", "the", "space", "of", "FCNs", ".", "Tuning", "optimization", "leads", "to", "more", "accurate", "networks", "and", "a", "means", "to", "learn", "skip", "architectures", "all", "-", "at", "-", "once", "instead", "of", "in", "stages", ".", "Experiments", "that", "mask", "foreground", "and", "background", "investigate", "the", "role", "of", "context", "and", "shape", ".", "Results", "on", "the", "object", "and", "scene", "labeling", "of", "PASCAL", "-", "Context", "reinforce", "merging", "object", "segmentation", "and", "scene", "parsing", "as", "unified", "pixelwise", "prediction", ".", "In", "the", "next", "section", ",", "we", "review", "related", "work", "on", "deep", "classification", "nets", ",", "FCNs", ",", "recent", "approaches", "to", "semantic", "segmentation", "using", "convnets", ",", "and", "extensions", "to", "FCNs", ".", "The", "following", "sections", "explain", "FCN", "design", ",", "introduce", "our", "architecture", "with", "in", "-", "network", "upsampling", "and", "skip", "layers", ",", "and", "describe", "our", "experimental", "framework", ".", "Next", ",", "we", "demonstrate", "improved", "accuracy", "on", "PASCAL", "VOC", "2011", "-", "2", ",", "NYUDv2", ",", "SIFT", "Flow", ",", "and", "PASCAL", "-", "Context", ".", "Finally", ",", "we", "analyze", "design", "choices", ",", "examine", "what", "cues", "can", "be", "learned", "by", "an", "FCN", ",", "and", "calculate", "recognition", "bounds", "for", "semantic", "segmentation", ".", "section", ":", "Related", "Work", "Our", "approach", "draws", "on", "recent", "successes", "of", "deep", "nets", "for", "image", "classification", "and", "transfer", "learning", ".", "Transfer", "was", "first", "demonstrated", "on", "various", "visual", "recognition", "tasks", ",", "then", "on", "detection", ",", "and", "on", "both", "instance", "and", "semantic", "segmentation", "in", "hybrid", "proposal", "-", "classifier", "models", ".", "We", "now", "re", "-", "architect", "and", "fine", "-", "tune", "classification", "nets", "to", "direct", ",", "dense", "prediction", "of", "semantic", "segmentation", ".", "We", "chart", "the", "space", "of", "FCNs", "and", "relate", "prior", "models", "both", "historical", "and", "recent", ".", "Fully", "convolutional", "networks", "To", "our", "knowledge", ",", "the", "idea", "of", "extending", "a", "convnet", "to", "arbitrary", "-", "sized", "inputs", "first", "appeared", "in", "Matan", "et", "al", ".", ",", "which", "extended", "the", "classic", "LeNet", "to", "recognize", "strings", "of", "digits", ".", "Because", "their", "net", "was", "limited", "to", "one", "-", "dimensional", "input", "strings", ",", "Matan", "et", "al", ".", "used", "Viterbi", "decoding", "to", "obtain", "their", "outputs", ".", "Wolf", "and", "Platt", "expand", "convnet", "outputs", "to", "2", "-", "dimensional", "maps", "of", "detection", "scores", "for", "the", "four", "corners", "of", "postal", "address", "blocks", ".", "Both", "of", "these", "historical", "works", "do", "inference", "and", "learning", "fully", "convolutionally", "for", "detection", ".", "Ning", "et", "al", ".", "define", "a", "convnet", "for", "coarse", "multiclass", "segmentation", "of", "C.", "elegans", "tissues", "with", "fully", "convolutional", "inference", ".", "Fully", "convolutional", "computation", "has", "also", "been", "exploited", "in", "the", "present", "era", "of", "many", "-", "layered", "nets", ".", "Sliding", "window", "detection", "by", "Sermanet", "et", "al", ".", ",", "semantic", "segmentation", "by", "Pinheiro", "and", "Collobert", ",", "and", "image", "restoration", "by", "Eigen", "et", "al", ".", "do", "fully", "convolutional", "inference", ".", "Fully", "convolutional", "training", "is", "rare", ",", "but", "used", "effectively", "by", "Tompson", "et", "al", ".", "to", "learn", "an", "end", "-", "to", "-", "end", "part", "detector", "and", "spatial", "model", "for", "pose", "estimation", ",", "although", "they", "do", "not", "exposit", "on", "or", "analyze", "this", "method", ".", "Dense", "prediction", "with", "convnets", "Several", "recent", "works", "have", "applied", "convnets", "to", "dense", "prediction", "problems", ",", "including", "semantic", "segmentation", "by", "Ning", "et", "al", ".", ",", "Farabet", "et", "al", ".", ",", "and", "Pinheiro", "and", "Collobert", ";", "boundary", "prediction", "for", "electron", "microscopy", "by", "Ciresan", "et", "al", ".", "and", "for", "natural", "images", "by", "a", "hybrid", "convnet", "/", "nearest", "neighbor", "model", "by", "Ganin", "and", "Lempitsky", ";", "and", "image", "restoration", "and", "depth", "estimation", "by", "Eigen", "et", "al", ".", ".", "Common", "elements", "of", "these", "approaches", "include", "small", "models", "restricting", "capacity", "and", "receptive", "fields", ";", "patchwise", "training", ";", "refinement", "by", "superpixel", "projection", ",", "random", "field", "regularization", ",", "filtering", ",", "or", "local", "classification", ";", "“", "interlacing", "”", "to", "obtain", "dense", "output", ";", "multi", "-", "scale", "pyramid", "processing", ";", "saturating", "nonlinearities", ";", "and", "ensembles", ",", "whereas", "our", "method", "does", "without", "this", "machinery", ".", "However", ",", "we", "do", "study", "patchwise", "training", "(", "Section", "[", "reference", "]", ")", "and", "“", "shift", "-", "and", "-", "stitch", "”", "dense", "output", "(", "Section", "[", "reference", "]", ")", "from", "the", "perspective", "of", "FCNs", ".", "We", "also", "discuss", "in", "-", "network", "upsampling", "(", "Section", "[", "reference", "]", ")", ",", "of", "which", "the", "fully", "connected", "prediction", "by", "Eigen", "et", "al", ".", "is", "a", "special", "case", ".", "Unlike", "these", "existing", "methods", ",", "we", "adapt", "and", "extend", "deep", "classification", "architectures", ",", "using", "image", "classification", "as", "supervised", "pre", "-", "training", ",", "and", "fine", "-", "tune", "fully", "convolutionally", "to", "learn", "simply", "and", "efficiently", "from", "whole", "image", "inputs", "and", "whole", "image", "ground", "thruths", ".", "Hariharan", "et", "al", ".", "and", "Gupta", "et", "al", ".", "likewise", "adapt", "deep", "classification", "nets", "to", "semantic", "segmentation", ",", "but", "do", "so", "in", "hybrid", "proposal", "-", "classifier", "models", ".", "These", "approaches", "fine", "-", "tune", "an", "R", "-", "CNN", "system", "by", "sampling", "bounding", "boxes", "and", "/", "or", "region", "proposals", "for", "detection", ",", "semantic", "segmentation", ",", "and", "instance", "segmentation", ".", "Neither", "method", "is", "learned", "end", "-", "to", "-", "end", ".", "They", "achieve", "the", "previous", "best", "segmentation", "results", "on", "PASCAL", "VOC", "and", "NYUDv2", "respectively", ",", "so", "we", "directly", "compare", "our", "standalone", ",", "end", "-", "to", "-", "end", "FCN", "to", "their", "semantic", "segmentation", "results", "in", "Section", "[", "reference", "]", ".", "Combining", "feature", "hierarchies", "We", "fuse", "features", "across", "layers", "to", "define", "a", "nonlinear", "local", "-", "to", "-", "global", "representation", "that", "we", "tune", "end", "-", "to", "-", "end", ".", "The", "Laplacian", "pyramid", "is", "a", "classic", "multi", "-", "scale", "representation", "made", "of", "fixed", "smoothing", "and", "differencing", ".", "The", "jet", "of", "Koenderink", "and", "van", "Doorn", "is", "a", "rich", ",", "local", "feature", "defined", "by", "compositions", "of", "partial", "derivatives", ".", "In", "the", "context", "of", "deep", "networks", ",", "Sermanet", "et", "al", ".", "fuse", "intermediate", "layers", "but", "discard", "resolution", "in", "doing", "so", ".", "In", "contemporary", "work", "Hariharan", "et", "al", ".", "and", "Mostajabi", "et", "al", ".", "also", "fuse", "multiple", "layers", "but", "do", "not", "learn", "end", "-", "to", "-", "end", "and", "rely", "on", "fixed", "bottom", "-", "up", "grouping", ".", "FCN", "extensions", "Following", "the", "conference", "version", "of", "this", "paper", ",", "FCNs", "have", "been", "extended", "to", "new", "tasks", "and", "data", ".", "Tasks", "include", "region", "proposals", ",", "contour", "detection", ",", "depth", "regression", ",", "optical", "flow", ",", "and", "weakly", "-", "supervised", "semantic", "segmentation", ".", "In", "addition", ",", "new", "works", "have", "improved", "the", "FCNs", "presented", "here", "to", "further", "advance", "the", "state", "-", "of", "-", "the", "-", "art", "in", "semantic", "segmentation", ".", "The", "DeepLab", "models", "raise", "output", "resolution", "by", "dilated", "convolution", "and", "dense", "CRF", "inference", ".", "The", "joint", "CRFasRNN", "model", "is", "an", "end", "-", "to", "-", "end", "integration", "of", "the", "CRF", "for", "further", "improvement", ".", "ParseNet", "normalizes", "features", "for", "fusion", "and", "captures", "context", "with", "global", "pooling", ".", "The", "“", "deconvolutional", "network", "”", "approach", "of", "restores", "resolution", "by", "proposals", ",", "stacks", "of", "learned", "deconvolution", ",", "and", "unpooling", ".", "U", "-", "Net", "combines", "skip", "layers", "and", "learned", "deconvolution", "for", "pixel", "labeling", "of", "microscopy", "images", ".", "The", "dilation", "architecture", "of", "makes", "thorough", "use", "of", "dilated", "convolution", "for", "pixel", "-", "precise", "output", "without", "a", "random", "field", "or", "skip", "layers", ".", "section", ":", "Fully", "Convolutional", "Networks", "Each", "layer", "output", "in", "a", "convnet", "is", "a", "three", "-", "dimensional", "array", "of", "size", ",", "where", "and", "are", "spatial", "dimensions", ",", "and", "is", "the", "feature", "or", "channel", "dimension", ".", "The", "first", "layer", "is", "the", "image", ",", "with", "pixel", "size", ",", "and", "channels", ".", "Locations", "in", "higher", "layers", "correspond", "to", "the", "locations", "in", "the", "image", "they", "are", "path", "-", "connected", "to", ",", "which", "are", "called", "their", "receptive", "fields", ".", "Convnets", "are", "inherently", "translation", "invariant", ".", "Their", "basic", "components", "(", "convolution", ",", "pooling", ",", "and", "activation", "functions", ")", "operate", "on", "local", "input", "regions", ",", "and", "depend", "only", "on", "relative", "spatial", "coordinates", ".", "Writing", "for", "the", "data", "vector", "at", "location", "in", "a", "particular", "layer", ",", "and", "for", "the", "following", "layer", ",", "these", "functions", "compute", "outputs", "by", "where", "is", "called", "the", "kernel", "size", ",", "is", "the", "stride", "or", "subsampling", "factor", ",", "and", "determines", "the", "layer", "type", ":", "a", "matrix", "multiplication", "for", "convolution", "or", "average", "pooling", ",", "a", "spatial", "max", "for", "max", "pooling", ",", "or", "an", "elementwise", "nonlinearity", "for", "an", "activation", "function", ",", "and", "so", "on", "for", "other", "types", "of", "layers", ".", "This", "functional", "form", "is", "maintained", "under", "composition", ",", "with", "kernel", "size", "and", "stride", "obeying", "the", "transformation", "rule", "While", "a", "general", "net", "computes", "a", "general", "nonlinear", "function", ",", "a", "net", "with", "only", "layers", "of", "this", "form", "computes", "a", "nonlinear", "filter", ",", "which", "we", "call", "a", "deep", "filter", "or", "fully", "convolutional", "network", ".", "An", "FCN", "naturally", "operates", "on", "an", "input", "of", "any", "size", ",", "and", "produces", "an", "output", "of", "corresponding", "(", "possibly", "resampled", ")", "spatial", "dimensions", ".", "A", "real", "-", "valued", "loss", "function", "composed", "with", "an", "FCN", "defines", "a", "task", ".", "If", "the", "loss", "function", "is", "a", "sum", "over", "the", "spatial", "dimensions", "of", "the", "final", "layer", ",", ",", "its", "parameter", "gradient", "will", "be", "a", "sum", "over", "the", "parameter", "gradients", "of", "each", "of", "its", "spatial", "components", ".", "Thus", "stochastic", "gradient", "descent", "on", "computed", "on", "whole", "images", "will", "be", "the", "same", "as", "stochastic", "gradient", "descent", "on", ",", "taking", "all", "of", "the", "final", "layer", "receptive", "fields", "as", "a", "minibatch", ".", "When", "these", "receptive", "fields", "overlap", "significantly", ",", "both", "feedforward", "computation", "and", "backpropagation", "are", "much", "more", "efficient", "when", "computed", "layer", "-", "by", "-", "layer", "over", "an", "entire", "image", "instead", "of", "independently", "patch", "-", "by", "-", "patch", ".", "We", "next", "explain", "how", "to", "convert", "classification", "nets", "into", "fully", "convolutional", "nets", "that", "produce", "coarse", "output", "maps", ".", "For", "pixelwise", "prediction", ",", "we", "need", "to", "connect", "these", "coarse", "outputs", "back", "to", "the", "pixels", ".", "Section", "[", "reference", "]", "describes", "a", "trick", "used", "for", "this", "purpose", "(", "e.g.", ",", "by", "“", "fast", "scanning", "”", ")", ".", "We", "explain", "this", "trick", "in", "terms", "of", "network", "modification", ".", "As", "an", "efficient", ",", "effective", "alternative", ",", "we", "upsample", "in", "Section", "[", "reference", "]", ",", "reusing", "our", "implementation", "of", "convolution", ".", "In", "Section", "[", "reference", "]", "we", "consider", "training", "by", "patchwise", "sampling", ",", "and", "give", "evidence", "in", "Section", "[", "reference", "]", "that", "our", "whole", "image", "training", "is", "faster", "and", "equally", "effective", ".", "subsection", ":", "Adapting", "classifiers", "for", "dense", "prediction", "Typical", "recognition", "nets", ",", "including", "LeNet", ",", "AlexNet", ",", "and", "its", "deeper", "successors", ",", "ostensibly", "take", "fixed", "-", "sized", "inputs", "and", "produce", "non", "-", "spatial", "outputs", ".", "The", "fully", "connected", "layers", "of", "these", "nets", "have", "fixed", "dimensions", "and", "throw", "away", "spatial", "coordinates", ".", "However", ",", "fully", "connected", "layers", "can", "also", "be", "viewed", "as", "convolutions", "with", "kernels", "that", "cover", "their", "entire", "input", "regions", ".", "Doing", "so", "casts", "these", "nets", "into", "fully", "convolutional", "networks", "that", "take", "input", "of", "any", "size", "and", "make", "spatial", "output", "maps", ".", "This", "transformation", "is", "illustrated", "in", "Figure", "[", "reference", "]", ".", "Furthermore", ",", "while", "the", "resulting", "maps", "are", "equivalent", "to", "the", "evaluation", "of", "the", "original", "net", "on", "particular", "input", "patches", ",", "the", "computation", "is", "highly", "amortized", "over", "the", "overlapping", "regions", "of", "those", "patches", ".", "For", "example", ",", "while", "AlexNet", "takes", "ms", "(", "on", "a", "typical", "GPU", ")", "to", "infer", "the", "classification", "scores", "of", "a", "image", ",", "the", "fully", "convolutional", "net", "takes", "ms", "to", "produce", "a", "grid", "of", "outputs", "from", "a", "image", ",", "which", "is", "more", "than", "times", "faster", "than", "the", "naïve", "approach", ".", "The", "spatial", "output", "maps", "of", "these", "convolutionalized", "models", "make", "them", "a", "natural", "choice", "for", "dense", "problems", "like", "semantic", "segmentation", ".", "With", "ground", "truth", "available", "at", "every", "output", "cell", ",", "both", "the", "forward", "and", "backward", "passes", "are", "straightforward", ",", "and", "both", "take", "advantage", "of", "the", "inherent", "computational", "efficiency", "(", "and", "aggressive", "optimization", ")", "of", "convolution", ".", "The", "corresponding", "backward", "times", "for", "the", "AlexNet", "example", "are", "ms", "for", "a", "single", "image", "and", "ms", "for", "a", "fully", "convolutional", "output", "map", ",", "resulting", "in", "a", "speedup", "similar", "to", "that", "of", "the", "forward", "pass", ".", "While", "our", "reinterpretation", "of", "classification", "nets", "as", "fully", "convolutional", "yields", "output", "maps", "for", "inputs", "of", "any", "size", ",", "the", "output", "dimensions", "are", "typically", "reduced", "by", "subsampling", ".", "The", "classification", "nets", "subsample", "to", "keep", "filters", "small", "and", "computational", "requirements", "reasonable", ".", "This", "coarsens", "the", "output", "of", "a", "fully", "convolutional", "version", "of", "these", "nets", ",", "reducing", "it", "from", "the", "size", "of", "the", "input", "by", "a", "factor", "equal", "to", "the", "pixel", "stride", "of", "the", "receptive", "fields", "of", "the", "output", "units", ".", "subsection", ":", "Shift", "-", "and", "-", "stitch", "is", "filter", "dilation", "Dense", "predictions", "can", "be", "obtained", "from", "coarse", "outputs", "by", "stitching", "together", "outputs", "from", "shifted", "versions", "of", "the", "input", ".", "If", "the", "output", "is", "downsampled", "by", "a", "factor", "of", ",", "shift", "the", "input", "pixels", "to", "the", "right", "and", "pixels", "down", ",", "once", "for", "every", "such", "that", ".", "Process", "each", "of", "these", "inputs", ",", "and", "interlace", "the", "outputs", "so", "that", "the", "predictions", "correspond", "to", "the", "pixels", "at", "the", "centers", "of", "their", "receptive", "fields", ".", "Although", "this", "transformation", "naïvely", "increases", "the", "cost", "by", "a", "factor", "of", ",", "there", "is", "a", "well", "-", "known", "trick", "for", "efficiently", "producing", "identical", "results", ".", "(", "This", "trick", "is", "also", "used", "in", "the", "algorithme", "à", "trous", "for", "wavelet", "transforms", "and", "related", "to", "the", "Noble", "identities", "from", "signal", "processing", ".", ")", "Consider", "a", "layer", "(", "convolution", "or", "pooling", ")", "with", "input", "stride", ",", "and", "a", "subsequent", "convolution", "layer", "with", "filter", "weights", "(", "eliding", "the", "irrelevant", "feature", "dimensions", ")", ".", "Setting", "the", "earlier", "layer", "’s", "input", "stride", "to", "one", "upsamples", "its", "output", "by", "a", "factor", "of", ".", "However", ",", "convolving", "the", "original", "filter", "with", "the", "upsampled", "output", "does", "not", "produce", "the", "same", "result", "as", "shift", "-", "and", "-", "stitch", ",", "because", "the", "original", "filter", "only", "sees", "a", "reduced", "portion", "of", "its", "(", "now", "upsampled", ")", "input", ".", "To", "produce", "the", "same", "result", ",", "dilate", "(", "or", "“", "rarefy", "”", ")", "the", "filter", "by", "forming", "(", "with", "and", "zero", "-", "based", ")", ".", "Reproducing", "the", "full", "net", "output", "of", "shift", "-", "and", "-", "stitch", "involves", "repeating", "this", "filter", "enlargement", "layer", "-", "by", "-", "layer", "until", "all", "subsampling", "is", "removed", ".", "(", "In", "practice", ",", "this", "can", "be", "done", "efficiently", "by", "processing", "subsampled", "versions", "of", "the", "upsampled", "input", ".", ")", "Simply", "decreasing", "subsampling", "within", "a", "net", "is", "a", "tradeoff", ":", "the", "filters", "see", "finer", "information", ",", "but", "have", "smaller", "receptive", "fields", "and", "take", "longer", "to", "compute", ".", "This", "dilation", "trick", "is", "another", "kind", "of", "tradeoff", ":", "the", "output", "is", "denser", "without", "decreasing", "the", "receptive", "field", "sizes", "of", "the", "filters", ",", "but", "the", "filters", "are", "prohibited", "from", "accessing", "information", "at", "a", "finer", "scale", "than", "their", "original", "design", ".", "Although", "we", "have", "done", "preliminary", "experiments", "with", "dilation", ",", "we", "do", "not", "use", "it", "in", "our", "model", ".", "We", "find", "learning", "through", "upsampling", ",", "as", "described", "in", "the", "next", "section", ",", "to", "be", "effective", "and", "efficient", ",", "especially", "when", "combined", "with", "the", "skip", "layer", "fusion", "described", "later", "on", ".", "For", "further", "detail", "regarding", "dilation", ",", "refer", "to", "the", "dilated", "FCN", "of", ".", "subsection", ":", "Upsampling", "is", "(", "fractionally", "strided", ")", "convolution", "Another", "way", "to", "connect", "coarse", "outputs", "to", "dense", "pixels", "is", "interpolation", ".", "For", "instance", ",", "simple", "bilinear", "interpolation", "computes", "each", "output", "from", "the", "nearest", "four", "inputs", "by", "a", "linear", "map", "that", "depends", "only", "on", "the", "relative", "positions", "of", "the", "input", "and", "output", "cells", ":", "where", "is", "the", "upsampling", "factor", ",", "and", "denotes", "the", "fractional", "part", ".", "In", "a", "sense", ",", "upsampling", "with", "factor", "is", "convolution", "with", "a", "fractional", "input", "stride", "of", ".", "So", "long", "as", "is", "integral", ",", "it", "’s", "natural", "to", "implement", "upsampling", "through", "“", "backward", "convolution", "”", "by", "reversing", "the", "forward", "and", "backward", "passes", "of", "more", "typical", "input", "-", "strided", "convolution", ".", "Thus", "upsampling", "is", "performed", "in", "-", "network", "for", "end", "-", "to", "-", "end", "learning", "by", "backpropagation", "from", "the", "pixelwise", "loss", ".", "Per", "their", "use", "in", "deconvolution", "networks", "(", "esp", ".", ")", ",", "these", "(", "convolution", ")", "layers", "are", "sometimes", "referred", "to", "as", "deconvolution", "layers", ".", "Note", "that", "the", "convolution", "filter", "in", "such", "a", "layer", "need", "not", "be", "fixed", "(", "e.g.", ",", "to", "bilinear", "upsampling", ")", ",", "but", "can", "be", "learned", ".", "A", "stack", "of", "deconvolution", "layers", "and", "activation", "functions", "can", "even", "learn", "a", "nonlinear", "upsampling", ".", "In", "our", "experiments", ",", "we", "find", "that", "in", "-", "network", "upsampling", "is", "fast", "and", "effective", "for", "learning", "dense", "prediction", ".", "subsection", ":", "Patchwise", "training", "is", "loss", "sampling", "In", "stochastic", "optimization", ",", "gradient", "computation", "is", "driven", "by", "the", "training", "distribution", ".", "Both", "patchwise", "training", "and", "fully", "convolutional", "training", "can", "be", "made", "to", "produce", "any", "distribution", "of", "the", "inputs", ",", "although", "their", "relative", "computational", "efficiency", "depends", "on", "overlap", "and", "minibatch", "size", ".", "Whole", "image", "fully", "convolutional", "training", "is", "identical", "to", "patchwise", "training", "where", "each", "batch", "consists", "of", "all", "the", "receptive", "fields", "of", "the", "output", "units", "for", "an", "image", "(", "or", "collection", "of", "images", ")", ".", "While", "this", "is", "more", "efficient", "than", "uniform", "sampling", "of", "patches", ",", "it", "reduces", "the", "number", "of", "possible", "batches", ".", "However", ",", "random", "sampling", "of", "patches", "within", "an", "image", "may", "be", "easily", "recovered", ".", "Restricting", "the", "loss", "to", "a", "randomly", "sampled", "subset", "of", "its", "spatial", "terms", "(", "or", ",", "equivalently", "applying", "a", "DropConnect", "mask", "between", "the", "output", "and", "the", "loss", ")", "excludes", "patches", "from", "the", "gradient", ".", "If", "the", "kept", "patches", "still", "have", "significant", "overlap", ",", "fully", "convolutional", "computation", "will", "still", "speed", "up", "training", ".", "If", "gradients", "are", "accumulated", "over", "multiple", "backward", "passes", ",", "batches", "can", "include", "patches", "from", "several", "images", ".", "If", "inputs", "are", "shifted", "by", "values", "up", "to", "the", "output", "stride", ",", "random", "selection", "of", "all", "possible", "patches", "is", "possible", "even", "though", "the", "output", "units", "lie", "on", "a", "fixed", ",", "strided", "grid", ".", "Sampling", "in", "patchwise", "training", "can", "correct", "class", "imbalance", "and", "mitigate", "the", "spatial", "correlation", "of", "dense", "patches", ".", "In", "fully", "convolutional", "training", ",", "class", "balance", "can", "also", "be", "achieved", "by", "weighting", "the", "loss", ",", "and", "loss", "sampling", "can", "be", "used", "to", "address", "spatial", "correlation", ".", "We", "explore", "training", "with", "sampling", "in", "Section", "[", "reference", "]", ",", "and", "do", "not", "find", "that", "it", "yields", "faster", "or", "better", "convergence", "for", "dense", "prediction", ".", "Whole", "image", "training", "is", "effective", "and", "efficient", ".", "section", ":", "Segmentation", "Architecture", "We", "cast", "ILSVRC", "classifiers", "into", "FCNs", "and", "augment", "them", "for", "dense", "prediction", "with", "in", "-", "network", "upsampling", "and", "a", "pixelwise", "loss", ".", "We", "train", "for", "segmentation", "by", "fine", "-", "tuning", ".", "Next", ",", "we", "add", "skips", "between", "layers", "to", "fuse", "coarse", ",", "semantic", "and", "local", ",", "appearance", "information", ".", "This", "skip", "architecture", "is", "learned", "end", "-", "to", "-", "end", "to", "refine", "the", "semantics", "and", "spatial", "precision", "of", "the", "output", ".", "For", "this", "investigation", ",", "we", "train", "and", "validate", "on", "the", "PASCAL", "VOC", "2011", "segmentation", "challenge", ".", "We", "train", "with", "a", "per", "-", "pixel", "softmax", "loss", "and", "validate", "with", "the", "standard", "metric", "of", "mean", "pixel", "intersection", "over", "union", ",", "with", "the", "mean", "taken", "over", "all", "classes", ",", "including", "background", ".", "The", "training", "ignores", "pixels", "that", "are", "masked", "out", "(", "as", "ambiguous", "or", "difficult", ")", "in", "the", "ground", "truth", ".", "subsection", ":", "From", "classifier", "to", "dense", "FCN", "We", "begin", "by", "convolutionalizing", "proven", "classification", "architectures", "as", "in", "Section", "[", "reference", "]", ".", "We", "consider", "the", "AlexNet", "architecture", "that", "won", "ILSVRC12", ",", "as", "well", "as", "the", "VGG", "nets", "and", "the", "GoogLeNet", "which", "did", "exceptionally", "well", "in", "ILSVRC14", ".", "We", "pick", "the", "VGG", "16", "-", "layer", "net", ",", "which", "we", "found", "to", "be", "equivalent", "to", "the", "19", "-", "layer", "net", "on", "this", "task", ".", "For", "GoogLeNet", ",", "we", "use", "only", "the", "final", "loss", "layer", ",", "and", "improve", "performance", "by", "discarding", "the", "final", "average", "pooling", "layer", ".", "We", "decapitate", "each", "net", "by", "discarding", "the", "final", "classifier", "layer", ",", "and", "convert", "all", "fully", "connected", "layers", "to", "convolutions", ".", "We", "append", "a", "convolution", "with", "channel", "dimension", "to", "predict", "scores", "for", "each", "of", "the", "PASCAL", "classes", "(", "including", "background", ")", "at", "each", "of", "the", "coarse", "output", "locations", ",", "followed", "by", "a", "(", "backward", ")", "convolution", "layer", "to", "bilinearly", "upsample", "the", "coarse", "outputs", "to", "pixelwise", "outputs", "as", "described", "in", "Section", "[", "reference", "]", ".", "Table", "[", "reference", "]", "compares", "the", "preliminary", "validation", "results", "along", "with", "the", "basic", "characteristics", "of", "each", "net", ".", "We", "report", "the", "best", "results", "achieved", "after", "convergence", "at", "a", "fixed", "learning", "rate", "(", "at", "least", "175", "epochs", ")", ".", "Our", "training", "for", "this", "comparison", "follows", "the", "practices", "for", "classification", "networks", ".", "We", "train", "by", "SGD", "with", "momentum", ".", "Gradients", "are", "accumulated", "over", "20", "images", ".", "We", "set", "fixed", "learning", "rates", "of", ",", ",", "and", "for", "FCN", "-", "AlexNet", ",", "FCN", "-", "VGG16", ",", "and", "FCN", "-", "GoogLeNet", ",", "respectively", ",", "chosen", "by", "line", "search", ".", "We", "use", "momentum", ",", "weight", "decay", "of", "or", ",", "and", "doubled", "learning", "rate", "for", "biases", ".", "We", "zero", "-", "initialize", "the", "class", "scoring", "layer", ",", "as", "random", "initialization", "yielded", "neither", "better", "performance", "nor", "faster", "convergence", ".", "Dropout", "is", "included", "where", "used", "in", "the", "original", "classifier", "nets", "(", "however", ",", "training", "without", "it", "made", "little", "to", "no", "difference", ")", ".", "Fine", "-", "tuning", "from", "classification", "to", "segmentation", "gives", "reasonable", "predictions", "from", "each", "net", ".", "Even", "the", "worst", "model", "achieved", "of", "the", "previous", "best", "performance", ".", "FCN", "-", "VGG16", "already", "appears", "to", "be", "better", "than", "previous", "methods", "at", "56.0", "mean", "IU", "on", "val", ",", "compared", "to", "52.6", "on", "test", ".", "Although", "VGG", "and", "GoogLeNet", "are", "similarly", "accurate", "as", "classifiers", ",", "our", "FCN", "-", "GoogLeNet", "did", "not", "match", "FCN", "-", "VGG16", ".", "We", "select", "FCN", "-", "VGG16", "as", "our", "base", "network", ".", "subsection", ":", "Image", "-", "to", "-", "image", "learning", "batch", "size", "pixel", "acc", ".", "mean", "acc", ".", "mean", "IU", "f.w", ".", "IU", "The", "image", "-", "to", "-", "image", "learning", "setting", "includes", "high", "effective", "batch", "size", "and", "correlated", "inputs", ".", "This", "optimization", "requires", "some", "attention", "to", "properly", "tune", "FCNs", ".", "We", "begin", "with", "the", "loss", ".", "We", "do", "not", "normalize", "the", "loss", ",", "so", "that", "every", "pixel", "has", "the", "same", "weight", "regardless", "of", "the", "batch", "and", "image", "dimensions", ".", "Thus", "we", "use", "a", "small", "learning", "rate", "since", "the", "loss", "is", "summed", "spatially", "over", "all", "pixels", ".", "We", "consider", "two", "regimes", "for", "batch", "size", ".", "In", "the", "first", ",", "gradients", "are", "accumulated", "over", "20", "images", ".", "Accumulation", "reduces", "the", "memory", "required", "and", "respects", "the", "different", "dimensions", "of", "each", "input", "by", "reshaping", "the", "network", ".", "We", "picked", "this", "batch", "size", "empirically", "to", "result", "in", "reasonable", "convergence", ".", "Learning", "in", "this", "way", "is", "similar", "to", "standard", "classification", "training", ":", "each", "minibatch", "contains", "several", "images", "and", "has", "a", "varied", "distribution", "of", "class", "labels", ".", "The", "nets", "compared", "in", "Table", "[", "reference", "]", "are", "optimized", "in", "this", "fashion", ".", "However", ",", "batching", "is", "not", "the", "only", "way", "to", "do", "image", "-", "wise", "learning", ".", "In", "the", "second", "regime", ",", "batch", "size", "one", "is", "used", "for", "online", "learning", ".", "Properly", "tuned", ",", "online", "learning", "achieves", "higher", "accuracy", "and", "faster", "convergence", "in", "both", "number", "of", "iterations", "and", "wall", "clock", "time", ".", "Additionally", ",", "we", "try", "a", "higher", "momentum", "of", ",", "which", "increases", "the", "weight", "on", "recent", "gradients", "in", "a", "similar", "way", "to", "batching", ".", "See", "Table", "[", "reference", "]", "for", "the", "comparison", "of", "accumulation", ",", "online", ",", "and", "high", "momentum", "or", "“", "heavy", "”", "learning", "(", "discussed", "further", "in", "Section", "[", "reference", "]", ")", ".", "subsection", ":", "Combining", "what", "and", "where", "We", "define", "a", "new", "fully", "convolutional", "net", "for", "segmentation", "that", "combines", "layers", "of", "the", "feature", "hierarchy", "and", "refines", "the", "spatial", "precision", "of", "the", "output", ".", "See", "Figure", "[", "reference", "]", ".", "While", "fully", "convolutionalized", "classifiers", "fine", "-", "tuned", "to", "semantic", "segmentation", "both", "recognize", "and", "localize", ",", "as", "shown", "in", "Section", "[", "reference", "]", ",", "these", "networks", "can", "be", "improved", "to", "make", "direct", "use", "of", "shallower", ",", "more", "local", "features", ".", "Even", "though", "these", "base", "networks", "score", "highly", "on", "the", "standard", "metrics", ",", "their", "output", "is", "dissatisfyingly", "coarse", "(", "see", "Figure", "[", "reference", "]", ")", ".", "The", "stride", "of", "the", "network", "prediction", "limits", "the", "scale", "of", "detail", "in", "the", "upsampled", "output", ".", "We", "address", "this", "by", "adding", "skips", "that", "fuse", "layer", "outputs", ",", "in", "particular", "to", "include", "shallower", "layers", "with", "finer", "strides", "in", "prediction", ".", "This", "turns", "a", "line", "topology", "into", "a", "DAG", ":", "edges", "skip", "ahead", "from", "shallower", "to", "deeper", "layers", ".", "It", "is", "natural", "to", "make", "more", "local", "predictions", "from", "shallower", "layers", "since", "their", "receptive", "fields", "are", "smaller", "and", "see", "fewer", "pixels", ".", "Once", "augmented", "with", "skips", ",", "the", "network", "makes", "and", "fuses", "predictions", "from", "several", "streams", "that", "are", "learned", "jointly", "and", "end", "-", "to", "-", "end", ".", "Combining", "fine", "layers", "and", "coarse", "layers", "lets", "the", "model", "make", "local", "predictions", "that", "respect", "global", "structure", ".", "This", "crossing", "of", "layers", "and", "resolutions", "is", "a", "learned", ",", "nonlinear", "counterpart", "to", "the", "multi", "-", "scale", "representation", "of", "the", "Laplacian", "pyramid", ".", "By", "analogy", "to", "the", "jet", "of", "Koenderick", "and", "van", "Doorn", ",", "we", "call", "our", "feature", "hierarchy", "the", "deep", "jet", ".", "Layer", "fusion", "is", "essentially", "an", "elementwise", "operation", ".", "However", ",", "the", "correspondence", "of", "elements", "across", "layers", "is", "complicated", "by", "resampling", "and", "padding", ".", "Thus", ",", "in", "general", ",", "layers", "to", "be", "fused", "must", "be", "aligned", "by", "scaling", "and", "cropping", ".", "We", "bring", "two", "layers", "into", "scale", "agreement", "by", "upsampling", "the", "lower", "-", "resolution", "layer", ",", "doing", "so", "in", "-", "network", "as", "explained", "in", "Section", "[", "reference", "]", ".", "Cropping", "removes", "any", "portion", "of", "the", "upsampled", "layer", "which", "extends", "beyond", "the", "other", "layer", "due", "to", "padding", ".", "This", "results", "in", "layers", "of", "equal", "dimensions", "in", "exact", "alignment", ".", "The", "offset", "of", "the", "cropped", "region", "depends", "on", "the", "resampling", "and", "padding", "parameters", "of", "all", "intermediate", "layers", ".", "Determining", "the", "crop", "that", "results", "in", "exact", "correspondence", "can", "be", "intricate", ",", "but", "it", "follows", "automatically", "from", "the", "network", "definition", "(", "and", "we", "include", "code", "for", "it", "in", "Caffe", ")", ".", "Having", "spatially", "aligned", "the", "layers", ",", "we", "next", "pick", "a", "fusion", "operation", ".", "We", "fuse", "features", "by", "concatenation", ",", "and", "immediately", "follow", "with", "classification", "by", "a", "“", "score", "layer", "”", "consisting", "of", "a", "convolution", ".", "Rather", "than", "storing", "concatenated", "features", "in", "memory", ",", "we", "commute", "the", "concatenation", "and", "subsequent", "classification", "(", "as", "both", "are", "linear", ")", ".", "Thus", ",", "our", "skips", "are", "implemented", "by", "first", "scoring", "each", "layer", "to", "be", "fused", "by", "convolution", ",", "carrying", "out", "any", "necessary", "interpolation", "and", "alignment", ",", "and", "then", "summing", "the", "scores", ".", "We", "also", "considered", "max", "fusion", ",", "but", "found", "learning", "to", "be", "difficult", "due", "to", "gradient", "switching", ".", "The", "score", "layer", "parameters", "are", "zero", "-", "initialized", "when", "a", "skip", "is", "added", ",", "so", "that", "they", "do", "not", "interfere", "with", "existing", "predictions", "of", "other", "streams", ".", "Once", "all", "layers", "have", "been", "fused", ",", "the", "final", "prediction", "is", "then", "upsampled", "back", "to", "image", "resolution", ".", "Skip", "Architectures", "for", "Segmentation", "We", "define", "a", "skip", "architecture", "to", "extend", "FCN", "-", "VGG16", "to", "a", "three", "-", "stream", "net", "with", "eight", "pixel", "stride", "shown", "in", "Figure", "[", "reference", "]", ".", "Adding", "a", "skip", "from", "pool4", "halves", "the", "stride", "by", "scoring", "from", "this", "stride", "sixteen", "layer", ".", "The", "interpolation", "layer", "of", "the", "skip", "is", "initialized", "to", "bilinear", "interpolation", ",", "but", "is", "not", "fixed", "so", "that", "it", "can", "be", "learned", "as", "described", "in", "Section", "[", "reference", "]", ".", "We", "call", "this", "two", "-", "stream", "net", "FCN", "-", "16s", ",", "and", "likewise", "define", "FCN", "-", "8s", "by", "adding", "a", "further", "skip", "from", "pool3", "to", "make", "stride", "eight", "predictions", ".", "(", "Note", "that", "predicting", "at", "stride", "eight", "does", "not", "significantly", "limit", "the", "maximum", "achievable", "mean", "IU", ";", "see", "Section", "[", "reference", "]", ".", ")", "We", "experiment", "with", "both", "staged", "training", "and", "all", "-", "at", "-", "once", "training", ".", "In", "the", "staged", "version", ",", "we", "learn", "the", "single", "-", "stream", "FCN", "-", "32s", ",", "then", "upgrade", "to", "the", "two", "-", "stream", "FCN", "-", "16s", "and", "continue", "learning", ",", "and", "finally", "upgrade", "to", "the", "three", "-", "stream", "FCN", "-", "8s", "and", "finish", "learning", ".", "At", "each", "stage", "the", "net", "is", "learned", "end", "-", "to", "-", "end", ",", "initialized", "with", "the", "parameters", "of", "the", "earlier", "net", ".", "The", "learning", "rate", "is", "dropped", "from", "FCN", "-", "32s", "to", "FCN", "-", "16s", "and", "more", "from", "FCN", "-", "16s", "to", "FCN", "-", "8s", ",", "which", "we", "found", "to", "be", "necessary", "for", "continued", "improvements", ".", "Learning", "all", "-", "at", "-", "once", "rather", "than", "in", "stages", "gives", "nearly", "equivalent", "results", ",", "while", "training", "is", "faster", "and", "less", "tedious", ".", "However", ",", "disparate", "feature", "scales", "make", "naïve", "training", "prone", "to", "divergence", ".", "To", "remedy", "this", "we", "scale", "each", "stream", "by", "a", "fixed", "constant", ",", "for", "a", "similar", "in", "-", "network", "effect", "to", "the", "staged", "learning", "rate", "adjustments", ".", "These", "constants", "are", "picked", "to", "approximately", "equalize", "average", "feature", "norms", "across", "streams", ".", "(", "Other", "normalization", "schemes", "should", "have", "similar", "effect", ".", ")", "With", "FCN", "-", "16s", "validation", "score", "improves", "to", "65.0", "mean", "IU", ",", "and", "FCN", "-", "8s", "brings", "a", "minor", "improvement", "to", "65.5", ".", "At", "this", "point", "our", "fusion", "improvements", "have", "met", "diminishing", "returns", ",", "so", "we", "do", "not", "continue", "fusing", "even", "shallower", "layers", ".", "To", "identify", "the", "contribution", "of", "the", "skips", "we", "compare", "scoring", "from", "the", "intermediate", "layers", "in", "isolation", ",", "which", "results", "in", "poor", "performance", ",", "or", "dropping", "the", "learning", "rate", "without", "adding", "skips", ",", "which", "gives", "negligible", "improvement", "in", "score", "without", "refining", "the", "visual", "quality", "of", "output", ".", "All", "skip", "comparisons", "are", "reported", "in", "Table", "[", "reference", "]", ".", "Figure", "[", "reference", "]", "shows", "the", "progressively", "finer", "structure", "of", "the", "output", ".", "pixel", "acc", ".", "mean", "acc", ".", "mean", "IU", "f.w", ".", "IU", "subsection", ":", "Experimental", "framework", "Fine", "-", "tuning", "We", "fine", "-", "tune", "all", "layers", "by", "backpropagation", "through", "the", "whole", "net", ".", "Fine", "-", "tuning", "the", "output", "classifier", "alone", "yields", "only", "73", "%", "of", "the", "full", "fine", "-", "tuning", "performance", "as", "compared", "in", "Table", "[", "reference", "]", ".", "Fine", "-", "tuning", "in", "stages", "takes", "36", "hours", "on", "a", "single", "GPU", ".", "Learning", "FCN", "-", "8s", "all", "-", "at", "-", "once", "takes", "half", "the", "time", "to", "reach", "comparable", "accuracy", ".", "Training", "from", "scratch", "gives", "substantially", "lower", "accuracy", ".", "More", "training", "data", "The", "PASCAL", "VOC", "2011", "segmentation", "training", "set", "labels", "1", ",", "112", "images", ".", "Hariharan", "et", "al", ".", "collected", "labels", "for", "a", "larger", "set", "of", "8", ",", "498", "PASCAL", "training", "images", ",", "which", "was", "used", "to", "train", "the", "previous", "best", "system", ",", "SDS", ".", "This", "training", "data", "improves", "the", "FCN", "-", "32s", "validation", "score", "from", "57.7", "to", "63.6", "mean", "IU", "and", "improves", "the", "FCN", "-", "AlexNet", "score", "from", "39.8", "to", "48.0", "mean", "IU", ".", "Loss", "The", "per", "-", "pixel", ",", "unnormalized", "softmax", "loss", "is", "a", "natural", "choice", "for", "segmenting", "images", "of", "any", "size", "into", "disjoint", "classes", ",", "so", "we", "train", "our", "nets", "with", "it", ".", "The", "softmax", "operation", "induces", "competition", "between", "classes", "and", "promotes", "the", "most", "confident", "prediction", ",", "but", "it", "is", "not", "clear", "that", "this", "is", "necessary", "or", "helpful", ".", "For", "comparison", ",", "we", "train", "with", "the", "sigmoid", "cross", "-", "entropy", "loss", "and", "find", "that", "it", "gives", "similar", "results", ",", "even", "though", "it", "normalizes", "each", "class", "prediction", "independently", ".", "Patch", "sampling", "As", "explained", "in", "Section", "[", "reference", "]", ",", "our", "whole", "image", "training", "effectively", "batches", "each", "image", "into", "a", "regular", "grid", "of", "large", ",", "overlapping", "patches", ".", "By", "contrast", ",", "prior", "work", "randomly", "samples", "patches", "over", "a", "full", "dataset", ",", "potentially", "resulting", "in", "higher", "variance", "batches", "that", "may", "accelerate", "convergence", ".", "We", "study", "this", "tradeoff", "by", "spatially", "sampling", "the", "loss", "in", "the", "manner", "described", "earlier", ",", "making", "an", "independent", "choice", "to", "ignore", "each", "final", "layer", "cell", "with", "some", "probability", ".", "To", "avoid", "changing", "the", "effective", "batch", "size", ",", "we", "simultaneously", "increase", "the", "number", "of", "images", "per", "batch", "by", "a", "factor", ".", "Note", "that", "due", "to", "the", "efficiency", "of", "convolution", ",", "this", "form", "of", "rejection", "sampling", "is", "still", "faster", "than", "patchwise", "training", "for", "large", "enough", "values", "of", "(", "e.g.", ",", "at", "least", "for", "according", "to", "the", "numbers", "in", "Section", "[", "reference", "]", ")", ".", "Figure", "[", "reference", "]", "shows", "the", "effect", "of", "this", "form", "of", "sampling", "on", "convergence", ".", "We", "find", "that", "sampling", "does", "not", "have", "a", "significant", "effect", "on", "convergence", "rate", "compared", "to", "whole", "image", "training", ",", "but", "takes", "significantly", "more", "time", "due", "to", "the", "larger", "number", "of", "images", "that", "need", "to", "be", "considered", "per", "batch", ".", "We", "therefore", "choose", "unsampled", ",", "whole", "image", "training", "in", "our", "other", "experiments", ".", "Class", "balancing", "Fully", "convolutional", "training", "can", "balance", "classes", "by", "weighting", "or", "sampling", "the", "loss", ".", "Although", "our", "labels", "are", "mildly", "unbalanced", "(", "about", "are", "background", ")", ",", "we", "find", "class", "balancing", "unnecessary", ".", "Dense", "Prediction", "The", "scores", "are", "upsampled", "to", "the", "input", "dimensions", "by", "backward", "convolution", "layers", "within", "the", "net", ".", "Final", "layer", "backward", "convolution", "weights", "are", "fixed", "to", "bilinear", "interpolation", ",", "while", "intermediate", "upsampling", "layers", "are", "initialized", "to", "bilinear", "interpolation", ",", "and", "then", "learned", ".", "This", "simple", ",", "end", "-", "to", "-", "end", "method", "is", "accurate", "and", "fast", ".", "Augmentation", "We", "tried", "augmenting", "the", "training", "data", "by", "randomly", "mirroring", "and", "“", "jittering", "”", "the", "images", "by", "translating", "them", "up", "to", "32", "pixels", "(", "the", "coarsest", "scale", "of", "prediction", ")", "in", "each", "direction", ".", "This", "yielded", "no", "noticeable", "improvement", ".", "Implementation", "All", "models", "are", "trained", "and", "tested", "with", "Caffe", "on", "a", "single", "NVIDIA", "Titan", "X.", "Our", "models", "and", "code", "are", "publicly", "available", "at", ".", "section", ":", "Results", "We", "test", "our", "FCN", "on", "semantic", "segmentation", "and", "scene", "parsing", ",", "exploring", "PASCAL", "VOC", ",", "NYUDv2", ",", "SIFT", "Flow", ",", "and", "PASCAL", "-", "Context", ".", "Although", "these", "tasks", "have", "historically", "distinguished", "between", "objects", "and", "regions", ",", "we", "treat", "both", "uniformly", "as", "pixel", "prediction", ".", "We", "evaluate", "our", "FCN", "skip", "architecture", "on", "each", "of", "these", "datasets", ",", "and", "then", "extend", "it", "to", "multi", "-", "modal", "input", "for", "NYUDv2", "and", "multi", "-", "task", "prediction", "for", "the", "semantic", "and", "geometric", "labels", "of", "SIFT", "Flow", ".", "All", "experiments", "follow", "the", "same", "network", "architecture", "and", "optimization", "settings", "decided", "on", "in", "Section", "[", "reference", "]", ".", "Metrics", "We", "report", "metrics", "from", "common", "semantic", "segmentation", "and", "scene", "parsing", "evaluations", "that", "are", "variations", "on", "pixel", "accuracy", "and", "region", "intersection", "over", "union", "(", "IU", ")", ":", "pixel", "accuracy", ":", "mean", "accuraccy", ":", "mean", "IU", ":", "frequency", "weighted", "IU", ":", "where", "is", "the", "number", "of", "pixels", "of", "class", "predicted", "to", "belong", "to", "class", ",", "there", "are", "different", "classes", ",", "and", "is", "the", "total", "number", "of", "pixels", "of", "class", ".", "PASCAL", "VOC", "Table", "[", "reference", "]", "gives", "the", "performance", "of", "our", "FCN", "-", "8s", "on", "the", "test", "sets", "of", "PASCAL", "VOC", "2011", "and", "2012", ",", "and", "compares", "it", "to", "the", "previous", "best", ",", "SDS", ",", "and", "the", "well", "-", "known", "R", "-", "CNN", ".", "We", "achieve", "the", "best", "results", "on", "mean", "IU", "by", "30", "%", "relative", ".", "Inference", "time", "is", "reduced", "(", "convnet", "only", ",", "ignoring", "proposals", "and", "refinement", ")", "or", "(", "overall", ")", ".", "NYUDv2", "is", "an", "RGB", "-", "D", "dataset", "collected", "using", "the", "Microsoft", "Kinect", ".", "It", "has", "1", ",", "449", "RGB", "-", "D", "images", ",", "with", "pixelwise", "labels", "that", "have", "been", "coalesced", "into", "a", "40", "class", "semantic", "segmentation", "task", "by", "Gupta", "et", "al", ".", ".", "We", "report", "results", "on", "the", "standard", "split", "of", "795", "training", "images", "and", "654", "testing", "images", ".", "Table", "[", "reference", "]", "gives", "the", "performance", "of", "several", "net", "variations", ".", "First", "we", "train", "our", "unmodified", "coarse", "model", "(", "FCN", "-", "32s", ")", "on", "RGB", "images", ".", "To", "add", "depth", "information", ",", "we", "train", "on", "a", "model", "upgraded", "to", "take", "four", "-", "channel", "RGB", "-", "D", "input", "(", "early", "fusion", ")", ".", "This", "provides", "little", "benefit", ",", "perhaps", "due", "to", "similar", "number", "of", "parameters", "or", "the", "difficulty", "of", "propagating", "meaningful", "gradients", "all", "the", "way", "through", "the", "net", ".", "Following", "the", "success", "of", "Gupta", "et", "al", ".", ",", "we", "try", "the", "three", "-", "dimensional", "HHA", "encoding", "of", "depth", "and", "train", "a", "net", "on", "just", "this", "information", ".", "To", "effectively", "combine", "color", "and", "depth", ",", "we", "define", "a", "“", "late", "fusion", "”", "of", "RGB", "and", "HHA", "that", "averages", "the", "final", "layer", "scores", "from", "both", "nets", "and", "learn", "the", "resulting", "two", "-", "stream", "net", "end", "-", "to", "-", "end", ".", "This", "late", "fusion", "RGB", "-", "HHA", "net", "is", "the", "most", "accurate", ".", "SIFT", "Flow", "is", "a", "dataset", "of", "2", ",", "688", "images", "with", "pixel", "labels", "for", "33", "semantic", "classes", "(", "“", "bridge", "”", ",", "“", "mountain", "”", ",", "“", "sun", "”", ")", ",", "as", "well", "as", "three", "geometric", "classes", "(", "“", "horizontal", "”", ",", "“", "vertical", "”", ",", "and", "“", "sky", "”", ")", ".", "An", "FCN", "can", "naturally", "learn", "a", "joint", "representation", "that", "simultaneously", "predicts", "both", "types", "of", "labels", ".", "We", "learn", "a", "two", "-", "headed", "version", "of", "FCN", "-", "32", "/", "16", "/", "8s", "with", "semantic", "and", "geometric", "prediction", "layers", "and", "losses", ".", "This", "net", "performs", "as", "well", "on", "both", "tasks", "as", "two", "independently", "trained", "nets", ",", "while", "learning", "and", "inference", "are", "essentially", "as", "fast", "as", "each", "independent", "net", "by", "itself", ".", "The", "results", "in", "Table", "[", "reference", "]", ",", "computed", "on", "the", "standard", "split", "into", "2", ",", "488", "training", "and", "200", "test", "images", ",", "show", "better", "performance", "on", "both", "tasks", ".", "pixel", "acc", ".", "mean", "acc", ".", "mean", "IU", "f.w", ".", "IU", "pixel", "acc", ".", "mean", "acc", ".", "mean", "IU", "f.w", ".", "IU", "geom", ".", "acc", ".", "pixel", "acc", ".", "mean", "acc", ".", "mean", "IU", "f.w", ".", "IU", "PASCAL", "-", "Context", "provides", "whole", "scene", "annotations", "of", "PASCAL", "VOC", "2010", ".", "While", "there", "are", "400", "+", "classes", ",", "we", "follow", "the", "59", "class", "task", "defined", "by", "that", "picks", "the", "most", "frequent", "classes", ".", "We", "train", "and", "evaluate", "on", "the", "training", "and", "val", "sets", "respectively", ".", "In", "Table", "[", "reference", "]", "we", "compare", "to", "the", "previous", "best", "result", "on", "this", "task", ".", "FCN", "-", "8s", "scores", "39.1", "mean", "IU", "for", "a", "relative", "improvement", "of", "more", "than", "10", "%", ".", "section", ":", "Analysis", "We", "examine", "the", "learning", "and", "inference", "of", "fully", "convolutional", "networks", ".", "Masking", "experiments", "investigate", "the", "role", "of", "context", "and", "shape", "by", "reducing", "the", "input", "to", "only", "foreground", ",", "only", "background", ",", "or", "shape", "alone", ".", "Defining", "a", "“", "null", "”", "background", "model", "checks", "the", "necessity", "of", "learning", "a", "background", "classifier", "for", "semantic", "segmentation", ".", "We", "detail", "an", "approximation", "between", "momentum", "and", "batch", "size", "to", "further", "tune", "whole", "image", "learning", ".", "Finally", ",", "we", "measure", "bounds", "on", "task", "accuracy", "for", "given", "output", "resolutions", "to", "show", "there", "is", "still", "much", "to", "improve", ".", "subsection", ":", "Cues", "Given", "the", "large", "receptive", "field", "size", "of", "an", "FCN", ",", "it", "is", "natural", "to", "wonder", "about", "the", "relative", "importance", "of", "foreground", "and", "background", "pixels", "in", "the", "prediction", ".", "Is", "foreground", "appearance", "sufficient", "for", "inference", ",", "or", "does", "the", "context", "influence", "the", "output", "?", "Conversely", ",", "can", "a", "network", "learn", "to", "recognize", "a", "class", "by", "its", "shape", "and", "context", "alone", "?", "Masking", "To", "explore", "these", "issues", "we", "experiment", "with", "masked", "versions", "of", "the", "standard", "PASCAL", "VOC", "segmentation", "challenge", ".", "We", "both", "mask", "input", "to", "networks", "trained", "on", "normal", "PASCAL", ",", "and", "learn", "new", "networks", "on", "the", "masked", "PASCAL", ".", "See", "Table", "[", "reference", "]", "for", "masked", "results", ".", "Masking", "the", "foreground", "at", "inference", "time", "is", "catastrophic", ".", "However", ",", "masking", "the", "foreground", "during", "learning", "yields", "a", "network", "capable", "of", "recognizing", "object", "segments", "without", "observing", "a", "single", "pixel", "of", "the", "labeled", "class", ".", "Masking", "the", "background", "has", "little", "effect", "overall", "but", "does", "lead", "to", "class", "confusion", "in", "certain", "cases", ".", "When", "the", "background", "is", "masked", "during", "both", "learning", "and", "inference", ",", "the", "network", "unsurprisingly", "achieves", "nearly", "perfect", "background", "accuracy", ";", "however", "certain", "classes", "are", "more", "confused", ".", "All", "-", "in", "-", "all", "this", "suggests", "that", "FCNs", "do", "incorporate", "context", "even", "though", "decisions", "are", "driven", "by", "foreground", "pixels", ".", "To", "separate", "the", "contribution", "of", "shape", ",", "we", "learn", "a", "net", "restricted", "to", "the", "simple", "input", "of", "foreground", "/", "background", "masks", ".", "The", "accuracy", "in", "this", "shape", "-", "only", "condition", "is", "lower", "than", "when", "only", "the", "foreground", "is", "masked", ",", "suggesting", "that", "the", "net", "is", "capable", "of", "learning", "context", "to", "boost", "recognition", ".", "Nonetheless", ",", "it", "is", "surprisingly", "accurate", ".", "See", "Figure", "[", "reference", "]", ".", "Background", "modeling", "It", "is", "standard", "in", "detection", "and", "semantic", "segmentation", "to", "have", "a", "background", "model", ".", "This", "model", "usually", "takes", "the", "same", "form", "as", "the", "models", "for", "the", "classes", "of", "interest", ",", "but", "is", "supervised", "by", "negative", "instances", ".", "In", "our", "experiments", "we", "have", "followed", "the", "same", "approach", ",", "learning", "parameters", "to", "score", "all", "classes", "including", "background", ".", "Is", "this", "actually", "necessary", ",", "or", "do", "class", "models", "suffice", "?", "To", "investigate", ",", "we", "define", "a", "net", "with", "a", "“", "null", "”", "background", "model", "that", "gives", "a", "constant", "score", "of", "zero", ".", "Instead", "of", "training", "with", "the", "softmax", "loss", ",", "which", "induces", "competition", "by", "normalizing", "across", "classes", ",", "we", "train", "with", "the", "sigmoid", "cross", "-", "entropy", "loss", ",", "which", "independently", "normalizes", "each", "score", ".", "For", "inference", "each", "pixel", "is", "assigned", "the", "highest", "scoring", "class", ".", "In", "all", "other", "respects", "the", "experiment", "is", "identical", "to", "our", "FCN", "-", "32s", "on", "PASCAL", "VOC", ".", "The", "null", "background", "net", "scores", "1", "point", "lower", "than", "the", "reference", "FCN", "-", "32s", "and", "a", "control", "FCN", "-", "32s", "trained", "on", "all", "classes", "including", "background", "with", "the", "sigmoid", "cross", "-", "entropy", "loss", ".", "To", "put", "this", "drop", "in", "perspective", ",", "note", "that", "discarding", "the", "background", "model", "in", "this", "way", "reduces", "the", "total", "number", "of", "parameters", "by", "less", "than", "0.1", "%", ".", "Nonetheless", ",", "this", "result", "suggests", "that", "learning", "a", "dedicated", "background", "model", "for", "semantic", "segmentation", "is", "not", "vital", ".", "Image", "Ground", "Truth", "Output", "Input", "subsection", ":", "Momentum", "and", "batch", "size", "In", "comparing", "optimization", "schemes", "for", "FCNs", ",", "we", "find", "that", "“", "heavy", "”", "online", "learning", "with", "high", "momentum", "trains", "more", "accurate", "models", "in", "less", "wall", "clock", "time", "(", "see", "Section", "[", "reference", "]", ")", ".", "Here", "we", "detail", "a", "relationship", "between", "momentum", "and", "batch", "size", "that", "motivates", "heavy", "learning", ".", "By", "writing", "the", "updates", "computed", "by", "gradient", "accumulation", "as", "a", "non", "-", "recursive", "sum", ",", "we", "will", "see", "that", "momentum", "and", "batch", "size", "can", "be", "approximately", "traded", "off", ",", "which", "suggests", "alternative", "training", "parameters", ".", "Let", "be", "the", "step", "taken", "by", "minibatch", "SGD", "with", "momentum", "at", "time", ",", "where", "is", "the", "loss", "for", "example", "and", "parameters", ",", "is", "the", "momentum", ",", "is", "the", "batch", "size", ",", "and", "is", "the", "learning", "rate", ".", "Expanding", "this", "recurrence", "as", "an", "infinite", "sum", "with", "geometric", "coefficients", ",", "we", "have", "In", "other", "words", ",", "each", "example", "is", "included", "in", "the", "sum", "with", "coefficient", ",", "where", "the", "index", "orders", "the", "examples", "from", "most", "recently", "considered", "to", "least", "recently", "considered", ".", "Approximating", "this", "expression", "by", "dropping", "the", "floor", ",", "we", "see", "that", "learning", "with", "momentum", "and", "batch", "size", "appears", "to", "be", "similar", "to", "learning", "with", "momentum", "and", "batch", "size", "if", ".", "Note", "that", "this", "is", "not", "an", "exact", "equivalence", ":", "a", "smaller", "batch", "size", "results", "in", "more", "frequent", "weight", "updates", ",", "and", "may", "make", "more", "learning", "progress", "for", "the", "same", "number", "of", "gradient", "computations", ".", "For", "typical", "FCN", "values", "of", "momentum", "and", "a", "batch", "size", "of", "20", "images", ",", "an", "approximately", "equivalent", "training", "regime", "uses", "momentum", "and", "a", "batch", "size", "of", "one", ",", "resulting", "in", "online", "learning", ".", "In", "practice", ",", "we", "find", "that", "online", "learning", "works", "well", "and", "yields", "better", "FCN", "models", "in", "less", "wall", "clock", "time", ".", "subsection", ":", "Upper", "bounds", "on", "IU", "FCNs", "achieve", "good", "performance", "on", "the", "mean", "IU", "segmentation", "metric", "even", "with", "spatially", "coarse", "semantic", "prediction", ".", "To", "better", "understand", "this", "metric", "and", "the", "limits", "of", "this", "approach", "with", "respect", "to", "it", ",", "we", "compute", "approximate", "upper", "bounds", "on", "performance", "with", "prediction", "at", "various", "resolutions", ".", "We", "do", "this", "by", "downsampling", "ground", "truth", "images", "and", "then", "upsampling", "back", "to", "simulate", "the", "best", "results", "obtainable", "with", "a", "particular", "downsampling", "factor", ".", "The", "following", "table", "gives", "the", "mean", "IU", "on", "a", "subset", "of", "PASCAL", "2011", "val", "for", "various", "downsampling", "factors", ".", "Pixel", "-", "perfect", "prediction", "is", "clearly", "not", "necessary", "to", "achieve", "mean", "IU", "well", "above", "state", "-", "of", "-", "the", "-", "art", ",", "and", ",", "conversely", ",", "mean", "IU", "is", "a", "not", "a", "good", "measure", "of", "fine", "-", "scale", "accuracy", ".", "The", "gaps", "between", "oracle", "and", "state", "-", "of", "-", "the", "-", "art", "accuracy", "at", "every", "stride", "suggest", "that", "recognition", "and", "not", "resolution", "is", "the", "bottleneck", "for", "this", "metric", ".", "section", ":", "Conclusion", "Fully", "convolutional", "networks", "are", "a", "rich", "class", "of", "models", "that", "address", "many", "pixelwise", "tasks", ".", "FCNs", "for", "semantic", "segmentation", "dramatically", "improve", "accuracy", "by", "transferring", "pre", "-", "trained", "classifier", "weights", ",", "fusing", "different", "layer", "representations", ",", "and", "learning", "end", "-", "to", "-", "end", "on", "whole", "images", ".", "End", "-", "to", "-", "end", ",", "pixel", "-", "to", "-", "pixel", "operation", "simultaneously", "simplifies", "and", "speeds", "up", "learning", "and", "inference", ".", "All", "code", "for", "this", "paper", "is", "open", "source", "in", "Caffe", ",", "and", "all", "models", "are", "freely", "available", "in", "the", "Caffe", "Model", "Zoo", ".", "Further", "works", "have", "demonstrated", "the", "generality", "of", "fully", "convolutional", "networks", "for", "a", "variety", "of", "image", "-", "to", "-", "image", "tasks", ".", "section", ":", "Acknowledgements", "This", "work", "was", "supported", "in", "part", "by", "DARPA", "’s", "MSEE", "and", "SMISC", "programs", ",", "NSF", "awards", "IIS", "-", "1427425", ",", "IIS", "-", "1212798", ",", "IIS", "-", "1116411", ",", "and", "the", "NSF", "GRFP", ",", "Toyota", ",", "and", "the", "Berkeley", "Vision", "and", "Learning", "Center", ".", "We", "gratefully", "acknowledge", "NVIDIA", "for", "GPU", "donation", ".", "We", "thank", "Bharath", "Hariharan", "and", "Saurabh", "Gupta", "for", "their", "advice", "and", "dataset", "tools", ".", "We", "thank", "Sergio", "Guadarrama", "for", "reproducing", "GoogLeNet", "in", "Caffe", ".", "We", "thank", "Jitendra", "Malik", "for", "his", "helpful", "comments", ".", "Thanks", "to", "Wei", "Liu", "for", "pointing", "out", "an", "issue", "wth", "our", "SIFT", "Flow", "mean", "IU", "computation", "and", "an", "error", "in", "our", "frequency", "weighted", "mean", "IU", "formula", ".", "bibliography", ":", "References"], "entity": [[2, 5, ["Fully", "Convolutional", "Networks"], "Method"], [6, 8, ["Semantic", "Segmentation"], "Task"], [8, 10, ["Convolutional", "networks"], "Method"], [12, 14, ["visual", "models"], "Method"], [23, 25, ["convolutional", "networks"], "Method"], [48, 50, ["semantic", "segmentation"], "Task"], [76, 77, ["inference"], "Task"], [78, 79, ["learning"], "Task"], [95, 99, ["spatially", "dense", "prediction", "tasks"], "Task"], [104, 106, ["prior", "models"], "Method"], [110, 112, ["classification", "networks"], "Method"], [113, 114, ["AlexNet"], "Method"], [116, 118, ["VGG", "net"], "Method"], [120, 121, ["GoogLeNet"], "Method"], [137, 139, ["segmentation", "task"], "Task"], [144, 146, ["skip", "architecture"], "Method"], [167, 171, ["accurate", "and", "detailed", "segmentations"], "Task"], [178, 179, ["segmentation"], "Task"], [180, 182, ["PASCAL", "VOC"], "Material"], [190, 192, ["mean", "IU"], "Metric"], [202, 205, ["PASCAL", "-", "Context"], "Material"], [207, 208, ["inference"], "Task"], [219, 221, ["Semantic", "Segmentation"], "Task"], [222, 224, ["Convolutional", "Networks"], "Method"], [225, 227, ["Deep", "Learning"], "Method"], [228, 230, ["Transfer", "Learning"], "Method"], [233, 235, ["Convolutional", "networks"], "Method"], [239, 240, ["recognition"], "Task"], [241, 242, ["Convnets"], "Method"], [257, 259, ["local", "tasks"], "Task"], [272, 276, ["part", "and", "keypoint", "prediction"], "Task"], [278, 280, ["local", "correspondence"], "Task"], [289, 293, ["coarse", "to", "fine", "inference"], "Task"], [306, 307, ["convnets"], "Method"], [308, 310, ["semantic", "segmentation"], "Task"], [342, 343, ["FCNs"], "Method"], [357, 359, ["semantic", "segmentation"], "Task"], [379, 380, ["FCNs"], "Method"], [389, 391, ["pixelwise", "prediction"], "Task"], [396, 400, ["supervised", "pre", "-", "training"], "Task"], [401, 404, ["Fully", "convolutional", "versions"], "Method"], [417, 418, ["learning"], "Task"], [419, 420, ["inference"], "Task"], [432, 435, ["dense", "feedforward", "computation"], "Method"], [436, 437, ["backpropagation"], "Method"], [438, 443, ["In", "-", "network", "upsampling", "layers"], "Method"], [444, 446, ["pixelwise", "prediction"], "Task"], [447, 450, ["learning", "in", "nets"], "Task"], [451, 452, ["subsampling"], "Method"], [474, 476, ["Patchwise", "training"], "Method"], [484, 487, ["fully", "convolutional", "training"], "Method"], [509, 513, ["post", "-", "hoc", "refinement"], "Method"], [514, 516, ["random", "fields"], "Method"], [517, 519, ["local", "classifiers"], "Method"], [526, 527, ["classification"], "Task"], [528, 530, ["dense", "prediction"], "Task"], [532, 534, ["classification", "nets"], "Method"], [535, 537, ["fully", "convolutional"], "Method"], [553, 555, ["small", "convnets"], "Method"], [556, 560, ["supervised", "pre", "-", "training"], "Method"], [561, 563, ["Semantic", "segmentation"], "Task"], [609, 611, ["deep", "networks"], "Method"], [612, 614, ["image", "classification"], "Method"], [615, 616, ["representations"], "Task"], [618, 622, ["accurate", ",", "pixelwise", "recognition"], "Task"], [633, 637, ["pre", "-", "trained", "networks"], "Method"], [638, 641, ["fully", "convolutional", "form"], "Method"], [647, 649, ["skip", "architecture"], "Method"], [659, 661, ["skip", "architecture"], "Method"], [725, 726, ["analysis"], "Task"], [744, 745, ["FCNs"], "Method"], [746, 748, ["Tuning", "optimization"], "Task"], [758, 760, ["skip", "architectures"], "Method"], [796, 799, ["merging", "object", "segmentation"], "Task"], [800, 802, ["scene", "parsing"], "Task"], [803, 806, ["unified", "pixelwise", "prediction"], "Task"], [817, 820, ["deep", "classification", "nets"], "Method"], [821, 822, ["FCNs"], "Method"], [826, 828, ["semantic", "segmentation"], "Task"], [829, 830, ["convnets"], "Method"], [834, 835, ["FCNs"], "Method"], [847, 854, ["in", "-", "network", "upsampling", "and", "skip", "layers"], "Method"], [866, 867, ["accuracy"], "Metric"], [880, 883, ["PASCAL", "-", "Context"], "Material"], [899, 900, ["FCN"], "Method"], [903, 905, ["recognition", "bounds"], "Method"], [906, 908, ["semantic", "segmentation"], "Task"], [920, 922, ["deep", "nets"], "Method"], [923, 925, ["image", "classification"], "Method"], [926, 928, ["transfer", "learning"], "Task"], [929, 930, ["Transfer"], "Task"], [935, 938, ["visual", "recognition", "tasks"], "Task"], [941, 942, ["detection"], "Task"], [948, 950, ["semantic", "segmentation"], "Task"], [951, 956, ["hybrid", "proposal", "-", "classifier", "models"], "Method"], [966, 968, ["classification", "nets"], "Method"], [982, 983, ["FCNs"], "Method"], [985, 987, ["prior", "models"], "Method"], [992, 995, ["Fully", "convolutional", "networks"], "Method"], [1004, 1005, ["convnet"], "Method"], [1022, 1023, ["LeNet"], "Method"], [1046, 1048, ["Viterbi", "decoding"], "Method"], [1057, 1059, ["convnet", "outputs"], "Method"], [1082, 1083, ["inference"], "Task"], [1084, 1087, ["learning", "fully", "convolutionally"], "Method"], [1088, 1089, ["detection"], "Task"], [1096, 1097, ["convnet"], "Method"], [1098, 1105, ["coarse", "multiclass", "segmentation", "of", "C.", "elegans", "tissues"], "Task"], [1106, 1109, ["fully", "convolutional", "inference"], "Method"], [1110, 1113, ["Fully", "convolutional", "computation"], "Method"], [1122, 1126, ["many", "-", "layered", "nets"], "Task"], [1136, 1138, ["semantic", "segmentation"], "Task"], [1144, 1146, ["image", "restoration"], "Task"], [1152, 1155, ["fully", "convolutional", "inference"], "Method"], [1156, 1159, ["Fully", "convolutional", "training"], "Method"], [1173, 1180, ["end", "-", "to", "-", "end", "part", "detector"], "Method"], [1181, 1183, ["spatial", "model"], "Method"], [1184, 1186, ["pose", "estimation"], "Task"], [1198, 1200, ["Dense", "prediction"], "Task"], [1201, 1202, ["convnets"], "Method"], [1207, 1208, ["convnets"], "Method"], [1209, 1212, ["dense", "prediction", "problems"], "Task"], [1214, 1216, ["semantic", "segmentation"], "Task"], [1232, 1234, ["boundary", "prediction"], "Task"], [1235, 1237, ["electron", "microscopy"], "Task"], [1248, 1254, ["hybrid", "convnet", "/", "nearest", "neighbor", "model"], "Method"], [1260, 1262, ["image", "restoration"], "Task"], [1263, 1265, ["depth", "estimation"], "Task"], [1277, 1284, ["small", "models", "restricting", "capacity", "and", "receptive", "fields"], "Method"], [1285, 1287, ["patchwise", "training"], "Method"], [1288, 1289, ["refinement"], "Method"], [1290, 1292, ["superpixel", "projection"], "Method"], [1293, 1296, ["random", "field", "regularization"], "Method"], [1297, 1298, ["filtering"], "Method"], [1300, 1302, ["local", "classification"], "Method"], [1311, 1316, ["multi", "-", "scale", "pyramid", "processing"], "Method"], [1336, 1338, ["patchwise", "training"], "Method"], [1364, 1365, ["FCNs"], "Method"], [1369, 1373, ["in", "-", "network", "upsampling"], "Task"], [1383, 1386, ["fully", "connected", "prediction"], "Method"], [1405, 1408, ["deep", "classification", "architectures"], "Method"], [1410, 1412, ["image", "classification"], "Method"], [1413, 1417, ["supervised", "pre", "-", "training"], "Task"], [1422, 1424, ["fully", "convolutionally"], "Method"], [1450, 1453, ["deep", "classification", "nets"], "Method"], [1454, 1456, ["semantic", "segmentation"], "Task"], [1461, 1466, ["hybrid", "proposal", "-", "classifier", "models"], "Method"], [1473, 1477, ["R", "-", "CNN", "system"], "Method"], [1484, 1486, ["region", "proposals"], "Method"], [1487, 1488, ["detection"], "Task"], [1489, 1491, ["semantic", "segmentation"], "Task"], [1493, 1495, ["instance", "segmentation"], "Task"], [1511, 1512, ["segmentation"], "Task"], [1514, 1516, ["PASCAL", "VOC"], "Material"], [1535, 1537, ["semantic", "segmentation"], "Task"], [1545, 1547, ["feature", "hierarchies"], "Method"], [1555, 1562, ["nonlinear", "local", "-", "to", "-", "global", "representation"], "Method"], [1572, 1574, ["Laplacian", "pyramid"], "Method"], [1577, 1581, ["multi", "-", "scale", "representation"], "Method"], [1583, 1585, ["fixed", "smoothing"], "Method"], [1586, 1587, ["differencing"], "Method"], [1603, 1607, ["compositions", "of", "partial", "derivatives"], "Method"], [1612, 1614, ["deep", "networks"], "Method"], [1657, 1662, ["fixed", "bottom", "-", "up", "grouping"], "Method"], [1673, 1674, ["FCNs"], "Method"], [1685, 1687, ["region", "proposals"], "Task"], [1691, 1693, ["depth", "regression"], "Task"], [1694, 1696, ["optical", "flow"], "Task"], [1712, 1713, ["FCNs"], "Method"], [1727, 1729, ["semantic", "segmentation"], "Task"], [1731, 1733, ["DeepLab", "models"], "Method"], [1737, 1739, ["dilated", "convolution"], "Method"], [1740, 1743, ["dense", "CRF", "inference"], "Method"], [1745, 1748, ["joint", "CRFasRNN", "model"], "Method"], [1758, 1759, ["CRF"], "Method"], [1767, 1768, ["fusion"], "Task"], [1772, 1774, ["global", "pooling"], "Method"], [1777, 1781, ["deconvolutional", "network", "”", "approach"], "Method"], [1782, 1784, ["restores", "resolution"], "Task"], [1787, 1791, ["stacks", "of", "learned", "deconvolution"], "Method"], [1793, 1794, ["unpooling"], "Method"], [1795, 1798, ["U", "-", "Net"], "Method"], [1799, 1801, ["skip", "layers"], "Method"], [1802, 1804, ["learned", "deconvolution"], "Method"], [1805, 1810, ["pixel", "labeling", "of", "microscopy", "images"], "Task"], [1812, 1814, ["dilation", "architecture"], "Method"], [1819, 1821, ["dilated", "convolution"], "Method"], [1836, 1839, ["Fully", "Convolutional", "Networks"], "Method"], [1844, 1845, ["convnet"], "Method"], [1907, 1908, ["Convnets"], "Method"], [1917, 1918, ["convolution"], "Method"], [1919, 1920, ["pooling"], "Method"], [1922, 1924, ["activation", "functions"], "Method"], [1983, 1985, ["matrix", "multiplication"], "Method"], [1986, 1987, ["convolution"], "Method"], [1988, 1990, ["average", "pooling"], "Method"], [1992, 1994, ["spatial", "max"], "Method"], [1995, 1997, ["max", "pooling"], "Method"], [2000, 2002, ["elementwise", "nonlinearity"], "Method"], [2004, 2006, ["activation", "function"], "Method"], [2031, 2033, ["transformation", "rule"], "Method"], [2035, 2037, ["general", "net"], "Method"], [2053, 2055, ["nonlinear", "filter"], "Method"], [2060, 2062, ["deep", "filter"], "Method"], [2068, 2069, ["FCN"], "Method"], [2092, 2097, ["real", "-", "valued", "loss", "function"], "Method"], [2100, 2101, ["FCN"], "Method"], [2137, 2139, ["spatial", "components"], "Method"], [2141, 2144, ["stochastic", "gradient", "descent"], "Method"], [2154, 2157, ["stochastic", "gradient", "descent"], "Method"], [2179, 2181, ["feedforward", "computation"], "Method"], [2182, 2183, ["backpropagation"], "Method"], [2213, 2215, ["classification", "nets"], "Method"], [2226, 2228, ["pixelwise", "prediction"], "Task"], [2257, 2259, ["fast", "scanning"], "Method"], [2269, 2271, ["network", "modification"], "Task"], [2291, 2292, ["convolution"], "Method"], [2300, 2301, ["training"], "Task"], [2302, 2304, ["patchwise", "sampling"], "Method"], [2316, 2318, ["image", "training"], "Task"], [2327, 2328, ["classifiers"], "Method"], [2329, 2331, ["dense", "prediction"], "Task"], [2332, 2334, ["recognition", "nets"], "Method"], [2336, 2337, ["LeNet"], "Method"], [2338, 2339, ["AlexNet"], "Method"], [2376, 2379, ["fully", "connected", "layers"], "Method"], [2384, 2387, ["convolutions", "with", "kernels"], "Method"], [2462, 2463, ["AlexNet"], "Method"], [2464, 2465, ["ms"], "Metric"], [2485, 2486, ["ms"], "Metric"], [2513, 2515, ["convolutionalized", "models"], "Method"], [2521, 2523, ["dense", "problems"], "Task"], [2524, 2526, ["semantic", "segmentation"], "Task"], [2552, 2554, ["computational", "efficiency"], "Metric"], [2557, 2558, ["optimization"], "Task"], [2560, 2561, ["convolution"], "Method"], [2571, 2572, ["ms"], "Metric"], [2577, 2578, ["ms"], "Metric"], [2580, 2584, ["fully", "convolutional", "output", "map"], "Method"], [2601, 2603, ["classification", "nets"], "Method"], [2604, 2606, ["fully", "convolutional"], "Method"], [2622, 2623, ["subsampling"], "Method"], [2625, 2627, ["classification", "nets"], "Method"], [2643, 2646, ["fully", "convolutional", "version"], "Method"], [2677, 2682, ["Shift", "-", "and", "-", "stitch"], "Method"], [2683, 2685, ["filter", "dilation"], "Method"], [2685, 2687, ["Dense", "predictions"], "Task"], [2763, 2764, ["cost"], "Metric"], [2794, 2796, ["wavelet", "transforms"], "Task"], [2803, 2805, ["signal", "processing"], "Task"], [2811, 2814, ["convolution", "or", "pooling"], "Method"], [2822, 2824, ["convolution", "layer"], "Method"], [2856, 2858, ["original", "filter"], "Method"], [2912, 2915, ["zero", "-", "based"], "Method"], [2923, 2928, ["shift", "-", "and", "-", "stitch"], "Method"], [2955, 2961, ["subsampled", "versions", "of", "the", "upsampled", "input"], "Method"], [2991, 2993, ["dilation", "trick"], "Method"], [3037, 3038, ["dilation"], "Method"], [3050, 3051, ["learning"], "Task"], [3052, 3053, ["upsampling"], "Method"], [3072, 3075, ["skip", "layer", "fusion"], "Method"], [3083, 3084, ["dilation"], "Task"], [3089, 3090, ["FCN"], "Method"], [3094, 3095, ["Upsampling"], "Method"], [3097, 3101, ["fractionally", "strided", ")", "convolution"], "Method"], [3117, 3119, ["bilinear", "interpolation"], "Method"], [3129, 3131, ["linear", "map"], "Method"], [3161, 3164, ["upsampling", "with", "factor"], "Method"], [3165, 3166, ["convolution"], "Method"], [3184, 3185, ["upsampling"], "Method"], [3187, 3189, ["backward", "convolution"], "Method"], [3200, 3204, ["input", "-", "strided", "convolution"], "Method"], [3206, 3207, ["upsampling"], "Method"], [3210, 3212, ["-", "network"], "Method"], [3213, 3219, ["end", "-", "to", "-", "end", "learning"], "Task"], [3220, 3221, ["backpropagation"], "Method"], [3230, 3232, ["deconvolution", "networks"], "Method"], [3239, 3242, ["convolution", ")", "layers"], "Method"], [3247, 3249, ["deconvolution", "layers"], "Method"], [3253, 3255, ["convolution", "filter"], "Method"], [3277, 3281, ["stack", "of", "deconvolution", "layers"], "Method"], [3282, 3284, ["activation", "functions"], "Method"], [3288, 3290, ["nonlinear", "upsampling"], "Method"], [3298, 3302, ["in", "-", "network", "upsampling"], "Method"], [3307, 3310, ["learning", "dense", "prediction"], "Task"], [3313, 3315, ["Patchwise", "training"], "Task"], [3316, 3318, ["loss", "sampling"], "Method"], [3319, 3321, ["stochastic", "optimization"], "Task"], [3322, 3324, ["gradient", "computation"], "Task"], [3332, 3334, ["patchwise", "training"], "Method"], [3335, 3338, ["fully", "convolutional", "training"], "Method"], [3352, 3354, ["computational", "efficiency"], "Metric"], [3361, 3366, ["Whole", "image", "fully", "convolutional", "training"], "Method"], [3369, 3371, ["patchwise", "training"], "Method"], [3400, 3404, ["uniform", "sampling", "of", "patches"], "Method"], [3415, 3419, ["random", "sampling", "of", "patches"], "Method"], [3469, 3472, ["fully", "convolutional", "computation"], "Method"], [3476, 3477, ["training"], "Task"], [3528, 3529, ["Sampling"], "Method"], [3530, 3532, ["patchwise", "training"], "Task"], [3534, 3536, ["class", "imbalance"], "Task"], [3546, 3549, ["fully", "convolutional", "training"], "Method"], [3550, 3552, ["class", "balance"], "Task"], [3562, 3564, ["loss", "sampling"], "Method"], [3576, 3577, ["sampling"], "Method"], [3593, 3594, ["convergence"], "Metric"], [3595, 3597, ["dense", "prediction"], "Task"], [3598, 3601, ["Whole", "image", "training"], "Task"], [3608, 3610, ["Segmentation", "Architecture"], "Method"], [3612, 3614, ["ILSVRC", "classifiers"], "Method"], [3615, 3616, ["FCNs"], "Method"], [3620, 3622, ["dense", "prediction"], "Task"], [3623, 3627, ["in", "-", "network", "upsampling"], "Task"], [3635, 3636, ["segmentation"], "Task"], [3637, 3640, ["fine", "-", "tuning"], "Task"], [3660, 3662, ["skip", "architecture"], "Method"], [3700, 3705, ["per", "-", "pixel", "softmax", "loss"], "Metric"], [3712, 3717, ["mean", "pixel", "intersection", "over", "union"], "Metric"], [3751, 3752, ["classifier"], "Method"], [3760, 3762, ["classification", "architectures"], "Method"], [3772, 3774, ["AlexNet", "architecture"], "Method"], [3776, 3777, ["ILSVRC12"], "Method"], [3782, 3784, ["VGG", "nets"], "Method"], [3786, 3787, ["GoogLeNet"], "Method"], [3792, 3793, ["ILSVRC14"], "Task"], [3797, 3802, ["VGG", "16", "-", "layer", "net"], "Method"], [3820, 3821, ["GoogLeNet"], "Method"], [3837, 3840, ["average", "pooling", "layer"], "Method"], [3849, 3851, ["classifier", "layer"], "Method"], [3855, 3858, ["fully", "connected", "layers"], "Method"], [3859, 3860, ["convolutions"], "Method"], [3864, 3868, ["convolution", "with", "channel", "dimension"], "Method"], [3875, 3877, ["PASCAL", "classes"], "Material"], [3893, 3897, ["backward", ")", "convolution", "layer"], "Method"], [3943, 3945, ["learning", "rate"], "Metric"], [3961, 3963, ["classification", "networks"], "Method"], [3967, 3968, ["SGD"], "Method"], [3969, 3970, ["momentum"], "Method"], [3981, 3983, ["learning", "rates"], "Metric"], [4005, 4007, ["line", "search"], "Method"], [4019, 4021, ["learning", "rate"], "Metric"], [4029, 4032, ["class", "scoring", "layer"], "Method"], [4034, 4036, ["random", "initialization"], "Method"], [4042, 4043, ["convergence"], "Metric"], [4044, 4045, ["Dropout"], "Method"], [4052, 4054, ["classifier", "nets"], "Method"], [4067, 4070, ["Fine", "-", "tuning"], "Task"], [4071, 4072, ["classification"], "Task"], [4073, 4074, ["segmentation"], "Task"], [4105, 4107, ["mean", "IU"], "Metric"], [4108, 4109, ["val"], "Metric"], [4117, 4118, ["VGG"], "Method"], [4119, 4120, ["GoogLeNet"], "Method"], [4124, 4125, ["classifiers"], "Method"], [4149, 4155, ["Image", "-", "to", "-", "image", "learning"], "Task"], [4155, 4159, ["batch", "size", "pixel", "acc"], "Metric"], [4160, 4162, ["mean", "acc"], "Metric"], [4167, 4168, ["IU"], "Metric"], [4169, 4176, ["image", "-", "to", "-", "image", "learning", "setting"], "Task"], [4186, 4187, ["optimization"], "Task"], [4193, 4194, ["FCNs"], "Method"], [4229, 4231, ["learning", "rate"], "Metric"], [4246, 4248, ["batch", "size"], "Task"], [4260, 4261, ["Accumulation"], "Method"], [4288, 4289, ["convergence"], "Metric"], [4290, 4291, ["Learning"], "Task"], [4298, 4300, ["classification", "training"], "Task"], [4331, 4332, ["batching"], "Method"], [4339, 4343, ["image", "-", "wise", "learning"], "Task"], [4355, 4357, ["online", "learning"], "Task"], [4361, 4363, ["online", "learning"], "Method"], [4365, 4366, ["accuracy"], "Metric"], [4368, 4369, ["convergence"], "Metric"], [4373, 4374, ["iterations"], "Metric"], [4375, 4378, ["wall", "clock", "time"], "Metric"], [4400, 4401, ["batching"], "Task"], [4411, 4412, ["accumulation"], "Task"], [4413, 4414, ["online"], "Task"], [4447, 4448, ["segmentation"], "Task"], [4471, 4474, ["fully", "convolutionalized", "classifiers"], "Method"], [4478, 4480, ["semantic", "segmentation"], "Task"], [4483, 4484, ["localize"], "Task"], [4538, 4540, ["network", "prediction"], "Method"], [4571, 4572, ["prediction"], "Task"], [4616, 4617, ["skips"], "Method"], [4665, 4667, ["nonlinear", "counterpart"], "Method"], [4669, 4677, ["multi", "-", "scale", "representation", "of", "the", "Laplacian", "pyramid"], "Method"], [4698, 4700, ["Layer", "fusion"], "Method"], [4703, 4705, ["elementwise", "operation"], "Method"], [4766, 4767, ["Cropping"], "Method"], [4831, 4833, ["network", "definition"], "Method"], [4841, 4842, ["Caffe"], "Task"], [4854, 4856, ["fusion", "operation"], "Method"], [4867, 4868, ["classification"], "Task"], [4871, 4873, ["score", "layer"], "Method"], [4877, 4878, ["convolution"], "Method"], [4893, 4894, ["classification"], "Task"], [4904, 4905, ["skips"], "Method"], [4916, 4917, ["convolution"], "Method"], [4935, 4937, ["max", "fusion"], "Task"], [4940, 4941, ["learning"], "Task"], [4946, 4948, ["gradient", "switching"], "Method"], [4985, 4986, ["prediction"], "Task"], [4994, 4996, ["Skip", "Architectures"], "Method"], [4997, 4998, ["Segmentation"], "Task"], [5001, 5003, ["skip", "architecture"], "Method"], [5012, 5014, ["stream", "net"], "Method"], [5042, 5044, ["interpolation", "layer"], "Method"], [5046, 5047, ["skip"], "Method"], [5050, 5052, ["bilinear", "interpolation"], "Task"], [5085, 5088, ["FCN", "-", "8s"], "Method"], [5115, 5117, ["mean", "IU"], "Metric"], [5129, 5131, ["staged", "training"], "Method"], [5132, 5138, ["all", "-", "at", "-", "once", "training"], "Task"], [5166, 5167, ["learning"], "Task"], [5181, 5182, ["learning"], "Task"], [5206, 5208, ["learning", "rate"], "Metric"], [5225, 5228, ["FCN", "-", "8s"], "Method"], [5255, 5256, ["training"], "Method"], [5268, 5270, ["naïve", "training"], "Method"], [5295, 5299, ["staged", "learning", "rate", "adjustments"], "Method"], [5315, 5317, ["normalization", "schemes"], "Method"], [5332, 5334, ["mean", "IU"], "Metric"], [5336, 5339, ["FCN", "-", "8s"], "Method"], [5393, 5395, ["learning", "rate"], "Metric"], [5408, 5412, ["visual", "quality", "of", "output"], "Metric"], [5437, 5439, ["pixel", "acc"], "Metric"], [5440, 5442, ["mean", "acc"], "Metric"], [5447, 5448, ["IU"], "Metric"], [5452, 5455, ["Fine", "-", "tuning"], "Task"], [5462, 5463, ["backpropagation"], "Method"], [5472, 5474, ["output", "classifier"], "Method"], [5482, 5485, ["fine", "-", "tuning"], "Metric"], [5494, 5499, ["Fine", "-", "tuning", "in", "stages"], "Task"], [5523, 5524, ["accuracy"], "Metric"], [5531, 5532, ["accuracy"], "Metric"], [5563, 5566, ["PASCAL", "training", "images"], "Material"], [5577, 5578, ["SDS"], "Method"], [5593, 5595, ["mean", "IU"], "Metric"], [5606, 5608, ["mean", "IU"], "Metric"], [5611, 5618, ["per", "-", "pixel", ",", "unnormalized", "softmax", "loss"], "Task"], [5623, 5625, ["segmenting", "images"], "Task"], [5641, 5643, ["softmax", "operation"], "Method"], [5673, 5678, ["sigmoid", "cross", "-", "entropy", "loss"], "Method"], [5691, 5693, ["class", "prediction"], "Method"], [5695, 5697, ["Patch", "sampling"], "Method"], [5707, 5709, ["image", "training"], "Task"], [5745, 5746, ["convergence"], "Metric"], [5804, 5805, ["convolution"], "Method"], [5809, 5811, ["rejection", "sampling"], "Method"], [5815, 5817, ["patchwise", "training"], "Method"], [5850, 5851, ["sampling"], "Method"], [5852, 5853, ["convergence"], "Task"], [5857, 5858, ["sampling"], "Method"], [5865, 5867, ["convergence", "rate"], "Metric"], [5869, 5872, ["whole", "image", "training"], "Task"], [5896, 5901, ["unsampled", ",", "whole", "image", "training"], "Task"], [5935, 5937, ["class", "balancing"], "Method"], [5939, 5941, ["Dense", "Prediction"], "Task"], [5950, 5953, ["backward", "convolution", "layers"], "Method"], [5965, 5967, ["bilinear", "interpolation"], "Method"], [5969, 5972, ["intermediate", "upsampling", "layers"], "Method"], [5996, 5997, ["Augmentation"], "Task"], [6044, 6045, ["Caffe"], "Method"], [6066, 6067, ["FCN"], "Method"], [6068, 6070, ["semantic", "segmentation"], "Task"], [6071, 6073, ["scene", "parsing"], "Task"], [6075, 6077, ["PASCAL", "VOC"], "Material"], [6084, 6087, ["PASCAL", "-", "Context"], "Material"], [6104, 6106, ["pixel", "prediction"], "Task"], [6124, 6128, ["multi", "-", "modal", "input"], "Method"], [6131, 6135, ["multi", "-", "task", "prediction"], "Task"], [6139, 6144, ["geometric", "labels", "of", "SIFT", "Flow"], "Task"], [6150, 6152, ["network", "architecture"], "Method"], [6169, 6171, ["semantic", "segmentation"], "Task"], [6172, 6175, ["scene", "parsing", "evaluations"], "Task"], [6179, 6181, ["pixel", "accuracy"], "Metric"], [6182, 6186, ["region", "intersection", "over", "union"], "Metric"], [6187, 6188, ["IU"], "Metric"], [6190, 6192, ["pixel", "accuracy"], "Metric"], [6193, 6195, ["mean", "accuraccy"], "Metric"], [6196, 6198, ["mean", "IU"], "Metric"], [6243, 6246, ["FCN", "-", "8s"], "Method"], [6251, 6253, ["PASCAL", "VOC"], "Material"], [6255, 6256, ["2012"], "Material"], [6265, 6266, ["SDS"], "Method"], [6272, 6275, ["R", "-", "CNN"], "Method"], [6282, 6284, ["mean", "IU"], "Metric"], [6289, 6291, ["Inference", "time"], "Metric"], [6294, 6295, ["convnet"], "Method"], [6382, 6385, ["unmodified", "coarse", "model"], "Method"], [6415, 6417, ["early", "fusion"], "Task"], [6457, 6464, ["three", "-", "dimensional", "HHA", "encoding", "of", "depth"], "Method"], [6484, 6486, ["late", "fusion"], "Method"], [6488, 6489, ["RGB"], "Method"], [6490, 6491, ["HHA"], "Method"], [6506, 6508, ["stream", "net"], "Method"], [6515, 6521, ["late", "fusion", "RGB", "-", "HHA", "net"], "Method"], [6526, 6528, ["SIFT", "Flow"], "Method"], [6579, 6580, ["FCN"], "Method"], [6584, 6586, ["joint", "representation"], "Method"], [6610, 6615, ["semantic", "and", "geometric", "prediction", "layers"], "Method"], [6616, 6617, ["losses"], "Method"], [6633, 6634, ["learning"], "Task"], [6635, 6636, ["inference"], "Task"], [6677, 6679, ["pixel", "acc"], "Metric"], [6680, 6682, ["mean", "acc"], "Metric"], [6691, 6693, ["mean", "acc"], "Metric"], [6701, 6702, ["acc"], "Metric"], [6703, 6705, ["pixel", "acc"], "Metric"], [6706, 6708, ["mean", "acc"], "Metric"], [6737, 6739, ["class", "task"], "Task"], [6776, 6779, ["FCN", "-", "8s"], "Method"], [6781, 6783, ["mean", "IU"], "Metric"], [6834, 6838, ["null", "”", "background", "model"], "Method"], [6844, 6846, ["background", "classifier"], "Method"], [6847, 6849, ["semantic", "segmentation"], "Task"], [6862, 6865, ["whole", "image", "learning"], "Task"], [6872, 6874, ["task", "accuracy"], "Metric"], [6898, 6899, ["FCN"], "Method"], [6916, 6917, ["prediction"], "Task"], [6923, 6924, ["inference"], "Task"], [6950, 6951, ["Masking"], "Task"], [6958, 6960, ["masked", "versions"], "Method"], [6976, 6978, ["normal", "PASCAL"], "Material"], [6985, 6987, ["masked", "PASCAL"], "Material"], [7018, 7021, ["recognizing", "object", "segments"], "Task"], [7055, 7056, ["learning"], "Task"], [7057, 7058, ["inference"], "Task"], [7065, 7067, ["background", "accuracy"], "Metric"], [7083, 7084, ["FCNs"], "Method"], [7119, 7120, ["accuracy"], "Metric"], [7147, 7148, ["recognition"], "Task"], [7162, 7164, ["Background", "modeling"], "Method"], [7168, 7169, ["detection"], "Task"], [7170, 7172, ["semantic", "segmentation"], "Task"], [7175, 7177, ["background", "model"], "Method"], [7227, 7229, ["class", "models"], "Method"], [7237, 7238, ["net"], "Method"], [7241, 7245, ["null", "”", "background", "model"], "Method"], [7273, 7278, ["sigmoid", "cross", "-", "entropy", "loss"], "Method"], [7310, 7312, ["PASCAL", "VOC"], "Material"], [7358, 7360, ["background", "model"], "Method"], [7384, 7386, ["background", "model"], "Method"], [7387, 7389, ["semantic", "segmentation"], "Task"], [7402, 7404, ["batch", "size"], "Metric"], [7406, 7408, ["optimization", "schemes"], "Method"], [7409, 7410, ["FCNs"], "Method"], [7417, 7419, ["online", "learning"], "Method"], [7451, 7453, ["heavy", "learning"], "Method"], [7460, 7462, ["gradient", "accumulation"], "Method"], [7464, 7468, ["non", "-", "recursive", "sum"], "Method"], [7495, 7497, ["minibatch", "SGD"], "Method"], [7523, 7525, ["learning", "rate"], "Metric"], [7592, 7593, ["momentum"], "Method"], [7637, 7638, ["momentum"], "Method"], [7649, 7651, ["training", "regime"], "Method"], [7652, 7653, ["momentum"], "Method"], [7662, 7664, ["online", "learning"], "Method"], [7671, 7673, ["online", "learning"], "Method"], [7682, 7685, ["wall", "clock", "time"], "Metric"], [7688, 7690, ["Upper", "bounds"], "Metric"], [7691, 7692, ["IU"], "Metric"], [7692, 7693, ["FCNs"], "Method"], [7704, 7708, ["spatially", "coarse", "semantic", "prediction"], "Task"], [7733, 7734, ["prediction"], "Task"], [7767, 7769, ["mean", "IU"], "Metric"], [7781, 7785, ["Pixel", "-", "perfect", "prediction"], "Task"], [7791, 7793, ["mean", "IU"], "Metric"], [7807, 7809, ["mean", "IU"], "Metric"], [7816, 7820, ["fine", "-", "scale", "accuracy"], "Metric"], [7833, 7834, ["accuracy"], "Metric"], [7839, 7840, ["recognition"], "Task"], [7842, 7843, ["resolution"], "Task"], [7853, 7856, ["Fully", "convolutional", "networks"], "Method"], [7865, 7867, ["pixelwise", "tasks"], "Task"], [7868, 7869, ["FCNs"], "Method"], [7870, 7872, ["semantic", "segmentation"], "Task"], [7874, 7875, ["accuracy"], "Metric"], [7885, 7887, ["layer", "representations"], "Method"], [7916, 7919, ["learning", "and", "inference"], "Task"], [7957, 7963, ["image", "-", "to", "-", "image", "tasks"], "Task"], [58, 60, ["fully", "convolutional"], "Method"], [87, 89, ["fully", "convolutional"], "Method"], [123, 125, ["fully", "convolutional"], "Method"], [173, 175, ["fully", "convolutional"], "Method"], [249, 251, ["image", "classification"], "Method"], [270, 271, ["detection"], "Task"], [338, 340, ["fully", "convolutional"], "Method"], [792, 795, ["PASCAL", "-", "Context"], "Material"], [840, 841, ["FCN"], "Method"], [868, 870, ["PASCAL", "VOC"], "Material"], [974, 976, ["semantic", "segmentation"], "Task"], [1065, 1066, ["detection"], "Task"], [1129, 1130, ["detection"], "Task"], [1532, 1533, ["FCN"], "Method"], [1663, 1664, ["FCN"], "Method"], [1689, 1690, ["detection"], "Task"], [1701, 1703, ["semantic", "segmentation"], "Task"], [2063, 2065, ["fully", "convolutional"], "Method"], [2216, 2218, ["fully", "convolutional"], "Method"], [2359, 2362, ["fully", "connected", "layers"], "Method"], [2400, 2402, ["fully", "convolutional"], "Method"], [2481, 2483, ["fully", "convolutional"], "Method"], [3690, 3692, ["PASCAL", "VOC"], "Material"], [3754, 3755, ["FCN"], "Method"], [3988, 3989, ["FCN"], "Method"], [3992, 3993, ["FCN"], "Method"], [3997, 3998, ["FCN"], "Method"], [4092, 4093, ["FCN"], "Method"], [4127, 4128, ["FCN"], "Method"], [4133, 4134, ["FCN"], "Method"], [4139, 4140, ["FCN"], "Method"], [4163, 4165, ["mean", "IU"], "Metric"], [4443, 4445, ["fully", "convolutional"], "Method"], [5005, 5006, ["FCN"], "Method"], [5078, 5079, ["FCN"], "Method"], [5150, 5151, ["FCN"], "Method"], [5161, 5162, ["FCN"], "Method"], [5176, 5177, ["FCN"], "Method"], [5211, 5212, ["FCN"], "Method"], [5215, 5216, ["FCN"], "Method"], [5221, 5222, ["FCN"], "Method"], [5324, 5325, ["FCN"], "Method"], [5443, 5445, ["mean", "IU"], "Metric"], [5508, 5509, ["FCN"], "Method"], [5537, 5539, ["PASCAL", "VOC"], "Material"], [5584, 5585, ["FCN"], "Method"], [5598, 5599, ["FCN"], "Method"], [5908, 5911, ["Fully", "convolutional", "training"], "Method"], [6110, 6111, ["FCN"], "Method"], [6201, 6202, ["IU"], "Metric"], [6232, 6234, ["PASCAL", "VOC"], "Material"], [6341, 6343, ["semantic", "segmentation"], "Task"], [6386, 6387, ["FCN"], "Method"], [6602, 6603, ["FCN"], "Method"], [6683, 6685, ["mean", "IU"], "Metric"], [6687, 6688, ["IU"], "Metric"], [6694, 6696, ["mean", "IU"], "Metric"], [6698, 6699, ["IU"], "Metric"], [6709, 6711, ["mean", "IU"], "Metric"], [6713, 6714, ["IU"], "Metric"], [6714, 6717, ["PASCAL", "-", "Context"], "Material"], [6722, 6724, ["PASCAL", "VOC"], "Material"], [6756, 6757, ["val"], "Metric"], [6803, 6805, ["fully", "convolutional"], "Method"], [6963, 6965, ["PASCAL", "VOC"], "Material"], [7306, 7307, ["FCN"], "Method"], [7324, 7325, ["FCN"], "Method"], [7330, 7331, ["FCN"], "Method"], [7634, 7635, ["FCN"], "Method"], [7678, 7679, ["FCN"], "Method"], [7698, 7700, ["mean", "IU"], "Metric"], [7775, 7776, ["val"], "Metric"], [7950, 7952, ["fully", "convolutional"], "Method"], [8064, 8066, ["mean", "IU"], "Metric"], [8074, 8076, ["mean", "IU"], "Metric"]]} {"doc_id": "088b649415378c4dde666f614ba24d08a36b6652", "text": ["document", ":", "Long", "Short", "-", "Term", "Memory", "-", "Networks", "for", "Machine", "Reading", "In", "this", "paper", "we", "address", "the", "question", "of", "how", "to", "render", "sequence", "-", "level", "networks", "better", "at", "handling", "structured", "input", ".", "We", "propose", "a", "machine", "reading", "simulator", "which", "processes", "text", "incrementally", "from", "left", "to", "right", "and", "performs", "shallow", "reasoning", "with", "memory", "and", "attention", ".", "The", "reader", "extends", "the", "Long", "Short", "-", "Term", "Memory", "architecture", "with", "a", "memory", "network", "in", "place", "of", "a", "single", "memory", "cell", ".", "This", "enables", "adaptive", "memory", "usage", "during", "recurrence", "with", "neural", "attention", ",", "offering", "a", "way", "to", "weakly", "induce", "relations", "among", "tokens", ".", "The", "system", "is", "initially", "designed", "to", "process", "a", "single", "sequence", "but", "we", "also", "demonstrate", "how", "to", "integrate", "it", "with", "an", "encoder", "-", "decoder", "architecture", ".", "Experiments", "on", "language", "modeling", ",", "sentiment", "analysis", ",", "and", "natural", "language", "inference", "show", "that", "our", "model", "matches", "or", "outperforms", "the", "state", "of", "the", "art", ".", "section", ":", "Introduction", "How", "can", "a", "sequence", "-", "level", "network", "induce", "relations", "which", "are", "presumed", "latent", "during", "text", "processing", "?", "How", "can", "a", "recurrent", "network", "attentively", "memorize", "longer", "sequences", "in", "a", "way", "that", "humans", "do", "?", "In", "this", "paper", "we", "design", "a", "machine", "reader", "that", "automatically", "learns", "to", "understand", "text", ".", "The", "term", "machine", "reading", "is", "related", "to", "a", "wide", "range", "of", "tasks", "from", "answering", "reading", "comprehension", "questions", ",", "to", "fact", "and", "relation", "extraction", ",", "ontology", "learning", ",", "and", "textual", "entailment", ".", "Rather", "than", "focusing", "on", "a", "specific", "task", ",", "we", "develop", "a", "general", "-", "purpose", "reading", "simulator", ",", "drawing", "inspiration", "from", "human", "language", "processing", "and", "the", "fact", "language", "comprehension", "is", "incremental", "with", "readers", "continuously", "extracting", "the", "meaning", "of", "utterances", "on", "a", "word", "-", "by", "-", "word", "basis", ".", "In", "order", "to", "understand", "texts", ",", "our", "machine", "reader", "should", "provide", "facilities", "for", "extracting", "and", "representing", "meaning", "from", "natural", "language", "text", ",", "storing", "meanings", "internally", ",", "and", "working", "with", "stored", "meanings", "to", "derive", "further", "consequences", ".", "Ideally", ",", "such", "a", "system", "should", "be", "robust", ",", "open", "-", "domain", ",", "and", "degrade", "gracefully", "in", "the", "presence", "of", "semantic", "representations", "which", "may", "be", "incomplete", ",", "inaccurate", ",", "or", "incomprehensible", ".", "It", "would", "also", "be", "desirable", "to", "simulate", "the", "behavior", "of", "English", "speakers", "who", "process", "text", "sequentially", ",", "from", "left", "to", "right", ",", "fixating", "nearly", "every", "word", "while", "they", "read", "and", "creating", "partial", "representations", "for", "sentence", "prefixes", ".", "Language", "modeling", "tools", "such", "as", "recurrent", "neural", "networks", "(", "RNN", ")", "bode", "well", "with", "human", "reading", "behavior", ".", "RNNs", "treat", "each", "sentence", "as", "a", "sequence", "of", "words", "and", "recursively", "compose", "each", "word", "with", "its", "previous", "memory", ",", "until", "the", "meaning", "of", "the", "whole", "sentence", "has", "been", "derived", ".", "In", "practice", ",", "however", ",", "sequence", "-", "level", "networks", "are", "met", "with", "at", "least", "three", "challenges", ".", "The", "first", "one", "concerns", "model", "training", "problems", "associated", "with", "vanishing", "and", "exploding", "gradients", ",", "which", "can", "be", "partially", "ameliorated", "with", "gated", "activation", "functions", ",", "such", "as", "the", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", ",", "and", "gradient", "clipping", ".", "The", "second", "issue", "relates", "to", "memory", "compression", "problems", ".", "As", "the", "input", "sequence", "gets", "compressed", "and", "blended", "into", "a", "single", "dense", "vector", ",", "sufficiently", "large", "memory", "capacity", "is", "required", "to", "store", "past", "information", ".", "As", "a", "result", ",", "the", "network", "generalizes", "poorly", "to", "long", "sequences", "while", "wasting", "memory", "on", "shorter", "ones", ".", "Finally", ",", "it", "should", "be", "acknowledged", "that", "sequence", "-", "level", "networks", "lack", "a", "mechanism", "for", "handling", "the", "structure", "of", "the", "input", ".", "This", "imposes", "an", "inductive", "bias", "which", "is", "at", "odds", "with", "the", "fact", "that", "language", "has", "inherent", "structure", ".", "In", "this", "paper", ",", "we", "develop", "a", "text", "processing", "system", "which", "addresses", "these", "limitations", "while", "maintaining", "the", "incremental", ",", "generative", "property", "of", "a", "recurrent", "language", "model", ".", "The", "The", "The", "The", "FBI", "FBI", "The", "The", "FBI", "FBI", "is", "is", "The", "The", "FBI", "FBI", "is", "is", "chasing", "chasing", "The", "The", "FBI", "FBI", "is", "is", "chasing", "chasing", "a", "a", "The", "The", "FBI", "FBI", "is", "is", "chasing", "chasing", "a", "a", "criminal", "criminal", "The", "The", "FBI", "FBI", "is", "is", "chasing", "chasing", "a", "a", "criminal", "criminal", "on", "on", "The", "The", "FBI", "FBI", "is", "is", "chasing", "chasing", "a", "a", "criminal", "criminal", "on", "on", "the", "the", "The", "The", "FBI", "FBI", "is", "is", "chasing", "chasing", "a", "a", "criminal", "criminal", "on", "on", "the", "the", "run", "run", "Recent", "attempts", "to", "render", "neural", "networks", "more", "structure", "aware", "have", "seen", "the", "incorporation", "of", "external", "memories", "in", "the", "context", "of", "recurrent", "neural", "networks", ".", "The", "idea", "is", "to", "use", "multiple", "memory", "slots", "outside", "the", "recurrence", "to", "piece", "-", "wise", "store", "representations", "of", "the", "input", ";", "read", "and", "write", "operations", "for", "each", "slot", "can", "be", "modeled", "as", "an", "attention", "mechanism", "with", "a", "recurrent", "controller", ".", "We", "also", "leverage", "memory", "and", "attention", "to", "empower", "a", "recurrent", "network", "with", "stronger", "memorization", "capability", "and", "more", "importantly", "the", "ability", "to", "discover", "relations", "among", "tokens", ".", "This", "is", "realized", "by", "inserting", "a", "memory", "network", "module", "in", "the", "update", "of", "a", "recurrent", "network", "together", "with", "attention", "for", "memory", "addressing", ".", "The", "attention", "acts", "as", "a", "weak", "inductive", "module", "discovering", "relations", "between", "input", "tokens", ",", "and", "is", "trained", "without", "direct", "supervision", ".", "As", "a", "point", "of", "departure", "from", "previous", "work", ",", "the", "memory", "network", "we", "employ", "is", "internal", "to", "the", "recurrence", ",", "thus", "strengthening", "the", "interaction", "of", "the", "two", "and", "leading", "to", "a", "representation", "learner", "which", "is", "able", "to", "reason", "over", "shallow", "structures", ".", "The", "resulting", "model", ",", "which", "we", "term", "Long", "Short", "-", "Term", "Memory", "-", "Network", "(", "LSTMN", ")", ",", "is", "a", "reading", "simulator", "that", "can", "be", "used", "for", "sequence", "processing", "tasks", ".", "Figure", "[", "reference", "]", "illustrates", "the", "reading", "behavior", "of", "the", "LSTMN", ".", "The", "model", "processes", "text", "incrementally", "while", "learning", "which", "past", "tokens", "in", "the", "memory", "and", "to", "what", "extent", "they", "relate", "to", "the", "current", "token", "being", "processed", ".", "As", "a", "result", ",", "the", "model", "induces", "undirected", "relations", "among", "tokens", "as", "an", "intermediate", "step", "of", "learning", "representations", ".", "We", "validate", "the", "performance", "of", "the", "LSTMN", "in", "language", "modeling", ",", "sentiment", "analysis", ",", "and", "natural", "language", "inference", ".", "In", "all", "cases", ",", "we", "train", "LSTMN", "models", "end", "-", "to", "-", "end", "with", "task", "-", "specific", "supervision", "signals", ",", "achieving", "performance", "comparable", "or", "better", "to", "state", "-", "of", "-", "the", "-", "art", "models", "and", "superior", "to", "vanilla", "LSTMs", ".", "section", ":", "Related", "Work", "Our", "machine", "reader", "is", "a", "recurrent", "neural", "network", "exhibiting", "two", "important", "properties", ":", "it", "is", "incremental", ",", "simulating", "human", "behavior", ",", "and", "performs", "shallow", "structure", "reasoning", "over", "input", "streams", ".", "Recurrent", "neural", "network", "(", "RNNs", ")", "have", "been", "successfully", "applied", "to", "various", "sequence", "modeling", "and", "sequence", "-", "to", "-", "sequence", "transduction", "tasks", ".", "The", "latter", "have", "assumed", "several", "guises", "in", "the", "literature", "such", "as", "machine", "translation", ",", "sentence", "compression", ",", "and", "reading", "comprehension", ".", "A", "key", "contributing", "factor", "to", "their", "success", "has", "been", "the", "ability", "to", "handle", "well", "-", "known", "problems", "with", "exploding", "or", "vanishing", "gradients", ",", "leading", "to", "models", "with", "gated", "activation", "functions", ",", "and", "more", "advanced", "architectures", "that", "enhance", "the", "information", "flow", "within", "the", "network", ".", "A", "remaining", "practical", "bottleneck", "for", "RNNs", "is", "memory", "compression", ":", "since", "the", "inputs", "are", "recursively", "combined", "into", "a", "single", "memory", "representation", "which", "is", "typically", "too", "small", "in", "terms", "of", "parameters", ",", "it", "becomes", "difficult", "to", "accurately", "memorize", "sequences", ".", "In", "the", "encoder", "-", "decoder", "architecture", ",", "this", "problem", "can", "be", "sidestepped", "with", "an", "attention", "mechanism", "which", "learns", "soft", "alignments", "between", "the", "decoding", "states", "and", "the", "encoded", "memories", ".", "In", "our", "model", ",", "memory", "and", "attention", "are", "added", "within", "a", "sequence", "encoder", "allowing", "the", "network", "to", "uncover", "lexical", "relations", "between", "tokens", ".", "The", "idea", "of", "introducing", "a", "structural", "bias", "to", "neural", "models", "is", "by", "no", "means", "new", ".", "For", "example", ",", "it", "is", "reflected", "in", "the", "work", "of", "socher", "-", "EtAl:2013:EMNLP", "who", "apply", "recursive", "neural", "networks", "for", "learning", "natural", "language", "representations", ".", "In", "the", "context", "of", "recurrent", "neural", "networks", ",", "efforts", "to", "build", "modular", ",", "structured", "neural", "models", "date", "back", "to", "das1992learning", "who", "connect", "a", "recurrent", "neural", "network", "with", "an", "external", "memory", "stack", "for", "learning", "context", "free", "grammars", ".", "Recently", ",", "weston2014memory", "propose", "Memory", "Networks", "to", "explicitly", "segregate", "memory", "storage", "from", "the", "computation", "of", "neural", "networks", "in", "general", ".", "Their", "model", "is", "trained", "end", "-", "to", "-", "end", "with", "a", "memory", "addressing", "mechanism", "closely", "related", "to", "soft", "attention", "and", "has", "been", "applied", "to", "machine", "translation", ".", "grefenstette2015learning", "define", "a", "set", "of", "differentiable", "data", "structures", "(", "stacks", ",", "queues", ",", "and", "dequeues", ")", "as", "memories", "controlled", "by", "a", "recurrent", "neural", "network", ".", "ke2016memory", "combine", "the", "LSTM", "with", "an", "external", "memory", "block", "component", "which", "interacts", "with", "its", "hidden", "state", ".", "kumar2015ask", "employ", "a", "structured", "neural", "network", "with", "episodic", "memory", "modules", "for", "natural", "language", "and", "also", "visual", "question", "answering", ".", "Similar", "to", "the", "above", "work", ",", "we", "leverage", "memory", "and", "attention", "in", "a", "recurrent", "neural", "network", "for", "inducing", "relations", "between", "tokens", "as", "a", "module", "in", "a", "larger", "network", "responsible", "for", "representation", "learning", ".", "As", "a", "property", "of", "soft", "attention", ",", "all", "intermediate", "relations", "we", "aim", "to", "capture", "are", "soft", "and", "differentiable", ".", "This", "is", "in", "contrast", "to", "shift", "-", "reduce", "type", "neural", "models", "where", "the", "intermediate", "decisions", "are", "hard", "and", "induction", "is", "more", "difficult", ".", "Finally", ",", "note", "that", "our", "model", "captures", "undirected", "lexical", "relations", "and", "is", "thus", "distinct", "from", "work", "on", "dependency", "grammar", "induction", "where", "the", "learned", "head", "-", "modifier", "relations", "are", "directed", ".", "section", ":", "The", "Machine", "Reader", "In", "this", "section", "we", "present", "our", "machine", "reader", "which", "is", "designed", "to", "process", "structured", "input", "while", "retaining", "the", "incrementality", "of", "a", "recurrent", "neural", "network", ".", "The", "core", "of", "our", "model", "is", "a", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "unit", "with", "an", "extended", "memory", "tape", "that", "explicitly", "simulates", "the", "human", "memory", "span", ".", "The", "model", "performs", "implicit", "relation", "analysis", "between", "tokens", "with", "an", "attention", "-", "based", "memory", "addressing", "mechanism", "at", "every", "time", "step", ".", "In", "the", "following", ",", "we", "first", "review", "the", "standard", "Long", "Short", "-", "Term", "Memory", "and", "then", "describe", "our", "model", ".", "subsection", ":", "Long", "Short", "-", "Term", "Memory", "A", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "recurrent", "neural", "network", "processes", "a", "variable", "-", "length", "sequence", "by", "incrementally", "adding", "new", "content", "into", "a", "single", "memory", "slot", ",", "with", "gates", "controlling", "the", "extent", "to", "which", "new", "content", "should", "be", "memorized", ",", "old", "content", "should", "be", "erased", ",", "and", "current", "content", "should", "be", "exposed", ".", "At", "time", "step", ",", "the", "memory", "and", "the", "hidden", "state", "are", "updated", "with", "the", "following", "equations", ":", "where", ",", ",", "and", "are", "gate", "activations", ".", "Compared", "to", "the", "standard", "RNN", ",", "the", "LSTM", "uses", "additive", "memory", "updates", "and", "it", "separates", "the", "memory", "from", "the", "hidden", "state", ",", "which", "interacts", "with", "the", "environment", "when", "making", "predictions", ".", "subsection", ":", "Long", "Short", "-", "Term", "Memory", "-", "Network", "The", "first", "question", "that", "arises", "with", "LSTMs", "is", "the", "extent", "to", "which", "they", "are", "able", "to", "memorize", "sequences", "under", "recursive", "compression", ".", "LSTMs", "can", "produce", "a", "list", "of", "state", "representations", "during", "composition", ",", "however", ",", "the", "next", "state", "is", "always", "computed", "from", "the", "current", "state", ".", "That", "is", "to", "say", ",", "given", "the", "current", "state", ",", "the", "next", "state", "is", "conditionally", "independent", "of", "states", "and", "tokens", ".", "While", "the", "recursive", "state", "update", "is", "performed", "in", "a", "Markov", "manner", ",", "it", "is", "assumed", "that", "LSTMs", "maintain", "unbounded", "memory", "(", "i.e.", ",", "the", "current", "state", "alone", "summarizes", "well", "the", "tokens", "it", "has", "seen", "so", "far", ")", ".", "This", "assumption", "may", "fail", "in", "practice", ",", "for", "example", "when", "the", "sequence", "is", "long", "or", "when", "the", "memory", "size", "is", "not", "large", "enough", ".", "Another", "undesired", "property", "of", "LSTMs", "concerns", "modeling", "structured", "input", ".", "An", "LSTM", "aggregates", "information", "on", "a", "token", "-", "by", "-", "token", "basis", "in", "sequential", "order", ",", "but", "there", "is", "no", "explicit", "mechanism", "for", "reasoning", "over", "structure", "and", "modeling", "relations", "between", "tokens", ".", "Our", "model", "aims", "to", "address", "both", "limitations", ".", "Our", "solution", "is", "to", "modify", "the", "standard", "LSTM", "structure", "by", "replacing", "the", "memory", "cell", "with", "a", "memory", "network", ".", "The", "resulting", "Long", "Short", "-", "Term", "Memory", "-", "Network", "(", "LSTMN", ")", "stores", "the", "contextual", "representation", "of", "each", "input", "token", "with", "a", "unique", "memory", "slot", "and", "the", "size", "of", "the", "memory", "grows", "with", "time", "until", "an", "upper", "bound", "of", "the", "memory", "span", "is", "reached", ".", "This", "design", "enables", "the", "LSTM", "to", "reason", "about", "relations", "between", "tokens", "with", "a", "neural", "attention", "layer", "and", "then", "perform", "non", "-", "Markov", "state", "updates", ".", "Although", "it", "is", "feasible", "to", "apply", "both", "write", "and", "read", "operations", "to", "the", "memories", "with", "attention", ",", "we", "concentrate", "on", "the", "latter", ".", "We", "conceptualize", "the", "read", "operation", "as", "attentively", "linking", "the", "current", "token", "to", "previous", "memories", "and", "selecting", "useful", "content", "when", "processing", "it", ".", "Although", "not", "the", "focus", "of", "this", "work", ",", "the", "significance", "of", "the", "write", "operation", "can", "be", "analogously", "justified", "as", "a", "way", "of", "incrementally", "updating", "previous", "memories", ",", "e.g.", ",", "to", "correct", "wrong", "interpretations", "when", "processing", "garden", "path", "sentences", ".", "The", "architecture", "of", "the", "LSTMN", "is", "shown", "in", "Figure", "[", "reference", "]", "and", "the", "formal", "definition", "is", "provided", "as", "follows", ".", "The", "model", "maintains", "two", "sets", "of", "vectors", "stored", "in", "a", "hidden", "state", "tape", "used", "to", "interact", "with", "the", "environment", "(", "e.g.", ",", "computing", "attention", ")", ",", "and", "a", "memory", "tape", "used", "to", "represent", "what", "is", "actually", "stored", "in", "memory", ".", "Therefore", ",", "each", "token", "is", "associated", "with", "a", "hidden", "vector", "and", "a", "memory", "vector", ".", "Let", "denote", "the", "current", "input", ";", "denotes", "the", "current", "memory", "tape", ",", "and", "the", "previous", "hidden", "tape", ".", "At", "time", "step", ",", "the", "model", "computes", "the", "relation", "between", "and", "through", "with", "an", "attention", "layer", ":", "This", "yields", "a", "probability", "distribution", "over", "the", "hidden", "state", "vectors", "of", "previous", "tokens", ".", "We", "can", "then", "compute", "an", "adaptive", "summary", "vector", "for", "the", "previous", "hidden", "tape", "and", "memory", "tape", "denoted", "by", "and", ",", "respectively", ":", "and", "use", "them", "for", "computing", "the", "values", "of", "and", "in", "the", "recurrent", "update", "as", ":", "where", ",", ",", "and", "are", "the", "new", "weight", "terms", "of", "the", "network", ".", "A", "key", "idea", "behind", "the", "LSTMN", "is", "to", "use", "attention", "for", "inducing", "relations", "between", "tokens", ".", "These", "relations", "are", "soft", "and", "differentiable", ",", "and", "components", "of", "a", "larger", "representation", "learning", "network", ".", "Although", "it", "is", "appealing", "to", "provide", "direct", "supervision", "for", "the", "attention", "layer", ",", "e.g.", ",", "with", "evidence", "collected", "from", "a", "dependency", "treebank", ",", "we", "treat", "it", "as", "a", "submodule", "being", "optimized", "within", "the", "larger", "network", "in", "a", "downstream", "task", ".", "It", "is", "also", "possible", "to", "have", "a", "more", "structured", "relational", "reasoning", "module", "by", "stacking", "multiple", "memory", "and", "hidden", "layers", "in", "an", "alternating", "fashion", ",", "resembling", "a", "stacked", "LSTM", "or", "a", "multi", "-", "hop", "memory", "network", ".", "This", "can", "be", "achieved", "by", "feeding", "the", "output", "of", "the", "lower", "layer", "as", "input", "to", "the", "upper", "layer", ".", "The", "attention", "at", "the", "th", "layer", "is", "computed", "as", ":", "Skip", "-", "connections", "can", "be", "applied", "to", "feed", "to", "upper", "layers", "as", "well", ".", "[", "Decoder", "with", "shallow", "attention", "fusion", ".", "]", "[", "Decoder", "with", "deep", "attention", "fusion", ".", "]", "section", ":", "Modeling", "Two", "Sequences", "with", "LSTMN", "Natural", "language", "processing", "tasks", "such", "as", "machine", "translation", "and", "textual", "entailment", "are", "concerned", "with", "modeling", "two", "sequences", "rather", "than", "a", "single", "one", ".", "A", "standard", "tool", "for", "modeling", "two", "sequences", "with", "recurrent", "networks", "is", "the", "encoder", "-", "decoder", "architecture", "where", "the", "second", "sequence", "(", "also", "known", "as", "the", "target", ")", "is", "being", "processed", "conditioned", "on", "the", "first", "one", "(", "also", "known", "as", "the", "source", ")", ".", "In", "this", "section", "we", "explain", "how", "to", "combine", "the", "LSTMN", "which", "applies", "attention", "for", "intra", "-", "relation", "reasoning", ",", "with", "the", "encoder", "-", "decoder", "network", "whose", "attention", "module", "learns", "the", "inter", "-", "alignment", "between", "two", "sequences", ".", "Figures", "[", "reference", "]", "a", "and", "[", "reference", "]", "b", "illustrate", "two", "types", "of", "combination", ".", "We", "describe", "the", "models", "more", "formally", "below", ".", "paragraph", ":", "Shallow", "Attention", "Fusion", "Shallow", "fusion", "simply", "treats", "the", "LSTMN", "as", "a", "separate", "module", "that", "can", "be", "readily", "used", "in", "an", "encoder", "-", "decoder", "architecture", ",", "in", "lieu", "of", "a", "standard", "RNN", "or", "LSTM", ".", "As", "shown", "in", "Figure", "[", "reference", "]", "a", ",", "both", "encoder", "and", "decoder", "are", "modeled", "as", "LSTMNs", "with", "intra", "-", "attention", ".", "Meanwhile", ",", "inter", "-", "attention", "is", "triggered", "when", "the", "decoder", "reads", "a", "target", "token", ",", "similar", "to", "the", "inter", "-", "attention", "introduced", "in", "bahdanau2014neural", ".", "paragraph", ":", "Deep", "Attention", "Fusion", "Deep", "fusion", "combines", "inter", "-", "and", "intra", "-", "attention", "(", "initiated", "by", "the", "decoder", ")", "when", "computing", "state", "updates", ".", "We", "use", "different", "notation", "to", "represent", "the", "two", "sets", "of", "attention", ".", "Following", "Section", "[", "reference", "]", ",", "and", "denote", "the", "target", "memory", "tape", "and", "hidden", "tape", ",", "which", "store", "representations", "of", "the", "target", "symbols", "that", "have", "been", "processed", "so", "far", ".", "The", "computation", "of", "intra", "-", "attention", "follows", "Equations", "(", "[", "reference", "]", ")", "–", "(", "[", "reference", "]", ")", ".", "Additionally", ",", "we", "use", "and", "to", "represent", "the", "source", "memory", "tape", "and", "hidden", "tape", ",", "with", "being", "the", "length", "of", "the", "source", "sequence", "conditioned", "upon", ".", "We", "compute", "inter", "-", "attention", "between", "the", "input", "at", "time", "step", "and", "tokens", "in", "the", "entire", "source", "sequence", "as", "follows", ":", "After", "that", "we", "compute", "the", "adaptive", "representation", "of", "the", "source", "memory", "tape", "and", "hidden", "tape", "as", ":", "We", "can", "then", "transfer", "the", "adaptive", "source", "representation", "to", "the", "target", "memory", "with", "another", "gating", "operation", ",", "analogous", "to", "the", "gates", "in", "Equation", "(", "[", "reference", "]", ")", ".", "The", "new", "target", "memory", "includes", "inter", "-", "alignment", ",", "intra", "-", "relation", ",", "and", "the", "new", "input", "information", ":", "As", "shown", "in", "the", "equations", "above", "and", "Figure", "[", "reference", "]", "b", ",", "the", "major", "change", "of", "deep", "fusion", "lies", "in", "the", "recurrent", "storage", "of", "the", "inter", "-", "alignment", "vector", "in", "the", "target", "memory", "network", ",", "as", "a", "way", "to", "help", "the", "target", "network", "review", "source", "information", ".", "section", ":", "Experiments", "In", "this", "section", "we", "present", "our", "experiments", "for", "evaluating", "the", "performance", "of", "the", "LSTMN", "machine", "reader", ".", "We", "start", "with", "language", "modeling", "as", "it", "is", "a", "natural", "testbed", "for", "our", "model", ".", "We", "then", "assess", "the", "model", "’s", "ability", "to", "extract", "meaning", "representations", "for", "generic", "sentence", "classification", "tasks", "such", "as", "sentiment", "analysis", ".", "Finally", ",", "we", "examine", "whether", "the", "LSTMN", "can", "recognize", "the", "semantic", "relationship", "between", "two", "sentences", "by", "applying", "it", "to", "a", "natural", "language", "inference", "task", ".", "Our", "code", "is", "available", "at", ".", "subsection", ":", "Language", "Modeling", "Our", "language", "modeling", "experiments", "were", "conducted", "on", "the", "English", "Penn", "Treebank", "dataset", ".", "Following", "common", "practice", ",", "we", "trained", "on", "sections", "0–20", "(", "1", "M", "words", ")", ",", "used", "sections", "21–22", "for", "validation", "(", "80", "K", "words", ")", ",", "and", "sections", "23–24", "(", "90", "K", "words", "for", "testing", ")", ".", "The", "dataset", "contains", "approximately", "1", "million", "tokens", "and", "a", "vocabulary", "size", "of", "10K.", "The", "average", "sentence", "length", "is", "21", ".", "We", "use", "perplexity", "as", "our", "evaluation", "metric", ":", ",", "where", "denotes", "the", "negative", "log", "likelihood", "of", "the", "entire", "test", "set", "and", "the", "corresponding", "number", "of", "tokens", ".", "We", "used", "stochastic", "gradient", "descent", "for", "optimization", "with", "an", "initial", "learning", "rate", "of", "0.65", ",", "which", "decays", "by", "a", "factor", "of", "0.85", "per", "epoch", "if", "no", "significant", "improvement", "has", "been", "observed", "on", "the", "validation", "set", ".", "We", "renormalize", "the", "gradient", "if", "its", "norm", "is", "greater", "than", "5", ".", "The", "mini", "-", "batch", "size", "was", "set", "to", "40", ".", "The", "dimensions", "of", "the", "word", "embeddings", "were", "set", "to", "150", "for", "all", "models", ".", "In", "this", "suite", "of", "experiments", "we", "compared", "the", "LSTMN", "against", "a", "variety", "of", "baselines", ".", "The", "first", "one", "is", "a", "Kneser", "-", "Ney", "5", "-", "gram", "language", "model", "(", "KN5", ")", "which", "generally", "serves", "as", "a", "non", "-", "neural", "baseline", "for", "the", "language", "modeling", "task", ".", "We", "also", "present", "perplexity", "results", "for", "the", "standard", "RNN", "and", "LSTM", "models", ".", "We", "also", "implemented", "more", "sophisticated", "LSTM", "architectures", ",", "such", "as", "a", "stacked", "LSTM", "(", "sLSTM", ")", ",", "a", "gated", "-", "feedback", "LSTM", "(", "gLSTM", ";", "chung2015gated", ")", "and", "a", "depth", "-", "gated", "LSTM", "(", "dLSTM", ";", "yao2015depth", ")", ".", "The", "gated", "-", "feedback", "LSTM", "has", "feedback", "gates", "connecting", "the", "hidden", "states", "across", "multiple", "time", "steps", "as", "an", "adaptive", "control", "of", "the", "information", "flow", ".", "The", "depth", "-", "gated", "LSTM", "uses", "a", "depth", "gate", "to", "connect", "memory", "cells", "of", "vertically", "adjacent", "layers", ".", "In", "general", ",", "both", "gLSTM", "and", "dLSTM", "are", "able", "to", "capture", "long", "-", "term", "dependencies", "to", "some", "degree", ",", "but", "they", "do", "not", "explicitly", "keep", "past", "memories", ".", "We", "set", "the", "number", "of", "layers", "to", "3", "in", "this", "experiment", ",", "mainly", "to", "agree", "with", "the", "language", "modeling", "experiments", "of", "chung2015gated", ".", "Also", "note", "that", "that", "there", "are", "no", "single", "-", "layer", "variants", "for", "gLSTM", "and", "dLSTM", ";", "they", "have", "to", "be", "implemented", "as", "multi", "-", "layer", "systems", ".", "The", "hidden", "unit", "size", "of", "the", "LSTMN", "and", "all", "comparison", "models", "(", "except", "KN5", ")", "was", "set", "to", "300", ".", "The", "results", "of", "the", "language", "modeling", "task", "are", "shown", "in", "Table", "[", "reference", "]", ".", "Perplexity", "results", "for", "KN5", "and", "RNN", "are", "taken", "from", "mikolov2015learning", ".", "As", "can", "be", "seen", ",", "the", "single", "-", "layer", "LSTMN", "outperforms", "these", "two", "baselines", "and", "the", "LSTM", "by", "a", "significant", "margin", ".", "Amongst", "all", "deep", "architectures", ",", "the", "three", "-", "layer", "LSTMN", "also", "performs", "best", ".", "We", "can", "study", "the", "memory", "activation", "mechanism", "of", "the", "machine", "reader", "by", "visualizing", "the", "attention", "scores", ".", "Figure", "[", "reference", "]", "shows", "four", "sentences", "sampled", "from", "the", "Penn", "Treebank", "validation", "set", ".", "Although", "we", "explicitly", "encourage", "the", "reader", "to", "attend", "to", "any", "memory", "slot", ",", "much", "attention", "focuses", "on", "recent", "memories", ".", "This", "agrees", "with", "the", "linguistic", "intuition", "that", "long", "-", "term", "dependencies", "are", "relatively", "rare", ".", "As", "illustrated", "in", "Figure", "[", "reference", "]", "the", "model", "captures", "some", "valid", "lexical", "relations", "(", "e.g.", ",", "the", "dependency", "between", "sits", "and", "at", ",", "sits", "and", "plays", ",", "everyone", "and", "is", ",", "is", "and", "watching", ")", ".", "Note", "that", "arcs", "here", "are", "undirected", "and", "are", "different", "from", "the", "directed", "arcs", "denoting", "head", "-", "modifier", "relations", "in", "dependency", "graphs", ".", "subsection", ":", "Sentiment", "Analysis", "Our", "second", "task", "concerns", "the", "prediction", "of", "sentiment", "labels", "of", "sentences", ".", "We", "used", "the", "Stanford", "Sentiment", "Treebank", ",", "which", "contains", "fine", "-", "grained", "sentiment", "labels", "(", "very", "positive", ",", "positive", ",", "neutral", ",", "negative", ",", "very", "negative", ")", "for", "11", ",", "855", "sentences", ".", "Following", "previous", "work", "on", "this", "dataset", ",", "we", "used", "8", ",", "544", "sentences", "for", "training", ",", "1", ",", "101", "for", "validation", ",", "and", "2", ",", "210", "for", "testing", ".", "The", "average", "sentence", "length", "is", "19.1", ".", "In", "addition", ",", "we", "also", "performed", "a", "binary", "classification", "task", "(", "positive", ",", "negative", ")", "after", "removing", "the", "neutral", "label", ".", "This", "resulted", "in", "6", ",", "920", "sentences", "for", "training", ",", "872", "for", "validation", "and", "1", ",", "821", "for", "testing", ".", "Table", "[", "reference", "]", "reports", "results", "on", "both", "fine", "-", "grained", "and", "binary", "classification", "tasks", ".", "We", "experimented", "with", "1", "-", "and", "2", "-", "layer", "LSTMNs", ".", "For", "the", "latter", "model", ",", "we", "predict", "the", "sentiment", "label", "of", "the", "sentence", "based", "on", "the", "averaged", "hidden", "vector", "passed", "to", "a", "2", "-", "layer", "neural", "network", "classifier", "with", "ReLU", "as", "the", "activation", "function", ".", "The", "memory", "size", "for", "both", "LSTMN", "models", "was", "set", "to", "168", "to", "be", "compatible", "with", "previous", "LSTM", "models", "applied", "to", "the", "same", "task", ".", "We", "used", "pre", "-", "trained", "300", "-", "D", "Glove", "840B", "vectors", "to", "initialize", "the", "word", "embeddings", ".", "The", "gradient", "for", "words", "with", "Glove", "embeddings", ",", "was", "scaled", "by", "0.35", "in", "the", "first", "epoch", "after", "which", "all", "word", "embeddings", "were", "updated", "normally", ".", "We", "used", "Adam", "for", "optimization", "with", "the", "two", "momentum", "parameters", "set", "to", "0.9", "and", "0.999", "respectively", ".", "The", "initial", "learning", "rate", "was", "set", "to", "2E", "-", "3", ".", "The", "regularization", "constant", "was", "1E", "-", "4", "and", "the", "mini", "-", "batch", "size", "was", "5", ".", "A", "dropout", "rate", "of", "0.5", "was", "applied", "to", "the", "neural", "network", "classifier", ".", "We", "compared", "our", "model", "with", "a", "wide", "range", "of", "top", "-", "performing", "systems", ".", "Most", "of", "these", "models", "(", "including", "ours", ")", "are", "LSTM", "variants", "(", "third", "block", "in", "Table", "[", "reference", "]", ")", ",", "recursive", "neural", "networks", "(", "first", "block", ")", ",", "or", "convolutional", "neural", "networks", "(", "CNNs", ";", "second", "block", ")", ".", "Recursive", "models", "assume", "the", "input", "sentences", "are", "represented", "as", "parse", "trees", "and", "can", "take", "advantage", "of", "annotations", "at", "the", "phrase", "level", ".", "LSTM", "-", "type", "models", "and", "CNNs", "are", "trained", "on", "sequential", "input", ",", "with", "the", "exception", "of", "CT", "-", "LSTM", "which", "operates", "over", "tree", "-", "structured", "network", "topologies", "such", "as", "constituent", "trees", ".", "For", "comparison", ",", "we", "also", "report", "the", "performance", "of", "the", "paragraph", "vector", "model", "(", "PV", ";", "le2014distributed", ";", "see", "Table", "[", "reference", "]", ",", "second", "block", ")", "which", "neither", "operates", "on", "trees", "nor", "sequences", "but", "learns", "distributed", "document", "representations", "parameterized", "directly", ".", "The", "results", "in", "Table", "[", "reference", "]", "show", "that", "both", "1", "-", "and", "2", "-", "layer", "LSTMNs", "outperform", "the", "LSTM", "baselines", "while", "achieving", "numbers", "comparable", "to", "state", "of", "the", "art", ".", "The", "number", "of", "layers", "for", "our", "models", "was", "set", "to", "be", "comparable", "to", "previously", "published", "results", ".", "On", "the", "fine", "-", "grained", "and", "binary", "classification", "tasks", "our", "2", "-", "layer", "LSTMN", "performs", "close", "to", "the", "best", "system", "T", "-", "CNN", ".", "Figure", "[", "reference", "]", "shows", "examples", "of", "intra", "-", "attention", "for", "sentiment", "words", ".", "Interestingly", ",", "the", "network", "learns", "to", "associate", "sentiment", "important", "words", "such", "as", "though", "and", "fantastic", "or", "not", "and", "good", ".", "subsection", ":", "Natural", "Language", "Inference", "The", "ability", "to", "reason", "about", "the", "semantic", "relationship", "between", "two", "sentences", "is", "an", "integral", "part", "of", "text", "understanding", ".", "We", "therefore", "evaluate", "our", "model", "on", "recognizing", "textual", "entailment", ",", "i.e.", ",", "whether", "two", "premise", "-", "hypothesis", "pairs", "are", "entailing", ",", "contradictory", ",", "or", "neutral", ".", "For", "this", "task", "we", "used", "the", "Stanford", "Natural", "Language", "Inference", "(", "SNLI", ")", "dataset", ",", "which", "contains", "premise", "-", "hypothesis", "pairs", "and", "target", "labels", "indicating", "their", "relation", ".", "After", "removing", "sentences", "with", "unknown", "labels", ",", "we", "end", "up", "with", "549", ",", "367", "pairs", "for", "training", ",", "9", ",", "842", "for", "development", "and", "9", ",", "824", "for", "testing", ".", "The", "vocabulary", "size", "is", "36", ",", "809", "and", "the", "average", "sentence", "length", "is", "22", ".", "We", "performed", "lower", "-", "casing", "and", "tokenization", "for", "the", "entire", "dataset", ".", "Recent", "approaches", "use", "two", "sequential", "LSTMs", "to", "encode", "the", "premise", "and", "the", "hypothesis", "respectively", ",", "and", "apply", "neural", "attention", "to", "reason", "about", "their", "logical", "relationship", ".", "Furthermore", ",", "rocktaschel2015reasoning", "show", "that", "a", "non", "-", "standard", "encoder", "-", "decoder", "architecture", "which", "processes", "the", "hypothesis", "conditioned", "on", "the", "premise", "results", "significantly", "boosts", "performance", ".", "We", "use", "a", "similar", "approach", "to", "tackle", "this", "task", "with", "LSTMNs", ".", "Specifically", ",", "we", "use", "two", "LSTMNs", "to", "read", "the", "premise", "and", "hypothesis", ",", "and", "then", "match", "them", "by", "comparing", "their", "hidden", "state", "tapes", ".", "We", "perform", "average", "pooling", "for", "the", "hidden", "state", "tape", "of", "each", "LSTMN", ",", "and", "concatenate", "the", "two", "averages", "to", "form", "the", "input", "to", "a", "2", "-", "layer", "neural", "network", "classifier", "with", "ReLU", "as", "the", "activation", "function", ".", "We", "used", "pre", "-", "trained", "300", "-", "D", "Glove", "840B", "vectors", "to", "initialize", "the", "word", "embeddings", ".", "Out", "-", "of", "-", "vocabulary", "(", "OOV", ")", "words", "were", "initialized", "randomly", "with", "Gaussian", "samples", "(", "=", "0", ",", "=", "1", ")", ".", "We", "only", "updated", "OOV", "vectors", "in", "the", "first", "epoch", ",", "after", "which", "all", "word", "embeddings", "were", "updated", "normally", ".", "The", "dropout", "rate", "was", "selected", "from", "[", "0.1", ",", "0.2", ",", "0.3", ",", "0.4", "]", ".", "We", "used", "Adam", "for", "optimization", "with", "the", "two", "momentum", "parameters", "set", "to", "0.9", "and", "0.999", "respectively", ",", "and", "the", "initial", "learning", "rate", "set", "to", "1E", "-", "3", ".", "The", "mini", "-", "batch", "size", "was", "set", "to", "16", "or", "32", ".", "For", "a", "fair", "comparison", "against", "previous", "work", ",", "we", "report", "results", "with", "different", "hidden", "/", "memory", "dimensions", "(", "i.e.", ",", "100", ",", "300", ",", "and", "450", ")", ".", "We", "compared", "variants", "of", "our", "model", "against", "different", "types", "of", "LSTMs", "(", "see", "the", "second", "block", "in", "Table", "[", "reference", "]", ")", ".", "Specifically", ",", "these", "include", "a", "model", "which", "encodes", "the", "premise", "and", "hypothesis", "independently", "with", "two", "LSTMs", ",", "a", "shared", "LSTM", ",", "a", "word", "-", "by", "-", "word", "attention", "model", ",", "and", "a", "matching", "LSTM", "(", "mLSTM", ";", "wang2015learning", ")", ".", "This", "model", "sequentially", "processes", "the", "hypothesis", ",", "and", "at", "each", "position", "tries", "to", "match", "the", "current", "word", "with", "an", "attention", "-", "weighted", "representation", "of", "the", "premise", "(", "rather", "than", "basing", "its", "predictions", "on", "whole", "sentence", "embeddings", ")", ".", "We", "also", "compared", "our", "models", "with", "a", "bag", "-", "of", "-", "words", "baseline", "which", "averages", "the", "pre", "-", "trained", "embeddings", "for", "the", "words", "in", "each", "sentence", "and", "concatenates", "them", "to", "create", "features", "for", "a", "logistic", "regression", "classifier", "(", "first", "block", "in", "Table", "[", "reference", "]", ")", ".", "LSTMNs", "achieve", "better", "performance", "compared", "to", "LSTMs", "(", "with", "and", "without", "attention", ";", "2nd", "block", "in", "Table", "[", "reference", "]", ")", ".", "We", "also", "observe", "that", "fusion", "is", "generally", "beneficial", ",", "and", "that", "deep", "fusion", "slightly", "improves", "over", "shallow", "fusion", ".", "One", "explanation", "is", "that", "with", "deep", "fusion", "the", "inter", "-", "attention", "vectors", "are", "recurrently", "memorized", "by", "the", "decoder", "with", "a", "gating", "operation", ",", "which", "also", "improves", "the", "information", "flow", "of", "the", "network", ".", "With", "standard", "training", ",", "our", "deep", "fusion", "yields", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "in", "this", "task", ".", "Although", "encouraging", ",", "this", "result", "should", "be", "interpreted", "with", "caution", "since", "our", "model", "has", "substantially", "more", "parameters", "compared", "to", "related", "systems", ".", "We", "could", "compare", "different", "models", "using", "the", "same", "number", "of", "total", "parameters", ".", "However", ",", "this", "would", "inevitably", "introduce", "other", "biases", ",", "e.g.", ",", "the", "number", "of", "hyper", "-", "parameters", "would", "become", "different", ".", "section", ":", "Conclusions", "In", "this", "paper", "we", "proposed", "a", "machine", "reading", "simulator", "to", "address", "the", "limitations", "of", "recurrent", "neural", "networks", "when", "processing", "inherently", "structured", "input", ".", "Our", "model", "is", "based", "on", "a", "Long", "Short", "-", "Term", "Memory", "architecture", "embedded", "with", "a", "memory", "network", ",", "explicitly", "storing", "contextual", "representations", "of", "input", "tokens", "without", "recursively", "compressing", "them", ".", "More", "importantly", ",", "an", "intra", "-", "attention", "mechanism", "is", "employed", "for", "memory", "addressing", ",", "as", "a", "way", "to", "induce", "undirected", "relations", "among", "tokens", ".", "The", "attention", "layer", "is", "not", "optimized", "with", "a", "direct", "supervision", "signal", "but", "with", "the", "entire", "network", "in", "downstream", "tasks", ".", "Experimental", "results", "across", "three", "tasks", "show", "that", "our", "model", "yields", "performance", "comparable", "or", "superior", "to", "state", "of", "the", "art", ".", "Although", "our", "experiments", "focused", "on", "LSTMs", ",", "the", "idea", "of", "building", "more", "structure", "aware", "neural", "models", "is", "general", "and", "can", "be", "applied", "to", "other", "types", "of", "networks", ".", "When", "direct", "supervision", "is", "provided", ",", "similar", "architectures", "can", "be", "adapted", "to", "tasks", "such", "as", "dependency", "parsing", "and", "relation", "extraction", ".", "In", "the", "future", ",", "we", "hope", "to", "develop", "more", "linguistically", "plausible", "neural", "architectures", "able", "to", "reason", "over", "nested", "structures", "and", "neural", "models", "that", "learn", "to", "discover", "compositionality", "with", "weak", "or", "indirect", "supervision", ".", "section", ":", "Acknowledgments", "We", "thank", "members", "of", "the", "ILCC", "at", "the", "School", "of", "Informatics", "and", "the", "anonymous", "reviewers", "for", "helpful", "comments", ".", "The", "support", "of", "the", "European", "Research", "Council", "under", "award", "number", "681760", "“", "Translating", "Multiple", "Modalities", "into", "Text", "”", "is", "gratefully", "acknowledged", ".", "bibliography", ":", "References"], "entity": [[2, 9, ["Long", "Short", "-", "Term", "Memory", "-", "Networks"], "Method"], [10, 12, ["Machine", "Reading"], "Task"], [23, 27, ["sequence", "-", "level", "networks"], "Method"], [36, 39, ["machine", "reading", "simulator"], "Method"], [49, 51, ["shallow", "reasoning"], "Method"], [68, 70, ["memory", "network"], "Method"], [84, 85, ["recurrence"], "Task"], [119, 123, ["encoder", "-", "decoder", "architecture"], "Method"], [126, 128, ["language", "modeling"], "Task"], [129, 131, ["sentiment", "analysis"], "Task"], [133, 136, ["natural", "language", "inference"], "Task"], [155, 159, ["sequence", "-", "level", "network"], "Method"], [166, 168, ["text", "processing"], "Task"], [172, 174, ["recurrent", "network"], "Method"], [191, 193, ["machine", "reader"], "Method"], [202, 204, ["machine", "reading"], "Task"], [213, 217, ["answering", "reading", "comprehension", "questions"], "Task"], [219, 223, ["fact", "and", "relation", "extraction"], "Task"], [224, 226, ["ontology", "learning"], "Task"], [228, 230, ["textual", "entailment"], "Task"], [242, 247, ["general", "-", "purpose", "reading", "simulator"], "Method"], [251, 254, ["human", "language", "processing"], "Task"], [256, 259, ["fact", "language", "comprehension"], "Task"], [285, 287, ["machine", "reader"], "Method"], [291, 295, ["extracting", "and", "representing", "meaning"], "Task"], [334, 336, ["semantic", "representations"], "Method"], [377, 379, ["partial", "representations"], "Method"], [383, 386, ["Language", "modeling", "tools"], "Method"], [388, 391, ["recurrent", "neural", "networks"], "Method"], [392, 393, ["RNN"], "Method"], [401, 402, ["RNNs"], "Method"], [436, 440, ["sequence", "-", "level", "networks"], "Method"], [452, 455, ["model", "training", "problems"], "Task"], [457, 461, ["vanishing", "and", "exploding", "gradients"], "Task"], [468, 471, ["gated", "activation", "functions"], "Method"], [485, 487, ["gradient", "clipping"], "Method"], [493, 496, ["memory", "compression", "problems"], "Task"], [547, 551, ["sequence", "-", "level", "networks"], "Method"], [587, 590, ["text", "processing", "system"], "Method"], [603, 606, ["recurrent", "language", "model"], "Method"], [615, 616, ["FBI"], "Method"], [701, 703, ["neural", "networks"], "Method"], [717, 720, ["recurrent", "neural", "networks"], "Method"], [754, 756, ["attention", "mechanism"], "Method"], [758, 760, ["recurrent", "controller"], "Method"], [770, 772, ["recurrent", "network"], "Method"], [793, 796, ["memory", "network", "module"], "Method"], [801, 803, ["recurrent", "network"], "Method"], [805, 806, ["attention"], "Method"], [807, 809, ["memory", "addressing"], "Task"], [811, 812, ["attention"], "Method"], [816, 818, ["inductive", "module"], "Method"], [841, 843, ["memory", "network"], "Method"], [862, 864, ["representation", "learner"], "Method"], [880, 887, ["Long", "Short", "-", "Term", "Memory", "-", "Network"], "Method"], [888, 889, ["LSTMN"], "Method"], [893, 895, ["reading", "simulator"], "Method"], [900, 903, ["sequence", "processing", "tasks"], "Task"], [914, 915, ["LSTMN"], "Method"], [958, 960, ["learning", "representations"], "Method"], [967, 968, ["LSTMN"], "Method"], [969, 971, ["language", "modeling"], "Task"], [972, 974, ["sentiment", "analysis"], "Task"], [976, 979, ["natural", "language", "inference"], "Task"], [986, 988, ["LSTMN", "models"], "Method"], [1025, 1027, ["machine", "reader"], "Method"], [1029, 1032, ["recurrent", "neural", "network"], "Method"], [1047, 1050, ["shallow", "structure", "reasoning"], "Method"], [1054, 1057, ["Recurrent", "neural", "network"], "Method"], [1058, 1059, ["RNNs"], "Method"], [1066, 1068, ["sequence", "modeling"], "Task"], [1069, 1076, ["sequence", "-", "to", "-", "sequence", "transduction", "tasks"], "Task"], [1088, 1090, ["machine", "translation"], "Task"], [1091, 1093, ["sentence", "compression"], "Task"], [1095, 1097, ["reading", "comprehension"], "Task"], [1147, 1148, ["RNNs"], "Method"], [1149, 1151, ["memory", "compression"], "Task"], [1161, 1163, ["memory", "representation"], "Method"], [1171, 1172, ["parameters"], "Metric"], [1183, 1187, ["encoder", "-", "decoder", "architecture"], "Method"], [1195, 1197, ["attention", "mechanism"], "Method"], [1221, 1223, ["sequence", "encoder"], "Method"], [1241, 1243, ["neural", "models"], "Method"], [1264, 1267, ["recursive", "neural", "networks"], "Method"], [1269, 1272, ["natural", "language", "representations"], "Task"], [1277, 1280, ["recurrent", "neural", "networks"], "Method"], [1284, 1289, ["modular", ",", "structured", "neural", "models"], "Method"], [1296, 1299, ["recurrent", "neural", "network"], "Method"], [1301, 1304, ["external", "memory", "stack"], "Method"], [1306, 1309, ["context", "free", "grammars"], "Method"], [1314, 1316, ["Memory", "Networks"], "Method"], [1325, 1327, ["neural", "networks"], "Method"], [1341, 1344, ["memory", "addressing", "mechanism"], "Method"], [1354, 1356, ["machine", "translation"], "Task"], [1378, 1381, ["recurrent", "neural", "network"], "Method"], [1385, 1386, ["LSTM"], "Method"], [1388, 1392, ["external", "memory", "block", "component"], "Method"], [1402, 1405, ["structured", "neural", "network"], "Method"], [1406, 1409, ["episodic", "memory", "modules"], "Method"], [1410, 1412, ["natural", "language"], "Task"], [1414, 1417, ["visual", "question", "answering"], "Task"], [1426, 1429, ["memory", "and", "attention"], "Method"], [1431, 1434, ["recurrent", "neural", "network"], "Method"], [1448, 1450, ["representation", "learning"], "Task"], [1475, 1481, ["shift", "-", "reduce", "type", "neural", "models"], "Method"], [1488, 1489, ["induction"], "Task"], [1510, 1513, ["dependency", "grammar", "induction"], "Task"], [1534, 1536, ["machine", "reader"], "Method"], [1549, 1552, ["recurrent", "neural", "network"], "Method"], [1560, 1565, ["Long", "Short", "-", "Term", "Memory"], "Method"], [1566, 1567, ["LSTM"], "Method"], [1585, 1588, ["implicit", "relation", "analysis"], "Task"], [1592, 1598, ["attention", "-", "based", "memory", "addressing", "mechanism"], "Method"], [1612, 1617, ["Long", "Short", "-", "Term", "Memory"], "Method"], [1625, 1630, ["Long", "Short", "-", "Term", "Memory"], "Method"], [1631, 1636, ["Long", "Short", "-", "Term", "Memory"], "Method"], [1637, 1638, ["LSTM"], "Method"], [1714, 1715, ["RNN"], "Method"], [1717, 1718, ["LSTM"], "Method"], [1719, 1722, ["additive", "memory", "updates"], "Method"], [1739, 1740, ["predictions"], "Task"], [1743, 1750, ["Long", "Short", "-", "Term", "Memory", "-", "Network"], "Method"], [1756, 1757, ["LSTMs"], "Method"], [1769, 1771, ["recursive", "compression"], "Method"], [1772, 1773, ["LSTMs"], "Method"], [1778, 1780, ["state", "representations"], "Method"], [1781, 1782, ["composition"], "Task"], [1819, 1822, ["recursive", "state", "update"], "Method"], [1826, 1828, ["Markov", "manner"], "Method"], [1833, 1834, ["LSTMs"], "Method"], [1883, 1884, ["LSTMs"], "Method"], [1885, 1888, ["modeling", "structured", "input"], "Task"], [1890, 1891, ["LSTM"], "Method"], [1945, 1947, ["memory", "network"], "Method"], [1950, 1957, ["Long", "Short", "-", "Term", "Memory", "-", "Network"], "Method"], [1958, 1959, ["LSTMN"], "Method"], [1962, 1964, ["contextual", "representation"], "Method"], [1997, 1998, ["LSTM"], "Method"], [2012, 2017, ["non", "-", "Markov", "state", "updates"], "Method"], [2044, 2046, ["read", "operation"], "Method"], [2075, 2077, ["write", "operation"], "Method"], [2085, 2089, ["incrementally", "updating", "previous", "memories"], "Task"], [2106, 2107, ["LSTMN"], "Method"], [2210, 2212, ["attention", "layer"], "Method"], [2282, 2283, ["LSTMN"], "Method"], [2288, 2292, ["inducing", "relations", "between", "tokens"], "Task"], [2305, 2308, ["representation", "learning", "network"], "Method"], [2346, 2348, ["downstream", "task"], "Task"], [2358, 2361, ["relational", "reasoning", "module"], "Method"], [2364, 2368, ["memory", "and", "hidden", "layers"], "Method"], [2375, 2377, ["stacked", "LSTM"], "Method"], [2379, 2384, ["multi", "-", "hop", "memory", "network"], "Method"], [2429, 2430, ["Decoder"], "Method"], [2431, 2434, ["shallow", "attention", "fusion"], "Method"], [2437, 2438, ["Decoder"], "Method"], [2439, 2442, ["deep", "attention", "fusion"], "Method"], [2446, 2449, ["Modeling", "Two", "Sequences"], "Task"], [2450, 2451, ["LSTMN"], "Method"], [2451, 2455, ["Natural", "language", "processing", "tasks"], "Task"], [2457, 2459, ["machine", "translation"], "Task"], [2460, 2462, ["textual", "entailment"], "Task"], [2478, 2481, ["modeling", "two", "sequences"], "Task"], [2482, 2484, ["recurrent", "networks"], "Method"], [2486, 2490, ["encoder", "-", "decoder", "architecture"], "Method"], [2526, 2527, ["LSTMN"], "Method"], [2529, 2530, ["attention"], "Method"], [2531, 2535, ["intra", "-", "relation", "reasoning"], "Task"], [2538, 2542, ["encoder", "-", "decoder", "network"], "Method"], [2543, 2545, ["attention", "module"], "Method"], [2580, 2583, ["Shallow", "Attention", "Fusion"], "Method"], [2583, 2585, ["Shallow", "fusion"], "Method"], [2588, 2589, ["LSTMN"], "Method"], [2600, 2604, ["encoder", "-", "decoder", "architecture"], "Method"], [2610, 2611, ["RNN"], "Method"], [2612, 2613, ["LSTM"], "Method"], [2626, 2627, ["decoder"], "Method"], [2630, 2635, ["LSTMNs", "with", "intra", "-", "attention"], "Method"], [2663, 2666, ["Deep", "Attention", "Fusion"], "Method"], [2666, 2668, ["Deep", "fusion"], "Method"], [2679, 2680, ["decoder"], "Method"], [2683, 2685, ["state", "updates"], "Task"], [2729, 2734, ["computation", "of", "intra", "-", "attention"], "Task"], [2800, 2802, ["adaptive", "representation"], "Method"], [2817, 2820, ["adaptive", "source", "representation"], "Method"], [2826, 2828, ["gating", "operation"], "Method"], [2877, 2879, ["deep", "fusion"], "Method"], [2882, 2884, ["recurrent", "storage"], "Method"], [2893, 2895, ["memory", "network"], "Method"], [2931, 2933, ["language", "modeling"], "Method"], [2952, 2954, ["meaning", "representations"], "Method"], [2955, 2959, ["generic", "sentence", "classification", "tasks"], "Task"], [2961, 2963, ["sentiment", "analysis"], "Task"], [2970, 2971, ["LSTMN"], "Method"], [2997, 2999, ["Language", "Modeling"], "Method"], [3000, 3002, ["language", "modeling"], "Method"], [3063, 3066, ["average", "sentence", "length"], "Metric"], [3071, 3072, ["perplexity"], "Method"], [3074, 3076, ["evaluation", "metric"], "Metric"], [3098, 3101, ["stochastic", "gradient", "descent"], "Method"], [3102, 3103, ["optimization"], "Task"], [3106, 3110, ["learning", "rate", "of", "0.65"], "Metric"], [3176, 3177, ["LSTMN"], "Method"], [3188, 3196, ["Kneser", "-", "Ney", "5", "-", "gram", "language", "model"], "Method"], [3197, 3198, ["KN5"], "Method"], [3204, 3208, ["non", "-", "neural", "baseline"], "Method"], [3210, 3213, ["language", "modeling", "task"], "Task"], [3238, 3240, ["stacked", "LSTM"], "Method"], [3241, 3242, ["sLSTM"], "Method"], [3250, 3251, ["gLSTM"], "Method"], [3261, 3262, ["dLSTM"], "Method"], [3313, 3314, ["gLSTM"], "Method"], [3315, 3316, ["dLSTM"], "Method"], [3354, 3356, ["language", "modeling"], "Method"], [3367, 3371, ["single", "-", "layer", "variants"], "Method"], [3372, 3373, ["gLSTM"], "Method"], [3374, 3375, ["dLSTM"], "Method"], [3382, 3386, ["multi", "-", "layer", "systems"], "Method"], [3393, 3394, ["LSTMN"], "Method"], [3396, 3398, ["comparison", "models"], "Method"], [3400, 3401, ["KN5"], "Method"], [3411, 3414, ["language", "modeling", "task"], "Task"], [3422, 3423, ["Perplexity"], "Method"], [3425, 3426, ["KN5"], "Method"], [3427, 3428, ["RNN"], "Method"], [3449, 3450, ["LSTM"], "Method"], [3457, 3459, ["deep", "architectures"], "Method"], [3473, 3476, ["memory", "activation", "mechanism"], "Method"], [3478, 3480, ["machine", "reader"], "Task"], [3483, 3485, ["attention", "scores"], "Metric"], [3597, 3599, ["Sentiment", "Analysis"], "Task"], [3604, 3610, ["prediction", "of", "sentiment", "labels", "of", "sentences"], "Task"], [3658, 3659, ["training"], "Task"], [3664, 3665, ["validation"], "Task"], [3671, 3672, ["testing"], "Task"], [3674, 3677, ["average", "sentence", "length"], "Metric"], [3687, 3690, ["binary", "classification", "task"], "Task"], [3709, 3710, ["training"], "Task"], [3713, 3714, ["validation"], "Task"], [3719, 3720, ["testing"], "Task"], [3729, 3732, ["fine", "-", "grained"], "Task"], [3733, 3736, ["binary", "classification", "tasks"], "Task"], [3743, 3747, ["2", "-", "layer", "LSTMNs"], "Method"], [3770, 3776, ["2", "-", "layer", "neural", "network", "classifier"], "Method"], [3777, 3778, ["ReLU"], "Method"], [3784, 3786, ["memory", "size"], "Metric"], [3788, 3790, ["LSTMN", "models"], "Method"], [3812, 3817, ["300", "-", "D", "Glove", "840B"], "Method"], [3851, 3852, ["Adam"], "Method"], [3853, 3854, ["optimization"], "Task"], [3858, 3859, ["parameters"], "Metric"], [3868, 3870, ["learning", "rate"], "Metric"], [3894, 3898, ["dropout", "rate", "of", "0.5"], "Metric"], [3902, 3905, ["neural", "network", "classifier"], "Method"], [3941, 3944, ["recursive", "neural", "networks"], "Method"], [3950, 3953, ["convolutional", "neural", "networks"], "Method"], [3954, 3955, ["CNNs"], "Method"], [3960, 3962, ["Recursive", "models"], "Method"], [3987, 3988, ["CNNs"], "Method"], [4004, 4009, ["tree", "-", "structured", "network", "topologies"], "Method"], [4024, 4027, ["paragraph", "vector", "model"], "Method"], [4028, 4029, ["PV"], "Method"], [4050, 4053, ["distributed", "document", "representations"], "Method"], [4069, 4073, ["2", "-", "layer", "LSTMNs"], "Method"], [4106, 4113, ["fine", "-", "grained", "and", "binary", "classification", "tasks"], "Task"], [4114, 4118, ["2", "-", "layer", "LSTMN"], "Method"], [4124, 4127, ["T", "-", "CNN"], "Method"], [4164, 4167, ["Natural", "Language", "Inference"], "Task"], [4183, 4185, ["text", "understanding"], "Task"], [4192, 4195, ["recognizing", "textual", "entailment"], "Task"], [4218, 4222, ["Stanford", "Natural", "Language", "Inference"], "Material"], [4223, 4224, ["SNLI"], "Material"], [4262, 4263, ["development"], "Task"], [4271, 4273, ["vocabulary", "size"], "Metric"], [4279, 4282, ["average", "sentence", "length"], "Metric"], [4287, 4290, ["lower", "-", "casing"], "Method"], [4291, 4292, ["tokenization"], "Task"], [4301, 4303, ["sequential", "LSTMs"], "Method"], [4314, 4316, ["neural", "attention"], "Method"], [4332, 4336, ["encoder", "-", "decoder", "architecture"], "Method"], [4359, 4360, ["LSTMNs"], "Method"], [4366, 4367, ["LSTMNs"], "Method"], [4387, 4389, ["average", "pooling"], "Method"], [4396, 4397, ["LSTMN"], "Method"], [4409, 4415, ["2", "-", "layer", "neural", "network", "classifier"], "Method"], [4416, 4417, ["ReLU"], "Method"], [4427, 4432, ["300", "-", "D", "Glove", "840B"], "Method"], [4482, 4484, ["dropout", "rate"], "Metric"], [4499, 4500, ["Adam"], "Method"], [4501, 4502, ["optimization"], "Task"], [4506, 4507, ["parameters"], "Metric"], [4517, 4519, ["learning", "rate"], "Metric"], [4559, 4560, ["300"], "Method"], [4562, 4563, ["450"], "Method"], [4575, 4576, ["LSTMs"], "Method"], [4603, 4604, ["LSTMs"], "Method"], [4610, 4617, ["word", "-", "by", "-", "word", "attention", "model"], "Method"], [4623, 4624, ["mLSTM"], "Method"], [4673, 4679, ["bag", "-", "of", "-", "words", "baseline"], "Method"], [4700, 4703, ["logistic", "regression", "classifier"], "Method"], [4713, 4714, ["LSTMNs"], "Method"], [4719, 4720, ["LSTMs"], "Method"], [4739, 4740, ["fusion"], "Task"], [4746, 4748, ["deep", "fusion"], "Method"], [4751, 4753, ["shallow", "fusion"], "Method"], [4759, 4761, ["deep", "fusion"], "Method"], [4771, 4772, ["decoder"], "Method"], [4774, 4776, ["gating", "operation"], "Method"], [4789, 4790, ["training"], "Method"], [4792, 4794, ["deep", "fusion"], "Method"], [4824, 4825, ["parameters"], "Metric"], [4841, 4842, ["parameters"], "Metric"], [4857, 4860, ["hyper", "-", "parameters"], "Metric"], [4873, 4876, ["machine", "reading", "simulator"], "Method"], [4881, 4884, ["recurrent", "neural", "networks"], "Method"], [4905, 4907, ["memory", "network"], "Method"], [4910, 4912, ["contextual", "representations"], "Method"], [4924, 4928, ["intra", "-", "attention", "mechanism"], "Method"], [4931, 4933, ["memory", "addressing"], "Task"], [4945, 4947, ["attention", "layer"], "Method"], [4961, 4963, ["downstream", "tasks"], "Task"], [4989, 4990, ["LSTMs"], "Method"], [4996, 5000, ["structure", "aware", "neural", "models"], "Method"], [5027, 5029, ["dependency", "parsing"], "Task"], [5030, 5032, ["relation", "extraction"], "Task"], [5042, 5046, ["linguistically", "plausible", "neural", "architectures"], "Method"], [5053, 5055, ["neural", "models"], "Method"], [60, 65, ["Long", "Short", "-", "Term", "Memory"], "Method"], [475, 480, ["Long", "Short", "-", "Term", "Memory"], "Method"], [481, 482, ["LSTM"], "Method"], [1018, 1019, ["LSTMs"], "Method"], [1936, 1937, ["LSTM"], "Method"], [2007, 2009, ["attention", "layer"], "Method"], [2319, 2321, ["attention", "layer"], "Method"], [2924, 2925, ["LSTMN"], "Method"], [2984, 2987, ["natural", "language", "inference"], "Task"], [3224, 3225, ["LSTM"], "Method"], [3232, 3233, ["LSTM"], "Method"], [3248, 3249, ["LSTM"], "Method"], [3259, 3260, ["LSTM"], "Method"], [3270, 3271, ["LSTM"], "Method"], [3295, 3296, ["LSTM"], "Method"], [3405, 3406, ["300"], "Method"], [3442, 3443, ["LSTMN"], "Method"], [3464, 3465, ["LSTMN"], "Method"], [3799, 3800, ["LSTM"], "Method"], [3929, 3930, ["LSTM"], "Method"], [3982, 3983, ["LSTM"], "Method"], [4000, 4001, ["LSTM"], "Method"], [4075, 4076, ["LSTM"], "Method"], [4607, 4608, ["LSTM"], "Method"], [4621, 4622, ["LSTM"], "Method"], [4896, 4901, ["Long", "Short", "-", "Term", "Memory"], "Method"]]} {"doc_id": "0910a4c470a410fac446f4026f7c8ef512ae7427", "text": ["document", ":", "Hierarchical", "Question", "-", "Image", "Co", "-", "Attention", "for", "Visual", "Question", "Answering", "A", "number", "of", "recent", "works", "have", "proposed", "attention", "models", "for", "Visual", "Question", "Answering", "(", "VQA", ")", "that", "generate", "spatial", "maps", "highlighting", "image", "regions", "relevant", "to", "answering", "the", "question", ".", "In", "this", "paper", ",", "we", "argue", "that", "in", "addition", "to", "modeling", "‘", "‘", "where", "to", "look", "’", "’", "or", "visual", "attention", ",", "it", "is", "equally", "important", "to", "model", "‘", "‘", "what", "words", "to", "listen", "to", "’", "’", "or", "question", "attention", ".", "We", "present", "a", "novel", "co", "-", "attention", "model", "for", "VQA", "that", "jointly", "reasons", "about", "image", "and", "question", "attention", ".", "In", "addition", ",", "our", "model", "reasons", "about", "the", "question", "(", "and", "consequently", "the", "image", "via", "the", "co", "-", "attention", "mechanism", ")", "in", "a", "hierarchical", "fashion", "via", "a", "novel", "1", "-", "dimensional", "convolution", "neural", "networks", "(", "CNN", ")", ".", "Our", "model", "improves", "the", "state", "-", "of", "-", "the", "-", "art", "on", "the", "VQA", "dataset", "from", "60.3", "%", "to", "60.5", "%", ",", "and", "from", "61.6", "%", "to", "63.3", "%", "on", "the", "COCO", "-", "QA", "dataset", ".", "By", "using", "ResNet", ",", "the", "performance", "is", "further", "improved", "to", "62.1", "%", "for", "VQA", "and", "65.4", "%", "for", "COCO", "-", "QA", ".", ".", "[", "itemize", "]", "leftmargin=20pt", "section", ":", "Introduction", "Visual", "Question", "Answering", "(", "VQA", ")", "has", "emerged", "as", "a", "prominent", "multi", "-", "discipline", "research", "problem", "in", "both", "academia", "and", "industry", ".", "To", "correctly", "answer", "visual", "questions", "about", "an", "image", ",", "the", "machine", "needs", "to", "understand", "both", "the", "image", "and", "question", ".", "Recently", ",", "visual", "attention", "based", "models", "have", "been", "explored", "for", "VQA", ",", "where", "the", "attention", "mechanism", "typically", "produces", "a", "spatial", "map", "highlighting", "image", "regions", "relevant", "to", "answering", "the", "question", ".", "So", "far", ",", "all", "attention", "models", "for", "VQA", "in", "literature", "have", "focused", "on", "the", "problem", "of", "identifying", "‘", "‘", "where", "to", "look", "’", "’", "or", "visual", "attention", ".", "In", "this", "paper", ",", "we", "argue", "that", "the", "problem", "of", "identifying", "‘", "‘", "which", "words", "to", "listen", "to", "’", "’", "or", "question", "attention", "is", "equally", "important", ".", "Consider", "the", "questions", "‘", "‘", "how", "many", "horses", "are", "in", "this", "image", "?", "’", "’", "and", "‘", "‘", "how", "many", "horses", "can", "you", "see", "in", "this", "image", "?", "\"", ".", "They", "have", "the", "same", "meaning", ",", "essentially", "captured", "by", "the", "first", "three", "words", ".", "A", "machine", "that", "attends", "to", "the", "first", "three", "words", "would", "arguably", "be", "more", "robust", "to", "linguistic", "variations", "irrelevant", "to", "the", "meaning", "and", "answer", "of", "the", "question", ".", "Motivated", "by", "this", "observation", ",", "in", "addition", "to", "reasoning", "about", "visual", "attention", ",", "we", "also", "address", "the", "problem", "of", "question", "attention", ".", "Specifically", ",", "we", "present", "a", "novel", "multi", "-", "modal", "attention", "model", "for", "VQA", "with", "the", "following", "two", "unique", "features", ":", "Co", "-", "Attention", ":", "We", "propose", "a", "novel", "mechanism", "that", "jointly", "reasons", "about", "visual", "attention", "and", "question", "attention", ",", "which", "we", "refer", "to", "as", "co", "-", "attention", ".", "Unlike", "previous", "works", ",", "which", "only", "focus", "on", "visual", "attention", ",", "our", "model", "has", "a", "natural", "symmetry", "between", "the", "image", "and", "question", ",", "in", "the", "sense", "that", "the", "image", "representation", "is", "used", "to", "guide", "the", "question", "attention", "and", "the", "question", "representation", "(", "s", ")", "are", "used", "to", "guide", "image", "attention", ".", "Question", "Hierarchy", ":", "We", "build", "a", "hierarchical", "architecture", "that", "co", "-", "attends", "to", "the", "image", "and", "question", "at", "three", "levels", ":", "(", "a", ")", "word", "level", ",", "(", "b", ")", "phrase", "level", "and", "(", "c", ")", "question", "level", ".", "At", "the", "word", "level", ",", "we", "embed", "the", "words", "to", "a", "vector", "space", "through", "an", "embedding", "matrix", ".", "At", "the", "phrase", "level", ",", "1", "-", "dimensional", "convolution", "neural", "networks", "are", "used", "to", "capture", "the", "information", "contained", "in", "unigrams", ",", "bigrams", "and", "trigrams", ".", "Specifically", ",", "we", "convolve", "word", "representations", "with", "temporal", "filters", "of", "varying", "support", ",", "and", "then", "combine", "the", "various", "n", "-", "gram", "responses", "by", "pooling", "them", "into", "a", "single", "phrase", "level", "representation", ".", "At", "the", "question", "level", ",", "we", "use", "recurrent", "neural", "networks", "to", "encode", "the", "entire", "question", ".", "For", "each", "level", "of", "the", "question", "representation", "in", "this", "hierarchy", ",", "we", "construct", "joint", "question", "and", "image", "co", "-", "attention", "maps", ",", "which", "are", "then", "combined", "recursively", "to", "ultimately", "predict", "a", "distribution", "over", "the", "answers", ".", "Overall", ",", "the", "main", "contributions", "of", "our", "work", "are", ":", "We", "propose", "a", "novel", "co", "-", "attention", "mechanism", "for", "VQA", "that", "jointly", "performs", "question", "-", "guided", "visual", "attention", "and", "image", "-", "guided", "question", "attention", ".", "We", "explore", "this", "mechanism", "with", "two", "strategies", ",", "parallel", "and", "alternating", "co", "-", "attention", ",", "which", "are", "described", "in", "Sec", ".", "[", "reference", "]", ";", "We", "propose", "a", "hierarchical", "architecture", "to", "represent", "the", "question", ",", "and", "consequently", "construct", "image", "-", "question", "co", "-", "attention", "maps", "at", "3", "different", "levels", ":", "word", "level", ",", "phrase", "level", "and", "question", "level", ".", "These", "co", "-", "attended", "features", "are", "then", "recursively", "combined", "from", "word", "level", "to", "question", "level", "for", "the", "final", "answer", "prediction", ";", "At", "the", "phrase", "level", ",", "we", "propose", "a", "novel", "convolution", "-", "pooling", "strategy", "to", "adaptively", "select", "the", "phrase", "sizes", "whose", "representations", "are", "passed", "to", "the", "question", "level", "representation", ";", "Finally", ",", "we", "evaluate", "our", "proposed", "model", "on", "two", "large", "datasets", ",", "VQA", "and", "COCO", "-", "QA", ".", "We", "also", "perform", "ablation", "studies", "to", "quantify", "the", "roles", "of", "different", "components", "in", "our", "model", ".", "section", ":", "Related", "Work", "Many", "recent", "works", "have", "proposed", "models", "for", "VQA", ".", "We", "compare", "and", "relate", "our", "proposed", "co", "-", "attention", "mechanism", "to", "other", "vision", "and", "language", "attention", "mechanisms", "in", "literature", ".", "Image", "attention", ".", "Instead", "of", "directly", "using", "the", "holistic", "entire", "-", "image", "embedding", "from", "the", "fully", "connected", "layer", "of", "a", "deep", "CNN", "(", "as", "in", ")", ",", "a", "number", "of", "recent", "works", "have", "explored", "image", "attention", "models", "for", "VQA", ".", "Zhu", "add", "spatial", "attention", "to", "the", "standard", "LSTM", "model", "for", "pointing", "and", "grounded", "QA", ".", "Andreas", "propose", "a", "compositional", "scheme", "that", "consists", "of", "a", "language", "parser", "and", "a", "number", "of", "neural", "modules", "networks", ".", "The", "language", "parser", "predicts", "which", "neural", "module", "network", "should", "be", "instantiated", "to", "answer", "the", "question", ".", "Some", "other", "works", "perform", "image", "attention", "multiple", "times", "in", "a", "stacked", "manner", ".", "In", ",", "the", "authors", "propose", "a", "stacked", "attention", "network", ",", "which", "runs", "multiple", "hops", "to", "infer", "the", "answer", "progressively", ".", "To", "capture", "fine", "-", "grained", "information", "from", "the", "question", ",", "Xu", "propose", "a", "multi", "-", "hop", "image", "attention", "scheme", ".", "It", "aligns", "words", "to", "image", "patches", "in", "the", "first", "hop", ",", "and", "then", "refers", "to", "the", "entire", "question", "for", "obtaining", "image", "attention", "maps", "in", "the", "second", "hop", ".", "In", ",", "the", "authors", "generate", "image", "regions", "with", "object", "proposals", "and", "then", "select", "the", "regions", "relevant", "to", "the", "question", "and", "answer", "choice", ".", "Xiong", "augments", "dynamic", "memory", "network", "with", "a", "new", "input", "fusion", "module", "and", "retrieves", "an", "answer", "from", "an", "attention", "based", "GRU", ".", "In", "concurrent", "work", ",", "collected", "‘", "human", "attention", "maps", "’", "that", "are", "used", "to", "evaluate", "the", "attention", "maps", "generated", "by", "attention", "models", "for", "VQA", ".", "Note", "that", "all", "of", "these", "approaches", "model", "visual", "attention", "alone", ",", "and", "do", "not", "model", "question", "attention", ".", "Moreover", ",", "model", "attention", "sequentially", ",", "i.e.", ",", "later", "attention", "is", "based", "on", "earlier", "attention", ",", "which", "is", "prone", "to", "error", "propagation", ".", "In", "contrast", ",", "we", "conduct", "co", "-", "attention", "at", "three", "levels", "independently", ".", "Language", "Attention", ".", "Though", "no", "prior", "work", "has", "explored", "question", "attention", "in", "VQA", ",", "there", "are", "some", "related", "works", "in", "natural", "language", "processing", "(", "NLP", ")", "in", "general", "that", "have", "modeled", "language", "attention", ".", "In", "order", "to", "overcome", "difficulty", "in", "translation", "of", "long", "sentences", ",", "Bahdanau", "propose", "RNNSearch", "to", "learn", "an", "alignment", "over", "the", "input", "sentences", ".", "In", ",", "the", "authors", "propose", "an", "attention", "model", "to", "circumvent", "the", "bottleneck", "caused", "by", "fixed", "width", "hidden", "vector", "in", "text", "reading", "and", "comprehension", ".", "A", "more", "fine", "-", "grained", "attention", "mechanism", "is", "proposed", "in", ".", "The", "authors", "employ", "a", "word", "-", "by", "-", "word", "neural", "attention", "mechanism", "to", "reason", "about", "the", "entailment", "in", "two", "sentences", ".", "Also", "focused", "on", "modeling", "sentence", "pairs", ",", "the", "authors", "in", "propose", "an", "attention", "-", "based", "bigram", "CNN", "for", "jointly", "performing", "attention", "between", "two", "CNN", "hierarchies", ".", "In", "their", "work", ",", "three", "attention", "schemes", "are", "proposed", "and", "evaluated", ".", "In", ",", "the", "authors", "propose", "a", "two", "-", "way", "attention", "mechanism", "to", "project", "the", "paired", "inputs", "into", "a", "common", "representation", "space", ".", "section", ":", "Method", "We", "begin", "by", "introducing", "the", "notation", "used", "in", "this", "paper", ".", "To", "ease", "understanding", ",", "our", "full", "model", "is", "described", "in", "parts", ".", "First", ",", "our", "hierarchical", "question", "representation", "is", "described", "in", "Sec", ".", "[", "reference", "]", "and", "the", "proposed", "co", "-", "attention", "mechanism", "is", "then", "described", "in", "Sec", ".", "[", "reference", "]", ".", "Finally", ",", "Sec", ".", "[", "reference", "]", "shows", "how", "to", "recursively", "combine", "the", "attended", "question", "and", "image", "features", "to", "output", "answers", ".", "subsection", ":", "Notation", "Given", "a", "question", "with", "words", ",", "its", "representation", "is", "denoted", "by", ",", "where", "is", "the", "feature", "vector", "for", "the", "-", "th", "word", ".", "We", "denote", ",", "and", "as", "word", "embedding", ",", "phrase", "embedding", "and", "question", "embedding", "at", "position", ",", "respectively", ".", "The", "image", "feature", "is", "denoted", "by", ",", "where", "is", "the", "feature", "vector", "at", "the", "spatial", "location", ".", "The", "co", "-", "attention", "features", "of", "image", "and", "question", "at", "each", "level", "in", "the", "hierarchy", "are", "denoted", "as", "and", "where", ".", "The", "weights", "in", "different", "modules", "/", "layers", "are", "denoted", "with", ",", "with", "appropriate", "sub", "/", "super", "-", "scripts", "as", "necessary", ".", "In", "the", "exposition", "that", "follows", ",", "we", "omit", "the", "bias", "term", "to", "avoid", "notational", "clutter", ".", "subsection", ":", "Question", "Hierarchy", "Given", "the", "1", "-", "hot", "encoding", "of", "the", "question", "words", ",", "we", "first", "embed", "the", "words", "to", "a", "vector", "space", "(", "learnt", "end", "-", "to", "-", "end", ")", "to", "get", ".", "To", "compute", "the", "phrase", "features", ",", "we", "apply", "1", "-", "D", "convolution", "on", "the", "word", "embedding", "vectors", ".", "Concretely", ",", "at", "each", "word", "location", ",", "we", "compute", "the", "inner", "product", "of", "the", "word", "vectors", "with", "filters", "of", "three", "window", "sizes", ":", "unigram", ",", "bigram", "and", "trigram", ".", "For", "the", "-", "th", "word", ",", "the", "convolution", "output", "with", "window", "size", "is", "given", "by", "where", "is", "the", "weight", "parameters", ".", "The", "word", "-", "level", "features", "are", "appropriately", "0", "-", "padded", "before", "feeding", "into", "bigram", "and", "trigram", "convolutions", "to", "maintain", "the", "length", "of", "the", "sequence", "after", "convolution", ".", "Given", "the", "convolution", "result", ",", "we", "then", "apply", "max", "-", "pooling", "across", "different", "n", "-", "grams", "at", "each", "word", "location", "to", "obtain", "phrase", "-", "level", "features", "Our", "pooling", "method", "differs", "from", "those", "used", "in", "previous", "works", "in", "that", "it", "adaptively", "selects", "different", "gram", "features", "at", "each", "time", "step", ",", "while", "preserving", "the", "original", "sequence", "length", "and", "order", ".", "We", "use", "a", "LSTM", "to", "encode", "the", "sequence", "after", "max", "-", "pooling", ".", "The", "corresponding", "question", "-", "level", "feature", "is", "the", "LSTM", "hidden", "vector", "at", "time", ".", "Our", "hierarchical", "representation", "of", "the", "question", "is", "depicted", "in", "Fig", ".", "[", "reference", "]", "(", "a", ")", ".", "subsection", ":", "Co", "-", "Attention", "We", "propose", "two", "co", "-", "attention", "mechanisms", "that", "differ", "in", "the", "order", "in", "which", "image", "and", "question", "attention", "maps", "are", "generated", ".", "The", "first", "mechanism", ",", "which", "we", "call", "parallel", "co", "-", "attention", ",", "generates", "image", "and", "question", "attention", "simultaneously", ".", "The", "second", "mechanism", ",", "which", "we", "call", "alternating", "co", "-", "attention", ",", "sequentially", "alternates", "between", "generating", "image", "and", "question", "attentions", ".", "See", "Fig", ".", "[", "reference", "]", ".", "These", "co", "-", "attention", "mechanisms", "are", "executed", "at", "all", "three", "levels", "of", "the", "question", "hierarchy", ".", "Parallel", "Co", "-", "Attention", ".", "Parallel", "co", "-", "attention", "attends", "to", "the", "image", "and", "question", "simultaneously", ".", "Similar", "to", ",", "we", "connect", "the", "image", "and", "question", "by", "calculating", "the", "similarity", "between", "image", "and", "question", "features", "at", "all", "pairs", "of", "image", "-", "locations", "and", "question", "-", "locations", ".", "Specifically", ",", "given", "an", "image", "feature", "map", ",", "and", "the", "question", "representation", ",", "the", "affinity", "matrix", "is", "calculated", "by", "where", "contains", "the", "weights", ".", "After", "computing", "this", "affinity", "matrix", ",", "one", "possible", "way", "of", "computing", "the", "image", "(", "or", "question", ")", "attention", "is", "to", "simply", "maximize", "out", "the", "affinity", "over", "the", "locations", "of", "other", "modality", ",", "and", ".", "Instead", "of", "choosing", "the", "max", "activation", ",", "we", "find", "that", "performance", "is", "improved", "if", "we", "consider", "this", "affinity", "matrix", "as", "a", "feature", "and", "learn", "to", "predict", "image", "and", "question", "attention", "maps", "via", "the", "following", "where", ",", "are", "the", "weight", "parameters", ".", "and", "are", "the", "attention", "probabilities", "of", "each", "image", "region", "and", "word", "respectively", ".", "The", "affinity", "matrix", "transforms", "question", "attention", "space", "to", "image", "attention", "space", "(", "vice", "versa", "for", ")", ".", "Based", "on", "the", "above", "attention", "weights", ",", "the", "image", "and", "question", "attention", "vectors", "are", "calculated", "as", "the", "weighted", "sum", "of", "the", "image", "features", "and", "question", "features", ",", "i.e.", ",", "The", "parallel", "co", "-", "attention", "is", "done", "at", "each", "level", "in", "the", "hierarchy", ",", "leading", "to", "and", "where", ".", "Alternating", "Co", "-", "Attention", ".", "In", "this", "attention", "mechanism", ",", "we", "sequentially", "alternate", "between", "generating", "image", "and", "question", "attention", ".", "Briefly", ",", "this", "consists", "of", "three", "steps", "(", "marked", "in", "Fig", ".", "[", "reference", "]", "b", ")", ":", "1", ")", "summarize", "the", "question", "into", "a", "single", "vector", ";", "2", ")", "attend", "to", "the", "image", "based", "on", "the", "question", "summary", ";", "3", ")", "attend", "to", "the", "question", "based", "on", "the", "attended", "image", "feature", ".", "Concretely", ",", "we", "define", "an", "attention", "operation", ",", "which", "takes", "the", "image", "(", "or", "question", ")", "features", "and", "attention", "guidance", "derived", "from", "question", "(", "or", "image", ")", "as", "inputs", ",", "and", "outputs", "the", "attended", "image", "(", "or", "question", ")", "vector", ".", "The", "operation", "can", "be", "expressed", "in", "the", "following", "steps", "where", "is", "a", "vector", "with", "all", "elements", "to", "be", "1", ".", "and", "are", "parameters", ".", "is", "the", "attention", "weight", "of", "feature", ".", "The", "alternating", "co", "-", "attention", "process", "is", "illustrated", "in", "Fig", ".", "[", "reference", "]", "(", "b", ")", ".", "At", "the", "first", "step", "of", "alternating", "co", "-", "attention", ",", ",", "and", "is", ";", "At", "the", "second", "step", ",", "where", "is", "the", "image", "features", ",", "and", "the", "guidance", "is", "intermediate", "attended", "question", "feature", "from", "the", "first", "step", ";", "Finally", ",", "we", "use", "the", "attended", "image", "feature", "as", "the", "guidance", "to", "attend", "the", "question", "again", ",", "i.e.", ",", "and", ".", "Similar", "to", "the", "parallel", "co", "-", "attention", ",", "the", "alternating", "co", "-", "attention", "is", "also", "done", "at", "each", "level", "of", "the", "hierarchy", ".", "subsection", ":", "Encoding", "for", "Predicting", "Answers", "Following", ",", "we", "treat", "VQA", "as", "a", "classification", "task", ".", "We", "predict", "the", "answer", "based", "on", "the", "co", "-", "attended", "image", "and", "question", "features", "from", "all", "three", "levels", ".", "We", "use", "a", "multi", "-", "layer", "perceptron", "(", "MLP", ")", "to", "recursively", "encode", "the", "attention", "features", "as", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "b", ")", ".", "where", "and", "are", "the", "weight", "parameters", ".", "is", "the", "concatenation", "operation", "on", "two", "vectors", ".", "is", "the", "probability", "of", "the", "final", "answer", ".", "section", ":", "Experiment", "subsection", ":", "Datasets", "We", "evaluate", "the", "proposed", "model", "on", "two", "datasets", ",", "the", "VQA", "dataset", "and", "the", "COCO", "-", "QA", "dataset", ".", "VQA", "dataset", "is", "the", "largest", "dataset", "for", "this", "problem", ",", "containing", "human", "annotated", "questions", "and", "answers", "on", "Microsoft", "COCO", "dataset", ".", "The", "dataset", "contains", "248", ",", "349", "training", "questions", ",", "121", ",", "512", "validation", "questions", ",", "244", ",", "302", "testing", "questions", ",", "and", "a", "total", "of", "6", ",", "141", ",", "630", "question", "-", "answers", "pairs", ".", "There", "are", "three", "sub", "-", "categories", "according", "to", "answer", "-", "types", "including", "yes", "/", "no", ",", "number", ",", "and", "other", ".", "Each", "question", "has", "10", "free", "-", "response", "answers", ".", "We", "use", "the", "top", "1000", "most", "frequent", "answers", "as", "the", "possible", "outputs", "similar", "to", ".", "This", "set", "of", "answers", "covers", "86.54", "%", "of", "the", "train", "+", "val", "answers", ".", "For", "testing", ",", "we", "train", "our", "model", "on", "VQA", "train", "+", "val", "and", "report", "the", "test", "-", "dev", "and", "test", "-", "standard", "results", "from", "the", "VQA", "evaluation", "server", ".", "We", "use", "the", "evaluation", "protocol", "of", "in", "the", "experiment", ".", "COCO", "-", "QA", "dataset", "is", "automatically", "generated", "from", "captions", "in", "the", "Microsoft", "COCO", "dataset", ".", "There", "are", "78", ",", "736", "train", "questions", "and", "38", ",", "948", "test", "questions", "in", "the", "dataset", ".", "These", "questions", "are", "based", "on", "8", ",", "000", "and", "4", ",", "000", "images", "respectively", ".", "There", "are", "four", "types", "of", "questions", "including", "object", ",", "number", ",", "color", ",", "and", "location", ".", "Each", "type", "takes", ",", ",", ",", "and", "of", "the", "whole", "dataset", ",", "respectively", ".", "All", "answers", "in", "this", "data", "set", "are", "single", "word", ".", "As", "in", ",", "we", "report", "classification", "accuracy", "as", "well", "as", "Wu", "-", "Palmer", "similarity", "(", "WUPS", ")", "in", "Table", "2", ".", "subsection", ":", "Setup", "We", "use", "Torch", "to", "develop", "our", "model", ".", "We", "use", "the", "Rmsprop", "optimizer", "with", "a", "base", "learning", "rate", "of", "4e", "-", "4", ",", "momentum", "0.99", "and", "weight", "-", "decay", "1e", "-", "8", ".", "We", "set", "batch", "size", "to", "be", "300", "and", "train", "for", "up", "to", "256", "epochs", "with", "early", "stopping", "if", "the", "validation", "accuracy", "has", "not", "improved", "in", "the", "last", "5", "epochs", ".", "For", "COCO", "-", "QA", ",", "the", "size", "of", "hidden", "layer", "is", "set", "to", "512", "and", "1024", "for", "VQA", "since", "it", "is", "a", "much", "larger", "dataset", ".", "All", "the", "other", "word", "embedding", "and", "hidden", "layers", "were", "vectors", "of", "size", "512", ".", "We", "apply", "dropout", "with", "probability", "on", "each", "layer", ".", "Following", ",", "we", "rescale", "the", "image", "to", ",", "and", "then", "take", "the", "activation", "from", "the", "last", "pooling", "layer", "of", "VGGNet", "or", "ResNet", "as", "its", "feature", ".", "subsection", ":", "Results", "and", "Analysis", "There", "are", "two", "test", "scenarios", "on", "VQA", ":", "open", "-", "ended", "and", "multiple", "-", "choice", ".", "The", "best", "performing", "method", "deeper", "LSTM", "Q", "+", "norm", "I", "from", "is", "used", "as", "our", "baseline", ".", "For", "open", "-", "ended", "test", "scenario", ",", "we", "compare", "our", "method", "with", "the", "recent", "proposed", "SMem", ",", "SAN", ",", "FDA", "and", "DMN", "+", ".", "For", "multiple", "choice", ",", "we", "compare", "with", "Region", "Sel", ".", "and", "FDA", ".", "We", "compare", "with", "2", "-", "VIS", "+", "BLSTM", ",", "IMG", "-", "CNN", "and", "SAN", "on", "COCO", "-", "QA", ".", "We", "use", "to", "refer", "to", "our", "parallel", "co", "-", "attention", ",", "for", "alternating", "co", "-", "attention", ".", "Table", "[", "reference", "]", "shows", "results", "on", "the", "VQA", "test", "sets", "for", "both", "open", "-", "ended", "and", "multiple", "-", "choice", "settings", ".", "We", "can", "see", "that", "our", "approach", "improves", "the", "state", "of", "art", "from", "60.4", "%", "(", "DMN", "+", ")", "to", "62.1", "%", "(", "+", "ResNet", ")", "on", "open", "-", "ended", "and", "from", "64.2", "%", "(", "FDA", ")", "to", "66.1", "%", "(", "+", "ResNet", ")", "on", "multiple", "-", "choice", ".", "Notably", ",", "for", "the", "question", "type", "Other", "and", "Num", ",", "we", "achieve", "3.4", "%", "and", "1.4", "%", "improvement", "on", "open", "-", "ended", "questions", ",", "and", "4.0", "%", "and", "1.1", "%", "on", "multiple", "-", "choice", "questions", ".", "As", "we", "can", "see", ",", "ResNet", "features", "outperform", "or", "match", "VGG", "features", "in", "all", "cases", ".", "Our", "improvements", "are", "not", "solely", "due", "to", "the", "use", "of", "a", "better", "CNN", ".", "Specifically", ",", "FDA", "also", "uses", "ResNet", ",", "but", "+", "ResNet", "outperforms", "it", "by", "1.8", "%", "on", "test", "-", "dev", ".", "SMem", "uses", "GoogLeNet", "and", "the", "rest", "all", "use", "VGGNet", ",", "and", "Ours", "+", "VGG", "outperforms", "them", "by", "0.2", "%", "on", "test", "-", "dev", "(", "DMN", "+", ")", ".", "Table", "[", "reference", "]", "shows", "results", "on", "the", "COCO", "-", "QA", "test", "set", ".", "Similar", "to", "the", "result", "on", "VQA", ",", "our", "model", "improves", "the", "state", "-", "of", "-", "the", "-", "art", "from", "61.6", "%", "(", "SAN", "(", "2", ",", "CNN", ")", ")", "to", "65.4", "%", "(", "+", "ResNet", ")", ".", "We", "observe", "that", "parallel", "co", "-", "attention", "performs", "better", "than", "alternating", "co", "-", "attention", "in", "this", "setup", ".", "Both", "attention", "mechanisms", "have", "their", "advantages", "and", "disadvantages", ":", "parallel", "co", "-", "attention", "is", "harder", "to", "train", "because", "of", "the", "dot", "product", "between", "image", "and", "text", "which", "compresses", "two", "vectors", "into", "a", "single", "value", ".", "On", "the", "other", "hand", ",", "alternating", "co", "-", "attention", "may", "suffer", "from", "errors", "being", "accumulated", "at", "each", "round", ".", "subsection", ":", "Ablation", "Study", "In", "this", "section", ",", "we", "perform", "ablation", "studies", "to", "quantify", "the", "role", "of", "each", "component", "in", "our", "model", ".", "Specifically", ",", "we", "re", "-", "train", "our", "approach", "by", "ablating", "certain", "components", ":", "Image", "Attention", "alone", ",", "where", "in", "a", "manner", "similar", "to", "previous", "works", ",", "we", "do", "not", "use", "any", "question", "attention", ".", "The", "goal", "of", "this", "comparison", "is", "to", "verify", "that", "our", "improvements", "are", "not", "the", "result", "of", "orthogonal", "contributions", ".", "(", "say", "better", "optimization", "or", "better", "CNN", "features", ")", ".", "Question", "Attention", "alone", ",", "where", "no", "image", "attention", "is", "performed", ".", "W", "/", "O", "Conv", ",", "where", "no", "convolution", "and", "pooling", "is", "performed", "to", "represent", "phrases", ".", "Instead", ",", "we", "stack", "another", "word", "embedding", "layer", "on", "the", "top", "of", "word", "level", "outputs", ".", "W", "/", "O", "W", "-", "Atten", ",", "where", "no", "word", "level", "co", "-", "attention", "is", "performed", ".", "We", "replace", "the", "word", "level", "attention", "with", "a", "uniform", "distribution", ".", "Phrase", "and", "question", "level", "co", "-", "attentions", "are", "still", "modeled", ".", "W", "/", "O", "P", "-", "Atten", ",", "where", "no", "phrase", "level", "co", "-", "attention", "is", "performed", ",", "and", "the", "phrase", "level", "attention", "is", "set", "to", "be", "uniform", ".", "Word", "and", "question", "level", "co", "-", "attentions", "are", "still", "modeled", ".", "W", "/", "O", "Q", "-", "Atten", ",", "where", "no", "question", "level", "co", "-", "attention", "is", "performed", ".", "We", "replace", "the", "question", "level", "attention", "with", "a", "uniform", "distribution", ".", "Word", "and", "phrase", "level", "co", "-", "attentions", "are", "still", "modeled", ".", "Table", "[", "reference", "]", "shows", "the", "comparison", "of", "our", "full", "approach", "w.r.t", "these", "ablations", "on", "the", "VQA", "validation", "set", "(", "test", "sets", "are", "not", "recommended", "to", "be", "used", "for", "such", "experiments", ")", ".", "The", "deeper", "LSTM", "Q", "+", "norm", "I", "baseline", "in", "antol2015vqa", "is", "also", "reported", "for", "comparison", ".", "We", "can", "see", "that", "image", "-", "attention", "-", "alone", "does", "improve", "performance", "over", "the", "holistic", "image", "feature", "(", "deeper", "LSTM", "Q", "+", "norm", "I", ")", ",", "which", "is", "consistent", "with", "findings", "of", "previous", "attention", "models", "for", "VQA", ".", "Ablation", "study", "on", "the", "VQA", "dataset", "using", "Oursa", "+", "VGG", ".", "validationMethodY", "/", "NNumOtherAllLSTM", "Q", "+", "I79.832.940.754.3Image", "Atten79.833.943.655.9Question", "Atten79.433.341.754.8W", "/", "O", "Q", "-", "Atten79.632.142.955.3W", "/", "O", "P", "-", "Atten79.534.145.456.7W", "/", "O", "W", "-", "Atten79.634.445.656.8Full", "Model79.635.045.757.0", "Comparing", "the", "full", "model", "ablated", "versions", "without", "word", ",", "phrase", ",", "question", "level", "attentions", "reveals", "a", "clear", "interesting", "trend", "–", "the", "attention", "mechanisms", "closest", "to", "the", "‘", "top", "’", "of", "the", "hierarchy", "(", "question", ")", "matter", "most", ",", "with", "a", "drop", "of", "1.7", "%", "in", "accuracy", "if", "not", "modeled", ";", "followed", "by", "the", "intermediate", "level", "(", "phrase", ")", ",", "with", "a", "drop", "of", "0.3", "%", ";", "finally", "followed", "by", "the", "‘", "bottom", "’", "of", "the", "hierarchy", "(", "word", ")", ",", "with", "a", "drop", "of", "0.2", "%", "in", "accuracy", ".", "We", "hypothesize", "that", "this", "is", "because", "the", "question", "level", "is", "the", "‘", "closest", "’", "to", "the", "answer", "prediction", "layers", "in", "our", "model", ".", "Note", "that", "all", "levels", "are", "important", ",", "and", "our", "final", "model", "significantly", "outperforms", "not", "using", "any", "linguistic", "attention", "(", "1.1", "%", "difference", "between", "Full", "Model", "and", "Image", "Atten", ")", ".", "The", "question", "attention", "alone", "model", "is", "better", "than", "LSTM", "Q", "+", "I", ",", "with", "an", "improvement", "of", "0.5", "%", "and", "worse", "than", "image", "attention", "alone", ",", "with", "a", "drop", "of", "1.1", "%", ".", "further", "improves", "if", "we", "performed", "alternating", "co", "-", "attention", "for", "one", "more", "round", ",", "with", "an", "improvement", "of", "0.3", "%", ".", "subsection", ":", "Qualitative", "Results", "We", "now", "visualize", "some", "co", "-", "attention", "maps", "generated", "by", "our", "method", "in", "Fig", ".", "[", "reference", "]", ".", "At", "the", "word", "level", ",", "our", "model", "attends", "mostly", "to", "the", "object", "regions", "in", "an", "image", ",", "e.g.", ",", "heads", ",", "bird", ".", "At", "the", "phrase", "level", ",", "the", "image", "attention", "has", "different", "patterns", "across", "images", ".", "For", "the", "first", "two", "images", ",", "the", "attention", "transfers", "from", "objects", "to", "background", "regions", ".", "For", "the", "third", "image", ",", "the", "attention", "becomes", "more", "focused", "on", "the", "objects", ".", "We", "suspect", "that", "this", "is", "caused", "by", "the", "different", "question", "types", ".", "On", "the", "question", "side", ",", "our", "model", "is", "capable", "of", "localizing", "the", "key", "phrases", "in", "the", "question", ",", "thus", "essentially", "discovering", "the", "question", "types", "in", "the", "dataset", ".", "For", "example", ",", "our", "model", "pays", "attention", "to", "the", "phrases", "‘", "‘", "what", "color", "’", "’", "and", "‘", "‘", "how", "many", "snowboarders", "’", "’", ".", "Our", "model", "successfully", "attends", "to", "the", "regions", "in", "images", "and", "phrases", "in", "the", "questions", "appropriate", "for", "answering", "the", "question", ",", "e.g.", ",", "‘", "‘", "color", "of", "the", "bird", "’", "’", "and", "bird", "region", ".", "Because", "our", "model", "performs", "co", "-", "attention", "at", "three", "levels", ",", "it", "often", "captures", "complementary", "information", "from", "each", "level", ",", "and", "then", "combines", "them", "to", "predict", "the", "answer", ".", "Q", ":", "what", "is", "the", "man", "holding", "a", "snowboard", "on", "top", "of", "a", "snow", "covered", "?", "A", ":", "mountain", "what", "is", "the", "man", "holding", "a", "snowboard", "on", "top", "of", "a", "snow", "covered", "what", "is", "the", "man", "holding", "a", "snowboard", "on", "top", "of", "a", "snow", "covered", "?", "what", "is", "the", "man", "holding", "a", "snowboard", "on", "top", "of", "a", "snow", "covered", "?", "Q", ":", "what", "is", "the", "color", "of", "the", "bird", "?", "A", ":", "white", "what", "is", "the", "color", "of", "the", "bird", "?", "what", "is", "the", "color", "of", "the", "bird", "?", "what", "is", "the", "color", "of", "the", "bird", "?", "Q", ":", "how", "many", "snowboarders", "in", "formation", "in", "the", "snow", ",", "four", "is", "sitting", "?", "A", ":", "5", "how", "many", "snowboarders", "in", "formation", "in", "the", "snow", ",", "four", "is", "sitting", "?", "how", "many", "snowboarders", "in", "formation", "in", "the", "snow", ",", "four", "is", "sitting", "?", "how", "many", "snowboarders", "in", "formation", "in", "the", "snow", ",", "four", "is", "sitting", "?", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "proposed", "a", "hierarchical", "co", "-", "attention", "model", "for", "visual", "question", "answering", ".", "Co", "-", "attention", "allows", "our", "model", "to", "attend", "to", "different", "regions", "of", "the", "image", "as", "well", "as", "different", "fragments", "of", "the", "question", ".", "We", "model", "the", "question", "hierarchically", "at", "three", "levels", "to", "capture", "information", "from", "different", "granularities", ".", "The", "ablation", "studies", "further", "demonstrate", "the", "roles", "of", "co", "-", "attention", "and", "question", "hierarchy", "in", "our", "final", "performance", ".", "Through", "visualizations", ",", "we", "can", "see", "that", "our", "model", "co", "-", "attends", "to", "interpretable", "regions", "of", "images", "and", "questions", "for", "predicting", "the", "answer", ".", "Though", "our", "model", "was", "evaluated", "on", "visual", "question", "answering", ",", "it", "can", "be", "potentially", "applied", "to", "other", "tasks", "involving", "vision", "and", "language", ".", "Acknowledgements", "This", "work", "was", "funded", "in", "part", "by", "NSF", "CAREER", "awards", "to", "DP", "and", "DB", ",", "an", "ONR", "YIP", "award", "to", "DP", ",", "ONR", "Grant", "N00014", "-", "14", "-", "1", "-", "0679", "to", "DB", ",", "a", "Sloan", "Fellowship", "to", "DP", ",", "ARO", "YIP", "awards", "to", "DB", "and", "DP", ",", "a", "Allen", "Distinguished", "Investigator", "award", "to", "DP", "from", "the", "Paul", "G.", "Allen", "Family", "Foundation", ",", "ICTAS", "Junior", "Faculty", "awards", "to", "DB", "and", "DP", ",", "Google", "Faculty", "Research", "Awards", "to", "DP", "and", "DB", ",", "AWS", "in", "Education", "Research", "grant", "to", "DB", ",", "and", "NVIDIA", "GPU", "donations", "to", "DB", ".", "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", "U.S.", "Government", "or", "any", "sponsor", ".", "Q", ":", "what", "is", "the", "color", "of", "the", "kitten", "?", "A", ":", "black", "Q", ":", "what", "are", "standing", "in", "tall", "dry", "grass", "look", "at", "the", "tourists", "?", "A", ":", "zebras", "Q", ":", "where", "is", "the", "woman", "while", "her", "baby", "is", "sleeping", "?", "A", ":", "kitchen", "Q", ":", "what", "seating", "area", "is", "on", "the", "right", "?", "A", ":", "park", "Q", ":", "is", "the", "person", "dressed", "properly", "for", "this", "sport", "?", "A", ":", "yes", "what", "is", "the", "color", "of", "the", "kitten", "?", "what", "are", "standing", "in", "tall", "dry", "grass", "look", "at", "the", "tourists", "?", "where", "is", "the", "woman", "while", "her", "baby", "is", "sleeping", "?", "what", "seating", "area", "is", "on", "the", "right", "?", "is", "the", "person", "dressed", "properly", "for", "the", "sport", "?", "what", "is", "the", "color", "of", "the", "kitten", "?", "what", "are", "standing", "in", "tall", "dry", "grass", "look", "at", "the", "tourists", "?", "where", "is", "the", "woman", "while", "her", "baby", "is", "sleeping", "?", "what", "seating", "area", "is", "on", "the", "right", "?", "is", "the", "person", "dressed", "properly", "for", "the", "sport", "?", "what", "is", "the", "color", "of", "the", "kitten", "?", "what", "are", "standing", "in", "tall", "dry", "grass", "look", "at", "the", "tourists", "?", "where", "is", "the", "woman", "while", "her", "baby", "is", "sleeping", "?", "what", "seating", "area", "is", "on", "the", "right", "?", "is", "the", "person", "dressed", "properly", "for", "the", "sport", "?", "Q", ":", "how", "many", "red", "motorcycles", "with", "riders", "in", "protective", "gear", "are", "on", "the", "street", "?", "A", ":", "two", "(", "three", ")", "Q", ":", "what", "is", "the", "color", "of", "the", "bus", "?", "A", ":", "green", "(", "red", ")", "Q", ":", "what", "is", "shown", "in", "different", "places", "?", "A", ":", "hydrant", "(", "toy", ")", "Q", ":", "do", "the", "doors", "open", "in", "or", "out", "?", "A", ":", "open", "(", "in", ")", "Q", ":", "is", "this", "an", "open", "market", "at", "night", "?", "A", ":", "yes", "(", "no", ")", "how", "many", "red", "motorcycles", "with", "riders", "in", "protective", "gear", "are", "on", "the", "street", "?", "what", "is", "the", "color", "of", "the", "bus", "?", "what", "is", "shown", "in", "different", "places", "?", "do", "the", "doors", "open", "in", "or", "out", "?", "is", "this", "an", "open", "market", "at", "night", "?", "how", "many", "red", "motorcycles", "with", "riders", "in", "protective", "gear", "are", "on", "the", "street", "?", "what", "is", "the", "color", "of", "the", "bus", "?", "what", "is", "shown", "in", "different", "places", "?", "do", "the", "doors", "open", "in", "or", "out", "?", "is", "this", "an", "open", "market", "at", "night", "?", "how", "many", "red", "motorcycles", "with", "riders", "in", "protective", "gear", "are", "on", "the", "street", "?", "what", "is", "the", "color", "of", "the", "bus", "?", "what", "is", "shown", "in", "different", "places", "?", "do", "the", "doors", "open", "in", "or", "out", "?", "is", "this", "an", "open", "market", "at", "night", "?", "bibliography", ":", "References"], "entity": [[2, 9, ["Hierarchical", "Question", "-", "Image", "Co", "-", "Attention"], "Method"], [10, 13, ["Visual", "Question", "Answering"], "Task"], [20, 22, ["attention", "models"], "Method"], [23, 26, ["Visual", "Question", "Answering"], "Task"], [27, 28, ["VQA"], "Task"], [31, 36, ["spatial", "maps", "highlighting", "image", "regions"], "Task"], [87, 91, ["co", "-", "attention", "model"], "Method"], [92, 93, ["VQA"], "Task"], [97, 101, ["image", "and", "question", "attention"], "Task"], [118, 122, ["co", "-", "attention", "mechanism"], "Method"], [125, 127, ["hierarchical", "fashion"], "Method"], [130, 136, ["1", "-", "dimensional", "convolution", "neural", "networks"], "Method"], [137, 138, ["CNN"], "Method"], [153, 155, ["VQA", "dataset"], "Material"], [178, 179, ["ResNet"], "Method"], [189, 190, ["VQA"], "Task"], [206, 209, ["Visual", "Question", "Answering"], "Task"], [210, 211, ["VQA"], "Task"], [217, 222, ["multi", "-", "discipline", "research", "problem"], "Task"], [250, 254, ["visual", "attention", "based", "models"], "Method"], [258, 259, ["VQA"], "Task"], [262, 264, ["attention", "mechanism"], "Method"], [282, 284, ["attention", "models"], "Method"], [285, 286, ["VQA"], "Task"], [303, 305, ["visual", "attention"], "Task"], [378, 379, ["machine"], "Method"], [414, 416, ["visual", "attention"], "Task"], [423, 425, ["question", "attention"], "Task"], [432, 437, ["multi", "-", "modal", "attention", "model"], "Method"], [438, 439, ["VQA"], "Task"], [446, 449, ["Co", "-", "Attention"], "Task"], [459, 461, ["visual", "attention"], "Task"], [462, 464, ["question", "attention"], "Task"], [470, 473, ["co", "-", "attention"], "Task"], [482, 484, ["visual", "attention"], "Task"], [502, 504, ["image", "representation"], "Method"], [513, 517, ["question", "representation", "(", "s"], "Method"], [522, 524, ["image", "attention"], "Task"], [531, 533, ["hierarchical", "architecture"], "Method"], [579, 581, ["embedding", "matrix"], "Method"], [587, 593, ["1", "-", "dimensional", "convolution", "neural", "networks"], "Method"], [611, 613, ["word", "representations"], "Method"], [614, 616, ["temporal", "filters"], "Method"], [635, 638, ["phrase", "level", "representation"], "Method"], [646, 649, ["recurrent", "neural", "networks"], "Method"], [660, 662, ["question", "representation"], "Method"], [705, 709, ["co", "-", "attention", "mechanism"], "Method"], [710, 711, ["VQA"], "Task"], [714, 719, ["question", "-", "guided", "visual", "attention"], "Task"], [720, 725, ["image", "-", "guided", "question", "attention"], "Task"], [734, 740, ["parallel", "and", "alternating", "co", "-", "attention"], "Method"], [754, 756, ["hierarchical", "architecture"], "Method"], [764, 771, ["image", "-", "question", "co", "-", "attention", "maps"], "Task"], [803, 805, ["answer", "prediction"], "Task"], [815, 819, ["convolution", "-", "pooling", "strategy"], "Method"], [831, 834, ["question", "level", "representation"], "Method"], [847, 848, ["VQA"], "Task"], [856, 858, ["ablation", "studies"], "Task"], [880, 881, ["VQA"], "Task"], [888, 892, ["co", "-", "attention", "mechanism"], "Method"], [894, 899, ["vision", "and", "language", "attention", "mechanisms"], "Task"], [902, 904, ["Image", "attention"], "Task"], [910, 915, ["holistic", "entire", "-", "image", "embedding"], "Method"], [917, 920, ["fully", "connected", "layer"], "Method"], [922, 924, ["deep", "CNN"], "Method"], [936, 939, ["image", "attention", "models"], "Method"], [940, 941, ["VQA"], "Task"], [949, 951, ["LSTM", "model"], "Method"], [952, 956, ["pointing", "and", "grounded", "QA"], "Task"], [960, 962, ["compositional", "scheme"], "Method"], [966, 968, ["language", "parser"], "Method"], [972, 975, ["neural", "modules", "networks"], "Method"], [977, 979, ["language", "parser"], "Method"], [981, 984, ["neural", "module", "network"], "Method"], [996, 998, ["image", "attention"], "Task"], [1011, 1014, ["stacked", "attention", "network"], "Method"], [1038, 1044, ["multi", "-", "hop", "image", "attention", "scheme"], "Method"], [1065, 1068, ["image", "attention", "maps"], "Task"], [1098, 1101, ["dynamic", "memory", "network"], "Method"], [1104, 1107, ["input", "fusion", "module"], "Method"], [1113, 1116, ["attention", "based", "GRU"], "Method"], [1137, 1139, ["attention", "models"], "Method"], [1140, 1141, ["VQA"], "Task"], [1149, 1151, ["visual", "attention"], "Task"], [1196, 1198, ["Language", "Attention"], "Task"], [1205, 1207, ["question", "attention"], "Task"], [1208, 1209, ["VQA"], "Task"], [1216, 1219, ["natural", "language", "processing"], "Task"], [1220, 1221, ["NLP"], "Task"], [1227, 1229, ["language", "attention"], "Task"], [1236, 1240, ["translation", "of", "long", "sentences"], "Task"], [1243, 1244, ["RNNSearch"], "Method"], [1259, 1261, ["attention", "model"], "Method"], [1272, 1276, ["text", "reading", "and", "comprehension"], "Task"], [1281, 1284, ["grained", "attention", "mechanism"], "Method"], [1292, 1300, ["word", "-", "by", "-", "word", "neural", "attention", "mechanism"], "Method"], [1313, 1315, ["sentence", "pairs"], "Task"], [1321, 1326, ["attention", "-", "based", "bigram", "CNN"], "Method"], [1329, 1330, ["attention"], "Task"], [1332, 1334, ["CNN", "hierarchies"], "Method"], [1340, 1342, ["attention", "schemes"], "Method"], [1355, 1358, ["way", "attention", "mechanism"], "Method"], [1398, 1401, ["hierarchical", "question", "representation"], "Method"], [1412, 1416, ["co", "-", "attention", "mechanism"], "Method"], [1573, 1577, ["1", "-", "hot", "encoding"], "Method"], [1610, 1614, ["1", "-", "D", "convolution"], "Method"], [1637, 1638, ["filters"], "Method"], [1643, 1644, ["unigram"], "Method"], [1645, 1646, ["bigram"], "Method"], [1647, 1648, ["trigram"], "Method"], [1683, 1687, ["bigram", "and", "trigram", "convolutions"], "Method"], [1705, 1708, ["max", "-", "pooling"], "Method"], [1719, 1723, ["phrase", "-", "level", "features"], "Method"], [1724, 1726, ["pooling", "method"], "Method"], [1758, 1759, ["LSTM"], "Method"], [1764, 1767, ["max", "-", "pooling"], "Method"], [1783, 1785, ["hierarchical", "representation"], "Method"], [1808, 1812, ["co", "-", "attention", "mechanisms"], "Method"], [1834, 1838, ["parallel", "co", "-", "attention"], "Task"], [1840, 1844, ["image", "and", "question", "attention"], "Task"], [1853, 1857, ["alternating", "co", "-", "attention"], "Task"], [1861, 1866, ["generating", "image", "and", "question", "attentions"], "Task"], [1875, 1879, ["co", "-", "attention", "mechanisms"], "Method"], [1890, 1894, ["Parallel", "Co", "-", "Attention"], "Task"], [1895, 1899, ["Parallel", "co", "-", "attention"], "Task"], [1947, 1949, ["question", "representation"], "Method"], [1973, 1979, ["image", "(", "or", "question", ")", "attention"], "Task"], [2021, 2026, ["image", "and", "question", "attention", "maps"], "Task"], [2096, 2100, ["parallel", "co", "-", "attention"], "Task"], [2114, 2118, ["Alternating", "Co", "-", "Attention"], "Task"], [2121, 2123, ["attention", "mechanism"], "Method"], [2128, 2133, ["generating", "image", "and", "question", "attention"], "Task"], [2192, 2194, ["attention", "operation"], "Method"], [2260, 2265, ["alternating", "co", "-", "attention", "process"], "Method"], [2282, 2286, ["alternating", "co", "-", "attention"], "Task"], [2339, 2343, ["parallel", "co", "-", "attention"], "Task"], [2345, 2349, ["alternating", "co", "-", "attention"], "Task"], [2363, 2365, ["Predicting", "Answers"], "Task"], [2369, 2370, ["VQA"], "Task"], [2372, 2374, ["classification", "task"], "Task"], [2397, 2403, ["multi", "-", "layer", "perceptron", "(", "MLP"], "Method"], [2461, 2463, ["VQA", "dataset"], "Material"], [2470, 2472, ["VQA", "dataset"], "Material"], [2487, 2490, ["Microsoft", "COCO", "dataset"], "Material"], [2593, 2597, ["VQA", "train", "+", "val"], "Material"], [2600, 2603, ["test", "-", "dev"], "Metric"], [2610, 2613, ["VQA", "evaluation", "server"], "Task"], [2635, 2638, ["Microsoft", "COCO", "dataset"], "Material"], [2716, 2718, ["classification", "accuracy"], "Metric"], [2721, 2725, ["Wu", "-", "Palmer", "similarity"], "Metric"], [2726, 2727, ["WUPS"], "Metric"], [2737, 2738, ["Torch"], "Method"], [2746, 2748, ["Rmsprop", "optimizer"], "Method"], [2750, 2753, ["base", "learning", "rate"], "Method"], [2761, 2764, ["weight", "-", "decay"], "Method"], [2787, 2789, ["validation", "accuracy"], "Metric"], [2815, 2816, ["VQA"], "Task"], [2827, 2829, ["word", "embedding"], "Method"], [2830, 2832, ["hidden", "layers"], "Method"], [2840, 2841, ["dropout"], "Method"], [2863, 2865, ["pooling", "layer"], "Method"], [2866, 2867, ["VGGNet"], "Method"], [2868, 2869, ["ResNet"], "Method"], [2884, 2885, ["VQA"], "Task"], [2886, 2889, ["open", "-", "ended"], "Material"], [2890, 2893, ["multiple", "-", "choice"], "Material"], [2899, 2904, ["LSTM", "Q", "+", "norm", "I"], "Method"], [2926, 2927, ["SMem"], "Method"], [2928, 2929, ["SAN"], "Method"], [2930, 2931, ["FDA"], "Method"], [2932, 2934, ["DMN", "+"], "Method"], [2936, 2938, ["multiple", "choice"], "Task"], [2942, 2944, ["Region", "Sel"], "Method"], [2951, 2956, ["2", "-", "VIS", "+", "BLSTM"], "Method"], [2957, 2960, ["IMG", "-", "CNN"], "Method"], [2961, 2962, ["SAN"], "Method"], [2973, 2977, ["parallel", "co", "-", "attention"], "Method"], [2979, 2983, ["alternating", "co", "-", "attention"], "Task"], [2992, 2995, ["VQA", "test", "sets"], "Material"], [3021, 3023, ["DMN", "+"], "Method"], [3029, 3030, ["ResNet"], "Method"], [3032, 3035, ["open", "-", "ended"], "Material"], [3040, 3041, ["FDA"], "Method"], [3047, 3048, ["ResNet"], "Method"], [3050, 3053, ["multiple", "-", "choice"], "Material"], [3073, 3077, ["open", "-", "ended", "questions"], "Material"], [3085, 3089, ["multiple", "-", "choice", "questions"], "Material"], [3095, 3096, ["ResNet"], "Method"], [3118, 3119, ["CNN"], "Method"], [3122, 3123, ["FDA"], "Method"], [3125, 3126, ["ResNet"], "Method"], [3129, 3130, ["ResNet"], "Method"], [3140, 3141, ["SMem"], "Method"], [3142, 3143, ["GoogLeNet"], "Method"], [3148, 3149, ["VGGNet"], "Method"], [3164, 3166, ["DMN", "+"], "Method"], [3187, 3188, ["VQA"], "Task"], [3204, 3209, ["SAN", "(", "2", ",", "CNN"], "Method"], [3216, 3217, ["ResNet"], "Method"], [3222, 3226, ["parallel", "co", "-", "attention"], "Method"], [3229, 3233, ["alternating", "co", "-", "attention"], "Method"], [3238, 3240, ["attention", "mechanisms"], "Method"], [3246, 3250, ["parallel", "co", "-", "attention"], "Task"], [3277, 3281, ["alternating", "co", "-", "attention"], "Task"], [3293, 3295, ["Ablation", "Study"], "Task"], [3301, 3303, ["ablation", "studies"], "Task"], [3327, 3329, ["Image", "Attention"], "Task"], [3370, 3371, ["optimization"], "Task"], [3377, 3379, ["Question", "Attention"], "Task"], [3388, 3392, ["W", "/", "O", "Conv"], "Method"], [3395, 3398, ["convolution", "and", "pooling"], "Method"], [3409, 3412, ["word", "embedding", "layer"], "Method"], [3445, 3447, ["uniform", "distribution"], "Method"], [3553, 3556, ["VQA", "validation", "set"], "Material"], [3571, 3578, ["deeper", "LSTM", "Q", "+", "norm", "I", "baseline"], "Method"], [3590, 3593, ["image", "-", "attention"], "Method"], [3605, 3610, ["LSTM", "Q", "+", "norm", "I"], "Method"], [3619, 3621, ["attention", "models"], "Method"], [3622, 3623, ["VQA"], "Task"], [3628, 3630, ["VQA", "dataset"], "Material"], [3631, 3634, ["Oursa", "+", "VGG"], "Method"], [3680, 3682, ["attention", "mechanisms"], "Method"], [3704, 3705, ["accuracy"], "Metric"], [3746, 3747, ["accuracy"], "Metric"], [3794, 3796, ["Full", "Model"], "Method"], [3797, 3799, ["Image", "Atten"], "Method"], [3802, 3806, ["question", "attention", "alone", "model"], "Method"], [3809, 3813, ["LSTM", "Q", "+", "I"], "Method"], [3823, 3826, ["image", "attention", "alone"], "Method"], [3839, 3843, ["alternating", "co", "-", "attention"], "Method"], [3907, 3909, ["image", "attention"], "Task"], [4236, 4241, ["hierarchical", "co", "-", "attention", "model"], "Method"], [4242, 4245, ["visual", "question", "answering"], "Task"], [4246, 4249, ["Co", "-", "attention"], "Method"], [4333, 4336, ["visual", "question", "answering"], "Task"], [4346, 4347, ["vision"], "Task"], [4348, 4349, ["language"], "Task"], [2912, 2915, ["open", "-", "ended"], "Material"], [2997, 3000, ["open", "-", "ended"], "Material"], [3001, 3004, ["multiple", "-", "choice"], "Material"]]} {"doc_id": "097851f362667a599208f4a06fa289dac2e91fbe", "text": ["We", "propose", "a", "meta", "-", "parameter", "free", ",", "off", "-", "the", "-", "shelf", ",", "simple", "and", "fast", "unsupervised", "feature", "learning", "algorithm", ",", "which", "exploits", "a", "new", "way", "of", "optimizing", "for", "sparsity", ".", "Experiments", "on", "STL", "-", "10", "show", "that", "the", "method", "presents", "state", "-", "of", "-", "the", "-", "art", "performance", "and", "provides", "discriminative", "features", "that", "generalize", "well", ".", "Nomoremeta", "-", "parametertuninginunsupervisedsparsefeaturelearning", "section", ":", "Introduction", "Significant", "effort", "has", "been", "devoted", "to", "handcraft", "appropriate", "feature", "representations", "of", "data", "in", "several", "fields", ".", "In", "tasks", "such", "as", "image", "classification", "and", "object", "recognition", ",", "unsupervised", "learned", "features", "have", "shown", "to", "compete", "well", "or", "even", "outperform", "manually", "designed", "ones", ".", "Unsupervised", "feature", "learning", "has", "also", "shown", "to", "be", "helpful", "in", "greedy", "layerwise", "pre", "-", "training", "of", "deep", "architectures", ".", "In", ",", "the", "author", "claims", "that", "potentially", "interesting", "research", "involves", "pre", "-", "training", "algorithms", ",", "which", "“", "[", "…", "]", "would", "be", "proficient", "at", "extracting", "good", "features", "but", "involving", "an", "easier", "optimization", "problem", ".", "”", "In", "addition", "to", "that", ",", "one", "of", "the", "main", "criticisms", "to", "state", "-", "of", "-", "the", "-", "art", "methods", "is", "that", "they", "require", "a", "significant", "amount", "of", "meta", "-", "parameters", ".", "As", "stated", "in", ",", "the", "tuning", "of", "these", "meta", "-", "parameters", "is", "a", "laborious", "task", "that", "requires", "expert", "knowledge", ",", "rules", "of", "thumb", "or", "extensive", "search", "and", ",", "whose", "setting", "can", "vary", "for", "different", "tasks", ".", "Therefore", ",", "there", "is", "great", "interest", "for", "meta", "-", "parameter", "free", "methods", "and", "automatic", "approaches", "to", "optimize", "the", "performance", "of", "learning", "algorithms", ".", "Nevertheless", ",", "little", "effort", "has", "been", "devoted", "to", "address", "this", "problem", "(", "see", "Table", "[", "reference", "]", "for", "a", "comparison", "of", "meta", "-", "parameters", "required", "by", "unsupervised", "feature", "learning", "methods", ")", ".", "To", "the", "best", "of", "our", "knowledge", ",", "work", "in", "this", "direction", "includes", "ICA", "and", "sparse", "filtering", ".", "Although", "ICA", "provides", "good", "results", "at", "object", "recognition", "tasks", ",", "the", "method", "scales", "poorly", "to", "large", "datasets", "and", "high", "input", "dimensionality", ".", "Computational", "complexity", "is", "also", "a", "major", "drawback", "of", "many", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "ICA", "requires", "an", "expensive", "orthogonalization", "to", "be", "computed", "at", "each", "iteration", ".", "Sparse", "coding", "has", "an", "expensive", "inference", ",", "which", "requires", "a", "prohibitive", "iterative", "optimization", ".", "Significant", "amount", "of", "work", "has", "been", "done", "in", "order", "to", "overcome", "this", "limitation", ".", "Predictive", "Sparse", "Decomposition", "(", "PSD", ")", "is", "a", "successful", "variant", "of", "sparse", "coding", ",", "which", "uses", "a", "predictor", "to", "approximate", "the", "sparse", "representation", "and", "solves", "the", "sparse", "coding", "computationally", "expensive", "encoding", "step", ".", "In", "this", "paper", ",", "we", "aim", "to", "solve", "some", "of", "the", "above", "-", "mentioned", "problems", ".", "We", "propose", "a", "meta", "-", "parameter", "free", ",", "off", "-", "the", "-", "shelf", ",", "simple", "and", "fast", "approach", ",", "which", "exploits", "a", "new", "way", "of", "optimizing", "for", "a", "sparsity", ",", "without", "explicitly", "modeling", "the", "data", "distribution", ".", "The", "method", "iteratively", "builds", "an", "ideally", "sparse", "target", "and", "optimizes", "the", "dictionary", "by", "minimizing", "the", "error", "between", "the", "system", "output", "and", "the", "ideally", "sparse", "target", ".", "Defining", "sparsity", "concepts", "in", "terms", "of", "expected", "output", "allows", "to", "exploit", "a", "new", "strategy", "in", "unsupervised", "training", ".", "It", "is", "worth", "stressing", "that", "many", "optimization", "strategies", "can", "be", "used", "to", "minimize", "the", "above", "-", "mentioned", "error", "and", "that", "parameters", "of", "these", "optimization", "techniques", "must", "not", "be", "considered", "as", "belonging", "to", "our", "approach", ".", "Experiments", "on", "STL", "-", "10", "dataset", "show", "that", "the", "method", "outperforms", "state", "-", "of", "-", "the", "-", "art", "methods", "in", "single", "layer", "image", "classification", ",", "providing", "discriminative", "features", "that", "generalize", "well", ".", "Linear", "feature", "extraction", "methods", "combined", "with", "sparse", "coding", "encodings", "are", "among", "best", "performers", "on", "object", "recognition", "datasets", ".", "The", "importance", "of", "properly", "combining", "training", "/", "encoding", "and", "encoding", "/", "pooling", "strategies", "has", "been", "argued", "in", "and", "respectively", ".", "Since", "the", "goal", "of", "this", "paper", "is", "to", "propose", "a", "new", "method", "for", "unsupervised", "feature", "learning", ",", "dealing", "with", "all", "the", "possible", "combinations", "of", "encoding", "and", "pooling", "could", "mask", "the", "benefits", "of", "the", "method", "that", "we", "propose", ".", "However", ",", "for", "the", "sake", "of", "fair", "comparison", "with", "the", "state", "-", "of", "-", "the", "-", "art", ",", "we", "test", "the", "method", "with", "sparse", "coding", "and", "soft", "-", "threshold", "encodings", "combined", "with", "sum", "pooling", ",", "following", "the", "experimental", "pipeline", "of", ".", "section", ":", "State", "-", "of", "-", "the", "-", "art", "Commonly", "used", "algorithms", "for", "unsupervised", "feature", "learning", "include", "Restricted", "Boltzmann", "Machines", "(", "RBM", ")", ",", "auto", "-", "encoders", ",", "sparse", "coding", "and", "hybrids", "such", "as", "PSD", ".", "Many", "other", "methods", "such", "as", "ICA", ",", "Reconstruction", "ICA", "(", "RICA", ")", ",", "Sparse", "Filtering", "and", "methods", "related", "to", "vector", "quantization", "such", "as", "Orthogonal", "Matching", "Pursuit", "(", "OMP", "-", "k", ")", "have", "also", "been", "used", "in", "the", "literature", "to", "extract", "unsupervised", "feature", "representations", ".", "These", "algorithms", "could", "be", "divided", "into", "two", "categories", ":", "explicitly", "modeling", "or", "not", "the", "input", "distribution", ".", "Sparse", "auto", "-", "encoders", ",", "sparse", "RBM", ",", "sparse", "coding", ",", "PSD", ",", "OMP", "-", "k", "and", "Reconstruction", "ICA", "(", "RICA", ")", "explicitly", "model", "the", "data", "distribution", "by", "minimizing", "the", "reconstruction", "error", ".", "Although", "learning", "a", "good", "approximation", "of", "the", "data", "distribution", "may", "be", "desirable", ",", "approaches", "such", "as", "sparse", "filtering", "show", "that", "this", "seems", "not", "so", "important", "if", "the", "goal", "is", "to", "have", "a", "discriminative", "sparse", "system", ".", "Sparse", "filtering", "does", "not", "attempt", "to", "explicitly", "model", "the", "input", "distribution", "but", "focuses", "on", "the", "properties", "of", "the", "output", "distribution", "instead", ".", "Sparsity", "is", "among", "the", "desirable", "properties", "of", "a", "good", "output", "representation", ".", "Sparse", "features", "consist", "of", "a", "large", "amount", "of", "outputs", ",", "which", "respond", "rarely", "and", "provide", "high", "responses", "when", "they", "do", "respond", ".", "Sparsity", "can", "be", "described", "in", "terms", "of", "population", "sparsity", "and", "lifetime", "sparsity", ".", "Both", "lifetime", "and", "population", "sparsity", "are", "important", "properties", "of", "the", "output", "distribution", ".", "On", "one", "hand", ",", "lifetime", "sparsity", "plays", "an", "important", "role", "in", "preventing", "bad", "solutions", "such", "as", "numerous", "dead", "outputs", ".", "There", "seems", "to", "be", "a", "consensus", "to", "overcome", "such", "degenerate", "solutions", ",", "which", "is", "to", "ensure", "similar", "statistics", "among", "outputs", ".", "On", "the", "other", "hand", ",", "population", "sparsity", "helps", "providing", "a", "simple", "interpretation", "of", "the", "input", "data", "such", "as", "the", "ones", "found", "in", "early", "visual", "areas", ".", "To", "the", "best", "of", "our", "knowledge", ",", "the", "definition", "of", "population", "sparsity", "remains", "ambiguous", ".", "State", "-", "of", "-", "the", "-", "art", "methods", "optimize", "either", "for", "one", "or", "both", "sparsity", "forms", "in", "their", "objective", "function", ".", "The", "great", "majority", "seeks", "sparsity", "using", "the", "penalty", "and", "does", "not", "optimize", "for", "an", "explicit", "level", "of", "sparsity", "in", "their", "outputs", ".", "Sparse", "auto", "-", "encoders", "optimize", "for", "a", "target", "activation", "allowing", "to", "deal", "with", "lifetime", "sparsity", ";", "nevertheless", ",", "the", "target", "activation", "requires", "tuning", "and", "does", "not", "explicitly", "control", "the", "level", "of", "population", "sparsity", ".", "OMP", "-", "k", "defines", "the", "level", "of", "population", "sparsity", "by", "setting", "to", "the", "maximum", "expected", "number", "of", "non", "-", "zero", "elements", "per", "output", "code", ",", "whereas", "the", "methods", "in", "do", "not", "explicitly", "define", "the", "proportion", "of", "outputs", "expected", "to", "be", "active", "at", "the", "same", "time", ".", "section", ":", "Method", "In", "this", "section", ",", "we", "describe", "how", "the", "proposed", "method", "learns", "a", "sparse", "feature", "representation", "of", "the", "data", "in", "terms", "of", "population", "and", "lifetime", "sparsity", ".", "The", "method", "iteratively", "builds", "an", "ideally", "sparse", "target", "and", "optimizes", "the", "dictionary", "by", "minimizing", "the", "error", "between", "the", "system", "output", "and", "the", "ideally", "sparse", "target", ".", "Subsection", "[", "reference", "]", "highlights", "the", "algorithm", "to", "enforce", "lifetime", "and", "population", "sparsity", "in", "the", "ideally", "sparse", "target", ".", "Subsection", "[", "reference", "]", "provides", "implementation", "details", "on", "the", "system", "and", "optimization", "strategies", "used", "to", "minimize", "the", "error", "between", "the", "system", "output", "and", "the", "ideally", "sparse", "target", ".", "subsection", ":", "Enforcing", "Population", "and", "Lifetime", "Sparsity", "by", "defining", "an", "ideal", "target", "We", "define", "population", "and", "lifetime", "sparsity", "as", "properties", "of", "an", "ideal", "sparse", "output", ".", "Given", "training", "samples", "and", "an", "output", "of", "dimensionality", ",", "we", "define", "the", "first", "property", "of", "the", "output", "as", ":", "Strong", "Lifetime", "Sparsity", ":", "The", "output", "vectors", "must", "be", "composed", "solely", "of", "active", "and", "inactive", "units", "(", "no", "intermediate", "values", "between", "two", "fixed", "scalars", "are", "allowed", ")", "and", "all", "outputs", "must", "activate", "for", "an", "equal", "number", "of", "inputs", ".", "Activation", "is", "exactly", "distributed", "among", "the", "outputs", ".", "Our", "Strong", "Lifetime", "Sparsity", "definition", "is", "a", "more", "strict", "requirement", "than", "the", "high", "dispersal", "concept", "introduced", "in", ",", "since", "they", "only", "require", "that", "“", "the", "mean", "squared", "activations", "of", "each", "feature", "(", "output", ")", "[", "…", "]", "should", "be", "roughly", "the", "same", "for", "all", "features", "(", "outputs", ")", "”", ".", "While", "high", "dispersal", "attempts", "to", "diversify", "the", "learned", "bases", ",", "it", "does", "not", "guarantee", "the", "output", "distribution", ",", "in", "the", "lifetime", "sense", ",", "to", "be", "composed", "of", "only", "a", "few", "activations", ".", "Furthermore", ",", "our", "definition", "ensures", "the", "absence", "of", "dead", "outputs", ".", "Given", "our", "definition", "of", "Strong", "Lifetime", "Sparsity", ",", "the", "population", "sparsity", "must", "require", "that", ",", "for", "each", "training", "sample", ",", "only", "one", "output", "element", "is", "active", ":", "Strong", "Population", "Sparsity", ":", "For", "each", "training", "sample", "only", "one", "output", "must", "be", "active", ".", "The", "rationale", "of", "our", "approach", "is", "to", "appropriately", "generate", "an", "ideal", "output", "target", "that", "fulfils", "properties", "(", "1", ")", "and", "(", "2", ")", ",", "and", "then", "learn", "the", "parameters", "of", "the", "system", "by", "minimizing", "the", "error", "between", "the", "output", "target", "and", "the", "output", "generated", "by", "the", "system", "during", "training", ".", "In", "this", "way", ",", "we", "seek", "a", "system", "optimized", "for", "both", "population", "and", "lifetime", "sparsity", "in", "an", "explicit", "way", ".", "The", "key", "component", "of", "our", "approach", "is", "how", "to", "define", "the", "ideal", "output", "target", "based", "on", "the", "above", "-", "mentioned", "properties", ".", "However", ",", "to", "ensure", "that", "the", "optimization", "of", "the", "system", "parameters", "converges", ",", "we", "add", "a", "third", "property", ":", "Minimal", "Perturbation", ":", "The", "ideal", "output", "target", "should", "be", "defined", "as", "the", "best", "approximation", "of", "the", "system", "output", "by", "means", "of", "error", "fulfilling", "properties", "(", "1", ")", "&", "(", "2", ")", ".", "Creating", "the", "output", "target", "that", "ensures", "the", "above", "-", "mentioned", "properties", "is", "analogous", "to", "solving", "an", "assignment", "problem", ".", "The", "Hungarian", "method", "is", "a", "combinatorial", "optimization", "algorithm", ",", "which", "solves", "the", "assignment", "problem", ".", "However", ",", "its", "computational", "cost", "is", "prohibitive", ".", "Therefore", ",", "in", "the", "next", "section", "we", "propose", "a", "simple", "and", "fast", "algorithm", "to", "generate", "the", "ideal", "output", "target", ",", "which", "ensures", "sparsity", "properties", "(", "1", ")", "and", "(", "2", ")", "and", "provides", "an", "approximate", "solution", "for", "minimal", "perturbation", "property", "(", "3", ")", ".", "subsubsection", ":", "Ideal", "target", "generation", ":", "the", "Enforcing", "Population", "and", "Lifetime", "Sparsity", "(", "EPLS", ")", "algorithm", "Let", "us", "assume", "that", "we", "have", "a", "system", ",", "which", "produces", "a", "row", "output", "vector", ".", "We", "use", "the", "notation", "to", "refer", "to", "one", "element", "of", ".", "We", "define", "an", "output", "matrix", "composed", "of", "output", "vectors", "of", "dimensionality", ",", "such", "that", ".", "Likewise", ",", "we", "define", "an", "ideal", "target", "output", "matrix", "of", "the", "same", "size", ".", "Algorithm", "[", "reference", "]", "details", "the", "EPLS", "algorithm", "to", "generate", "the", "ideal", "target", "from", ".", "For", "the", "sake", "of", "simplicity", ",", "every", "step", "of", "the", "algorithm", "where", "the", "subscript", "appears", "must", "be", "applied", ".", "EPLS", "[", "1", "]", ",", "a", ",", "N", "a", "T", "to", "active", "/", "inactive", "values", "of", "the", "corresponding", "function", ".", "Starting", "with", "no", "activation", "in", "(", "line", "1", ")", ",", "the", "algorithm", "proceeds", "as", "follows", ".", "A", "row", "vector", "from", "is", "processed", "at", "each", "iteration", "(", "line", "3", ")", ".", "The", "crucial", "step", "is", "performed", "in", "line", "4", ":", "the", "output", "that", "has", "to", "be", "activated", "in", "the", "row", "of", "is", "selected", "as", "the", "one", "that", "has", "the", "maximal", "activation", "value", "minus", "the", "inhibitor", ".", "The", "inhibitor", "can", "be", "seen", "as", "an", "accumulator", "that", "“", "counts", "”", "the", "number", "of", "times", "an", "output", "has", "been", "selected", ",", "increasing", "its", "inhibition", "progressively", "by", "until", "reaching", "maximal", "inhibition", ".", "This", "prevents", "the", "selection", "of", "an", "output", "that", "has", "already", "been", "activated", "times", ".", "The", "rationale", "behind", "the", "equation", "in", "line", "4", "is", "that", ",", "while", "selecting", "the", "maximal", "responses", "in", "the", "matrix", ",", "we", "have", "to", "take", "care", "to", "distribute", "them", "evenly", "among", "all", "outputs", "(", "in", "order", "to", "ensure", "Strong", "Lifetime", "Sparsity", ")", ".", "Using", "this", "strategy", ",", "it", "can", "be", "demonstrated", "that", "the", "resulting", "matrix", "perfectly", "fulfills", "properties", "(", "1", ")", "and", "(", "2", ")", ".", "In", "line", "5", ",", "the", "algorithm", "activates", "the", "element", "of", "row", "of", "the", "target", "matrix", ".", "By", "activating", "the", "“", "relative", "”", "maximum", ",", "we", "approximate", "property", "(", "3", ")", ".", "Finally", ",", "the", "inhibitor", "is", "updated", "in", "line", "6", ".", "subsection", ":", "System", "and", "Optimization", "strategies", "Let", "us", "assume", "that", "we", "have", "a", "system", "parameterized", "by", ",", "with", "activation", "function", ",", "which", "takes", "as", "input", "a", "data", "vector", "and", "produces", "an", "output", "vector", ".", "We", "use", "the", "same", "notation", "as", "in", "Section", "[", "reference", "]", "and", "define", "a", "data", "matrix", "composed", "of", "rows", "and", "columns", ",", "where", "is", "the", "input", "dimensionality", ".", "To", "compare", "our", "training", "strategy", "to", "previous", "well", "known", "systems", ",", "we", "tested", "our", "algorithm", "using", "where", "is", "a", "logistic", "non", "-", "linearity", ".", "subsubsection", ":", "Optimization", "strategy", "The", "system", "might", "be", "trained", "by", "means", "of", "an", "off", "-", "the", "-", "shelf", "mini", "-", "batch", "Stochastic", "Gradient", "Descent", "(", "SGD", ")", "method", "with", "adaptive", "learning", "rates", "such", "as", "variance", "-", "based", "SGD", "(", "vSGD", ")", ".", "Algorithm", "[", "reference", "]", "details", "the", "latter", "training", "process", ".", "The", "mini", "-", "batch", "size", "can", "be", "set", "to", "any", "value", ",", "in", "all", "the", "experiments", "we", "have", "set", ".", "Starting", "with", "set", "to", "small", "random", "numbers", "as", "in", "(", "line", "1", ")", ",", "at", "each", "epoch", "we", "shuffle", "the", "samples", "of", "the", "training", "set", "(", "line", "3", ")", ",", "reset", "the", "EPLS", "inhibitor", "to", "a", "flat", "activation", "(", "line", "4", ")", "and", "process", "all", "mini", "-", "batches", ".", "For", "each", "mini", "-", "batch", ",", "samples", "are", "selected", "(", "line", "6", ")", ".", "Then", ",", "the", "output", "is", "computed", "(", "line", "7", ")", "and", "the", "EPLS", "is", "invoked", "to", "compute", "and", "update", "(", "line", "8", ")", ".", "After", "that", ",", "the", "gradient", "of", "the", "error", "is", "computed", "(", "line", "9", ")", "and", "the", "learning", "rate", "is", "estimated", "as", "in", "(", "line", "10", ")", ".", "The", "system", "parameters", "are", "then", "updated", "to", "minimize", "the", "error", "(", "line", "11", ")", ".", "Finally", ",", "the", "bases", "in", "are", "limited", "to", "have", "unit", "norm", "to", "avoid", "degenerate", "solutions", "(", "line", "13", ")", ".", "This", "procedure", "is", "repeated", "until", "a", "stop", "condition", "is", "met", ";", "in", "our", "experiments", ",", "the", "training", "stops", "when", "the", "relative", "decrement", "error", "between", "epochs", "is", "small", "(", ")", ".", "When", "updating", "the", "system", "parameters", ",", "we", "assume", "that", "does", "not", "depend", "on", ",", "thus", ";", "we", "carried", "out", "experiments", "that", "show", "that", "this", "approximation", "does", "not", "significantly", "influence", "the", "gradient", "descent", "convergence", "nor", "the", "quality", "of", "the", "minimization", ".", "Moreover", ",", "this", "assumption", "makes", "the", "algorithm", "faster", ",", "since", "we", "remove", "the", "need", "of", "computing", "the", "numerical", "partial", "derivatives", "of", ".", "The", "mini", "-", "batch", "vSGD", "allows", "to", "scale", "the", "algorithm", "easily", ",", "especially", "with", "respect", "to", "the", "number", "of", "samples", ".", "Standard", "EPLS", "training", "[", "1", "]", "small", "random", "values", "D", "randomly", "flat", "activation", "mini", "-", "batch", "samples", "D", "(", "b", ")", "learning", "rate", "η", "as", "in", "(", ")", "the", "bases", "W", "in", "Γ", "to", "have", "unit", "norm", "condition", "verified", "section", ":", "Experiments", "The", "performance", "of", "training", "and", "encoding", "strategies", "in", "single", "layer", "networks", "has", "been", "extensively", "analyzed", "in", "the", "literature", "on", "STL", "-", "10", "dataset", ".", "STL", "-", "10", "dataset", "consists", "of", "96x96", "pixels", "color", "images", "belonging", "to", "10", "different", "classes", ".", "The", "dataset", "is", "divided", "into", "a", "large", "unlabeled", "training", "set", "containing", "100", "K", "images", "and", "smaller", "labeled", "training", "and", "test", "sets", ",", "containing", "5000", "and", "8000", "images", ",", "respectively", ".", "It", "has", "to", "be", "considered", "that", "in", "STL", "-", "10", ",", "the", "primary", "challenge", "is", "to", "make", "use", "of", "the", "unlabeled", "data", "(", "100", "K", "images", ")", ",", "which", "is", "100", "times", "bigger", "than", "the", "labeled", "data", "used", "to", "train", "the", "classifier", "(", "1000", "images", "per", "fold", ")", ".", "In", "this", "case", ",", "the", "supervised", "training", "must", "strongly", "rely", "on", "the", "ability", "of", "the", "unsupervised", "method", "to", "learn", "discriminative", "features", ".", "Moreover", ",", "since", "the", "unlabeled", "dataset", "contains", "other", "types", "of", "animals", "(", "bears", ",", "rabbits", ",", "etc", ".", ")", "and", "vehicles", "(", "trains", ",", "buses", ",", "etc", ".", ")", "in", "addition", "to", "the", "ones", "in", "the", "labeled", "set", ",", "the", "unsupervised", "method", "should", "be", "able", "to", "generalize", "well", ".", "To", "validate", "our", "method", ",", "we", "follow", "the", "experimental", "pipeline", "of", ".", "We", "first", "extract", "random", "patches", "and", "normalize", "them", "for", "local", "brightness", "and", "contrast", ".", "Note", "that", "EPLS", "does", "not", "require", "any", "whitening", "of", "the", "input", "data", ",", "since", "it", "decorrelates", "the", "data", "during", "the", "training", "by", "means", "of", "the", "imposed", "strong", "sparsity", "properties", "of", "the", "output", "target", ".", "Then", ",", "we", "apply", "the", "system", "to", "retrieve", "sparse", "features", "of", "patches", "covering", "the", "input", "image", ",", "pool", "them", "into", "4", "quadrants", "and", "finally", "train", "a", "SVM", "for", "classification", "purposes", ".", "We", "tune", "the", "SVM", "parameter", "using", "5", "-", "fold", "cross", "-", "validation", ".", "As", "in", ",", "we", "use", "a", "receptive", "field", "of", "10x10", "pixels", "and", "a", "stride", "of", "1", ".", "The", "number", "of", "outputs", "is", "set", "to", "for", "fair", "comparison", "with", "the", "other", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "We", "also", "provide", "the", "results", "of", "our", "method", "with", "sign", "split", "(", "x", ",", "using", "and", "for", "encoding", "as", "in", ")", "and", "using", "the", "sparse", "coding", "(", "SC", ")", "encoder", ",", "which", "found", "to", "be", "the", "best", "when", "small", "number", "of", "labeled", "data", "is", "available", ".", "For", "this", "encoder", ",", "we", "searched", "over", "the", "same", "set", "of", "parameter", "values", "as", ",", "i.e.", ",", ".", "The", "parameter", "is", "tuned", "to", "consider", "the", "use", "of", "sparse", "coding", "as", "encoder", "after", "the", "training", "and", ",", "thus", ",", "does", "not", "belong", "to", "the", "method", "that", "we", "propose", ".", "Table", "[", "reference", "]", "summarizes", "the", "results", "obtained", "on", "this", "dataset", "compared", "to", "other", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "When", "pairing", "each", "training", "method", "with", "its", "associated", "natural", "encoding", ",", "EPLS", "outperforms", "all", "the", "other", "methods", ".", "When", "pairing", "the", "training", "methods", "with", "sparse", "coding", ",", "EPLS", "outperforms", "the", "state", "-", "of", "-", "the", "-", "art", "best", "performer", "in", "single", "layer", "networks", "as", "well", ",", "achieving", "accuracy", ".", "Moreover", ",", "the", "standard", "deviation", "of", "the", "folds", "is", "lower", "than", "the", "one", "provided", "by", "OMP", "-", "1", "with", "sparse", "coding", "encoding", ".", "Results", "are", "even", "more", "impressive", "if", "we", "compare", "them", "to", "meta", "-", "parameter", "free", "algorithms", ".", "Figure", "[", "reference", "]", "shows", "a", "subset", "of", "100", "randomly", "selected", "bases", "learned", "by", "our", "method", ",", "10x10", "pixel", "receptive", "field", "and", "a", "system", "of", "outputs", ".", "As", "shown", "in", "the", "figure", ",", "the", "method", "learns", "not", "only", "common", "bases", "such", "as", "oriented", "edges", "/", "ridges", "in", "many", "directions", "and", "colors", "but", "also", "corner", "detectors", ",", "tri", "-", "banded", "colored", "filters", ",", "center", "surrounds", "and", "Laplacian", "of", "Gaussians", "among", "others", ".", "This", "suggests", "that", "enforcing", "lifetime", "sparsity", "helps", "the", "system", "to", "learn", "a", "set", "of", "complex", ",", "rich", "and", "diversified", "bases", ".", "section", ":", "Computational", "complexity", "The", "EPLS", "algorithm", "requires", "the", "computation", "of", ",", "which", "has", "cost", ",", "and", "therefore", "scales", "linearly", "on", "both", "and", ".", "Since", "we", "can", "use", "vSGD", "for", "optimization", ",", "the", "method", "scales", "linearly", "on", "given", "a", "fixed", "number", "of", "epochs", ".", "Finally", ",", "applying", "the", "activation", "function", ",", "the", "cost", "of", "computing", "the", "derivative", "is", "linear", "with", ",", "since", "we", "use", "a", "closed", "form", "for", ".", "The", "memory", "complexity", "is", "related", "to", "the", "mini", "-", "batch", "size", ".", "Consequently", ",", "the", "method", "can", "scale", "gracefully", "to", "very", "large", "datasets", ":", "theoretically", ",", "it", "requires", "to", "store", "in", "memory", "the", "mini", "-", "batch", "input", "data", "(", "elements", ")", ",", "output", "(", "elements", ")", ",", "target", "(", "elements", ")", "and", "the", "system", "parameters", "to", "optimize", "(", "elements", ")", ";", "a", "total", "amount", "of", "elements", ".", "section", ":", "Discussion", "Our", "results", "show", "that", "simultaneously", "enforcing", "both", "population", "and", "lifetime", "sparsity", "helps", "in", "learning", "discriminative", "dictionaries", ",", "which", "reflect", "in", "better", "performance", ",", "especially", "when", "compared", "to", "meta", "-", "parameter", "free", "methods", ".", "Experiments", "suggest", "that", "our", "algorithm", "is", "able", "to", "extract", "features", "that", "generalize", "well", "on", "unseen", "data", ".", "When", "comparing", "the", "performance", "STL", "-", "10", "dataset", ",", "our", "algorithm", "outperforms", "state", "-", "of", "-", "the", "-", "art", "best", "performers", ".", "Results", "suggest", "that", "our", "algorithm", "helps", "the", "classifier", "in", "generalizing", "with", "a", "few", "training", "examples", "(", "of", "the", "dataset", ")", ",", "gaining", "accuracy", "w.r.t", ".", "the", "state", "-", "of", "-", "the", "art", "best", "performer", "(", "OMP", "-", "1", "paired", "with", "sparse", "coding", ")", "with", "a", "lower", "standard", "deviation", "across", "folds", ",", "suggesting", "more", "robustness", "to", "variations", "in", "the", "training", "folds", ".", "It", "is", "important", "to", "highlight", "that", "OMP", "-", "1", "can", "be", "seen", "as", "a", "special", "case", "of", "our", "algorithm", ",", "where", "the", "activation", "function", "is", "and", "lifetime", "sparsity", "is", "not", "taken", "into", "account", "in", "the", "optimization", "process", "(", "potentially", "leading", "to", "dead", "outputs", ")", ".", "Our", "algorithm", "has", "several", "advantages", "over", "OMP", "-", "1", ":", "(", "1", ")", "It", "can", "use", "any", "activation", "function", ";", "(", "2", ")", "by", "enforcing", "lifetime", "sparsity", "it", "does", "not", "suffer", "of", "the", "dead", "output", "problem", ",", "thus", "not", "requiring", "ad", "-", "hoc", "tricks", "to", "avoid", "it", ";", "(", "3", ")", "it", "does", "not", "require", "whitening", ",", "which", "can", "be", "a", "problem", "if", "the", "input", "dimensionality", "is", "large", ".", "With", "our", "proposal", ",", "we", "advance", "in", "the", "meta", "-", "parameter", "free", "line", "of", "ICA", "and", "sparse", "filtering", ".", "It", "is", "clear", "that", "the", "advantage", "of", "sparse", "filtering", "over", "ICA", "comes", "from", "removing", "the", "orthogonality", "constraint", ",", "and", "imposing", "some", "sort", "of", "“", "competition", "”", "between", "outputs", ",", "which", "also", "permits", "overcomplete", "representations", ".", "Following", "this", "spirit", ",", "our", "algorithm", "imposes", "an", "even", "more", "strict", "form", "of", "competition", "to", "prevent", "dead", "outputs", "by", "means", "of", "Strong", "Lifetime", "Sparsity", "and", "confirms", "the", "trend", "of", "that", "data", "reconstruction", "seems", "not", "so", "important", "if", "the", "goal", "is", "to", "have", "a", "discriminative", "sparse", "system", ".", "Last", "and", "most", "importantly", ",", "it", "is", "worth", "highlighting", "five", "interesting", "properties", "of", "the", "EPLS", "algorithm", ".", "First", ",", "the", "method", "is", "meta", "-", "parameter", "free", ",", "which", "highly", "simplifies", "the", "training", "process", "for", "practitioners", ",", "especially", "when", "used", "as", "a", "greedy", "pre", "-", "training", "method", "in", "deep", "architectures", ".", "Second", ",", "the", "method", "is", "fast", "and", "scales", "linearly", "with", "the", "number", "of", "training", "samples", "and", "the", "input", "/", "output", "dimensionalities", ".", "Third", ",", "EPLS", "is", "easy", "to", "implement", ".", "We", "implemented", "the", "EPLS", "in", "Algorithm", "[", "reference", "]", "in", "less", "than", "50", "lines", "of", "C", "code", ".", "The", "mini", "-", "batch", "vSGD", "is", "a", "general", "purpose", "optimizer", ";", "our", "Matlab", "implementation", "of", "vSGD", "plus", "the", "EPLS", "mex", "source", "will", "be", "publicly", "available", "after", "publication", ".", "Fourth", ",", "the", "proposed", "learning", "strategy", "is", "not", "limited", "to", "perceptrons", ".", "Fifth", ",", "there", "is", "an", "interest", "in", "the", "literature", "in", "avoiding", "redundancy", "in", "the", "image", "representation", "by", "using", "the", "algorithms", "in", "a", "convolutional", "fashion", ".", "For", "this", "purpose", ",", "the", "EPLS", "can", "be", "slightly", "modified", "to", "apply", "the", "procedure", "to", "a", "whole", "image", "at", "once", "and", "consider", "the", "mini", "-", "batch", "size", "to", "be", "the", "image", "divided", "into", "patches", ".", "This", "aspect", "is", "not", "considered", "in", "the", "paper", "and", "is", "left", "for", "future", "investigation", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "introduced", "the", "Enforcing", "Population", "and", "Lifetime", "Sparsity", "method", ".", "The", "algorithm", "provides", "a", "meta", "-", "parameter", "free", ",", "off", "-", "the", "-", "shelf", ",", "simple", "and", "computationally", "efficient", "approach", "for", "unsupervised", "sparse", "feature", "learning", ".", "It", "seeks", "both", "lifetime", "and", "population", "sparsity", "in", "an", "explicit", "way", "in", "order", "to", "learn", "discriminative", "features", ",", "thus", "preventing", "dead", "outputs", ".", "Results", "show", "that", "the", "method", "significantly", "outperforms", "all", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "STL", "-", "10", "dataset", "with", "lower", "standard", "deviation", "across", "folds", ",", "suggesting", "more", "robustness", "across", "training", "sets", ".", "bibliography", ":", "References"], "entity": [[3, 21, ["meta", "-", "parameter", "free", ",", "off", "-", "the", "-", "shelf", ",", "simple", "and", "fast", "unsupervised", "feature", "learning", "algorithm"], "Method"], [30, 31, ["sparsity"], "Task"], [34, 37, ["STL", "-", "10"], "Material"], [72, 76, ["feature", "representations", "of", "data"], "Method"], [84, 86, ["image", "classification"], "Task"], [87, 89, ["object", "recognition"], "Task"], [105, 108, ["Unsupervised", "feature", "learning"], "Method"], [115, 123, ["greedy", "layerwise", "pre", "-", "training", "of", "deep", "architectures"], "Task"], [134, 138, ["pre", "-", "training", "algorithms"], "Method"], [155, 157, ["optimization", "problem"], "Task"], [233, 238, ["meta", "-", "parameter", "free", "methods"], "Method"], [239, 241, ["automatic", "approaches"], "Method"], [246, 248, ["learning", "algorithms"], "Method"], [275, 279, ["unsupervised", "feature", "learning", "methods"], "Method"], [293, 294, ["ICA"], "Method"], [295, 297, ["sparse", "filtering"], "Method"], [299, 300, ["ICA"], "Method"], [304, 307, ["object", "recognition", "tasks"], "Task"], [320, 322, ["Computational", "complexity"], "Metric"], [338, 339, ["ICA"], "Method"], [350, 352, ["Sparse", "coding"], "Method"], [355, 356, ["inference"], "Task"], [361, 363, ["iterative", "optimization"], "Method"], [378, 381, ["Predictive", "Sparse", "Decomposition"], "Method"], [382, 383, ["PSD"], "Method"], [389, 391, ["sparse", "coding"], "Method"], [395, 396, ["predictor"], "Method"], [399, 401, ["sparse", "representation"], "Method"], [404, 406, ["sparse", "coding"], "Method"], [408, 410, ["encoding", "step"], "Method"], [430, 433, ["meta", "-", "parameter"], "Method"], [479, 480, ["error"], "Metric"], [505, 507, ["unsupervised", "training"], "Task"], [514, 516, ["optimization", "strategies"], "Method"], [531, 533, ["optimization", "techniques"], "Method"], [545, 549, ["STL", "-", "10", "dataset"], "Material"], [575, 579, ["Linear", "feature", "extraction", "methods"], "Method"], [581, 584, ["sparse", "coding", "encodings"], "Method"], [598, 606, ["training", "/", "encoding", "and", "encoding", "/", "pooling", "strategies"], "Method"], [626, 629, ["unsupervised", "feature", "learning"], "Method"], [639, 640, ["pooling"], "Task"], [674, 676, ["sparse", "coding"], "Method"], [677, 681, ["soft", "-", "threshold", "encodings"], "Method"], [683, 685, ["sum", "pooling"], "Method"], [705, 708, ["unsupervised", "feature", "learning"], "Method"], [709, 712, ["Restricted", "Boltzmann", "Machines"], "Method"], [713, 714, ["RBM"], "Method"], [716, 719, ["auto", "-", "encoders"], "Method"], [720, 722, ["sparse", "coding"], "Method"], [723, 724, ["hybrids"], "Method"], [726, 727, ["PSD"], "Method"], [733, 734, ["ICA"], "Method"], [735, 737, ["Reconstruction", "ICA"], "Method"], [738, 739, ["RICA"], "Method"], [741, 743, ["Sparse", "Filtering"], "Method"], [747, 749, ["vector", "quantization"], "Method"], [751, 754, ["Orthogonal", "Matching", "Pursuit"], "Method"], [755, 758, ["OMP", "-", "k"], "Method"], [768, 771, ["unsupervised", "feature", "representations"], "Task"], [789, 793, ["Sparse", "auto", "-", "encoders"], "Method"], [794, 796, ["sparse", "RBM"], "Method"], [797, 799, ["sparse", "coding"], "Method"], [800, 801, ["PSD"], "Method"], [802, 805, ["OMP", "-", "k"], "Method"], [806, 808, ["Reconstruction", "ICA"], "Method"], [809, 810, ["RICA"], "Method"], [819, 821, ["reconstruction", "error"], "Metric"], [838, 840, ["sparse", "filtering"], "Method"], [854, 857, ["discriminative", "sparse", "system"], "Method"], [858, 860, ["Sparse", "filtering"], "Method"], [880, 881, ["Sparsity"], "Method"], [889, 891, ["output", "representation"], "Method"], [914, 915, ["Sparsity"], "Method"], [1017, 1019, ["population", "sparsity"], "Task"], [1040, 1042, ["objective", "function"], "Metric"], [1050, 1051, ["penalty"], "Method"], [1065, 1069, ["Sparse", "auto", "-", "encoders"], "Method"], [1099, 1102, ["OMP", "-", "k"], "Method"], [1160, 1166, ["sparse", "feature", "representation", "of", "the", "data"], "Method"], [1189, 1190, ["error"], "Metric"], [1230, 1232, ["optimization", "strategies"], "Method"], [1250, 1251, ["Population"], "Task"], [1252, 1254, ["Lifetime", "Sparsity"], "Task"], [1340, 1344, ["Strong", "Lifetime", "Sparsity", "definition"], "Method"], [1436, 1439, ["Strong", "Lifetime", "Sparsity"], "Task"], [1535, 1536, ["population"], "Task"], [1537, 1539, ["lifetime", "sparsity"], "Task"], [1633, 1635, ["assignment", "problem"], "Task"], [1637, 1639, ["Hungarian", "method"], "Method"], [1641, 1644, ["combinatorial", "optimization", "algorithm"], "Method"], [1648, 1650, ["assignment", "problem"], "Task"], [1654, 1656, ["computational", "cost"], "Metric"], [1693, 1695, ["approximate", "solution"], "Method"], [1705, 1708, ["Ideal", "target", "generation"], "Method"], [1710, 1719, ["Enforcing", "Population", "and", "Lifetime", "Sparsity", "(", "EPLS", ")", "algorithm"], "Method"], [1781, 1783, ["EPLS", "algorithm"], "Method"], [1809, 1810, ["EPLS"], "Method"], [2050, 2052, ["Optimization", "strategies"], "Method"], [2064, 2066, ["activation", "function"], "Method"], [2111, 2113, ["training", "strategy"], "Method"], [2134, 2136, ["Optimization", "strategy"], "Method"], [2145, 2160, ["off", "-", "the", "-", "shelf", "mini", "-", "batch", "Stochastic", "Gradient", "Descent", "(", "SGD", ")", "method"], "Method"], [2161, 2164, ["adaptive", "learning", "rates"], "Method"], [2166, 2170, ["variance", "-", "based", "SGD"], "Method"], [2171, 2172, ["vSGD"], "Method"], [2236, 2238, ["EPLS", "inhibitor"], "Method"], [2279, 2280, ["EPLS"], "Method"], [2307, 2309, ["learning", "rate"], "Metric"], [2373, 2376, ["relative", "decrement", "error"], "Metric"], [2413, 2416, ["gradient", "descent", "convergence"], "Metric"], [2418, 2419, ["quality"], "Metric"], [2421, 2422, ["minimization"], "Task"], [2446, 2450, ["mini", "-", "batch", "vSGD"], "Method"], [2467, 2469, ["EPLS", "training"], "Method"], [2487, 2490, ["learning", "rate", "η"], "Metric"], [2511, 2515, ["training", "and", "encoding", "strategies"], "Method"], [2516, 2519, ["single", "layer", "networks"], "Method"], [2527, 2531, ["STL", "-", "10", "dataset"], "Material"], [2532, 2535, ["STL", "-", "10"], "Material"], [2585, 2588, ["STL", "-", "10"], "Material"], [2619, 2620, ["classifier"], "Method"], [2632, 2634, ["supervised", "training"], "Task"], [2642, 2644, ["unsupervised", "method"], "Method"], [2689, 2691, ["unsupervised", "method"], "Method"], [2726, 2727, ["EPLS"], "Method"], [2784, 2785, ["SVM"], "Method"], [2786, 2788, ["classification", "purposes"], "Task"], [2792, 2794, ["SVM", "parameter"], "Method"], [2795, 2801, ["5", "-", "fold", "cross", "-", "validation"], "Method"], [2865, 2871, ["sparse", "coding", "(", "SC", ")", "encoder"], "Method"], [2914, 2916, ["sparse", "coding"], "Method"], [2917, 2918, ["encoder"], "Method"], [2966, 2968, ["natural", "encoding"], "Method"], [2969, 2970, ["EPLS"], "Method"], [2979, 2981, ["training", "methods"], "Method"], [2982, 2984, ["sparse", "coding"], "Method"], [2985, 2986, ["EPLS"], "Method"], [2998, 3001, ["single", "layer", "networks"], "Method"], [3005, 3006, ["accuracy"], "Metric"], [3022, 3029, ["OMP", "-", "1", "with", "sparse", "coding", "encoding"], "Method"], [3040, 3045, ["meta", "-", "parameter", "free", "algorithms"], "Method"], [3099, 3101, ["corner", "detectors"], "Method"], [3102, 3107, ["tri", "-", "banded", "colored", "filters"], "Method"], [3111, 3114, ["Laplacian", "of", "Gaussians"], "Method"], [3140, 3142, ["Computational", "complexity"], "Metric"], [3143, 3145, ["EPLS", "algorithm"], "Method"], [3166, 3167, ["vSGD"], "Method"], [3168, 3169, ["optimization"], "Task"], [3186, 3188, ["activation", "function"], "Method"], [3208, 3210, ["memory", "complexity"], "Metric"], [3291, 3293, ["discriminative", "dictionaries"], "Task"], [3304, 3309, ["meta", "-", "parameter", "free", "methods"], "Method"], [3331, 3335, ["STL", "-", "10", "dataset"], "Material"], [3356, 3357, ["classifier"], "Method"], [3371, 3372, ["accuracy"], "Metric"], [3384, 3387, ["OMP", "-", "1"], "Method"], [3389, 3391, ["sparse", "coding"], "Method"], [3416, 3419, ["OMP", "-", "1"], "Method"], [3445, 3447, ["optimization", "process"], "Task"], [3461, 3464, ["OMP", "-", "1"], "Method"], [3472, 3474, ["activation", "function"], "Method"], [3480, 3482, ["lifetime", "sparsity"], "Method"], [3488, 3491, ["dead", "output", "problem"], "Task"], [3538, 3539, ["ICA"], "Method"], [3540, 3542, ["sparse", "filtering"], "Method"], [3550, 3552, ["sparse", "filtering"], "Method"], [3553, 3554, ["ICA"], "Method"], [3575, 3577, ["overcomplete", "representations"], "Method"], [3608, 3610, ["data", "reconstruction"], "Task"], [3621, 3624, ["discriminative", "sparse", "system"], "Method"], [3639, 3641, ["EPLS", "algorithm"], "Method"], [3656, 3658, ["training", "process"], "Task"], [3659, 3660, ["practitioners"], "Task"], [3666, 3671, ["greedy", "pre", "-", "training", "method"], "Method"], [3672, 3674, ["deep", "architectures"], "Method"], [3699, 3700, ["EPLS"], "Method"], [3708, 3709, ["EPLS"], "Method"], [3724, 3728, ["mini", "-", "batch", "vSGD"], "Method"], [3730, 3733, ["general", "purpose", "optimizer"], "Method"], [3735, 3737, ["Matlab", "implementation"], "Method"], [3738, 3739, ["vSGD"], "Method"], [3741, 3744, ["EPLS", "mex", "source"], "Method"], [3755, 3757, ["learning", "strategy"], "Method"], [3761, 3762, ["perceptrons"], "Method"], [3777, 3779, ["image", "representation"], "Task"], [3785, 3787, ["convolutional", "fashion"], "Method"], [3793, 3794, ["EPLS"], "Method"], [3848, 3854, ["Enforcing", "Population", "and", "Lifetime", "Sparsity", "method"], "Method"], [3876, 3880, ["unsupervised", "sparse", "feature", "learning"], "Method"], [3921, 3925, ["STL", "-", "10", "dataset"], "Material"], [565, 567, ["image", "classification"], "Task"]]} {"doc_id": "0b0dc14b8a8dcccbfc62a38355fff2f6a361e9d2", "text": ["document", ":", "Effective", "LSTMs", "for", "Target", "-", "Dependent", "Sentiment", "Classification", "Target", "-", "dependent", "sentiment", "classification", "remains", "a", "challenge", ":", "modeling", "the", "semantic", "relatedness", "of", "a", "target", "with", "its", "context", "words", "in", "a", "sentence", ".", "Different", "context", "words", "have", "different", "influences", "on", "determining", "the", "sentiment", "polarity", "of", "a", "sentence", "towards", "the", "target", ".", "Therefore", ",", "it", "is", "desirable", "to", "integrate", "the", "connections", "between", "target", "word", "and", "context", "words", "when", "building", "a", "learning", "system", ".", "In", "this", "paper", ",", "we", "develop", "two", "target", "dependent", "long", "short", "-", "term", "memory", "(", "LSTM", ")", "models", ",", "where", "target", "information", "is", "automatically", "taken", "into", "account", ".", "We", "evaluate", "our", "methods", "on", "a", "benchmark", "dataset", "from", "Twitter", ".", "Empirical", "results", "show", "that", "modeling", "sentence", "representation", "with", "standard", "LSTM", "does", "not", "perform", "well", ".", "Incorporating", "target", "information", "into", "LSTM", "can", "significantly", "boost", "the", "classification", "accuracy", ".", "The", "target", "-", "dependent", "LSTM", "models", "achieve", "state", "-", "of", "-", "the", "-", "art", "performances", "without", "using", "syntactic", "parser", "or", "external", "sentiment", "lexicons", ".", "section", ":", "Introduction", "This", "work", "is", "licensed", "under", "a", "Creative", "Commons", "Attribution", "4.0", "International", "License", ".", "License", "details", ":", "Sentiment", "analysis", ",", "also", "known", "as", "opinion", "mining", ",", "is", "a", "fundamental", "task", "in", "natural", "language", "processing", "and", "computational", "linguistics", ".", "Sentiment", "analysis", "is", "crucial", "to", "understanding", "user", "generated", "text", "in", "social", "networks", "or", "product", "reviews", ",", "and", "has", "drawn", "a", "lot", "of", "attentions", "from", "both", "industry", "and", "academic", "communities", ".", "In", "this", "paper", ",", "we", "focus", "on", "target", "-", "dependent", "sentiment", "classification", ",", "which", "is", "a", "fundamental", "and", "extensively", "studied", "task", "in", "the", "field", "of", "sentiment", "analysis", ".", "Given", "a", "sentence", "and", "a", "target", "mention", ",", "the", "task", "calls", "for", "inferring", "the", "sentiment", "polarity", "(", "e.g.", "positive", ",", "negative", ",", "neutral", ")", "of", "the", "sentence", "towards", "the", "target", ".", "For", "example", ",", "let", "us", "consider", "the", "sentence", ":", "“", "I", "bought", "a", "new", "camera", ".", "The", "picture", "quality", "is", "amazing", "but", "the", "battery", "life", "is", "too", "short", "”", ".", "If", "the", "target", "string", "is", "picture", "quality", ",", "the", "expected", "sentiment", "polarity", "is", "“", "positive", "”", "as", "the", "sentence", "expresses", "a", "positive", "opinion", "towards", "picture", "quality", ".", "If", "we", "consider", "the", "target", "as", "battery", "life", ",", "the", "correct", "sentiment", "polarity", "should", "be", "“", "negative", "”", ".", "Target", "-", "dependent", "sentiment", "classification", "is", "typically", "regarded", "as", "a", "kind", "of", "text", "classification", "problem", "in", "literature", ".", "Majority", "of", "existing", "studies", "build", "sentiment", "classifiers", "with", "supervised", "machine", "learning", "approach", ",", "such", "as", "feature", "based", "Supported", "Vector", "Machine", "or", "neural", "network", "approaches", ".", "Despite", "the", "effectiveness", "of", "these", "approaches", ",", "we", "argue", "that", "target", "-", "dependent", "sentiment", "classification", "remains", "a", "challenge", ":", "how", "to", "effectively", "model", "the", "semantic", "relatedness", "of", "a", "target", "word", "with", "its", "context", "words", "in", "a", "sentence", ".", "One", "straight", "forward", "way", "to", "address", "this", "problem", "is", "to", "manually", "design", "a", "set", "of", "target", "-", "dependent", "features", ",", "and", "integrate", "them", "into", "existing", "feature", "-", "based", "SVM", ".", "However", ",", "feature", "engineering", "is", "labor", "intensive", "and", "the", "“", "sparse", "”", "and", "“", "discrete", "”", "features", "are", "clumsy", "in", "encoding", "side", "information", "like", "target", "-", "context", "relatedness", ".", "In", "addition", ",", "a", "person", "asked", "to", "do", "this", "task", "will", "naturally", "“", "look", "at", "”", "parts", "of", "relevant", "context", "words", "which", "are", "helpful", "to", "determine", "the", "sentiment", "polarity", "of", "a", "sentence", "towards", "the", "target", ".", "These", "motivate", "us", "to", "develop", "a", "powerful", "neural", "network", "approach", ",", "which", "is", "capable", "of", "learning", "continuous", "features", "(", "representations", ")", "without", "feature", "engineering", "and", "meanwhile", "capturing", "the", "intricate", "relatedness", "between", "target", "and", "context", "words", ".", "In", "this", "paper", ",", "we", "present", "neural", "network", "models", "to", "deal", "with", "target", "-", "dependent", "sentiment", "classification", ".", "The", "approach", "is", "an", "extension", "on", "long", "short", "-", "term", "memory", "(", "LSTM", ")", "by", "incorporating", "target", "information", ".", "Such", "target", "-", "dependent", "LSTM", "approach", "models", "the", "relatedness", "of", "a", "target", "word", "with", "its", "context", "words", ",", "and", "selects", "the", "relevant", "parts", "of", "contexts", "to", "infer", "the", "sentiment", "polarity", "towards", "the", "target", ".", "The", "model", "could", "be", "trained", "in", "an", "end", "-", "to", "-", "end", "way", "with", "standard", "backpropagation", ",", "where", "the", "loss", "function", "is", "cross", "-", "entropy", "error", "of", "supervised", "sentiment", "classification", ".", "We", "apply", "the", "neural", "model", "to", "target", "-", "dependent", "sentiment", "classification", "on", "a", "benchmark", "dataset", ".", "We", "compare", "with", "feature", "-", "based", "SVM", ",", "adaptive", "recursive", "neural", "network", "and", "lexicon", "-", "enhanced", "neural", "network", ".", "Empirical", "results", "show", "that", "the", "proposed", "approach", "without", "using", "syntactic", "parser", "or", "external", "sentiment", "lexicon", "obtains", "state", "-", "of", "-", "the", "-", "art", "classification", "accuracy", ".", "In", "addition", ",", "we", "find", "that", "modeling", "sentence", "with", "standard", "LSTM", "does", "not", "perform", "well", "on", "this", "target", "-", "dependent", "task", ".", "Integrating", "target", "information", "into", "LSTM", "could", "significantly", "improve", "the", "classification", "accuracy", ".", "section", ":", "The", "Approach", "We", "describe", "the", "proposed", "approach", "for", "target", "-", "dependent", "sentiment", "classification", "in", "this", "section", ".", "We", "first", "present", "a", "basic", "long", "short", "-", "term", "memory", "(", "LSTM", ")", "approach", ",", "which", "models", "the", "semantic", "representation", "of", "a", "sentence", "without", "considering", "the", "target", "word", "being", "evaluated", ".", "Afterwards", ",", "we", "extend", "LSTM", "by", "considering", "the", "target", "word", ",", "obtaining", "the", "Target", "-", "Dependent", "Long", "Short", "-", "Term", "Memory", "(", "TD", "-", "LSTM", ")", "model", ".", "Finally", ",", "we", "extend", "TD", "-", "LSTM", "with", "target", "connection", ",", "where", "the", "semantic", "relatedness", "of", "target", "with", "its", "context", "words", "are", "incorporated", ".", "subsection", ":", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "In", "this", "part", ",", "we", "describe", "a", "long", "short", "-", "term", "memory", "(", "LSTM", ")", "model", "for", "target", "-", "dependent", "sentiment", "classification", ".", "It", "is", "a", "basic", "version", "of", "our", "approach", ".", "In", "this", "setting", ",", "the", "target", "to", "be", "evaluated", "is", "ignored", "so", "that", "the", "task", "is", "considered", "in", "a", "target", "independent", "way", ".", "We", "use", "LSTM", "as", "it", "is", "a", "state", "-", "of", "-", "the", "-", "art", "performer", "for", "semantic", "composition", "in", "the", "area", "of", "sentiment", "analysis", ".", "It", "is", "capable", "of", "computing", "the", "representation", "of", "a", "longer", "expression", "(", "e.g.", "a", "sentence", ")", "from", "the", "representation", "of", "its", "children", "with", "multi", "levels", "of", "abstraction", ".", "The", "sentence", "representation", "can", "be", "naturally", "considered", "as", "the", "feature", "to", "predict", "the", "sentiment", "polarity", "of", "sentence", ".", "Specifically", ",", "each", "word", "is", "represented", "as", "a", "low", "dimensional", ",", "continuous", "and", "real", "-", "valued", "vector", ",", "also", "known", "as", "word", "embedding", ".", "All", "the", "word", "vectors", "are", "stacked", "in", "a", "word", "embedding", "matrix", ",", "where", "is", "the", "dimension", "of", "word", "vector", "and", "is", "vocabulary", "size", ".", "In", "this", "work", ",", "we", "pre", "-", "train", "the", "values", "of", "word", "vectors", "from", "text", "corpus", "with", "embedding", "learning", "algorithms", "to", "make", "better", "use", "of", "semantic", "and", "grammatical", "associations", "of", "words", ".", "We", "use", "LSTM", "to", "compute", "the", "vector", "of", "a", "sentence", "from", "the", "vectors", "of", "words", "it", "contains", ",", "an", "illustration", "of", "the", "model", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "LSTM", "is", "a", "kind", "of", "recurrent", "neural", "network", "(", "RNN", ")", ",", "which", "is", "capable", "of", "mapping", "vectors", "of", "words", "with", "variable", "length", "to", "a", "fixed", "-", "length", "vector", "by", "recursively", "transforming", "current", "word", "vector", "with", "the", "output", "vector", "of", "the", "previous", "step", ".", "The", "transition", "function", "of", "standard", "RNN", "is", "a", "linear", "layer", "followed", "by", "a", "pointwise", "non", "-", "linear", "layer", "such", "as", "hyperbolic", "tangent", "function", "(", ")", ".", "where", ",", ",", "is", "dimension", "of", "word", "vector", ".", "However", ",", "standard", "RNN", "suffers", "the", "problem", "of", "gradient", "vanishing", "or", "exploding", ",", "where", "gradients", "may", "grow", "or", "decay", "exponentially", "over", "long", "sequences", ".", "Many", "researchers", "use", "a", "more", "sophisticated", "and", "powerful", "LSTM", "cell", "as", "the", "transition", "function", ",", "so", "that", "long", "-", "distance", "semantic", "correlations", "in", "a", "sequence", "could", "be", "better", "modeled", ".", "Compared", "with", "standard", "RNN", ",", "LSTM", "cell", "contains", "three", "additional", "neural", "gates", ":", "an", "input", "gate", ",", "a", "forget", "gate", "and", "an", "output", "gate", ".", "These", "gates", "adaptively", "remember", "input", "vector", ",", "forget", "previous", "history", "and", "generate", "output", "vector", ".", "LSTM", "cell", "is", "calculated", "as", "follows", ".", "where", "stands", "for", "element", "-", "wise", "multiplication", ",", "is", "sigmoid", "function", ",", ",", ",", ",", ",", ",", "are", "the", "parameters", "of", "input", ",", "forget", "and", "output", "gates", ".", "After", "calculating", "the", "hidden", "vector", "of", "each", "position", ",", "we", "regard", "the", "last", "hidden", "vector", "as", "the", "sentence", "representation", ".", "We", "feed", "it", "to", "a", "linear", "layer", "whose", "output", "length", "is", "class", "number", ",", "and", "add", "a", "layer", "to", "output", "the", "probability", "of", "classifying", "the", "sentence", "as", "positive", ",", "negative", "or", "neutral", ".", "Softmax", "function", "is", "calculated", "as", "follows", ",", "where", "is", "the", "number", "of", "sentiment", "categories", ".", "subsection", ":", "Target", "-", "Dependent", "LSTM", "(", "TD", "-", "LSTM", ")", "The", "aforementioned", "LSTM", "model", "solves", "target", "-", "dependent", "sentiment", "classification", "in", "a", "target", "-", "independent", "way", ".", "That", "is", "to", "say", ",", "the", "feature", "representation", "used", "for", "sentiment", "classification", "remains", "the", "same", "without", "considering", "the", "target", "words", ".", "Let", "us", "again", "take", "“", "I", "bought", "a", "new", "camera", ".", "The", "picture", "quality", "is", "amazing", "but", "the", "battery", "life", "is", "too", "short", "”", "as", "an", "example", ".", "The", "representations", "of", "this", "sentence", "with", "regard", "to", "picture", "quality", "and", "battery", "life", "are", "identical", ".", "This", "is", "evidently", "problematic", "as", "the", "sentiment", "polarity", "labels", "towards", "these", "two", "targets", "are", "different", ".", "To", "take", "into", "account", "of", "the", "target", "information", ",", "we", "make", "a", "slight", "modification", "on", "the", "aforementioned", "LSTM", "model", "and", "introduce", "a", "target", "-", "dependent", "LSTM", "(", "TD", "-", "LSTM", ")", "in", "this", "subsection", ".", "The", "basic", "idea", "is", "to", "model", "the", "preceding", "and", "following", "contexts", "surrounding", "the", "target", "string", ",", "so", "that", "contexts", "in", "both", "directions", "could", "be", "used", "as", "feature", "representations", "for", "sentiment", "classification", ".", "We", "believe", "that", "capturing", "such", "target", "-", "dependent", "context", "information", "could", "improve", "the", "accuracy", "of", "target", "-", "dependent", "sentiment", "classification", ".", "Specifically", ",", "we", "use", "two", "LSTM", "neural", "networks", ",", "a", "left", "one", "LSTM", "and", "a", "right", "one", "LSTM", ",", "to", "model", "the", "preceding", "and", "following", "contexts", "respectively", ".", "An", "illustration", "of", "the", "model", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "The", "input", "of", "LSTM", "is", "the", "preceding", "contexts", "plus", "target", "string", ",", "and", "the", "input", "of", "LSTM", "is", "the", "following", "contexts", "plus", "target", "string", ".", "We", "run", "LSTM", "from", "left", "to", "right", ",", "and", "run", "LSTM", "from", "right", "to", "left", ".", "We", "favor", "this", "strategy", "as", "we", "believe", "that", "regarding", "target", "string", "as", "the", "last", "unit", "could", "better", "utilize", "the", "semantics", "of", "target", "string", "when", "using", "the", "composed", "representation", "for", "sentiment", "classification", ".", "Afterwards", ",", "we", "concatenate", "the", "last", "hidden", "vectors", "of", "LSTM", "and", "LSTM", ",", "and", "feed", "them", "to", "a", "layer", "to", "classify", "the", "sentiment", "polarity", "label", ".", "One", "could", "also", "try", "averaging", "or", "summing", "the", "last", "hidden", "vectors", "of", "LSTM", "and", "LSTM", "as", "alternatives", ".", "subsection", ":", "Target", "-", "Connection", "LSTM", "(", "TC", "-", "LSTM", ")", "Compared", "with", "LSTM", "model", ",", "target", "-", "dependent", "LSTM", "(", "TD", "-", "LSTM", ")", "could", "make", "better", "use", "of", "the", "target", "information", ".", "However", ",", "we", "think", "TD", "-", "LSTM", "is", "still", "not", "good", "enough", "because", "it", "does", "not", "capture", "the", "interactions", "between", "target", "word", "and", "its", "contexts", ".", "Furthermore", ",", "a", "person", "asked", "to", "do", "target", "-", "dependent", "sentiment", "classification", "will", "select", "the", "relevant", "context", "words", "which", "are", "helpful", "to", "determine", "the", "sentiment", "polarity", "of", "a", "sentence", "towards", "the", "target", ".", "Based", "on", "the", "consideration", "mentioned", "above", ",", "we", "go", "one", "step", "further", "and", "develop", "a", "target", "-", "connection", "long", "short", "-", "term", "memory", "(", "TC", "-", "LSTM", ")", ".", "This", "model", "extends", "TD", "-", "LSTM", "by", "incorporating", "an", "target", "connection", "component", ",", "which", "explicitly", "utilizes", "the", "connections", "between", "target", "word", "and", "each", "context", "word", "when", "composing", "the", "representation", "of", "a", "sentence", ".", "An", "overview", "of", "TC", "-", "LSTM", "is", "illustrated", "in", "Figure", "[", "reference", "]", ".", "The", "input", "of", "TC", "-", "LSTM", "is", "a", "sentence", "consisting", "of", "words", "and", "a", "target", "string", "occurs", "in", "the", "sentence", ".", "We", "represent", "target", "as", "because", "a", "target", "could", "be", "a", "word", "sequence", "of", "variable", "length", ",", "such", "as", "“", "google", "”", "or", "“", "harry", "potter", "”", ".", "When", "processing", "a", "sentence", ",", "we", "split", "it", "into", "three", "components", ":", "target", "words", ",", "preceding", "context", "words", "and", "following", "context", "words", ".", "We", "obtain", "target", "vector", "by", "averaging", "the", "vectors", "of", "words", "it", "contains", ",", "which", "has", "been", "proven", "to", "be", "simple", "and", "effective", "in", "representing", "named", "entities", ".", "When", "compute", "the", "hidden", "vectors", "of", "preceding", "and", "following", "context", "words", ",", "we", "use", "two", "separate", "long", "short", "-", "term", "memory", "models", ",", "which", "are", "similar", "with", "the", "strategy", "used", "in", "TD", "-", "LSTM", ".", "The", "difference", "is", "that", "in", "TC", "-", "LSTM", "the", "input", "at", "each", "position", "is", "the", "concatenation", "of", "word", "embedding", "and", "target", "vector", ",", "while", "in", "TD", "-", "LSTM", "the", "input", "at", "each", "position", "only", "includes", "the", "embedding", "of", "current", "word", ".", "We", "believe", "that", "TC", "-", "LSTM", "could", "make", "better", "use", "of", "the", "connection", "between", "target", "and", "each", "context", "word", "when", "building", "the", "representation", "of", "a", "sentence", ".", "subsection", ":", "Model", "Training", "We", "train", "LSTM", ",", "TD", "-", "LSTM", "and", "TC", "-", "LSTM", "in", "an", "end", "-", "to", "-", "end", "way", "in", "a", "supervised", "learning", "framework", ".", "The", "loss", "function", "is", "the", "cross", "-", "entropy", "error", "of", "sentiment", "classification", ".", "where", "is", "the", "training", "data", ",", "is", "the", "number", "of", "sentiment", "categories", ",", "means", "a", "sentence", ",", "is", "the", "probability", "of", "predicting", "as", "class", "given", "by", "the", "layer", ",", "indicates", "whether", "class", "is", "the", "correct", "sentiment", "category", ",", "whose", "value", "is", "1", "or", "0", ".", "We", "take", "the", "derivative", "of", "loss", "function", "through", "back", "-", "propagation", "with", "respect", "to", "all", "parameters", ",", "and", "update", "parameters", "with", "stochastic", "gradient", "descent", ".", "section", ":", "Experiment", "We", "apply", "the", "proposed", "method", "to", "target", "-", "dependent", "sentiment", "classification", "to", "evaluate", "its", "effectiveness", ".", "We", "describe", "experimental", "setting", "and", "empirical", "results", "in", "this", "section", ".", "subsection", ":", "Experimental", "Settings", "We", "conduct", "experiment", "in", "a", "supervised", "setting", "on", "a", "benchmark", "dataset", ".", "Each", "instance", "in", "the", "training", "/", "test", "set", "has", "a", "manually", "labeled", "sentiment", "polarity", ".", "Training", "set", "contains", "6", ",", "248", "sentences", "and", "test", "set", "has", "692", "sentences", ".", "The", "percentages", "of", "positive", ",", "negative", "and", "neutral", "in", "training", "and", "test", "sets", "are", "both", "25", "%", ",", "25", "%", ",", "50", "%", ".", "We", "train", "the", "model", "on", "training", "set", ",", "and", "evaluate", "the", "performance", "on", "test", "set", ".", "Evaluation", "metrics", "are", "accuracy", "and", "macro", "-", "F1", "score", "over", "positive", ",", "negative", "and", "neutral", "categories", ".", "subsection", ":", "Comparison", "to", "Other", "Methods", "We", "compare", "with", "several", "baseline", "methods", ",", "including", ":", "In", "SVM", "-", "indep", ",", "SVM", "classifier", "is", "built", "with", "target", "-", "independent", "features", ",", "such", "as", "unigram", ",", "bigram", ",", "punctuations", ",", "emoticons", ",", "hashtags", ",", "the", "numbers", "of", "positive", "or", "negative", "words", "in", "General", "Inquirer", "sentiment", "lexicon", ".", "In", "SVM", "-", "dep", ",", "target", "-", "dependent", "features", "are", "also", "concatenated", "as", "the", "feature", "representation", ".", "In", "Recursive", "NN", ",", "standard", "Recursive", "neural", "network", "is", "used", "for", "feature", "learning", "over", "a", "transfered", "target", "-", "dependent", "dependency", "tree", ".", "AdaRNN", "-", "w", "/", "oE", ",", "AdaRNN", "-", "w", "/", "E", "and", "AdaRNN", "-", "comb", "are", "different", "variations", "of", "adaptive", "recursive", "neural", "network", ",", "whose", "composition", "functions", "are", "adaptively", "selected", "according", "to", "the", "inputs", ".", "In", "Target", "-", "dep", ",", "SVM", "classifier", "is", "built", "based", "on", "rich", "target", "-", "independent", "and", "target", "-", "dependent", "features", ".", "In", "Target", "-", "dep", "+", ",", "sentiment", "lexicon", "features", "are", "further", "incorporated", ".", "The", "neural", "models", "developed", "in", "this", "paper", "are", "abbreviated", "as", "LSTM", ",", "TD", "-", "LSTM", "and", "TC", "-", "LSTM", ",", "which", "are", "described", "in", "the", "previous", "section", ".", "We", "use", "100", "-", "dimensional", "Glove", "vectors", "learned", "from", "Twitter", ",", "randomize", "the", "parameters", "with", "uniform", "distribution", ",", "set", "the", "clipping", "threshold", "of", "softmax", "layer", "as", "200", "and", "set", "learning", "rate", "as", "0.01", ".", "Experimental", "results", "of", "baseline", "models", "and", "our", "methods", "are", "given", "in", "Table", "[", "reference", "]", ".", "Comparing", "between", "SVM", "-", "indep", "and", "SVM", "-", "dep", ",", "we", "can", "find", "that", "incorporating", "target", "information", "can", "improve", "the", "classification", "accuracy", "of", "a", "basic", "SVM", "classifier", ".", "AdaRNN", "performs", "better", "than", "feature", "based", "SVM", "by", "making", "use", "of", "dependency", "parsing", "information", "and", "tree", "-", "structured", "semantic", "composition", ".", "We", "can", "find", "that", "target", "-", "dep", "is", "a", "strong", "performer", "even", "without", "using", "lexicon", "features", ".", "It", "benefits", "from", "rich", "automatic", "features", "generated", "from", "word", "embeddings", ".", "Among", "LSTM", "based", "models", "described", "in", "this", "paper", ",", "the", "basic", "LSTM", "approach", "performs", "worst", ".", "This", "is", "not", "surprising", "because", "this", "task", "requires", "understanding", "target", "-", "dependent", "text", "semantics", ",", "while", "the", "basic", "LSTM", "model", "does", "not", "capture", "any", "target", "information", "so", "that", "it", "predicts", "the", "same", "result", "for", "different", "targets", "in", "a", "sentence", ".", "TD", "-", "LSTM", "obtains", "a", "big", "improvement", "over", "LSTM", "when", "target", "signals", "are", "taken", "into", "consideration", ".", "This", "result", "demonstrates", "the", "importance", "of", "target", "information", "for", "target", "-", "dependent", "sentiment", "classification", ".", "By", "incorporating", "target", "-", "connection", "mechanism", ",", "TC", "-", "LSTM", "obtains", "the", "best", "performances", "and", "outperforms", "all", "baseline", "methods", "in", "term", "of", "classification", "accuracy", ".", "Comparing", "between", "Target", "-", "dep", "and", "Target", "-", "dep", ",", "we", "find", "that", "sentiment", "lexicon", "feature", "could", "further", "improve", "the", "classification", "accuracy", ".", "Our", "final", "model", "TC", "-", "LSTM", "without", "using", "sentiment", "lexicon", "information", "performs", "comparably", "with", "Target", "-", "dep", ".", "We", "believe", "that", "incorporation", "lexicon", "information", "in", "TC", "-", "LSTM", "could", "get", "further", "improvement", ".", "We", "leave", "this", "as", "a", "potential", "future", "work", ".", "subsection", ":", "Effects", "of", "Word", "Embeddings", "It", "is", "well", "accepted", "that", "a", "good", "word", "embedding", "is", "crucial", "to", "composing", "a", "powerful", "text", "representation", "at", "higher", "level", ".", "We", "therefore", "study", "the", "effects", "of", "different", "word", "embeddings", "on", "LSTM", ",", "TD", "-", "LSTM", "and", "TC", "-", "LSTM", "in", "this", "part", ".", "Since", "the", "benchmark", "dataset", "from", "comes", "from", "Twitter", ",", "we", "compare", "between", "sentiment", "-", "specific", "word", "embedding", "(", "SSWE", ")", "and", "Glove", "vectors", ".", "All", "these", "word", "vectors", "are", "50", "-", "dimensional", "and", "learned", "from", "Twitter", ".", "SSWE", ",", "SSWE", "and", "SSWE", "are", "different", "embedding", "learning", "algorithms", "introduced", "in", ".", "SSWE", "and", "SSWE", "learn", "word", "embeddings", "by", "only", "using", "sentiment", "of", "sentences", ".", "SSWE", "takes", "into", "account", "of", "sentiment", "of", "sentences", "and", "contexts", "of", "words", "simultaneously", ".", "From", "Figure", "[", "reference", "]", ",", "we", "can", "find", "that", "SSWE", "and", "SSWE", "perform", "worse", "than", "SSWE", ",", "which", "is", "consistent", "with", "the", "results", "reported", "on", "target", "-", "independent", "sentiment", "classification", "of", "tweets", ".", "This", "shows", "the", "importance", "of", "context", "information", "for", "word", "embedding", "learning", "as", "both", "SSWE", "and", "SSWE", "do", "not", "encode", "any", "word", "contexts", ".", "Glove", "and", "SSWE", "perform", "comparably", ",", "which", "indicates", "the", "importance", "of", "global", "context", "for", "estimating", "a", "good", "word", "representation", ".", "In", "addition", ",", "the", "target", "connection", "model", "TC", "-", "LSTM", "performs", "best", "when", "considering", "a", "specific", "word", "embedding", ".", "We", "compare", "between", "Glove", "vectors", "with", "different", "dimensions", "(", "50", "/", "100", "/", "200", ")", ".", "Classification", "accuracy", "and", "time", "cost", "are", "given", "in", "Figure", "[", "reference", "]", "and", "Table", "[", "reference", "]", ",", "respectively", ".", "We", "can", "find", "that", "100", "-", "dimensional", "word", "vectors", "perform", "better", "than", "50", "-", "dimensional", "word", "vectors", ",", "while", "200", "-", "dimensional", "word", "vectors", "do", "not", "show", "significant", "improvements", ".", "Furthermore", ",", "TD", "-", "LSTM", "and", "LSTM", "have", "similar", "time", "cost", ",", "while", "TD", "-", "LSTM", "gets", "higher", "classification", "accuracy", "as", "target", "information", "is", "incorporated", ".", "TC", "-", "LSTM", "performs", "slightly", "better", "than", "TD", "-", "LSTM", "while", "at", "the", "cost", "of", "longer", "training", "time", "because", "the", "parameter", "number", "of", "TC", "-", "LSTM", "is", "larger", ".", "subsection", ":", "Case", "Study", "In", "this", "section", ",", "we", "explore", "to", "what", "extent", "the", "target", "-", "dependent", "LSTM", "models", "including", "TD", "-", "LSTM", "and", "TC", "-", "LSTM", "improve", "the", "performance", "of", "a", "basic", "LSTM", "model", ".", "In", "Table", "[", "reference", "]", ",", "we", "list", "some", "examples", "whose", "polarity", "labels", "are", "incorrectly", "inferred", "by", "LSTM", "but", "correctly", "predicted", "by", "both", "TD", "-", "LSTM", "and", "TC", "-", "LSTM", ".", "We", "observe", "that", "LSTM", "model", "prefers", "to", "assigning", "the", "polarity", "of", "the", "entire", "sentence", "while", "ignoring", "the", "target", "to", "be", "evaluated", ".", "TD", "-", "LSTM", "and", "TC", "-", "LSTM", "could", "take", "into", "account", "of", "target", "information", "to", "some", "extend", ".", "For", "example", ",", "in", "the", "2nd", "example", "the", "opinion", "holder", "expresses", "a", "negative", "opinion", "about", "his", "work", ",", "but", "holds", "a", "neutral", "sentiment", "towards", "the", "target", "“", "lindsay", "lohan", "”", ".", "In", "the", "last", "example", ",", "the", "whole", "sentence", "expresses", "a", "neutral", "sentiment", "while", "it", "holds", "a", "positive", "opinion", "towards", "“", "google", "”", ".", "We", "analyse", "the", "error", "cases", "that", "both", "TD", "-", "LSTM", "and", "TC", "-", "LSTM", "can", "not", "well", "handle", ",", "and", "find", "that", "85.4", "%", "of", "the", "misclassified", "examples", "relate", "to", "neutral", "category", ".", "The", "positive", "instances", "are", "rarely", "misclassified", "as", "negative", ",", "and", "vice", "versa", ".", "A", "example", "of", "errors", "is", ":", "“", "freaky", "friday", "on", "television", "reminding", "me", "to", "think", "wtf", "happened", "to", "lindsay", "lohan", ",", "she", "was", "such", "a", "terrific", "actress", ",", "+", "my", "huge", "crush", "on", "haley", "hudson", ".", "”", ",", "which", "is", "incorrectly", "predicted", "as", "positive", "towards", "target", "“", "indsay", "lohan", "”", "in", "both", "TD", "-", "LSTM", "and", "TC", "-", "LSTM", ".", "subsection", ":", "Discussion", "In", "order", "to", "capture", "the", "semantic", "relatedness", "between", "target", "and", "context", "words", ",", "we", "extend", "TD", "-", "LSTM", "by", "adding", "a", "target", "connection", "component", ".", "One", "could", "also", "try", "other", "extensions", "to", "capture", "the", "connection", "between", "target", "and", "context", "words", ".", "For", "example", ",", "we", "also", "tried", "an", "attention", "-", "based", "LSTM", "model", ",", "which", "is", "inspired", "by", "the", "recent", "success", "of", "attention", "-", "based", "neural", "network", "in", "machine", "translation", "and", "document", "encoding", ".", "We", "implement", "the", "soft", "-", "attention", "mechanism", "to", "enhance", "TD", "-", "LSTM", ".", "We", "incorporate", "two", "attention", "layers", "for", "preceding", "LSTM", "and", "following", "LSTM", ",", "respectively", ".", "The", "output", "vector", "for", "each", "attention", "layer", "is", "the", "weighted", "average", "among", "hidden", "vectors", "of", "LSTM", ",", "where", "the", "weight", "of", "each", "hidden", "vector", "is", "calculated", "with", "a", "feedforward", "neural", "network", ".", "The", "outputs", "of", "preceding", "and", "following", "attention", "models", "are", "concatenated", "and", "fed", "to", "for", "sentiment", "classification", ".", "However", ",", "we", "can", "not", "obtain", "better", "result", "with", "such", "an", "attention", "model", ".", "The", "accuracy", "of", "this", "attention", "model", "is", "slightly", "lower", "than", "the", "standard", "LSTM", "model", "(", "around", "65", "%", ")", ",", "which", "means", "that", "the", "attention", "component", "has", "a", "negative", "impact", "on", "the", "model", ".", "A", "potential", "reason", "might", "be", "that", "the", "attention", "based", "LSTM", "has", "larger", "number", "of", "parameters", ",", "which", "can", "not", "be", "easily", "optimized", "with", "the", "small", "number", "of", "corpus", ".", "section", ":", "Related", "Work", "We", "briefly", "review", "existing", "studies", "on", "target", "-", "dependent", "sentiment", "classification", "and", "neural", "network", "approaches", "for", "sentiment", "classification", "in", "this", "section", ".", "subsection", ":", "Target", "-", "Dependent", "Sentiment", "Classification", "Target", "-", "dependent", "sentiment", "classification", "is", "typically", "regarded", "as", "a", "kind", "of", "text", "classification", "problem", "in", "literature", ".", "Therefore", ",", "standard", "text", "classification", "approach", "such", "as", "feature", "-", "based", "Supported", "Vector", "Machine", "can", "be", "naturally", "employed", "to", "build", "a", "sentiment", "classifier", ".", "Despite", "the", "effectiveness", "of", "feature", "engineering", ",", "it", "is", "labor", "intensive", "and", "unable", "to", "discover", "the", "discriminative", "or", "explanatory", "factors", "of", "data", ".", "To", "handle", "this", "problem", ",", "some", "recent", "studies", "use", "neural", "network", "methods", "and", "encode", "each", "sentence", "in", "continuous", "and", "low", "-", "dimensional", "vector", "space", "without", "feature", "engineering", ".", "Dong", "et", "al", ".", "Dong2014a", "transfer", "a", "dependency", "tree", "of", "a", "sentence", "into", "a", "target", "-", "specific", "recursive", "structure", ",", "and", "get", "higher", "level", "representation", "based", "on", "that", "structure", ".", "Vo", "and", "Zhang", "Vo2015", "use", "rich", "features", "including", "sentiment", "-", "specific", "word", "embedding", "and", "sentiment", "lexicons", ".", "Different", "from", "previous", "studies", ",", "the", "LSTM", "models", "developed", "in", "this", "work", "are", "purely", "data", "-", "driven", ",", "and", "do", "not", "rely", "on", "dependency", "parsing", "results", "or", "external", "sentiment", "lexicons", ".", "subsection", ":", "Neural", "Network", "for", "Sentiment", "Classification", "Neural", "network", "approaches", "have", "shown", "promising", "results", "on", "many", "sentence", "/", "document", "-", "level", "sentiment", "classification", ".", "The", "power", "of", "neural", "model", "lies", "in", "its", "ability", "in", "learning", "continuous", "text", "representation", "from", "data", "without", "any", "feature", "engineering", ".", "For", "sentence", "/", "document", "level", "sentiment", "classification", ",", "previous", "studies", "mostly", "have", "two", "steps", ".", "They", "first", "learn", "continuous", "word", "vector", "embeddings", "from", "data", ".", "Afterwards", ",", "semantic", "compositional", "approaches", "are", "used", "to", "compute", "the", "vector", "of", "a", "sentence", "/", "document", "from", "the", "vectors", "of", "its", "constituents", "based", "on", "the", "principle", "of", "compositionality", ".", "Representative", "compositional", "approaches", "to", "learn", "sentence", "representation", "include", "recursive", "neural", "networks", ",", "convolutional", "neural", "network", ",", "long", "short", "-", "term", "memory", "and", "tree", "-", "structured", "LSTM", ".", "There", "also", "exists", "some", "studies", "focusing", "on", "learning", "continuous", "representation", "of", "documents", ".", "section", ":", "Conclusion", "We", "develop", "target", "-", "specific", "long", "short", "term", "memory", "models", "for", "target", "-", "dependent", "sentiment", "classification", ".", "The", "approach", "captures", "the", "connection", "between", "target", "word", "and", "its", "contexts", "when", "generating", "the", "representation", "of", "a", "sentence", ".", "We", "train", "the", "model", "in", "an", "end", "-", "to", "-", "end", "way", "on", "a", "benchmark", "dataset", ",", "and", "show", "that", "incorporating", "target", "information", "could", "boost", "the", "performance", "of", "a", "long", "short", "-", "term", "memory", "model", ".", "The", "target", "-", "dependent", "LSTM", "model", "obtains", "state", "-", "of", "-", "the", "-", "art", "classification", "accuracy", ".", "section", ":", "Acknowledgements", "We", "greatly", "thank", "Yaming", "Sun", "for", "tremendously", "helpful", "discussions", ".", "This", "work", "was", "supported", "by", "the", "National", "High", "Technology", "Development", "863", "Program", "of", "China", "(", "No", ".", "2015AA015407", ")", ",", "National", "Natural", "Science", "Foundation", "of", "China", "(", "No", ".", "61632011", "and", "No.61273321", ")", ".", "According", "to", "the", "meaning", "given", "to", "this", "role", "by", "Harbin", "Institute", "of", "Technology", ",", "the", "contact", "author", "of", "this", "paper", "is", "Bing", "Qin", ".", "bibliography", ":", "References"], "entity": [[3, 4, ["LSTMs"], "Method"], [5, 10, ["Target", "-", "Dependent", "Sentiment", "Classification"], "Task"], [10, 15, ["Target", "-", "dependent", "sentiment", "classification"], "Task"], [70, 72, ["learning", "system"], "Method"], [80, 87, ["target", "dependent", "long", "short", "-", "term", "memory"], "Method"], [88, 89, ["LSTM"], "Method"], [117, 119, ["sentence", "representation"], "Method"], [121, 122, ["LSTM"], "Method"], [131, 132, ["LSTM"], "Method"], [156, 158, ["syntactic", "parser"], "Method"], [182, 184, ["Sentiment", "analysis"], "Task"], [188, 190, ["opinion", "mining"], "Task"], [196, 199, ["natural", "language", "processing"], "Task"], [200, 202, ["computational", "linguistics"], "Task"], [203, 205, ["Sentiment", "analysis"], "Task"], [240, 245, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [258, 260, ["sentiment", "analysis"], "Task"], [368, 373, ["Target", "-", "dependent", "sentiment", "classification"], "Task"], [380, 383, ["text", "classification", "problem"], "Task"], [391, 393, ["sentiment", "classifiers"], "Method"], [394, 398, ["supervised", "machine", "learning", "approach"], "Method"], [401, 406, ["feature", "based", "Supported", "Vector", "Machine"], "Method"], [407, 410, ["neural", "network", "approaches"], "Method"], [421, 426, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [474, 478, ["feature", "-", "based", "SVM"], "Method"], [481, 483, ["feature", "engineering"], "Method"], [551, 554, ["neural", "network", "approach"], "Method"], [566, 568, ["feature", "engineering"], "Method"], [586, 589, ["neural", "network", "models"], "Method"], [592, 597, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [666, 667, ["backpropagation"], "Method"], [673, 677, ["cross", "-", "entropy", "error"], "Metric"], [678, 681, ["supervised", "sentiment", "classification"], "Method"], [685, 687, ["neural", "model"], "Method"], [688, 693, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [701, 705, ["feature", "-", "based", "SVM"], "Method"], [706, 710, ["adaptive", "recursive", "neural", "network"], "Method"], [711, 716, ["lexicon", "-", "enhanced", "neural", "network"], "Method"], [726, 728, ["syntactic", "parser"], "Method"], [753, 754, ["LSTM"], "Method"], [760, 764, ["target", "-", "dependent", "task"], "Task"], [769, 770, ["LSTM"], "Method"], [787, 792, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [814, 816, ["semantic", "representation"], "Method"], [831, 832, ["LSTM"], "Method"], [840, 848, ["Target", "-", "Dependent", "Long", "Short", "-", "Term", "Memory"], "Method"], [849, 852, ["TD", "-", "LSTM"], "Method"], [859, 862, ["TD", "-", "LSTM"], "Method"], [881, 886, ["Long", "Short", "-", "Term", "Memory"], "Method"], [887, 888, ["LSTM"], "Method"], [896, 901, ["long", "short", "-", "term", "memory"], "Method"], [902, 903, ["LSTM"], "Method"], [906, 911, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [946, 947, ["LSTM"], "Method"], [960, 962, ["semantic", "composition"], "Task"], [966, 968, ["sentiment", "analysis"], "Task"], [998, 1000, ["sentence", "representation"], "Method"], [1010, 1014, ["sentiment", "polarity", "of", "sentence"], "Task"], [1036, 1038, ["word", "embedding"], "Method"], [1047, 1050, ["word", "embedding", "matrix"], "Method"], [1080, 1083, ["embedding", "learning", "algorithms"], "Method"], [1097, 1098, ["LSTM"], "Method"], [1126, 1127, ["LSTM"], "Method"], [1131, 1134, ["recurrent", "neural", "network"], "Method"], [1135, 1136, ["RNN"], "Method"], [1175, 1176, ["RNN"], "Method"], [1178, 1180, ["linear", "layer"], "Method"], [1183, 1188, ["pointwise", "non", "-", "linear", "layer"], "Method"], [1190, 1193, ["hyperbolic", "tangent", "function"], "Method"], [1208, 1209, ["RNN"], "Method"], [1213, 1215, ["gradient", "vanishing"], "Task"], [1216, 1217, ["exploding"], "Task"], [1237, 1239, ["LSTM", "cell"], "Method"], [1262, 1263, ["RNN"], "Method"], [1264, 1266, ["LSTM", "cell"], "Method"], [1299, 1301, ["LSTM", "cell"], "Method"], [1351, 1353, ["sentence", "representation"], "Method"], [1359, 1361, ["linear", "layer"], "Method"], [1404, 1408, ["Target", "-", "Dependent", "LSTM"], "Method"], [1409, 1412, ["TD", "-", "LSTM"], "Method"], [1415, 1417, ["LSTM", "model"], "Method"], [1418, 1423, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [1436, 1438, ["feature", "representation"], "Method"], [1440, 1442, ["sentiment", "classification"], "Task"], [1487, 1489, ["picture", "quality"], "Metric"], [1490, 1492, ["battery", "life"], "Metric"], [1528, 1530, ["LSTM", "model"], "Method"], [1533, 1537, ["target", "-", "dependent", "LSTM"], "Method"], [1538, 1541, ["TD", "-", "LSTM"], "Method"], [1572, 1574, ["feature", "representations"], "Method"], [1575, 1577, ["sentiment", "classification"], "Task"], [1591, 1592, ["accuracy"], "Metric"], [1593, 1598, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [1609, 1612, ["left", "one", "LSTM"], "Method"], [1643, 1644, ["LSTM"], "Method"], [1656, 1657, ["LSTM"], "Method"], [1667, 1668, ["LSTM"], "Method"], [1675, 1676, ["LSTM"], "Method"], [1707, 1709, ["composed", "representation"], "Method"], [1710, 1712, ["sentiment", "classification"], "Task"], [1722, 1723, ["LSTM"], "Method"], [1724, 1725, ["LSTM"], "Method"], [1751, 1752, ["LSTM"], "Method"], [1753, 1754, ["LSTM"], "Method"], [1759, 1763, ["Target", "-", "Connection", "LSTM"], "Method"], [1764, 1767, ["TC", "-", "LSTM"], "Method"], [1770, 1772, ["LSTM", "model"], "Method"], [1773, 1777, ["target", "-", "dependent", "LSTM"], "Method"], [1778, 1781, ["TD", "-", "LSTM"], "Method"], [1795, 1798, ["TD", "-", "LSTM"], "Method"], [1824, 1829, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [1874, 1877, ["TC", "-", "LSTM"], "Method"], [1882, 1885, ["TD", "-", "LSTM"], "Method"], [1888, 1891, ["target", "connection", "component"], "Method"], [1915, 1918, ["TC", "-", "LSTM"], "Method"], [1929, 1932, ["TC", "-", "LSTM"], "Method"], [2055, 2058, ["TD", "-", "LSTM"], "Method"], [2064, 2067, ["TC", "-", "LSTM"], "Method"], [2084, 2087, ["TD", "-", "LSTM"], "Method"], [2103, 2106, ["TC", "-", "LSTM"], "Method"], [2129, 2131, ["Model", "Training"], "Method"], [2133, 2134, ["LSTM"], "Method"], [2135, 2138, ["TD", "-", "LSTM"], "Method"], [2139, 2142, ["TC", "-", "LSTM"], "Method"], [2152, 2155, ["supervised", "learning", "framework"], "Method"], [2157, 2159, ["loss", "function"], "Metric"], [2161, 2165, ["cross", "-", "entropy", "error"], "Metric"], [2166, 2168, ["sentiment", "classification"], "Task"], [2222, 2225, ["back", "-", "propagation"], "Method"], [2235, 2238, ["stochastic", "gradient", "descent"], "Method"], [2248, 2253, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [2278, 2280, ["supervised", "setting"], "Task"], [2354, 2356, ["Evaluation", "metrics"], "Metric"], [2357, 2358, ["accuracy"], "Metric"], [2359, 2363, ["macro", "-", "F1", "score"], "Metric"], [2387, 2390, ["SVM", "-", "indep"], "Method"], [2391, 2393, ["SVM", "classifier"], "Method"], [2427, 2430, ["SVM", "-", "dep"], "Method"], [2440, 2442, ["feature", "representation"], "Method"], [2444, 2446, ["Recursive", "NN"], "Method"], [2448, 2451, ["Recursive", "neural", "network"], "Method"], [2454, 2456, ["feature", "learning"], "Task"], [2467, 2470, ["w", "/", "oE"], "Method"], [2471, 2476, ["AdaRNN", "-", "w", "/", "E"], "Method"], [2477, 2480, ["AdaRNN", "-", "comb"], "Method"], [2484, 2488, ["adaptive", "recursive", "neural", "network"], "Method"], [2501, 2504, ["Target", "-", "dep"], "Method"], [2505, 2507, ["SVM", "classifier"], "Method"], [2522, 2526, ["Target", "-", "dep", "+"], "Method"], [2535, 2537, ["neural", "models"], "Method"], [2544, 2545, ["LSTM"], "Method"], [2546, 2549, ["TD", "-", "LSTM"], "Method"], [2550, 2553, ["TC", "-", "LSTM"], "Method"], [2591, 2593, ["learning", "rate"], "Metric"], [2614, 2617, ["SVM", "-", "indep"], "Method"], [2618, 2621, ["SVM", "-", "dep"], "Method"], [2637, 2639, ["SVM", "classifier"], "Method"], [2640, 2641, ["AdaRNN"], "Method"], [2644, 2647, ["feature", "based", "SVM"], "Method"], [2665, 2668, ["target", "-", "dep"], "Task"], [2686, 2688, ["word", "embeddings"], "Method"], [2723, 2725, ["LSTM", "model"], "Method"], [2745, 2748, ["TD", "-", "LSTM"], "Method"], [2753, 2754, ["LSTM"], "Method"], [2771, 2776, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [2779, 2783, ["target", "-", "connection", "mechanism"], "Method"], [2784, 2787, ["TC", "-", "LSTM"], "Method"], [2850, 2853, ["TC", "-", "LSTM"], "Method"], [2880, 2882, ["word", "embedding"], "Method"], [2888, 2890, ["text", "representation"], "Method"], [2901, 2903, ["word", "embeddings"], "Method"], [2904, 2905, ["LSTM"], "Method"], [2906, 2909, ["TD", "-", "LSTM"], "Method"], [2910, 2913, ["TC", "-", "LSTM"], "Method"], [2929, 2934, ["sentiment", "-", "specific", "word", "embedding"], "Method"], [2935, 2936, ["SSWE"], "Method"], [2938, 2940, ["Glove", "vectors"], "Method"], [2954, 2955, ["SSWE"], "Method"], [2956, 2957, ["SSWE"], "Method"], [2958, 2959, ["SSWE"], "Method"], [2961, 2964, ["embedding", "learning", "algorithms"], "Method"], [2967, 2968, ["SSWE"], "Method"], [2969, 2970, ["SSWE"], "Method"], [2971, 2973, ["word", "embeddings"], "Method"], [2980, 2981, ["SSWE"], "Method"], [3004, 3005, ["SSWE"], "Method"], [3006, 3007, ["SSWE"], "Method"], [3010, 3011, ["SSWE"], "Method"], [3020, 3025, ["target", "-", "independent", "sentiment", "classification"], "Task"], [3036, 3039, ["word", "embedding", "learning"], "Task"], [3041, 3042, ["SSWE"], "Method"], [3043, 3044, ["SSWE"], "Method"], [3051, 3052, ["Glove"], "Method"], [3053, 3054, ["SSWE"], "Method"], [3068, 3070, ["word", "representation"], "Method"], [3075, 3078, ["target", "connection", "model"], "Method"], [3078, 3081, ["TC", "-", "LSTM"], "Method"], [3087, 3089, ["word", "embedding"], "Method"], [3109, 3111, ["time", "cost"], "Metric"], [3158, 3161, ["TD", "-", "LSTM"], "Method"], [3162, 3163, ["LSTM"], "Method"], [3165, 3167, ["time", "cost"], "Metric"], [3169, 3172, ["TD", "-", "LSTM"], "Method"], [3182, 3185, ["TC", "-", "LSTM"], "Method"], [3189, 3192, ["TD", "-", "LSTM"], "Method"], [3198, 3200, ["training", "time"], "Metric"], [3205, 3208, ["TC", "-", "LSTM"], "Method"], [3231, 3234, ["TD", "-", "LSTM"], "Method"], [3235, 3238, ["TC", "-", "LSTM"], "Method"], [3244, 3246, ["LSTM", "model"], "Method"], [3264, 3265, ["LSTM"], "Method"], [3270, 3273, ["TD", "-", "LSTM"], "Method"], [3274, 3277, ["TC", "-", "LSTM"], "Method"], [3281, 3283, ["LSTM", "model"], "Method"], [3300, 3303, ["TD", "-", "LSTM"], "Method"], [3304, 3307, ["TC", "-", "LSTM"], "Method"], [3379, 3382, ["TD", "-", "LSTM"], "Method"], [3383, 3386, ["TC", "-", "LSTM"], "Method"], [3470, 3473, ["TD", "-", "LSTM"], "Method"], [3474, 3477, ["TC", "-", "LSTM"], "Method"], [3496, 3499, ["TD", "-", "LSTM"], "Method"], [3502, 3505, ["target", "connection", "component"], "Method"], [3529, 3534, ["attention", "-", "based", "LSTM", "model"], "Method"], [3543, 3548, ["attention", "-", "based", "neural", "network"], "Method"], [3549, 3551, ["machine", "translation"], "Task"], [3552, 3554, ["document", "encoding"], "Task"], [3558, 3562, ["soft", "-", "attention", "mechanism"], "Method"], [3564, 3567, ["TD", "-", "LSTM"], "Method"], [3571, 3573, ["attention", "layers"], "Method"], [3574, 3576, ["preceding", "LSTM"], "Method"], [3577, 3579, ["following", "LSTM"], "Method"], [3587, 3589, ["attention", "layer"], "Method"], [3597, 3598, ["LSTM"], "Method"], [3610, 3613, ["feedforward", "neural", "network"], "Method"], [3617, 3622, ["preceding", "and", "following", "attention", "models"], "Method"], [3628, 3630, ["sentiment", "classification"], "Task"], [3642, 3644, ["attention", "model"], "Method"], [3646, 3647, ["accuracy"], "Metric"], [3649, 3651, ["attention", "model"], "Method"], [3657, 3659, ["LSTM", "model"], "Method"], [3669, 3671, ["attention", "component"], "Method"], [3686, 3689, ["attention", "based", "LSTM"], "Method"], [3718, 3723, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [3724, 3727, ["neural", "network", "approaches"], "Method"], [3728, 3730, ["sentiment", "classification"], "Task"], [3736, 3741, ["Target", "-", "Dependent", "Sentiment", "Classification"], "Task"], [3741, 3746, ["Target", "-", "dependent", "sentiment", "classification"], "Task"], [3753, 3756, ["text", "classification", "problem"], "Task"], [3762, 3765, ["text", "classification", "approach"], "Method"], [3767, 3773, ["feature", "-", "based", "Supported", "Vector", "Machine"], "Method"], [3780, 3782, ["sentiment", "classifier"], "Method"], [3787, 3789, ["feature", "engineering"], "Method"], [3815, 3818, ["neural", "network", "methods"], "Method"], [3831, 3833, ["feature", "engineering"], "Method"], [3857, 3859, ["level", "representation"], "Method"], [3872, 3877, ["sentiment", "-", "specific", "word", "embedding"], "Method"], [3878, 3880, ["sentiment", "lexicons"], "Method"], [3914, 3916, ["Neural", "Network"], "Method"], [3917, 3919, ["Sentiment", "Classification"], "Task"], [3919, 3922, ["Neural", "network", "approaches"], "Method"], [3928, 3935, ["sentence", "/", "document", "-", "level", "sentiment", "classification"], "Task"], [3939, 3941, ["neural", "model"], "Method"], [3947, 3950, ["continuous", "text", "representation"], "Task"], [3954, 3956, ["feature", "engineering"], "Method"], [3958, 3964, ["sentence", "/", "document", "level", "sentiment", "classification"], "Task"], [3975, 3979, ["continuous", "word", "vector", "embeddings"], "Method"], [3984, 3987, ["semantic", "compositional", "approaches"], "Method"], [4007, 4010, ["principle", "of", "compositionality"], "Method"], [4012, 4014, ["compositional", "approaches"], "Method"], [4016, 4018, ["sentence", "representation"], "Task"], [4019, 4022, ["recursive", "neural", "networks"], "Method"], [4023, 4026, ["convolutional", "neural", "network"], "Method"], [4027, 4032, ["long", "short", "-", "term", "memory"], "Method"], [4033, 4037, ["tree", "-", "structured", "LSTM"], "Method"], [4046, 4050, ["continuous", "representation", "of", "documents"], "Task"], [4056, 4064, ["target", "-", "specific", "long", "short", "term", "memory", "models"], "Method"], [4065, 4070, ["target", "-", "dependent", "sentiment", "classification"], "Task"], [137, 138, ["accuracy"], "Metric"], [143, 144, ["LSTM"], "Method"], [604, 609, ["long", "short", "-", "term", "memory"], "Method"], [610, 611, ["LSTM"], "Method"], [621, 622, ["LSTM"], "Method"], [741, 742, ["accuracy"], "Metric"], [775, 776, ["accuracy"], "Metric"], [801, 806, ["long", "short", "-", "term", "memory"], "Method"], [807, 808, ["LSTM"], "Method"], [1604, 1605, ["LSTM"], "Method"], [1616, 1617, ["LSTM"], "Method"], [1868, 1873, ["long", "short", "-", "term", "memory"], "Method"], [2040, 2045, ["long", "short", "-", "term", "memory"], "Method"], [2633, 2634, ["accuracy"], "Metric"], [2690, 2691, ["LSTM"], "Method"], [2700, 2701, ["LSTM"], "Method"], [2800, 2801, ["accuracy"], "Metric"], [2823, 2824, ["accuracy"], "Metric"], [2830, 2831, ["LSTM"], "Method"], [3107, 3108, ["accuracy"], "Metric"], [3175, 3176, ["accuracy"], "Metric"], [3228, 3229, ["LSTM"], "Method"], [3887, 3888, ["LSTM"], "Method"], [4119, 4124, ["long", "short", "-", "term", "memory"], "Method"], [4130, 4131, ["LSTM"], "Method"], [4141, 4142, ["accuracy"], "Metric"]]} {"doc_id": "0b4b6932d5df74b366d9235b40334bc40d719c72", "text": ["Published", "as", "a", "conference", "paper", "at", "ICLR", "2017", "TEMPORAL", "ENSEMBLING", "FOR", "SEMI", "-", "SUPERVISED", "LEARNING", "section", ":", "ABSTRACT", "In", "this", "paper", ",", "we", "present", "a", "simple", "and", "efficient", "method", "for", "training", "deep", "neural", "networks", "in", "a", "semi", "-", "supervised", "setting", "where", "only", "a", "small", "portion", "of", "training", "data", "is", "labeled", ".", "We", "introduce", "self", "-", "ensembling", ",", "where", "we", "form", "a", "consensus", "prediction", "of", "the", "unknown", "labels", "using", "the", "outputs", "of", "the", "network", "-", "in", "-", "training", "on", "different", "epochs", ",", "and", "most", "importantly", ",", "under", "different", "regularization", "and", "input", "augmentation", "conditions", ".", "This", "ensemble", "prediction", "can", "be", "expected", "to", "be", "a", "better", "predictor", "for", "the", "unknown", "labels", "than", "the", "output", "of", "the", "network", "at", "the", "most", "recent", "training", "epoch", ",", "and", "can", "thus", "be", "used", "as", "a", "target", "for", "training", ".", "Using", "our", "method", ",", "we", "set", "new", "records", "for", "two", "standard", "semi", "-", "supervised", "learning", "benchmarks", ",", "reducing", "the", "(", "non", "-", "augmented", ")", "classification", "error", "rate", "from", "18.44", "%", "to", "7.05", "%", "in", "SVHN", "with", "500", "labels", "and", "from", "18.63", "%", "to", "16.55", "%", "in", "CIFAR", "-", "10", "with", "4000", "labels", ",", "and", "further", "to", "5.12", "%", "and", "12.16", "%", "by", "enabling", "the", "standard", "augmentations", ".", "We", "additionally", "obtain", "a", "clear", "improvement", "in", "CIFAR", "-", "100", "classification", "accuracy", "by", "using", "random", "images", "from", "the", "Tiny", "Images", "dataset", "as", "unlabeled", "extra", "inputs", "during", "training", ".", "Finally", ",", "we", "demonstrate", "good", "tolerance", "to", "incorrect", "labels", ".", "section", ":", "INTRODUCTION", "It", "has", "long", "been", "known", "that", "an", "ensemble", "of", "multiple", "neural", "networks", "generally", "yields", "better", "predictions", "than", "a", "single", "network", "in", "the", "ensemble", ".", "This", "effect", "has", "also", "been", "indirectly", "exploited", "when", "training", "a", "single", "network", "through", "dropout", "[", "reference", "]", ",", "dropconnect", "[", "reference", "]", ",", "or", "stochastic", "depth", "[", "reference", "]", "regularization", "methods", ",", "and", "in", "swapout", "networks", "[", "reference", "]", ",", "where", "training", "always", "focuses", "on", "a", "particular", "subset", "of", "the", "network", ",", "and", "thus", "the", "complete", "network", "can", "be", "seen", "as", "an", "implicit", "ensemble", "of", "such", "trained", "sub", "-", "networks", ".", "We", "extend", "this", "idea", "by", "forming", "ensemble", "predictions", "during", "training", ",", "using", "the", "outputs", "of", "a", "single", "network", "on", "different", "training", "epochs", "and", "under", "different", "regularization", "and", "input", "augmentation", "conditions", ".", "Our", "training", "still", "operates", "on", "a", "single", "network", ",", "but", "the", "predictions", "made", "on", "different", "epochs", "correspond", "to", "an", "ensemble", "prediction", "of", "a", "large", "number", "of", "individual", "sub", "-", "networks", "because", "of", "dropout", "regularization", ".", "This", "ensemble", "prediction", "can", "be", "exploited", "for", "semi", "-", "supervised", "learning", "where", "only", "a", "small", "portion", "of", "training", "data", "is", "labeled", ".", "If", "we", "compare", "the", "ensemble", "prediction", "to", "the", "current", "output", "of", "the", "network", "being", "trained", ",", "the", "ensemble", "prediction", "is", "likely", "to", "be", "closer", "to", "the", "correct", ",", "unknown", "labels", "of", "the", "unlabeled", "inputs", ".", "Therefore", "the", "labels", "inferred", "this", "way", "can", "be", "used", "as", "training", "targets", "for", "the", "unlabeled", "inputs", ".", "Our", "method", "relies", "heavily", "on", "dropout", "regularization", "and", "versatile", "input", "augmentation", ".", "Indeed", ",", "without", "neither", ",", "there", "would", "be", "much", "less", "reason", "to", "place", "confidence", "in", "whatever", "labels", "are", "inferred", "for", "the", "unlabeled", "training", "data", ".", "We", "describe", "two", "ways", "to", "implement", "self", "-", "ensembling", ",", "Π", "-", "model", "and", "temporal", "ensembling", ".", "Both", "approaches", "surpass", "prior", "state", "-", "of", "-", "the", "-", "art", "results", "in", "semi", "-", "supervised", "learning", "by", "a", "considerable", "margin", ".", "We", "furthermore", "observe", "that", "self", "-", "ensembling", "improves", "the", "classification", "accuracy", "in", "fully", "labeled", "cases", "as", "well", ",", "and", "provides", "tolerance", "against", "incorrect", "labels", ".", "The", "recently", "introduced", "transform", "/", "stability", "loss", "of", "[", "reference", "]", "is", "based", "on", "the", "same", "principle", "as", "our", "work", ",", "and", "the", "Π", "-", "model", "can", "be", "seen", "as", "a", "special", "case", "of", "it", ".", "The", "Π", "-", "model", "can", "also", "be", "seen", "as", "a", "simplification", "of", "the", "Γ", "-", "model", "of", "the", "ladder", "network", "by", "[", "reference", "]", ",", "a", "previously", "presented", "network", "architecture", "for", "semi", "-", "supervised", "learning", ".", "Our", "temporal", "ensembling", "method", "has", "connections", "to", "the", "bootstrapping", "method", "of", "[", "reference", "]", "targeted", "for", "training", "with", "noisy", "labels", ".", "1", "|B|", "i∈", "(", "B∩L", ")", "log", "z", "i", "[", "y", "i", "]", "supervised", "loss", "component", "+", "w", "(", "t", ")", "1", "C|B|", "i∈B", "||z", "i", "−z", "i", "||", "2", "unsupervised", "loss", "component", "update", "θ", "using", ",", "e.g.", ",", "ADAM", "update", "network", "parameters", "end", "for", "end", "for", "return", "θ", "section", ":", "SELF", "-", "ENSEMBLING", "DURING", "TRAINING", "We", "present", "two", "implementations", "of", "self", "-", "ensembling", "during", "training", ".", "The", "first", "one", ",", "Π", "-", "model", ",", "encourages", "consistent", "network", "output", "between", "two", "realizations", "of", "the", "same", "input", "stimulus", ",", "under", "two", "different", "dropout", "conditions", ".", "The", "second", "method", ",", "temporal", "ensembling", ",", "simplifies", "and", "extends", "this", "by", "taking", "into", "account", "the", "network", "predictions", "over", "multiple", "previous", "training", "epochs", ".", "We", "shall", "describe", "our", "methods", "in", "the", "context", "of", "traditional", "image", "classification", "networks", ".", "Let", "the", "training", "data", "consist", "of", "total", "of", "N", "inputs", ",", "out", "of", "which", "M", "are", "labeled", ".", "The", "input", "stimuli", ",", "available", "for", "all", "training", "data", ",", "are", "denoted", "x", "i", ",", "where", "i", "∈", "{", "1", ".", ".", ".", "N", "}", ".", "Let", "set", "L", "contain", "the", "indices", "of", "the", "labeled", "inputs", ",", "|L|", "=", "M", ".", "For", "every", "i", "∈", "L", ",", "we", "have", "a", "known", "correct", "label", "y", "i", "∈", "{", "1", ".", ".", ".", "C", "}", ",", "where", "C", "is", "the", "number", "of", "different", "classes", ".", "section", ":", "Π", "-", "MODEL", "The", "structure", "of", "Π", "-", "model", "is", "shown", "in", "Figure", "1", "(", "top", ")", ",", "and", "the", "pseudocode", "in", "Algorithm", "1", ".", "During", "training", ",", "we", "evaluate", "the", "network", "for", "each", "training", "input", "x", "i", "twice", ",", "resulting", "in", "prediction", "vectors", "z", "i", "andz", "i", ".", "Our", "loss", "function", "consists", "of", "two", "components", ".", "The", "first", "component", "is", "the", "standard", "crossentropy", "loss", ",", "evaluated", "for", "labeled", "inputs", "only", ".", "The", "second", "component", ",", "evaluated", "for", "all", "inputs", ",", "penalizes", "different", "predictions", "for", "the", "same", "training", "input", "x", "i", "by", "taking", "the", "mean", "square", "difference", "between", "the", "prediction", "vectors", "z", "i", "andz", "i", ".", "1", "To", "combine", "the", "supervised", "and", "unsupervised", "loss", "terms", ",", "we", "scale", "the", "latter", "by", "time", "-", "dependent", "weighting", "function", "w", "(", "t", ")", ".", "By", "comparing", "the", "entire", "output", "vectors", "z", "i", "andz", "i", ",", "we", "effectively", "ask", "the", "\"", "dark", "knowledge", "\"", "[", "reference", "]", "between", "the", "two", "evaluations", "to", "be", "close", ",", "which", "is", "a", "much", "stronger", "requirement", "compared", "to", "asking", "that", "only", "the", "final", "classification", "remains", "the", "same", ",", "which", "is", "what", "happens", "in", "traditional", "training", ".", "It", "is", "important", "to", "notice", "that", ",", "because", "of", "dropout", "regularization", ",", "the", "network", "output", "during", "training", "is", "a", "stochastic", "variable", ".", "Thus", "two", "evaluations", "of", "the", "same", "input", "x", "i", "under", "same", "network", "weights", "θ", "yield", "different", "results", ".", "In", "addition", ",", "Gaussian", "noise", "and", "augmentations", "such", "as", "random", "translation", "are", "evaluated", "twice", ",", "resulting", "in", "additional", "variation", ".", "The", "combination", "of", "these", "effects", "explains", "the", "difference", "between", "the", "prediction", "vectors", "z", "i", "andz", "i", ".", "This", "difference", "can", "be", "seen", "as", "an", "error", "in", "classification", ",", "given", "that", "the", "original", "input", "x", "i", "was", "the", "same", ",", "and", "thus", "minimizing", "it", "is", "a", "reasonable", "goal", ".", "In", "our", "implementation", ",", "the", "unsupervised", "loss", "weighting", "function", "w", "(", "t", ")", "ramps", "up", ",", "starting", "from", "zero", ",", "along", "a", "Gaussian", "curve", "during", "the", "first", "80", "training", "epochs", ".", "See", "Appendix", "A", "for", "further", "details", "about", "this", "and", "other", "training", "parameters", ".", "In", "the", "beginning", "the", "total", "loss", "and", "the", "learning", "gradients", "are", "thus", "dominated", "by", "the", "supervised", "loss", "component", ",", "i.e.", ",", "the", "labeled", "data", "only", ".", "We", "have", "found", "it", "to", "be", "very", "important", "that", "the", "ramp", "-", "up", "of", "the", "unsupervised", "loss", "component", "is", "slow", "enough", "-", "otherwise", ",", "the", "network", "gets", "easily", "stuck", "in", "a", "degenerate", "solution", "where", "no", "meaningful", "classification", "of", "the", "data", "is", "obtained", ".", "Our", "approach", "is", "somewhat", "similar", "to", "the", "Γ", "-", "model", "of", "the", "ladder", "network", "by", "[", "reference", "]", ",", "but", "conceptually", "simpler", ".", "In", "the", "Π", "-", "model", ",", "the", "comparison", "is", "done", "directly", "on", "network", "outputs", ",", "i.e.", ",", "after", "softmax", "activation", ",", "and", "there", "is", "no", "auxiliary", "mapping", "between", "the", "two", "branches", "such", "as", "the", "learned", "denoising", "functions", "in", "the", "ladder", "network", "architecture", ".", "Furthermore", ",", "instead", "of", "having", "one", "\"", "clean", "\"", "and", "one", "\"", "corrupted", "\"", "branch", "as", "in", "Γ", "-", "model", ",", "we", "apply", "equal", "augmentation", "and", "noise", "to", "the", "inputs", "for", "both", "branches", ".", "As", "shown", "in", "Section", "3", ",", "the", "Π", "-", "model", "combined", "with", "a", "good", "convolutional", "network", "architecture", "provides", "a", "significant", "improvement", "over", "prior", "art", "in", "classification", "accuracy", ".", "section", ":", "TEMPORAL", "ENSEMBLING", "Analyzing", "how", "the", "Π", "-", "model", "works", ",", "we", "could", "equally", "well", "split", "the", "evaluation", "of", "the", "two", "branches", "in", "two", "separate", "phases", ":", "first", "classifying", "the", "training", "set", "once", "without", "updating", "the", "weights", "θ", ",", "and", "then", "training", "the", "network", "on", "the", "same", "inputs", "under", "different", "augmentations", "and", "dropout", ",", "using", "the", "just", "obtained", "predictions", "as", "targets", "for", "the", "unsupervised", "loss", "component", ".", "As", "the", "training", "targets", "obtained", "this", "way", "are", "based", "on", "a", "single", "evaluation", "of", "the", "network", ",", "they", "can", "be", "expected", "to", "be", "noisy", ".", "Temporal", "ensembling", "alleviates", "this", "by", "aggregating", "the", "predictions", "of", "multiple", "previous", "network", "evaluations", "into", "an", "ensemble", "prediction", ".", "It", "also", "lets", "us", "evaluate", "the", "network", "only", "once", "during", "training", ",", "gaining", "an", "approximate", "2x", "speedup", "over", "the", "Π", "-", "model", ".", "The", "structure", "of", "our", "temporal", "ensembling", "method", "is", "shown", "in", "Figure", "1", "(", "bottom", ")", ",", "and", "the", "pseudocode", "in", "Algorithm", "2", ".", "The", "main", "difference", "to", "the", "Π", "-", "model", "is", "that", "the", "network", "and", "augmentations", "are", "evaluated", "only", "once", "per", "input", "per", "epoch", ",", "and", "the", "target", "vectorsz", "for", "the", "unsupervised", "loss", "component", "are", "based", "on", "prior", "network", "evaluations", "instead", "of", "a", "second", "evaluation", "of", "the", "network", ".", "After", "every", "training", "epoch", ",", "the", "network", "outputs", "z", "i", "are", "accumulated", "into", "ensemble", "outputs", "Z", "i", "by", "updating", "Z", "i", "←", "αZ", "i", "+", "(", "1", "−", "α", ")", "z", "i", ",", "where", "α", "is", "a", "momentum", "term", "that", "controls", "how", "far", "the", "ensemble", "reaches", "into", "training", "history", ".", "Because", "of", "dropout", "regularization", "and", "stochastic", "augmentation", ",", "Z", "thus", "contains", "a", "weighted", "average", "of", "the", "outputs", "of", "an", "ensemble", "of", "networks", "f", "from", "previous", "training", "epochs", ",", "with", "recent", "epochs", "having", "larger", "weight", "than", "distant", "epochs", ".", "For", "generating", "the", "training", "targetsz", ",", "we", "need", "to", "correct", "for", "the", "startup", "bias", "in", "Z", "by", "dividing", "by", "factor", "(", "1", "−", "α", "t", ")", ".", "A", "similar", "bias", "correction", "has", "been", "used", "in", ",", "e.g.", ",", "[", "reference", "]", "and", "mean", "-", "only", "batch", "normalization", "[", "reference", "]", ".", "On", "the", "first", "training", "epoch", ",", "Z", "andz", "are", "zero", "as", "no", "data", "from", "previous", "epochs", "is", "available", ".", "For", "this", "reason", ",", "we", "specify", "the", "unsupervised", "weight", "ramp", "-", "up", "function", "w", "(", "t", ")", "to", "also", "be", "zero", "on", "the", "first", "training", "epoch", ".", "Algorithm", "2", "Temporal", "ensembling", "pseudocode", ".", "Note", "that", "the", "updates", "of", "Z", "andz", "could", "equally", "well", "be", "done", "inside", "the", "minibatch", "loop", ";", "in", "this", "pseudocode", "they", "occur", "between", "epochs", "for", "clarity", ".", "Require", ":", "x", "i", "=", "training", "stimuli", "Require", ":", "L", "=", "set", "of", "training", "input", "indices", "with", "known", "labels", "Require", ":", "y", "i", "=", "labels", "for", "labeled", "inputs", "i", "∈", "L", "Require", ":", "α", "=", "ensembling", "momentum", ",", "0", "≤", "α", "<", "1", "Require", ":", "w", "(", "t", ")", "=", "unsupervised", "weight", "ramp", "-", "up", "function", "Require", ":", "f", "θ", "(", "x", ")", "=", "stochastic", "neural", "network", "with", "trainable", "parameters", "θ", "Require", ":", "g", "(", "x", ")", "=", "stochastic", "input", "augmentation", "function", "unsupervised", "loss", "component", "update", "θ", "using", ",", "e.g.", ",", "ADAM", "update", "network", "parameters", "end", "for", "construct", "target", "vectors", "by", "bias", "correction", "end", "for", "return", "θ", "The", "benefits", "of", "temporal", "ensembling", "compared", "to", "Π", "-", "model", "are", "twofold", ".", "First", ",", "the", "training", "is", "faster", "because", "the", "network", "is", "evaluated", "only", "once", "per", "input", "on", "each", "epoch", ".", "Second", ",", "the", "training", "targets", "z", "can", "be", "expected", "to", "be", "less", "noisy", "than", "with", "Π", "-", "model", ".", "As", "shown", "in", "Section", "3", ",", "we", "indeed", "obtain", "somewhat", "better", "results", "with", "temporal", "ensembling", "than", "with", "Π", "-", "model", "in", "the", "same", "number", "of", "training", "epochs", ".", "The", "downside", "compared", "to", "Π", "-", "model", "is", "the", "need", "to", "store", "auxiliary", "data", "across", "epochs", ",", "and", "the", "new", "hyperparameter", "α", ".", "While", "the", "matrix", "Z", "can", "be", "fairly", "large", "when", "the", "dataset", "contains", "a", "large", "number", "of", "items", "and", "categories", ",", "its", "elements", "are", "accessed", "relatively", "infrequently", ".", "Thus", "it", "can", "be", "stored", ",", "e.g.", ",", "in", "a", "memory", "mapped", "file", ".", "An", "intriguing", "additional", "possibility", "of", "temporal", "ensembling", "is", "collecting", "other", "statistics", "from", "the", "network", "predictions", "z", "i", "besides", "the", "mean", ".", "For", "example", ",", "by", "tracking", "the", "second", "raw", "moment", "of", "the", "network", "outputs", ",", "we", "can", "estimate", "the", "variance", "of", "each", "output", "component", "z", "i", ",", "j", ".", "This", "makes", "it", "possible", "to", "reason", "about", "the", "uncertainty", "of", "network", "outputs", "in", "a", "principled", "way", "[", "reference", "]", ".", "Based", "on", "this", "information", ",", "we", "could", ",", "e.g.", ",", "place", "more", "weight", "on", "more", "certain", "predictions", "vs.", "uncertain", "ones", "in", "the", "unsupervised", "loss", "term", ".", "However", ",", "we", "leave", "the", "exploration", "of", "these", "avenues", "as", "future", "work", ".", "section", ":", "RESULTS", "Our", "network", "structure", "is", "given", "in", "Table", "5", ",", "and", "the", "test", "setup", "and", "all", "training", "parameters", "are", "detailed", "in", "Appendix", "A.", "We", "test", "the", "Π", "-", "model", "and", "temporal", "ensembling", "in", "two", "image", "classification", "tasks", ",", "CIFAR", "-", "10", "and", "SVHN", ",", "and", "report", "the", "mean", "and", "standard", "deviation", "of", "10", "runs", "using", "different", "random", "seeds", ".", "Although", "it", "is", "rarely", "stated", "explicitly", ",", "we", "believe", "that", "our", "comparison", "methods", "do", "not", "use", "input", "augmentation", ",", "i.e.", ",", "are", "limited", "to", "dropout", "and", "other", "forms", "of", "permutation", "-", "invariant", "noise", ".", "Therefore", "we", "report", "the", "error", "rates", "without", "augmentation", ",", "unless", "explicitly", "stated", "otherwise", ".", "Given", "that", "the", "ability", "of", "an", "algorithm", "to", "extract", "benefit", "from", "augmentation", "is", "also", "an", "important", "property", ",", "we", "report", "the", "classification", "accuracy", "using", "a", "standard", "set", "of", "augmentations", "as", "well", ".", "In", "purely", "supervised", "training", "the", "de", "facto", "standard", "way", "of", "augmenting", "the", "CIFAR", "-", "10", "dataset", "includes", "horizontal", "flips", "and", "random", "translations", ",", "while", "SVHN", "is", "limited", "to", "random", "translations", ".", "By", "using", "these", "same", "augmentations", "we", "can", "compare", "against", "the", "best", "fully", "supervised", "results", "as", "well", ".", "After", "all", ",", "the", "fully", "supervised", "results", "should", "indicate", "the", "upper", "bound", "of", "obtainable", "accuracy", ".", "[", "reference", "]", "20.40", "±", "0.47", "CatGAN", "[", "reference", "]", "19.58", "±", "0.58", "GAN", "of", "18.63", "±", "2.32", "Π", "-", "model", "16.55", "±", "0.29", "6.90", "±", "0.07", "Π", "-", "model", "with", "augmentation", "12.36", "±", "0.31", "5.56", "±", "0.10", "Temporal", "ensembling", "with", "augmentation", "12.16", "±", "0.24", "5.60", "±", "0.10", "Table", "2", ":", "SVHN", "results", "for", "500", "and", "1000", "labels", ",", "averages", "of", "10", "runs", "(", "4", "runs", "for", "all", "labels", ")", ".", "Error", "rate", "(", "%", ")", "with", "#", "labels", "36.02", "±", "0.10", "Virtual", "Adversarial", "[", "reference", "]", "24.63", "ADGM", "[", "reference", "]", "22.86", "SDGM", "[", "reference", "]", "16.61", "±", "0.24", "GAN", "of", "18.44", "±", "4.8", "8", ".", "section", ":", "CIFAR", "-", "10", "CIFAR", "-", "10", "is", "a", "dataset", "consisting", "of", "32", "×", "32", "pixel", "RGB", "images", "from", "ten", "classes", ".", "Table", "1", "shows", "a", "2.1", "percentage", "point", "reduction", "in", "classification", "error", "rate", "with", "4000", "labels", "(", "400", "per", "class", ")", "compared", "to", "earlier", "methods", "for", "the", "non", "-", "augmented", "Π", "-", "model", ".", "Enabling", "the", "standard", "set", "of", "augmentations", "further", "reduces", "the", "error", "rate", "by", "4.2", "percentage", "points", "to", "12.36", "%", ".", "Temporal", "ensembling", "is", "slightly", "better", "still", "at", "12.16", "%", ",", "while", "being", "twice", "as", "fast", "to", "train", ".", "This", "small", "improvement", "conceals", "the", "subtle", "fact", "that", "random", "horizontal", "flips", "need", "to", "be", "done", "independently", "for", "each", "epoch", "in", "temporal", "ensembling", ",", "while", "Π", "-", "model", "can", "randomize", "once", "per", "a", "pair", "of", "evaluations", ",", "which", "according", "to", "our", "measurements", "is", "∼0.5", "percentage", "points", "better", "than", "independent", "flips", ".", "A", "principled", "comparison", "with", "[", "reference", "]", "is", "difficult", "due", "to", "several", "reasons", ".", "They", "provide", "results", "only", "for", "a", "fairly", "extreme", "set", "of", "augmentations", "(", "translations", ",", "flipping", ",", "rotations", ",", "stretching", ",", "and", "shearing", ")", "on", "top", "of", "fractional", "max", "pooling", "[", "reference", "]", ",", "which", "introduces", "random", ",", "local", "stretching", "inside", "the", "network", ",", "and", "is", "known", "to", "improve", "classification", "results", "substantially", ".", "They", "quote", "an", "error", "rate", "of", "only", "13.60", "%", "for", "supervised", "-", "only", "training", "with", "4000", "labels", ",", "while", "our", "corresponding", "baseline", "is", "34.85", "%", ".", "This", "gap", "indicates", "a", "huge", "benefit", "from", "versatile", "augmentations", "and", "fractional", "max", "pooling", "-", "in", "fact", ",", "their", "baseline", "result", "is", "already", "better", "than", "any", "previous", "semisupervised", "results", ".", "By", "enabling", "semi", "-", "supervised", "learning", "they", "achieve", "a", "17", "%", "drop", "in", "classification", "error", "rate", "(", "from", "13.60", "%", "to", "11.29", "%", ")", ",", "while", "we", "see", "a", "much", "larger", "relative", "drop", "of", "65", "%", "(", "from", "34.85", "%", "to", "12.16", "%", ")", ".", "section", ":", "SVHN", "The", "street", "view", "house", "numbers", "(", "SVHN", ")", "dataset", "consists", "of", "32", "×", "32", "pixel", "RGB", "images", "of", "real", "-", "world", "house", "numbers", ",", "and", "the", "task", "is", "to", "classify", "the", "centermost", "digit", ".", "In", "SVHN", "we", "chose", "to", "use", "only", "the", ".", "Even", "with", "this", "choice", "our", "error", "rate", "with", "all", "labels", "is", "only", "3.05", "%", "without", "augmentation", ".", "Table", "2", "compares", "our", "method", "to", "the", "previous", "state", "-", "of", "-", "the", "-", "art", ".", "With", "the", "most", "commonly", "used", "1000", "labels", "we", "observe", "an", "improvement", "of", "2.7", "percentage", "points", ",", "from", "8.11", "%", "to", "5.43", "%", "without", "augmentation", ",", "and", "further", "to", "4.42", "%", "with", "standard", "augmentations", ".", "We", "also", "investigated", "the", "behavior", "with", "500", "labels", ",", "where", "we", "obtained", "an", "error", "rate", "less", "than", "half", "of", "without", "augmentations", ",", "with", "a", "significantly", "lower", "standard", "deviation", "as", "well", ".", "When", "augmentations", "were", "enabled", ",", "temporal", "ensembling", "further", "reduced", "the", "error", "rate", "to", "5.12", "%", ".", "In", "this", "test", "the", "difference", "between", "Π", "-", "model", "and", "temporal", "ensembling", "was", "quite", "significant", "at", "1.5", "percentage", "points", ".", "In", "[", "reference", "]", "provide", "results", "without", "augmentation", ",", "with", "the", "caveat", "that", "they", "use", "fractional", "max", "pooling", ",", "which", "is", "a", "very", "augmentation", "-", "like", "technique", "due", "to", "the", "random", ",", "local", "stretching", "it", "introduces", "inside", "the", "network", ".", "It", "leads", "to", "a", "superb", "error", "rate", "of", "2.28", "%", "in", "supervisedonly", "training", ",", "while", "our", "corresponding", "baseline", "is", "3.05", "%", "(", "or", "2.88", "%", "with", "translations", ")", ".", "Given", "that", "in", "a", "separate", "experiment", "our", "network", "matched", "the", "best", "published", "result", "for", "non", "-", "augmented", "SVHN", "when", "extra", "data", "is", "used", "(", "1.69", "%", "from", "[", "reference", "]", ")", ",", "this", "gap", "is", "quite", "surprising", ",", "and", "leads", "us", "to", "conclude", "that", "fractional", "max", "pooling", "leads", "to", "a", "powerful", "augmentation", "of", "the", "dataset", ",", "well", "beyond", "what", "simple", "translations", "can", "achieve", ".", "Our", "temporal", "ensembling", "technique", "obtains", "better", "error", "rates", "for", "both", "500", "and", "1000", "labels", "(", "5.12", "%", "and", "4.42", "%", ",", "respectively", ")", "compared", "to", "the", "6.03", "%", "reported", "by", "Sajjadi", "et", "al", ".", "for", "732", "labels", ".", "section", ":", "CIFAR", "-", "100", "AND", "TINY", "IMAGES", "The", "CIFAR", "-", "100", "dataset", "consists", "of", "32", "×", "32", "pixel", "RGB", "images", "from", "a", "hundred", "classes", ".", "We", "are", "not", "aware", "of", "previous", "semi", "-", "supervised", "results", "in", "this", "dataset", ",", "and", "chose", "10000", "labels", "for", "our", "experiments", ".", "Table", "3", "shows", "error", "rates", "of", "43.43", "%", "and", "38.65", "%", "without", "and", "with", "augmentation", ",", "respectively", ".", "These", "correspond", "to", "7.8", "and", "5.9", "percentage", "point", "improvements", "compared", "to", "supervised", "learning", "with", "labeled", "inputs", "only", ".", "We", "ran", "two", "additional", "tests", "using", "unlabeled", "extra", "data", "from", "Tiny", "Images", "dataset", "[", "reference", "]", ":", "one", "with", "randomly", "selected", "500k", "extra", "images", ",", "most", "not", "corresponding", "to", "any", "of", "the", "CIFAR", "-", "100", "categories", ",", "and", "another", "with", "a", "restricted", "set", "of", "237k", "images", "from", "the", "categories", "that", "correspond", "to", "those", "found", "in", "the", "CIFAR", "-", "100", "dataset", "(", "see", "appendix", "A", "for", "details", ")", ".", "The", "results", "are", "shown", "in", "Table", "4", ".", "The", "addition", "of", "randomly", "selected", ",", "unlabeled", "extra", "images", "improved", "the", "error", "rate", "by", "2.7", "percentage", "points", "(", "from", "26.30", "%", "to", "23.63", "%", ")", ",", "indicating", "a", "desirable", "ability", "to", "learn", "from", "random", "natural", "images", ".", "Temporal", "ensembling", "benefited", "much", "more", "from", "the", "extra", "data", "than", "the", "Π", "-", "model", ".", "Interestingly", ",", "restricting", "the", "extra", "data", "to", "categories", "that", "are", "present", "in", "CIFAR", "-", "100", "did", "not", "improve", "With", "standard", "supervised", "training", "(", "left", ")", "the", "classification", "accuracy", "suffers", "when", "even", "a", "small", "portion", "of", "the", "labels", "give", "disinformation", ",", "and", "the", "situation", "worsens", "quickly", "as", "the", "portion", "of", "randomized", "labels", "increases", "to", "50", "%", "or", "more", ".", "On", "the", "other", "hand", ",", "temporal", "ensembling", "(", "right", ")", "shows", "almost", "perfect", "resistance", "to", "disinformation", "when", "half", "of", "the", "labels", "are", "random", ",", "and", "retains", "over", "ninety", "percent", "classification", "accuracy", "even", "when", "80", "%", "of", "the", "labels", "are", "random", ".", "the", "classification", "accuracy", "further", ".", "This", "indicates", "that", "in", "order", "to", "train", "a", "better", "classifier", "by", "adding", "extra", "data", "as", "unlabeled", "inputs", ",", "it", "is", "enough", "to", "have", "the", "extra", "data", "roughly", "in", "the", "same", "space", "as", "the", "actual", "inputs", "-", "in", "our", "case", ",", "natural", "images", ".", "We", "hypothesize", "that", "it", "may", "even", "be", "possible", "to", "use", "properly", "crafted", "synthetic", "data", "as", "unlabeled", "inputs", "to", "obtain", "improved", "classifiers", ".", "In", "order", "to", "keep", "the", "training", "times", "tolerable", ",", "we", "limited", "the", "number", "of", "unlabeled", "inputs", "to", "50k", "per", "epoch", "in", "these", "tests", ",", "i.e.", ",", "on", "every", "epoch", "we", "trained", "using", "all", "50k", "labeled", "inputs", "from", "CIFAR", "-", "100", "and", "50k", "additional", "unlabeled", "inputs", "from", "Tiny", "Images", ".", "The", "50k", "unlabeled", "inputs", "were", "chosen", "randomly", "on", "each", "epoch", "from", "the", "500k", "or", "237k", "extra", "inputs", ".", "In", "temporal", "ensembling", ",", "after", "each", "epoch", "we", "updated", "only", "the", "rows", "of", "Z", "that", "corresponded", "to", "inputs", "used", "on", "that", "epoch", ".", "section", ":", "SUPERVISED", "LEARNING", "When", "all", "labels", "are", "used", "for", "traditional", "supervised", "training", ",", "our", "network", "approximately", "matches", "the", "state", "-", "of", "-", "the", "-", "art", "error", "rate", "for", "a", "single", "model", "in", "CIFAR", "-", "10", "with", "augmentation", "[", "reference", "][", "reference", "]", "at", "6.05", "%", ",", "and", "without", "augmentation", "[", "reference", "]", "at", "7.33", "%", ".", "The", "same", "is", "probably", "true", "for", "SVHN", "as", "well", ",", "but", "there", "the", "best", "published", "results", "rely", "on", "extra", "data", "that", "we", "chose", "not", "to", "use", ".", "Given", "this", "premise", ",", "it", "is", "perhaps", "somewhat", "surprising", "that", "our", "methods", "reduce", "the", "error", "rate", "also", "when", "all", "labels", "are", "used", "(", "Tables", "1", "and", "2", ")", ".", "We", "believe", "that", "this", "is", "an", "indication", "that", "the", "consistency", "requirement", "adds", "a", "degree", "of", "resistance", "to", "ambiguous", "labels", "that", "are", "fairly", "common", "in", "many", "classification", "tasks", ",", "and", "that", "it", "encourages", "features", "to", "be", "more", "invariant", "to", "stochastic", "sampling", ".", "section", ":", "TOLERANCE", "TO", "INCORRECT", "LABELS", "In", "a", "further", "test", "we", "studied", "the", "hypothesis", "that", "our", "methods", "add", "tolerance", "to", "incorrect", "labels", "by", "assigning", "a", "random", "label", "to", "a", "certain", "percentage", "of", "the", "training", "set", "before", "starting", "to", "train", ".", "Figure", "2", "shows", "the", "classification", "error", "graphs", "for", "standard", "supervised", "training", "and", "temporal", "ensembling", ".", "Clearly", "our", "methods", "provide", "considerable", "resistance", "to", "wrong", "labels", ",", "and", "we", "believe", "this", "is", "because", "the", "unsupervised", "loss", "term", "encourages", "the", "mapping", "function", "implemented", "by", "the", "network", "to", "be", "flat", "in", "the", "vicinity", "of", "all", "input", "data", "points", ",", "whereas", "the", "supervised", "loss", "term", "enforces", "the", "mapping", "function", "to", "have", "a", "specific", "value", "in", "the", "vicinity", "of", "the", "labeled", "input", "data", "points", ".", "This", "means", "that", "even", "the", "wrongly", "labeled", "inputs", "play", "a", "role", "in", "shaping", "the", "mapping", "function", "-", "the", "unsupervised", "loss", "term", "smooths", "the", "mapping", "function", "and", "thus", "also", "the", "decision", "boundaries", ",", "effectively", "fusing", "the", "inputs", "into", "coherent", "clusters", ",", "whereas", "the", "excess", "of", "correct", "labels", "in", "each", "class", "is", "sufficient", "for", "locking", "the", "clusters", "to", "the", "right", "output", "vectors", "through", "the", "supervised", "loss", "term", ".", "The", "difference", "to", "classical", "regularizers", "is", "that", "we", "induce", "smoothness", "only", "on", "the", "manifold", "of", "likely", "inputs", "instead", "of", "over", "the", "entire", "input", "domain", ".", "For", "further", "analysis", "about", "the", "importance", "of", "the", "gradient", "of", "the", "mapping", "function", ",", "see", "[", "reference", "]", ".", "section", ":", "RELATED", "WORK", "There", "is", "a", "large", "body", "of", "previous", "work", "on", "semi", "-", "supervised", "learning", "[", "reference", "]", ".", "In", "here", "we", "will", "concentrate", "on", "the", "ones", "that", "are", "most", "directly", "connected", "to", "our", "work", ".", "Γ", "-", "model", "is", "a", "subset", "of", "a", "ladder", "network", "[", "reference", "]", "that", "introduces", "lateral", "connections", "into", "an", "encoder", "-", "decoder", "type", "network", "architecture", ",", "targeted", "at", "semi", "-", "supervised", "learning", ".", "In", "Γ", "-", "model", ",", "all", "but", "the", "highest", "lateral", "connections", "in", "the", "ladder", "network", "are", "removed", ",", "and", "after", "pruning", "the", "unnecessary", "stages", ",", "the", "remaining", "network", "consists", "of", "two", "parallel", ",", "identical", "branches", ".", "One", "of", "the", "branches", "takes", "the", "original", "training", "inputs", ",", "whereas", "the", "other", "branch", "is", "given", "the", "same", "input", "corrupted", "with", "noise", ".", "The", "unsupervised", "loss", "term", "is", "computed", "as", "the", "squared", "difference", "between", "the", "(", "pre", "-", "activation", ")", "output", "of", "the", "clean", "branch", "and", "a", "denoised", "(", "pre", "-", "activation", ")", "output", "of", "the", "corrupted", "branch", ".", "The", "denoised", "estimate", "is", "computed", "from", "the", "output", "of", "the", "corrupted", "branch", "using", "a", "parametric", "nonlinearity", "that", "has", "10", "auxiliary", "trainable", "parameters", "per", "unit", ".", "Our", "Π", "-", "model", "differs", "from", "the", "Γ", "-", "model", "in", "removing", "the", "parametric", "nonlinearity", "and", "denoising", ",", "having", "two", "corrupted", "paths", ",", "and", "comparing", "the", "outputs", "of", "the", "network", "instead", "of", "pre", "-", "activation", "data", "of", "the", "final", "layer", ".", "[", "reference", "]", "recently", "introduced", "a", "new", "loss", "function", "for", "semi", "-", "supervised", "learning", ",", "so", "called", "transform", "/", "stability", "loss", ",", "which", "is", "founded", "on", "the", "same", "principle", "as", "our", "work", ".", "During", "training", ",", "they", "run", "augmentation", "and", "network", "evaluation", "n", "times", "for", "each", "minibatch", ",", "and", "then", "compute", "an", "unsupervised", "loss", "term", "as", "the", "sum", "of", "all", "pairwise", "squared", "distances", "between", "the", "obtained", "n", "network", "outputs", ".", "As", "such", ",", "their", "technique", "follows", "the", "general", "pseudo", "-", "ensemble", "agreement", "(", "PEA", ")", "regularization", "framework", "of", "[", "reference", "]", ".", "In", "addition", ",", "they", "employ", "a", "mutual", "exclusivity", "loss", "term", "[", "reference", "]", ")", "that", "we", "do", "not", "use", ".", "Our", "Π", "-", "model", "can", "be", "seen", "as", "a", "special", "case", "of", "the", "transform", "/", "stability", "loss", "obtained", "by", "setting", "n", "=", "2", ".", "The", "computational", "cost", "of", "training", "with", "transform", "/", "stability", "loss", "increases", "linearly", "as", "a", "function", "of", "n", ",", "whereas", "the", "efficiency", "of", "our", "temporal", "ensembling", "technique", "remains", "constant", "regardless", "of", "how", "large", "effective", "ensemble", "we", "obtain", "via", "the", "averaging", "of", "previous", "epochs", "'", "predictions", ".", "In", "bootstrap", "aggregating", ",", "or", "bagging", ",", "multiple", "networks", "are", "trained", "independently", "based", "on", "subsets", "of", "training", "data", "[", "reference", "]", ".", "This", "results", "in", "an", "ensemble", "that", "is", "more", "stable", "and", "accurate", "than", "the", "individual", "networks", ".", "Our", "approach", "can", "be", "seen", "as", "pulling", "the", "predictions", "from", "an", "implicit", "ensemble", "that", "is", "based", "on", "a", "single", "network", ",", "and", "the", "variability", "is", "a", "result", "of", "evaluating", "it", "under", "different", "dropout", "and", "augmentation", "conditions", "instead", "of", "training", "on", "different", "subsets", "of", "data", ".", "In", "work", "parallel", "to", "ours", ",", "[", "reference", "]", "store", "multiple", "snapshots", "of", "the", "network", "during", "training", ",", "hopefully", "corresponding", "to", "different", "local", "minima", ",", "and", "use", "them", "as", "an", "explicit", "ensemble", ".", "The", "general", "technique", "of", "inferring", "new", "labels", "from", "partially", "labeled", "data", "is", "often", "referred", "to", "as", "bootstrapping", "or", "self", "-", "training", ",", "and", "it", "was", "first", "proposed", "by", "[", "reference", "]", "in", "the", "context", "of", "linguistic", "analysis", ".", "[", "reference", "]", "analyze", "Yarowsky", "'s", "algorithm", "and", "propose", "a", "novel", "graph", "-", "based", "label", "propagation", "approach", ".", "Similarly", ",", "label", "propagation", "methods", "[", "reference", "]", "infer", "labels", "for", "unlabeled", "training", "data", "by", "comparing", "the", "associated", "inputs", "to", "labeled", "training", "inputs", "using", "a", "suitable", "distance", "metric", ".", "Our", "approach", "differs", "from", "this", "in", "two", "important", "ways", ".", "Firstly", ",", "we", "never", "compare", "training", "inputs", "against", "each", "other", ",", "but", "instead", "only", "rely", "on", "the", "unknown", "labels", "remaining", "constant", ",", "and", "secondly", ",", "we", "let", "the", "network", "produce", "the", "likely", "classifications", "for", "the", "unlabeled", "inputs", "instead", "of", "providing", "them", "through", "an", "outside", "process", ".", "In", "addition", "to", "partially", "labeled", "data", ",", "considerable", "amount", "of", "effort", "has", "been", "put", "into", "dealing", "with", "densely", "but", "inaccurately", "labeled", "data", ".", "This", "can", "be", "seen", "as", "a", "semi", "-", "supervised", "learning", "task", "where", "part", "of", "the", "training", "process", "is", "to", "identify", "the", "labels", "that", "are", "not", "to", "be", "trusted", ".", "For", "recent", "work", "in", "this", "area", ",", "see", ",", "e.g.", ",", "[", "reference", "]", "and", "[", "reference", "]", ".", "In", "this", "context", "of", "noisy", "labels", ",", "[", "reference", "]", "presented", "a", "simple", "bootstrapping", "method", "that", "trains", "a", "classifier", "with", "the", "target", "composed", "of", "a", "convex", "combination", "of", "the", "previous", "epoch", "output", "and", "the", "known", "but", "potentially", "noisy", "labels", ".", "Our", "temporal", "ensembling", "differs", "from", "this", "by", "taking", "into", "account", "the", "evaluations", "over", "multiple", "epochs", ".", "Generative", "Adversarial", "Networks", "(", "GAN", ")", "have", "been", "recently", "used", "for", "semi", "-", "supervised", "learning", "with", "promising", "results", "[", "reference", "][", "reference", "][", "reference", "]", ".", "It", "Additive", "Gaussian", "noise", "σ", "=", "0.15", "conv1a", "128", "filters", ",", "3", "×", "3", ",", "pad", "=", "'", "same", "'", ",", "LReLU", "(", "α", "=", "0.1", ")", "conv1b", "128", "filters", ",", "3", "×", "3", ",", "pad", "=", "'", "same", "'", ",", "LReLU", "(", "α", "=", "0.1", ")", "conv1c", "128", "filters", ",", "3", "×", "3", ",", "pad", "=", "'", "same", "'", ",", "LReLU", "(", "α", "=", "0.1", ")", "pool1", "Maxpool", "2", "×", "2", "pixels", "drop1", "Dropout", ",", "p", "=", "0.5", "conv2a", "256", "filters", ",", "3", "×", "3", ",", "pad", "=", "'", "same", "'", ",", "LReLU", "(", "α", "=", "0.1", ")", "conv2b", "256", "filters", ",", "3", "×", "3", ",", "pad", "=", "'", "same", "'", ",", "LReLU", "(", "α", "=", "0.1", ")", "conv2c", "256", "filters", ",", "3", "×", "3", ",", "pad", "=", "'", "same", "'", ",", "LReLU", "(", "α", "=", "0.1", ")", "pool2", "Maxpool", "2", "×", "2", "pixels", "drop2", "Dropout", ",", "p", "=", "0.5", "conv3a", "512", "filters", ",", "3", "×", "3", ",", "pad", "=", "'", "valid", "'", ",", "LReLU", "(", "α", "=", "0.1", ")", "conv3b", "256", "filters", ",", "1", "×", "1", ",", "LReLU", "(", "α", "=", "0.1", ")", "conv3c", "128", "filters", ",", "1", "×", "1", ",", "LReLU", "(", "α", "=", "0.1", ")", "pool3", "Global", "average", "pool", "(", "6", "×", "6", "→", "1×1", "pixels", ")", "dense", "Fully", "connected", "128", "→", "10", "output", "Softmax", "could", "be", "an", "interesting", "avenue", "for", "future", "work", "to", "incorporate", "a", "generative", "component", "to", "our", "solution", ".", "We", "also", "envision", "that", "our", "methods", "could", "be", "applied", "to", "regression", "-", "type", "learning", "tasks", ".", "[", "reference", "]", ".", "All", "data", "layers", "were", "initialized", "following", "[", "reference", "]", ",", "and", "we", "applied", "weight", "normalization", "and", "mean", "-", "only", "batch", "normalization", "[", "reference", "]", "with", "momentum", "0.999", "to", "all", "of", "them", ".", "We", "used", "leaky", "ReLU", "[", "reference", "]", "with", "α", "=", "0.1", "as", "the", "non", "-", "linearity", ",", "and", "chose", "to", "use", "max", "pooling", "instead", "of", "strided", "convolutions", "because", "it", "gave", "consistently", "better", "results", "in", "our", "experiments", ".", "All", "networks", "were", "trained", "using", "Adam", "(", "Kingma", "&", "Ba", ",", "2014", ")", "with", "a", "maximum", "learning", "rate", "of", "λ", "max", "=", "0.003", ",", "except", "for", "temporal", "ensembling", "in", "the", "SVHN", "case", "where", "a", "maximum", "learning", "rate", "of", "λ", "max", "=", "0.001", "worked", "better", ".", "Adam", "momentum", "parameters", "were", "set", "to", "β", "1", "=", "0.9", "and", "β", "2", "=", "0.999", "as", "suggested", "in", "the", "paper", ".", "The", "maximum", "value", "for", "the", "unsupervised", "loss", "component", "was", "set", "to", "w", "max", "·", "M", "/", "N", ",", "where", "M", "is", "the", "number", "of", "labeled", "inputs", "and", "N", "is", "the", "total", "number", "of", "training", "inputs", ".", "For", "Π", "-", "model", "runs", ",", "we", "used", "w", "max", "=", "100", "in", "all", "runs", "except", "for", "CIFAR", "-", "100", "with", "Tiny", "Images", "where", "we", "set", "w", "max", "=", "300", ".", "For", "temporal", "ensembling", "we", "used", "w", "max", "=", "30", "in", "most", "runs", ".", "For", "the", "corrupted", "label", "test", "in", "Section", "3.5", "we", "used", "w", "max", "=", "300", "for", "0", "%", "and", "20", "%", "corruption", ",", "and", "w", "max", "=", "3000", "for", "corruption", "of", "50", "%", "and", "higher", ".", "For", "basic", "CIFAR", "-", "100", "runs", "we", "used", "w", "max", "=", "100", ",", "and", "for", "CIFAR", "-", "100", "with", "Tiny", "Images", "we", "used", "w", "max", "=", "1000", ".", "The", "accumulation", "decay", "constant", "of", "temporal", "ensembling", "was", "set", "to", "α", "=", "0.6", "in", "all", "runs", ".", "In", "all", "runs", "we", "ramped", "up", "both", "the", "learning", "rate", "λ", "and", "unsupervised", "loss", "component", "weight", "w", "during", "the", "first", "80", "epochs", "using", "a", "Gaussian", "ramp", "-", "up", "curve", "exp", "[", "−5", "(", "1", "−", "T", ")", "2", "]", ",", "where", "T", "advances", "linearly", "from", "zero", "to", "one", "during", "the", "ramp", "-", "up", "period", ".", "In", "addition", "to", "ramp", "-", "up", ",", "we", "annealed", "the", "learning", "rate", "λ", "to", "zero", "and", "Adam", "β", "1", "to", "0.5", "during", "the", "last", "50", "epochs", ",", "but", "otherwise", "we", "did", "not", "decay", "them", "during", "training", ".", "The", "ramp", "-", "down", "curve", "was", "similar", "to", "the", "ramp", "-", "up", "curve", "but", "time", "-", "reversed", "and", "with", "a", "scaling", "constant", "of", "12.5", "instead", "of", "5", ".", "All", "networks", "were", "trained", "for", "300", "epochs", "with", "minibatch", "size", "of", "100", ".", "section", ":", "CIFAR", "-", "10", "Following", "previous", "work", "in", "fully", "supervised", "learning", ",", "we", "pre", "-", "processed", "the", "images", "using", "ZCA", "and", "augmented", "the", "dataset", "using", "horizontal", "flips", "and", "random", "translations", ".", "The", "translations", "were", "drawn", "from", "[", "−2", ",", "2", "]", "pixels", ",", "and", "were", "independently", "applied", "to", "both", "branches", "in", "the", "Π", "-", "model", ".", "section", ":", "SVHN", "We", "pre", "-", "processed", "the", "input", "images", "by", "biasing", "and", "scaling", "each", "input", "image", "to", "zero", "mean", "and", "unit", "variance", ".", "We", "used", "only", "the", "73257", "items", "in", "the", "official", "training", "set", ",", "i.e.", ",", "did", "not", "use", "the", "provided", "531131", "extra", "items", ".", "The", "training", "setups", "were", "otherwise", "similar", "to", "CIFAR", "-", "10", "except", "that", "horizontal", "flips", "were", "not", "used", ".", "section", ":", "Implementation", "Our", "implementation", "is", "written", "in", "Python", "using", "Theano", "[", "reference", "]", "and", "[", "reference", "]", ",", "and", "is", "available", "at", "https:", "//", "github.com", "/", "smlaine2", "/", "tempens", ".", "Model", "convergence", "As", "discussed", "in", "Section", "2.1", ",", "a", "slow", "ramp", "-", "up", "of", "the", "unsupervised", "cost", "is", "very", "important", "for", "getting", "the", "models", "to", "converge", ".", "Furthermore", ",", "in", "our", "very", "preliminary", "tests", "with", "250", "labels", "in", "SVHN", "we", "noticed", "that", "optimization", "tended", "to", "explode", "during", "the", "ramp", "-", "up", "period", ",", "and", "we", "eventually", "found", "that", "using", "a", "lower", "value", "for", "Adam", "β", "2", "parameter", "(", "e.g.", ",", "0.99", "instead", "of", "0.999", ")", "seems", "to", "help", "in", "this", "regard", ".", "We", "do", "not", "attempt", "to", "guarantee", "that", "the", "occurrence", "of", "labeled", "inputs", "during", "training", "would", "be", "somehow", "stratified", ";", "with", "bad", "luck", "there", "might", "be", "several", "consecutive", "minibatches", "without", "any", "labeled", "inputs", "when", "the", "label", "density", "is", "very", "low", ".", "Some", "previous", "work", "has", "identified", "this", "as", "a", "weakness", ",", "and", "have", "solved", "the", "issue", "by", "shuffling", "the", "input", "sequences", "in", "such", "a", "way", "that", "stratification", "is", "guaranteed", ",", "e.g.", "[", "reference", "]", "(", "confirmed", "from", "the", "authors", ")", ".", "This", "kind", "of", "stratification", "might", "further", "improve", "the", "convergence", "of", "our", "methods", "as", "well", ".", "Tiny", "Images", ",", "extra", "data", "from", "restricted", "categories", "The", "restricted", "extra", "data", "in", "Section", "3.3", "was", "extracted", "from", "Tiny", "Images", "by", "picking", "all", "images", "with", "labels", "corresponding", "to", "the", "100", "categories", "used", "in", "CIFAR", "-", "100", ".", "As", "the", "Tiny", "Images", "dataset", "does", "not", "contain", "CIFAR", "-", "100", "categories", "aquarium", "fish", "and", "maple", "tree", ",", "we", "used", "images", "with", "labels", "fish", "and", "maple", "instead", ".", "The", "result", "was", "a", "total", "of", "237", "203", "images", "that", "were", "used", "as", "unlabeled", "extra", "data", ".", "Table", "6", "shows", "the", "composition", "of", "this", "extra", "data", "set", ".", "It", "is", "worth", "noting", "that", "the", "CIFAR", "-", "100", "dataset", "itself", "is", "a", "subset", "of", "Tiny", "Images", ",", "and", "we", "did", "not", "explicitly", "prevent", "overlap", "between", "this", "extra", "set", "and", "CIFAR", "-", "100", ".", "This", "led", "to", "approximately", "a", "third", "of", "the", "CIFAR", "-", "100", "training", "and", "test", "images", "being", "present", "as", "unlabeled", "inputs", "in", "the", "extra", "set", ".", "The", "other", "test", "with", "500k", "extra", "entries", "picked", "randomly", "out", "of", "all", "79", "million", "images", "had", "a", "negligible", "overlap", "with", "CIFAR", "-", "100", ".", "section", ":", "section", ":", "ACKNOWLEDGEMENTS", "We", "thank", "the", "anonymous", "reviewers", ",", "Tero", "Karras", ",", "Pekka", "Jänis", ",", "Tim", "Salimans", ",", "Ian", "Goodfellow", ",", "as", "well", "as", "Harri", "Valpola", "and", "his", "colleagues", "at", "Curious", "AI", "for", "valuable", "suggestions", "that", "helped", "to", "improve", "this", "article", ".", "section", ":", "section", ":", ":", "The", "Tiny", "Images", "[", "reference", "]", "labels", "and", "image", "counts", "used", "in", "the", "CIFAR", "-", "100", "plus", "restricted", "extra", "data", "tests", "(", "rightmost", "column", "of", "Table", "4", ")", ".", "Note", "that", "the", "extra", "input", "images", "were", "supplied", "as", "unlabeled", "data", "for", "our", "networks", ",", "and", "the", "labels", "were", "used", "only", "for", "narrowing", "down", "the", "full", "set", "of", "79", "million", "images", ".", "section", ":"], "entity": [[8, 10, ["TEMPORAL", "ENSEMBLING"], "Task"], [11, 15, ["SEMI", "-", "SUPERVISED", "LEARNING"], "Task"], [31, 34, ["deep", "neural", "networks"], "Method"], [36, 40, ["semi", "-", "supervised", "setting"], "Task"], [53, 56, ["self", "-", "ensembling"], "Method"], [61, 63, ["consensus", "prediction"], "Method"], [94, 96, ["ensemble", "prediction"], "Method"], [143, 148, ["semi", "-", "supervised", "learning", "benchmarks"], "Task"], [152, 159, ["non", "-", "augmented", ")", "classification", "error", "rate"], "Metric"], [166, 167, ["SVHN"], "Material"], [178, 181, ["CIFAR", "-", "10"], "Material"], [197, 198, ["augmentations"], "Method"], [247, 252, ["ensemble", "of", "multiple", "neural", "networks"], "Method"], [277, 278, ["dropout"], "Method"], [282, 283, ["dropconnect"], "Method"], [288, 290, ["stochastic", "depth"], "Method"], [293, 295, ["regularization", "methods"], "Method"], [298, 300, ["swapout", "networks"], "Method"], [331, 334, ["sub", "-", "networks"], "Method"], [341, 343, ["ensemble", "predictions"], "Method"], [344, 345, ["training"], "Task"], [385, 387, ["ensemble", "prediction"], "Method"], [393, 396, ["sub", "-", "networks"], "Method"], [398, 400, ["dropout", "regularization"], "Method"], [402, 404, ["ensemble", "prediction"], "Method"], [408, 412, ["semi", "-", "supervised", "learning"], "Task"], [427, 429, ["ensemble", "prediction"], "Method"], [440, 442, ["ensemble", "prediction"], "Task"], [480, 482, ["dropout", "regularization"], "Method"], [483, 486, ["versatile", "input", "augmentation"], "Method"], [518, 521, ["self", "-", "ensembling"], "Method"], [522, 525, ["Π", "-", "model"], "Method"], [526, 528, ["temporal", "ensembling"], "Method"], [542, 546, ["semi", "-", "supervised", "learning"], "Task"], [555, 558, ["self", "-", "ensembling"], "Method"], [560, 562, ["classification", "accuracy"], "Metric"], [579, 583, ["transform", "/", "stability", "loss"], "Metric"], [599, 602, ["Π", "-", "model"], "Method"], [613, 616, ["Π", "-", "model"], "Method"], [625, 628, ["Γ", "-", "model"], "Method"], [630, 632, ["ladder", "network"], "Method"], [640, 642, ["network", "architecture"], "Method"], [643, 647, ["semi", "-", "supervised", "learning"], "Task"], [649, 652, ["temporal", "ensembling", "method"], "Method"], [656, 658, ["bootstrapping", "method"], "Method"], [682, 685, ["supervised", "loss", "component"], "Method"], [699, 704, ["unsupervised", "loss", "component", "update", "θ"], "Method"], [708, 712, ["ADAM", "update", "network", "parameters"], "Method"], [720, 723, ["SELF", "-", "ENSEMBLING"], "Method"], [724, 725, ["TRAINING"], "Task"], [730, 733, ["self", "-", "ensembling"], "Method"], [734, 735, ["training"], "Task"], [740, 743, ["Π", "-", "model"], "Method"], [767, 769, ["temporal", "ensembling"], "Method"], [797, 800, ["image", "classification", "networks"], "Method"], [894, 897, ["Π", "-", "MODEL"], "Method"], [900, 903, ["Π", "-", "model"], "Method"], [944, 946, ["loss", "function"], "Method"], [957, 959, ["crossentropy", "loss"], "Method"], [1015, 1020, ["time", "-", "dependent", "weighting", "function"], "Method"], [1020, 1024, ["w", "(", "t", ")"], "Method"], [1068, 1069, ["classification"], "Task"], [1090, 1092, ["dropout", "regularization"], "Method"], [1097, 1098, ["training"], "Method"], [1167, 1168, ["classification"], "Task"], [1194, 1198, ["unsupervised", "loss", "weighting", "function"], "Method"], [1198, 1202, ["w", "(", "t", ")"], "Method"], [1238, 1239, ["loss"], "Metric"], [1248, 1251, ["supervised", "loss", "component"], "Method"], [1274, 1277, ["unsupervised", "loss", "component"], "Method"], [1327, 1330, ["Π", "-", "model"], "Method"], [1364, 1367, ["ladder", "network", "architecture"], "Method"], [1385, 1388, ["Γ", "-", "model"], "Method"], [1391, 1393, ["equal", "augmentation"], "Method"], [1409, 1412, ["Π", "-", "model"], "Method"], [1416, 1419, ["convolutional", "network", "architecture"], "Method"], [1427, 1429, ["classification", "accuracy"], "Metric"], [1432, 1434, ["TEMPORAL", "ENSEMBLING"], "Method"], [1437, 1440, ["Π", "-", "model"], "Method"], [1494, 1497, ["unsupervised", "loss", "component"], "Method"], [1523, 1525, ["Temporal", "ensembling"], "Method"], [1534, 1536, ["network", "evaluations"], "Method"], [1538, 1540, ["ensemble", "prediction"], "Method"], [1551, 1552, ["training"], "Task"], [1560, 1563, ["Π", "-", "model"], "Method"], [1568, 1571, ["temporal", "ensembling", "method"], "Method"], [1592, 1595, ["Π", "-", "model"], "Method"], [1598, 1601, ["network", "and", "augmentations"], "Method"], [1616, 1619, ["unsupervised", "loss", "component"], "Method"], [1686, 1688, ["dropout", "regularization"], "Method"], [1689, 1691, ["stochastic", "augmentation"], "Method"], [1751, 1753, ["bias", "correction"], "Method"], [1764, 1769, ["mean", "-", "only", "batch", "normalization"], "Method"], [1821, 1824, ["Temporal", "ensembling", "pseudocode"], "Method"], [1902, 1908, ["unsupervised", "weight", "ramp", "-", "up", "function"], "Method"], [1916, 1919, ["stochastic", "neural", "network"], "Method"], [1920, 1922, ["trainable", "parameters"], "Method"], [1930, 1934, ["stochastic", "input", "augmentation", "function"], "Method"], [1934, 1939, ["unsupervised", "loss", "component", "update", "θ"], "Method"], [1943, 1947, ["ADAM", "update", "network", "parameters"], "Method"], [1953, 1955, ["bias", "correction"], "Method"], [1962, 1964, ["temporal", "ensembling"], "Method"], [1966, 1969, ["Π", "-", "model"], "Method"], [1975, 1976, ["training"], "Task"], [2006, 2009, ["Π", "-", "model"], "Method"], [2023, 2025, ["temporal", "ensembling"], "Method"], [2027, 2030, ["Π", "-", "model"], "Method"], [2042, 2045, ["Π", "-", "model"], "Method"], [2107, 2109, ["temporal", "ensembling"], "Task"], [2115, 2117, ["network", "predictions"], "Method"], [2214, 2216, ["network", "structure"], "Method"], [2238, 2241, ["Π", "-", "model"], "Method"], [2242, 2244, ["temporal", "ensembling"], "Method"], [2246, 2249, ["image", "classification", "tasks"], "Task"], [2250, 2253, ["CIFAR", "-", "10"], "Material"], [2254, 2255, ["SVHN"], "Material"], [2295, 2296, ["dropout"], "Method"], [2309, 2311, ["error", "rates"], "Metric"], [2312, 2313, ["augmentation"], "Task"], [2330, 2331, ["augmentation"], "Task"], [2340, 2342, ["classification", "accuracy"], "Metric"], [2347, 2348, ["augmentations"], "Method"], [2353, 2355, ["supervised", "training"], "Task"], [2363, 2367, ["CIFAR", "-", "10", "dataset"], "Material"], [2375, 2376, ["SVHN"], "Material"], [2413, 2414, ["accuracy"], "Metric"], [2421, 2422, ["CatGAN"], "Method"], [2453, 2455, ["Temporal", "ensembling"], "Method"], [2456, 2457, ["augmentation"], "Method"], [2466, 2467, ["SVHN"], "Material"], [2486, 2488, ["Error", "rate"], "Metric"], [2497, 2499, ["Virtual", "Adversarial"], "Method"], [2503, 2504, ["ADGM"], "Method"], [2508, 2509, ["SDGM"], "Method"], [2515, 2516, ["GAN"], "Method"], [2524, 2527, ["CIFAR", "-", "10"], "Material"], [2527, 2530, ["CIFAR", "-", "10"], "Material"], [2554, 2557, ["classification", "error", "rate"], "Metric"], [2583, 2584, ["augmentations"], "Method"], [2587, 2589, ["error", "rate"], "Metric"], [2597, 2599, ["Temporal", "ensembling"], "Method"], [2635, 2637, ["temporal", "ensembling"], "Task"], [2639, 2642, ["Π", "-", "model"], "Method"], [2705, 2708, ["fractional", "max", "pooling"], "Method"], [2727, 2728, ["classification"], "Task"], [2734, 2736, ["error", "rate"], "Metric"], [2741, 2745, ["supervised", "-", "only", "training"], "Task"], [2764, 2766, ["versatile", "augmentations"], "Method"], [2767, 2770, ["fractional", "max", "pooling"], "Method"], [2783, 2784, ["semisupervised"], "Method"], [2788, 2792, ["semi", "-", "supervised", "learning"], "Method"], [2799, 2802, ["classification", "error", "rate"], "Metric"], [2833, 2834, ["SVHN"], "Material"], [2835, 2839, ["street", "view", "house", "numbers"], "Material"], [2840, 2843, ["SVHN", ")", "dataset"], "Material"], [2869, 2870, ["SVHN"], "Material"], [2882, 2884, ["error", "rate"], "Metric"], [2892, 2893, ["augmentation"], "Method"], [2933, 2934, ["augmentation"], "Method"], [2942, 2943, ["augmentations"], "Method"], [2957, 2959, ["error", "rate"], "Metric"], [2964, 2965, ["augmentations"], "Method"], [2980, 2982, ["temporal", "ensembling"], "Method"], [2985, 2987, ["error", "rate"], "Metric"], [2997, 3000, ["Π", "-", "model"], "Method"], [3001, 3003, ["temporal", "ensembling"], "Method"], [3018, 3019, ["augmentation"], "Task"], [3026, 3029, ["fractional", "max", "pooling"], "Method"], [3055, 3058, ["superb", "error", "rate"], "Metric"], [3062, 3064, ["supervisedonly", "training"], "Task"], [3097, 3098, ["SVHN"], "Material"], [3124, 3127, ["fractional", "max", "pooling"], "Method"], [3145, 3148, ["temporal", "ensembling", "technique"], "Method"], [3150, 3152, ["error", "rates"], "Metric"], [3214, 3217, ["semi", "-", "supervised"], "Task"], [3233, 3235, ["error", "rates"], "Metric"], [3244, 3245, ["augmentation"], "Method"], [3259, 3261, ["supervised", "learning"], "Method"], [3353, 3355, ["error", "rate"], "Metric"], [3379, 3381, ["Temporal", "ensembling"], "Method"], [3390, 3393, ["Π", "-", "model"], "Method"], [3414, 3416, ["supervised", "training"], "Method"], [3420, 3422, ["classification", "accuracy"], "Metric"], [3481, 3483, ["classification", "accuracy"], "Metric"], [3494, 3496, ["classification", "accuracy"], "Metric"], [3507, 3508, ["classifier"], "Method"], [3561, 3562, ["classifiers"], "Method"], [3568, 3570, ["training", "times"], "Metric"], [3631, 3633, ["temporal", "ensembling"], "Task"], [3655, 3657, ["SUPERVISED", "LEARNING"], "Task"], [3664, 3666, ["supervised", "training"], "Task"], [3679, 3681, ["error", "rate"], "Metric"], [3686, 3689, ["CIFAR", "-", "10"], "Material"], [3690, 3691, ["augmentation"], "Method"], [3702, 3703, ["augmentation"], "Method"], [3716, 3717, ["SVHN"], "Material"], [3751, 3753, ["error", "rate"], "Metric"], [3791, 3793, ["classification", "tasks"], "Task"], [3804, 3806, ["stochastic", "sampling"], "Method"], [3851, 3854, ["classification", "error", "graphs"], "Metric"], [3856, 3858, ["supervised", "training"], "Task"], [3859, 3861, ["temporal", "ensembling"], "Task"], [3879, 3882, ["unsupervised", "loss", "term"], "Method"], [3904, 3907, ["supervised", "loss", "term"], "Method"], [3944, 3947, ["unsupervised", "loss", "term"], "Method"], [3988, 3991, ["supervised", "loss", "term"], "Method"], [3996, 3997, ["regularizers"], "Method"], [4049, 4053, ["semi", "-", "supervised", "learning"], "Task"], [4074, 4077, ["Γ", "-", "model"], "Method"], [4082, 4084, ["ladder", "network"], "Method"], [4093, 4099, ["encoder", "-", "decoder", "type", "network", "architecture"], "Method"], [4102, 4106, ["semi", "-", "supervised", "learning"], "Task"], [4108, 4111, ["Γ", "-", "model"], "Method"], [4120, 4122, ["ladder", "network"], "Method"], [4167, 4170, ["unsupervised", "loss", "term"], "Metric"], [4203, 4205, ["denoised", "estimate"], "Method"], [4216, 4218, ["parametric", "nonlinearity"], "Method"], [4228, 4231, ["Π", "-", "model"], "Method"], [4234, 4237, ["Γ", "-", "model"], "Method"], [4243, 4244, ["denoising"], "Method"], [4275, 4277, ["loss", "function"], "Method"], [4278, 4282, ["semi", "-", "supervised", "learning"], "Task"], [4285, 4289, ["transform", "/", "stability", "loss"], "Metric"], [4302, 4303, ["training"], "Task"], [4306, 4307, ["augmentation"], "Method"], [4308, 4310, ["network", "evaluation"], "Task"], [4320, 4323, ["unsupervised", "loss", "term"], "Metric"], [4346, 4350, ["pseudo", "-", "ensemble", "agreement"], "Method"], [4351, 4352, ["PEA"], "Method"], [4381, 4384, ["Π", "-", "model"], "Method"], [4405, 4407, ["computational", "cost"], "Metric"], [4408, 4409, ["training"], "Task"], [4410, 4414, ["transform", "/", "stability", "loss"], "Metric"], [4427, 4430, ["temporal", "ensembling", "technique"], "Method"], [4450, 4452, ["bootstrap", "aggregating"], "Task"], [4454, 4455, ["bagging"], "Task"], [4456, 4458, ["multiple", "networks"], "Method"], [4498, 4500, ["implicit", "ensemble"], "Method"], [4569, 4572, ["inferring", "new", "labels"], "Task"], [4581, 4582, ["bootstrapping"], "Method"], [4583, 4586, ["self", "-", "training"], "Task"], [4600, 4602, ["linguistic", "analysis"], "Task"], [4607, 4610, ["Yarowsky", "'s", "algorithm"], "Method"], [4614, 4620, ["graph", "-", "based", "label", "propagation", "approach"], "Method"], [4623, 4626, ["label", "propagation", "methods"], "Method"], [4647, 4649, ["distance", "metric"], "Metric"], [4703, 4705, ["outside", "process"], "Method"], [4735, 4740, ["semi", "-", "supervised", "learning", "task"], "Task"], [4744, 4746, ["training", "process"], "Task"], [4790, 4792, ["bootstrapping", "method"], "Method"], [4795, 4796, ["classifier"], "Method"], [4818, 4820, ["temporal", "ensembling"], "Method"], [4833, 4839, ["Generative", "Adversarial", "Networks", "(", "GAN", ")"], "Method"], [4844, 4848, ["semi", "-", "supervised", "learning"], "Task"], [4952, 4953, ["LReLU"], "Method"], [4992, 4993, ["LReLU"], "Method"], [5024, 5025, ["LReLU"], "Method"], [5077, 5078, ["Softmax"], "Method"], [5089, 5091, ["generative", "component"], "Method"], [5105, 5110, ["regression", "-", "type", "learning", "tasks"], "Task"], [5128, 5130, ["weight", "normalization"], "Method"], [5131, 5136, ["mean", "-", "only", "batch", "normalization"], "Method"], [5149, 5151, ["leaky", "ReLU"], "Method"], [5168, 5170, ["max", "pooling"], "Method"], [5172, 5174, ["strided", "convolutions"], "Method"], [5189, 5190, ["Adam"], "Method"], [5199, 5202, ["maximum", "learning", "rate"], "Metric"], [5210, 5212, ["temporal", "ensembling"], "Task"], [5214, 5215, ["SVHN"], "Material"], [5218, 5221, ["maximum", "learning", "rate"], "Metric"], [5255, 5258, ["unsupervised", "loss", "component"], "Method"], [5287, 5290, ["Π", "-", "model"], "Method"], [5318, 5320, ["temporal", "ensembling"], "Task"], [5332, 5335, ["corrupted", "label", "test"], "Task"], [5350, 5351, ["corruption"], "Metric"], [5358, 5359, ["corruption"], "Metric"], [5418, 5421, ["learning", "rate", "λ"], "Metric"], [5422, 5427, ["unsupervised", "loss", "component", "weight", "w"], "Method"], [5503, 5507, ["ramp", "-", "down", "curve"], "Method"], [5545, 5548, ["CIFAR", "-", "10"], "Material"], [5552, 5555, ["fully", "supervised", "learning"], "Task"], [5563, 5564, ["ZCA"], "Method"], [5596, 5599, ["Π", "-", "model"], "Method"], [5602, 5603, ["SVHN"], "Material"], [5654, 5657, ["CIFAR", "-", "10"], "Material"], [5673, 5674, ["Python"], "Method"], [5696, 5698, ["Model", "convergence"], "Method"], [5711, 5713, ["unsupervised", "cost"], "Metric"], [5734, 5735, ["SVHN"], "Material"], [5738, 5739, ["optimization"], "Task"], [5843, 5844, ["stratification"], "Task"], [5866, 5867, ["convergence"], "Metric"], [209, 211, ["classification", "accuracy"], "Metric"], [686, 690, ["w", "(", "t", ")"], "Method"], [1309, 1312, ["Γ", "-", "model"], "Method"], [1805, 1809, ["w", "(", "t", ")"], "Method"], [1897, 1901, ["w", "(", "t", ")"], "Method"], [2433, 2436, ["Π", "-", "model"], "Method"], [2442, 2445, ["Π", "-", "model"], "Method"], [2574, 2577, ["Π", "-", "model"], "Method"]]} {"doc_id": "0c5c93594647d77dcbad5f3ed71139b186a23a3c", "text": ["Improving", "Neural", "Machine", "Translation", "Models", "with", "Monolingual", "Data", "section", ":", "Abstract", "Neural", "Machine", "Translation", "(", "NMT", ")", "has", "obtained", "state", "-", "of", "-", "the", "art", "performance", "for", "several", "language", "pairs", ",", "while", "only", "using", "parallel", "data", "for", "training", ".", "Targetside", "monolingual", "data", "plays", "an", "important", "role", "in", "boosting", "fluency", "for", "phrasebased", "statistical", "machine", "translation", ",", "and", "we", "investigate", "the", "use", "of", "monolingual", "data", "for", "NMT", ".", "In", "contrast", "to", "previous", "work", ",", "which", "combines", "NMT", "models", "with", "separately", "trained", "language", "models", ",", "we", "note", "that", "encoder", "-", "decoder", "NMT", "architectures", "already", "have", "the", "capacity", "to", "learn", "the", "same", "information", "as", "a", "language", "model", ",", "and", "we", "explore", "strategies", "to", "train", "with", "monolingual", "data", "without", "changing", "the", "neural", "network", "architecture", ".", "By", "pairing", "monolingual", "training", "data", "with", "an", "automatic", "backtranslation", ",", "we", "can", "treat", "it", "as", "additional", "parallel", "training", "data", ",", "and", "we", "obtain", "substantial", "improvements", "on", "the", "WMT", "15", "task", "English↔German", "(", "+", "2.8", "-", "3.7", "BLEU", ")", ",", "and", "for", "the", "low", "-", "resourced", "IWSLT", "14", "task", "Turkish→English", "(", "+", "2.1", "-", "3.4", "BLEU", ")", ",", "obtaining", "new", "state", "-", "of", "-", "the", "-", "art", "results", ".", "We", "also", "show", "that", "fine", "-", "tuning", "on", "in", "-", "domain", "monolingual", "and", "parallel", "data", "gives", "substantial", "improvements", "for", "the", "IWSLT", "15", "task", "English→German", ".", "section", ":", "Introduction", "Neural", "Machine", "Translation", "(", "NMT", ")", "has", "obtained", "state", "-", "of", "-", "the", "art", "performance", "for", "several", "language", "pairs", ",", "while", "only", "using", "parallel", "data", "for", "training", ".", "Target", "-", "side", "monolingual", "data", "plays", "an", "important", "role", "in", "boosting", "fluency", "for", "phrase", "-", "based", "statisti", "-", "The", "research", "presented", "in", "this", "publication", "was", "conducted", "in", "cooperation", "with", "Samsung", "Electronics", "Polska", "sp", ".", "z", "o.o", ".", "-", "Samsung", "R", "&", "D", "Institute", "Poland", ".", "cal", "machine", "translation", ",", "and", "we", "investigate", "the", "use", "of", "monolingual", "data", "for", "NMT", ".", "Language", "models", "trained", "on", "monolingual", "data", "have", "played", "a", "central", "role", "in", "statistical", "machine", "translation", "since", "the", "first", "IBM", "models", "[", "reference", "]", ")", ".", "There", "are", "two", "major", "reasons", "for", "their", "importance", ".", "Firstly", ",", "word", "-", "based", "and", "phrase", "-", "based", "translation", "models", "make", "strong", "independence", "assumptions", ",", "with", "the", "probability", "of", "translation", "units", "estimated", "independently", "from", "context", ",", "and", "language", "models", ",", "by", "making", "different", "independence", "assumptions", ",", "can", "model", "how", "well", "these", "translation", "units", "fit", "together", ".", "Secondly", ",", "the", "amount", "of", "available", "monolingual", "data", "in", "the", "target", "language", "typically", "far", "exceeds", "the", "amount", "of", "parallel", "data", ",", "and", "models", "typically", "improve", "when", "trained", "on", "more", "data", ",", "or", "data", "more", "similar", "to", "the", "translation", "task", ".", "In", "(", "attentional", ")", "encoder", "-", "decoder", "architectures", "for", "neural", "machine", "translation", "[", "reference", "][", "reference", "]", ",", "the", "decoder", "is", "essentially", "an", "RNN", "language", "model", "that", "is", "also", "conditioned", "on", "source", "context", ",", "so", "the", "first", "rationale", ",", "adding", "a", "language", "model", "to", "compensate", "for", "the", "independence", "assumptions", "of", "the", "translation", "model", ",", "does", "not", "apply", ".", "However", ",", "the", "data", "argument", "is", "still", "valid", "in", "NMT", ",", "and", "we", "expect", "monolingual", "data", "to", "be", "especially", "helpful", "if", "parallel", "data", "is", "sparse", ",", "or", "a", "poor", "fit", "for", "the", "translation", "task", ",", "for", "instance", "because", "of", "a", "domain", "mismatch", ".", "In", "contrast", "to", "previous", "work", ",", "which", "integrates", "a", "separately", "trained", "RNN", "language", "model", "into", "the", "NMT", "model", "[", "reference", "]", ",", "we", "explore", "strategies", "to", "include", "monolingual", "training", "data", "in", "the", "training", "process", "without", "changing", "the", "neural", "network", "architecture", ".", "This", "makes", "our", "approach", "applicable", "to", "different", "NMT", "architectures", ".", "The", "main", "contributions", "of", "this", "paper", "are", "as", "follows", ":", "•", "we", "show", "that", "we", "can", "improve", "the", "machine", "translation", "quality", "of", "NMT", "systems", "by", "mixing", "monolingual", "target", "sentences", "into", "the", "training", "set", ".", "•", "we", "investigate", "two", "different", "methods", "to", "fill", "the", "source", "side", "of", "monolingual", "training", "instances", ":", "using", "a", "dummy", "source", "sentence", ",", "and", "using", "a", "source", "sentence", "obtained", "via", "backtranslation", ",", "which", "we", "call", "synthetic", ".", "We", "find", "that", "the", "latter", "is", "more", "effective", ".", "•", "we", "successfully", "adapt", "NMT", "models", "to", "a", "new", "domain", "by", "fine", "-", "tuning", "with", "either", "monolingual", "or", "parallel", "in", "-", "domain", "data", ".", "section", ":", "Neural", "Machine", "Translation", "We", "follow", "the", "neural", "machine", "translation", "architecture", "by", "[", "reference", "]", ",", "which", "we", "will", "briefly", "summarize", "here", ".", "However", ",", "we", "note", "that", "our", "approach", "is", "not", "specific", "to", "this", "architecture", ".", "The", "neural", "machine", "translation", "system", "is", "implemented", "as", "an", "encoder", "-", "decoder", "network", "with", "recurrent", "neural", "networks", ".", "The", "encoder", "is", "a", "bidirectional", "neural", "network", "with", "gated", "recurrent", "units", "[", "reference", "]", "that", "reads", "an", "input", "sequence", "x", "=", "(", "x", "1", ",", "...", ",", "x", "m", ")", "and", "calculates", "a", "forward", "sequence", "of", "hidden", "states", "(", "−", "→", "h", "1", ",", "...", ",", "−", "→", "h", "m", ")", ",", "and", "a", "backward", "sequence", "The", "hidden", "states", "−", "→", "h", "j", "and", "←", "−", "h", "j", "are", "concatenated", "to", "obtain", "the", "annotation", "vector", "h", "j", ".", "The", "decoder", "is", "a", "recurrent", "neural", "network", "that", "predicts", "a", "target", "sequence", "y", "=", "(", "y", "1", ",", "...", ",", "y", "n", ")", ".", "Each", "word", "y", "i", "is", "predicted", "based", "on", "a", "recurrent", "hidden", "state", "s", "i", ",", "the", "previously", "predicted", "word", "y", "i−1", ",", "and", "a", "context", "vector", "c", "i", ".", "c", "i", "is", "computed", "as", "a", "weighted", "sum", "of", "the", "annotations", "h", "j", ".", "The", "weight", "of", "each", "annotation", "h", "j", "is", "computed", "through", "an", "alignment", "model", "α", "ij", ",", "which", "models", "the", "probability", "that", "y", "i", "is", "aligned", "to", "x", "j", ".", "The", "alignment", "model", "is", "a", "singlelayer", "feedforward", "neural", "network", "that", "is", "learned", "jointly", "with", "the", "rest", "of", "the", "network", "through", "backpropagation", ".", "A", "detailed", "description", "can", "be", "found", "in", "[", "reference", "]", ".", "Training", "is", "performed", "on", "a", "parallel", "corpus", "with", "stochastic", "gradient", "descent", ".", "For", "translation", ",", "a", "beam", "search", "with", "small", "beam", "size", "is", "employed", ".", "section", ":", "NMT", "Training", "with", "Monolingual", "Training", "Data", "In", "machine", "translation", ",", "more", "monolingual", "data", "(", "or", "monolingual", "data", "more", "similar", "to", "the", "test", "set", ")", "serves", "to", "improve", "the", "estimate", "of", "the", "prior", "probability", "p", "(", "T", ")", "of", "the", "target", "sentence", "T", ",", "before", "taking", "the", "source", "sentence", "S", "into", "account", ".", "In", "contrast", "to", "[", "reference", "]", ",", "who", "train", "separate", "language", "models", "on", "monolingual", "training", "data", "and", "incorporate", "them", "into", "the", "neural", "network", "through", "shallow", "or", "deep", "fusion", ",", "we", "propose", "techniques", "to", "train", "the", "main", "NMT", "model", "with", "monolingual", "data", ",", "exploiting", "the", "fact", "that", "encoder", "-", "decoder", "neural", "networks", "already", "condition", "the", "probability", "distribution", "of", "the", "next", "target", "word", "on", "the", "previous", "target", "words", ".", "We", "describe", "two", "strategies", "to", "do", "this", ":", "providing", "monolingual", "training", "examples", "with", "an", "empty", "(", "or", "dummy", ")", "source", "sentence", ",", "or", "providing", "monolingual", "training", "data", "with", "a", "synthetic", "source", "sentence", "that", "is", "obtained", "from", "automatically", "translating", "the", "target", "sentence", "into", "the", "source", "language", ",", "which", "we", "will", "refer", "to", "as", "back", "-", "translation", ".", "section", ":", "Dummy", "Source", "Sentences", "The", "first", "technique", "we", "employ", "is", "to", "treat", "monolingual", "training", "examples", "as", "parallel", "examples", "with", "empty", "source", "side", ",", "essentially", "adding", "training", "examples", "whose", "context", "vector", "c", "i", "is", "uninformative", ",", "and", "for", "which", "the", "network", "has", "to", "fully", "rely", "on", "the", "previous", "target", "words", "for", "its", "prediction", ".", "This", "could", "be", "conceived", "as", "a", "form", "of", "dropout", "[", "reference", "]", ",", "with", "the", "difference", "that", "the", "training", "instances", "that", "have", "the", "context", "vector", "dropped", "out", "constitute", "novel", "training", "data", ".", "We", "can", "also", "conceive", "of", "this", "setup", "as", "multi", "-", "task", "learning", ",", "with", "the", "two", "tasks", "being", "translation", "when", "the", "source", "is", "known", ",", "and", "language", "modelling", "when", "it", "is", "unknown", ".", "During", "training", ",", "we", "use", "both", "parallel", "and", "monolingual", "training", "examples", "in", "the", "ratio", "1", "-", "to", "-", "1", ",", "and", "randomly", "shuffle", "them", ".", "We", "define", "an", "epoch", "as", "one", "iteration", "through", "the", "parallel", "data", "set", ",", "and", "resample", "from", "the", "monolingual", "data", "set", "for", "every", "epoch", ".", "We", "pair", "monolingual", "sentences", "with", "a", "single", "-", "word", "dummy", "source", "side", "<", "null", ">", "to", "allow", "processing", "of", "both", "parallel", "and", "monolingual", "training", "examples", "with", "the", "same", "network", "graph", ".", "1", "For", "monolingual", "minibatches", "2", ",", "we", "freeze", "the", "network", "parameters", "of", "the", "encoder", "and", "the", "attention", "model", ".", "One", "problem", "with", "this", "integration", "of", "monolin", "-", "gual", "data", "is", "that", "we", "can", "not", "arbitrarily", "increase", "the", "ratio", "of", "monolingual", "training", "instances", ",", "or", "finetune", "a", "model", "with", "only", "monolingual", "training", "data", ",", "because", "different", "output", "layer", "parameters", "are", "optimal", "for", "the", "two", "tasks", ",", "and", "the", "network", "'", "unlearns", "'", "its", "conditioning", "on", "the", "source", "context", "if", "the", "ratio", "of", "monolingual", "training", "instances", "is", "too", "high", ".", "section", ":", "Synthetic", "Source", "Sentences", "To", "ensure", "that", "the", "output", "layer", "remains", "sensitive", "to", "the", "source", "context", ",", "and", "that", "good", "parameters", "are", "not", "unlearned", "from", "monolingual", "data", ",", "we", "propose", "to", "pair", "monolingual", "training", "instances", "with", "a", "synthetic", "source", "sentence", "from", "which", "a", "context", "vector", "can", "be", "approximated", ".", "We", "obtain", "these", "through", "back", "-", "translation", ",", "i.e.", "an", "automatic", "translation", "of", "the", "monolingual", "target", "text", "into", "the", "source", "language", ".", "During", "training", ",", "we", "mix", "synthetic", "parallel", "text", "into", "the", "original", "(", "human", "-", "translated", ")", "parallel", "text", "and", "do", "not", "distinguish", "between", "the", "two", ":", "no", "network", "parameters", "are", "frozen", ".", "Importantly", ",", "only", "the", "source", "side", "of", "these", "additional", "training", "examples", "is", "synthetic", ",", "and", "the", "target", "side", "comes", "from", "the", "monolingual", "corpus", ".", "section", ":", "Evaluation", "We", "evaluate", "NMT", "training", "on", "parallel", "text", ",", "and", "with", "additional", "monolingual", "data", ",", "on", "English↔German", "and", "Turkish→English", ",", "using", "training", "and", "test", "data", "from", "WMT", "15", "for", "English↔German", ",", "IWSLT", "15", "for", "English→German", ",", "and", "IWSLT", "14", "for", "Turkish→English", ".", "section", ":", "Data", "and", "Methods", "We", "use", "Groundhog", "3", "as", "the", "implementation", "of", "the", "NMT", "system", "for", "all", "experiments", "[", "reference", "][", "reference", "]", ".", "We", "generally", "follow", "the", "settings", "and", "training", "procedure", "described", "by", "[", "reference", "]", ".", "For", "English↔German", ",", "we", "report", "case", "-", "sensitive", "BLEU", "on", "detokenized", "text", "with", "mteval", "-", "v13a.pl", "for", "comparison", "to", "official", "WMT", "and", "IWSLT", "results", ".", "For", "Turkish→English", ",", "we", "report", "case", "-", "sensitive", "BLEU", "on", "tokenized", "text", "with", "multi", "-", "bleu.perl", "for", "comparison", "to", "results", "by", "[", "reference", "]", ".", "[", "reference", "]", "determine", "the", "network", "vocabulary", "based", "on", "the", "parallel", "training", "data", ",", "and", "replace", "out", "-", "of", "-", "vocabulary", "words", "with", "a", "special", "UNK", "symbol", ".", "They", "remove", "monolingual", "sentences", "with", "more", "than", "10", "%", "UNK", "symbols", ".", "In", "contrast", ",", "we", "represent", "unseen", "words", "as", "sequences", "of", "subword", "units", "[", "reference", "]", ",", "and", "can", "represent", "any", "additional", "training", "data", "with", "the", "existing", "network", "vocabulary", "that", "was", "learned", "on", "the", "parallel", "data", ".", "In", "all", "experiments", ",", "the", "network", "vocabulary", "remains", "fixed", ".", "section", ":", "English↔German", "We", "use", "all", "parallel", "training", "data", "provided", "by", "WMT", "2015", "[", "reference", "]", "4", ".", "We", "use", "the", "News", "Crawl", "corpora", "as", "additional", "training", "data", "for", "the", "experiments", "with", "monolingual", "data", ".", "The", "amount", "of", "training", "data", "is", "shown", "in", "Table", "1", ".", "Baseline", "models", "are", "trained", "for", "a", "week", ".", "Ensembles", "are", "sampled", "from", "the", "last", "4", "saved", "models", "of", "training", "(", "saved", "at", "12h", "-", "intervals", ")", ".", "Each", "model", "is", "fine", "-", "tuned", "with", "fixed", "embeddings", "for", "12", "hours", ".", "For", "the", "experiments", "with", "synthetic", "parallel", "data", ",", "we", "back", "-", "translate", "a", "random", "sample", "of", "3", "600", "000", "sentences", "from", "the", "German", "monolingual", "data", "set", "into", "English", ".", "The", "German→English", "system", "used", "for", "this", "is", "the", "baseline", "system", "(", "parallel", ")", ".", "Translation", "took", "about", "a", "week", "on", "an", "NVIDIA", "Titan", "Black", "GPU", ".", "For", "experiments", "in", "German→English", ",", "we", "back", "-", "translate", "4", "200", "000", "monolingual", "English", "sentences", "into", "German", ",", "using", "the", "English→German", "system", "+", "synthetic", ".", "Note", "that", "we", "always", "use", "single", "models", "for", "backtranslation", ",", "not", "ensembles", ".", "We", "leave", "it", "to", "future", "work", "to", "explore", "how", "sensitive", "NMT", "training", "with", "synthetic", "data", "is", "to", "the", "quality", "of", "the", "backtranslation", ".", "We", "tokenize", "and", "truecase", "the", "training", "data", ",", "and", "represent", "rare", "words", "via", "BPE", "[", "reference", "]", ".", "Specifically", ",", "we", "follow", "[", "reference", "]", "in", "performing", "BPE", "on", "the", "joint", "vocabulary", "with", "89", "500", "merge", "operations", ".", "dataset", "sentences", "WIT", "160", "000", "SETimes", "160", "000", "Gigaword", "mono", "177", "000", "000", "Gigaword", "synth", "3", "200", "000", "Table", "2", ":", "Turkish→English", "training", "data", ".", "The", "network", "vocabulary", "size", "is", "90", "000", ".", "We", "also", "perform", "experiments", "on", "the", "IWSLT", "15", "test", "sets", "to", "investigate", "a", "cross", "-", "domain", "setting", ".", "[", "reference", "]", "The", "test", "sets", "consist", "of", "TED", "talk", "transcripts", ".", "As", "indomain", "training", "data", ",", "IWSLT", "provides", "the", "WIT", "3", "parallel", "corpus", "[", "reference", "]", ",", "which", "also", "consists", "of", "TED", "talks", ".", "section", ":", "Turkish→English", "We", "use", "data", "provided", "for", "the", "IWSLT", "14", "machine", "translation", "track", "[", "reference", "]", ",", "namely", "the", "WIT", "3", "parallel", "corpus", "[", "reference", "]", ",", "which", "consists", "of", "TED", "talks", ",", "and", "the", "SETimes", "corpus", "(", "Tyers", "and", "Alperen", ",", "2010", ")", ".", "[", "reference", "]", "After", "removal", "of", "sentence", "pairs", "which", "contain", "empty", "lines", "or", "lines", "with", "a", "length", "ratio", "above", "9", ",", "we", "retain", "320", "000", "sentence", "pairs", "of", "training", "data", ".", "For", "the", "experiments", "with", "monolingual", "training", "data", ",", "we", "use", "the", "English", "LDC", "Gigaword", "corpus", "(", "Fifth", "Edition", ")", ".", "The", "amount", "of", "training", "data", "is", "shown", "in", "Table", "2", ".", "With", "only", "320", "000", "sentences", "of", "parallel", "data", "available", "for", "training", ",", "this", "is", "a", "much", "lower", "-", "resourced", "translation", "setting", "than", "English↔German", ".", "[", "reference", "]", "segment", "the", "Turkish", "text", "with", "the", "morphology", "tool", "Zemberek", ",", "followed", "by", "a", "disambiguation", "of", "the", "morphological", "analysis", "[", "reference", "]", ",", "and", "removal", "of", "non", "-", "surface", "tokens", "produced", "by", "the", "analysis", ".", "We", "use", "the", "same", "preprocessing", "7", ".", "For", "both", "Turkish", "and", "English", ",", "we", "represent", "rare", "words", "(", "or", "morphemes", "in", "the", "case", "of", "Turkish", ")", "as", "character", "bigram", "sequences", "[", "reference", "]", ".", "The", "20", "000", "most", "frequent", "words", "(", "morphemes", ")", "are", "left", "unsegmented", ".", "The", "networks", "have", "a", "vocabulary", "size", "of", "23", "000", "symbols", ".", "To", "obtain", "a", "synthetic", "parallel", "training", "set", ",", "we", "back", "-", "translate", "a", "random", "sample", "of", "3", "200", "000", "sentences", "from", "Gigaword", ".", "We", "use", "an", "English→Turkish", "NMT", "system", "trained", "with", "the", "same", "settings", "as", "the", "Turkish→English", "baseline", "system", ".", "We", "found", "overfitting", "to", "be", "a", "bigger", "problem", "than", "with", "the", "larger", "English↔German", "data", "set", ",", "and", "follow", "[", "reference", "]", "in", "using", "Gaussian", "noise", "(", "stddev", "0.01", ")", "[", "reference", "]", ",", "and", "dropout", "on", "the", "output", "layer", "(", "p=0.5", ")", "[", "reference", "]", ".", "We", "also", "use", "early", "stopping", ",", "based", "on", "BLEU", "measured", "every", "three", "hours", "on", "tst2010", ",", "which", "we", "treat", "as", "development", "set", ".", "For", "Turkish→English", ",", "we", "use", "gradient", "clipping", "with", "threshold", "5", ",", "following", "[", "reference", "]", ",", "in", "contrast", "to", "the", "threshold", "1", "that", "we", "use", "for", "English↔German", ",", "following", "[", "reference", "]", ".", "Table", "3", "shows", "English→German", "results", "with", "WMT", "training", "and", "test", "data", ".", "We", "find", "that", "mixing", "parallel", "training", "data", "with", "monolingual", "data", "with", "a", "dummy", "source", "side", "in", "a", "ratio", "of", "1", "-", "1", "improves", "quality", "by", "0.4", "-", "0.5", "BLEU", "for", "the", "single", "system", ",", "1", "BLEU", "for", "the", "ensemble", ".", "We", "train", "the", "system", "for", "twice", "as", "long", "as", "the", "baseline", "to", "provide", "the", "training", "algorithm", "with", "a", "similar", "amount", "of", "parallel", "training", "instances", ".", "To", "ensure", "that", "the", "quality", "improvement", "is", "due", "to", "the", "monolingual", "training", "instances", ",", "and", "not", "just", "increased", "training", "time", ",", "we", "also", "continued", "training", "our", "baseline", "system", "for", "another", "week", ",", "but", "saw", "no", "improvements", "in", "BLEU", ".", "section", ":", "Results", "section", ":", "English→German", "WMT", "15", "Including", "synthetic", "data", "during", "training", "is", "very", "effective", ",", "and", "yields", "an", "improvement", "over", "our", "baseline", "by", "2.8", "-", "3.4", "BLEU", ".", "Our", "best", "ensemble", "system", "also", "outperforms", "a", "syntaxbased", "baseline", "[", "reference", "]", ")", "by", "1.2", "-", "2.1", "BLEU", ".", "We", "also", "substantially", "outperform", "NMT", "results", "reported", "by", "[", "reference", "]", "and", ",", "who", "previously", "reported", "SOTA", "result", ".", "[", "reference", "]", "We", "note", "that", "the", "difference", "is", "particularly", "large", "for", "single", "systems", ",", "since", "our", "ensemble", "is", "not", "as", "diverse", "as", "that", "of", ",", "who", "used", "8", "independently", "trained", "ensemble", "components", ",", "whereas", "we", "sampled", "4", "ensemble", "components", "from", "the", "same", "training", "run", ".", "test", "sets", ",", "which", "are", "news", "texts", ".", "We", "investigate", "if", "monolingual", "training", "data", "is", "especially", "valuable", "if", "it", "can", "be", "used", "to", "adapt", "a", "model", "to", "a", "new", "genre", "or", "domain", ",", "specifically", "adapting", "a", "system", "trained", "on", "WMT", "data", "to", "translating", "TED", "talks", ".", "Systems", "1", "and", "2", "correspond", "to", "systems", "in", "Table", "3", ",", "trained", "only", "on", "WMT", "data", ".", "System", "2", ",", "trained", "on", "parallel", "and", "synthetic", "WMT", "data", ",", "obtains", "a", "BLEU", "score", "of", "25.5", "on", "tst2015", ".", "We", "observe", "that", "even", "a", "small", "amount", "of", "fine", "-", "tuning", "9", ",", "i.e.", "continued", "training", "of", "an", "existing", "model", ",", "on", "WIT", "data", "can", "adapt", "a", "system", "trained", "on", "WMT", "data", "to", "the", "TED", "domain", ".", "By", "back", "-", "translating", "the", "monolingual", "WIT", "corpus", "(", "using", "a", "German→English", "system", "trained", "on", "WMT", "data", ",", "i.e.", "without", "in", "-", "domain", "knowledge", ")", ",", "we", "obtain", "the", "synthetic", "data", "set", "WIT", "synth", ".", "A", "single", "epoch", "of", "fine", "-", "tuning", "on", "WIT", "synth", "(", "system", "4", ")", "results", "in", "a", "BLEU", "score", "of", "26.7", "on", "tst2015", ",", "or", "an", "improvement", "of", "1.2", "BLEU", ".", "We", "observed", "no", "improvement", "from", "fine", "-", "tuning", "on", "WIT", "mono", ",", "the", "monolingual", "TED", "corpus", "with", "dummy", "input", "(", "system", "3", ")", ".", "section", ":", "English→German", "IWSLT", "15", "These", "adaptation", "experiments", "with", "monolingual", "data", "are", "slightly", "artificial", "in", "that", "parallel", "training", "data", "is", "available", ".", "System", "5", ",", "which", "is", "finetuned", "with", "the", "original", "WIT", "training", "data", ",", "obtains", "a", "BLEU", "of", "28.4", "on", "tst2015", ",", "which", "is", "an", "improve", "-", "[", "reference", "]", "We", "leave", "the", "word", "embeddings", "fixed", "for", "fine", "-", "tuning", ".", "BLEU", "name", "2014", "2015", "PBSMT", "28.8", "29.3", "NMT", "[", "reference", "]", "23.6", "-+", "shallow", "fusion", "23.7", "-+", "deep", "fusion", "24.0", "-", "parallel", "25.9", "26.7", "+", "synthetic", "29.5", "30.4", "+", "synthetic", "(", "ensemble", "of", "4", ")", "30.8", "31.6", "Table", "5", ":", "German→English", "translation", "performance", "(", "BLEU", ")", "on", "WMT", "training", "/", "test", "sets", "(", "newstest2014", ";", "newstest2015", ")", ".", "ment", "of", "2.9", "BLEU", ".", "While", "it", "is", "unsurprising", "that", "in", "-", "domain", "parallel", "data", "is", "most", "valuable", ",", "we", "find", "it", "encouraging", "that", "NMT", "domain", "adaptation", "with", "monolingual", "data", "is", "also", "possible", ",", "and", "effective", ",", "since", "there", "are", "settings", "where", "only", "monolingual", "in", "-", "domain", "data", "is", "available", ".", "The", "best", "results", "published", "on", "this", "dataset", "are", "by", ",", "obtained", "with", "an", "ensemble", "of", "8", "independently", "trained", "models", ".", "In", "a", "comparison", "of", "single", "-", "model", "results", ",", "we", "outperform", "their", "model", "on", "tst2013", "by", "1", "BLEU", ".", "section", ":", "German→English", "WMT", "15", "Results", "for", "German→English", "on", "the", "WMT", "15", "data", "sets", "are", "shown", "in", "Table", "5", ".", "Like", "for", "the", "reverse", "translation", "direction", ",", "we", "see", "substantial", "improvements", "(", "3.6", "-", "3.7", "BLEU", ")", "from", "adding", "monolingual", "training", "data", "with", "synthetic", "source", "sentences", ",", "which", "is", "substantially", "bigger", "than", "the", "improvement", "observed", "with", "deep", "fusion", "[", "reference", "]", ";", "our", "ensemble", "outperforms", "the", "previous", "state", "of", "the", "art", "on", "newstest2015", "by", "2.3", "BLEU", ".", "Table", "6", "shows", "results", "for", "Turkish→English", ".", "On", "average", ",", "we", "see", "an", "improvement", "of", "0.6", "BLEU", "on", "the", "test", "sets", "from", "adding", "monolingual", "data", "with", "a", "dummy", "source", "side", "in", "a", "1", "-", "1", "ratio", "10", ",", "although", "we", "note", "a", "high", "variance", "between", "different", "test", "sets", ".", "section", ":", "Turkish→English", "IWSLT", "14", "With", "synthetic", "training", "data", "(", "Gigaword", "synth", ")", ",", "we", "outperform", "the", "baseline", "by", "2.7", "BLEU", "on", "average", ",", "and", "also", "outperform", "results", "obtained", "via", "shallow", "or", "deep", "fusion", "by", "[", "reference", "]", "by", "0.5", "BLEU", "on", "average", ".", "To", "compare", "to", "what", "extent", "synthetic", "data", "has", "a", "regularization", "effect", ",", "even", "without", "novel", "training", "data", ",", "we", "also", "back", "-", "translate", "the", "target", "side", "of", "the", "parallel", "training", "text", "to", "obtain", "the", "training", "corpus", "parallel", "synth", ".", "Mixing", "the", "original", "parallel", "corpus", "with", "parallel", "synth", "(", "ratio", "1", "-", "1", ")", "gives", "some", "improvement", "over", "the", "baseline", "(", "1.7", "BLEU", "on", "average", ")", ",", "but", "the", "novel", "monolingual", "training", "data", "(", "Gigaword", "mono", ")", "gives", "higher", "improvements", ",", "despite", "being", "out", "-", "of", "-", "domain", "in", "relation", "to", "the", "test", "sets", ".", "We", "speculate", "that", "novel", "in", "-", "domain", "monolingual", "data", "would", "lead", "to", "even", "higher", "improvements", ".", "section", ":", "Back", "-", "translation", "Quality", "for", "Synthetic", "Data", "One", "question", "that", "our", "previous", "experiments", "leave", "open", "is", "how", "the", "quality", "of", "the", "automatic", "backtranslation", "affects", "training", "with", "synthetic", "data", ".", "To", "investigate", "this", "question", ",", "we", "back", "-", "translate", "the", "same", "German", "monolingual", "corpus", "with", "three", "different", "German→English", "systems", ":", "•", "with", "our", "baseline", "system", "and", "greedy", "decoding", "•", "with", "our", "baseline", "system", "and", "beam", "search", "(", "beam", "size", "12", ")", ".", "This", "is", "the", "same", "system", "used", "for", "the", "experiments", "in", "Table", "3", ".", "[", "reference", "]", "We", "also", "experimented", "with", "higher", "ratios", "of", "monolingual", "data", ",", "but", "this", "led", "to", "decreased", "BLEU", "scores", ".", "Table", "7", ":", "English→German", "translation", "performance", "(", "BLEU", ")", "on", "WMT", "training", "/", "test", "sets", "(", "newstest2014", ";", "newstest2015", ")", ".", "Systems", "differ", "in", "how", "the", "synthetic", "training", "data", "is", "obtained", ".", "Ensembles", "of", "4", "models", "(", "unless", "specified", "otherwise", ")", ".", "•", "with", "the", "German→English", "system", "that", "was", "itself", "trained", "with", "synthetic", "data", "(", "beam", "size", "12", ")", ".", "BLEU", "scores", "of", "the", "German→English", "systems", ",", "and", "of", "the", "resulting", "English→German", "systems", "that", "are", "trained", "on", "the", "different", "backtranslations", ",", "are", "shown", "in", "Table", "7", ".", "The", "quality", "of", "the", "German→English", "back", "-", "translation", "differs", "substantially", ",", "with", "a", "difference", "of", "6", "BLEU", "on", "newstest2015", ".", "Regarding", "the", "English→German", "systems", "trained", "on", "the", "different", "synthetic", "corpora", ",", "we", "find", "that", "the", "6", "BLEU", "difference", "in", "back", "-", "translation", "quality", "leads", "to", "a", "0.6", "-", "0.7", "BLEU", "difference", "in", "translation", "quality", ".", "This", "is", "balanced", "by", "the", "fact", "that", "we", "can", "increase", "the", "speed", "of", "back", "-", "translation", "by", "trading", "off", "some", "quality", ",", "for", "instance", "by", "reducing", "beam", "size", ",", "and", "we", "leave", "it", "to", "future", "research", "to", "explore", "how", "much", "the", "amount", "of", "synthetic", "data", "affects", "translation", "quality", ".", "We", "also", "show", "results", "for", "an", "ensemble", "of", "3", "models", "(", "the", "best", "single", "model", "of", "each", "training", "run", ")", ",", "and", "12", "models", "(", "all", "4", "models", "of", "each", "training", "run", ")", ".", "Thanks", "to", "the", "increased", "diversity", "of", "the", "ensemble", "components", ",", "these", "ensembles", "outperform", "the", "ensembles", "of", "4", "models", "that", "were", "all", "sampled", "from", "the", "same", "training", "run", ",", "and", "we", "obtain", "another", "improvement", "of", "0.8", "-", "1.0", "BLEU", ".", "section", ":", "Contrast", "to", "Phrase", "-", "based", "SMT", "The", "back", "-", "translation", "of", "monolingual", "target", "data", "into", "the", "source", "language", "to", "produce", "synthetic", "parallel", "text", "has", "been", "previously", "explored", "for", "phrase", "-", "based", "SMT", "[", "reference", "][", "reference", "]", ".", "While", "our", "approach", "is", "technically", "similar", ",", "synthetic", "parallel", "data", "fulfills", "novel", "name", "training", "BLEU", "data", "instances", "tst2011", "tst2012", "tst2013", "tst2014", "baseline", "[", "reference", "]", "18.4", "18.8", "19.9", "18.7", "deep", "fusion", "[", "reference", "]", "20", "Table", "8", ":", "Phrase", "-", "based", "SMT", "results", "(", "English→German", ")", "on", "WMT", "test", "sets", "(", "average", "of", "newstest201{4", ",", "5", "}", ")", ",", "and", "IWSLT", "test", "sets", "(", "average", "of", "tst201{3", ",", "4", ",", "5", "}", ")", ",", "and", "average", "BLEU", "gain", "from", "adding", "synthetic", "data", "for", "both", "PBSMT", "and", "NMT", ".", "roles", "in", "NMT", ".", "To", "explore", "the", "relative", "effectiveness", "of", "backtranslated", "data", "for", "phrase", "-", "based", "SMT", "and", "NMT", ",", "we", "train", "two", "phrase", "-", "based", "SMT", "systems", "with", "Moses", "[", "reference", "]", ",", "using", "only", "WMT", "parallel", ",", "or", "both", "WMT", "parallel", "and", "WMT", "synth_de", "for", "training", "the", "translation", "and", "reordering", "model", ".", "Both", "systems", "contain", "the", "same", "language", "model", ",", "a", "5", "-", "gram", "Kneser", "-", "Ney", "model", "trained", "on", "all", "available", "WMT", "data", ".", "We", "use", "the", "baseline", "features", "described", "by", ".", "Results", "are", "shown", "in", "Table", "8", ".", "In", "phrasebased", "SMT", ",", "we", "find", "that", "the", "use", "of", "back", "-", "translated", "training", "data", "has", "a", "moderate", "positive", "effect", "on", "the", "WMT", "test", "sets", "(", "+", "0.7", "BLEU", ")", ",", "but", "not", "on", "the", "IWSLT", "test", "sets", ".", "This", "is", "in", "line", "with", "the", "expectation", "that", "the", "main", "effect", "of", "back", "-", "translated", "data", "for", "phrase", "-", "based", "SMT", "is", "domain", "adaptation", "[", "reference", "]", ".", "Both", "the", "WMT", "test", "sets", "and", "the", "News", "Crawl", "corpora", "which", "we", "used", "as", "monolingual", "data", "come", "from", "the", "same", "source", ",", "a", "web", "crawl", "of", "newspaper", "articles", ".", "11", "In", "contrast", ",", "News", "Crawl", "is", "out", "-", "of", "-", "domain", "for", "the", "IWSLT", "test", "sets", ".", "[", "reference", "]", "The", "WMT", "test", "sets", "are", "held", "-", "out", "from", "News", "Crawl", ".", "In", "contrast", "to", "phrase", "-", "based", "SMT", ",", "which", "can", "make", "use", "of", "monolingual", "data", "via", "the", "language", "model", ",", "NMT", "has", "so", "far", "not", "been", "able", "to", "use", "monolingual", "data", "to", "great", "effect", ",", "and", "without", "requiring", "architectural", "changes", ".", "We", "find", "that", "the", "effect", "of", "synthetic", "parallel", "data", "is", "not", "limited", "to", "domain", "adaptation", ",", "and", "that", "even", "out", "-", "of", "-", "domain", "synthetic", "data", "improves", "NMT", "quality", ",", "as", "in", "our", "evaluation", "on", "IWSLT", ".", "The", "fact", "that", "the", "synthetic", "data", "is", "more", "effective", "on", "the", "WMT", "test", "sets", "(", "+", "2.9", "BLEU", ")", "than", "on", "the", "IWSLT", "test", "sets", "(", "+", "1.2", "BLEU", ")", "supports", "the", "hypothesis", "that", "domain", "adaptation", "contributes", "to", "the", "effectiveness", "of", "adding", "synthetic", "data", "to", "NMT", "training", ".", "It", "is", "an", "important", "finding", "that", "back", "-", "translated", "data", ",", "which", "is", "mainly", "effective", "for", "domain", "adaptation", "in", "phrase", "-", "based", "SMT", ",", "is", "more", "generally", "useful", "in", "NMT", ",", "and", "has", "positive", "effects", "that", "go", "beyond", "domain", "adaptation", ".", "In", "the", "next", "section", ",", "we", "will", "investigate", "further", "reasons", "for", "its", "effectiveness", ".", "section", ":", "Analysis", "We", "previously", "indicated", "that", "overfitting", "is", "a", "concern", "with", "our", "baseline", "system", ",", "especially", "on", "small", "data", "sets", "of", "several", "hundred", "thousand", "training", "sentences", ",", "despite", "the", "regularization", "employed", ".", "This", "overfitting", "is", "illustrated", "in", "Figure", "1", ",", "which", "plots", "training", "and", "development", "set", "cross", "-", "entropy", "by", "training", "time", "for", "Turkish→English", "models", ".", "For", "comparability", ",", "we", "measure", "training", "set", "crossentropy", "for", "all", "models", "on", "the", "same", "random", "sample", "of", "the", "parallel", "training", "set", ".", "We", "can", "see", "that", "the", "model", "trained", "on", "only", "parallel", "training", "data", "quickly", "overfits", ",", "while", "all", "three", "monolingual", "data", "sets", "(", "parallel", "synth", ",", "Gigaword", "mono", ",", "or", "Gigaword", "synth", ")", "delay", "overfitting", ",", "and", "give", "better", "perplexity", "on", "the", "development", "set", ".", "The", "best", "development", "set", "cross", "-", "entropy", "is", "reached", "by", "Gigaword", "synth", ".", "Figure", "2", "shows", "cross", "-", "entropy", "for", "English→German", ",", "comparing", "the", "system", "trained", "on", "only", "parallel", "data", "and", "the", "system", "that", "includes", "synthetic", "training", "data", ".", "Since", "more", "training", "data", "is", "available", "for", "English→German", ",", "there", "is", "no", "indication", "that", "overfitting", "happens", "during", "the", "first", "40", "million", "training", "instances", "(", "or", "7", "days", "of", "training", ")", ";", "while", "both", "systems", "obtain", "comparable", "training", "set", "cross", "-", "entropies", ",", "the", "system", "with", "synthetic", "data", "reaches", "a", "lower", "cross", "-", "entropy", "on", "the", "development", "set", ".", "One", "explanation", "for", "this", "is", "the", "domain", "effect", "discussed", "in", "the", "previous", "section", ".", "A", "central", "theoretical", "expectation", "is", "that", "monolingual", "target", "-", "side", "data", "improves", "the", "model", "'s", "flu", "-", "We", "compare", "the", "number", "of", "words", "in", "the", "system", "output", "for", "the", "newstest2015", "test", "set", "which", "are", "produced", "via", "subword", "units", ",", "and", "that", "do", "not", "occur", "in", "the", "parallel", "training", "corpus", ".", "We", "also", "count", "how", "many", "of", "them", "are", "attested", "in", "the", "full", "monolingual", "corpus", "or", "the", "reference", "translation", ",", "which", "we", "all", "consider", "'", "natural", "'", ".", "Additionally", ",", "the", "main", "authors", ",", "a", "native", "speaker", "of", "German", ",", "annotated", "a", "random", "subset", "(", "n", "=", "100", ")", "of", "unattested", "words", "of", "each", "system", "according", "to", "their", "naturalness", "13", ",", "distinguishing", "between", "natural", "German", "words", "(", "or", "names", ")", "such", "as", "Literatur|klassen", "'", "literature", "classes", "'", ",", "and", "nonsensical", "ones", "such", "as", "*", "As|best|atten", "(", "a", "missspelling", "of", "Astbestmatten", "'", "asbestos", "mats", "'", ")", ".", "In", "the", "results", "(", "Table", "9", ")", ",", "we", "see", "that", "the", "systems", "trained", "with", "additional", "monolingual", "or", "synthetic", "data", "have", "a", "higher", "proportion", "of", "novel", "words", "attested", "in", "the", "non", "-", "parallel", "data", ",", "and", "a", "higher", "proportion", "that", "is", "deemed", "natural", "by", "our", "annotator", ".", "This", "supports", "our", "expectation", "that", "additional", "monolingual", "data", "improves", "the", "(", "word", "-", "level", ")", "fluency", "of", "the", "NMT", "system", ".", "section", ":", "Related", "Work", "To", "our", "knowledge", ",", "the", "integration", "of", "monolingual", "data", "for", "pure", "neural", "machine", "translation", "architectures", "was", "first", "investigated", "by", "[", "reference", "]", ",", "who", "train", "monolingual", "language", "models", "independently", ",", "and", "then", "integrate", "them", "during", "decoding", "through", "rescoring", "of", "the", "beam", "(", "shallow", "fusion", ")", ",", "or", "by", "adding", "the", "recurrent", "hidden", "state", "of", "the", "language", "model", "to", "the", "decoder", "state", "of", "the", "encoder", "-", "decoder", "network", ",", "with", "an", "additional", "controller", "mechanism", "that", "controls", "the", "magnitude", "of", "the", "LM", "signal", "(", "deep", "fusion", ")", ".", "In", "deep", "fusion", ",", "the", "controller", "parameters", "and", "output", "parameters", "are", "tuned", "on", "further", "parallel", "training", "data", ",", "but", "the", "language", "model", "parameters", "are", "fixed", "during", "the", "finetuning", "stage", ".", "[", "reference", "]", "also", "report", "on", "experiments", "with", "reranking", "of", "NMT", "output", "with", "a", "5", "-", "gram", "language", "model", ",", "but", "improvements", "are", "small", "(", "between", "0.1", "-", "0.5", "BLEU", ")", ".", "The", "production", "of", "synthetic", "parallel", "texts", "bears", "resemblance", "to", "data", "augmentation", "techniques", "used", "in", "computer", "vision", ",", "where", "datasets", "are", "often", "augmented", "with", "rotated", ",", "scaled", ",", "or", "otherwise", "distorted", "variants", "of", "the", "(", "limited", ")", "training", "set", "[", "reference", "]", ".", "Another", "similar", "avenue", "of", "research", "is", "selftraining", "[", "reference", "][", "reference", "]", ".", "The", "main", "difference", "is", "that", "self", "-", "training", "typically", "refers", "to", "scenario", "where", "the", "training", "set", "is", "enhanced", "with", "training", "instances", "with", "artificially", "produced", "output", "labels", ",", "whereas", "we", "start", "with", "human", "-", "produced", "output", "(", "i.e.", "the", "translation", ")", ",", "and", "artificially", "produce", "an", "input", ".", "We", "expect", "that", "this", "is", "more", "robust", "towards", "noise", "in", "the", "automatic", "translation", ".", "Improving", "NMT", "with", "monolingual", "source", "data", ",", "following", "similar", "work", "on", "phrasebased", "SMT", "[", "reference", "]", ",", "remains", "possible", "future", "work", ".", "Domain", "adaptation", "of", "neural", "networks", "via", "continued", "training", "has", "been", "shown", "to", "be", "effective", "for", "neural", "language", "models", "by", "[", "reference", "]", ",", "and", "in", "work", "parallel", "to", "ours", ",", "for", "neural", "translation", "models", ".", "We", "are", "the", "first", "to", "show", "that", "we", "can", "effectively", "adapt", "neural", "translation", "models", "with", "monolingual", "data", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "propose", "two", "simple", "methods", "to", "use", "monolingual", "training", "data", "during", "training", "of", "NMT", "systems", ",", "with", "no", "changes", "to", "the", "network", "architecture", ".", "Providing", "training", "examples", "with", "dummy", "source", "context", "was", "successful", "to", "some", "extent", ",", "but", "we", "achieve", "substantial", "gains", "in", "all", "tasks", ",", "and", "new", "SOTA", "results", ",", "via", "back", "-", "translation", "of", "monolingual", "target", "data", "into", "the", "source", "language", ",", "and", "treating", "this", "synthetic", "data", "as", "additional", "training", "data", ".", "We", "also", "show", "that", "small", "amounts", "of", "indomain", "monolingual", "data", ",", "back", "-", "translated", "into", "the", "source", "language", ",", "can", "be", "effectively", "used", "for", "domain", "adaptation", ".", "In", "our", "analysis", ",", "we", "identified", "domain", "adaptation", "effects", ",", "a", "reduction", "of", "overfitting", ",", "and", "improved", "fluency", "as", "reasons", "for", "the", "effectiveness", "of", "using", "monolingual", "data", "for", "training", ".", "While", "our", "experiments", "did", "make", "use", "of", "monolingual", "training", "data", ",", "we", "only", "used", "a", "small", "random", "sample", "of", "the", "available", "data", ",", "especially", "for", "the", "experiments", "with", "synthetic", "parallel", "data", ".", "It", "is", "conceivable", "that", "larger", "synthetic", "data", "sets", ",", "or", "data", "sets", "obtained", "via", "data", "selection", ",", "will", "provide", "bigger", "performance", "benefits", ".", "Because", "we", "do", "not", "change", "the", "neural", "network", "architecture", "to", "integrate", "monolingual", "training", "data", ",", "our", "approach", "can", "be", "easily", "applied", "to", "other", "NMT", "systems", ".", "We", "expect", "that", "the", "effectiveness", "of", "our", "approach", "not", "only", "varies", "with", "the", "quality", "of", "the", "MT", "system", "used", "for", "back", "-", "translation", ",", "but", "also", "depends", "on", "the", "amount", "(", "and", "similarity", "to", "the", "test", "set", ")", "of", "available", "parallel", "and", "monolingual", "data", ",", "and", "the", "extent", "of", "overfitting", "of", "the", "baseline", "model", ".", "Future", "work", "will", "explore", "the", "effectiveness", "of", "our", "approach", "in", "more", "settings", ".", "section", ":", "section", ":", "Acknowledgments", "The", "research", "presented", "in", "this", "publication", "was", "conducted", "in", "cooperation", "with", "Samsung", "Electronics", "Polska", "sp", ".", "z", "o.o", ".", "-", "Samsung", "R", "&", "D", "Institute", "Poland", ".", "This", "project", "received", "funding", "from", "the", "European", "Union", "'s", "Horizon", "2020", "research", "and", "innovation", "programme", "under", "grant", "agreement", "645452", "(", "QT21", ")", ".", "section", ":"], "entity": [[1, 5, ["Neural", "Machine", "Translation", "Models"], "Method"], [6, 8, ["Monolingual", "Data"], "Method"], [11, 14, ["Neural", "Machine", "Translation"], "Task"], [15, 17, ["NMT", ")"], "Task"], [40, 42, ["monolingual", "data"], "Method"], [47, 49, ["boosting", "fluency"], "Task"], [50, 54, ["phrasebased", "statistical", "machine", "translation"], "Task"], [61, 63, ["monolingual", "data"], "Method"], [64, 65, ["NMT"], "Task"], [74, 76, ["NMT", "models"], "Method"], [79, 81, ["language", "models"], "Method"], [85, 90, ["encoder", "-", "decoder", "NMT", "architectures"], "Method"], [101, 103, ["language", "model"], "Method"], [111, 112, ["monolingual"], "Method"], [116, 119, ["neural", "network", "architecture"], "Method"], [122, 125, ["monolingual", "training", "data"], "Method"], [147, 151, ["WMT", "15", "task", "English↔German"], "Task"], [156, 157, ["BLEU"], "Metric"], [162, 169, ["low", "-", "resourced", "IWSLT", "14", "task", "Turkish→English"], "Task"], [174, 175, ["BLEU"], "Metric"], [199, 200, ["monolingual"], "Method"], [208, 212, ["IWSLT", "15", "task", "English→German"], "Task"], [216, 219, ["Neural", "Machine", "Translation"], "Task"], [220, 222, ["NMT", ")"], "Task"], [247, 249, ["monolingual", "data"], "Method"], [254, 256, ["boosting", "fluency"], "Task"], [257, 262, ["phrase", "-", "based", "statisti", "-"], "Task"], [290, 292, ["machine", "translation"], "Task"], [299, 301, ["monolingual", "data"], "Method"], [302, 303, ["NMT"], "Task"], [304, 306, ["Language", "models"], "Method"], [308, 310, ["monolingual", "data"], "Method"], [316, 319, ["statistical", "machine", "translation"], "Task"], [322, 324, ["IBM", "models"], "Method"], [340, 349, ["word", "-", "based", "and", "phrase", "-", "based", "translation", "models"], "Method"], [366, 368, ["language", "models"], "Method"], [391, 393, ["monolingual", "data"], "Method"], [422, 424, ["translation", "task"], "Task"], [427, 433, ["attentional", ")", "encoder", "-", "decoder", "architectures"], "Method"], [434, 437, ["neural", "machine", "translation"], "Task"], [444, 445, ["decoder"], "Method"], [448, 451, ["RNN", "language", "model"], "Method"], [466, 468, ["language", "model"], "Method"], [476, 478, ["translation", "model"], "Method"], [492, 493, ["NMT"], "Task"], [497, 499, ["monolingual", "data"], "Method"], [515, 517, ["translation", "task"], "Task"], [537, 540, ["RNN", "language", "model"], "Method"], [542, 544, ["NMT", "model"], "Method"], [553, 556, ["monolingual", "training", "data"], "Method"], [563, 566, ["neural", "network", "architecture"], "Method"], [574, 576, ["NMT", "architectures"], "Method"], [595, 598, ["machine", "translation", "quality"], "Metric"], [599, 601, ["NMT", "systems"], "Method"], [603, 606, ["monolingual", "target", "sentences"], "Method"], [623, 625, ["monolingual", "training"], "Method"], [640, 641, ["backtranslation"], "Method"], [660, 662, ["NMT", "models"], "Method"], [672, 673, ["monolingual"], "Method"], [682, 685, ["Neural", "Machine", "Translation"], "Task"], [688, 692, ["neural", "machine", "translation", "architecture"], "Method"], [719, 723, ["neural", "machine", "translation", "system"], "Method"], [727, 731, ["encoder", "-", "decoder", "network"], "Method"], [732, 735, ["recurrent", "neural", "networks"], "Method"], [737, 738, ["encoder"], "Method"], [740, 743, ["bidirectional", "neural", "network"], "Method"], [744, 747, ["gated", "recurrent", "units"], "Method"], [815, 816, ["decoder"], "Method"], [818, 821, ["recurrent", "neural", "network"], "Method"], [892, 896, ["alignment", "model", "α", "ij"], "Method"], [911, 913, ["alignment", "model"], "Method"], [915, 919, ["singlelayer", "feedforward", "neural", "network"], "Method"], [930, 931, ["backpropagation"], "Method"], [951, 954, ["stochastic", "gradient", "descent"], "Method"], [956, 957, ["translation"], "Task"], [959, 961, ["beam", "search"], "Method"], [970, 972, ["NMT", "Training"], "Task"], [973, 976, ["Monolingual", "Training", "Data"], "Method"], [977, 979, ["machine", "translation"], "Task"], [981, 983, ["monolingual", "data"], "Method"], [985, 987, ["monolingual", "data"], "Method"], [1032, 1034, ["language", "models"], "Method"], [1035, 1038, ["monolingual", "training", "data"], "Method"], [1043, 1045, ["neural", "network"], "Method"], [1046, 1050, ["shallow", "or", "deep", "fusion"], "Method"], [1058, 1060, ["NMT", "model"], "Method"], [1061, 1063, ["monolingual", "data"], "Method"], [1068, 1073, ["encoder", "-", "decoder", "neural", "networks"], "Method"], [1098, 1101, ["monolingual", "training", "examples"], "Method"], [1113, 1116, ["monolingual", "training", "data"], "Method"], [1141, 1144, ["back", "-", "translation"], "Task"], [1158, 1161, ["monolingual", "training", "examples"], "Method"], [1197, 1198, ["prediction"], "Task"], [1207, 1208, ["dropout"], "Method"], [1239, 1243, ["multi", "-", "task", "learning"], "Task"], [1249, 1250, ["translation"], "Task"], [1257, 1259, ["language", "modelling"], "Method"], [1265, 1266, ["training"], "Task"], [1272, 1275, ["monolingual", "training", "examples"], "Method"], [1306, 1308, ["monolingual", "data"], "Method"], [1315, 1317, ["monolingual", "sentences"], "Method"], [1335, 1338, ["monolingual", "training", "examples"], "Method"], [1341, 1343, ["network", "graph"], "Method"], [1357, 1358, ["encoder"], "Method"], [1360, 1362, ["attention", "model"], "Method"], [1383, 1386, ["monolingual", "training", "instances"], "Method"], [1393, 1396, ["monolingual", "training", "data"], "Method"], [1425, 1428, ["monolingual", "training", "instances"], "Method"], [1441, 1443, ["output", "layer"], "Method"], [1458, 1460, ["monolingual", "data"], "Method"], [1465, 1467, ["monolingual", "training"], "Method"], [1486, 1489, ["back", "-", "translation"], "Task"], [1492, 1494, ["automatic", "translation"], "Task"], [1496, 1499, ["monolingual", "target", "text"], "Method"], [1505, 1506, ["training"], "Task"], [1565, 1567, ["NMT", "training"], "Method"], [1574, 1576, ["monolingual", "data"], "Method"], [1618, 1620, ["NMT", "system"], "Method"], [1648, 1652, ["case", "-", "sensitive", "BLEU"], "Metric"], [1673, 1677, ["case", "-", "sensitive", "BLEU"], "Metric"], [1723, 1725, ["monolingual", "sentences"], "Method"], [1811, 1813, ["monolingual", "data"], "Method"], [1825, 1827, ["Baseline", "models"], "Method"], [1833, 1834, ["Ensembles"], "Method"], [1888, 1891, ["monolingual", "data", "set"], "Method"], [1895, 1897, ["German→English", "system"], "Method"], [1908, 1909, ["Translation"], "Task"], [1940, 1944, ["English→German", "system", "+", "synthetic"], "Method"], [1953, 1954, ["backtranslation"], "Task"], [1968, 1970, ["NMT", "training"], "Method"], [1994, 1995, ["BPE"], "Method"], [2008, 2009, ["BPE"], "Method"], [2016, 2018, ["merge", "operations"], "Method"], [2065, 2069, ["cross", "-", "domain", "setting"], "Task"], [2114, 2119, ["IWSLT", "14", "machine", "translation", "track"], "Task"], [2186, 2189, ["monolingual", "training", "data"], "Method"], [2232, 2234, ["translation", "setting"], "Task"], [2246, 2249, ["morphology", "tool", "Zemberek"], "Method"], [2253, 2254, ["disambiguation"], "Task"], [2256, 2258, ["morphological", "analysis"], "Method"], [2358, 2361, ["English→Turkish", "NMT", "system"], "Method"], [2368, 2371, ["Turkish→English", "baseline", "system"], "Method"], [2406, 2407, ["dropout"], "Method"], [2426, 2427, ["BLEU"], "Metric"], [2432, 2433, ["tst2010"], "Method"], [2446, 2448, ["gradient", "clipping"], "Method"], [2494, 2496, ["monolingual", "data"], "Method"], [2509, 2510, ["quality"], "Metric"], [2514, 2515, ["BLEU"], "Metric"], [2521, 2522, ["BLEU"], "Metric"], [2540, 2542, ["training", "algorithm"], "Method"], [2561, 2564, ["monolingual", "training", "instances"], "Method"], [2569, 2571, ["training", "time"], "Metric"], [2588, 2589, ["BLEU"], "Metric"], [2618, 2619, ["BLEU"], "Metric"], [2627, 2629, ["syntaxbased", "baseline"], "Method"], [2637, 2638, ["BLEU"], "Metric"], [2643, 2644, ["NMT"], "Method"], [2689, 2691, ["ensemble", "components"], "Method"], [2696, 2698, ["ensemble", "components"], "Method"], [2715, 2718, ["monolingual", "training", "data"], "Method"], [2780, 2782, ["BLEU", "score"], "Metric"], [2835, 2837, ["German→English", "system"], "Method"], [2863, 2866, ["fine", "-", "tuning"], "Method"], [2867, 2869, ["WIT", "synth"], "Task"], [2876, 2878, ["BLEU", "score"], "Metric"], [2888, 2889, ["BLEU"], "Metric"], [2895, 2898, ["fine", "-", "tuning"], "Method"], [2923, 2925, ["monolingual", "data"], "Method"], [2951, 2952, ["BLEU"], "Metric"], [2972, 2975, ["fine", "-", "tuning"], "Task"], [2976, 2977, ["BLEU"], "Metric"], [2980, 2981, ["PBSMT"], "Method"], [2983, 2984, ["NMT"], "Method"], [2990, 2991, ["fusion"], "Method"], [2994, 2995, ["fusion"], "Method"], [3016, 3019, ["German→English", "translation", "performance"], "Metric"], [3020, 3021, ["BLEU"], "Metric"], [3037, 3038, ["BLEU"], "Metric"], [3058, 3061, ["NMT", "domain", "adaptation"], "Method"], [3062, 3064, ["monolingual", "data"], "Method"], [3077, 3082, ["monolingual", "in", "-", "domain", "data"], "Method"], [3102, 3104, ["trained", "models"], "Method"], [3122, 3123, ["BLEU"], "Metric"], [3147, 3150, ["reverse", "translation", "direction"], "Task"], [3159, 3160, ["BLEU"], "Metric"], [3163, 3166, ["monolingual", "training", "data"], "Method"], [3180, 3182, ["deep", "fusion"], "Method"], [3199, 3200, ["BLEU"], "Metric"], [3217, 3218, ["BLEU"], "Metric"], [3224, 3226, ["monolingual", "data"], "Method"], [3270, 3271, ["BLEU"], "Metric"], [3280, 3284, ["shallow", "or", "deep", "fusion"], "Method"], [3290, 3291, ["BLEU"], "Metric"], [3355, 3356, ["BLEU"], "Metric"], [3363, 3366, ["monolingual", "training", "data"], "Method"], [3395, 3397, ["monolingual", "data"], "Method"], [3406, 3410, ["Back", "-", "translation", "Quality"], "Metric"], [3427, 3429, ["automatic", "backtranslation"], "Method"], [3430, 3431, ["training"], "Task"], [3447, 3449, ["monolingual", "corpus"], "Method"], [3452, 3454, ["German→English", "systems"], "Method"], [3461, 3463, ["greedy", "decoding"], "Method"], [3469, 3471, ["beam", "search"], "Method"], [3500, 3502, ["monolingual", "data"], "Method"], [3508, 3510, ["BLEU", "scores"], "Metric"], [3514, 3517, ["English→German", "translation", "performance"], "Metric"], [3518, 3519, ["BLEU"], "Metric"], [3556, 3558, ["German→English", "system"], "Method"], [3571, 3573, ["BLEU", "scores"], "Metric"], [3575, 3577, ["German→English", "systems"], "Method"], [3582, 3584, ["English→German", "systems"], "Method"], [3602, 3606, ["German→English", "back", "-", "translation"], "Task"], [3614, 3615, ["BLEU"], "Metric"], [3620, 3622, ["English→German", "systems"], "Method"], [3634, 3636, ["BLEU", "difference"], "Metric"], [3637, 3641, ["back", "-", "translation", "quality"], "Metric"], [3647, 3649, ["BLEU", "difference"], "Metric"], [3650, 3652, ["translation", "quality"], "Metric"], [3666, 3669, ["back", "-", "translation"], "Task"], [3673, 3674, ["quality"], "Metric"], [3699, 3701, ["translation", "quality"], "Metric"], [3743, 3745, ["ensemble", "components"], "Method"], [3772, 3774, ["1.0", "BLEU"], "Metric"], [3779, 3783, ["Phrase", "-", "based", "SMT"], "Method"], [3784, 3787, ["back", "-", "translation"], "Task"], [3788, 3791, ["monolingual", "target", "data"], "Method"], [3805, 3809, ["phrase", "-", "based", "SMT"], "Task"], [3844, 3846, ["deep", "fusion"], "Method"], [3853, 3857, ["Phrase", "-", "based", "SMT"], "Task"], [3890, 3893, ["average", "BLEU", "gain"], "Metric"], [3899, 3900, ["PBSMT"], "Method"], [3901, 3902, ["NMT"], "Task"], [3905, 3906, ["NMT"], "Task"], [3916, 3920, ["phrase", "-", "based", "SMT"], "Task"], [3921, 3922, ["NMT"], "Task"], [3926, 3931, ["phrase", "-", "based", "SMT", "systems"], "Method"], [3932, 3933, ["Moses"], "Method"], [3939, 3941, ["WMT", "parallel"], "Method"], [3944, 3946, ["WMT", "parallel"], "Method"], [3947, 3949, ["WMT", "synth_de"], "Method"], [3952, 3956, ["translation", "and", "reordering", "model"], "Method"], [3962, 3964, ["language", "model"], "Method"], [3966, 3973, ["5", "-", "gram", "Kneser", "-", "Ney", "model"], "Method"], [3996, 3998, ["phrasebased", "SMT"], "Task"], [4017, 4020, ["WMT", "test", "sets"], "Metric"], [4023, 4024, ["BLEU"], "Metric"], [4051, 4055, ["phrase", "-", "based", "SMT"], "Task"], [4056, 4058, ["domain", "adaptation"], "Task"], [4076, 4078, ["monolingual", "data"], "Method"], [4095, 4097, ["News", "Crawl"], "Task"], [4127, 4131, ["phrase", "-", "based", "SMT"], "Method"], [4137, 4139, ["monolingual", "data"], "Method"], [4141, 4143, ["language", "model"], "Method"], [4144, 4145, ["NMT"], "Task"], [4153, 4155, ["monolingual", "data"], "Method"], [4178, 4180, ["domain", "adaptation"], "Task"], [4192, 4194, ["NMT", "quality"], "Metric"], [4219, 4220, ["BLEU"], "Metric"], [4230, 4231, ["BLEU"], "Metric"], [4236, 4238, ["domain", "adaptation"], "Method"], [4247, 4249, ["NMT", "training"], "Task"], [4266, 4268, ["domain", "adaptation"], "Task"], [4269, 4273, ["phrase", "-", "based", "SMT"], "Task"], [4279, 4280, ["NMT"], "Task"], [4288, 4290, ["domain", "adaptation"], "Method"], [4335, 4336, ["regularization"], "Method"], [4348, 4355, ["training", "and", "development", "set", "cross", "-", "entropy"], "Metric"], [4356, 4358, ["training", "time"], "Metric"], [4359, 4361, ["Turkish→English", "models"], "Method"], [4367, 4370, ["training", "set", "crossentropy"], "Metric"], [4402, 4405, ["monolingual", "data", "sets"], "Method"], [4422, 4423, ["perplexity"], "Metric"], [4432, 4435, ["cross", "-", "entropy"], "Metric"], [4438, 4440, ["Gigaword", "synth"], "Method"], [4444, 4447, ["cross", "-", "entropy"], "Metric"], [4503, 4508, ["training", "set", "cross", "-", "entropies"], "Metric"], [4517, 4520, ["cross", "-", "entropy"], "Metric"], [4545, 4550, ["monolingual", "target", "-", "side", "data"], "Method"], [4601, 4603, ["monolingual", "corpus"], "Method"], [4700, 4701, ["monolingual"], "Method"], [4737, 4739, ["monolingual", "data"], "Method"], [4742, 4747, ["word", "-", "level", ")", "fluency"], "Metric"], [4749, 4751, ["NMT", "system"], "Method"], [4763, 4765, ["monolingual", "data"], "Method"], [4767, 4771, ["neural", "machine", "translation", "architectures"], "Task"], [4791, 4792, ["decoding"], "Task"], [4798, 4800, ["shallow", "fusion"], "Method"], [4811, 4813, ["language", "model"], "Method"], [4819, 4823, ["encoder", "-", "decoder", "network"], "Method"], [4827, 4829, ["controller", "mechanism"], "Method"], [4838, 4840, ["deep", "fusion"], "Method"], [4843, 4845, ["deep", "fusion"], "Task"], [4869, 4871, ["finetuning", "stage"], "Method"], [4880, 4884, ["reranking", "of", "NMT", "output"], "Task"], [4886, 4891, ["5", "-", "gram", "language", "model"], "Method"], [4901, 4902, ["BLEU"], "Metric"], [4905, 4910, ["production", "of", "synthetic", "parallel", "texts"], "Task"], [4913, 4916, ["data", "augmentation", "techniques"], "Method"], [4918, 4920, ["computer", "vision"], "Task"], [4964, 4967, ["self", "-", "training"], "Task"], [5017, 5019, ["automatic", "translation"], "Task"], [5021, 5022, ["NMT"], "Task"], [5023, 5026, ["monolingual", "source", "data"], "Method"], [5031, 5033, ["phrasebased", "SMT"], "Task"], [5042, 5047, ["Domain", "adaptation", "of", "neural", "networks"], "Method"], [5048, 5050, ["continued", "training"], "Method"], [5057, 5060, ["neural", "language", "models"], "Task"], [5073, 5076, ["neural", "translation", "models"], "Task"], [5088, 5091, ["neural", "translation", "models"], "Method"], [5092, 5094, ["monolingual", "data"], "Method"], [5109, 5112, ["monolingual", "training", "data"], "Method"], [5115, 5117, ["NMT", "systems"], "Task"], [5123, 5125, ["network", "architecture"], "Method"], [5154, 5157, ["back", "-", "translation"], "Task"], [5158, 5161, ["monolingual", "target", "data"], "Method"], [5184, 5186, ["monolingual", "data"], "Method"], [5200, 5202, ["domain", "adaptation"], "Task"], [5216, 5217, ["overfitting"], "Metric"], [5220, 5221, ["fluency"], "Metric"], [5228, 5230, ["monolingual", "data"], "Method"], [5231, 5232, ["training"], "Task"], [5240, 5243, ["monolingual", "training", "data"], "Method"], [5279, 5281, ["data", "selection"], "Method"], [5294, 5297, ["neural", "network", "architecture"], "Method"], [5299, 5302, ["monolingual", "training", "data"], "Method"], [5311, 5313, ["NMT", "systems"], "Method"], [5330, 5332, ["MT", "system"], "Method"], [5334, 5337, ["back", "-", "translation"], "Task"], [5356, 5358, ["monolingual", "data"], "Method"], [5366, 5368, ["baseline", "model"], "Method"], [1346, 1347, ["monolingual"], "Method"], [1557, 1558, ["monolingual"], "Method"], [1932, 1933, ["monolingual"], "Method"], [2829, 2830, ["monolingual"], "Method"], [2903, 2904, ["monolingual"], "Method"], [4781, 4782, ["monolingual"], "Method"]]} {"doc_id": "0c7f1d285ce069b2f7a807a4b2750695098bffe6", "text": ["We", "present", "a", "semi", "-", "supervised", "learning", "framework", "based", "on", "graph", "embeddings", ".", "Given", "a", "graph", "between", "instances", ",", "we", "train", "an", "embedding", "for", "each", "instance", "to", "jointly", "predict", "the", "class", "label", "and", "the", "neighborhood", "context", "in", "the", "graph", ".", "We", "develop", "both", "transductive", "and", "inductive", "variants", "of", "our", "method", ".", "In", "the", "transductive", "variant", "of", "our", "method", ",", "the", "class", "labels", "are", "determined", "by", "both", "the", "learned", "embeddings", "and", "input", "feature", "vectors", ",", "while", "in", "the", "inductive", "variant", ",", "the", "embeddings", "are", "defined", "as", "a", "parametric", "function", "of", "the", "feature", "vectors", ",", "so", "predictions", "can", "be", "made", "on", "instances", "not", "seen", "during", "training", ".", "On", "a", "large", "and", "diverse", "set", "of", "benchmark", "tasks", ",", "including", "text", "classification", ",", "distantly", "supervised", "entity", "extraction", ",", "and", "entity", "classification", ",", "we", "show", "improved", "performance", "over", "many", "of", "the", "existing", "models", ".", "RevisitingSemi", "-", "SupervisedLearningwithGraphEmbeddings", "section", ":", "Introduction", "Semi", "-", "supervised", "learning", "aims", "to", "leverage", "unlabeled", "data", "to", "improve", "performance", ".", "A", "large", "number", "of", "semi", "-", "supervised", "learning", "algorithms", "jointly", "optimize", "two", "training", "objective", "functions", ":", "the", "supervised", "loss", "over", "labeled", "data", "and", "the", "unsupervised", "loss", "over", "both", "labeled", "and", "unlabeled", "data", ".", "Graph", "-", "based", "semi", "-", "supervised", "learning", "defines", "the", "loss", "function", "as", "a", "weighted", "sum", "of", "the", "supervised", "loss", "over", "labeled", "instances", "and", "a", "graph", "Laplacian", "regularization", "term", ".", "The", "graph", "Laplacian", "regularization", "is", "based", "on", "the", "assumption", "that", "nearby", "nodes", "in", "a", "graph", "are", "likely", "to", "have", "the", "same", "labels", ".", "Graph", "Laplacian", "regularization", "is", "effective", "because", "it", "constrains", "the", "labels", "to", "be", "consistent", "with", "the", "graph", "structure", ".", "Recently", "developed", "unsupervised", "representation", "learning", "methods", "learn", "embeddings", "that", "predict", "a", "distributional", "context", ",", "e.g.", "a", "word", "embedding", "might", "predict", "nearby", "context", "words", ",", "or", "a", "node", "embedding", "might", "predict", "nearby", "nodes", "in", "a", "graph", ".", "Embeddings", "trained", "with", "distributional", "context", "can", "be", "used", "to", "boost", "the", "performance", "of", "related", "tasks", ".", "For", "example", ",", "word", "embeddings", "trained", "from", "a", "language", "model", "can", "be", "applied", "to", "part", "-", "of", "-", "speech", "tagging", ",", "chunking", "and", "named", "entity", "recognition", ".", "In", "this", "paper", "we", "consider", "not", "word", "embeddings", "but", "graph", "embeddings", ".", "Existing", "results", "show", "that", "graph", "embeddings", "are", "effective", "at", "classifying", "the", "nodes", "in", "a", "graph", ",", "such", "as", "user", "behavior", "prediction", "in", "a", "social", "network", ".", "However", ",", "the", "graph", "embeddings", "are", "usually", "learned", "separately", "from", "the", "supervised", "task", ",", "and", "hence", "do", "not", "leverage", "the", "label", "information", "in", "a", "specific", "task", ".", "Hence", "graph", "embeddings", "are", "in", "some", "sense", "complementary", "to", "graph", "Laplacian", "regularization", "that", "does", "not", "produce", "useful", "features", "itself", "and", "might", "not", "be", "able", "to", "fully", "leverage", "the", "distributional", "information", "encoded", "in", "the", "graph", "structure", ".", "The", "main", "highlight", "of", "our", "work", "is", "to", "incorporate", "embedding", "techniques", "into", "the", "graph", "-", "based", "semi", "-", "supervised", "learning", "setting", ".", "We", "propose", "a", "novel", "graph", "-", "based", "semi", "-", "supervised", "learning", "framework", ",", "Planetoid", "(", "Predicting", "Labels", "And", "Neighbors", "with", "Embeddings", "Transductively", "Or", "Inductively", "from", "Data", ")", ".", "The", "embedding", "of", "an", "instance", "is", "jointly", "trained", "to", "predict", "the", "class", "label", "of", "the", "instance", "and", "the", "context", "in", "the", "graph", ".", "We", "then", "concatenate", "the", "embeddings", "and", "the", "hidden", "layers", "of", "the", "original", "classifier", "and", "feed", "them", "to", "a", "softmax", "layer", "when", "making", "the", "prediction", ".", "Since", "the", "embeddings", "are", "learned", "based", "on", "the", "graph", "structure", ",", "the", "above", "method", "is", "transductive", ",", "which", "means", "we", "can", "only", "predict", "instances", "that", "are", "already", "observed", "in", "the", "graph", "at", "training", "time", ".", "In", "many", "cases", ",", "however", ",", "it", "may", "be", "desirable", "to", "have", "an", "inductive", "approach", ",", "where", "predictions", "can", "be", "made", "on", "instances", "unobserved", "in", "the", "graph", "seen", "at", "training", "time", ".", "To", "address", "this", "issue", ",", "we", "further", "develop", "an", "inductive", "variant", "of", "our", "framework", ",", "where", "we", "define", "the", "embeddings", "as", "a", "parameterized", "function", "of", "input", "feature", "vectors", ";", "i.e.", ",", "the", "embeddings", "can", "be", "viewed", "as", "hidden", "layers", "of", "a", "neural", "network", ".", "To", "demonstrate", "the", "effectiveness", "of", "our", "proposed", "approach", ",", "we", "conducted", "experiments", "on", "five", "datasets", "for", "three", "tasks", ",", "including", "text", "classification", ",", "distantly", "supervised", "entity", "extraction", ",", "and", "entity", "classification", ".", "Our", "inductive", "method", "outperforms", "the", "second", "best", "inductive", "method", "by", "up", "to", "points", "and", "on", "average", "points", "in", "terms", "of", "accuracy", ".", "The", "best", "of", "our", "inductive", "and", "transductive", "methods", "outperforms", "the", "best", "of", "all", "the", "other", "compared", "methods", "by", "up", "to", "and", "on", "average", ".", "section", ":", "Related", "Work", "subsection", ":", "Semi", "-", "Supervised", "Learning", "Let", "and", "be", "the", "number", "of", "labeled", "and", "unlabeled", "instances", ".", "Let", "and", "denote", "the", "feature", "vectors", "of", "labeled", "and", "unlabeled", "instances", "respectively", ".", "The", "labels", "are", "also", "given", ".", "Based", "on", "both", "labeled", "and", "unlabeled", "instances", ",", "the", "problem", "of", "semi", "-", "supervised", "learning", "is", "defined", "as", "learning", "a", "classifier", ".", "There", "are", "two", "learning", "paradigms", ",", "transductive", "learning", "and", "inductive", "learning", ".", "Transductive", "learning", "only", "aims", "to", "apply", "the", "classifier", "on", "the", "unlabeled", "instances", "observed", "at", "training", "time", ",", "and", "the", "classifier", "does", "not", "generalize", "to", "unobserved", "instances", ".", "For", "instance", ",", "transductive", "support", "vector", "machine", "(", "TSVM", ")", "maximizes", "the", "“", "unlabeled", "data", "margin", "”", "based", "on", "the", "low", "-", "density", "separation", "assumption", "that", "a", "good", "decision", "hyperplane", "lies", "on", "a", "sparse", "area", "of", "the", "feature", "space", ".", "Inductive", "learning", ",", "on", "the", "other", "hand", ",", "aims", "to", "learn", "a", "parameterized", "classifier", "that", "is", "generalizable", "to", "unobserved", "instances", ".", "subsection", ":", "Graph", "-", "Based", "Semi", "-", "Supervised", "Learning", "In", "addition", "to", "labeled", "and", "unlabeled", "instances", ",", "a", "graph", ",", "denoted", "as", "a", "matrix", ",", "is", "also", "given", "to", "graph", "-", "based", "semi", "-", "supervised", "learning", "methods", ".", "Each", "entry", "indicates", "the", "similarity", "between", "instance", "and", ",", "which", "can", "be", "either", "labeled", "or", "unlabeled", ".", "The", "graph", "can", "either", "be", "derived", "from", "distances", "between", "instances", ",", "or", "be", "explicitly", "derived", "from", "external", "data", ",", "such", "as", "a", "knowledge", "graph", "or", "a", "citation", "network", "between", "documents", ".", "In", "this", "paper", ",", "we", "mainly", "focus", "on", "the", "setting", "that", "a", "graph", "is", "explicitly", "given", "and", "represents", "additional", "information", "not", "present", "in", "the", "feature", "vectors", "(", "e.g.", ",", "the", "graph", "edges", "correspond", "to", "hyperlinks", "between", "documents", ",", "rather", "than", "distances", "between", "the", "bag", "-", "of", "-", "words", "representation", "of", "a", "document", ")", ".", "Graph", "-", "based", "semi", "-", "supervised", "learning", "is", "based", "on", "the", "assumption", "that", "nearby", "nodes", "tend", "to", "have", "the", "same", "labels", ".", "Generally", ",", "the", "loss", "function", "of", "graph", "-", "based", "semi", "-", "supervised", "learning", "in", "the", "binary", "case", "can", "be", "written", "as", "In", "Eq", ".", "(", "[", "reference", "]", ")", ",", "the", "first", "term", "is", "the", "standard", "supervised", "loss", "function", ",", "where", "can", "be", "log", "loss", ",", "squared", "loss", "or", "hinge", "loss", ".", "The", "second", "term", "is", "the", "graph", "Laplacian", "regularization", ",", "which", "incurs", "a", "large", "penalty", "when", "similar", "nodes", "with", "a", "large", "are", "predicted", "to", "have", "different", "labels", ".", "The", "graph", "Laplacian", "matrix", "is", "defined", "as", ",", "where", "is", "a", "diagonal", "matrix", "with", "each", "entry", "defined", "as", ".", "is", "a", "constant", "weighting", "factor", ".", "(", "Note", "that", "we", "omit", "the", "parameter", "regularization", "terms", "for", "simplicity", ".", ")", "Various", "graph", "-", "based", "semi", "-", "supervised", "learning", "algorithms", "define", "the", "loss", "functions", "as", "variants", "of", "Eq", ".", "(", "[", "reference", "]", ")", ".", "Label", "propagation", "forces", "to", "agree", "with", "labeled", "instances", ";", "is", "a", "label", "lookup", "table", "for", "unlabeled", "instances", "in", "the", "graph", ",", "and", "can", "be", "obtained", "with", "a", "closed", "-", "form", "solution", ".", "Learning", "with", "local", "and", "global", "consistency", "defines", "as", "squared", "loss", "and", "as", "a", "label", "lookup", "table", ";", "it", "does", "not", "force", "to", "agree", "with", "labeled", "instances", ".", "Modified", "Adsorption", "(", "MAD", ")", "is", "a", "variant", "of", "label", "propagation", "that", "allows", "prediction", "on", "labeled", "instances", "to", "vary", "and", "incorporates", "node", "uncertainty", ".", "Manifold", "regularization", "parameterizes", "in", "the", "Reproducing", "Kernel", "Hilbert", "Space", "(", "RKHS", ")", "with", "being", "squared", "loss", "or", "hinge", "loss", ".", "Since", "is", "a", "parameterized", "classifier", ",", "manifold", "regularization", "is", "inductive", "and", "can", "naturally", "handle", "unobserved", "instances", ".", "Semi", "-", "supervised", "embedding", "extends", "the", "regularization", "term", "in", "Eq", ".", "(", "[", "reference", "]", ")", "to", "be", ",", "where", "represents", "embeddings", "of", "instances", ",", "which", "can", "be", "the", "output", "labels", ",", "hidden", "layers", "or", "auxiliary", "embeddings", "in", "a", "neural", "network", ".", "By", "extending", "the", "regularization", "from", "to", ",", "this", "method", "imposes", "stronger", "constraints", "on", "a", "neural", "network", ".", "Iterative", "classification", "algorithm", "(", "ICA", ")", "uses", "a", "local", "classifier", "that", "takes", "the", "labels", "of", "neighbor", "nodes", "as", "input", ",", "and", "employs", "an", "iterative", "process", "between", "estimating", "the", "local", "classifier", "and", "assigning", "new", "labels", ".", "subsection", ":", "Learning", "Embeddings", "Extensive", "research", "was", "done", "on", "learning", "graph", "embeddings", ".", "A", "probabilistic", "generative", "model", "was", "proposed", "to", "learn", "node", "embeddings", "that", "generate", "the", "edges", "in", "a", "graph", ".", "A", "clustering", "method", "was", "proposed", "to", "learn", "latent", "social", "states", "in", "a", "social", "network", "to", "predict", "social", "ties", ".", "More", "recently", ",", "a", "number", "of", "embedding", "learning", "methods", "are", "based", "on", "the", "Skipgram", "model", ",", "which", "is", "a", "variant", "of", "the", "softmax", "model", ".", "Given", "an", "instance", "and", "its", "context", ",", "the", "objective", "of", "Skipgram", "is", "usually", "formulated", "as", "minimizing", "the", "log", "loss", "of", "predicting", "the", "context", "using", "the", "embedding", "of", "an", "instance", "as", "input", "features", ".", "Formally", ",", "let", "be", "a", "set", "of", "pairs", "of", "instance", "and", "context", ",", "the", "loss", "function", "can", "be", "written", "as", "where", "is", "the", "set", "of", "all", "possible", "context", ",", "’s", "are", "parameters", "of", "the", "Skipgram", "model", ",", "and", "is", "the", "embedding", "of", "instance", ".", "Skipgram", "was", "first", "introduced", "to", "learn", "representations", "of", "words", ",", "known", "as", "word2vec", ".", "In", "word2vec", ",", "for", "each", "training", "pair", ",", "the", "instance", "is", "the", "current", "word", "whose", "embedding", "is", "under", "estimation", ";", "the", "context", "is", "each", "of", "the", "surrounding", "words", "of", "within", "a", "fixed", "window", "size", "in", "a", "sentence", ";", "the", "context", "space", "is", "the", "vocabulary", "of", "the", "corpus", ".", "Skipgram", "was", "later", "extended", "to", "learn", "graph", "embeddings", ".", "Deepwalk", "uses", "the", "embedding", "of", "a", "node", "to", "predict", "the", "context", "in", "the", "graph", ",", "where", "the", "context", "is", "generated", "by", "random", "walk", ".", "More", "specifically", ",", "for", "each", "training", "pair", ",", "the", "instance", "is", "the", "current", "node", "whose", "embedding", "is", "under", "estimation", ";", "the", "context", "is", "each", "of", "the", "neighbor", "nodes", "within", "a", "fixed", "window", "size", "in", "a", "generated", "random", "walk", "sequence", ";", "the", "context", "space", "is", "all", "the", "nodes", "in", "the", "graph", ".", "LINE", "extends", "the", "model", "to", "have", "multiple", "context", "spaces", "for", "modeling", "both", "first", "and", "second", "order", "proximity", ".", "Although", "Skipgram", "-", "like", "models", "for", "graphs", "have", "received", "much", "recent", "attention", ",", "many", "other", "models", "exist", ".", "TransE", "learns", "the", "embeddings", "of", "entities", "in", "a", "knowledge", "graph", "jointly", "with", "their", "relations", ".", "Autoencoders", "were", "used", "to", "learn", "graph", "embeddings", "for", "clustering", "on", "graphs", ".", "subsection", ":", "Comparison", "We", "compare", "our", "approach", "in", "this", "paper", "with", "other", "methods", "in", "semi", "-", "supervised", "learning", "and", "embedding", "learning", "in", "Table", "[", "reference", "]", ".", "Unlike", "our", "approach", ",", "conventional", "graph", "Laplacian", "based", "methods", "impose", "regularization", "on", "the", "labels", "but", "do", "not", "learn", "embeddings", ".", "Semi", "-", "supervised", "embedding", "method", "learns", "embeddings", "in", "a", "neural", "network", ",", "but", "our", "approach", "is", "different", "from", "this", "method", "in", "that", "instead", "of", "imposing", "regularization", ",", "we", "use", "the", "embeddings", "to", "predict", "the", "context", "in", "the", "graph", ".", "Graph", "embedding", "methods", "encode", "the", "graph", "structure", "into", "embeddings", ";", "however", ",", "different", "from", "our", "approach", ",", "these", "methods", "are", "purely", "unsupervised", "and", "do", "not", "leverage", "label", "information", "for", "a", "specific", "task", ".", "Moreover", ",", "these", "methods", "are", "transductive", "and", "can", "not", "be", "directly", "generalized", "to", "instances", "unseen", "at", "training", "time", ".", "section", ":", "Semi", "-", "Supervised", "Learning", "with", "Graph", "Embeddings", "Following", "the", "notations", "in", "the", "previous", "section", ",", "the", "input", "to", "our", "method", "includes", "labeled", "instances", ",", ",", "unlabeled", "instances", "and", "a", "graph", "denoted", "as", "a", "matrix", ".", "Each", "instance", "has", "an", "embedding", "denoted", "as", ".", "We", "formulate", "our", "framework", "based", "on", "feed", "-", "forward", "neural", "networks", ".", "Given", "the", "input", "feature", "vector", ",", "the", "-", "th", "hidden", "layer", "of", "the", "network", "is", "denoted", "as", ",", "which", "is", "a", "nonlinear", "function", "of", "the", "previous", "hidden", "layer", "defined", "as", ":", "where", "and", "are", "parameters", "of", "the", "-", "th", "layer", ",", "and", ".", "We", "adopt", "rectified", "linear", "unit", "as", "the", "nonlinear", "function", "in", "this", "work", ".", "The", "loss", "function", "of", "our", "framework", "can", "be", "expressed", "as", "where", "is", "a", "supervised", "loss", "of", "predicting", "the", "labels", ",", "and", "is", "an", "unsupervised", "loss", "of", "predicting", "the", "graph", "context", ".", "In", "the", "following", "sections", ",", "we", "first", "formulate", "by", "introducing", "how", "to", "sample", "context", "from", "the", "graph", ",", "and", "then", "formulate", "to", "form", "our", "semi", "-", "supervised", "learning", "framework", ".", "subsection", ":", "Sampling", "Context", "We", "formulate", "the", "unsupervised", "loss", "as", "a", "variant", "of", "Eq", ".", "(", "[", "reference", "]", ")", ".", "Given", "a", "graph", ",", "the", "basic", "idea", "of", "our", "approach", "is", "to", "sample", "pairs", "of", "instance", "and", "context", ",", "and", "then", "formulate", "the", "loss", "using", "the", "log", "loss", "as", "in", "Eq", ".", "(", "[", "reference", "]", ")", ".", "We", "first", "present", "the", "formulation", "of", "by", "introducing", "negative", "sampling", ",", "and", "then", "discuss", "how", "to", "sample", "pairs", "of", "instance", "and", "context", ".", "It", "is", "usually", "intractable", "to", "directly", "optimize", "Eq", ".", "(", "[", "reference", "]", ")", "due", "to", "normalization", "over", "the", "whole", "context", "space", ".", "Negative", "sampling", "was", "introduced", "to", "address", "this", "issue", ",", "which", "samples", "negative", "examples", "to", "approximate", "the", "normalization", "term", ".", "In", "our", "case", ",", "we", "are", "sampling", "from", "a", "distribution", ",", "where", "and", "denote", "instance", "and", "context", "respectively", ",", "means", "is", "a", "positive", "pair", "and", "means", "negative", ".", "Given", ",", "we", "minimize", "the", "cross", "entropy", "loss", "of", "classifying", "the", "pair", "to", "a", "binary", "label", ":", "where", "is", "the", "sigmoid", "function", "defined", "as", ",", "and", "is", "an", "indicator", "function", "that", "outputs", "when", "the", "argument", "is", "true", ",", "otherwise", ".", "Therefore", ",", "the", "unsupervised", "loss", "with", "negative", "sampling", "can", "be", "written", "as", "The", "distribution", "is", "conditioned", "on", "labels", "and", "the", "graph", ".", "However", ",", "since", "they", "are", "the", "input", "to", "our", "algorithm", "and", "kept", "fixed", ",", "we", "drop", "the", "conditioning", "in", "our", "notation", ".", "We", "now", "define", "the", "distribution", "directly", "using", "a", "sampling", "process", ",", "which", "is", "illustrated", "in", "Algorithm", "[", "reference", "]", ".", "There", "are", "two", "types", "of", "context", "that", "are", "sampled", "in", "this", "algorithm", ".", "The", "first", "type", "of", "context", "is", "based", "on", "the", "graph", ",", "which", "encodes", "the", "structure", "(", "distributional", ")", "information", ",", "and", "the", "second", "type", "of", "context", "is", "based", "on", "the", "labels", ",", "which", "we", "use", "to", "inject", "label", "information", "into", "the", "embeddings", ".", "We", "use", "a", "parameter", "to", "control", "the", "ratio", "of", "positive", "and", "negative", "samples", ",", "and", "use", "to", "control", "the", "ratio", "of", "two", "types", "of", "context", ".", "With", "probability", ",", "we", "sample", "the", "context", "based", "on", "the", "graph", ".", "We", "first", "uniformly", "sample", "a", "random", "walk", "sequence", ".", "More", "specifically", ",", "we", "uniformly", "sample", "the", "first", "instance", "from", "the", "set", ".", "Given", "the", "previous", "instance", ",", "the", "next", "instance", "is", "sampled", "with", "probability", ".", "With", "probability", ",", "we", "sample", "a", "positive", "pair", "from", "the", "set", ",", "where", "is", "another", "parameter", "determining", "the", "window", "size", ".", "With", "probability", ",", "we", "uniformly", "corrupt", "the", "context", "to", "sample", "a", "negative", "pair", ".", "With", "probability", ",", "we", "sample", "the", "context", "based", "on", "the", "class", "labels", ".", "Positive", "pairs", "have", "the", "same", "labels", "and", "negative", "pairs", "have", "different", "labels", ".", "Only", "labeled", "instances", "are", "sampled", ".", "Our", "random", "walk", "based", "sampling", "method", "is", "built", "upon", "Deepwalk", ".", "In", "contrast", "to", "their", "method", ",", "our", "method", "handles", "real", "-", "valued", ",", "incorporates", "negative", "sampling", ",", "and", "explicitly", "samples", "from", "labels", "with", "probability", "to", "inject", "supervised", "information", ".", "An", "example", "of", "sampling", "when", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "[", "tb", "]", "Sampling", "Context", "Distribution", "⁢p", "(", "i", ",", "c", ",", "γ", ")", "Input", ":", "graph", ",", "labels", ",", "parameters", "Initialize", "triplet", "if", "then", "else", "Uniformly", "sample", "a", "random", "walk", "of", "length", "Uniformly", "sample", "with", ",", "if", "then", "uniformly", "sample", "from", "Uniformly", "sample", "with", "Uniformly", "sample", "with", "return", "subsection", ":", "Transductive", "Formulation", "In", "this", "section", ",", "we", "present", "a", "method", "that", "infers", "the", "labels", "of", "unlabeled", "instances", "without", "generalizing", "to", "unobserved", "instances", ".", "Transductive", "learning", "usually", "performs", "better", "than", "inductive", "learning", "because", "transductive", "learning", "can", "leverage", "the", "unlabeled", "test", "data", "when", "training", "the", "model", ".", "We", "apply", "layers", "on", "the", "input", "feature", "vector", "to", "obtain", ",", "and", "layers", "on", "the", "embedding", "to", "obtain", ",", "as", "illustrated", "in", "Figure", "[", "reference", "]", ".", "The", "two", "hidden", "layers", "are", "concatenated", ",", "and", "fed", "to", "a", "softmax", "layer", "to", "predict", "the", "class", "label", "of", "the", "instance", ".", "More", "specifically", ",", "the", "probability", "of", "predicting", "the", "label", "is", "written", "as", ":", "where", "denotes", "concatenation", "of", "two", "row", "vectors", ",", "the", "super", "script", "denotes", "the", "transpose", "of", "vector", ",", "and", "represents", "the", "model", "parameter", ".", "Combined", "with", "Eq", ".", "(", "[", "reference", "]", ")", ",", "the", "loss", "function", "of", "transductive", "learning", "is", "defined", "as", ":", "where", "the", "first", "term", "is", "defined", "by", "Eq", ".", "(", "[", "reference", "]", ")", ",", "and", "is", "a", "constant", "weighting", "factor", ".", "The", "first", "term", "is", "the", "loss", "function", "of", "class", "label", "prediction", "and", "the", "second", "term", "is", "the", "loss", "function", "of", "context", "prediction", ".", "This", "formulation", "is", "transductive", "because", "the", "prediction", "of", "label", "depends", "on", "the", "embedding", ",", "which", "can", "only", "be", "learned", "for", "instances", "observed", "in", "the", "graph", "during", "training", "time", ".", "subsection", ":", "Inductive", "Formulation", "While", "we", "consider", "transductive", "learning", "in", "the", "above", "formulation", ",", "in", "many", "cases", ",", "it", "is", "desirable", "to", "learn", "a", "classifier", "that", "can", "generalize", "to", "unobserved", "instances", ",", "especially", "for", "large", "-", "scale", "tasks", ".", "For", "example", ",", "machine", "reading", "systems", "very", "frequently", "encounter", "novel", "entities", "on", "the", "Web", "and", "it", "is", "not", "practical", "to", "train", "a", "semi", "-", "supervised", "learning", "system", "on", "the", "entire", "Web", ".", "However", ",", "since", "learning", "graph", "embeddings", "is", "transductive", "in", "nature", ",", "it", "is", "not", "straightforward", "to", "do", "it", "in", "an", "inductive", "setting", ".", "Perozzi", "et", "al", ".", "perozzi2014deepwalk", "addressed", "this", "issue", "by", "retraining", "the", "embeddings", "incrementally", ",", "which", "is", "time", "consuming", "and", "does", "not", "scale", "(", "and", "not", "inductive", "essentially", ")", ".", "To", "make", "the", "method", "inductive", ",", "the", "prediction", "of", "label", "should", "only", "depend", "on", "the", "input", "feature", "vector", ".", "Therefore", ",", "we", "define", "the", "embedding", "as", "a", "parameterized", "function", "of", "feature", ",", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "Similar", "to", "the", "transductive", "formulation", ",", "we", "apply", "layers", "on", "the", "input", "feature", "vector", "to", "obtain", ".", "However", ",", "rather", "than", "using", "a", "“", "free", "”", "embedding", ",", "we", "apply", "layers", "on", "the", "input", "feature", "vector", "and", "define", "it", "as", "the", "embedding", ".", "Then", "another", "layers", "are", "applied", "on", "the", "embedding", ",", "denoted", "as", "where", ".", "The", "embedding", "in", "this", "formulation", "can", "be", "viewed", "as", "a", "hidden", "layer", "that", "is", "a", "parameterized", "function", "of", "the", "feature", ".", "With", "the", "above", "formulation", ",", "the", "label", "only", "depends", "on", "the", "feature", ".", "More", "specifically", ",", "Replacing", "in", "Eq", ".", "(", "[", "reference", "]", ")", "with", ",", "the", "loss", "function", "of", "inductive", "learning", "is", "where", "the", "first", "term", "is", "defined", "by", "Eq", ".", "(", "[", "reference", "]", ")", ".", "subsection", ":", "Training", "We", "adopt", "stochastic", "gradient", "descent", "(", "SGD", ")", "to", "train", "our", "model", "in", "the", "mini", "-", "batch", "mode", ".", "We", "first", "sample", "a", "batch", "of", "labeled", "instances", "and", "take", "a", "gradient", "step", "to", "optimize", "the", "loss", "function", "of", "class", "label", "prediction", ".", "We", "then", "sample", "a", "batch", "of", "context", "and", "take", "another", "gradient", "step", "to", "optimize", "the", "loss", "function", "of", "context", "prediction", ".", "We", "repeat", "the", "above", "procedures", "for", "and", "iterations", "respectively", "to", "approximate", "the", "weighting", "factor", ".", "Algorithm", "[", "reference", "]", "illustrates", "the", "SGD", "-", "based", "training", "algorithm", "for", "the", "transductive", "formulation", ".", "Similarly", ",", "we", "can", "replace", "with", "in", "to", "obtain", "the", "training", "algorithm", "for", "the", "inductive", "formulation", ".", "Let", "denote", "all", "model", "parameters", ".", "We", "update", "both", "embeddings", "and", "parameters", "in", "transductive", "learning", ",", "and", "update", "only", "parameters", "in", "inductive", "learning", ".", "Before", "the", "joint", "training", "procedure", ",", "we", "apply", "a", "number", "of", "training", "iterations", "that", "optimize", "the", "unsupervised", "loss", "alone", "and", "use", "the", "learned", "embeddings", "as", "initialization", "for", "joint", "training", ".", "[", "tb", "]", "Model", "Training", "(", "Transductive", ")", "Input", ":", ",", ",", ",", ",", "batch", "iterations", "and", "sizes", "to", "Sample", "a", "batch", "of", "labeled", "instances", "of", "size", "Take", "a", "gradient", "step", "for", "to", "Sample", "a", "batch", "of", "context", "from", "of", "size", "Take", "a", "gradient", "step", "for", "stopping", "section", ":", "Experiments", "In", "our", "experiments", ",", "Planetoid", "-", "T", "and", "Planetoid", "-", "I", "denote", "the", "transductive", "and", "inductive", "formulation", "of", "our", "approach", ".", "We", "compare", "our", "approach", "with", "label", "propagation", "(", "LP", ")", ",", "semi", "-", "supervised", "embedding", "(", "SemiEmb", ")", ",", "manifold", "regularization", "(", "ManiReg", ")", ",", "TSVM", ",", "and", "graph", "embeddings", "(", "GraphEmb", ")", ".", "Another", "baseline", "method", ",", "denoted", "as", "Feat", ",", "is", "a", "linear", "softmax", "model", "that", "takes", "only", "the", "feature", "vectors", "as", "input", ".", "We", "also", "derive", "a", "variant", "Planetoid", "-", "G", "that", "learns", "embeddings", "to", "jointly", "predict", "class", "labels", "and", "graph", "context", "without", "use", "of", "feature", "vectors", ".", "The", "architecture", "of", "Planetoid", "-", "G", "is", "similar", "to", "Figure", "[", "reference", "]", "except", "that", "the", "input", "feature", "and", "the", "corresponding", "hidden", "layers", "are", "removed", ".", "Among", "the", "above", "methods", ",", "LP", ",", "GraphEmb", "and", "Planetoid", "-", "G", "do", "not", "use", "the", "features", ",", "while", "TSVM", "and", "Feat", "do", "not", "use", "the", "graph", ".", "We", "include", "these", "methods", "into", "our", "experimental", "settings", "to", "better", "evaluate", "our", "approach", ".", "Our", "preliminary", "experiments", "on", "the", "text", "classification", "datasets", "show", "that", "the", "performance", "of", "our", "model", "is", "not", "very", "sensitive", "to", "specific", "choices", "of", "the", "network", "architecture", ".", "We", "adapt", "the", "implementation", "of", "GraphEmb", "to", "our", "Skipgram", "implementation", ".", "We", "use", "the", "Junto", "library", "for", "label", "propagation", ",", "and", "SVMLight", "for", "TSVM", ".", "We", "also", "use", "our", "own", "implementation", "of", "ManiReg", "and", "SemiEmb", "by", "modifying", "the", "symbolic", "objective", "function", "in", "Planetoid", ".", "In", "all", "of", "our", "experiments", ",", "we", "set", "the", "model", "hyper", "-", "parameters", "to", ",", ",", ",", "and", "for", "Planetoid", ".", "We", "use", "the", "same", ",", "and", "for", "GraphEmb", ",", "and", "the", "same", "and", "for", "ManiReg", "and", "SemiEmb", ".", "We", "tune", ",", ",", ",", "the", "learning", "rate", "and", "hyper", "-", "parameters", "in", "other", "models", "based", "on", "an", "additional", "data", "split", "with", "a", "different", "random", "seed", ".", "The", "statistics", "for", "five", "of", "our", "benchmark", "datasets", "are", "reported", "in", "Table", "[", "reference", "]", ".", "For", "each", "dataset", ",", "we", "split", "all", "instances", "into", "three", "parts", ",", "labeled", "data", ",", "unlabeled", "data", ",", "and", "test", "data", ".", "Inductive", "methods", "are", "trained", "on", "the", "labeled", "and", "unlabeled", "data", ",", "and", "tested", "on", "the", "test", "data", ".", "Transductive", "methods", ",", "on", "the", "other", "hand", ",", "are", "trained", "on", "the", "labeled", ",", "unlabeled", "data", ",", "and", "test", "data", "without", "labels", ".", "subsection", ":", "Text", "Classification", "We", "first", "considered", "three", "text", "classification", "datasets", ",", "Citeseer", ",", "Cora", "and", "Pubmed", ".", "Each", "dataset", "contains", "bag", "-", "of", "-", "words", "representation", "of", "documents", "and", "citation", "links", "between", "the", "documents", ".", "We", "treat", "the", "bag", "-", "of", "-", "words", "as", "feature", "vectors", ".", "We", "construct", "the", "graph", "based", "on", "the", "citation", "links", ";", "if", "document", "cites", ",", "then", "we", "set", ".", "The", "goal", "is", "to", "classify", "each", "document", "into", "one", "class", ".", "We", "randomly", "sample", "instances", "for", "each", "class", "as", "labeled", "data", ",", "instances", "as", "test", "data", ",", "and", "the", "rest", "are", "used", "as", "unlabeled", "data", ".", "The", "same", "data", "splits", "are", "used", "for", "different", "methods", ",", "and", "we", "compute", "the", "average", "accuracy", "for", "comparison", ".", "The", "experimental", "results", "are", "reported", "in", "Table", "[", "reference", "]", ".", "Among", "the", "inductive", "methods", ",", "Planetoid", "-", "I", "achieves", "the", "best", "performance", "on", "all", "the", "three", "datasets", "with", "the", "improvement", "of", "up", "to", "on", "Pubmed", ",", "which", "indicates", "that", "our", "embedding", "techniques", "are", "more", "effective", "than", "graph", "Laplacian", "regularization", ".", "Among", "the", "transductive", "methods", ",", "Planetoid", "-", "T", "achieves", "the", "best", "performance", "on", "Cora", "and", "Pubmed", ",", "while", "TSVM", "performs", "the", "best", "on", "Citeseer", ".", "However", ",", "TSVM", "does", "not", "perform", "well", "on", "Cora", "and", "Pubmed", ".", "Planetoid", "-", "I", "slightly", "outperforms", "Planetoid", "-", "T", "on", "Citeseer", "and", "Pubmed", ",", "while", "Planetoid", "-", "T", "gets", "up", "to", "improvement", "over", "Planetoid", "-", "I", "on", "Cora", ".", "We", "conjecture", "that", "in", "Planetoid", "-", "I", ",", "the", "feature", "vectors", "impose", "constraints", "on", "the", "learned", "embeddings", ",", "since", "they", "are", "represented", "by", "a", "parameterized", "function", "of", "the", "input", "feature", "vectors", ".", "If", "such", "constraints", "are", "appropriate", ",", "as", "is", "the", "case", "on", "Citeseer", "and", "Pubmed", ",", "it", "improves", "the", "non", "-", "convex", "optimization", "of", "embedding", "learning", "and", "leads", "to", "better", "performance", ".", "However", ",", "if", "such", "constraints", "rule", "out", "the", "optimal", "embeddings", ",", "the", "inductive", "model", "will", "suffer", ".", "Planetoid", "-", "G", "consistently", "outperforms", "GraphEmb", "on", "all", "three", "datasets", ",", "which", "indicates", "that", "joint", "training", "with", "label", "information", "can", "improve", "the", "performance", "over", "training", "the", "supervised", "and", "unsupervised", "objectives", "separately", ".", "Figure", "[", "reference", "]", "displays", "the", "-", "D", "embedding", "spaces", "on", "the", "Cora", "dataset", "using", "t", "-", "SNE", ".", "Note", "that", "different", "classes", "are", "better", "separated", "in", "the", "embedding", "space", "of", "Planetoid", "-", "T", "than", "that", "of", "GraphEmb", "and", "SemiEmb", ",", "which", "is", "consistent", "with", "our", "empirical", "findings", ".", "We", "also", "observe", "similar", "results", "for", "the", "other", "two", "datasets", ".", "subsection", ":", "Distantly", "-", "Supervised", "Entity", "Extraction", "We", "next", "considered", "the", "DIEL", "(", "Distant", "Information", "Extraction", "using", "coordinate", "-", "term", "Lists", ")", "dataset", ".", "The", "DIEL", "dataset", "contains", "pre", "-", "extracted", "features", "for", "each", "entity", "mention", "in", "text", ",", "and", "a", "graph", "that", "connects", "entity", "mentions", "to", "coordinate", "lists", ".", "The", "goal", "is", "to", "extract", "medical", "entities", "from", "text", "given", "feature", "vectors", "and", "the", "graph", ".", "We", "follow", "the", "exact", "experimental", "setup", "as", "in", "the", "original", "DIEL", "paper", ",", "including", "data", "splits", "of", "different", "runs", ",", "preprocessing", "of", "entity", "mentions", "and", "coordinate", "lists", ",", "and", "evaluation", ".", "We", "treat", "the", "top", "-", "entities", "given", "by", "a", "model", "as", "positive", "instances", ",", "and", "compute", "recall@", "for", "evaluation", "(", "is", "set", "to", "following", "the", "DIEL", "paper", ")", ".", "We", "report", "the", "average", "result", "of", "10", "runs", "in", "Table", "[", "reference", "]", ",", "where", "Feat", "refers", "to", "a", "result", "obtained", "by", "SVM", "(", "referred", "to", "as", "DS", "-", "Baseline", "in", "the", "DIEL", "paper", ")", ".", "The", "result", "of", "LP", "was", "also", "taken", "from", ".", "DIEL", "in", "Table", "[", "reference", "]", "refers", "to", "the", "method", "proposed", "by", "the", "original", "paper", ",", "which", "is", "an", "improved", "version", "of", "label", "propagation", "that", "trains", "classifiers", "on", "feature", "vectors", "based", "on", "the", "output", "of", "label", "propagation", ".", "We", "did", "not", "include", "TSVM", "into", "the", "comparison", "since", "it", "does", "not", "scale", ".", "Since", "we", "use", "Freebase", "as", "ground", "truth", "and", "some", "entities", "are", "not", "present", "in", "text", ",", "the", "upper", "bound", "of", "recall", "as", "shown", "in", "Table", "[", "reference", "]", "is", ".", "Both", "Planetoid", "-", "I", "and", "Planetoid", "-", "T", "significantly", "outperform", "all", "other", "methods", ".", "Each", "of", "Planetoid", "-", "I", "and", "Planetoid", "-", "T", "achieves", "the", "best", "performance", "in", "5", "out", "of", "10", "runs", ",", "and", "they", "give", "a", "similar", "recall", "on", "average", ",", "which", "indicates", "that", "there", "is", "no", "significant", "difference", "between", "these", "two", "methods", "on", "this", "dataset", ".", "Planetoid", "-", "G", "clearly", "outperforms", "GraphEmb", ",", "which", "again", "shows", "the", "benefit", "of", "joint", "training", ".", "subsection", ":", "Entity", "Classification", "We", "sorted", "out", "an", "entity", "classification", "dataset", "from", "the", "knowledge", "base", "of", "Never", "Ending", "Language", "Learning", "(", "NELL", ")", "and", "a", "hierarchical", "entity", "classification", "dataset", "that", "links", "NELL", "entities", "to", "text", "in", "ClueWeb09", ".", "We", "extracted", "the", "entities", "and", "the", "relations", "between", "entities", "from", "the", "NELL", "knowledge", "base", ",", "and", "then", "obtained", "text", "description", "by", "linking", "the", "entities", "to", "ClueWeb09", ".", "We", "use", "text", "bag", "-", "of", "-", "words", "representation", "as", "feature", "vectors", "of", "the", "entities", ".", "We", "next", "describe", "how", "to", "construct", "the", "graph", "based", "on", "the", "knowledge", "base", ".", "We", "first", "remove", "relations", "that", "are", "not", "populated", "in", "NELL", ",", "including", "“", "generalizations", "”", ",", "“", "haswikipediaurl", "”", ",", "and", "“", "atdate", "”", ".", "In", "the", "knowledge", "base", ",", "each", "relation", "is", "denoted", "as", "a", "triplet", ",", "where", ",", ",", "denote", "head", "entity", ",", "relation", ",", "and", "tail", "entity", "respectively", ".", "We", "treat", "each", "entity", "as", "a", "node", "in", "the", "graph", ",", "and", "each", "relation", "is", "split", "as", "two", "nodes", "and", "in", "the", "graph", ".", "For", "each", ",", "we", "add", "two", "edges", "in", "the", "graph", ",", "and", ".", "We", "removed", "all", "classes", "with", "less", "than", "entities", ".", "The", "goal", "is", "to", "classify", "the", "entities", "in", "the", "knowledge", "base", "into", "one", "of", "the", "classes", "given", "the", "feature", "vectors", "and", "the", "graph", ".", "Let", "be", "the", "labeling", "rate", ".", "We", "set", "to", ",", ",", "and", ".", "instances", "are", "labeled", "for", "a", "class", "with", "entities", ",", "so", "each", "class", "has", "at", "least", "one", "entity", "in", "the", "labeled", "data", ".", "We", "report", "the", "results", "in", "Table", "[", "reference", "]", ".", "We", "did", "not", "include", "TSVM", "since", "it", "does", "not", "scale", "to", "such", "a", "large", "number", "of", "classes", "with", "the", "one", "-", "vs", "-", "rest", "scheme", ".", "Adding", "feature", "vectors", "does", "not", "improve", "the", "performance", "of", "Planetoid", "-", "T", ",", "so", "we", "set", "the", "feature", "vectors", "for", "Planetoid", "-", "T", "to", "be", "all", "empty", ",", "and", "therefore", "Planetoid", "-", "T", "is", "equivalent", "to", "Planetoid", "-", "G", "in", "this", "case", ".", "Planetoid", "-", "I", "significantly", "outperforms", "the", "best", "of", "the", "other", "compared", "inductive", "methods", "—", "i.e", ".", ",", "SemiEmb", "—", "by", ",", ",", "and", "respectively", "with", "three", "labeling", "rates", ".", "As", "the", "labeling", "rate", "decreases", ",", "the", "improvement", "of", "Planetoid", "-", "I", "over", "SemiEmb", "becomes", "more", "significant", ".", "Graph", "structure", "is", "more", "informative", "than", "features", "in", "this", "dataset", ",", "so", "inductive", "methods", "perform", "worse", "than", "transductive", "methods", ".", "Planetoid", "-", "G", "outperforms", "GraphEmb", "by", ",", "and", ".", "section", ":", "Conclusion", "Our", "contribution", "is", "three", "-", "fold", ":", "a", ")", "incontrast", "to", "previous", "semi", "-", "supervised", "learning", "approaches", "that", "largely", "depend", "on", "graph", "Laplacian", "regularization", ",", "we", "propose", "a", "novel", "approach", "by", "joint", "training", "of", "classification", "and", "graph", "context", "prediction", ";", "b", ")", "since", "it", "is", "difficult", "to", "generalize", "graph", "embeddings", "to", "novel", "instances", ",", "we", "design", "a", "novel", "inductive", "approach", "that", "conditions", "embeddings", "on", "input", "features", ";", "c", ")", "we", "empirically", "show", "substantial", "improvement", "over", "existing", "methods", "(", "up", "to", "and", "on", "average", ")", ",", "and", "even", "more", "significant", "improvement", "in", "the", "inductive", "setting", "(", "up", "to", "and", "on", "average", ")", ".", "Our", "experimental", "results", "on", "five", "benchmark", "datasets", "also", "show", "that", "a", ")", "joint", "training", "gives", "improvement", "over", "unsupervised", "learning", ";", "b", ")", "predicting", "graph", "context", "is", "more", "effective", "than", "graph", "Laplacian", "regularization", ";", "c", ")", "the", "performance", "of", "the", "inductive", "variant", "depends", "on", "the", "informativeness", "of", "feature", "vectors", ".", "One", "direction", "of", "future", "work", "would", "be", "to", "apply", "our", "framework", "to", "more", "complex", "networks", ",", "including", "recurrent", "networks", ".", "It", "would", "also", "be", "interesting", "to", "experiment", "with", "datasets", "where", "a", "graph", "is", "computed", "based", "on", "distances", "between", "feature", "vectors", ".", "section", ":", "Acknowledgements", "This", "work", "was", "funded", "by", "the", "NSF", "under", "grants", "CCF", "-", "1414030", "and", "IIS", "-", "1250956", ",", "and", "by", "Google", ".", "bibliography", ":", "References"], "entity": [[3, 8, ["semi", "-", "supervised", "learning", "framework"], "Method"], [10, 12, ["graph", "embeddings"], "Method"], [22, 23, ["embedding"], "Method"], [77, 79, ["inductive", "variant"], "Method"], [116, 118, ["text", "classification"], "Task"], [119, 123, ["distantly", "supervised", "entity", "extraction"], "Task"], [125, 127, ["entity", "classification"], "Task"], [145, 149, ["Semi", "-", "supervised", "learning"], "Method"], [162, 167, ["semi", "-", "supervised", "learning", "algorithms"], "Method"], [170, 173, ["training", "objective", "functions"], "Method"], [191, 198, ["Graph", "-", "based", "semi", "-", "supervised", "learning"], "Method"], [204, 206, ["weighted", "sum"], "Method"], [215, 219, ["graph", "Laplacian", "regularization", "term"], "Method"], [221, 224, ["graph", "Laplacian", "regularization"], "Method"], [243, 246, ["Graph", "Laplacian", "regularization"], "Method"], [263, 267, ["unsupervised", "representation", "learning", "methods"], "Method"], [268, 269, ["embeddings"], "Method"], [287, 289, ["node", "embedding"], "Method"], [297, 298, ["Embeddings"], "Method"], [310, 312, ["related", "tasks"], "Task"], [316, 318, ["word", "embeddings"], "Method"], [321, 323, ["language", "model"], "Method"], [327, 333, ["part", "-", "of", "-", "speech", "tagging"], "Task"], [334, 335, ["chunking"], "Task"], [336, 339, ["named", "entity", "recognition"], "Task"], [346, 348, ["word", "embeddings"], "Method"], [349, 351, ["graph", "embeddings"], "Method"], [356, 358, ["graph", "embeddings"], "Method"], [370, 373, ["user", "behavior", "prediction"], "Task"], [375, 377, ["social", "network"], "Task"], [389, 391, ["supervised", "task"], "Task"], [406, 408, ["graph", "embeddings"], "Method"], [414, 417, ["graph", "Laplacian", "regularization"], "Method"], [450, 452, ["embedding", "techniques"], "Method"], [454, 462, ["graph", "-", "based", "semi", "-", "supervised", "learning", "setting"], "Task"], [467, 475, ["graph", "-", "based", "semi", "-", "supervised", "learning", "framework"], "Method"], [476, 477, ["Planetoid"], "Method"], [526, 527, ["classifier"], "Method"], [532, 534, ["softmax", "layer"], "Method"], [537, 538, ["prediction"], "Task"], [554, 555, ["transductive"], "Method"], [587, 589, ["inductive", "approach"], "Method"], [615, 617, ["inductive", "variant"], "Method"], [647, 649, ["neural", "network"], "Method"], [670, 672, ["text", "classification"], "Task"], [673, 677, ["distantly", "supervised", "entity", "extraction"], "Task"], [679, 681, ["entity", "classification"], "Task"], [683, 685, ["inductive", "method"], "Method"], [689, 691, ["inductive", "method"], "Method"], [702, 703, ["accuracy"], "Metric"], [734, 738, ["Semi", "-", "Supervised", "Learning"], "Task"], [779, 783, ["semi", "-", "supervised", "learning"], "Task"], [788, 789, ["classifier"], "Method"], [793, 795, ["learning", "paradigms"], "Method"], [799, 801, ["inductive", "learning"], "Method"], [802, 804, ["Transductive", "learning"], "Method"], [809, 810, ["classifier"], "Method"], [821, 822, ["classifier"], "Method"], [837, 838, ["TSVM"], "Method"], [849, 854, ["low", "-", "density", "separation", "assumption"], "Method"], [869, 871, ["Inductive", "learning"], "Method"], [881, 883, ["parameterized", "classifier"], "Method"], [892, 899, ["Graph", "-", "Based", "Semi", "-", "Supervised", "Learning"], "Method"], [919, 927, ["graph", "-", "based", "semi", "-", "supervised", "learning", "methods"], "Method"], [971, 973, ["citation", "network"], "Method"], [1019, 1025, ["bag", "-", "of", "-", "words", "representation"], "Method"], [1030, 1037, ["Graph", "-", "based", "semi", "-", "supervised", "learning"], "Method"], [1055, 1057, ["loss", "function"], "Method"], [1058, 1065, ["graph", "-", "based", "semi", "-", "supervised", "learning"], "Method"], [1067, 1069, ["binary", "case"], "Task"], [1098, 1100, ["squared", "loss"], "Method"], [1101, 1103, ["hinge", "loss"], "Method"], [1109, 1112, ["graph", "Laplacian", "regularization"], "Method"], [1170, 1178, ["graph", "-", "based", "semi", "-", "supervised", "learning", "algorithms"], "Method"], [1193, 1195, ["Label", "propagation"], "Task"], [1220, 1224, ["closed", "-", "form", "solution"], "Method"], [1225, 1226, ["Learning"], "Method"], [1252, 1254, ["Modified", "Adsorption"], "Method"], [1255, 1256, ["MAD"], "Method"], [1261, 1263, ["label", "propagation"], "Method"], [1265, 1266, ["prediction"], "Task"], [1276, 1278, ["Manifold", "regularization"], "Task"], [1281, 1285, ["Reproducing", "Kernel", "Hilbert", "Space"], "Method"], [1286, 1287, ["RKHS"], "Method"], [1299, 1301, ["parameterized", "classifier"], "Method"], [1302, 1304, ["manifold", "regularization"], "Method"], [1313, 1317, ["Semi", "-", "supervised", "embedding"], "Task"], [1352, 1354, ["neural", "network"], "Method"], [1369, 1371, ["neural", "network"], "Method"], [1372, 1375, ["Iterative", "classification", "algorithm"], "Method"], [1376, 1377, ["ICA"], "Method"], [1380, 1382, ["local", "classifier"], "Method"], [1395, 1397, ["iterative", "process"], "Method"], [1400, 1402, ["local", "classifier"], "Method"], [1409, 1411, ["Learning", "Embeddings"], "Task"], [1416, 1419, ["learning", "graph", "embeddings"], "Task"], [1421, 1424, ["probabilistic", "generative", "model"], "Method"], [1439, 1441, ["clustering", "method"], "Method"], [1463, 1466, ["embedding", "learning", "methods"], "Method"], [1470, 1472, ["Skipgram", "model"], "Method"], [1479, 1481, ["softmax", "model"], "Method"], [1492, 1493, ["Skipgram"], "Task"], [1499, 1505, ["log", "loss", "of", "predicting", "the", "context"], "Task"], [1549, 1551, ["Skipgram", "model"], "Method"], [1555, 1558, ["embedding", "of", "instance"], "Method"], [1559, 1560, ["Skipgram"], "Method"], [1565, 1568, ["representations", "of", "words"], "Task"], [1621, 1622, ["Skipgram"], "Method"], [1627, 1629, ["graph", "embeddings"], "Task"], [1630, 1631, ["Deepwalk"], "Method"], [1651, 1653, ["random", "walk"], "Method"], [1717, 1722, ["first", "and", "second", "order", "proximity"], "Task"], [1724, 1728, ["Skipgram", "-", "like", "models"], "Method"], [1729, 1730, ["graphs"], "Method"], [1741, 1742, ["TransE"], "Method"], [1756, 1757, ["Autoencoders"], "Method"], [1761, 1763, ["graph", "embeddings"], "Method"], [1764, 1767, ["clustering", "on", "graphs"], "Task"], [1782, 1786, ["semi", "-", "supervised", "learning"], "Task"], [1787, 1789, ["embedding", "learning"], "Task"], [1800, 1804, ["graph", "Laplacian", "based", "methods"], "Method"], [1815, 1820, ["Semi", "-", "supervised", "embedding", "method"], "Method"], [1821, 1822, ["embeddings"], "Method"], [1824, 1826, ["neural", "network"], "Method"], [1854, 1857, ["Graph", "embedding", "methods"], "Method"], [1892, 1893, ["transductive"], "Method"], [1908, 1912, ["Semi", "-", "Supervised", "Learning"], "Task"], [1913, 1915, ["Graph", "Embeddings"], "Task"], [1957, 1962, ["feed", "-", "forward", "neural", "networks"], "Method"], [2008, 2011, ["rectified", "linear", "unit"], "Method"], [2020, 2022, ["loss", "function"], "Method"], [2032, 2038, ["supervised", "loss", "of", "predicting", "the", "labels"], "Task"], [2042, 2049, ["unsupervised", "loss", "of", "predicting", "the", "graph", "context"], "Task"], [2074, 2079, ["semi", "-", "supervised", "learning", "framework"], "Method"], [2087, 2089, ["unsupervised", "loss"], "Task"], [2147, 2149, ["negative", "sampling"], "Task"], [2185, 2187, ["Negative", "sampling"], "Task"], [2237, 2240, ["cross", "entropy", "loss"], "Metric"], [2275, 2277, ["unsupervised", "loss"], "Task"], [2278, 2280, ["negative", "sampling"], "Task"], [2324, 2326, ["sampling", "process"], "Method"], [2533, 2538, ["random", "walk", "based", "sampling", "method"], "Method"], [2541, 2542, ["Deepwalk"], "Method"], [2557, 2559, ["negative", "sampling"], "Method"], [2614, 2616, ["random", "walk"], "Method"], [2636, 2638, ["Transductive", "Formulation"], "Method"], [2659, 2661, ["Transductive", "learning"], "Method"], [2665, 2667, ["inductive", "learning"], "Method"], [2696, 2697, ["embedding"], "Method"], [2719, 2721, ["softmax", "layer"], "Method"], [2777, 2779, ["loss", "function"], "Method"], [2813, 2815, ["loss", "function"], "Method"], [2816, 2819, ["class", "label", "prediction"], "Task"], [2828, 2830, ["context", "prediction"], "Task"], [2837, 2840, ["prediction", "of", "label"], "Task"], [2862, 2864, ["Inductive", "Formulation"], "Method"], [2884, 2885, ["classifier"], "Method"], [2894, 2898, ["large", "-", "scale", "tasks"], "Task"], [2902, 2905, ["machine", "reading", "systems"], "Task"], [2921, 2926, ["semi", "-", "supervised", "learning", "system"], "Method"], [2934, 2937, ["learning", "graph", "embeddings"], "Task"], [2951, 2953, ["inductive", "setting"], "Task"], [2990, 2993, ["prediction", "of", "label"], "Task"], [3031, 3032, ["layers"], "Method"], [3047, 3050, ["free", "”", "embedding"], "Method"], [3073, 3074, ["embedding"], "Task"], [3080, 3081, ["embedding"], "Task"], [3089, 3091, ["hidden", "layer"], "Method"], [3128, 3130, ["loss", "function"], "Method"], [3131, 3133, ["inductive", "learning"], "Method"], [3154, 3157, ["stochastic", "gradient", "descent"], "Method"], [3158, 3160, ["SGD", ")"], "Method"], [3166, 3170, ["mini", "-", "batch", "mode"], "Task"], [3182, 3184, ["gradient", "step"], "Method"], [3187, 3193, ["loss", "function", "of", "class", "label", "prediction"], "Task"], [3204, 3206, ["gradient", "step"], "Method"], [3209, 3214, ["loss", "function", "of", "context", "prediction"], "Task"], [3236, 3241, ["SGD", "-", "based", "training", "algorithm"], "Method"], [3256, 3258, ["training", "algorithm"], "Method"], [3260, 3262, ["inductive", "formulation"], "Method"], [3284, 3286, ["inductive", "learning"], "Task"], [3289, 3292, ["joint", "training", "procedure"], "Method"], [3298, 3300, ["training", "iterations"], "Method"], [3314, 3316, ["joint", "training"], "Task"], [3320, 3322, ["Model", "Training"], "Method"], [3323, 3324, ["Transductive"], "Task"], [3346, 3348, ["gradient", "step"], "Method"], [3371, 3374, ["Planetoid", "-", "T"], "Method"], [3375, 3378, ["Planetoid", "-", "I"], "Method"], [3380, 3381, ["transductive"], "Method"], [3382, 3384, ["inductive", "formulation"], "Method"], [3393, 3395, ["label", "propagation"], "Method"], [3396, 3397, ["LP"], "Method"], [3399, 3406, ["semi", "-", "supervised", "embedding", "(", "SemiEmb", ")"], "Method"], [3407, 3409, ["manifold", "regularization"], "Method"], [3410, 3411, ["ManiReg"], "Method"], [3413, 3414, ["TSVM"], "Method"], [3416, 3418, ["graph", "embeddings"], "Method"], [3419, 3420, ["GraphEmb"], "Method"], [3428, 3429, ["Feat"], "Method"], [3432, 3435, ["linear", "softmax", "model"], "Method"], [3449, 3452, ["Planetoid", "-", "G"], "Method"], [3472, 3475, ["Planetoid", "-", "G"], "Method"], [3500, 3501, ["LP"], "Method"], [3502, 3503, ["GraphEmb"], "Method"], [3504, 3507, ["Planetoid", "-", "G"], "Method"], [3514, 3515, ["TSVM"], "Method"], [3516, 3517, ["Feat"], "Method"], [3561, 3563, ["network", "architecture"], "Method"], [3569, 3570, ["GraphEmb"], "Method"], [3572, 3574, ["Skipgram", "implementation"], "Method"], [3578, 3580, ["Junto", "library"], "Method"], [3581, 3583, ["label", "propagation"], "Task"], [3585, 3586, ["SVMLight"], "Method"], [3587, 3588, ["TSVM"], "Method"], [3596, 3597, ["ManiReg"], "Method"], [3598, 3599, ["SemiEmb"], "Method"], [3606, 3607, ["Planetoid"], "Method"], [3627, 3628, ["Planetoid"], "Method"], [3636, 3637, ["GraphEmb"], "Method"], [3643, 3644, ["ManiReg"], "Method"], [3645, 3646, ["SemiEmb"], "Method"], [3653, 3655, ["learning", "rate"], "Metric"], [3712, 3714, ["Inductive", "methods"], "Method"], [3730, 3732, ["Transductive", "methods"], "Method"], [3755, 3757, ["Text", "Classification"], "Task"], [3765, 3766, ["Citeseer"], "Material"], [3767, 3768, ["Cora"], "Material"], [3769, 3770, ["Pubmed"], "Material"], [3774, 3782, ["bag", "-", "of", "-", "words", "representation", "of", "documents"], "Method"], [3887, 3889, ["inductive", "methods"], "Method"], [3890, 3893, ["Planetoid", "-", "I"], "Method"], [3909, 3910, ["Pubmed"], "Material"], [3915, 3917, ["embedding", "techniques"], "Method"], [3921, 3924, ["graph", "Laplacian", "regularization"], "Method"], [3930, 3933, ["Planetoid", "-", "T"], "Method"], [3938, 3939, ["Cora"], "Material"], [3940, 3941, ["Pubmed"], "Material"], [3943, 3944, ["TSVM"], "Method"], [3948, 3949, ["Citeseer"], "Material"], [3952, 3953, ["TSVM"], "Method"], [3958, 3959, ["Cora"], "Material"], [3960, 3961, ["Pubmed"], "Material"], [3962, 3965, ["Planetoid", "-", "I"], "Method"], [3967, 3970, ["Planetoid", "-", "T"], "Method"], [3971, 3972, ["Citeseer"], "Material"], [3973, 3974, ["Pubmed"], "Material"], [3976, 3979, ["Planetoid", "-", "T"], "Method"], [3984, 3987, ["Planetoid", "-", "I"], "Method"], [3988, 3989, ["Cora"], "Material"], [3994, 3997, ["Planetoid", "-", "I"], "Method"], [4033, 4034, ["Citeseer"], "Material"], [4035, 4036, ["Pubmed"], "Material"], [4040, 4047, ["non", "-", "convex", "optimization", "of", "embedding", "learning"], "Task"], [4065, 4067, ["inductive", "model"], "Method"], [4070, 4073, ["Planetoid", "-", "G"], "Method"], [4075, 4076, ["GraphEmb"], "Method"], [4084, 4086, ["joint", "training"], "Method"], [4114, 4116, ["Cora", "dataset"], "Material"], [4117, 4120, ["t", "-", "SNE"], "Method"], [4133, 4136, ["Planetoid", "-", "T"], "Method"], [4139, 4140, ["GraphEmb"], "Method"], [4141, 4142, ["SemiEmb"], "Method"], [4164, 4169, ["Distantly", "-", "Supervised", "Entity", "Extraction"], "Task"], [4173, 4178, ["DIEL", "(", "Distant", "Information", "Extraction"], "Task"], [4275, 4276, ["recall@"], "Metric"], [4277, 4278, ["evaluation"], "Task"], [4310, 4311, ["SVM"], "Method"], [4315, 4318, ["DS", "-", "Baseline"], "Method"], [4333, 4334, ["DIEL"], "Method"], [4355, 4357, ["label", "propagation"], "Method"], [4359, 4360, ["classifiers"], "Method"], [4368, 4370, ["label", "propagation"], "Method"], [4375, 4376, ["TSVM"], "Method"], [4405, 4406, ["recall"], "Metric"], [4416, 4419, ["Planetoid", "-", "I"], "Method"], [4420, 4423, ["Planetoid", "-", "T"], "Method"], [4431, 4434, ["Planetoid", "-", "I"], "Method"], [4435, 4438, ["Planetoid", "-", "T"], "Method"], [4454, 4455, ["recall"], "Metric"], [4474, 4477, ["Planetoid", "-", "G"], "Method"], [4479, 4480, ["GraphEmb"], "Method"], [4487, 4489, ["joint", "training"], "Method"], [4492, 4494, ["Entity", "Classification"], "Task"], [4506, 4510, ["Never", "Ending", "Language", "Learning"], "Material"], [4511, 4512, ["NELL"], "Material"], [4521, 4523, ["NELL", "entities"], "Material"], [4557, 4564, ["text", "bag", "-", "of", "-", "words", "representation"], "Method"], [4582, 4584, ["knowledge", "base"], "Material"], [4594, 4595, ["NELL"], "Material"], [4612, 4614, ["knowledge", "base"], "Material"], [4710, 4712, ["labeling", "rate"], "Metric"], [4756, 4757, ["TSVM"], "Method"], [4771, 4777, ["one", "-", "vs", "-", "rest", "scheme"], "Method"], [4787, 4790, ["Planetoid", "-", "T"], "Method"], [4798, 4801, ["Planetoid", "-", "T"], "Method"], [4808, 4811, ["Planetoid", "-", "T"], "Method"], [4814, 4817, ["Planetoid", "-", "G"], "Method"], [4821, 4824, ["Planetoid", "-", "I"], "Method"], [4832, 4834, ["inductive", "methods"], "Method"], [4847, 4849, ["labeling", "rates"], "Metric"], [4852, 4854, ["labeling", "rate"], "Metric"], [4859, 4862, ["Planetoid", "-", "I"], "Method"], [4863, 4864, ["SemiEmb"], "Method"], [4880, 4882, ["inductive", "methods"], "Method"], [4888, 4891, ["Planetoid", "-", "G"], "Method"], [4892, 4893, ["GraphEmb"], "Method"], [4912, 4917, ["semi", "-", "supervised", "learning", "approaches"], "Method"], [4921, 4924, ["graph", "Laplacian", "regularization"], "Method"], [4931, 4939, ["joint", "training", "of", "classification", "and", "graph", "context", "prediction"], "Task"], [4958, 4960, ["inductive", "approach"], "Method"], [4992, 4994, ["inductive", "setting"], "Task"], [5014, 5016, ["joint", "training"], "Method"], [5019, 5021, ["unsupervised", "learning"], "Method"], [5024, 5027, ["predicting", "graph", "context"], "Task"], [5031, 5034, ["graph", "Laplacian", "regularization"], "Method"], [5041, 5043, ["inductive", "variant"], "Method"], [5068, 5070, ["recurrent", "networks"], "Method"], [43, 44, ["transductive"], "Method"], [53, 54, ["transductive"], "Method"], [710, 711, ["transductive"], "Method"], [796, 797, ["transductive"], "Method"], [832, 833, ["transductive"], "Method"], [2668, 2669, ["transductive"], "Method"], [2780, 2781, ["transductive"], "Method"], [2834, 2835, ["transductive"], "Method"], [2867, 2868, ["transductive"], "Method"], [2938, 2939, ["transductive"], "Method"], [3026, 3027, ["transductive"], "Method"], [3243, 3244, ["transductive"], "Method"], [3276, 3277, ["transductive"], "Method"], [3542, 3544, ["text", "classification"], "Task"], [3761, 3763, ["text", "classification"], "Task"], [3870, 3871, ["accuracy"], "Metric"], [3927, 3928, ["transductive"], "Method"], [4498, 4500, ["entity", "classification"], "Task"], [4503, 4505, ["knowledge", "base"], "Material"], [4516, 4518, ["entity", "classification"], "Task"], [4540, 4542, ["knowledge", "base"], "Material"], [4692, 4694, ["knowledge", "base"], "Material"], [4885, 4886, ["transductive"], "Method"]]} {"doc_id": "0e6b990f41fc0e7fdd0662f88b5ad024375bd77b", "text": ["Learning", "Deep", "Parsimonious", "Representations", "In", "this", "paper", "we", "aim", "at", "facilitating", "generalization", "for", "deep", "networks", "while", "supporting", "interpretability", "of", "the", "learned", "representations", ".", "Towards", "this", "goal", ",", "we", "propose", "a", "clustering", "based", "regularization", "that", "encourages", "parsimonious", "representations", ".", "Our", "k", "-", "means", "style", "objective", "is", "easy", "to", "optimize", "and", "flexible", ",", "supporting", "various", "forms", "of", "clustering", ",", "such", "as", "sample", "clustering", ",", "spatial", "clustering", ",", "as", "well", "as", "co", "-", "clustering", ".", "We", "demonstrate", "the", "effectiveness", "of", "our", "approach", "on", "the", "tasks", "of", "unsupervised", "learning", ",", "classification", ",", "fine", "grained", "categorization", ",", "and", "zero", "-", "shot", "learning", ".", "1", "Introduction", "In", "recent", "years", ",", "deep", "neural", "networks", "have", "been", "shown", "to", "perform", "extremely", "well", "on", "a", "variety", "of", "tasks", "including", "classification", "[", "21", "]", ",", "semantic", "segmentation", "[", "13", "]", ",", "machine", "translation", "[", "27", "]", "and", "speech", "recognition", "[", "16", "]", ".", "This", "has", "led", "to", "their", "adoption", "across", "many", "areas", "such", "as", "computer", "vision", ",", "natural", "language", "processing", "and", "robotics", "[", "16", ",", "21", ",", "22", ",", "27", "]", ".", "Three", "major", "advances", "are", "responsible", "for", "the", "recent", "success", "of", "neural", "networks", ":", "the", "increase", "in", "available", "computational", "resources", ",", "access", "to", "large", "scale", "data", "sets", ",", "and", "several", "algorithmic", "improvements", ".", "Many", "of", "these", "algorithmic", "advances", "are", "related", "to", "regularization", ",", "which", "is", "key", "to", "prevent", "overfitting", "and", "improve", "generalization", "of", "the", "learned", "classifier", ",", "as", "the", "current", "trend", "is", "to", "increase", "the", "capacity", "of", "neural", "nets", ".", "For", "example", ",", "batch", "normalization", "[", "18", "]", "is", "used", "to", "normalize", "intermediate", "representations", "which", "can", "be", "interpreted", "as", "imposing", "constraints", ".", "In", "contrast", ",", "dropout", "[", "26", "]", "removes", "a", "fraction", "of", "the", "learned", "representations", "at", "random", "to", "prevent", "co", "-", "adaptation", ".", "Learning", "of", "de", "-", "correlated", "activations", "[", "6", "]", "shares", "a", "similar", "idea", "since", "it", "explicitly", "discourages", "correlation", "between", "the", "units", ".", "In", "this", "paper", "we", "propose", "a", "new", "type", "of", "regularization", "that", "encourages", "the", "network", "representations", "to", "form", "clusters", ".", "As", "a", "consequence", ",", "the", "learned", "feature", "space", "is", "compactly", "representable", ",", "facilitating", "generalization", ".", "Furthermore", ",", "clustering", "supports", "interpretability", "of", "the", "learned", "representations", ".", "We", "formulate", "our", "regularization", "with", "a", "k", "-", "means", "style", "objective", "which", "is", "easy", "to", "optimize", ",", "and", "investigate", "different", "types", "of", "clusterings", ",", "including", "sample", "clustering", ",", "spatial", "clustering", ",", "and", "co", "-", "clustering", ".", "We", "demonstrate", "the", "generalization", "performance", "of", "our", "proposed", "method", "in", "several", "settings", ":", "autoencoders", "trained", "on", "the", "MNIST", "dataset", "[", "23", "]", ",", "classification", "on", "CIFAR10", "and", "CIFAR100", "[", "20", "]", ",", "as", "well", "as", "fine", "-", "grained", "classification", "and", "zero", "-", "shot", "learning", "on", "the", "CUB", "-", "200", "-", "2011", "dataset", "[", "34", "]", ".", "We", "show", "that", "our", "approach", "leads", "to", "significant", "wins", "in", "all", "these", "scenarios", ".", "In", "addition", ",", "we", "are", "able", "to", "demonstrate", "on", "the", "CUB", "-", "200", "-", "2011", "dataset", "that", "the", "network", "representation", "captures", "meaningful", "part", "representations", "even", "though", "it", "is", "not", "explicitly", "trained", "to", "do", "so", ".", "2", "Related", "Work", "Standard", "neural", "network", "regularization", "involves", "penalties", "on", "the", "weights", "based", "on", "the", "norm", "of", "the", "parameters", "[", "29", ",", "30", "]", ".", "Also", "popular", "are", "regularization", "methods", "applied", "to", "intermediate", "representations", ",", "29th", "Conference", "on", "Neural", "Information", "Processing", "Systems", "(", "NIPS", "2016", ")", ",", "Barcelona", ",", "Spain", ".", "such", "as", "Dropout", "[", "26", "]", ",", "Drop", "-", "Connect", "[", "32", "]", ",", "Maxout", "[", "10", "]", "and", "DeCov", "[", "6", "]", ".", "These", "approaches", "share", "the", "aim", "of", "preventing", "the", "activations", "in", "the", "network", "to", "be", "correlated", ".", "Our", "work", "can", "be", "seen", "as", "a", "different", "form", "of", "regularization", ",", "where", "we", "encourage", "parsimonious", "representations", ".", "A", "variety", "of", "approaches", "have", "applied", "clustering", "to", "the", "parameters", "of", "the", "neural", "network", "with", "the", "aim", "of", "compressing", "the", "network", ".", "Compression", "rates", "of", "more", "than", "an", "order", "of", "magnitude", "were", "demonstrated", "in", "[", "11", "]", "without", "sacrificing", "accuracy", ".", "In", "the", "same", "spirit", "hash", "functions", "were", "exploited", "in", "[", "5", "]", ".", "Early", "approaches", "to", "compression", "include", "biased", "weight", "decay", "[", "12", "]", "and", "[", "14", ",", "24", "]", ",", "which", "prunes", "the", "network", "based", "on", "the", "Hessian", "of", "the", "loss", "function", ".", "Recently", ",", "various", "combinations", "of", "clustering", "with", "representation", "learning", "have", "been", "proposed", ".", "We", "categorize", "them", "broadly", "into", "two", "areas", ":", "(", "i", ")", "work", "that", "applies", "clustering", "after", "having", "learned", "a", "representation", ",", "and", "(", "ii", ")", "approaches", "that", "jointly", "optimize", "the", "learning", "and", "clustering", "objectives", ".", "[", "4", "]", "combines", "deep", "belief", "networks", "(", "DBN", ")", "with", "non", "-", "parametric", "maximum", "-", "margin", "clustering", "in", "a", "posthoc", "manner", ":", "A", "DBN", "is", "trained", "layer", "-", "wise", "to", "obtain", "an", "intermediate", "representation", "of", "the", "data", ";", "non", "-", "parametric", "maximum", "-", "margin", "clustering", "is", "then", "applied", "to", "the", "data", "representation", ".", "Another", "line", "of", "work", "utilizes", "an", "embedding", "of", "the", "deep", "network", ",", "which", "can", "be", "based", "on", "annotated", "data", "[", "15", "]", ",", "or", "from", "a", "learned", "unsupervised", "method", "such", "as", "a", "stacked", "auto", "-", "encoder", "[", "28", "]", ".", "In", "these", "approaches", ",", "the", "network", "is", "trained", "to", "approximate", "the", "embedding", ",", "and", "subsequently", "either", "k", "-", "means", "or", "spectral", "clustering", "is", "performed", "to", "partition", "the", "space", ".", "An", "alternative", "is", "to", "use", "non", "-", "negative", "matrix", "factorization", ",", "which", "represents", "a", "given", "data", "matrix", "as", "the", "product", "of", "components", "[", "31", "]", ".", "This", "deep", "non", "-", "negative", "matrix", "factorization", "is", "trained", "using", "the", "reconstruction", "loss", "rather", "than", "a", "clustering", "objective", ".", "Nonetheless", ",", "it", "was", "shown", "that", "factors", "lower", "in", "the", "hierarchy", "have", "superior", "clustering", "performance", "on", "lowlevel", "concepts", "while", "factors", "later", "in", "the", "hierarchy", "cluster", "high", "-", "level", "concepts", ".", "The", "aforementioned", "approaches", "differ", "from", "our", "proposed", "technique", ",", "since", "we", "aim", "at", "jointly", "learning", "a", "representation", "that", "is", "parsimonious", "via", "a", "clustering", "regularization", ".", "Also", "related", "are", "approaches", "that", "utilize", "sparse", "coding", ".", "Wang", "et", "al", ".", "[", "33", "]", "unrolls", "the", "iterations", "forming", "the", "sparse", "codes", "and", "optimizes", "end", "-", "to", "-", "end", "the", "involved", "parameters", "using", "a", "clustering", "objective", "as", "loss", "function", "[", "33", "]", ".", "The", "proposed", "framework", "is", "further", "augmented", "by", "clustering", "objectives", "applied", "to", "intermediate", "representations", ",", "which", "act", "as", "feature", "regularization", "within", "the", "unrolled", "optimization", ".", "They", "found", "that", "features", "lower", "in", "the", "unrolled", "hierarchy", "cluster", "low", "-", "level", "concepts", ",", "while", "features", "later", "in", "the", "hierarchy", "capture", "high", "-", "level", "concepts", ".", "Our", "method", "differs", "in", "that", "we", "use", "convolutional", "neural", "networks", "rather", "than", "unrolling", "a", "sparse", "coding", "optimization", ".", "In", "the", "context", "of", "unsupervised", "clustering", "[", "35", "]", "exploited", "agglomerative", "clustering", "as", "a", "regularizer", ";", "this", "approach", "was", "formulated", "as", "a", "recurrent", "network", ".", "In", "contrast", "we", "employ", "a", "k", "-", "means", "like", "clustering", "objective", "which", "simplifies", "the", "optimization", "significantly", "and", "does", "not", "require", "a", "recurrent", "procedure", ".", "Furthermore", ",", "we", "investigate", "both", "unsupervised", "and", "supervised", "learning", ".", "3", "Learning", "Deep", "Parsimonious", "Representations", "In", "this", "section", ",", "we", "introduce", "our", "new", "clustering", "based", "regularization", "which", "not", "only", "encourages", "the", "neural", "network", "to", "learn", "more", "compact", "representations", ",", "but", "also", "enables", "interpretability", "of", "the", "neural", "network", ".", "We", "first", "show", "that", "by", "exploiting", "different", "unfoldings", "of", "the", "representation", "tensor", ",", "we", "obtain", "multiple", "types", "of", "clusterings", ",", "each", "possessing", "different", "properties", ".", "We", "then", "devise", "an", "efficient", "online", "update", "to", "jointly", "learn", "the", "clustering", "with", "the", "parameters", "of", "the", "neural", "network", ".", "3.1", "Clustering", "of", "Representations", "We", "first", "introduce", "some", "notation", ".", "We", "refer", "to", "[", "K", "]", "as", "the", "set", "of", "K", "positive", "integers", ",", "i.e.", ",", "[", "K", "]", "=", "{", "1", ",", "2", ",", "...", ",", "K}.", "We", "use", "S\\A", "to", "denote", "the", "set", "S", "with", "elements", "from", "the", "set", "A", "removed", ".", "A", "tensor", "is", "a", "multilinear", "map", "over", "a", "set", "of", "vector", "spaces", ".", "In", "tensor", "terminology", ",", "n", "-", "mode", "vectors", "of", "a", "D", "-", "order", "tensor", "Y", "∈", "RI1×I2×···×ID", "are", "In", "-", "dimensional", "vectors", "obtained", "from", "Y", "by", "varying", "the", "index", "in", "Indimension", ",", "while", "keeping", "all", "other", "indices", "fixed", ".", "An", "n", "-", "mode", "matrix", "unfolding", "of", "a", "tensor", "is", "a", "matrix", "which", "has", "all", "n", "-", "mode", "vectors", "as", "its", "columns", "[", "7", "]", ".", "Formally", "we", "use", "the", "operator", "T", "{", "In}×{Ij", "|j∈", "[", "D", "]", "\\n", "}", "to", "denote", "the", "n", "-", "mode", "matrix", "unfolding", ",", "which", "returns", "a", "matrix", "of", "size", "In", "×", "∏", "j∈", "[", "D", "]", "\\n", "Ij", ".", "Similarly", ",", "we", "definee", "T", "{", "Ii", ",", "Ij}×{Ik|k∈", "[", "D", "]", "\\{i", ",", "j", "}", "}", "to", "be", "an", "(", "i", ",", "j", ")-", "mode", "matrix", "unfolding", "operator", ".", "In", "this", "case", "a", "column", "vector", "is", "a", "concatenation", "of", "one", "i", "-", "mode", "vector", "and", "one", "j", "-", "mode", "vector", ".", "We", "denote", "the", "m", "-", "th", "row", "vector", "of", "a", "matrix", "X", "as", "Xm", ".", "In", "this", "paper", "we", "assume", "the", "representation", "of", "one", "layer", "within", "a", "neural", "network", "to", "be", "a", "4", "-", "D", "tensor", "Y", "∈", "RN×C×H×W", ",", "where", "N", ",", "C", ",", "H", "and", "W", "are", "the", "number", "of", "samples", "within", "a", "mini", "-", "batch", ",", "the", "number", "of", "hidden", "units", ",", "the", "height", "and", "width", "of", "the", "representation", "respectively", ".", "Note", "that", "C", ",", "H", "and", "W", "can", "vary", "between", "layers", ",", "and", "in", "the", "case", "of", "a", "fully", "connected", "layer", ",", "the", "dimensions", "along", "height", "and", "width", "become", "a", "singleton", "and", "the", "tensor", "degenerates", "to", "a", "matrix", ".", "Let", "L", "be", "the", "loss", "function", "of", "a", "neural", "network", ".", "In", "addition", ",", "we", "refer", "to", "the", "clustering", "regularization", "of", "a", "single", "layer", "viaR.", "The", "final", "objective", "is", "L", "+", "λR", ",", "where", "λ", "adjusts", "the", "importance", "of", "the", "clustering", "regularization", ".", "Note", "that", "we", "can", "add", "a", "regularization", "term", "for", "any", "subset", "of", "layers", ",", "but", "we", "focus", "on", "a", "single", "layer", "for", "notational", "simplicity", ".", "In", "what", "follows", ",", "we", "show", "three", "different", "types", "of", "clustering", ",", "each", "possessing", "different", "properties", ".", "In", "our", "framework", "any", "variant", "can", "be", "applied", "to", "any", "layer", ".", "(", "A", ")", "Sample", "Clustering", ":", "We", "first", "investigate", "clustering", "along", "the", "sample", "dimension", ".", "Since", "the", "cluster", "assignments", "of", "different", "layers", "are", "not", "linked", ",", "each", "layer", "is", "free", "to", "cluster", "examples", "in", "a", "different", "way", ".", "For", "example", ",", "in", "a", "ConvNet", ",", "bottom", "layer", "representations", "may", "focus", "on", "low", "-", "level", "visual", "cues", ",", "such", "as", "color", "and", "edges", ",", "while", "top", "layer", "features", "may", "focus", "on", "high", "-", "level", "attributes", "which", "have", "a", "more", "semantic", "meaning", ".", "We", "refer", "the", "reader", "to", "Fig", ".", "1", "(", "a", ")", "for", "an", "illustration", ".", "In", "particular", ",", "given", "the", "representation", "tensor", "Y", ",", "we", "first", "unfold", "it", "into", "a", "matrix", "T", "{", "N}×{H", ",", "W", ",", "C}", "(", "Y", ")", "∈", "RN×HWC", ".", "We", "then", "encourage", "the", "samples", "to", "cluster", "as", "follows", ":", "Rsample", "(", "Y", ",", "µ", ")", "=", "1", "2NCHW", "N∑", "n=1", "∥∥∥T", "{", "N}×{H", ",", "W", ",", "C}", "(", "Y", ")", "n", "−", "µzn∥∥∥2", ",", "(", "1", ")", "where", "µ", "is", "a", "matrix", "of", "size", "K", "×", "HWC", "encoding", "all", "cluster", "centers", ",", "with", "K", "the", "total", "number", "of", "clusters", ".", "zn", "∈", "[", "K", "]", "is", "a", "discrete", "latent", "variable", "corresponding", "to", "the", "n", "-", "th", "sample", ".", "It", "indicates", "which", "cluster", "this", "sample", "belongs", "to", ".", "Note", "that", "for", "a", "fully", "connected", "layer", ",", "the", "formulation", "is", "the", "same", "except", "that", "T", "{", "N}×{H", ",", "W", ",", "C}", "(", "Y", ")", "n", "and", "µzn", "are", "C", "-", "sized", "vectors", "since", "H", "=", "W", "=", "1", "in", "this", "case", ".", "(", "B", ")", "Spatial", "Clustering", ":", "The", "representation", "of", "one", "sample", "can", "be", "regarded", "as", "a", "C", "-", "channel", "“", "image", ".", "”", "Each", "spatial", "location", "within", "that", "“", "image", "”", "can", "be", "thought", "of", "as", "a", "“", "pixel", ",", "”", "and", "is", "a", "vector", "of", "size", "C", "(", "shown", "as", "a", "colored", "bar", "in", "Fig", ".", "1", ")", ".", "For", "a", "ConvNet", ",", "every", "“", "pixel", "”", "has", "a", "corresponding", "receptive", "field", "covering", "a", "local", "region", "in", "the", "input", "image", ".", "Therefore", ",", "by", "clustering", "“", "pixels", "”", "of", "all", "images", "during", "learning", ",", "we", "expect", "to", "model", "local", "parts", "shared", "by", "multiple", "objects", "or", "scenes", ".", "To", "achieve", "this", ",", "we", "adopt", "the", "unfolding", "operator", "T", "{", "N", ",", "H", ",", "W}×{C}", "(", "Y", ")", "and", "use", "Rspatial", "(", "Y", ",", "µ", ")", "=", "1", "2NCHW", "NHW∑", "i=1", "‖T", "{", "N", ",", "H", ",", "W}×{C}", "(", "Y", ")", "i", "−", "µzi‖2", ".", "(", "2", ")", "Note", "that", "although", "we", "use", "the", "analogy", "of", "a", "“", "pixel", ",", "”", "when", "using", "text", "data", "a", "“", "pixel", "”", "may", "corresponds", "to", "words", ".", "For", "spatial", "clustering", "the", "dimension", "of", "the", "matrix", "µ", "is", "K", "×", "C.", "(", "C", ")", "Channel", "Co", "-", "Clustering", ":", "This", "regularizer", "groups", "the", "channels", "of", "different", "samples", "directly", ",", "thus", "co", "-", "clustering", "samples", "and", "filters", ".", "We", "expect", "this", "type", "of", "regularization", "to", "model", "re", "-", "occurring", "Algorithm", "1", ":", "Learning", "Parsimonious", "Representations", "1", ":", "Initialization", ":", "Maximum", "training", "iteration", "R", ",", "batch", "size", "B", ",", "smooth", "weight", "α", ",", "set", "of", "clustering", "layers", "S", "and", "set", "of", "cluster", "centers", "{", "µ0k|k", "∈", "[", "K", "]", "}", ",", "update", "period", "M", "2", ":", "For", "iteration", "t", "=", "1", ",", "2", ",", "...", ",", "R", ":", "3", ":", "For", "layer", "l", "=", "1", ",", "2", ",", "...", ",", "L", ":", "4", ":", "Compute", "the", "output", "representation", "of", "layer", "l", "as", "x.", "5", ":", "If", "l", "∈", "S", ":", "6", ":", "Assigning", "cluster", "zn", "=", "argmin", "k", "‖Xn", "−", "µt−1k", "‖2", ",", "∀n", "∈", "[", "B", "]", ".", "7", ":", "Compute", "cluster", "center", "µ̂k", "=", "1|Nk|", "∑", "n∈Nk", "Xn", ",", "where", "Nk", "=", "[", "B", "]", "⋂", "{", "n|zn", "=", "k}.", "8", ":", "Smooth", "cluster", "center", "µtk", "=", "αµ̂k", "+", "(", "1−", "α", ")", "µ", "t−1", "k", "9", ":", "End", "10", ":", "End", "11", ":", "Compute", "the", "gradients", "with", "cluster", "centers", "µtk", "fixed", ".", "12", ":", "Update", "weights", ".", "13", ":", "Update", "drifted", "cluster", "centers", "using", "Kmeans", "++", "every", "M", "iterations", ".", "14", ":", "End", "patterns", "shared", "not", "only", "among", "different", "samples", "but", "also", "within", "each", "sample", ".", "Relying", "on", "the", "unfolding", "operator", "T", "{", "N", ",", "C}×{H", ",", "W}", "(", "Y", ")", ",", "we", "formulate", "this", "type", "of", "clustering", "objective", "as", "Rchannel", "(", "Y", ",", "µ", ")", "=", "1", "2NCHW", "NC∑", "i=1", "‖T", "{", "N", ",", "C}×{H", ",", "W}", "(", "Y", ")", "i", "−", "µzi‖2", ".", "(", "3", ")", "Note", "that", "the", "dimension", "of", "the", "matrix", "µ", "is", "K", "×HW", "in", "this", "case", ".", "3.2", "Efficient", "Online", "Update", "We", "now", "derive", "an", "efficient", "online", "update", "to", "jointly", "learn", "the", "weights", "while", "clustering", "the", "representations", "of", "the", "neural", "network", ".", "In", "particular", ",", "we", "illustrate", "the", "sample", "clustering", "case", "while", "noting", "that", "the", "other", "types", "can", "be", "derived", "easily", "by", "applying", "the", "corresponding", "unfolding", "operator", ".", "For", "ease", "of", "notation", ",", "we", "denote", "the", "unfolded", "matrix", "T", "{", "N}×{H", ",", "W", ",", "C}", "(", "Y", ")", "as", "X.", "The", "gradient", "of", "the", "clustering", "regularization", "layer", "w.r.t", ".", "its", "input", "representation", "X", "can", "be", "expressed", "as", ",", "∂R", "∂Xn", "=", "1", "NCHW", "Xn", "−", "µzn", "−", "1Qzn", "∑", "zp", "=", "zn", ",", "∀p∈", "[", "N", "]", "(", "Xn", "−", "µzp", ")", ",", "(", "4", ")", "where", "Qzn", "is", "the", "number", "of", "samples", "which", "belong", "to", "the", "zn", "-", "th", "cluster", ".", "This", "gradient", "is", "then", "backpropagated", "through", "the", "network", "to", "obtain", "the", "gradient", "w.r.t", ".", "the", "parameters", "of", "the", "network", ".", "The", "time", "and", "space", "complexity", "of", "the", "gradient", "computation", "of", "one", "regularization", "layer", "are", "max", "(", "O", "(", "KCHW", ")", ",", "O", "(", "NCHW", ")", ")", "and", "O", "(", "NCHW", ")", "respectively", ".", "Note", "that", "we", "can", "cache", "the", "centered", "data", "Xn", "−", "µzn", "in", "the", "forward", "pass", "to", "speed", "up", "the", "gradient", "computation", ".", "The", "overall", "learning", "algorithm", "of", "our", "framework", "is", "summarized", "in", "Alg", ".", "1", ".", "In", "the", "forward", "pass", ",", "we", "first", "compute", "the", "representation", "of", "the", "n", "-", "th", "sample", "as", "Xn", "for", "each", "layer", ".", "We", "then", "infer", "the", "latent", "cluster", "label", "zn", "for", "each", "sample", "based", "on", "the", "distance", "to", "the", "cluster", "centers", "µt−1k", "from", "the", "last", "time", "step", "t−", "1", ",", "and", "assign", "the", "sample", "to", "the", "cluster", "center", "which", "has", "the", "smallest", "distance", ".", "Once", "all", "the", "cluster", "assignments", "are", "computed", ",", "we", "estimate", "the", "cluster", "centers", "µ̂k", "based", "on", "the", "new", "labels", "of", "the", "current", "batch", ".", "We", "then", "combine", "the", "estimate", "based", "on", "the", "current", "batch", "with", "the", "former", "cluster", "center", ".", "This", "is", "done", "via", "an", "online", "update", ".", "We", "found", "an", "online", "update", "together", "with", "the", "random", "restart", "strategy", "to", "work", "well", "in", "practice", ",", "as", "the", "learning", "of", "the", "neural", "network", "proceeds", "one", "mini", "-", "batch", "at", "a", "time", ",", "and", "as", "it", "is", "too", "expensive", "to", "recompute", "the", "cluster", "assignment", "for", "all", "data", "samples", "in", "every", "iteration", ".", "Since", "we", "trust", "our", "current", "cluster", "center", "estimate", "more", "than", "older", "ones", ",", "we", "smooth", "the", "estimation", "by", "using", "an", "exponential", "moving", "average", ".", "The", "cluster", "center", "estimate", "at", "iteration", "t", "is", "obtained", "via", "µtk", "=", "αµ̂k", "+", "(", "1−", "α", ")", "µ", "t−1", "k", ",", "where", "α", "is", "a", "smoothing", "weight", ".", "However", ",", "as", "the", "representation", "learned", "by", "the", "neural", "network", "may", "go", "through", "drastic", "changes", ",", "especially", "in", "the", "beginning", "of", "training", ",", "some", "of", "the", "cluster", "centers", "may", "quickly", "be", "less", "favored", "and", "the", "number", "of", "incoming", "samples", "assigned", "to", "it", "will", "be", "largely", "reduced", ".", "To", "overcome", "this", "issue", ",", "we", "exploit", "the", "Kmeans", "++", "[", "3", "]", "procedure", "to", "re", "-", "sample", "the", "cluster", "center", "from", "the", "current", "mini", "-", "batch", ".", "Specifically", ",", "denoting", "the", "the", "distance", "between", "sample", "Xn", "and", "its", "nearest", "cluster", "center", "as", "dn", ",", "the", "probability", "of", "taking", "Xn", "as", "the", "new", "cluster", "center", "is", "d2n", "/", "∑", "i", "d", "2", "i", ".", "After", "sampling", ",", "we", "replace", "the", "old", "cluster", "center", "with", "the", "new", "one", "and", "continue", "the", "learning", "process", ".", "In", "practice", ",", "at", "the", "end", "of", "every", "epoch", ",", "we", "apply", "the", "kmeans", "++", "update", "to", "cluster", "centers", "for", "which", "the", "number", "of", "assigned", "samples", "is", "small", ".", "See", "Alg", ".", "1", "for", "an", "outline", "of", "the", "steps", ".", "The", "overall", "procedure", "stabilizes", "the", "optimization", "and", "also", "increases", "the", "diversity", "of", "the", "cluster", "centers", ".", "In", "the", "backward", "pass", ",", "we", "fix", "the", "latest", "estimation", "of", "the", "cluster", "centers", "µtk", "and", "compute", "the", "gradient", "of", "loss", "function", "and", "the", "gradient", "of", "the", "clustering", "objective", "based", "on", "Eq", ".", "(", "4", ")", ".", "Then", "we", "back", "-", "propagate", "all", "the", "gradients", "and", "update", "the", "weights", ".", "4", "Experiments", "In", "this", "section", ",", "we", "conduct", "experiments", "on", "unsupervised", ",", "supervised", "and", "zero", "-", "shot", "learning", "on", "several", "datasets", ".", "Our", "implementation", "based", "on", "TensorFlow", "[", "9", "]", "is", "publicly", "available.1", "For", "initializing", "the", "cluster", "centers", "before", "training", ",", "we", "randomly", "choose", "them", "from", "the", "representations", "obtained", "with", "the", "initial", "network", ".", "4.1", "Autoencoder", "on", "MNIST", "We", "first", "test", "our", "method", "on", "the", "unsupervised", "learning", "task", "of", "training", "an", "autoencoder", ".", "Our", "architecture", "is", "identical", "to", "[", "17", "]", ".", "For", "ease", "of", "training", "we", "did", "not", "tie", "the", "weights", "between", "the", "encoder", "and", "the", "decoder", ".", "We", "use", "the", "squared", "`", "2", "reconstruction", "error", "as", "the", "loss", "function", "and", "SGD", "with", "momentum", ".", "The", "standard", "training", "-", "test", "-", "split", "is", "used", ".", "We", "compute", "the", "mean", "reconstruction", "error", "over", "all", "test", "images", "and", "repeat", "the", "experiments", "4", "times", "with", "different", "random", "initializations", ".", "We", "compare", "the", "baseline", "model", ",", "i.e.", ",", "a", "plain", "autoencoder", ",", "with", "one", "that", "employs", "our", "sample", "-", "clustering", "regularization", "on", "all", "layers", "except", "the", "top", "fully", "connected", "layer", ".", "Sample", "clustering", "was", "chosen", "since", "this", "autoencoder", "only", "contains", "fully", "connected", "layers", ".", "The", "number", "of", "clusters", "and", "the", "regularization", "weight", "λ", "of", "all", "layers", "are", "set", "to", "100", "and", "1.0e−2", "respectively", ".", "For", "both", "models", "the", "same", "learning", "rate", "and", "momentum", "are", "used", ".", "Our", "exact", "parameter", "choices", "are", "detailed", "in", "the", "Appendix", ".", "As", "shown", "in", "Table", "1", ",", "our", "regularization", "facilitates", "generalization", "as", "it", "suffers", "less", "from", "overfitting", ".", "Specifically", ",", "applying", "our", "regularization", "results", "in", "lower", "test", "set", "error", "despite", "slightly", "higher", "training", "error", ".", "More", "importantly", ",", "the", "standard", "deviation", "of", "the", "error", "is", "one", "order", "of", "magnitude", "smaller", "for", "both", "training", "and", "testing", "when", "applying", "our", "regularization", ".", "This", "indicates", "that", "our", "sample", "-", "clustering", "regularization", "stabilizes", "the", "model", ".", "1https:", "//", "github.com", "/", "lrjconan", "/", "deep_parsimonious", "4.2", "CIFAR10", "and", "CIFAR100", "In", "this", "section", ",", "we", "explore", "the", "CIFAR10", "and", "CIFAR100", "datasets", "[", "20", "]", ".", "CIFAR10", "consists", "of", "60", ",", "000", "32×", "32", "images", "assigned", "to", "10", "categories", ",", "while", "CIFAR100", "differentiates", "between", "100", "classes", ".", "We", "use", "the", "standard", "split", "on", "both", "datasets", ".", "The", "quick", "CIFAR10", "architecture", "of", "Caffe", "[", "19", "]", "is", "used", "for", "benchmarking", "both", "datasets", ".", "It", "consists", "of", "3", "convolutional", "layers", "and", "1", "fully", "connected", "layer", "followed", "by", "a", "softmax", "layer", ".", "The", "detailed", "parameters", "are", "publicly", "available", "on", "the", "Caffe", "[", "19", "]", "website", ".", "We", "report", "mean", "accuracy", "averaged", "over", "4", "trials", ".", "For", "fully", "connected", "layers", "we", "use", "the", "sample", "-", "clustering", "objective", ".", "For", "convolutional", "layers", ",", "we", "provide", "the", "results", "of", "all", "three", "clustering", "objectives", ",", "which", "we", "refer", "to", "as", "‘", "sample", "-", "clustering", ",", "’", "‘", "spatial", "-", "clustering", ",", "’", "and", "‘", "channel", "-", "co", "-", "clustering", "’", "respectively", ".", "We", "set", "all", "hyper", "-", "parameters", "based", "on", "cross", "-", "validation", ".", "Specifically", ",", "the", "number", "of", "cluster", "centers", "are", "set", "to", "100", "for", "all", "layers", "for", "both", "CIFAR10", "and", "CIFAR100", ".", "λ", "is", "set", "to", "1.0e−3", "and", "1.0e−2", "for", "the", "first", "two", "convolutional", "and", "the", "remaining", "layers", "respectively", "in", "CIFAR10", ";", "for", "CIFAR100", ",", "λ", "is", "set", "to", "10", "and", "1", "for", "the", "first", "convolutional", "layer", "and", "the", "remaining", "layers", "respectively", ".", "The", "smoothness", "parameter", "α", "is", "set", "to", "0.9", "and", "0.95", "for", "CIFAR10", "and", "CIFAR100", "respectively", ".", "Generalization", ":", "In", "Table", "2", "we", "compare", "our", "framework", "to", "some", "recent", "regularizers", ",", "like", "DeCov", "[", "6", "]", ",", "Dropout", "[", "26", "]", "and", "the", "baseline", "results", "obtained", "using", "Caffe", ".", "We", "again", "observe", "that", "all", "of", "our", "methods", "achieve", "better", "generalization", "performance", ".", "Visualization", ":", "To", "demonstrate", "the", "interpretability", "of", "our", "learned", "network", ",", "we", "visualize", "sampleclustering", "and", "spatial", "-", "clustering", "in", "Fig", ".", "2", ",", "showing", "the", "top", "-", "10", "ranked", "images", "and", "parts", "per", "cluster", ".", "In", "the", "case", "of", "sample", "-", "clustering", ",", "for", "each", "cluster", "we", "rank", "all", "its", "assigned", "images", "based", "on", "the", "distance", "to", "the", "cluster", "center", ".", "We", "chose", "to", "show", "2", "clusters", "from", "the", "4th", "fully", "connected", "layer", ".", "In", "the", "case", "of", "spatial", "-", "clustering", ",", "we", "rank", "all", "“", "pixels", "”", "belonging", "to", "one", "cluster", "based", "on", "the", "distance", "to", "the", "cluster", "center", ".", "Note", "that", "we", "have", "one", "part", "(", "i.e.", ",", "one", "receptive", "field", "region", "in", "the", "input", "image", ")", "for", "each", "“", "pixel", ".", "”", "We", "chose", "to", "show", "2", "clusters", "from", "the", "2nd", "convolutional", "layer", ".", "The", "receptive", "field", "of", "the", "2nd", "convolutional", "layer", "is", "of", "size", "18", "×", "18", "in", "the", "original", "32", "×", "32", "sized", "image", ".", "We", "observe", "that", "clusterings", "of", "the", "fully", "connected", "layer", "representations", "encode", "high", "-", "level", "semantic", "meaning", ".", "In", "contrast", ",", "clusterings", "of", "the", "convolutional", "layer", "representations", "encode", "attributes", "like", "shape", ".", "Note", "that", "some", "parts", "are", "uninformative", "which", "may", "be", "due", "to", "the", "fact", "that", "images", "in", "CIFAR10", "are", "very", "small", ".", "Additional", "clusters", "and", "visualizations", "on", "CIFAR100", "are", "shown", "in", "the", "Appendix", ".", "Quantitative", "Evaluation", "of", "Parsimonious", "Representation", ":", "We", "quantitatively", "evaluate", "our", "learned", "parsimonious", "representation", "on", "CIFAR100", ".", "Since", "only", "the", "image", "category", "is", "provided", "as", "ground", "truth", ",", "we", "investigate", "sample", "clustering", "using", "the", "4th", "fully", "connected", "layer", "where", "representations", "capture", "semantic", "meaning", ".", "In", "particular", ",", "we", "apply", "K", "-", "means", "clustering", "to", "the", "learned", "representation", "extracted", "from", "the", "model", "with", "and", "without", "sample", "clustering", "respectively", ".", "For", "both", "cases", ",", "we", "set", "the", "number", "of", "clusters", "to", "be", "100", "and", "control", "the", "random", "seed", "to", "be", "the", "same", ".", "The", "most", "frequent", "class", "label", "within", "one", "cluster", "is", "assigned", "to", "all", "of", "its", "members", ".", "Then", "we", "compute", "the", "normalized", "mutual", "information", "(", "NMI", ")", "[", "25", "]", "to", "measure", "the", "clustering", "accuracy", ".", "The", "average", "results", "over", "10", "runs", "are", "shown", "in", "Table", "3", ".", "Our", "representations", "achieve", "significantly", "better", "clustering", "quality", "compared", "to", "the", "baseline", "which", "suggests", "that", "they", "are", "distributed", "in", "a", "more", "compact", "way", "in", "the", "feature", "space", ".", "4.3", "CUB", "-", "200", "-", "2011", "Next", "we", "test", "our", "framework", "on", "the", "Caltech", "-", "UCSD", "birds", "dataset", "[", "34", "]", "which", "contains", "11", ",", "788", "images", "of", "200", "different", "categories", ".", "We", "follow", "the", "dataset", "split", "provided", "by", "[", "34", "]", "and", "the", "common", "practice", "of", "cropping", "the", "image", "using", "the", "ground", "-", "truth", "bounding", "box", "annotation", "of", "the", "birds", "[", "8", ",", "36", "]", ".", "We", "use", "Alex", "-", "Net", "[", "21", "]", "pretrained", "on", "ImageNet", "as", "the", "base", "model", "and", "adapt", "the", "last", "layer", "to", "fit", "classification", "of", "200", "categories", ".", "We", "resize", "the", "image", "to", "227×", "227", "to", "fit", "the", "input", "size", ".", "We", "add", "clusterings", "to", "all", "layers", "except", "the", "softmax", "-", "layer", ".", "Based", "on", "cross", "-", "validation", ",", "the", "number", "of", "clusters", "are", "set", "to", "200", "for", "all", "layers", ".", "For", "convolutional", "layers", ",", "we", "set", "λ", "to", "1.0e−5", "for", "the", "first", "(", "bottom", ")", "2", "and", "use", "1.0e−4", "for", "the", "remaining", "ones", ".", "For", "fully", "connected", "layers", ",", "we", "set", "λ", "to", "1.0e−3", "and", "α", "is", "equal", "to", "0.5", ".", "We", "apply", "Kmeans", "++", "to", "replace", "cluster", "centers", "with", "less", "than", "10", "assigned", "samples", "at", "the", "end", "of", "every", "epoch", ".", "Generalization", ":", "We", "investigate", "the", "impact", "of", "our", "parsimonious", "representation", "on", "generalization", "performance", ".", "We", "compare", "with", "the", "DeCAF", "result", "reported", "in", "[", "8", "]", ",", "which", "used", "the", "same", "network", "to", "extract", "a", "representation", "and", "applied", "logistic", "regression", "on", "top", "for", "fine", "-", "tuning", ".", "We", "also", "fine", "-", "tune", "Alex", "-", "Net", "which", "uses", "weight", "-", "decay", "and", "Dropout", ",", "and", "report", "the", "best", "result", "we", "achieved", "in", "Table", "4", ".", "We", "observe", "that", "for", "the", "Alex", "-", "Net", "architecture", "our", "clustering", "improves", "the", "generalization", "compare", "to", "direct", "fine", "-", "tuning", "and", "the", "DeCAF", "result", ".", "Note", "that", "Alex", "-", "Net", "pretrained", "on", "ImageNet", "easily", "overfits", "on", "this", "dataset", "as", "all", "training", "accuracies", "reach", "100", "percent", ".", "Visualization", ":", "To", "visualize", "the", "sample", "-", "clustering", "and", "spatial", "-", "clustering", "we", "follow", "the", "setting", "employed", "when", "evaluating", "on", "the", "CIFAR", "dataset", ".", "For", "the", "selected", "cluster", "center", "we", "show", "the", "10", "closest", "images", "in", "Fig", ".", "3", ".", "For", "sample", "clustering", ",", "2", "clusters", "from", "the", "3rd", "convolutional", "layer", "and", "the", "7th", "fully", "connected", "layer", "are", "chosen", "for", "visualization", ".", "For", "spatial", "clustering", ",", "2", "clusters", "from", "the", "2nd", "and", "3rd", "convolutional", "layers", "are", "chosen", "for", "visualization", ".", "More", "clusters", "are", "shown", "in", "the", "Appendix", ".", "The", "receptive", "fields", "of", "pixels", "from", "the", "2nd", "and", "3rd", "convolutional", "layers", "are", "of", "sizes", "59", "×", "59", "and", "123", "×", "123", "in", "the", "resized", "227", "×", "227", "image", ".", "We", "observe", "that", "cluster", "centers", "of", "sample", "clustering", "applied", "to", "layers", "lower", "in", "the", "network", "capture", "pose", "and", "shape", "information", ",", "while", "cluster", "centers", "from", "top", "layers", "model", "the", "fine", "-", "grained", "categories", "of", "birds", ".", "For", "spatial", "clustering", ",", "cluster", "centers", "from", "different", "layers", "capture", "parts", "of", "birds", "in", "different", "scales", ",", "like", "the", "beak", ",", "chest", ",", "etc", ".", "4.4", "Zero", "-", "Shot", "Learning", "We", "also", "investigate", "a", "zero", "-", "shot", "setting", "on", "the", "CUB", "dataset", "to", "see", "whether", "our", "parsimonious", "representation", "is", "applicable", "to", "unseen", "categories", ".", "We", "follow", "the", "setting", "in", "[", "1", ",", "2", "]", "and", "use", "the", "same", "split", "where", "100", ",", "50", "and", "50", "classes", "are", "used", "as", "training", ",", "validation", "and", "testing", "(", "unseen", "classes", ")", ".", "We", "use", "a", "pre", "-", "trained", "Alex", "-", "Net", "as", "the", "baseline", "model", "and", "extract", "4096", "-", "dimension", "representations", "from", "the", "7th", "fully", "connected", "(", "fc", ")", "layer", ".", "We", "compare", "sample", "-", "clustering", "against", "other", "recent", "methods", "which", "also", "report", "results", "of", "using", "7th", "fc", "feature", "of", "Alex", "-", "Net", ".", "Given", "these", "features", ",", "we", "learn", "the", "output", "embedding", "W", "via", "the", "same", "unregularized", "structured", "SVM", "as", "in", "[", "1", ",", "2", "]", ":", "min", "W", "1", "N", "N∑", "n=1", "max", "y∈Y", "{", "0", ",", "∆", "(", "yn", ",", "y", ")", "+", "x", ">", "nW", "[", "φ", "(", "y", ")", "−", "φ", "(", "yn", ")", "]", ")", "}", ",", "(", "5", ")", "where", "xn", "and", "yn", "are", "the", "feature", "and", "class", "label", "of", "the", "n", "-", "th", "sample", "and", "∆", "is", "the", "0", "-", "1", "loss", "function", "..", "φ", "is", "the", "class", "-", "attribute", "matrix", "provided", "by", "the", "CUB", "dataset", ",", "where", "each", "entry", "is", "a", "real", "-", "valued", "score", "indicating", "how", "likely", "a", "human", "thinks", "one", "attribute", "is", "present", "in", "a", "given", "class", ".", "We", "tune", "the", "hyperparameters", "on", "the", "validation", "set", "and", "report", "results", "in", "terms", "of", "top", "-", "1", "accuracy", "averaged", "over", "the", "unseen", "classes", ".", "As", "shown", "in", "Table", "5", "our", "approach", "significantly", "outperforms", "other", "approaches", ".", "5", "Conclusions", "We", "have", "proposed", "a", "novel", "clustering", "based", "regularization", "which", "encourages", "parsimonious", "representations", ",", "while", "being", "easy", "to", "optimize", ".", "We", "have", "demonstrated", "the", "effectiveness", "of", "our", "approach", "on", "a", "variety", "of", "tasks", "including", "unsupervised", "learning", ",", "classification", ",", "fine", "grained", "categorization", ",", "and", "zero", "-", "shot", "learning", ".", "In", "the", "future", "we", "plan", "to", "apply", "our", "approach", "to", "even", "larger", "networks", ",", "e.g.", ",", "residual", "nets", ",", "and", "develop", "a", "probabilistic", "formulation", "which", "provides", "a", "soft", "clustering", ".", "Acknowledgments", "This", "work", "was", "partially", "supported", "by", "ONR", "-", "N00014", "-", "14", "-", "1", "-", "0232", ",", "NVIDIA", "and", "the", "Intelligence", "Advanced", "Research", "Projects", "Activity", "(", "IARPA", ")", "via", "Department", "of", "Interior", "/", "Interior", "Business", "Center", "(", "DoI", "/", "IBC", ")", "contract", "number", "D16PC00003", ".", "The", "U.S.", "Government", "is", "authorized", "to", "reproduce", "and", "distribute", "reprints", "for", "Governmental", "purposes", "notwithstanding", "any", "copyright", "annotation", "thereon", ".", "Disclaimer", ":", "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", "IARPA", ",", "DoI", "/", "IBC", ",", "or", "the", "U.S.", "Government", "."], "entity": [[1, 4, ["Deep", "Parsimonious", "Representations"], "Method"], [11, 12, ["generalization"], "Task"], [13, 15, ["deep", "networks"], "Method"], [20, 22, ["learned", "representations"], "Method"], [35, 37, ["parsimonious", "representations"], "Method"], [39, 44, ["k", "-", "means", "style", "objective"], "Method"], [55, 56, ["clustering"], "Method"], [59, 61, ["sample", "clustering"], "Method"], [62, 64, ["spatial", "clustering"], "Method"], [68, 71, ["co", "-", "clustering"], "Method"], [83, 85, ["unsupervised", "learning"], "Task"], [86, 87, ["classification"], "Task"], [88, 91, ["fine", "grained", "categorization"], "Task"], [93, 97, ["zero", "-", "shot", "learning"], "Task"], [104, 107, ["deep", "neural", "networks"], "Method"], [120, 121, ["classification"], "Task"], [125, 127, ["semantic", "segmentation"], "Task"], [131, 133, ["machine", "translation"], "Task"], [137, 139, ["speech", "recognition"], "Task"], [154, 156, ["computer", "vision"], "Task"], [157, 160, ["natural", "language", "processing"], "Task"], [161, 162, ["robotics"], "Task"], [182, 184, ["neural", "networks"], "Method"], [212, 213, ["regularization"], "Task"], [222, 223, ["generalization"], "Task"], [225, 227, ["learned", "classifier"], "Method"], [238, 240, ["neural", "nets"], "Method"], [244, 246, ["batch", "normalization"], "Method"], [266, 267, ["dropout"], "Method"], [281, 284, ["co", "-", "adaptation"], "Task"], [285, 291, ["Learning", "of", "de", "-", "correlated", "activations"], "Task"], [316, 317, ["regularization"], "Method"], [320, 322, ["network", "representations"], "Method"], [339, 340, ["generalization"], "Task"], [343, 344, ["clustering"], "Method"], [348, 350, ["learned", "representations"], "Method"], [357, 362, ["k", "-", "means", "style", "objective"], "Method"], [373, 374, ["clusterings"], "Method"], [376, 378, ["sample", "clustering"], "Method"], [379, 381, ["spatial", "clustering"], "Method"], [383, 386, ["co", "-", "clustering"], "Method"], [390, 391, ["generalization"], "Task"], [400, 401, ["autoencoders"], "Method"], [410, 411, ["classification"], "Task"], [427, 431, ["zero", "-", "shot", "learning"], "Method"], [433, 439, ["CUB", "-", "200", "-", "2011", "dataset"], "Material"], [467, 473, ["CUB", "-", "200", "-", "2011", "dataset"], "Material"], [475, 477, ["network", "representation"], "Method"], [478, 481, ["meaningful", "part", "representations"], "Method"], [496, 499, ["neural", "network", "regularization"], "Method"], [520, 522, ["regularization", "methods"], "Method"], [524, 526, ["intermediate", "representations"], "Method"], [530, 534, ["Neural", "Information", "Processing", "Systems"], "Task"], [535, 537, ["NIPS", "2016"], "Task"], [545, 546, ["Dropout"], "Method"], [550, 553, ["Drop", "-", "Connect"], "Method"], [557, 558, ["Maxout"], "Method"], [562, 563, ["DeCov"], "Method"], [593, 594, ["regularization"], "Task"], [598, 600, ["parsimonious", "representations"], "Method"], [607, 608, ["clustering"], "Method"], [613, 615, ["neural", "network"], "Method"], [623, 625, ["Compression", "rates"], "Metric"], [640, 641, ["accuracy"], "Metric"], [658, 659, ["compression"], "Task"], [660, 663, ["biased", "weight", "decay"], "Method"], [691, 692, ["clustering"], "Method"], [693, 695, ["representation", "learning"], "Method"], [713, 714, ["clustering"], "Method"], [738, 741, ["deep", "belief", "networks"], "Method"], [742, 743, ["DBN"], "Method"], [758, 759, ["DBN"], "Method"], [767, 769, ["intermediate", "representation"], "Method"], [785, 787, ["data", "representation"], "Method"], [794, 799, ["embedding", "of", "the", "deep", "network"], "Method"], [815, 817, ["unsupervised", "method"], "Method"], [820, 824, ["stacked", "auto", "-", "encoder"], "Method"], [844, 847, ["k", "-", "means"], "Method"], [848, 850, ["spectral", "clustering"], "Method"], [862, 867, ["non", "-", "negative", "matrix", "factorization"], "Method"], [884, 890, ["deep", "non", "-", "negative", "matrix", "factorization"], "Method"], [899, 901, ["clustering", "objective"], "Method"], [915, 916, ["clustering"], "Method"], [948, 949, ["representation"], "Task"], [954, 956, ["clustering", "regularization"], "Method"], [963, 965, ["sparse", "coding"], "Method"], [978, 980, ["sparse", "codes"], "Method"], [992, 994, ["clustering", "objective"], "Method"], [1008, 1010, ["clustering", "objectives"], "Method"], [1012, 1014, ["intermediate", "representations"], "Method"], [1018, 1020, ["feature", "regularization"], "Method"], [1022, 1024, ["unrolled", "optimization"], "Method"], [1059, 1062, ["convolutional", "neural", "networks"], "Method"], [1066, 1069, ["sparse", "coding", "optimization"], "Method"], [1080, 1082, ["agglomerative", "clustering"], "Method"], [1084, 1085, ["regularizer"], "Method"], [1092, 1094, ["recurrent", "network"], "Method"], [1100, 1106, ["k", "-", "means", "like", "clustering", "objective"], "Method"], [1109, 1110, ["optimization"], "Task"], [1116, 1118, ["recurrent", "procedure"], "Method"], [1124, 1128, ["unsupervised", "and", "supervised", "learning"], "Method"], [1131, 1134, ["Deep", "Parsimonious", "Representations"], "Method"], [1150, 1152, ["neural", "network"], "Method"], [1164, 1166, ["neural", "network"], "Method"], [1177, 1179, ["representation", "tensor"], "Method"], [1185, 1186, ["clusterings"], "Method"], [1197, 1199, ["online", "update"], "Method"], [1203, 1204, ["clustering"], "Method"], [1209, 1211, ["neural", "network"], "Method"], [1213, 1216, ["Clustering", "of", "Representations"], "Method"], [1267, 1268, ["tensor"], "Method"], [1270, 1272, ["multilinear", "map"], "Method"], [1280, 1282, ["tensor", "terminology"], "Task"], [1319, 1324, ["n", "-", "mode", "matrix", "unfolding"], "Method"], [1326, 1327, ["tensor"], "Method"], [1361, 1366, ["n", "-", "mode", "matrix", "unfolding"], "Method"], [1462, 1464, ["neural", "network"], "Method"], [1467, 1472, ["4", "-", "D", "tensor", "Y"], "Method"], [1556, 1558, ["neural", "network"], "Method"], [1566, 1568, ["clustering", "regularization"], "Method"], [1570, 1572, ["single", "layer"], "Method"], [1626, 1627, ["clustering"], "Method"], [1648, 1650, ["Sample", "Clustering"], "Method"], [1654, 1655, ["clustering"], "Method"], [1688, 1689, ["ConvNet"], "Method"], [1690, 1693, ["bottom", "layer", "representations"], "Method"], [1904, 1906, ["Spatial", "Clustering"], "Method"], [1963, 1964, ["ConvNet"], "Method"], [1994, 1995, ["learning"], "Task"], [2016, 2018, ["unfolding", "operator"], "Method"], [2030, 2033, ["Rspatial", "(", "Y"], "Method"], [2085, 2087, ["spatial", "clustering"], "Method"], [2100, 2104, ["Channel", "Co", "-", "Clustering"], "Method"], [2106, 2107, ["regularizer"], "Method"], [2128, 2129, ["regularization"], "Method"], [2137, 2140, ["Learning", "Parsimonious", "Representations"], "Task"], [2311, 2313, ["Kmeans", "++"], "Method"], [2336, 2338, ["unfolding", "operator"], "Method"], [2354, 2356, ["clustering", "objective"], "Method"], [2357, 2360, ["Rchannel", "(", "Y"], "Method"], [2402, 2404, ["Online", "Update"], "Task"], [2409, 2411, ["online", "update"], "Method"], [2419, 2424, ["representations", "of", "the", "neural", "network"], "Method"], [2448, 2450, ["unfolding", "operator"], "Method"], [2556, 2560, ["time", "and", "space", "complexity"], "Metric"], [2562, 2564, ["gradient", "computation"], "Method"], [2566, 2568, ["regularization", "layer"], "Method"], [2607, 2609, ["gradient", "computation"], "Task"], [2612, 2614, ["learning", "algorithm"], "Method"], [2733, 2735, ["online", "update"], "Method"], [2739, 2741, ["online", "update"], "Method"], [2744, 2747, ["random", "restart", "strategy"], "Method"], [2758, 2760, ["neural", "network"], "Method"], [2808, 2811, ["exponential", "moving", "average"], "Method"], [2813, 2816, ["cluster", "center", "estimate"], "Method"], [2849, 2851, ["neural", "network"], "Method"], [2896, 2898, ["Kmeans", "++"], "Method"], [2968, 2970, ["learning", "process"], "Method"], [2984, 2987, ["kmeans", "++", "update"], "Method"], [3016, 3017, ["optimization"], "Task"], [3054, 3056, ["clustering", "objective"], "Method"], [3087, 3095, ["unsupervised", ",", "supervised", "and", "zero", "-", "shot", "learning"], "Task"], [3103, 3104, ["TensorFlow"], "Method"], [3128, 3130, ["initial", "network"], "Method"], [3132, 3135, ["Autoencoder", "on", "MNIST"], "Method"], [3142, 3145, ["unsupervised", "learning", "task"], "Task"], [3148, 3149, ["autoencoder"], "Method"], [3171, 3172, ["encoder"], "Method"], [3179, 3184, ["squared", "`", "2", "reconstruction", "error"], "Metric"], [3189, 3190, ["SGD"], "Method"], [3191, 3192, ["momentum"], "Method"], [3206, 3209, ["mean", "reconstruction", "error"], "Metric"], [3234, 3235, ["autoencoder"], "Method"], [3261, 3262, ["autoencoder"], "Method"], [3293, 3295, ["learning", "rate"], "Metric"], [3317, 3318, ["regularization"], "Method"], [3319, 3320, ["generalization"], "Task"], [3331, 3332, ["regularization"], "Method"], [3335, 3338, ["test", "set", "error"], "Metric"], [3341, 3343, ["training", "error"], "Metric"], [3352, 3353, ["error"], "Metric"], [3367, 3368, ["regularization"], "Method"], [3387, 3390, ["deep_parsimonious", "4.2", "CIFAR10"], "Method"], [3438, 3441, ["quick", "CIFAR10", "architecture"], "Method"], [3442, 3443, ["Caffe"], "Method"], [3457, 3459, ["convolutional", "layers"], "Method"], [3461, 3464, ["fully", "connected", "layer"], "Method"], [3467, 3469, ["softmax", "layer"], "Method"], [3487, 3488, ["accuracy"], "Metric"], [3506, 3508, ["convolutional", "layers"], "Method"], [3516, 3518, ["clustering", "objectives"], "Method"], [3531, 3534, ["spatial", "-", "clustering"], "Method"], [3538, 3543, ["channel", "-", "co", "-", "clustering"], "Method"], [3554, 3557, ["cross", "-", "validation"], "Metric"], [3611, 3613, ["convolutional", "layer"], "Method"], [3635, 3636, ["Generalization"], "Task"], [3647, 3648, ["regularizers"], "Method"], [3650, 3651, ["DeCov"], "Method"], [3655, 3656, ["Dropout"], "Method"], [3665, 3666, ["Caffe"], "Method"], [3677, 3679, ["generalization", "performance"], "Metric"], [3680, 3681, ["Visualization"], "Task"], [3688, 3690, ["learned", "network"], "Method"], [3750, 3753, ["fully", "connected", "layer"], "Method"], [3758, 3761, ["spatial", "-", "clustering"], "Method"], [3813, 3816, ["2nd", "convolutional", "layer"], "Method"], [3822, 3825, ["2nd", "convolutional", "layer"], "Method"], [3843, 3844, ["clusterings"], "Method"], [3846, 3850, ["fully", "connected", "layer", "representations"], "Method"], [3860, 3861, ["clusterings"], "Method"], [3863, 3866, ["convolutional", "layer", "representations"], "Method"], [3897, 3898, ["CIFAR100"], "Task"], [3904, 3906, ["Quantitative", "Evaluation"], "Task"], [3907, 3909, ["Parsimonious", "Representation"], "Method"], [3915, 3917, ["parsimonious", "representation"], "Method"], [3933, 3935, ["sample", "clustering"], "Method"], [3937, 3941, ["4th", "fully", "connected", "layer"], "Method"], [3952, 3956, ["K", "-", "means", "clustering"], "Method"], [3967, 3969, ["sample", "clustering"], "Method"], [4014, 4017, ["normalized", "mutual", "information"], "Metric"], [4018, 4019, ["NMI"], "Method"], [4137, 4140, ["Alex", "-", "Net"], "Method"], [4157, 4158, ["classification"], "Task"], [4177, 4178, ["clusterings"], "Method"], [4183, 4186, ["softmax", "-", "layer"], "Method"], [4189, 4192, ["cross", "-", "validation"], "Metric"], [4206, 4208, ["convolutional", "layers"], "Method"], [4248, 4250, ["Kmeans", "++"], "Method"], [4267, 4268, ["Generalization"], "Task"], [4275, 4277, ["parsimonious", "representation"], "Method"], [4278, 4279, ["generalization"], "Task"], [4285, 4286, ["DeCAF"], "Method"], [4304, 4306, ["logistic", "regression"], "Method"], [4309, 4312, ["fine", "-", "tuning"], "Task"], [4318, 4321, ["Alex", "-", "Net"], "Method"], [4323, 4326, ["weight", "-", "decay"], "Method"], [4327, 4328, ["Dropout"], "Method"], [4345, 4349, ["Alex", "-", "Net", "architecture"], "Method"], [4350, 4351, ["clustering"], "Method"], [4353, 4354, ["generalization"], "Task"], [4356, 4360, ["direct", "fine", "-", "tuning"], "Task"], [4367, 4370, ["Alex", "-", "Net"], "Method"], [4372, 4373, ["ImageNet"], "Method"], [4380, 4382, ["training", "accuracies"], "Metric"], [4386, 4387, ["Visualization"], "Task"], [4427, 4429, ["sample", "clustering"], "Method"], [4434, 4437, ["3rd", "convolutional", "layer"], "Method"], [4440, 4443, ["fully", "connected", "layer"], "Method"], [4446, 4447, ["visualization"], "Task"], [4449, 4451, ["spatial", "clustering"], "Method"], [4458, 4461, ["3rd", "convolutional", "layers"], "Method"], [4464, 4465, ["visualization"], "Task"], [4483, 4486, ["3rd", "convolutional", "layers"], "Method"], [4510, 4512, ["sample", "clustering"], "Method"], [4541, 4543, ["spatial", "clustering"], "Method"], [4566, 4570, ["Zero", "-", "Shot", "Learning"], "Method"], [4574, 4578, ["zero", "-", "shot", "setting"], "Task"], [4580, 4582, ["CUB", "dataset"], "Material"], [4586, 4588, ["parsimonious", "representation"], "Method"], [4635, 4638, ["Alex", "-", "Net"], "Method"], [4644, 4648, ["4096", "-", "dimension", "representations"], "Method"], [4651, 4657, ["fully", "connected", "(", "fc", ")", "layer"], "Method"], [4677, 4680, ["Alex", "-", "Net"], "Method"], [4694, 4697, ["unregularized", "structured", "SVM"], "Method"], [4779, 4781, ["CUB", "dataset"], "Material"], [4877, 4879, ["unsupervised", "learning"], "Task"], [4880, 4881, ["classification"], "Task"], [4882, 4885, ["fine", "grained", "categorization"], "Task"], [4887, 4891, ["zero", "-", "shot", "learning"], "Task"], [4914, 4916, ["probabilistic", "formulation"], "Method"], [4919, 4921, ["soft", "clustering"], "Method"], [30, 31, ["clustering"], "Method"], [425, 426, ["classification"], "Task"], [731, 732, ["clustering"], "Method"], [751, 752, ["clustering"], "Method"], [779, 780, ["clustering"], "Method"], [1075, 1076, ["clustering"], "Method"], [1142, 1143, ["clustering"], "Method"], [1588, 1589, ["clustering"], "Method"], [1986, 1987, ["clustering"], "Method"], [2118, 2119, ["clustering"], "Method"], [2159, 2160, ["clustering"], "Method"], [2417, 2418, ["clustering"], "Method"], [2432, 2433, ["clustering"], "Method"], [2477, 2478, ["clustering"], "Method"], [3243, 3244, ["clustering"], "Method"], [3256, 3257, ["clustering"], "Method"], [3375, 3376, ["clustering"], "Method"], [3502, 3503, ["clustering"], "Method"], [3527, 3528, ["clustering"], "Method"], [3697, 3698, ["clustering"], "Method"], [3721, 3722, ["clustering"], "Method"], [4026, 4027, ["clustering"], "Method"], [4027, 4028, ["accuracy"], "Metric"], [4046, 4047, ["clustering"], "Method"], [4393, 4394, ["clustering"], "Method"], [4397, 4398, ["clustering"], "Method"], [4662, 4663, ["clustering"], "Method"], [4823, 4824, ["accuracy"], "Metric"], [4849, 4850, ["clustering"], "Method"]]} {"doc_id": "11f411eee106efac9f677149e655252463811b09", "text": ["document", ":", "Mode", "Seeking", "Generative", "Adversarial", "Networks", "for", "Diverse", "Image", "Synthesis", "Most", "conditional", "generation", "tasks", "expect", "diverse", "outputs", "given", "a", "single", "conditional", "context", ".", "However", ",", "conditional", "generative", "adversarial", "networks", "(", "cGANs", ")", "often", "focus", "on", "the", "prior", "conditional", "information", "and", "ignore", "the", "input", "noise", "vectors", ",", "which", "contribute", "to", "the", "output", "variations", ".", "Recent", "attempts", "to", "resolve", "the", "mode", "collapse", "issue", "for", "cGANs", "are", "usually", "task", "-", "specific", "and", "computationally", "expensive", ".", "In", "this", "work", ",", "we", "propose", "a", "simple", "yet", "effective", "regularization", "term", "to", "address", "the", "mode", "collapse", "issue", "for", "cGANs", ".", "The", "proposed", "method", "explicitly", "maximizes", "the", "ratio", "of", "the", "distance", "between", "generated", "images", "with", "respect", "to", "the", "corresponding", "latent", "codes", ",", "thus", "encouraging", "the", "generators", "to", "explore", "more", "minor", "modes", "during", "training", ".", "This", "mode", "seeking", "regularization", "term", "is", "readily", "applicable", "to", "various", "conditional", "generation", "tasks", "without", "imposing", "training", "overhead", "or", "modifying", "the", "original", "network", "structures", ".", "We", "validate", "the", "proposed", "algorithm", "on", "three", "conditional", "image", "synthesis", "tasks", "including", "categorical", "generation", ",", "image", "-", "to", "-", "image", "translation", ",", "and", "text", "-", "to", "-", "image", "synthesis", "with", "different", "baseline", "models", ".", "Both", "qualitative", "and", "quantitative", "results", "demonstrate", "the", "effectiveness", "of", "the", "proposed", "regularization", "method", "for", "improving", "diversity", "without", "loss", "of", "quality", ".", "section", ":", "Introduction", "Generative", "adversarial", "networks", "(", "GANs", ")", "have", "been", "shown", "to", "capture", "complex", "and", "high", "-", "dimensional", "image", "data", "with", "numerous", "applications", "effectively", ".", "Built", "upon", "GANs", ",", "conditional", "GANs", "(", "cGANs", ")", "take", "external", "information", "as", "additional", "inputs", ".", "For", "image", "synthesis", ",", "cGANs", "can", "be", "applied", "to", "various", "tasks", "with", "different", "conditional", "contexts", ".", "With", "class", "labels", ",", "cGANs", "can", "be", "applied", "to", "categorical", "image", "generation", ".", "With", "text", "sentences", ",", "cGANs", "can", "be", "applied", "to", "text", "-", "to", "-", "image", "synthesis", ".", "With", "images", ",", "cGANs", "have", "been", "used", "in", "tasks", "including", "image", "-", "to", "-", "image", "translation", ",", "semantic", "manipulation", "and", "style", "transfer", ".", "For", "most", "conditional", "generation", "tasks", ",", "the", "mappings", "are", "in", "nature", "multimodal", ",", "i.e.", ",", "a", "single", "input", "context", "corresponds", "to", "multiple", "plausible", "outputs", ".", "A", "straightforward", "approach", "to", "handle", "multimodality", "is", "to", "take", "random", "noise", "vectors", "along", "with", "the", "conditional", "contexts", "as", "inputs", ",", "where", "the", "contexts", "determine", "the", "main", "content", "and", "noise", "vectors", "are", "responsible", "for", "variations", ".", "For", "instance", ",", "in", "the", "dog", "-", "to", "-", "cat", "image", "-", "to", "-", "image", "translation", "task", ",", "the", "input", "dog", "images", "decide", "contents", "like", "orientations", "of", "heads", "and", "positions", "of", "facial", "landmarks", ",", "while", "the", "noise", "vectors", "help", "the", "generation", "of", "different", "species", ".", "However", ",", "cGANs", "usually", "suffer", "from", "the", "mode", "collapse", "problem", ",", "where", "generators", "only", "produce", "samples", "from", "a", "single", "or", "few", "modes", "of", "the", "distribution", "and", "ignore", "other", "modes", ".", "The", "noise", "vectors", "are", "ignored", "or", "of", "minor", "impacts", ",", "since", "cGANs", "pay", "more", "attention", "to", "learn", "from", "the", "high", "-", "dimensional", "and", "structured", "conditional", "contexts", ".", "There", "are", "two", "main", "approaches", "to", "address", "the", "mode", "collapse", "problem", "in", "GANs", ".", "A", "number", "of", "methods", "focus", "on", "discriminators", "by", "introducing", "different", "divergence", "metrics", "and", "optimization", "process", ".", "The", "other", "methods", "use", "auxiliary", "networks", "such", "as", "multiple", "generators", "and", "additional", "encoders", ".", "However", ",", "mode", "collapse", "is", "relatively", "less", "studied", "in", "cGANs", ".", "Some", "recent", "efforts", "have", "been", "made", "in", "the", "image", "-", "to", "-", "image", "translation", "task", "to", "improve", "diversity", ".", "Similar", "to", "the", "second", "category", "with", "the", "unconditional", "setting", ",", "these", "approaches", "introduce", "additional", "encoders", "and", "loss", "functions", "to", "encourage", "the", "one", "-", "to", "-", "one", "relationship", "between", "the", "output", "and", "the", "latent", "code", ".", "These", "methods", "either", "entail", "heavy", "computational", "overheads", "on", "training", "or", "require", "auxiliary", "networks", "that", "are", "often", "task", "-", "specific", "that", "can", "not", "be", "easily", "extended", "to", "other", "frameworks", ".", "In", "this", "work", ",", "we", "propose", "a", "mode", "seeking", "regularization", "method", "that", "can", "be", "applied", "to", "cGANs", "for", "various", "tasks", "to", "alleviate", "the", "mode", "collapse", "problem", ".", "Given", "two", "latent", "vectors", "and", "the", "corresponding", "output", "images", ",", "we", "propose", "to", "maximize", "the", "ratio", "of", "the", "distance", "between", "images", "with", "respect", "to", "the", "distance", "between", "latent", "vectors", ".", "In", "other", "words", ",", "this", "regularization", "term", "encourages", "generators", "to", "generate", "dissimilar", "images", "during", "training", ".", "As", "a", "result", ",", "generators", "can", "explore", "the", "target", "distribution", ",", "and", "enhance", "the", "chances", "of", "generating", "samples", "from", "different", "modes", ".", "On", "the", "other", "hand", ",", "we", "can", "train", "the", "discriminators", "with", "dissimilar", "generated", "samples", "to", "provide", "gradients", "from", "minor", "modes", "that", "are", "likely", "to", "be", "ignored", "otherwise", ".", "This", "mode", "seeking", "regularization", "method", "incurs", "marginal", "computational", "overheads", "and", "can", "be", "easily", "embedded", "in", "different", "cGAN", "frameworks", "to", "improve", "the", "diversity", "of", "synthesized", "images", ".", "We", "validate", "the", "proposed", "regularization", "algorithm", "through", "an", "extensive", "evaluation", "of", "three", "conditional", "image", "synthesis", "tasks", "with", "different", "baseline", "models", ".", "First", ",", "for", "categorical", "image", "generation", ",", "we", "apply", "the", "proposed", "method", "on", "DCGAN", "using", "the", "CIFAR", "-", "10", "dataset", ".", "Second", ",", "for", "image", "-", "to", "-", "image", "translation", ",", "we", "embed", "the", "proposed", "regularization", "scheme", "in", "Pix2Pix", "and", "DRIT", "using", "the", "facades", ",", "maps", ",", "Yosemite", ",", "and", "cat", "dog", "datasets", ".", "Third", ",", "for", "text", "-", "to", "-", "image", "synthesis", ",", "we", "incorporate", "StackGAN", "++", "with", "the", "proposed", "regularization", "term", "using", "the", "CUB", "-", "200", "-", "2011", "dataset", ".", "We", "evaluate", "the", "diversity", "of", "synthesized", "images", "using", "perceptual", "distance", "metrics", ".", "However", ",", "the", "diversity", "metric", "alone", "can", "not", "guarantee", "the", "similarity", "between", "the", "distribution", "of", "generated", "images", "and", "the", "distribution", "of", "real", "data", ".", "Therefore", ",", "we", "adopt", "two", "recently", "proposed", "bin", "-", "based", "metrics", ",", "the", "Number", "of", "Statistically", "-", "Different", "Bins", "(", "NDB", ")", "metric", "which", "determines", "the", "relative", "proportions", "of", "samples", "fallen", "into", "clusters", "predetermined", "by", "real", "data", ",", "and", "the", "Jensen", "-", "Shannon", "Divergence", "(", "JSD", ")", "distance", "which", "measures", "the", "similarity", "between", "bin", "distributions", ".", "Furthermore", ",", "to", "verify", "that", "we", "do", "not", "achieve", "diversity", "at", "the", "expense", "of", "realism", ",", "we", "evaluate", "our", "method", "with", "the", "Fréchet", "Inception", "Distance", "(", "FID", ")", "as", "the", "metric", "for", "quality", ".", "Experimental", "results", "demonstrate", "that", "the", "proposed", "regularization", "method", "can", "facilitate", "existing", "models", "from", "various", "applications", "achieving", "better", "diversity", "without", "loss", "of", "image", "quality", ".", "Figure", "[", "reference", "]", "shows", "the", "effectiveness", "of", "the", "proposed", "regularization", "method", "for", "existing", "models", ".", "The", "main", "contributions", "of", "this", "work", "are", ":", "We", "propose", "a", "simple", "yet", "effective", "mode", "seeking", "regularization", "method", "to", "address", "the", "mode", "collapse", "problem", "in", "cGANs", ".", "This", "regularization", "scheme", "can", "be", "readily", "extended", "into", "existing", "frameworks", "with", "marginal", "training", "overheads", "and", "modifications", ".", "We", "demonstrate", "the", "generalizability", "of", "the", "proposed", "regularization", "method", "on", "three", "different", "conditional", "generation", "tasks", ":", "categorical", "generation", ",", "image", "-", "to", "-", "image", "translation", ",", "and", "text", "-", "to", "-", "image", "synthesis", ".", "Extensive", "experiments", "show", "that", "the", "proposed", "method", "can", "facilitate", "existing", "models", "from", "different", "tasks", "achieving", "better", "diversity", "without", "sacrificing", "visual", "quality", "of", "the", "generated", "images", ".", "Our", "code", "and", "pre", "-", "trained", "models", "are", "available", "at", ".", "section", ":", "Related", "Work", "paragraph", ":", "Conditional", "generative", "adversarial", "networks", ".", "Generative", "adversarial", "networks", "have", "been", "widely", "used", "for", "image", "synthesis", ".", "With", "adversarial", "training", ",", "generators", "are", "encouraged", "to", "capture", "the", "distribution", "of", "real", "images", ".", "On", "the", "basis", "of", "GANs", ",", "conditional", "GANs", "synthesize", "images", "based", "on", "various", "contexts", ".", "For", "instances", ",", "cGANs", "can", "generate", "high", "-", "resolution", "images", "conditioned", "on", "low", "-", "resolution", "images", ",", "translate", "images", "between", "different", "visual", "domains", ",", "generate", "images", "with", "desired", "style", ",", "and", "synthesize", "images", "according", "to", "sentences", ".", "Although", "cGANs", "have", "achieved", "success", "in", "various", "applications", ",", "existing", "approaches", "suffer", "from", "the", "mode", "collapse", "problem", ".", "Since", "the", "conditional", "contexts", "provide", "strong", "structural", "prior", "information", "for", "the", "output", "images", "and", "have", "higher", "dimensions", "than", "the", "input", "noise", "vectors", ",", "generators", "tend", "to", "ignore", "the", "input", "noise", "vectors", ",", "which", "are", "responsible", "for", "the", "variation", "of", "generated", "images", ".", "As", "a", "result", ",", "the", "generators", "are", "prone", "to", "produce", "images", "with", "similar", "appearances", ".", "In", "this", "work", ",", "we", "aim", "to", "address", "the", "mode", "collapse", "problem", "for", "cGANs", ".", "paragraph", ":", "Reducing", "mode", "collapse", ".", "Some", "methods", "focus", "on", "the", "discriminator", "with", "different", "optimization", "process", "and", "divergence", "metrics", "to", "stabilize", "the", "training", "process", ".", "The", "minibatch", "discrimination", "scheme", "allows", "the", "discriminator", "to", "discriminate", "between", "whole", "mini", "-", "batches", "of", "samples", "instead", "of", "between", "individual", "samples", ".", "In", ",", "Durugkar", "et", "al", ".", "use", "multiple", "discriminators", "to", "address", "this", "issue", ".", "The", "other", "methods", "use", "auxiliary", "networks", "to", "alleviate", "the", "mode", "collapse", "issue", ".", "ModeGAN", "and", "VEEGAN", "enforce", "the", "bijection", "mapping", "between", "the", "input", "noise", "vectors", "and", "generated", "images", "with", "additional", "encoder", "networks", ".", "Multiple", "generators", "and", "weight", "-", "sharing", "generators", "are", "developed", "to", "capture", "more", "modes", "of", "the", "distribution", ".", "However", ",", "these", "approaches", "either", "entail", "heavy", "computational", "overheads", "or", "require", "modifications", "of", "the", "network", "structure", ",", "and", "may", "not", "be", "easily", "applicable", "to", "cGANs", ".", "In", "the", "field", "of", "cGANs", ",", "some", "efforts", "have", "been", "recently", "made", "to", "address", "the", "mode", "collapse", "issue", "on", "the", "image", "-", "to", "-", "image", "translation", "task", ".", "Similar", "to", "ModeGAN", "and", "VEEGAN", ",", "additional", "encoders", "are", "introduced", "to", "provide", "a", "bijection", "constraint", "between", "the", "generated", "images", "and", "input", "noise", "vectors", ".", "However", ",", "these", "approaches", "require", "other", "task", "-", "specific", "networks", "and", "objective", "functions", ".", "The", "additional", "components", "make", "the", "methods", "less", "generalizable", "and", "incur", "extra", "computational", "loads", "on", "training", ".", "In", "contrast", ",", "we", "propose", "a", "simple", "regularization", "term", "that", "imposes", "no", "training", "overheads", "and", "requires", "no", "modifications", "of", "the", "network", "structure", ".", "Therefore", ",", "the", "proposed", "method", "can", "be", "readily", "applied", "to", "various", "conditional", "generation", "tasks", ".", "section", ":", "Diverse", "Conditional", "Image", "Synthesis", "subsection", ":", "Preliminaries", "The", "training", "process", "of", "GANs", "can", "be", "formulated", "as", "a", "mini", "-", "max", "problem", ":", "a", "discriminator", "learns", "to", "be", "a", "classifier", "by", "assigning", "higher", "discriminative", "values", "to", "the", "real", "data", "samples", "and", "lower", "ones", "to", "the", "generated", "ones", ".", "Meanwhile", ",", "a", "generator", "aims", "to", "fool", "by", "synthesizing", "realistic", "examples", ".", "Through", "adversarial", "training", ",", "the", "gradients", "from", "will", "guide", "toward", "generating", "samples", "with", "the", "distribution", "similar", "to", "the", "real", "data", "one", ".", "The", "mode", "collapse", "problem", "with", "GANs", "is", "well", "known", "in", "the", "literature", ".", "Several", "methods", "attribute", "the", "missing", "mode", "to", "the", "lack", "of", "penalty", "when", "this", "issue", "occurs", ".", "Since", "all", "modes", "usually", "have", "similar", "discriminative", "values", ",", "larger", "modes", "are", "likely", "to", "be", "favored", "through", "the", "training", "process", "based", "on", "gradient", "descent", ".", "On", "the", "other", "hand", ",", "it", "is", "difficult", "to", "generate", "samples", "from", "minor", "modes", ".", "The", "mode", "missing", "problem", "becomes", "worse", "in", "cGANs", ".", "Generally", ",", "conditional", "contexts", "are", "high", "-", "dimensional", "and", "structured", "(", "e.g.", ",", "images", "and", "sentences", ")", "as", "opposed", "to", "the", "noise", "vectors", ".", "As", "such", ",", "the", "generators", "are", "likely", "to", "focus", "on", "the", "contexts", "and", "ignore", "the", "noise", "vectors", ",", "which", "account", "for", "diversity", ".", "subsection", ":", "Mode", "Seeking", "GANs", "In", "this", "work", ",", "we", "propose", "to", "alleviate", "the", "missing", "mode", "problem", "from", "the", "generator", "perspective", ".", "Figure", "[", "reference", "]", "illustrates", "the", "main", "ideas", "of", "our", "approach", ".", "Let", "a", "latent", "vector", "from", "the", "latent", "code", "space", "be", "mapped", "to", "the", "image", "space", ".", "When", "mode", "collapse", "occurs", ",", "the", "mapped", "images", "are", "collapsed", "into", "a", "few", "modes", ".", "Furthermore", ",", "when", "two", "latent", "codes", "and", "are", "closer", ",", "the", "mapped", "images", "and", "are", "more", "likely", "to", "be", "collapsed", "into", "the", "same", "mode", ".", "To", "address", "this", "issue", ",", "we", "propose", "a", "mode", "seeking", "regularization", "term", "to", "directly", "maximize", "the", "ratio", "of", "the", "distance", "between", "and", "with", "respect", "to", "the", "distance", "between", "and", ",", "where", "denotes", "the", "distance", "metric", ".", "The", "regularization", "term", "offers", "a", "virtuous", "circle", "for", "training", "cGANs", ".", "It", "encourages", "the", "generator", "to", "explore", "the", "image", "space", "and", "enhances", "the", "chances", "for", "generating", "samples", "of", "minor", "modes", ".", "On", "the", "other", "hand", ",", "the", "discriminator", "is", "forced", "to", "pay", "attention", "to", "generated", "samples", "from", "minor", "modes", ".", "Figure", "[", "reference", "]", "shows", "a", "mode", "collapse", "situation", "where", "two", "close", "samples", ",", "and", ",", "are", "mapped", "onto", "the", "same", "mode", ".", "However", ",", "with", "the", "proposed", "regularization", "term", ",", "is", "mapped", "to", ",", "which", "belongs", "to", "an", "unexplored", "mode", ".", "With", "the", "adversarial", "mechanism", ",", "the", "generator", "will", "thus", "have", "better", "chances", "to", "generate", "samples", "of", "in", "the", "following", "training", "steps", ".", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "the", "proposed", "regularization", "term", "can", "be", "easily", "integrated", "with", "existing", "cGANs", "by", "appending", "it", "to", "the", "original", "objective", "function", ".", "where", "denotes", "the", "original", "objective", "function", "and", "the", "weights", "to", "control", "the", "importance", "of", "the", "regularization", ".", "Here", ",", "can", "be", "as", "a", "simple", "loss", "function", ".", "For", "example", ",", "in", "categorical", "generation", "task", ",", "where", "denote", "class", "labels", ",", "real", "images", ",", "and", "noise", "vectors", ",", "respectively", ".", "In", "image", "-", "to", "-", "image", "translation", "task", ",", "where", "denotes", "input", "images", "and", "is", "the", "typical", "GAN", "loss", ".", "can", "be", "arbitrary", "complex", "objective", "function", "from", "any", "task", ",", "as", "shown", "in", "Figure", "[", "reference", "]", "(", "b", ")", ".", "We", "name", "the", "proposed", "method", "as", "Mode", "Seeking", "GANs", "(", "MSGANs", ")", ".", "[", "Proposed", "regularization", "]", "[", "Applying", "proposed", "regularization", "on", "StackGAN", "++", "]", "section", ":", "Experiments", "We", "evaluate", "the", "proposed", "regularization", "method", "through", "extensive", "quantitative", "and", "qualitative", "evaluation", ".", "We", "apply", "MSGANs", "to", "the", "baseline", "models", "from", "three", "representative", "conditional", "image", "synthesis", "tasks", ":", "categorical", "generation", ",", "image", "-", "to", "-", "image", "translation", ",", "and", "text", "-", "to", "-", "image", "synthesis", ".", "Note", "that", "we", "augment", "the", "original", "objective", "functions", "with", "the", "proposed", "regularization", "term", "while", "maintaining", "original", "network", "architectures", "and", "hyper", "-", "parameters", ".", "We", "employ", "norm", "distance", "as", "our", "distance", "metrics", "for", "both", "and", "and", "set", "the", "hyper", "-", "parameter", "in", "all", "experiments", ".", "More", "implementation", "and", "evaluation", "details", ",", "please", "refer", "to", "the", "appendixes", ".", "subsection", ":", "Evaluation", "Metrics", "We", "conduct", "evaluations", "using", "the", "following", "metrics", ".", "FID", ".", "To", "evaluate", "the", "quality", "of", "the", "generated", "images", ",", "we", "use", "FID", "to", "measure", "the", "distance", "between", "the", "generated", "distribution", "and", "the", "real", "one", "through", "features", "extracted", "by", "Inception", "Network", ".", "Lower", "FID", "values", "indicate", "better", "quality", "of", "the", "generated", "images", ".", "LPIPS", ".", "To", "evaluate", "diversity", ",", "we", "employ", "LPIPS", "following", ".", "LIPIS", "measures", "the", "average", "feature", "distances", "between", "generated", "samples", ".", "Higher", "LPIPS", "score", "indicates", "better", "diversity", "among", "the", "generated", "images", ".", "NDB", "and", "JSD", ".", "To", "measure", "the", "similarity", "between", "the", "distribution", "between", "real", "images", "and", "generated", "one", ",", "we", "adopt", "two", "bin", "-", "based", "metrics", ",", "NDB", "and", "JSD", ",", "proposed", "in", ".", "These", "metrics", "evaluate", "the", "extent", "of", "mode", "missing", "of", "generative", "models", ".", "Following", ",", "the", "training", "samples", "are", "first", "clustered", "using", "K", "-", "means", "into", "different", "bins", "which", "can", "be", "viewed", "as", "modes", "of", "the", "real", "data", "distribution", ".", "Then", "each", "generated", "sample", "is", "assigned", "to", "the", "bin", "of", "its", "nearest", "neighbor", ".", "We", "calculate", "the", "bin", "-", "proportions", "of", "the", "training", "samples", "and", "the", "synthesized", "samples", "to", "evaluate", "the", "difference", "between", "the", "generated", "distribution", "and", "the", "real", "data", "distribution", ".", "NDB", "score", "and", "JSD", "of", "the", "bin", "-", "proportion", "are", "then", "computed", "to", "measure", "the", "mode", "collapse", ".", "Lower", "NDB", "score", "and", "JSD", "mean", "the", "generated", "data", "distribution", "approaches", "the", "real", "data", "distribution", "better", "by", "fitting", "more", "modes", ".", "Please", "refer", "to", "for", "more", "details", ".", "subsection", ":", "Conditioned", "on", "Class", "Label", "We", "first", "validate", "the", "proposed", "method", "on", "categorical", "generation", ".", "In", "categorical", "generation", ",", "networks", "take", "class", "labels", "as", "conditional", "contexts", "to", "synthesize", "images", "of", "different", "categories", ".", "We", "apply", "the", "regularization", "term", "to", "the", "baseline", "framework", "DCGAN", ".", "We", "conduct", "experiments", "on", "the", "CIFAR", "-", "10", "dataset", "which", "includes", "images", "of", "ten", "categories", ".", "Since", "images", "in", "the", "CIFAR", "-", "10", "dataset", "are", "of", "size", "and", "upsampling", "degrades", "the", "image", "quality", ",", "we", "do", "not", "compute", "LPIPS", "in", "this", "task", ".", "Table", "[", "reference", "]", "and", "Table", "[", "reference", "]", "present", "the", "results", "of", "NDB", ",", "JS", ",", "and", "FID", ".", "MSGAN", "mitigates", "the", "mode", "collapse", "issue", "in", "most", "classes", "while", "maintaining", "image", "quality", ".", "subsection", ":", "Conditioned", "on", "Image", "Image", "-", "to", "-", "image", "translation", "aims", "to", "learn", "the", "mapping", "between", "two", "visual", "domains", ".", "Conditioned", "on", "images", "from", "the", "source", "domain", ",", "models", "attempt", "to", "synthesize", "corresponding", "images", "in", "the", "target", "domain", ".", "Despite", "the", "multimodal", "nature", "of", "the", "image", "-", "to", "-", "image", "translation", "task", ",", "early", "work", "abandons", "noise", "vectors", "and", "performs", "one", "-", "to", "-", "one", "mapping", "since", "the", "latent", "codes", "are", "easily", "ignored", "during", "training", "as", "shown", "in", ".", "To", "achieve", "multimodality", ",", "several", "recent", "attempts", "introduce", "additional", "encoder", "networks", "and", "objective", "functions", "to", "impose", "a", "bijection", "constraint", "between", "the", "latent", "code", "space", "and", "the", "image", "space", ".", "To", "demonstrate", "the", "generalizability", ",", "we", "apply", "the", "proposed", "method", "to", "a", "unimodal", "model", "Pix2Pix", "using", "paired", "training", "data", "and", "a", "multimodal", "model", "DRIT", "using", "unpaired", "images", ".", "subsubsection", ":", "Conditioned", "on", "Paired", "Images", "We", "take", "Pix2Pix", "as", "the", "baseline", "model", ".", "We", "also", "compare", "MSGAN", "to", "BicycleGAN", "which", "generates", "diverse", "images", "with", "paired", "training", "images", ".", "For", "fair", "comparisons", ",", "architectures", "of", "the", "generator", "and", "the", "discriminator", "in", "all", "methods", "follow", "the", "ones", "in", "BicycleGAN", ".", "We", "conduct", "experiments", "on", "the", "facades", "and", "maps", "datasets", ".", "MSGAN", "obtains", "consistent", "improvements", "on", "all", "metrics", "over", "Pix2Pix", ".", "Moreover", ",", "MSGAN", "demonstrates", "comparable", "diversity", "to", "BicycleGAN", ",", "which", "applies", "an", "additional", "encoder", "network", ".", "Figure", ".", "[", "reference", "]", "and", "Table", ".", "[", "reference", "]", "demonstrate", "the", "qualitative", "and", "quantitative", "results", ",", "respectively", ".", "subsubsection", ":", "Conditioned", "on", "Unpaired", "Images", "We", "choose", "DRIT", ",", "one", "of", "the", "state", "-", "of", "-", "the", "-", "art", "frameworks", "to", "generate", "diverse", "images", "with", "unpaired", "training", "data", ",", "as", "the", "baseline", "framework", ".", "Though", "DRIT", "synthesizes", "diverse", "images", "in", "most", "cases", ",", "mode", "collapse", "occurs", "in", "some", "challenging", "shape", "-", "variation", "cases", "(", "e.g.", ",", "translation", "between", "cats", "and", "dogs", ")", ".", "To", "demonstrate", "the", "robustness", "of", "the", "proposed", "method", ",", "we", "evaluate", "on", "the", "shape", "-", "preserving", "Yosemite", "(", "summer", "winter", ")", "dataset", "and", "the", "cat", "dog", "dataset", "that", "requires", "shape", "variations", ".", "As", "the", "quantitative", "results", "exhibited", "in", "Table", ".", "[", "reference", "]", ",", "MSGAN", "performs", "favorably", "against", "DRIT", "in", "all", "metrics", "on", "both", "datasets", ".", "Especially", "in", "the", "challenging", "cat", "dog", "dataset", ",", "MSGAN", "obtains", "substantial", "diversity", "gains", ".", "From", "the", "statistical", "point", "of", "view", ",", "we", "visualize", "the", "bin", "proportions", "of", "the", "dog", "-", "to", "-", "cat", "translation", "in", "Figure", ".", "[", "reference", "]", ".", "The", "graph", "shows", "the", "severe", "mode", "collapse", "issue", "of", "DRIT", "and", "the", "substantial", "improvement", "with", "the", "proposed", "regularization", "term", ".", "Qualitatively", ",", "Figure", ".", "[", "reference", "]", "shows", "that", "MSGAN", "discovers", "more", "modes", "without", "the", "loss", "of", "visual", "quality", ".", "subsection", ":", "Conditioned", "on", "Text", "Text", "-", "to", "-", "image", "synthesis", "targets", "at", "generating", "images", "conditioned", "on", "text", "descriptions", ".", "We", "integrate", "the", "proposed", "regularization", "term", "on", "StackGAN", "++", "using", "the", "CUB", "-", "200", "-", "2011", "dataset", ".", "To", "improve", "diversity", ",", "StackGAN", "++", "introduces", "a", "Conditioning", "Augmentation", "(", "CA", ")", "module", "that", "re", "-", "parameterizes", "text", "descriptions", "into", "text", "codes", "of", "the", "Gaussian", "distribution", ".", "Instead", "of", "applying", "the", "regularization", "term", "on", "the", "semantically", "meaningful", "text", "codes", ",", "we", "focus", "on", "exploiting", "the", "latent", "codes", "randomly", "sampled", "from", "the", "prior", "distribution", ".", "However", ",", "for", "a", "fair", "comparison", ",", "we", "evaluation", "MSGAN", "against", "StackGAN", "++", "in", "two", "settings", ":", "1", ")", "Perform", "generation", "without", "fixing", "text", "codes", "for", "text", "descriptions", ".", "In", "this", "case", ",", "text", "codes", "also", "provide", "variations", "for", "output", "images", ".", "2", ")", "Perform", "generation", "with", "fixed", "text", "codes", ".", "In", "this", "setting", ",", "the", "effects", "of", "text", "codes", "are", "excluded", ".", "Table", ".", "[", "reference", "]", "presents", "quantitative", "comparisons", "between", "MSGAN", "and", "StackGAN", "++", ".", "MSGAN", "improves", "the", "diversity", "of", "StackGAN", "++", "and", "maintains", "visual", "quality", ".", "To", "better", "illustrate", "the", "role", "that", "latent", "codes", "play", "for", "the", "diversity", ",", "we", "show", "qualitative", "comparisons", "with", "the", "text", "codes", "fixed", ".", "In", "this", "setting", ",", "we", "do", "not", "consider", "the", "diversity", "resulting", "from", "CA", ".", "Figure", ".", "[", "reference", "]", "illustrates", "that", "latent", "codes", "of", "StackGAN", "++", "have", "minor", "effects", "on", "the", "variations", "of", "the", "image", ".", "On", "the", "contrary", ",", "latent", "codes", "of", "MSGAN", "contribute", "to", "various", "appearances", "and", "poses", "of", "birds", ".", "subsection", ":", "Interpolation", "of", "Latent", "Space", "in", "MSGANs", "We", "perform", "linear", "interpolation", "between", "two", "given", "latent", "codes", "and", "generate", "corresponding", "images", "to", "have", "a", "better", "understanding", "of", "how", "well", "MSGANs", "exploit", "the", "latent", "space", ".", "Figure", ".", "[", "reference", "]", "shows", "the", "interpolation", "results", "on", "the", "dog", "-", "to", "-", "cat", "translation", "and", "the", "text", "-", "to", "-", "image", "synthesis", "task", ".", "In", "the", "dog", "-", "to", "-", "cat", "translation", ",", "we", "can", "see", "the", "coat", "colors", "and", "patterns", "varies", "smoothly", "along", "with", "the", "latent", "vectors", ".", "In", "the", "text", "-", "to", "-", "image", "synthesis", ",", "both", "orientations", "of", "birds", "and", "the", "appearances", "of", "footholds", "change", "gradually", "with", "the", "variations", "of", "the", "latent", "codes", ".", "section", ":", "Conclusions", "In", "this", "work", ",", "we", "present", "a", "simple", "but", "effective", "mode", "seeking", "regularization", "term", "on", "the", "generator", "to", "address", "the", "model", "collapse", "in", "cGANs", ".", "By", "maximizing", "the", "distance", "between", "generated", "images", "with", "respect", "to", "that", "between", "the", "corresponding", "latent", "codes", ",", "the", "regularization", "term", "forces", "the", "generators", "to", "explore", "more", "minor", "modes", ".", "The", "proposed", "regularization", "method", "can", "be", "readily", "integrated", "with", "existing", "cGANs", "framework", "without", "imposing", "training", "overheads", "and", "modifications", "of", "network", "structures", ".", "We", "demonstrate", "the", "generalizability", "of", "the", "proposed", "method", "on", "three", "different", "conditional", "generation", "tasks", "including", "categorical", "generation", ",", "image", "-", "to", "-", "image", "translation", ",", "and", "text", "-", "to", "-", "image", "synthesis", ".", "Both", "qualitative", "and", "quantitative", "results", "show", "that", "the", "proposed", "regularization", "term", "facilitates", "the", "baseline", "frameworks", "improving", "the", "diversity", "without", "sacrificing", "visual", "quality", "of", "the", "generated", "images", ".", "bibliography", ":", "References", "section", ":", "Implementation", "Details", "Table", "[", "reference", "]", "summarizes", "the", "datasets", "and", "baseline", "models", "used", "on", "various", "tasks", ".", "For", "all", "of", "the", "baseline", "methods", ",", "we", "incorporate", "the", "original", "objective", "functions", "with", "the", "proposed", "regularization", "term", ".", "Note", "that", "we", "remain", "the", "original", "network", "architecture", "design", "and", "use", "the", "default", "setting", "of", "hyper", "-", "parameters", "for", "the", "training", ".", "DCGAN", ".", "Since", "the", "images", "in", "the", "CIFAR", "-", "10", "dataset", "are", "of", "size", ",", "we", "modify", "the", "structure", "of", "the", "generator", "and", "discriminator", "in", "DCGAN", ",", "as", "shown", "in", "Table", "[", "reference", "]", ".", "We", "use", "the", "batch", "size", "of", ",", "learning", "rate", "of", "and", "Adam", "optimizer", "with", "and", "to", "train", "both", "the", "baseline", "and", "MSGAN", "network", ".", "Pix2Pix", ".", "We", "adopt", "the", "generator", "and", "discriminator", "in", "BicycleGAN", "to", "build", "the", "Pix2Pix", "model", ".", "Same", "as", "BicycleGAN", ",", "we", "use", "a", "U", "-", "Net", "network", "for", "the", "generator", ",", "and", "inject", "the", "latent", "codes", "into", "every", "layer", "of", "the", "generator", ".", "The", "architecture", "of", "the", "discriminator", "is", "a", "two", "-", "scale", "PatchGAN", "network", ".", "For", "the", "training", ",", "both", "Pix2Pix", "and", "MSGAN", "framework", "use", "the", "same", "hyper", "-", "parameters", "as", "the", "officially", "released", "version", ".", "DRIT", ".", "DRIT", "involves", "two", "stages", "of", "image", "-", "to", "-", "image", "translations", "in", "the", "training", "process", ".", "We", "only", "apply", "the", "mode", "seeking", "regularization", "term", "to", "generators", "in", "the", "first", "stage", ",", "which", "is", "modified", "on", "the", "officially", "released", "code", ".", "StackGAN", "++", ".", "StackGAN", "++", "is", "a", "tree", "-", "like", "structure", "with", "multiple", "generators", "and", "discriminators", ".", "We", "use", "the", "output", "images", "from", "the", "last", "generator", "and", "input", "latent", "codes", "to", "calculate", "the", "mode", "seeking", "regularization", "term", ".", "The", "implementation", "is", "based", "on", "the", "officially", "released", "code", ".", "section", ":", "Evaluation", "Details", "We", "employ", "the", "official", "implementation", "of", "FID", ",", "NDB", "and", "JSD", ",", "and", "LPIPS", ".", "For", "NDB", "and", "JSD", ",", "we", "use", "the", "K", "-", "means", "method", "on", "training", "samples", "to", "obtain", "the", "clusters", ".", "Then", "the", "generated", "samples", "are", "assigned", "to", "the", "nearest", "cluster", "to", "compute", "the", "bin", "proportions", ".", "As", "suggested", "by", "the", "author", "of", ",", "there", "are", "at", "least", "training", "samples", "for", "each", "cluster", ".", "Therefore", ",", "we", "cluster", "the", "number", "of", "bins", "in", "all", "tasks", ",", "where", "denotes", "the", "number", "of", "training", "samples", "for", "computing", "the", "clusters", ".", "We", "have", "verified", "that", "the", "performance", "is", "consistent", "within", "a", "large", "range", "of", ".", "For", "evaluation", ",", "we", "randomly", "generate", "images", "for", "a", "given", "conditional", "context", "on", "various", "tasks", ".", "We", "conduct", "five", "independent", "trials", "and", "report", "the", "mean", "and", "standard", "derivation", "based", "on", "the", "result", "of", "each", "trial", ".", "More", "evaluation", "details", "of", "one", "trial", "are", "presented", "as", "follows", ".", "Conditioned", "on", "Class", "Label", ".", "We", "randomly", "generate", "images", "for", "each", "class", "label", ".", "We", "use", "all", "the", "training", "samples", "and", "the", "generated", "samples", "to", "compute", "FID", ".", "For", "NDB", "and", "JSD", ",", "we", "employ", "the", "training", "samples", "in", "each", "class", "to", "calculate", "clusters", ".", "Conditioned", "on", "Image", ".", "We", "randomly", "generate", "images", "for", "each", "input", "image", "in", "the", "test", "set", ".", "For", "LPIPS", ",", "we", "randomly", "select", "pairs", "of", "the", "images", "of", "each", "context", "in", "the", "test", "set", "to", "compute", "LPIPS", "and", "average", "all", "the", "values", "for", "this", "trial", ".", "Then", ",", "we", "randomly", "choose", "input", "images", "and", "their", "corresponding", "generated", "images", "to", "form", "generated", "samples", ".", "We", "use", "the", "generated", "samples", "and", "all", "samples", "in", "training", "set", "to", "compute", "FID", ".", "For", "NDB", "and", "JSD", ",", "we", "employ", "all", "the", "training", "samples", "for", "clustering", "and", "choose", "bins", "for", "facades", ",", "and", "bins", "for", "other", "datasets", ".", "Conditioned", "on", "Text", ".", "We", "randomly", "select", "sentences", "and", "generate", "images", "for", "each", "sentence", ",", "which", "forms", "generated", "samples", ".", "Then", ",", "we", "randomly", "select", "samples", "for", "computing", "FID", ",", "and", "clustering", "them", "into", "bins", "for", "NDB", "and", "JSD", ".", "For", "LPIPS", ",", "we", "randomly", "choose", "pairs", "for", "each", "sentence", "and", "average", "the", "values", "of", "all", "the", "pairs", "for", "this", "trial", ".", "section", ":", "Ablation", "Study", "on", "the", "Regularization", "Term", "subsection", ":", "The", "Weighting", "Parameter", "To", "analyze", "the", "influence", "of", "the", "regularization", "term", ",", "we", "conduct", "an", "ablation", "study", "by", "varying", "the", "weighting", "parameter", "on", "image", "-", "to", "-", "image", "translation", "task", "using", "the", "facades", "dataset", ".", "Figure", "[", "reference", "]", "presents", "the", "qualitative", "and", "quantitative", "results", ".", "It", "can", "be", "observed", "that", "increasing", "improves", "both", "the", "quality", "and", "diversity", "of", "the", "generated", "images", ".", "Nevertheless", ",", "as", "the", "weighting", "parameter", "becomes", "larger", "than", "a", "threshold", "value", "(", ")", ",", "the", "training", "becomes", "unstable", ",", "which", "yields", "low", "quality", ",", "and", "even", "low", "diversity", "synthesized", "images", ".", "As", "a", "result", ",", "we", "empirically", "set", "the", "weighting", "parameter", "for", "all", "experiments", ".", "subsection", ":", "The", "Design", "Choice", "of", "the", "Distance", "Metric", "We", "have", "explored", "other", "design", "choice", "of", "the", "distance", "metric", ".", "We", "conduct", "experiments", "using", "discriminator", "feature", "distance", "in", "our", "regularization", "term", "in", "a", "way", "similar", "to", "feature", "matching", "loss", ",", "where", "denotes", "the", "layer", "of", "the", "discriminator", ".", "We", "apply", "it", "to", "Pix2Pix", "on", "the", "facades", "dataset", ".", "Table", ".", "[", "reference", "]", "shows", "that", "MSGAN", "using", "feature", "distance", "also", "obtains", "improvement", "over", "Pix2Pix", ".", "However", ",", "MSGAN", "using", "distance", "has", "higher", "diversity", ".", "Therefore", ",", "we", "employ", "MSGAN", "using", "distance", "for", "all", "experiments", ".", "section", ":", "Computational", "Overheads", "We", "compare", "MSGAN", "with", "Pix2Pix", ",", "BicycleGAN", "in", "terms", "of", "training", "time", ",", "memory", "consumption", ",", "and", "model", "parameters", "on", "an", "NVIDIA", "TITAN", "X", "GPU", ".", "Table", ".", "[", "reference", "]", "shows", "that", "our", "method", "incurs", "marginal", "overheads", ".", "However", ",", "BicycleGAN", "requires", "longer", "time", "per", "iteration", "and", "larger", "memory", "with", "an", "additional", "encoder", "and", "another", "discriminator", "network", ".", "section", ":", "Additional", "Results", "We", "present", "more", "results", "of", "categorical", "generation", ",", "image", "-", "to", "-", "image", "translation", ",", "and", "text", "-", "to", "-", "image", "synthesis", "in", "Figure", "[", "reference", "]", ",", "Figure", "[", "reference", "]", ",", "Figure", "[", "reference", "]", ",", "Figure", "[", "reference", "]", ",", "Figure", "[", "reference", "]", ",", "and", "Figure", "[", "reference", "]", ",", "respectively", "."], "entity": [[2, 7, ["Mode", "Seeking", "Generative", "Adversarial", "Networks"], "Method"], [8, 11, ["Diverse", "Image", "Synthesis"], "Task"], [12, 15, ["conditional", "generation", "tasks"], "Task"], [26, 30, ["conditional", "generative", "adversarial", "networks"], "Method"], [31, 33, ["cGANs", ")"], "Method"], [59, 62, ["mode", "collapse", "issue"], "Task"], [63, 64, ["cGANs"], "Method"], [83, 85, ["regularization", "term"], "Method"], [88, 91, ["mode", "collapse", "issue"], "Task"], [92, 93, ["cGANs"], "Task"], [128, 132, ["mode", "seeking", "regularization", "term"], "Method"], [137, 140, ["conditional", "generation", "tasks"], "Task"], [158, 162, ["conditional", "image", "synthesis", "tasks"], "Task"], [163, 165, ["categorical", "generation"], "Task"], [166, 172, ["image", "-", "to", "-", "image", "translation"], "Task"], [174, 180, ["text", "-", "to", "-", "image", "synthesis"], "Task"], [182, 184, ["baseline", "models"], "Method"], [196, 198, ["regularization", "method"], "Method"], [200, 201, ["diversity"], "Metric"], [202, 205, ["loss", "of", "quality"], "Metric"], [209, 212, ["Generative", "adversarial", "networks"], "Method"], [213, 214, ["GANs"], "Method"], [234, 235, ["GANs"], "Method"], [236, 238, ["conditional", "GANs"], "Method"], [239, 240, ["cGANs"], "Method"], [249, 251, ["image", "synthesis"], "Task"], [252, 253, ["cGANs"], "Method"], [268, 269, ["cGANs"], "Method"], [273, 276, ["categorical", "image", "generation"], "Task"], [281, 282, ["cGANs"], "Method"], [286, 292, ["text", "-", "to", "-", "image", "synthesis"], "Task"], [296, 297, ["cGANs"], "Method"], [303, 309, ["image", "-", "to", "-", "image", "translation"], "Task"], [310, 312, ["semantic", "manipulation"], "Task"], [313, 315, ["style", "transfer"], "Task"], [318, 321, ["conditional", "generation", "tasks"], "Task"], [381, 393, ["dog", "-", "to", "-", "cat", "image", "-", "to", "-", "image", "translation", "task"], "Task"], [423, 424, ["cGANs"], "Method"], [428, 431, ["mode", "collapse", "problem"], "Task"], [433, 434, ["generators"], "Method"], [462, 463, ["cGANs"], "Method"], [486, 489, ["mode", "collapse", "problem"], "Task"], [490, 491, ["GANs"], "Task"], [498, 499, ["discriminators"], "Method"], [502, 504, ["divergence", "metrics"], "Metric"], [505, 507, ["optimization", "process"], "Method"], [512, 514, ["auxiliary", "networks"], "Method"], [516, 518, ["multiple", "generators"], "Method"], [520, 521, ["encoders"], "Method"], [524, 526, ["mode", "collapse"], "Task"], [531, 532, ["cGANs"], "Method"], [541, 548, ["image", "-", "to", "-", "image", "translation", "task"], "Task"], [559, 561, ["unconditional", "setting"], "Task"], [566, 567, ["encoders"], "Method"], [592, 594, ["computational", "overheads"], "Metric"], [595, 596, ["training"], "Task"], [598, 600, ["auxiliary", "networks"], "Method"], [623, 627, ["mode", "seeking", "regularization", "method"], "Method"], [632, 633, ["cGANs"], "Method"], [639, 642, ["mode", "collapse", "problem"], "Task"], [678, 680, ["regularization", "term"], "Method"], [720, 721, ["discriminators"], "Method"], [740, 744, ["mode", "seeking", "regularization", "method"], "Method"], [746, 748, ["computational", "overheads"], "Metric"], [755, 757, ["cGAN", "frameworks"], "Method"], [769, 771, ["regularization", "algorithm"], "Method"], [777, 781, ["conditional", "image", "synthesis", "tasks"], "Task"], [783, 785, ["baseline", "models"], "Method"], [789, 792, ["categorical", "image", "generation"], "Task"], [799, 800, ["DCGAN"], "Method"], [802, 806, ["CIFAR", "-", "10", "dataset"], "Material"], [810, 816, ["image", "-", "to", "-", "image", "translation"], "Task"], [821, 823, ["regularization", "scheme"], "Method"], [824, 827, ["Pix2Pix", "and", "DRIT"], "Method"], [843, 849, ["text", "-", "to", "-", "image", "synthesis"], "Task"], [852, 854, ["StackGAN", "++"], "Method"], [857, 859, ["regularization", "term"], "Method"], [876, 879, ["perceptual", "distance", "metrics"], "Metric"], [883, 885, ["diversity", "metric"], "Metric"], [911, 915, ["bin", "-", "based", "metrics"], "Method"], [924, 927, ["NDB", ")", "metric"], "Method"], [944, 948, ["Jensen", "-", "Shannon", "Divergence"], "Metric"], [949, 950, ["JSD"], "Metric"], [974, 975, ["realism"], "Metric"], [982, 985, ["Fréchet", "Inception", "Distance"], "Metric"], [986, 987, ["FID"], "Metric"], [990, 991, ["metric"], "Metric"], [992, 993, ["quality"], "Metric"], [1000, 1002, ["regularization", "method"], "Method"], [1015, 1017, ["image", "quality"], "Metric"], [1028, 1030, ["regularization", "method"], "Method"], [1048, 1052, ["mode", "seeking", "regularization", "method"], "Method"], [1055, 1058, ["mode", "collapse", "problem"], "Task"], [1059, 1060, ["cGANs"], "Task"], [1062, 1064, ["regularization", "scheme"], "Method"], [1085, 1087, ["regularization", "method"], "Method"], [1090, 1093, ["conditional", "generation", "tasks"], "Task"], [1094, 1096, ["categorical", "generation"], "Task"], [1097, 1103, ["image", "-", "to", "-", "image", "translation"], "Task"], [1105, 1111, ["text", "-", "to", "-", "image", "synthesis"], "Task"], [1128, 1129, ["diversity"], "Metric"], [1131, 1133, ["visual", "quality"], "Metric"], [1155, 1159, ["Conditional", "generative", "adversarial", "networks"], "Method"], [1160, 1163, ["Generative", "adversarial", "networks"], "Method"], [1168, 1170, ["image", "synthesis"], "Task"], [1172, 1174, ["adversarial", "training"], "Method"], [1175, 1176, ["generators"], "Method"], [1190, 1191, ["GANs"], "Method"], [1192, 1194, ["conditional", "GANs"], "Method"], [1204, 1205, ["cGANs"], "Method"], [1239, 1240, ["cGANs"], "Method"], [1252, 1255, ["mode", "collapse", "problem"], "Task"], [1322, 1325, ["mode", "collapse", "problem"], "Task"], [1326, 1327, ["cGANs"], "Task"], [1330, 1333, ["Reducing", "mode", "collapse"], "Task"], [1339, 1340, ["discriminator"], "Method"], [1342, 1344, ["optimization", "process"], "Method"], [1345, 1347, ["divergence", "metrics"], "Metric"], [1350, 1352, ["training", "process"], "Task"], [1354, 1357, ["minibatch", "discrimination", "scheme"], "Method"], [1383, 1384, ["discriminators"], "Method"], [1393, 1395, ["auxiliary", "networks"], "Method"], [1398, 1401, ["mode", "collapse", "issue"], "Task"], [1402, 1403, ["ModeGAN"], "Method"], [1419, 1421, ["encoder", "networks"], "Method"], [1422, 1424, ["Multiple", "generators"], "Method"], [1425, 1429, ["weight", "-", "sharing", "generators"], "Method"], [1463, 1464, ["cGANs"], "Method"], [1469, 1470, ["cGANs"], "Task"], [1480, 1483, ["mode", "collapse", "issue"], "Task"], [1485, 1492, ["image", "-", "to", "-", "image", "translation", "task"], "Task"], [1495, 1496, ["ModeGAN"], "Method"], [1497, 1498, ["VEEGAN"], "Method"], [1500, 1501, ["encoders"], "Method"], [1542, 1544, ["computational", "loads"], "Metric"], [1545, 1546, ["training"], "Task"], [1554, 1556, ["regularization", "term"], "Method"], [1581, 1584, ["conditional", "generation", "tasks"], "Task"], [1587, 1591, ["Diverse", "Conditional", "Image", "Synthesis"], "Task"], [1598, 1599, ["GANs"], "Method"], [1604, 1608, ["mini", "-", "max", "problem"], "Task"], [1610, 1611, ["discriminator"], "Method"], [1615, 1616, ["classifier"], "Method"], [1637, 1638, ["generator"], "Method"], [1647, 1649, ["adversarial", "training"], "Method"], [1669, 1672, ["mode", "collapse", "problem"], "Task"], [1673, 1674, ["GANs"], "Method"], [1715, 1717, ["training", "process"], "Method"], [1719, 1721, ["gradient", "descent"], "Method"], [1738, 1741, ["mode", "missing", "problem"], "Task"], [1744, 1745, ["cGANs"], "Method"], [1795, 1798, ["Mode", "Seeking", "GANs"], "Method"], [1807, 1810, ["missing", "mode", "problem"], "Task"], [1812, 1814, ["generator", "perspective"], "Method"], [1844, 1846, ["mode", "collapse"], "Task"], [1891, 1895, ["mode", "seeking", "regularization", "term"], "Method"], [1916, 1918, ["distance", "metric"], "Metric"], [1920, 1922, ["regularization", "term"], "Method"], [1928, 1929, ["cGANs"], "Method"], [1956, 1957, ["discriminator"], "Method"], [1997, 1999, ["regularization", "term"], "Method"], [2013, 2015, ["adversarial", "mechanism"], "Method"], [2017, 2018, ["generator"], "Method"], [2043, 2045, ["regularization", "term"], "Method"], [2051, 2052, ["cGANs"], "Method"], [2085, 2087, ["loss", "function"], "Method"], [2111, 2118, ["image", "-", "to", "-", "image", "translation", "task"], "Task"], [2127, 2129, ["GAN", "loss"], "Metric"], [2157, 2160, ["Mode", "Seeking", "GANs"], "Method"], [2161, 2162, ["MSGANs"], "Method"], [2166, 2167, ["regularization"], "Method"], [2171, 2172, ["regularization"], "Method"], [2173, 2175, ["StackGAN", "++"], "Method"], [2183, 2185, ["regularization", "method"], "Method"], [2187, 2191, ["quantitative", "and", "qualitative", "evaluation"], "Metric"], [2194, 2195, ["MSGANs"], "Method"], [2202, 2206, ["conditional", "image", "synthesis", "tasks"], "Task"], [2207, 2209, ["categorical", "generation"], "Task"], [2210, 2216, ["image", "-", "to", "-", "image", "translation"], "Task"], [2218, 2224, ["text", "-", "to", "-", "image", "synthesis"], "Task"], [2250, 2252, ["norm", "distance"], "Method"], [2254, 2256, ["distance", "metrics"], "Metric"], [2283, 2285, ["Evaluation", "Metrics"], "Metric"], [2293, 2294, ["FID"], "Metric"], [2306, 2307, ["FID"], "Metric"], [2310, 2311, ["distance"], "Metric"], [2323, 2325, ["Inception", "Network"], "Method"], [2337, 2338, ["LPIPS"], "Method"], [2345, 2347, ["LPIPS", "following"], "Method"], [2348, 2349, ["LIPIS"], "Method"], [2359, 2361, ["LPIPS", "score"], "Metric"], [2369, 2370, ["NDB"], "Method"], [2371, 2372, ["JSD"], "Metric"], [2390, 2394, ["bin", "-", "based", "metrics"], "Method"], [2395, 2396, ["NDB"], "Method"], [2397, 2398, ["JSD"], "Metric"], [2408, 2413, ["mode", "missing", "of", "generative", "models"], "Method"], [2423, 2426, ["K", "-", "means"], "Method"], [2483, 2485, ["NDB", "score"], "Metric"], [2486, 2487, ["JSD"], "Metric"], [2502, 2504, ["NDB", "score"], "Metric"], [2505, 2506, ["JSD"], "Metric"], [2509, 2511, ["data", "distribution"], "Method"], [2542, 2544, ["categorical", "generation"], "Task"], [2546, 2548, ["categorical", "generation"], "Task"], [2566, 2568, ["regularization", "term"], "Method"], [2570, 2573, ["baseline", "framework", "DCGAN"], "Method"], [2579, 2583, ["CIFAR", "-", "10", "dataset"], "Material"], [2594, 2598, ["CIFAR", "-", "10", "dataset"], "Material"], [2605, 2607, ["image", "quality"], "Metric"], [2612, 2613, ["LPIPS"], "Method"], [2632, 2633, ["JS"], "Method"], [2635, 2636, ["FID"], "Metric"], [2637, 2638, ["MSGAN"], "Method"], [2640, 2643, ["mode", "collapse", "issue"], "Task"], [2648, 2650, ["image", "quality"], "Metric"], [2656, 2662, ["Image", "-", "to", "-", "image", "translation"], "Task"], [2666, 2671, ["mapping", "between", "two", "visual", "domains"], "Task"], [2697, 2704, ["image", "-", "to", "-", "image", "translation", "task"], "Task"], [2712, 2718, ["one", "-", "to", "-", "one", "mapping"], "Task"], [2740, 2742, ["encoder", "networks"], "Method"], [2743, 2745, ["objective", "functions"], "Method"], [2772, 2775, ["unimodal", "model", "Pix2Pix"], "Method"], [2781, 2784, ["multimodal", "model", "DRIT"], "Method"], [2788, 2789, ["subsubsection"], "Method"], [2796, 2797, ["Pix2Pix"], "Method"], [2805, 2806, ["MSGAN"], "Method"], [2807, 2808, ["BicycleGAN"], "Method"], [2824, 2825, ["generator"], "Method"], [2827, 2828, ["discriminator"], "Method"], [2847, 2848, ["MSGAN"], "Method"], [2853, 2854, ["metrics"], "Metric"], [2855, 2856, ["Pix2Pix"], "Method"], [2859, 2860, ["MSGAN"], "Method"], [2862, 2863, ["diversity"], "Metric"], [2864, 2865, ["BicycleGAN"], "Method"], [2870, 2872, ["encoder", "network"], "Method"], [2893, 2894, ["subsubsection"], "Method"], [2901, 2902, ["DRIT"], "Method"], [2929, 2930, ["DRIT"], "Method"], [2937, 2939, ["mode", "collapse"], "Task"], [2943, 2947, ["shape", "-", "variation", "cases"], "Task"], [3001, 3002, ["MSGAN"], "Method"], [3005, 3006, ["DRIT"], "Method"], [3021, 3022, ["MSGAN"], "Method"], [3063, 3064, ["DRIT"], "Method"], [3071, 3073, ["regularization", "term"], "Method"], [3083, 3084, ["MSGAN"], "Method"], [3089, 3093, ["loss", "of", "visual", "quality"], "Metric"], [3099, 3105, ["Text", "-", "to", "-", "image", "synthesis"], "Task"], [3118, 3120, ["regularization", "term"], "Method"], [3121, 3123, ["StackGAN", "++"], "Method"], [3136, 3138, ["StackGAN", "++"], "Method"], [3140, 3146, ["Conditioning", "Augmentation", "(", "CA", ")", "module"], "Method"], [3157, 3159, ["Gaussian", "distribution"], "Method"], [3196, 3197, ["MSGAN"], "Method"], [3198, 3200, ["StackGAN", "++"], "Method"], [3207, 3208, ["generation"], "Task"], [3232, 3233, ["generation"], "Task"], [3259, 3260, ["MSGAN"], "Method"], [3261, 3263, ["StackGAN", "++"], "Method"], [3264, 3265, ["MSGAN"], "Method"], [3269, 3271, ["StackGAN", "++"], "Method"], [3273, 3275, ["visual", "quality"], "Metric"], [3320, 3322, ["latent", "codes"], "Method"], [3323, 3325, ["StackGAN", "++"], "Method"], [3342, 3343, ["MSGAN"], "Method"], [3354, 3358, ["Interpolation", "of", "Latent", "Space"], "Task"], [3359, 3360, ["MSGANs"], "Task"], [3362, 3364, ["linear", "interpolation"], "Method"], [3381, 3382, ["MSGANs"], "Method"], [3398, 3404, ["dog", "-", "to", "-", "cat", "translation"], "Task"], [3406, 3413, ["text", "-", "to", "-", "image", "synthesis", "task"], "Task"], [3416, 3422, ["dog", "-", "to", "-", "cat", "translation"], "Task"], [3441, 3447, ["text", "-", "to", "-", "image", "synthesis"], "Task"], [3480, 3484, ["mode", "seeking", "regularization", "term"], "Method"], [3486, 3487, ["generator"], "Method"], [3490, 3494, ["model", "collapse", "in", "cGANs"], "Task"], [3513, 3515, ["regularization", "term"], "Method"], [3526, 3528, ["regularization", "method"], "Method"], [3534, 3536, ["cGANs", "framework"], "Method"], [3557, 3560, ["conditional", "generation", "tasks"], "Task"], [3561, 3563, ["categorical", "generation"], "Task"], [3564, 3570, ["image", "-", "to", "-", "image", "translation"], "Task"], [3572, 3578, ["text", "-", "to", "-", "image", "synthesis"], "Task"], [3588, 3590, ["regularization", "term"], "Method"], [3596, 3597, ["diversity"], "Metric"], [3599, 3601, ["visual", "quality"], "Metric"], [3653, 3656, ["network", "architecture", "design"], "Method"], [3669, 3670, ["DCGAN"], "Method"], [3676, 3680, ["CIFAR", "-", "10", "dataset"], "Material"], [3690, 3693, ["generator", "and", "discriminator"], "Method"], [3694, 3695, ["DCGAN"], "Method"], [3711, 3713, ["learning", "rate"], "Metric"], [3715, 3717, ["Adam", "optimizer"], "Method"], [3723, 3727, ["baseline", "and", "MSGAN", "network"], "Method"], [3728, 3729, ["Pix2Pix"], "Method"], [3733, 3736, ["generator", "and", "discriminator"], "Method"], [3737, 3738, ["BicycleGAN"], "Method"], [3741, 3743, ["Pix2Pix", "model"], "Method"], [3746, 3747, ["BicycleGAN"], "Method"], [3751, 3755, ["U", "-", "Net", "network"], "Method"], [3757, 3758, ["generator"], "Method"], [3775, 3776, ["discriminator"], "Method"], [3778, 3783, ["two", "-", "scale", "PatchGAN", "network"], "Method"], [3786, 3787, ["training"], "Task"], [3789, 3793, ["Pix2Pix", "and", "MSGAN", "framework"], "Method"], [3805, 3806, ["DRIT"], "Method"], [3807, 3808, ["DRIT"], "Method"], [3812, 3818, ["image", "-", "to", "-", "image", "translations"], "Task"], [3820, 3822, ["training", "process"], "Task"], [3827, 3831, ["mode", "seeking", "regularization", "term"], "Method"], [3847, 3849, ["StackGAN", "++"], "Method"], [3850, 3852, ["StackGAN", "++"], "Method"], [3854, 3858, ["tree", "-", "like", "structure"], "Method"], [3862, 3863, ["discriminators"], "Method"], [3880, 3884, ["mode", "seeking", "regularization", "term"], "Method"], [3905, 3906, ["FID"], "Metric"], [3912, 3913, ["LPIPS"], "Method"], [3915, 3916, ["NDB"], "Task"], [3917, 3918, ["JSD"], "Metric"], [3922, 3926, ["K", "-", "means", "method"], "Method"], [4078, 4079, ["FID"], "Metric"], [4081, 4082, ["NDB"], "Task"], [4083, 4084, ["JSD"], "Metric"], [4115, 4116, ["LPIPS"], "Method"], [4133, 4134, ["LPIPS"], "Method"], [4173, 4174, ["FID"], "Metric"], [4176, 4177, ["NDB"], "Task"], [4178, 4179, ["JSD"], "Metric"], [4187, 4188, ["clustering"], "Task"], [4228, 4229, ["FID"], "Metric"], [4236, 4237, ["NDB"], "Task"], [4238, 4239, ["JSD"], "Metric"], [4241, 4242, ["LPIPS"], "Method"], [4264, 4266, ["Ablation", "Study"], "Task"], [4268, 4270, ["Regularization", "Term"], "Method"], [4287, 4289, ["ablation", "study"], "Task"], [4295, 4302, ["image", "-", "to", "-", "image", "translation", "task"], "Task"], [4327, 4328, ["quality"], "Metric"], [4351, 4352, ["training"], "Method"], [4388, 4390, ["Distance", "Metric"], "Metric"], [4398, 4400, ["distance", "metric"], "Metric"], [4417, 4420, ["feature", "matching", "loss"], "Task"], [4446, 4447, ["MSGAN"], "Method"], [4448, 4450, ["feature", "distance"], "Method"], [4454, 4455, ["Pix2Pix"], "Method"], [4458, 4461, ["MSGAN", "using", "distance"], "Method"], [4469, 4470, ["MSGAN"], "Method"], [4471, 4472, ["distance"], "Method"], [4478, 4480, ["Computational", "Overheads"], "Metric"], [4482, 4483, ["MSGAN"], "Method"], [4484, 4485, ["Pix2Pix"], "Method"], [4486, 4487, ["BicycleGAN"], "Method"], [4490, 4492, ["training", "time"], "Metric"], [4493, 4495, ["memory", "consumption"], "Metric"], [4497, 4499, ["model", "parameters"], "Method"], [4516, 4518, ["marginal", "overheads"], "Metric"], [4521, 4522, ["BicycleGAN"], "Method"], [4533, 4534, ["encoder"], "Method"], [4536, 4538, ["discriminator", "network"], "Method"], [4548, 4550, ["categorical", "generation"], "Task"], [4551, 4557, ["image", "-", "to", "-", "image", "translation"], "Task"], [4559, 4565, ["text", "-", "to", "-", "image", "synthesis"], "Task"], [2092, 2094, ["categorical", "generation"], "Task"], [2327, 2328, ["FID"], "Metric"], [3909, 3910, ["JSD"], "Metric"]]} {"doc_id": "1275125b71b999b7c49bc554c6ec404a764eb299", "text": ["document", ":", "HD", "-", "CNN", ":", "Hierarchical", "Deep", "Convolutional", "Neural", "Network", "for", "Large", "Scale", "Visual", "Recognition", "In", "image", "classification", ",", "visual", "separability", "between", "different", "object", "categories", "is", "highly", "uneven", ",", "and", "some", "categories", "are", "more", "difficult", "to", "distinguish", "than", "others", ".", "Such", "difficult", "categories", "demand", "more", "dedicated", "classifiers", ".", "However", ",", "existing", "deep", "convolutional", "neural", "networks", "(", "CNN", ")", "are", "trained", "as", "flat", "N", "-", "way", "classifiers", ",", "and", "few", "efforts", "have", "been", "made", "to", "leverage", "the", "hierarchical", "structure", "of", "categories", ".", "In", "this", "paper", ",", "we", "introduce", "hierarchical", "deep", "CNNs", "(", "HD", "-", "CNNs", ")", "by", "embedding", "deep", "CNNs", "into", "a", "category", "hierarchy", ".", "An", "HD", "-", "CNN", "separates", "easy", "classes", "using", "a", "coarse", "category", "classifier", "while", "distinguishing", "difficult", "classes", "using", "fine", "category", "classifiers", ".", "During", "HD", "-", "CNN", "training", ",", "component", "-", "wise", "pretraining", "is", "followed", "by", "global", "finetuning", "with", "a", "multinomial", "logistic", "loss", "regularized", "by", "a", "coarse", "category", "consistency", "term", ".", "In", "addition", ",", "conditional", "executions", "of", "fine", "category", "classifiers", "and", "layer", "parameter", "compression", "make", "HD", "-", "CNNs", "scalable", "for", "large", "-", "scale", "visual", "recognition", ".", "We", "achieve", "state", "-", "of", "-", "the", "-", "art", "results", "on", "both", "CIFAR100", "and", "large", "-", "scale", "ImageNet", "1000", "-", "class", "benchmark", "datasets", ".", "In", "our", "experiments", ",", "we", "build", "up", "three", "different", "HD", "-", "CNNs", "and", "they", "lower", "the", "top", "-", "1", "error", "of", "the", "standard", "CNNs", "by", ",", "and", ",", "respectively", ".", ".22", ".72", "section", ":", "Introduction", "Deep", "CNNs", "are", "well", "suited", "for", "large", "-", "scale", "learning", "based", "visual", "recognition", "tasks", "because", "of", "its", "highly", "scalable", "training", "algorithm", ",", "which", "only", "needs", "to", "cache", "a", "small", "chunk", "(", "mini", "-", "batch", ")", "of", "the", "potentially", "huge", "volume", "of", "training", "data", "during", "sequential", "scans", "(", "epochs", ")", ".", "They", "have", "achieved", "increasingly", "better", "performance", "in", "recent", "years", ".", "As", "datasets", "become", "bigger", "and", "the", "number", "of", "object", "categories", "becomes", "larger", ",", "one", "of", "the", "complications", "that", "come", "along", "is", "that", "visual", "separability", "between", "different", "object", "categories", "is", "highly", "uneven", ".", "Some", "categories", "are", "much", "harder", "to", "distinguish", "than", "others", ".", "Take", "the", "categories", "in", "CIFAR100", "as", "an", "example", ".", "It", "is", "easy", "to", "tell", "an", "Apple", "from", "a", "Bus", ",", "but", "harder", "to", "tell", "an", "Apple", "from", "an", "Orange", ".", "In", "fact", ",", "both", "Apples", "and", "Oranges", "belong", "to", "the", "same", "coarse", "category", "fruit", "and", "vegetables", "while", "Buses", "belong", "to", "another", "coarse", "category", "vehicles", "1", ",", "as", "defined", "within", "CIFAR100", ".", "Nonetheless", ",", "most", "deep", "CNN", "models", "nowadays", "are", "flat", "N", "-", "way", "classifiers", ",", "which", "share", "a", "set", "of", "fully", "connected", "layers", ".", "This", "makes", "us", "wonder", "whether", "such", "a", "flat", "structure", "is", "adequate", "for", "distinguishing", "all", "the", "difficult", "categories", ".", "A", "very", "natural", "and", "intuitive", "alternative", "organizes", "classifiers", "in", "a", "hierarchical", "manner", "according", "to", "the", "divide", "-", "and", "-", "conquer", "strategy", ".", "Although", "hierarchical", "classification", "has", "been", "proven", "effective", "for", "conventional", "linear", "classifiers", ",", "few", "attempts", "have", "been", "made", "to", "exploit", "category", "hierarchies", "in", "deep", "CNN", "models", ".", "Since", "deep", "CNN", "models", "are", "large", "models", "themselves", ",", "organizing", "them", "hierarchically", "imposes", "the", "following", "challenges", ".", "First", ",", "instead", "of", "a", "handcrafted", "category", "hierarchy", ",", "how", "can", "we", "learn", "such", "a", "category", "hierarchy", "from", "the", "training", "data", "itself", "so", "that", "cascaded", "inferences", "in", "a", "hierarchical", "classifier", "will", "not", "degrade", "the", "overall", "accuracy", "while", "dedicated", "fine", "category", "classifiers", "exist", "for", "hard", "-", "to", "-", "distinguish", "categories", "?", "Second", ",", "a", "hierarchical", "CNN", "classifier", "consists", "of", "multiple", "CNN", "models", "at", "different", "levels", ".", "How", "can", "we", "leverage", "the", "commonalities", "among", "these", "models", "and", "effectively", "train", "them", "all", "?", "Third", ",", "it", "would", "also", "be", "slower", "and", "more", "memory", "-", "consuming", "to", "run", "a", "hierarchical", "CNN", "classifier", "on", "a", "novel", "testing", "image", ".", "How", "can", "we", "alleviate", "such", "limitations", "?", "In", "this", "paper", ",", "we", "propose", "a", "generic", "and", "principled", "hierarchical", "architecture", ",", "Hierarchical", "Deep", "Convolutional", "Neural", "Network", "(", "HD", "-", "CNN", ")", ",", "that", "decomposes", "an", "image", "classification", "task", "into", "two", "steps", ".", "An", "HD", "-", "CNN", "first", "uses", "a", "coarse", "category", "CNN", "classifier", "to", "separate", "easy", "classes", "from", "one", "another", ".", "More", "challenging", "classes", "are", "routed", "downstream", "to", "fine", "category", "classifiers", "that", "focus", "on", "confusing", "classes", ".", "We", "adopt", "a", "module", "design", "principle", "and", "every", "HD", "-", "CNN", "is", "built", "upon", "a", "building", "block", "CNN", ".", "The", "building", "block", "can", "be", "chosen", "to", "be", "any", "of", "the", "currently", "top", "ranked", "single", "CNNs", ".", "Thus", "HD", "-", "CNNs", "can", "always", "benefit", "from", "the", "progress", "of", "single", "CNN", "design", ".", "An", "HD", "-", "CNN", "follows", "the", "coarse", "-", "to", "-", "fine", "classification", "paradigm", "and", "probabilistically", "integrates", "predictions", "from", "the", "fine", "category", "classifiers", ".", "Compared", "with", "the", "building", "block", "CNN", ",", "the", "corresponding", "HD", "-", "CNN", "can", "achieve", "lower", "error", "at", "the", "cost", "of", "a", "manageable", "increase", "in", "memory", "footprint", "and", "classification", "time", ".", "In", "summary", ",", "this", "paper", "has", "the", "following", "contributions", ".", "First", ",", "we", "introduce", "a", "novel", "hierarchical", "architecture", ",", "called", "HD", "-", "CNN", ",", "for", "image", "classification", ".", "Second", ",", "we", "develop", "a", "scheme", "for", "learning", "the", "two", "-", "level", "organization", "of", "coarse", "and", "fine", "categories", ",", "and", "demonstrate", "various", "components", "of", "an", "HD", "-", "CNN", "can", "be", "independently", "pretrained", ".", "The", "complete", "HD", "-", "CNN", "is", "further", "fine", "-", "tuned", "using", "a", "multinomial", "logistic", "loss", "regularized", "by", "a", "coarse", "category", "consistency", "term", ".", "Third", ",", "we", "make", "the", "HD", "-", "CNN", "scalable", "by", "compressing", "the", "layer", "parameters", "and", "conditionally", "executing", "the", "fine", "category", "classifiers", ".", "We", "have", "performed", "evaluations", "on", "the", "medium", "-", "scale", "CIFAR100", "dataset", "and", "the", "large", "-", "scale", "ImageNet", "1000", "-", "class", "dataset", ",", "and", "our", "method", "has", "achieved", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "both", "of", "them", ".", "section", ":", "Related", "Work", "Our", "work", "is", "inspired", "by", "progresses", "in", "CNN", "design", "and", "efforts", "on", "integrating", "a", "category", "hierarchy", "with", "linear", "classifiers", ".", "The", "main", "novelty", "of", "our", "method", "is", "a", "new", "scalable", "HD", "-", "CNN", "architecture", "that", "integrates", "a", "category", "hierarchy", "with", "deep", "CNNs", ".", "subsection", ":", "Convolutional", "Neural", "Networks", "CNN", "-", "based", "models", "hold", "state", "-", "of", "-", "the", "-", "art", "performance", "in", "various", "computer", "vision", "tasks", ",", "including", "image", "classifcation", ",", "object", "detection", ",", "and", "image", "parsing", ".", "Recently", ",", "there", "has", "been", "considerable", "interest", "in", "enhancing", "CNN", "components", ",", "including", "pooling", "layers", ",", "activation", "units", ",", "and", "nonlinear", "layers", ".", "These", "enhancements", "either", "improve", "CNN", "training", ",", "or", "expand", "the", "network", "learning", "capacity", ".", "This", "work", "boosts", "CNN", "performance", "from", "an", "orthogonal", "angle", "and", "does", "not", "redesign", "a", "specific", "part", "within", "any", "existing", "CNN", "model", ".", "Instead", ",", "we", "design", "a", "novel", "generic", "hierarchical", "architecture", "that", "uses", "an", "existing", "CNN", "model", "as", "a", "building", "block", ".", "We", "embed", "multiple", "building", "blocks", "into", "a", "larger", "hierarchical", "deep", "CNN", ".", "subsection", ":", "Category", "Hierarchy", "for", "Visual", "Recognition", "In", "visual", "recognition", ",", "there", "is", "a", "vast", "literature", "exploiting", "category", "hierarchical", "structures", ".", "For", "classification", "with", "a", "large", "number", "of", "classes", "using", "linear", "classifiers", ",", "a", "common", "strategy", "is", "to", "build", "a", "hierarchy", "or", "taxonomy", "of", "classifiers", "so", "that", "the", "number", "of", "classifiers", "evaluated", "given", "a", "testing", "image", "scales", "sub", "-", "linearly", "in", "the", "number", "of", "classes", ".", "The", "hierarchy", "can", "be", "either", "predefined", "or", "learnt", "by", "top", "-", "down", "and", "bottom", "-", "up", "approaches", ".", "In", ",", "the", "predefined", "category", "hierarchy", "of", "ImageNet", "dataset", "is", "utilized", "to", "achieve", "the", "trade", "-", "offs", "between", "classification", "accuracy", "and", "specificity", ".", "In", ",", "a", "hierarchical", "label", "tree", "is", "constructed", "to", "probabilistically", "combine", "predictions", "from", "leaf", "nodes", ".", "Such", "hierarchical", "classifier", "achieves", "significant", "speedup", "at", "the", "cost", "of", "certain", "accuracy", "loss", ".", "One", "of", "the", "earliest", "attempts", "to", "introduce", "a", "category", "hierarchy", "in", "CNN", "-", "based", "methods", "is", "reported", "in", "but", "their", "main", "goal", "is", "transferring", "knowledge", "between", "classes", "to", "improve", "the", "results", "for", "classes", "with", "insufficient", "training", "examples", ".", "In", ",", "various", "label", "relations", "are", "encoded", "in", "a", "hierarchy", ".", "Improved", "accuracy", "is", "achieved", "only", "when", "a", "subset", "of", "training", "images", "are", "relabeled", "with", "internal", "nodes", "in", "the", "hierarchical", "class", "tree", ".", "They", "are", "not", "able", "to", "improve", "the", "accuracy", "in", "the", "original", "setting", "where", "all", "training", "images", "are", "labeled", "with", "leaf", "nodes", ".", "In", ",", "a", "hierarchy", "of", "CNNs", "is", "introduced", "but", "they", "experimented", "with", "only", "two", "coarse", "categories", "mainly", "due", "to", "scalability", "constraints", ".", "HD", "-", "CNN", "exploits", "the", "category", "hierarchy", "in", "a", "novel", "way", "that", "we", "embed", "deep", "CNNs", "into", "the", "hierarchy", "in", "a", "scalable", "manner", "and", "achieves", "superior", "classification", "results", "over", "the", "standard", "CNN", ".", "section", ":", "Overview", "of", "HD", "-", "CNN", "subsection", ":", "Notations", "The", "following", "notations", "are", "used", "below", ".", "A", "dataset", "consists", "of", "images", ".", "and", "denote", "the", "image", "data", "and", "label", ",", "respectively", ".", "There", "are", "fine", "categories", "of", "images", "in", "the", "dataset", ".", "We", "will", "learn", "a", "category", "hierarchy", "with", "coarse", "categories", ".", "subsection", ":", "HD", "-", "CNN", "Architecture", "HD", "-", "CNN", "is", "designed", "to", "mimic", "the", "structure", "of", "category", "hierarchy", "where", "fine", "categories", "are", "divided", "into", "coarse", "categories", "as", "in", "Fig", "[", "reference", "]", "(", "a", ")", ".", "It", "performs", "end", "-", "to", "-", "end", "classification", "as", "illustrated", "in", "Fig", "[", "reference", "]", "(", "b", ")", ".", "It", "mainly", "comprises", "four", "parts", ",", "namely", "shared", "layers", ",", "a", "single", "coarse", "category", "component", ",", "multiple", "fine", "category", "components", "and", "a", "single", "probabilistic", "averaging", "layer", ".", "On", "the", "left", "side", "of", "Fig", "[", "reference", "]", "(", "b", ")", "are", "the", "shared", "layers", ".", "They", "receive", "raw", "image", "pixel", "as", "input", "and", "extract", "low", "-", "level", "features", ".", "The", "configuration", "of", "shared", "layers", "are", "set", "to", "be", "the", "same", "as", "the", "preceding", "layers", "in", "the", "building", "block", "net", ".", "On", "the", "top", "of", "Fig", "[", "reference", "]", "(", "b", ")", "are", "independent", "layers", "of", "coarse", "category", "component", "which", "reuses", "the", "configuration", "of", "rear", "layers", "from", "the", "building", "block", "CNN", "and", "produces", "a", "fine", "prediction", "for", "an", "image", ".", "To", "produce", "a", "prediction", "over", "coarse", "categories", ",", "we", "append", "a", "fine", "-", "to", "-", "coarse", "aggregation", "layer", "which", "aggregates", "fine", "predictions", "into", "coarse", "ones", "when", "a", "mapping", "from", "fine", "categories", "to", "coarse", "ones", "is", "given", ".", "The", "coarse", "category", "probabilities", "serve", "two", "purposes", ".", "First", ",", "they", "are", "used", "as", "weights", "for", "combining", "the", "predictions", "made", "by", "fine", "category", "components", ".", "Second", ",", "when", "thresholded", ",", "they", "enable", "conditional", "executions", "of", "fine", "category", "components", "whose", "corresponding", "coarse", "probabilities", "are", "sufficiently", "large", ".", "In", "the", "bottom", "-", "right", "of", "Fig", "[", "reference", "]", "(", "b", ")", "are", "independent", "layers", "of", "a", "set", "of", "fine", "category", "classifiers", ",", "each", "of", "which", "makes", "fine", "category", "predictions", ".", "As", "each", "fine", "component", "only", "excels", "in", "classifying", "a", "small", "set", "of", "categories", ",", "they", "produce", "a", "fine", "prediction", "over", "a", "partial", "set", "of", "categories", ".", "The", "probabilities", "of", "other", "fine", "categories", "absent", "in", "the", "partial", "set", "are", "implicitly", "set", "to", "zero", ".", "The", "layer", "configurations", "are", "mostly", "copied", "from", "the", "building", "block", "CNN", "except", "that", "in", "the", "final", "classification", "layer", "the", "number", "of", "filters", "is", "set", "to", "be", "the", "size", "of", "partial", "set", "instead", "of", "the", "full", "categories", ".", "Both", "coarse", "category", "component", "and", "fine", "category", "components", "share", "common", "layers", ".", "The", "reason", "is", "three", "-", "fold", ".", "First", ",", "it", "is", "shown", "in", "that", "preceding", "layers", "in", "deep", "networks", "response", "to", "class", "-", "agnostic", "low", "-", "level", "features", "such", "as", "corners", "and", "edges", ",", "while", "rear", "layers", "extract", "more", "class", "-", "specific", "features", "such", "as", "dog", "face", "and", "bird", "’s", "legs", ".", "Since", "low", "-", "level", "features", "are", "useful", "for", "both", "coarse", "and", "fine", "classification", "tasks", ",", "we", "allow", "the", "preceding", "layers", "to", "be", "shared", "by", "both", "coarse", "and", "fine", "components", ".", "Second", ",", "it", "reduces", "both", "the", "total", "floating", "point", "operations", "and", "the", "memory", "footprint", "of", "network", "execution", ".", "Both", "are", "of", "practical", "significance", "to", "deploy", "HD", "-", "CNN", "in", "real", "applications", ".", "Last", "but", "not", "the", "least", ",", "it", "can", "decrease", "the", "number", "of", "HD", "-", "CNN", "parameters", "which", "is", "critical", "to", "the", "success", "of", "HD", "-", "CNN", "training", ".", "On", "the", "right", "side", "of", "Fig", "[", "reference", "]", "(", "b", ")", "is", "the", "probabilistic", "averaging", "layer", "which", "receives", "fine", "category", "predictions", "as", "well", "as", "coarse", "category", "prediction", "and", "produces", "a", "weighted", "average", "as", "the", "final", "prediction", ".", "where", "is", "the", "probability", "of", "coarse", "category", "for", "the", "image", "predicted", "by", "the", "coarse", "category", "component", ".", "is", "the", "fine", "category", "prediction", "made", "by", "the", "fine", "category", "component", ".", "We", "stress", "that", "both", "coarse", "and", "fine", "category", "components", "reuse", "the", "layer", "configurations", "from", "the", "building", "block", "CNN", ".", "This", "flexible", "modular", "design", "allows", "us", "to", "choose", "the", "best", "module", "CNN", "as", "the", "building", "block", ",", "depending", "on", "the", "task", "at", "hand", ".", "section", ":", "Learning", "a", "Category", "Hierarchy", "Our", "goal", "of", "building", "a", "category", "hierarchy", "is", "grouping", "confusing", "fine", "categories", "into", "the", "same", "coarse", "category", "for", "which", "a", "dedicated", "fine", "category", "classifier", "will", "be", "trained", ".", "We", "employ", "a", "top", "-", "down", "approach", "to", "learn", "the", "hierarchy", "from", "the", "training", "data", ".", "We", "randomly", "sample", "a", "held", "-", "out", "set", "of", "images", "with", "balanced", "class", "distribution", "from", "the", "training", "set", ".", "The", "rest", "of", "the", "training", "set", "is", "used", "to", "train", "a", "building", "block", "net", ".", "We", "obtain", "a", "confusion", "matrix", "by", "evaluating", "the", "net", "on", "the", "held", "-", "out", "set", ".", "A", "distance", "matrix", "is", "derived", "as", "and", "its", "diagonal", "entries", "are", "set", "to", "be", "zero", ".", "is", "further", "transformed", "by", "to", "be", "symmetric", ".", "The", "entry", "measures", "how", "easy", "it", "is", "to", "discriminate", "categories", "and", ".", "Spectral", "clustering", "is", "performed", "on", "to", "cluster", "fine", "categories", "into", "coarse", "categories", ".", "The", "result", "is", "a", "two", "-", "level", "category", "hierarchy", "representing", "a", "many", "-", "to", "-", "one", "mapping", "from", "fine", "to", "coarse", "categories", ".", "Here", ",", "the", "coarse", "categories", "are", "disjoint", ".", "Overlapping", "Coarse", "Categories", "With", "disjoint", "coarse", "categories", ",", "the", "overall", "classification", "depends", "heavily", "on", "the", "coarse", "category", "classifier", ".", "If", "an", "image", "is", "routed", "to", "an", "incorrect", "fine", "category", "classifier", ",", "then", "the", "mistake", "can", "not", "be", "corrected", "as", "the", "probability", "of", "ground", "truth", "label", "is", "implicitly", "set", "to", "zero", "there", ".", "Removing", "the", "separability", "constraint", "between", "coarse", "categories", "can", "make", "the", "HD", "-", "CNN", "less", "dependent", "on", "the", "coarse", "category", "classifier", ".", "Therefore", ",", "we", "add", "more", "fine", "categories", "to", "the", "coarse", "categories", ".", "For", "a", "certain", "fine", "classifier", ",", "we", "prefer", "to", "add", "those", "fine", "categories", "that", "are", "likely", "to", "be", "misclassfied", "into", "the", "coarse", "category", ".", "Therefore", ",", "we", "estimate", "the", "likelihood", "that", "an", "image", "in", "fine", "category", "is", "misclassified", "into", "a", "coarse", "category", "on", "the", "held", "-", "out", "set", ".", "is", "the", "coarse", "category", "probability", "which", "is", "obtained", "by", "aggregating", "fine", "category", "probabilities", "according", "to", "the", "mapping", ":", ".", "We", "threshold", "the", "likelihood", "using", "a", "parametric", "variable", "and", "add", "to", "the", "partial", "set", "all", "fine", "categories", "such", "that", ".", "Note", "that", "each", "branching", "component", "gives", "a", "full", "set", "prediction", "when", "and", "a", "disjoint", "set", "prediction", "when", ".", "With", "overlapping", "coarse", "categories", ",", "the", "category", "hierarchy", "mapping", "is", "extended", "to", "be", "a", "many", "-", "to", "-", "many", "mapping", "and", "the", "coarse", "category", "predictions", "are", "updated", "accordingly", ".", "Note", "the", "sum", "of", "exceeds", "and", "hence", "we", "perform", "normalization", ".", "The", "use", "of", "overlapping", "coarse", "categories", "was", "also", "shown", "to", "be", "useful", "for", "hierarchical", "linear", "classifiers", ".", "section", ":", "HD", "-", "CNN", "Training", "As", "we", "embed", "fine", "category", "components", "into", "HD", "-", "CNN", ",", "the", "number", "of", "parameters", "in", "rear", "layers", "grows", "linearly", "in", "the", "number", "of", "coarse", "categories", ".", "Given", "the", "same", "amount", "of", "training", "data", ",", "this", "increases", "the", "training", "complexity", "and", "the", "risk", "of", "over", "-", "fitting", ".", "On", "the", "other", "hand", ",", "the", "training", "images", "within", "the", "stochastic", "gradient", "descent", "mini", "-", "batch", "are", "probabilistically", "routed", "to", "different", "fine", "category", "components", ".", "It", "requires", "larger", "mini", "-", "batch", "to", "ensure", "parameter", "gradients", "in", "the", "fine", "category", "components", "are", "estimated", "by", "a", "sufficiently", "large", "number", "of", "training", "samples", ".", "Large", "training", "mini", "-", "batch", "both", "increases", "the", "training", "memory", "footprint", "and", "slows", "down", "the", "training", "process", ".", "Therefore", ",", "we", "decompose", "the", "HD", "-", "CNN", "training", "into", "multiple", "steps", "instead", "of", "training", "the", "complete", "HD", "-", "CNN", "from", "scratch", "as", "outlined", "in", "Algorithm", "[", "reference", "]", ".", "subsection", ":", "Pretraining", "HD", "-", "CNN", "We", "sequentially", "pretrain", "the", "coarse", "category", "component", "and", "fine", "category", "components", ".", "subsubsection", ":", "Initializing", "the", "Coarse", "Category", "Component", "We", "first", "pretrain", "a", "building", "block", "CNN", "using", "the", "training", "set", ".", "As", "both", "the", "preceding", "and", "rear", "layers", "in", "coarse", "category", "component", "resemble", "the", "layers", "in", "the", "building", "block", "CNN", ",", "we", "copy", "the", "weights", "of", "into", "coarse", "category", "component", "for", "initialization", "purpose", ".", "subsubsection", ":", "Pretraining", "the", "Rear", "Layers", "of", "Fine", "Category", "Components", "Fine", "category", "components", "can", "be", "independently", "pretrained", "in", "parallel", ".", "Each", "should", "specialize", "in", "classifying", "fine", "categories", "within", "the", "coarse", "category", ".", "Therefore", ",", "the", "pretraining", "of", "each", "only", "uses", "images", "from", "the", "coarse", "category", ".", "The", "shared", "preceding", "layers", "are", "already", "initialized", "and", "kept", "fixed", "in", "this", "stage", ".", "For", "each", ",", "we", "initialize", "all", "the", "rear", "layers", "except", "the", "last", "convolutional", "layer", "by", "copying", "the", "learned", "parameters", "from", "the", "pretrained", "model", ".", "[", "!", "t", "]", "HD", "-", "CNN", "training", "algorithm", "[", "1", "]", "HD", "-", "CNN", "Training", "Step", "1", ":", "Pretrain", "HD", "-", "CNN", "Step", "1.1", ":", "Initialize", "coarse", "category", "component", "Step", "1.2", ":", "Pretrain", "fine", "category", "components", "Step", "2", ":", "Fine", "-", "tune", "the", "complete", "HD", "-", "CNN", "subsection", ":", "Fine", "-", "tuning", "HD", "-", "CNN", "After", "both", "coarse", "category", "component", "and", "fine", "category", "components", "are", "properly", "pretrained", ",", "we", "fine", "-", "tune", "the", "complete", "HD", "-", "CNN", ".", "Once", "the", "category", "hierarchy", "as", "well", "as", "the", "associated", "mapping", "is", "learnt", ",", "each", "fine", "category", "component", "focuses", "on", "classifying", "a", "fixed", "subset", "of", "fine", "categories", ".", "During", "fine", "-", "tuning", ",", "the", "semantics", "of", "coarse", "categories", "predicted", "by", "the", "coarse", "category", "component", "should", "be", "kept", "consistent", "with", "those", "associated", "with", "fine", "category", "components", ".", "Thus", "we", "add", "a", "coarse", "category", "consistency", "term", "to", "regularize", "the", "conventional", "multinomial", "logistic", "loss", ".", "Coarse", "category", "consistency", "The", "learnt", "fine", "-", "to", "-", "coarse", "category", "mapping", "provides", "a", "way", "to", "specify", "the", "target", "coarse", "category", "distribution", ".", "Specifically", ",", "is", "set", "to", "be", "the", "fraction", "of", "all", "the", "training", "images", "within", "the", "coarse", "category", "under", "the", "assumption", "the", "distribution", "over", "coarse", "categories", "across", "the", "training", "dataset", "is", "close", "to", "that", "within", "a", "training", "mini", "-", "batch", ".", "The", "final", "loss", "function", "we", "use", "for", "fine", "-", "tuning", "the", "HD", "-", "CNN", "is", "shown", "below", ".", "where", "is", "the", "size", "of", "training", "mini", "-", "batch", ".", "is", "a", "regularization", "constant", "and", "is", "set", "to", ".", "section", ":", "HD", "-", "CNN", "Testing", "As", "we", "add", "fine", "category", "components", "into", "the", "HD", "-", "CNN", ",", "the", "number", "of", "parameters", ",", "memory", "footprint", "and", "execution", "time", "in", "rear", "layers", ",", "all", "scale", "linearly", "in", "the", "number", "of", "coarse", "categories", ".", "To", "ensure", "HD", "-", "CNN", "is", "scalable", "for", "large", "-", "scale", "visual", "recognition", ",", "we", "develop", "conditional", "execution", "and", "layer", "parameter", "compression", "techniques", ".", "Conditional", "Execution", ".", "At", "test", "time", ",", "for", "a", "given", "image", ",", "it", "is", "not", "necessary", "to", "evaluate", "all", "fine", "category", "classifiers", "as", "most", "of", "them", "have", "insignificant", "weights", "as", "in", "Eqn", "[", "reference", "]", ".", "Their", "contributions", "to", "the", "final", "prediction", "are", "negligible", ".", "Conditional", "executions", "of", "the", "top", "weighted", "fine", "components", "can", "accelerate", "the", "HD", "-", "CNN", "classification", ".", "Therefore", ",", "we", "threshold", "using", "a", "parametric", "variable", "and", "reset", "to", "zero", "when", ".", "Those", "fine", "category", "classifiers", "with", "are", "not", "evaluated", ".", "Parameter", "Compression", ".", "In", "HD", "-", "CNN", ",", "the", "number", "of", "parameters", "in", "rear", "layers", "of", "fine", "category", "classifiers", "grows", "linearly", "in", "the", "number", "of", "coarse", "categories", ".", "Thus", "we", "compress", "the", "layer", "parameters", "at", "test", "time", "to", "reduce", "memory", "footprint", ".", "Specifically", ",", "we", "choose", "the", "Product", "Quantization", "approach", "to", "compress", "the", "parameter", "matrix", "by", "first", "partitioning", "it", "horizontally", "into", "segments", "of", "width", ".", ".", "Then", "-", "means", "clustering", "is", "used", "to", "cluster", "the", "rows", "in", ".", "By", "only", "storing", "the", "nearest", "cluster", "indices", "in", "a", "8", "-", "bit", "integer", "matrix", "and", "cluster", "centers", "in", "a", "single", "-", "precision", "floating", "number", "matrix", ",", "we", "can", "achieve", "a", "compression", "factor", ".", "The", "hyperparameters", "for", "parameter", "compression", "are", ".", "section", ":", "Experiments", "subsection", ":", "Overview", "We", "evaluate", "HD", "-", "CNN", "on", "the", "benchmark", "datasets", "CIFAR100", "and", "ImageNet", ".", "HD", "-", "CNN", "is", "implemented", "on", "the", "widely", "deployed", "Caffe", "software", ".", "The", "network", "is", "trained", "by", "back", "propagation", ".", "We", "run", "all", "the", "testing", "experiments", "on", "a", "single", "NVIDIA", "Tesla", "K40c", "card", ".", "subsection", ":", "CIFAR100", "Dataset", "The", "CIFAR100", "dataset", "consists", "of", "100", "classes", "of", "natural", "images", ".", "There", "are", "50", "K", "training", "images", "and", "10", "K", "testing", "images", ".", "We", "follow", "to", "preprocess", "the", "datasets", "(", "e.g.", "global", "contrast", "normalization", "and", "ZCA", "whitening", ")", ".", "Randomly", "cropped", "and", "flipped", "image", "patches", "of", "size", "are", "used", "for", "training", ".", "We", "adopt", "a", "NIN", "network", "with", "three", "stacked", "layers", ".", "We", "denote", "it", "as", "CIFAR100", "-", "NIN", "which", "will", "be", "the", "HD", "-", "CNN", "building", "block", ".", "Fine", "category", "components", "share", "preceding", "layers", "from", "conv1", "to", "pool1", "which", "accounts", "for", "of", "the", "total", "parameters", "and", "of", "the", "total", "floating", "point", "operations", ".", "The", "remaining", "layers", "are", "used", "as", "independent", "layers", ".", "For", "building", "the", "category", "hierarchy", ",", "we", "randomly", "choose", "10", "K", "images", "from", "the", "training", "set", "as", "held", "-", "out", "set", ".", "Fine", "categories", "within", "the", "same", "coarse", "categories", "are", "visually", "more", "similar", ".", "We", "pretrain", "the", "rear", "layers", "of", "fine", "category", "components", ".", "The", "initial", "learning", "rate", "is", "and", "it", "is", "decreased", "by", "a", "factor", "of", "10", "every", "6", "K", "iterations", ".", "Fine", "-", "tuning", "is", "performed", "for", "20", "K", "iterations", "with", "large", "mini", "-", "batches", "of", "size", "256", ".", "The", "initial", "learning", "rate", "is", "and", "is", "reduced", "by", "a", "factor", "of", "10", "once", "after", "10", "K", "iterations", ".", "For", "evaluation", ",", "we", "use", "10", "-", "view", "testing", ".", "We", "extract", "five", "patches", "(", "the", "4", "corner", "patches", "and", "the", "center", "patch", ")", "as", "well", "as", "their", "horizontal", "reflections", "and", "average", "their", "predictions", ".", "The", "CIFAR100", "-", "NIN", "net", "obtains", "testing", "error", ".", "Our", "HD", "-", "CNN", "achieves", "testing", "error", "of", "which", "improves", "the", "building", "block", "net", "by", ".", "Category", "hierarchy", ".", "During", "the", "construction", "of", "the", "category", "hierarchy", ",", "the", "number", "of", "coarse", "categories", "can", "be", "adjusted", "by", "the", "clustering", "algorithm", ".", "We", "can", "also", "make", "the", "coarse", "categories", "either", "disjoint", "or", "overlapping", "by", "varying", "the", "hyperparameter", ".", "Thus", "we", "investigate", "their", "impacts", "on", "the", "classification", "error", ".", "We", "experiment", "with", "5", ",", "9", ",", "14", "and", "19", "coarse", "categories", "and", "vary", "the", "value", "of", ".", "The", "best", "results", "are", "obtained", "with", "9", "overlapping", "coarse", "categories", "and", "as", "shown", "in", "Fig", "[", "reference", "]", "left", ".", "A", "histogram", "of", "fine", "category", "occurrences", "in", "9", "overlapping", "coarse", "categories", "is", "shown", "in", "Fig", "[", "reference", "]", "right", ".", "The", "optimal", "value", "of", "coarse", "category", "number", "and", "hyperparameter", "are", "dataset", "dependent", ",", "mainly", "affected", "by", "the", "inherent", "hierarchy", "within", "the", "categories", ".", ".30", ".18", "Shared", "layers", ".", "The", "use", "of", "shared", "layers", "makes", "both", "computational", "complexity", "and", "memory", "footprint", "of", "HD", "-", "CNN", "sublinear", "in", "the", "number", "of", "fine", "category", "classifiers", "when", "compared", "to", "the", "building", "block", "net", ".", "Our", "HD", "-", "CNN", "with", "9", "fine", "category", "classifiers", "based", "on", "CIFAR100", "-", "NIN", "consumes", "less", "than", "three", "times", "as", "much", "memory", "as", "the", "building", "block", "net", "without", "parameter", "compression", ".", "We", "also", "want", "to", "investigate", "the", "impact", "of", "the", "use", "of", "shared", "layers", "on", "the", "classification", "error", ",", "memory", "footprint", "and", "the", "net", "execution", "time", "(", "Table", "[", "reference", "]", ")", ".", "We", "build", "another", "HD", "-", "CNN", "where", "coarse", "category", "component", "and", "all", "fine", "category", "components", "use", "independent", "preceding", "layers", "initialized", "from", "a", "pretrained", "building", "block", "net", ".", "Under", "the", "single", "-", "view", "testing", "where", "only", "a", "central", "cropping", "is", "used", ",", "we", "observe", "a", "minor", "error", "increase", "from", "to", ".", "But", "using", "shared", "layers", "dramatically", "reduces", "the", "memory", "footprint", "from", "MB", "to", "MB", "and", "testing", "time", "from", "seconds", "to", "seconds", ".", "Conditional", "executions", ".", "By", "varying", "the", "hyperparameter", ",", "we", "can", "effectively", "affect", "the", "number", "of", "fine", "category", "components", "that", "will", "be", "executed", ".", "There", "is", "a", "trade", "-", "off", "between", "execution", "time", "and", "classification", "error", ".", "A", "larger", "value", "of", "leads", "to", "higher", "accuracy", "at", "the", "cost", "of", "executing", "more", "components", "for", "fine", "categorization", ".", "By", "enabling", "conditional", "executions", "with", "hyperparameter", ",", "we", "obtain", "a", "substantial", "X", "speed", "up", "with", "merely", "a", "minor", "increase", "in", "error", "from", "to", "(", "Table", "[", "reference", "]", ")", ".", "The", "testing", "time", "of", "HD", "-", "CNN", "is", "about", "times", "as", "much", "as", "that", "of", "the", "building", "block", "net", ".", "Parameter", "compression", ".", "As", "fine", "category", "CNNs", "have", "independent", "layers", "from", "conv2", "to", "cccp6", ",", "we", "compress", "them", "and", "reduce", "the", "memory", "footprint", "from", "MB", "to", "MB", "with", "a", "minor", "increase", "in", "error", "from", "to", ".", "Comparison", "with", "a", "strong", "baseline", ".", "As", "our", "HD", "-", "CNN", "memory", "footprint", "is", "about", "two", "times", "as", "much", "as", "the", "building", "block", "model", "(", "Table", "[", "reference", "]", ")", ",", "it", "is", "necessary", "to", "compare", "a", "stronger", "baseline", "of", "similar", "complexity", "with", "HD", "-", "CNN", ".", "We", "adapt", "CIFAR100", "-", "NIN", "and", "double", "the", "number", "of", "filters", "in", "all", "convolutional", "layers", "which", "accordingly", "increases", "the", "memory", "footprint", "by", "three", "times", ".", "We", "denote", "it", "as", "CIFAR100", "-", "NIN", "-", "double", "and", "obtain", "error", "which", "is", "lower", "than", "that", "of", "the", "building", "block", "net", "but", "is", "higher", "than", "that", "of", "HD", "-", "CNN", ".", "Comparison", "with", "model", "averaging", ".", "HD", "-", "CNN", "is", "fundamentally", "different", "from", "model", "averaging", ".", "In", "model", "averaging", ",", "all", "models", "are", "capable", "of", "classifying", "the", "full", "set", "of", "the", "categories", "and", "each", "one", "is", "trained", "independently", ".", "The", "main", "sources", "of", "their", "prediction", "differences", "are", "different", "initializations", ".", "In", "HD", "-", "CNN", ",", "each", "fine", "category", "classifier", "only", "excels", "at", "classifying", "a", "partial", "set", "of", "categories", ".", "To", "compare", "HD", "-", "CNN", "with", "model", "averaging", ",", "we", "independently", "train", "two", "CIFAR100", "-", "NIN", "networks", "and", "take", "their", "averaged", "prediction", "as", "the", "final", "prediction", ".", "We", "obtain", "an", "error", "of", ",", "which", "is", "about", "higher", "than", "that", "of", "HD", "-", "CNN", "(", "Table", "[", "reference", "]", ")", ".", "Note", "that", "HD", "-", "CNN", "is", "orthogonal", "to", "the", "model", "averaging", "and", "an", "ensemble", "of", "HD", "-", "CNN", "networks", "can", "further", "improve", "the", "performance", ".", "Coarse", "category", "consistency", ".", "To", "verify", "the", "effectiveness", "of", "coarse", "category", "consistency", "term", "in", "our", "loss", "function", "(", "[", "reference", "]", ")", ",", "we", "fine", "-", "tune", "a", "HD", "-", "CNN", "using", "the", "traditional", "multinomial", "logistic", "loss", "function", ".", "The", "testing", "error", "is", ",", "which", "is", "higher", "than", "that", "of", "a", "HD", "-", "CNN", "fine", "-", "tuned", "with", "coarse", "category", "consistency", "(", "Table", "[", "reference", "]", ")", ".", "Comparison", "with", "state", "-", "of", "-", "the", "-", "art", ".", "Our", "HD", "-", "CNN", "improves", "on", "the", "current", "two", "best", "methods", "and", "by", "and", "respectively", "and", "sets", "new", "state", "-", "of", "-", "the", "-", "art", "results", "on", "CIFAR100", "(", "Table", "[", "reference", "]", ")", ".", "subsection", ":", "ImageNet", "1000", "-", "class", "Dataset", "The", "ILSVRC", "-", "2012", "ImageNet", "dataset", "consists", "of", "about", "million", "training", "images", ",", "validation", "images", ".", "To", "demonstrate", "the", "generality", "of", "HD", "-", "CNN", ",", "we", "experiment", "with", "two", "different", "building", "block", "nets", ".", "In", "both", "cases", ",", "HD", "-", "CNNs", "achieve", "significantly", "lower", "testing", "errors", "than", "the", "building", "block", "nets", ".", "subsubsection", ":", "Network", "-", "In", "-", "Network", "Building", "Block", "Net", "(", "a", ")", "(", "b", ")", "(", "c", ")", "(", "d", ")", "(", "e", ")", "(", "f", ")", "(", "g", ")", "We", "choose", "a", "public", "4", "-", "layer", "NIN", "net", "as", "our", "first", "building", "block", "as", "it", "has", "greatly", "reduced", "number", "of", "parameters", "compared", "to", "AlexNet", "but", "similar", "error", "rates", ".", "It", "is", "denoted", "as", "ImageNet", "-", "NIN", ".", "In", "HD", "-", "CNN", ",", "various", "components", "share", "preceding", "layers", "from", "conv1", "to", "pool3", "which", "account", "for", "of", "the", "total", "parameters", "and", "of", "the", "total", "floating", "point", "operations", ".", "We", "follow", "the", "training", "and", "testing", "protocols", "as", "in", ".", "Original", "images", "are", "resized", "to", ".", "Randomly", "cropped", "and", "horizontally", "reflected", "patches", "are", "used", "for", "training", ".", "At", "test", "time", ",", "the", "net", "makes", "a", "10", "-", "view", "averaged", "prediction", ".", "We", "train", "ImageNet", "-", "NIN", "for", "45", "epochs", ".", "The", "top", "-", "1", "and", "top", "-", "5", "errors", "are", "and", ".", "To", "build", "the", "category", "hierarchy", ",", "we", "take", "100", "K", "training", "images", "as", "the", "held", "-", "out", "set", "and", "find", "89", "overlapping", "coarse", "categories", ".", "Each", "fine", "category", "CNN", "is", "fine", "-", "tuned", "for", "40", "K", "iterations", "while", "the", "initial", "learning", "rate", "is", "decreased", "by", "a", "factor", "of", "every", "15", "K", "iterations", ".", "Fine", "-", "tuning", "the", "complete", "HD", "-", "CNN", "is", "not", "performed", "as", "the", "required", "mini", "-", "batch", "size", "is", "significantly", "higher", "than", "that", "for", "the", "building", "block", "net", ".", "Nevertheless", ",", "we", "still", "achieve", "top", "-", "1", "and", "top", "-", "5", "errors", "of", "and", "and", "improve", "the", "building", "block", "net", "by", "and", ",", "respectively", "(", "Table", "[", "reference", "]", ")", ".", "The", "class", "-", "wise", "top", "-", "5", "error", "improvement", "over", "the", "building", "block", "net", "is", "shown", "in", "Fig", "[", "reference", "]", "left", ".", "Case", "studies", "We", "want", "to", "investigate", "how", "HD", "-", "CNN", "corrects", "the", "mistakes", "made", "by", "the", "building", "block", "net", ".", "In", "Fig", "[", "reference", "]", ",", "we", "collect", "four", "testing", "cases", ".", "In", "the", "first", "case", ",", "the", "building", "block", "net", "fails", "to", "predict", "the", "label", "of", "the", "tiny", "hermit", "crab", "in", "the", "top", "5", "guesses", ".", "In", "HD", "-", "CNN", ",", "two", "coarse", "categories", "and", "receive", "most", "of", "the", "coarse", "probability", "mass", ".", "The", "fine", "category", "component", "specializes", "in", "classifying", "crab", "breeds", "and", "strongly", "suggests", "the", "ground", "truth", "label", ".", "By", "combining", "the", "predictions", "from", "the", "top", "fine", "category", "classifiers", ",", "the", "HD", "-", "CNN", "predicts", "hermit", "crab", "as", "the", "most", "probable", "label", ".", "In", "the", "second", "case", ",", "the", "ImageNet", "-", "NIN", "confuses", "the", "ground", "truth", "hand", "blower", "with", "other", "objects", "of", "close", "shapes", "and", "appearances", ",", "such", "as", "plunger", "and", "barbell", ".", "For", "HD", "-", "CNN", ",", "the", "coarse", "category", "component", "is", "also", "not", "confident", "about", "which", "coarse", "category", "the", "object", "belongs", "to", "and", "thus", "assigns", "even", "probability", "mass", "to", "the", "top", "coarse", "categories", ".", "For", "the", "top", "3", "fine", "category", "classifiers", ",", "strongly", "predicts", "ground", "truth", "label", "while", "the", "other", "two", "and", "rank", "the", "ground", "truth", "label", "at", "the", "2nd", "and", "4th", "place", "respectively", ".", "Overall", ",", "the", "HD", "-", "CNN", "ranks", "the", "ground", "truth", "label", "at", "the", "1st", "place", ".", "This", "demonstrates", "HD", "-", "CNN", "needs", "to", "rely", "on", "multiple", "fine", "category", "classifiers", "to", "make", "correct", "predictions", "for", "difficult", "cases", ".", "Overlapping", "coarse", "categories", ".To", "investigate", "the", "impact", "of", "overlapping", "coarse", "categories", "on", "the", "classification", ",", "we", "train", "another", "HD", "-", "CNN", "with", "89", "fine", "category", "classifiers", "using", "disjoint", "coarse", "categories", ".", "It", "achieves", "top", "-", "1", "and", "top", "-", "5", "errors", "of", "and", "respectively", ",", "which", "is", "higher", "than", "those", "of", "the", "HD", "-", "CNN", "using", "overlapping", "coarse", "category", "hierarchy", "by", "and", "(", "Table", "[", "reference", "]", ")", ".", "Conditional", "executions", ".", "By", "varying", "the", "hyperparameter", ",", "we", "can", "control", "the", "number", "of", "fine", "category", "components", "that", "will", "be", "executed", ".", "There", "is", "a", "trade", "-", "off", "between", "execution", "time", "and", "classification", "error", "as", "shown", "in", "Fig", "[", "reference", "]", "right", ".", "A", "larger", "value", "of", "leads", "to", "lower", "error", "at", "the", "cost", "of", "more", "executed", "fine", "category", "components", ".", "By", "enabling", "conditional", "executions", "with", "hyperparameter", ",", "we", "obtain", "a", "substantial", "X", "speed", "up", "with", "merely", "a", "minor", "increase", "of", "single", "-", "view", "testing", "top", "-", "5", "error", "from", "to", "(", "Table", "[", "reference", "]", ")", ".", "With", "such", "speedup", ",", "the", "HD", "-", "CNN", "testing", "time", "is", "less", "than", "times", "as", "much", "as", "that", "of", "the", "building", "block", "net", ".", "Parameter", "compression", ".", "We", "compress", "independent", "layers", "conv4", "and", "cccp7", "as", "they", "account", "for", "of", "the", "parameters", "in", "ImageNet", "-", "NIN", ".", "Their", "parameter", "matrices", "are", "of", "size", "and", "and", "we", "use", "compression", "hyperparameters", "and", ".", "The", "compression", "factors", "are", "and", ".", "The", "compression", "decreases", "the", "memory", "footprint", "from", "MB", "to", "MB", "and", "merely", "increases", "the", "top", "-", "5", "error", "from", "to", "under", "single", "-", "view", "testing", "(", "Table", "[", "reference", "]", ")", ".", ".23", ".27", "Comparison", "with", "model", "averaging", ".", "As", "the", "HD", "-", "CNN", "memory", "footprint", "is", "about", "three", "times", "as", "much", "as", "the", "building", "block", "net", ",", "we", "independently", "train", "three", "ImageNet", "-", "NIN", "nets", "and", "average", "their", "predictions", ".", "We", "obtain", "top", "-", "5", "error", "which", "is", "lower", "than", "the", "building", "block", "but", "is", "higher", "than", "that", "of", "HD", "-", "CNN", "(", "Table", "[", "reference", "]", ")", ".", "subsubsection", ":", "VGG", "-", "16", "-", "layer", "Building", "Block", "Net", "The", "second", "building", "block", "net", "we", "use", "is", "a", "16", "-", "layer", "CNN", "from", ".", "We", "denote", "it", "as", "ImageNet", "-", "VGG", "-", "16", "-", "layer", ".", "The", "layers", "from", "conv1_1", "to", "pool4", "are", "shared", "and", "they", "account", "for", "of", "the", "total", "parameters", "and", "of", "the", "total", "floating", "number", "operations", ".", "The", "remaining", "layers", "are", "used", "as", "independent", "layers", "in", "coarse", "and", "fine", "category", "classifiers", ".", "We", "follow", "the", "training", "and", "testing", "protocols", "as", "in", ".", "For", "training", ",", "we", "first", "sample", "a", "size", "from", "the", "range", "and", "resize", "the", "image", "so", "that", "the", "length", "of", "short", "edge", "is", ".", "Then", "a", "randomly", "cropped", "and", "flipped", "patch", "of", "size", "is", "used", "for", "training", ".", "For", "testing", ",", "dense", "evaluation", "is", "performed", "on", "three", "scales", "and", "the", "averaged", "prediction", "is", "used", "as", "the", "final", "prediction", ".", "Please", "refer", "to", "for", "more", "training", "and", "testing", "details", ".", "On", "ImageNet", "validation", "set", ",", "ImageNet", "-", "VGG", "-", "16", "-", "layer", "achieves", "top", "-", "1", "and", "top", "-", "5", "errors", "and", "respectively", ".", "We", "build", "a", "category", "hierarchy", "with", "84", "overlapping", "coarse", "categories", ".", "We", "implement", "multi", "-", "GPU", "training", "on", "Caffe", "by", "exploiting", "data", "parallelism", "and", "train", "the", "fine", "category", "classifiers", "on", "two", "NVIDIA", "Tesla", "K40c", "cards", ".", "The", "initial", "learning", "rate", "is", "and", "it", "is", "decreased", "by", "a", "factor", "of", "10", "every", "4", "K", "iterations", ".", "HD", "-", "CNN", "fine", "-", "tuning", "is", "not", "performed", ".", "Due", "to", "large", "memory", "footprint", "of", "the", "building", "block", "net", "(", "Table", "[", "reference", "]", ")", ",", "the", "HD", "-", "CNN", "with", "84", "fine", "category", "classifiers", "can", "not", "fit", "into", "the", "memory", "directly", ".", "Therefore", ",", "we", "compress", "the", "parameters", "in", "layers", "fc6", "and", "fc7", "as", "they", "account", "for", "over", "of", "the", "parameters", ".", "Parameter", "matrices", "in", "fc6", "and", "fc7", "are", "of", "size", "and", ".", "Their", "compression", "hyperparameters", "are", "and", ".", "The", "compression", "factors", "are", "and", "respectively", ".", "The", "HD", "-", "CNN", "obtains", "top", "-", "1", "and", "top", "-", "5", "errors", "and", "on", "ImageNet", "validation", "set", "and", "improves", "over", "ImageNet", "-", "VGG", "-", "16", "-", "layer", "by", "and", "respectively", ".", "Comparison", "with", "state", "-", "of", "-", "the", "-", "art", ".", "Currently", ",", "the", "two", "best", "nets", "on", "ImageNet", "dataset", "are", "GoogLeNet", "(", "Table", "[", "reference", "]", ")", "and", "VGG", "19", "-", "layer", "network", ".", "Using", "multi", "-", "scale", "multi", "-", "crop", "testing", ",", "a", "single", "GoogLeNet", "net", "achieves", "top", "-", "5", "error", ".", "With", "multi", "-", "scale", "dense", "evaluation", ",", "a", "single", "VGG", "19", "-", "layer", "net", "obtains", "top", "-", "1", "and", "top", "-", "5", "errors", "and", "and", "improves", "top", "-", "5", "error", "of", "GoogLeNet", "by", ".", "Our", "HD", "-", "CNN", "decreases", "top", "-", "1", "and", "top", "-", "5", "errors", "of", "VGG", "19", "-", "layer", "net", "by", "and", "respectively", ".", "Furthermore", ",", "HD", "-", "CNN", "slightly", "outperforms", "the", "results", "of", "averaging", "the", "predictions", "from", "VGG", "-", "16", "-", "layer", "and", "VGG", "-", "19", "-", "layer", "nets", ".", "section", ":", "Conclusions", "and", "Future", "Work", "We", "demonstrated", "that", "HD", "-", "CNN", "is", "a", "flexible", "deep", "CNN", "architecture", "to", "improve", "over", "existing", "deep", "CNN", "models", ".", "We", "showed", "this", "empirically", "on", "both", "CIFAR", "-", "100", "and", "Image", "-", "Net", "datasets", "using", "three", "different", "building", "block", "nets", ".", "As", "part", "of", "future", "work", ",", "we", "plan", "to", "extend", "HD", "-", "CNN", "architectures", "to", "those", "with", "more", "than", "2", "hierarchical", "levels", "and", "also", "verify", "our", "empirical", "results", "in", "a", "theoretical", "framework", ".", "bibliography", ":", "References"], "entity": [[6, 11, ["Hierarchical", "Deep", "Convolutional", "Neural", "Network"], "Method"], [12, 16, ["Large", "Scale", "Visual", "Recognition"], "Task"], [17, 19, ["image", "classification"], "Task"], [20, 22, ["visual", "separability"], "Task"], [47, 48, ["classifiers"], "Method"], [52, 56, ["deep", "convolutional", "neural", "networks"], "Method"], [57, 59, ["CNN", ")"], "Method"], [62, 67, ["flat", "N", "-", "way", "classifiers"], "Method"], [88, 91, ["hierarchical", "deep", "CNNs"], "Method"], [92, 95, ["HD", "-", "CNNs"], "Method"], [122, 125, ["fine", "category", "classifiers"], "Method"], [132, 136, ["component", "-", "wise", "pretraining"], "Method"], [139, 141, ["global", "finetuning"], "Method"], [143, 146, ["multinomial", "logistic", "loss"], "Method"], [157, 163, ["conditional", "executions", "of", "fine", "category", "classifiers"], "Method"], [164, 167, ["layer", "parameter", "compression"], "Method"], [168, 171, ["HD", "-", "CNNs"], "Method"], [173, 178, ["large", "-", "scale", "visual", "recognition"], "Task"], [191, 192, ["CIFAR100"], "Material"], [212, 215, ["HD", "-", "CNNs"], "Method"], [219, 223, ["top", "-", "1", "error"], "Metric"], [226, 227, ["CNNs"], "Method"], [244, 252, ["large", "-", "scale", "learning", "based", "visual", "recognition", "tasks"], "Task"], [257, 259, ["training", "algorithm"], "Method"], [344, 345, ["CIFAR100"], "Material"], [399, 400, ["CIFAR100"], "Material"], [410, 414, ["N", "-", "way", "classifiers"], "Method"], [449, 450, ["classifiers"], "Method"], [457, 463, ["divide", "-", "and", "-", "conquer", "strategy"], "Method"], [465, 467, ["hierarchical", "classification"], "Method"], [473, 475, ["linear", "classifiers"], "Method"], [531, 533, ["cascaded", "inferences"], "Method"], [535, 537, ["hierarchical", "classifier"], "Method"], [542, 543, ["accuracy"], "Metric"], [545, 548, ["fine", "category", "classifiers"], "Method"], [625, 630, ["generic", "and", "principled", "hierarchical", "architecture"], "Method"], [631, 636, ["Hierarchical", "Deep", "Convolutional", "Neural", "Network"], "Method"], [678, 681, ["fine", "category", "classifiers"], "Method"], [690, 693, ["module", "design", "principle"], "Method"], [721, 722, ["CNNs"], "Method"], [724, 727, ["HD", "-", "CNNs"], "Method"], [757, 760, ["fine", "category", "classifiers"], "Method"], [776, 777, ["error"], "Metric"], [785, 787, ["memory", "footprint"], "Metric"], [788, 790, ["classification", "time"], "Metric"], [807, 809, ["hierarchical", "architecture"], "Method"], [816, 818, ["image", "classification"], "Task"], [864, 867, ["multinomial", "logistic", "loss"], "Method"], [893, 896, ["fine", "category", "classifiers"], "Method"], [958, 960, ["linear", "classifiers"], "Method"], [986, 989, ["Convolutional", "Neural", "Networks"], "Method"], [1004, 1007, ["computer", "vision", "tasks"], "Task"], [1009, 1011, ["image", "classifcation"], "Task"], [1012, 1014, ["object", "detection"], "Task"], [1016, 1018, ["image", "parsing"], "Task"], [1032, 1034, ["pooling", "layers"], "Method"], [1035, 1037, ["activation", "units"], "Method"], [1039, 1041, ["nonlinear", "layers"], "Method"], [1052, 1055, ["network", "learning", "capacity"], "Method"], [1059, 1060, ["CNN"], "Method"], [1084, 1087, ["generic", "hierarchical", "architecture"], "Method"], [1095, 1097, ["building", "block"], "Method"], [1115, 1117, ["Visual", "Recognition"], "Task"], [1118, 1120, ["visual", "recognition"], "Task"], [1132, 1133, ["classification"], "Task"], [1140, 1142, ["linear", "classifiers"], "Method"], [1152, 1155, ["taxonomy", "of", "classifiers"], "Method"], [1160, 1161, ["classifiers"], "Method"], [1185, 1193, ["top", "-", "down", "and", "bottom", "-", "up", "approaches"], "Method"], [1212, 1214, ["classification", "accuracy"], "Metric"], [1215, 1216, ["specificity"], "Metric"], [1234, 1236, ["hierarchical", "classifier"], "Method"], [1244, 1246, ["accuracy", "loss"], "Metric"], [1255, 1257, ["category", "hierarchy"], "Method"], [1297, 1298, ["accuracy"], "Metric"], [1325, 1326, ["accuracy"], "Metric"], [1388, 1389, ["classification"], "Metric"], [1393, 1394, ["CNN"], "Method"], [1486, 1492, ["end", "-", "to", "-", "end", "classification"], "Task"], [1510, 1512, ["shared", "layers"], "Method"], [1520, 1523, ["fine", "category", "components"], "Method"], [1526, 1529, ["probabilistic", "averaging", "layer"], "Method"], [1615, 1617, ["fine", "prediction"], "Task"], [1679, 1682, ["fine", "category", "components"], "Method"], [1690, 1696, ["conditional", "executions", "of", "fine", "category", "components"], "Method"], [1724, 1727, ["fine", "category", "classifiers"], "Method"], [1732, 1735, ["fine", "category", "predictions"], "Task"], [1738, 1740, ["fine", "component"], "Method"], [1795, 1797, ["classification", "layer"], "Method"], [1821, 1824, ["fine", "category", "components"], "Method"], [1845, 1847, ["deep", "networks"], "Method"], [1863, 1865, ["rear", "layers"], "Method"], [1889, 1890, ["coarse"], "Task"], [1891, 1894, ["fine", "classification", "tasks"], "Task"], [1917, 1920, ["floating", "point", "operations"], "Metric"], [1922, 1924, ["memory", "footprint"], "Metric"], [1925, 1927, ["network", "execution"], "Task"], [1939, 1941, ["real", "applications"], "Task"], [1984, 1987, ["probabilistic", "averaging", "layer"], "Method"], [2001, 2003, ["weighted", "average"], "Method"], [2027, 2030, ["fine", "category", "prediction"], "Task"], [2033, 2036, ["fine", "category", "component"], "Method"], [2084, 2086, ["Category", "Hierarchy"], "Task"], [2091, 2093, ["category", "hierarchy"], "Task"], [2107, 2110, ["fine", "category", "classifier"], "Method"], [2117, 2121, ["top", "-", "down", "approach"], "Method"], [2160, 2163, ["building", "block", "net"], "Method"], [2167, 2169, ["confusion", "matrix"], "Metric"], [2216, 2218, ["Spectral", "clustering"], "Method"], [2236, 2238, ["category", "hierarchy"], "Method"], [2270, 2271, ["classification"], "Task"], [2287, 2290, ["fine", "category", "classifier"], "Method"], [2348, 2350, ["fine", "classifier"], "Method"], [2436, 2438, ["branching", "component"], "Method"], [2457, 2460, ["category", "hierarchy", "mapping"], "Method"], [2489, 2490, ["normalization"], "Method"], [2504, 2507, ["hierarchical", "linear", "classifiers"], "Method"], [2517, 2520, ["fine", "category", "components"], "Method"], [2552, 2554, ["training", "complexity"], "Metric"], [2558, 2561, ["over", "-", "fitting"], "Task"], [2572, 2575, ["stochastic", "gradient", "descent"], "Method"], [2575, 2578, ["mini", "-", "batch"], "Method"], [2583, 2586, ["fine", "category", "components"], "Method"], [2614, 2618, ["training", "mini", "-", "batch"], "Method"], [2621, 2624, ["training", "memory", "footprint"], "Metric"], [2628, 2630, ["training", "process"], "Method"], [2675, 2678, ["fine", "category", "components"], "Method"], [2679, 2680, ["subsubsection"], "Method"], [2683, 2686, ["Coarse", "Category", "Component"], "Method"], [2728, 2730, ["initialization", "purpose"], "Task"], [2731, 2732, ["subsubsection"], "Method"], [2735, 2741, ["Rear", "Layers", "of", "Fine", "Category", "Components"], "Method"], [2741, 2744, ["Fine", "category", "components"], "Method"], [2812, 2814, ["pretrained", "model"], "Method"], [2849, 2852, ["fine", "category", "components"], "Method"], [2877, 2880, ["fine", "category", "components"], "Method"], [2908, 2911, ["fine", "category", "component"], "Method"], [2922, 2925, ["fine", "-", "tuning"], "Task"], [2945, 2948, ["fine", "category", "components"], "Method"], [2961, 2964, ["multinomial", "logistic", "loss"], "Method"], [2965, 2968, ["Coarse", "category", "consistency"], "Metric"], [3030, 3032, ["loss", "function"], "Metric"], [3074, 3077, ["fine", "category", "components"], "Method"], [3088, 3090, ["memory", "footprint"], "Metric"], [3091, 3093, ["execution", "time"], "Metric"], [3115, 3120, ["large", "-", "scale", "visual", "recognition"], "Task"], [3123, 3130, ["conditional", "execution", "and", "layer", "parameter", "compression", "techniques"], "Method"], [3131, 3133, ["Conditional", "Execution"], "Task"], [3150, 3153, ["fine", "category", "classifiers"], "Method"], [3172, 3173, ["prediction"], "Task"], [3181, 3184, ["weighted", "fine", "components"], "Method"], [3187, 3191, ["HD", "-", "CNN", "classification"], "Task"], [3207, 3210, ["fine", "category", "classifiers"], "Method"], [3215, 3217, ["Parameter", "Compression"], "Method"], [3231, 3234, ["fine", "category", "classifiers"], "Method"], [3262, 3265, ["Product", "Quantization", "approach"], "Method"], [3282, 3285, ["-", "means", "clustering"], "Method"], [3323, 3325, ["compression", "factor"], "Metric"], [3329, 3331, ["parameter", "compression"], "Task"], [3361, 3363, ["Caffe", "software"], "Method"], [3369, 3371, ["back", "propagation"], "Method"], [3425, 3427, ["ZCA", "whitening"], "Method"], [3445, 3447, ["NIN", "network"], "Method"], [3449, 3451, ["stacked", "layers"], "Method"], [3469, 3472, ["Fine", "category", "components"], "Method"], [3476, 3477, ["conv1"], "Method"], [3506, 3508, ["category", "hierarchy"], "Task"], [3540, 3546, ["rear", "layers", "of", "fine", "category", "components"], "Method"], [3549, 3551, ["learning", "rate"], "Metric"], [3566, 3569, ["Fine", "-", "tuning"], "Task"], [3586, 3588, ["learning", "rate"], "Metric"], [3608, 3612, ["10", "-", "view", "testing"], "Task"], [3644, 3646, ["testing", "error"], "Metric"], [3652, 3654, ["testing", "error"], "Metric"], [3658, 3661, ["building", "block", "net"], "Method"], [3684, 3686, ["clustering", "algorithm"], "Method"], [3710, 3712, ["classification", "error"], "Metric"], [3806, 3808, ["computational", "complexity"], "Metric"], [3809, 3811, ["memory", "footprint"], "Metric"], [3820, 3823, ["fine", "category", "classifiers"], "Method"], [3827, 3830, ["building", "block", "net"], "Method"], [3837, 3840, ["fine", "category", "classifiers"], "Method"], [3855, 3858, ["building", "block", "net"], "Method"], [3859, 3861, ["parameter", "compression"], "Method"], [3877, 3879, ["classification", "error"], "Metric"], [3880, 3882, ["memory", "footprint"], "Metric"], [3884, 3887, ["net", "execution", "time"], "Metric"], [3906, 3909, ["fine", "category", "components"], "Method"], [3911, 3913, ["preceding", "layers"], "Method"], [3916, 3920, ["pretrained", "building", "block", "net"], "Method"], [3923, 3927, ["single", "-", "view", "testing"], "Task"], [3951, 3953, ["memory", "footprint"], "Metric"], [3958, 3960, ["testing", "time"], "Metric"], [3995, 3997, ["execution", "time"], "Metric"], [3998, 4000, ["classification", "error"], "Metric"], [4008, 4009, ["accuracy"], "Metric"], [4017, 4019, ["fine", "categorization"], "Task"], [4025, 4026, ["hyperparameter"], "Method"], [4031, 4034, ["X", "speed", "up"], "Metric"], [4040, 4041, ["error"], "Metric"], [4051, 4053, ["testing", "time"], "Metric"], [4066, 4069, ["building", "block", "net"], "Method"], [4070, 4072, ["Parameter", "compression"], "Method"], [4083, 4084, ["cccp6"], "Method"], [4102, 4103, ["error"], "Metric"], [4127, 4130, ["building", "block", "model"], "Method"], [4147, 4148, ["complexity"], "Metric"], [4166, 4168, ["convolutional", "layers"], "Method"], [4172, 4174, ["memory", "footprint"], "Metric"], [4182, 4187, ["CIFAR100", "-", "NIN", "-", "double"], "Method"], [4189, 4190, ["error"], "Metric"], [4197, 4200, ["building", "block", "net"], "Method"], [4212, 4214, ["model", "averaging"], "Method"], [4222, 4224, ["model", "averaging"], "Method"], [4226, 4228, ["model", "averaging"], "Method"], [4265, 4268, ["fine", "category", "classifier"], "Method"], [4284, 4286, ["model", "averaging"], "Method"], [4308, 4309, ["error"], "Metric"], [4337, 4339, ["model", "averaging"], "Method"], [4353, 4356, ["Coarse", "category", "consistency"], "Metric"], [4368, 4370, ["loss", "function"], "Method"], [4387, 4391, ["multinomial", "logistic", "loss", "function"], "Method"], [4393, 4395, ["testing", "error"], "Metric"], [4458, 4459, ["CIFAR100"], "Material"], [4503, 4506, ["building", "block", "nets"], "Method"], [4511, 4514, ["HD", "-", "CNNs"], "Method"], [4517, 4519, ["testing", "errors"], "Metric"], [4521, 4524, ["building", "block", "nets"], "Method"], [4527, 4535, ["Network", "-", "In", "-", "Network", "Building", "Block", "Net"], "Method"], [4559, 4565, ["public", "4", "-", "layer", "NIN", "net"], "Method"], [4580, 4581, ["AlexNet"], "Method"], [4583, 4585, ["error", "rates"], "Metric"], [4590, 4593, ["ImageNet", "-", "NIN"], "Method"], [4602, 4604, ["preceding", "layers"], "Method"], [4658, 4663, ["10", "-", "view", "averaged", "prediction"], "Method"], [4666, 4669, ["ImageNet", "-", "NIN"], "Method"], [4678, 4682, ["top", "-", "5", "errors"], "Metric"], [4725, 4727, ["learning", "rate"], "Metric"], [4738, 4741, ["Fine", "-", "tuning"], "Task"], [4763, 4766, ["building", "block", "net"], "Method"], [4800, 4808, ["class", "-", "wise", "top", "-", "5", "error", "improvement"], "Metric"], [4810, 4813, ["building", "block", "net"], "Method"], [4838, 4841, ["building", "block", "net"], "Method"], [4860, 4863, ["building", "block", "net"], "Method"], [4897, 4900, ["fine", "category", "component"], "Method"], [4902, 4905, ["classifying", "crab", "breeds"], "Task"], [4920, 4923, ["fine", "category", "classifiers"], "Method"], [4943, 4946, ["ImageNet", "-", "NIN"], "Method"], [5004, 5007, ["fine", "category", "classifiers"], "Method"], [5057, 5060, ["fine", "category", "classifiers"], "Method"], [5081, 5082, ["classification"], "Task"], [5091, 5094, ["fine", "category", "classifiers"], "Method"], [5166, 5168, ["execution", "time"], "Metric"], [5169, 5171, ["classification", "error"], "Metric"], [5187, 5188, ["error"], "Metric"], [5194, 5197, ["fine", "category", "components"], "Method"], [5203, 5204, ["hyperparameter"], "Method"], [5209, 5212, ["X", "speed", "up"], "Metric"], [5220, 5226, ["view", "testing", "top", "-", "5", "error"], "Metric"], [5255, 5258, ["building", "block", "net"], "Method"], [5259, 5261, ["Parameter", "compression"], "Method"], [5264, 5267, ["independent", "layers", "conv4"], "Method"], [5268, 5269, ["cccp7"], "Method"], [5277, 5280, ["ImageNet", "-", "NIN"], "Method"], [5296, 5298, ["compression", "factors"], "Metric"], [5302, 5303, ["compression"], "Method"], [5305, 5307, ["memory", "footprint"], "Metric"], [5315, 5319, ["top", "-", "5", "error"], "Metric"], [5337, 5339, ["model", "averaging"], "Method"], [5355, 5358, ["building", "block", "net"], "Method"], [5374, 5378, ["top", "-", "5", "error"], "Metric"], [5383, 5385, ["building", "block"], "Method"], [5403, 5411, ["VGG", "-", "16", "-", "layer", "Building", "Block", "Net"], "Method"], [5413, 5416, ["building", "block", "net"], "Method"], [5430, 5437, ["ImageNet", "-", "VGG", "-", "16", "-", "layer"], "Method"], [5528, 5530, ["dense", "evaluation"], "Task"], [5537, 5539, ["averaged", "prediction"], "Method"], [5561, 5568, ["ImageNet", "-", "VGG", "-", "16", "-", "layer"], "Method"], [5569, 5572, ["top", "-", "1"], "Metric"], [5573, 5577, ["top", "-", "5", "errors"], "Metric"], [5593, 5597, ["multi", "-", "GPU", "training"], "Method"], [5598, 5599, ["Caffe"], "Method"], [5606, 5609, ["fine", "category", "classifiers"], "Method"], [5611, 5615, ["NVIDIA", "Tesla", "K40c", "cards"], "Method"], [5618, 5620, ["learning", "rate"], "Metric"], [5652, 5655, ["building", "block", "net"], "Method"], [5668, 5671, ["fine", "category", "classifiers"], "Method"], [5686, 5688, ["layers", "fc6"], "Method"], [5689, 5690, ["fc7"], "Method"], [5702, 5703, ["fc6"], "Method"], [5704, 5705, ["fc7"], "Method"], [5711, 5713, ["compression", "hyperparameters"], "Method"], [5717, 5719, ["compression", "factors"], "Metric"], [5744, 5751, ["ImageNet", "-", "VGG", "-", "16", "-", "layer"], "Method"], [5783, 5788, ["VGG", "19", "-", "layer", "network"], "Method"], [5790, 5797, ["multi", "-", "scale", "multi", "-", "crop", "testing"], "Task"], [5800, 5802, ["GoogLeNet", "net"], "Method"], [5803, 5807, ["top", "-", "5", "error"], "Metric"], [5811, 5814, ["scale", "dense", "evaluation"], "Task"], [5817, 5822, ["VGG", "19", "-", "layer", "net"], "Method"], [5834, 5838, ["top", "-", "5", "error"], "Metric"], [5839, 5840, ["GoogLeNet"], "Method"], [5856, 5861, ["VGG", "19", "-", "layer", "net"], "Method"], [5879, 5891, ["VGG", "-", "16", "-", "layer", "and", "VGG", "-", "19", "-", "layer", "nets"], "Method"], [5924, 5927, ["CIFAR", "-", "100"], "Material"], [5935, 5938, ["building", "block", "nets"], "Method"], [4, 5, ["CNN"], "Method"], [99, 100, ["CNNs"], "Method"], [108, 109, ["CNN"], "Method"], [114, 115, ["coarse"], "Task"], [129, 130, ["CNN"], "Method"], [149, 150, ["coarse"], "Task"], [239, 240, ["CNNs"], "Method"], [381, 382, ["coarse"], "Task"], [391, 392, ["coarse"], "Task"], [405, 406, ["CNN"], "Method"], [487, 488, ["CNN"], "Method"], [492, 493, ["CNN"], "Method"], [561, 562, ["CNN"], "Method"], [566, 567, ["CNN"], "Method"], [603, 604, ["CNN"], "Method"], [639, 640, ["CNN"], "Method"], [645, 647, ["image", "classification"], "Task"], [655, 656, ["CNN"], "Method"], [659, 660, ["coarse"], "Task"], [661, 662, ["CNN"], "Method"], [697, 698, ["CNN"], "Method"], [704, 705, ["CNN"], "Method"], [735, 736, ["CNN"], "Method"], [741, 742, ["CNN"], "Method"], [744, 745, ["coarse"], "Task"], [766, 767, ["CNN"], "Method"], [772, 773, ["CNN"], "Method"], [813, 814, ["CNN"], "Method"], [833, 834, ["coarse"], "Task"], [846, 847, ["CNN"], "Method"], [856, 857, ["CNN"], "Method"], [870, 871, ["coarse"], "Task"], [882, 883, ["CNN"], "Method"], [906, 907, ["CIFAR100"], "Material"], [948, 949, ["CNN"], "Method"], [973, 974, ["CNN"], "Method"], [982, 983, ["CNNs"], "Method"], [989, 990, ["CNN"], "Method"], [1028, 1029, ["CNN"], "Method"], [1046, 1047, ["CNN"], "Method"], [1075, 1076, ["CNN"], "Method"], [1091, 1092, ["CNN"], "Method"], [1108, 1109, ["CNN"], "Method"], [1258, 1259, ["CNN"], "Method"], [1345, 1346, ["CNNs"], "Method"], [1354, 1355, ["coarse"], "Task"], [1364, 1365, ["CNN"], "Method"], [1377, 1378, ["CNNs"], "Method"], [1401, 1402, ["CNN"], "Method"], [1445, 1446, ["coarse"], "Task"], [1452, 1453, ["CNN"], "Method"], [1456, 1457, ["CNN"], "Method"], [1472, 1473, ["coarse"], "Task"], [1515, 1516, ["coarse"], "Task"], [1597, 1598, ["coarse"], "Task"], [1611, 1612, ["CNN"], "Method"], [1626, 1627, ["coarse"], "Task"], [1636, 1637, ["coarse"], "Task"], [1644, 1645, ["coarse"], "Task"], [1653, 1654, ["coarse"], "Task"], [1659, 1660, ["coarse"], "Task"], [1698, 1699, ["coarse"], "Task"], [1789, 1790, ["CNN"], "Method"], [1817, 1818, ["coarse"], "Task"], [1905, 1906, ["coarse"], "Task"], [1937, 1938, ["CNN"], "Method"], [1956, 1957, ["CNN"], "Method"], [1967, 1968, ["CNN"], "Method"], [1995, 1996, ["coarse"], "Task"], [2013, 2014, ["coarse"], "Task"], [2021, 2022, ["coarse"], "Task"], [2041, 2042, ["coarse"], "Task"], [2054, 2055, ["CNN"], "Method"], [2067, 2068, ["CNN"], "Method"], [2101, 2102, ["coarse"], "Task"], [2226, 2227, ["coarse"], "Task"], [2249, 2250, ["coarse"], "Task"], [2255, 2256, ["coarse"], "Task"], [2265, 2266, ["coarse"], "Task"], [2275, 2276, ["coarse"], "Task"], [2317, 2318, ["coarse"], "Task"], [2324, 2325, ["CNN"], "Method"], [2329, 2330, ["coarse"], "Task"], [2342, 2343, ["coarse"], "Task"], [2366, 2367, ["coarse"], "Task"], [2385, 2386, ["coarse"], "Task"], [2396, 2397, ["coarse"], "Task"], [2453, 2454, ["coarse"], "Task"], [2473, 2474, ["coarse"], "Task"], [2495, 2496, ["coarse"], "Task"], [2512, 2513, ["CNN"], "Method"], [2523, 2524, ["CNN"], "Method"], [2538, 2539, ["coarse"], "Task"], [2638, 2639, ["CNN"], "Method"], [2650, 2651, ["CNN"], "Method"], [2666, 2667, ["CNN"], "Method"], [2671, 2672, ["coarse"], "Task"], [2692, 2693, ["CNN"], "Method"], [2706, 2707, ["coarse"], "Task"], [2716, 2717, ["CNN"], "Method"], [2724, 2725, ["coarse"], "Task"], [2760, 2761, ["coarse"], "Task"], [2774, 2775, ["coarse"], "Task"], [2821, 2822, ["CNN"], "Method"], [2829, 2830, ["CNN"], "Method"], [2837, 2838, ["CNN"], "Method"], [2842, 2843, ["coarse"], "Task"], [2862, 2863, ["CNN"], "Method"], [2870, 2871, ["CNN"], "Method"], [2873, 2874, ["coarse"], "Task"], [2892, 2893, ["CNN"], "Method"], [2929, 2930, ["coarse"], "Task"], [2934, 2935, ["coarse"], "Task"], [2953, 2954, ["coarse"], "Task"], [2974, 2975, ["coarse"], "Task"], [2984, 2985, ["coarse"], "Task"], [3003, 3004, ["coarse"], "Task"], [3011, 3012, ["coarse"], "Task"], [3041, 3042, ["CNN"], "Method"], [3069, 3070, ["CNN"], "Method"], [3081, 3082, ["CNN"], "Method"], [3104, 3105, ["coarse"], "Task"], [3111, 3112, ["CNN"], "Method"], [3221, 3222, ["CNN"], "Method"], [3240, 3241, ["coarse"], "Task"], [3343, 3344, ["CNN"], "Method"], [3348, 3349, ["CIFAR100"], "Material"], [3354, 3355, ["CNN"], "Method"], [3388, 3389, ["CIFAR100"], "Material"], [3391, 3392, ["CIFAR100"], "Material"], [3456, 3457, ["CIFAR100"], "Material"], [3465, 3466, ["CNN"], "Method"], [3530, 3531, ["coarse"], "Task"], [3639, 3640, ["CIFAR100"], "Material"], [3650, 3651, ["CNN"], "Method"], [3677, 3678, ["coarse"], "Task"], [3692, 3693, ["coarse"], "Task"], [3723, 3724, ["coarse"], "Task"], [3739, 3740, ["coarse"], "Task"], [3760, 3761, ["coarse"], "Task"], [3775, 3776, ["coarse"], "Task"], [3814, 3815, ["CNN"], "Method"], [3834, 3835, ["CNN"], "Method"], [3842, 3843, ["CIFAR100"], "Material"], [3899, 3900, ["CNN"], "Method"], [3901, 3902, ["coarse"], "Task"], [4056, 4057, ["CNN"], "Method"], [4076, 4077, ["CNNs"], "Method"], [4116, 4117, ["CNN"], "Method"], [4151, 4152, ["CNN"], "Method"], [4155, 4156, ["CIFAR100"], "Material"], [4208, 4209, ["CNN"], "Method"], [4217, 4218, ["CNN"], "Method"], [4262, 4263, ["CNN"], "Method"], [4282, 4283, ["CNN"], "Method"], [4291, 4292, ["CIFAR100"], "Material"], [4320, 4321, ["CNN"], "Method"], [4332, 4333, ["CNN"], "Method"], [4345, 4346, ["CNN"], "Method"], [4362, 4363, ["coarse"], "Task"], [4383, 4384, ["CNN"], "Method"], [4406, 4407, ["CNN"], "Method"], [4411, 4412, ["coarse"], "Task"], [4434, 4435, ["CNN"], "Method"], [4496, 4497, ["CNN"], "Method"], [4597, 4598, ["CNN"], "Method"], [4674, 4677, ["top", "-", "1"], "Metric"], [4707, 4708, ["coarse"], "Task"], [4713, 4714, ["CNN"], "Method"], [4745, 4746, ["CNN"], "Method"], [4772, 4775, ["top", "-", "1"], "Metric"], [4776, 4780, ["top", "-", "5", "errors"], "Metric"], [4831, 4832, ["CNN"], "Method"], [4882, 4883, ["CNN"], "Method"], [4885, 4886, ["coarse"], "Task"], [4892, 4893, ["coarse"], "Task"], [4927, 4928, ["CNN"], "Method"], [4970, 4971, ["CNN"], "Method"], [4973, 4974, ["coarse"], "Task"], [4982, 4983, ["coarse"], "Task"], [4997, 4998, ["coarse"], "Task"], [5036, 5037, ["CNN"], "Method"], [5051, 5052, ["CNN"], "Method"], [5069, 5070, ["coarse"], "Task"], [5077, 5078, ["coarse"], "Task"], [5088, 5089, ["CNN"], "Method"], [5096, 5097, ["coarse"], "Task"], [5101, 5104, ["top", "-", "1"], "Metric"], [5105, 5109, ["top", "-", "5", "errors"], "Metric"], [5122, 5123, ["CNN"], "Method"], [5125, 5126, ["coarse"], "Task"], [5242, 5243, ["CNN"], "Method"], [5344, 5345, ["CNN"], "Method"], [5363, 5366, ["ImageNet", "-", "NIN"], "Method"], [5393, 5394, ["CNN"], "Method"], [5423, 5424, ["CNN"], "Method"], [5471, 5472, ["coarse"], "Task"], [5588, 5589, ["coarse"], "Task"], [5637, 5638, ["CNN"], "Method"], [5665, 5666, ["CNN"], "Method"], [5726, 5727, ["CNN"], "Method"], [5728, 5731, ["top", "-", "1"], "Metric"], [5732, 5736, ["top", "-", "5", "errors"], "Metric"], [5823, 5826, ["top", "-", "1"], "Metric"], [5827, 5831, ["top", "-", "5", "errors"], "Metric"], [5845, 5846, ["CNN"], "Method"], [5847, 5850, ["top", "-", "1"], "Metric"], [5851, 5855, ["top", "-", "5", "errors"], "Metric"], [5869, 5870, ["CNN"], "Method"], [5903, 5904, ["CNN"], "Method"], [5908, 5909, ["CNN"], "Method"], [5915, 5916, ["CNN"], "Method"], [5951, 5952, ["CNN"], "Method"]]} {"doc_id": "12806c298e01083a79db77927530367d85939907", "text": ["document", ":", "An", "Empirical", "Evaluation", "of", "Deep", "Learning", "on", "Highway", "Driving", "Numerous", "groups", "have", "applied", "a", "variety", "of", "deep", "learning", "techniques", "to", "computer", "vision", "problems", "in", "highway", "perception", "scenarios", ".", "In", "this", "paper", ",", "we", "presented", "a", "number", "of", "empirical", "evaluations", "of", "recent", "deep", "learning", "advances", ".", "Computer", "vision", ",", "combined", "with", "deep", "learning", ",", "has", "the", "potential", "to", "bring", "about", "a", "relatively", "inexpensive", ",", "robust", "solution", "to", "autonomous", "driving", ".", "To", "prepare", "deep", "learning", "for", "industry", "uptake", "and", "practical", "applications", ",", "neural", "networks", "will", "require", "large", "data", "sets", "that", "represent", "all", "possible", "driving", "environments", "and", "scenarios", ".", "We", "collect", "a", "large", "data", "set", "of", "highway", "data", "and", "apply", "deep", "learning", "and", "computer", "vision", "algorithms", "to", "problems", "such", "as", "car", "and", "lane", "detection", ".", "We", "show", "how", "existing", "convolutional", "neural", "networks", "(", "CNNs", ")", "can", "be", "used", "to", "perform", "lane", "and", "vehicle", "detection", "while", "running", "at", "frame", "rates", "required", "for", "a", "real", "-", "time", "system", ".", "Our", "results", "lend", "credence", "to", "the", "hypothesis", "that", "deep", "learning", "holds", "promise", "for", "autonomous", "driving", ".", "section", ":", "Introduction", "Since", "the", "DARPA", "Grand", "Challenges", "for", "autonomous", "vehicles", ",", "there", "has", "been", "an", "explosion", "in", "applications", "and", "research", "for", "self", "-", "driving", "cars", ".", "Among", "the", "different", "environments", "for", "self", "-", "driving", "cars", ",", "highway", "and", "urban", "roads", "are", "on", "opposite", "ends", "of", "the", "spectrum", ".", "In", "general", ",", "highways", "tend", "to", "be", "more", "predictable", "and", "orderly", ",", "with", "road", "surfaces", "typically", "well", "-", "maintained", "and", "lanes", "well", "-", "marked", ".", "In", "contrast", ",", "residential", "or", "urban", "driving", "environments", "feature", "a", "much", "higher", "degree", "of", "unpredictability", "with", "many", "generic", "objects", ",", "inconsistent", "lane", "-", "markings", ",", "and", "elaborate", "traffic", "flow", "patterns", ".", "The", "relative", "regularity", "and", "structure", "of", "highways", "has", "facilitated", "some", "of", "the", "first", "practical", "applications", "of", "autonomous", "driving", "technology", ".", "Many", "automakers", "have", "begun", "pursuing", "highway", "auto", "-", "pilot", "solutions", "designed", "to", "mitigate", "driver", "stress", "and", "fatigue", "and", "to", "provide", "additional", "safety", "features", ";", "for", "example", ",", "certain", "advanced", "-", "driver", "assistance", "systems", "(", "ADAS", ")", "can", "both", "keep", "cars", "within", "their", "lane", "and", "perform", "front", "-", "view", "car", "detection", ".", "Currently", ",", "the", "human", "drivers", "retain", "liability", "and", ",", "as", "such", ",", "must", "keep", "their", "hands", "on", "the", "steering", "wheel", "and", "prepare", "to", "control", "the", "vehicle", "in", "the", "event", "of", "any", "unexpected", "obstacle", "or", "catastrophic", "incident", ".", "Financial", "considerations", "contribute", "to", "a", "substantial", "performance", "gap", "between", "commercially", "available", "auto", "-", "pilot", "systems", "and", "fully", "self", "-", "driving", "cars", "developed", "by", "Google", "and", "others", ".", "Namely", ",", "today", "’s", "self", "-", "driving", "cars", "are", "equipped", "with", "expensive", "but", "critical", "sensors", ",", "such", "as", "LIDAR", ",", "radar", "and", "high", "-", "precision", "GPS", "coupled", "with", "highly", "detailed", "maps", ".", "In", "today", "’s", "production", "-", "grade", "autonomous", "vehicles", ",", "critical", "sensors", "include", "radar", ",", "sonar", ",", "and", "cameras", ".", "Long", "-", "range", "vehicle", "detection", "typically", "requires", "radar", ",", "while", "nearby", "car", "detection", "can", "be", "solved", "with", "sonar", ".", "Computer", "vision", "can", "play", "an", "important", "a", "role", "in", "lane", "detection", "as", "well", "as", "redundant", "object", "detection", "at", "moderate", "distances", ".", "Radar", "works", "reasonably", "well", "for", "detecting", "vehicles", ",", "but", "has", "difficulty", "distinguishing", "between", "different", "metal", "objects", "and", "thus", "can", "register", "false", "positives", "on", "objects", "such", "as", "tin", "cans", ".", "Also", ",", "radar", "provides", "little", "orientation", "information", "and", "has", "a", "higher", "variance", "on", "the", "lateral", "position", "of", "objects", ",", "making", "the", "localization", "difficult", "on", "sharp", "bends", ".", "The", "utility", "of", "sonar", "is", "both", "compromised", "at", "high", "speeds", "and", ",", "even", "at", "slow", "speeds", ",", "is", "limited", "to", "a", "working", "distance", "of", "about", "meters", ".", "Compared", "to", "sonar", "and", "radar", ",", "cameras", "generate", "a", "richer", "set", "of", "features", "at", "a", "fraction", "of", "the", "cost", ".", "By", "advancing", "computer", "vision", ",", "cameras", "could", "serve", "as", "a", "reliable", "redundant", "sensor", "for", "autonomous", "driving", ".", "Despite", "its", "potential", ",", "computer", "vision", "has", "yet", "to", "assume", "a", "significant", "role", "in", "today", "’s", "self", "-", "driving", "cars", ".", "Classic", "computer", "vision", "techniques", "simply", "have", "not", "provided", "the", "robustness", "required", "for", "production", "grade", "automotives", ";", "these", "techniques", "require", "intensive", "hand", "engineering", ",", "road", "modeling", ",", "and", "special", "case", "handling", ".", "Considering", "the", "seemingly", "infinite", "number", "of", "specific", "driving", "situations", ",", "environments", ",", "and", "unexpected", "obstacles", ",", "the", "task", "of", "scaling", "classic", "computer", "vision", "to", "robust", ",", "human", "-", "level", "performance", "would", "prove", "monumental", "and", "is", "likely", "to", "be", "unrealistic", ".", "Deep", "learning", ",", "or", "neural", "networks", ",", "represents", "an", "alternative", "approach", "to", "computer", "vision", ".", "It", "shows", "considerable", "promise", "as", "a", "solution", "to", "the", "shortcomings", "of", "classic", "computer", "vision", ".", "Recent", "progress", "in", "the", "field", "has", "advanced", "the", "feasibility", "of", "deep", "learning", "applications", "to", "solve", "complex", ",", "real", "-", "world", "problems", ";", "industry", "has", "responded", "by", "increasing", "uptake", "of", "such", "technology", ".", "Deep", "learning", "is", "data", "centric", ",", "requiring", "heavy", "computation", "but", "minimal", "hand", "-", "engineering", ".", "In", "the", "last", "few", "years", ",", "an", "increase", "in", "available", "storage", "and", "compute", "capabilities", "have", "enabled", "deep", "learning", "to", "achieve", "success", "in", "supervised", "perception", "tasks", ",", "such", "as", "image", "detection", ".", "A", "neural", "network", ",", "after", "training", "for", "days", "or", "even", "weeks", "on", "a", "large", "data", "set", ",", "can", "be", "capable", "of", "inference", "in", "real", "-", "time", "with", "a", "model", "size", "that", "is", "no", "larger", "than", "a", "few", "hundred", "MB", ".", "State", "-", "of", "-", "the", "-", "art", "neural", "networks", "for", "computer", "vision", "require", "very", "large", "training", "sets", "coupled", "with", "extensive", "networks", "capable", "of", "modeling", "such", "immense", "volumes", "of", "data", ".", "For", "example", ",", "the", "ILSRVC", "data", "-", "set", ",", "where", "neural", "networks", "achieve", "top", "results", ",", "contains", "million", "images", "in", "over", "categories", ".", "By", "using", "expensive", "existing", "sensors", "which", "are", "currently", "used", "for", "self", "-", "driving", "applications", ",", "such", "as", "LIDAR", "and", "mm", "-", "accurate", "GPS", ",", "and", "calibrating", "them", "with", "cameras", ",", "we", "can", "create", "a", "video", "data", "set", "containing", "labeled", "lane", "-", "markings", "and", "annotated", "vehicles", "with", "location", "and", "relative", "speed", ".", "By", "building", "a", "labeled", "data", "set", "in", "all", "types", "of", "driving", "situations", "(", "rain", ",", "snow", ",", "night", ",", "day", ",", "etc", ".", ")", ",", "we", "can", "evaluate", "neural", "networks", "on", "this", "data", "to", "determine", "if", "it", "is", "robust", "in", "every", "driving", "environment", "and", "situation", "for", "which", "we", "have", "training", "data", ".", "In", "this", "paper", ",", "we", "detail", "empirical", "evaluation", "on", "the", "data", "set", "we", "collect", ".", "In", "addition", ",", "we", "explain", "the", "neural", "network", "that", "we", "applied", "for", "detecting", "lanes", "and", "cars", ",", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "section", ":", "Related", "Work", "Recently", ",", "computer", "vision", "has", "been", "expected", "to", "player", "a", "larger", "role", "within", "autonomous", "driving", ".", "However", ",", "due", "to", "its", "history", "of", "relatively", "low", "precision", ",", "it", "is", "typically", "used", "in", "conjunction", "with", "either", "other", "sensors", "or", "other", "road", "models", ".", "Cho", "et", "al", ".", "uses", "multiple", "sensors", ",", "such", "as", "LIDAR", ",", "radar", ",", "and", "computer", "vision", "for", "object", "detection", ".", "They", "then", "fuse", "these", "sensors", "together", "in", "a", "Kalman", "filter", "using", "motion", "models", "on", "the", "objects", ".", "Held", "et", "al", ".", ",", "uses", "only", "a", "deformable", "parts", "based", "model", "on", "images", "to", "get", "the", "detections", ",", "then", "uses", "road", "models", "to", "filter", "out", "false", "positives", ".", "Carafii", "et", "al", ".", "uses", "a", "WaldBoost", "detector", "along", "with", "a", "tracker", "to", "generate", "pixel", "space", "detections", "in", "real", "time", ".", "Jazayeri", "et", "al", ".", "relies", "on", "temporal", "information", "of", "features", "for", "detection", ",", "and", "then", "filters", "out", "false", "positives", "with", "a", "front", "-", "view", "motion", "model", ".", "In", "contrast", "to", "these", "object", "detectors", ",", "we", "do", "not", "use", "any", "road", "or", "motion", "-", "based", "models", ";", "instead", "we", "rely", "only", "on", "the", "robustness", "of", "a", "neural", "network", "to", "make", "reasonable", "predictions", ".", "In", "addition", ",", "we", "currently", "do", "not", "rely", "on", "any", "temporal", "features", ",", "and", "the", "detector", "operates", "independently", "on", "single", "frames", "from", "a", "monocular", "camera", ".", "To", "make", "up", "for", "the", "lack", "of", "other", "sensors", ",", "which", "estimate", "object", "depth", ",", "we", "train", "the", "neural", "network", "to", "predict", "depth", "based", "on", "labels", "extracted", "from", "radar", "returns", ".", "Although", "the", "model", "only", "predicts", "a", "single", "depth", "value", "for", "each", "object", ",", "Eigen", "et", "al", ".", "have", "shown", "how", "a", "neural", "network", "can", "predict", "entire", "depth", "maps", "from", "single", "images", ".", "The", "network", "we", "train", "likely", "learns", "some", "model", "of", "the", "road", "for", "object", "detection", "and", "depth", "predictions", ",", "but", "it", "is", "never", "explicitly", "engineered", "and", "instead", "learns", "from", "the", "annotations", "alone", ".", "Before", "the", "wide", "spread", "adoption", "of", "Convolutional", "Neural", "Networks", "(", "CNNs", ")", "within", "computer", "vision", ",", "deformable", "parts", "based", "models", "were", "the", "most", "successful", "methods", "for", "detection", ".", "After", "the", "popular", "CNN", "model", "AlexNet", "was", "proposed", ",", "state", "-", "of", "-", "the", "-", "art", "detection", "shifted", "towards", "CNNs", "for", "feature", "extraction", ".", "Girshick", "et", "al", ".", "developed", "R", "-", "CNN", ",", "a", "two", "part", "system", "which", "used", "Selective", "Search", "to", "propose", "regions", "and", "AlexNet", "to", "classify", "them", ".", "R", "-", "CNN", "achieved", "state", "-", "of", "-", "the", "-", "art", "on", "Pascal", "by", "a", "large", "margin", ";", "however", ",", "due", "to", "its", "nearly", "classification", "queries", "and", "inefficient", "re", "-", "use", "of", "convolutions", ",", "it", "remains", "impractical", "for", "real", "-", "time", "implementations", ".", "Szegedy", "et", "al", ".", "presented", "a", "more", "scalable", "alternative", "to", "R", "-", "CNN", ",", "that", "relies", "on", "the", "CNN", "to", "propose", "higher", "quality", "regions", "compared", "to", "Selective", "Search", ".", "This", "reduces", "the", "number", "of", "region", "proposals", "down", "to", "as", "low", "as", "while", "keeping", "the", "mAP", "competitive", "with", "Selective", "Search", ".", "An", "even", "faster", "approach", "to", "image", "detection", "called", "Overfeat", "was", "presented", "by", "Sermanet", "et", "al", ".", ".", "By", "using", "a", "regular", "pattern", "of", "“", "region", "proposals", "”", ",", "Overfeat", "can", "efficiently", "reuse", "convolution", "computations", "from", "each", "layer", ",", "requiring", "only", "a", "single", "forward", "pass", "for", "inference", ".", "For", "our", "empirical", "evaluation", ",", "we", "use", "a", "straight", "-", "forward", "application", "of", "Overfeat", ",", "due", "to", "its", "efficiencies", ",", "and", "combine", "this", "with", "labels", "similar", "to", "the", "ones", "proposed", "by", "Szegedy", "et", "al", ".", ".", "We", "describe", "the", "model", "and", "similarities", "in", "the", "next", "section", ".", "section", ":", "Real", "Time", "Vehicle", "Detection", "Convolutional", "Neural", "Networks", "(", "CNNs", ")", "have", "had", "the", "largest", "success", "in", "image", "recognition", "in", "the", "past", "3", "years", ".", "From", "these", "image", "recognition", "systems", ",", "a", "number", "of", "detection", "networks", "were", "adapted", ",", "leading", "to", "further", "advances", "in", "image", "detection", ".", "While", "the", "improvements", "have", "been", "staggering", ",", "not", "much", "consideration", "had", "been", "given", "to", "the", "real", "-", "time", "detection", "performance", "required", "for", "some", "applications", ".", "In", "this", "paper", ",", "we", "present", "a", "detection", "system", "capable", "of", "operating", "at", "greater", "than", "Hz", "using", "nothing", "but", "a", "laptop", "GPU", ".", "Due", "to", "the", "requirements", "of", "highway", "driving", ",", "we", "need", "to", "ensure", "that", "the", "system", "used", "can", "detect", "cars", "more", "than", "m", "away", "and", "can", "operate", "at", "speeds", "greater", "than", "Hz", ";", "this", "distance", "requires", "higher", "image", "resolutions", "than", "is", "typically", "used", ",", "and", "in", "our", "case", "is", ".", "We", "use", "the", "Overfeat", "CNN", "detector", ",", "which", "is", "very", "scalable", ",", "and", "simulates", "a", "sliding", "window", "detector", "in", "a", "single", "forward", "pass", "in", "the", "network", "by", "efficiently", "reusing", "convolutional", "results", "on", "each", "layer", ".", "Other", "detection", "systems", ",", "such", "as", "R", "-", "CNN", ",", "rely", "on", "selecting", "as", "many", "as", "candidate", "windows", ",", "where", "each", "is", "evaluated", "independently", "and", "does", "not", "reuse", "convolutional", "results", ".", "In", "our", "implementation", ",", "we", "make", "a", "few", "minor", "modifications", "to", "Overfeat", "’s", "labels", "in", "order", "to", "handle", "occlusions", "of", "cars", ",", "predictions", "of", "lanes", ",", "and", "accelerate", "performance", "during", "inference", ".", "We", "will", "first", "provide", "a", "brief", "overview", "of", "the", "original", "implementation", "and", "will", "then", "address", "the", "modifications", ".", "Overfeat", "converts", "an", "image", "recognition", "CNN", "into", "a", "“", "sliding", "window", "”", "detector", "by", "providing", "a", "larger", "resolution", "image", "and", "transforming", "the", "fully", "connected", "layers", "into", "convolutional", "layers", ".", "Then", ",", "after", "converting", "the", "fully", "connected", "layer", ",", "which", "would", "have", "produced", "a", "single", "final", "feature", "vector", ",", "to", "a", "convolutional", "layer", ",", "a", "grid", "of", "final", "feature", "vectors", "is", "produced", ".", "Each", "of", "the", "resulting", "feature", "vectors", "represents", "a", "slightly", "different", "context", "view", "location", "within", "the", "original", "pixel", "space", ".", "To", "determine", "the", "stride", "of", "this", "window", "in", "pixel", "space", ",", "it", "is", "possible", "to", "simply", "multiply", "the", "strides", "on", "each", "convolutional", "or", "pool", "layer", "together", ".", "The", "network", "we", "used", "has", "a", "stride", "size", "of", "pixels", ".", "Each", "final", "feature", "vector", "in", "this", "grid", "can", "predict", "the", "presence", "of", "an", "object", ";", "once", "an", "object", "is", "detected", ",", "those", "same", "features", "are", "then", "used", "to", "predict", "a", "single", "bounding", "box", "through", "regression", ".", "The", "classifier", "will", "predict", "“", "no", "-", "object", "”", "if", "it", "can", "not", "discern", "any", "part", "of", "an", "object", "within", "its", "entire", "input", "view", ".", "This", "causes", "large", "ambiguities", "for", "the", "classifier", ",", "which", "can", "only", "predict", "a", "single", "object", ",", "as", "two", "different", "objects", "could", "can", "easily", "appear", "in", "the", "context", "view", "of", "the", "final", "feature", "vector", ",", "which", "is", "typically", "larger", "than", "of", "the", "input", "image", "resolution", ".", "The", "network", "we", "used", "has", "a", "context", "view", "of", "pixels", "in", "size", ".", "To", "ensure", "that", "all", "objects", "in", "the", "image", "are", "classified", "at", "least", "once", ",", "many", "different", "context", "views", "are", "taken", "of", "the", "image", "by", "using", "skip", "gram", "kernels", "to", "reduce", "the", "stride", "of", "the", "context", "views", "and", "by", "using", "up", "to", "four", "different", "scales", "of", "the", "input", "image", ".", "The", "classifier", "is", "then", "trained", "to", "activate", "when", "an", "object", "appears", "anywhere", "within", "its", "entire", "context", "view", ".", "In", "the", "original", "Overfeat", "paper", ",", "this", "results", "in", "different", "context", "views", "(", "or", "final", "feature", "vectors", ")", ",", "where", "each", "one", "is", "likely", "to", "become", "active", "(", "create", "a", "bounding", "box", ")", ".", "This", "creates", "two", "problems", "for", "our", "empirical", "evaluation", ".", "Due", "to", "the", "L2", "loss", "between", "the", "predicted", "bounding", "box", "and", "actual", "bounding", "proposed", "by", "Sermanet", "et", "al", ".", ",", "the", "ambiguity", "of", "having", "two", "valid", "bounding", "box", "locations", "to", "predict", "when", "two", "objects", "appear", ",", "is", "incorrectly", "handled", "by", "the", "network", "by", "predicting", "a", "box", "in", "the", "center", "of", "the", "two", "objects", "to", "minimize", "its", "expected", "loss", ".", "These", "boxes", "tend", "to", "cause", "a", "problem", "for", "the", "bounding", "box", "merging", "algorithm", ",", "which", "incorrectly", "decides", "that", "there", "must", "be", "a", "third", "object", "between", "the", "two", "ground", "truth", "objects", ".", "This", "could", "cause", "problems", "for", "an", "ADAS", "system", "which", "falsely", "believes", "there", "is", "a", "car", "where", "there", "is", "not", ",", "and", "emergency", "breaking", "is", "falsely", "applied", ".", "In", "addition", ",", "the", "merging", "algorithm", ",", "used", "only", "during", "inference", ",", "operates", "in", "where", "is", "the", "number", "of", "bounding", "boxes", "proposed", ".", "Because", "the", "bounding", "box", "merging", "is", "not", "as", "easily", "parallelizable", "as", "the", "CNN", ",", "this", "merging", "may", "become", "the", "bottleneck", "of", "a", "real", "-", "time", "system", "in", "the", "case", "of", "an", "inefficient", "implementation", "or", "too", "many", "predicted", "bounding", "boxes", ".", "In", "our", "evaluations", ",", "we", "use", "a", "mask", "detector", "as", "described", "in", "Szegedy", "et", "al", ".", "to", "improve", "some", "of", "the", "issues", "with", "Overfeat", "as", "described", "above", ".", "Szegedy", "et", "al", ".", "proposes", "a", "CNN", "that", "takes", "an", "image", "as", "input", "and", "outputs", "an", "object", "mask", "through", "regression", ",", "highlighting", "the", "object", "location", ".", "The", "idea", "of", "a", "mask", "detector", "is", "shown", "in", "Fig", "[", "reference", "]", ".", "To", "distinguish", "multiple", "nearby", "objects", ",", "different", "part", "-", "detectors", "output", "object", "masks", ",", "from", "which", "bounding", "boxes", "are", "then", "extracted", ".", "The", "detector", "they", "propose", "must", "take", "many", "crops", "of", "the", "image", ",", "and", "run", "multiple", "CNNs", "for", "each", "part", "on", "every", "crop", ".", "Their", "resulting", "implementation", "takes", "roughly", "-", "seconds", "per", "frame", "per", "class", "using", "a", "12", "-", "core", "machine", ",", "which", "would", "be", "too", "slow", "for", "our", "application", ".", "We", "combine", "these", "ideas", "by", "using", "the", "efficient", "“", "sliding", "window", "”", "detector", "of", "Overfeat", "to", "produce", "an", "object", "mask", "and", "perform", "bounding", "box", "regression", ".", "This", "is", "shown", "in", "Fig", "[", "reference", "]", ".", "In", "this", "implementation", ",", "we", "use", "a", "single", "image", "resolution", "of", "with", "no", "skip", "gram", "kernels", ".", "To", "help", "the", "ambiguity", "problem", ",", "and", "reduce", "the", "number", "of", "bounding", "boxes", "predicted", ",", "we", "alter", "the", "detector", "on", "the", "top", "layer", "to", "only", "activate", "within", "a", "pixel", "region", "at", "the", "center", "of", "its", "context", "view", ",", "as", "shown", "in", "the", "first", "box", "in", "Fig", "[", "reference", "]", ".", "Because", "it", "’s", "highly", "unlikely", "that", "any", "two", "different", "object", "’s", "bounding", "boxes", "appear", "in", "a", "pixel", "region", ",", "compared", "to", "the", "entire", "context", "view", "with", "Overfeat", ",", "the", "bounding", "box", "regressor", "will", "no", "longer", "have", "to", "arbitrarily", "choose", "between", "two", "valid", "objects", "in", "its", "context", "view", ".", "In", "addition", ",", "because", "the", "requirement", "for", "the", "detector", "to", "fire", "is", "stricter", ",", "this", "produces", "many", "fewer", "bounding", "boxes", "which", "greatly", "reduces", "our", "run", "-", "time", "performance", "during", "inference", ".", "Although", "these", "changes", "helped", ",", "ambiguity", "was", "still", "a", "common", "problem", "on", "the", "border", "of", "bounding", "boxes", "in", "the", "cases", "of", "occlusion", ".", "This", "ambiguity", "results", "in", "a", "false", "bounding", "box", "being", "predicted", "between", "the", "two", "ground", "truth", "bounding", "boxes", ".", "To", "fix", "this", "problem", ",", "the", "bounding", "boxes", "were", "first", "shrunk", "by", "before", "creating", "the", "detection", "mask", "label", ".", "This", "added", "the", "additional", "requirement", "that", "the", "center", "-", "pixel", "region", "of", "the", "detector", "window", "had", "to", "be", "within", "the", "center", "region", "of", "the", "object", "before", "activating", ".", "The", "bounding", "box", "regressor", "however", ",", "still", "predicts", "the", "original", "bounding", "box", "before", "shrinking", ".", "This", "also", "further", "reduces", "the", "number", "of", "active", "bounding", "boxes", "as", "input", "to", "our", "merging", "algorithm", ".", "We", "also", "found", "that", "switching", "from", "L2", "to", "L1", "loss", "on", "the", "bounding", "box", "regressions", "results", "in", "better", "performance", ".", "To", "merge", "the", "bounding", "boxes", "together", ",", "we", "used", "OpenCV", "’s", "efficient", "implementation", "of", "groupRectangles", ",", "which", "clusters", "the", "bounding", "boxes", "based", "on", "a", "similarity", "metric", "in", ".", "The", "lower", "layers", "of", "our", "CNN", "we", "use", "for", "feature", "extraction", "is", "similar", "to", "the", "one", "proposed", "by", "Krizhevsky", "et", "al", ".", ".", "Our", "modifications", "to", "the", "network", "occurs", "on", "the", "dense", "layers", "which", "are", "converted", "to", "convolution", ",", "as", "described", "in", "Sermanet", "et", "al", ".", ".", "When", "using", "our", "larger", "image", "sizes", "of", "this", "changes", "the", "previous", "final", "feature", "response", "maps", "of", "size", "to", ".", "As", "stated", "earlier", ",", "each", "of", "these", "feature", "vectors", "sees", "a", "context", "region", "of", "pixels", ",", "and", "the", "stride", "between", "them", "is", "pixels", ";", "however", ",", "we", "want", "each", "making", "predictions", "at", "a", "resolution", "of", "pixels", ",", "which", "would", "leave", "gaps", "in", "our", "input", "image", ".", "To", "fix", "this", ",", "we", "use", "each", "feature", "as", "input", "to", "softmax", "classifiers", ",", "which", "are", "arranged", "in", "an", "grid", "each", "predicting", "if", "an", "object", "is", "within", "a", "different", "pixel", "region", ".", "This", "allows", "for", "the", "feature", "vector", "to", "cover", "the", "full", "stride", "size", "of", "pixels", ";", "the", "end", "result", "is", "a", "grid", "mask", "detector", "of", "size", "where", "each", "element", "is", "pixels", "which", "covers", "the", "entire", "input", "image", "of", "size", ".", "subsection", ":", "Lane", "Detection", "The", "CNN", "used", "for", "vehicle", "detection", "can", "be", "easily", "extended", "for", "lane", "boundary", "detection", "by", "adding", "an", "additional", "class", ".", "Whereas", "the", "regression", "for", "the", "vehicle", "class", "predicts", "a", "five", "dimensional", "value", "(", "four", "for", "the", "bounding", "box", "and", "one", "for", "depth", ")", ",", "the", "lane", "regression", "predicts", "six", "dimensions", ".", "Similar", "to", "the", "vehicle", "detector", ",", "the", "first", "four", "dimensions", "indicate", "the", "two", "end", "points", "of", "a", "local", "line", "segment", "of", "the", "lane", "boundary", ".", "The", "remaining", "two", "dimensions", "indicate", "the", "depth", "of", "the", "endpoints", "with", "respect", "to", "the", "camera", ".", "Fig", "[", "reference", "]", "visualizes", "the", "lane", "boundary", "ground", "truth", "label", "overlaid", "on", "an", "example", "image", ".", "The", "green", "tiles", "indicate", "locations", "where", "the", "detector", "is", "trained", "to", "fire", ",", "and", "the", "line", "segments", "represented", "by", "the", "regression", "labels", "are", "explicitly", "drawn", ".", "The", "line", "segments", "have", "their", "ends", "connected", "to", "form", "continuous", "splines", ".", "The", "depth", "of", "the", "line", "segments", "are", "color", "-", "coded", "such", "that", "the", "closest", "segments", "are", "red", "and", "the", "furthest", "ones", "are", "blue", ".", "Due", "to", "our", "data", "collection", "methods", "for", "lane", "labels", ",", "we", "are", "able", "to", "obtain", "ground", "truth", "in", "spite", "of", "objects", "that", "occlude", "them", ".", "This", "forces", "the", "neural", "network", "to", "learn", "more", "than", "a", "simple", "paint", "detector", ",", "and", "must", "use", "context", "to", "predict", "lanes", "where", "there", "are", "occlusions", ".", "Similar", "to", "the", "vehicle", "detector", ",", "we", "use", "L1", "loss", "to", "train", "the", "regressor", ".", "We", "use", "mini", "-", "batch", "stochastic", "gradient", "descent", "for", "optimization", ".", "The", "learning", "rate", "is", "controlled", "by", "a", "variant", "of", "the", "momentum", "scheduler", ".", "To", "obtain", "semantic", "lane", "information", ",", "we", "use", "DBSCAN", "to", "cluster", "the", "line", "segments", "into", "lanes", ".", "Fig", "[", "reference", "]", "shows", "our", "lane", "predictions", "after", "DBSCAN", "clustering", ".", "Different", "lanes", "are", "represented", "by", "different", "colors", ".", "Since", "our", "regressor", "outputs", "depths", "as", "well", ",", "we", "can", "predict", "the", "lane", "shapes", "in", "3D", "using", "inverse", "camera", "perspective", "mapping", ".", "section", ":", "Experimental", "Setup", "subsection", ":", "Data", "Collection", "Our", "Research", "Vehicle", "is", "a", "2014", "Infiniti", "Q50", ".", "The", "car", "currently", "uses", "the", "following", "sensors", ":", "6x", "Point", "Grey", "Flea3", "cameras", ",", "1x", "Velodyne", "HDL32E", "lidar", ",", "and", "1x", "Novatel", "SPAN", "-", "SE", "Receiver", ".", "We", "also", "have", "access", "to", "the", "Q50", "built", "-", "in", "Continental", "mid", "-", "range", "radar", "system", ".", "The", "sensors", "are", "connected", "to", "a", "Linux", "PC", "with", "a", "Core", "i7", "-", "4770k", "processor", ".", "Once", "the", "raw", "videos", "are", "collected", ",", "we", "annotate", "the", "3D", "locations", "for", "vehicles", "and", "lanes", "as", "well", "as", "the", "relative", "speed", "of", "all", "the", "vehicles", ".", "To", "get", "vehicle", "annotations", ",", "we", "follow", "the", "conventional", "approach", "of", "using", "Amazon", "Mechanical", "Turk", "to", "get", "accurate", "bounding", "box", "locations", "within", "pixel", "space", ".", "Then", ",", "we", "match", "bounding", "boxes", "and", "radar", "returns", "to", "obtain", "the", "distance", "and", "relative", "speed", "of", "the", "vehicles", ".", "Unlike", "vehicles", "that", "can", "be", "annotated", "with", "bounding", "boxes", ",", "highway", "lane", "borders", "often", "need", "to", "be", "annotated", "as", "curves", "of", "various", "shapes", ".", "This", "makes", "frame", "-", "level", "labelling", "not", "only", "tedious", "and", "inefficient", ",", "but", "also", "prone", "to", "human", "errors", ".", "Fortunately", ",", "lane", "markings", "can", "be", "considered", "as", "“static”", "objects", "that", "do", "not", "change", "their", "geolocations", "very", "often", ".", "We", "follow", "the", "process", "descried", "in", "to", "create", "LIDAR", "maps", "of", "the", "environment", "using", "the", "Velodyne", "and", "GNSS", "systems", ".", "Using", "these", "maps", ",", "labeling", "is", "straight", "forward", ".", "First", ",", "we", "filter", "the", "3D", "point", "clouds", "based", "on", "lidar", "return", "intensity", "and", "position", "to", "obtain", "the", "left", "and", "right", "boundaries", "of", "the", "ego", "-", "lane", ".", "Then", ",", "we", "replicate", "the", "left", "and", "right", "ego", "-", "lane", "boundaries", "to", "obtain", "initial", "guesses", "for", "all", "the", "lane", "boundaries", ".", "A", "human", "annotator", "inspects", "the", "generated", "lane", "boundaries", "and", "makes", "appropriate", "corrections", "using", "our", "3D", "labelling", "tool", ".", "For", "completeness", ",", "we", "describe", "each", "of", "these", "steps", "in", "details", ".", "subsubsection", ":", "Ego", "-", "lane", "boundary", "generation", "Since", "we", "do", "not", "change", "lanes", "during", "data", "collection", "drives", ",", "the", "GPS", "trajectory", "of", "our", "research", "vehicle", "already", "gives", "a", "decent", "estimate", "of", "the", "shape", "of", "the", "road", ".", "We", "can", "then", "easily", "locate", "the", "ego", "-", "lane", "boundaries", "using", "a", "few", "heuristic", "filters", ".", "Noting", "that", "lane", "boundaries", "on", "highways", "are", "usually", "marked", "with", "retro", "-", "reflective", "materials", ",", "we", "first", "filter", "out", "low", "-", "reflectivity", "surfaces", "such", "as", "asphalt", "in", "our", "3D", "point", "cloud", "maps", "and", "only", "consider", "points", "with", "high", "enough", "laser", "return", "intensities", ".", "We", "then", "filter", "out", "other", "reflective", "surfaces", "such", "as", "cars", "and", "traffic", "signs", "by", "only", "considering", "points", "whose", "heights", "are", "close", "enough", "the", "ground", "plane", ".", "Lastly", ",", "assuming", "our", "car", "drives", "close", "to", "the", "center", "of", "the", "lane", ",", "we", "filter", "out", "ground", "paint", "other", "than", "the", "ego", "-", "lane", "boundaries", ",", "such", "as", "other", "lane", "boundaries", ",", "car", "-", "pool", "or", "directional", "signs", ",", "by", "only", "considering", "markings", "whose", "absolute", "lateral", "distances", "from", "the", "car", "are", "smaller", "than", "2.2", "meters", "and", "greater", "than", "1.4", "meters", ".", "We", "can", "also", "distinguish", "the", "left", "boundary", "from", "the", "right", "one", "using", "the", "sign", "of", "the", "lateral", "distance", ".", "After", "obtaining", "the", "points", "in", "the", "left", "and", "right", "boundaries", ",", "we", "fit", "a", "piecewise", "linear", "curve", "similar", "to", "the", "GPS", "trajectory", "to", "each", "boundary", ".", "subsubsection", ":", "Semi", "-", "automatic", "generation", "of", "multiple", "lane", "boundaries", "We", "observe", "that", "the", "width", "of", "lanes", "during", "a", "single", "data", "collection", "run", "stays", "constant", "most", "of", "the", "time", ",", "with", "occasional", "exceptions", "such", "as", "merges", "and", "splits", ".", "Therefore", ",", "if", "we", "predefine", "the", "number", "of", "lanes", "to", "the", "left", "and", "right", "of", "the", "car", "for", "a", "single", "run", ",", "we", "can", "make", "a", "good", "initial", "guess", "of", "all", "the", "lane", "boundaries", "by", "shifting", "the", "auto", "-", "generated", "ego", "-", "lane", "boundaries", "laterally", "by", "multiples", "of", "the", "lane", "width", ".", "We", "will", "then", "rely", "on", "human", "annotators", "to", "fix", "the", "exception", "cases", ".", "subsection", ":", "Data", "Set", "At", "the", "time", "of", "this", "writing", "our", "annotated", "data", "-", "set", "consists", "of", "days", "of", "driving", "in", "the", "San", "Francisco", "Bay", "Area", "during", "the", "months", "of", "April", "-", "June", "for", "a", "few", "hours", "each", "day", ".", "The", "vehicle", "annotated", "data", "is", "sampled", "at", "and", "contains", "nearly", "thousand", "frames", "with", "thousand", "bounding", "boxes", ".", "The", "lane", "annotated", "data", "is", "sampled", "at", "and", "contains", "over", "thousand", "frames", ".", "During", "training", ",", "translation", "and", "7", "different", "perspective", "distortions", "are", "applied", "to", "the", "raw", "data", "sets", ".", "Fig", "[", "reference", "]", "shows", "an", "example", "image", "after", "perspective", "distortions", "are", "applied", ".", "Note", "that", "we", "apply", "the", "same", "perspective", "distortion", "to", "the", "ground", "truth", "labels", "so", "that", "they", "match", "correctly", "with", "the", "distorted", "image", ".", "subsection", ":", "Results", "The", "detection", "network", "used", "is", "capable", "of", "running", "at", "Hz", "using", "a", "desktop", "PC", "equipped", "with", "a", "GTX", "780", "Ti", ".", "When", "using", "a", "mobile", "GPU", ",", "such", "as", "the", "Tegra", "K1", ",", "we", "were", "capable", "of", "running", "the", "network", "at", "Hz", ",", "and", "would", "expect", "the", "system", "to", "run", "at", "Hz", "using", "the", "Nvidia", "PX1", "chipset", ".", "Our", "lane", "detection", "test", "set", "consists", "of", "22", "video", "clips", "collected", "using", "both", "left", "and", "right", "cameras", "during", "11", "different", "data", "collection", "runs", ",", "which", "correspond", "to", "about", "50", "minutes", "of", "driving", ".", "We", "evaluate", "detection", "results", "for", "four", "lane", "boundaries", ",", "namely", ",", "the", "left", "and", "right", "boundaries", "of", "the", "ego", "lane", ",", "plus", "the", "outer", "boundaries", "of", "the", "two", "adjacent", "lanes", ".", "For", "each", "of", "these", "lane", "boundaries", ",", "we", "further", "break", "down", "the", "evaluation", "by", "longitudinal", "distances", ",", "which", "range", "from", "15", "to", "80", "meters", "ahead", "of", "the", "car", ",", "spaced", "by", "5", "meters", ".", "Thus", ",", "there", "are", "at", "maximum", "positions", "at", "which", "we", "evaluate", "the", "detection", "results", ".", "We", "pair", "up", "the", "prediction", "and", "ground", "truth", "points", "at", "each", "of", "these", "locations", "using", "greedy", "nearest", "neighbor", "matching", ".", "True", "positives", ",", "false", "positives", "and", "false", "negatives", "are", "accumulated", "at", "every", "evaluation", "location", "in", "a", "standard", "way", ":", "A", "true", "positive", "is", "counted", "when", "the", "matched", "prediction", "and", "ground", "truth", "differ", "by", "less", "than", "0.5", "meter", ".", "If", "the", "matched", "prediction", "and", "ground", "truth", "differ", "by", "more", "than", "0.5", "meter", ",", "both", "false", "positive", "and", "false", "negative", "counts", "are", "incremented", ".", "Fig", "[", "reference", "]", "shows", "a", "visualization", "of", "this", "evaluation", "method", "on", "one", "image", ".", "The", "blue", "dots", "are", "true", "positives", ".", "The", "red", "dots", "are", "false", "positives", ",", "and", "the", "yellow", "ones", "are", "false", "negatives", ".", "Fig", "[", "reference", "]", "shows", "the", "aggregated", "precision", ",", "recall", "and", "F1", "score", "on", "all", "test", "videos", ".", "For", "the", "ego", "-", "lane", "boundaries", ",", "we", "obtain", "F1", "score", "up", "to", "50", "meters", ".", "Recall", "starts", "to", "drop", "fast", "beyond", "65", "meters", ",", "mainly", "because", "the", "resolution", "of", "the", "image", "can", "not", "capture", "the", "width", "of", "the", "lane", "markings", "at", "that", "distance", ".", "For", "the", "adjacent", "lanes", ",", "recall", "is", "low", "for", "the", "nearest", "point", "because", "it", "is", "outside", "the", "field", "of", "view", "of", "the", "camera", ".", "[", "b", "]", "1.6", "in", "[", "b", "]", "1.6", "in", "[", "b", "]", "1.6", "in", "[", "b", "]", "1.6", "in", "The", "vehicle", "detection", "test", "set", "consists", "of", "13", "video", "clips", "collected", "from", "a", "single", "day", ",", "which", "corresponds", "to", "1", "hour", "and", "30", "mins", "of", "driving", ".", "The", "accuracy", "of", "the", "vehicle", "bounding", "box", "predictions", "were", "measured", "using", "Intersection", "Over", "Union", "(", "IOU", ")", "against", "the", "ground", "truth", "boxes", "from", "Amazon", "Mechanical", "Turk", "(", "AMT", ")", ".", "A", "bounding", "box", "prediction", "matched", "with", "ground", "truth", "if", "IOU", ".", "The", "performance", "of", "our", "car", "detection", "as", "a", "function", "of", "depth", "can", "be", "seen", "in", "Fig", "[", "reference", "]", ".", "Nearby", "false", "positives", "can", "cause", "the", "largest", "problems", "for", "ADAS", "systems", "which", "could", "cause", "the", "system", "to", "needlessly", "apply", "the", "brakes", ".", "In", "our", "system", ",", "we", "found", "overpasses", "and", "shading", "effects", "to", "cause", "the", "largest", "problems", ".", "Two", "examples", "of", "these", "situations", "are", "shown", "in", "Fig", "[", "reference", "]", ".", "[", "b", "]", "1.7", "in", "[", "b", "]", "1.7", "in", "As", "a", "baseline", "to", "our", "car", "detector", ",", "we", "compared", "the", "detection", "results", "to", "the", "Continental", "mid", "-", "range", "radar", "within", "our", "data", "collection", "vehicle", ".", "While", "matching", "radar", "returns", "to", "ground", "truth", "bounding", "boxes", ",", "we", "found", "that", "although", "radar", "had", "nearly", "precision", ",", "false", "positives", "were", "being", "introduced", "through", "errors", "in", "radar", "/", "camera", "calibration", ".", "Therefore", ",", "to", "ensure", "a", "fair", "comparison", "we", "matched", "every", "radar", "return", "to", "a", "ground", "truth", "bounding", "box", "even", "if", "IOU", ",", "giving", "our", "radar", "returns", "precision", ".", "This", "comparison", "is", "shown", "in", "Fig", "[", "reference", "]", ",", "the", "F1", "score", "for", "radar", "is", "simply", "the", "recall", ".", "In", "addition", "to", "the", "bounding", "box", "locations", ",", "we", "measured", "the", "accuracy", "of", "the", "predicted", "depth", "by", "using", "radar", "returns", "as", "ground", "truth", ".", "The", "standard", "error", "in", "the", "depth", "predictions", "as", "a", "function", "of", "depth", "can", "be", "seen", "in", "Fig", "[", "reference", "]", ".", "For", "a", "qualitative", "review", "of", "the", "detection", "system", ",", "we", "have", "uploaded", "a", "hour", "video", "of", "the", "vehicle", "detector", "ran", "on", "our", "test", "set", ".", "This", "may", "be", "found", "at", "youtu.be", "/", "GJ0cZBkHoHc", ".", "A", "short", "video", "of", "our", "lane", "detector", "may", "also", "be", "found", "online", "at", "youtu.be", "/", "_", "_", "f5pqqp6aM.", "In", "these", "videos", ",", "we", "evaluate", "the", "detector", "on", "every", "frame", "independently", "and", "display", "the", "raw", "detections", ",", "without", "the", "use", "of", "any", "Kalman", "filters", "or", "road", "models", ".", "The", "red", "locations", "in", "the", "video", "correspond", "to", "the", "mask", "detectors", "that", "are", "activated", ".", "This", "network", "was", "only", "trained", "on", "the", "rear", "view", "of", "cars", "traveling", "in", "the", "same", "direction", ",", "which", "is", "why", "cars", "across", "the", "highway", "barrier", "are", "commonly", "missed", ".", "We", "have", "open", "sourced", "the", "code", "for", "the", "vehicle", "and", "lane", "detector", "online", "at", "github.com", "/", "brodyh", "/", "caffe", ".", "Our", "repository", "was", "forked", "from", "the", "original", "Caffe", "code", "base", "from", "the", "BVLC", "group", ".", "section", ":", "Conclusion", "By", "using", "Camera", ",", "Lidar", ",", "Radar", ",", "and", "GPS", "we", "built", "a", "highway", "data", "set", "consisting", "of", "thousand", "image", "frames", "with", "vehicle", "bounding", "boxes", "and", "over", "thousand", "image", "frames", "with", "lane", "annotations", ".", "We", "then", "trained", "on", "this", "data", "using", "a", "CNN", "architecture", "capable", "of", "detecting", "all", "lanes", "and", "cars", "in", "a", "single", "forward", "pass", ".", "Using", "a", "single", "GTX", "780", "Ti", "our", "system", "runs", "at", "Hz", ",", "which", "is", "more", "than", "adequate", "for", "real", "-", "time", "use", ".", "Our", "results", "show", "existing", "CNN", "algorithms", "are", "capable", "of", "good", "performance", "in", "highway", "lane", "and", "vehicle", "detection", ".", "Future", "work", "will", "focus", "on", "acquiring", "frame", "level", "annotations", "that", "will", "allow", "us", "to", "develop", "new", "neural", "networks", "capable", "of", "using", "temporal", "information", "across", "frames", ".", "section", ":", "Acknowledgment", "This", "research", "was", "funded", "in", "part", "by", "Nissan", "who", "generously", "donated", "the", "car", "used", "for", "data", "collection", ".", "We", "thank", "our", "colleagues", "Yuta", "Yoshihata", "from", "Nissan", "who", "provided", "technical", "support", "and", "expertise", "on", "vehicles", "that", "assisted", "the", "research", ".", "In", "addition", ",", "the", "authors", "would", "like", "to", "thank", "the", "author", "of", "Overfeat", ",", "Pierre", "Sermanet", ",", "for", "their", "helpful", "suggestions", "on", "image", "detection", ".", "bibliography", ":", "References"], "entity": [[6, 8, ["Deep", "Learning"], "Task"], [9, 11, ["Highway", "Driving"], "Task"], [18, 21, ["deep", "learning", "techniques"], "Method"], [22, 25, ["computer", "vision", "problems"], "Task"], [26, 29, ["highway", "perception", "scenarios"], "Task"], [43, 46, ["deep", "learning", "advances"], "Method"], [47, 49, ["Computer", "vision"], "Task"], [52, 54, ["deep", "learning"], "Method"], [68, 70, ["autonomous", "driving"], "Task"], [73, 75, ["deep", "learning"], "Method"], [76, 78, ["industry", "uptake"], "Task"], [79, 81, ["practical", "applications"], "Task"], [82, 84, ["neural", "networks"], "Method"], [109, 115, ["deep", "learning", "and", "computer", "vision", "algorithms"], "Method"], [119, 120, ["car"], "Task"], [121, 123, ["lane", "detection"], "Task"], [128, 131, ["convolutional", "neural", "networks"], "Method"], [132, 133, ["CNNs"], "Method"], [139, 140, ["lane"], "Task"], [141, 143, ["vehicle", "detection"], "Task"], [146, 148, ["frame", "rates"], "Metric"], [151, 155, ["real", "-", "time", "system"], "Task"], [164, 166, ["deep", "learning"], "Method"], [169, 171, ["autonomous", "driving"], "Task"], [177, 180, ["DARPA", "Grand", "Challenges"], "Task"], [181, 183, ["autonomous", "vehicles"], "Task"], [209, 213, ["highway", "and", "urban", "roads"], "Task"], [293, 296, ["autonomous", "driving", "technology"], "Task"], [302, 307, ["highway", "auto", "-", "pilot", "solutions"], "Task"], [325, 330, ["advanced", "-", "driver", "assistance", "systems"], "Method"], [331, 332, ["ADAS"], "Method"], [342, 347, ["front", "-", "view", "car", "detection"], "Task"], [396, 400, ["auto", "-", "pilot", "systems"], "Method"], [430, 431, ["LIDAR"], "Method"], [432, 433, ["radar"], "Method"], [436, 438, ["precision", "GPS"], "Method"], [447, 452, ["production", "-", "grade", "autonomous", "vehicles"], "Task"], [453, 455, ["critical", "sensors"], "Task"], [463, 468, ["Long", "-", "range", "vehicle", "detection"], "Task"], [470, 471, ["radar"], "Task"], [473, 476, ["nearby", "car", "detection"], "Task"], [480, 481, ["sonar"], "Method"], [482, 484, ["Computer", "vision"], "Task"], [491, 493, ["lane", "detection"], "Task"], [503, 504, ["Radar"], "Method"], [508, 510, ["detecting", "vehicles"], "Task"], [534, 535, ["radar"], "Method"], [553, 554, ["localization"], "Task"], [562, 563, ["sonar"], "Method"], [588, 589, ["sonar"], "Method"], [590, 591, ["radar"], "Method"], [592, 593, ["cameras"], "Method"], [608, 610, ["computer", "vision"], "Task"], [617, 619, ["redundant", "sensor"], "Method"], [620, 622, ["autonomous", "driving"], "Task"], [627, 629, ["computer", "vision"], "Task"], [645, 648, ["computer", "vision", "techniques"], "Method"], [653, 654, ["robustness"], "Metric"], [656, 659, ["production", "grade", "automotives"], "Task"], [664, 666, ["hand", "engineering"], "Task"], [667, 669, ["road", "modeling"], "Task"], [671, 674, ["special", "case", "handling"], "Task"], [696, 698, ["computer", "vision"], "Method"], [715, 717, ["Deep", "learning"], "Method"], [719, 721, ["neural", "networks"], "Method"], [727, 729, ["computer", "vision"], "Task"], [742, 744, ["computer", "vision"], "Task"], [755, 758, ["deep", "learning", "applications"], "Task"], [762, 766, ["real", "-", "world", "problems"], "Task"], [777, 779, ["Deep", "learning"], "Method"], [780, 782, ["data", "centric"], "Task"], [808, 810, ["deep", "learning"], "Method"], [814, 817, ["supervised", "perception", "tasks"], "Task"], [820, 822, ["image", "detection"], "Task"], [824, 826, ["neural", "network"], "Method"], [844, 845, ["inference"], "Task"], [846, 849, ["real", "-", "time"], "Task"], [870, 872, ["neural", "networks"], "Method"], [873, 875, ["computer", "vision"], "Task"], [903, 905, ["neural", "networks"], "Method"], [926, 930, ["self", "-", "driving", "applications"], "Task"], [935, 939, ["mm", "-", "accurate", "GPS"], "Method"], [995, 997, ["neural", "networks"], "Method"], [1040, 1042, ["neural", "network"], "Method"], [1046, 1048, ["detecting", "lanes"], "Task"], [1049, 1050, ["cars"], "Task"], [1065, 1067, ["computer", "vision"], "Task"], [1076, 1078, ["autonomous", "driving"], "Task"], [1088, 1089, ["precision"], "Metric"], [1102, 1104, ["road", "models"], "Method"], [1117, 1118, ["radar"], "Method"], [1120, 1122, ["computer", "vision"], "Task"], [1123, 1125, ["object", "detection"], "Task"], [1134, 1136, ["Kalman", "filter"], "Method"], [1137, 1139, ["motion", "models"], "Method"], [1151, 1155, ["deformable", "parts", "based", "model"], "Method"], [1164, 1166, ["road", "models"], "Method"], [1183, 1184, ["tracker"], "Method"], [1186, 1189, ["pixel", "space", "detections"], "Task"], [1204, 1205, ["detection"], "Task"], [1214, 1219, ["front", "-", "view", "motion", "model"], "Method"], [1224, 1226, ["object", "detectors"], "Method"], [1232, 1238, ["road", "or", "motion", "-", "based", "models"], "Method"], [1248, 1250, ["neural", "network"], "Method"], [1299, 1301, ["neural", "network"], "Method"], [1333, 1335, ["neural", "network"], "Method"], [1337, 1340, ["entire", "depth", "maps"], "Task"], [1354, 1355, ["road"], "Task"], [1382, 1385, ["Convolutional", "Neural", "Networks"], "Method"], [1386, 1387, ["CNNs"], "Method"], [1389, 1391, ["computer", "vision"], "Task"], [1392, 1396, ["deformable", "parts", "based", "models"], "Method"], [1402, 1403, ["detection"], "Task"], [1420, 1421, ["detection"], "Task"], [1423, 1424, ["CNNs"], "Method"], [1425, 1427, ["feature", "extraction"], "Task"], [1433, 1436, ["R", "-", "CNN"], "Method"], [1439, 1441, ["part", "system"], "Method"], [1449, 1450, ["AlexNet"], "Method"], [1454, 1457, ["R", "-", "CNN"], "Method"], [1466, 1467, ["Pascal"], "Task"], [1478, 1480, ["classification", "queries"], "Metric"], [1486, 1487, ["convolutions"], "Method"], [1492, 1496, ["real", "-", "time", "implementations"], "Task"], [1507, 1510, ["R", "-", "CNN"], "Method"], [1515, 1516, ["CNN"], "Method"], [1523, 1525, ["Selective", "Search"], "Method"], [1541, 1542, ["mAP"], "Method"], [1544, 1546, ["Selective", "Search"], "Method"], [1552, 1554, ["image", "detection"], "Task"], [1555, 1556, ["Overfeat"], "Method"], [1575, 1576, ["Overfeat"], "Method"], [1579, 1581, ["convolution", "computations"], "Method"], [1589, 1591, ["forward", "pass"], "Method"], [1592, 1593, ["inference"], "Task"], [1607, 1608, ["Overfeat"], "Method"], [1643, 1647, ["Real", "Time", "Vehicle", "Detection"], "Task"], [1647, 1650, ["Convolutional", "Neural", "Networks"], "Method"], [1651, 1652, ["CNNs"], "Method"], [1659, 1661, ["image", "recognition"], "Task"], [1669, 1672, ["image", "recognition", "systems"], "Method"], [1676, 1678, ["detection", "networks"], "Method"], [1686, 1688, ["image", "detection"], "Task"], [1704, 1708, ["real", "-", "time", "detection"], "Task"], [1721, 1723, ["detection", "system"], "Method"], [1734, 1736, ["laptop", "GPU"], "Method"], [1742, 1744, ["highway", "driving"], "Task"], [1822, 1824, ["detection", "systems"], "Method"], [1827, 1830, ["R", "-", "CNN"], "Method"], [1863, 1864, ["Overfeat"], "Method"], [1874, 1877, ["predictions", "of", "lanes"], "Task"], [1882, 1883, ["inference"], "Task"], [1902, 1903, ["Overfeat"], "Method"], [1924, 1927, ["fully", "connected", "layers"], "Method"], [1928, 1930, ["convolutional", "layers"], "Method"], [1936, 1939, ["fully", "connected", "layer"], "Method"], [1952, 1954, ["convolutional", "layer"], "Method"], [2004, 2008, ["convolutional", "or", "pool", "layer"], "Method"], [2055, 2056, ["regression"], "Method"], [2058, 2059, ["classifier"], "Method"], [2088, 2089, ["classifier"], "Method"], [2165, 2168, ["skip", "gram", "kernels"], "Method"], [2190, 2191, ["classifier"], "Method"], [2210, 2211, ["Overfeat"], "Method"], [2318, 2322, ["bounding", "box", "merging", "algorithm"], "Method"], [2346, 2348, ["ADAS", "system"], "Method"], [2371, 2373, ["merging", "algorithm"], "Method"], [2377, 2378, ["inference"], "Task"], [2392, 2395, ["bounding", "box", "merging"], "Method"], [2402, 2403, ["CNN"], "Method"], [2405, 2406, ["merging"], "Task"], [2412, 2416, ["real", "-", "time", "system"], "Task"], [2437, 2439, ["mask", "detector"], "Method"], [2453, 2454, ["Overfeat"], "Method"], [2464, 2465, ["CNN"], "Method"], [2477, 2478, ["regression"], "Method"], [2488, 2490, ["mask", "detector"], "Method"], [2505, 2508, ["part", "-", "detectors"], "Method"], [2535, 2536, ["CNNs"], "Method"], [2556, 2560, ["12", "-", "core", "machine"], "Method"], [2584, 2585, ["Overfeat"], "Method"], [2592, 2595, ["bounding", "box", "regression"], "Method"], [2618, 2621, ["skip", "gram", "kernels"], "Method"], [2625, 2627, ["ambiguity", "problem"], "Task"], [2698, 2699, ["Overfeat"], "Method"], [2701, 2704, ["bounding", "box", "regressor"], "Method"], [2744, 2747, ["run", "-", "time"], "Metric"], [2749, 2750, ["inference"], "Task"], [2840, 2843, ["bounding", "box", "regressor"], "Method"], [2868, 2870, ["merging", "algorithm"], "Method"], [2883, 2886, ["bounding", "box", "regressions"], "Method"], [2900, 2901, ["OpenCV"], "Method"], [2905, 2906, ["groupRectangles"], "Method"], [2915, 2917, ["similarity", "metric"], "Metric"], [2924, 2925, ["CNN"], "Method"], [2928, 2930, ["feature", "extraction"], "Task"], [2950, 2952, ["dense", "layers"], "Method"], [2956, 2957, ["convolution"], "Method"], [2992, 2994, ["feature", "vectors"], "Method"], [3042, 3044, ["softmax", "classifiers"], "Method"], [3083, 3086, ["grid", "mask", "detector"], "Method"], [3104, 3106, ["Lane", "Detection"], "Task"], [3107, 3108, ["CNN"], "Method"], [3110, 3112, ["vehicle", "detection"], "Task"], [3131, 3133, ["vehicle", "class"], "Task"], [3160, 3162, ["vehicle", "detector"], "Method"], [3222, 3223, ["detector"], "Method"], [3280, 3283, ["data", "collection", "methods"], "Method"], [3305, 3307, ["neural", "network"], "Method"], [3313, 3315, ["paint", "detector"], "Method"], [3331, 3333, ["vehicle", "detector"], "Method"], [3336, 3338, ["L1", "loss"], "Method"], [3341, 3342, ["regressor"], "Method"], [3345, 3351, ["mini", "-", "batch", "stochastic", "gradient", "descent"], "Method"], [3352, 3353, ["optimization"], "Task"], [3355, 3357, ["learning", "rate"], "Metric"], [3364, 3366, ["momentum", "scheduler"], "Method"], [3375, 3376, ["DBSCAN"], "Method"], [3421, 3425, ["inverse", "camera", "perspective", "mapping"], "Task"], [3440, 3442, ["Infiniti", "Q50"], "Method"], [3452, 3456, ["Point", "Grey", "Flea3", "cameras"], "Method"], [3458, 3461, ["Velodyne", "HDL32E", "lidar"], "Method"], [3464, 3469, ["Novatel", "SPAN", "-", "SE", "Receiver"], "Method"], [3480, 3486, ["Continental", "mid", "-", "range", "radar", "system"], "Task"], [3493, 3495, ["Linux", "PC"], "Method"], [3532, 3534, ["vehicle", "annotations"], "Task"], [3601, 3605, ["frame", "-", "level", "labelling"], "Method"], [3645, 3650, ["LIDAR", "maps", "of", "the", "environment"], "Task"], [3652, 3656, ["Velodyne", "and", "GNSS", "systems"], "Method"], [3661, 3662, ["labeling"], "Task"], [3730, 3733, ["3D", "labelling", "tool"], "Method"], [3796, 3798, ["heuristic", "filters"], "Method"], [3827, 3831, ["3D", "point", "cloud", "maps"], "Task"], [3963, 3966, ["piecewise", "linear", "curve"], "Method"], [4207, 4209, ["detection", "network"], "Method"], [4223, 4226, ["GTX", "780", "Ti"], "Method"], [4236, 4238, ["Tegra", "K1"], "Method"], [4260, 4263, ["Nvidia", "PX1", "chipset"], "Method"], [4265, 4267, ["lane", "detection"], "Task"], [4299, 4300, ["detection"], "Task"], [4374, 4375, ["detection"], "Task"], [4392, 4396, ["greedy", "nearest", "neighbor", "matching"], "Method"], [4403, 4405, ["false", "negatives"], "Metric"], [4453, 4456, ["false", "negative", "counts"], "Metric"], [4502, 4504, ["aggregated", "precision"], "Metric"], [4505, 4506, ["recall"], "Metric"], [4507, 4509, ["F1", "score"], "Metric"], [4523, 4525, ["F1", "score"], "Metric"], [4530, 4531, ["Recall"], "Metric"], [4564, 4565, ["recall"], "Metric"], [4604, 4606, ["vehicle", "detection"], "Task"], [4631, 4632, ["accuracy"], "Metric"], [4634, 4638, ["vehicle", "bounding", "box", "predictions"], "Task"], [4641, 4644, ["Intersection", "Over", "Union"], "Metric"], [4645, 4646, ["IOU"], "Metric"], [4661, 4664, ["bounding", "box", "prediction"], "Method"], [4700, 4702, ["ADAS", "systems"], "Task"], [4757, 4759, ["car", "detector"], "Method"], [4763, 4764, ["detection"], "Task"], [4767, 4772, ["Continental", "mid", "-", "range", "radar"], "Method"], [4792, 4793, ["radar"], "Method"], [4795, 4796, ["precision"], "Metric"], [4805, 4809, ["radar", "/", "camera", "calibration"], "Method"], [4834, 4837, ["radar", "returns", "precision"], "Metric"], [4849, 4851, ["F1", "score"], "Metric"], [4852, 4853, ["radar"], "Task"], [4856, 4857, ["recall"], "Metric"], [4869, 4870, ["accuracy"], "Metric"], [4909, 4911, ["detection", "system"], "Method"], [4920, 4922, ["vehicle", "detector"], "Method"], [4942, 4944, ["lane", "detector"], "Method"], [4978, 4980, ["Kalman", "filters"], "Method"], [4981, 4983, ["road", "models"], "Method"], [4993, 4995, ["mask", "detectors"], "Method"], [5068, 5069, ["Camera"], "Method"], [5070, 5071, ["Lidar"], "Method"], [5072, 5073, ["Radar"], "Method"], [5075, 5076, ["GPS"], "Method"], [5158, 5160, ["highway", "lane"], "Task"], [5161, 5163, ["vehicle", "detection"], "Task"], [5180, 5182, ["neural", "networks"], "Method"], [5244, 5245, ["Overfeat"], "Method"], [5254, 5256, ["image", "detection"], "Task"], [197, 198, ["cars"], "Task"], [207, 208, ["cars"], "Task"], [267, 268, ["lane"], "Task"], [336, 337, ["cars"], "Task"], [339, 340, ["lane"], "Task"], [405, 406, ["cars"], "Task"], [419, 420, ["cars"], "Task"], [497, 499, ["object", "detection"], "Task"], [642, 643, ["cars"], "Task"], [955, 956, ["lane"], "Task"], [1179, 1180, ["detector"], "Method"], [1270, 1271, ["detector"], "Method"], [1356, 1358, ["object", "detection"], "Task"], [1407, 1408, ["CNN"], "Method"], [1755, 1756, ["cars"], "Task"], [1789, 1790, ["Overfeat"], "Method"], [1790, 1791, ["CNN"], "Method"], [1791, 1792, ["detector"], "Method"], [1803, 1804, ["detector"], "Method"], [1872, 1873, ["cars"], "Task"], [1907, 1908, ["CNN"], "Method"], [1914, 1915, ["detector"], "Method"], [2354, 2355, ["car"], "Task"], [2521, 2522, ["detector"], "Method"], [2582, 2583, ["detector"], "Method"], [2640, 2641, ["detector"], "Method"], [2728, 2729, ["detector"], "Method"], [2824, 2825, ["detector"], "Method"], [3117, 3118, ["lane"], "Task"], [3151, 3152, ["lane"], "Task"], [3179, 3180, ["lane"], "Task"], [3204, 3205, ["lane"], "Task"], [3284, 3285, ["lane"], "Task"], [3370, 3371, ["lane"], "Task"], [3390, 3391, ["lane"], "Task"], [3393, 3394, ["DBSCAN"], "Method"], [3416, 3417, ["lane"], "Task"], [3444, 3445, ["car"], "Task"], [3586, 3587, ["lane"], "Task"], [3620, 3621, ["lane"], "Task"], [3692, 3693, ["lane"], "Task"], [3704, 3705, ["lane"], "Task"], [3713, 3714, ["lane"], "Task"], [3722, 3723, ["lane"], "Task"], [3750, 3751, ["lane"], "Task"], [3791, 3792, ["lane"], "Task"], [3801, 3802, ["lane"], "Task"], [3851, 3852, ["cars"], "Task"], [3872, 3873, ["car"], "Task"], [3880, 3881, ["lane"], "Task"], [3892, 3893, ["lane"], "Task"], [3898, 3899, ["lane"], "Task"], [3901, 3902, ["car"], "Task"], [3918, 3919, ["car"], "Task"], [3983, 3984, ["lane"], "Task"], [4030, 4031, ["car"], "Task"], [4046, 4047, ["lane"], "Task"], [4056, 4057, ["lane"], "Task"], [4063, 4064, ["lane"], "Task"], [4137, 4138, ["lane"], "Task"], [4303, 4304, ["lane"], "Task"], [4316, 4317, ["lane"], "Task"], [4332, 4333, ["lane"], "Task"], [4355, 4356, ["car"], "Task"], [4518, 4519, ["lane"], "Task"], [4553, 4554, ["lane"], "Task"], [4669, 4670, ["IOU"], "Metric"], [4675, 4676, ["car"], "Task"], [4830, 4831, ["IOU"], "Metric"], [4962, 4963, ["detector"], "Method"], [5009, 5010, ["cars"], "Task"], [5019, 5020, ["cars"], "Task"], [5038, 5039, ["lane"], "Task"], [5039, 5040, ["detector"], "Method"], [5097, 5098, ["lane"], "Task"], [5108, 5109, ["CNN"], "Method"], [5116, 5117, ["cars"], "Task"], [5150, 5151, ["CNN"], "Method"], [5205, 5206, ["car"], "Task"]]} {"doc_id": "148eb2b649ce0f66214fceba5c061b417ab88d79", "text": ["document", ":", "Inferencing", "Based", "on", "Unsupervised", "Learning", "of", "Disentangled", "Representations", "Combining", "Generative", "Adversarial", "Networks", "(", "GANs", ")", "with", "encoders", "that", "learn", "to", "encode", "data", "points", "has", "shown", "promising", "results", "in", "learning", "data", "representations", "in", "an", "unsupervised", "way", ".", "We", "propose", "a", "framework", "that", "combines", "an", "encoder", "and", "a", "generator", "to", "learn", "disentangled", "representations", "which", "encode", "meaningful", "information", "about", "the", "data", "distribution", "without", "the", "need", "for", "any", "labels", ".", "While", "current", "approaches", "focus", "mostly", "on", "the", "generative", "aspects", "of", "GANs", ",", "our", "framework", "can", "be", "used", "to", "perform", "inference", "on", "both", "real", "and", "generated", "data", "points", ".", "Experiments", "on", "several", "data", "sets", "show", "that", "the", "encoder", "learns", "interpretable", ",", "disentangled", "representations", "which", "encode", "descriptive", "properties", "and", "can", "be", "used", "to", "sample", "images", "that", "exhibit", "specific", "characteristics", ".", "0", "AcceptedasaconferencepaperattheEuropeanSymposiumonArtificialNeural", "Networks", ",", "ComputationalIntelligenceandMachineLearning", "(", "ESANN", ")", "2018", "section", ":", "Introduction", "Learning", "meaningful", "representations", "of", "data", "is", "an", "important", "step", "for", "models", "to", "understand", "the", "world", ".", "Recently", ",", "the", "Generative", "Adversarial", "Network", "(", "GAN", ")", "has", "been", "proposed", "as", "a", "method", "that", "can", "learn", "characteristics", "of", "data", "distributions", "without", "the", "need", "for", "labels", ".", "GANs", "traditionally", "consist", "of", "a", "generator", ",", "which", "generates", "data", "from", "randomly", "sampled", "vectors", ",", "and", "a", "discriminator", ",", "which", "tries", "to", "distinguish", "generated", "data", "from", "real", "data", ".", "During", "training", ",", "the", "generator", "learns", "to", "generate", "realistic", "data", "samples", ",", "while", "the", "discriminator", "becomes", "better", "at", "distinguishing", "between", "the", "generated", "and", "the", "real", "data", ".", "As", "a", "result", ",", "both", "the", "generator", "and", "the", "discriminator", "learn", "characteristics", "about", "the", "underlying", "data", "distribution", "without", "the", "need", "for", "any", "labels", ".", "One", "desirable", "characteristic", "of", "learned", "representations", "is", "disentanglement", ",", "which", "means", "that", "different", "parts", "of", "the", "representation", "encode", "different", "factors", "of", "the", "data", "-", "generating", "distribution", ".", "This", "makes", "representations", "more", "interpretable", ",", "easier", "to", "modify", ",", "and", "is", "a", "useful", "property", "for", "many", "tasks", "such", "as", "classification", ",", "clustering", ",", "or", "image", "captioning", ".", "To", "achieve", "this", ",", "Chen", "et", "al", ".", "introduced", "a", "GAN", "variant", "in", "which", "the", "generator", "’s", "input", "is", "split", "into", "two", "parts", "and", ".", "Here", ",", "encodes", "unstructured", "noise", "while", "encodes", "meaningful", ",", "data", "-", "generating", "factors", ".", "Through", "enforcing", "high", "mutual", "information", "between", "and", "and", "the", "generated", "images", "the", "generator", "is", "trained", "using", "the", "inputs", "as", "meaningful", "encodings", "for", "certain", "image", "characteristics", ".", "For", "example", ",", "a", "ten", "-", "dimensional", "categorical", "code", "for", "could", "represent", "the", "ten", "different", "digit", "classes", "in", "the", "MNIST", "data", "set", ".", "Since", "no", "labels", "are", "provided", "the", "generator", "has", "to", "learn", "by", "itself", "which", "image", "characteristics", "can", "be", "represented", "through", ".", "One", "drawback", "of", "this", "model", "is", "that", "the", "only", "way", "to", "perform", "inference", ",", "i.e.", "map", "real", "data", "samples", "into", "a", "(", "disentangled", ")", "representation", ",", "is", "to", "use", "the", "discriminator", ".", "However", ",", "there", "is", "no", "guarantee", "that", "the", "discriminator", "learns", "good", "representations", "of", "the", "data", "in", "general", ",", "as", "it", "is", "trained", "to", "discriminate", "between", "real", "and", "generated", "data", "and", "may", "therefore", "focus", "only", "on", "features", "that", "are", "helpful", "for", "discriminating", "these", "two", ",", "but", "are", "not", "necessarily", "descriptive", "of", "the", "data", "distribution", "in", "general", ".", "Zhang", "et", "al", ".", "tried", "to", "enforce", "disentangled", "representations", "in", "order", "to", "improve", "the", "controllability", "of", "the", "generator", ".", "The", "latent", "representation", "is", "split", "up", "into", "two", "parts", "encoding", "meaningful", "information", "and", "unknown", "factors", "of", "variation", ".", "Two", "additional", "inference", "networks", "are", "introduced", "to", "enforce", "the", "disentanglement", "between", "the", "two", "parts", "of", "the", "latent", "representation", ".", "While", "this", "setup", "yields", "a", "better", "controllability", "over", "the", "generative", "process", "it", "depends", "on", "labeled", "samples", "for", "its", "training", "objective", "and", "can", "not", "discover", "unknown", "data", "-", "generating", "factors", ",", "but", "only", "encodes", "known", "factors", "of", "variation", "(", "obtained", "through", "labels", ")", "in", "its", "disentangled", "representation", ".", "Donahue", "et", "al", ".", "and", "Dumoulin", "et", "al", ".", "introduced", "an", "extension", "which", "includes", "an", "encoder", "that", "learns", "the", "encodings", "of", "real", "data", "samples", ".", "The", "discriminator", "gets", "as", "input", "both", "the", "data", "sample", "(", "either", "real", "or", "generated", ")", "and", "the", "according", "representation", "(", "either", "or", ")", "and", "has", "to", "classify", "them", "as", "either", "coming", "from", "the", "generator", "or", "the", "encoder", ".", "The", "generator", "and", "the", "encoder", "try", "to", "fool", "the", "discriminator", "into", "misclassifying", "the", "samples", ".", "As", "a", "result", ",", "the", "encoder", "learns", "to", "approximate", "the", "inverse", "of", "the", "generator", "and", "can", "be", "used", "to", "map", "real", "data", "samples", "into", "representations", "for", "other", "applications", ".", "However", ",", "in", "these", "approaches", "the", "representations", "follow", "a", "simple", "prior", ",", "e.g.", "a", "Gaussian", "or", "uniform", "distribution", ",", "and", "do", "not", "exhibit", "any", "disentangled", "properties", ".", "Our", "model", ",", "the", "Bidirectional", "-", "InfoGAN", ",", "integrates", "some", "of", "these", "approaches", "by", "extending", "traditional", "GANs", "with", "an", "encoder", "that", "learns", "disentangled", "representations", "in", "an", "unsupervised", "setting", ".", "After", "training", ",", "the", "encoder", "can", "map", "data", "points", "to", "meaningful", ",", "disentangled", "representations", "which", "can", "potentially", "be", "used", "for", "different", "tasks", "such", "as", "classification", ",", "clustering", ",", "or", "image", "captioning", ".", "Compared", "to", "the", "InfoGAN", "we", "introduce", "an", "encoder", "to", "mitigate", "the", "problems", "of", "using", "a", "discriminator", "for", "both", "the", "adversarial", "loss", "and", "the", "inference", "task", ".", "Unlike", "the", "Structured", "GAN", "our", "training", "procedure", "is", "completely", "unsupervised", ",", "can", "detect", "unknown", "data", "-", "generating", "factors", ",", "and", "only", "introduces", "one", "additional", "inference", "network", "(", "the", "encoder", ")", ".", "In", "contrast", "to", "the", "Bidirectional", "GAN", "we", "replace", "the", "simple", "prior", "on", "the", "latent", "representation", "with", "a", "distribution", "that", "is", "amenable", "to", "disentangled", "representations", "and", "introduce", "an", "additional", "loss", "for", "the", "encoder", "and", "the", "generator", "to", "achieve", "disentangled", "representations", ".", "On", "the", "MNIST", ",", "CelebA", ",", "and", "SVHN", "data", "sets", "we", "show", "that", "the", "encoder", "does", "learn", "interpretable", "representations", "which", "encode", "meaningful", "properties", "of", "the", "data", "distribution", ".", "Using", "these", "we", "can", "sample", "images", "that", "exhibit", "certain", "characteristics", ",", "e.g.", "digit", "identity", "and", "specific", "stroke", "widths", "for", "the", "MNIST", "data", "set", ",", "or", "different", "hair", "colors", "and", "clothing", "accessories", "in", "the", "CelebA", "data", "set", ".", "every", "picture", "/", ".style", "=", "line", "width=0.2", "mm", "[", "scale=0.39", "]", "[", "]", "at", "(", "2.5", ",", "9.5", ")", "features", ";", "fill", "=", "white", ",", "rounded", "corners=0.2", "cm", "]", "(", "0.6", ",", "1.5", ")", "rectangle", "(", "4.6", ",", "2.8", ")", "node", "[", "pos=.5", "]", ";", "[", "fill", "=", "white", ",", "rounded", "corners=0.2", "cm", "]", "(", "0.6", ",-", "0.2", ")", "rectangle", "(", "4.6", ",", "1.1", ")", "node", "[", "pos=.5", "]", ";", "[", "fill", "=", "gray", ",", "opacity=0.3", ",", "rounded", "corners=0.2", "cm", "]", "(", "0.6", ",", "6.65", ")", "rectangle", "(", "4.6", ",", "7.95", ")", "node", "[", "pos=.5", ",", "opacity=1", "]", ";", "[", "-", "¿", "]", "(", "4.6", ",", "7.3", ")", "–", "(", "7", ",", "7.3", ")", ";", "[", "-", "¿", "]", "(", "9", ",", "7.3", ")", "–", "(", "10.8", ",", "7.3", ")", ";", "(", "18.5", ",", "5.75", ")", "ellipse", "(", "2.7", "and", "1", ")", "node", "[", "]", ";", "[", "]", "at", "(", "13", ",", "9.5", ")", "data", ";", "fill", "=", "gray", ",", "opacity=0.3", ",", "rounded", "corners=0.2", "cm", "]", "(", "10.9", ",", "1.5", ")", "rectangle", "(", "14.9", ",", "2.8", ")", "node", "[", "pos=.5", ",", "opacity=1", "]", ";", "[", "-", "¿", "]", "(", "10.9", ",", "2.15", ")", "–", "(", "10.5", ",", "2.15", ")", "–", "(", "10.5", ",", "1.7", ")", "–", "(", "9", ",", "1.7", ")", ";", "[", "fill", "=", "white", ",", "rounded", "corners=0.2", "cm", "]", "(", "10.9", ",-", "0.2", ")", "rectangle", "(", "14.9", ",", "1.1", ")", "node", "[", "pos=.5", "]", ";", "[", "-", "¿", "]", "(", "10.9", ",", "0.45", ")", "–", "(", "10.5", ",", "0.45", ")", "–", "(", "10.5", ",", "0.9", ")", "–", "(", "9", ",", "0.9", ")", ";", "[", "fill", "=", "white", ",", "rounded", "corners=0.2", "cm", "]", "(", "10.9", ",", "6.65", ")", "rectangle", "(", "14.9", ",", "7.95", ")", "node", "[", "pos=.5", "]", ";", "[", "-", "¿", "]", "(", "2.5", ",", "6.6", ")", "–", "(", "2.5", ",", "5.8", ")", "–", "(", "15.8", ",", "5.8", ")", ";", "[", "]", "(", "13", ",", "6.6", ")", "–", "(", "13", ",", "5.8", ")", ";", "[", "]", "(", "13", ",", "2.8", ")", "–", "(", "13", ",", "3.4", ")", ";", "[", "-", "¿", "]", "(", "2.5", ",", "2.8", ")", "–", "(", "2.5", ",", "3.4", ")", "–", "(", "15.8", ",", "3.4", ")", "node", "[", "pos=.4", ",", "above", ",", "inner", "sep=0.5pt", "]", "if", "image", "node", "[", "pos=.4", ",", "below", ",", "inner", "sep=0.5pt", "]", "is", "real", ";", "[", "]", "(", "13", ",-", "1", ")", "–", "(", "13", ",-", "0.2", ")", ";", "[", "-", "¿", "]", "(", "2.5", ",-", "0.2", ")", "–", "(", "2.5", ",-", "1", ")", "–", "(", "15.8", ",-", "1", ")", "node", "[", "pos=.4", ",", "above", ",", "inner", "sep=0.5pt", "]", "if", "image", "node", "[", "pos=.4", ",", "below", ",", "inner", "sep=0.5pt", "]", "is", "generated", ";", "(", "18.5", ",", "3.35", ")", "ellipse", "(", "2.7", "and", "1", ")", "node", "[", "]", ";", "(", "19.8", ",-", "1", ")", "ellipse", "(", "4", "and", "1", ")", "node", "[", "]", ";", "[", "fill", "=", "yellow!30", "]", "(", "27.5", ",-", "1", ")", "ellipse", "(", "1.9", "and", "0.9", ")", "node", "[", "]", "node", "[", "above", ",", "inner", "sep=11pt", "]", "mutual", "information", ";", "[", "-", "¿", "]", "(", "23.8", ",-", "1", ")", "–", "(", "25.6", ",-", "1", ")", ";", "[", "fill", "=", "red!10", "]", "(", "7", ",", "8.3", ")", "–", "(", "8", ",", "8.3", ")", "–", "(", "9", ",", "7.3", ")", "–", "(", "8", ",", "6.3", ")", "–", "(", "7", ",", "6.3", ")", "–", "(", "7", ",", "8.3", ")", ";", "[", "]", "at", "(", "7.75", ",", "7.3", ")", ";", "[", "fill", "=", "red!10", "]", "(", "8", ",", "2.3", ")", "–", "(", "9", ",", "2.3", ")", "–", "(", "9", ",", "0.3", ")", "–", "(", "8", ",", "0.3", ")", "–", "(", "7", ",", "1.3", ")", "–", "(", "8", ",", "2.3", ")", ";", "[", "]", "at", "(", "8.25", ",", "1.3", ")", ";", "[", "-", "¿", "]", "(", "7", ",", "1.3", ")", "–", "(", "6", ",", "1.3", ")", "–", "(", "6", ",", "2.1", ")", "–", "(", "4.7", ",", "2.1", ")", ";", "[", "-", "¿", "]", "(", "6", ",", "1.3", ")", "–", "(", "6", ",", "0.5", ")", "–", "(", "4.7", ",", "0.5", ")", ";", "[", "fill", "=", "red!10", "]", "(", "23.5", ",", "5.6", ")", "–", "(", "22.5", ",", "5.6", ")", "–", "(", "22.5", ",", "3.6", ")", "–", "(", "23.5", ",", "3.6", ")", "–", "(", "24.5", ",", "4.6", ")", "–", "(", "23.5", ",", "5.6", ")", ";", "[", "]", "at", "(", "23.25", ",", "4.6", ")", ";", "[", "-", "¿", "]", "(", "21.2", ",", "5.75", ")", "–", "(", "21.5", ",", "5.75", ")", "–", "(", "21.5", ",", "5.2", ")", "–", "(", "22.5", ",", "5.2", ")", ";", "[", "-", "¿", "]", "(", "21.2", ",", "3.45", ")", "–", "(", "21.5", ",", "3.45", ")", "–", "(", "21.5", ",", "4.2", ")", "–", "(", "22.5", ",", "4.2", ")", ";", "[", "fill", "=", "yellow!30", "]", "(", "28", ",", "4.6", ")", "ellipse", "(", "2.5", "and", "0.9", ")", "node", "[", "]", "node", "[", "above", ",", "inner", "sep=11pt", "]", "adversarial", "cost", ";", "[", "-", "¿", "]", "(", "24.5", ",", "4.6", ")", "–", "(", "25.5", ",", "4.6", ")", ";", "section", ":", "Methodology", "Our", "model", ",", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "consists", "of", "a", "generator", ",", "a", "discriminator", ",", "and", "an", "encoder", ",", "which", "are", "implemented", "as", "neural", "networks", ".", "The", "input", "vector", "that", "is", "given", "to", "the", "generator", "is", "made", "up", "of", "two", "parts", ".", "Here", ",", "is", "sampled", "from", "a", "uniform", "distribution", ",", ",", "and", "is", "used", "to", "represent", "unstructured", "noise", "in", "the", "images", ".", "On", "the", "other", "hand", ",", "is", "the", "part", "of", "the", "representation", "that", "encodes", "meaningful", "information", "in", "a", "disentangled", "manner", "and", "is", "made", "up", "of", "both", "categorical", "values", "and", "continuous", "values", ".", "takes", "as", "input", "and", "transforms", "it", "into", "an", "image", ",", "i.e.", ".", "is", "a", "convolutional", "network", "that", "gets", "as", "input", "either", "real", "or", "fake", "images", "and", "encodes", "them", "into", "a", "latent", "representation", ".", "gets", "as", "input", "an", "image", "and", "the", "corresponding", "representation", "concatenated", "along", "the", "channel", "axis", ".", "It", "then", "tries", "to", "classify", "the", "pair", "as", "coming", "either", "from", "the", "generator", "or", "the", "encoder", ",", "i.e.", ",", "while", "both", "and", "try", "to", "fool", "the", "discriminator", "into", "misclassifying", "its", "input", ".", "As", "a", "result", "the", "original", "GAN", "minimax", "game", "is", "extended", "and", "becomes", ":", "where", "is", "the", "adversarial", "cost", "as", "depicted", "in", "Fig", ".", "[", "reference", "]", ".", "[", "b", "]", "0.8", "[", "b", "]", "0.8", "In", "order", "to", "force", "the", "generator", "to", "use", "the", "information", "provided", "in", "we", "maximize", "the", "mutual", "information", "between", "and", ".", "Maximizing", "the", "mutual", "information", "directly", "is", "hard", ",", "as", "it", "requires", "the", "posterior", "and", "we", "therefore", "follow", "the", "approach", "by", "Chen", "et", "al", ".", "and", "define", "an", "auxiliary", "distribution", "to", "approximate", ".", "We", "then", "maximize", "the", "lower", "bound", ",", "where", "is", "the", "mutual", "information", "depicted", "in", "Fig", ".", "[", "reference", "]", ".", "For", "simplicity", "reasons", "we", "fix", "the", "distribution", "over", "and", ",", "therefore", ",", "the", "entropy", "term", "is", "treated", "as", "a", "constant", ".", "In", "our", "case", "is", "the", "encoder", "network", "which", "gets", "images", "generated", "by", "as", "input", "and", "is", "trained", "to", "approximate", "the", "unknown", "posterior", ".", "For", "categorical", "we", "use", "the", "softmax", "nonlinearity", "to", "represent", "while", "we", "treat", "the", "posterior", "of", "continuous", "as", "a", "factored", "Gaussian", ".", "Given", "this", "structure", ",", "the", "minimax", "game", "for", "the", "Bidirectional", "-", "InfoGAN", "(", "BInfoGAN", ")", "is", "then", "where", "determines", "the", "strength", "of", "the", "impact", "of", "the", "mutual", "information", "criterion", "and", "is", "set", "to", "in", "all", "our", "experiments", ".", "section", ":", "Experiments", "We", "perform", "experiments", "on", "the", "MNIST", ",", "the", "CelebA", ",", "and", "the", "SVHN", "data", "set", ".", "While", "the", "final", "performance", "of", "the", "model", "is", "likely", "influenced", "by", "choosing", "the", "“", "optimal", "”", "characteristics", "for", "this", "is", "usually", "not", "possible", ",", "since", "we", "do", "not", "know", "all", "data", "-", "generating", "factors", "beforehand", ".", "When", "choosing", "the", "characteristics", "and", "dimensionality", "of", "the", "disentangled", "vector", "we", "therefore", "mostly", "stick", "with", "the", "values", "previously", "chosen", "by", "Chen", "et", "al", ".", ".", "For", "further", "information", "on", "the", "network", "architectures", "and", "more", "examples", "of", "the", "learned", "characteristics", "on", "the", "different", "data", "sets", "see", "our", "Git", ":", ".", "[", "b", "]", "0.8", "[", "b", "]", "0.8", "On", "the", "MNIST", "data", "set", "we", "model", "the", "latent", "code", "with", "one", "categorical", "variable", "and", "two", "continuous", "variables", ".", "During", "the", "optimization", "process", "and", "without", "the", "use", "of", "any", "labels", "the", "encoder", "learns", "to", "use", "to", "encode", "different", "digit", "classes", ",", "while", "and", "encode", "stroke", "width", "and", "digit", "rotation", ".", "Fig", ".", "[", "reference", "]", "shows", "images", "randomly", "sampled", "from", "the", "test", "set", "according", "to", "the", "ten", "different", "categorical", "values", ".", "We", "can", "see", "that", "the", "encoder", "has", "learned", "to", "reliably", "assign", "a", "different", "categorical", "value", "for", "different", "digits", ".", "Indeed", ",", "by", "manually", "matching", "the", "different", "categories", "in", "to", "a", "digit", "type", ",", "we", "achieve", "a", "test", "set", "accuracy", "of", "96.61", "%", "(", ",", "averaged", "over", "10", "independent", "runs", ")", "without", "ever", "using", "labels", "during", "the", "training", ",", "compared", "to", "Chen", "et", "al", ".", "(", "unsupervised", ")", "with", "an", "accuracy", "of", "95", "%", ",", "and", "Zhang", "et", "al", ".", "(", "semi", "-", "supervised", ",", "20", "labels", ")", "with", "an", "accuracy", "of", "96", "%", ".", "Fig", ".", "[", "reference", "]", "shows", "images", "sampled", "from", "the", "test", "set", "for", "different", "values", "of", "and", ".", "We", "see", "that", "we", "can", "use", "the", "encodings", "from", "to", "now", "sample", "for", "digits", "with", "certain", "characteristics", "such", "as", "stroke", "width", "and", "rotation", ",", "even", "though", "this", "information", "was", "not", "explicitly", "provided", "during", "training", ".", "On", "the", "CelebA", "data", "set", "the", "latent", "code", "is", "modeled", "with", "four", "categorical", "codes", "and", "four", "continuous", "variables", ".", "Again", ",", "the", "encoder", "learns", "to", "associate", "certain", "image", "characteristics", "with", "specific", "codes", "in", ".", "This", "includes", "characteristics", "such", "as", "the", "presence", "of", "glasses", ",", "hair", "color", ",", "and", "background", "color", "and", "is", "visualized", "in", "Fig", ".", "[", "reference", "]", ".", "On", "the", "SVHN", "data", "set", "we", "use", "the", "same", "network", "architecture", "and", "latent", "code", "representations", "as", "for", "the", "CelebA", "data", "set", ".", "Again", ",", "the", "encoder", "learns", "interpretable", ",", "disentangled", "representations", "encoding", "characteristics", "such", "as", "image", "background", ",", "contrast", "and", "digit", "type", ".", "See", "Fig", ".", "[", "reference", "]", "for", "examples", "sampled", "from", "the", "SVHN", "test", "set", ".", "These", "results", "indicate", "that", "the", "Bidirectional", "-", "InfoGAN", "is", "indeed", "capable", "of", "mapping", "data", "points", "into", "disentangled", "representations", "that", "encode", "meaningful", "characteristics", "in", "a", "completely", "unsupervised", "manner", ".", "section", ":", "Conclusion", "We", "showed", "that", "an", "encoder", "coupled", "with", "a", "generator", "in", "a", "Generative", "Adversarial", "Network", "can", "learn", "disentangled", "representations", "of", "the", "data", "without", "the", "need", "for", "any", "explicit", "labels", ".", "Using", "the", "encoder", "network", "we", "maximize", "the", "mutual", "information", "between", "certain", "parts", "of", "the", "generator", "’s", "input", "and", "the", "images", "that", "are", "generated", "from", "it", ".", "Through", "this", "the", "generator", "learns", "to", "associate", "certain", "image", "characteristics", "with", "specific", "parts", "of", "its", "input", ".", "Additionally", ",", "the", "adversarial", "cost", "from", "the", "discriminator", "forces", "both", "the", "generator", "to", "generate", "realistic", "looking", "images", "and", "the", "encoder", "to", "approximate", "the", "inverse", "of", "the", "generator", ",", "leading", "to", "disentangled", "representations", "that", "can", "be", "used", "for", "inference", ".", "The", "learned", "characteristics", "are", "often", "meaningful", "and", "humanly", "interpretable", ",", "and", "can", "potentially", "help", "with", "other", "tasks", "such", "as", "classification", "and", "clustering", ".", "Additionally", ",", "our", "method", "can", "be", "used", "as", "a", "pre", "-", "training", "step", "on", "unlabeled", "data", "sets", ",", "where", "it", "can", "lead", "to", "better", "representations", "for", "the", "final", "task", ".", "However", ",", "currently", "we", "have", "no", "influence", "over", "which", "characteristics", "are", "learned", "in", "the", "unsupervised", "setting", "which", "means", "that", "the", "model", "can", "also", "learn", "characteristics", "or", "features", "that", "are", "meaningless", "or", "not", "interpretable", "by", "humans", ".", "In", "the", "future", ",", "this", "can", "be", "mitigated", "by", "combining", "our", "approach", "with", "semi", "-", "supervised", "approaches", ",", "in", "which", "we", "can", "supply", "a", "limited", "amount", "of", "labels", "for", "the", "characteristics", "we", "are", "interested", "in", "to", "exert", "more", "control", "over", "which", "data", "-", "generating", "factors", "are", "learned", "while", "still", "being", "able", "to", "discover", "“", "new", "”", "generating", "factors", "which", "do", "not", "have", "to", "be", "known", "or", "specified", "beforehand", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["Inferencing"], "Task"], [5, 10, ["Unsupervised", "Learning", "of", "Disentangled", "Representations"], "Method"], [11, 14, ["Generative", "Adversarial", "Networks"], "Method"], [15, 16, ["GANs"], "Method"], [18, 19, ["encoders"], "Method"], [31, 33, ["data", "representations"], "Task"], [45, 46, ["encoder"], "Method"], [48, 49, ["generator"], "Method"], [51, 53, ["disentangled", "representations"], "Method"], [75, 77, ["generative", "aspects"], "Method"], [78, 79, ["GANs"], "Method"], [87, 88, ["inference"], "Task"], [104, 105, ["encoder"], "Method"], [108, 110, ["disentangled", "representations"], "Method"], [139, 143, ["meaningful", "representations", "of", "data"], "Task"], [157, 160, ["Generative", "Adversarial", "Network"], "Method"], [161, 162, ["GAN"], "Method"], [182, 183, ["GANs"], "Method"], [187, 188, ["generator"], "Method"], [199, 200, ["discriminator"], "Method"], [215, 216, ["generator"], "Method"], [225, 226, ["discriminator"], "Method"], [244, 245, ["generator"], "Method"], [247, 248, ["discriminator"], "Method"], [266, 268, ["learned", "representations"], "Method"], [309, 310, ["classification"], "Task"], [311, 312, ["clustering"], "Task"], [314, 316, ["image", "captioning"], "Task"], [401, 404, ["MNIST", "data", "set"], "Material"], [437, 438, ["inference"], "Task"], [455, 456, ["discriminator"], "Method"], [465, 466, ["discriminator"], "Method"], [533, 535, ["latent", "representation"], "Method"], [552, 554, ["inference", "networks"], "Method"], [566, 568, ["latent", "representation"], "Method"], [578, 580, ["generative", "process"], "Method"], [613, 615, ["disentangled", "representation"], "Method"], [631, 632, ["encoder"], "Method"], [642, 643, ["discriminator"], "Method"], [674, 675, ["generator"], "Method"], [677, 678, ["encoder"], "Method"], [680, 681, ["generator"], "Method"], [683, 684, ["encoder"], "Method"], [688, 689, ["discriminator"], "Method"], [699, 700, ["encoder"], "Method"], [707, 708, ["generator"], "Method"], [754, 757, ["Bidirectional", "-", "InfoGAN"], "Method"], [766, 767, ["GANs"], "Method"], [769, 770, ["encoder"], "Method"], [772, 774, ["disentangled", "representations"], "Method"], [776, 778, ["unsupervised", "setting"], "Task"], [783, 784, ["encoder"], "Method"], [791, 793, ["disentangled", "representations"], "Method"], [803, 804, ["classification"], "Task"], [805, 806, ["clustering"], "Task"], [808, 810, ["image", "captioning"], "Task"], [814, 815, ["InfoGAN"], "Method"], [818, 819, ["encoder"], "Method"], [826, 827, ["discriminator"], "Method"], [830, 832, ["adversarial", "loss"], "Task"], [834, 836, ["inference", "task"], "Task"], [842, 844, ["training", "procedure"], "Method"], [861, 863, ["inference", "network"], "Method"], [865, 866, ["encoder"], "Method"], [872, 874, ["Bidirectional", "GAN"], "Method"], [881, 883, ["latent", "representation"], "Method"], [890, 892, ["disentangled", "representations"], "Method"], [899, 900, ["encoder"], "Method"], [902, 903, ["generator"], "Method"], [905, 907, ["disentangled", "representations"], "Method"], [910, 911, ["MNIST"], "Material"], [922, 923, ["encoder"], "Method"], [925, 927, ["interpretable", "representations"], "Method"], [956, 959, ["MNIST", "data", "set"], "Material"], [1772, 1774, ["adversarial", "cost"], "Metric"], [1808, 1809, ["generator"], "Method"], [1811, 1812, ["discriminator"], "Method"], [1815, 1816, ["encoder"], "Method"], [1821, 1823, ["neural", "networks"], "Method"], [1832, 1833, ["generator"], "Method"], [1906, 1908, ["convolutional", "network"], "Method"], [1922, 1924, ["latent", "representation"], "Method"], [1952, 1953, ["generator"], "Method"], [1955, 1956, ["encoder"], "Method"], [2054, 2056, ["auxiliary", "distribution"], "Method"], [2105, 2107, ["encoder", "network"], "Method"], [2141, 2143, ["factored", "Gaussian"], "Method"], [2149, 2151, ["minimax", "game"], "Method"], [2153, 2156, ["Bidirectional", "-", "InfoGAN"], "Method"], [2157, 2158, ["BInfoGAN"], "Method"], [2170, 2173, ["mutual", "information", "criterion"], "Metric"], [2190, 2191, ["MNIST"], "Material"], [2267, 2269, ["network", "architectures"], "Method"], [2296, 2299, ["MNIST", "data", "set"], "Material"], [2315, 2317, ["optimization", "process"], "Task"], [2325, 2326, ["encoder"], "Method"], [2403, 2404, ["accuracy"], "Metric"], [2434, 2435, ["accuracy"], "Metric"], [2445, 2448, ["semi", "-", "supervised"], "Task"], [2454, 2455, ["accuracy"], "Metric"], [2534, 2535, ["encoder"], "Method"], [2581, 2583, ["network", "architecture"], "Method"], [2584, 2587, ["latent", "code", "representations"], "Method"], [2597, 2598, ["encoder"], "Method"], [2635, 2638, ["Bidirectional", "-", "InfoGAN"], "Method"], [2646, 2648, ["disentangled", "representations"], "Method"], [2655, 2657, ["unsupervised", "manner"], "Method"], [2665, 2666, ["encoder"], "Method"], [2669, 2670, ["generator"], "Method"], [2672, 2675, ["Generative", "Adversarial", "Network"], "Method"], [2677, 2679, ["disentangled", "representations"], "Method"], [2692, 2694, ["encoder", "network"], "Method"], [2740, 2741, ["discriminator"], "Method"], [2744, 2745, ["generator"], "Method"], [2752, 2753, ["encoder"], "Method"], [2763, 2765, ["disentangled", "representations"], "Method"], [2770, 2771, ["inference"], "Task"], [2804, 2808, ["pre", "-", "training", "step"], "Task"], [2839, 2841, ["unsupervised", "setting"], "Task"], [2874, 2878, ["semi", "-", "supervised", "approaches"], "Method"], [327, 328, ["GAN"], "Method"], [840, 841, ["GAN"], "Method"], [1977, 1978, ["GAN"], "Method"], [2791, 2792, ["classification"], "Task"]]} {"doc_id": "14cb3de564377cc4ba2b9bfe8ebd8c0a7c2787ce", "text": ["document", ":", "Taking", "a", "Deeper", "Look", "at", "Pedestrians", "In", "this", "paper", "we", "study", "the", "use", "of", "convolutional", "neural", "networks", "(", "convnets", ")", "for", "the", "task", "of", "pedestrian", "detection", ".", "Despite", "their", "recent", "diverse", "successes", ",", "convnets", "historically", "underperform", "compared", "to", "other", "pedestrian", "detectors", ".", "We", "deliberately", "omit", "explicitly", "modelling", "the", "problem", "into", "the", "network", "(", "e.g.", "parts", "or", "occlusion", "modelling", ")", "and", "show", "that", "we", "can", "reach", "competitive", "performance", "without", "bells", "and", "whistles", ".", "In", "a", "wide", "range", "of", "experiments", "we", "analyse", "small", "and", "big", "convnets", ",", "their", "architectural", "choices", ",", "parameters", ",", "and", "the", "influence", "of", "different", "training", "data", ",", "including", "pre", "-", "training", "on", "surrogate", "tasks", ".", "We", "present", "the", "best", "convnet", "detectors", "on", "the", "Caltech", "and", "KITTI", "dataset", ".", "On", "Caltech", "our", "convnets", "reach", "top", "performance", "both", "for", "the", "Caltech1x", "and", "Caltech10x", "training", "setup", ".", "Using", "additional", "data", "at", "training", "time", "our", "strongest", "convnet", "model", "is", "competitive", "even", "to", "detectors", "that", "use", "additional", "data", "(", "optical", "flow", ")", "at", "test", "time", ".", "=", "-", "1", "section", ":", "Introduction", "In", "recent", "years", "the", "field", "of", "computer", "vision", "has", "seen", "an", "explosion", "of", "success", "stories", "involving", "convolutional", "neural", "networks", "(", "convnets", ")", ".", "Such", "architectures", "currently", "provide", "top", "results", "for", "general", "object", "classification", ",", "general", "object", "detection", ",", "feature", "matching", ",", "stereo", "matching", ",", "scene", "recognition", ",", "pose", "estimation", ",", "action", "recognition", "and", "many", "other", "tasks", ".", "Pedestrian", "detection", "is", "a", "canonical", "case", "of", "object", "detection", "with", "relevant", "applications", "in", "car", "safety", ",", "surveillance", ",", "and", "robotics", ".", "A", "diverse", "set", "of", "ideas", "has", "been", "explored", "for", "this", "problem", "and", "established", "benchmark", "datasets", "are", "available", ".", "We", "would", "like", "to", "know", "if", "the", "success", "of", "convnets", "is", "transferable", "to", "the", "pedestrian", "detection", "task", ".", "Previous", "work", "on", "neural", "networks", "for", "pedestrian", "detection", "has", "relied", "on", "special", "-", "purpose", "designs", ",", "e.g.", "hand", "-", "crafted", "features", ",", "part", "and", "occlusion", "modelling", ".", "Although", "these", "proposed", "methods", "perform", "ably", ",", "current", "top", "methods", "are", "all", "based", "on", "decision", "trees", "learned", "via", "Adaboost", ".", "In", "this", "work", "we", "revisit", "the", "question", ",", "and", "show", "that", "both", "small", "and", "large", "vanilla", "convnets", "can", "reach", "top", "performance", "on", "the", "challenging", "Caltech", "pedestrians", "dataset", ".", "We", "provide", "extensive", "experiments", "regarding", "the", "details", "of", "training", ",", "network", "parameters", ",", "and", "different", "proposal", "methods", ".", "subsection", ":", "Related", "work", "Despite", "the", "popularity", "of", "the", "task", "of", "pedestrian", "detection", ",", "only", "few", "works", "have", "applied", "deep", "neural", "networks", "to", "this", "task", ":", "we", "are", "aware", "of", "only", "six", ".", "The", "first", "paper", "using", "convnets", "for", "pedestrian", "detection", "focuses", "on", "how", "to", "handle", "the", "limited", "training", "data", "(", "they", "use", "the", "INRIA", "dataset", ",", "which", "provides", "614", "positives", "and", "1218", "negative", "images", "for", "training", ")", ".", "First", ",", "each", "layer", "is", "initialized", "using", "a", "form", "of", "convolutional", "sparse", "coding", ",", "and", "the", "entire", "network", "is", "subsequently", "fine", "-", "tuned", "for", "the", "detection", "task", ".", "They", "propose", "an", "architecture", "that", "uses", "features", "from", "the", "last", "and", "second", "last", "layer", "for", "detection", ".", "This", "method", "is", "named", "ConvNet", ".", "A", "different", "line", "of", "work", "extends", "a", "deformable", "parts", "model", "(", "DPM", ")", "with", "a", "stack", "of", "Restricted", "Boltzmann", "Machines", "(", "RBMs", ")", "trained", "to", "reason", "about", "parts", "and", "occlusion", "(", "DBN", "-", "Isol", ")", ".", "This", "model", "was", "extended", "to", "account", "for", "person", "-", "to", "-", "person", "relations", "(", "DBN", "-", "Mut", ")", "and", "finally", "to", "jointly", "optimize", "all", "these", "aspects", ":", "JointDeep", "jointly", "optimizes", "features", ",", "parts", "deformations", ",", "occlusions", ",", "and", "person", "-", "to", "-", "person", "relations", ".", "The", "MultiSDP", "network", "feeds", "each", "layer", "with", "contextual", "features", "computed", "at", "different", "scales", "around", "the", "candidate", "pedestrian", "detection", ".", "Finally", "SDN", ",", "the", "current", "best", "performing", "convnet", "for", "pedestrian", "detection", ",", "uses", "additional", "‘", "‘", "switchable", "layers", "’", "’", "(", "RBM", "variants", ")", "to", "automatically", "learn", "both", "low", "-", "level", "features", "and", "high", "-", "level", "parts", "(", "e.g.", "‘", "‘", "head", "’", "’", ",", "‘", "‘", "legs", "’", "’", ",", "etc", ".", ")", ".", "Note", "that", "none", "of", "the", "existing", "papers", "rely", "on", "a", "‘", "‘", "straightforward", "’", "’", "convolutional", "network", "similar", "to", "the", "original", "LeNet", "(", "layers", "of", "convolutions", ",", "non", "-", "linearities", ",", "pooling", ",", "inner", "products", ",", "and", "a", "softmax", "on", "top", ")", ".", "We", "will", "revisit", "this", "decision", "in", "this", "paper", ".", "paragraph", ":", "Object", "detection", "Other", "than", "pedestrian", "detection", ",", "related", "convnets", "have", "been", "used", "for", "detection", "of", "ImageNet", "and", "Pascal", "VOC", "categories", ".", "The", "most", "successful", "general", "object", "detectors", "are", "based", "on", "variants", "of", "the", "R", "-", "CNN", "framework", ".", "Given", "an", "input", "image", ",", "a", "reduced", "set", "of", "detection", "proposals", "is", "created", ",", "and", "these", "are", "then", "evaluated", "via", "a", "convnet", ".", "This", "is", "essentially", "a", "two", "-", "stage", "cascade", "sliding", "window", "method", ".", "See", "for", "a", "review", "of", "recent", "proposal", "methods", ".", "paragraph", ":", "Detection", "proposals", "The", "most", "popular", "proposal", "method", "for", "generic", "objects", "is", "SelectiveSearch", ".", "The", "recent", "review", "also", "points", "out", "EdgeBoxes", "as", "a", "fast", "and", "effective", "method", ".", "For", "pedestrian", "detection", "DBN", "-", "Isol", "and", "DBN", "-", "Mut", "use", "DPM", "for", "proposals", ".", "JointDeep", ",", "MultiSDP", ",", "and", "SDN", "use", "a", "HOG", "+", "CSS", "+", "linear", "SVM", "detector", "(", "similar", "to", ")", "for", "proposals", ".", "Only", "ConvNet", "applies", "a", "convnet", "in", "a", "sliding", "fashion", ".", "paragraph", ":", "Decision", "forests", "Most", "methods", "proposed", "for", "pedestrian", "detection", "do", "not", "use", "convnets", "for", "detection", ".", "Leaving", "aside", "methods", "that", "use", "optical", "flow", ",", "the", "current", "top", "performing", "methods", "(", "on", "Caltech", "and", "KITTI", "datasets", ")", "are", "SquaresChnFtrs", ",", "InformedHaar", ",", "SpatialPooling", ",", "LDCF", ",", "and", "Regionlets", ".", "All", "of", "them", "are", "boosted", "decision", "forests", ",", "and", "can", "be", "considered", "variants", "of", "the", "integral", "channels", "features", "architecture", ".", "Regionlets", "and", "SpatialPooling", "use", "an", "large", "set", "of", "features", ",", "including", "HOG", ",", "LBP", "and", "CSS", ",", "while", "SquaresChnFtrs", ",", "InformedHaar", ",", "and", "LDCF", "build", "over", "HOG", "+", "LUV", ".", "On", "the", "Caltech", "benchmark", ",", "the", "best", "convnet", "(", "SDN", ")", "is", "outperformed", "by", "all", "aforementioned", "methods", ".", "paragraph", ":", "Input", "to", "convnets", "It", "is", "important", "to", "highlight", "that", "ConvNet", "learns", "to", "predict", "from", "YUV", "input", "pixels", ",", "whereas", "all", "other", "methods", "use", "additional", "hand", "-", "crafted", "features", ".", "DBN", "-", "Isol", "and", "DBN", "-", "Mut", "use", "HOG", "features", "as", "input", ".", "MultiSDP", "uses", "HOG", "+", "CSS", "features", "as", "input", ".", "JointDeep", "and", "SDN", "uses", "YUV", "+", "Gradients", "as", "input", "(", "and", "HOG", "+", "CSS", "for", "the", "detection", "proposals", ")", ".", "We", "will", "show", "in", "our", "experiments", "that", "good", "performance", "can", "be", "reached", "using", "RGB", "alone", ",", "but", "we", "also", "show", "that", "more", "sophisticated", "inputs", "systematically", "improve", "detection", "quality", ".", "Our", "data", "indicates", "that", "the", "antagonism", "‘", "‘", "hand", "-", "crafted", "features", "versus", "convnets", "’", "’", "is", "an", "illusion", ".", "subsection", ":", "Contributions", "In", "this", "paper", "we", "propose", "to", "revisit", "pedestrian", "detection", "with", "convolutional", "neural", "networks", "by", "carefully", "exploring", "the", "design", "space", "(", "number", "of", "layers", ",", "filter", "sizes", ",", "etc", ".", ")", ",", "and", "the", "critical", "implementation", "choices", "(", "training", "data", "preprocessing", ",", "effect", "of", "detections", "proposal", ",", "etc", ".", ")", ".", "We", "show", "that", "both", "small", "(", "parameters", ")", "and", "large", "(", "parameters", ")", "networks", "can", "reach", "good", "performance", "when", "trained", "from", "scratch", "(", "even", "when", "using", "the", "same", "data", "as", "previous", "methods", ")", ".", "We", "also", "show", "the", "benefits", "of", "using", "extended", "and", "external", "data", ",", "which", "leads", "to", "the", "strongest", "single", "-", "frame", "detector", "on", "Caltech", ".", "We", "report", "the", "best", "known", "performance", "for", "a", "convnet", "on", "the", "challenging", "Caltech", "dataset", "(", "improving", "by", "more", "than", "percent", "points", ")", ",", "and", "the", "first", "convnet", "results", "on", "the", "KITTI", "dataset", ".", "section", ":", "Training", "data", "It", "is", "well", "known", "that", "for", "convnets", "the", "volume", "of", "training", "data", "is", "quite", "important", "to", "reach", "good", "performance", ".", "Below", "are", "the", "datasets", "we", "consider", "along", "the", "paper", ".", "paragraph", ":", "Caltech", "The", "Caltech", "dataset", "and", "its", "associated", "benchmark", "is", "one", "of", "the", "most", "popular", "pedestrian", "detection", "datasets", ".", "It", "consists", "of", "videos", "captured", "from", "a", "car", "traversing", "U.S.", "streets", "under", "good", "weather", "conditions", ".", "The", "standard", "training", "set", "in", "the", "‘", "‘", "Reasonable", "’", "’", "setting", "consists", "of", "frames", "with", "annotated", "pedestrians", ",", "and", "the", "test", "set", "covers", "frames", "with", "pedestrians", ".", "paragraph", ":", "Caltech", "validation", "set", "In", "our", "experiments", "we", "also", "use", "Caltech", "training", "data", "for", "validation", ".", "For", "those", "experiments", "we", "use", "one", "of", "the", "suggested", "validation", "splits", ":", "the", "first", "five", "training", "videos", "are", "used", "for", "validation", "training", "and", "the", "sixth", "training", "video", "for", "validation", "testing", ".", "paragraph", ":", "Caltech10x", "Because", "the", "Caltech", "dataset", "videos", "are", "fully", "annotated", ",", "the", "amount", "of", "training", "data", "can", "be", "increased", "by", "resampling", "the", "videos", ".", "Inspired", "by", ",", "we", "increase", "the", "training", "data", "tenfold", "by", "sampling", "one", "out", "of", "three", "frames", "(", "instead", "of", "one", "out", "of", "thirty", "frames", "in", "the", "standard", "setup", ")", ".", "This", "yields", "annotated", "pedestrians", "for", "training", ",", "extracted", "from", "frames", ".", "paragraph", ":", "KITTI", "The", "KITTI", "dataset", "consists", "of", "videos", "captured", "from", "a", "car", "traversing", "German", "streets", ",", "also", "under", "good", "weather", "conditions", ".", "Although", "similar", "in", "appearance", "to", "Caltech", ",", "it", "has", "been", "shown", "to", "have", "different", "statistics", "(", "see", ")", ".", "Its", "training", "set", "contains", "pedestrians", "(", "taller", "than", "pixels", ")", "over", "frames", ",", "and", "its", "test", "set", "frames", ".", "paragraph", ":", "ImageNet", ",", "Places", "In", "section", "[", "reference", "]", "we", "will", "consider", "using", "large", "convnets", "that", "can", "exploit", "pre", "-", "training", "for", "surrogate", "tasks", ".", "We", "consider", "two", "such", "tasks", "(", "and", "their", "associated", "datasets", ")", ",", "the", "ImageNet", "2012", "classification", "of", "a", "thousand", "object", "categories", "and", "the", "classification", "of", "scene", "categories", ".", "The", "datasets", "provide", "and", "annotated", "images", "for", "training", ",", "respectively", ".", "section", ":", "From", "decision", "forests", "to", "neural", "networks", "Before", "diving", "into", "the", "experiments", ",", "it", "is", "worth", "noting", "that", "the", "proposal", "method", "we", "are", "using", "can", "be", "converted", "into", "a", "convnet", "so", "that", "the", "overall", "system", "can", "be", "seen", "as", "a", "cascade", "of", "two", "neural", "networks", ".", "SquaresChnFtrs", "is", "a", "decision", "forest", ",", "where", "each", "tree", "node", "pools", "and", "thresholds", "information", "from", "one", "out", "of", "several", "feature", "channels", ".", "As", "mentioned", "in", "section", "[", "reference", "]", "it", "is", "common", "practice", "to", "learn", "pedestrian", "detection", "convnets", "on", "handcrafted", "features", ",", "thus", "the", "feature", "channels", "need", "not", "be", "part", "of", "the", "conversion", ".", "In", "this", "case", ",", "a", "decision", "node", "can", "be", "realised", "using", "(", "i", ")", "a", "fully", "connected", "layer", "with", "constant", "non", "-", "zero", "weights", "corresponding", "to", "the", "original", "pooling", "region", "and", "zero", "weights", "elsewhere", ",", "(", "ii", ")", "a", "bias", "term", "that", "applies", "the", "threshold", ",", "(", "iii", ")", "and", "a", "sigmoid", "non", "-", "linearity", "that", "yields", "a", "decision", ".", "A", "two", "-", "layer", "network", "is", "sufficient", "to", "model", "a", "level", "-", "2", "decision", "tree", "given", "the", "three", "simulated", "node", "outputs", ".", "Finally", ",", "the", "weighted", "sum", "over", "the", "tree", "decisions", "can", "be", "modelled", "with", "yet", "another", "fully", "-", "connected", "layer", ".", "The", "mapping", "from", "SquaresChnFtrs", "to", "a", "deep", "neural", "network", "is", "exact", ":", "evaluating", "the", "same", "inputs", "it", "will", "return", "the", "exact", "same", "outputs", ".", "What", "is", "special", "about", "the", "resulting", "network", "is", "that", "it", "has", "not", "been", "trained", "by", "back", "-", "propagation", ",", "but", "by", "Adaboost", ".", "This", "network", "already", "performs", "better", "than", "the", "best", "known", "convnet", "on", "Caltech", ",", "SDN", ".", "Unfortunately", ",", "experiments", "to", "soften", "the", "non", "-", "linearities", "and", "use", "back", "-", "propagation", "to", "fine", "-", "tune", "the", "model", "parameters", "did", "not", "show", "significant", "improvements", ".", "section", ":", "Vanilla", "convolutional", "networks", "In", "our", "experience", "many", "convnet", "architectures", "and", "training", "hyper", "-", "parameters", "do", "not", "enable", "effective", "learning", "for", "diverse", "and", "challenging", "tasks", ".", "It", "is", "thus", "considered", "best", "practice", "to", "start", "exploration", "from", "architectures", "and", "parameters", "that", "are", "known", "to", "work", "well", "and", "progressively", "adapt", "it", "to", "the", "task", "at", "hand", ".", "This", "is", "the", "strategy", "of", "the", "following", "sections", ".", "In", "this", "section", "we", "first", "consider", "CifarNet", ",", "a", "small", "network", "designed", "to", "solve", "the", "CIFAR", "-", "10", "classification", "problem", "(", "objects", "categories", ",", "colour", "images", "of", ")", ".", "In", "section", "[", "reference", "]", "we", "consider", "AlexNet", ",", "a", "network", "that", "has", "times", "more", "parameters", "than", "CifarNet", "and", "designed", "to", "solve", "the", "ILSVRC2012", "classification", "problem", "(", "objects", "categories", ",", "colour", "images", "of", "VGA", "resolution", ")", ".", "Both", "of", "these", "networks", "were", "introduced", "in", "and", "are", "re", "-", "implemented", "in", "the", "open", "source", "Caffe", "project", ".", "Although", "pedestrian", "detection", "is", "quite", "a", "different", "task", "than", "CIFAR", "-", "10", ",", "we", "decide", "to", "start", "our", "exploration", "from", "the", "CifarNet", ",", "which", "provides", "fair", "performance", "on", "CIFAR", "-", "10", ".", "Its", "architecture", "is", "depicted", "in", "figure", "[", "reference", "]", ",", "unless", "otherwise", "specified", "we", "use", "raw", "RGB", "input", ".", "We", "first", "discuss", "how", "to", "use", "the", "CifarNet", "network", "(", "section", "[", "reference", "]", ")", ".", "This", "naive", "approach", "already", "improves", "over", "the", "best", "known", "convnets", "(", "section", "[", "reference", "]", ")", ".", "Sections", "[", "reference", "]", "and", "[", "reference", "]", "explore", "the", "design", "space", "around", "CifarNet", "and", "further", "push", "the", "detection", "quality", ".", "All", "models", "in", "this", "section", "are", "trained", "using", "Caltech", "data", "only", "(", "see", "section", "[", "reference", "]", ")", ".", "subsection", ":", "How", "to", "use", "CifarNet", "?", "Given", "an", "initial", "network", "specification", ",", "there", "are", "still", "several", "design", "choices", "that", "affect", "the", "final", "detection", "quality", ".", "We", "discuss", "some", "of", "them", "in", "the", "following", "paragraphs", ".", "paragraph", ":", "Detection", "proposals", "Unless", "otherwise", "specified", "we", "use", "the", "SquaresChnFtrs", "detector", "to", "generate", "proposals", "because", ",", "at", "the", "time", "of", "writing", ",", "it", "is", "the", "best", "performing", "pedestrian", "detector", "(", "on", "Caltech", ")", "with", "source", "code", "available", ".", "In", "figure", "[", "reference", "]", "we", "compare", "SquaresChnFtrs", "against", "EdgeBoxes", ",", "a", "state", "of", "the", "art", "class", "-", "agnostic", "proposal", "method", ".", "Using", "class", "-", "specific", "proposals", "allows", "to", "reduce", "the", "number", "of", "proposals", "by", "three", "orders", "of", "magnitude", ".", "paragraph", ":", "Thresholds", "for", "positive", "and", "negative", "samples", "Given", "both", "training", "proposals", "and", "ground", "truth", "(", "GT", ")", "annotations", ",", "we", "now", "consider", "which", "training", "label", "to", "assign", "to", "each", "proposal", ".", "A", "proposal", "is", "considered", "to", "be", "a", "positive", "example", "if", "it", "exceeds", "a", "certain", "Intersection", "-", "over", "-", "Union", "(", "IoU", ")", "threshold", "for", "at", "least", "one", "GT", "annotation", ".", "It", "is", "considered", "negative", "if", "it", "does", "not", "exceed", "a", "second", "IoU", "threshold", "for", "any", "GT", "annotation", ",", "and", "is", "ignored", "otherwise", ".", "We", "find", "that", "using", "GT", "annotations", "as", "positives", "is", "beneficial", "(", "i.e.", "not", "applying", "significant", "jitter", ")", ".", "paragraph", ":", "Model", "window", "size", "A", "typical", "choice", "for", "pedestrian", "detectors", "is", "a", "model", "window", "of", "in", "which", "the", "pedestrian", "occupies", "an", "area", "of", ".", "It", "is", "unclear", "that", "this", "is", "the", "ideal", "input", "size", "for", "convnets", ".", "Despite", "CifarNet", "being", "designed", "to", "operate", "over", ",", "table", "[", "reference", "]", "shows", "that", "a", "model", "size", "of", "indeed", "works", "best", ".", "We", "experimented", "with", "other", "variants", "(", "stretching", "versus", "cropping", ",", "larger", "context", "border", ")", "with", "no", "clear", "improvement", ".", "paragraph", ":", "Training", "batch", "In", "a", "detection", "setup", ",", "training", "samples", "are", "typically", "highly", "imbalanced", "towards", "the", "background", "class", ".", "Although", "in", "our", "validation", "setup", "the", "imbalance", "is", "limited", "(", "see", "table", "[", "reference", "]", ")", ",", "we", "found", "it", "beneficial", "throughout", "our", "experiments", "to", "enforce", "a", "strict", "ratio", "of", "positive", "to", "negative", "examples", "per", "batch", "of", "the", "stochastic", "gradient", "descend", "optimisation", ".", "The", "final", "performance", "is", "not", "sensitive", "to", "this", "parameter", "as", "long", "as", "some", "ratio", "(", "vs.", "None", ")", "is", "maintained", ".", "We", "use", "a", "ratio", "of", ".", "In", "the", "supplementary", "material", "we", "detail", "all", "other", "training", "parameters", ".", "subsection", ":", "How", "far", "can", "we", "get", "with", "the", "CifarNet", "?", "Given", "the", "parameter", "selection", "on", "the", "validation", "set", "from", "previous", "sections", ",", "how", "does", "CifarNet", "compare", "to", "previously", "reported", "convnet", "results", "on", "the", "Caltech", "test", "set", "?", "In", "table", "[", "reference", "]", "and", "figure", "[", "reference", "]", ",", "we", "see", "that", "our", "naive", "network", "right", "away", "improves", "over", "the", "best", "known", "convnet", "(", "versus", "SDN", ")", ".", "To", "decouple", "the", "contribution", "of", "our", "strong", "SquaresChnFtrs", "proposals", "to", "the", "CifarNet", "performance", ",", "we", "also", "train", "a", "CifarNet", "using", "the", "proposal", "from", "JointDeep", ".", "When", "using", "the", "same", "detection", "proposals", "at", "training", "and", "test", "time", ",", "the", "vanilla", "CifarNet", "already", "improves", "over", "both", "custom", "-", "designed", "JointDeep", "and", "SDN", ".", "Our", "CifarNet", "results", "are", "surprisingly", "close", "to", "the", "best", "known", "pedestrian", "detector", "trained", "on", "Caltech1x", "(", "versus", "SpatialPooling", ")", ".", "subsection", ":", "Exploring", "different", "architectures", "Encouraged", "by", "our", "initial", "results", ",", "we", "proceed", "to", "explore", "different", "parameters", "for", "the", "CifarNet", "architecture", ".", "subsubsection", ":", "Number", "and", "size", "of", "convolutional", "filters", "Using", "the", "Caltech", "validation", "set", "we", "perform", "a", "swipe", "of", "convolutional", "filter", "sizes", "(", ",", ",", "or", "pixels", ")", "and", "number", "of", "filters", "at", "each", "layer", "(", ",", ",", "or", "filters", ")", ".", "We", "include", "the", "full", "table", "in", "the", "supplementary", "material", ".", "We", "observe", "that", "using", "large", "filter", "sizes", "hurts", "quality", ",", "while", "the", "varying", "the", "number", "of", "filters", "shows", "less", "impact", ".", "Although", "some", "fluctuation", "in", "miss", "-", "rate", "is", "observed", ",", "overall", "there", "is", "no", "clear", "trend", "indicating", "that", "a", "configuration", "is", "clearly", "better", "than", "another", ".", "Thus", ",", "for", "sake", "of", "simplicity", ",", "we", "keep", "using", "CifarNet", "(", "-", "-", "filters", "of", "pixel", ")", "in", "the", "subsequent", "experiments", ".", "subsubsection", ":", "Number", "and", "type", "of", "layers", "In", "table", "[", "reference", "]", "we", "evaluate", "the", "effect", "of", "changing", "the", "number", "and", "type", "of", "layers", ",", "while", "keeping", "other", "CifarNet", "parameters", "fix", ".", "Besides", "convolutional", "layers", "(", "CONV", ")", "and", "fully", "-", "connected", "layers", "(", "FC", ")", ",", "we", "also", "consider", "locally", "-", "connected", "layers", "(", "LC", ")", ",", "and", "concatenating", "features", "across", "layers", "(", "CONCAT23", ")", "(", "used", "in", "ConvNet", ")", ".", "None", "of", "the", "considered", "architecture", "changes", "improves", "over", "the", "original", "three", "convolutional", "layers", "of", "CifarNet", ".", "subsection", ":", "Input", "channels", "As", "discussed", "in", "section", "[", "reference", "]", ",", "the", "majority", "of", "previous", "convnets", "for", "pedestrian", "detection", "use", "gradient", "and", "colour", "features", "as", "input", ",", "instead", "of", "raw", "RGB", ".", "In", "table", "[", "reference", "]", "we", "evaluate", "the", "effect", "of", "different", "input", "features", "over", "CifarNet", ".", "It", "seems", "that", "HOG", "+", "L", "channel", "provide", "a", "small", "advantage", "over", "RGB", ".", "For", "purposes", "of", "direct", "comparison", "with", "the", "large", "networks", ",", "in", "the", "next", "sections", "we", "keep", "using", "raw", "RGB", "as", "input", "for", "our", "CifarNet", "experiments", ".", "We", "report", "the", "CifarNet", "test", "set", "results", "in", "section", "[", "reference", "]", ".", "section", ":", "Large", "convolutional", "network", "One", "appealing", "characteristic", "of", "convnets", "is", "their", "ability", "to", "scale", "in", "size", "of", "training", "data", "volume", ".", "In", "this", "section", "we", "explore", "larger", "networks", "trained", "with", "more", "data", ".", "We", "base", "our", "experiments", "on", "the", "R", "-", "CNN", "[", "]", "approach", ",", "which", "is", "currently", "one", "of", "the", "best", "performer", "on", "the", "Pascal", "VOC", "detection", "task", ".", "We", "are", "thus", "curious", "to", "evaluate", "its", "performance", "for", "pedestrian", "detection", ".", "subsection", ":", "Surrogate", "tasks", "for", "improved", "detections", "The", "R", "-", "CNN", "approach", "(", "‘", "‘", "Regions", "with", "CNN", "features", "’", "’", ")", "wraps", "the", "large", "network", "previously", "trained", "for", "the", "ImageNet", "classification", "task", ",", "which", "we", "refer", "to", "as", "AlexNet", "(", "see", "figure", "[", "reference", "]", ")", ".", "We", "also", "use", "‘", "‘", "AlexNet", "’", "’", "as", "shorthand", "for", "‘", "‘", "R", "-", "CNN", "with", "AlexNet", "’", "’", "with", "the", "distinction", "made", "clear", "by", "the", "context", ".", "During", "R", "-", "CNN", "training", "AlexNet", "is", "fine", "-", "tuned", "for", "the", "(", "pedestrian", ")", "detection", "task", ",", "and", "in", "a", "second", "step", ",", "the", "softmax", "output", "is", "replaced", "by", "a", "linear", "SVM", ".", "Unless", "otherwise", "specified", ",", "we", "use", "the", "default", "parameters", "of", "the", "open", "source", ",", "Caffe", "based", ",", "R", "-", "CNN", "implementation", ".", "Like", "in", "the", "previous", "sections", ",", "we", "use", "SquaresChnFtrs", "for", "detection", "proposals", ".", "paragraph", ":", "Pre", "-", "training", "If", "we", "only", "train", "the", "top", "layer", "SVM", ",", "without", "fine", "-", "tuning", "the", "lower", "layers", "of", "AlexNet", ",", "we", "obtain", "on", "the", "Caltech", "test", "set", ".", "This", "is", "already", "surprisingly", "close", "to", "the", "best", "known", "convnet", "for", "the", "task", "(", "SDN", ")", ".", "When", "fine", "-", "tuning", "all", "layers", "on", "Caltech", ",", "the", "test", "set", "performance", "increases", "dramatically", ",", "reaching", ".", "This", "confirms", "the", "effectiveness", "of", "the", "general", "R", "-", "CNN", "recipe", "for", "detection", "(", "train", "AlexNet", "on", "ImageNet", ",", "fine", "-", "tune", "for", "the", "task", "of", "interest", ")", ".", "In", "table", "[", "reference", "]", "we", "investigate", "the", "influence", "of", "the", "pre", "-", "training", "task", "by", "considering", "AlexNets", "that", "have", "been", "trained", "for", "scene", "recognition", "(", "‘", "‘", "Places", "’", "’", ",", "see", "section", "[", "reference", "]", ")", "and", "on", "both", "Places", "and", "ImageNet", "(", "‘", "‘", "Hybrid", "’", "’", ")", ".", "‘", "‘", "Places", "’", "’", "provides", "results", "close", "to", "ImageNet", ",", "suggesting", "that", "the", "exact", "pre", "-", "training", "task", "is", "not", "critical", "and", "that", "there", "is", "nothing", "special", "about", "ImageNet", ".", "paragraph", ":", "Caltech10x", "Due", "to", "the", "large", "number", "of", "parameters", "of", "AlexNet", ",", "we", "consider", "providing", "additional", "training", "data", "using", "Caltech10x", "for", "fine", "-", "tuning", "the", "network", "(", "see", "section", "[", "reference", "]", ")", ".", "Despite", "the", "strong", "correlation", "across", "training", "samples", ",", "we", "do", "observe", "further", "improvement", "(", "see", "table", "[", "reference", "]", ")", ".", "Interestingly", ",", "the", "bulk", "of", "the", "improvement", "is", "due", "to", "more", "pedestrians", "(", "Positives10x", ",", "uses", "positives", "from", "Caltech10x", "and", "negatives", "from", "Caltech1x", ")", ".", "Our", "top", "result", ",", ",", "makes", "our", "AlexNet", "setup", "the", "best", "reported", "single", "-", "frame", "detector", "on", "Caltech", "(", "i.e.", "no", "optical", "flow", ")", ".", "subsection", ":", "Caltech", "-", "only", "training", "To", "compare", "with", "CifarNet", ",", "and", "to", "verify", "whether", "pre", "-", "training", "is", "necessary", "at", "all", ",", "we", "train", "AlexNet", "‘", "‘", "from", "scratch", "’", "’", "using", "solely", "the", "Caltech", "training", "data", ".", "We", "collect", "results", "in", "table", "[", "reference", "]", ".", "Training", "AlexNet", "solely", "on", "Caltech", ",", "yields", ",", "which", "improves", "over", "the", "proposals", "(", "SquaresChnFtrs", ")", "and", "the", "previous", "best", "known", "convnet", "on", "Caltech", "(", "SDN", ")", ".", "Using", "Caltech10x", "further", "improves", "the", "performance", ",", "down", "to", ".", "Although", "these", "numbers", "are", "inferior", "than", "the", "ones", "obtained", "with", "ImageNet", "pre", "-", "training", "(", ",", "see", "table", "[", "reference", "]", ")", ",", "we", "can", "get", "surprisingly", "competitive", "results", "using", "only", "pedestrian", "data", "despite", "the", "free", "parameters", "of", "the", "AlexNet", "model", ".", "AlexNet", "with", "Caltech10x", "is", "second", "best", "known", "single", "-", "frame", "pedestrian", "detector", "on", "Caltech", "(", "best", "known", "is", "LDCF", ",", "which", "also", "uses", "Caltech10x", ")", ".", "subsection", ":", "Additional", "experiments", "paragraph", ":", "How", "many", "layers", "?", "So", "far", "all", "experiments", "use", "the", "default", "parameters", "of", "R", "-", "CNN", ".", "Previous", "works", "have", "reported", "that", ",", "depending", "on", "the", "task", ",", "using", "features", "from", "lower", "AlexNet", "layers", "can", "provide", "better", "results", ".", "Table", "[", "reference", "]", "reports", "Caltech", "validation", "results", "when", "training", "the", "SVM", "output", "layer", "on", "top", "of", "layers", "four", "to", "seven", "(", "see", "figure", "[", "reference", "]", ")", ".", "We", "report", "results", "when", "using", "the", "default", "parameters", "and", "parameters", "that", "have", "been", "optimised", "by", "grid", "search", "(", "detailed", "grid", "search", "included", "in", "supplementary", "material", ")", ".", "We", "observe", "a", "negligible", "difference", "between", "default", "and", "optimized", "parameter", "(", "at", "most", "percent", "points", ")", ".", "Results", "for", "default", "parameters", "exhibit", "a", "slight", "trend", "of", "better", "performance", "for", "higher", "levels", ".", "These", "validation", "set", "results", "indicate", "that", ",", "for", "pedestrian", "detection", ",", "the", "R", "-", "CNN", "default", "parameters", "are", "a", "good", "choice", "overall", ".", "paragraph", ":", "Effect", "of", "proposal", "method", "When", "comparing", "the", "performance", "of", "AlexNet", "fine", "-", "tuned", "on", "Caltech1x", "to", "the", "proposal", "method", ",", "we", "see", "an", "improvement", "of", "(", "percent", "points", ")", "in", "miss", "-", "rate", ".", "In", "table", "[", "reference", "]", "we", "study", "the", "impact", "of", "using", "weaker", "or", "stronger", "proposals", ".", "Both", "ACF", "and", "SquaresChnFtrs", "provide", "source", "code", ",", "allowing", "us", "to", "generate", "training", "proposals", ".", "Katamari", "and", "SpatialPooling", "+", "[", "]", "are", "current", "top", "performers", "on", "the", "Caltech", "dataset", ",", "both", "using", "optical", "flow", ",", "i.e.", "additional", "information", "at", "test", "time", ".", "There", "is", "a", "gap", "between", "the", "detectors", "ACF", ",", "SquaresChnFtrs", ",", "and", "Katamari", "/", "SpatialPooling", ",", "allowing", "us", "to", "cover", "different", "operating", "points", ".", "The", "results", "of", "table", "[", "reference", "]", "indicate", "that", ",", "despite", "the", "gap", ",", "there", "is", "no", "noticeable", "difference", "between", "AlexNet", "models", "trained", "with", "ACF", "or", "SquaresChnFtrs", ".", "It", "is", "seems", "that", "as", "long", "as", "the", "proposals", "are", "not", "random", "(", "see", "top", "row", "of", "table", "[", "reference", "]", ")", ",", "the", "obtained", "quality", "is", "rather", "stable", ".", "The", "results", "also", "indicate", "that", "the", "quality", "improvement", "from", "AlexNet", "saturates", "around", ".", "Using", "stronger", "proposals", "does", "not", "lead", "to", "further", "improvement", ".", "This", "means", "that", "the", "discriminative", "power", "of", "our", "trained", "AlexNet", "is", "on", "par", "with", "the", "best", "known", "models", "on", "the", "Caltech", "dataset", ",", "but", "does", "not", "overtake", "them", ".", "paragraph", ":", "KITTI", "test", "set", "In", "figure", "[", "reference", "]", "we", "show", "performance", "of", "the", "AlexNet", "in", "context", "of", "the", "KITTI", "pedestrian", "detection", "benchmark", ".", "The", "network", "is", "pre", "-", "trained", "on", "ImageNet", "and", "fine", "-", "tuned", "using", "KITTI", "training", "data", ".", "SquaresChnFtrs", "reaches", "(", "average", "precision", ")", ",", "which", "the", "AlexNet", "can", "improve", "to", ".", "These", "are", "the", "first", "published", "results", "for", "convnets", "on", "the", "KITTI", "pedestrian", "detection", "dataset", ".", "subsection", ":", "Error", "analysis", "Results", "from", "the", "previous", "section", "are", "encouraging", ",", "but", "not", "as", "good", "as", "could", "be", "expected", "from", "looking", "at", "improvements", "on", "Pascal", "VOC", ".", "So", "what", "bounds", "performance", "?", "The", "proposal", "method", "?", "The", "localization", "quality", "of", "the", "convnet", "?", "Looking", "at", "the", "highest", "scoring", "false", "positives", "paints", "a", "picture", "of", "localization", "errors", "of", "the", "proposal", "method", ",", "the", "R", "-", "CNN", ",", "and", "even", "the", "ground", "truth", ".", "To", "quantify", "this", "effect", "we", "rerun", "the", "Caltech", "evaluation", "but", "remove", "all", "false", "positives", "that", "touch", "an", "annotation", ".", "This", "experiment", "provides", "an", "upper", "bound", "on", "performance", "when", "solving", "localisation", "issues", "in", "detectors", "and", "doing", "perfect", "non", "-", "maximum", "suppression", ".", "We", "see", "a", "surprisingly", "consistent", "improvement", "for", "all", "methods", "of", "about", "MR", ".", "This", "means", "that", "the", "intuition", "we", "gathered", "from", "looking", "at", "false", "positives", "is", "wrong", "and", "actually", "almost", "all", "of", "the", "mistakes", "that", "worsen", "the", "MR", "are", "actually", "background", "windows", "that", "are", "mistaken", "for", "pedestrians", ".", "What", "is", "striking", "about", "this", "result", "is", "that", "this", "is", "not", "just", "the", "case", "for", "our", "R", "-", "CNN", "experiments", "on", "detection", "proposals", "but", "also", "for", "methods", "that", "are", "trained", "as", "a", "sliding", "window", "detector", ".", "section", ":", "Small", "or", "big", "convnet", "?", "Since", "we", "have", "analysed", "the", "CifarNet", "and", "AlexNet", "separately", ",", "we", "compare", "their", "performance", "in", "this", "section", "side", "by", "side", ".", "Table", "[", "reference", "]", "shows", "performance", "on", "the", "Caltech", "test", "set", "for", "models", "that", "have", "been", "trained", "only", "on", "Caltech1x", "and", "Caltech10x", ".", "With", "less", "training", "data", "the", "CifarNet", "reaches", "MR", ",", "performing", "2", "percent", "points", "better", "than", "the", "AlexNet", ".", "On", "Caltech10x", ",", "we", "find", "the", "CifarNet", "performance", "improved", "to", ",", "while", "the", "AlexNet", "improves", "to", "MR", ".", "The", "trend", "confirms", "the", "intuition", "that", "models", "with", "lower", "capacity", "saturate", "earlier", "when", "increasing", "the", "amount", "of", "training", "data", "than", "models", "with", "higher", "capacity", ".", "We", "can", "also", "conclude", "that", "the", "AlexNet", "would", "profit", "from", "better", "regularisation", "when", "training", "on", "Caltech1x", ".", "paragraph", ":", "Timing", "The", "runtime", "during", "detection", "is", "about", "3ms", "per", "proposal", "window", ".", "This", "is", "too", "slow", "for", "sliding", "window", "detection", ",", "but", "given", "a", "fast", "proposal", "method", "that", "has", "high", "recall", "with", "less", "than", "windows", "per", "image", ",", "scoring", "takes", "about", "300ms", "per", "image", ".", "In", "our", "experience", "SquaresChnFtrs", "runs", "in", "2s", "per", "image", ",", "so", "proposing", "detections", "takes", "most", "of", "the", "detection", "time", ".", "section", ":", "Takeaways", "Previous", "work", "suggests", "that", "convnets", "for", "pedestrian", "detection", "underperform", ",", "despite", "having", "involved", "architectures", "(", "see", "for", "a", "survey", "of", "pedestrian", "detection", ")", ".", "In", "this", "paper", "we", "showed", "that", "neither", "has", "to", "be", "the", "case", ".", "We", "present", "a", "wide", "range", "of", "experiments", "with", "two", "off", "-", "the", "-", "shelf", "models", "that", "reach", "competitive", "performance", ":", "the", "small", "CifarNet", "and", "the", "big", "AlexNet", ".", "We", "present", "two", "networks", "that", "are", "trained", "on", "Caltech", "only", ",", "which", "outperform", "all", "previously", "published", "convnets", "on", "Caltech", ".", "The", "CifarNet", "shows", "better", "performance", "than", "related", "work", ",", "even", "when", "using", "the", "same", "training", "data", "as", "the", "respective", "methods", "(", "section", "[", "reference", "]", ")", ".", "Despite", "its", "size", ",", "the", "AlexNet", "also", "improves", "over", "all", "convnets", "even", "when", "it", "is", "trained", "on", "Caltech", "only", "(", "section", "[", "reference", "]", ")", ".", "We", "push", "the", "state", "of", "the", "art", "for", "pedestrian", "detectors", "that", "have", "been", "trained", "on", "Caltech1x", "and", "Caltech10x", ".", "The", "CifarNet", "is", "the", "best", "single", "-", "frame", "pedestrian", "detector", "that", "has", "been", "trained", "on", "Caltech1x", "(", "section", "[", "reference", "]", ")", ",", "while", "AlexNet", "is", "the", "best", "single", "-", "frame", "pedestrian", "detector", "trained", "on", "Caltech10x", "(", "section", "[", "reference", "]", ")", ".", "In", "figure", "[", "reference", "]", ",", "we", "include", "include", "all", "published", "methods", "on", "Caltech", "into", "the", "comparison", ",", "which", "also", "adds", "methods", "that", "use", "additional", "information", "at", "test", "time", ".", "The", "AlexNet", "that", "has", "been", "pre", "-", "trained", "on", "ImageNet", "reaches", "competitive", "results", "to", "the", "best", "published", "methods", ",", "but", "without", "using", "additional", "information", "at", "test", "time", "(", "section", "[", "reference", "]", ")", ".", "We", "report", "first", "results", "for", "convnets", "on", "the", "KITTI", "pedestrian", "detection", "benchmark", ".", "The", "AlexNet", "improves", "over", "the", "proposal", "method", "alone", ",", "delivering", "encouraging", "results", "to", "further", "push", "KITTI", "performance", "with", "convnets", ".", "section", ":", "Conclusion", "We", "have", "presented", "extensive", "and", "systematic", "experimental", "evidence", "on", "the", "effectiveness", "of", "convnets", "for", "pedestrian", "detection", ".", "Compared", "to", "previous", "convnets", "applied", "to", "pedestrian", "detection", "our", "approach", "avoids", "custom", "designs", ".", "When", "using", "the", "exact", "same", "proposals", "and", "training", "data", "as", "previous", "approaches", "our", "‘", "‘", "vanilla", "’", "’", "networks", "outperform", "previous", "results", ".", "We", "have", "shown", "that", "with", "pre", "-", "training", "on", "surrogate", "tasks", ",", "convnets", "can", "reach", "top", "performance", "on", "this", "task", ".", "Interestingly", "we", "have", "shown", "that", "even", "without", "pre", "-", "training", "competitive", "results", "can", "be", "achieved", ",", "and", "this", "result", "is", "quite", "insensitive", "to", "the", "model", "size", "(", "from", "to", "parameters", ")", ".", "Our", "experiments", "also", "detail", "which", "parameters", "are", "most", "critical", "to", "achieve", "top", "performance", ".", "We", "report", "the", "best", "known", "results", "for", "convnets", "on", "both", "the", "challenging", "Caltech", "and", "KITTI", "datasets", ".", "Our", "experience", "with", "convnets", "indicate", "that", "they", "show", "good", "promise", "on", "pedestrian", "detection", ",", "and", "that", "reported", "best", "practices", "do", "transfer", "to", "said", "task", ".", "That", "being", "said", ",", "on", "this", "more", "mature", "field", "we", "do", "not", "observe", "the", "large", "improvement", "seen", "on", "datasets", "such", "as", "Pascal", "VOC", "and", "ImageNet", ".", "paragraph", ":", "Acknowledgements", "We", "thank", "Shanshan", "Zhang", "for", "the", "help", "provided", "setting", "up", "some", "of", "the", "experiments", ".", "bibliography", ":", "References", "appendix", ":", "CifarNet", "training", ",", "the", "devil", "is", "in", "the", "details", "Training", "neural", "networks", "is", "sensitive", "to", "a", "large", "number", "of", "parameters", ",", "including", "the", "learning", "rate", ",", "how", "the", "network", "weights", "are", "initialised", ",", "the", "type", "of", "regularisation", "applied", "to", "the", "weights", ",", "and", "the", "training", "batch", "size", ".", "It", "is", "difficult", "to", "isolate", "the", "effects", "of", "the", "individual", "parameters", ",", "and", "the", "best", "parameters", "will", "largely", "depend", "on", "the", "specific", "setup", ".", "We", "report", "here", "the", "parameters", "we", "used", ".", "We", "train", "CifarNet", "via", "stochastic", "gradient", "descent", "(", "SGD", ")", "with", "a", "learning", "rate", "of", ",", "a", "momentum", "of", ",", "and", "a", "batch", "size", "of", ".", "After", "epochs", ",", "we", "reduce", "the", "learning", "rate", "by", "a", "factor", "of", "and", "train", "for", "an", "additional", "epochs", ".", "Reducing", "the", "learning", "rate", "even", "further", "did", "not", "improve", "the", "classification", "accuracy", ".", "The", "other", "learning", "rate", "policies", "we", "explored", "yielded", "inferior", "performance", "(", "e.g.", "gradually", "reducing", "the", "learning", "rate", "each", "training", "iteration", ")", ".", "Careful", "tuning", "of", "the", "learning", "rate", "while", "adjusting", "the", "batch", "size", "was", "critical", ".", "Other", "than", "the", "softmax", "classification", "loss", ",", "the", "training", "loss", "includes", "a", "regularisation", "of", "the", "network", "weights", ".", "In", "the", "objective", "function", ",", "this", "regularization", "term", "has", "a", "weight", "of", "for", "all", "layers", "but", "the", "last", "one", "(", "softmax", "weights", ")", ",", "which", "receives", "weight", ".", "This", "parameter", "is", "referred", "in", "Caffe", "as", "‘", "‘", "weight", "decay", "’", "’", ".", "The", "network", "weights", "are", "initialised", "by", "drawing", "values", "from", "a", "Gaussian", "distribution", "with", "standard", "deviation", ",", "with", "the", "exception", "of", "the", "first", "layer", ",", "for", "which", "we", "set", ".", "appendix", ":", "Grid", "search", "around", "CifarNet", "Table", "[", "reference", "]", "shows", "the", "detection", "quality", "of", "different", "variants", "of", "CifarNet", "obtained", "by", "changing", "the", "number", "and", "size", "of", "the", "convolutional", "filters", "of", "each", "layer", ".", "See", "related", "section", "4.3.1", "of", "the", "main", "paper", ".", "Since", "different", "training", "rounds", "have", "different", "random", "initial", "weights", ",", "we", "train", "four", "networks", "for", "each", "parameter", "set", "and", "average", "the", "results", ".", "We", "report", "both", "mean", "and", "standard", "deviation", "of", "the", "miss", "rate", "on", "our", "validation", "set", ".", "We", "observe", "that", "using", "either", "too", "small", "or", "too", "large", "filter", "sizes", "throughout", "the", "network", "hurts", "quality", ".", "The", "network", "width", "also", "seems", "to", "matter", ",", "a", "network", "too", "narrow", "or", "too", "wide", "can", "negatively", "impact", "classification", "accuracy", ".", "All", "and", "all", "the", "‘", "‘", "middle", "-", "section", "’", "’", "of", "the", "table", "shows", "only", "small", "fluctuations", "in", "miss", "-", "rate", "(", "specially", "when", "considering", "the", "variance", ")", ".", "In", "addition", "to", "filter", "size", "and", "layer", "width", ",", "we", "also", "experimented", "with", "different", "types", "of", "pooling", "layers", "(", "max", "-", "pooling", "versus", "mean", "-", "pooling", ")", ",", "see", "figure", "2", "of", "main", "paper", ".", "Other", "than", "on", "the", "first", "layer", ",", "replacing", "mean", "-", "pooling", "with", "max", "-", "pooling", "hurts", "performance", ".", "The", "results", "of", "table", "indicate", "that", "there", "is", "no", "set", "of", "parameters", "close", "to", "CifarNet", "with", "a", "clear", "advantage", "over", "the", "default", "CifarNet", "parameters", ".", "When", "going", "too", "far", "from", "CifarNet", "parameters", ",", "classification", "accuracy", "plunges", ".", "appendix", ":", "Grid", "search", "for", "AlexNet", "Table", "[", "reference", "]", "presents", "the", "swipe", "of", "parameters", "used", "to", "construct", "the", "‘", "‘", "Best", "parameters", "’", "’", "entries", "in", "table", "8", "of", "the", "main", "paper", ".", "We", "vary", "the", "criterion", "to", "select", "negative", "samples", "and", "the", "SVM", "regularization", "parameters", ".", "Defaults", "are", "parameters", "are", ",", "and", ".", "Overall", "we", "notice", "that", "neither", "parameter", "is", "very", "sensitive", "(", "percent", "points", "fluctuations", ")", ".", "When", "C", "is", "far", "from", "optimal", "large", "degradation", "is", "observed", "(", "per", "cent", "points", ")", ".", "As", "seen", "in", "table", "8", "of", "the", "main", "paper", "the", "gap", "between", "default", "and", "tuned", "parameters", "is", "rather", "small", "(", "percent", "points", ")", ".", "[", "layer", "fc7", "]", "[", "layer", "fc6", "]", "[", "layer", "pool5", "]", "[", "layer", "conv4", "]", "appendix", ":", "Datasets", "statistics", "In", "figure", "[", "reference", "]", "we", "plot", "the", "height", "distribution", "for", "pedestrians", "in", "Caltech", "and", "KITTI", "training", "sets", ".", "Although", "the", "datasets", "are", "visually", "similar", ",", "the", "height", "distributions", "are", "somewhat", "dissimilar", "(", "for", "reference", "ImageNet", "and", "Pascal", "distributions", "are", "more", "look", "alike", "among", "each", "other", ")", ".", "It", "was", "shown", "in", "that", "models", "trained", "in", "each", "dataset", ",", "do", "not", "transfer", "well", "across", "each", "other", "(", "compared", "to", "models", "trained", "on", "the", "smaller", "INRIA", "dataset", ")", ".", "[", "Caltech", "Reasonable", "training", "set", "]", "[", "KITTI", "training", "set", "]", "appendix", ":", "Proposals", "statistics", "In", "figures", "[", "reference", "]", "and", "[", "reference", "]", "we", "show", "statistics", "of", "different", "detectors", "on", "the", "Caltech", "test", "set", ",", "including", "the", "ones", "we", "use", "as", "proposals", "in", "our", "experiments", ".", "These", "figures", "complement", "table", "9", "of", "the", "main", "paper", ".", "Our", "initial", "experiments", "indicated", "that", "it", "is", "important", "to", "keep", "a", "low", "number", "of", "average", "proposals", "per", "image", "in", "order", "to", "reduce", "the", "false", "positives", "rate", "(", "post", "re", "-", "scoring", ")", ".", "This", "is", "in", "contrast", "to", "common", "practice", "when", "using", "class", "-", "agnostic", "proposal", "methods", ",", "where", "using", "more", "windows", "is", "considered", "better", "because", "they", "provide", "higher", "recall", ".", "We", "filter", "proposals", "via", "a", "threshold", "on", "the", "detection", "score", ".", "As", "can", "be", "seen", "in", "figure", "[", "reference", "]", "a", "recall", "higher", "than", "can", "be", "achieved", "with", "only", "proposals", "per", "image", "on", "average", "(", "for", "Intersection", "-", "over", "-", "Union", "threshold", "above", ",", "the", "evaluation", "criterion", ")", ".", "The", "average", "number", "of", "proposals", "per", "image", "is", "quite", "low", "because", "most", "frames", "of", "the", "Caltech", "test", "set", "do", "not", "contain", "any", "pedestrian", ".", "In", "figure", "[", "reference", "]", "we", "show", "the", "number", "of", "false", "positives", "at", "different", "overlap", "levels", "with", "the", "ground", "truth", "annotations", ".", "The", "bump", "around", "IoU", ",", "most", "visible", "for", "SpatialPooling", "and", "LDCF", ",", "is", "an", "artefact", "of", "the", "non", "-", "maximum", "suppression", "method", "used", "by", "each", "method", ".", "Both", "these", "method", "obtain", "high", "quality", "detection", ",", "thus", "they", "must", "assign", "(", "very", ")", "low", "-", "scores", "to", "these", "false", "positives", "windows", ".", "To", "further", "improve", "quality", "the", "re", "-", "scoring", "method", "must", "do", "the", "same", ".", "When", "using", "a", "method", "for", "proposals", "one", "desires", "to", "have", "high", "recall", "with", "high", "overlap", "with", "the", "ground", "truth", "(", "figure", "[", "reference", "]", ")", ",", "as", "well", "has", "having", "false", "positives", "with", "low", "overlap", "with", "the", "ground", "truth", "(", "figure", "[", "reference", "]", ")", ".", "False", "positive", "proposals", "overlapping", "true", "pedestrians", "will", "have", "pieces", "of", "persons", ",", "which", "might", "confuse", "the", "re", "-", "scoring", "classifier", ".", "Classifying", "fully", "centred", "persons", "versus", "random", "background", "is", "assumed", "to", "be", "easier", "task", ".", "In", "table", "9", "of", "the", "main", "paper", "we", "see", "that", "AlexNet", "reaches", "top", "detection", "quality", "by", "improving", "over", "LDCF", ",", "SquaresChnFtrs", ",", "and", "Katamari", "."], "entity": [[16, 19, ["convolutional", "neural", "networks"], "Method"], [20, 21, ["convnets"], "Method"], [26, 28, ["pedestrian", "detection"], "Task"], [35, 36, ["convnets"], "Method"], [41, 43, ["pedestrian", "detectors"], "Method"], [58, 60, ["occlusion", "modelling"], "Method"], [102, 105, ["pre", "-", "training"], "Method"], [106, 108, ["surrogate", "tasks"], "Task"], [113, 115, ["convnet", "detectors"], "Method"], [117, 118, ["Caltech"], "Material"], [123, 124, ["Caltech"], "Material"], [125, 126, ["convnets"], "Method"], [132, 133, ["Caltech1x"], "Material"], [134, 135, ["Caltech10x"], "Material"], [146, 148, ["convnet", "model"], "Method"], [152, 153, ["detectors"], "Method"], [177, 179, ["computer", "vision"], "Task"], [187, 190, ["convolutional", "neural", "networks"], "Method"], [191, 192, ["convnets"], "Method"], [201, 204, ["general", "object", "classification"], "Task"], [205, 208, ["general", "object", "detection"], "Task"], [209, 211, ["feature", "matching"], "Task"], [212, 214, ["stereo", "matching"], "Task"], [215, 217, ["scene", "recognition"], "Task"], [218, 220, ["pose", "estimation"], "Task"], [221, 223, ["action", "recognition"], "Task"], [228, 230, ["Pedestrian", "detection"], "Task"], [235, 237, ["object", "detection"], "Task"], [241, 243, ["car", "safety"], "Task"], [244, 245, ["surveillance"], "Task"], [247, 248, ["robotics"], "Task"], [276, 277, ["convnets"], "Method"], [288, 290, ["neural", "networks"], "Method"], [291, 293, ["pedestrian", "detection"], "Task"], [296, 300, ["special", "-", "purpose", "designs"], "Method"], [307, 311, ["part", "and", "occlusion", "modelling"], "Method"], [320, 322, ["top", "methods"], "Method"], [330, 331, ["Adaboost"], "Method"], [347, 349, ["vanilla", "convnets"], "Method"], [368, 369, ["training"], "Task"], [375, 377, ["proposal", "methods"], "Method"], [389, 391, ["pedestrian", "detection"], "Task"], [397, 400, ["deep", "neural", "networks"], "Method"], [415, 416, ["convnets"], "Method"], [417, 419, ["pedestrian", "detection"], "Task"], [457, 460, ["convolutional", "sparse", "coding"], "Method"], [472, 474, ["detection", "task"], "Task"], [487, 489, ["last", "layer"], "Method"], [490, 491, ["detection"], "Task"], [496, 497, ["ConvNet"], "Method"], [505, 508, ["deformable", "parts", "model"], "Method"], [509, 510, ["DPM"], "Method"], [513, 518, ["stack", "of", "Restricted", "Boltzmann", "Machines"], "Method"], [519, 520, ["RBMs"], "Method"], [529, 532, ["DBN", "-", "Isol"], "Method"], [541, 547, ["person", "-", "to", "-", "person", "relations"], "Task"], [548, 551, ["DBN", "-", "Mut"], "Method"], [580, 582, ["MultiSDP", "network"], "Method"], [599, 600, ["SDN"], "Method"], [605, 606, ["convnet"], "Method"], [607, 609, ["pedestrian", "detection"], "Task"], [619, 621, ["RBM", "variants"], "Method"], [668, 670, ["convolutional", "network"], "Method"], [678, 679, ["convolutions"], "Method"], [684, 685, ["pooling"], "Method"], [686, 688, ["inner", "products"], "Method"], [707, 709, ["Object", "detection"], "Task"], [711, 713, ["pedestrian", "detection"], "Task"], [715, 716, ["convnets"], "Method"], [731, 734, ["general", "object", "detectors"], "Method"], [740, 744, ["R", "-", "CNN", "framework"], "Method"], [766, 767, ["convnet"], "Method"], [774, 779, ["stage", "cascade", "sliding", "window", "method"], "Method"], [791, 792, ["Detection"], "Task"], [799, 801, ["generic", "objects"], "Task"], [802, 803, ["SelectiveSearch"], "Method"], [825, 828, ["DBN", "-", "Mut"], "Method"], [829, 830, ["DPM"], "Method"], [833, 834, ["JointDeep"], "Method"], [835, 836, ["MultiSDP"], "Method"], [838, 839, ["SDN"], "Method"], [856, 857, ["ConvNet"], "Method"], [859, 860, ["convnet"], "Method"], [867, 869, ["Decision", "forests"], "Task"], [873, 875, ["pedestrian", "detection"], "Task"], [878, 879, ["convnets"], "Method"], [880, 881, ["detection"], "Task"], [887, 889, ["optical", "flow"], "Method"], [897, 898, ["Caltech"], "Material"], [903, 904, ["SquaresChnFtrs"], "Method"], [905, 906, ["InformedHaar"], "Method"], [907, 908, ["SpatialPooling"], "Method"], [909, 910, ["LDCF"], "Method"], [912, 913, ["Regionlets"], "Method"], [918, 921, ["boosted", "decision", "forests"], "Method"], [929, 933, ["integral", "channels", "features", "architecture"], "Method"], [934, 935, ["Regionlets"], "Method"], [936, 937, ["SpatialPooling"], "Method"], [949, 950, ["CSS"], "Method"], [952, 953, ["SquaresChnFtrs"], "Method"], [954, 955, ["InformedHaar"], "Method"], [957, 958, ["LDCF"], "Method"], [971, 972, ["convnet"], "Method"], [973, 974, ["SDN"], "Method"], [993, 994, ["ConvNet"], "Method"], [1013, 1016, ["DBN", "-", "Isol"], "Method"], [1017, 1020, ["DBN", "-", "Mut"], "Method"], [1026, 1027, ["MultiSDP"], "Method"], [1028, 1029, ["HOG"], "Method"], [1030, 1031, ["CSS"], "Method"], [1035, 1036, ["JointDeep"], "Method"], [1037, 1038, ["SDN"], "Method"], [1039, 1040, ["YUV"], "Method"], [1041, 1042, ["Gradients"], "Method"], [1051, 1053, ["detection", "proposals"], "Task"], [1068, 1069, ["RGB"], "Method"], [1081, 1083, ["detection", "quality"], "Metric"], [1097, 1098, ["convnets"], "Method"], [1114, 1116, ["pedestrian", "detection"], "Task"], [1117, 1120, ["convolutional", "neural", "networks"], "Method"], [1210, 1212, ["frame", "detector"], "Method"], [1213, 1214, ["Caltech"], "Material"], [1223, 1224, ["convnet"], "Method"], [1258, 1259, ["convnets"], "Method"], [1298, 1300, ["pedestrian", "detection"], "Task"], [1361, 1362, ["validation"], "Task"], [1383, 1385, ["validation", "training"], "Task"], [1391, 1393, ["validation", "testing"], "Task"], [1535, 1537, ["large", "convnets"], "Method"], [1540, 1543, ["pre", "-", "training"], "Method"], [1544, 1546, ["surrogate", "tasks"], "Task"], [1560, 1568, ["ImageNet", "2012", "classification", "of", "a", "thousand", "object", "categories"], "Task"], [1570, 1574, ["classification", "of", "scene", "categories"], "Task"], [1589, 1591, ["decision", "forests"], "Task"], [1592, 1594, ["neural", "networks"], "Method"], [1606, 1608, ["proposal", "method"], "Method"], [1616, 1617, ["convnet"], "Method"], [1627, 1632, ["cascade", "of", "two", "neural", "networks"], "Method"], [1633, 1634, ["SquaresChnFtrs"], "Method"], [1636, 1638, ["decision", "forest"], "Method"], [1702, 1705, ["fully", "connected", "layer"], "Method"], [1726, 1728, ["bias", "term"], "Method"], [1738, 1742, ["sigmoid", "non", "-", "linearity"], "Method"], [1748, 1752, ["two", "-", "layer", "network"], "Method"], [1784, 1788, ["fully", "-", "connected", "layer"], "Method"], [1792, 1793, ["SquaresChnFtrs"], "Method"], [1795, 1798, ["deep", "neural", "network"], "Method"], [1828, 1831, ["back", "-", "propagation"], "Method"], [1834, 1835, ["Adaboost"], "Method"], [1845, 1846, ["convnet"], "Method"], [1847, 1848, ["Caltech"], "Material"], [1862, 1865, ["back", "-", "propagation"], "Method"], [1880, 1883, ["Vanilla", "convolutional", "networks"], "Method"], [1887, 1889, ["convnet", "architectures"], "Method"], [1898, 1899, ["learning"], "Method"], [1900, 1904, ["diverse", "and", "challenging", "tasks"], "Task"], [1949, 1950, ["CifarNet"], "Method"], [1952, 1954, ["small", "network"], "Method"], [1958, 1963, ["CIFAR", "-", "10", "classification", "problem"], "Task"], [1964, 1966, ["objects", "categories"], "Task"], [1979, 1980, ["AlexNet"], "Method"], [1989, 1990, ["CifarNet"], "Method"], [1995, 1998, ["ILSVRC2012", "classification", "problem"], "Task"], [1999, 2001, ["objects", "categories"], "Task"], [2029, 2031, ["pedestrian", "detection"], "Task"], [2049, 2050, ["CifarNet"], "Method"], [2104, 2105, ["convnets"], "Method"], [2125, 2126, ["CifarNet"], "Method"], [2130, 2132, ["detection", "quality"], "Metric"], [2157, 2158, ["CifarNet"], "Method"], [2175, 2177, ["detection", "quality"], "Metric"], [2190, 2191, ["Detection"], "Task"], [2198, 2200, ["SquaresChnFtrs", "detector"], "Method"], [2216, 2218, ["pedestrian", "detector"], "Method"], [2220, 2221, ["Caltech"], "Material"], [2234, 2235, ["SquaresChnFtrs"], "Method"], [2236, 2237, ["EdgeBoxes"], "Method"], [2243, 2248, ["class", "-", "agnostic", "proposal", "method"], "Method"], [2313, 2322, ["Intersection", "-", "over", "-", "Union", "(", "IoU", ")", "threshold"], "Metric"], [2379, 2381, ["pedestrian", "detectors"], "Task"], [2406, 2407, ["convnets"], "Method"], [2409, 2410, ["CifarNet"], "Method"], [2455, 2457, ["detection", "setup"], "Task"], [2507, 2509, ["stochastic", "gradient"], "Method"], [2509, 2511, ["descend", "optimisation"], "Task"], [2575, 2576, ["CifarNet"], "Method"], [2580, 2581, ["convnet"], "Method"], [2603, 2605, ["naive", "network"], "Method"], [2612, 2613, ["convnet"], "Method"], [2615, 2616, ["SDN"], "Method"], [2624, 2627, ["strong", "SquaresChnFtrs", "proposals"], "Method"], [2629, 2630, ["CifarNet"], "Method"], [2636, 2637, ["CifarNet"], "Method"], [2641, 2642, ["JointDeep"], "Method"], [2665, 2666, ["JointDeep"], "Method"], [2667, 2668, ["SDN"], "Method"], [2670, 2671, ["CifarNet"], "Method"], [2679, 2681, ["pedestrian", "detector"], "Method"], [2683, 2684, ["Caltech1x"], "Material"], [2686, 2687, ["SpatialPooling"], "Method"], [2717, 2719, ["convolutional", "filters"], "Method"], [2770, 2771, ["quality"], "Metric"], [2787, 2790, ["miss", "-", "rate"], "Metric"], [2819, 2820, ["CifarNet"], "Method"], [2865, 2867, ["convolutional", "layers"], "Method"], [2868, 2869, ["CONV"], "Method"], [2871, 2875, ["fully", "-", "connected", "layers"], "Method"], [2876, 2877, ["FC"], "Method"], [2882, 2886, ["locally", "-", "connected", "layers"], "Method"], [2887, 2888, ["LC"], "Method"], [2901, 2902, ["ConvNet"], "Method"], [2915, 2917, ["convolutional", "layers"], "Method"], [2918, 2919, ["CifarNet"], "Method"], [2936, 2937, ["convnets"], "Method"], [2938, 2940, ["pedestrian", "detection"], "Task"], [2967, 2968, ["CifarNet"], "Method"], [2990, 2992, ["large", "networks"], "Method"], [3024, 3027, ["Large", "convolutional", "network"], "Method"], [3031, 3032, ["convnets"], "Method"], [3062, 3068, ["R", "-", "CNN", "[", "]", "approach"], "Method"], [3079, 3083, ["Pascal", "VOC", "detection", "task"], "Task"], [3093, 3095, ["pedestrian", "detection"], "Task"], [3098, 3100, ["Surrogate", "tasks"], "Method"], [3102, 3103, ["detections"], "Task"], [3104, 3108, ["R", "-", "CNN", "approach"], "Method"], [3120, 3122, ["large", "network"], "Method"], [3126, 3129, ["ImageNet", "classification", "task"], "Task"], [3135, 3136, ["AlexNet"], "Method"], [3149, 3150, ["AlexNet"], "Method"], [3157, 3160, ["R", "-", "CNN"], "Method"], [3161, 3162, ["AlexNet"], "Method"], [3185, 3190, ["(", "pedestrian", ")", "detection", "task"], "Task"], [3204, 3206, ["linear", "SVM"], "Method"], [3218, 3228, ["open", "source", ",", "Caffe", "based", ",", "R", "-", "CNN", "implementation"], "Method"], [3237, 3238, ["SquaresChnFtrs"], "Method"], [3239, 3241, ["detection", "proposals"], "Task"], [3244, 3247, ["Pre", "-", "training"], "Task"], [3252, 3255, ["top", "layer", "SVM"], "Method"], [3264, 3265, ["AlexNet"], "Method"], [3283, 3284, ["convnet"], "Method"], [3288, 3289, ["SDN"], "Method"], [3298, 3299, ["Caltech"], "Material"], [3316, 3320, ["R", "-", "CNN", "recipe"], "Method"], [3321, 3322, ["detection"], "Task"], [3324, 3325, ["AlexNet"], "Method"], [3349, 3353, ["pre", "-", "training", "task"], "Task"], [3355, 3356, ["AlexNets"], "Method"], [3361, 3363, ["scene", "recognition"], "Task"], [3405, 3409, ["pre", "-", "training", "task"], "Task"], [3432, 3433, ["AlexNet"], "Method"], [3441, 3442, ["Caltech10x"], "Material"], [3495, 3496, ["Caltech10x"], "Material"], [3499, 3500, ["Caltech1x"], "Material"], [3514, 3518, ["single", "-", "frame", "detector"], "Method"], [3519, 3520, ["Caltech"], "Material"], [3536, 3537, ["CifarNet"], "Method"], [3542, 3545, ["pre", "-", "training"], "Task"], [3552, 3553, ["AlexNet"], "Method"], [3576, 3577, ["AlexNet"], "Method"], [3579, 3580, ["Caltech"], "Material"], [3589, 3590, ["SquaresChnFtrs"], "Method"], [3596, 3597, ["convnet"], "Method"], [3598, 3599, ["Caltech"], "Material"], [3600, 3601, ["SDN"], "Method"], [3604, 3605, ["Caltech10x"], "Material"], [3623, 3627, ["ImageNet", "pre", "-", "training"], "Method"], [3644, 3646, ["pedestrian", "data"], "Material"], [3655, 3656, ["AlexNet"], "Method"], [3657, 3658, ["Caltech10x"], "Material"], [3662, 3667, ["single", "-", "frame", "pedestrian", "detector"], "Method"], [3668, 3669, ["Caltech"], "Material"], [3673, 3674, ["LDCF"], "Method"], [3678, 3679, ["Caltech10x"], "Material"], [3700, 3703, ["R", "-", "CNN"], "Method"], [3719, 3720, ["AlexNet"], "Method"], [3737, 3740, ["SVM", "output", "layer"], "Method"], [3770, 3772, ["grid", "search"], "Method"], [3774, 3776, ["grid", "search"], "Method"], [3822, 3824, ["pedestrian", "detection"], "Task"], [3826, 3831, ["R", "-", "CNN", "default", "parameters"], "Method"], [3853, 3854, ["Caltech1x"], "Material"], [3856, 3858, ["proposal", "method"], "Method"], [3869, 3872, ["miss", "-", "rate"], "Metric"], [3890, 3891, ["ACF"], "Method"], [3892, 3893, ["SquaresChnFtrs"], "Method"], [3904, 3905, ["Katamari"], "Method"], [3906, 3908, ["SpatialPooling", "+"], "Method"], [3937, 3939, ["detectors", "ACF"], "Method"], [3940, 3941, ["SquaresChnFtrs"], "Method"], [3943, 3946, ["Katamari", "/", "SpatialPooling"], "Method"], [3979, 3980, ["ACF"], "Method"], [3981, 3982, ["SquaresChnFtrs"], "Method"], [4008, 4009, ["quality"], "Metric"], [4019, 4021, ["quality", "improvement"], "Metric"], [4022, 4023, ["AlexNet"], "Method"], [4045, 4046, ["AlexNet"], "Method"], [4080, 4081, ["AlexNet"], "Method"], [4107, 4108, ["SquaresChnFtrs"], "Method"], [4110, 4112, ["average", "precision"], "Metric"], [4116, 4117, ["AlexNet"], "Method"], [4128, 4129, ["convnets"], "Task"], [4138, 4140, ["Error", "analysis"], "Method"], [4174, 4176, ["localization", "quality"], "Metric"], [4178, 4179, ["convnet"], "Method"], [4191, 4193, ["localization", "errors"], "Metric"], [4195, 4197, ["proposal", "method"], "Method"], [4199, 4202, ["R", "-", "CNN"], "Method"], [4238, 4240, ["localisation", "issues"], "Task"], [4241, 4242, ["detectors"], "Task"], [4244, 4249, ["perfect", "non", "-", "maximum", "suppression"], "Task"], [4261, 4262, ["MR"], "Metric"], [4287, 4288, ["MR"], "Metric"], [4314, 4318, ["R", "-", "CNN", "experiments"], "Method"], [4319, 4321, ["detection", "proposals"], "Task"], [4330, 4333, ["sliding", "window", "detector"], "Method"], [4346, 4347, ["CifarNet"], "Method"], [4348, 4349, ["AlexNet"], "Method"], [4381, 4382, ["Caltech1x"], "Material"], [4383, 4384, ["Caltech10x"], "Material"], [4390, 4391, ["CifarNet"], "Method"], [4392, 4393, ["MR"], "Method"], [4401, 4402, ["AlexNet"], "Method"], [4404, 4405, ["Caltech10x"], "Material"], [4416, 4417, ["AlexNet"], "Method"], [4419, 4420, ["MR"], "Metric"], [4452, 4453, ["AlexNet"], "Method"], [4457, 4458, ["regularisation"], "Method"], [4461, 4462, ["Caltech1x"], "Material"], [4469, 4470, ["detection"], "Task"], [4482, 4485, ["sliding", "window", "detection"], "Task"], [4490, 4492, ["proposal", "method"], "Method"], [4495, 4496, ["recall"], "Metric"], [4503, 4504, ["scoring"], "Task"], [4527, 4529, ["detection", "time"], "Metric"], [4537, 4538, ["convnets"], "Method"], [4539, 4541, ["pedestrian", "detection"], "Task"], [4553, 4555, ["pedestrian", "detection"], "Task"], [4583, 4585, ["shelf", "models"], "Method"], [4614, 4615, ["convnets"], "Method"], [4616, 4617, ["Caltech"], "Material"], [4619, 4620, ["CifarNet"], "Method"], [4650, 4651, ["AlexNet"], "Method"], [4655, 4656, ["convnets"], "Method"], [4662, 4663, ["Caltech"], "Material"], [4679, 4681, ["pedestrian", "detectors"], "Task"], [4686, 4687, ["Caltech1x"], "Material"], [4688, 4689, ["Caltech10x"], "Material"], [4691, 4692, ["CifarNet"], "Method"], [4695, 4700, ["single", "-", "frame", "pedestrian", "detector"], "Method"], [4705, 4706, ["Caltech1x"], "Material"], [4714, 4715, ["AlexNet"], "Method"], [4718, 4723, ["single", "-", "frame", "pedestrian", "detector"], "Method"], [4746, 4747, ["Caltech"], "Material"], [4764, 4765, ["AlexNet"], "Method"], [4772, 4773, ["ImageNet"], "Method"], [4802, 4803, ["convnets"], "Task"], [4811, 4812, ["AlexNet"], "Method"], [4815, 4817, ["proposal", "method"], "Method"], [4825, 4826, ["KITTI"], "Task"], [4828, 4829, ["convnets"], "Method"], [4845, 4846, ["convnets"], "Method"], [4847, 4849, ["pedestrian", "detection"], "Task"], [4853, 4854, ["convnets"], "Method"], [4856, 4858, ["pedestrian", "detection"], "Task"], [4879, 4883, ["vanilla", "’", "’", "networks"], "Method"], [4892, 4895, ["pre", "-", "training"], "Method"], [4896, 4898, ["surrogate", "tasks"], "Task"], [4899, 4900, ["convnets"], "Method"], [4915, 4918, ["pre", "-", "training"], "Method"], [4961, 4962, ["convnets"], "Task"], [4966, 4967, ["Caltech"], "Material"], [4974, 4975, ["convnets"], "Method"], [4982, 4984, ["pedestrian", "detection"], "Task"], [5054, 5057, ["Training", "neural", "networks"], "Method"], [5068, 5070, ["learning", "rate"], "Metric"], [5081, 5082, ["regularisation"], "Method"], [5127, 5128, ["CifarNet"], "Method"], [5129, 5135, ["stochastic", "gradient", "descent", "(", "SGD", ")"], "Method"], [5137, 5139, ["learning", "rate"], "Metric"], [5157, 5159, ["learning", "rate"], "Metric"], [5172, 5174, ["learning", "rate"], "Metric"], [5180, 5182, ["classification", "accuracy"], "Metric"], [5185, 5188, ["learning", "rate", "policies"], "Method"], [5198, 5200, ["learning", "rate"], "Metric"], [5209, 5211, ["learning", "rate"], "Metric"], [5227, 5229, ["training", "loss"], "Metric"], [5243, 5245, ["regularization", "term"], "Method"], [5289, 5291, ["Gaussian", "distribution"], "Method"], [5310, 5312, ["Grid", "search"], "Task"], [5313, 5314, ["CifarNet"], "Method"], [5320, 5322, ["detection", "quality"], "Metric"], [5326, 5327, ["CifarNet"], "Method"], [5383, 5385, ["miss", "rate"], "Metric"], [5406, 5407, ["quality"], "Metric"], [5426, 5428, ["classification", "accuracy"], "Metric"], [5448, 5451, ["miss", "-", "rate"], "Metric"], [5475, 5477, ["pooling", "layers"], "Method"], [5478, 5481, ["max", "-", "pooling"], "Method"], [5482, 5485, ["mean", "-", "pooling"], "Method"], [5502, 5505, ["mean", "-", "pooling"], "Method"], [5506, 5509, ["max", "-", "pooling"], "Method"], [5526, 5527, ["CifarNet"], "Method"], [5545, 5547, ["classification", "accuracy"], "Metric"], [5551, 5553, ["Grid", "search"], "Task"], [5554, 5555, ["AlexNet"], "Method"], [5692, 5693, ["Caltech"], "Material"], [5837, 5840, ["false", "positives", "rate"], "Metric"], [5841, 5845, ["post", "re", "-", "scoring"], "Task"], [5856, 5861, ["class", "-", "agnostic", "proposal", "methods"], "Method"], [5873, 5874, ["recall"], "Metric"], [5883, 5885, ["detection", "score"], "Metric"], [5896, 5897, ["recall"], "Metric"], [5911, 5917, ["Intersection", "-", "over", "-", "Union", "threshold"], "Metric"], [5920, 5922, ["evaluation", "criterion"], "Metric"], [5978, 5979, ["SpatialPooling"], "Method"], [5980, 5981, ["LDCF"], "Method"], [5987, 5992, ["non", "-", "maximum", "suppression", "method"], "Method"], [6003, 6004, ["detection"], "Task"], [6026, 6030, ["re", "-", "scoring", "method"], "Method"], [6040, 6041, ["proposals"], "Task"], [6046, 6047, ["recall"], "Metric"], [6097, 6101, ["re", "-", "scoring", "classifier"], "Method"], [6102, 6106, ["Classifying", "fully", "centred", "persons"], "Task"], [6126, 6127, ["AlexNet"], "Method"], [6129, 6131, ["detection", "quality"], "Metric"], [6134, 6135, ["LDCF"], "Method"], [6136, 6137, ["SquaresChnFtrs"], "Method"], [6139, 6140, ["Katamari"], "Method"], [281, 283, ["pedestrian", "detection"], "Task"], [356, 357, ["Caltech"], "Material"], [561, 562, ["JointDeep"], "Method"], [595, 597, ["pedestrian", "detection"], "Task"], [819, 821, ["pedestrian", "detection"], "Task"], [841, 842, ["HOG"], "Method"], [843, 844, ["CSS"], "Method"], [945, 946, ["HOG"], "Method"], [960, 961, ["HOG"], "Method"], [966, 967, ["Caltech"], "Material"], [998, 999, ["YUV"], "Method"], [1021, 1022, ["HOG"], "Method"], [1046, 1047, ["HOG"], "Method"], [1048, 1049, ["CSS"], "Method"], [1227, 1228, ["Caltech"], "Material"], [1284, 1285, ["Caltech"], "Material"], [1286, 1287, ["Caltech"], "Material"], [1348, 1349, ["Caltech"], "Material"], [1357, 1358, ["Caltech"], "Material"], [1396, 1397, ["Caltech10x"], "Material"], [1399, 1400, ["Caltech"], "Material"], [1488, 1489, ["Caltech"], "Material"], [1668, 1670, ["pedestrian", "detection"], "Task"], [1849, 1850, ["SDN"], "Method"], [2086, 2087, ["CifarNet"], "Method"], [2141, 2142, ["Caltech"], "Material"], [2559, 2560, ["CifarNet"], "Method"], [2584, 2585, ["Caltech"], "Material"], [2657, 2658, ["CifarNet"], "Method"], [2708, 2709, ["CifarNet"], "Method"], [2721, 2722, ["Caltech"], "Material"], [2860, 2861, ["CifarNet"], "Method"], [2972, 2973, ["HOG"], "Method"], [3006, 3007, ["CifarNet"], "Method"], [3012, 3013, ["CifarNet"], "Method"], [3178, 3179, ["AlexNet"], "Method"], [3270, 3271, ["Caltech"], "Material"], [3423, 3424, ["Caltech10x"], "Material"], [3509, 3510, ["AlexNet"], "Method"], [3529, 3530, ["Caltech"], "Material"], [3562, 3563, ["Caltech"], "Material"], [3652, 3653, ["AlexNet"], "Method"], [3731, 3732, ["Caltech"], "Material"], [3848, 3849, ["AlexNet"], "Method"], [3916, 3917, ["Caltech"], "Material"], [3975, 3976, ["AlexNet"], "Method"], [4056, 4057, ["Caltech"], "Material"], [4086, 4088, ["pedestrian", "detection"], "Task"], [4132, 4134, ["pedestrian", "detection"], "Task"], [4216, 4217, ["Caltech"], "Material"], [4370, 4371, ["Caltech"], "Material"], [4409, 4410, ["CifarNet"], "Method"], [4592, 4593, ["CifarNet"], "Method"], [4596, 4597, ["AlexNet"], "Method"], [4606, 4607, ["Caltech"], "Material"], [4725, 4726, ["Caltech10x"], "Material"], [4806, 4808, ["pedestrian", "detection"], "Task"], [5045, 5046, ["CifarNet"], "Method"], [5534, 5535, ["CifarNet"], "Method"], [5542, 5543, ["CifarNet"], "Method"], [5758, 5759, ["Caltech"], "Material"], [5789, 5790, ["Caltech"], "Material"], [5939, 5940, ["Caltech"], "Material"]]} {"doc_id": "1510cf4b8abea80b9f352325ca4c132887de21a0", "text": ["Many", "machine", "learning", "algorithms", "require", "the", "input", "to", "be", "represented", "as", "a", "fixed", "-", "length", "feature", "vector", ".", "When", "it", "comes", "to", "texts", ",", "one", "of", "the", "most", "common", "fixed", "-", "length", "features", "is", "bag", "-", "of", "-", "words", ".", "Despite", "their", "popularity", ",", "bag", "-", "of", "-", "words", "features", "have", "two", "major", "weaknesses", ":", "they", "lose", "the", "ordering", "of", "the", "words", "and", "they", "also", "ignore", "semantics", "of", "the", "words", ".", "For", "example", ",", "“", "powerful", ",", "”", "“", "strong", "”", "and", "“", "Paris", "”", "are", "equally", "distant", ".", "In", "this", "paper", ",", "we", "propose", "Paragraph", "Vector", ",", "an", "unsupervised", "algorithm", "that", "learns", "fixed", "-", "length", "feature", "representations", "from", "variable", "-", "length", "pieces", "of", "texts", ",", "such", "as", "sentences", ",", "paragraphs", ",", "and", "documents", ".", "Our", "algorithm", "represents", "each", "document", "by", "a", "dense", "vector", "which", "is", "trained", "to", "predict", "words", "in", "the", "document", ".", "Its", "construction", "gives", "our", "algorithm", "the", "potential", "to", "overcome", "the", "weaknesses", "of", "bag", "-", "of", "-", "words", "models", ".", "Empirical", "results", "show", "that", "Paragraph", "Vectors", "outperform", "bag", "-", "of", "-", "words", "models", "as", "well", "as", "other", "techniques", "for", "text", "representations", ".", "Finally", ",", "we", "achieve", "new", "state", "-", "of", "-", "the", "-", "art", "results", "on", "several", "text", "classification", "and", "sentiment", "analysis", "tasks", ".", "DistributedRepresentationsofSentencesandDocuments", "section", ":", "Introduction", "Text", "classification", "and", "clustering", "play", "an", "important", "role", "in", "many", "applications", ",", "e.g", ",", "document", "retrieval", ",", "web", "search", ",", "spam", "filtering", ".", "At", "the", "heart", "of", "these", "applications", "is", "machine", "learning", "algorithms", "such", "as", "logistic", "regression", "or", "K", "-", "means", ".", "These", "algorithms", "typically", "require", "the", "text", "input", "to", "be", "represented", "as", "a", "fixed", "-", "length", "vector", ".", "Perhaps", "the", "most", "common", "fixed", "-", "length", "vector", "representation", "for", "texts", "is", "the", "bag", "-", "of", "-", "words", "or", "bag", "-", "of", "-", "n", "-", "grams", "due", "to", "its", "simplicity", ",", "efficiency", "and", "often", "surprising", "accuracy", ".", "However", ",", "the", "bag", "-", "of", "-", "words", "(", "BOW", ")", "has", "many", "disadvantages", ".", "The", "word", "order", "is", "lost", ",", "and", "thus", "different", "sentences", "can", "have", "exactly", "the", "same", "representation", ",", "as", "long", "as", "the", "same", "words", "are", "used", ".", "Even", "though", "bag", "-", "of", "-", "n", "-", "grams", "considers", "the", "word", "order", "in", "short", "context", ",", "it", "suffers", "from", "data", "sparsity", "and", "high", "dimensionality", ".", "Bag", "-", "of", "-", "words", "and", "bag", "-", "of", "-", "n", "-", "grams", "have", "very", "little", "sense", "about", "the", "semantics", "of", "the", "words", "or", "more", "formally", "the", "distances", "between", "the", "words", ".", "This", "means", "that", "words", "“", "powerful", ",", "”", "“", "strong", "”", "and", "“", "Paris", "”", "are", "equally", "distant", "despite", "the", "fact", "that", "semantically", ",", "“", "powerful", "”", "should", "be", "closer", "to", "“", "strong", "”", "than", "“", "Paris", ".", "”", "In", "this", "paper", ",", "we", "propose", "Paragraph", "Vector", ",", "an", "unsupervised", "framework", "that", "learns", "continuous", "distributed", "vector", "representations", "for", "pieces", "of", "texts", ".", "The", "texts", "can", "be", "of", "variable", "-", "length", ",", "ranging", "from", "sentences", "to", "documents", ".", "The", "name", "Paragraph", "Vector", "is", "to", "emphasize", "the", "fact", "that", "the", "method", "can", "be", "applied", "to", "variable", "-", "length", "pieces", "of", "texts", ",", "anything", "from", "a", "phrase", "or", "sentence", "to", "a", "large", "document", ".", "In", "our", "model", ",", "the", "vector", "representation", "is", "trained", "to", "be", "useful", "for", "predicting", "words", "in", "a", "paragraph", ".", "More", "precisely", ",", "we", "concatenate", "the", "paragraph", "vector", "with", "several", "word", "vectors", "from", "a", "paragraph", "and", "predict", "the", "following", "word", "in", "the", "given", "context", ".", "Both", "word", "vectors", "and", "paragraph", "vectors", "are", "trained", "by", "the", "stochastic", "gradient", "descent", "and", "backpropagation", ".", "While", "paragraph", "vectors", "are", "unique", "among", "paragraphs", ",", "the", "word", "vectors", "are", "shared", ".", "At", "prediction", "time", ",", "the", "paragraph", "vectors", "are", "inferred", "by", "fixing", "the", "word", "vectors", "and", "training", "the", "new", "paragraph", "vector", "until", "convergence", ".", "Our", "technique", "is", "inspired", "by", "the", "recent", "work", "in", "learning", "vector", "representations", "of", "words", "using", "neural", "networks", ".", "In", "their", "formulation", ",", "each", "word", "is", "represented", "by", "a", "vector", "which", "is", "concatenated", "or", "averaged", "with", "other", "word", "vectors", "in", "a", "context", ",", "and", "the", "resulting", "vector", "is", "used", "to", "predict", "other", "words", "in", "the", "context", ".", "For", "example", ",", "the", "neural", "network", "language", "model", "proposed", "in", "uses", "the", "concatenation", "of", "several", "previous", "word", "vectors", "to", "form", "the", "input", "of", "a", "neural", "network", ",", "and", "tries", "to", "predict", "the", "next", "word", ".", "The", "outcome", "is", "that", "after", "the", "model", "is", "trained", ",", "the", "word", "vectors", "are", "mapped", "into", "a", "vector", "space", "such", "that", "semantically", "similar", "words", "have", "similar", "vector", "representations", "(", "e.g.", ",", "“", "strong", "”", "is", "close", "to", "“", "powerful", "”", ")", ".", "Following", "these", "successful", "techniques", ",", "researchers", "have", "tried", "to", "extend", "the", "models", "to", "go", "beyond", "word", "level", "to", "achieve", "phrase", "-", "level", "or", "sentence", "-", "level", "representations", ".", "For", "instance", ",", "a", "simple", "approach", "is", "using", "a", "weighted", "average", "of", "all", "the", "words", "in", "the", "document", ".", "A", "more", "sophisticated", "approach", "is", "combining", "the", "word", "vectors", "in", "an", "order", "given", "by", "a", "parse", "tree", "of", "a", "sentence", ",", "using", "matrix", "-", "vector", "operations", ".", "Both", "approaches", "have", "weaknesses", ".", "The", "first", "approach", ",", "weighted", "averaging", "of", "word", "vectors", ",", "loses", "the", "word", "order", "in", "the", "same", "way", "as", "the", "standard", "bag", "-", "of", "-", "words", "models", "do", ".", "The", "second", "approach", ",", "using", "a", "parse", "tree", "to", "combine", "word", "vectors", ",", "has", "been", "shown", "to", "work", "for", "only", "sentences", "because", "it", "relies", "on", "parsing", ".", "Paragraph", "Vector", "is", "capable", "of", "constructing", "representations", "of", "input", "sequences", "of", "variable", "length", ".", "Unlike", "some", "of", "the", "previous", "approaches", ",", "it", "is", "general", "and", "applicable", "to", "texts", "of", "any", "length", ":", "sentences", ",", "paragraphs", ",", "and", "documents", ".", "It", "does", "not", "require", "task", "-", "specific", "tuning", "of", "the", "word", "weighting", "function", "nor", "does", "it", "rely", "on", "the", "parse", "trees", ".", "Further", "in", "the", "paper", ",", "we", "will", "present", "experiments", "on", "several", "benchmark", "datasets", "that", "demonstrate", "the", "advantages", "of", "Paragraph", "Vector", ".", "For", "example", ",", "on", "sentiment", "analysis", "task", ",", "we", "achieve", "new", "state", "-", "of", "-", "the", "-", "art", "results", ",", "better", "than", "complex", "methods", ",", "yielding", "a", "relative", "improvement", "of", "more", "than", "16", "%", "in", "terms", "of", "error", "rate", ".", "On", "a", "text", "classification", "task", ",", "our", "method", "convincingly", "beats", "bag", "-", "of", "-", "words", "models", ",", "giving", "a", "relative", "improvement", "of", "about", "30", "%", ".", "section", ":", "Algorithms", "We", "start", "by", "discussing", "previous", "methods", "for", "learning", "word", "vectors", ".", "These", "methods", "are", "the", "inspiration", "for", "our", "Paragraph", "Vector", "methods", ".", "subsection", ":", "Learning", "Vector", "Representation", "of", "Words", "This", "section", "introduces", "the", "concept", "of", "distributed", "vector", "representation", "of", "words", ".", "A", "well", "known", "framework", "for", "learning", "the", "word", "vectors", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "The", "task", "is", "to", "predict", "a", "word", "given", "the", "other", "words", "in", "a", "context", ".", "In", "this", "framework", ",", "every", "word", "is", "mapped", "to", "a", "unique", "vector", ",", "represented", "by", "a", "column", "in", "a", "matrix", ".", "The", "column", "is", "indexed", "by", "position", "of", "the", "word", "in", "the", "vocabulary", ".", "The", "concatenation", "or", "sum", "of", "the", "vectors", "is", "then", "used", "as", "features", "for", "prediction", "of", "the", "next", "word", "in", "a", "sentence", ".", "More", "formally", ",", "given", "a", "sequence", "of", "training", "words", ",", "the", "objective", "of", "the", "word", "vector", "model", "is", "to", "maximize", "the", "average", "log", "probability", "The", "prediction", "task", "is", "typically", "done", "via", "a", "multiclass", "classifier", ",", "such", "as", "softmax", ".", "There", ",", "we", "have", "Each", "of", "is", "un", "-", "normalized", "log", "-", "probability", "for", "each", "output", "word", ",", "computed", "as", "where", "are", "the", "softmax", "parameters", ".", "is", "constructed", "by", "a", "concatenation", "or", "average", "of", "word", "vectors", "extracted", "from", ".", "In", "practice", ",", "hierarchical", "softmax", "is", "preferred", "to", "softmax", "for", "fast", "training", ".", "In", "our", "work", ",", "the", "structure", "of", "the", "hierarical", "softmax", "is", "a", "binary", "Huffman", "tree", ",", "where", "short", "codes", "are", "assigned", "to", "frequent", "words", ".", "This", "is", "a", "good", "speedup", "trick", "because", "common", "words", "are", "accessed", "quickly", ".", "This", "use", "of", "binary", "Huffman", "code", "for", "the", "hierarchy", "is", "the", "same", "with", ".", "The", "neural", "network", "based", "word", "vectors", "are", "usually", "trained", "using", "stochastic", "gradient", "descent", "where", "the", "gradient", "is", "obtained", "via", "backpropagation", ".", "This", "type", "of", "models", "is", "commonly", "known", "as", "neural", "language", "models", ".", "A", "particular", "implementation", "of", "neural", "network", "based", "algorithm", "for", "training", "the", "word", "vectors", "is", "available", "at", "code.google.com", "/", "p", "/", "word2vec", "/", ".", "After", "the", "training", "converges", ",", "words", "with", "similar", "meaning", "are", "mapped", "to", "a", "similar", "position", "in", "the", "vector", "space", ".", "For", "example", ",", "“", "powerful", "”", "and", "“", "strong", "”", "are", "close", "to", "each", "other", ",", "whereas", "“", "powerful", "”", "and", "“", "Paris", "”", "are", "more", "distant", ".", "The", "difference", "between", "word", "vectors", "also", "carry", "meaning", ".", "For", "example", ",", "the", "word", "vectors", "can", "be", "used", "to", "answer", "analogy", "questions", "using", "simple", "vector", "algebra", ":", "“", "King", "”", "-", "“", "man", "”", "+", "“", "woman", "”", "=", "“", "Queen", "”", ".", "It", "is", "also", "possible", "to", "learn", "a", "linear", "matrix", "to", "translate", "words", "and", "phrases", "between", "languages", ".", "These", "properties", "make", "word", "vectors", "attractive", "for", "many", "natural", "language", "processing", "tasks", "such", "as", "language", "modeling", ",", "natural", "language", "understanding", ",", "statistical", "machine", "translation", ",", "image", "understanding", "and", "relational", "extraction", ".", "subsection", ":", "Paragraph", "Vector", ":", "A", "distributed", "memory", "model", "Our", "approach", "for", "learning", "paragraph", "vectors", "is", "inspired", "by", "the", "methods", "for", "learning", "the", "word", "vectors", ".", "The", "inspiration", "is", "that", "the", "word", "vectors", "are", "asked", "to", "contribute", "to", "a", "prediction", "task", "about", "the", "next", "word", "in", "the", "sentence", ".", "So", "despite", "the", "fact", "that", "the", "word", "vectors", "are", "initialized", "randomly", ",", "they", "can", "eventually", "capture", "semantics", "as", "an", "indirect", "result", "of", "the", "prediction", "task", ".", "We", "will", "use", "this", "idea", "in", "our", "paragraph", "vectors", "in", "a", "similar", "manner", ".", "The", "paragraph", "vectors", "are", "also", "asked", "to", "contribute", "to", "the", "prediction", "task", "of", "the", "next", "word", "given", "many", "contexts", "sampled", "from", "the", "paragraph", ".", "In", "our", "Paragraph", "Vector", "framework", "(", "see", "Figure", "[", "reference", "]", ")", ",", "every", "paragraph", "is", "mapped", "to", "a", "unique", "vector", ",", "represented", "by", "a", "column", "in", "matrix", "and", "every", "word", "is", "also", "mapped", "to", "a", "unique", "vector", ",", "represented", "by", "a", "column", "in", "matrix", ".", "The", "paragraph", "vector", "and", "word", "vectors", "are", "averaged", "or", "concatenated", "to", "predict", "the", "next", "word", "in", "a", "context", ".", "In", "the", "experiments", ",", "we", "use", "concatenation", "as", "the", "method", "to", "combine", "the", "vectors", ".", "More", "formally", ",", "the", "only", "change", "in", "this", "model", "compared", "to", "the", "word", "vector", "framework", "is", "in", "equation", "[", "reference", "]", ",", "where", "is", "constructed", "from", "and", ".", "The", "paragraph", "token", "can", "be", "thought", "of", "as", "another", "word", ".", "It", "acts", "as", "a", "memory", "that", "remembers", "what", "is", "missing", "from", "the", "current", "context", "–", "or", "the", "topic", "of", "the", "paragraph", ".", "For", "this", "reason", ",", "we", "often", "call", "this", "model", "the", "Distributed", "Memory", "Model", "of", "Paragraph", "Vectors", "(", "PV", "-", "DM", ")", ".", "The", "contexts", "are", "fixed", "-", "length", "and", "sampled", "from", "a", "sliding", "window", "over", "the", "paragraph", ".", "The", "paragraph", "vector", "is", "shared", "across", "all", "contexts", "generated", "from", "the", "same", "paragraph", "but", "not", "across", "paragraphs", ".", "The", "word", "vector", "matrix", ",", "however", ",", "is", "shared", "across", "paragraphs", ".", "I.e.", ",", "the", "vector", "for", "“", "powerful", "”", "is", "the", "same", "for", "all", "paragraphs", ".", "The", "paragraph", "vectors", "and", "word", "vectors", "are", "trained", "using", "stochastic", "gradient", "descent", "and", "the", "gradient", "is", "obtained", "via", "backpropagation", ".", "At", "every", "step", "of", "stochastic", "gradient", "descent", ",", "one", "can", "sample", "a", "fixed", "-", "length", "context", "from", "a", "random", "paragraph", ",", "compute", "the", "error", "gradient", "from", "the", "network", "in", "Figure", "[", "reference", "]", "and", "use", "the", "gradient", "to", "update", "the", "parameters", "in", "our", "model", ".", "At", "prediction", "time", ",", "one", "needs", "to", "perform", "an", "inference", "step", "to", "compute", "the", "paragraph", "vector", "for", "a", "new", "paragraph", ".", "This", "is", "also", "obtained", "by", "gradient", "descent", ".", "In", "this", "step", ",", "the", "parameters", "for", "the", "rest", "of", "the", "model", ",", "the", "word", "vectors", "and", "the", "softmax", "weights", ",", "are", "fixed", ".", "Suppose", "that", "there", "are", "paragraphs", "in", "the", "corpus", ",", "words", "in", "the", "vocabulary", ",", "and", "we", "want", "to", "learn", "paragraph", "vectors", "such", "that", "each", "paragraph", "is", "mapped", "to", "dimensions", "and", "each", "word", "is", "mapped", "to", "dimensions", ",", "then", "the", "model", "has", "the", "total", "of", "parameters", "(", "excluding", "the", "softmax", "parameters", ")", ".", "Even", "though", "the", "number", "of", "parameters", "can", "be", "large", "when", "is", "large", ",", "the", "updates", "during", "training", "are", "typically", "sparse", "and", "thus", "efficient", ".", "After", "being", "trained", ",", "the", "paragraph", "vectors", "can", "be", "used", "as", "features", "for", "the", "paragraph", "(", "e.g.", ",", "in", "lieu", "of", "or", "in", "addition", "to", "bag", "-", "of", "-", "words", ")", ".", "We", "can", "feed", "these", "features", "directly", "to", "conventional", "machine", "learning", "techniques", "such", "as", "logistic", "regression", ",", "support", "vector", "machines", "or", "K", "-", "means", ".", "In", "summary", ",", "the", "algorithm", "itself", "has", "two", "key", "stages", ":", "1", ")", "training", "to", "get", "word", "vectors", ",", "softmax", "weights", "and", "paragraph", "vectors", "on", "already", "seen", "paragraphs", ";", "and", "2", ")", "“", "the", "inference", "stage", "”", "to", "get", "paragraph", "vectors", "for", "new", "paragraphs", "(", "never", "seen", "before", ")", "by", "adding", "more", "columns", "in", "and", "gradient", "descending", "on", "while", "holding", "fixed", ".", "We", "use", "to", "make", "a", "prediction", "about", "some", "particular", "labels", "using", "a", "standard", "classifier", ",", "e.g.", ",", "logistic", "regression", ".", "paragraph", ":", "Advantages", "of", "paragraph", "vectors", ":", "An", "important", "advantage", "of", "paragraph", "vectors", "is", "that", "they", "are", "learned", "from", "unlabeled", "data", "and", "thus", "can", "work", "well", "for", "tasks", "that", "do", "not", "have", "enough", "labeled", "data", ".", "Paragraph", "vectors", "also", "address", "some", "of", "the", "key", "weaknesses", "of", "bag", "-", "of", "-", "words", "models", ".", "First", ",", "they", "inherit", "an", "important", "property", "of", "the", "word", "vectors", ":", "the", "semantics", "of", "the", "words", ".", "In", "this", "space", ",", "“", "powerful", "”", "is", "closer", "to", "“", "strong", "”", "than", "to", "“", "Paris", ".", "”", "The", "second", "advantage", "of", "the", "paragraph", "vectors", "is", "that", "they", "take", "into", "consideration", "the", "word", "order", ",", "at", "least", "in", "a", "small", "context", ",", "in", "the", "same", "way", "that", "an", "n", "-", "gram", "model", "with", "a", "large", "n", "would", "do", ".", "This", "is", "important", ",", "because", "the", "n", "-", "gram", "model", "preserves", "a", "lot", "of", "information", "of", "the", "paragraph", ",", "including", "the", "word", "order", ".", "That", "said", ",", "our", "model", "is", "perhaps", "better", "than", "a", "bag", "-", "of", "-", "n", "-", "grams", "model", "because", "a", "bag", "of", "n", "-", "grams", "model", "would", "create", "a", "very", "high", "-", "dimensional", "representation", "that", "tends", "to", "generalize", "poorly", ".", "subsection", ":", "Paragraph", "Vector", "without", "word", "ordering", ":", "Distributed", "bag", "of", "words", "The", "above", "method", "considers", "the", "concatenation", "of", "the", "paragraph", "vector", "with", "the", "word", "vectors", "to", "predict", "the", "next", "word", "in", "a", "text", "window", ".", "Another", "way", "is", "to", "ignore", "the", "context", "words", "in", "the", "input", ",", "but", "force", "the", "model", "to", "predict", "words", "randomly", "sampled", "from", "the", "paragraph", "in", "the", "output", ".", "In", "reality", ",", "what", "this", "means", "is", "that", "at", "each", "iteration", "of", "stochastic", "gradient", "descent", ",", "we", "sample", "a", "text", "window", ",", "then", "sample", "a", "random", "word", "from", "the", "text", "window", "and", "form", "a", "classification", "task", "given", "the", "Paragraph", "Vector", ".", "This", "technique", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "We", "name", "this", "version", "the", "Distributed", "Bag", "of", "Words", "version", "of", "Paragraph", "Vector", "(", "PV", "-", "DBOW", ")", ",", "as", "opposed", "to", "Distributed", "Memory", "version", "of", "Paragraph", "Vector", "(", "PV", "-", "DM", ")", "in", "previous", "section", ".", "In", "addition", "to", "being", "conceptually", "simple", ",", "this", "model", "requires", "to", "store", "less", "data", ".", "We", "only", "need", "to", "store", "the", "softmax", "weights", "as", "opposed", "to", "both", "softmax", "weights", "and", "word", "vectors", "in", "the", "previous", "model", ".", "This", "model", "is", "also", "similar", "to", "the", "Skip", "-", "gram", "model", "in", "word", "vectors", ".", "In", "our", "experiments", ",", "each", "paragraph", "vector", "is", "a", "combination", "of", "two", "vectors", ":", "one", "learned", "by", "the", "standard", "paragraph", "vector", "with", "distributed", "memory", "(", "PV", "-", "DM", ")", "and", "one", "learned", "by", "the", "paragraph", "vector", "with", "distributed", "bag", "of", "words", "(", "PV", "-", "DBOW", ")", ".", "PV", "-", "DM", "alone", "usually", "works", "well", "for", "most", "tasks", "(", "with", "state", "-", "of", "-", "art", "performances", ")", ",", "but", "its", "combination", "with", "PV", "-", "DBOW", "is", "usually", "more", "consistent", "across", "many", "tasks", "that", "we", "try", "and", "therefore", "strongly", "recommended", ".", "section", ":", "Experiments", "We", "perform", "experiments", "to", "better", "understand", "the", "behavior", "of", "the", "paragraph", "vectors", ".", "To", "achieve", "this", ",", "we", "benchmark", "Paragraph", "Vector", "on", "two", "text", "understanding", "problems", "that", "require", "fixed", "-", "length", "vector", "representations", "of", "paragraphs", ":", "sentiment", "analysis", "and", "information", "retrieval", ".", "For", "sentiment", "analysis", ",", "we", "use", "two", "datasets", ":", "Stanford", "sentiment", "treebank", "dataset", "and", "IMDB", "dataset", ".", "Documents", "in", "these", "datasets", "differ", "significantly", "in", "lengths", ":", "every", "example", "in", "Socher", "et", "al", ".", "’s", "dataset", "is", "a", "single", "sentence", "while", "every", "example", "in", "Maas", "et", "al", ".", "’s", "dataset", "consists", "of", "several", "sentences", ".", "We", "also", "test", "our", "method", "on", "an", "information", "retrieval", "task", ",", "where", "the", "goal", "is", "to", "decide", "if", "a", "document", "should", "be", "retrieved", "given", "a", "query", ".", "subsection", ":", "Sentiment", "Analysis", "with", "the", "Stanford", "Sentiment", "Treebank", "Dataset", "paragraph", ":", "Dataset", ":", "This", "dataset", "was", "first", "proposed", "by", "and", "subsequently", "extended", "by", "as", "a", "benchmark", "for", "sentiment", "analysis", ".", "It", "has", "11855", "sentences", "taken", "from", "the", "movie", "review", "site", "Rotten", "Tomatoes", ".", "The", "dataset", "consists", "of", "three", "sets", ":", "8544", "sentences", "for", "training", ",", "2210", "sentences", "for", "test", "and", "1101", "sentences", "for", "validation", "(", "or", "development", ")", ".", "Every", "sentence", "in", "the", "dataset", "has", "a", "label", "which", "goes", "from", "very", "negative", "to", "very", "positive", "in", "the", "scale", "from", "0.0", "to", "1.0", ".", "The", "labels", "are", "generated", "by", "human", "annotators", "using", "Amazon", "Mechanical", "Turk", ".", "The", "dataset", "comes", "with", "detailed", "labels", "for", "sentences", ",", "and", "subphrases", "in", "the", "same", "scale", ".", "To", "achieve", "this", ",", "Socher", "et", "al", ".", "used", "the", "Stanford", "Parser", "to", "parse", "each", "sentence", "to", "subphrases", ".", "The", "subphrases", "were", "then", "labeled", "by", "human", "annotators", "in", "the", "same", "way", "as", "the", "sentences", "were", "labeled", ".", "In", "total", ",", "there", "are", "239", ",", "232", "labeled", "phrases", "in", "the", "dataset", ".", "The", "dataset", "can", "be", "downloaded", "at", ":", "http:", "//", "nlp", ".", "Stanford.edu", "/", "sentiment", "/", "paragraph", ":", "Tasks", "and", "Baselines", ":", "In", ",", "the", "authors", "propose", "two", "ways", "of", "benchmarking", ".", "First", ",", "one", "could", "consider", "a", "5", "-", "way", "fine", "-", "grained", "classification", "task", "where", "the", "labels", "are", "{", "Very", "Negative", ",", "Negative", ",", "Neutral", ",", "Positive", ",", "Very", "Positive", "}", "or", "a", "2", "-", "way", "coarse", "-", "grained", "classification", "task", "where", "the", "labels", "are", "{", "Negative", ",", "Positive}.", "The", "other", "axis", "of", "variation", "is", "in", "terms", "of", "whether", "we", "should", "label", "the", "entire", "sentence", "or", "all", "phrases", "in", "the", "sentence", ".", "In", "this", "work", "we", "only", "consider", "labeling", "the", "full", "sentences", ".", "Socher", "et", "al", ".", "apply", "several", "methods", "to", "this", "dataset", "and", "find", "that", "their", "Recursive", "Neural", "Tensor", "Network", "works", "much", "better", "than", "bag", "-", "of", "-", "words", "model", ".", "It", "can", "be", "argued", "that", "this", "is", "because", "movie", "reviews", "are", "often", "short", "and", "compositionality", "plays", "an", "important", "role", "in", "deciding", "whether", "the", "review", "is", "positive", "or", "negative", ",", "as", "well", "as", "similarity", "between", "words", "does", "given", "the", "rather", "tiny", "size", "of", "the", "training", "set", ".", "paragraph", ":", "Experimental", "protocols", ":", "We", "follow", "the", "experimental", "protocols", "as", "described", "in", ".", "To", "make", "use", "of", "the", "available", "labeled", "data", ",", "in", "our", "model", ",", "each", "subphrase", "is", "treated", "as", "an", "independent", "sentence", "and", "we", "learn", "the", "representations", "for", "all", "the", "subphrases", "in", "the", "training", "set", ".", "After", "learning", "the", "vector", "representations", "for", "training", "sentences", "and", "their", "subphrases", ",", "we", "feed", "them", "to", "a", "logistic", "regression", "to", "learn", "a", "predictor", "of", "the", "movie", "rating", ".", "At", "test", "time", ",", "we", "freeze", "the", "vector", "representation", "for", "each", "word", ",", "and", "learn", "the", "representations", "for", "the", "sentences", "using", "gradient", "descent", ".", "Once", "the", "vector", "representations", "for", "the", "test", "sentences", "are", "learned", ",", "we", "feed", "them", "through", "the", "logistic", "regression", "to", "predict", "the", "movie", "rating", ".", "In", "our", "experiments", ",", "we", "cross", "validate", "the", "window", "size", "using", "the", "validation", "set", ",", "and", "the", "optimal", "window", "size", "is", "8", ".", "The", "vector", "presented", "to", "the", "classifier", "is", "a", "concatenation", "of", "two", "vectors", ",", "one", "from", "PV", "-", "DBOW", "and", "one", "from", "PV", "-", "DM", ".", "In", "PV", "-", "DBOW", ",", "the", "learned", "vector", "representations", "have", "400", "dimensions", ".", "In", "PV", "-", "DM", ",", "the", "learned", "vector", "representations", "have", "400", "dimensions", "for", "both", "words", "and", "paragraphs", ".", "To", "predict", "the", "8", "-", "th", "word", ",", "we", "concatenate", "the", "paragraph", "vectors", "and", "7", "word", "vectors", ".", "Special", "characters", "such", "as", ",", ".", "!", "?", "are", "treated", "as", "a", "normal", "word", ".", "If", "the", "paragraph", "has", "less", "than", "9", "words", ",", "we", "pre", "-", "pad", "with", "a", "special", "NULL", "word", "symbol", ".", "paragraph", ":", "Results", ":", "We", "report", "the", "error", "rates", "of", "different", "methods", "in", "Table", "[", "reference", "]", ".", "The", "first", "highlight", "for", "this", "Table", "is", "that", "bag", "-", "of", "-", "words", "or", "bag", "-", "of", "-", "n", "-", "grams", "models", "(", "NB", ",", "SVM", ",", "BiNB", ")", "perform", "poorly", ".", "Simply", "averaging", "the", "word", "vectors", "(", "in", "a", "bag", "-", "of", "-", "words", "fashion", ")", "does", "not", "improve", "the", "results", ".", "This", "is", "because", "bag", "-", "of", "-", "words", "models", "do", "not", "consider", "how", "each", "sentence", "is", "composed", "(", "e.g.", ",", "word", "ordering", ")", "and", "therefore", "fail", "to", "recognize", "many", "sophisticated", "linguistic", "phenomena", ",", "for", "instance", "sarcasm", ".", "The", "results", "also", "show", "that", "more", "advanced", "methods", "(", "such", "as", "Recursive", "Neural", "Network", ")", ",", "which", "require", "parsing", "and", "take", "into", "account", "the", "compositionality", ",", "perform", "much", "better", ".", "Our", "method", "performs", "better", "than", "all", "these", "baselines", ",", "e.g.", ",", "recursive", "networks", ",", "despite", "the", "fact", "that", "it", "does", "not", "require", "parsing", ".", "On", "the", "coarse", "-", "grained", "classification", "task", ",", "our", "method", "has", "an", "absolute", "improvement", "of", "2.4", "%", "in", "terms", "of", "error", "rates", ".", "This", "translates", "to", "16", "%", "relative", "improvement", ".", "subsection", ":", "Beyond", "One", "Sentence", ":", "Sentiment", "Analysis", "with", "IMDB", "dataset", "Some", "of", "the", "previous", "techniques", "only", "work", "on", "sentences", ",", "but", "not", "paragraphs", "/", "documents", "with", "several", "sentences", ".", "For", "instance", ",", "Recursive", "Neural", "Tensor", "Network", "is", "based", "on", "the", "parsing", "over", "each", "sentence", "and", "it", "is", "unclear", "how", "to", "combine", "the", "representations", "over", "many", "sentences", ".", "Such", "techniques", "therefore", "are", "restricted", "to", "work", "on", "sentences", "but", "not", "paragraphs", "or", "documents", ".", "Our", "method", "does", "not", "require", "parsing", ",", "thus", "it", "can", "produce", "a", "representation", "for", "a", "long", "document", "consisting", "of", "many", "sentences", ".", "This", "advantage", "makes", "our", "method", "more", "general", "than", "some", "of", "the", "other", "approaches", ".", "The", "following", "experiment", "on", "IMDB", "dataset", "demonstrates", "this", "advantage", ".", "paragraph", ":", "Dataset", ":", "The", "IMDB", "dataset", "was", "first", "proposed", "by", "Maas", "et", "al", ".", "as", "a", "benchmark", "for", "sentiment", "analysis", ".", "The", "dataset", "consists", "of", "100", ",", "000", "movie", "reviews", "taken", "from", "IMDB", ".", "One", "key", "aspect", "of", "this", "dataset", "is", "that", "each", "movie", "review", "has", "several", "sentences", ".", "The", "100", ",", "000", "movie", "reviews", "are", "divided", "into", "three", "datasets", ":", "25", ",", "000", "labeled", "training", "instances", ",", "25", ",", "000", "labeled", "test", "instances", "and", "50", ",", "000", "unlabeled", "training", "instances", ".", "There", "are", "two", "types", "of", "labels", ":", "Positive", "and", "Negative", ".", "These", "labels", "are", "balanced", "in", "both", "the", "training", "and", "the", "test", "set", ".", "The", "dataset", "can", "be", "downloaded", "at", "http:", "//", "ai", ".", "Stanford.edu", "/", "amaas", "/", "data", "/", "sentiment", "/", "index.html", "paragraph", ":", "Experimental", "protocols", ":", "We", "learn", "the", "word", "vectors", "and", "paragraph", "vectors", "using", "75", ",", "000", "training", "documents", "(", "25", ",", "000", "labeled", "and", "50", ",", "000", "unlabeled", "instances", ")", ".", "The", "paragraph", "vectors", "for", "the", "25", ",", "000", "labeled", "instances", "are", "then", "fed", "through", "a", "neural", "network", "with", "one", "hidden", "layer", "with", "50", "units", "and", "a", "logistic", "classifier", "to", "learn", "to", "predict", "the", "sentiment", ".", "At", "test", "time", ",", "given", "a", "test", "sentence", ",", "we", "again", "freeze", "the", "rest", "of", "the", "network", "and", "learn", "the", "paragraph", "vectors", "for", "the", "test", "reviews", "by", "gradient", "descent", ".", "Once", "the", "vectors", "are", "learned", ",", "we", "feed", "them", "through", "the", "neural", "network", "to", "predict", "the", "sentiment", "of", "the", "reviews", ".", "The", "hyperparameters", "of", "our", "paragraph", "vector", "model", "are", "selected", "in", "the", "same", "manner", "as", "in", "the", "previous", "task", ".", "In", "particular", ",", "we", "cross", "validate", "the", "window", "size", ",", "and", "the", "optimal", "window", "size", "is", "10", "words", ".", "The", "vector", "presented", "to", "the", "classifier", "is", "a", "concatenation", "of", "two", "vectors", ",", "one", "from", "PV", "-", "DBOW", "and", "one", "from", "PV", "-", "DM", ".", "In", "PV", "-", "DBOW", ",", "the", "learned", "vector", "representations", "have", "400", "dimensions", ".", "In", "PV", "-", "DM", ",", "the", "learned", "vector", "representations", "have", "400", "dimensions", "for", "both", "words", "and", "documents", ".", "To", "predict", "the", "10", "-", "th", "word", ",", "we", "concatenate", "the", "paragraph", "vectors", "and", "word", "vectors", ".", "Special", "characters", "such", "as", ",", ".", "!", "?", "are", "treated", "as", "a", "normal", "word", ".", "If", "the", "document", "has", "less", "than", "9", "words", ",", "we", "pre", "-", "pad", "with", "a", "special", "NULL", "word", "symbol", ".", "paragraph", ":", "Results", ":", "The", "results", "of", "Paragraph", "Vector", "and", "other", "baselines", "are", "reported", "in", "Table", "[", "reference", "]", ".", "As", "can", "be", "seen", "from", "the", "Table", ",", "for", "long", "documents", ",", "bag", "-", "of", "-", "words", "models", "perform", "quite", "well", "and", "it", "is", "difficult", "to", "improve", "upon", "them", "using", "word", "vectors", ".", "The", "most", "significant", "improvement", "happened", "in", "2012", "in", "the", "work", "of", "where", "they", "combine", "a", "Restricted", "Boltzmann", "Machines", "model", "with", "bag", "-", "of", "-", "words", ".", "The", "combination", "of", "two", "models", "yields", "an", "improvement", "approximately", "1.5", "%", "in", "terms", "of", "error", "rates", ".", "Another", "significant", "improvement", "comes", "from", "the", "work", "of", ".", "Among", "many", "variations", "they", "tried", ",", "NBSVM", "on", "bigram", "features", "works", "the", "best", "and", "yields", "a", "considerable", "improvement", "of", "2", "%", "in", "terms", "of", "the", "error", "rate", ".", "The", "method", "described", "in", "this", "paper", "is", "the", "only", "approach", "that", "goes", "significantly", "beyond", "the", "barrier", "of", "10", "%", "error", "rate", ".", "It", "achieves", "7.42", "%", "which", "is", "another", "1.3", "%", "absolute", "improvement", "(", "or", "15", "%", "relative", "improvement", ")", "over", "the", "best", "previous", "result", "of", ".", "subsection", ":", "Information", "Retrieval", "with", "Paragraph", "Vectors", "We", "turn", "our", "attention", "to", "an", "information", "retrieval", "task", "which", "requires", "fixed", "-", "length", "representations", "of", "paragraphs", ".", "Here", ",", "we", "have", "a", "dataset", "of", "paragraphs", "in", "the", "first", "10", "results", "returned", "by", "a", "search", "engine", "given", "each", "of", "1", ",", "000", ",", "000", "most", "popular", "queries", ".", "Each", "of", "these", "paragraphs", "is", "also", "known", "as", "a", "“", "snippet", "”", "which", "summarizes", "the", "content", "of", "a", "web", "page", "and", "how", "a", "web", "page", "matches", "the", "query", ".", "From", "such", "collection", ",", "we", "derive", "a", "new", "dataset", "to", "test", "vector", "representations", "of", "paragraphs", ".", "For", "each", "query", ",", "we", "create", "a", "triplet", "of", "paragraphs", ":", "the", "two", "paragraphs", "are", "results", "of", "the", "same", "query", ",", "whereas", "the", "third", "paragraph", "is", "a", "randomly", "sampled", "paragraph", "from", "the", "rest", "of", "the", "collection", "(", "returned", "as", "the", "result", "of", "a", "different", "query", ")", ".", "Our", "goal", "is", "to", "identify", "which", "of", "the", "three", "paragraphs", "are", "results", "of", "the", "same", "query", ".", "To", "achieve", "this", ",", "we", "will", "use", "paragraph", "vectors", "and", "compute", "the", "distances", "the", "paragraphs", ".", "A", "better", "representation", "is", "one", "that", "achieves", "a", "small", "distance", "for", "pairs", "of", "paragraphs", "of", "the", "same", "query", "and", "a", "larg", "distance", "for", "pairs", "of", "paragraphs", "of", "different", "queries", ".", "Here", "is", "a", "sample", "of", "three", "paragraphs", ",", "where", "the", "first", "paragraph", "should", "be", "closer", "to", "the", "second", "paragraph", "than", "the", "third", "paragraph", ":", "Paragraph", "1", ":", "calls", "from", "(", "000", ")", "000", "-", "0000", ".", "3913", "calls", "reported", "from", "this", "number", ".", "according", "to", "4", "reports", "the", "identity", "of", "this", "caller", "is", "american", "airlines", ".", "Paragraph", "2", ":", "do", "you", "want", "to", "find", "out", "who", "called", "you", "from", "+", "1", "000", "-", "000", "-", "0000", ",", "+", "1", "0000000000", "or", "(", "000", ")", "000", "-", "0000", "?", "see", "reports", "and", "share", "information", "you", "have", "about", "this", "caller", "Paragraph", "3", ":", "allina", "health", "clinic", "patients", "for", "your", "convenience", ",", "you", "can", "pay", "your", "allina", "health", "clinic", "bill", "online", ".", "pay", "your", "clinic", "bill", "now", ",", "question", "and", "answers", "…", "The", "triplets", "are", "split", "into", "three", "sets", ":", "80", "%", "for", "training", ",", "10", "%", "for", "validation", ",", "and", "10", "%", "for", "testing", ".", "Any", "method", "that", "requires", "learning", "will", "be", "trained", "on", "the", "training", "set", ",", "while", "its", "hyperparameters", "will", "be", "selected", "on", "the", "validation", "set", ".", "We", "benchmark", "four", "methods", "to", "compute", "features", "for", "paragraphs", ":", "bag", "-", "of", "-", "words", ",", "bag", "-", "of", "-", "bigrams", ",", "averaging", "word", "vectors", "and", "Paragraph", "Vector", ".", "To", "improve", "bag", "-", "of", "-", "bigrams", ",", "we", "also", "learn", "a", "weighting", "matrix", "such", "that", "the", "distance", "between", "the", "first", "two", "paragraphs", "is", "minimized", "whereas", "the", "distance", "between", "the", "first", "and", "the", "third", "paragraph", "is", "maximized", "(", "the", "weighting", "factor", "between", "the", "two", "losses", "is", "a", "hyperparameter", ")", ".", "We", "record", "the", "number", "of", "times", "when", "each", "method", "produces", "smaller", "distance", "for", "the", "first", "two", "paragraphs", "than", "the", "first", "and", "the", "third", "paragraph", ".", "An", "error", "is", "made", "if", "a", "method", "does", "not", "produce", "that", "desirable", "distance", "metric", "on", "a", "triplet", "of", "paragraphs", ".", "The", "results", "of", "Paragraph", "Vector", "and", "other", "baselines", "are", "reported", "in", "Table", "[", "reference", "]", ".", "In", "this", "task", ",", "we", "find", "that", "TF", "-", "IDF", "weighting", "performs", "better", "than", "raw", "counts", ",", "and", "therefore", "we", "only", "report", "the", "results", "of", "methods", "with", "TF", "-", "IDF", "weighting", ".", "The", "results", "show", "that", "Paragraph", "Vector", "works", "well", "and", "gives", "a", "32", "%", "relative", "improvement", "in", "terms", "of", "error", "rate", ".", "The", "fact", "that", "the", "paragraph", "vector", "method", "significantly", "outperforms", "bag", "of", "words", "and", "bigrams", "suggests", "that", "our", "proposed", "method", "is", "useful", "for", "capturing", "the", "semantics", "of", "the", "input", "text", ".", "subsection", ":", "Some", "further", "observations", "We", "perform", "further", "experiments", "to", "understand", "various", "aspects", "of", "the", "models", ".", "Here", "’s", "some", "observations", "PV", "-", "DM", "is", "consistently", "better", "than", "PV", "-", "DBOW", ".", "PV", "-", "DM", "alone", "can", "achieve", "results", "close", "to", "many", "results", "in", "this", "paper", "(", "see", "Table", "[", "reference", "]", ")", ".", "For", "example", ",", "in", "IMDB", ",", "PV", "-", "DM", "only", "achieves", "7.63", "%", ".", "The", "combination", "of", "PV", "-", "DM", "and", "PV", "-", "DBOW", "often", "work", "consistently", "better", "(", "7.42", "%", "in", "IMDB", ")", "and", "therefore", "recommended", ".", "Using", "concatenation", "in", "PV", "-", "DM", "is", "often", "better", "than", "sum", ".", "In", "IMDB", ",", "PV", "-", "DM", "with", "sum", "can", "only", "achieve", "8.06", "%", ".", "Perhaps", ",", "this", "is", "because", "the", "model", "loses", "the", "ordering", "information", ".", "It", "’s", "better", "to", "cross", "validate", "the", "window", "size", ".", "A", "good", "guess", "of", "window", "size", "in", "many", "applications", "is", "between", "5", "and", "12", ".", "In", "IMDB", ",", "varying", "the", "window", "sizes", "between", "5", "and", "12", "causes", "the", "error", "rate", "to", "fluctuate", "0.7", "%", ".", "Paragraph", "Vector", "can", "be", "expensive", ",", "but", "it", "can", "be", "done", "in", "parallel", "at", "test", "time", ".", "On", "average", ",", "our", "implementation", "takes", "30", "minutes", "to", "compute", "the", "paragraph", "vectors", "of", "the", "IMDB", "test", "set", ",", "using", "a", "16", "core", "machine", "(", "25", ",", "000", "documents", ",", "each", "document", "on", "average", "has", "230", "words", ")", ".", "section", ":", "Related", "Work", "Distributed", "representations", "for", "words", "were", "first", "proposed", "in", "and", "have", "become", "a", "successful", "paradigm", ",", "especially", "for", "statistical", "language", "modeling", ".", "Word", "vectors", "have", "been", "used", "in", "NLP", "applications", "such", "as", "word", "representation", ",", "named", "entity", "recognition", ",", "word", "sense", "disambiguation", ",", "parsing", ",", "tagging", "and", "machine", "translation", ".", "Representing", "phrases", "is", "a", "recent", "trend", "and", "received", "much", "attention", ".", "In", "this", "direction", ",", "autoencoder", "-", "style", "models", "have", "also", "been", "used", "to", "model", "paragraphs", ".", "Distributed", "representations", "of", "phrases", "and", "sentences", "are", "also", "the", "focus", "of", "Socher", "et", "al", ".", ".", "Their", "methods", "typically", "require", "parsing", "and", "is", "shown", "to", "work", "for", "sentence", "-", "level", "representations", ".", "And", "it", "is", "not", "obvious", "how", "to", "extend", "their", "methods", "beyond", "single", "sentences", ".", "Their", "methods", "are", "also", "supervised", "and", "thus", "require", "more", "labeled", "data", "to", "work", "well", ".", "Paragraph", "Vector", ",", "in", "contrast", ",", "is", "mostly", "unsupervised", "and", "thus", "can", "work", "well", "with", "less", "labeled", "data", ".", "Our", "approach", "of", "computing", "the", "paragraph", "vectors", "via", "gradient", "descent", "bears", "resemblance", "to", "a", "successful", "paradigm", "in", "computer", "vision", "known", "as", "Fisher", "kernels", ".", "The", "basic", "construction", "of", "Fisher", "kernels", "is", "the", "gradient", "vector", "over", "an", "unsupervised", "generative", "model", ".", "section", ":", "Discussion", "We", "described", "Paragraph", "Vector", ",", "an", "unsupervised", "learning", "algorithm", "that", "learns", "vector", "representations", "for", "variable", "-", "length", "pieces", "of", "texts", "such", "as", "sentences", "and", "documents", ".", "The", "vector", "representations", "are", "learned", "to", "predict", "the", "surrounding", "words", "in", "contexts", "sampled", "from", "the", "paragraph", ".", "Our", "experiments", "on", "several", "text", "classification", "tasks", "such", "as", "Stanford", "Treebank", "and", "IMDB", "sentiment", "analysis", "datasets", "show", "that", "the", "method", "is", "competitive", "with", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "The", "good", "performance", "demonstrates", "the", "merits", "of", "Paragraph", "Vector", "in", "capturing", "the", "semantics", "of", "paragraphs", ".", "In", "fact", ",", "paragraph", "vectors", "have", "the", "potential", "to", "overcome", "many", "weaknesses", "of", "bag", "-", "of", "-", "words", "models", ".", "Although", "the", "focus", "of", "this", "work", "is", "to", "represent", "texts", ",", "our", "method", "can", "be", "applied", "to", "learn", "representations", "for", "sequential", "data", ".", "In", "non", "-", "text", "domains", "where", "parsing", "is", "not", "available", ",", "we", "expect", "Paragraph", "Vector", "to", "be", "a", "strong", "alternative", "to", "bag", "-", "of", "-", "words", "and", "bag", "-", "of", "-", "n", "-", "grams", "models", ".", "bibliography", ":", "References"], "entity": [[1, 4, ["machine", "learning", "algorithms"], "Method"], [44, 50, ["bag", "-", "of", "-", "words", "features"], "Method"], [95, 97, ["Paragraph", "Vector"], "Method"], [99, 101, ["unsupervised", "algorithm"], "Method"], [103, 108, ["fixed", "-", "length", "feature", "representations"], "Method"], [156, 162, ["bag", "-", "of", "-", "words", "models"], "Method"], [167, 169, ["Paragraph", "Vectors"], "Method"], [170, 176, ["bag", "-", "of", "-", "words", "models"], "Method"], [182, 184, ["text", "representations"], "Task"], [200, 206, ["text", "classification", "and", "sentiment", "analysis", "tasks"], "Task"], [211, 213, ["Text", "classification"], "Task"], [214, 215, ["clustering"], "Task"], [225, 227, ["document", "retrieval"], "Task"], [228, 230, ["web", "search"], "Task"], [231, 233, ["spam", "filtering"], "Task"], [241, 244, ["machine", "learning", "algorithms"], "Method"], [246, 248, ["logistic", "regression"], "Method"], [249, 252, ["K", "-", "means"], "Method"], [274, 279, ["fixed", "-", "length", "vector", "representation"], "Method"], [289, 296, ["bag", "-", "of", "-", "n", "-", "grams"], "Method"], [305, 306, ["accuracy"], "Metric"], [310, 318, ["bag", "-", "of", "-", "words", "(", "BOW", ")"], "Method"], [350, 357, ["bag", "-", "of", "-", "n", "-", "grams"], "Method"], [380, 387, ["bag", "-", "of", "-", "n", "-", "grams"], "Method"], [451, 453, ["Paragraph", "Vector"], "Method"], [455, 457, ["unsupervised", "framework"], "Method"], [459, 463, ["continuous", "distributed", "vector", "representations"], "Method"], [522, 524, ["vector", "representation"], "Method"], [530, 532, ["predicting", "words"], "Task"], [571, 574, ["stochastic", "gradient", "descent"], "Method"], [575, 576, ["backpropagation"], "Method"], [623, 628, ["learning", "vector", "representations", "of", "words"], "Task"], [629, 631, ["neural", "networks"], "Method"], [674, 678, ["neural", "network", "language", "model"], "Method"], [694, 696, ["neural", "network"], "Method"], [770, 774, ["sentence", "-", "level", "representations"], "Method"], [784, 786, ["weighted", "average"], "Method"], [816, 820, ["matrix", "-", "vector", "operations"], "Method"], [830, 835, ["weighted", "averaging", "of", "word", "vectors"], "Method"], [847, 853, ["bag", "-", "of", "-", "words", "models"], "Method"], [880, 881, ["parsing"], "Task"], [882, 884, ["Paragraph", "Vector"], "Method"], [931, 934, ["word", "weighting", "function"], "Method"], [961, 963, ["Paragraph", "Vector"], "Method"], [968, 971, ["sentiment", "analysis", "task"], "Task"], [1001, 1003, ["error", "rate"], "Metric"], [1006, 1009, ["text", "classification", "task"], "Task"], [1014, 1020, ["bag", "-", "of", "-", "words", "models"], "Method"], [1040, 1043, ["learning", "word", "vectors"], "Task"], [1057, 1062, ["Learning", "Vector", "Representation", "of", "Words"], "Task"], [1068, 1073, ["distributed", "vector", "representation", "of", "words"], "Method"], [1153, 1154, ["prediction"], "Task"], [1176, 1179, ["word", "vector", "model"], "Method"], [1183, 1186, ["average", "log", "probability"], "Metric"], [1187, 1189, ["prediction", "task"], "Task"], [1194, 1196, ["multiclass", "classifier"], "Method"], [1199, 1200, ["softmax"], "Method"], [1231, 1237, ["concatenation", "or", "average", "of", "word", "vectors"], "Method"], [1243, 1245, ["hierarchical", "softmax"], "Method"], [1248, 1249, ["softmax"], "Method"], [1250, 1252, ["fast", "training"], "Task"], [1261, 1263, ["hierarical", "softmax"], "Method"], [1265, 1268, ["binary", "Huffman", "tree"], "Method"], [1294, 1297, ["binary", "Huffman", "code"], "Method"], [1306, 1311, ["neural", "network", "based", "word", "vectors"], "Method"], [1315, 1318, ["stochastic", "gradient", "descent"], "Method"], [1324, 1325, ["backpropagation"], "Method"], [1334, 1337, ["neural", "language", "models"], "Method"], [1342, 1346, ["neural", "network", "based", "algorithm"], "Method"], [1429, 1431, ["analogy", "questions"], "Task"], [1433, 1435, ["vector", "algebra"], "Method"], [1472, 1474, ["word", "vectors"], "Method"], [1477, 1481, ["natural", "language", "processing", "tasks"], "Task"], [1483, 1485, ["language", "modeling"], "Task"], [1486, 1489, ["natural", "language", "understanding"], "Task"], [1490, 1493, ["statistical", "machine", "translation"], "Task"], [1494, 1496, ["image", "understanding"], "Task"], [1497, 1499, ["relational", "extraction"], "Task"], [1502, 1504, ["Paragraph", "Vector"], "Method"], [1506, 1509, ["distributed", "memory", "model"], "Method"], [1512, 1515, ["learning", "paragraph", "vectors"], "Task"], [1539, 1541, ["prediction", "task"], "Task"], [1572, 1574, ["prediction", "task"], "Task"], [1599, 1601, ["prediction", "task"], "Task"], [1684, 1685, ["concatenation"], "Method"], [1705, 1708, ["word", "vector", "framework"], "Method"], [1764, 1770, ["Distributed", "Memory", "Model", "of", "Paragraph", "Vectors"], "Method"], [1771, 1774, ["PV", "-", "DM"], "Method"], [1846, 1849, ["stochastic", "gradient", "descent"], "Method"], [1855, 1856, ["backpropagation"], "Method"], [1861, 1864, ["stochastic", "gradient", "descent"], "Method"], [1911, 1913, ["inference", "step"], "Method"], [1928, 1930, ["gradient", "descent"], "Method"], [2023, 2024, ["training"], "Task"], [2071, 2074, ["machine", "learning", "techniques"], "Method"], [2076, 2078, ["logistic", "regression"], "Method"], [2079, 2082, ["support", "vector", "machines"], "Method"], [2083, 2086, ["K", "-", "means"], "Method"], [2100, 2101, ["training"], "Task"], [2121, 2123, ["inference", "stage"], "Method"], [2162, 2163, ["classifier"], "Method"], [2166, 2168, ["logistic", "regression"], "Method"], [2180, 2182, ["paragraph", "vectors"], "Method"], [2205, 2207, ["Paragraph", "vectors"], "Method"], [2215, 2221, ["bag", "-", "of", "-", "words", "models"], "Method"], [2264, 2266, ["paragraph", "vectors"], "Method"], [2289, 2293, ["n", "-", "gram", "model"], "Method"], [2306, 2310, ["n", "-", "gram", "model"], "Method"], [2334, 2342, ["bag", "-", "of", "-", "n", "-", "grams", "model"], "Method"], [2344, 2350, ["bag", "of", "n", "-", "grams", "model"], "Method"], [2356, 2358, ["dimensional", "representation"], "Method"], [2369, 2371, ["word", "ordering"], "Method"], [2372, 2376, ["Distributed", "bag", "of", "words"], "Method"], [2440, 2443, ["stochastic", "gradient", "descent"], "Method"], [2462, 2464, ["classification", "task"], "Task"], [2484, 2492, ["Distributed", "Bag", "of", "Words", "version", "of", "Paragraph", "Vector"], "Method"], [2493, 2496, ["PV", "-", "DBOW"], "Method"], [2508, 2511, ["PV", "-", "DM"], "Method"], [2560, 2564, ["Skip", "-", "gram", "model"], "Method"], [2587, 2589, ["paragraph", "vector"], "Method"], [2590, 2592, ["distributed", "memory"], "Method"], [2593, 2596, ["PV", "-", "DM"], "Method"], [2602, 2604, ["paragraph", "vector"], "Method"], [2605, 2609, ["distributed", "bag", "of", "words"], "Method"], [2610, 2613, ["PV", "-", "DBOW"], "Method"], [2615, 2618, ["PV", "-", "DM"], "Method"], [2639, 2642, ["PV", "-", "DBOW"], "Method"], [2679, 2681, ["Paragraph", "Vector"], "Method"], [2683, 2686, ["text", "understanding", "problems"], "Task"], [2688, 2695, ["fixed", "-", "length", "vector", "representations", "of", "paragraphs"], "Method"], [2696, 2698, ["sentiment", "analysis"], "Task"], [2699, 2701, ["information", "retrieval"], "Task"], [2703, 2705, ["sentiment", "analysis"], "Task"], [2763, 2766, ["information", "retrieval", "task"], "Task"], [2785, 2787, ["Sentiment", "Analysis"], "Task"], [2811, 2813, ["sentiment", "analysis"], "Task"], [2847, 2848, ["validation"], "Task"], [2850, 2851, ["development"], "Task"], [2915, 2917, ["Stanford", "Parser"], "Method"], [2985, 2986, ["benchmarking"], "Task"], [2993, 3001, ["5", "-", "way", "fine", "-", "grained", "classification", "task"], "Task"], [3020, 3028, ["2", "-", "way", "coarse", "-", "grained", "classification", "task"], "Task"], [3084, 3088, ["Recursive", "Neural", "Tensor", "Network"], "Method"], [3092, 3098, ["bag", "-", "of", "-", "words", "model"], "Method"], [3197, 3199, ["vector", "representations"], "Method"], [3211, 3213, ["logistic", "regression"], "Method"], [3219, 3221, ["movie", "rating"], "Task"], [3229, 3231, ["vector", "representation"], "Method"], [3243, 3245, ["gradient", "descent"], "Method"], [3248, 3250, ["vector", "representations"], "Method"], [3262, 3264, ["logistic", "regression"], "Method"], [3267, 3269, ["movie", "rating"], "Task"], [3298, 3299, ["classifier"], "Method"], [3308, 3311, ["PV", "-", "DBOW"], "Method"], [3314, 3317, ["PV", "-", "DM"], "Method"], [3319, 3322, ["PV", "-", "DBOW"], "Method"], [3325, 3327, ["vector", "representations"], "Method"], [3332, 3335, ["PV", "-", "DM"], "Method"], [3338, 3340, ["vector", "representations"], "Method"], [3409, 3411, ["error", "rates"], "Metric"], [3434, 3442, ["bag", "-", "of", "-", "n", "-", "grams", "models"], "Method"], [3443, 3444, ["NB"], "Method"], [3445, 3446, ["SVM"], "Method"], [3447, 3448, ["BiNB"], "Method"], [3460, 3466, ["bag", "-", "of", "-", "words", "fashion"], "Method"], [3476, 3482, ["bag", "-", "of", "-", "words", "models"], "Method"], [3521, 3524, ["Recursive", "Neural", "Network"], "Method"], [3528, 3529, ["parsing"], "Method"], [3551, 3553, ["recursive", "networks"], "Method"], [3562, 3563, ["parsing"], "Task"], [3566, 3571, ["coarse", "-", "grained", "classification", "task"], "Task"], [3584, 3586, ["error", "rates"], "Metric"], [3601, 3603, ["Sentiment", "Analysis"], "Task"], [3628, 3632, ["Recursive", "Neural", "Tensor", "Network"], "Method"], [3636, 3637, ["parsing"], "Method"], [3673, 3674, ["parsing"], "Task"], [3733, 3735, ["sentiment", "analysis"], "Task"], [3887, 3889, ["neural", "network"], "Method"], [3891, 3893, ["hidden", "layer"], "Method"], [3898, 3900, ["logistic", "classifier"], "Method"], [3934, 3936, ["gradient", "descent"], "Method"], [3948, 3950, ["neural", "network"], "Method"], [3959, 3960, ["hyperparameters"], "Method"], [3962, 3965, ["paragraph", "vector", "model"], "Method"], [4001, 4002, ["classifier"], "Method"], [4011, 4014, ["PV", "-", "DBOW"], "Method"], [4017, 4020, ["PV", "-", "DM"], "Method"], [4022, 4025, ["PV", "-", "DBOW"], "Method"], [4028, 4030, ["vector", "representations"], "Method"], [4035, 4038, ["PV", "-", "DM"], "Method"], [4041, 4043, ["vector", "representations"], "Method"], [4111, 4113, ["Paragraph", "Vector"], "Method"], [4136, 4142, ["bag", "-", "of", "-", "words", "models"], "Method"], [4172, 4176, ["Restricted", "Boltzmann", "Machines", "model"], "Method"], [4177, 4182, ["bag", "-", "of", "-", "words"], "Method"], [4197, 4199, ["error", "rates"], "Metric"], [4215, 4216, ["NBSVM"], "Method"], [4234, 4236, ["error", "rate"], "Metric"], [4256, 4258, ["error", "rate"], "Metric"], [4286, 4288, ["Information", "Retrieval"], "Task"], [4297, 4300, ["information", "retrieval", "task"], "Task"], [4302, 4308, ["fixed", "-", "length", "representations", "of", "paragraphs"], "Method"], [4325, 4327, ["search", "engine"], "Method"], [4379, 4383, ["vector", "representations", "of", "paragraphs"], "Method"], [4639, 4640, ["validation"], "Metric"], [4651, 4652, ["learning"], "Task"], [4681, 4686, ["bag", "-", "of", "-", "words"], "Task"], [4687, 4692, ["bag", "-", "of", "-", "bigrams"], "Method"], [4702, 4707, ["bag", "-", "of", "-", "bigrams"], "Task"], [4787, 4789, ["distance", "metric"], "Metric"], [4798, 4800, ["Paragraph", "Vector"], "Method"], [4818, 4822, ["TF", "-", "IDF", "weighting"], "Method"], [4838, 4842, ["TF", "-", "IDF", "weighting"], "Method"], [4847, 4849, ["Paragraph", "Vector"], "Method"], [4861, 4863, ["error", "rate"], "Metric"], [4868, 4871, ["paragraph", "vector", "method"], "Method"], [4877, 4878, ["bigrams"], "Method"], [4915, 4918, ["PV", "-", "DM"], "Method"], [4922, 4925, ["PV", "-", "DBOW"], "Method"], [4952, 4953, ["IMDB"], "Task"], [4954, 4957, ["PV", "-", "DM"], "Method"], [4980, 4981, ["IMDB"], "Metric"], [4987, 4988, ["concatenation"], "Method"], [4989, 4992, ["PV", "-", "DM"], "Method"], [4996, 4997, ["sum"], "Method"], [4999, 5000, ["IMDB"], "Task"], [5001, 5004, ["PV", "-", "DM"], "Method"], [5005, 5006, ["sum"], "Method"], [5050, 5051, ["IMDB"], "Task"], [5062, 5064, ["error", "rate"], "Metric"], [5069, 5071, ["Paragraph", "Vector"], "Method"], [5129, 5131, ["Distributed", "representations"], "Method"], [5146, 5149, ["statistical", "language", "modeling"], "Task"], [5150, 5152, ["Word", "vectors"], "Method"], [5156, 5158, ["NLP", "applications"], "Task"], [5160, 5162, ["word", "representation"], "Task"], [5163, 5166, ["named", "entity", "recognition"], "Task"], [5167, 5170, ["word", "sense", "disambiguation"], "Task"], [5171, 5172, ["parsing"], "Task"], [5173, 5174, ["tagging"], "Task"], [5175, 5177, ["machine", "translation"], "Task"], [5178, 5180, ["Representing", "phrases"], "Task"], [5193, 5197, ["autoencoder", "-", "style", "models"], "Method"], [5205, 5207, ["Distributed", "representations"], "Method"], [5225, 5226, ["parsing"], "Task"], [5232, 5236, ["sentence", "-", "level", "representations"], "Task"], [5266, 5268, ["Paragraph", "Vector"], "Method"], [5274, 5275, ["unsupervised"], "Method"], [5293, 5295, ["gradient", "descent"], "Method"], [5302, 5304, ["computer", "vision"], "Task"], [5306, 5308, ["Fisher", "kernels"], "Method"], [5313, 5315, ["Fisher", "kernels"], "Method"], [5317, 5319, ["gradient", "vector"], "Method"], [5321, 5324, ["unsupervised", "generative", "model"], "Method"], [5330, 5332, ["Paragraph", "Vector"], "Method"], [5334, 5337, ["unsupervised", "learning", "algorithm"], "Method"], [5339, 5341, ["vector", "representations"], "Method"], [5355, 5357, ["vector", "representations"], "Method"], [5375, 5378, ["text", "classification", "tasks"], "Task"], [5410, 5412, ["Paragraph", "Vector"], "Method"], [5415, 5418, ["semantics", "of", "paragraphs"], "Task"], [5432, 5438, ["bag", "-", "of", "-", "words", "models"], "Method"], [5463, 5467, ["non", "-", "text", "domains"], "Task"], [5468, 5469, ["parsing"], "Task"], [5475, 5477, ["Paragraph", "Vector"], "Method"], [5483, 5488, ["bag", "-", "of", "-", "words"], "Method"], [5489, 5497, ["bag", "-", "of", "-", "n", "-", "grams", "models"], "Method"], [485, 487, ["Paragraph", "Vector"], "Method"], [1051, 1053, ["Paragraph", "Vector"], "Method"], [1615, 1617, ["Paragraph", "Vector"], "Method"], [2366, 2368, ["Paragraph", "Vector"], "Method"], [2466, 2468, ["Paragraph", "Vector"], "Method"], [2505, 2507, ["Paragraph", "Vector"], "Method"], [4697, 4699, ["Paragraph", "Vector"], "Method"], [4926, 4929, ["PV", "-", "DM"], "Method"], [4965, 4968, ["PV", "-", "DM"], "Method"], [4969, 4972, ["PV", "-", "DBOW"], "Method"]]} {"doc_id": "1594d954abc650bce2db445c52a76e49655efb0c", "text": ["document", ":", "Recurrent", "Neural", "Network", "Grammars", "We", "introduce", "recurrent", "neural", "network", "grammars", ",", "probabilistic", "models", "of", "sentences", "with", "explicit", "phrase", "structure", ".", "We", "explain", "efficient", "inference", "procedures", "that", "allow", "application", "to", "both", "parsing", "and", "language", "modeling", ".", "Experiments", "show", "that", "they", "provide", "better", "parsing", "in", "English", "than", "any", "single", "previously", "published", "supervised", "generative", "model", "and", "better", "language", "modeling", "than", "state", "-", "of", "-", "the", "-", "art", "sequential", "RNNs", "in", "English", "and", "Chinese", ".", "Due", "to", "an", "implentation", "bug", "in", "the", "RNNG", "’s", "recursive", "composition", "function", ",", "the", "results", "reported", "in", "Dyer", "et", "al", ".", "(", "2016", ")", "did", "not", "correspond", "to", "the", "model", "as", "it", "was", "presented", ".", "This", "corrigendum", "describes", "the", "buggy", "implementation", "and", "reports", "results", "with", "a", "corrected", "implementation", ".", "After", "correction", ",", "on", "the", "PTB", "§", "23", "and", "CTB", "5.1", "test", "sets", ",", "respectively", ",", "the", "generative", "model", "achieves", "language", "modeling", "perplexities", "of", "105.2", "and", "148.5", ",", "and", "phrase", "-", "structure", "parsing", "F1", "of", "93.3", "and", "86.9", ",", "a", "new", "state", "of", "the", "art", "in", "phrase", "-", "structure", "parsing", "for", "both", "languages", ".", "*", "subsubsection", ":", "This", "is", "modified", "version", "of", "a", "paper", "originally", "published", "at", "NAACL", "2016", "that", "contains", "a", "corrigendum", "at", "the", "end", ",", "with", "improved", "results", "after", "fixing", "an", "implementation", "bug", "in", "the", "RNNG", "composition", "function", ".", "section", ":", "Introduction", "Sequential", "recurrent", "neural", "networks", "(", "RNNs", ")", "are", "remarkably", "effective", "models", "of", "natural", "language", ".", "In", "the", "last", "few", "years", ",", "language", "model", "results", "that", "substantially", "improve", "over", "long", "-", "established", "state", "-", "of", "-", "the", "-", "art", "baselines", "have", "been", "obtained", "using", "RNNs", "as", "well", "as", "in", "various", "conditional", "language", "modeling", "tasks", "such", "as", "machine", "translation", ",", "image", "caption", "generation", ",", "and", "dialogue", "generation", ".", "Despite", "these", "impressive", "results", ",", "sequential", "models", "are", "a", "priori", "inappropriate", "models", "of", "natural", "language", ",", "since", "relationships", "among", "words", "are", "largely", "organized", "in", "terms", "of", "latent", "nested", "structures", "rather", "than", "sequential", "surface", "order", ".", "In", "this", "paper", ",", "we", "introduce", "recurrent", "neural", "network", "grammars", "(", "RNNGs", ";", "§", "[", "reference", "]", ")", ",", "a", "new", "generative", "probabilistic", "model", "of", "sentences", "that", "explicitly", "models", "nested", ",", "hierarchical", "relationships", "among", "words", "and", "phrases", ".", "RNNGs", "operate", "via", "a", "recursive", "syntactic", "process", "reminiscent", "of", "probabilistic", "context", "-", "free", "grammar", "generation", ",", "but", "decisions", "are", "parameterized", "using", "RNNs", "that", "condition", "on", "the", "entire", "syntactic", "derivation", "history", ",", "greatly", "relaxing", "context", "-", "free", "independence", "assumptions", ".", "The", "foundation", "of", "this", "work", "is", "a", "top", "-", "down", "variant", "of", "transition", "-", "based", "parsing", "(", "§", "[", "reference", "]", ")", ".", "We", "give", "two", "variants", "of", "the", "algorithm", ",", "one", "for", "parsing", "(", "given", "an", "observed", "sentence", ",", "transform", "it", "into", "a", "tree", ")", ",", "and", "one", "for", "generation", ".", "While", "several", "transition", "-", "based", "neural", "models", "of", "syntactic", "generation", "exist", ",", "these", "have", "relied", "on", "structure", "building", "operations", "based", "on", "parsing", "actions", "in", "shift", "-", "reduce", "and", "left", "-", "corner", "parsers", "which", "operate", "in", "a", "largely", "bottom", "-", "up", "fashion", ".", "While", "this", "construction", "is", "appealing", "because", "inference", "is", "relatively", "straightforward", ",", "it", "limits", "the", "use", "of", "top", "-", "down", "grammar", "information", ",", "which", "is", "helpful", "for", "generation", ".", "RNNGs", "maintain", "the", "algorithmic", "convenience", "of", "transition", "-", "based", "parsing", "but", "incorporate", "top", "-", "down", "(", "i.e.", ",", "root", "-", "to", "-", "terminal", ")", "syntactic", "information", "(", "§", "[", "reference", "]", ")", ".", "The", "top", "-", "down", "transition", "set", "that", "RNNGs", "are", "based", "on", "lends", "itself", "to", "discriminative", "modeling", "as", "well", ",", "where", "sequences", "of", "transitions", "are", "modeled", "conditional", "on", "the", "full", "input", "sentence", "along", "with", "the", "incrementally", "constructed", "syntactic", "structures", ".", "Similar", "to", "previously", "published", "discriminative", "bottom", "-", "up", "transition", "-", "based", "parsers", ",", "greedy", "prediction", "with", "our", "model", "yields", "a", "linear", "-", "time", "deterministic", "parser", "(", "provided", "an", "upper", "bound", "on", "the", "number", "of", "actions", "taken", "between", "processing", "subsequent", "terminal", "symbols", "is", "imposed", ")", ";", "however", ",", "our", "algorithm", "generates", "arbitrary", "tree", "structures", "directly", ",", "without", "the", "binarization", "required", "by", "shift", "-", "reduce", "parsers", ".", "The", "discriminative", "model", "also", "lets", "us", "use", "ancestor", "sampling", "to", "obtain", "samples", "of", "parse", "trees", "for", "sentences", ",", "and", "this", "is", "used", "to", "solve", "a", "second", "practical", "challenge", "with", "RNNGs", ":", "approximating", "the", "marginal", "likelihood", "and", "MAP", "tree", "of", "a", "sentence", "under", "the", "generative", "model", ".", "We", "present", "a", "simple", "importance", "sampling", "algorithm", "which", "uses", "samples", "from", "the", "discriminative", "parser", "to", "solve", "inference", "problems", "in", "the", "generative", "model", "(", "§", "[", "reference", "]", ")", ".", "Experiments", "show", "that", "RNNGs", "are", "effective", "for", "both", "language", "modeling", "and", "parsing", "(", "§", "[", "reference", "]", ")", ".", "Our", "generative", "model", "obtains", "(", "i", ")", "the", "best", "-", "known", "parsing", "results", "using", "a", "single", "supervised", "generative", "model", "and", "(", "ii", ")", "better", "perplexities", "in", "language", "modeling", "than", "state", "-", "of", "-", "the", "-", "art", "sequential", "LSTM", "language", "models", ".", "Surprisingly", "—", "although", "in", "line", "with", "previous", "parsing", "results", "showing", "the", "effectiveness", "of", "generative", "models", "—parsing", "with", "the", "generative", "model", "obtains", "significantly", "better", "results", "than", "parsing", "with", "the", "discriminative", "model", ".", "section", ":", "RNN", "Grammars", "Formally", ",", "an", "RNNG", "is", "a", "triple", "consisting", "of", "a", "finite", "set", "of", "nonterminal", "symbols", "(", ")", ",", "a", "finite", "set", "of", "terminal", "symbols", "(", ")", "such", "that", ",", "and", "a", "collection", "of", "neural", "network", "parameters", ".", "It", "does", "not", "explicitly", "define", "rules", "since", "these", "are", "implicitly", "characterized", "by", ".", "The", "algorithm", "that", "the", "grammar", "uses", "to", "generate", "trees", "and", "strings", "in", "the", "language", "is", "characterized", "in", "terms", "of", "a", "transition", "-", "based", "algorithm", ",", "which", "is", "outlined", "in", "the", "next", "section", ".", "In", "the", "section", "after", "that", ",", "the", "semantics", "of", "the", "parameters", "that", "are", "used", "to", "turn", "this", "into", "a", "stochastic", "algorithm", "that", "generates", "pairs", "of", "trees", "and", "strings", "are", "discussed", ".", "section", ":", "Top", "-", "down", "Parsing", "and", "Generation", "RNNGs", "are", "based", "on", "a", "top", "-", "down", "generation", "algorithm", "that", "relies", "on", "a", "stack", "data", "structure", "of", "partially", "completed", "syntactic", "constituents", ".", "To", "emphasize", "the", "similarity", "of", "our", "algorithm", "to", "more", "familiar", "bottom", "-", "up", "shift", "-", "reduce", "recognition", "algorithms", ",", "we", "first", "present", "the", "parsing", "(", "rather", "than", "generation", ")", "version", "of", "our", "algorithm", "(", "§", "[", "reference", "]", ")", "and", "then", "present", "modifications", "to", "turn", "it", "into", "a", "generator", "(", "§", "[", "reference", "]", ")", ".", "subsection", ":", "Parser", "Transitions", "The", "parsing", "algorithm", "transforms", "a", "sequence", "of", "words", "into", "a", "parse", "tree", "using", "two", "data", "structures", "(", "a", "stack", "and", "an", "input", "buffer", ")", ".", "As", "with", "the", "bottom", "-", "up", "algorithm", "of", ",", "our", "algorithm", "begins", "with", "the", "stack", "(", ")", "empty", "and", "the", "complete", "sequence", "of", "words", "in", "the", "input", "buffer", "(", ")", ".", "The", "buffer", "contains", "unprocessed", "terminal", "symbols", ",", "and", "the", "stack", "contains", "terminal", "symbols", ",", "“", "open", "”", "nonterminal", "symbols", ",", "and", "completed", "constituents", ".", "At", "each", "timestep", ",", "one", "of", "the", "following", "three", "classes", "of", "operations", "(", "Fig", ".", "[", "reference", "]", ")", "is", "selected", "by", "a", "classifier", ",", "based", "on", "the", "current", "contents", "on", "the", "stack", "and", "buffer", ":", "introduces", "an", "“", "open", "nonterminal", "”", "X", "onto", "the", "top", "of", "the", "stack", ".", "Open", "nonterminals", "are", "written", "as", "a", "nonterminal", "symbol", "preceded", "by", "an", "open", "parenthesis", ",", "e.g.", ",", "“", "(", "VP", "”", ",", "and", "they", "represent", "a", "nonterminal", "whose", "child", "nodes", "have", "not", "yet", "been", "fully", "constructed", ".", "Open", "nonterminals", "are", "“", "closed", "”", "to", "form", "complete", "constituents", "by", "subsequent", "reduce", "operations", ".", "shift", "removes", "the", "terminal", "symbol", "from", "the", "front", "of", "the", "input", "buffer", ",", "and", "pushes", "it", "onto", "the", "top", "of", "the", "stack", ".", "reduce", "repeatedly", "pops", "completed", "subtrees", "or", "terminal", "symbols", "from", "the", "stack", "until", "an", "open", "nonterminal", "is", "encountered", ",", "and", "then", "this", "open", "NT", "is", "popped", "and", "used", "as", "the", "label", "of", "a", "new", "constituent", "that", "has", "the", "popped", "subtrees", "as", "its", "children", ".", "This", "new", "completed", "constituent", "is", "pushed", "onto", "the", "stack", "as", "a", "single", "composite", "item", ".", "A", "single", "reduce", "operation", "can", "thus", "create", "constituents", "with", "an", "unbounded", "number", "of", "children", ".", "The", "parsing", "algorithm", "terminates", "when", "there", "is", "a", "single", "completed", "constituent", "on", "the", "stack", "and", "the", "buffer", "is", "empty", ".", "Fig", ".", "[", "reference", "]", "shows", "an", "example", "parse", "using", "our", "transition", "set", ".", "Note", "that", "in", "this", "paper", "we", "do", "not", "model", "preterminal", "symbols", "(", "i.e.", ",", "part", "-", "of", "-", "speech", "tags", ")", "and", "our", "examples", "therefore", "do", "not", "include", "them", ".", "Our", "transition", "set", "is", "closely", "related", "to", "the", "operations", "used", "in", "Earley", "’s", "algorithm", "which", "likewise", "introduces", "nonterminals", "symbols", "with", "its", "predict", "operation", "and", "later", "complete", "s", "them", "after", "consuming", "terminal", "symbols", "one", "at", "a", "time", "using", "scan", ".", "It", "is", "likewise", "closely", "related", "to", "the", "“", "linearized", "”", "parse", "trees", "proposed", "by", "and", "to", "the", "top", "-", "down", ",", "left", "-", "to", "-", "right", "decompositions", "of", "trees", "used", "in", "previous", "generative", "parsing", "and", "language", "modeling", "work", ".", "A", "further", "connection", "is", "to", "parsing", "which", "uses", "an", "unbounded", "lookahead", "(", "compactly", "represented", "by", "a", "DFA", ")", "to", "distinguish", "between", "parse", "alternatives", "in", "a", "top", "-", "down", "parser", ";", "however", ",", "our", "parser", "uses", "an", "RNN", "encoding", "of", "the", "lookahead", "rather", "than", "a", "DFA", ".", "Input", ":", "The", "hungry", "cat", "meows", ".", "paragraph", ":", "Constraints", "on", "parser", "transitions", ".", "To", "guarantee", "that", "only", "well", "-", "formed", "phrase", "-", "structure", "trees", "are", "produced", "by", "the", "parser", ",", "we", "impose", "the", "following", "constraints", "on", "the", "transitions", "that", "can", "be", "applied", "at", "each", "step", "which", "are", "a", "function", "of", "the", "parser", "state", "where", "is", "the", "number", "of", "open", "nonterminals", "on", "the", "stack", ":", "The", "operation", "can", "only", "be", "applied", "if", "is", "not", "empty", "and", ".", "The", "shift", "operation", "can", "only", "be", "applied", "if", "is", "not", "empty", "and", ".", "The", "reduce", "operation", "can", "only", "be", "applied", "if", "the", "top", "of", "the", "stack", "is", "not", "an", "open", "nonterminal", "symbol", ".", "The", "reduce", "operation", "can", "only", "be", "applied", "if", "or", "if", "the", "buffer", "is", "empty", ".", "To", "designate", "the", "set", "of", "valid", "parser", "transitions", ",", "we", "write", ".", "subsection", ":", "Generator", "Transitions", "The", "parsing", "algorithm", "that", "maps", "from", "sequences", "of", "words", "to", "parse", "trees", "can", "be", "adapted", "with", "minor", "changes", "to", "produce", "an", "algorithm", "that", "stochastically", "generates", "trees", "and", "terminal", "symbols", ".", "Two", "changes", "are", "required", ":", "(", "i", ")", "there", "is", "no", "input", "buffer", "of", "unprocessed", "words", ",", "rather", "there", "is", "an", "output", "buffer", "(", ")", ",", "and", "(", "ii", ")", "instead", "of", "a", "shift", "operation", "there", "are", "operations", "which", "generate", "terminal", "symbol", "and", "add", "it", "to", "the", "top", "of", "the", "stack", "and", "the", "output", "buffer", ".", "At", "each", "timestep", "an", "action", "is", "stochastically", "selected", "according", "to", "a", "conditional", "distribution", "that", "depends", "on", "the", "current", "contents", "of", "and", ".", "The", "algorithm", "terminates", "when", "a", "single", "completed", "constituent", "remains", "on", "the", "stack", ".", "Fig", ".", "[", "reference", "]", "shows", "an", "example", "generation", "sequence", ".", "paragraph", ":", "Constraints", "on", "generator", "transitions", ".", "The", "generation", "algorithm", "also", "requires", "slightly", "modified", "constraints", ".", "These", "are", ":", "The", "operation", "can", "only", "be", "applied", "if", ".", "The", "reduce", "operation", "can", "only", "be", "applied", "if", "the", "top", "of", "the", "stack", "is", "not", "an", "open", "nonterminal", "symbol", "and", ".", "To", "designate", "the", "set", "of", "valid", "generator", "transitions", ",", "we", "write", ".", "This", "transition", "set", "generates", "trees", "using", "nearly", "the", "same", "structure", "building", "actions", "and", "stack", "configurations", "as", "the", "“", "top", "-", "down", "PDA", "”", "construction", "proposed", "by", ",", "albeit", "without", "the", "restriction", "that", "the", "trees", "be", "in", "Chomsky", "normal", "form", ".", "subsection", ":", "Transition", "Sequences", "from", "Trees", "Any", "parse", "tree", "can", "be", "converted", "to", "a", "sequence", "of", "transitions", "via", "a", "depth", "-", "first", ",", "left", "-", "to", "-", "right", "traversal", "of", "a", "parse", "tree", ".", "Since", "there", "is", "a", "unique", "depth", "-", "first", ",", "left", "-", "ro", "-", "right", "traversal", "of", "a", "tree", ",", "there", "is", "exactly", "one", "transition", "sequence", "of", "each", "tree", ".", "For", "a", "tree", "and", "a", "sequence", "of", "symbols", ",", "we", "write", "to", "indicate", "the", "corresponding", "sequence", "of", "generation", "transitions", ",", "and", "to", "indicate", "the", "parser", "transitions", ".", "subsection", ":", "Runtime", "Analysis", "A", "detailed", "analysis", "of", "the", "algorithmic", "properties", "of", "our", "top", "-", "down", "parser", "is", "beyond", "the", "scope", "of", "this", "paper", ";", "however", ",", "we", "briefly", "state", "several", "facts", ".", "Assuming", "the", "availability", "of", "constant", "time", "push", "and", "pop", "operations", ",", "the", "runtime", "is", "linear", "in", "the", "number", "of", "the", "nodes", "in", "the", "parse", "tree", "that", "is", "generated", "by", "the", "parser", "/", "generator", "(", "intuitively", ",", "this", "is", "true", "since", "although", "an", "individual", "reduce", "operation", "may", "require", "applying", "a", "number", "of", "pops", "that", "is", "linear", "in", "the", "number", "of", "input", "symbols", ",", "the", "total", "number", "of", "pop", "operations", "across", "an", "entire", "parse", "/", "generation", "run", "will", "also", "be", "linear", ")", ".", "Since", "there", "is", "no", "way", "to", "bound", "the", "number", "of", "output", "nodes", "in", "a", "parse", "tree", "as", "a", "function", "of", "the", "number", "of", "input", "words", ",", "stating", "the", "runtime", "complexity", "of", "the", "parsing", "algorithm", "as", "a", "function", "of", "the", "input", "size", "requires", "further", "assumptions", ".", "Assuming", "our", "fixed", "constraint", "on", "maximum", "depth", ",", "it", "is", "linear", ".", "subsection", ":", "Comparison", "to", "Other", "Models", "Our", "generation", "algorithm", "algorithm", "differs", "from", "previous", "stack", "-", "based", "parsing", "/", "generation", "algorithms", "in", "two", "ways", ".", "First", ",", "it", "constructs", "rooted", "tree", "structures", "top", "down", "(", "rather", "than", "bottom", "up", ")", ",", "and", "second", ",", "the", "transition", "operators", "are", "capable", "of", "directly", "generating", "arbitrary", "tree", "structures", "rather", "than", ",", "e.g.", ",", "assuming", "binarized", "trees", ",", "as", "is", "the", "case", "in", "much", "prior", "work", "that", "has", "used", "transition", "-", "based", "algorithms", "to", "produce", "phrase", "-", "structure", "trees", ".", "section", ":", "Generative", "Model", "RNNGs", "use", "the", "generator", "transition", "set", "just", "presented", "to", "define", "a", "joint", "distribution", "on", "syntax", "trees", "(", ")", "and", "words", "(", ")", ".", "This", "distribution", "is", "defined", "as", "a", "sequence", "model", "over", "generator", "transitions", "that", "is", "parameterized", "using", "a", "continuous", "space", "embedding", "of", "the", "algorithm", "state", "at", "each", "time", "step", "(", ")", ";", "i.e.", ",", "and", "where", "action", "-", "specific", "embeddings", "and", "bias", "vector", "are", "parameters", "in", ".", "The", "representation", "of", "the", "algorithm", "state", "at", "time", ",", ",", "is", "computed", "by", "combining", "the", "representation", "of", "the", "generator", "’s", "three", "data", "structures", ":", "the", "output", "buffer", "(", ")", ",", "represented", "by", "an", "embedding", ",", "the", "stack", "(", ")", ",", "represented", "by", "an", "embedding", ",", "and", "the", "history", "of", "actions", "(", ")", "taken", "by", "the", "generator", ",", "represented", "by", "an", "embedding", ",", "where", "and", "are", "parameters", ".", "Refer", "to", "Figure", "[", "reference", "]", "for", "an", "illustration", "of", "the", "architecture", ".", "The", "output", "buffer", ",", "stack", ",", "and", "history", "are", "sequences", "that", "grow", "unboundedly", ",", "and", "to", "obtain", "representations", "of", "them", "we", "use", "recurrent", "neural", "networks", "to", "“", "encode", "”", "their", "contents", ".", "Since", "the", "output", "buffer", "and", "history", "of", "actions", "are", "only", "appended", "to", "and", "only", "contain", "symbols", "from", "a", "finite", "alphabet", ",", "it", "is", "straightforward", "to", "apply", "a", "standard", "RNN", "encoding", "architecture", ".", "The", "stack", "(", ")", "is", "more", "complicated", "for", "two", "reasons", ".", "First", ",", "the", "elements", "of", "the", "stack", "are", "more", "complicated", "objects", "than", "symbols", "from", "a", "discrete", "alphabet", ":", "open", "nonterminals", ",", "terminals", ",", "and", "full", "trees", ",", "are", "all", "present", "on", "the", "stack", ".", "Second", ",", "it", "is", "manipulated", "using", "both", "push", "and", "pop", "operations", ".", "To", "efficiently", "obtain", "representations", "of", "under", "push", "and", "pop", "operations", ",", "we", "use", "stack", "LSTMs", ".", "To", "represent", "complex", "parse", "trees", ",", "we", "define", "a", "new", "syntactic", "composition", "function", "that", "recursively", "defines", "representations", "of", "trees", ".", "subsection", ":", "Syntactic", "Composition", "Function", "When", "a", "reduce", "operation", "is", "executed", ",", "the", "parser", "pops", "a", "sequence", "of", "completed", "subtrees", "and", "/", "or", "tokens", "(", "together", "with", "their", "vector", "embeddings", ")", "from", "the", "stack", "and", "makes", "them", "children", "of", "the", "most", "recent", "open", "nonterminal", "on", "the", "stack", ",", "“", "completing", "”", "the", "constituent", ".", "To", "compute", "an", "embedding", "of", "this", "new", "subtree", ",", "we", "use", "a", "composition", "function", "based", "on", "bidirectional", "LSTMs", ",", "which", "is", "illustrated", "in", "Fig", ".", "[", "reference", "]", ".", "The", "first", "vector", "read", "by", "the", "LSTM", "in", "both", "the", "forward", "and", "reverse", "directions", "is", "an", "embedding", "of", "the", "label", "on", "the", "constituent", "being", "constructed", "(", "in", "the", "figure", ",", "NP", ")", ".", "This", "is", "followed", "by", "the", "embeddings", "of", "the", "child", "subtrees", "(", "or", "tokens", ")", "in", "forward", "or", "reverse", "order", ".", "Intuitively", ",", "this", "order", "serves", "to", "“", "notify", "”", "each", "LSTM", "what", "sort", "of", "head", "it", "should", "be", "looking", "for", "as", "it", "processes", "the", "child", "node", "embeddings", ".", "The", "final", "state", "of", "the", "forward", "and", "reverse", "LSTMs", "are", "concatenated", ",", "passed", "through", "an", "affine", "transformation", "and", "a", "nonlinearity", "to", "become", "the", "subtree", "embedding", ".", "Because", "each", "of", "the", "child", "node", "embeddings", "(", ",", ",", "in", "Fig", ".", "[", "reference", "]", ")", "is", "computed", "similarly", "(", "if", "it", "corresponds", "to", "an", "internal", "node", ")", ",", "this", "composition", "function", "is", "a", "kind", "of", "recursive", "neural", "network", ".", "subsection", ":", "Word", "Generation", "To", "reduce", "the", "size", "of", ",", "word", "generation", "is", "broken", "into", "two", "parts", ".", "First", ",", "the", "decision", "to", "generate", "is", "made", "(", "by", "predicting", "gen", "as", "an", "action", ")", ",", "and", "then", "choosing", "the", "word", ",", "conditional", "on", "the", "current", "parser", "state", ".", "To", "further", "reduce", "the", "computational", "complexity", "of", "modeling", "the", "generation", "of", "a", "word", ",", "we", "use", "a", "class", "-", "factored", "softmax", ".", "By", "using", "classes", "for", "a", "vocabulary", "of", "size", ",", "this", "prediction", "step", "runs", "in", "time", "rather", "than", "the", "of", "the", "full", "-", "vocabulary", "softmax", ".", "To", "obtain", "clusters", ",", "we", "use", "the", "greedy", "agglomerative", "clustering", "algorithm", "of", ".", "subsection", ":", "Training", "The", "parameters", "in", "the", "model", "are", "learned", "to", "maximize", "the", "likelihood", "of", "a", "corpus", "of", "trees", ".", "subsection", ":", "Discriminative", "Parsing", "Model", "A", "discriminative", "parsing", "model", "can", "be", "obtained", "by", "replacing", "the", "embedding", "of", "at", "each", "time", "step", "with", "an", "embedding", "of", "the", "input", "buffer", ".", "To", "train", "this", "model", ",", "the", "conditional", "likelihood", "of", "each", "sequence", "of", "actions", "given", "the", "input", "string", "is", "maximized", ".", "section", ":", "Inference", "via", "Importance", "Sampling", "Our", "generative", "model", "defines", "a", "joint", "distribution", "on", "trees", "(", ")", "and", "sequences", "of", "words", "(", ")", ".", "To", "evaluate", "this", "as", "a", "language", "model", ",", "it", "is", "necessary", "to", "compute", "the", "marginal", "probability", ".", "And", ",", "to", "evaluate", "the", "model", "as", "a", "parser", ",", "we", "need", "to", "be", "able", "to", "find", "the", "MAP", "parse", "tree", ",", "i.e.", ",", "the", "tree", "that", "maximizes", ".", "However", ",", "because", "of", "the", "unbounded", "dependencies", "across", "the", "sequence", "of", "parsing", "actions", "in", "our", "model", ",", "exactly", "solving", "either", "of", "these", "inference", "problems", "is", "intractable", ".", "To", "obtain", "estimates", "of", "these", ",", "we", "use", "a", "variant", "of", "importance", "sampling", ".", "Our", "importance", "sampling", "algorithm", "uses", "a", "conditional", "proposal", "distribution", "with", "the", "following", "properties", ":", "(", "i", ")", ";", "(", "ii", ")", "samples", "can", "be", "obtained", "efficiently", ";", "and", "(", "iii", ")", "the", "conditional", "probabilities", "of", "these", "samples", "are", "known", ".", "While", "many", "such", "distributions", "are", "available", ",", "the", "discriminatively", "trained", "variant", "of", "our", "parser", "(", "§", "[", "reference", "]", ")", "fulfills", "these", "requirements", ":", "sequences", "of", "actions", "can", "be", "sampled", "using", "a", "simple", "ancestral", "sampling", "approach", ",", "and", ",", "since", "parse", "trees", "and", "action", "sequences", "exist", "in", "a", "one", "-", "to", "-", "one", "relationship", ",", "the", "product", "of", "the", "action", "probabilities", "is", "the", "conditional", "probability", "of", "the", "parse", "tree", "under", ".", "We", "therefore", "use", "our", "discriminative", "parser", "as", "our", "proposal", "distribution", ".", "Importance", "sampling", "uses", "importance", "weights", ",", "which", "we", "define", "as", ",", "to", "compute", "this", "estimate", ".", "Under", "this", "definition", ",", "we", "can", "derive", "the", "estimator", "as", "follows", ":", "We", "now", "replace", "this", "expectation", "with", "its", "Monte", "Carlo", "estimate", "as", "follows", ",", "using", "samples", "from", ":", "To", "obtain", "an", "estimate", "of", "the", "MAP", "tree", ",", "we", "choose", "the", "sampled", "tree", "with", "the", "highest", "probability", "under", "the", "joint", "model", ".", "section", ":", "Experiments", "We", "present", "results", "of", "our", "two", "models", "both", "on", "parsing", "(", "discriminative", "and", "generative", ")", "and", "as", "a", "language", "model", "(", "generative", "only", ")", "in", "English", "and", "Chinese", ".", "paragraph", ":", "Data", ".", "For", "English", ",", "§", "2–21", "of", "the", "Penn", "Treebank", "are", "used", "as", "training", "corpus", "for", "both", ",", "with", "§", "24", "held", "out", "as", "validation", ",", "and", "§", "23", "used", "for", "evaluation", ".", "Singleton", "words", "in", "the", "training", "corpus", "with", "unknown", "word", "classes", "using", "the", "the", "Berkeley", "parser", "’s", "mapping", "rules", ".", "Orthographic", "case", "distinctions", "are", "preserved", ",", "and", "numbers", "(", "beyond", "singletons", ")", "are", "not", "normalized", ".", "For", "Chinese", ",", "we", "use", "the", "Penn", "Chinese", "Treebank", "Version", "5.1", "(", "CTB", ")", ".", "For", "the", "Chinese", "experiments", ",", "we", "use", "a", "single", "unknown", "word", "class", ".", "Corpus", "statistics", "are", "given", "in", "Table", "[", "reference", "]", ".", "paragraph", ":", "Model", "and", "training", "parameters", ".", "For", "the", "discriminative", "model", ",", "we", "used", "hidden", "dimensions", "of", "128", "and", "2", "-", "layer", "LSTMs", "(", "larger", "numbers", "of", "dimensions", "reduced", "validation", "set", "performance", ")", ".", "For", "the", "generative", "model", ",", "we", "used", "256", "dimensions", "and", "2", "-", "layer", "LSTMs", ".", "For", "both", "models", ",", "we", "tuned", "the", "dropout", "rate", "to", "maximize", "validation", "set", "likelihood", ",", "obtaining", "optimal", "rates", "of", "0.2", "(", "discriminative", ")", "and", "0.3", "(", "generative", ")", ".", "For", "the", "sequential", "LSTM", "baseline", "for", "the", "language", "model", ",", "we", "also", "found", "an", "optimal", "dropout", "rate", "of", "0.3", ".", "For", "training", "we", "used", "stochastic", "gradient", "descent", "with", "a", "learning", "rate", "of", "0.1", ".", "All", "parameters", "were", "initialized", "according", "to", "recommendations", "given", "by", ".", "paragraph", ":", "English", "parsing", "results", ".", "Table", "[", "reference", "]", "(", "last", "two", "rows", ")", "gives", "the", "performance", "of", "our", "parser", "on", "Section", "23", ",", "as", "well", "as", "the", "performance", "of", "several", "representative", "models", ".", "For", "the", "discriminative", "model", ",", "we", "used", "a", "greedy", "decoding", "rule", "as", "opposed", "to", "beam", "search", "in", "some", "shift", "-", "reduce", "baselines", ".", "For", "the", "generative", "model", ",", "we", "obtained", "100", "independent", "samples", "from", "a", "flattened", "distribution", "of", "the", "discriminative", "parser", "(", "by", "exponentiating", "each", "probability", "by", "and", "renormalizing", ")", "and", "reranked", "them", "according", "to", "the", "generative", "model", ".", "paragraph", ":", "Chinese", "parsing", "results", ".", "Chinese", "parsing", "results", "were", "obtained", "with", "the", "same", "methodology", "as", "in", "English", "and", "show", "the", "same", "pattern", "(", "Table", "[", "reference", "]", ")", ".", "paragraph", ":", "Language", "model", "results", ".", "We", "report", "held", "-", "out", "per", "-", "word", "perplexities", "of", "three", "language", "models", ",", "both", "sequential", "and", "syntactic", ".", "Log", "probabilities", "are", "normalized", "by", "the", "number", "of", "words", "(", "excluding", "the", "stop", "symbol", ")", ",", "inverted", ",", "and", "exponentiated", "to", "yield", "the", "perplexity", ".", "Results", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "section", ":", "Discussion", "It", "is", "clear", "from", "our", "experiments", "that", "the", "proposed", "generative", "model", "is", "quite", "effective", "both", "as", "a", "parser", "and", "as", "a", "language", "model", ".", "This", "is", "the", "result", "of", "(", "i", ")", "relaxing", "conventional", "independence", "assumptions", "(", "e.g.", ",", "context", "-", "freeness", ")", "and", "(", "ii", ")", "inferring", "continuous", "representations", "of", "symbols", "alongside", "non", "-", "linear", "models", "of", "their", "syntactic", "relationships", ".", "The", "most", "significant", "question", "that", "remains", "is", "why", "the", "discriminative", "model", "—", "which", "has", "more", "information", "available", "to", "it", "than", "the", "generative", "model", "—", "performs", "worse", "than", "the", "generative", "model", ".", "This", "pattern", "has", "been", "observed", "before", "in", "neural", "parsing", "by", ",", "who", "hypothesized", "that", "larger", ",", "unstructured", "conditioning", "contexts", "are", "harder", "to", "learn", "from", ",", "and", "provide", "opportunities", "to", "overfit", ".", "Our", "discriminative", "model", "conditions", "on", "the", "entire", "history", ",", "stack", ",", "and", "buffer", ",", "while", "our", "generative", "model", "only", "accesses", "the", "history", "and", "stack", ".", "The", "fully", "discriminative", "model", "of", "was", "able", "to", "obtain", "results", "similar", "to", "those", "of", "our", "generative", "model", "(", "albeit", "using", "much", "larger", "training", "sets", "obtained", "through", "semisupervision", ")", "but", "similar", "results", "to", "those", "of", "our", "discriminative", "parser", "using", "the", "same", "data", ".", "In", "light", "of", "their", "results", ",", "we", "believe", "Henderson", "’s", "hypothesis", "is", "correct", ",", "and", "that", "generative", "models", "should", "be", "considered", "as", "a", "more", "statistically", "efficient", "method", "for", "learning", "neural", "networks", "from", "small", "data", ".", "section", ":", "Related", "Work", "Our", "language", "model", "combines", "work", "from", "two", "modeling", "traditions", ":", "(", "i", ")", "recurrent", "neural", "network", "language", "models", "and", "(", "ii", ")", "syntactic", "language", "modeling", ".", "Recurrent", "neural", "network", "language", "models", "use", "RNNs", "to", "compute", "representations", "of", "an", "unbounded", "history", "of", "words", "in", "a", "left", "-", "to", "-", "right", "language", "model", ".", "Syntactic", "language", "models", "jointly", "generate", "a", "syntactic", "structure", "and", "a", "sequence", "of", "words", ".", "There", "is", "an", "extensive", "literature", "here", ",", "but", "one", "strand", "of", "work", "has", "emphasized", "a", "bottom", "-", "up", "generation", "of", "the", "tree", ",", "using", "variants", "of", "shift", "-", "reduce", "parser", "actions", "to", "define", "the", "probability", "space", ".", "The", "neural", "-", "network", "–", "based", "model", "of", "is", "particularly", "similar", "to", "ours", "in", "using", "an", "unbounded", "history", "in", "a", "neural", "network", "architecture", "to", "parameterize", "generative", "parsing", "based", "on", "a", "left", "-", "corner", "model", ".", "Dependency", "-", "only", "language", "models", "have", "also", "been", "explored", ".", "Modeling", "generation", "top", "-", "down", "as", "a", "rooted", "branching", "process", "that", "recursively", "rewrites", "nonterminals", "has", "been", "explored", "by", "and", ".", "Of", "particular", "note", "is", "the", "work", "of", ",", "which", "uses", "random", "forests", "and", "hand", "-", "engineered", "features", "over", "the", "entire", "syntactic", "derivation", "history", "to", "make", "decisions", "over", "the", "next", "action", "to", "take", ".", "The", "neural", "networks", "we", "use", "to", "model", "sentences", "are", "structured", "according", "to", "the", "syntax", "of", "the", "sentence", "being", "generated", ".", "Syntactically", "structured", "neural", "architectures", "have", "been", "explored", "in", "a", "number", "of", "applications", ",", "including", "discriminative", "parsing", ",", "sentiment", "analysis", ",", "and", "sentence", "representation", ".", "However", ",", "these", "models", "have", "been", ",", "without", "exception", ",", "discriminative", ";", "this", "is", "the", "first", "work", "to", "use", "syntactically", "structured", "neural", "models", "to", "generate", "language", ".", "Earlier", "work", "has", "demonstrated", "that", "sequential", "RNNs", "have", "the", "capacity", "to", "recognize", "context", "-", "free", "(", "and", "beyond", ")", "languages", ".", "In", "contrast", ",", "our", "work", "may", "be", "understood", "as", "a", "way", "of", "incorporating", "a", "context", "-", "free", "inductive", "bias", "into", "the", "model", "structure", ".", "section", ":", "Outlook", "RNNGs", "can", "be", "combined", "with", "a", "particle", "filter", "inference", "scheme", "(", "rather", "than", "the", "importance", "sampling", "method", "based", "on", "a", "discriminative", "parser", ",", "§", "[", "reference", "]", ")", "to", "produce", "a", "left", "-", "to", "-", "right", "marginalization", "algorithm", "that", "runs", "in", "expected", "linear", "time", ".", "Thus", ",", "they", "could", "be", "used", "in", "applications", "that", "require", "language", "models", ".", "A", "second", "possibility", "is", "to", "replace", "the", "sequential", "generation", "architectures", "found", "in", "many", "neural", "network", "transduction", "problems", "that", "produce", "sentences", "conditioned", "on", "some", "input", ".", "Previous", "work", "in", "machine", "translation", "has", "showed", "that", "conditional", "syntactic", "models", "can", "function", "quite", "well", "without", "the", "computationally", "expensive", "marginalization", "process", "at", "decoding", "time", ".", "A", "third", "consideration", "regarding", "how", "RNNGs", ",", "human", "sentence", "processing", "takes", "place", "in", "a", "left", "-", "to", "-", "right", ",", "incremental", "order", ".", "While", "an", "RNNG", "is", "not", "a", "processing", "model", "(", "it", "is", "a", "grammar", ")", ",", "the", "fact", "that", "it", "is", "left", "-", "to", "-", "right", "opens", "up", "several", "possibilities", "for", "developing", "new", "sentence", "processing", "models", "based", "on", "an", "explicit", "grammars", ",", "similar", "to", "the", "processing", "model", "of", ".", "Finally", ",", "although", "we", "considered", "only", "the", "supervised", "learning", "scenario", ",", "RNNGs", "are", "joint", "models", "that", "could", "be", "trained", "without", "trees", ",", "for", "example", ",", "using", "expectation", "maximization", ".", "section", ":", "Conclusion", "We", "introduced", "recurrent", "neural", "network", "grammars", ",", "a", "probabilistic", "model", "of", "phrase", "-", "structure", "trees", "that", "can", "be", "trained", "generatively", "and", "used", "as", "a", "language", "model", "or", "a", "parser", ",", "and", "a", "corresponding", "discriminative", "model", "that", "can", "be", "used", "as", "a", "parser", ".", "Apart", "from", "out", "-", "of", "-", "vocabulary", "preprocessing", ",", "the", "approach", "requires", "no", "feature", "design", "or", "transformations", "to", "treebank", "data", ".", "The", "generative", "model", "outperforms", "every", "previously", "published", "parser", "built", "on", "a", "single", "supervised", "generative", "model", "in", "English", ",", "and", "a", "bit", "behind", "the", "best", "-", "reported", "generative", "model", "in", "Chinese", ".", "As", "language", "models", ",", "RNNGs", "outperform", "the", "best", "single", "-", "sentence", "language", "models", ".", "section", ":", "Acknowledgments", "We", "thank", "Brendan", "O’Connor", ",", "Swabha", "Swayamdipta", ",", "and", "Brian", "Roark", "for", "feedback", "on", "drafts", "of", "this", "paper", ",", "and", "Jan", "Buys", ",", "Phil", "Blunsom", ",", "and", "Yue", "Zhang", "for", "help", "with", "data", "preparation", ".", "This", "work", "was", "sponsored", "in", "part", "by", "the", "Defense", "Advanced", "Research", "Projects", "Agency", "(", "DARPA", ")", "Information", "Innovation", "Office", "(", "I2O", ")", "under", "the", "Low", "Resource", "Languages", "for", "Emergent", "Incidents", "(", "LORELEI", ")", "program", "issued", "by", "DARPA", "/", "I2O", "under", "Contract", "No", ".", "HR0011", "-", "15", "-", "C", "-", "0114", ";", "it", "was", "also", "supported", "in", "part", "by", "Contract", "No", ".", "W911NF", "-", "15", "-", "1", "-", "0543", "with", "the", "DARPA", "and", "the", "Army", "Research", "Office", "(", "ARO", ")", ".", "Approved", "for", "public", "release", ",", "distribution", "unlimited", ".", "The", "views", "expressed", "are", "those", "of", "the", "authors", "and", "do", "not", "reflect", "the", "official", "policy", "or", "position", "of", "the", "Department", "of", "Defense", "or", "the", "U.S.", "Government", ".", "Miguel", "Ballesteros", "was", "supported", "by", "the", "European", "Commission", "under", "the", "contract", "numbers", "FP7", "-", "ICT", "-", "610411", "(", "project", "MULTISENSOR", ")", "and", "H2020", "-", "RIA", "-", "645012", "(", "project", "KRISTINA", ")", ".", "bibliography", ":", "References", "Chris", "Dyer⁢", "♠", "♣", "Adhiguna", "Kuncoro", "♠", "Miguel", "Ballesteros", "♢", "Noah", "A.", "Smith", "♡", "School", "of", "Computer", "Science", ",", "Carnegie", "Mellon", "University", ",", "Pittsburgh", ",", "PA", ",", "USA", "NLP", "Group", ",", "Pompeu", "Fabra", "University", ",", "Barcelona", ",", "Spain", "Google", "DeepMind", ",", "London", ",", "UK", "Computer", "Science", "&", "Engineering", ",", "University", "of", "Washington", ",", "Seattle", ",", "WA", ",", "USA", "{", "cdyer", ",", "akuncoro}@cs.cmu.edumiguel.ballesteros@upf.edu", ",", "nasmith@cs.washington.edu", "section", ":", "RNNG", "Composition", "Function", "and", "Implementation", "Error", "The", "composition", "function", "reduces", "a", "completed", "constituent", "into", "a", "single", "vector", "representation", "using", "a", "bidirectional", "LSTM", "(", "Figure", "[", "reference", "]", ")", "over", "embeddings", "of", "the", "constituent", "’s", "children", "as", "well", "as", "an", "embedding", "of", "the", "resulting", "nonterminal", "symbol", "type", ".", "The", "implementation", "error", "(", "Figure", "[", "reference", "]", ")", "composed", "the", "constituent", "(", "NP", "the", "hungry", "cat", ")", "by", "reading", "the", "sequence", "“", "NP", "the", "hungry", "NP", "”", ",", "that", "is", ",", "it", "discarded", "the", "rightmost", "child", "of", "every", "constituent", "and", "replaced", "it", "with", "a", "second", "copy", "of", "the", "constituent", "’s", "nonterminal", "symbol", ".", "This", "error", "occurs", "for", "every", "constituent", "and", "means", "crucial", "information", "is", "not", "properly", "propagated", "upwards", "in", "the", "tree", ".", "section", ":", "Results", "after", "Correction", "The", "implementation", "error", "affected", "both", "the", "generative", "and", "discriminative", "RNNGs", ".", "We", "summarize", "corrected", "English", "phrase", "-", "structure", "PTB", "§", "23", "parsing", "result", "in", "Table", "[", "reference", "]", ",", "Chinese", "(", "CTB", "5.1", "§", "271–300", ")", "in", "Table", "[", "reference", "]", "(", "achieving", "the", "the", "best", "reported", "result", "on", "both", "datasets", ")", ",", "and", "English", "and", "Chinese", "language", "modeling", "perplexities", "in", "Table", "[", "reference", "]", ".", "The", "considerable", "improvement", "in", "parsing", "accuracy", "indicates", "that", "properly", "composing", "the", "constituent", "and", "propagating", "information", "upwards", "is", "crucial", ".", "Despite", "slightly", "higher", "language", "modeling", "perplexity", "on", "PTB", "§", "23", ",", "the", "fixed", "RNNG", "still", "outperforms", "a", "highly", "optimized", "sequential", "LSTM", "baseline", "."], "entity": [[2, 6, ["Recurrent", "Neural", "Network", "Grammars"], "Method"], [8, 12, ["recurrent", "neural", "network", "grammars"], "Method"], [13, 17, ["probabilistic", "models", "of", "sentences"], "Method"], [25, 27, ["inference", "procedures"], "Method"], [32, 33, ["parsing"], "Task"], [34, 36, ["language", "modeling"], "Task"], [43, 44, ["parsing"], "Task"], [51, 54, ["supervised", "generative", "model"], "Method"], [56, 58, ["language", "modeling"], "Method"], [66, 68, ["sequential", "RNNs"], "Method"], [80, 85, ["RNNG", "’s", "recursive", "composition", "function"], "Method"], [127, 128, ["PTB"], "Material"], [139, 141, ["generative", "model"], "Method"], [142, 145, ["language", "modeling", "perplexities"], "Method"], [155, 156, ["F1"], "Metric"], [209, 212, ["RNNG", "composition", "function"], "Method"], [216, 220, ["Sequential", "recurrent", "neural", "networks"], "Method"], [221, 222, ["RNNs"], "Method"], [237, 239, ["language", "model"], "Method"], [259, 260, ["RNNs"], "Method"], [265, 269, ["conditional", "language", "modeling", "tasks"], "Task"], [271, 273, ["machine", "translation"], "Task"], [274, 277, ["image", "caption", "generation"], "Task"], [279, 281, ["dialogue", "generation"], "Task"], [287, 289, ["sequential", "models"], "Method"], [295, 297, ["natural", "language"], "Task"], [323, 327, ["recurrent", "neural", "network", "grammars"], "Method"], [328, 329, ["RNNGs"], "Method"], [338, 343, ["generative", "probabilistic", "model", "of", "sentences"], "Method"], [355, 356, ["RNNGs"], "Method"], [359, 362, ["recursive", "syntactic", "process"], "Method"], [364, 370, ["probabilistic", "context", "-", "free", "grammar", "generation"], "Method"], [376, 377, ["RNNs"], "Method"], [401, 405, ["top", "-", "down", "variant"], "Method"], [427, 428, ["parsing"], "Task"], [444, 445, ["generation"], "Task"], [448, 453, ["transition", "-", "based", "neural", "models"], "Method"], [454, 456, ["syntactic", "generation"], "Task"], [462, 465, ["structure", "building", "operations"], "Method"], [470, 478, ["shift", "-", "reduce", "and", "left", "-", "corner", "parsers"], "Method"], [494, 495, ["inference"], "Task"], [514, 515, ["generation"], "Task"], [516, 517, ["RNNGs"], "Method"], [556, 557, ["RNNGs"], "Method"], [563, 565, ["discriminative", "modeling"], "Method"], [592, 600, ["discriminative", "bottom", "-", "up", "transition", "-", "based", "parsers"], "Method"], [601, 603, ["greedy", "prediction"], "Method"], [608, 613, ["linear", "-", "time", "deterministic", "parser"], "Method"], [645, 646, ["binarization"], "Method"], [648, 652, ["shift", "-", "reduce", "parsers"], "Method"], [654, 656, ["discriminative", "model"], "Method"], [660, 662, ["ancestor", "sampling"], "Method"], [682, 683, ["RNNGs"], "Method"], [696, 698, ["generative", "model"], "Method"], [703, 706, ["importance", "sampling", "algorithm"], "Method"], [711, 713, ["discriminative", "parser"], "Method"], [715, 717, ["inference", "problems"], "Task"], [719, 721, ["generative", "model"], "Method"], [731, 732, ["RNNGs"], "Method"], [736, 738, ["language", "modeling"], "Task"], [739, 740, ["parsing"], "Task"], [748, 750, ["generative", "model"], "Method"], [758, 759, ["parsing"], "Task"], [763, 766, ["supervised", "generative", "model"], "Method"], [771, 772, ["perplexities"], "Method"], [773, 775, ["language", "modeling"], "Method"], [783, 787, ["sequential", "LSTM", "language", "models"], "Method"], [795, 796, ["parsing"], "Task"], [801, 803, ["generative", "models"], "Method"], [806, 808, ["generative", "model"], "Method"], [813, 814, ["parsing"], "Task"], [816, 818, ["discriminative", "model"], "Method"], [821, 823, ["RNN", "Grammars"], "Method"], [826, 827, ["RNNG"], "Method"], [856, 859, ["neural", "network", "parameters"], "Method"], [877, 878, ["grammar"], "Method"], [893, 897, ["transition", "-", "based", "algorithm"], "Method"], [925, 927, ["stochastic", "algorithm"], "Method"], [939, 943, ["Top", "-", "down", "Parsing"], "Task"], [944, 945, ["Generation"], "Task"], [945, 946, ["RNNGs"], "Method"], [950, 955, ["top", "-", "down", "generation", "algorithm"], "Method"], [978, 986, ["bottom", "-", "up", "shift", "-", "reduce", "recognition", "algorithms"], "Method"], [991, 992, ["parsing"], "Task"], [1016, 1017, ["generator"], "Method"], [1056, 1060, ["bottom", "-", "up", "algorithm"], "Method"], [1131, 1132, ["classifier"], "Method"], [1206, 1208, ["reduce", "operations"], "Method"], [1209, 1210, ["shift"], "Method"], [1292, 1294, ["reduce", "operation"], "Method"], [1380, 1383, ["Earley", "’s", "algorithm"], "Method"], [1443, 1445, ["language", "modeling"], "Task"], [1452, 1453, ["parsing"], "Task"], [1463, 1464, ["DFA"], "Method"], [1472, 1476, ["top", "-", "down", "parser"], "Method"], [1480, 1481, ["parser"], "Method"], [1483, 1485, ["RNN", "encoding"], "Method"], [1487, 1488, ["lookahead"], "Method"], [1491, 1492, ["DFA"], "Method"], [1522, 1523, ["parser"], "Method"], [1571, 1573, ["shift", "operation"], "Method"], [1584, 1586, ["reduce", "operation"], "Method"], [1604, 1606, ["reduce", "operation"], "Method"], [1632, 1634, ["Generator", "Transitions"], "Method"], [1697, 1699, ["shift", "operation"], "Method"], [1774, 1776, ["generation", "algorithm"], "Method"], [1794, 1796, ["reduce", "operation"], "Method"], [1844, 1850, ["top", "-", "down", "PDA", "”", "construction"], "Method"], [1885, 1888, ["depth", "-", "first"], "Method"], [1958, 1960, ["Runtime", "Analysis"], "Task"], [1969, 1973, ["top", "-", "down", "parser"], "Method"], [2001, 2002, ["runtime"], "Metric"], [2019, 2022, ["parser", "/", "generator"], "Method"], [2032, 2034, ["reduce", "operation"], "Method"], [2060, 2063, ["parse", "/", "generation"], "Task"], [2098, 2100, ["runtime", "complexity"], "Metric"], [2134, 2137, ["generation", "algorithm", "algorithm"], "Method"], [2171, 2173, ["transition", "operators"], "Method"], [2201, 2205, ["transition", "-", "based", "algorithms"], "Method"], [2214, 2216, ["Generative", "Model"], "Method"], [2216, 2217, ["RNNGs"], "Method"], [2245, 2247, ["sequence", "model"], "Method"], [2248, 2250, ["generator", "transitions"], "Method"], [2255, 2258, ["continuous", "space", "embedding"], "Method"], [2302, 2303, ["generator"], "Method"], [2305, 2307, ["data", "structures"], "Method"], [2317, 2318, ["embedding"], "Method"], [2327, 2328, ["embedding"], "Method"], [2339, 2340, ["generator"], "Method"], [2344, 2345, ["embedding"], "Method"], [2386, 2389, ["recurrent", "neural", "networks"], "Method"], [2424, 2427, ["RNN", "encoding", "architecture"], "Method"], [2498, 2500, ["stack", "LSTMs"], "Method"], [2511, 2514, ["syntactic", "composition", "function"], "Method"], [2523, 2526, ["Syntactic", "Composition", "Function"], "Method"], [2528, 2530, ["reduce", "operation"], "Method"], [2534, 2535, ["parser"], "Method"], [2587, 2589, ["composition", "function"], "Method"], [2591, 2593, ["bidirectional", "LSTMs"], "Method"], [2610, 2611, ["LSTM"], "Method"], [2667, 2668, ["LSTM"], "Method"], [2690, 2694, ["forward", "and", "reverse", "LSTMs"], "Method"], [2700, 2702, ["affine", "transformation"], "Method"], [2708, 2710, ["subtree", "embedding"], "Method"], [2742, 2744, ["composition", "function"], "Method"], [2748, 2751, ["recursive", "neural", "network"], "Method"], [2754, 2756, ["Word", "Generation"], "Task"], [2762, 2764, ["word", "generation"], "Task"], [2804, 2806, ["computational", "complexity"], "Metric"], [2809, 2813, ["generation", "of", "a", "word"], "Task"], [2817, 2821, ["class", "-", "factored", "softmax"], "Method"], [2832, 2834, ["prediction", "step"], "Method"], [2842, 2846, ["full", "-", "vocabulary", "softmax"], "Method"], [2854, 2858, ["greedy", "agglomerative", "clustering", "algorithm"], "Method"], [2882, 2885, ["Discriminative", "Parsing", "Model"], "Method"], [2931, 2932, ["Inference"], "Task"], [2933, 2935, ["Importance", "Sampling"], "Method"], [2936, 2938, ["generative", "model"], "Method"], [2958, 2960, ["language", "model"], "Method"], [2978, 2979, ["parser"], "Method"], [3021, 3023, ["inference", "problems"], "Task"], [3037, 3039, ["importance", "sampling"], "Method"], [3041, 3044, ["importance", "sampling", "algorithm"], "Method"], [3046, 3049, ["conditional", "proposal", "distribution"], "Method"], [3088, 3091, ["discriminatively", "trained", "variant"], "Method"], [3093, 3094, ["parser"], "Method"], [3113, 3116, ["ancestral", "sampling", "approach"], "Method"], [3155, 3157, ["discriminative", "parser"], "Method"], [3159, 3161, ["proposal", "distribution"], "Method"], [3162, 3164, ["Importance", "sampling"], "Method"], [3197, 3200, ["Monte", "Carlo", "estimate"], "Method"], [3227, 3229, ["joint", "model"], "Method"], [3242, 3243, ["parsing"], "Task"], [3244, 3247, ["discriminative", "and", "generative"], "Method"], [3251, 3253, ["language", "model"], "Method"], [3254, 3256, ["generative", "only"], "Method"], [3273, 3275, ["Penn", "Treebank"], "Material"], [3311, 3316, ["Berkeley", "parser", "’s", "mapping", "rules"], "Method"], [3380, 3382, ["discriminative", "model"], "Method"], [3390, 3394, ["2", "-", "layer", "LSTMs"], "Method"], [3407, 3409, ["generative", "model"], "Method"], [3415, 3419, ["2", "-", "layer", "LSTMs"], "Method"], [3427, 3429, ["dropout", "rate"], "Metric"], [3431, 3434, ["validation", "set", "likelihood"], "Metric"], [3451, 3454, ["sequential", "LSTM", "baseline"], "Method"], [3456, 3458, ["language", "model"], "Method"], [3464, 3468, ["dropout", "rate", "of", "0.3"], "Metric"], [3470, 3471, ["training"], "Task"], [3473, 3476, ["stochastic", "gradient", "descent"], "Method"], [3478, 3482, ["learning", "rate", "of", "0.1"], "Metric"], [3513, 3514, ["parser"], "Method"], [3530, 3532, ["discriminative", "model"], "Method"], [3536, 3539, ["greedy", "decoding", "rule"], "Method"], [3542, 3544, ["beam", "search"], "Method"], [3546, 3550, ["shift", "-", "reduce", "baselines"], "Method"], [3553, 3555, ["generative", "model"], "Method"], [3567, 3569, ["discriminative", "parser"], "Method"], [3584, 3586, ["generative", "model"], "Method"], [3619, 3621, ["Language", "model"], "Method"], [3630, 3632, ["word", "perplexities"], "Metric"], [3634, 3636, ["language", "models"], "Method"], [3638, 3641, ["sequential", "and", "syntactic"], "Task"], [3688, 3690, ["generative", "model"], "Method"], [3696, 3697, ["parser"], "Method"], [3700, 3702, ["language", "model"], "Method"], [3727, 3731, ["continuous", "representations", "of", "symbols"], "Method"], [3732, 3736, ["non", "-", "linear", "models"], "Method"], [3750, 3752, ["discriminative", "model"], "Method"], [3762, 3764, ["generative", "model"], "Method"], [3769, 3771, ["generative", "model"], "Method"], [3804, 3806, ["discriminative", "model"], "Method"], [3819, 3821, ["generative", "model"], "Method"], [3829, 3832, ["fully", "discriminative", "model"], "Method"], [3843, 3845, ["generative", "model"], "Method"], [3854, 3855, ["semisupervision"], "Method"], [3863, 3865, ["discriminative", "parser"], "Method"], [3886, 3888, ["generative", "models"], "Method"], [3899, 3901, ["neural", "networks"], "Method"], [3910, 3912, ["language", "model"], "Method"], [3922, 3927, ["recurrent", "neural", "network", "language", "models"], "Method"], [3931, 3934, ["syntactic", "language", "modeling"], "Method"], [3935, 3940, ["Recurrent", "neural", "network", "language", "models"], "Method"], [3941, 3942, ["RNNs"], "Method"], [3957, 3960, ["right", "language", "model"], "Method"], [3961, 3964, ["Syntactic", "language", "models"], "Method"], [3990, 3997, ["bottom", "-", "up", "generation", "of", "the", "tree"], "Task"], [4001, 4006, ["shift", "-", "reduce", "parser", "actions"], "Method"], [4013, 4019, ["neural", "-", "network", "–", "based", "model"], "Method"], [4032, 4035, ["neural", "network", "architecture"], "Method"], [4042, 4046, ["left", "-", "corner", "model"], "Method"], [4047, 4052, ["Dependency", "-", "only", "language", "models"], "Method"], [4057, 4062, ["Modeling", "generation", "top", "-", "down"], "Task"], [4064, 4067, ["rooted", "branching", "process"], "Method"], [4087, 4089, ["random", "forests"], "Method"], [4111, 4113, ["neural", "networks"], "Method"], [4130, 4134, ["Syntactically", "structured", "neural", "architectures"], "Method"], [4147, 4149, ["sentiment", "analysis"], "Task"], [4151, 4153, ["sentence", "representation"], "Task"], [4173, 4177, ["syntactically", "structured", "neural", "models"], "Method"], [4186, 4188, ["sequential", "RNNs"], "Method"], [4229, 4230, ["RNNGs"], "Method"], [4235, 4239, ["particle", "filter", "inference", "scheme"], "Method"], [4243, 4246, ["importance", "sampling", "method"], "Method"], [4249, 4251, ["discriminative", "parser"], "Method"], [4260, 4267, ["left", "-", "to", "-", "right", "marginalization", "algorithm"], "Method"], [4284, 4286, ["language", "models"], "Method"], [4294, 4297, ["sequential", "generation", "architectures"], "Method"], [4300, 4304, ["neural", "network", "transduction", "problems"], "Method"], [4315, 4317, ["machine", "translation"], "Task"], [4320, 4323, ["conditional", "syntactic", "models"], "Method"], [4331, 4333, ["marginalization", "process"], "Method"], [4342, 4343, ["RNNGs"], "Method"], [4344, 4347, ["human", "sentence", "processing"], "Task"], [4362, 4363, ["RNNG"], "Method"], [4366, 4368, ["processing", "model"], "Method"], [4372, 4373, ["grammar"], "Method"], [4392, 4395, ["sentence", "processing", "models"], "Method"], [4398, 4400, ["explicit", "grammars"], "Method"], [4404, 4406, ["processing", "model"], "Method"], [4415, 4418, ["supervised", "learning", "scenario"], "Task"], [4419, 4420, ["RNNGs"], "Method"], [4421, 4423, ["joint", "models"], "Method"], [4434, 4436, ["expectation", "maximization"], "Method"], [4442, 4446, ["recurrent", "neural", "network", "grammars"], "Method"], [4448, 4455, ["probabilistic", "model", "of", "phrase", "-", "structure", "trees"], "Method"], [4464, 4466, ["language", "model"], "Method"], [4468, 4469, ["parser"], "Method"], [4473, 4475, ["discriminative", "model"], "Method"], [4481, 4482, ["parser"], "Method"], [4485, 4491, ["out", "-", "of", "-", "vocabulary", "preprocessing"], "Method"], [4496, 4498, ["feature", "design"], "Method"], [4499, 4500, ["transformations"], "Method"], [4501, 4503, ["treebank", "data"], "Material"], [4505, 4507, ["generative", "model"], "Method"], [4511, 4512, ["parser"], "Method"], [4516, 4519, ["supervised", "generative", "model"], "Method"], [4530, 4532, ["generative", "model"], "Method"], [4536, 4538, ["language", "models"], "Method"], [4539, 4540, ["RNNGs"], "Method"], [4543, 4548, ["single", "-", "sentence", "language", "models"], "Method"], [4584, 4586, ["data", "preparation"], "Task"], [4803, 4806, ["RNNG", "Composition", "Function"], "Method"], [4810, 4812, ["composition", "function"], "Method"], [4819, 4821, ["vector", "representation"], "Method"], [4823, 4825, ["bidirectional", "LSTM"], "Method"], [4929, 4931, ["implementation", "error"], "Metric"], [4982, 4988, ["English", "and", "Chinese", "language", "modeling", "perplexities"], "Task"], [5016, 5019, ["language", "modeling", "perplexity"], "Metric"], [5020, 5021, ["PTB"], "Material"], [5025, 5027, ["fixed", "RNNG"], "Method"], [5032, 5035, ["sequential", "LSTM", "baseline"], "Method"], [154, 155, ["parsing"], "Task"], [171, 172, ["parsing"], "Task"], [409, 410, ["parsing"], "Task"], [467, 468, ["parsing"], "Task"], [525, 526, ["parsing"], "Task"], [1029, 1030, ["parsing"], "Task"], [1306, 1307, ["parsing"], "Task"], [1441, 1442, ["parsing"], "Task"], [1635, 1636, ["parsing"], "Task"], [2102, 2103, ["parsing"], "Task"], [2143, 2144, ["parsing"], "Task"], [2887, 2888, ["parsing"], "Task"], [3010, 3011, ["parsing"], "Task"], [3496, 3497, ["parsing"], "Task"], [3590, 3591, ["parsing"], "Task"], [3594, 3595, ["parsing"], "Task"], [3780, 3781, ["parsing"], "Task"], [4038, 4039, ["parsing"], "Task"], [4145, 4146, ["parsing"], "Task"], [4937, 4938, ["RNNGs"], "Method"], [4946, 4947, ["PTB"], "Material"], [4949, 4950, ["parsing"], "Task"], [4998, 4999, ["parsing"], "Task"]]} {"doc_id": "15c10b24ef645d83ff4059affd86945c33e00328", "text": ["document", ":", "Snips", "Voice", "Platform", ":", "an", "embedded", "Spoken", "Language", "Understanding", "system", "for", "private", "-", "by", "-", "design", "voice", "interfaces", "This", "paper", "presents", "the", "machine", "learning", "architecture", "of", "the", "Snips", "Voice", "Platform", ",", "a", "software", "solution", "to", "perform", "Spoken", "Language", "Understanding", "on", "microprocessors", "typical", "of", "IoT", "devices", ".", "The", "embedded", "inference", "is", "fast", "and", "accurate", "while", "enforcing", "privacy", "by", "design", ",", "as", "no", "personal", "user", "data", "is", "ever", "collected", ".", "Focusing", "on", "Automatic", "Speech", "Recognition", "and", "Natural", "Language", "Understanding", ",", "we", "detail", "our", "approach", "to", "training", "high", "-", "performance", "Machine", "Learning", "models", "that", "are", "small", "enough", "to", "run", "in", "real", "-", "time", "on", "small", "devices", ".", "Additionally", ",", "we", "describe", "a", "data", "generation", "procedure", "that", "provides", "sufficient", ",", "high", "-", "quality", "training", "data", "without", "compromising", "user", "privacy", ".", "section", ":", "Introduction", "Over", "the", "last", "years", ",", "thanks", "in", "part", "to", "steady", "improvements", "brought", "by", "deep", "learning", "approaches", "to", "speech", "recognition", ",", "voice", "interfaces", "have", "greatly", "evolved", "from", "spotting", "limited", "and", "predetermined", "keywords", "to", "understanding", "arbitrary", "formulations", "of", "a", "given", "intention", ".", "They", "also", "became", "much", "more", "reliable", ",", "with", "state", "-", "of", "-", "the", "-", "art", "speech", "recognition", "engines", "reaching", "human", "level", "in", "English", ".", "This", "achievement", "unlocked", "many", "practical", "applications", "of", "voice", "assistants", "which", "are", "now", "used", "in", "many", "fields", "from", "customer", "support", ",", "to", "autonomous", "cars", ",", "or", "smart", "homes", ".", "In", "particular", ",", "smart", "speaker", "adoption", "by", "the", "public", "is", "on", "the", "rise", ",", "with", "a", "recent", "study", "showing", "that", "nearly", "20", "%", "of", "U.S.", "adults", "reported", "having", "a", "smart", "speaker", "at", "home", ".", "These", "recent", "developments", "however", "raise", "questions", "about", "user", "privacy", "–", "especially", "since", "unique", "speaker", "identification", "is", "an", "active", "field", "of", "research", "using", "voice", "as", "a", "sensitive", "biometric", "feature", ".", "The", "CNIL", "(", "French", "Data", "Protection", "Authority", ")", "advises", "owners", "of", "connected", "speakers", "to", "switch", "off", "the", "microphone", "when", "possible", "and", "to", "warn", "guests", "of", "the", "presence", "of", "such", "a", "device", "in", "their", "home", ".", "The", "General", "Data", "Protection", "Regulation", "which", "harmonizes", "data", "privacy", "laws", "across", "the", "European", "Union", "indeed", "requires", "companies", "to", "ask", "for", "explicit", "consent", "before", "collecting", "user", "data", ".", "Some", "of", "the", "most", "popular", "commercial", "solutions", "for", "voice", "assistants", "include", "Microsoft", "’s", "Cortana", ",", "Google", "’s", "DialogFlow", ",", "IBM", "’s", "Watson", ",", "or", "Amazon", "Alexa", ".", "In", "this", "paper", ",", "we", "introduce", "a", "competing", "solution", ",", "the", "Snips", "Voice", "Platform", "which", ",", "unlike", "the", "previous", "ones", ",", "is", "completely", "cloud", "independent", "and", "runs", "offline", "on", "typical", "IoT", "microprocessors", ",", "thus", "guaranteeing", "privacy", "by", "design", ",", "with", "no", "user", "data", "ever", "collected", "nor", "stored", ".", "The", "Natural", "Language", "Understanding", "component", "of", "the", "platform", "is", "already", "open", "source", ",", "while", "the", "other", "components", "will", "be", "opensourced", "in", "the", "future", ".", "The", "aim", "of", "this", "paper", "is", "to", "contribute", "to", "the", "collective", "effort", "towards", "ever", "more", "private", "and", "efficient", "cloud", "-", "independent", "voice", "interfaces", ".", "To", "this", "end", ",", "we", "devote", "this", "introduction", "to", "a", "brief", "description", "of", "the", "Snips", "Ecosystem", "and", "of", "some", "of", "the", "design", "principles", "behind", "the", "Snips", "Voice", "Platform", ".", "subsection", ":", "The", "Snips", "Ecosystem", "The", "Snips", "ecosystem", "comprises", "a", "web", "console", "to", "build", "voice", "assistants", "and", "train", "the", "corresponding", "Spoken", "Language", "Understanding", "(", "SLU", ")", "engine", ",", "made", "of", "an", "Automatic", "Speech", "Recognition", "(", "ASR", ")", "engine", "and", "a", "Natural", "Language", "Understanding", "(", "NLU", ")", "engine", ".", "The", "console", "can", "be", "used", "as", "a", "self", "-", "service", "development", "environment", "by", "businesses", "or", "individuals", ",", "or", "through", "professional", "services", ".", "The", "Snips", "Voice", "Platform", "is", "free", "for", "non", "-", "commercial", "use", ".", "Since", "its", "launch", "in", "Summer", "2017", ",", "over", "23", ",", "000", "Snips", "voice", "assistants", "have", "been", "created", "by", "over", "13", ",", "000", "developers", ".", "The", "languages", "currently", "supported", "by", "the", "Snips", "platform", "are", "English", ",", "French", "and", "German", ",", "with", "additional", "NLU", "support", "for", "Spanish", "and", "Korean", ".", "More", "languages", "are", "added", "regularly", ".", "An", "assistant", "is", "composed", "of", "a", "set", "of", "skills", "–", "e.g.", "SmartLights", ",", "SmartThermostat", ",", "or", "SmartOven", "skills", "for", "a", "SmartHome", "assistant", "–", "that", "may", "be", "either", "selected", "from", "preexisting", "ones", "in", "a", "skill", "store", "or", "created", "from", "scratch", "on", "the", "web", "console", ".", "A", "given", "skill", "may", "contain", "several", "intents", ",", "or", "user", "intention", "–", "e.g.", "SwitchLightOn", "and", "SwitchLightOff", "for", "a", "SmartLights", "skill", ".", "Finally", ",", "a", "given", "intent", "is", "bound", "to", "a", "list", "of", "entities", "that", "must", "be", "extracted", "from", "the", "user", "’s", "query", "–", "e.g.", "room", "for", "the", "SwitchLightOn", "intent", ".", "We", "call", "slot", "the", "particular", "value", "of", "an", "entity", "in", "a", "query", "–", "e.g.", "kitchen", "for", "the", "entity", "room", ".", "When", "a", "user", "speaks", "to", "the", "assistant", ",", "the", "SLU", "engine", "trained", "on", "the", "different", "skills", "will", "handle", "the", "request", "by", "successively", "converting", "speech", "into", "text", ",", "classifying", "the", "user", "’s", "intent", ",", "and", "extracting", "the", "relevant", "slots", ".", "Once", "the", "user", "’s", "request", "has", "been", "processed", "and", "based", "on", "the", "information", "that", "has", "been", "extracted", "from", "the", "query", "and", "fed", "to", "the", "device", ",", "a", "dialog", "management", "component", "is", "responsible", "for", "providing", "a", "feedback", "to", "the", "user", ",", "or", "performing", "an", "action", ".", "It", "may", "take", "multiple", "forms", ",", "such", "as", "an", "audio", "response", "via", "speech", "synthesis", "or", "a", "direct", "action", "on", "a", "connected", "device", "–", "e.g.", "actually", "turning", "on", "the", "lights", "for", "a", "SmartLights", "skill", ".", "Figure", "[", "reference", "]", "illustrates", "the", "typical", "interaction", "flow", ".", "subsection", ":", "A", "private", "-", "by", "-", "design", "embedded", "platform", "The", "Privacy", "by", "Design", "principle", "sets", "privacy", "as", "the", "default", "standard", "in", "the", "design", "and", "engineering", "of", "a", "system", ".", "In", "the", "context", "of", "voice", "assistants", ",", "that", "can", "be", "deployed", "anywhere", "including", "users", "’", "homes", ",", "this", "principle", "calls", "for", "a", "strong", "interpretation", "to", "protect", "users", "against", "any", "future", "misuse", "of", "their", "private", "data", ".", "In", "the", "following", ",", "we", "call", "private", "-", "by", "-", "design", "a", "system", "that", "does", "not", "transfer", "user", "data", "to", "any", "remote", "location", ",", "such", "as", "cloud", "servers", ".", "Within", "the", "Snips", "ecosystem", ",", "the", "SLU", "components", "are", "trained", "on", "servers", ",", "but", "the", "inference", "happens", "directly", "on", "the", "device", "once", "the", "assistant", "has", "been", "deployed", "-", "no", "data", "from", "the", "user", "is", "ever", "collected", "nor", "stored", ".", "This", "design", "choice", "adds", "engineering", "complexity", "as", "most", "IoT", "devices", "run", "on", "specific", "hardware", "with", "limited", "memory", "and", "computing", "power", ".", "Cross", "-", "platform", "support", "is", "also", "a", "requirement", "in", "the", "IoT", "industry", ",", "since", "IoT", "devices", "are", "powered", "by", "many", "different", "hardware", "boards", ",", "with", "sustained", "innovation", "in", "that", "field", ".", "For", "these", "reasons", ",", "the", "Snips", "Voice", "Platform", "has", "been", "built", "with", "portability", "and", "footprint", "in", "mind", ".", "Its", "embedded", "inference", "runs", "on", "common", "IoT", "hardware", "as", "light", "as", "the", "Raspberry", "Pi", "3", "(", "CPU", "with", "1.4", "GHz", "and", "1", "GB", "of", "RAM", ")", ",", "a", "popular", "choice", "among", "developers", "and", "therefore", "our", "reference", "hardware", "setting", "throughout", "this", "paper", ".", "Other", "Linux", "boards", "are", "also", "supported", ",", "such", "as", "IMX.7D", ",", "i", ".", "MX8", "M", ",", "DragonBoard", "410c", ",", "and", "Jetson", "TX2", ".", "The", "Snips", "SDK", "for", "Android", "works", "with", "devices", "with", "Android", "5", "and", "ARM", "CPU", ",", "while", "the", "iOS", "SDK", "targets", "iOS", "11", "and", "newer", ".", "For", "efficiency", "and", "portability", "reasons", ",", "the", "algorithms", "have", "been", "re", "-", "implemented", "whenever", "needed", "in", "Rust", "–", "a", "modern", "programming", "language", "offering", "high", "performance", ",", "low", "memory", "overhead", ",", "and", "cross", "-", "compilation", ".", "SLU", "engines", "are", "usually", "broken", "down", "into", "two", "parts", ":", "Automatic", "Speech", "Recognition", "(", "ASR", ")", "and", "Natural", "Language", "Understanding", "(", "NLU", ")", ".", "The", "ASR", "engine", "translates", "a", "spoken", "utterance", "into", "text", "through", "an", "acoustic", "model", ",", "mapping", "raw", "audio", "to", "a", "phonetic", "representation", ",", "and", "a", "Language", "Model", "(", "LM", ")", ",", "mapping", "this", "phonetic", "representation", "to", "text", ".", "The", "NLU", "then", "extracts", "intent", "and", "slots", "from", "the", "decoded", "query", ".", "As", "discussed", "in", "section", "[", "reference", "]", ",", "LM", "and", "NLU", "have", "to", "be", "mutually", "consistent", "in", "order", "to", "optimize", "the", "accuracy", "of", "the", "SLU", "engine", ".", "It", "is", "therefore", "useful", "to", "introduce", "a", "language", "modeling", "component", "composed", "of", "the", "LM", "and", "NLU", ".", "Figure", "[", "reference", "]", "describes", "the", "building", "blocks", "of", "the", "SLU", "pipeline", ".", "As", "stated", "above", ",", "ASR", "engines", "relying", "on", "large", "deep", "learning", "models", "have", "improved", "drastically", "over", "the", "past", "few", "years", ".", "Yet", ",", "they", "still", "have", "a", "major", "drawback", "today", ".", "For", "example", ",", "the", "model", "achieving", "human", "parity", "in", "is", "a", "combination", "of", "several", "neural", "networks", ",", "each", "containing", "several", "hundreds", "of", "millions", "of", "parameters", ",", "and", "large", "-", "vocabulary", "language", "models", "made", "of", "several", "millions", "of", "n", "-", "grams", ".", "The", "size", "of", "these", "models", ",", "along", "with", "the", "computational", "resources", "necessary", "to", "run", "them", "in", "real", "-", "time", ",", "make", "them", "unfit", "for", "deployment", "on", "small", "devices", ",", "so", "that", "solutions", "implementing", "them", "are", "bound", "to", "rely", "on", "the", "cloud", "for", "speech", "recognition", ".", "Enforcing", "privacy", "by", "design", "therefore", "implies", "developing", "new", "tools", "to", "build", "reliable", "SLU", "engines", "that", "are", "constrained", "in", "size", "and", "computational", "requirements", ",", "which", "we", "detail", "in", "this", "paper", ".", "In", "section", "[", "reference", "]", ",", "we", "describe", "strategies", "to", "obtain", "small", "(", "10", "MB", ")", "and", "robust", "acoustic", "models", "trained", "on", "general", "speech", "corpora", "of", "a", "few", "hundred", "to", "a", "few", "thousand", "hours", ".", "Section", "[", "reference", "]", "is", "devoted", "to", "a", "description", "of", "the", "language", "modeling", "approach", "of", "the", "Snips", "SLU", "engine", ".", "Notably", ",", "we", "show", "how", "to", "ensure", "consistency", "between", "the", "language", "model", "of", "the", "ASR", "engine", "and", "the", "NLU", "engine", "while", "specializing", "them", "to", "a", "particular", "use", "case", ".", "The", "resulting", "SLU", "engine", "is", "lightweight", "and", "fast", "to", "execute", ",", "making", "it", "fit", "for", "deployment", "on", "small", "devices", "and", "the", "NLU", "component", "is", "open", "source", ".", "In", "section", "[", "reference", "]", ",", "we", "illustrate", "the", "high", "generalization", "accuracy", "of", "the", "SLU", "engine", "in", "the", "context", "of", "real", "-", "word", "voice", "assistants", ".", "Finally", ",", "we", "discuss", "in", "section", "[", "reference", "]", "a", "data", "generation", "procedure", "to", "automatically", "create", "training", "sets", "replacing", "user", "data", ".", "section", ":", "Acoustic", "model", "The", "acoustic", "model", "is", "the", "first", "step", "of", "the", "SLU", "pipeline", ",", "and", "is", "therefore", "crucial", "to", "its", "functioning", ".", "If", "the", "decoding", "contains", "errors", ",", "it", "might", "compromise", "the", "subsequent", "steps", "and", "trigger", "a", "different", "action", "than", "that", "intended", "by", "the", "user", ".", "The", "acoustic", "model", "is", "responsible", "for", "converting", "raw", "audio", "data", "to", "what", "can", "approximately", "be", "interpreted", "as", "phone", "probabilities", ",", "i.e.", "context", "-", "dependent", "clustered", "Hidden", "Markov", "Model", "(", "HMM", ")", "state", "probabilities", ".", "These", "probabilities", "are", "then", "fed", "to", "a", "language", "model", ",", "which", "decodes", "a", "sequence", "of", "words", "corresponding", "to", "the", "user", "utterance", ".", "The", "acoustic", "and", "language", "models", "are", "thus", "closely", "related", "in", "the", "Automatic", "Speech", "Recognition", "(", "ASR", ")", "engine", ",", "but", "are", "often", "designed", "and", "trained", "separately", ".", "The", "construction", "of", "the", "language", "model", "used", "in", "the", "SLU", "engine", "is", "detailed", "in", "section", "[", "reference", "]", ".", "In", "this", "section", ",", "we", "present", "the", "acoustic", "model", ".", "First", ",", "we", "give", "details", "about", "how", "the", "training", "data", "is", "collected", ",", "processed", ",", "cleaned", ",", "and", "augmented", ".", "Then", ",", "we", "present", "the", "acoustic", "model", "itself", "(", "a", "hybrid", "of", "Neural", "Networks", "and", "Hidden", "Markov", "Models", ",", "or", "NN", "/", "HMM", ")", "and", "how", "it", "is", "trained", ".", "Finally", ",", "we", "present", "the", "performance", "of", "the", "acoustic", "model", "in", "a", "large", "-", "vocabulary", "setup", ",", "in", "terms", "of", "word", "error", "rate", "(", "WER", ")", ",", "speed", ",", "and", "memory", "usage", ".", "subsection", ":", "Data", "paragraph", ":", "Training", "data", ".", "To", "train", "the", "acoustic", "model", ",", "we", "need", "several", "hundreds", "to", "thousands", "of", "hours", "of", "audio", "data", "with", "corresponding", "transcripts", ".", "The", "data", "is", "collected", "from", "public", "or", "commercial", "sources", ".", "A", "realignment", "of", "transcripts", "to", "the", "audio", "is", "performed", "to", "match", "transcripts", "to", "timestamps", ".", "This", "additionally", "helps", "in", "removing", "transcription", "errors", "that", "might", "be", "present", "in", "the", "data", ".", "The", "result", "is", "a", "set", "of", "audio", "extracts", "and", "matching", "transcripts", ",", "with", "lengths", "suitable", "for", "acoustic", "training", "(", "up", "to", "a", "few", "dozen", "seconds", ")", ".", "This", "data", "is", "split", "in", "a", "training", ",", "testing", ",", "and", "development", "sets", ".", "paragraph", ":", "Data", "augmentation", ".", "One", "of", "the", "main", "issues", "regarding", "the", "training", "of", "the", "acoustic", "model", "is", "the", "lack", "of", "data", "corresponding", "to", "real", "usage", "scenari", ".", "Most", "of", "the", "available", "training", "data", "is", "clear", "close", "-", "field", "speech", ",", "but", "voice", "assistants", "will", "often", "be", "used", "in", "noisy", "conditions", "(", "music", ",", "television", ",", "environment", "noise", ")", ",", "from", "a", "distance", "of", "several", "meters", "in", "far", "-", "field", "conditions", ",", "and", "in", "rooms", "or", "cars", "with", "reverberation", ".", "From", "a", "machine", "learning", "perspective", ",", "data", "corresponding", "to", "real", "usage", "of", "the", "system", "–", "-", "or", "in", "-", "domain", "data", "–", "is", "extremely", "valuable", ".", "Since", "spoken", "utterances", "from", "the", "user", "are", "not", "collected", "by", "our", "platform", "for", "privacy", "reasons", ",", "noisy", "and", "reverberant", "conditions", "are", "simulated", "by", "augmenting", "the", "data", ".", "Thousands", "of", "virtual", "rooms", "of", "different", "sizes", "are", "thus", "generated", "with", "random", "microphone", "and", "speaker", "locations", ",", "and", "the", "rerecording", "of", "the", "original", "data", "in", "those", "conditions", "is", "simulated", "using", "a", "method", "close", "to", "that", "presented", "in", ".", "subsection", ":", "Model", "training", "Acoustic", "models", "are", "hybrid", "NN", "/", "HMM", "models", ".", "More", "specifically", ",", "they", "are", "a", "custom", "version", "of", "the", "s5", "training", "recipe", "of", "the", "Kaldi", "toolkit", ".", "40", "MFCC", "features", "are", "extracted", "from", "the", "audio", "signal", "with", "windows", "of", "size", "25ms", "every", "10ms", ".", "Models", "with", "a", "variable", "number", "of", "layers", "and", "neurons", "can", "be", "trained", ",", "which", "will", "impact", "their", "accuracy", "and", "computational", "cost", ".", "We", "can", "thus", "train", "different", "model", "architectures", "depending", "on", "the", "target", "hardware", "and", "the", "desired", "application", "accuracy", ".", "In", "the", "following", "evaluation", "(", "section", "[", "reference", "]", ")", ",", "we", "present", "performance", "results", "of", "a", "model", "targeted", "for", "the", "Raspberry", "Pi", "3", ".", "First", ",", "a", "speaker", "-", "adaptive", "Gaussian", "Mixture", "Model", "Hidden", "Markov", "Model", "(", "GMM", "-", "HMM", ")", "is", "trained", "on", "the", "speech", "corpus", "to", "obtain", "a", "context", "-", "dependent", "bootstrapping", "model", "with", "which", "we", "align", "the", "full", "dataset", "and", "extract", "lattices", "to", "prepare", "the", "neural", "network", "training", ".", "We", "train", "a", "deep", "neural", "network", ",", "consisting", "of", "time", "-", "delay", "layers", "similar", "to", "those", "presented", "in", ",", "and", "long", "short", "-", "term", "memory", "layers", "similar", "to", "those", "of", ".", "The", "architecture", ",", "close", "to", "that", "of", ",", "is", "summarized", "in", "Figure", "[", "reference", "]", "and", "Table", "[", "reference", "]", ".", "The", "Raspberry", "Pi", "3", "model", "uses", "7", "layers", ",", "and", "is", "trained", "with", "the", "lattice", "-", "free", "Maximum", "Mutual", "Information", "(", "MMI", ")", "criterion", ",", "using", "natural", "gradient", "descent", ",", "with", "a", "learning", "rate", "of", "0.0005", "and", "the", "backstitching", "trick", ".", "We", "follow", "the", "approach", "described", "in", "to", "create", "fast", "acoustic", "models", ",", "namely", "an", "HMM", "topology", "with", "one", "state", "for", "each", "of", "the", "1", ",", "700", "context", "-", "dependent", "senones", ",", "operating", "at", "a", "third", "of", "the", "original", "frame", "rate", ".", "subsection", ":", "Acoustic", "model", "evaluation", "In", "this", "section", ",", "we", "present", "an", "evaluation", "of", "our", "acoustic", "model", "for", "English", ".", "Our", "goal", "is", "the", "design", "of", "an", "end", "-", "to", "-", "end", "SLU", "pipeline", "which", "runs", "in", "real", "-", "time", "on", "small", "embedded", "devices", ",", "but", "has", "state", "-", "of", "-", "the", "-", "art", "accuracy", ".", "This", "requires", "tradeoffs", "between", "speed", "and", "the", "generality", "of", "SLU", "task", ".", "More", "precisely", ",", "we", "use", "domain", "-", "adapted", "language", "models", "described", "in", "section", "[", "reference", "]", ",", "to", "compensate", "for", "the", "decrease", "of", "accuracy", "of", "smaller", "acoustic", "models", ".", "However", "in", "order", "to", "assess", "the", "quality", "of", "the", "acoustic", "model", "in", "a", "more", "general", "setting", ",", "the", "evaluation", "of", "this", "section", "is", "carried", "out", "in", "a", "large", "vocabulary", "setup", ",", "on", "the", "LibriSpeech", "evaluation", "dataset", ",", "chosen", "because", "it", "is", "freely", "available", "and", "widely", "used", "in", "state", "-", "of", "-", "the", "-", "art", "comparisons", ".", "The", "language", "model", "for", "the", "large", "-", "vocabulary", "evaluation", "is", "also", "freely", "available", "online", ".", "It", "is", "a", "pruned", "trigram", "LM", "with", "a", "vocabulary", "of", "200k", "words", ",", "trained", "on", "the", "content", "of", "public", "domain", "books", ".", "Two", "sets", "of", "experiments", "are", "reported", ".", "In", "the", "first", "set", ",", "the", "models", "are", "trained", "only", "on", "the", "LibriSpeech", "training", "set", "(", "or", "on", "a", "subset", "of", "it", ")", ".", "It", "allows", "us", "to", "validate", "our", "training", "approach", "and", "keep", "track", "of", "how", "the", "models", "we", "develop", "compare", "to", "the", "state", "of", "the", "art", "when", "trained", "on", "public", "data", ".", "Then", ",", "the", "performance", "of", "the", "model", "in", "terms", "of", "speed", "and", "memory", "usage", "is", "studied", ",", "which", "allows", "us", "to", "select", "a", "good", "tradeoff", "for", "the", "targeted", "Raspberry", "Pi", "3", "setting", ".", "subsubsection", ":", "Model", "architecture", "trained", "and", "evaluated", "on", "LibriSpeech", "To", "evaluate", "the", "impact", "of", "the", "dataset", "and", "model", "sizes", "on", "the", "model", "accuracy", ",", "neural", "networks", "of", "different", "sizes", "are", "trained", "on", "different", "subsets", "of", "the", "LibriSpeech", "dataset", ",", "with", "and", "without", "data", "augmentation", ".", "The", "results", "obtained", "with", "nnet", "-", "512", "are", "reported", "in", "Table", "[", "reference", "]", ".", "The", "Num", ".", "hours", "column", "corresponds", "to", "the", "number", "of", "training", "hours", "(", "460h", "in", "the", "train", "-", "clean", "split", "of", "the", "LibriSpeech", "dataset", "and", "500h", "in", "the", "train", "-", "other", "split", ")", ".", "The", "data", "augmentation", "was", "only", "applied", "to", "the", "clean", "data", ".", "For", "example", "460x2", "means", "460h", "of", "clean", "data", "+", "460h", "of", "augmented", "data", ".", "We", "observe", "that", "adding", "data", "does", "not", "have", "much", "impact", "on", "LibriSpeech", "’s", "clean", "test", "sets", "(", "dev", "-", "clean", "and", "test", "-", "clean", ")", ".", "The", "WER", "however", "decreases", "when", "adding", "data", "on", "the", "datasets", "marked", "as", "other", "(", "dev", "-", "other", ",", "test", "-", "other", ")", ".", "In", "general", "(", "not", "shown", "in", "those", "tests", ")", ",", "adding", "more", "data", "and", "using", "data", "augmentation", "increases", "significantly", "the", "performance", "on", "noisy", "and", "reverberant", "conditions", ".", "In", "the", "next", "experiment", ",", "the", "neural", "networks", "are", "trained", "with", "the", "same", "architecture", "but", "different", "layer", "sizes", "on", "the", "460x6", "+", "500", "hours", "dataset", ".", "Results", "are", "reported", "in", "Table", "[", "reference", "]", ".", "This", "shows", "that", "larger", "models", "are", "capable", "of", "fitting", "the", "data", "and", "generalizing", "better", ",", "as", "expected", ".", "This", "allows", "us", "to", "choose", "the", "best", "tradeoff", "between", "precision", "and", "computational", "cost", "depending", "on", "each", "target", "hardware", "and", "assistant", "needs", ".", "subsubsection", ":", "Online", "recognition", "performance", "While", "it", "is", "possible", "to", "get", "closer", "to", "the", "state", "of", "the", "art", "using", "larger", "neural", "network", "architectures", ",", "their", "associated", "memory", "and", "computational", "costs", "would", "prohibit", "their", "deployment", "on", "small", "devices", ".", "In", "section", "[", "reference", "]", ",", "we", "show", "how", "carefully", "adapting", "the", "LM", "allows", "to", "reach", "high", "end", "-", "to", "-", "end", "accuracies", "using", "the", "acoustic", "models", "described", "here", ".", "We", "now", "report", "experiments", "on", "the", "processing", "speed", "of", "these", "models", "on", "our", "target", "Raspberry", "Pi", "3", "hardware", "setting", ".", "We", "trained", "models", "with", "various", "sizes", "enjoying", "a", "faster", "-", "than", "-", "real", "-", "time", "processing", "factor", ",", "to", "account", "for", "additional", "processing", "time", "(", "necessitated", "e.g.", "by", "the", "LM", "decoding", "or", "the", "NLU", "engine", ")", ",", "and", "chose", "a", "model", "with", "a", "good", "compromise", "of", "accuracy", "to", "real", "-", "time", "factor", "and", "model", "size", "(", "on", "disk", "and", "in", "RAM", ")", ".", "As", "a", "reference", ",", "in", "terms", "of", "model", "size", "(", "as", "reported", "in", "Table", "[", "reference", "]", ")", "nnet", "-", "256", "is", "nearly", "six", "times", "smaller", "than", "nnet", "-", "768", ",", "with", "2.6", "M", "parameters", "vs", "15.4", "M", ",", "representing", "10", "MB", "vs", "59", "MB", "on", "disk", ".", "The", "gain", "is", "similar", "in", "RAM", ".", "In", "terms", "of", "speed", ",", "the", "nnet", "-", "256", "is", "6", "to", "10", "times", "faster", "than", "the", "nnet", "-", "768", ".", "These", "tradeoffs", "and", "comparison", "with", "other", "trained", "models", "led", "us", "to", "select", "the", "nnet", "-", "256", ".", "It", "has", "a", "reasonable", "speed", "and", "memory", "footprint", ",", "and", "the", "loss", "in", "accuracy", "is", "compensated", "by", "the", "adapted", "LM", "and", "robust", "NLU", ".", "This", "network", "architecture", "and", "size", "will", "be", "the", "one", "used", "in", "the", "subsequent", "experiments", ".", "The", "different", "architecture", "variations", "presented", "in", "this", "section", "were", "chosen", "for", "the", "sake", "of", "comparison", "and", "demonstration", ".", "This", "experimental", "comparison", ",", "along", "with", "optional", "layer", "factorization", "(", "similar", "to", ")", "or", "weight", "quantization", "are", "carried", "out", "for", "each", "target", "hardware", "setting", ",", "but", "this", "analysis", "is", "out", "of", "the", "scope", "of", "this", "paper", ".", "section", ":", "Language", "Modeling", "We", "now", "turn", "to", "the", "description", "of", "the", "language", "modeling", "component", "of", "the", "Snips", "platform", ",", "which", "is", "responsible", "for", "the", "extraction", "of", "the", "intent", "and", "slots", "from", "the", "output", "of", "the", "acoustic", "model", ".", "This", "component", "is", "made", "up", "of", "two", "closely", "-", "interacting", "parts", ".", "The", "first", "is", "the", "language", "model", "(", "LM", ")", ",", "that", "turns", "the", "predictions", "of", "the", "acoustic", "model", "into", "likely", "sentences", ",", "taking", "into", "account", "the", "probability", "of", "co", "-", "occurrence", "of", "words", ".", "The", "second", "is", "the", "Natural", "Language", "Understanding", "(", "NLU", ")", "model", ",", "that", "extracts", "intent", "and", "slots", "from", "the", "prediction", "of", "the", "Automatic", "Speech", "Recognition", "(", "ASR", ")", "engine", ".", "In", "typical", "commercial", "large", "vocabulary", "speech", "recognition", "systems", ",", "the", "LM", "component", "is", "usually", "the", "largest", "in", "size", ",", "and", "can", "take", "up", "to", "terabytes", "of", "storage", ".", "Indeed", ",", "to", "account", "for", "the", "high", "variability", "of", "general", "spoken", "language", ",", "large", "vocabulary", "language", "models", "need", "to", "be", "trained", "on", "very", "large", "text", "corpora", ".", "The", "size", "of", "these", "models", "also", "has", "an", "impact", "on", "decoding", "performance", ":", "the", "search", "space", "of", "the", "ASR", "is", "expanded", ",", "making", "speech", "recognition", "harder", "and", "more", "computationally", "demanding", ".", "Additionally", ",", "the", "performance", "of", "an", "ASR", "engine", "on", "a", "given", "domain", "will", "strongly", "depend", "on", "the", "perplexity", "of", "its", "LM", "on", "queries", "from", "this", "domain", ",", "making", "the", "choice", "of", "the", "training", "text", "corpus", "critical", ".", "This", "question", "is", "sometimes", "addressed", "through", "massive", "use", "of", "users", "’", "private", "data", ".", "One", "option", "to", "overcome", "these", "challenges", "is", "to", "specialize", "the", "language", "model", "of", "the", "assistant", "to", "a", "certain", "domain", ",", "e.g.", "by", "restricting", "its", "vocabulary", "as", "well", "as", "the", "variety", "of", "the", "queries", "it", "should", "model", ".", "While", "this", "approach", "appears", "to", "restrict", "the", "range", "of", "queries", "that", "can", "be", "made", "to", "an", "assistant", ",", "we", "argue", "that", "it", "does", "not", "impair", "the", "usability", "of", "the", "resulting", "assistant", ".", "In", "fact", ",", "while", "the", "performance", "of", "an", "ASR", "engine", "alone", "can", "be", "measured", "using", "e.g.", "the", "word", "error", "rate", "as", "in", "the", "previous", "section", ",", "we", "assess", "the", "performance", "of", "the", "SLU", "system", "through", "its", "end", "-", "to", "-", "end", ",", "speech", "-", "to", "-", "meaning", "accuracy", ",", "i.e.", "its", "ability", "to", "correctly", "predict", "the", "intent", "and", "slots", "of", "a", "spoken", "utterance", ".", "As", "a", "consequence", ",", "it", "is", "sufficient", "for", "the", "LM", "to", "correctly", "model", "the", "sentences", "that", "are", "in", "the", "domain", "that", "the", "NLU", "supports", ".", "The", "size", "of", "the", "model", "is", "thus", "greatly", "reduced", ",", "and", "the", "decoding", "speed", "increases", ".", "The", "resulting", "ASR", "is", "particularly", "robust", "within", "the", "use", "case", ",", "with", "an", "accuracy", "unreachable", "under", "our", "hardware", "constraints", "for", "an", "all", "-", "purpose", ",", "general", "ASR", "model", ".", "In", "the", "following", ",", "we", "detail", "the", "implementation", "of", "this", "design", "principle", ",", "allowing", "the", "Snips", "SLU", "component", "to", "run", "efficiently", "on", "small", "devices", "with", "high", "accuracy", ",", "and", "illustrate", "its", "performance", "on", "two", "real", "-", "world", "assistants", ".", "subsection", ":", "Data", "In", "application", "of", "the", "principles", "outlined", "above", ",", "we", "use", "the", "same", "data", "to", "train", "both", "LM", "and", "NLU", ".", "The", "next", "section", "is", "devoted", "to", "a", "description", "of", "this", "dataset", ".", "The", "generation", "of", "this", "dataset", "is", "discussed", "in", "section", "[", "reference", "]", ".", "subsubsection", ":", "Training", "dataset", "The", "dataset", "used", "to", "train", "both", "the", "LM", "and", "NLU", "contains", "written", "queries", "exemplifying", "intents", "that", "depend", "on", "entities", ".", "Entities", "are", "bound", "to", "an", "intent", "and", "used", "to", "describe", "all", "the", "possible", "values", "for", "a", "given", "attribute", ".", "For", "example", ",", "in", "the", "case", "of", "a", "SmartLights", "assistant", "handling", "connected", "lights", ",", "these", "entities", "are", "room", ",", "brightness", "and", "color", ".", "They", "are", "required", "by", "the", "assistant", "logic", "to", "execute", "the", "right", "action", ".", "Another", "example", "dealing", "with", "weather", "-", "related", "queries", "is", "described", "in", "section", "[", "reference", "]", ".", "An", "intent", "often", "has", "several", "entities", "that", "it", "can", "share", "with", "other", "intents", ".", "For", "instance", ",", "the", "room", "entity", "is", "used", "by", "several", "intents", "(", "SwitchLightOn", "and", "SwitchLightOff", ")", ",", "since", "the", "user", "might", "want", "to", "specify", "the", "room", "for", "both", "switching", "on", "and", "switching", "off", "the", "lights", ".", "Entities", "can", "be", "of", "two", "types", ",", "either", "custom", "or", "built", "-", "in", ".", "Custom", "entities", "are", "user", "-", "defined", "entities", "that", "can", "be", "exhaustively", "specified", "by", "a", "list", "of", "values", "(", "e.g.", "room", ":", "kitchen", ",", "bedroom", ",", "etc", ".", "and", "color", ":", "blue", ",", "red", ",", "etc", ".", ")", ".", "Built", "-", "in", "entities", "are", "common", "entities", "that", "can", "not", "be", "easily", "listed", "exhaustively", "by", "a", "user", ",", "and", "are", "therefore", "provided", "by", "the", "platform", "(", "numbers", ",", "ordinals", ",", "amounts", "with", "unit", ",", "date", "and", "times", ",", "durations", ",", "etc", ".", ")", ".", "In", "our", "SmartLights", "example", ",", "the", "entity", "brightness", "can", "be", "any", "number", "between", "0", "and", "100", ",", "so", "that", "the", "built", "-", "in", "entity", "type", "snips", "/", "number", "can", "be", "used", ".", "A", "query", "is", "the", "written", "expression", "of", "an", "intent", ".", "For", "instance", ",", "the", "query", "‘", "‘", "set", "the", "kitchen", "lights", "intensity", "to", "65", "’’", "is", "associated", "with", "the", "intent", "SetLightBrightness", ".", "Slot", "labeling", "is", "done", "by", "specifying", "chunks", "of", "the", "query", "that", "should", "be", "bound", "to", "a", "given", "entity", ".", "Using", "the", "same", "examples", ",", "the", "slots", "associated", "with", "the", "room", "and", "brightness", "entities", "in", "the", "query", "can", "be", "specified", "as", "follows", ":", "‘", "‘", "set", "the", "(", "kitchen", ")[", "room", "]", "lights", "intensity", "to", "(", "65", ")[", "brightness", "]", "’’", ".", "The", "number", "of", "queries", "per", "intent", "ranges", "from", "a", "few", "ones", "to", "several", "thousands", "depending", "on", "the", "variability", "needed", "to", "cover", "most", "common", "wordings", ".", "subsubsection", ":", "Normalization", "One", "key", "challenge", "related", "to", "end", "-", "to", "-", "end", "SLU", "is", "data", "consistency", "between", "training", "and", "inference", ".", "The", "dataset", "described", "above", "is", "collected", "via", "the", "console", "where", "no", "specific", "writing", "system", ",", "nor", "cleaning", "rules", "regarding", "non", "-", "alphanumeric", "characters", "are", "enforced", ".", "Before", "training", "the", "LM", ",", "this", "dataset", "therefore", "needs", "to", "be", "verbalized", ":", "entity", "values", "and", "user", "queries", "are", "tokenized", ",", "normalized", "to", "a", "canonical", "form", ",", "and", "verbalized", "to", "match", "entries", "from", "a", "lexicon", ".", "For", "instance", ",", "numbers", "and", "dates", "are", "spelled", "out", ",", "so", "that", "their", "pronunciation", "can", "be", "generated", "from", "their", "written", "form", ".", "Importantly", ",", "we", "apply", "the", "same", "preprocessing", "before", "training", "the", "NLU", ".", "This", "step", "ensures", "consistency", "when", "it", "comes", "to", "inference", ".", "More", "precisely", ",", "it", "guarantees", "that", "the", "words", "output", "by", "the", "ASR", "match", "those", "seen", "by", "the", "NLU", "during", "training", ".", "The", "normalization", "pipeline", "is", "used", "to", "handle", "languages", "specificities", ",", "through", "the", "use", "of", "a", "class", "-", "based", "tokenizer", "that", "allows", "support", "for", "case", "-", "by", "-", "case", "verbalization", "for", "each", "token", "class", ".", "For", "instance", ",", "numeric", "values", "are", "transliterated", "to", "words", ",", "punctuation", "tokens", "skipped", ",", "while", "quantities", "with", "units", "such", "as", "amounts", "of", "money", "require", "a", "more", "advanced", "verbalization", "(", "in", "English", ",", "‘", "‘", "$", "25", "’’", "should", "be", "verbalized", "as", "‘", "‘", "twenty", "five", "dollars", "’", "’", ")", ".", "The", "tokenizer", "is", "implemented", "as", "a", "character", "-", "level", "finite", "state", "transducer", ",", "and", "is", "designed", "to", "be", "easily", "extensible", "to", "accommodate", "new", "token", "types", "as", "more", "languages", "are", "supported", ".", "subsection", ":", "Language", "model", "The", "mapping", "from", "the", "output", "of", "the", "acoustic", "model", "to", "likely", "word", "sequences", "is", "done", "via", "a", "Viterbi", "search", "in", "a", "weighted", "Finite", "State", "Transducer", "(", "wFST", ")", ",", "called", "ASR", "decoding", "graph", "in", "the", "following", ".", "Formally", ",", "the", "decoding", "graph", "may", "be", "written", "as", "the", "composition", "of", "four", "wFSTs", ",", "where", "denotes", "transducer", "composition", "(", "see", "section", "[", "reference", "]", ")", ",", "represents", "Hidden", "Markov", "Models", "(", "HMMs", ")", "modeling", "context", "-", "dependent", "phones", ",", "represents", "the", "context", "-", "dependency", ",", "is", "the", "lexicon", "and", "is", "the", "LM", ",", "typically", "a", "bigram", "or", "a", "trigram", "model", "represented", "as", "a", "wFST", ".", "Determinization", "and", "minimization", "operations", "are", "also", "applied", "at", "each", "step", "in", "order", "to", "compute", "equivalent", "optimized", "transducers", "with", "less", "states", ",", "allowing", "the", "composition", "and", "the", "inference", "to", "run", "faster", ".", "More", "detailed", "definitions", "of", "the", "previous", "classical", "transducers", "are", "beyond", "the", "scope", "of", "this", "paper", ",", "and", "we", "refer", "the", "interested", "reader", "to", "and", "references", "therein", ".", "In", "the", "following", ",", "we", "focus", "on", "the", "construction", "of", "the", "G", "transducer", ",", "encoding", "the", "LM", ",", "from", "the", "domain", "-", "specific", "dataset", "presented", "above", ".", "subsubsection", ":", "Language", "Model", "Adaptation", "As", "explained", "earlier", ",", "the", "ASR", "engine", "is", "required", "to", "understand", "arbitrary", "formulations", "of", "a", "finite", "set", "of", "intents", "described", "in", "the", "dataset", ".", "In", "particular", ",", "it", "should", "be", "able", "to", "generalize", "to", "unseen", "queries", "within", "the", "same", "domain", ",", "and", "allow", "entity", "values", "to", "be", "interchangeable", ".", "The", "generalization", "properties", "of", "the", "ASR", "engine", "are", "preserved", "by", "using", "a", "statistical", "n", "-", "gram", "LM", "allowing", "to", "mix", "parts", "of", "the", "training", "queries", "to", "create", "new", "ones", ",", "and", "by", "using", "class", "-", "based", "language", "modeling", "where", "the", "value", "of", "each", "entity", "may", "be", "replaced", "by", "any", "other", ".", "We", "now", "detail", "the", "resulting", "LM", "construction", "strategy", ".", "The", "first", "step", "in", "building", "the", "LM", "is", "the", "creation", "of", "patterns", "abstracting", "the", "type", "of", "queries", "the", "user", "may", "make", "to", "the", "assistant", ".", "Starting", "from", "the", "dataset", "described", "above", ",", "we", "replace", "all", "occurrences", "of", "each", "entity", "by", "a", "symbol", "for", "the", "entity", ".", "For", "example", ",", "the", "query", "‘", "‘", "Play", "some", "music", "by", "(", "The", "Rolling", "Stones", ")[", "artist", "]", "’’", "is", "abstracted", "to", "‘", "‘", "Play", "some", "music", "by", "ARTIST", "’’", ".", "An", "n", "-", "gram", "model", "is", "then", "trained", "on", "the", "resulting", "set", "of", "patterns", ",", "which", "is", "then", "converted", "to", "a", "wFST", "called", ".", "Next", ",", "for", "each", "entity", "where", "and", "is", "the", "number", "of", "entities", ",", "an", "acceptor", "is", "defined", "to", "encode", "the", "values", "the", "entity", "can", "take", ".", "The", "construction", "of", "depends", "on", "the", "type", "of", "the", "entity", ".", "For", "custom", "entities", ",", "whose", "values", "are", "listed", "exhaustively", "in", "the", "dataset", ",", "can", "be", "defined", "either", "as", "a", "union", "of", "acceptors", "of", "the", "different", "values", "of", "the", "entity", ",", "or", "as", "an", "n", "-", "gram", "model", "trained", "specifically", "on", "the", "values", "of", "the", "entity", ".", "For", "built", "-", "in", "entities", "such", "as", "numbers", "or", "dates", "and", "times", ",", "is", "a", "wFST", "representation", "of", "a", "generative", "grammar", "describing", "the", "construction", "of", "any", "instance", "of", "the", "entity", ".", "The", "LM", "transducer", "is", "then", "defined", "as", "where", "Replace", "denotes", "wFST", "replacement", ".", "For", "instance", ",", "in", "the", "example", "above", ",", "the", "arcs", "of", "carrying", "the", "‘", "‘", "ARTIST", "’’", "symbol", "are", "expanded", "into", "the", "wFST", "representing", "the", "‘", "‘", "artist", "’’", "entity", ".", "This", "process", "is", "represented", "on", "a", "simple", "LM", "on", "Figure", "[", "reference", "]", ".", "The", "resulting", "allows", "the", "ASR", "to", "generalize", "to", "unseen", "queries", "and", "to", "swap", "entity", "values", ".", "Continuing", "with", "the", "simple", "example", "introduced", "above", ",", "the", "query", "‘", "‘", "Play", "me", "some", "music", "by", "The", "Beatles", "’’", "has", "the", "same", "weight", "as", "‘", "‘", "Play", "me", "some", "music", "by", "The", "Rolling", "Stones", "’’", "in", "the", "LM", ",", "while", "the", "sentence", "‘", "‘", "Play", "music", "by", "The", "Rolling", "Stones", "’’", "also", "has", "a", "finite", "weight", "thanks", "to", "the", "n", "-", "gram", "back", "-", "off", "mechanism", ".", "The", "lexicon", "transducer", "encodes", "the", "pronunciations", "of", "all", "the", "words", "in", "both", "and", ".", "The", "pronunciations", "are", "obtained", "from", "large", "base", "dictionaries", ",", "with", "a", "fall", "-", "back", "to", "a", "statistical", "grapheme", "-", "to", "-", "phoneme", "(", "G2P", ")", "system", "to", "generate", "the", "missing", "pronunciations", ".", "subsubsection", ":", "Dynamic", "Language", "Model", "The", "standard", "way", "to", "compute", "the", "decoding", "graph", "is", "to", "perform", "compositions", "from", "right", "to", "left", "with", "the", "following", "formula", "where", "each", "composition", "is", "followed", "by", "a", "determinization", "and", "a", "minimization", ".", "The", "order", "in", "which", "the", "compositions", "are", "done", "is", "important", ",", "as", "the", "composition", "is", "known", "to", "be", "intractable", "when", "is", "not", "deterministic", ",", "as", "is", "the", "case", "when", "G", "is", "a", "wFST", "representing", "an", "n", "-", "gram", "model", ".", "We", "will", "refer", "to", "the", "result", "of", "equation", "[", "reference", "]", "as", "a", "static", "model", "in", "the", "following", ".", "In", "the", "context", "of", "embedded", "inference", ",", "a", "major", "drawback", "of", "this", "standard", "method", "is", "the", "necessity", "to", "compute", "and", "load", "the", "static", "HCLG", "decoding", "graph", "in", "memory", "in", "order", "to", "perform", "speech", "recognition", ".", "The", "size", "of", "this", "decoding", "graph", "can", "claim", "a", "large", "chunk", "of", "the", "GB", "of", "RAM", "available", "on", "a", "Raspberry", "Pi", "3", ",", "or", "even", "be", "too", "big", "for", "smaller", "devices", ".", "Additionally", ",", "since", "LMs", "are", "trained", "synchronously", "in", "the", "Snips", "web", "console", "after", "the", "user", "has", "created", "their", "dataset", "(", "see", "section", "[", "reference", "]", ")", ",", "it", "is", "important", "for", "the", "decoding", "graph", "to", "be", "generated", "as", "fast", "as", "possible", ".", "For", "these", "reasons", ",", "a", "dynamic", "language", "model", ",", "composing", "the", "various", "transducers", "upon", "request", "instead", "of", "ahead", "of", "time", ",", "is", "employed", ".", "This", "is", "achieved", "by", "replacing", "the", "compositions", "of", "equation", "(", "[", "reference", "]", ")", "by", "delayed", "(", "or", "lazy", ")", "ones", ".", "Consequently", ",", "the", "states", "and", "transitions", "of", "the", "complete", "HCLG", "decoding", "graph", "are", "not", "computed", "at", "the", "time", "of", "creation", ",", "but", "rather", "at", "runtime", "during", "the", "inference", ",", "notably", "speeding", "up", "the", "building", "of", "the", "decoding", "graph", ".", "Additionally", ",", "employing", "lazy", "composition", "allows", "to", "break", "the", "decoding", "graph", "into", "two", "pieces", "(", "HCL", "on", "one", "hand", ",", "and", "G", "on", "the", "other", ")", ".", "The", "sum", "of", "the", "sizes", "of", "these", "pieces", "is", "typically", "several", "times", "smaller", "than", "the", "equivalent", ",", "statically", "-", "composed", "HCLG", ".", "In", "order", "to", "preserve", "the", "decoding", "speed", "of", "the", "ASR", "engine", "using", "a", "dynamic", "language", "model", ",", "a", "better", "composition", "algorithm", "using", "lazy", "look", "-", "ahead", "operations", "must", "be", "used", ".", "Indeed", ",", "a", "naive", "lazy", "composition", "typically", "creates", "many", "non", "co", "-", "accessible", "states", "in", "the", "resulting", "wFST", ",", "wasting", "both", "time", "and", "memory", ".", "In", "the", "case", "of", "a", "static", "decoding", "graph", ",", "these", "states", "are", "removed", "through", "a", "final", "optimization", "step", "of", "the", "HCLG", "that", "can", "not", "be", "applied", "in", "the", "dynamic", "case", "because", "the", "full", "HCLG", "is", "never", "built", ".", "This", "issue", "can", "be", "addressed", "through", "the", "use", "of", "composition", "filters", ".", "In", "particular", ",", "the", "use", "of", "look", "-", "ahead", "filters", "followed", "by", "label", "-", "reachability", "filters", "with", "weights", "and", "labels", "pushing", "allows", "to", "discard", "inaccessible", "and", "costly", "decoding", "hypotheses", "early", "in", "the", "decoding", ".", "The", "lexicon", "can", "therefore", "be", "composed", "with", "the", "language", "model", "while", "simultaneously", "optimizing", "the", "resulting", "transducer", ".", "Finally", ",", "the", "Replace", "operation", "of", "equation", "(", "[", "reference", "]", ")", "is", "also", "delayed", ".", "This", "allows", "to", "further", "break", "the", "decoding", "graph", "into", "smaller", "distinct", "pieces", ":", "the", "transducer", "mapping", "the", "output", "of", "the", "acoustic", "model", "to", "words", ",", "the", "query", "language", "model", ",", "and", "the", "entities", "languages", "models", ".", "Formally", ",", "at", "runtime", ",", "the", "dynamic", "decoding", "graph", "is", "created", "using", "the", "following", "formula", "where", "the", "HCL", "transducer", "is", "computed", "beforehand", "using", "regular", "transducer", "compositions", "(", "i.e.", ")", "and", "denotes", "the", "delayed", "transducer", "composition", "with", "composition", "filters", ".", "These", "improvements", "yield", "real", "time", "decoding", "on", "a", "Raspberry", "Pi", "3", "with", "a", "small", "overhead", "compared", "to", "a", "static", "model", ",", "and", "preserve", "decoding", "accuracy", "while", "reducing", "drastically", "the", "size", "of", "the", "model", "on", "disk", ".", "Additionally", ",", "this", "greatly", "reduces", "the", "training", "time", "of", "the", "LM", ".", "Finally", ",", "breaking", "down", "the", "LM", "into", "smaller", ",", "separate", "parts", "makes", "it", "possible", "to", "efficiently", "update", "it", ".", "It", "particular", ",", "performing", "on", "-", "device", "injection", "of", "new", "values", "in", "the", "LM", "becomes", "straightforward", ",", "enabling", "users", "to", "locally", "customize", "their", "SLU", "engine", "without", "going", "through", "the", "Snips", "web", "console", ".", "This", "feature", "is", "described", "in", "the", "following", ".", "subsubsection", ":", "On", "-", "device", "personalization", "Using", "contextual", "information", "in", "ASR", "is", "a", "promising", "approach", "to", "improving", "the", "recognition", "results", "by", "biasing", "the", "language", "model", "towards", "a", "user", "-", "specific", "vocabulary", ".", "A", "straightforward", "way", "of", "customizing", "the", "LM", "previously", "described", "is", "to", "update", "the", "list", "of", "values", "each", "entity", "can", "take", ".", "For", "instance", ",", "if", "we", "consider", "an", "assistant", "dedicated", "to", "making", "phone", "calls", "(", "‘", "‘", "call", "(", "Jane", "Doe", ")[", "contact", "]", "’", "’", ")", ",", "the", "user", "’s", "list", "of", "contacts", "could", "be", "added", "to", "the", "values", "of", "the", "entity", "‘", "‘", "contact", "’’", "in", "an", "embedded", "way", ",", "without", "this", "sensitive", "data", "ever", "leaving", "the", "device", ".", "This", "operation", "is", "called", "entity", "injection", "in", "the", "following", ".", "In", "order", "to", "perform", "entity", "injection", ",", "two", "modifications", "of", "the", "decoding", "graph", "are", "necessary", ".", "First", ",", "the", "new", "words", "and", "their", "pronunciations", "are", "added", "to", "the", "transducer", ".", "Second", ",", "the", "new", "values", "are", "added", "to", "the", "corresponding", "entity", "wFST", ".", "The", "pronunciations", "of", "the", "words", "already", "supported", "by", "the", "ASR", "are", "cached", "to", "avoid", "recomputing", "them", "on", "-", "device", ".", "Pronunciations", "for", "words", "absent", "from", "the", "transducer", "are", "computed", "via", "an", "embedded", "G2P.", "The", "updated", "transducer", "can", "then", "be", "fully", "recompiled", "and", "optimized", ".", "The", "procedure", "for", "adding", "a", "new", "value", "to", "varies", "depending", "on", "whether", "a", "union", "of", "word", "acceptors", "or", "an", "n", "-", "gram", "model", "is", "used", ".", "In", "the", "former", "case", ",", "an", "acceptor", "of", "the", "new", "value", "is", "created", "and", "its", "union", "with", "is", "computed", ".", "In", "the", "latter", "case", ",", "we", "update", "the", "n", "-", "gram", "counts", "with", "the", "new", "values", "and", "recompute", "using", "an", "embedded", "n", "-", "gram", "engine", ".", "The", "time", "required", "for", "the", "complete", "entity", "injection", "procedure", "just", "described", "ranges", "from", "a", "few", "seconds", "for", "small", "assistants", ",", "to", "a", "few", "dozen", "seconds", "for", "larger", "assistants", "supporting", "a", "vocabulary", "comprising", "tens", "of", "thousands", "of", "words", ".", "Breaking", "down", "the", "decoding", "graph", "into", "smaller", ",", "computationally", "manageable", "pieces", ",", "therefore", "allows", "to", "modify", "the", "model", "directly", "on", "device", "in", "order", "to", "provide", "a", "personalized", "user", "experience", "and", "increase", "the", "overall", "accuracy", "of", "the", "SLU", "component", ".", "subsubsection", ":", "Confidence", "scoring", "An", "important", "challenge", "of", "specialized", "SLU", "systems", "trained", "on", "small", "amounts", "of", "domain", "-", "specific", "text", "data", "is", "the", "ability", "to", "detect", "out", "-", "of", "-", "vocabulary", "(", "OOV", ")", "words", ".", "Indeed", ",", "while", "a", "sufficient", "amount", "of", "specific", "training", "data", "may", "guarantee", "sampling", "the", "important", "words", "which", "allow", "to", "discriminate", "between", "different", "intents", ",", "it", "will", "in", "general", "prove", "unable", "to", "correctly", "sample", "filler", "words", "from", "general", "spoken", "language", ".", "As", "a", "consequence", ",", "a", "specialized", "ASR", "such", "as", "the", "one", "described", "in", "the", "previous", "sections", "will", "tend", "to", "approximate", "unknown", "words", "using", "phonetically", "related", "ones", "from", "its", "vocabulary", ",", "potentially", "harming", "the", "subsequent", "NLU", ".", "One", "way", "of", "addressing", "this", "issue", "is", "to", "extract", "a", "word", "-", "level", "confidence", "score", "from", "the", "ASR", ",", "assigning", "a", "probability", "for", "the", "word", "to", "be", "correctly", "decoded", ".", "Confidence", "scoring", "is", "a", "notoriously", "hard", "problem", "in", "speech", "recognition", ".", "Our", "approach", "is", "based", "on", "the", "so", "-", "called", "‘", "‘", "confusion", "network", "’", "’", "representation", "of", "the", "hypotheses", "of", "the", "ASR", "(", "see", "Figure", "[", "reference", "]", ")", ".", "A", "confusion", "network", "is", "a", "graph", "encoding", ",", "for", "each", "speech", "segment", "in", "an", "utterance", ",", "the", "competing", "decoding", "hypotheses", "along", "with", "their", "posterior", "probability", ",", "thus", "providing", "a", "richer", "output", "than", "the", "-", "best", "decoding", "hypothesis", ".", "In", "particular", ",", "confusion", "networks", "in", "conjunction", "with", "NLU", "systems", "typically", "improve", "end", "-", "to", "-", "end", "performance", "in", "speech", "-", "to", "-", "meaning", "tasks", ".", "In", "the", "following", ",", "we", "restrict", "our", "use", "of", "confusion", "networks", "to", "a", "greedy", "decoder", "that", "outputs", ",", "for", "each", "speech", "segment", ",", "the", "most", "probable", "decoded", "word", "along", "with", "its", "probability", ".", "In", "this", "context", ",", "our", "strategy", "for", "identifying", "OOVs", "is", "to", "set", "a", "threshold", "on", "this", "word", "-", "level", "probability", ".", "Below", "this", "threshold", ",", "the", "word", "is", "declared", "misunderstood", ".", "In", "practice", ",", "a", "post", "-", "processing", "step", "is", "used", "to", "remove", "these", "words", "from", "the", "decoded", "sentence", ",", "replacing", "them", "with", "a", "special", "OOV", "symbol", ".", "This", "allows", "the", "SLU", "pipeline", "to", "proceed", "with", "the", "words", "the", "ASR", "has", "understood", "with", "sufficient", "probability", ",", "leaving", "out", "the", "filler", "words", "which", "are", "unimportant", "to", "extract", "the", "intent", "and", "the", "slots", "from", "the", "query", ",", "thus", "preserving", "the", "generalization", "properties", "of", "the", "SLU", "in", "the", "presence", "of", "unknown", "filler", "words", "(", "see", "section", "[", "reference", "]", "for", "a", "quantitative", "evaluation", ")", ".", "Finally", ",", "we", "may", "define", "a", "sentence", "-", "level", "confidence", "by", "simply", "taking", "the", "geometric", "mean", "of", "the", "word", "-", "level", "confidence", "scores", ".", "subsection", ":", "Natural", "Language", "Understanding", "The", "Natural", "Language", "Understanding", "component", "of", "the", "Snips", "Voice", "Platform", "extracts", "structured", "data", "from", "queries", "written", "in", "natural", "language", ".", "Snips", "NLU", "–", "a", "Python", "library", "–", "can", "be", "used", "for", "training", "and", "inference", ",", "with", "a", "Rust", "implementation", "focusing", "solely", "on", "inference", ".", "Both", "have", "been", "recently", "open", "-", "sourced", ".", "Three", "tasks", "are", "successively", "performed", ".", "Intent", "Classification", "consists", "in", "extracting", "the", "intent", "expressed", "in", "the", "query", "(", "e.g.", "SetTemperature", "or", "SwitchLightOn", ")", ".", "Once", "the", "intent", "is", "known", ",", "Slot", "Filling", "aims", "to", "extract", "the", "slots", ",", "i.e.", "the", "values", "of", "the", "entities", "present", "in", "the", "query", ".", "Finally", ",", "Entity", "Resolution", "focuses", "on", "built", "-", "in", "entities", ",", "such", "as", "date", "and", "times", ",", "durations", ",", "temperatures", ",", "for", "which", "Snips", "provides", "an", "extra", "resolution", "step", ".", "It", "basically", "transforms", "entity", "values", "such", "as", "\"", "tomorrow", "evening", "\"", "into", "formatted", "values", "such", "as", "\"", "2018", "-", "04", "-", "19", "19:00:00", "+", "00:00", "\"", ".", "Snippet", "[", "reference", "]", "illustrates", "a", "typical", "output", "of", "the", "NLU", "component", ".", "subsubsection", ":", "Models", "The", "Snips", "NLU", "pipeline", "(", "Figure", "[", "reference", "]", ")", "contains", "a", "main", "component", ",", "the", "NLU", "Engine", ",", "which", "itself", "is", "composed", "of", "several", "components", ".", "A", "first", "component", "is", "the", "Intent", "Parser", ",", "which", "performs", "both", "intent", "classification", "and", "slot", "filling", ".", "It", "does", "not", "resolve", "entity", "values", ".", "The", "NLU", "Engine", "calls", "two", "intent", "parsers", "successively", ":", "a", "deterministic", "intent", "parser", "a", "probabilistic", "intent", "parser", "The", "second", "one", "is", "called", "only", "when", "nothing", "is", "extracted", "by", "the", "first", "one", ".", "paragraph", ":", "Deterministic", "Intent", "Parser", ".", "The", "goal", "of", "the", "deterministic", "intent", "parser", "is", "to", "provide", "robustness", "and", "a", "predictable", "experience", "for", "the", "user", "as", "it", "is", "guaranteed", "to", "achieve", "a", "F1", "-", "score", "on", "the", "training", "examples", ".", "Its", "implementation", "relies", "on", "regular", "expressions", ".", "The", "queries", "contained", "in", "the", "training", "data", "are", "used", "to", "build", "patterns", "covering", "all", "combinations", "of", "entity", "values", ".", "Let", "us", "consider", ",", "for", "instance", ",", "the", "training", "sample", ":", "set", "the", "[", "kitchen", "](", "room", ")", "lights", "to", "[", "blue", "](", "color", ")", "Let", "us", "assume", "that", "the", "set", "of", "possible", "values", "for", "the", "room", "entity", "are", "kitchen", ",", "hall", ",", "bedroom", "and", "those", "for", "the", "color", "entity", "are", "blue", ",", "yellow", ",", "red", ".", "A", "representation", "of", "the", "generated", "pattern", "for", "this", "sample", "is", ":", "set", "the", "(", "?", "P", "<", "room", ">", "kitchen|hall|bedroom", ")", "lights", "to", "(", "?", "P", "<", "color", ">", "blue|yellow|red", ")", "paragraph", ":", "Probabilistic", "Intent", "Parser", ".", "The", "probabilistic", "intent", "parser", "aims", "at", "extending", "parsing", "beyond", "training", "examples", "and", "recognizing", "variations", "which", "do", "not", "appear", "in", "the", "training", "data", ".", "It", "provides", "the", "generalization", "power", "that", "the", "deterministic", "parser", "lacks", ".", "This", "parser", "runs", "in", "two", "cascaded", "steps", ":", "intent", "classification", "and", "slot", "filling", ".", "The", "intent", "classification", "is", "implemented", "with", "a", "logistic", "regression", "trained", "on", "the", "queries", "from", "every", "intent", ".", "The", "slot", "-", "filling", "step", "consists", "in", "several", "linear", "-", "chain", "Conditional", "Random", "Fields", "(", "CRFs", ")", ",", "each", "of", "them", "being", "trained", "for", "a", "specific", "intent", ".", "Once", "the", "intent", "is", "extracted", "by", "the", "intent", "classifier", ",", "the", "corresponding", "slot", "filler", "is", "used", "to", "extract", "slots", "from", "the", "query", ".", "The", "choice", "of", "CRFs", "for", "the", "slot", "-", "filling", "step", "results", "from", "careful", "considerations", "and", "experiments", ".", "They", "are", "indeed", "a", "standard", "approach", "for", "this", "task", ",", "and", "are", "known", "to", "have", "low", "generalization", "error", ".", "Recently", ",", "more", "computationally", "demanding", "approaches", "based", "on", "deep", "learning", "models", "have", "been", "proposed", ".", "Our", "experiments", "however", "showed", "that", "these", "approaches", "do", "not", "yield", "any", "significant", "gain", "in", "accuracy", "in", "the", "typical", "training", "size", "regime", "of", "custom", "voice", "assistants", "(", "a", "few", "hundred", "queries", ")", ".", "The", "lightest", "option", "was", "therefore", "favored", ".", "On", "top", "of", "the", "classical", "features", "used", "in", "slot", "-", "filling", "tasks", "such", "as", "n", "-", "grams", ",", "case", ",", "shape", ",", "etc", ".", ",", "additional", "features", "are", "crafted", ".", "In", "this", "kind", "of", "task", ",", "it", "appears", "that", "leveraging", "external", "knowledge", "is", "crucial", ".", "Hence", ",", "we", "apply", "a", "built", "-", "in", "entity", "extractor", "(", "see", "next", "paragraph", "about", "Entity", "Resolution", ")", "to", "build", "features", "that", "indicate", "whether", "or", "not", "a", "token", "in", "the", "sentence", "is", "part", "of", "a", "built", "-", "in", "entity", ".", "The", "value", "of", "the", "feature", "is", "the", "corresponding", "entity", ",", "if", "one", "is", "found", ",", "augmented", "with", "a", "BILOU", "coding", "scheme", ",", "indicating", "the", "position", "of", "the", "token", "in", "the", "matching", "entity", "value", ".", "We", "find", "empirically", "that", "the", "presence", "of", "such", "features", "improves", "the", "overall", "accuracy", ",", "thanks", "to", "the", "robustness", "of", "the", "built", "-", "in", "entities", "extractor", ".", "The", "problem", "of", "data", "sparsity", "is", "addressed", "by", "integrating", "features", "based", "on", "word", "clusters", ".", "More", "specifically", ",", "we", "use", "Brown", "clusters", "released", "by", "the", "authors", "of", ",", "as", "well", "as", "word", "clusters", "built", "from", "word2vec", "embeddings", "using", "k", "-", "means", "clustering", ".", "We", "find", "that", "the", "use", "of", "these", "features", "helps", "in", "reducing", "generalization", "error", ",", "by", "bringing", "the", "effective", "size", "of", "the", "vocabulary", "from", "typically", "50", "K", "words", "down", "to", "a", "few", "hundred", "word", "clusters", ".", "Finally", ",", "gazetteer", "features", "are", "built", ",", "based", "on", "entity", "values", "provided", "in", "the", "training", "data", ".", "One", "gazetteer", "is", "created", "per", "entity", "type", ",", "and", "used", "to", "match", "tokens", "via", "a", "BILOU", "coding", "scheme", ".", "Table", "[", "reference", "]", "displays", "some", "examples", "of", "features", "used", "in", "Snips", "NLU", ".", "Overfitting", "is", "avoided", "by", "dropping", "a", "fraction", "of", "the", "features", "during", "training", ".", "More", "precisely", ",", "each", "feature", "is", "assigned", "a", "dropout", "probability", ".", "For", "each", "training", "example", ",", "we", "compute", "the", "features", "and", "then", "erase", "feature", "with", "probability", ".", "Without", "this", "mechanism", ",", "we", "typically", "observed", "that", "the", "CRF", "learns", "to", "tag", "every", "value", "matching", "the", "entity", "gazetteer", "while", "discarding", "all", "those", "absent", "from", "it", ".", "paragraph", ":", "Entity", "resolution", ".", "The", "last", "step", "of", "the", "NLU", "pipeline", "consists", "in", "resolving", "slot", "values", "(", "e.g.", "from", "raw", "strings", "to", "ISO", "formatted", "values", "for", "date", "and", "time", "entities", ")", ".", "Entity", "values", "that", "can", "be", "resolved", "(", "e.g.", "dates", ",", "temperatures", ",", "numbers", ")", "correspond", "to", "the", "built", "-", "in", "entities", "introduced", "in", "section", "[", "reference", "]", ",", "and", "are", "supported", "natively", "without", "requiring", "training", "examples", ".", "The", "resolution", "is", "done", "with", "Rustling", ",", "an", "in", "-", "house", "re", "-", "implementation", "of", "Facebook", "’s", "Duckling", "library", "in", "Rust", ",", "which", "we", "also", "open", "sourced", ",", "with", "modifications", "to", "make", "its", "runtime", "more", "stable", "with", "regards", "to", "the", "length", "of", "the", "sentences", "parsed", ".", "subsubsection", ":", "Evaluation", "Snips", "NLU", "is", "evaluated", "and", "compared", "to", "various", "NLU", "services", "on", "two", "datasets", ":", "a", "previously", "published", "comparison", ",", "and", "an", "in", "-", "house", "open", "dataset", ".", "The", "latter", "has", "been", "made", "freely", "accessible", "on", "GitHub", "to", "promote", "transparency", "and", "reproducibility", ".", "paragraph", ":", "Evaluation", "on", "Braun", "et", "al", ".", ",", "2017", ".", "In", "January", "2018", ",", "we", "evaluated", "Snips", "NLU", "on", "a", "previously", "published", "comparison", "between", "various", "NLU", "services", ":", "a", "few", "of", "the", "main", "cloud", "-", "based", "solutions", "(", "Microsoft", "’s", "Luis", ",", "IBM", "Watson", ",", "API.AI", "now", "Google", "’s", "Dialogflow", ")", ",", "and", "the", "open", "-", "source", "platform", "Rasa", "NLU", ".", "For", "the", "raw", "results", "and", "methodology", ",", "see", ".", "The", "main", "metric", "used", "in", "this", "benchmark", "is", "the", "average", "F1", "-", "score", "of", "intent", "classification", "and", "slot", "filling", ".", "The", "data", "consists", "in", "three", "corpora", ".", "Two", "of", "the", "corpora", "were", "extracted", "from", "StackExchange", ",", "one", "from", "a", "Telegram", "chatbot", ".", "The", "exact", "same", "splits", "as", "in", "the", "original", "paper", "were", "used", "for", "the", "Ubuntu", "and", "Web", "Applications", "corpora", ".", "At", "the", "date", "we", "ran", "the", "evaluation", ",", "the", "train", "and", "test", "splits", "were", "not", "explicit", "for", "the", "Chatbot", "dataset", "(", "although", "they", "were", "added", "later", "on", ")", ".", "In", "that", "case", ",", "we", "ran", "a", "5", "-", "fold", "cross", "-", "validation", ".", "The", "results", "are", "presented", "in", "Table", "[", "reference", "]", ".", "Figure", "[", "reference", "]", "presents", "the", "average", "results", "on", "the", "three", "corpora", ",", "corresponding", "to", "the", "overall", "section", "of", "Table", "[", "reference", "]", ".", "For", "Rasa", ",", "we", "considered", "all", "three", "possible", "backends", "(", "Spacy", ",", "SKLearn", "+", "MITIE", ",", "MITIE", ")", ",", "see", "the", "abovementioned", "GitHub", "repository", "for", "more", "details", ".", "However", ",", "only", "Spacy", "was", "run", "on", "all", "3", "datasets", ",", "for", "train", "time", "reasons", ".", "For", "fairness", ",", "the", "latest", "version", "of", "Rasa", "NLU", "is", "also", "displayed", ".", "Results", "show", "that", "Snips", "NLU", "ranks", "second", "highest", "overall", ".", "paragraph", ":", "Evaluation", "on", "an", "in", "-", "house", "open", "dataset", ".", "In", "June", "2017", ",", "Snips", "NLU", "was", "evaluated", "on", "an", "in", "-", "house", "dataset", "of", "over", "16", "K", "crowdsourced", "queries", "(", "freely", "available", ")", "distributed", "among", "7", "user", "intents", "of", "various", "complexity", ":", "SearchCreativeWork", "(", "e.g.", "Find", "me", "the", "I", ",", "Robot", "television", "show", ")", ",", "GetWeather", "(", "e.g.", "Is", "it", "windy", "in", "Boston", ",", "MA", "right", "now", "?", ")", ",", "BookRestaurant", "(", "e.g.", "I", "want", "to", "book", "a", "highly", "rated", "restaurant", "in", "Paris", "tomorrow", "night", ")", ",", "PlayMusic", "(", "e.g.", "Play", "the", "last", "track", "from", "Beyoncé", "off", "Spotify", ")", ",", "AddToPlaylist", "(", "e.g.", "Add", "Diamonds", "to", "my", "roadtrip", "playlist", ")", "RateBook", "(", "e.g.", "Give", "6", "stars", "to", "Of", "Mice", "and", "Men", ")", "SearchScreeningEvent", "(", "e.g.", "Check", "the", "showtimes", "for", "Wonder", "Woman", "in", "Paris", ")", "The", "full", "ontology", "is", "available", "on", "Table", "[", "reference", "]", "in", "Appendix", ".", "In", "this", "experiment", ",", "the", "comparison", "is", "done", "separately", "on", "each", "intent", "to", "focus", "on", "slot", "filling", "(", "rather", "than", "intent", "classification", ")", ".", "The", "main", "metric", "used", "in", "this", "benchmark", "is", "the", "average", "F1", "-", "score", "of", "slot", "filling", "on", "all", "slots", ".", "Three", "training", "sets", "of", "70", "and", "2000", "queries", "have", "been", "drawn", "from", "the", "total", "pool", "of", "queries", "to", "gain", "in", "statistical", "relevance", ".", "Validation", "sets", "consist", "in", "100", "queries", "per", "intent", ".", "Five", "different", "cloud", "-", "based", "providers", "are", "compared", "to", "Snips", "NLU", "(", "Microsoft", "’s", "Luis", ",", "API.AI", "now", "Google", "’s", "Dialogflow", ",", "Facebook", "’s", "Wit.ai", ",", "and", "Amazon", "Alexa", ")", ".", "For", "more", "details", "about", "the", "specific", "methodology", "for", "each", "provider", "and", "access", "to", "the", "full", "dataset", ",", "see", ".", "Each", "solution", "is", "trained", "and", "evaluated", "on", "the", "exact", "same", "datasets", ".", "Table", "[", "reference", "]", "shows", "the", "precision", ",", "recall", "and", "F1", "-", "score", "averaged", "on", "all", "slots", "and", "on", "all", "intents", ".", "Results", "specific", "to", "each", "intent", "are", "available", "in", "Tables", "[", "reference", "]", "&", "[", "reference", "]", "in", "Appendix", ".", "Snips", "NLU", "is", "as", "accurate", "or", "better", "than", "competing", "cloud", "-", "based", "solutions", "in", "slot", "filling", ",", "regardless", "of", "the", "training", "set", "size", ".", "subsubsection", ":", "Embedded", "performance", "Using", "Rust", "for", "the", "NLU", "inference", "pipeline", "allows", "to", "keep", "the", "memory", "footprint", "and", "the", "inference", "runtime", "very", "low", ".", "Memory", "usage", "has", "been", "optimized", ",", "with", "model", "sizes", "ranging", "from", "a", "few", "hundred", "kilobytes", "of", "RAM", "for", "common", "cases", "to", "a", "few", "megabytes", "for", "the", "most", "complex", "assistants", ".", "They", "are", "therefore", "fit", "for", "deployment", "on", "a", "Raspberry", "Pi", "or", "a", "mobile", "app", ",", "and", "more", "powerful", "servers", "can", "handle", "hundreds", "of", "parallel", "instances", ".", "Using", "the", "embedded", "Snips", "Voice", "platform", "significantly", "reduces", "the", "inference", "runtime", "compared", "to", "a", "roundtrip", "to", "a", "cloud", "service", ",", "as", "displayed", "on", "Table", "[", "reference", "]", ".", "section", ":", "End", "-", "to", "-", "end", "Evaluation", "In", "this", "section", ",", "we", "evaluate", "the", "performance", "of", "the", "Snips", "Spoken", "Language", "Understanding", "(", "SLU", ")", "pipeline", "in", "an", "end", "-", "to", "-", "end", ",", "speech", "-", "to", "-", "meaning", "setting", ".", "To", "this", "end", ",", "we", "consider", "two", "real", "-", "world", "assistants", "of", "different", "sizes", ",", "namely", "SmartLights", "and", "Weather", ".", "The", "SmartLights", "assistant", "specializes", "in", "interacting", "with", "light", "devices", "supporting", "different", "colors", "and", "levels", "of", "brightness", ",", "and", "positioned", "in", "various", "rooms", ".", "The", "Weather", "assistant", "is", "targeted", "at", "weather", "queries", "in", "general", ",", "and", "supports", "various", "types", "of", "formulations", "and", "places", ".", "Tables", "[", "reference", "]", "and", "[", "reference", "]", "sum", "up", "the", "constitution", "and", "size", "of", "the", "datasets", "corresponding", "to", "these", "two", "assistants", ",", "while", "tables", "[", "reference", "]", "and", "[", "reference", "]", "describe", "their", "entities", ".", "Note", "in", "particular", "the", "use", "of", "the", "built", "-", "in", "‘", "‘", "snips", "/", "number", "’’", "(", "respectively", "‘", "‘", "snips", "/", "datetime", "’’", ")", "entity", "to", "define", "the", "brightness", "(", "resp", ".", "datetime", ")", "slot", ",", "which", "allows", "the", "assistant", "to", "generalize", "to", "values", "absent", "from", "the", "dataset", ".", "We", "are", "interested", "in", "computing", "end", "-", "to", "-", "end", "metrics", "quantifying", "the", "ability", "of", "the", "assistants", "to", "extract", "intent", "and", "slots", "from", "spoken", "utterances", ".", "We", "create", "a", "test", "set", "by", "crowdsourcing", "a", "spoken", "corpus", "corresponding", "to", "the", "queries", "of", "each", "dataset", ".", "For", "each", "sentence", "of", "the", "speech", "corpus", ",", "we", "apply", "the", "ASR", "engine", "followed", "by", "the", "NLU", "engine", ",", "and", "compare", "the", "predicted", "output", "to", "the", "ground", "true", "intent", "and", "slots", "in", "the", "dataset", ".", "In", "the", "following", ",", "we", "present", "our", "results", "in", "terms", "of", "the", "classical", "precision", ",", "recall", ",", "and", "F1", "scores", ".", "subsection", ":", "Language", "Model", "Generalization", "Error", "To", "be", "able", "to", "understand", "arbitrary", "formulations", "of", "an", "intent", ",", "the", "SLU", "engine", "must", "be", "able", "to", "generalize", "to", "unseen", "queries", "in", "the", "same", "domain", ".", "To", "test", "the", "generalization", "ability", "of", "the", "Snips", "SLU", "components", ",", "we", "use", "5", "-", "fold", "cross", "-", "validation", ",", "and", "successively", "train", "the", "LM", "and", "NLU", "on", "four", "fifth", "of", "the", "dataset", ",", "testing", "on", "the", "last", ",", "unseen", ",", "fifth", "of", "the", "data", ".", "The", "training", "procedure", "is", "identical", "to", "the", "one", "detailed", "in", "section", "[", "reference", "]", ".", "We", "note", "that", "all", "the", "values", "of", "the", "entities", "are", "always", "included", "in", "the", "training", "set", ".", "Tables", "[", "reference", "]", "and", "[", "reference", "]", "sum", "up", "the", "results", "of", "this", "experiment", ",", "highlighting", "in", "particular", "the", "modest", "effect", "of", "the", "introduction", "of", "the", "ASR", "engine", "compared", "to", "the", "accuracy", "of", "the", "NLU", "evaluated", "directly", "on", "the", "ground", "true", "query", ".", "Because", "unseen", "test", "queries", "may", "contain", "out", "of", "vocabulary", "words", "absent", "from", "the", "training", "splits", ",", "the", "ability", "of", "the", "SLU", "to", "generalize", "in", "this", "setting", "relies", "heavily", "on", "the", "identification", "of", "unknown", "words", "through", "the", "strategy", "detailed", "in", "section", "[", "reference", "]", ".", "As", "noted", "earlier", "and", "confirmed", "by", "these", "results", ",", "this", "confidence", "scoring", "strategy", "also", "allows", "to", "favor", "precision", "over", "recall", "by", "rejecting", "uncertain", "words", "that", "may", "be", "misinterpreted", "by", "the", "NLU", ".", "Figure", "[", "reference", "]", "illustrates", "the", "correlation", "between", "the", "sentence", "-", "level", "confidence", "score", "defined", "in", "section", "[", "reference", "]", "and", "the", "word", "error", "rate", ".", "While", "noisy", ",", "the", "confidence", "allows", "to", "detect", "misunderstood", "queries", ",", "and", "can", "be", "mixed", "with", "the", "intent", "classification", "probability", "output", "by", "the", "NLU", "to", "reject", "dubious", "queries", ",", "thus", "promoting", "precision", "over", "recall", ".", "subsection", ":", "Embedded", "Performance", "The", "embedded", "SLU", "components", "corresponding", "to", "the", "assistants", "described", "in", "the", "previous", "section", "are", "trained", "in", "under", "thirty", "seconds", "through", "the", "Snips", "web", "console", "(", "see", "section", "[", "reference", "]", ")", ".", "The", "resulting", "language", "models", "have", "a", "size", "of", "the", "order", "of", "the", "megabyte", "for", "the", "SmartLights", "assistant", "(", "MB", "in", "total", ",", "with", "the", "acoustic", "model", ")", ",", "and", "MB", "for", "the", "Weather", "assistant", "(", "MB", "in", "total", ")", ".", "The", "SLU", "components", "run", "faster", "than", "real", "time", "on", "a", "single", "core", "on", "a", "Raspberry", "Pi", "3", ",", "as", "well", "as", "on", "the", "smaller", "NXP", "imx7D.", "section", ":", "Training", "models", "without", "user", "data", "The", "private", "-", "by", "-", "design", "approach", "described", "in", "the", "previous", "sections", "requires", "to", "train", "high", "-", "performance", "machine", "learning", "models", "without", "access", "to", "users", "queries", ".", "This", "problem", "is", "especially", "critical", "for", "the", "specialized", "language", "modeling", "components", "–", "Language", "Model", "and", "Natural", "Language", "Understanding", "engine", "–", "as", "both", "need", "to", "be", "trained", "on", "an", "assistant", "-", "specific", "dataset", ".", "A", "solution", "is", "to", "develop", "a", "data", "generation", "pipeline", ".", "Once", "the", "scope", "of", "an", "assistant", "has", "been", "defined", ",", "a", "mix", "of", "crowdsourcing", "and", "semi", "-", "supervised", "machine", "learning", "is", "used", "to", "generate", "thousands", "of", "high", "-", "quality", "training", "examples", ",", "mimicking", "user", "data", "collection", "without", "compromising", "on", "privacy", ".", "The", "aim", "of", "this", "section", "is", "to", "describe", "the", "data", "generation", "pipeline", ",", "and", "to", "demonstrate", "its", "impact", "on", "the", "performance", "of", "the", "NLU", ".", "subsection", ":", "Data", "generation", "pipeline", "A", "first", "simple", "approach", "to", "data", "generation", "is", "grammar", "-", "based", "generation", ",", "which", "consists", "in", "breaking", "down", "a", "written", "query", "into", "consecutive", "semantic", "blocks", "and", "requires", "enumerating", "every", "possible", "pattern", "in", "the", "chosen", "language", ".", "While", "this", "method", "guarantees", "an", "exact", "slot", "and", "intent", "supervision", ",", "queries", "generated", "in", "this", "way", "are", "highly", "correlated", ":", "their", "diversity", "is", "limited", "to", "the", "expressive", "power", "of", "the", "used", "grammar", "and", "the", "imagination", "of", "the", "person", "having", "created", "it", ".", "Moreover", ",", "the", "pattern", "definition", "and", "enumeration", "can", "be", "very", "time", "consuming", "and", "requires", "an", "extensive", "knowledge", "of", "the", "given", "language", ".", "This", "approach", "is", "therefore", "unfit", "for", "generating", "queries", "in", "natural", "language", ".", "On", "the", "other", "hand", ",", "crowdsourcing", "–", "widely", "used", "in", "Natural", "Language", "Processing", "research", "–", "ensures", "diversity", "in", "formulation", "by", "sampling", "queries", "from", "a", "large", "number", "of", "demographically", "diverse", "contributors", ".", "However", ",", "the", "accuracy", "of", "intent", "and", "slot", "supervision", "decreases", "as", "soon", "as", "humans", "are", "in", "the", "loop", ".", "Any", "mislabeling", "of", "a", "query", "’s", "intent", "or", "slots", "has", "a", "strong", "impact", "on", "the", "end", "-", "to", "-", "end", "performance", "of", "the", "SLU", ".", "To", "guarantee", "a", "fast", "and", "accurate", "generation", "of", "training", "data", "for", "the", "language", "modeling", "components", ",", "we", "complement", "crowdsourcing", "with", "machine", "-", "learning", "-", "based", "disambiguation", "techniques", ".", "We", "new", "detail", "the", "implementation", "of", "the", "data", "generation", "pipeline", ".", "subsubsection", ":", "Crowdsourcing", "Crowdsourcing", "tasks", "were", "originally", "submitted", "to", "Amazon", "Mechanical", "Turk", ",", "a", "widely", "used", "platform", "in", "non", "-", "expert", "annotations", "for", "natural", "language", "tasks", ".", "While", "a", "sufficient", "number", "of", "English", "-", "speaking", "contributors", "can", "be", "reached", "easily", ",", "other", "languages", "such", "as", "French", ",", "German", "or", "Japanese", "suffer", "from", "a", "comparatively", "smaller", "available", "crowd", ".", "Local", "crowdsourcing", "platforms", "therefore", "had", "to", "be", "integrated", ".", "A", "text", "query", "generation", "task", "consists", "in", "generating", "an", "example", "of", "user", "query", "matching", "a", "provided", "set", "of", "intent", "and", "slots", "–", "e.g.", "the", "following", "set", ":", "Intent", ":", "The", "user", "wants", "to", "switch", "the", "lights", "on", ";", "slot", ":", "(", "bedroom", ")[", "room", "]", "could", "result", "in", "the", "generated", "query", "‘", "‘", "I", "want", "lights", "in", "the", "bedroom", "right", "now", "!", "’’", ".", "Fixing", "entity", "values", "reduces", "the", "task", "to", "a", "sentence", "generation", "and", "removes", "the", "need", "for", "a", "slot", "labeling", "step", ",", "limiting", "the", "sources", "of", "error", ".", "Diversity", "is", "enforced", "by", "both", "submitting", "this", "task", "to", "the", "widest", "possible", "crowd", "while", "limiting", "the", "number", "of", "available", "tasks", "per", "contributor", "and", "by", "selecting", "large", "sets", "of", "slot", "values", ".", "Each", "generated", "query", "goes", "through", "a", "validation", "process", "taking", "the", "form", "of", "a", "second", "crowdsourcing", "task", ",", "where", "at", "least", "two", "out", "of", "three", "new", "contributors", "must", "confirm", "its", "formulation", ",", "spelling", ",", "and", "intent", ".", "Majority", "voting", "is", "indeed", "a", "simple", "and", "straightforward", "approach", "for", "quality", "assessment", "in", "crowdsourcing", ".", "A", "custom", "dashboard", "hosted", "on", "our", "servers", "has", "been", "developed", "to", "optimize", "the", "contributor", "’s", "workflow", ",", "with", "clear", "descriptions", "of", "the", "task", ".", "The", "dashboard", "also", "prevents", "a", "contributor", "from", "submitting", "a", "query", "that", "does", "not", "contain", "the", "imposed", "entity", "values", ",", "with", "a", "fuzzy", "matching", "rule", "allowing", "for", "inflections", "in", "all", "supported", "languages", "(", "conjugation", ",", "plural", ",", "gender", ",", "compounds", ",", "etc", ".", ")", ".", "subsubsection", ":", "Disambiguation", "While", "the", "previous", "validation", "step", "allows", "to", "filter", "out", "most", "spelling", "and", "formulation", "mistakes", ",", "it", "does", "not", "always", "guarantee", "the", "correctness", "of", "the", "intent", "or", "the", "absence", "of", "spurious", "entities", ".", "Indeed", ",", "in", "a", "first", "type", "of", "errors", ",", "the", "intent", "of", "the", "query", "may", "not", "match", "the", "provided", "one", "–", "e.g.", "Switch", "off", "the", "lights", "when", "the", "intent", "SwitchLightOn", "was", "required", ".", "In", "a", "second", "type", "of", "errors", ",", "spurious", "entities", "may", "be", "added", "by", "the", "contributor", ",", "so", "that", "they", "are", "not", "labeled", "as", "such", "–", "e.g.", "‘", "‘", "I", "want", "lights", "in", "the", "guest", "[", "bedroom", "](", "room", ")", "at", "60", "right", "now", "!", "’’", "when", "only", "[", "bedroom", "](", "room", ")", "was", "mentioned", "in", "the", "task", "specifications", ".", "An", "unlabeled", "entity", "has", "a", "particularly", "strong", "impact", "on", "the", "CRF", "features", "in", "the", "NLU", "component", ",", "and", "limits", "the", "ability", "of", "the", "LM", "to", "generalize", ".", "These", "errors", "in", "the", "training", "queries", "must", "be", "fixed", "to", "achieve", "a", "high", "accuracy", "of", "the", "SLU", ".", "To", "do", "so", ",", "we", "perform", "a", "3", "-", "fold", "cross", "validation", "of", "the", "NLU", "engine", "on", "this", "dataset", ".", "This", "yields", "predicted", "intents", "and", "slots", "for", "each", "sentence", "in", "the", "dataset", ".", "By", "repeating", "this", "procedure", "several", "times", ",", "we", "obtain", "several", "predictions", "for", "each", "sentence", ".", "We", "then", "apply", "majority", "voting", "on", "these", "predictions", "to", "detect", "missing", "slots", "and", "wrong", "intents", ".", "Slots", "may", "therefore", "be", "extended", "–", "e.g.", "(", "bedroom", ")[", "room", "]", "(", "guest", "bedroom", ")[", "room", "]", "in", "the", "previous", "example", "–", "or", "added", "–", "(", "60", ")[", "intensity", "]", "–", "and", "ill", "-", "formed", "queries", "(", "with", "regard", "to", "spelling", "or", "intent", ")", "are", "filtered", "-", "out", ".", "subsection", ":", "Evaluation", "We", "illustrate", "the", "impact", "of", "data", "generation", "on", "the", "SLU", "performance", "on", "the", "specific", "case", "of", "the", "slot", "-", "filling", "task", "in", "the", "NLU", "component", ".", "The", "same", "in", "-", "house", "open", "dataset", "of", "over", "16", "K", "crowdsourced", "query", "presented", "in", "Section", "[", "reference", "]", "is", "used", ".", "Unsurprisingly", ",", "slot", "-", "filling", "performance", "drastically", "increases", "with", "the", "number", "of", "training", "samples", ".", "The", "F1", "-", "scores", "averaged", "over", "all", "slots", "are", "computed", ",", "depending", "on", "the", "number", "of", "training", "queries", "per", "intent", ".", "An", "NLU", "engine", "has", "been", "trained", "on", "each", "individual", "intent", ".", "Training", "queries", "are", "freely", "available", "on", "GitHub", ".", "The", "data", "has", "been", "generated", "with", "our", "data", "generation", "pipeline", ".", "Figure", "[", "reference", "]", "shows", "the", "influence", "of", "the", "number", "of", "training", "samples", "on", "the", "performance", "of", "the", "slot", "-", "filling", "task", "of", "the", "NLU", "component", ".", "Compared", "to", "10", "training", "queries", ",", "the", "gain", "in", "performance", "with", "500", "queries", "is", "of", "32", "%", "absolute", "on", "average", ",", "ranging", "from", "22", "%", "for", "the", "RateBook", "intent", "(", "from", "0.76", "to", "0.98", ")", "to", "44", "%", "for", "the", "GetWeather", "intent", "(", "from", "0.44", "to", "0.88", ")", ".", "This", "gain", "indeed", "strongly", "depends", "on", "the", "intent", "’s", "complexity", ",", "which", "is", "mainly", "defined", "by", "its", "entities", "(", "number", "of", "entities", ",", "built", "-", "in", "or", "custom", ",", "number", "of", "entity", "values", ",", "etc", ".", ")", ".", "While", "a", "few", "tens", "of", "training", "queries", "might", "suffice", "for", "some", "simple", "use", "cases", "(", "such", "as", "RateBook", ")", ",", "more", "complicated", "intents", "with", "larger", "sets", "of", "entity", "values", "(", "PlayMusic", "for", "instance", ")", "require", "larger", "training", "datasets", ".", "While", "it", "is", "easy", "to", "manually", "generate", "up", "to", "50", "queries", ",", "being", "able", "to", "come", "up", "with", "hundreds", "or", "thousands", "of", "diverse", "formulations", "of", "the", "same", "intent", "is", "nearly", "impossible", ".", "For", "private", "-", "by", "-", "design", "assistants", "that", "do", "not", "gather", "user", "queries", ",", "the", "ability", "to", "generate", "enough", "queries", "is", "key", "to", "training", "efficient", "machine", "learning", "models", ".", "Moreover", ",", "being", "able", "to", "generate", "training", "data", "allows", "us", "to", "validate", "the", "performance", "of", "our", "models", "before", "deploying", "them", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "have", "described", "the", "design", "of", "the", "Snips", "Voice", "Platform", ",", "a", "Spoken", "Language", "Understanding", "solution", "that", "can", "be", "embedded", "in", "small", "devices", "and", "runs", "entirely", "offline", ".", "In", "compliance", "with", "the", "privacy", "-", "by", "-", "design", "principle", ",", "assistants", "created", "through", "the", "Snips", "Voice", "Platform", "never", "send", "user", "queries", "to", "the", "cloud", "and", "offer", "state", "-", "of", "-", "the", "-", "art", "performance", ".", "Focusing", "on", "the", "Automatic", "Speech", "Recognition", "and", "Natural", "Language", "Understanding", "engines", ",", "we", "have", "described", "the", "challenges", "of", "embedding", "high", "-", "performance", "machine", "learning", "models", "on", "small", "IoT", "devices", ".", "On", "the", "acoustic", "modeling", "side", ",", "we", "have", "shown", "how", "small", "-", "sized", "neural", "networks", "can", "be", "trained", "that", "enjoy", "near", "state", "-", "of", "-", "the", "-", "art", "accuracy", "while", "running", "in", "real", "-", "time", "on", "small", "devices", ".", "On", "the", "language", "modeling", "side", ",", "we", "have", "described", "how", "to", "train", "the", "language", "model", "of", "the", "ASR", "and", "the", "NLU", "in", "a", "consistent", "way", ",", "efficiently", "specializing", "them", "to", "a", "particular", "use", "case", ".", "We", "have", "also", "demonstrated", "the", "accuracy", "of", "the", "resulting", "SLU", "engine", "on", "real", "-", "world", "assistants", ".", "Finally", ",", "we", "have", "shown", "how", "sufficient", ",", "high", "-", "quality", "training", "data", "can", "be", "obtained", "without", "compromising", "user", "privacy", "through", "a", "combination", "of", "crowdsourcing", "and", "machine", "learning", ".", "We", "hope", "that", "the", "present", "paper", "can", "contribute", "to", "a", "larger", "effort", "towards", "ever", "more", "private", "and", "ubiquitous", "artificial", "intelligence", ".", "Future", "research", "directions", "will", "include", "private", "analytics", ",", "allowing", "to", "receive", "privacy", "-", "preserving", "feedback", "from", "assistant", "usage", ",", "and", "federated", "learning", ",", "as", "a", "complement", "to", "data", "generation", ".", "section", ":", "Acknowledgments", "All", "of", "the", "work", "presented", "here", "has", "been", "done", "closely", "together", "with", "the", "engineering", "teams", "at", "Snips", ".", "We", "are", "grateful", "to", "the", "crowd", "of", "contributors", "who", "regularly", "work", "with", "us", "on", "the", "data", "generation", "pipeline", ".", "We", "are", "indebted", "to", "the", "community", "of", "users", "of", "the", "Snips", "Voice", "Platform", "for", "valuable", "feedback", "and", "contributions", ".", "section", ":", "Appendix", ":", "NLU", "benchmark", "on", "an", "in", "-", "house", "dataset", "bibliography", ":", "References"], "entity": [[2, 5, ["Snips", "Voice", "Platform"], "Method"], [7, 12, ["embedded", "Spoken", "Language", "Understanding", "system"], "Method"], [13, 20, ["private", "-", "by", "-", "design", "voice", "interfaces"], "Task"], [24, 27, ["machine", "learning", "architecture"], "Method"], [29, 32, ["Snips", "Voice", "Platform"], "Method"], [34, 36, ["software", "solution"], "Method"], [38, 41, ["Spoken", "Language", "Understanding"], "Task"], [45, 47, ["IoT", "devices"], "Task"], [49, 51, ["embedded", "inference"], "Method"], [72, 75, ["Automatic", "Speech", "Recognition"], "Task"], [76, 79, ["Natural", "Language", "Understanding"], "Method"], [89, 92, ["Machine", "Learning", "models"], "Method"], [111, 114, ["data", "generation", "procedure"], "Method"], [144, 147, ["deep", "learning", "approaches"], "Method"], [148, 150, ["speech", "recognition"], "Task"], [151, 153, ["voice", "interfaces"], "Task"], [202, 204, ["voice", "assistants"], "Task"], [212, 214, ["customer", "support"], "Task"], [216, 218, ["autonomous", "cars"], "Task"], [220, 222, ["smart", "homes"], "Task"], [226, 229, ["smart", "speaker", "adoption"], "Task"], [269, 272, ["unique", "speaker", "identification"], "Task"], [287, 288, ["CNIL"], "Method"], [323, 326, ["Data", "Protection", "Regulation"], "Method"], [356, 358, ["voice", "assistants"], "Task"], [386, 389, ["Snips", "Voice", "Platform"], "Method"], [405, 407, ["IoT", "microprocessors"], "Method"], [492, 494, ["design", "principles"], "Method"], [496, 499, ["Snips", "Voice", "Platform"], "Method"], [510, 512, ["web", "console"], "Method"], [514, 516, ["voice", "assistants"], "Task"], [540, 543, ["Natural", "Language", "Understanding"], "Method"], [549, 550, ["console"], "Method"], [555, 560, ["self", "-", "service", "development", "environment"], "Task"], [571, 574, ["Snips", "Voice", "Platform"], "Method"], [593, 594, ["Snips"], "Method"], [637, 638, ["assistant"], "Method"], [656, 658, ["SmartHome", "assistant"], "Task"], [816, 819, ["dialog", "management", "component"], "Method"], [846, 848, ["speech", "synthesis"], "Method"], [885, 888, ["design", "embedded", "platform"], "Method"], [889, 893, ["Privacy", "by", "Design", "principle"], "Method"], [901, 907, ["design", "and", "engineering", "of", "a", "system"], "Task"], [912, 914, ["voice", "assistants"], "Task"], [950, 955, ["private", "-", "by", "-", "design"], "Method"], [988, 989, ["inference"], "Task"], [1016, 1018, ["engineering", "complexity"], "Metric"], [1033, 1037, ["Cross", "-", "platform", "support"], "Task"], [1043, 1045, ["IoT", "industry"], "Task"], [1069, 1072, ["Snips", "Voice", "Platform"], "Method"], [1083, 1085, ["embedded", "inference"], "Method"], [1133, 1134, ["IMX.7D"], "Method"], [1144, 1146, ["Jetson", "TX2"], "Method"], [1164, 1166, ["iOS", "SDK"], "Method"], [1188, 1189, ["Rust"], "Method"], [1192, 1194, ["programming", "language"], "Method"], [1198, 1201, ["low", "memory", "overhead"], "Metric"], [1203, 1206, ["cross", "-", "compilation"], "Method"], [1217, 1220, ["Automatic", "Speech", "Recognition"], "Task"], [1221, 1222, ["ASR"], "Method"], [1224, 1227, ["Natural", "Language", "Understanding"], "Method"], [1228, 1229, ["NLU"], "Method"], [1242, 1244, ["acoustic", "model"], "Method"], [1250, 1252, ["phonetic", "representation"], "Method"], [1255, 1257, ["Language", "Model"], "Method"], [1258, 1259, ["LM"], "Method"], [1263, 1265, ["phonetic", "representation"], "Method"], [1269, 1270, ["NLU"], "Method"], [1288, 1289, ["LM"], "Method"], [1290, 1291, ["NLU"], "Method"], [1301, 1302, ["accuracy"], "Metric"], [1314, 1317, ["language", "modeling", "component"], "Method"], [1320, 1321, ["LM"], "Method"], [1322, 1323, ["NLU"], "Method"], [1345, 1349, ["large", "deep", "learning", "models"], "Method"], [1374, 1377, ["human", "parity", "in"], "Task"], [1382, 1384, ["neural", "networks"], "Method"], [1395, 1400, ["large", "-", "vocabulary", "language", "models"], "Method"], [1449, 1450, ["cloud"], "Method"], [1451, 1453, ["speech", "recognition"], "Task"], [1454, 1456, ["Enforcing", "privacy"], "Task"], [1502, 1504, ["acoustic", "models"], "Method"], [1530, 1533, ["language", "modeling", "approach"], "Method"], [1549, 1551, ["language", "model"], "Method"], [1585, 1587, ["small", "devices"], "Task"], [1605, 1607, ["generalization", "accuracy"], "Metric"], [1615, 1620, ["real", "-", "word", "voice", "assistants"], "Task"], [1631, 1634, ["data", "generation", "procedure"], "Method"], [1645, 1647, ["Acoustic", "model"], "Method"], [1648, 1650, ["acoustic", "model"], "Method"], [1692, 1694, ["acoustic", "model"], "Method"], [1712, 1719, ["context", "-", "dependent", "clustered", "Hidden", "Markov", "Model"], "Method"], [1720, 1724, ["HMM", ")", "state", "probabilities"], "Method"], [1732, 1734, ["language", "model"], "Method"], [1748, 1752, ["acoustic", "and", "language", "models"], "Method"], [1758, 1761, ["Automatic", "Speech", "Recognition"], "Task"], [1762, 1763, ["ASR"], "Method"], [1778, 1780, ["language", "model"], "Method"], [1800, 1802, ["acoustic", "model"], "Method"], [1828, 1830, ["acoustic", "model"], "Method"], [1833, 1837, ["hybrid", "of", "Neural", "Networks"], "Method"], [1838, 1841, ["Hidden", "Markov", "Models"], "Method"], [1843, 1846, ["NN", "/", "HMM"], "Method"], [1861, 1863, ["acoustic", "model"], "Method"], [1865, 1869, ["large", "-", "vocabulary", "setup"], "Task"], [1873, 1876, ["word", "error", "rate"], "Metric"], [1877, 1878, ["WER"], "Metric"], [1880, 1881, ["speed"], "Metric"], [1883, 1885, ["memory", "usage"], "Metric"], [1897, 1899, ["acoustic", "model"], "Method"], [1945, 1947, ["transcription", "errors"], "Task"], [1971, 1973, ["acoustic", "training"], "Task"], [1998, 2000, ["Data", "augmentation"], "Task"], [2011, 2013, ["acoustic", "model"], "Method"], [2078, 2081, ["machine", "learning", "perspective"], "Method"], [2169, 2171, ["Model", "training"], "Method"], [2171, 2173, ["Acoustic", "models"], "Method"], [2174, 2179, ["hybrid", "NN", "/", "HMM", "models"], "Method"], [2190, 2193, ["s5", "training", "recipe"], "Method"], [2195, 2197, ["Kaldi", "toolkit"], "Method"], [2232, 2233, ["accuracy"], "Metric"], [2234, 2236, ["computational", "cost"], "Metric"], [2242, 2244, ["model", "architectures"], "Method"], [2252, 2254, ["application", "accuracy"], "Metric"], [2276, 2278, ["Raspberry", "Pi"], "Task"], [2286, 2292, ["Gaussian", "Mixture", "Model", "Hidden", "Markov", "Model"], "Method"], [2293, 2296, ["GMM", "-", "HMM"], "Method"], [2301, 2303, ["speech", "corpus"], "Material"], [2306, 2311, ["context", "-", "dependent", "bootstrapping", "model"], "Method"], [2324, 2327, ["neural", "network", "training"], "Method"], [2331, 2334, ["deep", "neural", "network"], "Method"], [2337, 2341, ["time", "-", "delay", "layers"], "Method"], [2348, 2354, ["long", "short", "-", "term", "memory", "layers"], "Method"], [2381, 2385, ["Raspberry", "Pi", "3", "model"], "Method"], [2394, 2400, ["lattice", "-", "free", "Maximum", "Mutual", "Information"], "Metric"], [2401, 2404, ["MMI", ")", "criterion"], "Metric"], [2406, 2409, ["natural", "gradient", "descent"], "Method"], [2412, 2414, ["learning", "rate"], "Metric"], [2418, 2420, ["backstitching", "trick"], "Method"], [2429, 2432, ["fast", "acoustic", "models"], "Method"], [2435, 2437, ["HMM", "topology"], "Method"], [2459, 2461, ["frame", "rate"], "Metric"], [2464, 2467, ["Acoustic", "model", "evaluation"], "Method"], [2477, 2479, ["acoustic", "model"], "Method"], [2516, 2517, ["accuracy"], "Metric"], [2535, 2540, ["domain", "-", "adapted", "language", "models"], "Method"], [2553, 2554, ["accuracy"], "Metric"], [2556, 2558, ["acoustic", "models"], "Method"], [2568, 2570, ["acoustic", "model"], "Method"], [2586, 2589, ["large", "vocabulary", "setup"], "Task"], [2592, 2595, ["LibriSpeech", "evaluation", "dataset"], "Material"], [2616, 2618, ["language", "model"], "Method"], [2620, 2624, ["large", "-", "vocabulary", "evaluation"], "Task"], [2671, 2674, ["LibriSpeech", "training", "set"], "Material"], [2723, 2727, ["speed", "and", "memory", "usage"], "Metric"], [2741, 2745, ["Raspberry", "Pi", "3", "setting"], "Task"], [2748, 2750, ["Model", "architecture"], "Method"], [2754, 2755, ["LibriSpeech"], "Material"], [2767, 2769, ["model", "accuracy"], "Metric"], [2770, 2772, ["neural", "networks"], "Method"], [2782, 2784, ["LibriSpeech", "dataset"], "Material"], [2788, 2790, ["data", "augmentation"], "Method"], [2795, 2798, ["nnet", "-", "512"], "Method"], [2828, 2830, ["LibriSpeech", "dataset"], "Material"], [2841, 2843, ["data", "augmentation"], "Task"], [2848, 2850, ["clean", "data"], "Material"], [2857, 2859, ["clean", "data"], "Material"], [2862, 2864, ["augmented", "data"], "Material"], [2876, 2881, ["LibriSpeech", "’s", "clean", "test", "sets"], "Material"], [2882, 2885, ["dev", "-", "clean"], "Material"], [2886, 2889, ["test", "-", "clean"], "Material"], [2892, 2893, ["WER"], "Metric"], [2929, 2931, ["data", "augmentation"], "Method"], [2947, 2949, ["neural", "networks"], "Method"], [3003, 3004, ["precision"], "Metric"], [3005, 3007, ["computational", "cost"], "Metric"], [3018, 3020, ["Online", "recognition"], "Task"], [3036, 3039, ["neural", "network", "architectures"], "Method"], [3042, 3046, ["memory", "and", "computational", "costs"], "Metric"], [3066, 3067, ["LM"], "Method"], [3075, 3077, ["end", "accuracies"], "Metric"], [3079, 3081, ["acoustic", "models"], "Method"], [3090, 3092, ["processing", "speed"], "Metric"], [3098, 3100, ["Raspberry", "Pi"], "Method"], [3116, 3121, ["real", "-", "time", "processing", "factor"], "Metric"], [3126, 3128, ["processing", "time"], "Metric"], [3150, 3151, ["accuracy"], "Metric"], [3152, 3156, ["real", "-", "time", "factor"], "Metric"], [3157, 3159, ["model", "size"], "Metric"], [3174, 3176, ["model", "size"], "Metric"], [3185, 3188, ["nnet", "-", "256"], "Method"], [3194, 3197, ["nnet", "-", "768"], "Method"], [3225, 3226, ["speed"], "Metric"], [3228, 3231, ["nnet", "-", "256"], "Method"], [3239, 3242, ["nnet", "-", "768"], "Method"], [3256, 3259, ["nnet", "-", "256"], "Method"], [3264, 3268, ["speed", "and", "memory", "footprint"], "Metric"], [3273, 3274, ["accuracy"], "Metric"], [3278, 3280, ["adapted", "LM"], "Method"], [3281, 3283, ["robust", "NLU"], "Method"], [3285, 3287, ["network", "architecture"], "Method"], [3323, 3326, ["optional", "layer", "factorization"], "Method"], [3331, 3333, ["weight", "quantization"], "Method"], [3356, 3358, ["Language", "Modeling"], "Method"], [3366, 3369, ["language", "modeling", "component"], "Method"], [3379, 3385, ["extraction", "of", "the", "intent", "and", "slots"], "Task"], [3390, 3392, ["acoustic", "model"], "Method"], [3409, 3411, ["language", "model"], "Method"], [3421, 3423, ["acoustic", "model"], "Method"], [3443, 3446, ["Natural", "Language", "Understanding"], "Method"], [3447, 3448, ["NLU"], "Method"], [3461, 3464, ["Automatic", "Speech", "Recognition"], "Task"], [3465, 3466, ["ASR"], "Method"], [3510, 3514, ["large", "vocabulary", "language", "models"], "Method"], [3534, 3535, ["decoding"], "Task"], [3542, 3543, ["ASR"], "Method"], [3547, 3549, ["speech", "recognition"], "Task"], [3616, 3618, ["language", "model"], "Method"], [3692, 3695, ["word", "error", "rate"], "Metric"], [3717, 3723, ["speech", "-", "to", "-", "meaning", "accuracy"], "Metric"], [3748, 3749, ["LM"], "Method"], [3761, 3762, ["NLU"], "Method"], [3776, 3778, ["decoding", "speed"], "Metric"], [3782, 3783, ["ASR"], "Method"], [3793, 3794, ["accuracy"], "Metric"], [3835, 3836, ["accuracy"], "Metric"], [3867, 3868, ["LM"], "Method"], [3869, 3870, ["NLU"], "Method"], [3907, 3908, ["LM"], "Method"], [3909, 3910, ["NLU"], "Method"], [3979, 3983, ["weather", "-", "related", "queries"], "Task"], [4201, 4203, ["Slot", "labeling"], "Task"], [4289, 4290, ["Normalization"], "Method"], [4302, 4304, ["data", "consistency"], "Task"], [4305, 4306, ["training"], "Task"], [4307, 4308, ["inference"], "Task"], [4321, 4323, ["writing", "system"], "Method"], [4338, 4339, ["LM"], "Method"], [4403, 4404, ["NLU"], "Method"], [4413, 4414, ["inference"], "Task"], [4426, 4427, ["ASR"], "Method"], [4432, 4433, ["NLU"], "Method"], [4437, 4439, ["normalization", "pipeline"], "Method"], [4451, 4455, ["class", "-", "based", "tokenizer"], "Method"], [4459, 4465, ["case", "-", "by", "-", "case", "verbalization"], "Task"], [4521, 4522, ["tokenizer"], "Method"], [4526, 4532, ["character", "-", "level", "finite", "state", "transducer"], "Method"], [4553, 4555, ["Language", "model"], "Method"], [4562, 4564, ["acoustic", "model"], "Method"], [4572, 4574, ["Viterbi", "search"], "Method"], [4576, 4580, ["weighted", "Finite", "State", "Transducer"], "Method"], [4581, 4582, ["wFST"], "Method"], [4605, 4606, ["wFSTs"], "Method"], [4609, 4611, ["transducer", "composition"], "Method"], [4620, 4623, ["Hidden", "Markov", "Models"], "Method"], [4624, 4625, ["HMMs"], "Method"], [4644, 4645, ["LM"], "Method"], [4648, 4649, ["bigram"], "Method"], [4651, 4653, ["trigram", "model"], "Method"], [4656, 4657, ["wFST"], "Method"], [4658, 4662, ["Determinization", "and", "minimization", "operations"], "Method"], [4684, 4685, ["inference"], "Task"], [4695, 4697, ["classical", "transducers"], "Method"], [4727, 4729, ["G", "transducer"], "Method"], [4732, 4733, ["LM"], "Method"], [4745, 4748, ["Language", "Model", "Adaptation"], "Method"], [4798, 4800, ["generalization", "properties"], "Metric"], [4830, 4835, ["class", "-", "based", "language", "modeling"], "Method"], [4863, 4864, ["LM"], "Method"], [4935, 4939, ["n", "-", "gram", "model"], "Method"], [5028, 5032, ["n", "-", "gram", "model"], "Method"], [5056, 5058, ["wFST", "representation"], "Method"], [5060, 5062, ["generative", "grammar"], "Method"], [5080, 5081, ["Replace"], "Method"], [5082, 5084, ["wFST", "replacement"], "Method"], [5107, 5108, ["wFST"], "Method"], [5123, 5124, ["LM"], "Method"], [5134, 5135, ["ASR"], "Method"], [5206, 5213, ["n", "-", "gram", "back", "-", "off", "mechanism"], "Method"], [5215, 5217, ["lexicon", "transducer"], "Method"], [5245, 5250, ["grapheme", "-", "to", "-", "phoneme"], "Method"], [5251, 5252, ["G2P"], "Method"], [5262, 5265, ["Dynamic", "Language", "Model"], "Method"], [5295, 5296, ["minimization"], "Method"], [5329, 5330, ["wFST"], "Method"], [5332, 5336, ["n", "-", "gram", "model"], "Method"], [5350, 5352, ["static", "model"], "Method"], [5360, 5362, ["embedded", "inference"], "Task"], [5388, 5390, ["speech", "recognition"], "Task"], [5395, 5397, ["decoding", "graph"], "Method"], [5426, 5427, ["LMs"], "Method"], [5432, 5433, ["Snips"], "Method"], [5470, 5473, ["dynamic", "language", "model"], "Method"], [5495, 5498, ["compositions", "of", "equation"], "Method"], [5520, 5523, ["HCLG", "decoding", "graph"], "Method"], [5538, 5539, ["inference"], "Task"], [5547, 5549, ["decoding", "graph"], "Method"], [5553, 5555, ["lazy", "composition"], "Method"], [5597, 5598, ["HCLG"], "Method"], [5604, 5606, ["decoding", "speed"], "Metric"], [5612, 5615, ["dynamic", "language", "model"], "Method"], [5618, 5620, ["composition", "algorithm"], "Method"], [5634, 5636, ["lazy", "composition"], "Method"], [5647, 5648, ["wFST"], "Method"], [5671, 5673, ["optimization", "step"], "Method"], [5675, 5676, ["HCLG"], "Method"], [5683, 5685, ["dynamic", "case"], "Task"], [5688, 5689, ["HCLG"], "Method"], [5702, 5704, ["composition", "filters"], "Method"], [5711, 5715, ["look", "-", "ahead", "filters"], "Method"], [5717, 5721, ["label", "-", "reachability", "filters"], "Method"], [5737, 5738, ["decoding"], "Task"], [5747, 5749, ["language", "model"], "Method"], [5754, 5755, ["transducer"], "Method"], [5759, 5761, ["Replace", "operation"], "Method"], [5786, 5788, ["transducer", "mapping"], "Method"], [5792, 5794, ["acoustic", "model"], "Method"], [5798, 5801, ["query", "language", "model"], "Method"], [5804, 5807, ["entities", "languages", "models"], "Method"], [5825, 5827, ["HCL", "transducer"], "Method"], [5831, 5834, ["regular", "transducer", "compositions"], "Method"], [5840, 5843, ["delayed", "transducer", "composition"], "Method"], [5844, 5846, ["composition", "filters"], "Method"], [5850, 5853, ["real", "time", "decoding"], "Task"], [5865, 5867, ["static", "model"], "Method"], [5870, 5872, ["decoding", "accuracy"], "Metric"], [5889, 5891, ["training", "time"], "Metric"], [5893, 5894, ["LM"], "Method"], [5900, 5901, ["LM"], "Method"], [5918, 5925, ["on", "-", "device", "injection", "of", "new", "values"], "Task"], [5943, 5944, ["Snips"], "Method"], [5957, 5961, ["On", "-", "device", "personalization"], "Task"], [5965, 5966, ["ASR"], "Method"], [5973, 5974, ["recognition"], "Task"], [5978, 5980, ["language", "model"], "Method"], [5993, 5994, ["LM"], "Method"], [6072, 6074, ["entity", "injection"], "Task"], [6082, 6084, ["entity", "injection"], "Task"], [6106, 6107, ["transducer"], "Method"], [6130, 6131, ["ASR"], "Method"], [6147, 6148, ["transducer"], "Method"], [6152, 6154, ["embedded", "G2P."], "Method"], [6178, 6182, ["union", "of", "word", "acceptors"], "Method"], [6184, 6188, ["n", "-", "gram", "model"], "Method"], [6231, 6236, ["embedded", "n", "-", "gram", "engine"], "Method"], [6243, 6246, ["entity", "injection", "procedure"], "Task"], [6308, 6309, ["accuracy"], "Metric"], [6316, 6318, ["Confidence", "scoring"], "Method"], [6340, 6345, ["out", "-", "of", "-", "vocabulary"], "Method"], [6346, 6347, ["OOV"], "Method"], [6396, 6397, ["ASR"], "Method"], [6424, 6425, ["NLU"], "Method"], [6436, 6441, ["word", "-", "level", "confidence", "score"], "Metric"], [6443, 6444, ["ASR"], "Method"], [6456, 6458, ["Confidence", "scoring"], "Task"], [6464, 6466, ["speech", "recognition"], "Task"], [6478, 6480, ["confusion", "network"], "Method"], [6488, 6489, ["ASR"], "Method"], [6498, 6500, ["confusion", "network"], "Method"], [6502, 6504, ["graph", "encoding"], "Method"], [6538, 6540, ["confusion", "networks"], "Method"], [6554, 6560, ["speech", "-", "to", "-", "meaning", "tasks"], "Task"], [6570, 6572, ["confusion", "networks"], "Method"], [6574, 6576, ["greedy", "decoder"], "Method"], [6602, 6603, ["OOVs"], "Task"], [6629, 6633, ["post", "-", "processing", "step"], "Method"], [6696, 6697, ["SLU"], "Method"], [6722, 6726, ["sentence", "-", "level", "confidence"], "Metric"], [6730, 6739, ["geometric", "mean", "of", "the", "word", "-", "level", "confidence", "scores"], "Metric"], [6742, 6745, ["Natural", "Language", "Understanding"], "Method"], [6752, 6755, ["Snips", "Voice", "Platform"], "Method"], [6765, 6767, ["Snips", "NLU"], "Method"], [6769, 6771, ["Python", "library"], "Method"], [6776, 6779, ["training", "and", "inference"], "Task"], [6782, 6784, ["Rust", "implementation"], "Method"], [6787, 6788, ["inference"], "Task"], [6803, 6805, ["Intent", "Classification"], "Task"], [6827, 6829, ["Slot", "Filling"], "Task"], [6848, 6850, ["Entity", "Resolution"], "Task"], [6869, 6870, ["Snips"], "Method"], [6903, 6904, ["Snippet"], "Method"], [6951, 6953, ["Intent", "Parser"], "Method"], [6957, 6959, ["intent", "classification"], "Task"], [6960, 6962, ["slot", "filling"], "Task"], [6975, 6977, ["intent", "parsers"], "Method"], [6980, 6983, ["deterministic", "intent", "parser"], "Method"], [6984, 6987, ["probabilistic", "intent", "parser"], "Method"], [7004, 7007, ["Deterministic", "Intent", "Parser"], "Method"], [7012, 7015, ["deterministic", "intent", "parser"], "Method"], [7018, 7019, ["robustness"], "Metric"], [7157, 7160, ["Probabilistic", "Intent", "Parser"], "Method"], [7162, 7165, ["probabilistic", "intent", "parser"], "Method"], [7168, 7169, ["parsing"], "Task"], [7191, 7193, ["deterministic", "parser"], "Method"], [7196, 7197, ["parser"], "Method"], [7203, 7205, ["intent", "classification"], "Task"], [7206, 7208, ["slot", "filling"], "Task"], [7210, 7212, ["intent", "classification"], "Task"], [7216, 7218, ["logistic", "regression"], "Method"], [7227, 7231, ["slot", "-", "filling", "step"], "Method"], [7234, 7240, ["linear", "-", "chain", "Conditional", "Random", "Fields"], "Method"], [7241, 7242, ["CRFs"], "Method"], [7261, 7263, ["intent", "classifier"], "Method"], [7283, 7287, ["slot", "-", "filling", "step"], "Task"], [7310, 7312, ["generalization", "error"], "Metric"], [7321, 7324, ["deep", "learning", "models"], "Method"], [7342, 7343, ["accuracy"], "Metric"], [7350, 7353, ["custom", "voice", "assistants"], "Task"], [7375, 7379, ["slot", "-", "filling", "tasks"], "Task"], [7420, 7422, ["entity", "extractor"], "Method"], [7427, 7429, ["Entity", "Resolution"], "Task"], [7470, 7473, ["BILOU", "coding", "scheme"], "Method"], [7498, 7499, ["accuracy"], "Metric"], [7503, 7504, ["robustness"], "Metric"], [7506, 7511, ["built", "-", "in", "entities", "extractor"], "Method"], [7515, 7517, ["data", "sparsity"], "Task"], [7524, 7526, ["word", "clusters"], "Method"], [7547, 7549, ["word2vec", "embeddings"], "Method"], [7550, 7554, ["k", "-", "means", "clustering"], "Method"], [7566, 7568, ["generalization", "error"], "Metric"], [7622, 7625, ["BILOU", "coding", "scheme"], "Method"], [7637, 7639, ["Snips", "NLU"], "Method"], [7689, 7690, ["CRF"], "Method"], [7709, 7711, ["Entity", "resolution"], "Task"], [7778, 7779, ["resolution"], "Task"], [7782, 7783, ["Rustling"], "Method"], [7788, 7796, ["re", "-", "implementation", "of", "Facebook", "’s", "Duckling", "library"], "Method"], [7797, 7798, ["Rust"], "Method"], [7826, 7828, ["Snips", "NLU"], "Method"], [7885, 7887, ["Snips", "NLU"], "Method"], [7902, 7906, ["cloud", "-", "based", "solutions"], "Method"], [7907, 7910, ["Microsoft", "’s", "Luis"], "Method"], [7911, 7913, ["IBM", "Watson"], "Method"], [7914, 7915, ["API.AI"], "Method"], [7916, 7919, ["Google", "’s", "Dialogflow"], "Method"], [7948, 7952, ["average", "F1", "-", "score"], "Metric"], [7953, 7955, ["intent", "classification"], "Task"], [7956, 7958, ["slot", "filling"], "Task"], [7973, 7974, ["StackExchange"], "Method"], [8009, 8013, ["train", "and", "test", "splits"], "Metric"], [8036, 8042, ["5", "-", "fold", "cross", "-", "validation"], "Method"], [8078, 8079, ["Rasa"], "Method"], [8087, 8088, ["Spacy"], "Method"], [8089, 8090, ["SKLearn"], "Method"], [8091, 8092, ["MITIE"], "Method"], [8093, 8094, ["MITIE"], "Method"], [8108, 8109, ["Spacy"], "Method"], [8137, 8139, ["Snips", "NLU"], "Method"], [8159, 8161, ["Snips", "NLU"], "Method"], [8308, 8310, ["slot", "filling"], "Task"], [8313, 8315, ["intent", "classification"], "Task"], [8326, 8330, ["average", "F1", "-", "score"], "Metric"], [8331, 8333, ["slot", "filling"], "Method"], [8371, 8375, ["cloud", "-", "based", "providers"], "Method"], [8378, 8380, ["Snips", "NLU"], "Method"], [8381, 8384, ["Microsoft", "’s", "Luis"], "Method"], [8385, 8386, ["API.AI"], "Method"], [8387, 8390, ["Google", "’s", "Dialogflow"], "Method"], [8391, 8394, ["Facebook", "’s", "Wit.ai"], "Method"], [8396, 8398, ["Amazon", "Alexa"], "Method"], [8437, 8438, ["precision"], "Metric"], [8439, 8440, ["recall"], "Metric"], [8441, 8444, ["F1", "-", "score"], "Metric"], [8472, 8474, ["Snips", "NLU"], "Method"], [8481, 8485, ["cloud", "-", "based", "solutions"], "Method"], [8486, 8488, ["slot", "filling"], "Task"], [8501, 8502, ["Rust"], "Method"], [8511, 8513, ["memory", "footprint"], "Metric"], [8515, 8517, ["inference", "runtime"], "Metric"], [8520, 8522, ["Memory", "usage"], "Metric"], [8558, 8560, ["Raspberry", "Pi"], "Task"], [8562, 8564, ["mobile", "app"], "Task"], [8585, 8587, ["inference", "runtime"], "Metric"], [8590, 8591, ["roundtrip"], "Method"], [8622, 8626, ["Snips", "Spoken", "Language", "Understanding"], "Method"], [8627, 8628, ["SLU"], "Method"], [8632, 8644, ["end", "-", "to", "-", "end", ",", "speech", "-", "to", "-", "meaning", "setting"], "Task"], [8661, 8662, ["SmartLights"], "Method"], [8663, 8664, ["Weather"], "Method"], [8694, 8696, ["weather", "queries"], "Task"], [8810, 8811, ["assistants"], "Method"], [8843, 8845, ["speech", "corpus"], "Material"], [8885, 8887, ["classical", "precision"], "Metric"], [8888, 8889, ["recall"], "Metric"], [8891, 8893, ["F1", "scores"], "Metric"], [8896, 8900, ["Language", "Model", "Generalization", "Error"], "Metric"], [8930, 8932, ["generalization", "ability"], "Metric"], [8940, 8946, ["5", "-", "fold", "cross", "-", "validation"], "Method"], [8951, 8952, ["LM"], "Method"], [8953, 8954, ["NLU"], "Method"], [9037, 9038, ["accuracy"], "Metric"], [9040, 9041, ["NLU"], "Method"], [9069, 9070, ["SLU"], "Method"], [9103, 9106, ["confidence", "scoring", "strategy"], "Method"], [9110, 9111, ["precision"], "Metric"], [9112, 9113, ["recall"], "Metric"], [9134, 9139, ["sentence", "-", "level", "confidence", "score"], "Metric"], [9147, 9150, ["word", "error", "rate"], "Metric"], [9174, 9175, ["NLU"], "Method"], [9182, 9183, ["precision"], "Metric"], [9184, 9185, ["recall"], "Metric"], [9188, 9190, ["Embedded", "Performance"], "Metric"], [9211, 9212, ["Snips"], "Method"], [9224, 9226, ["language", "models"], "Method"], [9246, 9248, ["acoustic", "model"], "Method"], [9290, 9292, ["Training", "models"], "Method"], [9300, 9302, ["design", "approach"], "Method"], [9313, 9316, ["machine", "learning", "models"], "Method"], [9330, 9333, ["language", "modeling", "components"], "Method"], [9334, 9336, ["Language", "Model"], "Method"], [9361, 9364, ["data", "generation", "pipeline"], "Method"], [9378, 9379, ["crowdsourcing"], "Method"], [9380, 9385, ["semi", "-", "supervised", "machine", "learning"], "Method"], [9415, 9418, ["data", "generation", "pipeline"], "Method"], [9429, 9430, ["NLU"], "Method"], [9433, 9436, ["Data", "generation", "pipeline"], "Method"], [9441, 9443, ["data", "generation"], "Task"], [9444, 9448, ["grammar", "-", "based", "generation"], "Method"], [9517, 9519, ["pattern", "definition"], "Method"], [9520, 9521, ["enumeration"], "Task"], [9542, 9547, ["generating", "queries", "in", "natural", "language"], "Task"], [9553, 9554, ["crowdsourcing"], "Method"], [9558, 9562, ["Natural", "Language", "Processing", "research"], "Task"], [9566, 9567, ["formulation"], "Task"], [9582, 9583, ["accuracy"], "Metric"], [9586, 9588, ["slot", "supervision"], "Task"], [9621, 9622, ["SLU"], "Method"], [9635, 9638, ["language", "modeling", "components"], "Method"], [9641, 9642, ["crowdsourcing"], "Method"], [9643, 9650, ["machine", "-", "learning", "-", "based", "disambiguation", "techniques"], "Method"], [9658, 9661, ["data", "generation", "pipeline"], "Method"], [9665, 9667, ["Crowdsourcing", "tasks"], "Task"], [9680, 9684, ["non", "-", "expert", "annotations"], "Task"], [9685, 9688, ["natural", "language", "tasks"], "Task"], [9730, 9734, ["text", "query", "generation", "task"], "Task"], [9740, 9742, ["user", "query"], "Task"], [9801, 9803, ["sentence", "generation"], "Task"], [9809, 9812, ["slot", "labeling", "step"], "Method"], [9856, 9858, ["validation", "process"], "Method"], [9886, 9888, ["Majority", "voting"], "Method"], [9896, 9898, ["quality", "assessment"], "Task"], [9899, 9900, ["crowdsourcing"], "Method"], [9903, 9904, ["dashboard"], "Method"], [9946, 9949, ["fuzzy", "matching", "rule"], "Method"], [9971, 9972, ["Disambiguation"], "Task"], [10119, 10120, ["LM"], "Method"], [10136, 10137, ["accuracy"], "Metric"], [10139, 10140, ["SLU"], "Method"], [10148, 10153, ["3", "-", "fold", "cross", "validation"], "Method"], [10192, 10194, ["majority", "voting"], "Method"], [10263, 10265, ["data", "generation"], "Task"], [10267, 10268, ["SLU"], "Method"], [10275, 10279, ["slot", "-", "filling", "task"], "Task"], [10308, 10311, ["slot", "-", "filling"], "Method"], [10322, 10325, ["F1", "-", "scores"], "Metric"], [10368, 10371, ["data", "generation", "pipeline"], "Method"], [10390, 10394, ["slot", "-", "filling", "task"], "Task"], [10426, 10428, ["RateBook", "intent"], "Metric"], [10558, 10564, ["private", "-", "by", "-", "design", "assistants"], "Task"], [10582, 10585, ["machine", "learning", "models"], "Method"], [10621, 10624, ["Snips", "Voice", "Platform"], "Method"], [10626, 10630, ["Spoken", "Language", "Understanding", "solution"], "Method"], [10657, 10660, ["Snips", "Voice", "Platform"], "Method"], [10681, 10684, ["Automatic", "Speech", "Recognition"], "Task"], [10700, 10703, ["machine", "learning", "models"], "Method"], [10704, 10707, ["small", "IoT", "devices"], "Task"], [10710, 10712, ["acoustic", "modeling"], "Task"], [10718, 10723, ["small", "-", "sized", "neural", "networks"], "Method"], [10736, 10737, ["accuracy"], "Metric"], [10749, 10751, ["language", "modeling"], "Task"], [10760, 10762, ["language", "model"], "Method"], [10764, 10765, ["ASR"], "Method"], [10767, 10768, ["NLU"], "Method"], [10787, 10788, ["accuracy"], "Metric"], [10823, 10824, ["crowdsourcing"], "Method"], [10825, 10827, ["machine", "learning"], "Method"], [10845, 10848, ["ubiquitous", "artificial", "intelligence"], "Task"], [10854, 10856, ["private", "analytics"], "Task"], [10869, 10871, ["federated", "learning"], "Method"], [10876, 10878, ["data", "generation"], "Task"], [10898, 10899, ["Snips"], "Method"], [10915, 10918, ["data", "generation", "pipeline"], "Task"], [10929, 10932, ["Snips", "Voice", "Platform"], "Method"], [186, 188, ["speech", "recognition"], "Task"], [372, 374, ["Amazon", "Alexa"], "Method"], [424, 427, ["Natural", "Language", "Understanding"], "Method"], [485, 486, ["Snips"], "Method"], [503, 504, ["Snips"], "Method"], [506, 507, ["Snips"], "Method"], [524, 525, ["SLU"], "Method"], [531, 534, ["Automatic", "Speech", "Recognition"], "Task"], [535, 536, ["ASR"], "Method"], [544, 545, ["NLU"], "Method"], [612, 613, ["Snips"], "Method"], [623, 624, ["NLU"], "Method"], [647, 648, ["SmartLights"], "Method"], [698, 699, ["SmartLights"], "Method"], [759, 760, ["SLU"], "Method"], [865, 866, ["SmartLights"], "Method"], [975, 976, ["Snips"], "Method"], [979, 980, ["SLU"], "Method"], [1148, 1149, ["Snips"], "Method"], [1207, 1208, ["SLU"], "Method"], [1232, 1233, ["ASR"], "Method"], [1304, 1305, ["SLU"], "Method"], [1334, 1335, ["SLU"], "Method"], [1341, 1342, ["ASR"], "Method"], [1466, 1467, ["SLU"], "Method"], [1535, 1536, ["Snips"], "Method"], [1536, 1537, ["SLU"], "Method"], [1553, 1554, ["ASR"], "Method"], [1557, 1558, ["NLU"], "Method"], [1570, 1571, ["SLU"], "Method"], [1589, 1590, ["NLU"], "Method"], [1609, 1610, ["SLU"], "Method"], [1656, 1657, ["SLU"], "Method"], [1783, 1784, ["SLU"], "Method"], [2494, 2495, ["SLU"], "Method"], [2527, 2528, ["SLU"], "Method"], [2635, 2636, ["LM"], "Method"], [3133, 3134, ["LM"], "Method"], [3137, 3138, ["NLU"], "Method"], [3371, 3372, ["Snips"], "Method"], [3412, 3413, ["LM"], "Method"], [3474, 3476, ["speech", "recognition"], "Task"], [3479, 3480, ["LM"], "Method"], [3561, 3562, ["ASR"], "Method"], [3575, 3576, ["LM"], "Method"], [3683, 3684, ["ASR"], "Method"], [3707, 3708, ["SLU"], "Method"], [3806, 3807, ["ASR"], "Method"], [3824, 3825, ["Snips"], "Method"], [3825, 3826, ["SLU"], "Method"], [3947, 3948, ["SmartLights"], "Method"], [4139, 4140, ["SmartLights"], "Method"], [4300, 4301, ["SLU"], "Method"], [4585, 4586, ["ASR"], "Method"], [4753, 4754, ["ASR"], "Method"], [4802, 4803, ["ASR"], "Method"], [4813, 4814, ["LM"], "Method"], [4853, 4854, ["LM"], "Method"], [5073, 5074, ["LM"], "Method"], [5184, 5185, ["LM"], "Method"], [5608, 5609, ["ASR"], "Method"], [5927, 5928, ["LM"], "Method"], [5937, 5938, ["SLU"], "Method"], [6311, 6312, ["SLU"], "Method"], [6323, 6324, ["SLU"], "Method"], [6543, 6544, ["NLU"], "Method"], [6649, 6650, ["OOV"], "Method"], [6655, 6656, ["SLU"], "Method"], [6663, 6664, ["ASR"], "Method"], [6746, 6749, ["Natural", "Language", "Understanding"], "Method"], [6913, 6914, ["NLU"], "Method"], [6920, 6921, ["Snips"], "Method"], [6921, 6922, ["NLU"], "Method"], [6935, 6936, ["NLU"], "Method"], [6971, 6972, ["NLU"], "Method"], [7717, 7718, ["NLU"], "Method"], [7834, 7835, ["NLU"], "Method"], [7894, 7895, ["NLU"], "Method"], [7928, 7929, ["NLU"], "Method"], [8129, 8130, ["NLU"], "Method"], [8504, 8505, ["NLU"], "Method"], [8579, 8580, ["Snips"], "Method"], [8666, 8667, ["SmartLights"], "Method"], [8689, 8690, ["Weather"], "Method"], [8826, 8827, ["crowdsourcing"], "Method"], [8849, 8850, ["ASR"], "Method"], [8854, 8855, ["NLU"], "Method"], [8912, 8913, ["SLU"], "Method"], [8934, 8935, ["Snips"], "Method"], [8935, 8936, ["SLU"], "Method"], [9032, 9033, ["ASR"], "Method"], [9123, 9124, ["NLU"], "Method"], [9192, 9193, ["SLU"], "Method"], [9237, 9238, ["SmartLights"], "Method"], [9254, 9255, ["Weather"], "Method"], [9263, 9264, ["SLU"], "Method"], [9337, 9340, ["Natural", "Language", "Understanding"], "Method"], [9721, 9722, ["crowdsourcing"], "Method"], [9864, 9865, ["crowdsourcing"], "Method"], [10110, 10111, ["NLU"], "Method"], [10155, 10156, ["NLU"], "Method"], [10281, 10282, ["NLU"], "Method"], [10343, 10344, ["NLU"], "Method"], [10396, 10397, ["NLU"], "Method"], [10685, 10688, ["Natural", "Language", "Understanding"], "Method"], [10791, 10792, ["SLU"], "Method"], [10942, 10943, ["NLU"], "Method"]]} {"doc_id": "16028acc465c434a43599f0f900c91f38d690e02", "text": ["CornerNet", ":", "Detecting", "Objects", "as", "Paired", "Keypoints", "section", ":", "Abstract", "We", "propose", "CornerNet", ",", "a", "new", "approach", "to", "object", "detection", "where", "we", "detect", "an", "object", "bounding", "box", "as", "a", "pair", "of", "keypoints", ",", "the", "top", "-", "left", "corner", "and", "the", "bottom", "-", "right", "corner", ",", "using", "a", "single", "convolution", "neural", "network", ".", "By", "detecting", "objects", "as", "paired", "keypoints", ",", "we", "eliminate", "the", "need", "for", "designing", "a", "set", "of", "anchor", "boxes", "commonly", "used", "in", "prior", "single", "-", "stage", "detectors", ".", "In", "addition", "to", "our", "novel", "formulation", ",", "we", "introduce", "corner", "pooling", ",", "a", "new", "type", "of", "pooling", "layer", "that", "helps", "the", "network", "better", "localize", "corners", ".", "Experiments", "show", "that", "CornerNet", "achieves", "a", "42.2", "%", "AP", "on", "MS", "COCO", ",", "outperforming", "all", "existing", "one", "-", "stage", "detectors", ".", "section", ":", "Introduction", "Object", "detectors", "based", "on", "convolutional", "neural", "networks", "(", "ConvNets", ")", "[", "reference", "][", "reference", "][", "reference", "]", "have", "achieved", "state", "-", "of", "-", "the", "-", "art", "results", "on", "various", "challenging", "benchmarks", "[", "reference", "][", "reference", "][", "reference", "]", ".", "A", "common", "component", "of", "state", "-", "of", "-", "the", "-", "art", "approaches", "is", "anchor", "boxes", "[", "reference", "][", "reference", "]", ",", "which", "are", "boxes", "of", "various", "sizes", "and", "aspect", "ratios", "that", "serve", "as", "detection", "candidates", ".", "Anchor", "boxes", "are", "extensively", "used", "in", "one", "-", "stage", "detectors", "[", "reference", "][", "reference", "][", "reference", "]", ",", "which", "can", "achieve", "results", "highly", "competitive", "with", "two", "-", "stage", "detectors", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "while", "being", "more", "efficient", ".", "One", "-", "stage", "detectors", "place", "anchor", "boxes", "densely", "over", "an", "image", "and", "generate", "final", "box", "predictions", "by", "scoring", "anchor", "boxes", "and", "refining", "their", "coordinates", "through", "regression", ".", "But", "the", "use", "of", "anchor", "boxes", "has", "two", "drawbacks", ".", "First", ",", "we", "typically", "need", "a", "very", "large", "set", "of", "anchor", "boxes", ",", "e.g.", "more", "than", "40k", "in", "DSSD", "and", "more", "than", "100k", "in", "RetinaNet", ".", "This", "is", "because", "the", "detector", "is", "trained", "to", "classify", "whether", "each", "anchor", "box", "sufficiently", "overlaps", "with", "a", "ground", "truth", "box", ",", "and", "a", "large", "number", "of", "anchor", "boxes", "is", "needed", "to", "ensure", "sufficient", "overlap", "with", "most", "ground", "truth", "boxes", ".", "As", "a", "result", ",", "only", "a", "tiny", "fraction", "of", "anchor", "boxes", "will", "overlap", "with", "ground", "truth", ";", "this", "creates", "a", "huge", "imbalance", "between", "positive", "and", "negative", "anchor", "boxes", "and", "slows", "down", "training", ".", "Second", ",", "the", "use", "of", "anchor", "boxes", "introduces", "many", "hyperparameters", "and", "design", "choices", ".", "These", "include", "how", "many", "boxes", ",", "what", "sizes", ",", "and", "what", "aspect", "ratios", ".", "Such", "choices", "have", "largely", "been", "made", "via", "ad", "-", "hoc", "heuristics", ",", "and", "can", "become", "even", "more", "complicated", "when", "combined", "with", "multiscale", "architectures", "where", "a", "single", "network", "makes", "separate", "predictions", "at", "multiple", "resolutions", ",", "with", "each", "scale", "using", "different", "features", "and", "its", "own", "set", "of", "anchor", "boxes", "[", "reference", "][", "reference", "][", "reference", "]", ".", "In", "this", "paper", "we", "introduce", "CornerNet", ",", "a", "new", "onestage", "approach", "to", "object", "detection", "that", "does", "away", "with", "anchor", "boxes", ".", "We", "detect", "an", "object", "as", "a", "pair", "of", "keypointsthe", "top", "-", "left", "corner", "and", "bottom", "-", "right", "corner", "of", "the", "bounding", "box", ".", "We", "use", "a", "single", "convolutional", "network", "to", "predict", "a", "heatmap", "for", "the", "top", "-", "left", "corners", "of", "all", "instances", "of", "the", "same", "object", "category", ",", "a", "heatmap", "for", "all", "bottomright", "corners", ",", "and", "an", "embedding", "vector", "for", "each", "detected", "corner", ".", "The", "embeddings", "serve", "to", "group", "a", "pair", "of", "corners", "that", "belong", "to", "the", "same", "object", "-", "the", "network", "is", "trained", "to", "predict", "similar", "embeddings", "for", "them", ".", "Our", "ap", "-", "section", ":", "ConvNet", "Embeddings", "Heatmaps", "Top", "-", "Left", "Corners", "Bottom", "-", "Right", "Corners", "Fig", ".", "1", "We", "detect", "an", "object", "as", "a", "pair", "of", "bounding", "box", "corners", "grouped", "together", ".", "A", "convolutional", "network", "outputs", "a", "heatmap", "for", "all", "top", "-", "left", "corners", ",", "a", "heatmap", "for", "all", "bottom", "-", "right", "corners", ",", "and", "an", "embedding", "vector", "for", "each", "detected", "corner", ".", "The", "network", "is", "trained", "to", "predict", "similar", "embeddings", "for", "corners", "that", "belong", "to", "the", "same", "object", ".", "proach", "greatly", "simplifies", "the", "output", "of", "the", "network", "and", "eliminates", "the", "need", "for", "designing", "anchor", "boxes", ".", "Our", "approach", "is", "inspired", "by", "the", "associative", "embedding", "method", "proposed", "by", ",", "who", "detect", "and", "group", "keypoints", "in", "the", "context", "of", "multiperson", "human", "-", "pose", "estimation", ".", "Fig", ".", "1", "illustrates", "the", "overall", "pipeline", "of", "our", "approach", ".", "Another", "novel", "component", "of", "CornerNet", "is", "corner", "pooling", ",", "a", "new", "type", "of", "pooling", "layer", "that", "helps", "a", "convolutional", "network", "better", "localize", "corners", "of", "bounding", "boxes", ".", "A", "corner", "of", "a", "bounding", "box", "is", "often", "outside", "the", "object", "-", "consider", "the", "case", "of", "a", "circle", "as", "well", "as", "the", "examples", "in", "Fig", ".", "2", ".", "In", "such", "cases", "a", "corner", "can", "not", "be", "localized", "based", "on", "local", "evidence", ".", "Instead", ",", "to", "determine", "whether", "there", "is", "a", "top", "-", "left", "corner", "at", "a", "pixel", "location", ",", "we", "need", "to", "look", "horizontally", "towards", "the", "right", "for", "the", "topmost", "boundary", "of", "the", "object", ",", "and", "look", "vertically", "towards", "the", "bottom", "for", "the", "leftmost", "boundary", ".", "This", "motivates", "our", "corner", "pooling", "layer", ":", "it", "takes", "in", "two", "feature", "maps", ";", "at", "each", "pixel", "location", "it", "max", "-", "pools", "all", "feature", "vectors", "to", "the", "right", "from", "the", "first", "feature", "map", ",", "maxpools", "all", "feature", "vectors", "directly", "below", "from", "the", "second", "feature", "map", ",", "and", "then", "adds", "the", "two", "pooled", "results", "together", ".", "An", "example", "is", "shown", "in", "Fig", ".", "3", ".", "We", "hypothesize", "two", "reasons", "why", "detecting", "corners", "would", "work", "better", "than", "bounding", "box", "centers", "or", "proposals", ".", "First", ",", "the", "center", "of", "a", "box", "can", "be", "harder", "to", "localize", "because", "it", "depends", "on", "all", "4", "sides", "of", "the", "object", ",", "whereas", "locating", "a", "corner", "depends", "on", "2", "sides", "and", "is", "thus", "easier", ",", "and", "even", "more", "so", "with", "corner", "pooling", ",", "which", "encodes", "some", "explicit", "prior", "knowledge", "about", "the", "definition", "of", "corners", ".", "Second", ",", "corners", "provide", "a", "more", "efficient", "way", "of", "densely", "discretizing", "the", "space", "of", "boxes", ":", "we", "just", "need", "O", "(", "wh", ")", "corners", "to", "represent", "O", "(", "w", "2", "h", "2", ")", "possible", "anchor", "boxes", ".", "We", "demonstrate", "the", "effectiveness", "of", "CornerNet", "on", "MS", "COCO", "[", "reference", "]", ".", "CornerNet", "achieves", "a", "42.2", "%", "AP", ",", "outperforming", "all", "existing", "one", "-", "stage", "detectors", ".", "In", "addition", ",", "through", "ablation", "studies", "we", "show", "that", "corner", "pooling", "is", "critical", "to", "the", "superior", "performance", "of", "CornerNet", ".", "Code", "is", "available", "at", "https:", "//", "github.com", "/", "princeton", "-", "vl", "/", "CornerNet", ".", "section", ":", "Related", "Works", "section", ":", "Two", "-", "stage", "object", "detectors", "Two", "-", "stage", "approach", "was", "first", "introduced", "and", "popularized", "by", "R", "-", "CNN", "[", "reference", "]", ".", "Two", "-", "stage", "detectors", "generate", "a", "sparse", "set", "of", "regions", "of", "interest", "(", "RoIs", ")", "and", "classify", "each", "of", "them", "by", "a", "network", ".", "R", "-", "CNN", "generates", "RoIs", "using", "a", "low", "level", "vision", "algorithm", "[", "reference", "][", "reference", "]", ".", "Each", "region", "is", "then", "extracted", "from", "the", "image", "and", "processed", "by", "a", "ConvNet", "independently", ",", "which", "creates", "lots", "of", "redundant", "computations", ".", "Later", ",", "SPP", "[", "reference", "]", "and", "Fast", "-", "RCNN", "[", "reference", "]", "improve", "R", "-", "CNN", "by", "designing", "a", "special", "pooling", "layer", "that", "pools", "each", "region", "from", "feature", "maps", "instead", ".", "However", ",", "both", "still", "rely", "on", "separate", "proposal", "algorithms", "and", "can", "not", "be", "trained", "end", "-", "to", "-", "end", ".", "Faster", "-", "RCNN", "[", "reference", "]", "does", "away", "low", "level", "proposal", "algorithms", "by", "introducing", "a", "region", "proposal", "network", "(", "RPN", ")", ",", "which", "generates", "proposals", "from", "a", "set", "of", "pre", "-", "determined", "candidate", "boxes", ",", "usually", "known", "as", "anchor", "boxes", ".", "This", "not", "only", "makes", "the", "detectors", "more", "efficient", "but", "also", "allows", "the", "detectors", "to", "be", "trained", "end", "-", "toend", ".", "R", "-", "FCN", "[", "reference", "]", "further", "improves", "the", "efficiency", "of", "Faster", "-", "RCNN", "by", "replacing", "the", "fully", "connected", "sub", "-", "detection", "network", "with", "a", "fully", "convolutional", "subdetection", "network", ".", "Other", "works", "focus", "on", "incorporating", "sub", "-", "category", "information", "[", "reference", "]", ",", "generating", "object", "proposals", "at", "multiple", "scales", "with", "more", "contextual", "information", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ",", "selecting", "better", "features", "[", "reference", "]", ",", "improving", "speed", ",", "cascade", "procedure", "[", "reference", "]", "and", "better", "training", "procedure", ".", "section", ":", "One", "-", "stage", "object", "detectors", "On", "the", "other", "hand", ",", "YOLO", "and", "SSD", "[", "reference", "]", "have", "popularized", "the", "one", "-", "stage", "approach", ",", "which", "removes", "the", "RoI", "pooling", "step", "and", "detects", "objects", "in", "a", "single", "network", ".", "One", "-", "stage", "detectors", "are", "usually", "more", "computationally", "efficient", "than", "twostage", "detectors", "while", "maintaining", "competitive", "performance", "on", "different", "challenging", "benchmarks", ".", "SSD", "places", "anchor", "boxes", "densely", "over", "feature", "maps", "from", "multiple", "scales", ",", "directly", "classifies", "and", "refines", "each", "anchor", "box", ".", "YOLO", "predicts", "bounding", "box", "coordinates", "directly", "from", "an", "image", ",", "and", "is", "later", "improved", "in", "YOLO9000", "by", "switching", "to", "anchor", "boxes", ".", "DSSD", "and", "RON", "[", "reference", "]", "adopt", "networks", "similar", "to", "the", "hourglass", "network", "[", "reference", "]", ",", "enabling", "them", "to", "combine", "low", "-", "level", "and", "high", "-", "level", "features", "via", "skip", "connections", "to", "predict", "bounding", "boxes", "more", "accurately", ".", "However", ",", "these", "one", "-", "stage", "detectors", "are", "still", "outperformed", "by", "the", "two", "-", "stage", "detectors", "until", "the", "introduction", "of", "RetinaNet", ".", "In", ",", "the", "authors", "suggest", "that", "the", "dense", "anchor", "boxes", "create", "a", "huge", "imbalance", "between", "positive", "and", "negative", "anchor", "boxes", "during", "training", ".", "This", "imbalance", "causes", "the", "training", "to", "be", "inefficient", "and", "hence", "the", "performance", "to", "be", "suboptimal", ".", "They", "propose", "a", "new", "loss", ",", "Focal", "Loss", ",", "to", "dynamically", "adjust", "the", "weights", "of", "each", "anchor", "box", "and", "show", "that", "their", "onestage", "detector", "can", "outperform", "the", "two", "-", "stage", "detectors", ".", "RefineDet", "proposes", "to", "filter", "the", "an", "-", "chor", "boxes", "to", "reduce", "the", "number", "of", "negative", "boxes", ",", "and", "to", "coarsely", "adjust", "the", "anchor", "boxes", ".", "DeNet", "[", "reference", "]", ")", "is", "a", "two", "-", "stage", "detector", "which", "generates", "RoIs", "without", "using", "anchor", "boxes", ".", "It", "first", "determines", "how", "likely", "each", "location", "belongs", "to", "either", "the", "top", "-", "left", ",", "top", "-", "right", ",", "bottomleft", "or", "bottom", "-", "right", "corner", "of", "a", "bounding", "box", ".", "It", "then", "generates", "RoIs", "by", "enumerating", "all", "possible", "corner", "combinations", ",", "and", "follows", "the", "standard", "two", "-", "stage", "approach", "to", "classify", "each", "RoI.", "Our", "approach", "is", "very", "different", "from", "DeNet", ".", "First", ",", "DeNet", "does", "not", "identify", "if", "two", "corners", "are", "from", "the", "same", "objects", "and", "relies", "on", "a", "sub", "-", "detection", "network", "to", "reject", "poor", "RoIs", ".", "In", "contrast", ",", "our", "approach", "is", "a", "one", "-", "stage", "approach", "which", "detects", "and", "groups", "the", "corners", "using", "a", "single", "ConvNet", ".", "Second", ",", "DeNet", "selects", "features", "at", "manually", "determined", "locations", "relative", "to", "a", "region", "for", "classification", ",", "while", "our", "approach", "does", "not", "require", "any", "feature", "selection", "step", ".", "Third", ",", "we", "introduce", "corner", "pooling", ",", "a", "novel", "type", "of", "layer", "to", "enhance", "corner", "detection", ".", "Point", "Linking", "Network", "(", "PLN", ")", "is", "an", "one", "-", "stage", "detector", "without", "anchor", "boxes", ".", "It", "first", "predicts", "the", "locations", "of", "the", "four", "corners", "and", "the", "center", "of", "a", "bounding", "box", ".", "Then", ",", "at", "each", "corner", "location", ",", "it", "predicts", "how", "likely", "each", "pixel", "location", "in", "the", "image", "is", "the", "center", ".", "Similarly", ",", "at", "the", "center", "location", ",", "it", "predicts", "how", "likely", "each", "pixel", "location", "belongs", "to", "either", "the", "top", "-", "left", ",", "top", "-", "right", ",", "bottom", "-", "left", "or", "bottom", "-", "right", "corner", ".", "It", "combines", "the", "predictions", "from", "each", "corner", "and", "center", "pair", "to", "generate", "a", "bounding", "box", ".", "Finally", ",", "it", "merges", "the", "four", "bounding", "boxes", "to", "give", "a", "bounding", "box", ".", "CornerNet", "is", "very", "different", "from", "PLN", ".", "First", ",", "CornerNet", "groups", "the", "corners", "by", "predicting", "embedding", "vectors", ",", "while", "PLN", "groups", "the", "corner", "and", "center", "by", "predicting", "pixel", "locations", ".", "Second", ",", "CornerNet", "uses", "corner", "pooling", "to", "better", "localize", "the", "corners", ".", "Our", "approach", "is", "inspired", "by", "on", "Associative", "Embedding", "in", "the", "context", "of", "multi", "-", "person", "pose", "estimation", ".", "Newell", "et", "al", ".", "propose", "an", "approach", "that", "detects", "and", "groups", "human", "joints", "in", "a", "single", "network", ".", "In", "their", "approach", "each", "detected", "human", "joint", "has", "an", "embedding", "vector", ".", "The", "joints", "are", "grouped", "based", "on", "the", "distances", "between", "their", "embeddings", ".", "To", "the", "best", "of", "our", "knowledge", ",", "we", "are", "the", "first", "to", "formulate", "the", "task", "of", "object", "detection", "as", "a", "task", "of", "detecting", "and", "grouping", "corners", "with", "embeddings", ".", "Another", "novelty", "of", "ours", "is", "the", "corner", "pooling", "layers", "that", "help", "better", "localize", "the", "corners", ".", "We", "also", "significantly", "modify", "the", "hourglass", "architecture", "and", "add", "our", "novel", "variant", "of", "focal", "loss", "to", "help", "better", "train", "the", "network", ".", "section", ":", "CornerNet", "section", ":", "Overview", "In", "CornerNet", ",", "we", "detect", "an", "object", "as", "a", "pair", "of", "keypointsthe", "top", "-", "left", "corner", "and", "bottom", "-", "right", "corner", "of", "the", "bounding", "box", ".", "A", "convolutional", "network", "predicts", "two", "sets", "of", "heatmaps", "to", "represent", "the", "locations", "of", "corners", "of", "different", "object", "categories", ",", "one", "set", "for", "the", "top", "-", "left", "corners", "and", "the", "other", "for", "the", "bottom", "-", "right", "corners", ".", "The", "network", "also", "predicts", "an", "embedding", "vector", "for", "each", "detected", "corner", "such", "that", "the", "distance", "between", "the", "embeddings", "of", "two", "corners", "from", "the", "same", "object", "is", "small", ".", "To", "produce", "tighter", "bounding", "boxes", ",", "the", "network", "also", "predicts", "offsets", "to", "slightly", "adjust", "the", "locations", "of", "the", "corners", ".", "With", "the", "predicted", "heatmaps", ",", "embeddings", "and", "offsets", ",", "we", "apply", "a", "simple", "post", "-", "processing", "algorithm", "to", "obtain", "the", "final", "bounding", "boxes", ".", "Fig", ".", "4", "provides", "an", "overview", "of", "CornerNet", ".", "We", "use", "the", "hourglass", "network", "[", "reference", "]", "as", "the", "backbone", "network", "of", "CornerNet", ".", "The", "hourglass", "network", "is", "followed", "by", "two", "prediction", "modules", ".", "One", "module", "is", "for", "the", "top", "-", "left", "corners", ",", "while", "the", "other", "one", "is", "for", "the", "bottomright", "corners", ".", "Each", "module", "has", "its", "own", "corner", "pooling", "module", "to", "pool", "features", "from", "the", "hourglass", "network", "before", "predicting", "the", "heatmaps", ",", "embeddings", "and", "offsets", ".", "Unlike", "many", "other", "object", "detectors", ",", "we", "do", "not", "use", "features", "from", "different", "scales", "to", "detect", "objects", "of", "different", "sizes", ".", "We", "only", "apply", "both", "modules", "to", "the", "output", "of", "the", "hourglass", "network", ".", "section", ":", "Detecting", "Corners", "We", "predict", "two", "sets", "of", "heatmaps", ",", "one", "for", "top", "-", "left", "corners", "and", "one", "for", "bottom", "-", "right", "corners", ".", "Each", "set", "of", "heatmaps", "has", "C", "channels", ",", "where", "C", "is", "the", "number", "of", "categories", ",", "and", "is", "of", "size", "H", "×", "W", ".", "There", "is", "no", "background", "channel", ".", "Each", "channel", "is", "a", "binary", "mask", "indicating", "the", "locations", "of", "the", "corners", "for", "a", "class", ".", "For", "each", "corner", ",", "there", "is", "one", "ground", "-", "truth", "positive", "location", ",", "and", "all", "other", "locations", "are", "negative", ".", "During", "training", ",", "instead", "of", "equally", "penalizing", "negative", "locations", ",", "we", "reduce", "the", "penalty", "given", "to", "negative", "locations", "within", "a", "radius", "of", "the", "positive", "location", ".", "This", "is", "because", "a", "pair", "of", "false", "corner", "detections", ",", "if", "they", "are", "close", "to", "their", "respective", "ground", "truth", "locations", ",", "can", "still", "produce", "a", "box", "that", "sufficiently", "overlaps", "the", "ground", "-", "truth", "box", "(", "Fig", ".", "5", ")", ".", "We", "determine", "the", "radius", "by", "the", "size", "of", "an", "object", "by", "ensuring", "that", "a", "pair", "of", "points", "within", "the", "radius", "would", "generate", "a", "bounding", "box", "with", "at", "least", "t", "IoU", "with", "the", "ground", "-", "truth", "annotation", "(", "we", "set", "t", "to", "0.3", "in", "all", "experiments", ")", ".", "Given", "the", "radius", ",", "the", "amount", "of", "penalty", "reduction", "is", "given", "by", "an", "unnormalized", "2D", "Gaussian", ",", "Fig", ".", "4", "Overview", "of", "CornerNet", ".", "The", "backbone", "network", "is", "followed", "by", "two", "prediction", "modules", ",", "one", "for", "the", "top", "-", "left", "corners", "and", "the", "other", "for", "the", "bottom", "-", "right", "corners", ".", "Using", "the", "predictions", "from", "both", "modules", ",", "we", "locate", "and", "group", "the", "corners", ".", "e", "−", "x", "2", "+", "y", "2", "2σ", "2", ",", "whose", "center", "is", "at", "the", "positive", "location", "and", "whose", "σ", "is", "1", "/", "3", "of", "the", "radius", ".", "Let", "p", "cij", "be", "the", "score", "at", "location", "(", "i", ",", "j", ")", "for", "class", "c", "in", "the", "predicted", "heatmaps", ",", "and", "let", "y", "cij", "be", "the", "\"", "groundtruth", "\"", "heatmap", "augmented", "with", "the", "unnormalized", "Gaussians", ".", "We", "design", "a", "variant", "of", "focal", "loss", ":", "where", "N", "is", "the", "number", "of", "objects", "in", "an", "image", ",", "and", "α", "and", "β", "are", "the", "hyper", "-", "parameters", "which", "control", "the", "contribution", "of", "each", "point", "(", "we", "set", "α", "to", "2", "and", "β", "to", "4", "in", "all", "experiments", ")", ".", "With", "the", "Gaussian", "bumps", "encoded", "in", "y", "cij", ",", "the", "(", "1", "−", "y", "cij", ")", "term", "reduces", "the", "penalty", "around", "the", "ground", "truth", "locations", ".", "Many", "networks", "[", "reference", "]", ")", "involve", "downsampling", "layers", "to", "gather", "global", "information", "and", "to", "reduce", "memory", "usage", ".", "When", "they", "are", "applied", "to", "an", "image", "fully", "convolutionally", ",", "the", "size", "of", "the", "output", "is", "usually", "smaller", "than", "the", "image", ".", "Hence", ",", "a", "location", "(", "x", ",", "y", ")", "in", "the", "image", "is", "mapped", "to", "the", "location", "x", "n", ",", "y", "n", "in", "the", "heatmaps", ",", "where", "n", "is", "the", "downsampling", "factor", ".", "When", "we", "remap", "the", "locations", "from", "the", "heatmaps", "to", "the", "input", "image", ",", "some", "precision", "may", "be", "lost", ",", "which", "can", "greatly", "affect", "the", "IoU", "of", "small", "bounding", "boxes", "with", "their", "ground", "truths", ".", "To", "address", "this", "issue", "we", "predict", "location", "offsets", "to", "slightly", "adjust", "the", "corner", "locations", "before", "remapping", "them", "to", "the", "input", "resolution", ".", "where", "o", "k", "is", "the", "offset", ",", "x", "k", "and", "y", "k", "are", "the", "x", "and", "y", "coordinate", "for", "corner", "k.", "In", "particular", ",", "we", "predict", "one", "set", "of", "offsets", "shared", "by", "the", "top", "-", "left", "corners", "of", "all", "categories", ",", "and", "another", "set", "shared", "by", "the", "bottom", "-", "right", "corners", ".", "For", "training", ",", "we", "apply", "the", "smooth", "L1", "Loss", "[", "reference", "]", "at", "ground", "-", "truth", "corner", "locations", ":", "section", ":", "Grouping", "Corners", "Multiple", "objects", "may", "appear", "in", "an", "image", ",", "and", "thus", "multiple", "top", "-", "left", "and", "bottom", "-", "right", "corners", "may", "be", "detected", ".", "We", "need", "to", "determine", "if", "a", "pair", "of", "the", "top", "-", "left", "corner", "and", "bottom", "-", "right", "corner", "is", "from", "the", "same", "bounding", "box", ".", "Our", "approach", "is", "inspired", "by", "the", "Associative", "Embedding", "method", "proposed", "by", "for", "the", "task", "of", "multi", "-", "person", "pose", "estimation", ".", "Newell", "et", "al", ".", "detect", "all", "human", "joints", "and", "generate", "an", "embedding", "for", "each", "detected", "joint", ".", "They", "group", "the", "joints", "based", "on", "the", "distances", "between", "the", "embeddings", ".", "The", "idea", "of", "associative", "embedding", "is", "also", "applicable", "to", "our", "task", ".", "The", "network", "predicts", "an", "embedding", "vector", "for", "each", "detected", "corner", "such", "that", "if", "a", "top", "-", "left", "corner", "and", "a", "bottom", "-", "right", "corner", "belong", "to", "the", "same", "bounding", "box", ",", "the", "distance", "between", "their", "embeddings", "should", "be", "small", ".", "We", "can", "then", "group", "the", "corners", "based", "on", "the", "distances", "between", "the", "embeddings", "of", "the", "top", "-", "left", "and", "bottom", "-", "right", "corners", ".", "The", "actual", "values", "of", "the", "embeddings", "are", "unimportant", ".", "Only", "the", "distances", "between", "the", "embeddings", "are", "used", "to", "group", "the", "corners", ".", "We", "follow", "and", "use", "embeddings", "of", "1", "dimension", ".", "Let", "e", "t", "k", "be", "the", "embedding", "for", "the", "top", "-", "left", "corner", "of", "object", "k", "and", "e", "b", "k", "for", "the", "bottom", "-", "right", "corner", ".", "As", "in", ",", "we", "use", "the", "\"", "pull", "\"", "loss", "to", "train", "the", "network", "to", "group", "the", "corners", "and", "the", "\"", "push", "\"", "loss", "to", "separate", "the", "corners", ":", "where", "e", "k", "is", "the", "average", "of", "e", "t", "k", "and", "e", "b", "k", "and", "we", "set", "∆", "to", "be", "1", "in", "all", "our", "experiments", ".", "Similar", "to", "the", "offset", "loss", ",", "we", "only", "apply", "the", "losses", "at", "the", "ground", "-", "truth", "corner", "location", ".", "section", ":", "Corner", "Pooling", "As", "shown", "in", "Fig", ".", "2", ",", "there", "is", "often", "no", "local", "visual", "evidence", "for", "the", "presence", "of", "corners", ".", "To", "determine", "if", "a", "pixel", "is", "a", "top", "-", "left", "corner", ",", "we", "need", "to", "look", "horizontally", "towards", "the", "right", "for", "the", "topmost", "boundary", "of", "an", "object", "and", "vertically", "towards", "the", "bottom", "for", "the", "leftmost", "boundary", ".", "We", "thus", "propose", "corner", "pooling", "to", "better", "localize", "the", "corners", "by", "encoding", "explicit", "prior", "knowledge", ".", "Suppose", "we", "want", "to", "determine", "if", "a", "pixel", "at", "location", "(", "i", ",", "j", ")", "is", "a", "top", "-", "left", "corner", ".", "Let", "f", "t", "and", "f", "l", "be", "the", "feature", "maps", "that", "are", "the", "inputs", "to", "the", "top", "-", "left", "corner", "pooling", "layer", ",", "and", "let", "f", "tij", "and", "f", "lij", "be", "the", "vectors", "at", "location", "(", "i", ",", "j", ")", "in", "f", "t", "and", "f", "l", "respectively", ".", "With", "H", "×", "W", "feature", "maps", ",", "the", "corner", "pooling", "layer", "first", "max", "-", "pools", "all", "feature", "vectors", "between", "(", "i", ",", "j", ")", "and", "(", "i", ",", "H", ")", "in", "f", "t", "to", "a", "feature", "vector", "t", "ij", ",", "and", "max", "-", "pools", "all", "feature", "vectors", "between", "(", "i", ",", "j", ")", "and", "(", "W", ",", "j", ")", "in", "f", "l", "to", "a", "feature", "vector", "l", "ij", ".", "Finally", ",", "it", "adds", "t", "ij", "and", "l", "ij", "together", ".", "This", "computation", "can", "be", "expressed", "by", "the", "following", "equations", ":", "where", "we", "apply", "an", "elementwise", "max", "operation", ".", "Both", "t", "ij", "and", "l", "ij", "can", "be", "computed", "efficiently", "by", "dynamic", "programming", "as", "shown", "Fig", ".", "8", ".", "We", "define", "bottom", "-", "right", "corner", "pooling", "layer", "in", "a", "similar", "way", ".", "It", "max", "-", "pools", "all", "feature", "vectors", "between", "(", "0", ",", "j", ")", "and", "(", "i", ",", "j", ")", ",", "and", "all", "feature", "vectors", "between", "(", "i", ",", "0", ")", "and", "(", "i", ",", "j", ")", "before", "adding", "the", "pooled", "results", ".", "The", "corner", "pooling", "layers", "are", "used", "in", "the", "prediction", "modules", "to", "predict", "heatmaps", ",", "embeddings", "and", "offsets", ".", "The", "architecture", "of", "the", "prediction", "module", "is", "shown", "in", "Fig", ".", "7", ".", "The", "first", "part", "of", "the", "module", "is", "a", "modified", "version", "of", "the", "residual", "block", ".", "In", "this", "modified", "residual", "block", ",", "we", "replace", "the", "first", "3", "×", "3", "convolution", "module", "with", "a", "corner", "pooling", "module", ",", "which", "first", "processes", "the", "features", "from", "the", "backbone", "network", "by", "two", "3", "×", "3", "convolution", "modules", "1", "with", "128", "channels", "and", "then", "applies", "a", "corner", "pooling", "layer", ".", "Following", "the", "design", "of", "a", "residual", "block", ",", "we", "then", "feed", "the", "pooled", "features", "into", "a", "3", "×", "3", "Conv", "-", "BN", "layer", "with", "256", "channels", "and", "add", "back", "the", "projection", "shortcut", ".", "The", "modified", "residual", "block", "is", "followed", "by", "a", "3×3", "convolution", "module", "with", "256", "channels", ",", "and", "3", "Conv", "-", "ReLU", "-", "Conv", "layers", "to", "produce", "the", "heatmaps", ",", "embeddings", "and", "offsets", ".", "section", ":", "Hourglass", "Network", "CornerNet", "uses", "the", "hourglass", "network", "[", "reference", "]", "as", "its", "backbone", "network", ".", "The", "hourglass", "network", "was", "first", "introduced", "for", "the", "human", "pose", "estimation", "task", ".", "It", "is", "a", "fully", "convolutional", "neural", "network", "that", "consists", "of", "one", "or", "more", "hourglass", "modules", ".", "An", "hourglass", "module", "first", "downsamples", "the", "input", "features", "by", "a", "series", "of", "convolution", "and", "max", "pooling", "layers", ".", "It", "then", "upsamples", "the", "features", "back", "to", "the", "original", "resolution", "by", "a", "series", "of", "upsampling", "and", "convolution", "layers", ".", "Since", "details", "are", "lost", "in", "the", "max", "pooling", "layers", ",", "skip", "layers", "are", "added", "to", "bring", "back", "the", "details", "to", "the", "upsampled", "features", ".", "The", "hourglass", "module", "captures", "both", "global", "and", "local", "features", "in", "a", "single", "unified", "structure", ".", "When", "multiple", "hourglass", "modules", "are", "stacked", "in", "the", "network", ",", "the", "hourglass", "modules", "can", "reprocess", "the", "features", "to", "capture", "higher", "-", "level", "of", "information", ".", "These", "properties", "make", "the", "hourglass", "network", "an", "ideal", "choice", "for", "object", "detection", "as", "well", ".", "In", "fact", ",", "many", "current", "detectors", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ")", "already", "adopted", "networks", "similar", "to", "the", "hourglass", "network", ".", "Our", "hourglass", "network", "consists", "of", "two", "hourglasses", ",", "and", "we", "make", "some", "modifications", "to", "the", "architecture", "of", "the", "hourglass", "module", ".", "Instead", "of", "using", "max", "pool", "-", "9", "10", "Fig", ".", "6", "The", "top", "-", "left", "corner", "pooling", "layer", "can", "be", "implemented", "very", "efficiently", ".", "We", "scan", "from", "right", "to", "left", "for", "the", "horizontal", "max", "-", "pooling", "and", "from", "bottom", "to", "top", "for", "the", "vertical", "max", "-", "pooling", ".", "We", "then", "add", "two", "max", "-", "pooled", "feature", "maps", ".", "Fig", ".", "7", "The", "prediction", "module", "starts", "with", "a", "modified", "residual", "block", ",", "in", "which", "we", "replace", "the", "first", "convolution", "module", "with", "our", "corner", "pooling", "module", ".", "The", "modified", "residual", "block", "is", "then", "followed", "by", "a", "convolution", "module", ".", "We", "have", "multiple", "branches", "for", "predicting", "the", "heatmaps", ",", "embeddings", "and", "offsets", ".", "ing", ",", "we", "simply", "use", "stride", "2", "to", "reduce", "feature", "resolution", ".", "We", "reduce", "feature", "resolutions", "5", "times", "and", "increase", "the", "number", "of", "feature", "channels", "along", "the", "way", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "When", "we", "upsample", "the", "features", ",", "we", "apply", "2", "residual", "modules", "followed", "by", "a", "nearest", "neighbor", "upsampling", ".", "Every", "skip", "connection", "also", "consists", "of", "2", "residual", "modules", ".", "There", "are", "4", "residual", "modules", "with", "512", "channels", "in", "the", "middle", "of", "an", "hourglass", "module", ".", "Before", "the", "hourglass", "modules", ",", "we", "reduce", "the", "image", "resolution", "by", "4", "times", "using", "a", "7", "×", "7", "convolution", "module", "with", "stride", "2", "and", "128", "channels", "followed", "by", "a", "residual", "block", "with", "stride", "2", "and", "256", "channels", ".", "Following", "[", "reference", "]", ",", "we", "also", "add", "intermediate", "supervision", "in", "training", ".", "However", ",", "we", "do", "not", "add", "back", "the", "intermediate", "predictions", "to", "the", "network", "as", "we", "find", "that", "this", "hurts", "the", "performance", "of", "the", "network", ".", "We", "apply", "a", "1", "×", "1", "Conv", "-", "BN", "module", "to", "both", "the", "input", "and", "output", "of", "the", "first", "hourglass", "module", ".", "We", "then", "merge", "them", "by", "element", "-", "wise", "addition", "followed", "by", "a", "ReLU", "and", "a", "residual", "block", "with", "256", "channels", ",", "which", "is", "then", "used", "as", "the", "input", "to", "the", "second", "hourglass", "module", ".", "The", "depth", "of", "the", "hourglass", "network", "is", "104", ".", "Unlike", "many", "other", "stateof", "-", "the", "-", "art", "detectors", ",", "we", "only", "use", "the", "features", "from", "the", "last", "layer", "of", "the", "whole", "network", "to", "make", "predictions", ".", "section", ":", "Experiments", "section", ":", "Training", "Details", "We", "implement", "CornerNet", "in", "PyTorch", "[", "reference", "]", ".", "The", "network", "is", "randomly", "initialized", "under", "the", "default", "setting", "of", "PyTorch", "with", "no", "pretraining", "on", "any", "external", "dataset", ".", "As", "we", "apply", "focal", "loss", ",", "we", "follow", "to", "set", "the", "biases", "in", "the", "convolution", "layers", "that", "predict", "the", "corner", "heatmaps", ".", "During", "training", ",", "we", "set", "the", "input", "resolution", "of", "the", "network", "to", "511", "×", "511", ",", "which", "leads", "to", "an", "output", "resolution", "of", "128", "×", "128", ".", "To", "reduce", "overfitting", ",", "we", "adopt", "standard", "data", "augmentation", "techniques", "including", "random", "horizontal", "flipping", ",", "random", "scaling", ",", "random", "cropping", "and", "random", "color", "jittering", ",", "which", "includes", "adjusting", "the", "brightness", ",", "saturation", "and", "contrast", "of", "an", "image", ".", "Finally", ",", "we", "apply", "PCA", "[", "reference", "]", "to", "the", "input", "image", ".", "We", "use", "Adam", "[", "reference", "]", "to", "optimize", "the", "full", "training", "loss", ":", "where", "α", ",", "β", "and", "γ", "are", "the", "weights", "for", "the", "pull", ",", "push", "and", "offset", "loss", "respectively", ".", "We", "set", "both", "α", "and", "β", "to", "0.1", "and", "γ", "to", "1", ".", "We", "find", "that", "1", "or", "larger", "values", "of", "α", "and", "β", "lead", "to", "poor", "performance", ".", "We", "use", "a", "batch", "size", "of", "49", "and", "train", "the", "network", "on", "10", "Titan", "X", "(", "PASCAL", ")", "GPUs", "(", "4", "images", "on", "the", "master", "GPU", ",", "5", "images", "per", "GPU", "for", "the", "rest", "of", "the", "GPUs", ")", ".", "To", "conserve", "GPU", "resources", ",", "in", "our", "ablation", "experiments", ",", "we", "train", "the", "networks", "for", "250k", "iterations", "with", "a", "learning", "rate", "of", "2.5", "×", "10", "−4", ".", "When", "we", "compare", "our", "results", "with", "other", "detectors", ",", "we", "train", "the", "networks", "for", "an", "extra", "250k", "iterations", "and", "reduce", "the", "learning", "rate", "to", "2.5", "×", "10", "−5", "for", "the", "last", "50k", "iterations", ".", "section", ":", "Testing", "Details", "During", "testing", ",", "we", "use", "a", "simple", "post", "-", "processing", "algorithm", "to", "generate", "bounding", "boxes", "from", "the", "heatmaps", ",", "embeddings", "and", "offsets", ".", "We", "first", "apply", "non", "-", "maximal", "suppression", "(", "NMS", ")", "by", "using", "a", "3×3", "max", "pooling", "layer", "on", "the", "corner", "heatmaps", ".", "Then", "we", "pick", "the", "top", "100", "top", "-", "left", "and", "top", "100", "bottom", "-", "right", "corners", "from", "the", "heatmaps", ".", "The", "corner", "locations", "are", "adjusted", "by", "the", "corresponding", "offsets", ".", "We", "calculate", "the", "L1", "distances", "between", "the", "embeddings", "of", "the", "top", "-", "left", "and", "bottom", "-", "right", "corners", ".", "Pairs", "that", "have", "distances", "greater", "than", "0.5", "or", "contain", "corners", "from", "different", "categories", "are", "rejected", ".", "The", "average", "scores", "of", "the", "top", "-", "left", "and", "bottom", "-", "right", "corners", "are", "used", "as", "the", "detection", "scores", ".", "Instead", "of", "resizing", "an", "image", "to", "a", "fixed", "size", ",", "we", "maintain", "the", "original", "resolution", "of", "the", "image", "and", "pad", "it", "with", "zeros", "before", "feeding", "it", "to", "CornerNet", ".", "Both", "the", "original", "and", "flipped", "images", "are", "used", "for", "testing", ".", "We", "combine", "the", "detections", "from", "the", "original", "and", "flipped", "images", ",", "and", "apply", "soft", "-", "nms", "[", "reference", "]", "to", "suppress", "redundant", "detections", ".", "Only", "the", "top", "100", "detections", "are", "reported", ".", "The", "average", "inference", "time", "is", "244ms", "per", "image", "on", "a", "Titan", "X", "(", "PASCAL", ")", "GPU", ".", "section", ":", "MS", "COCO", "We", "evaluate", "CornerNet", "on", "the", "very", "challenging", "MS", "COCO", "dataset", "[", "reference", "]", ".", "MS", "COCO", "contains", "80k", "images", "for", "training", ",", "40k", "for", "validation", "and", "20k", "for", "testing", ".", "All", "images", "in", "the", "training", "set", "and", "35k", "images", "in", "the", "validation", "set", "are", "used", "for", "training", ".", "The", "remaining", "5k", "images", "in", "validation", "set", "are", "used", "for", "hyper", "-", "parameter", "searching", "and", "ablation", "study", ".", "All", "results", "on", "the", "test", "set", "are", "submitted", "to", "an", "external", "server", "for", "evaluation", ".", "To", "provide", "fair", "comparisons", "with", "other", "detectors", ",", "we", "report", "our", "main", "results", "on", "the", "test", "-", "dev", "set", ".", "MS", "COCO", "uses", "average", "precisions", "(", "APs", ")", "at", "different", "IoUs", "and", "APs", "for", "different", "object", "sizes", "as", "the", "main", "evaluation", "metrics", ".", "section", ":", "Ablation", "Study", "section", ":", "Corner", "Pooling", "Corner", "pooling", "is", "a", "key", "component", "of", "CornerNet", ".", "To", "understand", "its", "contribution", "to", "performance", ",", "we", "train", "another", "network", "without", "corner", "pooling", "but", "with", "the", "same", "number", "of", "parameters", ".", "Tab", ".", "1", "shows", "that", "adding", "corner", "pooling", "gives", "significant", "improvement", ":", "2.0", "%", "on", "AP", ",", "2.1", "%", "on", "AP", "50", "and", "2.1", "%", "on", "AP", "75", ".", "We", "also", "see", "that", "corner", "pooling", "is", "especially", "helpful", "for", "medium", "and", "large", "objects", ",", "improving", "their", "APs", "by", "2.4", "%", "and", "3.6", "%", "respectively", ".", "This", "is", "expected", "because", "the", "topmost", ",", "bottommost", ",", "leftmost", ",", "rightmost", "boundaries", "of", "medium", "and", "large", "objects", "are", "likely", "to", "be", "further", "away", "from", "the", "corner", "locations", ".", "Fig", ".", "8", "shows", "four", "qualitative", "examples", "with", "and", "without", "corner", "pooling", ".", "section", ":", "Stability", "of", "Corner", "Pooling", "over", "Larger", "Area", "Corner", "pooling", "pools", "over", "different", "sizes", "of", "area", "in", "different", "quadrants", "of", "an", "image", ".", "For", "example", ",", "the", "top", "-", "left", "corner", "pooling", "pools", "over", "larger", "areas", "both", "horizontally", "and", "vertically", "in", "the", "upper", "-", "left", "quadrant", "of", "an", "image", ",", "compared", "to", "the", "lower", "-", "right", "quadrant", ".", "Therefore", ",", "the", "location", "of", "a", "corner", "may", "affect", "the", "stability", "of", "the", "corner", "pooling", ".", "We", "evaluate", "the", "performance", "of", "our", "network", "on", "detecting", "both", "the", "top", "-", "left", "and", "bottom", "-", "right", "corners", "in", "different", "quadrants", "of", "an", "image", ".", "Detecting", "corners", "can", "be", "seen", "as", "a", "binary", "classification", "task", "i.e.", "the", "groundtruth", "location", "of", "a", "corner", "is", "positive", ",", "and", "any", "location", "outside", "of", "a", "small", "radius", "of", "the", "corner", "is", "negative", ".", "We", "measure", "the", "performance", "using", "mAPs", "over", "all", "categories", "on", "the", "MS", "COCO", "validation", "set", ".", "Tab", ".", "3", "shows", "that", "without", "corner", "pooling", ",", "the", "topleft", "corner", "mAPs", "of", "upper", "-", "left", "and", "lower", "-", "right", "quadrant", "are", "66.1", "%", "and", "60.8", "%", "respectively", ".", "Top", "-", "left", "corner", "pooling", "improves", "the", "mAPs", "by", "3.1", "%", "(", "to", "69.2", "%", ")", "and", "2.7", "%", "(", "to", "63.5", "%", ")", "respectively", ".", "Similarly", ",", "bottomright", "corner", "pooling", "improves", "the", "bottom", "-", "right", "corner", "mAPs", "of", "upper", "-", "left", "quadrant", "by", "2.8", "%", "(", "from", "53.4", "%", "to", "56.2", "%", ")", ",", "and", "lower", "-", "right", "quadrant", "by", "2.6", "%", "(", "from", "65.0", "%", "to", "67.6", "%", ")", ".", "Corner", "pooling", "gives", "similar", "improvement", "to", "corners", "at", "different", "quadrants", ",", "show", "that", "corner", "pooling", "is", "effective", "and", "stable", "over", "both", "small", "and", "large", "areas", ".", "section", ":", "Reducing", "Penalty", "to", "Negative", "Locations", "We", "reduce", "the", "penalty", "given", "to", "negative", "locations", "around", "a", "positive", "location", ",", "within", "a", "radius", "determined", "by", "the", "size", "of", "the", "object", "(", "Sec", ".", "3.2", ")", ".", "To", "understand", "how", "this", "helps", "train", "CornerNet", ",", "we", "train", "one", "network", "with", "no", "penalty", "reduction", "and", "another", "network", "with", "a", "fixed", "radius", "of", "2.5", ".", "We", "compare", "them", "with", "CornerNet", "on", "the", "validation", "set", ".", "Tab", ".", "2", "shows", "that", "a", "fixed", "radius", "improves", "AP", "over", "the", "baseline", "by", "2.7", "%", ",", "AP", "m", "by", "1.5", "%", "and", "AP", "l", "by", "5.3", "%", ".", "Object", "-", "dependent", "radius", "further", "improves", "the", "AP", "by", "2.8", "%", ",", "AP", "m", "by", "2.0", "%", "and", "AP", "l", "by", "5.8", "%", ".", "In", "addition", ",", "we", "see", "that", "the", "penalty", "reduction", "especially", "benefits", "medium", "and", "large", "objects", ".", "section", ":", "Hourglass", "Network", "CornerNet", "uses", "the", "hourglass", "network", "[", "reference", "]", "as", "its", "backbone", "network", ".", "Since", "the", "hourglass", "network", "is", "not", "commonly", "used", "in", "other", "state", "-", "of", "-", "the", "-", "art", "detectors", ",", "we", "perform", "an", "experiment", "to", "study", "the", "contribution", "of", "the", "hourglass", "network", "in", "CornerNet", ".", "We", "train", "a", "CornerNet", "in", "which", "we", "replace", "the", "hourglass", "network", "with", "FPN", "(", "w", "/", "ResNet", "-", "101", ")", ",", "which", "is", "more", "commonly", "used", "in", "state", "-", "of", "-", "the", "-", "art", "object", "detectors", ".", "We", "only", "use", "the", "final", "output", "of", "FPN", "for", "predictions", ".", "Meanwhile", ",", "we", "train", "an", "anchor", "box", "based", "detector", "which", "uses", "the", "hourglass", "network", "as", "its", "backbone", ".", "Each", "hourglass", "module", "predicts", "anchor", "boxes", "at", "multiple", "resolutions", "by", "using", "features", "at", "multiple", "scales", "during", "upsampling", "stage", ".", "We", "follow", "the", "anchor", "box", "design", "in", "RetinaNet", "and", "add", "intermediate", "supervisions", "during", "training", ".", "In", "both", "experiments", ",", "we", "initialize", "the", "networks", "from", "scratch", "and", "follow", "the", "same", "training", "procedure", "as", "we", "train", "CornerNet", "(", "Sec", ".", "4.1", ")", ".", "Tab", ".", "4", "shows", "that", "CornerNet", "with", "hourglass", "network", "outperforms", "CornerNet", "with", "FPN", "by", "8.2", "%", "AP", ",", "and", "the", "anchor", "box", "based", "detector", "with", "hourglass", "network", "by", "5.5", "%", "AP", ".", "The", "results", "suggest", "that", "the", "choice", "of", "the", "backbone", "network", "is", "important", "and", "the", "hourglass", "network", "is", "crucial", "to", "the", "performance", "of", "CornerNet", ".", "section", ":", "Quality", "of", "the", "Bounding", "Boxes", "A", "good", "detector", "should", "predict", "high", "quality", "bounding", "boxes", "that", "cover", "objects", "tightly", ".", "To", "understand", "the", "quality", "of", "the", "bounding", "boxes", "predicted", "by", "CornerNet", ",", "we", "evaluate", "the", "performance", "of", "CornerNet", "at", "multiple", "IoU", "thresholds", ",", "and", "compare", "the", "results", "with", "other", "state", "-", "of", "-", "the", "-", "art", "detectors", ",", "including", "RetinaNet", ",", "Cascade", "R", "-", "CNN", "[", "reference", "]", "and", "IoU", "-", "Net", "[", "reference", "]", ".", "Tab", ".", "5", "shows", "that", "CornerNet", "achieves", "a", "much", "higher", "AP", "at", "0.9", "IoU", "than", "other", "detectors", ",", "outperforming", "Cascade", "R", "-", "CNN", "+", "IoU", "-", "Net", "by", "3.9", "%", ",", "Cascade", "R", "-", "CNN", "by", "7.6", "%", "and", "RetinaNet", "2", "by", "7.3", "%", ".", "This", "suggests", "that", "Cor", "-", "2", "We", "use", "the", "best", "model", "publicly", "available", "on", "https:", "//", "github.com", "/", "facebookresearch", "/", "Detectron", "/", "blob", "/", "master", "/", "MODEL_ZOO.md", "nerNet", "is", "able", "to", "generate", "bounding", "boxes", "of", "higher", "quality", "compared", "to", "other", "state", "-", "of", "-", "the", "-", "art", "detectors", ".", "section", ":", "Error", "Analysis", "CornerNet", "simultaneously", "outputs", "heatmaps", ",", "offsets", ",", "and", "embeddings", ",", "all", "of", "which", "affect", "detection", "performance", ".", "An", "object", "will", "be", "missed", "if", "either", "corner", "is", "missed", ";", "precise", "offsets", "are", "needed", "to", "generate", "tight", "bounding", "boxes", ";", "incorrect", "embeddings", "will", "result", "in", "many", "false", "bounding", "boxes", ".", "To", "understand", "how", "each", "part", "contributes", "to", "the", "final", "error", ",", "we", "perform", "an", "error", "analysis", "by", "replacing", "the", "predicted", "heatmaps", "and", "offsets", "with", "the", "ground", "-", "truth", "values", "and", "evaluting", "performance", "on", "the", "validation", "set", ".", "Tab", ".", "6", "shows", "that", "using", "the", "ground", "-", "truth", "corner", "heatmaps", "alone", "improves", "the", "AP", "from", "38.4", "%", "to", "73.1", "%", ".", "section", ":", "AP", "s", ",", "AP", "m", "and", "AP", "l", "also", "increase", "by", "42.3", "%", ",", "40.7", "%", "and", "30.0", "%", "respectively", ".", "If", "we", "replace", "the", "predicted", "offsets", "with", "the", "ground", "-", "truth", "offsets", ",", "the", "AP", "further", "increases", "by", "13.0", "%", "to", "86.1", "%", ".", "This", "suggests", "that", "although", "there", "is", "still", "ample", "room", "for", "improvement", "in", "both", "detecting", "and", "grouping", "corners", ",", "the", "main", "bottleneck", "is", "detecting", "corners", ".", "Fig", ".", "9", "shows", "some", "qualitative", "examples", "where", "the", "corner", "locations", "or", "embeddings", "are", "incorrect", ".", "section", ":", "Comparisons", "with", "state", "-", "of", "-", "the", "-", "art", "detectors", "We", "compare", "CornerNet", "with", "other", "state", "-", "of", "-", "the", "-", "art", "detectors", "on", "MS", "COCO", "test", "-", "dev", "(", "Tab", ".", "7", ")", ".", "With", "multiscale", "evaluation", ",", "CornerNet", "achieves", "an", "AP", "of", "42.2", "%", ",", "the", "state", "of", "the", "art", "among", "existing", "one", "-", "stage", "methods", "and", "competitive", "with", "two", "-", "stage", "methods", ".", "section", ":", "Conclusion", "We", "have", "presented", "CornerNet", ",", "a", "new", "approach", "to", "object", "detection", "that", "detects", "bounding", "boxes", "as", "pairs", "of", "corners", ".", "We", "evaluate", "CornerNet", "on", "MS", "COCO", "and", "demonstrate", "competitive", "results", ".", "section", ":", "section", ":", "Acknowledgements", "This", "work", "is", "partially", "supported", "by", "a", "grant", "from", "Toyota", "Research", "Institute", "and", "a", "DARPA", "grant", "FA8750", "-", "18", "-", "2", "-", "0019", ".", "This", "article", "solely", "reflects", "the", "opinions", "and", "conclusions", "of", "its", "authors", ".", "section", ":"], "entity": [[0, 1, ["CornerNet"], "Method"], [2, 4, ["Detecting", "Objects"], "Task"], [12, 13, ["CornerNet"], "Method"], [18, 20, ["object", "detection"], "Task"], [48, 51, ["convolution", "neural", "network"], "Method"], [74, 78, ["single", "-", "stage", "detectors"], "Method"], [88, 90, ["corner", "pooling"], "Method"], [95, 97, ["pooling", "layer"], "Method"], [108, 109, ["CornerNet"], "Method"], [113, 114, ["AP"], "Metric"], [115, 117, ["MS", "COCO"], "Material"], [121, 125, ["one", "-", "stage", "detectors"], "Method"], [129, 131, ["Object", "detectors"], "Method"], [133, 136, ["convolutional", "neural", "networks"], "Method"], [137, 138, ["ConvNets"], "Method"], [210, 214, ["one", "-", "stage", "detectors"], "Method"], [231, 233, ["stage", "detectors"], "Method"], [272, 273, ["regression"], "Method"], [302, 303, ["DSSD"], "Method"], [381, 382, ["training"], "Task"], [418, 422, ["ad", "-", "hoc", "heuristics"], "Method"], [432, 434, ["multiscale", "architectures"], "Method"], [471, 472, ["CornerNet"], "Method"], [475, 477, ["onestage", "approach"], "Method"], [478, 480, ["object", "detection"], "Task"], [514, 516, ["convolutional", "network"], "Method"], [583, 586, ["ConvNet", "Embeddings", "Heatmaps"], "Method"], [612, 614, ["convolutional", "network"], "Method"], [682, 685, ["associative", "embedding", "method"], "Method"], [691, 693, ["group", "keypoints"], "Task"], [697, 702, ["multiperson", "human", "-", "pose", "estimation"], "Task"], [718, 719, ["CornerNet"], "Method"], [720, 722, ["corner", "pooling"], "Method"], [727, 729, ["pooling", "layer"], "Method"], [732, 734, ["convolutional", "network"], "Method"], [896, 898, ["detecting", "corners"], "Task"], [949, 951, ["corner", "pooling"], "Method"], [1006, 1007, ["CornerNet"], "Method"], [1008, 1010, ["MS", "COCO"], "Material"], [1014, 1015, ["CornerNet"], "Method"], [1019, 1020, ["AP"], "Metric"], [1024, 1028, ["one", "-", "stage", "detectors"], "Method"], [1038, 1040, ["corner", "pooling"], "Method"], [1047, 1048, ["CornerNet"], "Method"], [1061, 1062, ["CornerNet"], "Method"], [1072, 1074, ["object", "detectors"], "Method"], [1084, 1087, ["R", "-", "CNN"], "Method"], [1093, 1095, ["stage", "detectors"], "Method"], [1115, 1118, ["R", "-", "CNN"], "Method"], [1119, 1120, ["RoIs"], "Method"], [1122, 1126, ["low", "level", "vision", "algorithm"], "Method"], [1144, 1145, ["ConvNet"], "Method"], [1156, 1157, ["SPP"], "Method"], [1161, 1164, ["Fast", "-", "RCNN"], "Method"], [1168, 1171, ["R", "-", "CNN"], "Method"], [1175, 1177, ["pooling", "layer"], "Method"], [1193, 1195, ["proposal", "algorithms"], "Method"], [1206, 1209, ["Faster", "-", "RCNN"], "Method"], [1214, 1218, ["low", "level", "proposal", "algorithms"], "Method"], [1221, 1224, ["region", "proposal", "network"], "Method"], [1225, 1226, ["RPN"], "Method"], [1267, 1270, ["R", "-", "FCN"], "Method"], [1278, 1281, ["Faster", "-", "RCNN"], "Method"], [1284, 1290, ["fully", "connected", "sub", "-", "detection", "network"], "Method"], [1292, 1296, ["fully", "convolutional", "subdetection", "network"], "Method"], [1338, 1339, ["speed"], "Metric"], [1340, 1342, ["cascade", "procedure"], "Method"], [1354, 1357, ["stage", "object", "detectors"], "Method"], [1364, 1365, ["SSD"], "Method"], [1371, 1375, ["one", "-", "stage", "approach"], "Method"], [1379, 1382, ["RoI", "pooling", "step"], "Method"], [1390, 1394, ["One", "-", "stage", "detectors"], "Method"], [1400, 1402, ["twostage", "detectors"], "Method"], [1411, 1412, ["SSD"], "Method"], [1453, 1454, ["DSSD"], "Method"], [1455, 1456, ["RON"], "Method"], [1464, 1466, ["hourglass", "network"], "Method"], [1495, 1499, ["one", "-", "stage", "detectors"], "Method"], [1506, 1508, ["stage", "detectors"], "Method"], [1559, 1561, ["Focal", "Loss"], "Method"], [1575, 1577, ["onestage", "detector"], "Method"], [1582, 1584, ["stage", "detectors"], "Method"], [1619, 1621, ["stage", "detector"], "Method"], [1692, 1693, ["DeNet"], "Method"], [1708, 1712, ["sub", "-", "detection", "network"], "Method"], [1726, 1728, ["stage", "approach"], "Method"], [1737, 1738, ["ConvNet"], "Method"], [1753, 1754, ["classification"], "Task"], [1762, 1765, ["feature", "selection", "step"], "Method"], [1770, 1772, ["corner", "pooling"], "Method"], [1780, 1782, ["corner", "detection"], "Task"], [1783, 1786, ["Point", "Linking", "Network"], "Method"], [1787, 1788, ["PLN"], "Method"], [1791, 1795, ["one", "-", "stage", "detector"], "Method"], [1796, 1798, ["anchor", "boxes"], "Method"], [1902, 1903, ["CornerNet"], "Method"], [1907, 1908, ["PLN"], "Method"], [1921, 1922, ["PLN"], "Method"], [1934, 1935, ["CornerNet"], "Method"], [1936, 1938, ["corner", "pooling"], "Method"], [1950, 1952, ["Associative", "Embedding"], "Task"], [1956, 1961, ["multi", "-", "person", "pose", "estimation"], "Task"], [2020, 2022, ["object", "detection"], "Task"], [2026, 2030, ["detecting", "and", "grouping", "corners"], "Task"], [2031, 2032, ["embeddings"], "Task"], [2054, 2056, ["hourglass", "architecture"], "Method"], [2062, 2064, ["focal", "loss"], "Method"], [2104, 2106, ["convolutional", "network"], "Method"], [2201, 2205, ["post", "-", "processing", "algorithm"], "Method"], [2219, 2220, ["CornerNet"], "Method"], [2224, 2226, ["hourglass", "network"], "Method"], [2231, 2233, ["backbone", "network"], "Method"], [2234, 2235, ["CornerNet"], "Method"], [2237, 2239, ["hourglass", "network"], "Method"], [2243, 2245, ["prediction", "modules"], "Method"], [2279, 2281, ["hourglass", "network"], "Method"], [2293, 2295, ["object", "detectors"], "Method"], [2321, 2323, ["hourglass", "network"], "Method"], [2326, 2328, ["Detecting", "Corners"], "Task"], [2333, 2334, ["heatmaps"], "Method"], [2416, 2417, ["training"], "Task"], [2541, 2544, ["unnormalized", "2D", "Gaussian"], "Method"], [2550, 2551, ["CornerNet"], "Method"], [2553, 2555, ["backbone", "network"], "Method"], [2559, 2561, ["prediction", "modules"], "Method"], [2644, 2646, ["y", "cij"], "Method"], [2655, 2657, ["unnormalized", "Gaussians"], "Method"], [2663, 2665, ["focal", "loss"], "Method"], [2714, 2716, ["y", "cij"], "Method"], [2741, 2743, ["downsampling", "layers"], "Method"], [2759, 2762, ["image", "fully", "convolutionally"], "Task"], [2917, 2918, ["training"], "Task"], [2993, 2996, ["Associative", "Embedding", "method"], "Method"], [3002, 3007, ["multi", "-", "person", "pose", "estimation"], "Task"], [3019, 3020, ["embedding"], "Method"], [3040, 3042, ["associative", "embedding"], "Method"], [3247, 3249, ["Corner", "Pooling"], "Method"], [3309, 3311, ["corner", "pooling"], "Method"], [3486, 3489, ["elementwise", "max", "operation"], "Method"], [3501, 3503, ["dynamic", "programming"], "Method"], [3572, 3574, ["prediction", "modules"], "Method"], [3586, 3588, ["prediction", "module"], "Method"], [3607, 3609, ["residual", "block"], "Method"], [3620, 3625, ["3", "×", "3", "convolution", "module"], "Method"], [3638, 3640, ["backbone", "network"], "Method"], [3643, 3647, ["×", "3", "convolution", "modules"], "Method"], [3675, 3682, ["3", "×", "3", "Conv", "-", "BN", "layer"], "Method"], [3700, 3703, ["3×3", "convolution", "module"], "Method"], [3709, 3715, ["Conv", "-", "ReLU", "-", "Conv", "layers"], "Method"], [3726, 3728, ["Hourglass", "Network"], "Method"], [3728, 3729, ["CornerNet"], "Method"], [3731, 3733, ["hourglass", "network"], "Method"], [3738, 3740, ["backbone", "network"], "Method"], [3742, 3744, ["hourglass", "network"], "Method"], [3749, 3753, ["human", "pose", "estimation", "task"], "Task"], [3757, 3761, ["fully", "convolutional", "neural", "network"], "Method"], [3767, 3769, ["hourglass", "modules"], "Method"], [3771, 3773, ["hourglass", "module"], "Method"], [3782, 3787, ["convolution", "and", "max", "pooling", "layers"], "Method"], [3802, 3806, ["upsampling", "and", "convolution", "layers"], "Method"], [3813, 3816, ["max", "pooling", "layers"], "Method"], [3817, 3819, ["skip", "layers"], "Method"], [3832, 3834, ["hourglass", "module"], "Method"], [3848, 3850, ["hourglass", "modules"], "Method"], [3857, 3859, ["hourglass", "modules"], "Method"], [3875, 3877, ["hourglass", "network"], "Method"], [3881, 3883, ["object", "detection"], "Task"], [3908, 3910, ["hourglass", "network"], "Method"], [3912, 3914, ["hourglass", "network"], "Method"], [3917, 3918, ["hourglasses"], "Method"], [3929, 3931, ["hourglass", "module"], "Method"], [3935, 3938, ["max", "pool", "-"], "Method"], [3964, 3968, ["horizontal", "max", "-", "pooling"], "Task"], [3975, 3979, ["vertical", "max", "-", "pooling"], "Method"], [3984, 3989, ["max", "-", "pooled", "feature", "maps"], "Method"], [3994, 3996, ["prediction", "module"], "Method"], [4009, 4011, ["convolution", "module"], "Method"], [4026, 4028, ["convolution", "module"], "Method"], [4091, 4093, ["residual", "modules"], "Method"], [4096, 4099, ["nearest", "neighbor", "upsampling"], "Method"], [4107, 4109, ["residual", "modules"], "Method"], [4113, 4115, ["residual", "modules"], "Method"], [4123, 4125, ["hourglass", "module"], "Method"], [4128, 4130, ["hourglass", "modules"], "Method"], [4142, 4146, ["×", "7", "convolution", "module"], "Method"], [4175, 4176, ["training"], "Task"], [4206, 4212, ["×", "1", "Conv", "-", "BN", "module"], "Method"], [4221, 4223, ["hourglass", "module"], "Method"], [4229, 4233, ["element", "-", "wise", "addition"], "Method"], [4236, 4237, ["ReLU"], "Method"], [4255, 4257, ["hourglass", "module"], "Method"], [4262, 4264, ["hourglass", "network"], "Method"], [4275, 4276, ["detectors"], "Method"], [4303, 4304, ["CornerNet"], "Method"], [4320, 4321, ["PyTorch"], "Method"], [4323, 4324, ["pretraining"], "Method"], [4343, 4345, ["convolution", "layers"], "Method"], [4385, 4388, ["data", "augmentation", "techniques"], "Method"], [4389, 4392, ["random", "horizontal", "flipping"], "Method"], [4393, 4395, ["random", "scaling"], "Method"], [4420, 4421, ["PCA"], "Method"], [4431, 4432, ["Adam"], "Method"], [4438, 4441, ["full", "training", "loss"], "Metric"], [4548, 4550, ["learning", "rate"], "Metric"], [4563, 4564, ["detectors"], "Method"], [4577, 4579, ["learning", "rate"], "Metric"], [4601, 4605, ["post", "-", "processing", "algorithm"], "Method"], [4620, 4624, ["non", "-", "maximal", "suppression"], "Method"], [4625, 4626, ["NMS"], "Method"], [4630, 4634, ["3×3", "max", "pooling", "layer"], "Method"], [4721, 4723, ["detection", "scores"], "Metric"], [4777, 4780, ["soft", "-", "nms"], "Method"], [4798, 4800, ["inference", "time"], "Metric"], [4815, 4817, ["MS", "COCO"], "Material"], [4819, 4820, ["CornerNet"], "Method"], [4831, 4833, ["MS", "COCO"], "Material"], [4841, 4842, ["validation"], "Task"], [4875, 4879, ["hyper", "-", "parameter", "searching"], "Task"], [4880, 4882, ["ablation", "study"], "Task"], [4918, 4920, ["MS", "COCO"], "Material"], [4921, 4923, ["average", "precisions"], "Metric"], [4924, 4925, ["APs"], "Metric"], [4938, 4940, ["evaluation", "metrics"], "Metric"], [4943, 4945, ["Ablation", "Study"], "Task"], [4947, 4949, ["Corner", "Pooling"], "Method"], [4949, 4951, ["Corner", "pooling"], "Method"], [4956, 4957, ["CornerNet"], "Method"], [4970, 4972, ["corner", "pooling"], "Method"], [4986, 4988, ["corner", "pooling"], "Method"], [4995, 4996, ["AP"], "Metric"], [5013, 5015, ["corner", "pooling"], "Method"], [5026, 5027, ["APs"], "Metric"], [5074, 5076, ["corner", "pooling"], "Method"], [5081, 5083, ["Corner", "Pooling"], "Method"], [5086, 5088, ["Corner", "pooling"], "Method"], [5178, 5180, ["Detecting", "corners"], "Task"], [5185, 5188, ["binary", "classification", "task"], "Task"], [5217, 5218, ["mAPs"], "Metric"], [5234, 5236, ["corner", "pooling"], "Method"], [5261, 5263, ["corner", "pooling"], "Method"], [5330, 5332, ["Corner", "pooling"], "Method"], [5343, 5345, ["corner", "pooling"], "Method"], [5398, 5399, ["CornerNet"], "Method"], [5406, 5408, ["penalty", "reduction"], "Method"], [5422, 5423, ["CornerNet"], "Method"], [5437, 5438, ["AP"], "Metric"], [5445, 5447, ["AP", "m"], "Metric"], [5451, 5452, ["AP"], "Metric"], [5464, 5465, ["AP"], "Metric"], [5469, 5471, ["AP", "m"], "Metric"], [5475, 5477, ["AP", "l"], "Metric"], [5488, 5490, ["penalty", "reduction"], "Task"], [5499, 5501, ["Hourglass", "Network"], "Method"], [5501, 5502, ["CornerNet"], "Method"], [5504, 5506, ["hourglass", "network"], "Method"], [5511, 5513, ["backbone", "network"], "Method"], [5516, 5518, ["hourglass", "network"], "Method"], [5543, 5545, ["hourglass", "network"], "Method"], [5546, 5547, ["CornerNet"], "Method"], [5551, 5552, ["CornerNet"], "Method"], [5557, 5559, ["hourglass", "network"], "Method"], [5560, 5561, ["FPN"], "Method"], [5564, 5567, ["ResNet", "-", "101"], "Method"], [5582, 5584, ["object", "detectors"], "Method"], [5592, 5593, ["FPN"], "Method"], [5594, 5595, ["predictions"], "Task"], [5601, 5605, ["anchor", "box", "based", "detector"], "Method"], [5608, 5610, ["hourglass", "network"], "Method"], [5615, 5617, ["hourglass", "module"], "Method"], [5630, 5632, ["upsampling", "stage"], "Method"], [5636, 5639, ["anchor", "box", "design"], "Method"], [5640, 5641, ["RetinaNet"], "Method"], [5667, 5668, ["CornerNet"], "Method"], [5684, 5685, ["CornerNet"], "Method"], [5686, 5687, ["FPN"], "Method"], [5690, 5691, ["AP"], "Metric"], [5694, 5698, ["anchor", "box", "based", "detector"], "Method"], [5699, 5701, ["hourglass", "network"], "Method"], [5704, 5705, ["AP"], "Metric"], [5714, 5716, ["backbone", "network"], "Method"], [5720, 5722, ["hourglass", "network"], "Method"], [5728, 5729, ["CornerNet"], "Method"], [5761, 5762, ["CornerNet"], "Method"], [5768, 5769, ["CornerNet"], "Method"], [5787, 5788, ["detectors"], "Method"], [5790, 5791, ["RetinaNet"], "Method"], [5792, 5796, ["Cascade", "R", "-", "CNN"], "Method"], [5800, 5803, ["IoU", "-", "Net"], "Method"], [5812, 5813, ["CornerNet"], "Method"], [5817, 5818, ["AP"], "Metric"], [5820, 5821, ["IoU"], "Metric"], [5826, 5830, ["Cascade", "R", "-", "CNN"], "Method"], [5831, 5834, ["IoU", "-", "Net"], "Metric"], [5838, 5842, ["Cascade", "R", "-", "CNN"], "Method"], [5846, 5848, ["RetinaNet", "2"], "Method"], [5855, 5858, ["Cor", "-", "2"], "Method"], [5903, 5905, ["Error", "Analysis"], "Task"], [5905, 5906, ["CornerNet"], "Method"], [5919, 5920, ["detection"], "Task"], [5967, 5969, ["error", "analysis"], "Method"], [6005, 6006, ["AP"], "Metric"], [6050, 6051, ["AP"], "Metric"], [6073, 6077, ["detecting", "and", "grouping", "corners"], "Task"], [6082, 6084, ["detecting", "corners"], "Task"], [6115, 6116, ["CornerNet"], "Method"], [6139, 6141, ["multiscale", "evaluation"], "Task"], [6142, 6143, ["CornerNet"], "Method"], [6145, 6146, ["AP"], "Metric"], [6157, 6161, ["one", "-", "stage", "methods"], "Method"], [6166, 6168, ["stage", "methods"], "Method"], [6175, 6176, ["CornerNet"], "Method"], [6181, 6183, ["object", "detection"], "Task"], [6194, 6195, ["CornerNet"], "Method"], [6196, 6198, ["MS", "COCO"], "Material"], [830, 832, ["corner", "pooling"], "Method"], [1911, 1912, ["CornerNet"], "Method"], [2039, 2041, ["corner", "pooling"], "Method"], [2073, 2074, ["CornerNet"], "Method"], [2078, 2079, ["CornerNet"], "Method"], [2271, 2273, ["corner", "pooling"], "Method"], [3363, 3365, ["corner", "pooling"], "Method"], [3400, 3402, ["corner", "pooling"], "Method"], [3514, 3516, ["corner", "pooling"], "Method"], [3565, 3567, ["corner", "pooling"], "Method"], [3627, 3629, ["corner", "pooling"], "Method"], [3655, 3657, ["corner", "pooling"], "Method"], [3947, 3949, ["corner", "pooling"], "Method"], [4013, 4015, ["corner", "pooling"], "Method"], [4751, 4752, ["CornerNet"], "Method"], [4824, 4826, ["MS", "COCO"], "Material"], [4930, 4931, ["APs"], "Metric"], [5000, 5001, ["AP"], "Metric"], [5006, 5007, ["AP"], "Metric"], [5108, 5110, ["corner", "pooling"], "Method"], [5149, 5151, ["corner", "pooling"], "Method"], [5223, 5225, ["MS", "COCO"], "Material"], [5240, 5241, ["mAPs"], "Metric"], [5265, 5266, ["mAPs"], "Metric"], [5287, 5289, ["corner", "pooling"], "Method"], [5295, 5296, ["mAPs"], "Metric"], [5679, 5680, ["CornerNet"], "Method"], [6015, 6016, ["AP"], "Metric"], [6018, 6019, ["AP"], "Metric"], [6021, 6022, ["AP"], "Metric"], [6127, 6129, ["MS", "COCO"], "Material"]]} {"doc_id": "170cc3b72dd2793920f8b0bbd0926d9244c49562", "text": ["document", ":", "Deep", "Exploration", "via", "Bootstrapped", "DQN", "Efficient", "exploration", "remains", "a", "major", "challenge", "for", "reinforcement", "learning", "(", "RL", ")", ".", "Common", "dithering", "strategies", "for", "exploration", ",", "such", "as", "-", "greedy", ",", "do", "not", "carry", "out", "temporally", "-", "extended", "(", "or", "deep", ")", "exploration", ";", "this", "can", "lead", "to", "exponentially", "larger", "data", "requirements", ".", "However", ",", "most", "algorithms", "for", "statistically", "efficient", "RL", "are", "not", "computationally", "tractable", "in", "complex", "environments", ".", "Randomized", "value", "functions", "offer", "a", "promising", "approach", "to", "efficient", "exploration", "with", "generalization", ",", "but", "existing", "algorithms", "are", "not", "compatible", "with", "nonlinearly", "parameterized", "value", "functions", ".", "As", "a", "first", "step", "towards", "addressing", "such", "contexts", "we", "develop", "bootstrapped", "DQN", ".", "We", "demonstrate", "that", "bootstrapped", "DQN", "can", "combine", "deep", "exploration", "with", "deep", "neural", "networks", "for", "exponentially", "faster", "learning", "than", "any", "dithering", "strategy", ".", "In", "the", "Arcade", "Learning", "Environment", "bootstrapped", "DQN", "substantially", "improves", "learning", "speed", "and", "cumulative", "performance", "across", "most", "games", ".", "section", ":", "Introduction", "We", "study", "the", "reinforcement", "learning", "(", "RL", ")", "problem", "where", "an", "agent", "interacts", "with", "an", "unknown", "environment", ".", "The", "agent", "takes", "a", "sequence", "of", "actions", "in", "order", "to", "maximize", "cumulative", "rewards", ".", "Unlike", "standard", "planning", "problems", ",", "an", "RL", "agent", "does", "not", "begin", "with", "perfect", "knowledge", "of", "the", "environment", ",", "but", "learns", "through", "experience", ".", "This", "leads", "to", "a", "fundamental", "trade", "-", "off", "of", "exploration", "versus", "exploitation", ";", "the", "agent", "may", "improve", "its", "future", "rewards", "by", "exploring", "poorly", "understood", "states", "and", "actions", ",", "but", "this", "may", "require", "sacrificing", "immediate", "rewards", ".", "To", "learn", "efficiently", "an", "agent", "should", "explore", "only", "when", "there", "are", "valuable", "learning", "opportunities", ".", "Further", ",", "since", "any", "action", "may", "have", "long", "term", "consequences", ",", "the", "agent", "should", "reason", "about", "the", "informational", "value", "of", "possible", "observation", "sequences", ".", "Without", "this", "sort", "of", "temporally", "extended", "(", "deep", ")", "exploration", ",", "learning", "times", "can", "worsen", "by", "an", "exponential", "factor", ".", "The", "theoretical", "RL", "literature", "offers", "a", "variety", "of", "provably", "-", "efficient", "approaches", "to", "deep", "exploration", ".", "However", ",", "most", "of", "these", "are", "designed", "for", "Markov", "decision", "processes", "(", "MDPs", ")", "with", "small", "finite", "state", "spaces", ",", "while", "others", "require", "solving", "computationally", "intractable", "planning", "tasks", ".", "These", "algorithms", "are", "not", "practical", "in", "complex", "environments", "where", "an", "agent", "must", "generalize", "to", "operate", "effectively", ".", "For", "this", "reason", ",", "large", "-", "scale", "applications", "of", "RL", "have", "relied", "upon", "statistically", "inefficient", "strategies", "for", "exploration", "or", "even", "no", "exploration", "at", "all", ".", "We", "review", "related", "literature", "in", "more", "detail", "in", "Section", "[", "reference", "]", ".", "Common", "dithering", "strategies", ",", "such", "as", "-", "greedy", ",", "approximate", "the", "value", "of", "an", "action", "by", "a", "single", "number", ".", "Most", "of", "the", "time", "they", "pick", "the", "action", "with", "the", "highest", "estimate", ",", "but", "sometimes", "they", "choose", "another", "action", "at", "random", ".", "In", "this", "paper", ",", "we", "consider", "an", "alternative", "approach", "to", "efficient", "exploration", "inspired", "by", "Thompson", "sampling", ".", "These", "algorithms", "have", "some", "notion", "of", "uncertainty", "and", "instead", "maintain", "a", "distribution", "over", "possible", "values", ".", "They", "explore", "by", "randomly", "select", "a", "policy", "according", "to", "the", "probability", "it", "is", "the", "optimal", "policy", ".", "Recent", "work", "has", "shown", "that", "randomized", "value", "functions", "can", "implement", "something", "similar", "to", "Thompson", "sampling", "without", "the", "need", "for", "an", "intractable", "exact", "posterior", "update", ".", "However", ",", "this", "work", "is", "restricted", "to", "linearly", "-", "parameterized", "value", "functions", ".", "We", "present", "a", "natural", "extension", "of", "this", "approach", "that", "enables", "use", "of", "complex", "non", "-", "linear", "generalization", "methods", "such", "as", "deep", "neural", "networks", ".", "We", "show", "that", "the", "bootstrap", "with", "random", "initialization", "can", "produce", "reasonable", "uncertainty", "estimates", "for", "neural", "networks", "at", "low", "computational", "cost", ".", "Bootstrapped", "DQN", "leverages", "these", "uncertainty", "estimates", "for", "efficient", "(", "and", "deep", ")", "exploration", ".", "We", "demonstrate", "that", "these", "benefits", "can", "extend", "to", "large", "scale", "problems", "that", "are", "not", "designed", "to", "highlight", "deep", "exploration", ".", "Bootstrapped", "DQN", "substantially", "reduces", "learning", "times", "and", "improves", "performance", "across", "most", "games", ".", "This", "algorithm", "is", "computationally", "efficient", "and", "parallelizable", ";", "on", "a", "single", "machine", "our", "implementation", "runs", "roughly", "%", "slower", "than", "DQN", ".", "section", ":", "Uncertainty", "for", "neural", "networks", "Deep", "neural", "networks", "(", "DNN", ")", "represent", "the", "state", "of", "the", "art", "in", "many", "supervised", "and", "reinforcement", "learning", "domains", ".", "We", "want", "an", "exploration", "strategy", "that", "is", "statistically", "computationally", "efficient", "together", "with", "a", "DNN", "representation", "of", "the", "value", "function", ".", "To", "explore", "efficiently", ",", "the", "first", "step", "to", "quantify", "uncertainty", "in", "value", "estimates", "so", "that", "the", "agent", "can", "judge", "potential", "benefits", "of", "exploratory", "actions", ".", "The", "neural", "network", "literature", "presents", "a", "sizable", "body", "of", "work", "on", "uncertainty", "quantification", "founded", "on", "parametric", "Bayesian", "inference", ".", "We", "actually", "found", "the", "simple", "non", "-", "parametric", "bootstrap", "with", "random", "initialization", "more", "effective", "in", "our", "experiments", ",", "but", "the", "main", "ideas", "of", "this", "paper", "would", "apply", "with", "any", "other", "approach", "to", "uncertainty", "in", "DNNs", ".", "The", "bootstrap", "princple", "is", "to", "approximate", "a", "population", "distribution", "by", "a", "sample", "distribution", ".", "In", "its", "most", "common", "form", ",", "the", "bootstrap", "takes", "as", "input", "a", "data", "set", "and", "an", "estimator", ".", "To", "generate", "a", "sample", "from", "the", "bootstrapped", "distribution", ",", "a", "data", "set", "of", "cardinality", "equal", "to", "that", "of", "is", "sampled", "uniformly", "with", "replacement", "from", ".", "The", "bootstrap", "sample", "estimate", "is", "then", "taken", "to", "be", ".", "The", "bootstrap", "is", "widely", "hailed", "as", "a", "great", "advance", "of", "20th", "century", "applied", "statistics", "and", "even", "comes", "with", "theoretical", "guarantees", ".", "In", "Figure", "[", "reference", "]", "we", "present", "an", "efficient", "and", "scalable", "method", "for", "generating", "bootstrap", "samples", "from", "a", "large", "and", "deep", "neural", "network", ".", "The", "network", "consists", "of", "a", "shared", "architecture", "with", "bootstrapped", "‘", "‘", "heads", "’", "’", "branching", "off", "independently", ".", "Each", "head", "is", "trained", "only", "on", "its", "bootstrapped", "sub", "-", "sample", "of", "the", "data", "and", "represents", "a", "single", "bootstrap", "sample", ".", "The", "shared", "network", "learns", "a", "joint", "feature", "representation", "across", "all", "the", "data", ",", "which", "can", "provide", "significant", "computational", "advantages", "at", "the", "cost", "of", "lower", "diversity", "between", "heads", ".", "This", "type", "of", "bootstrap", "can", "be", "trained", "efficiently", "in", "a", "single", "forward", "/", "backward", "pass", ";", "it", "can", "be", "thought", "of", "as", "a", "data", "-", "dependent", "dropout", ",", "where", "the", "dropout", "mask", "for", "each", "head", "is", "fixed", "for", "each", "data", "point", ".", "[", "b", "]", "0.31", "[", "b", "]", "0.31", "[", "b", "]", "0.31", "Figure", "[", "reference", "]", "presents", "an", "example", "of", "uncertainty", "estimates", "from", "bootstrapped", "neural", "networks", "on", "a", "regression", "task", "with", "noisy", "data", ".", "We", "trained", "a", "fully", "-", "connected", "2", "-", "layer", "neural", "networks", "with", "50", "rectified", "linear", "units", "(", "ReLU", ")", "in", "each", "layer", "on", "50", "bootstrapped", "samples", "from", "the", "data", ".", "As", "is", "standard", ",", "we", "initialize", "these", "networks", "with", "random", "parameter", "values", ",", "this", "induces", "an", "important", "initial", "diversity", "in", "the", "models", ".", "We", "were", "unable", "to", "generate", "effective", "uncertainty", "estimates", "for", "this", "problem", "using", "the", "dropout", "approach", "in", "prior", "literature", ".", "Further", "details", "are", "provided", "in", "Appendix", "[", "reference", "]", ".", "section", ":", "Bootstrapped", "DQN", "For", "a", "policy", "we", "define", "the", "value", "of", "an", "action", "in", "state", ",", "where", "is", "a", "discount", "factor", "that", "balances", "immediate", "versus", "future", "rewards", ".", "This", "expectation", "indicates", "that", "the", "initial", "state", "is", ",", "the", "initial", "action", "is", ",", "and", "thereafter", "actions", "are", "selected", "by", "the", "policy", ".", "The", "optimal", "value", "is", ".", "To", "scale", "to", "large", "problems", ",", "we", "learn", "a", "parameterized", "estimate", "of", "the", "Q", "-", "value", "function", "rather", "than", "a", "tabular", "encoding", ".", "We", "use", "a", "neural", "network", "to", "estimate", "this", "value", ".", "The", "Q", "-", "learning", "update", "from", "state", ",", "action", ",", "reward", "and", "new", "state", "is", "given", "by", "where", "is", "the", "scalar", "learning", "rate", "and", "is", "the", "target", "value", ".", "are", "target", "network", "parameters", "fixed", ".", "Several", "important", "modifications", "to", "the", "Q", "-", "learning", "update", "improve", "stability", "for", "DQN", ".", "First", "the", "algorithm", "learns", "from", "sampled", "transitions", "from", "an", "experience", "buffer", ",", "rather", "than", "learning", "fully", "online", ".", "Second", "the", "algorithm", "uses", "a", "target", "network", "with", "parameters", "that", "are", "copied", "from", "the", "learning", "network", "only", "every", "time", "steps", "and", "then", "kept", "fixed", "in", "between", "updates", ".", "Double", "DQN", "modifies", "the", "target", "and", "helps", "further", ":", "Bootstrapped", "DQN", "modifies", "DQN", "to", "approximate", "a", "distribution", "over", "Q", "-", "values", "via", "the", "bootstrap", ".", "At", "the", "start", "of", "each", "episode", ",", "bootstrapped", "DQN", "samples", "a", "single", "Q", "-", "value", "function", "from", "its", "approximate", "posterior", ".", "The", "agent", "then", "follows", "the", "policy", "which", "is", "optimal", "for", "that", "sample", "for", "the", "duration", "of", "the", "episode", ".", "This", "is", "a", "natural", "adaptation", "of", "the", "Thompson", "sampling", "heuristic", "to", "RL", "that", "allows", "for", "temporally", "extended", "(", "or", "deep", ")", "exploration", ".", "We", "implement", "this", "algorithm", "efficiently", "by", "building", "up", "bootstrapped", "estimates", "of", "the", "Q", "-", "value", "function", "in", "parallel", "as", "in", "Figure", "[", "reference", "]", ".", "Importantly", ",", "each", "one", "of", "these", "value", "function", "function", "heads", "is", "trained", "against", "its", "own", "target", "network", ".", "This", "means", "that", "each", "provide", "a", "temporally", "extended", "(", "and", "consistent", ")", "estimate", "of", "the", "value", "uncertainty", "via", "TD", "estimates", ".", "In", "order", "to", "keep", "track", "of", "which", "data", "belongs", "to", "which", "bootstrap", "head", "we", "store", "flags", "indicating", "which", "heads", "are", "privy", "to", "which", "data", ".", "We", "approximate", "a", "bootstrap", "sample", "by", "selecting", "uniformly", "at", "random", "and", "following", "for", "the", "duration", "of", "that", "episode", ".", "We", "present", "a", "detailed", "algorithm", "for", "our", "implementation", "of", "bootstrapped", "DQN", "in", "Appendix", "[", "reference", "]", ".", "section", ":", "Related", "work", "The", "observation", "that", "temporally", "extended", "exploration", "is", "necessary", "for", "efficient", "reinforcement", "learning", "is", "not", "new", ".", "For", "any", "prior", "distribution", "over", "MDPs", ",", "the", "optimal", "exploration", "strategy", "is", "available", "through", "dynamic", "programming", "in", "the", "Bayesian", "belief", "state", "space", ".", "However", ",", "the", "exact", "solution", "is", "intractable", "even", "for", "very", "simple", "systems", ".", "Many", "successful", "RL", "applications", "focus", "on", "generalization", "and", "planning", "but", "address", "exploration", "only", "via", "inefficient", "exploration", "or", "even", "none", "at", "all", ".", "However", ",", "such", "exploration", "strategies", "can", "be", "highly", "inefficient", ".", "Many", "exploration", "strategies", "are", "guided", "by", "the", "principle", "of", "‘", "‘", "optimism", "in", "the", "face", "of", "uncertainty", "’", "’", "(", "OFU", ")", ".", "These", "algorithms", "add", "an", "exploration", "bonus", "to", "values", "of", "state", "-", "action", "pairs", "that", "may", "lead", "to", "useful", "learning", "and", "select", "actions", "to", "maximize", "these", "adjusted", "values", ".", "This", "approach", "was", "first", "proposed", "for", "finite", "-", "armed", "bandits", ",", "but", "the", "principle", "has", "been", "extended", "successfully", "across", "bandits", "with", "generalization", "and", "tabular", "RL", ".", "Except", "for", "particular", "deterministic", "contexts", ",", "OFU", "methods", "that", "lead", "to", "efficient", "RL", "in", "complex", "domains", "have", "been", "computationally", "intractable", ".", "The", "work", "of", "aims", "to", "add", "an", "effective", "bonus", "through", "a", "variation", "of", "DQN", ".", "The", "resulting", "algorithm", "relies", "on", "a", "large", "number", "of", "hand", "-", "tuned", "parameters", "and", "is", "only", "suitable", "for", "application", "to", "deterministic", "problems", ".", "We", "compare", "our", "results", "on", "Atari", "to", "theirs", "in", "Appendix", "[", "reference", "]", "and", "find", "that", "bootstrapped", "DQN", "offers", "a", "significant", "improvement", "over", "previous", "methods", ".", "Perhaps", "the", "oldest", "heuristic", "for", "balancing", "exploration", "with", "exploitation", "is", "given", "by", "Thompson", "sampling", ".", "This", "bandit", "algorithm", "takes", "a", "single", "sample", "from", "the", "posterior", "at", "every", "time", "step", "and", "chooses", "the", "action", "which", "is", "optimal", "for", "that", "time", "step", ".", "To", "apply", "the", "Thompson", "sampling", "principle", "to", "RL", ",", "an", "agent", "should", "sample", "a", "value", "function", "from", "its", "posterior", ".", "Naive", "applications", "of", "Thompson", "sampling", "to", "RL", "which", "resample", "every", "timestep", "can", "be", "extremely", "inefficient", ".", "The", "agent", "must", "also", "commit", "to", "this", "sample", "for", "several", "time", "steps", "in", "order", "to", "achieve", "deep", "exploration", ".", "The", "algorithm", "PSRL", "does", "exactly", "this", ",", "with", "state", "of", "the", "art", "guarantees", ".", "However", ",", "this", "algorithm", "still", "requires", "solving", "a", "single", "known", "MDP", ",", "which", "will", "usually", "be", "intractable", "for", "large", "systems", ".", "Our", "new", "algorithm", ",", "bootstrapped", "DQN", ",", "approximates", "this", "approach", "to", "exploration", "via", "randomized", "value", "functions", "sampled", "from", "an", "approximate", "posterior", ".", "Recently", ",", "authors", "have", "proposed", "the", "RLSVI", "algorithm", "which", "accomplishes", "this", "for", "linearly", "parameterized", "value", "functions", ".", "Surprisingly", ",", "RLSVI", "recovers", "state", "of", "the", "art", "guarantees", "in", "the", "setting", "with", "tabular", "basis", "functions", ",", "but", "its", "performance", "is", "crucially", "dependent", "upon", "a", "suitable", "linear", "representation", "of", "the", "value", "function", ".", "We", "extend", "these", "ideas", "to", "produce", "an", "algorithm", "that", "can", "simultaneously", "perform", "generalization", "and", "exploration", "with", "a", "flexible", "nonlinear", "value", "function", "representation", ".", "Our", "method", "is", "simple", ",", "general", "and", "compatible", "with", "almost", "all", "advances", "in", "deep", "RL", "at", "low", "computational", "cost", "and", "with", "few", "tuning", "parameters", ".", "section", ":", "Deep", "Exploration", "Uncertainty", "estimates", "allow", "an", "agent", "to", "direct", "its", "exploration", "at", "potentially", "informative", "states", "and", "actions", ".", "In", "bandits", ",", "this", "choice", "of", "directed", "exploration", "rather", "than", "dithering", "generally", "categorizes", "efficient", "algorithms", ".", "The", "story", "in", "RL", "is", "not", "as", "simple", ",", "directed", "exploration", "is", "not", "enough", "to", "guarantee", "efficiency", ";", "the", "exploration", "must", "also", "be", "deep", ".", "Deep", "exploration", "means", "exploration", "which", "is", "directed", "over", "multiple", "time", "steps", ";", "it", "can", "also", "be", "called", "‘", "‘", "planning", "to", "learn", "’", "’", "or", "‘", "‘", "far", "-", "sighted", "’", "’", "exploration", ".", "Unlike", "bandit", "problems", ",", "which", "balance", "actions", "which", "are", "immediately", "rewarding", "or", "immediately", "informative", ",", "RL", "settings", "require", "planning", "over", "several", "time", "steps", ".", "For", "exploitation", ",", "this", "means", "that", "an", "efficient", "agent", "must", "consider", "the", "future", "rewards", "over", "several", "time", "steps", "and", "not", "simply", "the", "myopic", "rewards", ".", "In", "exactly", "the", "same", "way", ",", "efficient", "exploration", "may", "require", "taking", "actions", "which", "are", "neither", "immediately", "rewarding", ",", "nor", "immediately", "informative", ".", "To", "illustrate", "this", "distinction", ",", "consider", "a", "simple", "deterministic", "chain", "with", "three", "step", "horizon", "starting", "from", "state", ".", "This", "MDP", "is", "known", "to", "the", "agent", "a", "priori", ",", "with", "deterministic", "actions", "‘", "‘", "left", "’", "’", "and", "‘", "‘", "right", "’", "’", ".", "All", "states", "have", "zero", "reward", ",", "except", "for", "the", "leftmost", "state", "which", "has", "known", "reward", "and", "the", "rightmost", "state", "which", "is", "unknown", ".", "In", "order", "to", "reach", "either", "a", "rewarding", "state", "or", "an", "informative", "state", "within", "three", "steps", "from", "the", "agent", "must", "plan", "a", "consistent", "strategy", "over", "several", "time", "steps", ".", "Figure", "[", "reference", "]", "depicts", "the", "planning", "and", "look", "ahead", "trees", "for", "several", "algorithmic", "approaches", "in", "this", "example", "MDP", ".", "The", "action", "‘", "‘", "left", "’", "’", "is", "gray", ",", "the", "action", "‘", "‘", "right", "’", "’", "is", "black", ".", "Rewarding", "states", "are", "depicted", "as", "red", ",", "informative", "states", "as", "blue", ".", "Dashed", "lines", "indicate", "that", "the", "agent", "can", "plan", "ahead", "for", "either", "rewards", "or", "information", ".", "Unlike", "bandit", "algorithms", ",", "an", "RL", "agent", "can", "plan", "to", "exploit", "future", "rewards", ".", "Only", "an", "RL", "agent", "with", "deep", "exploration", "can", "plan", "to", "learn", ".", "[", "b", "]", "0.99", "[", "b", "]", "0.23", "[", "b", "]", "0.23", "[", "b", "]", "0.23", "[", "b", "]", "0.23", "subsection", ":", "Testing", "for", "deep", "exploration", "We", "now", "present", "a", "series", "of", "didactic", "computational", "experiments", "designed", "to", "highlight", "the", "need", "for", "deep", "exploration", ".", "These", "environments", "can", "be", "described", "by", "chains", "of", "length", "in", "Figure", "[", "reference", "]", ".", "Each", "episode", "of", "interaction", "lasts", "steps", "after", "which", "point", "the", "agent", "resets", "to", "the", "initial", "state", ".", "These", "are", "toy", "problems", "intended", "to", "be", "expository", "rather", "than", "entirely", "realistic", ".", "Balancing", "a", "well", "known", "and", "mildly", "successful", "strategy", "versus", "an", "unknown", ",", "but", "potentially", "more", "rewarding", ",", "approach", "can", "emerge", "in", "many", "practical", "applications", ".", "These", "environments", "may", "be", "described", "by", "a", "finite", "tabular", "MDP", ".", "However", ",", "we", "consider", "algorithms", "which", "interact", "with", "the", "MDP", "only", "through", "raw", "pixel", "features", ".", "We", "consider", "two", "feature", "mappings", "and", "in", ".", "We", "present", "results", "for", ",", "which", "worked", "better", "for", "all", "DQN", "variants", "due", "to", "better", "generalization", ",", "but", "the", "difference", "was", "relatively", "small", "-", "see", "Appendix", "[", "reference", "]", ".", "Thompson", "DQN", "is", "the", "same", "as", "bootstrapped", "DQN", ",", "but", "resamples", "every", "timestep", ".", "Ensemble", "DQN", "uses", "the", "same", "architecture", "as", "bootstrapped", "DQN", ",", "but", "with", "an", "ensemble", "policy", ".", "We", "say", "that", "the", "algorithm", "has", "successfully", "learned", "the", "optimal", "policy", "when", "it", "has", "successfully", "completed", "one", "hundred", "episodes", "with", "optimal", "reward", "of", ".", "For", "each", "chain", "length", ",", "we", "ran", "each", "learning", "algorithm", "for", "2000", "episodes", "across", "three", "seeds", ".", "We", "plot", "the", "median", "time", "to", "learn", "in", "Figure", "[", "reference", "]", ",", "together", "with", "a", "conservative", "lower", "bound", "of", "on", "the", "expected", "time", "to", "learn", "for", "any", "shallow", "exploration", "strategy", ".", "Only", "bootstrapped", "DQN", "demonstrates", "a", "graceful", "scaling", "to", "long", "chains", "which", "require", "deep", "exploration", ".", "subsection", ":", "How", "does", "bootstrapped", "DQN", "drive", "deep", "exploration", "?", "Bootstrapped", "DQN", "explores", "in", "a", "manner", "similar", "to", "the", "provably", "-", "efficient", "algorithm", "PSRL", "but", "it", "uses", "a", "bootstrapped", "neural", "network", "to", "approximate", "a", "posterior", "sample", "for", "the", "value", ".", "Unlike", "PSRL", ",", "bootstrapped", "DQN", "directly", "samples", "a", "value", "function", "and", "so", "does", "not", "require", "further", "planning", "steps", ".", "This", "algorithm", "is", "similar", "to", "RLSVI", ",", "which", "is", "also", "provably", "-", "efficient", ",", "but", "with", "a", "neural", "network", "instead", "of", "linear", "value", "function", "and", "bootstrap", "instead", "of", "Gaussian", "sampling", ".", "The", "analysis", "for", "the", "linear", "setting", "suggests", "that", "this", "nonlinear", "approach", "will", "work", "well", "so", "long", "as", "the", "distribution", "remains", "stochastically", "optimistic", ",", "or", "at", "least", "as", "spread", "out", "as", "the", "‘", "‘", "correct", "’", "’", "posterior", ".", "Bootstrapped", "DQN", "relies", "upon", "random", "initialization", "of", "the", "network", "weights", "as", "a", "prior", "to", "induce", "diversity", ".", "Surprisingly", ",", "we", "found", "this", "initial", "diversity", "was", "enough", "to", "maintain", "diverse", "generalization", "to", "new", "and", "unseen", "states", "for", "large", "and", "deep", "neural", "networks", ".", "This", "is", "effective", "for", "our", "experimental", "setting", ",", "but", "will", "not", "work", "in", "all", "situations", ".", "In", "general", "it", "may", "be", "necessary", "to", "maintain", "some", "more", "rigorous", "notion", "of", "‘", "‘", "prior", "’", "’", ",", "potentially", "through", "the", "use", "of", "artificial", "prior", "data", "to", "maintain", "diversity", ".", "One", "potential", "explanation", "for", "the", "efficacy", "of", "simple", "random", "initialization", "is", "that", "unlike", "supervised", "learning", "or", "bandits", ",", "where", "all", "networks", "fit", "the", "same", "data", ",", "each", "of", "our", "heads", "has", "a", "unique", "target", "network", ".", "This", ",", "together", "with", "stochastic", "minibatch", "and", "flexible", "nonlinear", "representations", ",", "means", "that", "even", "small", "differences", "at", "initialization", "may", "become", "bigger", "as", "they", "refit", "to", "unique", "TD", "errors", ".", "Bootstrapped", "DQN", "does", "not", "require", "that", "any", "single", "network", "is", "initialized", "to", "the", "correct", "policy", "of", "‘", "‘", "right", "’", "’", "at", "every", "step", ",", "which", "would", "be", "exponentially", "unlikely", "for", "large", "chains", ".", "For", "the", "algorithm", "to", "be", "successful", "in", "this", "example", "we", "only", "require", "that", "the", "networks", "generalize", "in", "a", "diverse", "way", "to", "the", "actions", "they", "have", "never", "chosen", "in", "the", "states", "they", "have", "not", "visited", "very", "often", ".", "Imagine", "that", ",", "in", "the", "example", "above", ",", "the", "network", "has", "made", "it", "as", "far", "as", "state", ",", "but", "never", "observed", "the", "action", "right", ".", "As", "long", "as", "one", "head", "imagines", "then", "TD", "bootstrapping", "can", "propagate", "this", "signal", "back", "to", "through", "the", "target", "network", "to", "drive", "deep", "exploration", ".", "The", "expected", "time", "for", "these", "estimates", "at", "to", "propagate", "to", "at", "least", "one", "head", "grows", "gracefully", "in", ",", "even", "for", "relatively", "small", ",", "as", "our", "experiments", "show", ".", "We", "expand", "upon", "this", "intuition", "with", "a", "video", "designed", "to", "highlight", "how", "bootstrapped", "DQN", "demonstrates", "deep", "exploration", ".", "We", "present", "further", "evaluation", "on", "a", "difficult", "stochastic", "MDP", "in", "Appendix", "[", "reference", "]", ".", "section", ":", "Arcade", "Learning", "Environment", "We", "now", "evaluate", "our", "algorithm", "across", "49", "Atari", "games", "on", "the", "Arcade", "Learning", "Environment", ".", "Importantly", ",", "and", "unlike", "the", "experiments", "in", "Section", "[", "reference", "]", ",", "these", "domains", "are", "not", "specifically", "designed", "to", "showcase", "our", "algorithm", ".", "In", "fact", ",", "many", "Atari", "games", "are", "structured", "so", "that", "small", "rewards", "always", "indicate", "part", "of", "an", "optimal", "policy", ".", "This", "may", "be", "crucial", "for", "the", "strong", "performance", "observed", "by", "dithering", "strategies", ".", "We", "find", "that", "exploration", "via", "bootstrapped", "DQN", "produces", "significant", "gains", "versus", "-", "greedy", "in", "this", "setting", ".", "Bootstrapped", "DQN", "reaches", "peak", "performance", "roughly", "similar", "to", "DQN", ".", "However", ",", "our", "improved", "exploration", "mean", "we", "reach", "human", "performance", "on", "average", "30", "%", "faster", "across", "all", "games", ".", "This", "translates", "to", "significantly", "improved", "cumulative", "rewards", "through", "learning", ".", "We", "follow", "the", "setup", "of", "for", "our", "network", "architecture", "and", "benchmark", "our", "performance", "against", "their", "algorithm", ".", "Our", "network", "structure", "is", "identical", "to", "the", "convolutional", "structure", "of", "DQN", "except", "we", "split", "10", "separate", "bootstrap", "heads", "after", "the", "convolutional", "layer", "as", "per", "Figure", "[", "reference", "]", ".", "Recently", ",", "several", "authors", "have", "provided", "architectural", "and", "algorithmic", "improvements", "to", "DDQN", ".", "We", "do", "not", "compare", "our", "results", "to", "these", "since", "their", "advances", "are", "orthogonal", "to", "our", "concern", "and", "could", "easily", "be", "incorporated", "to", "our", "bootstrapped", "DQN", "design", ".", "Full", "details", "of", "our", "experimental", "set", "up", "are", "available", "in", "Appendix", "[", "reference", "]", ".", "subsection", ":", "Implementing", "bootstrapped", "DQN", "at", "scale", "We", "now", "examine", "how", "to", "generate", "online", "bootstrap", "samples", "for", "DQN", "in", "a", "computationally", "efficient", "manner", ".", "We", "focus", "on", "three", "key", "questions", ":", "how", "many", "heads", "do", "we", "need", ",", "how", "should", "we", "pass", "gradients", "to", "the", "shared", "network", "and", "how", "should", "we", "bootstrap", "data", "online", "?", "We", "make", "significant", "compromises", "in", "order", "to", "maintain", "computational", "cost", "comparable", "to", "DQN", ".", "Figure", "[", "reference", "]", "presents", "the", "cumulative", "reward", "of", "bootstrapped", "DQN", "on", "the", "game", "Breakout", ",", "for", "different", "number", "of", "heads", ".", "More", "heads", "leads", "to", "faster", "learning", ",", "but", "even", "a", "small", "number", "of", "heads", "captures", "most", "of", "the", "benefits", "of", "bootstrapped", "DQN", ".", "We", "choose", ".", "[", "b", "]", "0.4", "[", "b", "]", "0.4", "The", "shared", "network", "architecture", "allows", "us", "to", "train", "this", "combined", "network", "via", "backpropagation", ".", "Feeding", "network", "heads", "to", "the", "shared", "convolutional", "network", "effectively", "increases", "the", "learning", "rate", "for", "this", "portion", "of", "the", "network", ".", "In", "some", "games", ",", "this", "leads", "to", "premature", "and", "sub", "-", "optimal", "convergence", ".", "We", "found", "the", "best", "final", "scores", "by", "normalizing", "the", "gradients", "by", ",", "but", "this", "also", "leads", "to", "slower", "early", "learning", ".", "See", "Appendix", "[", "reference", "]", "for", "more", "details", ".", "4mu", "plus", "2mu", "minus", "4mu=0mu", "3", "mu=0mu", "5mu", "plus", "5mu=0mu", "To", "implement", "an", "online", "bootstrap", "we", "use", "an", "independent", "Bernoulli", "mask", "for", "each", "head", "in", "each", "episode", ".", "These", "flags", "are", "stored", "in", "the", "memory", "replay", "buffer", "and", "identify", "which", "heads", "are", "trained", "on", "which", "data", ".", "However", ",", "when", "trained", "using", "a", "shared", "minibatch", "the", "algorithm", "will", "also", "require", "an", "effective", "more", "iterations", ";", "this", "is", "undesirable", "computationally", ".", "Surprisingly", ",", "we", "found", "the", "algorithm", "performed", "similarly", "irrespective", "of", "and", "all", "outperformed", "DQN", ",", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "This", "is", "strange", "and", "we", "discuss", "this", "phenomenon", "in", "Appendix", "[", "reference", "]", ".", "However", ",", "in", "light", "of", "this", "empirical", "observation", "for", "Atari", ",", "we", "chose", "to", "save", "on", "minibatch", "passes", ".", "As", "a", "result", "bootstrapped", "DQN", "runs", "at", "similar", "computational", "speed", "to", "vanilla", "DQN", "on", "identical", "hardware", ".", "subsection", ":", "Efficient", "exploration", "in", "Atari", "We", "find", "that", "Bootstrapped", "DQN", "drives", "efficient", "exploration", "in", "several", "Atari", "games", ".", "For", "the", "same", "amount", "of", "game", "experience", ",", "bootstrapped", "DQN", "generally", "outperforms", "DQN", "with", "-", "greedy", "exploration", ".", "Figure", "[", "reference", "]", "demonstrates", "this", "effect", "for", "a", "diverse", "selection", "of", "games", ".", "On", "games", "where", "DQN", "performs", "well", ",", "bootstrapped", "DQN", "typically", "performs", "better", ".", "Bootstrapped", "DQN", "does", "not", "reach", "human", "performance", "on", "Amidar", "(", "DQN", "does", ")", "but", "does", "on", "Beam", "Rider", "and", "Battle", "Zone", "(", "DQN", "does", "not", ")", ".", "To", "summarize", "this", "improvement", "in", "learning", "time", "we", "consider", "the", "number", "of", "frames", "required", "to", "reach", "human", "performance", ".", "If", "bootstrapped", "DQN", "reaches", "human", "performance", "in", "frames", "of", "DQN", "we", "say", "it", "has", "improved", "by", ".", "Figure", "[", "reference", "]", "shows", "that", "Bootstrapped", "DQN", "typically", "reaches", "human", "performance", "significantly", "faster", ".", "On", "most", "games", "where", "DQN", "does", "not", "reach", "human", "performance", ",", "bootstrapped", "DQN", "does", "not", "solve", "the", "problem", "by", "itself", ".", "On", "some", "challenging", "Atari", "games", "where", "deep", "exploration", "is", "conjectured", "to", "be", "important", "our", "results", "are", "not", "entirely", "successful", ",", "but", "still", "promising", ".", "In", "Frostbite", ",", "bootstrapped", "DQN", "reaches", "the", "second", "level", "much", "faster", "than", "DQN", "but", "network", "instabilities", "cause", "the", "performance", "to", "crash", ".", "In", "Montezuma", "’s", "Revenge", ",", "bootstrapped", "DQN", "reaches", "the", "first", "key", "after", "20", "m", "frames", "(", "DQN", "never", "observes", "a", "reward", "even", "after", "200", "m", "frames", ")", "but", "does", "not", "properly", "learn", "from", "this", "experience", ".", "Our", "results", "suggest", "that", "improved", "exploration", "may", "help", "to", "solve", "these", "remaining", "games", ",", "but", "also", "highlight", "the", "importance", "of", "other", "problems", "like", "network", "instability", ",", "reward", "clipping", "and", "temporally", "extended", "rewards", ".", "subsection", ":", "Overall", "performance", "Bootstrapped", "DQN", "is", "able", "to", "learn", "much", "faster", "than", "DQN", ".", "Figure", "[", "reference", "]", "shows", "that", "bootstrapped", "DQN", "also", "improves", "upon", "the", "final", "score", "across", "most", "games", ".", "However", ",", "the", "real", "benefits", "to", "efficient", "exploration", "mean", "that", "bootstrapped", "DQN", "outperforms", "DQN", "by", "orders", "of", "magnitude", "in", "terms", "of", "the", "cumulative", "rewards", "through", "learning", "(", "Figure", "[", "reference", "]", ".", "In", "both", "figures", "we", "normalize", "performance", "relative", "to", "a", "fully", "random", "policy", ".", "The", "most", "similar", "work", "to", "ours", "presents", "several", "other", "approaches", "to", "improved", "exploration", "in", "Atari", "they", "optimize", "for", "AUC", "-", "20", ",", "a", "normalized", "version", "of", "the", "cumulative", "returns", "after", "20", "m", "frames", ".", "According", "to", "their", "metric", ",", "averaged", "across", "the", "14", "games", "they", "consider", ",", "we", "improve", "upon", "both", "base", "DQN", "(", "0.29", ")", "and", "their", "best", "method", "(", "0.37", ")", "to", "obtain", "0.62", "via", "bootstrapped", "DQN", ".", "We", "present", "these", "results", "together", "with", "results", "tables", "across", "all", "49", "games", "in", "Appendix", "[", "reference", "]", ".", "subsection", ":", "Visualizing", "bootstrapped", "DQN", "We", "now", "present", "some", "more", "insight", "to", "how", "bootstrapped", "DQN", "drives", "deep", "exploration", "in", "Atari", ".", "In", "each", "game", ",", "although", "each", "head", "learns", "a", "high", "scoring", "policy", ",", "the", "policies", "they", "find", "are", "quite", "distinct", ".", "In", "the", "video", "we", "show", "the", "evolution", "of", "these", "policies", "simultaneously", "for", "several", "games", ".", "Although", "each", "head", "performs", "well", ",", "they", "each", "follow", "a", "unique", "policy", ".", "By", "contrast", ",", "-", "greedy", "strategies", "are", "almost", "indistinguishable", "for", "small", "values", "of", "and", "totally", "ineffectual", "for", "larger", "values", ".", "We", "believe", "that", "this", "deep", "exploration", "is", "key", "to", "improved", "learning", ",", "since", "diverse", "experiences", "allow", "for", "better", "generalization", ".", "Disregarding", "exploration", ",", "bootstrapped", "DQN", "may", "be", "beneficial", "as", "a", "purely", "exploitative", "policy", ".", "We", "can", "combine", "all", "the", "heads", "into", "a", "single", "ensemble", "policy", ",", "for", "example", "by", "choosing", "the", "action", "with", "the", "most", "votes", "across", "heads", ".", "This", "approach", "might", "have", "several", "benefits", ".", "First", ",", "we", "find", "that", "the", "ensemble", "policy", "can", "often", "outperform", "any", "individual", "policy", ".", "Second", ",", "the", "distribution", "of", "votes", "across", "heads", "to", "give", "a", "measure", "of", "the", "uncertainty", "in", "the", "optimal", "policy", ".", "Unlike", "vanilla", "DQN", ",", "bootstrapped", "DQN", "can", "know", "what", "it", "does", "n’t", "know", ".", "In", "an", "application", "where", "executing", "a", "poorly", "-", "understood", "action", "is", "dangerous", "this", "could", "be", "crucial", ".", "In", "the", "video", "we", "visualize", "this", "ensemble", "policy", "across", "several", "games", ".", "We", "find", "that", "the", "uncertainty", "in", "this", "policy", "is", "surprisingly", "interpretable", ":", "all", "heads", "agree", "at", "clearly", "crucial", "decision", "points", ",", "but", "remain", "diverse", "at", "other", "less", "important", "steps", ".", "section", ":", "Closing", "remarks", "In", "this", "paper", "we", "present", "bootstrapped", "DQN", "as", "an", "algorithm", "for", "efficient", "reinforcement", "learning", "in", "complex", "environments", ".", "We", "demonstrate", "that", "the", "bootstrap", "can", "produce", "useful", "uncertainty", "estimates", "for", "deep", "neural", "networks", ".", "Bootstrapped", "DQN", "is", "computationally", "tractable", "and", "also", "naturally", "scalable", "to", "massive", "parallel", "systems", ".", "We", "believe", "that", ",", "beyond", "our", "specific", "implementation", ",", "randomized", "value", "functions", "represent", "a", "promising", "alternative", "to", "dithering", "for", "exploration", ".", "Bootstrapped", "DQN", "practically", "combines", "efficient", "generalization", "with", "exploration", "for", "complex", "nonlinear", "value", "functions", ".", "bibliography", ":", "References", "APPENDICES", "appendix", ":", "Uncertainty", "for", "neural", "networks", "In", "this", "appendix", "we", "discuss", "some", "of", "the", "experimental", "setup", "to", "qualitatively", "evaluate", "uncertainty", "methods", "for", "deep", "neural", "networks", ".", "To", "do", "this", ",", "we", "generated", "twenty", "noisy", "regression", "pairs", "with", ":", "where", "are", "drawn", "uniformly", "from", "and", ".", "We", "set", "and", ".", "None", "of", "these", "numerical", "choices", "were", "important", "except", "to", "represent", "a", "highly", "nonlinear", "function", "with", "lots", "of", "noise", "and", "several", "clear", "regions", "where", "we", "should", "be", "uncertain", ".", "We", "present", "the", "regression", "data", "together", "with", "an", "indication", "of", "the", "generating", "distribution", "in", "Figure", "[", "reference", "]", ".", "Interestingly", ",", "we", "did", "not", "find", "that", "using", "dropout", "produced", "satisfying", "confidence", "intervals", "for", "this", "task", ".", "We", "present", "one", "example", "of", "this", "dropout", "posterior", "estimate", "in", "Figure", "[", "reference", "]", ".", "[", "b", "]", "0.45", "[", "b", "]", "0.45", "These", "results", "are", "unsatisfactory", "for", "several", "reasons", ".", "First", ",", "the", "network", "extrapolates", "the", "mean", "posterior", "far", "outside", "the", "range", "of", "any", "actual", "data", "for", ".", "We", "believe", "this", "is", "because", "dropout", "only", "perturbs", "locally", "from", "a", "single", "neural", "network", "fit", ",", "unlike", "bootstrap", ".", "Second", ",", "the", "posterior", "samples", "from", "the", "dropout", "approximation", "are", "very", "spiky", "and", "do", "not", "look", "like", "any", "sensible", "posterior", "sample", ".", "Third", ",", "the", "network", "collapses", "to", "almost", "zero", "uncertainty", "in", "regions", "with", "data", ".", "We", "spent", "some", "time", "altering", "our", "dropout", "scheme", "to", "fix", "this", "effect", ",", "which", "might", "be", "undesirable", "for", "stochastic", "domains", "and", "we", "believed", "might", "be", "an", "artefact", "of", "our", "implementation", ".", "However", ",", "after", "further", "thought", "we", "believe", "this", "to", "be", "an", "effect", "which", "you", "would", "expect", "for", "dropout", "posterior", "approximations", ".", "In", "Figure", "[", "reference", "]", "we", "present", "a", "didactic", "example", "taken", "from", "the", "author", "’s", "website", ".", "On", "the", "right", "hand", "side", "of", "the", "plot", "we", "generate", "noisy", "data", "with", "wildly", "different", "values", ".", "Training", "a", "neural", "network", "using", "MSE", "criterion", "means", "that", "the", "network", "will", "surely", "converge", "to", "the", "mean", "of", "the", "noisy", "data", ".", "Any", "dropout", "samples", "remain", "highly", "concentrated", "around", "this", "mean", ".", "By", "contrast", ",", "bootstrapped", "neural", "networks", "may", "include", "different", "subsets", "of", "this", "noisy", "data", "and", "so", "may", "produce", "a", "more", "intuitive", "uncertainty", "estimates", "for", "our", "settings", ".", "Note", "this", "is", "n’t", "necessarily", "a", "failure", "of", "dropout", "to", "approximate", "a", "Gaussian", "process", "posterior", ",", "but", "this", "artefact", "could", "be", "shared", "by", "any", "homoskedastic", "posterior", ".", "The", "authors", "of", "propose", "a", "heteroskedastic", "variant", "which", "can", "help", ",", "but", "does", "not", "address", "the", "fundamental", "issue", "that", "for", "large", "networks", "trained", "to", "convergence", "all", "dropout", "samples", "may", "converge", "to", "every", "single", "datapoint", "…", "even", "the", "outliers", ".", "In", "this", "paper", "we", "focus", "on", "the", "bootstrap", "approach", "to", "uncertainty", "for", "neural", "networks", ".", "We", "like", "its", "simplicity", ",", "connections", "to", "established", "statistical", "methodology", "and", "empirical", "good", "performance", ".", "However", ",", "the", "key", "insights", "of", "this", "paper", "is", "the", "use", "of", "deep", "exploration", "via", "randomized", "value", "functions", ".", "This", "is", "compatible", "with", "any", "approximate", "posterior", "estimator", "for", "deep", "neural", "networks", ".", "We", "believe", "that", "this", "area", "of", "uncertainty", "estimates", "for", "neural", "networks", "remains", "an", "important", "area", "of", "research", "in", "its", "own", "right", ".", "Bootstrapped", "uncertainty", "estimates", "for", "the", "Q", "-", "value", "functions", "have", "another", "crucial", "advantage", "over", "dropout", "which", "does", "not", "appear", "in", "the", "supervised", "problem", ".", "Unlike", "random", "dropout", "masks", "trained", "against", "random", "target", "networks", ",", "our", "implementation", "of", "bootstrap", "DQN", "trains", "against", "its", "own", "temporally", "consistent", "target", "network", ".", "This", "means", "that", "our", "bootstrap", "estimates", "(", "in", "the", "sense", "of", ")", ",", "are", "able", "to", "‘", "‘", "bootstrap", "’", "’", "(", "in", "the", "TD", "sense", "of", ")", "on", "their", "own", "estimates", "of", "the", "long", "run", "value", ".", "This", "is", "important", "to", "quantify", "the", "long", "run", "uncertainty", "over", "Q", "and", "drive", "deep", "exploration", ".", "appendix", ":", "Bootstrapped", "DQN", "implementation", "Algorithm", "[", "reference", "]", "gives", "a", "full", "description", "of", "Bootstrapped", "DQN", ".", "It", "captures", "two", "modes", "of", "operation", "where", "either", "neural", "networks", "are", "used", "to", "estimate", "the", "-", "value", "functions", ",", "or", "where", "one", "neural", "network", "with", "heads", "is", "used", "to", "estimate", "-", "value", "functions", ".", "In", "both", "cases", ",", "as", "this", "is", "largely", "a", "parameterisation", "issue", ",", "we", "denote", "the", "value", "function", "networks", "as", ",", "where", "is", "output", "of", "the", "th", "network", "or", "the", "th", "head", ".", "A", "core", "idea", "to", "the", "full", "bootstrapped", "DQN", "algorithm", "is", "the", "bootstrap", "mask", ".", "The", "mask", "decides", ",", "for", "each", "value", "function", ",", "whether", "or", "not", "it", "should", "train", "upon", "the", "experience", "generated", "at", "step", ".", "In", "its", "simplest", "form", "is", "a", "binary", "vector", "of", "length", ",", "masking", "out", "or", "including", "each", "value", "function", "for", "training", "on", "that", "time", "step", "of", "experience", "(", "i.e.", ",", "should", "it", "receive", "gradients", "from", "the", "corresponding", "tuple", ")", ".", "The", "masking", "distribution", "is", "responsible", "for", "generating", "each", ".", "For", "example", ",", "when", "yields", "whose", "components", "are", "independently", "drawn", "from", "a", "bernoulli", "distribution", "with", "parameter", "then", "this", "corresponds", "to", "the", "double", "-", "or", "-", "nothing", "bootstrap", ".", "On", "the", "other", "hand", ",", "if", "yields", "a", "mask", "with", "all", "ones", ",", "then", "the", "algorithm", "reduces", "to", "an", "ensemble", "method", ".", "Poisson", "masks", "provides", "the", "most", "natural", "parallel", "with", "the", "standard", "non", "-", "parameteric", "boostrap", "since", "as", ".", "Exponential", "masks", "closely", "resemble", "the", "standard", "Bayesian", "nonparametric", "posterior", "of", "a", "Dirichlet", "process", ".", "Bootstrapped", "DQN", "[", "1", "]", "Input", ":", "Value", "function", "networks", "with", "outputs", ".", "Masking", "distribution", ".", "Let", "be", "a", "replay", "buffer", "storing", "experience", "for", "training", ".", "each", "episode", "Obtain", "initial", "state", "from", "environment", "Pick", "a", "value", "function", "to", "act", "using", "step", "until", "end", "of", "episode", "Pick", "an", "action", "according", "to", "Receive", "state", "and", "reward", "from", "environment", ",", "having", "taking", "action", "Sample", "bootstrap", "mask", "Add", "to", "replay", "buffer", "Periodically", ",", "the", "replay", "buffer", "is", "played", "back", "to", "update", "the", "parameters", "of", "the", "value", "function", "network", ".", "The", "gradients", "of", "the", "th", "value", "function", "for", "the", "th", "tuple", "in", "the", "replay", "buffer", ",", "is", ":", "where", "is", "given", "by", "(", "[", "reference", "]", ")", ".", "Note", "that", "the", "mask", "modulates", "the", "gradient", ",", "giving", "rise", "to", "the", "bootstrap", "behaviour", ".", "appendix", ":", "Experiments", "for", "deep", "exploration", "subsection", ":", "Bootstrap", "methodology", "A", "naive", "implementation", "of", "bootstrapped", "DQN", "builds", "up", "complete", "networks", "with", "distinct", "memory", "buffers", ".", "This", "method", "is", "parallelizable", "up", "to", "many", "machines", ",", "however", "we", "wanted", "to", "produce", "an", "algorithm", "that", "was", "efficient", "even", "on", "a", "single", "machine", ".", "To", "do", "this", ",", "we", "implemented", "the", "bootstrap", "heads", "in", "a", "single", "larger", "network", ",", "like", "Figure", "[", "reference", "]", "but", "without", "any", "shared", "network", ".", "We", "implement", "bootstrap", "by", "masking", "each", "episode", "of", "data", "according", "to", ".", "In", "Figure", "[", "reference", "]", "we", "demonstrate", "that", "bootstrapped", "DQN", "can", "implement", "deep", "exploration", "even", "with", "relatively", "small", "values", "of", ".", "However", ",", "the", "results", "are", "more", "robust", "and", "scalable", "with", "larger", ".", "We", "run", "our", "experiments", "on", "the", "example", "from", "Figure", "[", "reference", "]", ".", "Surprisingly", ",", "this", "method", "is", "even", "effective", "with", "and", "complete", "data", "sharing", "between", "heads", ".", "This", "degenerate", "full", "sharing", "of", "information", "turns", "out", "to", "be", "remarkably", "efficient", "for", "training", "large", "and", "deep", "neural", "networks", ".", "We", "discuss", "this", "phenomenon", "more", "in", "Appendix", "[", "reference", "]", ".", "Generating", "good", "estimates", "for", "uncertainty", "is", "not", "enough", "for", "efficient", "exploration", ".", "In", "Figure", "[", "reference", "]", "we", "see", "that", "other", "methods", "trained", "with", "the", "same", "network", "architecture", "are", "totally", "ineffective", "at", "implementing", "deep", "exploration", ".", "The", "-", "greedy", "policy", "follows", "just", "one", "-", "value", "estimate", ".", "We", "allow", "this", "policy", "to", "be", "evaluated", "without", "dithering", ".", "The", "ensemble", "policy", "is", "trained", "exactly", "as", "per", "bootstrapped", "DQN", "except", "at", "each", "stage", "the", "algorithm", "follows", "the", "policy", "which", "is", "majority", "vote", "of", "the", "bootstrap", "heads", ".", "Thompson", "sampling", "is", "the", "same", "as", "bootstrapped", "DQN", "except", "a", "new", "head", "is", "sampled", "every", "timestep", ",", "rather", "than", "every", "episode", ".", "We", "can", "see", "that", "only", "bootstrapped", "DQN", "demonstrates", "efficient", "and", "deep", "exploration", "in", "this", "domain", ".", "subsection", ":", "A", "difficult", "stochastic", "MDP", "Figure", "[", "reference", "]", "shows", "that", "bootstrapped", "DQN", "can", "implement", "effective", "(", "and", "deep", ")", "exploration", "where", "similar", "deep", "RL", "architectures", "fail", ".", "However", ",", "since", "the", "underlying", "system", "is", "a", "small", "and", "finite", "MDP", "there", "may", "be", "several", "other", "simpler", "strategies", "which", "would", "also", "solve", "this", "problem", ".", "We", "will", "now", "consider", "a", "difficult", "variant", "of", "this", "chain", "system", "with", "significant", "stochastic", "noise", "in", "transitions", "as", "depicted", "in", "Figure", "[", "reference", "]", ".", "Action", "‘", "‘", "left", "’", "’", "deterministically", "moves", "the", "agent", "left", ",", "but", "action", "‘", "‘", "right", "’", "’", "is", "only", "successful", "50", "%", "of", "the", "time", "and", "otherwise", "also", "moves", "left", ".", "The", "agent", "interacts", "with", "the", "MDP", "in", "episodes", "of", "length", "and", "begins", "each", "episode", "at", ".", "Once", "again", "the", "optimal", "policy", "is", "to", "head", "right", ".", "Bootstrapped", "DQN", "is", "unique", "amongst", "scalable", "approaches", "to", "efficient", "exploration", "with", "deep", "RL", "in", "stochastic", "domains", ".", "For", "benchmark", "performance", "we", "implement", "three", "algorithms", "which", ",", "unlike", "bootstrapped", "DQN", ",", "will", "receive", "the", "true", "tabular", "representation", "for", "the", "MDP", ".", "These", "algorithms", "are", "based", "on", "three", "state", "of", "the", "art", "approaches", "to", "exploration", "via", "dithering", "(", "-", "greedy", ")", ",", "optimism", "and", "posterior", "sampling", ".", "We", "discuss", "the", "choice", "of", "these", "benchmarks", "in", "Appendix", "[", "reference", "]", ".", "[", "b", "]", "0.45", "[", "b", "]", "0.45", "In", "Figure", "[", "reference", "]", "we", "present", "the", "empirical", "regret", "of", "each", "algorithm", "averaged", "over", "10", "seeds", "over", "the", "first", "two", "thousand", "episodes", ".", "The", "empirical", "regret", "is", "the", "cumulative", "difference", "between", "the", "expected", "rewards", "of", "the", "optimal", "policy", "and", "the", "realized", "rewards", "of", "each", "algorithm", ".", "We", "find", "that", "bootstrapped", "DQN", "achieves", "similar", "performance", "to", "state", "of", "the", "art", "efficient", "exploration", "schemes", "such", "as", "PSRL", "even", "without", "prior", "knowledge", "of", "the", "tabular", "MDP", "structure", "and", "in", "noisy", "environments", ".", "Most", "telling", "is", "how", "much", "better", "bootstrapped", "DQN", "does", "than", "the", "state", "of", "the", "art", "optimistic", "algorithm", "UCRL2", ".", "Although", "Figure", "[", "reference", "]", "seems", "to", "suggest", "UCRL2", "incurs", "linear", "regret", ",", "actually", "it", "follows", "its", "bounds", "where", "is", "the", "number", "of", "states", "and", "is", "the", "number", "of", "actions", ".", "For", "the", "example", "in", "Figure", "[", "reference", "]", "we", "attempted", "to", "display", "our", "performance", "compared", "to", "several", "benchmark", "tabula", "rasa", "approaches", "to", "exploration", ".", "There", "are", "many", "other", "algorithms", "we", "could", "have", "considered", ",", "but", "for", "a", "short", "paper", "we", "chose", "to", "focus", "against", "the", "most", "common", "approach", "(", "-", "greedy", ")", "the", "pre", "-", "eminent", "optimistic", "approach", "(", "UCRL2", ")", "and", "posterior", "sampling", "(", "PSRL", ")", ".", "Other", "common", "heuristic", "approaches", ",", "such", "as", "optimistic", "initialization", "for", "Q", "-", "learning", "can", "be", "tuned", "to", "work", "well", "on", "this", "domain", ",", "however", "the", "precise", "parameters", "are", "sensitive", "to", "the", "underlying", "MDP", ".", "To", "make", "a", "general", "-", "purpose", "version", "of", "this", "heuristic", "essentially", "leads", "to", "optimistic", "algorithms", ".", "Since", "UCRL2", "is", "originally", "designed", "for", "infinite", "-", "horizon", "MDPs", ",", "we", "use", "the", "natural", "adaptation", "of", "this", "algorithm", ",", "which", "has", "state", "of", "the", "art", "guarantees", "in", "finite", "horizon", "MDPs", "as", "well", ".", "Figure", "[", "reference", "]", "displays", "the", "empirical", "regret", "of", "these", "algorithms", "together", "with", "bootstrapped", "DQN", "on", "the", "example", "from", "Figure", "[", "reference", "]", ".", "It", "is", "somewhat", "disconcerting", "that", "UCRL2", "appears", "to", "incur", "linear", "regret", ",", "but", "it", "is", "proven", "to", "satisfy", "near", "-", "optimal", "regret", "bounds", ".", "Actually", ",", "as", "we", "show", "in", "Figure", "[", "reference", "]", ",", "the", "algorithm", "produces", "regret", "which", "scales", "very", "similarly", "to", "its", "established", "bounds", ".", "Similarly", ",", "even", "for", "this", "tiny", "problem", "size", ",", "the", "recent", "analysis", "that", "proves", "a", "near", "optimal", "sample", "complexity", "in", "fixed", "horizon", "problems", "only", "guarantees", "that", "we", "will", "have", "fewer", "than", "suboptimal", "episodes", ".", "While", "these", "bounds", "may", "be", "acceptable", "in", "worst", "case", "scaling", ",", "they", "are", "not", "of", "much", "practical", "use", ".", "subsection", ":", "One", "-", "hot", "features", "In", "Figure", "[", "reference", "]", "we", "include", "the", "mean", "performance", "of", "bootstrapped", "DQN", "with", "one", "-", "hot", "feature", "encodings", ".", "We", "found", "that", ",", "using", "these", "features", ",", "bootstrapped", "DQN", "learned", "the", "optimal", "policy", "for", "most", "seeds", ",", "but", "was", "somewhat", "less", "robust", "than", "the", "thermometer", "encoding", ".", "Two", "out", "of", "ten", "seeds", "failed", "to", "learn", "the", "optimal", "policy", "within", "2000", "episodes", ",", "this", "is", "presented", "in", "Figure", "[", "reference", "]", ".", "appendix", ":", "Experiments", "for", "Atari", "subsection", ":", "Experimental", "setup", "We", "use", "the", "same", "49", "Atari", "games", "as", "for", "our", "experiments", ".", "Each", "step", "of", "the", "agent", "corresponds", "to", "four", "steps", "of", "the", "emulator", ",", "where", "the", "same", "action", "is", "repeated", ",", "the", "reward", "values", "of", "the", "agents", "are", "clipped", "between", "-", "1", "and", "1", "for", "stability", ".", "We", "evaluate", "our", "agents", "and", "report", "performance", "based", "upon", "the", "raw", "scores", ".", "The", "convolutional", "part", "of", "the", "network", "used", "is", "identical", "to", "the", "one", "used", "in", ".", "The", "input", "to", "the", "network", "is", "4x84x84", "tensor", "with", "a", "rescaled", ",", "grayscale", "version", "of", "the", "last", "four", "observations", ".", "The", "first", "convolutional", "(", "conv", ")", "layer", "has", "32", "filters", "of", "size", "8", "with", "a", "stride", "of", "4", ".", "The", "second", "conv", "layer", "has", "64", "filters", "of", "size", "4", "with", "stride", "2", ".", "The", "last", "conv", "layer", "has", "64", "filters", "of", "size", "3", ".", "We", "split", "the", "network", "beyond", "the", "final", "layer", "into", "distinct", "heads", ",", "each", "one", "is", "fully", "connected", "and", "identical", "to", "the", "single", "head", "of", "DQN", ".", "This", "consists", "of", "a", "fully", "connected", "layer", "to", "512", "units", "followed", "by", "another", "fully", "connected", "layer", "to", "the", "Q", "-", "Values", "for", "each", "action", ".", "The", "fully", "connected", "layers", "all", "use", "Rectified", "Linear", "Units", "(", "ReLU", ")", "as", "a", "non", "-", "linearity", ".", "We", "normalize", "gradients", "that", "flow", "from", "each", "head", ".", "We", "trained", "the", "networks", "with", "RMSProp", "with", "a", "momentum", "of", "0.95", "and", "a", "learning", "rate", "of", "0.00025", "as", "in", ".", "The", "discount", "was", "set", "to", ",", "the", "number", "of", "steps", "between", "target", "updates", "was", "set", "to", "steps", ".", "We", "trained", "the", "agents", "for", "a", "total", "of", "50", "m", "steps", "per", "game", ",", "which", "corresponds", "to", "200", "m", "frames", ".", "The", "agents", "were", "every", "1", "m", "frames", ",", "for", "evaluation", "in", "bootstrapped", "DQN", "we", "use", "an", "ensemble", "voting", "policy", ".", "The", "experience", "replay", "contains", "the", "1", "m", "most", "recent", "transitions", ".", "We", "update", "the", "network", "every", "steps", "by", "randomly", "sampling", "a", "minibatch", "of", "transitions", "from", "the", "replay", "buffer", "to", "use", "the", "exact", "same", "minibatch", "schedule", "as", "DQN", ".", "For", "training", "we", "used", "an", "-", "greedy", "policy", "with", "being", "annealed", "linearly", "from", "to", "over", "the", "first", "1", "m", "timesteps", ".", "subsection", ":", "Gradient", "normalization", "in", "bootstrap", "heads", "Most", "literature", "in", "deep", "RL", "for", "Atari", "focuses", "on", "learning", "the", "best", "single", "evaluation", "policy", ",", "with", "particular", "attention", "to", "whether", "this", "above", "or", "below", "human", "performance", ".", "This", "is", "unusual", "for", "the", "RL", "literature", ",", "which", "typically", "focuses", "upon", "cumulative", "or", "final", "performance", ".", "Bootstrapped", "DQN", "makes", "significant", "improvements", "to", "the", "cumulative", "rewards", "of", "DQN", "on", "Atari", ",", "as", "we", "display", "in", "Figure", "[", "reference", "]", ",", "while", "the", "peak", "performance", "is", "much", "more", "We", "found", "that", "using", "bootstrapped", "DQN", "without", "gradient", "normalization", "on", "each", "head", "typically", "learned", "even", "faster", "than", "our", "implementation", "with", "rescaling", ",", "but", "it", "was", "somewhat", "prone", "to", "premature", "and", "suboptimal", "convergence", ".", "We", "present", "an", "example", "of", "this", "phenomenon", "in", "Figure", "[", "reference", "]", ".", "We", "found", "that", ",", "in", "order", "to", "better", "the", "benchmark", "‘", "‘", "best", "’", "’", "policies", "reported", "by", "DQN", ",", "it", "was", "very", "helpful", "for", "us", "to", "use", "the", "gradient", "normalization", ".", "However", ",", "it", "is", "not", "entirely", "clear", "whether", "this", "represents", "an", "improvement", "for", "all", "settings", ".", "In", "Figures", "[", "reference", "]", "and", "[", "reference", "]", "we", "present", "the", "cumulative", "rewards", "of", "the", "same", "algorithms", "on", "Beam", "Rider", ".", "[", "b", "]", "0.45", "[", "b", "]", "0.45", "Where", "an", "RL", "system", "is", "deployed", "to", "learn", "with", "real", "interactions", ",", "cumulative", "rewards", "present", "a", "better", "measure", "for", "performance", ".", "In", "these", "settings", "the", "benefits", "of", "gradient", "normalization", "are", "less", "clear", ".", "However", ",", "even", "with", "normalization", "bootstrapped", "DQN", "significantly", "outperforms", "DQN", "in", "terms", "of", "cumulative", "rewards", ".", "This", "is", "reflected", "most", "clearly", "in", "Figure", "[", "reference", "]", "and", "Table", "[", "reference", "]", ".", "subsection", ":", "Sharing", "data", "in", "bootstrap", "heads", "In", "this", "setting", "all", "network", "heads", "share", "all", "the", "data", ",", "so", "they", "are", "not", "actually", "a", "traditional", "bootstrap", "at", "all", ".", "This", "is", "different", "from", "the", "regression", "task", "in", "Section", "[", "reference", "]", ",", "where", "bootstrapped", "data", "was", "essential", "to", "obtain", "meaningful", "uncertainty", "estimates", ".", "We", "have", "several", "theories", "for", "why", "the", "networks", "maintain", "significant", "diversity", "even", "without", "data", "bootstrapping", "in", "this", "setting", ".", "We", "build", "upon", "the", "intuition", "of", "Section", "[", "reference", "]", ".", "First", ",", "they", "all", "train", "on", "different", "target", "networks", ".", "This", "means", "that", "even", "when", "facing", "the", "same", "datapoint", "this", "can", "still", "lead", "to", "drastically", "different", "Q", "-", "value", "updates", ".", "Second", ",", "Atari", "is", "a", "deterministic", "environment", ",", "any", "transition", "observation", "is", "the", "unique", "correct", "datapoint", "for", "this", "setting", ".", "Third", ",", "the", "networks", "are", "deep", "and", "initialized", "from", "different", "random", "values", "so", "they", "will", "likely", "find", "quite", "diverse", "generalization", "even", "when", "they", "agree", "on", "given", "data", ".", "Finally", ",", "since", "all", "variants", "of", "DQN", "take", "many", "many", "frames", "to", "update", "their", "policy", ",", "it", "is", "likely", "that", "even", "using", "they", "would", "still", "populate", "their", "replay", "memory", "with", "identical", "datapoints", ".", "This", "means", "using", "to", "save", "on", "minibatch", "passes", "seems", "like", "a", "reasonable", "compromise", "and", "it", "does", "n’t", "seem", "to", "negatively", "affect", "performance", "too", "much", "in", "this", "setting", ".", "More", "research", "is", "needed", "to", "examine", "exactly", "where", "/", "when", "this", "data", "sharing", "is", "important", ".", "subsection", ":", "Results", "tables", "In", "Table", "[", "reference", "]", "the", "average", "score", "achieved", "by", "the", "agents", "during", "the", "most", "successful", "evaluation", "period", ",", "compared", "to", "human", "performance", "and", "a", "uniformly", "random", "policy", ".", "DQN", "is", "our", "implementation", "of", "DQN", "with", "the", "hyperparameters", "specified", "above", ",", "using", "the", "double", "Q", "-", "Learning", "update", ".", ".", "We", "find", "that", "peak", "final", "performance", "is", "similar", "under", "bootstrapped", "DQN", "to", "previous", "benchmarks", ".", "To", "compare", "the", "benefits", "of", "exploration", "via", "bootstrapped", "DQN", "we", "benchmark", "our", "performance", "against", "the", "most", "similar", "prior", "work", "on", "incentivizing", "exploration", "in", "Atari", ".", "To", "do", "this", ",", "we", "compute", "the", "AUC", "-", "100", "measure", "specified", "in", "this", "work", ".", "We", "present", "these", "results", "in", "Table", "[", "reference", "]", "compare", "to", "their", "best", "performing", "strategy", "as", "well", "as", "their", "implementation", "of", "DQN", ".", "Importantly", ",", "bootstrapped", "DQN", "outperforms", "this", "prior", "work", "significantly", ".", "We", "now", "compare", "our", "method", "against", "the", "results", "in", ".", "In", "this", "paper", "they", "introduce", "a", "new", "measure", "of", "performance", "called", "AUC", "-", "100", ",", "which", "is", "something", "similar", "to", "normalized", "cumulative", "rewards", "up", "to", "20", "million", "frames", ".", "Table", "[", "reference", "]", "displays", "the", "results", "for", "our", "reference", "DQN", "and", "bootstrapped", "DQN", "as", "Boot", "-", "DQN", ".", "We", "reproduce", "their", "reference", "results", "for", "DQN", "as", "DQN", "*", "and", "their", "best", "performing", "algorithm", ",", "Dynamic", "AE", ".", "We", "also", "present", "bootstrapped", "DQN", "without", "head", "rescaling", "as", "Boot", "-", "DQN", "+", ".", "We", "see", "that", ",", "on", "average", ",", "both", "bootstrapped", "DQN", "implementations", "outperform", "Dynamic", "AE", ",", "the", "best", "algorithm", "from", "previous", "work", ".", "The", "only", "game", "in", "which", "Dynamic", "AE", "produces", "best", "results", "is", "Bowling", ",", "but", "this", "difference", "in", "Bowling", "is", "dominated", "by", "the", "implementation", "of", "DQN", "*", "vs", "DQN", ".", "Bootstrapped", "DQN", "still", "gives", "over", "100", "%", "improvement", "over", "its", "relevant", "DQN", "baseline", ".", "Overall", "it", "is", "clear", "that", "Boot", "-", "DQN", "+", "(", "bootstrapped", "DQN", "without", "rescaling", ")", "performs", "best", "in", "terms", "of", "AUC", "-", "100", "metric", ".", "Averaged", "across", "the", "14", "games", "it", "is", "over", "50", "%", "better", "than", "the", "next", "best", "competitor", ",", "which", "is", "bootstrapped", "DQN", "with", "gradient", "normalization", ".", "However", ",", "in", "terms", "of", "peak", "performance", "over", "200", "m", "frames", "Boot", "-", "DQN", "generally", "reached", "higher", "scores", ".", "Boot", "-", "DQN", "+", "sometimes", "plateaud", "early", "as", "in", "Figure", "[", "reference", "]", ".", "This", "highlights", "an", "important", "distinction", "between", "evaluation", "based", "on", "best", "learned", "policy", "versus", "cumulative", "rewards", ",", "as", "we", "discuss", "in", "Appendix", "[", "reference", "]", ".", "Bootstrapped", "DQN", "displays", "the", "biggest", "improvements", "over", "DQN", "when", "doing", "well", "during", "learning", "is", "important", "."], "entity": [[2, 4, ["Deep", "Exploration"], "Task"], [5, 7, ["Bootstrapped", "DQN"], "Method"], [7, 9, ["Efficient", "exploration"], "Task"], [14, 16, ["reinforcement", "learning"], "Task"], [17, 18, ["RL"], "Task"], [21, 23, ["dithering", "strategies"], "Method"], [24, 25, ["exploration"], "Task"], [28, 30, ["-", "greedy"], "Method"], [69, 72, ["Randomized", "value", "functions"], "Method"], [78, 79, ["exploration"], "Task"], [80, 81, ["generalization"], "Task"], [104, 106, ["bootstrapped", "DQN"], "Method"], [110, 112, ["bootstrapped", "DQN"], "Method"], [114, 116, ["deep", "exploration"], "Method"], [117, 120, ["deep", "neural", "networks"], "Method"], [123, 124, ["learning"], "Task"], [126, 128, ["dithering", "strategy"], "Method"], [131, 134, ["Arcade", "Learning", "Environment"], "Method"], [134, 136, ["bootstrapped", "DQN"], "Method"], [138, 140, ["learning", "speed"], "Metric"], [184, 186, ["planning", "problems"], "Task"], [214, 215, ["exploration"], "Task"], [216, 217, ["exploitation"], "Task"], [284, 290, ["temporally", "extended", "(", "deep", ")", "exploration"], "Method"], [291, 293, ["learning", "times"], "Metric"], [313, 315, ["deep", "exploration"], "Task"], [324, 327, ["Markov", "decision", "processes"], "Task"], [328, 329, ["MDPs"], "Method"], [340, 344, ["computationally", "intractable", "planning", "tasks"], "Task"], [375, 378, ["statistically", "inefficient", "strategies"], "Method"], [379, 380, ["exploration"], "Task"], [401, 403, ["dithering", "strategies"], "Method"], [406, 408, ["-", "greedy"], "Method"], [452, 454, ["efficient", "exploration"], "Task"], [456, 458, ["Thompson", "sampling"], "Method"], [497, 500, ["randomized", "value", "functions"], "Method"], [505, 507, ["Thompson", "sampling"], "Method"], [543, 548, ["non", "-", "linear", "generalization", "methods"], "Method"], [550, 553, ["deep", "neural", "networks"], "Method"], [558, 559, ["bootstrap"], "Method"], [560, 562, ["random", "initialization"], "Method"], [565, 567, ["uncertainty", "estimates"], "Metric"], [568, 570, ["neural", "networks"], "Method"], [572, 574, ["computational", "cost"], "Metric"], [575, 577, ["Bootstrapped", "DQN"], "Method"], [579, 581, ["uncertainty", "estimates"], "Method"], [583, 588, ["(", "and", "deep", ")", "exploration"], "Task"], [597, 600, ["large", "scale", "problems"], "Task"], [606, 608, ["deep", "exploration"], "Task"], [609, 611, ["Bootstrapped", "DQN"], "Method"], [613, 615, ["learning", "times"], "Metric"], [641, 642, ["DQN"], "Method"], [647, 649, ["neural", "networks"], "Method"], [649, 652, ["Deep", "neural", "networks"], "Method"], [653, 655, ["DNN", ")"], "Method"], [672, 674, ["exploration", "strategy"], "Method"], [682, 688, ["DNN", "representation", "of", "the", "value", "function"], "Method"], [700, 702, ["value", "estimates"], "Task"], [715, 718, ["neural", "network", "literature"], "Method"], [725, 727, ["uncertainty", "quantification"], "Task"], [729, 732, ["parametric", "Bayesian", "inference"], "Method"], [743, 745, ["random", "initialization"], "Method"], [765, 766, ["uncertainty"], "Task"], [767, 768, ["DNNs"], "Task"], [780, 782, ["sample", "distribution"], "Method"], [790, 791, ["bootstrap"], "Method"], [799, 800, ["estimator"], "Method"], [807, 809, ["bootstrapped", "distribution"], "Method"], [837, 838, ["bootstrap"], "Method"], [875, 880, ["large", "and", "deep", "neural", "network"], "Method"], [886, 888, ["shared", "architecture"], "Method"], [921, 923, ["shared", "network"], "Method"], [925, 928, ["joint", "feature", "representation"], "Method"], [951, 952, ["bootstrap"], "Method"], [959, 963, ["forward", "/", "backward", "pass"], "Method"], [971, 975, ["data", "-", "dependent", "dropout"], "Method"], [1010, 1012, ["uncertainty", "estimates"], "Method"], [1013, 1016, ["bootstrapped", "neural", "networks"], "Method"], [1018, 1020, ["regression", "task"], "Task"], [1027, 1035, ["fully", "-", "connected", "2", "-", "layer", "neural", "networks"], "Method"], [1037, 1040, ["rectified", "linear", "units"], "Method"], [1041, 1042, ["ReLU"], "Method"], [1083, 1085, ["uncertainty", "estimates"], "Task"], [1090, 1092, ["dropout", "approach"], "Method"], [1108, 1110, ["Bootstrapped", "DQN"], "Method"], [1172, 1180, ["parameterized", "estimate", "of", "the", "Q", "-", "value", "function"], "Method"], [1183, 1185, ["tabular", "encoding"], "Method"], [1189, 1191, ["neural", "network"], "Method"], [1197, 1201, ["Q", "-", "learning", "update"], "Method"], [1216, 1219, ["scalar", "learning", "rate"], "Metric"], [1236, 1240, ["Q", "-", "learning", "update"], "Method"], [1241, 1242, ["stability"], "Task"], [1243, 1244, ["DQN"], "Method"], [1268, 1270, ["target", "network"], "Method"], [1277, 1279, ["learning", "network"], "Method"], [1300, 1302, ["Bootstrapped", "DQN"], "Method"], [1303, 1304, ["DQN"], "Method"], [1314, 1315, ["bootstrap"], "Method"], [1323, 1325, ["bootstrapped", "DQN"], "Method"], [1328, 1332, ["Q", "-", "value", "function"], "Method"], [1363, 1366, ["Thompson", "sampling", "heuristic"], "Method"], [1367, 1368, ["RL"], "Task"], [1371, 1378, ["temporally", "extended", "(", "or", "deep", ")", "exploration"], "Task"], [1387, 1389, ["bootstrapped", "estimates"], "Method"], [1391, 1395, ["Q", "-", "value", "function"], "Method"], [1410, 1414, ["value", "function", "function", "heads"], "Method"], [1440, 1442, ["TD", "estimates"], "Method"], [1471, 1473, ["bootstrap", "sample"], "Method"], [1496, 1498, ["bootstrapped", "DQN"], "Method"], [1511, 1514, ["temporally", "extended", "exploration"], "Task"], [1518, 1520, ["reinforcement", "learning"], "Task"], [1529, 1530, ["MDPs"], "Method"], [1533, 1535, ["exploration", "strategy"], "Method"], [1538, 1540, ["dynamic", "programming"], "Method"], [1566, 1569, ["generalization", "and", "planning"], "Task"], [1571, 1572, ["exploration"], "Task"], [1585, 1587, ["exploration", "strategies"], "Method"], [1593, 1595, ["exploration", "strategies"], "Method"], [1649, 1653, ["finite", "-", "armed", "bandits"], "Task"], [1662, 1663, ["bandits"], "Method"], [1675, 1677, ["OFU", "methods"], "Method"], [1681, 1682, ["RL"], "Task"], [1703, 1704, ["DQN"], "Method"], [1725, 1727, ["deterministic", "problems"], "Task"], [1744, 1746, ["bootstrapped", "DQN"], "Method"], [1759, 1763, ["balancing", "exploration", "with", "exploitation"], "Task"], [1766, 1768, ["Thompson", "sampling"], "Method"], [1770, 1772, ["bandit", "algorithm"], "Method"], [1798, 1801, ["Thompson", "sampling", "principle"], "Method"], [1802, 1803, ["RL"], "Task"], [1818, 1820, ["Thompson", "sampling"], "Method"], [1821, 1822, ["RL"], "Task"], [1847, 1849, ["deep", "exploration"], "Task"], [1852, 1853, ["PSRL"], "Method"], [1874, 1875, ["MDP"], "Method"], [1889, 1891, ["bootstrapped", "DQN"], "Method"], [1896, 1897, ["exploration"], "Task"], [1913, 1915, ["RLSVI", "algorithm"], "Method"], [1926, 1927, ["RLSVI"], "Method"], [1932, 1933, ["guarantees"], "Metric"], [1950, 1956, ["linear", "representation", "of", "the", "value", "function"], "Method"], [1969, 1972, ["generalization", "and", "exploration"], "Task"], [1975, 1979, ["nonlinear", "value", "function", "representation"], "Method"], [1996, 1999, ["low", "computational", "cost"], "Metric"], [2007, 2009, ["Deep", "Exploration"], "Task"], [2009, 2011, ["Uncertainty", "estimates"], "Task"], [2026, 2027, ["bandits"], "Method"], [2031, 2033, ["directed", "exploration"], "Method"], [2035, 2036, ["dithering"], "Method"], [2038, 2040, ["efficient", "algorithms"], "Method"], [2044, 2045, ["RL"], "Task"], [2050, 2052, ["directed", "exploration"], "Task"], [2060, 2061, ["exploration"], "Task"], [2066, 2068, ["Deep", "exploration"], "Task"], [2069, 2070, ["exploration"], "Task"], [2098, 2099, ["exploration"], "Task"], [2101, 2103, ["bandit", "problems"], "Task"], [2125, 2126, ["exploitation"], "Task"], [2156, 2157, ["exploration"], "Task"], [2179, 2181, ["deterministic", "chain"], "Method"], [2190, 2191, ["MDP"], "Method"], [2258, 2260, ["consistent", "strategy"], "Method"], [2271, 2276, ["planning", "and", "look", "ahead", "trees"], "Task"], [2278, 2280, ["algorithmic", "approaches"], "Method"], [2283, 2284, ["MDP"], "Method"], [2333, 2335, ["bandit", "algorithms"], "Method"], [2351, 2353, ["deep", "exploration"], "Method"], [2382, 2384, ["deep", "exploration"], "Task"], [2399, 2401, ["deep", "exploration"], "Task"], [2436, 2438, ["toy", "problems"], "Task"], [2479, 2482, ["finite", "tabular", "MDP"], "Method"], [2492, 2493, ["MDP"], "Method"], [2522, 2523, ["generalization"], "Task"], [2543, 2545, ["bootstrapped", "DQN"], "Method"], [2558, 2560, ["bootstrapped", "DQN"], "Method"], [2564, 2566, ["ensemble", "policy"], "Method"], [2576, 2578, ["optimal", "policy"], "Method"], [2599, 2601, ["learning", "algorithm"], "Method"], [2611, 2613, ["median", "time"], "Metric"], [2636, 2639, ["shallow", "exploration", "strategy"], "Method"], [2641, 2643, ["bootstrapped", "DQN"], "Method"], [2659, 2661, ["bootstrapped", "DQN"], "Method"], [2662, 2664, ["deep", "exploration"], "Task"], [2665, 2667, ["Bootstrapped", "DQN"], "Method"], [2677, 2679, ["algorithm", "PSRL"], "Method"], [2683, 2686, ["bootstrapped", "neural", "network"], "Method"], [2696, 2697, ["PSRL"], "Method"], [2698, 2700, ["bootstrapped", "DQN"], "Method"], [2703, 2705, ["value", "function"], "Method"], [2719, 2720, ["RLSVI"], "Method"], [2731, 2733, ["neural", "network"], "Method"], [2735, 2738, ["linear", "value", "function"], "Method"], [2739, 2740, ["bootstrap"], "Method"], [2742, 2744, ["Gaussian", "sampling"], "Method"], [2749, 2751, ["linear", "setting"], "Task"], [2754, 2756, ["nonlinear", "approach"], "Method"], [2783, 2785, ["Bootstrapped", "DQN"], "Method"], [2787, 2789, ["random", "initialization"], "Method"], [2819, 2824, ["large", "and", "deep", "neural", "networks"], "Method"], [2880, 2882, ["random", "initialization"], "Method"], [2885, 2887, ["supervised", "learning"], "Method"], [2888, 2889, ["bandits"], "Method"], [2912, 2914, ["stochastic", "minibatch"], "Method"], [2915, 2918, ["flexible", "nonlinear", "representations"], "Method"], [2937, 2939, ["Bootstrapped", "DQN"], "Method"], [3040, 3042, ["TD", "bootstrapping"], "Method"], [3054, 3056, ["deep", "exploration"], "Task"], [3097, 3099, ["bootstrapped", "DQN"], "Method"], [3100, 3102, ["deep", "exploration"], "Task"], [3110, 3112, ["stochastic", "MDP"], "Task"], [3120, 3123, ["Arcade", "Learning", "Environment"], "Method"], [3130, 3132, ["Atari", "games"], "Material"], [3134, 3137, ["Arcade", "Learning", "Environment"], "Task"], [3165, 3167, ["Atari", "games"], "Material"], [3178, 3180, ["optimal", "policy"], "Method"], [3191, 3193, ["dithering", "strategies"], "Method"], [3199, 3201, ["bootstrapped", "DQN"], "Method"], [3211, 3213, ["Bootstrapped", "DQN"], "Method"], [3219, 3220, ["DQN"], "Method"], [3248, 3249, ["learning"], "Task"], [3257, 3259, ["network", "architecture"], "Method"], [3268, 3270, ["network", "structure"], "Method"], [3274, 3276, ["convolutional", "structure"], "Method"], [3277, 3278, ["DQN"], "Method"], [3287, 3289, ["convolutional", "layer"], "Method"], [3307, 3308, ["DDQN"], "Method"], [3354, 3356, ["bootstrapped", "DQN"], "Method"], [3368, 3369, ["DQN"], "Method"], [3414, 3416, ["computational", "cost"], "Metric"], [3418, 3419, ["DQN"], "Method"], [3429, 3431, ["bootstrapped", "DQN"], "Method"], [3433, 3435, ["game", "Breakout"], "Task"], [3447, 3448, ["learning"], "Task"], [3462, 3464, ["bootstrapped", "DQN"], "Method"], [3477, 3480, ["shared", "network", "architecture"], "Method"], [3485, 3487, ["combined", "network"], "Method"], [3488, 3489, ["backpropagation"], "Method"], [3491, 3493, ["network", "heads"], "Method"], [3495, 3498, ["shared", "convolutional", "network"], "Method"], [3501, 3503, ["learning", "rate"], "Metric"], [3517, 3523, ["premature", "and", "sub", "-", "optimal", "convergence"], "Task"], [3542, 3544, ["early", "learning"], "Task"], [3607, 3609, ["shared", "minibatch"], "Method"], [3637, 3638, ["DQN"], "Method"], [3670, 3671, ["Atari"], "Material"], [3683, 3685, ["bootstrapped", "DQN"], "Method"], [3688, 3690, ["computational", "speed"], "Metric"], [3700, 3701, ["exploration"], "Task"], [3702, 3703, ["Atari"], "Material"], [3706, 3708, ["Bootstrapped", "DQN"], "Method"], [3710, 3711, ["exploration"], "Task"], [3713, 3715, ["Atari", "games"], "Material"], [3724, 3726, ["bootstrapped", "DQN"], "Method"], [3751, 3752, ["DQN"], "Method"], [3755, 3757, ["bootstrapped", "DQN"], "Method"], [3761, 3763, ["Bootstrapped", "DQN"], "Method"], [3769, 3770, ["Amidar"], "Method"], [3771, 3772, ["DQN"], "Method"], [3777, 3779, ["Beam", "Rider"], "Task"], [3780, 3782, ["Battle", "Zone"], "Task"], [3783, 3784, ["DQN"], "Method"], [3793, 3795, ["learning", "time"], "Metric"], [3808, 3810, ["bootstrapped", "DQN"], "Method"], [3830, 3832, ["Bootstrapped", "DQN"], "Method"], [3843, 3844, ["DQN"], "Method"], [3850, 3852, ["bootstrapped", "DQN"], "Method"], [3863, 3865, ["Atari", "games"], "Material"], [3866, 3868, ["deep", "exploration"], "Task"], [3885, 3886, ["Frostbite"], "Method"], [3887, 3889, ["bootstrapped", "DQN"], "Method"], [3896, 3897, ["DQN"], "Method"], [3907, 3910, ["Montezuma", "’s", "Revenge"], "Task"], [3911, 3913, ["bootstrapped", "DQN"], "Method"], [3947, 3948, ["exploration"], "Task"], [3979, 3981, ["Bootstrapped", "DQN"], "Method"], [3988, 3989, ["DQN"], "Method"], [3997, 3998, ["DQN"], "Method"], [4003, 4004, ["score"], "Metric"], [4015, 4016, ["exploration"], "Task"], [4018, 4020, ["bootstrapped", "DQN"], "Method"], [4021, 4022, ["DQN"], "Method"], [4030, 4032, ["cumulative", "rewards"], "Metric"], [4033, 4034, ["learning"], "Task"], [4049, 4052, ["fully", "random", "policy"], "Method"], [4065, 4066, ["exploration"], "Task"], [4067, 4068, ["Atari"], "Material"], [4071, 4074, ["AUC", "-", "20"], "Metric"], [4105, 4106, ["DQN"], "Method"], [4120, 4122, ["bootstrapped", "DQN"], "Method"], [4144, 4146, ["bootstrapped", "DQN"], "Method"], [4154, 4156, ["bootstrapped", "DQN"], "Method"], [4157, 4159, ["deep", "exploration"], "Task"], [4160, 4161, ["Atari"], "Material"], [4214, 4217, ["-", "greedy", "strategies"], "Method"], [4235, 4237, ["deep", "exploration"], "Method"], [4241, 4242, ["learning"], "Task"], [4249, 4250, ["generalization"], "Task"], [4252, 4253, ["exploration"], "Method"], [4254, 4256, ["bootstrapped", "DQN"], "Method"], [4262, 4264, ["exploitative", "policy"], "Method"], [4274, 4276, ["ensemble", "policy"], "Method"], [4303, 4305, ["ensemble", "policy"], "Method"], [4309, 4311, ["individual", "policy"], "Method"], [4329, 4331, ["optimal", "policy"], "Task"], [4336, 4338, ["bootstrapped", "DQN"], "Method"], [4369, 4371, ["ensemble", "policy"], "Method"], [4414, 4416, ["bootstrapped", "DQN"], "Method"], [4421, 4423, ["reinforcement", "learning"], "Task"], [4424, 4426, ["complex", "environments"], "Task"], [4431, 4432, ["bootstrap"], "Method"], [4435, 4437, ["uncertainty", "estimates"], "Task"], [4438, 4441, ["deep", "neural", "networks"], "Method"], [4442, 4444, ["Bootstrapped", "DQN"], "Method"], [4452, 4455, ["massive", "parallel", "systems"], "Task"], [4465, 4468, ["randomized", "value", "functions"], "Method"], [4475, 4476, ["exploration"], "Task"], [4482, 4483, ["generalization"], "Method"], [4484, 4485, ["exploration"], "Task"], [4499, 4501, ["neural", "networks"], "Method"], [4514, 4516, ["uncertainty", "methods"], "Method"], [4517, 4520, ["deep", "neural", "networks"], "Method"], [4583, 4585, ["generating", "distribution"], "Method"], [4599, 4600, ["dropout"], "Method"], [4614, 4617, ["dropout", "posterior", "estimate"], "Method"], [4662, 4663, ["dropout"], "Method"], [4669, 4672, ["neural", "network", "fit"], "Method"], [4674, 4675, ["bootstrap"], "Method"], [4683, 4685, ["dropout", "approximation"], "Method"], [4718, 4720, ["dropout", "scheme"], "Method"], [4760, 4763, ["dropout", "posterior", "approximations"], "Method"], [4800, 4802, ["neural", "network"], "Method"], [4803, 4805, ["MSE", "criterion"], "Metric"], [4833, 4836, ["bootstrapped", "neural", "networks"], "Method"], [4851, 4853, ["uncertainty", "estimates"], "Metric"], [4865, 4866, ["dropout"], "Method"], [4869, 4872, ["Gaussian", "process", "posterior"], "Method"], [4889, 4891, ["heteroskedastic", "variant"], "Method"], [4904, 4906, ["large", "networks"], "Method"], [4930, 4932, ["bootstrap", "approach"], "Method"], [4933, 4934, ["uncertainty"], "Task"], [4935, 4937, ["neural", "networks"], "Task"], [4946, 4948, ["statistical", "methodology"], "Method"], [4965, 4967, ["deep", "exploration"], "Method"], [4968, 4971, ["randomized", "value", "functions"], "Method"], [4977, 4980, ["approximate", "posterior", "estimator"], "Method"], [4981, 4984, ["deep", "neural", "networks"], "Method"], [4991, 4993, ["uncertainty", "estimates"], "Task"], [4994, 4996, ["neural", "networks"], "Method"], [5007, 5010, ["Bootstrapped", "uncertainty", "estimates"], "Method"], [5012, 5016, ["Q", "-", "value", "functions"], "Method"], [5021, 5022, ["dropout"], "Method"], [5028, 5030, ["supervised", "problem"], "Task"], [5032, 5035, ["random", "dropout", "masks"], "Method"], [5037, 5040, ["random", "target", "networks"], "Method"], [5050, 5054, ["temporally", "consistent", "target", "network"], "Method"], [5059, 5061, ["bootstrap", "estimates"], "Method"], [5106, 5108, ["deep", "exploration"], "Task"], [5123, 5125, ["Bootstrapped", "DQN"], "Method"], [5134, 5136, ["neural", "networks"], "Method"], [5148, 5150, ["neural", "network"], "Method"], [5151, 5152, ["heads"], "Method"], [5169, 5171, ["parameterisation", "issue"], "Task"], [5175, 5178, ["value", "function", "networks"], "Method"], [5203, 5205, ["bootstrap", "mask"], "Method"], [5268, 5270, ["masking", "distribution"], "Method"], [5323, 5325, ["ensemble", "method"], "Method"], [5326, 5328, ["Poisson", "masks"], "Method"], [5336, 5340, ["non", "-", "parameteric", "boostrap"], "Method"], [5349, 5356, ["Bayesian", "nonparametric", "posterior", "of", "a", "Dirichlet", "process"], "Method"], [5357, 5359, ["Bootstrapped", "DQN"], "Method"], [5364, 5367, ["Value", "function", "networks"], "Method"], [5376, 5378, ["replay", "buffer"], "Method"], [5438, 5441, ["value", "function", "network"], "Method"], [5489, 5491, ["deep", "exploration"], "Task"], [5493, 5495, ["Bootstrap", "methodology"], "Method"], [5499, 5501, ["bootstrapped", "DQN"], "Method"], [5503, 5505, ["complete", "networks"], "Method"], [5542, 5544, ["bootstrap", "heads"], "Method"], [5563, 5564, ["bootstrap"], "Method"], [5581, 5583, ["bootstrapped", "DQN"], "Method"], [5585, 5587, ["deep", "exploration"], "Task"], [5648, 5653, ["large", "and", "deep", "neural", "networks"], "Task"], [5675, 5676, ["exploration"], "Task"], [5691, 5693, ["network", "architecture"], "Method"], [5698, 5700, ["deep", "exploration"], "Task"], [5702, 5705, ["-", "greedy", "policy"], "Method"], [5723, 5725, ["ensemble", "policy"], "Method"], [5730, 5732, ["bootstrapped", "DQN"], "Method"], [5750, 5752, ["Thompson", "sampling"], "Method"], [5756, 5758, ["bootstrapped", "DQN"], "Method"], [5777, 5779, ["bootstrapped", "DQN"], "Method"], [5782, 5784, ["deep", "exploration"], "Task"], [5792, 5794, ["stochastic", "MDP"], "Method"], [5800, 5802, ["bootstrapped", "DQN"], "Method"], [5807, 5810, ["deep", ")", "exploration"], "Task"], [5827, 5829, ["finite", "MDP"], "Method"], [5852, 5854, ["chain", "system"], "Method"], [5906, 5907, ["MDP"], "Method"], [5927, 5929, ["Bootstrapped", "DQN"], "Method"], [5936, 5937, ["exploration"], "Task"], [5941, 5943, ["stochastic", "domains"], "Task"], [5954, 5956, ["bootstrapped", "DQN"], "Method"], [5961, 5963, ["tabular", "representation"], "Method"], [5965, 5966, ["MDP"], "Method"], [5979, 5980, ["exploration"], "Task"], [5981, 5982, ["dithering"], "Method"], [5987, 5988, ["optimism"], "Method"], [5989, 5991, ["posterior", "sampling"], "Method"], [6021, 6023, ["empirical", "regret"], "Metric"], [6038, 6040, ["empirical", "regret"], "Metric"], [6050, 6052, ["optimal", "policy"], "Method"], [6063, 6065, ["bootstrapped", "DQN"], "Method"], [6074, 6076, ["exploration", "schemes"], "Method"], [6078, 6079, ["PSRL"], "Method"], [6099, 6101, ["bootstrapped", "DQN"], "Method"], [6108, 6111, ["optimistic", "algorithm", "UCRL2"], "Method"], [6120, 6121, ["UCRL2"], "Method"], [6161, 6164, ["tabula", "rasa", "approaches"], "Method"], [6165, 6166, ["exploration"], "Task"], [6196, 6201, ["pre", "-", "eminent", "optimistic", "approach"], "Method"], [6202, 6203, ["UCRL2"], "Method"], [6205, 6207, ["posterior", "sampling"], "Method"], [6208, 6209, ["PSRL"], "Method"], [6213, 6215, ["heuristic", "approaches"], "Method"], [6218, 6220, ["optimistic", "initialization"], "Method"], [6221, 6224, ["Q", "-", "learning"], "Method"], [6243, 6244, ["MDP"], "Method"], [6258, 6260, ["optimistic", "algorithms"], "Method"], [6262, 6263, ["UCRL2"], "Method"], [6267, 6271, ["infinite", "-", "horizon", "MDPs"], "Task"], [6289, 6292, ["finite", "horizon", "MDPs"], "Task"], [6301, 6303, ["empirical", "regret"], "Metric"], [6308, 6310, ["bootstrapped", "DQN"], "Method"], [6324, 6325, ["UCRL2"], "Method"], [6357, 6358, ["regret"], "Metric"], [6382, 6386, ["near", "optimal", "sample", "complexity"], "Metric"], [6387, 6390, ["fixed", "horizon", "problems"], "Task"], [6408, 6411, ["worst", "case", "scaling"], "Task"], [6437, 6439, ["bootstrapped", "DQN"], "Method"], [6440, 6445, ["one", "-", "hot", "feature", "encodings"], "Method"], [6454, 6456, ["bootstrapped", "DQN"], "Method"], [6458, 6460, ["optimal", "policy"], "Method"], [6471, 6473, ["thermometer", "encoding"], "Method"], [6483, 6485, ["optimal", "policy"], "Method"], [6512, 6514, ["Atari", "games"], "Material"], [6530, 6531, ["emulator"], "Method"], [6565, 6567, ["raw", "scores"], "Metric"], [6569, 6571, ["convolutional", "part"], "Method"], [6605, 6610, ["convolutional", "(", "conv", ")", "layer"], "Method"], [6624, 6626, ["conv", "layer"], "Method"], [6638, 6640, ["conv", "layer"], "Method"], [6671, 6672, ["DQN"], "Method"], [6677, 6680, ["fully", "connected", "layer"], "Method"], [6686, 6689, ["fully", "connected", "layer"], "Method"], [6699, 6702, ["fully", "connected", "layers"], "Method"], [6704, 6707, ["Rectified", "Linear", "Units"], "Method"], [6708, 6709, ["ReLU"], "Method"], [6730, 6731, ["RMSProp"], "Method"], [6738, 6740, ["learning", "rate"], "Metric"], [6795, 6797, ["bootstrapped", "DQN"], "Method"], [6800, 6803, ["ensemble", "voting", "policy"], "Method"], [6840, 6841, ["DQN"], "Method"], [6843, 6844, ["training"], "Task"], [6847, 6850, ["-", "greedy", "policy"], "Method"], [6865, 6867, ["Gradient", "normalization"], "Method"], [6868, 6870, ["bootstrap", "heads"], "Method"], [6876, 6877, ["Atari"], "Material"], [6883, 6885, ["evaluation", "policy"], "Task"], [6915, 6917, ["Bootstrapped", "DQN"], "Method"], [6922, 6924, ["cumulative", "rewards"], "Metric"], [6925, 6926, ["DQN"], "Method"], [6927, 6928, ["Atari"], "Material"], [6949, 6951, ["bootstrapped", "DQN"], "Method"], [6952, 6954, ["gradient", "normalization"], "Method"], [6965, 6966, ["rescaling"], "Method"], [6976, 6977, ["convergence"], "Metric"], [7020, 7022, ["gradient", "normalization"], "Method"], [7058, 7060, ["Beam", "Rider"], "Task"], [7096, 7098, ["gradient", "normalization"], "Method"], [7111, 7112, ["DQN"], "Method"], [7115, 7117, ["cumulative", "rewards"], "Metric"], [7168, 7170, ["regression", "task"], "Task"], [7184, 7186, ["uncertainty", "estimates"], "Task"], [7200, 7202, ["data", "bootstrapping"], "Method"], [7224, 7226, ["target", "networks"], "Method"], [7302, 7303, ["DQN"], "Method"], [7368, 7370, ["data", "sharing"], "Task"], [7384, 7385, ["score"], "Metric"], [7402, 7405, ["uniformly", "random", "policy"], "Method"], [7406, 7407, ["DQN"], "Method"], [7411, 7412, ["DQN"], "Method"], [7420, 7425, ["double", "Q", "-", "Learning", "update"], "Method"], [7436, 7438, ["bootstrapped", "DQN"], "Method"], [7447, 7448, ["exploration"], "Task"], [7449, 7451, ["bootstrapped", "DQN"], "Method"], [7462, 7464, ["incentivizing", "exploration"], "Task"], [7465, 7466, ["Atari"], "Material"], [7474, 7478, ["AUC", "-", "100", "measure"], "Metric"], [7504, 7505, ["DQN"], "Method"], [7508, 7510, ["bootstrapped", "DQN"], "Method"], [7535, 7536, ["performance"], "Metric"], [7537, 7540, ["AUC", "-", "100"], "Method"], [7565, 7566, ["DQN"], "Method"], [7567, 7569, ["bootstrapped", "DQN"], "Method"], [7570, 7573, ["Boot", "-", "DQN"], "Method"], [7582, 7584, ["DQN", "*"], "Method"], [7590, 7592, ["Dynamic", "AE"], "Method"], [7596, 7598, ["bootstrapped", "DQN"], "Method"], [7602, 7606, ["Boot", "-", "DQN", "+"], "Method"], [7619, 7621, ["Dynamic", "AE"], "Method"], [7634, 7636, ["Dynamic", "AE"], "Method"], [7640, 7641, ["Bowling"], "Task"], [7646, 7647, ["Bowling"], "Task"], [7653, 7655, ["DQN", "*"], "Method"], [7656, 7657, ["DQN"], "Method"], [7658, 7660, ["Bootstrapped", "DQN"], "Method"], [7677, 7681, ["Boot", "-", "DQN", "+"], "Method"], [7692, 7696, ["AUC", "-", "100", "metric"], "Metric"], [7716, 7718, ["bootstrapped", "DQN"], "Method"], [7719, 7721, ["gradient", "normalization"], "Method"], [7735, 7736, ["DQN"], "Method"], [7741, 7742, ["Boot"], "Method"], [7743, 7745, ["DQN", "+"], "Method"], [7761, 7762, ["evaluation"], "Task"], [7780, 7782, ["Bootstrapped", "DQN"], "Method"], [7787, 7788, ["DQN"], "Method"], [7792, 7793, ["learning"], "Task"], [60, 61, ["RL"], "Task"], [153, 155, ["reinforcement", "learning"], "Task"], [156, 157, ["RL"], "Task"], [188, 189, ["RL"], "Task"], [302, 303, ["RL"], "Task"], [371, 372, ["RL"], "Task"], [665, 667, ["reinforcement", "learning"], "Task"], [741, 742, ["bootstrap"], "Method"], [770, 771, ["bootstrap"], "Method"], [827, 828, ["bootstrap"], "Method"], [871, 872, ["bootstrap"], "Method"], [917, 918, ["bootstrap"], "Method"], [1292, 1293, ["DQN"], "Method"], [1454, 1455, ["bootstrap"], "Method"], [1562, 1563, ["RL"], "Task"], [1667, 1668, ["RL"], "Task"], [1733, 1734, ["Atari"], "Material"], [1994, 1995, ["RL"], "Task"], [2115, 2116, ["RL"], "Task"], [2337, 2338, ["RL"], "Task"], [2348, 2349, ["RL"], "Task"], [2517, 2518, ["DQN"], "Method"], [2538, 2539, ["DQN"], "Method"], [2552, 2553, ["DQN"], "Method"], [3283, 3284, ["bootstrap"], "Method"], [3332, 3334, ["bootstrapped", "DQN"], "Method"], [3365, 3366, ["bootstrap"], "Method"], [3402, 3403, ["bootstrap"], "Method"], [3568, 3569, ["bootstrap"], "Method"], [3692, 3693, ["DQN"], "Method"], [3728, 3729, ["DQN"], "Method"], [3816, 3817, ["DQN"], "Method"], [3922, 3923, ["DQN"], "Method"], [4334, 4335, ["DQN"], "Method"], [4477, 4479, ["Bootstrapped", "DQN"], "Method"], [5044, 5045, ["bootstrap"], "Method"], [5045, 5046, ["DQN"], "Method"], [5073, 5074, ["bootstrap"], "Method"], [5111, 5113, ["Bootstrapped", "DQN"], "Method"], [5198, 5200, ["bootstrapped", "DQN"], "Method"], [5302, 5303, ["bootstrap"], "Method"], [5418, 5419, ["bootstrap"], "Method"], [5482, 5483, ["bootstrap"], "Method"], [5747, 5748, ["bootstrap"], "Method"], [5813, 5814, ["RL"], "Task"], [5939, 5940, ["RL"], "Task"], [6502, 6503, ["Atari"], "Material"], [6874, 6875, ["RL"], "Task"], [6903, 6904, ["RL"], "Task"], [7009, 7010, ["DQN"], "Method"], [7071, 7072, ["RL"], "Task"], [7107, 7109, ["bootstrapped", "DQN"], "Method"], [7139, 7140, ["bootstrap"], "Method"], [7159, 7160, ["bootstrap"], "Method"], [7250, 7251, ["Atari"], "Material"], [7580, 7581, ["DQN"], "Method"], [7615, 7617, ["bootstrapped", "DQN"], "Method"], [7669, 7670, ["DQN"], "Method"], [7682, 7684, ["bootstrapped", "DQN"], "Method"], [7733, 7734, ["Boot"], "Method"]]} {"doc_id": "1762baa638866a13dcc6d146fd5a49b36cbd9c30", "text": ["document", ":", "Semi", "-", "Supervised", "Classification", "with", "Graph", "Convolutional", "Networks", "We", "present", "a", "scalable", "approach", "for", "semi", "-", "supervised", "learning", "on", "graph", "-", "structured", "data", "that", "is", "based", "on", "an", "efficient", "variant", "of", "convolutional", "neural", "networks", "which", "operate", "directly", "on", "graphs", ".", "We", "motivate", "the", "choice", "of", "our", "convolutional", "architecture", "via", "a", "localized", "first", "-", "order", "approximation", "of", "spectral", "graph", "convolutions", ".", "Our", "model", "scales", "linearly", "in", "the", "number", "of", "graph", "edges", "and", "learns", "hidden", "layer", "representations", "that", "encode", "both", "local", "graph", "structure", "and", "features", "of", "nodes", ".", "In", "a", "number", "of", "experiments", "on", "citation", "networks", "and", "on", "a", "knowledge", "graph", "dataset", "we", "demonstrate", "that", "our", "approach", "outperforms", "related", "methods", "by", "a", "significant", "margin", ".", "section", ":", "Introduction", "We", "consider", "the", "problem", "of", "classifying", "nodes", "(", "such", "as", "documents", ")", "in", "a", "graph", "(", "such", "as", "a", "citation", "network", ")", ",", "where", "labels", "are", "only", "available", "for", "a", "small", "subset", "of", "nodes", ".", "This", "problem", "can", "be", "framed", "as", "graph", "-", "based", "semi", "-", "supervised", "learning", ",", "where", "label", "information", "is", "smoothed", "over", "the", "graph", "via", "some", "form", "of", "explicit", "graph", "-", "based", "regularization", "zhu2003semi", ",", "zhou2004learning", ",", "belkin2006manifold", ",", "weston2012deep", ",", "e.g.", "by", "using", "a", "graph", "Laplacian", "regularization", "term", "in", "the", "loss", "function", ":", "Here", ",", "denotes", "the", "supervised", "loss", "w.r.t", ".", "the", "labeled", "part", "of", "the", "graph", ",", "can", "be", "a", "neural", "network", "-", "like", "differentiable", "function", ",", "is", "a", "weighing", "factor", "and", "is", "a", "matrix", "of", "node", "feature", "vectors", ".", "denotes", "the", "unnormalized", "graph", "Laplacian", "of", "an", "undirected", "graph", "with", "nodes", ",", "edges", ",", "an", "adjacency", "matrix", "(", "binary", "or", "weighted", ")", "and", "a", "degree", "matrix", ".", "The", "formulation", "of", "Eq", ".", "[", "reference", "]", "relies", "on", "the", "assumption", "that", "connected", "nodes", "in", "the", "graph", "are", "likely", "to", "share", "the", "same", "label", ".", "This", "assumption", ",", "however", ",", "might", "restrict", "modeling", "capacity", ",", "as", "graph", "edges", "need", "not", "necessarily", "encode", "node", "similarity", ",", "but", "could", "contain", "additional", "information", ".", "In", "this", "work", ",", "we", "encode", "the", "graph", "structure", "directly", "using", "a", "neural", "network", "model", "and", "train", "on", "a", "supervised", "target", "for", "all", "nodes", "with", "labels", ",", "thereby", "avoiding", "explicit", "graph", "-", "based", "regularization", "in", "the", "loss", "function", ".", "Conditioning", "on", "the", "adjacency", "matrix", "of", "the", "graph", "will", "allow", "the", "model", "to", "distribute", "gradient", "information", "from", "the", "supervised", "loss", "and", "will", "enable", "it", "to", "learn", "representations", "of", "nodes", "both", "with", "and", "without", "labels", ".", "Our", "contributions", "are", "two", "-", "fold", ".", "Firstly", ",", "we", "introduce", "a", "simple", "and", "well", "-", "behaved", "layer", "-", "wise", "propagation", "rule", "for", "neural", "network", "models", "which", "operate", "directly", "on", "graphs", "and", "show", "how", "it", "can", "be", "motivated", "from", "a", "first", "-", "order", "approximation", "of", "spectral", "graph", "convolutions", "hammond2011wavelets", ".", "Secondly", ",", "we", "demonstrate", "how", "this", "form", "of", "a", "graph", "-", "based", "neural", "network", "model", "can", "be", "used", "for", "fast", "and", "scalable", "semi", "-", "supervised", "classification", "of", "nodes", "in", "a", "graph", ".", "Experiments", "on", "a", "number", "of", "datasets", "demonstrate", "that", "our", "model", "compares", "favorably", "both", "in", "classification", "accuracy", "and", "efficiency", "(", "measured", "in", "wall", "-", "clock", "time", ")", "against", "state", "-", "of", "-", "the", "-", "art", "methods", "for", "semi", "-", "supervised", "learning", ".", "section", ":", "Fast", "Approximate", "Convolutions", "on", "Graphs", "In", "this", "section", ",", "we", "provide", "theoretical", "motivation", "for", "a", "specific", "graph", "-", "based", "neural", "network", "model", "that", "we", "will", "use", "in", "the", "rest", "of", "this", "paper", ".", "We", "consider", "a", "multi", "-", "layer", "Graph", "Convolutional", "Network", "(", "GCN", ")", "with", "the", "following", "layer", "-", "wise", "propagation", "rule", ":", "Here", ",", "is", "the", "adjacency", "matrix", "of", "the", "undirected", "graph", "with", "added", "self", "-", "connections", ".", "is", "the", "identity", "matrix", ",", "and", "is", "a", "layer", "-", "specific", "trainable", "weight", "matrix", ".", "denotes", "an", "activation", "function", ",", "such", "as", "the", ".", "is", "the", "matrix", "of", "activations", "in", "the", "layer", ";", ".", "In", "the", "following", ",", "we", "show", "that", "the", "form", "of", "this", "propagation", "rule", "can", "be", "motivated", "via", "a", "first", "-", "order", "approximation", "of", "localized", "spectral", "filters", "on", "graphs", "hammond2011wavelets", ",", "defferrard2016convolutional", ".", "subsection", ":", "Spectral", "Graph", "Convolutions", "We", "consider", "spectral", "convolutions", "on", "graphs", "defined", "as", "the", "multiplication", "of", "a", "signal", "(", "a", "scalar", "for", "every", "node", ")", "with", "a", "filter", "parameterized", "by", "in", "the", "Fourier", "domain", ",", "i.e.", ":", "where", "is", "the", "matrix", "of", "eigenvectors", "of", "the", "normalized", "graph", "Laplacian", ",", "with", "a", "diagonal", "matrix", "of", "its", "eigenvalues", "and", "being", "the", "graph", "Fourier", "transform", "of", ".", "We", "can", "understand", "as", "a", "function", "of", "the", "eigenvalues", "of", ",", "i.e.", ".", "Evaluating", "Eq", ".", "[", "reference", "]", "is", "computationally", "expensive", ",", "as", "multiplication", "with", "the", "eigenvector", "matrix", "is", ".", "Furthermore", ",", "computing", "the", "eigendecomposition", "of", "in", "the", "first", "place", "might", "be", "prohibitively", "expensive", "for", "large", "graphs", ".", "To", "circumvent", "this", "problem", ",", "it", "was", "suggested", "in", "that", "can", "be", "well", "-", "approximated", "by", "a", "truncated", "expansion", "in", "terms", "of", "Chebyshev", "polynomials", "up", "to", "order", ":", "with", "a", "rescaled", ".", "denotes", "the", "largest", "eigenvalue", "of", ".", "is", "now", "a", "vector", "of", "Chebyshev", "coefficients", ".", "The", "Chebyshev", "polynomials", "are", "recursively", "defined", "as", ",", "with", "and", ".", "The", "reader", "is", "referred", "to", "for", "an", "in", "-", "depth", "discussion", "of", "this", "approximation", ".", "Going", "back", "to", "our", "definition", "of", "a", "convolution", "of", "a", "signal", "with", "a", "filter", ",", "we", "now", "have", ":", "with", ";", "as", "can", "easily", "be", "verified", "by", "noticing", "that", ".", "Note", "that", "this", "expression", "is", "now", "-", "localized", "since", "it", "is", "a", "-", "order", "polynomial", "in", "the", "Laplacian", ",", "i.e.", "it", "depends", "only", "on", "nodes", "that", "are", "at", "maximum", "steps", "away", "from", "the", "central", "node", "(", "-", "order", "neighborhood", ")", ".", "The", "complexity", "of", "evaluating", "Eq", ".", "[", "reference", "]", "is", ",", "i.e.", "linear", "in", "the", "number", "of", "edges", ".", "use", "this", "-", "localized", "convolution", "to", "define", "a", "convolutional", "neural", "network", "on", "graphs", ".", "subsection", ":", "Layer", "-", "Wise", "Linear", "Model", "A", "neural", "network", "model", "based", "on", "graph", "convolutions", "can", "therefore", "be", "built", "by", "stacking", "multiple", "convolutional", "layers", "of", "the", "form", "of", "Eq", ".", "[", "reference", "]", ",", "each", "layer", "followed", "by", "a", "point", "-", "wise", "non", "-", "linearity", ".", "Now", ",", "imagine", "we", "limited", "the", "layer", "-", "wise", "convolution", "operation", "to", "(", "see", "Eq", ".", "[", "reference", "]", ")", ",", "i.e.", "a", "function", "that", "is", "linear", "w.r.t", ".", "and", "therefore", "a", "linear", "function", "on", "the", "graph", "Laplacian", "spectrum", ".", "In", "this", "way", ",", "we", "can", "still", "recover", "a", "rich", "class", "of", "convolutional", "filter", "functions", "by", "stacking", "multiple", "such", "layers", ",", "but", "we", "are", "not", "limited", "to", "the", "explicit", "parameterization", "given", "by", ",", "e.g.", ",", "the", "Chebyshev", "polynomials", ".", "We", "intuitively", "expect", "that", "such", "a", "model", "can", "alleviate", "the", "problem", "of", "overfitting", "on", "local", "neighborhood", "structures", "for", "graphs", "with", "very", "wide", "node", "degree", "distributions", ",", "such", "as", "social", "networks", ",", "citation", "networks", ",", "knowledge", "graphs", "and", "many", "other", "real", "-", "world", "graph", "datasets", ".", "Additionally", ",", "for", "a", "fixed", "computational", "budget", ",", "this", "layer", "-", "wise", "linear", "formulation", "allows", "us", "to", "build", "deeper", "models", ",", "a", "practice", "that", "is", "known", "to", "improve", "modeling", "capacity", "on", "a", "number", "of", "domains", "he2015deep", ".", "In", "this", "linear", "formulation", "of", "a", "GCN", "we", "further", "approximate", ",", "as", "we", "can", "expect", "that", "neural", "network", "parameters", "will", "adapt", "to", "this", "change", "in", "scale", "during", "training", ".", "Under", "these", "approximations", "Eq", ".", "[", "reference", "]", "simplifies", "to", ":", "with", "two", "free", "parameters", "and", ".", "The", "filter", "parameters", "can", "be", "shared", "over", "the", "whole", "graph", ".", "Successive", "application", "of", "filters", "of", "this", "form", "then", "effectively", "convolve", "the", "-", "order", "neighborhood", "of", "a", "node", ",", "where", "is", "the", "number", "of", "successive", "filtering", "operations", "or", "convolutional", "layers", "in", "the", "neural", "network", "model", ".", "In", "practice", ",", "it", "can", "be", "beneficial", "to", "constrain", "the", "number", "of", "parameters", "further", "to", "address", "overfitting", "and", "to", "minimize", "the", "number", "of", "operations", "(", "such", "as", "matrix", "multiplications", ")", "per", "layer", ".", "This", "leaves", "us", "with", "the", "following", "expression", ":", "with", "a", "single", "parameter", ".", "Note", "that", "now", "has", "eigenvalues", "in", "the", "range", ".", "Repeated", "application", "of", "this", "operator", "can", "therefore", "lead", "to", "numerical", "instabilities", "and", "exploding", "/", "vanishing", "gradients", "when", "used", "in", "a", "deep", "neural", "network", "model", ".", "To", "alleviate", "this", "problem", ",", "we", "introduce", "the", "following", "renormalization", "trick", ":", ",", "with", "and", ".", "We", "can", "generalize", "this", "definition", "to", "a", "signal", "with", "input", "channels", "(", "i.e.", "a", "-", "dimensional", "feature", "vector", "for", "every", "node", ")", "and", "filters", "or", "feature", "maps", "as", "follows", ":", "where", "is", "now", "a", "matrix", "of", "filter", "parameters", "and", "is", "the", "convolved", "signal", "matrix", ".", "This", "filtering", "operation", "has", "complexity", ",", "as", "can", "be", "efficiently", "implemented", "as", "a", "product", "of", "a", "sparse", "matrix", "with", "a", "dense", "matrix", ".", "section", ":", "Semi", "-", "Supervised", "Node", "Classification", "Having", "introduced", "a", "simple", ",", "yet", "flexible", "model", "for", "efficient", "information", "propagation", "on", "graphs", ",", "we", "can", "return", "to", "the", "problem", "of", "semi", "-", "supervised", "node", "classification", ".", "As", "outlined", "in", "the", "introduction", ",", "we", "can", "relax", "certain", "assumptions", "typically", "made", "in", "graph", "-", "based", "semi", "-", "supervised", "learning", "by", "conditioning", "our", "model", "both", "on", "the", "data", "and", "on", "the", "adjacency", "matrix", "of", "the", "underlying", "graph", "structure", ".", "We", "expect", "this", "setting", "to", "be", "especially", "powerful", "in", "scenarios", "where", "the", "adjacency", "matrix", "contains", "information", "not", "present", "in", "the", "data", ",", "such", "as", "citation", "links", "between", "documents", "in", "a", "citation", "network", "or", "relations", "in", "a", "knowledge", "graph", ".", "The", "overall", "model", ",", "a", "multi", "-", "layer", "GCN", "for", "semi", "-", "supervised", "learning", ",", "is", "schematically", "depicted", "in", "Figure", "[", "reference", "]", ".", "[", "b", "]", "0.67", "[", "b", "]", "0.33", "subsection", ":", "Example", "In", "the", "following", ",", "we", "consider", "a", "two", "-", "layer", "GCN", "for", "semi", "-", "supervised", "node", "classification", "on", "a", "graph", "with", "a", "symmetric", "adjacency", "matrix", "(", "binary", "or", "weighted", ")", ".", "We", "first", "calculate", "in", "a", "pre", "-", "processing", "step", ".", "Our", "forward", "model", "then", "takes", "the", "simple", "form", ":", "Here", ",", "is", "an", "input", "-", "to", "-", "hidden", "weight", "matrix", "for", "a", "hidden", "layer", "with", "feature", "maps", ".", "is", "a", "hidden", "-", "to", "-", "output", "weight", "matrix", ".", "The", "softmax", "activation", "function", ",", "defined", "as", "with", ",", "is", "applied", "row", "-", "wise", ".", "For", "semi", "-", "supervised", "multi", "-", "class", "classification", ",", "we", "then", "evaluate", "the", "cross", "-", "entropy", "error", "over", "all", "labeled", "examples", ":", "where", "is", "the", "set", "of", "node", "indices", "that", "have", "labels", ".", "The", "neural", "network", "weights", "and", "are", "trained", "using", "gradient", "descent", ".", "In", "this", "work", ",", "we", "perform", "batch", "gradient", "descent", "using", "the", "full", "dataset", "for", "every", "training", "iteration", ",", "which", "is", "a", "viable", "option", "as", "long", "as", "datasets", "fit", "in", "memory", ".", "Using", "a", "sparse", "representation", "for", ",", "memory", "requirement", "is", ",", "i.e.", "linear", "in", "the", "number", "of", "edges", ".", "Stochasticity", "in", "the", "training", "process", "is", "introduced", "via", "dropout", "srivastava2014dropout", ".", "We", "leave", "memory", "-", "efficient", "extensions", "with", "mini", "-", "batch", "stochastic", "gradient", "descent", "for", "future", "work", ".", "subsection", ":", "Implementation", "In", "practice", ",", "we", "make", "use", "of", "TensorFlow", "tensorflow2015", "-", "whitepaper", "for", "an", "efficient", "GPU", "-", "based", "implementation", "of", "Eq", ".", "[", "reference", "]", "using", "sparse", "-", "dense", "matrix", "multiplications", ".", "The", "computational", "complexity", "of", "evaluating", "Eq", ".", "[", "reference", "]", "is", "then", ",", "i.e.", "linear", "in", "the", "number", "of", "graph", "edges", ".", "section", ":", "Related", "Work", "Our", "model", "draws", "inspiration", "both", "from", "the", "field", "of", "graph", "-", "based", "semi", "-", "supervised", "learning", "and", "from", "recent", "work", "on", "neural", "networks", "that", "operate", "on", "graphs", ".", "In", "what", "follows", ",", "we", "provide", "a", "brief", "overview", "on", "related", "work", "in", "both", "fields", ".", "subsection", ":", "Graph", "-", "Based", "Semi", "-", "Supervised", "Learning", "A", "large", "number", "of", "approaches", "for", "semi", "-", "supervised", "learning", "using", "graph", "representations", "have", "been", "proposed", "in", "recent", "years", ",", "most", "of", "which", "fall", "into", "two", "broad", "categories", ":", "methods", "that", "use", "some", "form", "of", "explicit", "graph", "Laplacian", "regularization", "and", "graph", "embedding", "-", "based", "approaches", ".", "Prominent", "examples", "for", "graph", "Laplacian", "regularization", "include", "label", "propagation", "zhu2003semi", ",", "manifold", "regularization", "belkin2006manifold", "and", "deep", "semi", "-", "supervised", "embedding", "weston2012deep", ".", "Recently", ",", "attention", "has", "shifted", "to", "models", "that", "learn", "graph", "embeddings", "with", "methods", "inspired", "by", "the", "skip", "-", "gram", "model", "mikolov2013distributed", ".", "DeepWalk", "perozzi2014deepwalk", "learns", "embeddings", "via", "the", "prediction", "of", "the", "local", "neighborhood", "of", "nodes", ",", "sampled", "from", "random", "walks", "on", "the", "graph", ".", "LINE", "tang2015line", "and", "node2vec", "grovernode2vec", "extend", "DeepWalk", "with", "more", "sophisticated", "random", "walk", "or", "breadth", "-", "first", "search", "schemes", ".", "For", "all", "these", "methods", ",", "however", ",", "a", "multi", "-", "step", "pipeline", "including", "random", "walk", "generation", "and", "semi", "-", "supervised", "training", "is", "required", "where", "each", "step", "has", "to", "be", "optimized", "separately", ".", "Planetoid", "yang2016revisiting", "alleviates", "this", "by", "injecting", "label", "information", "in", "the", "process", "of", "learning", "embeddings", ".", "subsection", ":", "Neural", "Networks", "on", "Graphs", "Neural", "networks", "that", "operate", "on", "graphs", "have", "previously", "been", "introduced", "in", "as", "a", "form", "of", "recurrent", "neural", "network", ".", "Their", "framework", "requires", "the", "repeated", "application", "of", "contraction", "maps", "as", "propagation", "functions", "until", "node", "representations", "reach", "a", "stable", "fixed", "point", ".", "This", "restriction", "was", "later", "alleviated", "in", "by", "introducing", "modern", "practices", "for", "recurrent", "neural", "network", "training", "to", "the", "original", "graph", "neural", "network", "framework", ".", "introduced", "a", "convolution", "-", "like", "propagation", "rule", "on", "graphs", "and", "methods", "for", "graph", "-", "level", "classification", ".", "Their", "approach", "requires", "to", "learn", "node", "degree", "-", "specific", "weight", "matrices", "which", "does", "not", "scale", "to", "large", "graphs", "with", "wide", "node", "degree", "distributions", ".", "Our", "model", "instead", "uses", "a", "single", "weight", "matrix", "per", "layer", "and", "deals", "with", "varying", "node", "degrees", "through", "an", "appropriate", "normalization", "of", "the", "adjacency", "matrix", "(", "see", "Section", "[", "reference", "]", ")", ".", "A", "related", "approach", "to", "node", "classification", "with", "a", "graph", "-", "based", "neural", "network", "was", "recently", "introduced", "in", ".", "They", "report", "complexity", ",", "limiting", "the", "range", "of", "possible", "applications", ".", "In", "a", "different", "yet", "related", "model", ",", "convert", "graphs", "locally", "into", "sequences", "that", "are", "fed", "into", "a", "conventional", "1D", "convolutional", "neural", "network", ",", "which", "requires", "the", "definition", "of", "a", "node", "ordering", "in", "a", "pre", "-", "processing", "step", ".", "Our", "method", "is", "based", "on", "spectral", "graph", "convolutional", "neural", "networks", ",", "introduced", "in", "and", "later", "extended", "by", "with", "fast", "localized", "convolutions", ".", "In", "contrast", "to", "these", "works", ",", "we", "consider", "here", "the", "task", "of", "transductive", "node", "classification", "within", "networks", "of", "significantly", "larger", "scale", ".", "We", "show", "that", "in", "this", "setting", ",", "a", "number", "of", "simplifications", "(", "see", "Section", "[", "reference", "]", ")", "can", "be", "introduced", "to", "the", "original", "frameworks", "of", "and", "that", "improve", "scalability", "and", "classification", "performance", "in", "large", "-", "scale", "networks", ".", "section", ":", "Experiments", "We", "test", "our", "model", "in", "a", "number", "of", "experiments", ":", "semi", "-", "supervised", "document", "classification", "in", "citation", "networks", ",", "semi", "-", "supervised", "entity", "classification", "in", "a", "bipartite", "graph", "extracted", "from", "a", "knowledge", "graph", ",", "an", "evaluation", "of", "various", "graph", "propagation", "models", "and", "a", "run", "-", "time", "analysis", "on", "random", "graphs", ".", "subsection", ":", "Datasets", "We", "closely", "follow", "the", "experimental", "setup", "in", ".", "Dataset", "statistics", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "In", "the", "citation", "network", "datasets", "—", "Citeseer", ",", "Cora", "and", "Pubmed", "sen2008collective", "—", "nodes", "are", "documents", "and", "edges", "are", "citation", "links", ".", "Label", "rate", "denotes", "the", "number", "of", "labeled", "nodes", "that", "are", "used", "for", "training", "divided", "by", "the", "total", "number", "of", "nodes", "in", "each", "dataset", ".", "NELL", "carlson2010toward", ",", "yang2016revisiting", "is", "a", "bipartite", "graph", "dataset", "extracted", "from", "a", "knowledge", "graph", "with", "55", ",", "864", "relation", "nodes", "and", "9", ",", "891", "entity", "nodes", ".", "paragraph", ":", "Citation", "networks", "We", "consider", "three", "citation", "network", "datasets", ":", "Citeseer", ",", "Cora", "and", "Pubmed", "sen2008collective", ".", "The", "datasets", "contain", "sparse", "bag", "-", "of", "-", "words", "feature", "vectors", "for", "each", "document", "and", "a", "list", "of", "citation", "links", "between", "documents", ".", "We", "treat", "the", "citation", "links", "as", "(", "undirected", ")", "edges", "and", "construct", "a", "binary", ",", "symmetric", "adjacency", "matrix", ".", "Each", "document", "has", "a", "class", "label", ".", "For", "training", ",", "we", "only", "use", "20", "labels", "per", "class", ",", "but", "all", "feature", "vectors", ".", "paragraph", ":", "NELL", "NELL", "is", "a", "dataset", "extracted", "from", "the", "knowledge", "graph", "introduced", "in", "carlson2010toward", ".", "A", "knowledge", "graph", "is", "a", "set", "of", "entities", "connected", "with", "directed", ",", "labeled", "edges", "(", "relations", ")", ".", "We", "follow", "the", "pre", "-", "processing", "scheme", "as", "described", "in", ".", "We", "assign", "separate", "relation", "nodes", "and", "for", "each", "entity", "pair", "as", "and", ".", "Entity", "nodes", "are", "described", "by", "sparse", "feature", "vectors", ".", "We", "extend", "the", "number", "of", "features", "in", "NELL", "by", "assigning", "a", "unique", "one", "-", "hot", "representation", "for", "every", "relation", "node", ",", "effectively", "resulting", "in", "a", "61", ",", "278", "-", "dim", "sparse", "feature", "vector", "per", "node", ".", "The", "semi", "-", "supervised", "task", "here", "considers", "the", "extreme", "case", "of", "only", "a", "single", "labeled", "example", "per", "class", "in", "the", "training", "set", ".", "We", "construct", "a", "binary", ",", "symmetric", "adjacency", "matrix", "from", "this", "graph", "by", "setting", "entries", ",", "if", "one", "or", "more", "edges", "are", "present", "between", "nodes", "and", ".", "paragraph", ":", "Random", "graphs", "We", "simulate", "random", "graph", "datasets", "of", "various", "sizes", "for", "experiments", "where", "we", "measure", "training", "time", "per", "epoch", ".", "For", "a", "dataset", "with", "nodes", "we", "create", "a", "random", "graph", "assigning", "edges", "uniformly", "at", "random", ".", "We", "take", "the", "identity", "matrix", "as", "input", "feature", "matrix", ",", "thereby", "implicitly", "taking", "a", "featureless", "approach", "where", "the", "model", "is", "only", "informed", "about", "the", "identity", "of", "each", "node", ",", "specified", "by", "a", "unique", "one", "-", "hot", "vector", ".", "We", "add", "dummy", "labels", "for", "every", "node", ".", "subsection", ":", "Experimental", "Set", "-", "Up", "Unless", "otherwise", "noted", ",", "we", "train", "a", "two", "-", "layer", "GCN", "as", "described", "in", "Section", "[", "reference", "]", "and", "evaluate", "prediction", "accuracy", "on", "a", "test", "set", "of", "1", ",", "000", "labeled", "examples", ".", "We", "provide", "additional", "experiments", "using", "deeper", "models", "with", "up", "to", "10", "layers", "in", "Appendix", "[", "reference", "]", ".", "We", "choose", "the", "same", "dataset", "splits", "as", "in", "with", "an", "additional", "validation", "set", "of", "500", "labeled", "examples", "for", "hyperparameter", "optimization", "(", "dropout", "rate", "for", "all", "layers", ",", "L2", "regularization", "factor", "for", "the", "first", "GCN", "layer", "and", "number", "of", "hidden", "units", ")", ".", "We", "do", "not", "use", "the", "validation", "set", "labels", "for", "training", ".", "For", "the", "citation", "network", "datasets", ",", "we", "optimize", "hyperparameters", "on", "Cora", "only", "and", "use", "the", "same", "set", "of", "parameters", "for", "Citeseer", "and", "Pubmed", ".", "We", "train", "all", "models", "for", "a", "maximum", "of", "200", "epochs", "(", "training", "iterations", ")", "using", "Adam", "kingma2014adam", "with", "a", "learning", "rate", "of", "and", "early", "stopping", "with", "a", "window", "size", "of", ",", "i.e.", "we", "stop", "training", "if", "the", "validation", "loss", "does", "not", "decrease", "for", "10", "consecutive", "epochs", ".", "We", "initialize", "weights", "using", "the", "initialization", "described", "in", "and", "accordingly", "(", "row", "-)", "normalize", "input", "feature", "vectors", ".", "On", "the", "random", "graph", "datasets", ",", "we", "use", "a", "hidden", "layer", "size", "of", "32", "units", "and", "omit", "regularization", "(", "i.e.", "neither", "dropout", "nor", "L2", "regularization", ")", ".", "subsection", ":", "Baselines", "We", "compare", "against", "the", "same", "baseline", "methods", "as", "in", ",", "i.e.", "label", "propagation", "(", "LP", ")", "zhu2003semi", ",", "semi", "-", "supervised", "embedding", "(", "SemiEmb", ")", "weston2012deep", ",", "manifold", "regularization", "(", "ManiReg", ")", "belkin2006manifold", "and", "skip", "-", "gram", "based", "graph", "embeddings", "(", "DeepWalk", ")", "perozzi2014deepwalk", ".", "We", "omit", "TSVM", "joachims1999transductive", ",", "as", "it", "does", "not", "scale", "to", "the", "large", "number", "of", "classes", "in", "one", "of", "our", "datasets", ".", "We", "further", "compare", "against", "the", "iterative", "classification", "algorithm", "(", "ICA", ")", "proposed", "in", "in", "conjunction", "with", "two", "logistic", "regression", "classifiers", ",", "one", "for", "local", "node", "features", "alone", "and", "one", "for", "relational", "classification", "using", "local", "features", "and", "an", "aggregation", "operator", "as", "described", "in", ".", "We", "first", "train", "the", "local", "classifier", "using", "all", "labeled", "training", "set", "nodes", "and", "use", "it", "to", "bootstrap", "class", "labels", "of", "unlabeled", "nodes", "for", "relational", "classifier", "training", ".", "We", "run", "iterative", "classification", "(", "relational", "classifier", ")", "with", "a", "random", "node", "ordering", "for", "10", "iterations", "on", "all", "unlabeled", "nodes", "(", "bootstrapped", "using", "the", "local", "classifier", ")", ".", "L2", "regularization", "parameter", "and", "aggregation", "operator", "(", "count", "vs.", "prop", ",", "see", ")", "are", "chosen", "based", "on", "validation", "set", "performance", "for", "each", "dataset", "separately", ".", "Lastly", ",", "we", "compare", "against", "Planetoid", "yang2016revisiting", ",", "where", "we", "always", "choose", "their", "best", "-", "performing", "model", "variant", "(", "transductive", "vs.", "inductive", ")", "as", "a", "baseline", ".", "section", ":", "Results", "subsection", ":", "Semi", "-", "Supervised", "Node", "Classification", "Results", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "Reported", "numbers", "denote", "classification", "accuracy", "in", "percent", ".", "For", "ICA", ",", "we", "report", "the", "mean", "accuracy", "of", "100", "runs", "with", "random", "node", "orderings", ".", "Results", "for", "all", "other", "baseline", "methods", "are", "taken", "from", "the", "Planetoid", "paper", "yang2016revisiting", ".", "Planetoid", "*", "denotes", "the", "best", "model", "for", "the", "respective", "dataset", "out", "of", "the", "variants", "presented", "in", "their", "paper", ".", "We", "further", "report", "wall", "-", "clock", "training", "time", "in", "seconds", "until", "convergence", "(", "in", "brackets", ")", "for", "our", "method", "(", "incl", ".", "evaluation", "of", "validation", "error", ")", "and", "for", "Planetoid", ".", "For", "the", "latter", ",", "we", "used", "an", "implementation", "provided", "by", "the", "authors", "and", "trained", "on", "the", "same", "hardware", "(", "with", "GPU", ")", "as", "our", "GCN", "model", ".", "We", "trained", "and", "tested", "our", "model", "on", "the", "same", "dataset", "splits", "as", "in", "and", "report", "mean", "accuracy", "of", "100", "runs", "with", "random", "weight", "initializations", ".", "We", "used", "the", "following", "sets", "of", "hyperparameters", "for", "Citeseer", ",", "Cora", "and", "Pubmed", ":", "0.5", "(", "dropout", "rate", ")", ",", "(", "L2", "regularization", ")", "and", "(", "number", "of", "hidden", "units", ")", ";", "and", "for", "NELL", ":", "0.1", "(", "dropout", "rate", ")", ",", "(", "L2", "regularization", ")", "and", "(", "number", "of", "hidden", "units", ")", ".", "In", "addition", ",", "we", "report", "performance", "of", "our", "model", "on", "10", "randomly", "drawn", "dataset", "splits", "of", "the", "same", "size", "as", "in", ",", "denoted", "by", "GCN", "(", "rand", ".", "splits", ")", ".", "Here", ",", "we", "report", "mean", "and", "standard", "error", "of", "prediction", "accuracy", "on", "the", "test", "set", "split", "in", "percent", ".", "subsection", ":", "Evaluation", "of", "Propagation", "Model", "We", "compare", "different", "variants", "of", "our", "proposed", "per", "-", "layer", "propagation", "model", "on", "the", "citation", "network", "datasets", ".", "We", "follow", "the", "experimental", "set", "-", "up", "described", "in", "the", "previous", "section", ".", "Results", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "The", "propagation", "model", "of", "our", "original", "GCN", "model", "is", "denoted", "by", "renormalization", "trick", "(", "in", "bold", ")", ".", "In", "all", "other", "cases", ",", "the", "propagation", "model", "of", "both", "neural", "network", "layers", "is", "replaced", "with", "the", "model", "specified", "under", "propagation", "model", ".", "Reported", "numbers", "denote", "mean", "classification", "accuracy", "for", "100", "repeated", "runs", "with", "random", "weight", "matrix", "initializations", ".", "In", "case", "of", "multiple", "variables", "per", "layer", ",", "we", "impose", "L2", "regularization", "on", "all", "weight", "matrices", "of", "the", "first", "layer", ".", "subsection", ":", "Training", "Time", "per", "Epoch", "Here", ",", "we", "report", "results", "for", "the", "mean", "training", "time", "per", "epoch", "(", "forward", "pass", ",", "cross", "-", "entropy", "calculation", ",", "backward", "pass", ")", "for", "100", "epochs", "on", "simulated", "random", "graphs", ",", "measured", "in", "seconds", "wall", "-", "clock", "time", ".", "See", "Section", "[", "reference", "]", "for", "a", "detailed", "description", "of", "the", "random", "graph", "dataset", "used", "in", "these", "experiments", ".", "We", "compare", "results", "on", "a", "GPU", "and", "on", "a", "CPU", "-", "only", "implementation", "in", "TensorFlow", "tensorflow2015", "-", "whitepaper", ".", "Figure", "[", "reference", "]", "summarizes", "the", "results", ".", "section", ":", "Discussion", "subsection", ":", "Semi", "-", "Supervised", "Model", "In", "the", "experiments", "demonstrated", "here", ",", "our", "method", "for", "semi", "-", "supervised", "node", "classification", "outperforms", "recent", "related", "methods", "by", "a", "significant", "margin", ".", "Methods", "based", "on", "graph", "-", "Laplacian", "regularization", "zhu2003semi", ",", "belkin2006manifold", ",", "weston2012deep", "are", "most", "likely", "limited", "due", "to", "their", "assumption", "that", "edges", "encode", "mere", "similarity", "of", "nodes", ".", "Skip", "-", "gram", "based", "methods", "on", "the", "other", "hand", "are", "limited", "by", "the", "fact", "that", "they", "are", "based", "on", "a", "multi", "-", "step", "pipeline", "which", "is", "difficult", "to", "optimize", ".", "Our", "proposed", "model", "can", "overcome", "both", "limitations", ",", "while", "still", "comparing", "favorably", "in", "terms", "of", "efficiency", "(", "measured", "in", "wall", "-", "clock", "time", ")", "to", "related", "methods", ".", "Propagation", "of", "feature", "information", "from", "neighboring", "nodes", "in", "every", "layer", "improves", "classification", "performance", "in", "comparison", "to", "methods", "like", "ICA", "lu2003link", ",", "where", "only", "label", "information", "is", "aggregated", ".", "We", "have", "further", "demonstrated", "that", "the", "proposed", "renormalized", "propagation", "model", "(", "Eq", ".", "[", "reference", "]", ")", "offers", "both", "improved", "efficiency", "(", "fewer", "parameters", "and", "operations", ",", "such", "as", "multiplication", "or", "addition", ")", "and", "better", "predictive", "performance", "on", "a", "number", "of", "datasets", "compared", "to", "a", "naïve", "-", "order", "model", "(", "Eq", ".", "[", "reference", "]", ")", "or", "higher", "-", "order", "graph", "convolutional", "models", "using", "Chebyshev", "polynomials", "(", "Eq", ".", "[", "reference", "]", ")", ".", "subsection", ":", "Limitations", "and", "Future", "Work", "Here", ",", "we", "describe", "several", "limitations", "of", "our", "current", "model", "and", "outline", "how", "these", "might", "be", "overcome", "in", "future", "work", ".", "paragraph", ":", "Memory", "requirement", "In", "the", "current", "setup", "with", "full", "-", "batch", "gradient", "descent", ",", "memory", "requirement", "grows", "linearly", "in", "the", "size", "of", "the", "dataset", ".", "We", "have", "shown", "that", "for", "large", "graphs", "that", "do", "not", "fit", "in", "GPU", "memory", ",", "training", "on", "CPU", "can", "still", "be", "a", "viable", "option", ".", "Mini", "-", "batch", "stochastic", "gradient", "descent", "can", "alleviate", "this", "issue", ".", "The", "procedure", "of", "generating", "mini", "-", "batches", ",", "however", ",", "should", "take", "into", "account", "the", "number", "of", "layers", "in", "the", "GCN", "model", ",", "as", "the", "-", "order", "neighborhood", "for", "a", "GCN", "with", "layers", "has", "to", "be", "stored", "in", "memory", "for", "an", "exact", "procedure", ".", "For", "very", "large", "and", "densely", "connected", "graph", "datasets", ",", "further", "approximations", "might", "be", "necessary", ".", "paragraph", ":", "Directed", "edges", "and", "edge", "features", "Our", "framework", "currently", "does", "not", "naturally", "support", "edge", "features", "and", "is", "limited", "to", "undirected", "graphs", "(", "weighted", "or", "unweighted", ")", ".", "Results", "on", "NELL", "however", "show", "that", "it", "is", "possible", "to", "handle", "both", "directed", "edges", "and", "edge", "features", "by", "representing", "the", "original", "directed", "graph", "as", "an", "undirected", "bipartite", "graph", "with", "additional", "nodes", "that", "represent", "edges", "in", "the", "original", "graph", "(", "see", "Section", "[", "reference", "]", "for", "details", ")", ".", "paragraph", ":", "Limiting", "assumptions", "Through", "the", "approximations", "introduced", "in", "Section", "[", "reference", "]", ",", "we", "implicitly", "assume", "locality", "(", "dependence", "on", "the", "-", "order", "neighborhood", "for", "a", "GCN", "with", "layers", ")", "and", "equal", "importance", "of", "self", "-", "connections", "vs.", "edges", "to", "neighboring", "nodes", ".", "For", "some", "datasets", ",", "however", ",", "it", "might", "be", "beneficial", "to", "introduce", "a", "trade", "-", "off", "parameter", "in", "the", "definition", "of", ":", "This", "parameter", "now", "plays", "a", "similar", "role", "as", "the", "trade", "-", "off", "parameter", "between", "supervised", "and", "unsupervised", "loss", "in", "the", "typical", "semi", "-", "supervised", "setting", "(", "see", "Eq", ".", "[", "reference", "]", ")", ".", "Here", ",", "however", ",", "it", "can", "be", "learned", "via", "gradient", "descent", ".", "section", ":", "Conclusion", "We", "have", "introduced", "a", "novel", "approach", "for", "semi", "-", "supervised", "classification", "on", "graph", "-", "structured", "data", ".", "Our", "GCN", "model", "uses", "an", "efficient", "layer", "-", "wise", "propagation", "rule", "that", "is", "based", "on", "a", "first", "-", "order", "approximation", "of", "spectral", "convolutions", "on", "graphs", ".", "Experiments", "on", "a", "number", "of", "network", "datasets", "suggest", "that", "the", "proposed", "GCN", "model", "is", "capable", "of", "encoding", "both", "graph", "structure", "and", "node", "features", "in", "a", "way", "useful", "for", "semi", "-", "supervised", "classification", ".", "In", "this", "setting", ",", "our", "model", "outperforms", "several", "recently", "proposed", "methods", "by", "a", "significant", "margin", ",", "while", "being", "computationally", "efficient", ".", "subsubsection", ":", "Acknowledgments", "We", "would", "like", "to", "thank", "Christos", "Louizos", ",", "Taco", "Cohen", ",", "Joan", "Bruna", ",", "Zhilin", "Yang", ",", "Dave", "Herman", ",", "Pramod", "Sinha", "and", "Abdul", "-", "Saboor", "Sheikh", "for", "helpful", "discussions", ".", "This", "research", "was", "funded", "by", "SAP", ".", "bibliography", ":", "References", "appendix", ":", "Relation", "to", "Weisfeiler", "-", "Lehman", "Algorithm", "A", "neural", "network", "model", "for", "graph", "-", "structured", "data", "should", "ideally", "be", "able", "to", "learn", "representations", "of", "nodes", "in", "a", "graph", ",", "taking", "both", "the", "graph", "structure", "and", "feature", "description", "of", "nodes", "into", "account", ".", "A", "well", "-", "studied", "framework", "for", "the", "unique", "assignment", "of", "node", "labels", "given", "a", "graph", "and", "(", "optionally", ")", "discrete", "initial", "node", "labels", "is", "provided", "by", "the", "1", "-", "dim", "Weisfeiler", "-", "Lehman", "(", "WL", "-", "1", ")", "algorithm", "weisfeiler1968reduction", ":", "[", "H", "]", "Initial", "node", "coloring", "Final", "node", "coloring", "t", "0", "stable", "node", "coloring", "is", "reached", "WL", "-", "1", "algorithm", "Here", ",", "denotes", "the", "coloring", "(", "label", "assignment", ")", "of", "node", "(", "at", "iteration", ")", "and", "is", "its", "set", "of", "neighboring", "node", "indices", "(", "irrespective", "of", "whether", "the", "graph", "includes", "self", "-", "connections", "for", "every", "node", "or", "not", ")", ".", "is", "a", "hash", "function", ".", "For", "an", "in", "-", "depth", "mathematical", "discussion", "of", "the", "WL", "-", "1", "algorithm", "see", ",", "e.g.", ",", ".", "We", "can", "replace", "the", "hash", "function", "in", "Algorithm", "[", "reference", "]", "with", "a", "neural", "network", "layer", "-", "like", "differentiable", "function", "with", "trainable", "parameters", "as", "follows", ":", "where", "is", "an", "appropriately", "chosen", "normalization", "constant", "for", "the", "edge", ".", "Further", ",", "we", "can", "take", "now", "to", "be", "a", "vector", "of", "activations", "of", "node", "in", "the", "neural", "network", "layer", ".", "is", "a", "layer", "-", "specific", "weight", "matrix", "and", "denotes", "a", "differentiable", ",", "non", "-", "linear", "activation", "function", ".", "By", "choosing", ",", "where", "denotes", "the", "degree", "of", "node", ",", "we", "recover", "the", "propagation", "rule", "of", "our", "Graph", "Convolutional", "Network", "(", "GCN", ")", "model", "in", "vector", "form", "(", "see", "Eq", ".", "[", "reference", "]", ")", ".", "This", "—", "loosely", "speaking", "—", "allows", "us", "to", "interpret", "our", "GCN", "model", "as", "a", "differentiable", "and", "parameterized", "generalization", "of", "the", "1", "-", "dim", "Weisfeiler", "-", "Lehman", "algorithm", "on", "graphs", ".", "subsection", ":", "Node", "Embeddings", "with", "Random", "Weights", "From", "the", "analogy", "with", "the", "Weisfeiler", "-", "Lehman", "algorithm", ",", "we", "can", "understand", "that", "even", "an", "untrained", "GCN", "model", "with", "random", "weights", "can", "serve", "as", "a", "powerful", "feature", "extractor", "for", "nodes", "in", "a", "graph", ".", "As", "an", "example", ",", "consider", "the", "following", "3", "-", "layer", "GCN", "model", ":", "with", "weight", "matrices", "initialized", "at", "random", "using", "the", "initialization", "described", "in", ".", ",", "and", "are", "defined", "as", "in", "Section", "[", "reference", "]", ".", "We", "apply", "this", "model", "on", "Zachary", "’s", "karate", "club", "network", "zachary1977information", ".", "This", "graph", "contains", "34", "nodes", ",", "connected", "by", "154", "(", "undirected", "and", "unweighted", ")", "edges", ".", "Every", "node", "is", "labeled", "by", "one", "of", "four", "classes", ",", "obtained", "via", "modularity", "-", "based", "clustering", "brandes2008modularity", ".", "See", "Figure", "[", "reference", "]", "for", "an", "illustration", ".", "[", "b", "]", "0.5", "[", "b", "]", "0.5", "We", "take", "a", "featureless", "approach", "by", "setting", ",", "where", "is", "the", "by", "identity", "matrix", ".", "is", "the", "number", "of", "nodes", "in", "the", "graph", ".", "Note", "that", "nodes", "are", "randomly", "ordered", "(", "i.e.", "ordering", "contains", "no", "information", ")", ".", "Furthermore", ",", "we", "choose", "a", "hidden", "layer", "dimensionality", "of", "and", "a", "two", "-", "dimensional", "output", "(", "so", "that", "the", "output", "can", "immediately", "be", "visualized", "in", "a", "2", "-", "dim", "plot", ")", ".", "Figure", "[", "reference", "]", "shows", "a", "representative", "example", "of", "node", "embeddings", "(", "outputs", ")", "obtained", "from", "an", "untrained", "GCN", "model", "applied", "to", "the", "karate", "club", "network", ".", "These", "results", "are", "comparable", "to", "embeddings", "obtained", "from", "DeepWalk", "perozzi2014deepwalk", ",", "which", "uses", "a", "more", "expensive", "unsupervised", "training", "procedure", ".", "subsection", ":", "Semi", "-", "Supervised", "Node", "Embeddings", "On", "this", "simple", "example", "of", "a", "GCN", "applied", "to", "the", "karate", "club", "network", "it", "is", "interesting", "to", "observe", "how", "embeddings", "react", "during", "training", "on", "a", "semi", "-", "supervised", "classification", "task", ".", "Such", "a", "visualization", "(", "see", "Figure", "[", "reference", "]", ")", "provides", "insights", "into", "how", "the", "GCN", "model", "can", "make", "use", "of", "the", "graph", "structure", "(", "and", "of", "features", "extracted", "from", "the", "graph", "structure", "at", "later", "layers", ")", "to", "learn", "embeddings", "that", "are", "useful", "for", "a", "classification", "task", ".", "We", "consider", "the", "following", "semi", "-", "supervised", "learning", "setup", ":", "we", "add", "a", "softmax", "layer", "on", "top", "of", "our", "model", "(", "Eq", ".", "[", "reference", "]", ")", "and", "train", "using", "only", "a", "single", "labeled", "example", "per", "class", "(", "i.e.", "a", "total", "number", "of", "4", "labeled", "nodes", ")", ".", "We", "train", "for", "300", "training", "iterations", "using", "Adam", "kingma2014adam", "with", "a", "learning", "rate", "of", "on", "a", "cross", "-", "entropy", "loss", ".", "Figure", "[", "reference", "]", "shows", "the", "evolution", "of", "node", "embeddings", "over", "a", "number", "of", "training", "iterations", ".", "The", "model", "succeeds", "in", "linearly", "separating", "the", "communities", "based", "on", "minimal", "supervision", "and", "the", "graph", "structure", "alone", ".", "A", "video", "of", "the", "full", "training", "process", "can", "be", "found", "on", "our", "website", ".", "[", "b", "]", "0.5", "[", "b", "]", "0.5", "[", "b", "]", "0.5", "[", "b", "]", "0.5", "[", "b", "]", "0.5", "[", "b", "]", "0.5", "appendix", ":", "Experiments", "on", "Model", "Depth", "In", "these", "experiments", ",", "we", "investigate", "the", "influence", "of", "model", "depth", "(", "number", "of", "layers", ")", "on", "classification", "performance", ".", "We", "report", "results", "on", "a", "5", "-", "fold", "cross", "-", "validation", "experiment", "on", "the", "Cora", ",", "Citeseer", "and", "Pubmed", "datasets", "sen2008collective", "using", "all", "labels", ".", "In", "addition", "to", "the", "standard", "GCN", "model", "(", "Eq", ".", "[", "reference", "]", ")", ",", "we", "report", "results", "on", "a", "model", "variant", "where", "we", "use", "residual", "connections", "he2015deep", "between", "hidden", "layers", "to", "facilitate", "training", "of", "deeper", "models", "by", "enabling", "the", "model", "to", "carry", "over", "information", "from", "the", "previous", "layer", "’s", "input", ":", "On", "each", "cross", "-", "validation", "split", ",", "we", "train", "for", "400", "epochs", "(", "without", "early", "stopping", ")", "using", "the", "Adam", "optimizer", "kingma2014adam", "with", "a", "learning", "rate", "of", ".", "Other", "hyperparameters", "are", "chosen", "as", "follows", ":", "0.5", "(", "dropout", "rate", ",", "first", "and", "last", "layer", ")", ",", "(", "L2", "regularization", ",", "first", "layer", ")", ",", "16", "(", "number", "of", "units", "for", "each", "hidden", "layer", ")", "and", "0.01", "(", "learning", "rate", ")", ".", "Results", "are", "summarized", "in", "Figure", "[", "reference", "]", ".", "[", "b", "]", "0.33", "[", "b", "]", "0.33", "[", "b", "]", "0.33", "For", "the", "datasets", "considered", "here", ",", "best", "results", "are", "obtained", "with", "a", "2", "-", "or", "3", "-", "layer", "model", ".", "We", "observe", "that", "for", "models", "deeper", "than", "7", "layers", ",", "training", "without", "the", "use", "of", "residual", "connections", "can", "become", "difficult", ",", "as", "the", "effective", "context", "size", "for", "each", "node", "increases", "by", "the", "size", "of", "its", "-", "order", "neighborhood", "(", "for", "a", "model", "with", "layers", ")", "with", "each", "additional", "layer", ".", "Furthermore", ",", "overfitting", "can", "become", "an", "issue", "as", "the", "number", "of", "parameters", "increases", "with", "model", "depth", "."], "entity": [[2, 6, ["Semi", "-", "Supervised", "Classification"], "Task"], [7, 10, ["Graph", "Convolutional", "Networks"], "Method"], [16, 20, ["semi", "-", "supervised", "learning"], "Task"], [33, 36, ["convolutional", "neural", "networks"], "Method"], [48, 50, ["convolutional", "architecture"], "Method"], [74, 77, ["hidden", "layer", "representations"], "Method"], [94, 96, ["citation", "networks"], "Task"], [123, 125, ["classifying", "nodes"], "Task"], [137, 139, ["citation", "network"], "Task"], [159, 166, ["graph", "-", "based", "semi", "-", "supervised", "learning"], "Method"], [179, 184, ["explicit", "graph", "-", "based", "regularization"], "Method"], [223, 229, ["neural", "network", "-", "like", "differentiable", "function"], "Method"], [334, 337, ["neural", "network", "model"], "Method"], [352, 356, ["graph", "-", "based", "regularization"], "Method"], [413, 418, ["layer", "-", "wise", "propagation", "rule"], "Method"], [419, 422, ["neural", "network", "models"], "Method"], [436, 444, ["first", "-", "order", "approximation", "of", "spectral", "graph", "convolutions"], "Method"], [455, 461, ["graph", "-", "based", "neural", "network", "model"], "Method"], [465, 477, ["fast", "and", "scalable", "semi", "-", "supervised", "classification", "of", "nodes", "in", "a", "graph"], "Task"], [492, 494, ["classification", "accuracy"], "Metric"], [495, 496, ["efficiency"], "Metric"], [499, 503, ["wall", "-", "clock", "time"], "Metric"], [514, 518, ["semi", "-", "supervised", "learning"], "Task"], [521, 526, ["Fast", "Approximate", "Convolutions", "on", "Graphs"], "Task"], [537, 543, ["graph", "-", "based", "neural", "network", "model"], "Method"], [560, 563, ["Graph", "Convolutional", "Network"], "Method"], [564, 565, ["GCN"], "Method"], [569, 574, ["layer", "-", "wise", "propagation", "rule"], "Method"], [599, 605, ["layer", "-", "specific", "trainable", "weight", "matrix"], "Method"], [636, 638, ["propagation", "rule"], "Method"], [643, 651, ["first", "-", "order", "approximation", "of", "localized", "spectral", "filters"], "Method"], [652, 653, ["graphs"], "Method"], [684, 685, ["filter"], "Method"], [716, 719, ["graph", "Fourier", "transform"], "Method"], [734, 736, ["Evaluating", "Eq"], "Task"], [787, 789, ["truncated", "expansion"], "Method"], [792, 794, ["Chebyshev", "polynomials"], "Method"], [817, 819, ["Chebyshev", "polynomials"], "Method"], [849, 853, ["convolution", "of", "a", "signal"], "Method"], [855, 856, ["filter"], "Method"], [914, 915, ["complexity"], "Metric"], [917, 918, ["Eq"], "Task"], [934, 937, ["-", "localized", "convolution"], "Method"], [948, 953, ["Layer", "-", "Wise", "Linear", "Model"], "Method"], [954, 957, ["neural", "network", "model"], "Method"], [959, 961, ["graph", "convolutions"], "Method"], [967, 970, ["multiple", "convolutional", "layers"], "Method"], [985, 991, ["point", "-", "wise", "non", "-", "linearity"], "Method"], [998, 1003, ["layer", "-", "wise", "convolution", "operation"], "Method"], [1044, 1047, ["convolutional", "filter", "functions"], "Method"], [1125, 1130, ["layer", "-", "wise", "linear", "formulation"], "Method"], [1134, 1136, ["deeper", "models"], "Method"], [1155, 1157, ["linear", "formulation"], "Method"], [1159, 1160, ["GCN"], "Method"], [1169, 1172, ["neural", "network", "parameters"], "Method"], [1213, 1214, ["filters"], "Method"], [1233, 1236, ["successive", "filtering", "operations"], "Method"], [1237, 1239, ["convolutional", "layers"], "Method"], [1241, 1244, ["neural", "network", "model"], "Method"], [1272, 1274, ["matrix", "multiplications"], "Method"], [1320, 1324, ["deep", "neural", "network", "model"], "Method"], [1334, 1336, ["renormalization", "trick"], "Method"], [1387, 1389, ["filtering", "operation"], "Method"], [1390, 1391, ["complexity"], "Metric"], [1402, 1404, ["sparse", "matrix"], "Method"], [1411, 1416, ["Semi", "-", "Supervised", "Node", "Classification"], "Task"], [1438, 1443, ["semi", "-", "supervised", "node", "classification"], "Task"], [1458, 1465, ["graph", "-", "based", "semi", "-", "supervised", "learning"], "Task"], [1533, 1537, ["semi", "-", "supervised", "learning"], "Task"], [1570, 1575, ["semi", "-", "supervised", "node", "classification"], "Task"], [1594, 1598, ["pre", "-", "processing", "step"], "Task"], [1600, 1602, ["forward", "model"], "Method"], [1621, 1623, ["hidden", "layer"], "Method"], [1653, 1660, ["semi", "-", "supervised", "multi", "-", "class", "classification"], "Task"], [1665, 1669, ["cross", "-", "entropy", "error"], "Metric"], [1686, 1689, ["neural", "network", "weights"], "Method"], [1693, 1695, ["gradient", "descent"], "Method"], [1702, 1705, ["batch", "gradient", "descent"], "Method"], [1729, 1731, ["sparse", "representation"], "Method"], [1733, 1735, ["memory", "requirement"], "Metric"], [1748, 1750, ["training", "process"], "Method"], [1753, 1754, ["dropout"], "Method"], [1763, 1769, ["mini", "-", "batch", "stochastic", "gradient", "descent"], "Method"], [1783, 1784, ["TensorFlow"], "Method"], [1790, 1796, ["GPU", "-", "based", "implementation", "of", "Eq"], "Method"], [1801, 1806, ["sparse", "-", "dense", "matrix", "multiplications"], "Method"], [1808, 1810, ["computational", "complexity"], "Metric"], [1811, 1813, ["evaluating", "Eq"], "Task"], [1842, 1849, ["graph", "-", "based", "semi", "-", "supervised", "learning"], "Task"], [1854, 1856, ["neural", "networks"], "Method"], [1879, 1886, ["Graph", "-", "Based", "Semi", "-", "Supervised", "Learning"], "Method"], [1892, 1896, ["semi", "-", "supervised", "learning"], "Task"], [1897, 1899, ["graph", "representations"], "Task"], [1921, 1925, ["explicit", "graph", "Laplacian", "regularization"], "Method"], [1926, 1931, ["graph", "embedding", "-", "based", "approaches"], "Method"], [1935, 1938, ["graph", "Laplacian", "regularization"], "Task"], [1939, 1941, ["label", "propagation"], "Task"], [1943, 1945, ["manifold", "regularization"], "Method"], [1947, 1952, ["deep", "semi", "-", "supervised", "embedding"], "Method"], [1963, 1965, ["graph", "embeddings"], "Method"], [1970, 1974, ["skip", "-", "gram", "model"], "Method"], [1976, 1977, ["DeepWalk"], "Method"], [1979, 1980, ["embeddings"], "Method"], [1998, 1999, ["LINE"], "Method"], [2001, 2002, ["node2vec"], "Method"], [2004, 2005, ["DeepWalk"], "Method"], [2008, 2016, ["random", "walk", "or", "breadth", "-", "first", "search", "schemes"], "Method"], [2025, 2029, ["multi", "-", "step", "pipeline"], "Method"], [2030, 2033, ["random", "walk", "generation"], "Method"], [2034, 2038, ["semi", "-", "supervised", "training"], "Method"], [2049, 2050, ["Planetoid"], "Method"], [2061, 2063, ["learning", "embeddings"], "Task"], [2066, 2068, ["Neural", "Networks"], "Method"], [2070, 2072, ["Neural", "networks"], "Method"], [2085, 2088, ["recurrent", "neural", "network"], "Method"], [2096, 2098, ["contraction", "maps"], "Method"], [2099, 2101, ["propagation", "functions"], "Method"], [2102, 2104, ["node", "representations"], "Method"], [2121, 2125, ["recurrent", "neural", "network", "training"], "Method"], [2128, 2132, ["graph", "neural", "network", "framework"], "Method"], [2135, 2140, ["convolution", "-", "like", "propagation", "rule"], "Method"], [2141, 2142, ["graphs"], "Method"], [2145, 2149, ["graph", "-", "level", "classification"], "Task"], [2180, 2184, ["weight", "matrix", "per", "layer"], "Method"], [2193, 2198, ["normalization", "of", "the", "adjacency", "matrix"], "Method"], [2210, 2212, ["node", "classification"], "Task"], [2214, 2219, ["graph", "-", "based", "neural", "network"], "Method"], [2226, 2227, ["complexity"], "Metric"], [2253, 2257, ["1D", "convolutional", "neural", "network"], "Method"], [2278, 2283, ["spectral", "graph", "convolutional", "neural", "networks"], "Method"], [2291, 2294, ["fast", "localized", "convolutions"], "Method"], [2307, 2310, ["transductive", "node", "classification"], "Task"], [2346, 2347, ["scalability"], "Metric"], [2348, 2349, ["classification"], "Task"], [2351, 2355, ["large", "-", "scale", "networks"], "Task"], [2369, 2374, ["semi", "-", "supervised", "document", "classification"], "Task"], [2375, 2377, ["citation", "networks"], "Task"], [2378, 2383, ["semi", "-", "supervised", "entity", "classification"], "Task"], [2397, 2400, ["graph", "propagation", "models"], "Method"], [2402, 2406, ["run", "-", "time", "analysis"], "Method"], [2437, 2438, ["Citeseer"], "Material"], [2439, 2440, ["Cora"], "Material"], [2441, 2442, ["Pubmed"], "Material"], [2453, 2455, ["Label", "rate"], "Metric"], [2506, 2508, ["Citation", "networks"], "Method"], [2515, 2516, ["Citeseer"], "Material"], [2517, 2518, ["Cora"], "Material"], [2519, 2520, ["Pubmed"], "Material"], [2572, 2573, ["training"], "Task"], [2589, 2590, ["NELL"], "Material"], [2590, 2591, ["NELL"], "Material"], [2624, 2628, ["pre", "-", "processing", "scheme"], "Method"], [2661, 2662, ["NELL"], "Material"], [2666, 2670, ["one", "-", "hot", "representation"], "Method"], [2691, 2695, ["semi", "-", "supervised", "task"], "Task"], [2756, 2758, ["training", "time"], "Metric"], [2791, 2793, ["featureless", "approach"], "Method"], [2849, 2851, ["prediction", "accuracy"], "Metric"], [2867, 2869, ["deeper", "models"], "Method"], [2898, 2900, ["hyperparameter", "optimization"], "Method"], [2941, 2942, ["hyperparameters"], "Method"], [2943, 2944, ["Cora"], "Material"], [2953, 2954, ["Citeseer"], "Material"], [2955, 2956, ["Pubmed"], "Material"], [2972, 2974, ["Adam", "kingma2014adam"], "Method"], [2976, 2978, ["learning", "rate"], "Metric"], [2994, 2996, ["validation", "loss"], "Metric"], [3039, 3040, ["regularization"], "Method"], [3043, 3044, ["dropout"], "Method"], [3045, 3047, ["L2", "regularization"], "Method"], [3063, 3065, ["label", "propagation"], "Task"], [3066, 3067, ["LP"], "Method"], [3070, 3074, ["semi", "-", "supervised", "embedding"], "Task"], [3075, 3076, ["SemiEmb"], "Method"], [3079, 3081, ["manifold", "regularization"], "Method"], [3086, 3092, ["skip", "-", "gram", "based", "graph", "embeddings"], "Method"], [3093, 3094, ["DeepWalk"], "Method"], [3099, 3100, ["TSVM"], "Method"], [3124, 3127, ["iterative", "classification", "algorithm"], "Method"], [3128, 3129, ["ICA"], "Task"], [3136, 3139, ["logistic", "regression", "classifiers"], "Method"], [3149, 3151, ["relational", "classification"], "Task"], [3156, 3158, ["aggregation", "operator"], "Method"], [3166, 3168, ["local", "classifier"], "Method"], [3185, 3188, ["relational", "classifier", "training"], "Method"], [3191, 3193, ["iterative", "classification"], "Method"], [3194, 3196, ["relational", "classifier"], "Method"], [3213, 3215, ["local", "classifier"], "Method"], [3247, 3248, ["Planetoid"], "Method"], [3261, 3264, ["transductive", "vs.", "inductive"], "Method"], [3274, 3279, ["Semi", "-", "Supervised", "Node", "Classification"], "Task"], [3291, 3293, ["classification", "accuracy"], "Metric"], [3297, 3298, ["ICA"], "Task"], [3303, 3304, ["accuracy"], "Metric"], [3322, 3324, ["Planetoid", "paper"], "Method"], [3348, 3353, ["wall", "-", "clock", "training", "time"], "Metric"], [3367, 3371, ["evaluation", "of", "validation", "error"], "Metric"], [3374, 3375, ["Planetoid"], "Method"], [3400, 3402, ["GCN", "model"], "Method"], [3419, 3420, ["accuracy"], "Metric"], [3424, 3427, ["random", "weight", "initializations"], "Method"], [3436, 3437, ["Citeseer"], "Material"], [3438, 3439, ["Cora"], "Material"], [3440, 3441, ["Pubmed"], "Material"], [3449, 3451, ["L2", "regularization"], "Method"], [3462, 3463, ["NELL"], "Material"], [3471, 3473, ["L2", "regularization"], "Method"], [3506, 3507, ["GCN"], "Method"], [3520, 3521, ["error"], "Metric"], [3522, 3524, ["prediction", "accuracy"], "Metric"], [3536, 3538, ["Propagation", "Model"], "Method"], [3545, 3550, ["per", "-", "layer", "propagation", "model"], "Method"], [3579, 3581, ["propagation", "model"], "Method"], [3584, 3586, ["GCN", "model"], "Method"], [3589, 3591, ["renormalization", "trick"], "Method"], [3602, 3604, ["propagation", "model"], "Method"], [3606, 3609, ["neural", "network", "layers"], "Method"], [3616, 3618, ["propagation", "model"], "Method"], [3623, 3625, ["classification", "accuracy"], "Metric"], [3630, 3634, ["random", "weight", "matrix", "initializations"], "Method"], [3645, 3647, ["L2", "regularization"], "Method"], [3658, 3660, ["Training", "Time"], "Metric"], [3669, 3674, ["mean", "training", "time", "per", "epoch"], "Metric"], [3675, 3677, ["forward", "pass"], "Metric"], [3678, 3682, ["cross", "-", "entropy", "calculation"], "Method"], [3683, 3685, ["backward", "pass"], "Method"], [3726, 3727, ["GPU"], "Method"], [3730, 3734, ["CPU", "-", "only", "implementation"], "Method"], [3753, 3757, ["Semi", "-", "Supervised", "Model"], "Method"], [3766, 3771, ["semi", "-", "supervised", "node", "classification"], "Task"], [3783, 3787, ["graph", "-", "Laplacian", "regularization"], "Method"], [3808, 3813, ["Skip", "-", "gram", "based", "methods"], "Method"], [3828, 3832, ["multi", "-", "step", "pipeline"], "Method"], [3853, 3854, ["efficiency"], "Metric"], [3857, 3861, ["wall", "-", "clock", "time"], "Metric"], [3877, 3878, ["classification"], "Task"], [3901, 3904, ["renormalized", "propagation", "model"], "Method"], [3929, 3930, ["predictive"], "Task"], [3939, 3943, ["naïve", "-", "order", "model"], "Method"], [3951, 3957, ["higher", "-", "order", "graph", "convolutional", "models"], "Method"], [3958, 3960, ["Chebyshev", "polynomials"], "Method"], [4038, 4039, ["CPU"], "Method"], [4046, 4052, ["Mini", "-", "batch", "stochastic", "gradient", "descent"], "Method"], [4060, 4064, ["generating", "mini", "-", "batches"], "Task"], [4077, 4079, ["GCN", "model"], "Method"], [4087, 4088, ["GCN"], "Method"], [4169, 4172, ["undirected", "bipartite", "graph"], "Method"], [4279, 4283, ["semi", "-", "supervised", "setting"], "Task"], [4301, 4303, ["gradient", "descent"], "Method"], [4314, 4318, ["semi", "-", "supervised", "classification"], "Task"], [4325, 4327, ["GCN", "model"], "Method"], [4330, 4335, ["layer", "-", "wise", "propagation", "rule"], "Method"], [4361, 4363, ["GCN", "model"], "Method"], [4378, 4382, ["semi", "-", "supervised", "classification"], "Task"], [4452, 4456, ["Weisfeiler", "-", "Lehman", "Algorithm"], "Method"], [4457, 4460, ["neural", "network", "model"], "Method"], [4498, 4503, ["unique", "assignment", "of", "node", "labels"], "Task"], [4518, 4530, ["1", "-", "dim", "Weisfeiler", "-", "Lehman", "(", "WL", "-", "1", ")", "algorithm"], "Method"], [4535, 4538, ["Initial", "node", "coloring"], "Method"], [4548, 4552, ["WL", "-", "1", "algorithm"], "Method"], [4594, 4596, ["hash", "function"], "Method"], [4606, 4610, ["WL", "-", "1", "algorithm"], "Method"], [4619, 4621, ["hash", "function"], "Method"], [4628, 4635, ["neural", "network", "layer", "-", "like", "differentiable", "function"], "Method"], [4668, 4671, ["neural", "network", "layer"], "Method"], [4707, 4710, ["Graph", "Convolutional", "Network"], "Method"], [4711, 4712, ["GCN"], "Method"], [4736, 4738, ["GCN", "model"], "Method"], [4740, 4744, ["differentiable", "and", "parameterized", "generalization"], "Method"], [4746, 4753, ["1", "-", "dim", "Weisfeiler", "-", "Lehman", "algorithm"], "Method"], [4754, 4755, ["graphs"], "Method"], [4768, 4772, ["Weisfeiler", "-", "Lehman", "algorithm"], "Method"], [4779, 4782, ["untrained", "GCN", "model"], "Method"], [4790, 4792, ["feature", "extractor"], "Method"], [4805, 4810, ["3", "-", "layer", "GCN", "model"], "Method"], [4874, 4878, ["modularity", "-", "based", "clustering"], "Method"], [4900, 4902, ["featureless", "approach"], "Method"], [4984, 4987, ["untrained", "GCN", "model"], "Method"], [4990, 4993, ["karate", "club", "network"], "Method"], [5002, 5003, ["DeepWalk"], "Method"], [5010, 5013, ["unsupervised", "training", "procedure"], "Method"], [5016, 5021, ["Semi", "-", "Supervised", "Node", "Embeddings"], "Task"], [5027, 5028, ["GCN"], "Method"], [5031, 5034, ["karate", "club", "network"], "Method"], [5046, 5051, ["semi", "-", "supervised", "classification", "task"], "Task"], [5067, 5069, ["GCN", "model"], "Method"], [5091, 5092, ["embeddings"], "Task"], [5097, 5099, ["classification", "task"], "Task"], [5104, 5109, ["semi", "-", "supervised", "learning", "setup"], "Task"], [5113, 5115, ["softmax", "layer"], "Method"], [5155, 5157, ["Adam", "kingma2014adam"], "Method"], [5159, 5161, ["learning", "rate"], "Metric"], [5164, 5168, ["cross", "-", "entropy", "loss"], "Metric"], [5246, 5248, ["Model", "Depth"], "Task"], [5265, 5266, ["classification"], "Task"], [5273, 5280, ["5", "-", "fold", "cross", "-", "validation", "experiment"], "Metric"], [5282, 5283, ["Cora"], "Material"], [5284, 5285, ["Citeseer"], "Material"], [5286, 5288, ["Pubmed", "datasets"], "Material"], [5298, 5300, ["GCN", "model"], "Method"], [5328, 5330, ["deeper", "models"], "Method"], [5347, 5351, ["cross", "-", "validation", "split"], "Metric"], [5364, 5366, ["Adam", "optimizer"], "Method"], [5369, 5371, ["learning", "rate"], "Metric"], [5392, 5394, ["L2", "regularization"], "Method"], [5412, 5414, ["learning", "rate"], "Metric"], [5449, 5456, ["2", "-", "or", "3", "-", "layer", "model"], "Method"], [40, 41, ["graphs"], "Method"], [53, 61, ["first", "-", "order", "approximation", "of", "spectral", "graph", "convolutions"], "Method"], [426, 427, ["graphs"], "Method"], [667, 668, ["graphs"], "Method"], [768, 769, ["graphs"], "Method"], [944, 945, ["graphs"], "Method"], [1089, 1090, ["graphs"], "Method"], [1106, 1107, ["graphs"], "Method"], [1429, 1430, ["graphs"], "Method"], [1531, 1532, ["GCN"], "Method"], [1568, 1569, ["GCN"], "Method"], [1859, 1860, ["graphs"], "Method"], [2075, 2076, ["graphs"], "Method"], [2167, 2168, ["graphs"], "Method"], [2243, 2244, ["graphs"], "Method"], [2408, 2409, ["graphs"], "Method"], [2477, 2478, ["NELL"], "Material"], [2742, 2743, ["graphs"], "Method"], [2839, 2840, ["GCN"], "Method"], [2901, 2902, ["dropout"], "Method"], [2913, 2914, ["GCN"], "Method"], [3326, 3327, ["Planetoid"], "Method"], [3444, 3445, ["dropout"], "Method"], [3466, 3467, ["dropout"], "Method"], [3692, 3693, ["graphs"], "Method"], [3735, 3736, ["TensorFlow"], "Method"], [3884, 3885, ["ICA"], "Task"], [4007, 4009, ["gradient", "descent"], "Method"], [4027, 4028, ["graphs"], "Method"], [4137, 4138, ["graphs"], "Method"], [4146, 4147, ["NELL"], "Material"], [4219, 4220, ["GCN"], "Method"], [4348, 4349, ["graphs"], "Method"], [5382, 5383, ["dropout"], "Method"]]} {"doc_id": "1827de6fa9c9c1b3d647a9d707042e89cf94abf0", "text": ["document", ":", "Batch", "Normalization", ":", "Accelerating", "Deep", "Network", "Training", "by", "Reducing", "Internal", "Covariate", "Shift", "Training", "Deep", "Neural", "Networks", "is", "complicated", "by", "the", "fact", "that", "the", "distribution", "of", "each", "layer", "’s", "inputs", "changes", "during", "training", ",", "as", "the", "parameters", "of", "the", "previous", "layers", "change", ".", "This", "slows", "down", "the", "training", "by", "requiring", "lower", "learning", "rates", "and", "careful", "parameter", "initialization", ",", "and", "makes", "it", "notoriously", "hard", "to", "train", "models", "with", "saturating", "nonlinearities", ".", "We", "refer", "to", "this", "phenomenon", "as", "internal", "covariate", "shift", ",", "and", "address", "the", "problem", "by", "normalizing", "layer", "inputs", ".", "Our", "method", "draws", "its", "strength", "from", "making", "normalization", "a", "part", "of", "the", "model", "architecture", "and", "performing", "the", "normalization", "for", "each", "training", "mini", "-", "batch", ".", "Batch", "Normalization", "allows", "us", "to", "use", "much", "higher", "learning", "rates", "and", "be", "less", "careful", "about", "initialization", ".", "It", "also", "acts", "as", "a", "regularizer", ",", "in", "some", "cases", "eliminating", "the", "need", "for", "Dropout", ".", "Applied", "to", "a", "state", "-", "of", "-", "the", "-", "art", "image", "classification", "model", ",", "Batch", "Normalization", "achieves", "the", "same", "accuracy", "with", "14", "times", "fewer", "training", "steps", ",", "and", "beats", "the", "original", "model", "by", "a", "significant", "margin", ".", "Using", "an", "ensemble", "of", "batch", "-", "normalized", "networks", ",", "we", "improve", "upon", "the", "best", "published", "result", "on", "ImageNet", "classification", ":", "reaching", "4.9", "%", "top", "-", "5", "validation", "error", "(", "and", "4.8", "%", "test", "error", ")", ",", "exceeding", "the", "accuracy", "of", "human", "raters", ".", "section", ":", "Introduction", "Deep", "learning", "has", "dramatically", "advanced", "the", "state", "of", "the", "art", "in", "vision", ",", "speech", ",", "and", "many", "other", "areas", ".", "Stochastic", "gradient", "descent", "(", "SGD", ")", "has", "proved", "to", "be", "an", "effective", "way", "of", "training", "deep", "networks", ",", "and", "SGD", "variants", "such", "as", "momentum", "and", "Adagrad", "have", "been", "used", "to", "achieve", "state", "of", "the", "art", "performance", ".", "SGD", "optimizes", "the", "parameters", "of", "the", "network", ",", "so", "as", "to", "minimize", "the", "loss", "where", "is", "the", "training", "data", "set", ".", "With", "SGD", ",", "the", "training", "proceeds", "in", "steps", ",", "and", "at", "each", "step", "we", "consider", "a", "mini", "-", "batch", "of", "size", ".", "The", "mini", "-", "batch", "is", "used", "to", "approximate", "the", "gradient", "of", "the", "loss", "function", "with", "respect", "to", "the", "parameters", ",", "by", "computing", "Using", "mini", "-", "batches", "of", "examples", ",", "as", "opposed", "to", "one", "example", "at", "a", "time", ",", "is", "helpful", "in", "several", "ways", ".", "First", ",", "the", "gradient", "of", "the", "loss", "over", "a", "mini", "-", "batch", "is", "an", "estimate", "of", "the", "gradient", "over", "the", "training", "set", ",", "whose", "quality", "improves", "as", "the", "batch", "size", "increases", ".", "Second", ",", "computation", "over", "a", "batch", "can", "be", "much", "more", "efficient", "than", "computations", "for", "individual", "examples", ",", "due", "to", "the", "parallelism", "afforded", "by", "the", "modern", "computing", "platforms", ".", "While", "stochastic", "gradient", "is", "simple", "and", "effective", ",", "it", "requires", "careful", "tuning", "of", "the", "model", "hyper", "-", "parameters", ",", "specifically", "the", "learning", "rate", "used", "in", "optimization", ",", "as", "well", "as", "the", "initial", "values", "for", "the", "model", "parameters", ".", "The", "training", "is", "complicated", "by", "the", "fact", "that", "the", "inputs", "to", "each", "layer", "are", "affected", "by", "the", "parameters", "of", "all", "preceding", "layers", "–", "so", "that", "small", "changes", "to", "the", "network", "parameters", "amplify", "as", "the", "network", "becomes", "deeper", ".", "The", "change", "in", "the", "distributions", "of", "layers", "’", "inputs", "presents", "a", "problem", "because", "the", "layers", "need", "to", "continuously", "adapt", "to", "the", "new", "distribution", ".", "When", "the", "input", "distribution", "to", "a", "learning", "system", "changes", ",", "it", "is", "said", "to", "experience", "covariate", "shift", ".", "This", "is", "typically", "handled", "via", "domain", "adaptation", ".", "However", ",", "the", "notion", "of", "covariate", "shift", "can", "be", "extended", "beyond", "the", "learning", "system", "as", "a", "whole", ",", "to", "apply", "to", "its", "parts", ",", "such", "as", "a", "sub", "-", "network", "or", "a", "layer", ".", "Consider", "a", "network", "computing", "where", "and", "are", "arbitrary", "transformations", ",", "and", "the", "parameters", "are", "to", "be", "learned", "so", "as", "to", "minimize", "the", "loss", ".", "Learning", "can", "be", "viewed", "as", "if", "the", "inputs", "are", "fed", "into", "the", "sub", "-", "network", "For", "example", ",", "a", "gradient", "descent", "step", "(", "for", "batch", "size", "and", "learning", "rate", ")", "is", "exactly", "equivalent", "to", "that", "for", "a", "stand", "-", "alone", "network", "with", "input", ".", "Therefore", ",", "the", "input", "distribution", "properties", "that", "make", "training", "more", "efficient", "–", "such", "as", "having", "the", "same", "distribution", "between", "the", "training", "and", "test", "data", "–", "apply", "to", "training", "the", "sub", "-", "network", "as", "well", ".", "As", "such", "it", "is", "advantageous", "for", "the", "distribution", "of", "to", "remain", "fixed", "over", "time", ".", "Then", ",", "does", "not", "have", "to", "readjust", "to", "compensate", "for", "the", "change", "in", "the", "distribution", "of", ".", "Fixed", "distribution", "of", "inputs", "to", "a", "sub", "-", "network", "would", "have", "positive", "consequences", "for", "the", "layers", "outside", "the", "sub", "-", "network", ",", "as", "well", ".", "Consider", "a", "layer", "with", "a", "sigmoid", "activation", "function", "where", "is", "the", "layer", "input", ",", "the", "weight", "matrix", "and", "bias", "vector", "are", "the", "layer", "parameters", "to", "be", "learned", ",", "and", ".", "As", "increases", ",", "tends", "to", "zero", ".", "This", "means", "that", "for", "all", "dimensions", "of", "except", "those", "with", "small", "absolute", "values", ",", "the", "gradient", "flowing", "down", "to", "will", "vanish", "and", "the", "model", "will", "train", "slowly", ".", "However", ",", "since", "is", "affected", "by", "and", "the", "parameters", "of", "all", "the", "layers", "below", ",", "changes", "to", "those", "parameters", "during", "training", "will", "likely", "move", "many", "dimensions", "of", "into", "the", "saturated", "regime", "of", "the", "nonlinearity", "and", "slow", "down", "the", "convergence", ".", "This", "effect", "is", "amplified", "as", "the", "network", "depth", "increases", ".", "In", "practice", ",", "the", "saturation", "problem", "and", "the", "resulting", "vanishing", "gradients", "are", "usually", "addressed", "by", "using", "Rectified", "Linear", "Units", ",", "careful", "initialization", ",", "and", "small", "learning", "rates", ".", "If", ",", "however", ",", "we", "could", "ensure", "that", "the", "distribution", "of", "nonlinearity", "inputs", "remains", "more", "stable", "as", "the", "network", "trains", ",", "then", "the", "optimizer", "would", "be", "less", "likely", "to", "get", "stuck", "in", "the", "saturated", "regime", ",", "and", "the", "training", "would", "accelerate", ".", "We", "refer", "to", "the", "change", "in", "the", "distributions", "of", "internal", "nodes", "of", "a", "deep", "network", ",", "in", "the", "course", "of", "training", ",", "as", "Internal", "Covariate", "Shift", ".", "Eliminating", "it", "offers", "a", "promise", "of", "faster", "training", ".", "We", "propose", "a", "new", "mechanism", ",", "which", "we", "call", "Batch", "Normalization", ",", "that", "takes", "a", "step", "towards", "reducing", "internal", "covariate", "shift", ",", "and", "in", "doing", "so", "dramatically", "accelerates", "the", "training", "of", "deep", "neural", "nets", ".", "It", "accomplishes", "this", "via", "a", "normalization", "step", "that", "fixes", "the", "means", "and", "variances", "of", "layer", "inputs", ".", "Batch", "Normalization", "also", "has", "a", "beneficial", "effect", "on", "the", "gradient", "flow", "through", "the", "network", ",", "by", "reducing", "the", "dependence", "of", "gradients", "on", "the", "scale", "of", "the", "parameters", "or", "of", "their", "initial", "values", ".", "This", "allows", "us", "to", "use", "much", "higher", "learning", "rates", "without", "the", "risk", "of", "divergence", ".", "Furthermore", ",", "batch", "normalization", "regularizes", "the", "model", "and", "reduces", "the", "need", "for", "Dropout", ".", "Finally", ",", "Batch", "Normalization", "makes", "it", "possible", "to", "use", "saturating", "nonlinearities", "by", "preventing", "the", "network", "from", "getting", "stuck", "in", "the", "saturated", "modes", ".", "In", "Sec", ".", "[", "reference", "]", ",", "we", "apply", "Batch", "Normalization", "to", "the", "best", "-", "performing", "ImageNet", "classification", "network", ",", "and", "show", "that", "we", "can", "match", "its", "performance", "using", "only", "7", "%", "of", "the", "training", "steps", ",", "and", "can", "further", "exceed", "its", "accuracy", "by", "a", "substantial", "margin", ".", "Using", "an", "ensemble", "of", "such", "networks", "trained", "with", "Batch", "Normalization", ",", "we", "achieve", "the", "top", "-", "5", "error", "rate", "that", "improves", "upon", "the", "best", "known", "results", "on", "ImageNet", "classification", ".", "section", ":", "Towards", "Reducing", "Internal", "Covariate", "Shift", "We", "define", "Internal", "Covariate", "Shift", "as", "the", "change", "in", "the", "distribution", "of", "network", "activations", "due", "to", "the", "change", "in", "network", "parameters", "during", "training", ".", "To", "improve", "the", "training", ",", "we", "seek", "to", "reduce", "the", "internal", "covariate", "shift", ".", "By", "fixing", "the", "distribution", "of", "the", "layer", "inputs", "as", "the", "training", "progresses", ",", "we", "expect", "to", "improve", "the", "training", "speed", ".", "It", "has", "been", "long", "known", "that", "the", "network", "training", "converges", "faster", "if", "its", "inputs", "are", "whitened", "–", "i.e.", ",", "linearly", "transformed", "to", "have", "zero", "means", "and", "unit", "variances", ",", "and", "decorrelated", ".", "As", "each", "layer", "observes", "the", "inputs", "produced", "by", "the", "layers", "below", ",", "it", "would", "be", "advantageous", "to", "achieve", "the", "same", "whitening", "of", "the", "inputs", "of", "each", "layer", ".", "By", "whitening", "the", "inputs", "to", "each", "layer", ",", "we", "would", "take", "a", "step", "towards", "achieving", "the", "fixed", "distributions", "of", "inputs", "that", "would", "remove", "the", "ill", "effects", "of", "the", "internal", "covariate", "shift", ".", "We", "could", "consider", "whitening", "activations", "at", "every", "training", "step", "or", "at", "some", "interval", ",", "either", "by", "modifying", "the", "network", "directly", "or", "by", "changing", "the", "parameters", "of", "the", "optimization", "algorithm", "to", "depend", "on", "the", "network", "activation", "values", ".", "However", ",", "if", "these", "modifications", "are", "interspersed", "with", "the", "optimization", "steps", ",", "then", "the", "gradient", "descent", "step", "may", "attempt", "to", "update", "the", "parameters", "in", "a", "way", "that", "requires", "the", "normalization", "to", "be", "updated", ",", "which", "reduces", "the", "effect", "of", "the", "gradient", "step", ".", "For", "example", ",", "consider", "a", "layer", "with", "the", "input", "that", "adds", "the", "learned", "bias", ",", "and", "normalizes", "the", "result", "by", "subtracting", "the", "mean", "of", "the", "activation", "computed", "over", "the", "training", "data", ":", "where", ",", "is", "the", "set", "of", "values", "of", "over", "the", "training", "set", ",", "and", ".", "If", "a", "gradient", "descent", "step", "ignores", "the", "dependence", "of", "on", ",", "then", "it", "will", "update", ",", "where", ".", "Then", ".", "Thus", ",", "the", "combination", "of", "the", "update", "to", "and", "subsequent", "change", "in", "normalization", "led", "to", "no", "change", "in", "the", "output", "of", "the", "layer", "nor", ",", "consequently", ",", "the", "loss", ".", "As", "the", "training", "continues", ",", "will", "grow", "indefinitely", "while", "the", "loss", "remains", "fixed", ".", "This", "problem", "can", "get", "worse", "if", "the", "normalization", "not", "only", "centers", "but", "also", "scales", "the", "activations", ".", "We", "have", "observed", "this", "empirically", "in", "initial", "experiments", ",", "where", "the", "model", "blows", "up", "when", "the", "normalization", "parameters", "are", "computed", "outside", "the", "gradient", "descent", "step", ".", "The", "issue", "with", "the", "above", "approach", "is", "that", "the", "gradient", "descent", "optimization", "does", "not", "take", "into", "account", "the", "fact", "that", "the", "normalization", "takes", "place", ".", "To", "address", "this", "issue", ",", "we", "would", "like", "to", "ensure", "that", ",", "for", "any", "parameter", "values", ",", "the", "network", "always", "produces", "activations", "with", "the", "desired", "distribution", ".", "Doing", "so", "would", "allow", "the", "gradient", "of", "the", "loss", "with", "respect", "to", "the", "model", "parameters", "to", "account", "for", "the", "normalization", ",", "and", "for", "its", "dependence", "on", "the", "model", "parameters", ".", "Let", "again", "be", "a", "layer", "input", ",", "treated", "as", "a", "vector", ",", "and", "be", "the", "set", "of", "these", "inputs", "over", "the", "training", "data", "set", ".", "The", "normalization", "can", "then", "be", "written", "as", "a", "transformation", "which", "depends", "not", "only", "on", "the", "given", "training", "example", "but", "on", "all", "examples", "–", "each", "of", "which", "depends", "on", "if", "is", "generated", "by", "another", "layer", ".", "For", "backpropagation", ",", "we", "would", "need", "to", "compute", "the", "Jacobians", "ignoring", "the", "latter", "term", "would", "lead", "to", "the", "explosion", "described", "above", ".", "Within", "this", "framework", ",", "whitening", "the", "layer", "inputs", "is", "expensive", ",", "as", "it", "requires", "computing", "the", "covariance", "matrix", "and", "its", "inverse", "square", "root", ",", "to", "produce", "the", "whitened", "activations", ",", "as", "well", "as", "the", "derivatives", "of", "these", "transforms", "for", "backpropagation", ".", "This", "motivates", "us", "to", "seek", "an", "alternative", "that", "performs", "input", "normalization", "in", "a", "way", "that", "is", "differentiable", "and", "does", "not", "require", "the", "analysis", "of", "the", "entire", "training", "set", "after", "every", "parameter", "update", ".", "Some", "of", "the", "previous", "approaches", "(", "e.g.", ")", "use", "statistics", "computed", "over", "a", "single", "training", "example", ",", "or", ",", "in", "the", "case", "of", "image", "networks", ",", "over", "different", "feature", "maps", "at", "a", "given", "location", ".", "However", ",", "this", "changes", "the", "representation", "ability", "of", "a", "network", "by", "discarding", "the", "absolute", "scale", "of", "activations", ".", "We", "want", "to", "a", "preserve", "the", "information", "in", "the", "network", ",", "by", "normalizing", "the", "activations", "in", "a", "training", "example", "relative", "to", "the", "statistics", "of", "the", "entire", "training", "data", ".", "section", ":", "Normalization", "via", "Mini", "-", "Batch", "Statistics", "Since", "the", "full", "whitening", "of", "each", "layer", "’s", "inputs", "is", "costly", "and", "not", "everywhere", "differentiable", ",", "we", "make", "two", "necessary", "simplifications", ".", "The", "first", "is", "that", "instead", "of", "whitening", "the", "features", "in", "layer", "inputs", "and", "outputs", "jointly", ",", "we", "will", "normalize", "each", "scalar", "feature", "independently", ",", "by", "making", "it", "have", "the", "mean", "of", "zero", "and", "the", "variance", "of", "1", ".", "For", "a", "layer", "with", "-", "dimensional", "input", ",", "we", "will", "normalize", "each", "dimension", "where", "the", "expectation", "and", "variance", "are", "computed", "over", "the", "training", "data", "set", ".", "As", "shown", "in", ",", "such", "normalization", "speeds", "up", "convergence", ",", "even", "when", "the", "features", "are", "not", "decorrelated", ".", "Note", "that", "simply", "normalizing", "each", "input", "of", "a", "layer", "may", "change", "what", "the", "layer", "can", "represent", ".", "For", "instance", ",", "normalizing", "the", "inputs", "of", "a", "sigmoid", "would", "constrain", "them", "to", "the", "linear", "regime", "of", "the", "nonlinearity", ".", "To", "address", "this", ",", "we", "make", "sure", "that", "the", "transformation", "inserted", "in", "the", "network", "can", "represent", "the", "identity", "transform", ".", "To", "accomplish", "this", ",", "we", "introduce", ",", "for", "each", "activation", ",", "a", "pair", "of", "parameters", ",", "which", "scale", "and", "shift", "the", "normalized", "value", ":", "These", "parameters", "are", "learned", "along", "with", "the", "original", "model", "parameters", ",", "and", "restore", "the", "representation", "power", "of", "the", "network", ".", "Indeed", ",", "by", "setting", "and", ",", "we", "could", "recover", "the", "original", "activations", ",", "if", "that", "were", "the", "optimal", "thing", "to", "do", ".", "In", "the", "batch", "setting", "where", "each", "training", "step", "is", "based", "on", "the", "entire", "training", "set", ",", "we", "would", "use", "the", "whole", "set", "to", "normalize", "activations", ".", "However", ",", "this", "is", "impractical", "when", "using", "stochastic", "optimization", ".", "Therefore", ",", "we", "make", "the", "second", "simplification", ":", "since", "we", "use", "mini", "-", "batches", "in", "stochastic", "gradient", "training", ",", "each", "mini", "-", "batch", "produces", "estimates", "of", "the", "mean", "and", "variance", "of", "each", "activation", ".", "This", "way", ",", "the", "statistics", "used", "for", "normalization", "can", "fully", "participate", "in", "the", "gradient", "backpropagation", ".", "Note", "that", "the", "use", "of", "mini", "-", "batches", "is", "enabled", "by", "computation", "of", "per", "-", "dimension", "variances", "rather", "than", "joint", "covariances", ";", "in", "the", "joint", "case", ",", "regularization", "would", "be", "required", "since", "the", "mini", "-", "batch", "size", "is", "likely", "to", "be", "smaller", "than", "the", "number", "of", "activations", "being", "whitened", ",", "resulting", "in", "singular", "covariance", "matrices", ".", "Consider", "a", "mini", "-", "batch", "of", "size", ".", "Since", "the", "normalization", "is", "applied", "to", "each", "activation", "independently", ",", "let", "us", "focus", "on", "a", "particular", "activation", "and", "omit", "for", "clarity", ".", "We", "have", "values", "of", "this", "activation", "in", "the", "mini", "-", "batch", ",", "Let", "the", "normalized", "values", "be", ",", "and", "their", "linear", "transformations", "be", ".", "We", "refer", "to", "the", "transform", "as", "the", "Batch", "Normalizing", "Transform", ".", "We", "present", "the", "BN", "Transform", "in", "Algorithm", "[", "reference", "]", ".", "In", "the", "algorithm", ",", "is", "a", "constant", "added", "to", "the", "mini", "-", "batch", "variance", "for", "numerical", "stability", ".", "Batch", "Normalizing", "Transform", ",", "applied", "to", "activation", "x", "over", "a", "mini", "-", "batch", ".", "The", "BN", "transform", "can", "be", "added", "to", "a", "network", "to", "manipulate", "any", "activation", ".", "In", "the", "notation", ",", "we", "indicate", "that", "the", "parameters", "and", "are", "to", "be", "learned", ",", "but", "it", "should", "be", "noted", "that", "the", "BN", "transform", "does", "not", "independently", "process", "the", "activation", "in", "each", "training", "example", ".", "Rather", ",", "depends", "both", "on", "the", "training", "example", "and", "the", "other", "examples", "in", "the", "mini", "-", "batch", ".", "The", "scaled", "and", "shifted", "values", "are", "passed", "to", "other", "network", "layers", ".", "The", "normalized", "activations", "are", "internal", "to", "our", "transformation", ",", "but", "their", "presence", "is", "crucial", ".", "The", "distributions", "of", "values", "of", "any", "has", "the", "expected", "value", "of", "and", "the", "variance", "of", ",", "as", "long", "as", "the", "elements", "of", "each", "mini", "-", "batch", "are", "sampled", "from", "the", "same", "distribution", ",", "and", "if", "we", "neglect", ".", "This", "can", "be", "seen", "by", "observing", "that", "and", ",", "and", "taking", "expectations", ".", "Each", "normalized", "activation", "can", "be", "viewed", "as", "an", "input", "to", "a", "sub", "-", "network", "composed", "of", "the", "linear", "transform", ",", "followed", "by", "the", "other", "processing", "done", "by", "the", "original", "network", ".", "These", "sub", "-", "network", "inputs", "all", "have", "fixed", "means", "and", "variances", ",", "and", "although", "the", "joint", "distribution", "of", "these", "normalized", "can", "change", "over", "the", "course", "of", "training", ",", "we", "expect", "that", "the", "introduction", "of", "normalized", "inputs", "accelerates", "the", "training", "of", "the", "sub", "-", "network", "and", ",", "consequently", ",", "the", "network", "as", "a", "whole", ".", "During", "training", "we", "need", "to", "backpropagate", "the", "gradient", "of", "loss", "through", "this", "transformation", ",", "as", "well", "as", "compute", "the", "gradients", "with", "respect", "to", "the", "parameters", "of", "the", "BN", "transform", ".", "We", "use", "chain", "rule", ",", "as", "follows", "(", "before", "simplification", ")", ":", "Thus", ",", "BN", "transform", "is", "a", "differentiable", "transformation", "that", "introduces", "normalized", "activations", "into", "the", "network", ".", "This", "ensures", "that", "as", "the", "model", "is", "training", ",", "layers", "can", "continue", "learning", "on", "input", "distributions", "that", "exhibit", "less", "internal", "covariate", "shift", ",", "thus", "accelerating", "the", "training", ".", "Furthermore", ",", "the", "learned", "affine", "transform", "applied", "to", "these", "normalized", "activations", "allows", "the", "BN", "transform", "to", "represent", "the", "identity", "transformation", "and", "preserves", "the", "network", "capacity", ".", "subsection", ":", "Training", "and", "Inference", "with", "Batch", "-", "Normalized", "Networks", "To", "Batch", "-", "Normalize", "a", "network", ",", "we", "specify", "a", "subset", "of", "activations", "and", "insert", "the", "BN", "transform", "for", "each", "of", "them", ",", "according", "to", "Alg", ".", "[", "reference", "]", ".", "Any", "layer", "that", "previously", "received", "as", "the", "input", ",", "now", "receives", ".", "A", "model", "employing", "Batch", "Normalization", "can", "be", "trained", "using", "batch", "gradient", "descent", ",", "or", "Stochastic", "Gradient", "Descent", "with", "a", "mini", "-", "batch", "size", ",", "or", "with", "any", "of", "its", "variants", "such", "as", "Adagrad", ".", "The", "normalization", "of", "activations", "that", "depends", "on", "the", "mini", "-", "batch", "allows", "efficient", "training", ",", "but", "is", "neither", "necessary", "nor", "desirable", "during", "inference", ";", "we", "want", "the", "output", "to", "depend", "only", "on", "the", "input", ",", "deterministically", ".", "For", "this", ",", "once", "the", "network", "has", "been", "trained", ",", "we", "use", "the", "normalization", "using", "the", "population", ",", "rather", "than", "mini", "-", "batch", ",", "statistics", ".", "Neglecting", ",", "these", "normalized", "activations", "have", "the", "same", "mean", "0", "and", "variance", "1", "as", "during", "training", ".", "We", "use", "the", "unbiased", "variance", "estimate", ",", "where", "the", "expectation", "is", "over", "training", "mini", "-", "batches", "of", "size", "and", "are", "their", "sample", "variances", ".", "Using", "moving", "averages", "instead", ",", "we", "can", "track", "the", "accuracy", "of", "a", "model", "as", "it", "trains", ".", "Since", "the", "means", "and", "variances", "are", "fixed", "during", "inference", ",", "the", "normalization", "is", "simply", "a", "linear", "transform", "applied", "to", "each", "activation", ".", "It", "may", "further", "be", "composed", "with", "the", "scaling", "by", "and", "shift", "by", ",", "to", "yield", "a", "single", "linear", "transform", "that", "replaces", ".", "Algorithm", "[", "reference", "]", "summarizes", "the", "procedure", "for", "training", "batch", "-", "normalized", "networks", ".", "Training", "a", "Batch", "-", "Normalized", "Network", "[", "1", "]", "Batch", "-", "normalized", "network", "for", "inference", ",", "Training", "BN", "network", "Add", "transformation", "to", "(", "Alg", ".", "[", "reference", "]", ")", "Modify", "each", "layer", "in", "with", "input", "to", "take", "instead", "Train", "to", "optimize", "the", "parameters", "For", "clarity", ",", ",", "etc", ".", "Process", "multiple", "training", "mini", "-", "batches", ",", "each", "of", "size", ",", "and", "average", "over", "them", ":", "In", ",", "replace", "the", "transform", "with", "subsection", ":", "Batch", "-", "Normalized", "Convolutional", "Networks", "Batch", "Normalization", "can", "be", "applied", "to", "any", "set", "of", "activations", "in", "the", "network", ".", "Here", ",", "we", "focus", "on", "transforms", "that", "consist", "of", "an", "affine", "transformation", "followed", "by", "an", "element", "-", "wise", "nonlinearity", ":", "where", "and", "are", "learned", "parameters", "of", "the", "model", ",", "and", "is", "the", "nonlinearity", "such", "as", "sigmoid", "or", "ReLU", ".", "This", "formulation", "covers", "both", "fully", "-", "connected", "and", "convolutional", "layers", ".", "We", "add", "the", "BN", "transform", "immediately", "before", "the", "nonlinearity", ",", "by", "normalizing", ".", "We", "could", "have", "also", "normalized", "the", "layer", "inputs", ",", "but", "since", "is", "likely", "the", "output", "of", "another", "nonlinearity", ",", "the", "shape", "of", "its", "distribution", "is", "likely", "to", "change", "during", "training", ",", "and", "constraining", "its", "first", "and", "second", "moments", "would", "not", "eliminate", "the", "covariate", "shift", ".", "In", "contrast", ",", "is", "more", "likely", "to", "have", "a", "symmetric", ",", "non", "-", "sparse", "distribution", ",", "that", "is", "“", "more", "Gaussian", "”", ";", "normalizing", "it", "is", "likely", "to", "produce", "activations", "with", "a", "stable", "distribution", ".", "Note", "that", ",", "since", "we", "normalize", ",", "the", "bias", "can", "be", "ignored", "since", "its", "effect", "will", "be", "canceled", "by", "the", "subsequent", "mean", "subtraction", "(", "the", "role", "of", "the", "bias", "is", "subsumed", "by", "in", "Alg", ".", "[", "reference", "]", ")", ".", "Thus", ",", "is", "replaced", "with", "where", "the", "BN", "transform", "is", "applied", "independently", "to", "each", "dimension", "of", ",", "with", "a", "separate", "pair", "of", "learned", "parameters", ",", "per", "dimension", ".", "For", "convolutional", "layers", ",", "we", "additionally", "want", "the", "normalization", "to", "obey", "the", "convolutional", "property", "–", "so", "that", "different", "elements", "of", "the", "same", "feature", "map", ",", "at", "different", "locations", ",", "are", "normalized", "in", "the", "same", "way", ".", "To", "achieve", "this", ",", "we", "jointly", "normalize", "all", "the", "activations", "in", "a", "mini", "-", "batch", ",", "over", "all", "locations", ".", "In", "Alg", ".", "[", "reference", "]", ",", "we", "let", "be", "the", "set", "of", "all", "values", "in", "a", "feature", "map", "across", "both", "the", "elements", "of", "a", "mini", "-", "batch", "and", "spatial", "locations", "–", "so", "for", "a", "mini", "-", "batch", "of", "size", "and", "feature", "maps", "of", "size", ",", "we", "use", "the", "effective", "mini", "-", "batch", "of", "size", ".", "We", "learn", "a", "pair", "of", "parameters", "and", "per", "feature", "map", ",", "rather", "than", "per", "activation", ".", "Alg", ".", "[", "reference", "]", "is", "modified", "similarly", ",", "so", "that", "during", "inference", "the", "BN", "transform", "applies", "the", "same", "linear", "transformation", "to", "each", "activation", "in", "a", "given", "feature", "map", ".", "subsection", ":", "Batch", "Normalization", "enables", "higher", "learning", "rates", "In", "traditional", "deep", "networks", ",", "too", "-", "high", "learning", "rate", "may", "result", "in", "the", "gradients", "that", "explode", "or", "vanish", ",", "as", "well", "as", "getting", "stuck", "in", "poor", "local", "minima", ".", "Batch", "Normalization", "helps", "address", "these", "issues", ".", "By", "normalizing", "activations", "throughout", "the", "network", ",", "it", "prevents", "small", "changes", "to", "the", "parameters", "from", "amplifying", "into", "larger", "and", "suboptimal", "changes", "in", "activations", "in", "gradients", ";", "for", "instance", ",", "it", "prevents", "the", "training", "from", "getting", "stuck", "in", "the", "saturated", "regimes", "of", "nonlinearities", ".", "Batch", "Normalization", "also", "makes", "training", "more", "resilient", "to", "the", "parameter", "scale", ".", "Normally", ",", "large", "learning", "rates", "may", "increase", "the", "scale", "of", "layer", "parameters", ",", "which", "then", "amplify", "the", "gradient", "during", "backpropagation", "and", "lead", "to", "the", "model", "explosion", ".", "However", ",", "with", "Batch", "Normalization", ",", "backpropagation", "through", "a", "layer", "is", "unaffected", "by", "the", "scale", "of", "its", "parameters", ".", "Indeed", ",", "for", "a", "scalar", ",", "and", "we", "can", "show", "that", "The", "scale", "does", "not", "affect", "the", "layer", "Jacobian", "nor", ",", "consequently", ",", "the", "gradient", "propagation", ".", "Moreover", ",", "larger", "weights", "lead", "to", "smaller", "gradients", ",", "and", "Batch", "Normalization", "will", "stabilize", "the", "parameter", "growth", ".", "We", "further", "conjecture", "that", "Batch", "Normalization", "may", "lead", "the", "layer", "Jacobians", "to", "have", "singular", "values", "close", "to", "1", ",", "which", "is", "known", "to", "be", "beneficial", "for", "training", ".", "Consider", "two", "consecutive", "layers", "with", "normalized", "inputs", ",", "and", "the", "transformation", "between", "these", "normalized", "vectors", ":", ".", "If", "we", "assume", "that", "and", "are", "Gaussian", "and", "uncorrelated", ",", "and", "that", "is", "a", "linear", "transformation", "for", "the", "given", "model", "parameters", ",", "then", "both", "and", "have", "unit", "covariances", ",", "and", ".", "Thus", ",", ",", "and", "so", "all", "singular", "values", "of", "are", "equal", "to", "1", ",", "which", "preserves", "the", "gradient", "magnitudes", "during", "backpropagation", ".", "In", "reality", ",", "the", "transformation", "is", "not", "linear", ",", "and", "the", "normalized", "values", "are", "not", "guaranteed", "to", "be", "Gaussian", "nor", "independent", ",", "but", "we", "nevertheless", "expect", "Batch", "Normalization", "to", "help", "make", "gradient", "propagation", "better", "behaved", ".", "The", "precise", "effect", "of", "Batch", "Normalization", "on", "gradient", "propagation", "remains", "an", "area", "of", "further", "study", ".", "subsection", ":", "Batch", "Normalization", "regularizes", "the", "model", "When", "training", "with", "Batch", "Normalization", ",", "a", "training", "example", "is", "seen", "in", "conjunction", "with", "other", "examples", "in", "the", "mini", "-", "batch", ",", "and", "the", "training", "network", "no", "longer", "producing", "deterministic", "values", "for", "a", "given", "training", "example", ".", "In", "our", "experiments", ",", "we", "found", "this", "effect", "to", "be", "advantageous", "to", "the", "generalization", "of", "the", "network", ".", "Whereas", "Dropout", "is", "typically", "used", "to", "reduce", "overfitting", ",", "in", "a", "batch", "-", "normalized", "network", "we", "found", "that", "it", "can", "be", "either", "removed", "or", "reduced", "in", "strength", ".", "section", ":", "Experiments", "subsection", ":", "Activations", "over", "time", "To", "verify", "the", "effects", "of", "internal", "covariate", "shift", "on", "training", ",", "and", "the", "ability", "of", "Batch", "Normalization", "to", "combat", "it", ",", "we", "considered", "the", "problem", "of", "predicting", "the", "digit", "class", "on", "the", "MNIST", "dataset", ".", "We", "used", "a", "very", "simple", "network", ",", "with", "a", "28x28", "binary", "image", "as", "input", ",", "and", "3", "fully", "-", "connected", "hidden", "layers", "with", "100", "activations", "each", ".", "Each", "hidden", "layer", "computes", "with", "sigmoid", "nonlinearity", ",", "and", "the", "weights", "initialized", "to", "small", "random", "Gaussian", "values", ".", "The", "last", "hidden", "layer", "is", "followed", "by", "a", "fully", "-", "connected", "layer", "with", "10", "activations", "(", "one", "per", "class", ")", "and", "cross", "-", "entropy", "loss", ".", "We", "trained", "the", "network", "for", "50000", "steps", ",", "with", "60", "examples", "per", "mini", "-", "batch", ".", "We", "added", "Batch", "Normalization", "to", "each", "hidden", "layer", "of", "the", "network", ",", "as", "in", "Sec", ".", "[", "reference", "]", ".", "We", "were", "interested", "in", "the", "comparison", "between", "the", "baseline", "and", "batch", "-", "normalized", "networks", ",", "rather", "than", "achieving", "the", "state", "of", "the", "art", "performance", "on", "MNIST", "(", "which", "the", "described", "architecture", "does", "not", ")", ".", "Figure", "[", "reference", "]", "(", "a", ")", "shows", "the", "fraction", "of", "correct", "predictions", "by", "the", "two", "networks", "on", "held", "-", "out", "test", "data", ",", "as", "training", "progresses", ".", "The", "batch", "-", "normalized", "network", "enjoys", "the", "higher", "test", "accuracy", ".", "To", "investigate", "why", ",", "we", "studied", "inputs", "to", "the", "sigmoid", ",", "in", "the", "original", "network", "N", "and", "batch", "-", "normalized", "network", "(", "Alg", ".", "[", "reference", "]", ")", "over", "the", "course", "of", "training", ".", "In", "Fig", ".", "[", "reference", "]", "(", "b", ",", "c", ")", "we", "show", ",", "for", "one", "typical", "activation", "from", "the", "last", "hidden", "layer", "of", "each", "network", ",", "how", "its", "distribution", "evolves", ".", "The", "distributions", "in", "the", "original", "network", "change", "significantly", "over", "time", ",", "both", "in", "their", "mean", "and", "the", "variance", ",", "which", "complicates", "the", "training", "of", "the", "subsequent", "layers", ".", "In", "contrast", ",", "the", "distributions", "in", "the", "batch", "-", "normalized", "network", "are", "much", "more", "stable", "as", "training", "progresses", ",", "which", "aids", "the", "training", ".", "subsection", ":", "ImageNet", "classification", "We", "applied", "Batch", "Normalization", "to", "a", "new", "variant", "of", "the", "Inception", "network", ",", "trained", "on", "the", "ImageNet", "classification", "task", ".", "The", "network", "has", "a", "large", "number", "of", "convolutional", "and", "pooling", "layers", ",", "with", "a", "softmax", "layer", "to", "predict", "the", "image", "class", ",", "out", "of", "1000", "possibilities", ".", "Convolutional", "layers", "use", "ReLU", "as", "the", "nonlinearity", ".", "The", "main", "difference", "to", "the", "network", "described", "in", "is", "that", "the", "convolutional", "layers", "are", "replaced", "by", "two", "consecutive", "layers", "of", "convolutions", "with", "up", "to", "filters", ".", "The", "network", "contains", "parameters", ",", "and", ",", "other", "than", "the", "top", "softmax", "layer", ",", "has", "no", "fully", "-", "connected", "layers", ".", "More", "details", "are", "given", "in", "the", "Appendix", ".", "We", "refer", "to", "this", "model", "as", "Inception", "in", "the", "rest", "of", "the", "text", ".", "The", "model", "was", "trained", "using", "a", "version", "of", "Stochastic", "Gradient", "Descent", "with", "momentum", ",", "using", "the", "mini", "-", "batch", "size", "of", "32", ".", "The", "training", "was", "performed", "using", "a", "large", "-", "scale", ",", "distributed", "architecture", "(", "similar", "to", ")", ".", "All", "networks", "are", "evaluated", "as", "training", "progresses", "by", "computing", "the", "validation", "accuracy", ",", "i.e.", "the", "probability", "of", "predicting", "the", "correct", "label", "out", "of", "1000", "possibilities", ",", "on", "a", "held", "-", "out", "set", ",", "using", "a", "single", "crop", "per", "image", ".", "In", "our", "experiments", ",", "we", "evaluated", "several", "modifications", "of", "Inception", "with", "Batch", "Normalization", ".", "In", "all", "cases", ",", "Batch", "Normalization", "was", "applied", "to", "the", "input", "of", "each", "nonlinearity", ",", "in", "a", "convolutional", "way", ",", "as", "described", "in", "section", "[", "reference", "]", ",", "while", "keeping", "the", "rest", "of", "the", "architecture", "constant", ".", "subsubsection", ":", "Accelerating", "BN", "Networks", "Simply", "adding", "Batch", "Normalization", "to", "a", "network", "does", "not", "take", "full", "advantage", "of", "our", "method", ".", "To", "do", "so", ",", "we", "further", "changed", "the", "network", "and", "its", "training", "parameters", ",", "as", "follows", ":", "Increase", "learning", "rate", ".", "In", "a", "batch", "-", "normalized", "model", ",", "we", "have", "been", "able", "to", "achieve", "a", "training", "speedup", "from", "higher", "learning", "rates", ",", "with", "no", "ill", "side", "effects", "(", "Sec", ".", "[", "reference", "]", ")", ".", "Remove", "Dropout", ".", "As", "described", "in", "Sec", ".", "[", "reference", "]", ",", "Batch", "Normalization", "fulfills", "some", "of", "the", "same", "goals", "as", "Dropout", ".", "Removing", "Dropout", "from", "Modified", "BN", "-", "Inception", "speeds", "up", "training", ",", "without", "increasing", "overfitting", ".", "Reduce", "the", "L2", "weight", "regularization", ".", "While", "in", "Inception", "an", "loss", "on", "the", "model", "parameters", "controls", "overfitting", ",", "in", "Modified", "BN", "-", "Inception", "the", "weight", "of", "this", "loss", "is", "reduced", "by", "a", "factor", "of", "5", ".", "We", "find", "that", "this", "improves", "the", "accuracy", "on", "the", "held", "-", "out", "validation", "data", ".", "Accelerate", "the", "learning", "rate", "decay", ".", "In", "training", "Inception", ",", "learning", "rate", "was", "decayed", "exponentially", ".", "Because", "our", "network", "trains", "faster", "than", "Inception", ",", "we", "lower", "the", "learning", "rate", "6", "times", "faster", ".", "Remove", "Local", "Response", "Normalization", "While", "Inception", "and", "other", "networks", "benefit", "from", "it", ",", "we", "found", "that", "with", "Batch", "Normalization", "it", "is", "not", "necessary", ".", "Shuffle", "training", "examples", "more", "thoroughly", ".", "We", "enabled", "within", "-", "shard", "shuffling", "of", "the", "training", "data", ",", "which", "prevents", "the", "same", "examples", "from", "always", "appearing", "in", "a", "mini", "-", "batch", "together", ".", "This", "led", "to", "about", "1", "%", "improvements", "in", "the", "validation", "accuracy", ",", "which", "is", "consistent", "with", "the", "view", "of", "Batch", "Normalization", "as", "a", "regularizer", "(", "Sec", ".", "[", "reference", "]", ")", ":", "the", "randomization", "inherent", "in", "our", "method", "should", "be", "most", "beneficial", "when", "it", "affects", "an", "example", "differently", "each", "time", "it", "is", "seen", ".", "Reduce", "the", "photometric", "distortions", ".", "Because", "batch", "-", "normalized", "networks", "train", "faster", "and", "observe", "each", "training", "example", "fewer", "times", ",", "we", "let", "the", "trainer", "focus", "on", "more", "“", "real", "”", "images", "by", "distorting", "them", "less", ".", "subsubsection", ":", "Single", "-", "Network", "Classification", "We", "evaluated", "the", "following", "networks", ",", "all", "trained", "on", "the", "LSVRC2012", "training", "data", ",", "and", "tested", "on", "the", "validation", "data", ":", "Inception", ":", "the", "network", "described", "at", "the", "beginning", "of", "Section", "[", "reference", "]", ",", "trained", "with", "the", "initial", "learning", "rate", "of", "0.0015", ".", "BN", "-", "Baseline", ":", "Same", "as", "Inception", "with", "Batch", "Normalization", "before", "each", "nonlinearity", ".", "BN", "-", "x5", ":", "Inception", "with", "Batch", "Normalization", "and", "the", "modifications", "in", "Sec", ".", "[", "reference", "]", ".", "The", "initial", "learning", "rate", "was", "increased", "by", "a", "factor", "of", "5", ",", "to", "0.0075", ".", "The", "same", "learning", "rate", "increase", "with", "original", "Inception", "caused", "the", "model", "parameters", "to", "reach", "machine", "infinity", ".", "BN", "-", "x30", ":", "Like", "BN", "-", "x5", ",", "but", "with", "the", "initial", "learning", "rate", "0.045", "(", "30", "times", "that", "of", "Inception", ")", ".", "BN", "-", "x5", "-", "Sigmoid", ":", "Like", "BN", "-", "x5", ",", "but", "with", "sigmoid", "nonlinearity", "instead", "of", "ReLU", ".", "We", "also", "attempted", "to", "train", "the", "original", "Inception", "with", "sigmoid", ",", "but", "the", "model", "remained", "at", "the", "accuracy", "equivalent", "to", "chance", ".", "In", "Figure", "[", "reference", "]", ",", "we", "show", "the", "validation", "accuracy", "of", "the", "networks", ",", "as", "a", "function", "of", "the", "number", "of", "training", "steps", ".", "Inception", "reached", "the", "accuracy", "of", "72.2", "%", "after", "training", "steps", ".", "The", "Figure", "[", "reference", "]", "shows", ",", "for", "each", "network", ",", "the", "number", "of", "training", "steps", "required", "to", "reach", "the", "same", "72.2", "%", "accuracy", ",", "as", "well", "as", "the", "maximum", "validation", "accuracy", "reached", "by", "the", "network", "and", "the", "number", "of", "steps", "to", "reach", "it", ".", "By", "only", "using", "Batch", "Normalization", "(", "BN", "-", "Baseline", ")", ",", "we", "match", "the", "accuracy", "of", "Inception", "in", "less", "than", "half", "the", "number", "of", "training", "steps", ".", "By", "applying", "the", "modifications", "in", "Sec", ".", "[", "reference", "]", ",", "we", "significantly", "increase", "the", "training", "speed", "of", "the", "network", ".", "BN", "-", "x5", "needs", "14", "times", "fewer", "steps", "than", "Inception", "to", "reach", "the", "72.2", "%", "accuracy", ".", "Interestingly", ",", "increasing", "the", "learning", "rate", "further", "(", "BN", "-", "x30", ")", "causes", "the", "model", "to", "train", "somewhat", "slower", "initially", ",", "but", "allows", "it", "to", "reach", "a", "higher", "final", "accuracy", ".", "It", "reaches", "74.8", "%", "after", "steps", ",", "i.e.", "5", "times", "fewer", "steps", "than", "required", "by", "Inception", "to", "reach", "72.2", "%", ".", "We", "also", "verified", "that", "the", "reduction", "in", "internal", "covariate", "shift", "allows", "deep", "networks", "with", "Batch", "Normalization", "to", "be", "trained", "when", "sigmoid", "is", "used", "as", "the", "nonlinearity", ",", "despite", "the", "well", "-", "known", "difficulty", "of", "training", "such", "networks", ".", "Indeed", ",", "BN", "-", "x5", "-", "Sigmoid", "achieves", "the", "accuracy", "of", "69.8", "%", ".", "Without", "Batch", "Normalization", ",", "Inception", "with", "sigmoid", "never", "achieves", "better", "than", "accuracy", ".", "subsubsection", ":", "Ensemble", "Classification", "The", "current", "reported", "best", "results", "on", "the", "ImageNet", "Large", "Scale", "Visual", "Recognition", "Competition", "are", "reached", "by", "the", "Deep", "Image", "ensemble", "of", "traditional", "models", "and", "the", "ensemble", "model", "of", ".", "The", "latter", "reports", "the", "top", "-", "5", "error", "of", "4.94", "%", ",", "as", "evaluated", "by", "the", "ILSVRC", "server", ".", "Here", "we", "report", "a", "top", "-", "5", "validation", "error", "of", "4.9", "%", ",", "and", "test", "error", "of", "4.82", "%", "(", "according", "to", "the", "ILSVRC", "server", ")", ".", "This", "improves", "upon", "the", "previous", "best", "result", ",", "and", "exceeds", "the", "estimated", "accuracy", "of", "human", "raters", "according", "to", ".", "For", "our", "ensemble", ",", "we", "used", "6", "networks", ".", "Each", "was", "based", "on", "BN", "-", "x30", ",", "modified", "via", "some", "of", "the", "following", ":", "increased", "initial", "weights", "in", "the", "convolutional", "layers", ";", "using", "Dropout", "(", "with", "the", "Dropout", "probability", "of", "5", "%", "or", "10", "%", ",", "vs.", "40", "%", "for", "the", "original", "Inception", ")", ";", "and", "using", "non", "-", "convolutional", ",", "per", "-", "activation", "Batch", "Normalization", "with", "last", "hidden", "layers", "of", "the", "model", ".", "Each", "network", "achieved", "its", "maximum", "accuracy", "after", "about", "training", "steps", ".", "The", "ensemble", "prediction", "was", "based", "on", "the", "arithmetic", "average", "of", "class", "probabilities", "predicted", "by", "the", "constituent", "networks", ".", "The", "details", "of", "ensemble", "and", "multicrop", "inference", "are", "similar", "to", ".", "We", "demonstrate", "in", "Fig", ".", "[", "reference", "]", "that", "batch", "normalization", "allows", "us", "to", "set", "new", "state", "-", "of", "-", "the", "-", "art", "by", "a", "healthy", "margin", "on", "the", "ImageNet", "classification", "challenge", "benchmarks", ".", "section", ":", "Conclusion", "We", "have", "presented", "a", "novel", "mechanism", "for", "dramatically", "accelerating", "the", "training", "of", "deep", "networks", ".", "It", "is", "based", "on", "the", "premise", "that", "covariate", "shift", ",", "which", "is", "known", "to", "complicate", "the", "training", "of", "machine", "learning", "systems", ",", "also", "applies", "to", "sub", "-", "networks", "and", "layers", ",", "and", "removing", "it", "from", "internal", "activations", "of", "the", "network", "may", "aid", "in", "training", ".", "Our", "proposed", "method", "draws", "its", "power", "from", "normalizing", "activations", ",", "and", "from", "incorporating", "this", "normalization", "in", "the", "network", "architecture", "itself", ".", "This", "ensures", "that", "the", "normalization", "is", "appropriately", "handled", "by", "any", "optimization", "method", "that", "is", "being", "used", "to", "train", "the", "network", ".", "To", "enable", "stochastic", "optimization", "methods", "commonly", "used", "in", "deep", "network", "training", ",", "we", "perform", "the", "normalization", "for", "each", "mini", "-", "batch", ",", "and", "backpropagate", "the", "gradients", "through", "the", "normalization", "parameters", ".", "Batch", "Normalization", "adds", "only", "two", "extra", "parameters", "per", "activation", ",", "and", "in", "doing", "so", "preserves", "the", "representation", "ability", "of", "the", "network", ".", "We", "presented", "an", "algorithm", "for", "constructing", ",", "training", ",", "and", "performing", "inference", "with", "batch", "-", "normalized", "networks", ".", "The", "resulting", "networks", "can", "be", "trained", "with", "saturating", "nonlinearities", ",", "are", "more", "tolerant", "to", "increased", "training", "rates", ",", "and", "often", "do", "not", "require", "Dropout", "for", "regularization", ".", "Merely", "adding", "Batch", "Normalization", "to", "a", "state", "-", "of", "-", "the", "-", "art", "image", "classification", "model", "yields", "a", "substantial", "speedup", "in", "training", ".", "By", "further", "increasing", "the", "learning", "rates", ",", "removing", "Dropout", ",", "and", "applying", "other", "modifications", "afforded", "by", "Batch", "Normalization", ",", "we", "reach", "the", "previous", "state", "of", "the", "art", "with", "only", "a", "small", "fraction", "of", "training", "steps", "–", "and", "then", "beat", "the", "state", "of", "the", "art", "in", "single", "-", "network", "image", "classification", ".", "Furthermore", ",", "by", "combining", "multiple", "models", "trained", "with", "Batch", "Normalization", ",", "we", "perform", "better", "than", "the", "best", "known", "system", "on", "ImageNet", ",", "by", "a", "significant", "margin", ".", "Interestingly", ",", "our", "method", "bears", "similarity", "to", "the", "standardization", "layer", "of", ",", "though", "the", "two", "methods", "stem", "from", "very", "different", "goals", ",", "and", "perform", "different", "tasks", ".", "The", "goal", "of", "Batch", "Normalization", "is", "to", "achieve", "a", "stable", "distribution", "of", "activation", "values", "throughout", "training", ",", "and", "in", "our", "experiments", "we", "apply", "it", "before", "the", "nonlinearity", "since", "that", "is", "where", "matching", "the", "first", "and", "second", "moments", "is", "more", "likely", "to", "result", "in", "a", "stable", "distribution", ".", "On", "the", "contrary", ",", "apply", "the", "standardization", "layer", "to", "the", "output", "of", "the", "nonlinearity", ",", "which", "results", "in", "sparser", "activations", ".", "In", "our", "large", "-", "scale", "image", "classification", "experiments", ",", "we", "have", "not", "observed", "the", "nonlinearity", "inputs", "to", "be", "sparse", ",", "neither", "with", "nor", "without", "Batch", "Normalization", ".", "Other", "notable", "differentiating", "characteristics", "of", "Batch", "Normalization", "include", "the", "learned", "scale", "and", "shift", "that", "allow", "the", "BN", "transform", "to", "represent", "identity", "(", "the", "standardization", "layer", "did", "not", "require", "this", "since", "it", "was", "followed", "by", "the", "learned", "linear", "transform", "that", ",", "conceptually", ",", "absorbs", "the", "necessary", "scale", "and", "shift", ")", ",", "handling", "of", "convolutional", "layers", ",", "deterministic", "inference", "that", "does", "not", "depend", "on", "the", "mini", "-", "batch", ",", "and", "batch", "-", "normalizing", "each", "convolutional", "layer", "in", "the", "network", ".", "In", "this", "work", ",", "we", "have", "not", "explored", "the", "full", "range", "of", "possibilities", "that", "Batch", "Normalization", "potentially", "enables", ".", "Our", "future", "work", "includes", "applications", "of", "our", "method", "to", "Recurrent", "Neural", "Networks", ",", "where", "the", "internal", "covariate", "shift", "and", "the", "vanishing", "or", "exploding", "gradients", "may", "be", "especially", "severe", ",", "and", "which", "would", "allow", "us", "to", "more", "thoroughly", "test", "the", "hypothesis", "that", "normalization", "improves", "gradient", "propagation", "(", "Sec", ".", "[", "reference", "]", ")", ".", "We", "plan", "to", "investigate", "whether", "Batch", "Normalization", "can", "help", "with", "domain", "adaptation", ",", "in", "its", "traditional", "sense", "–", "i.e.", "whether", "the", "normalization", "performed", "by", "the", "network", "would", "allow", "it", "to", "more", "easily", "generalize", "to", "new", "data", "distributions", ",", "perhaps", "with", "just", "a", "recomputation", "of", "the", "population", "means", "and", "variances", "(", "Alg", ".", "[", "reference", "]", ")", ".", "Finally", ",", "we", "believe", "that", "further", "theoretical", "analysis", "of", "the", "algorithm", "would", "allow", "still", "more", "improvements", "and", "applications", ".", "bibliography", ":", "References", "section", ":", "Appendix", "subsection", ":", "Variant", "of", "the", "Inception", "Model", "Used", "Figure", "[", "reference", "]", "documents", "the", "changes", "that", "were", "performed", "compared", "to", "the", "architecture", "with", "respect", "to", "the", "GoogleNet", "archictecture", ".", "For", "the", "interpretation", "of", "this", "table", ",", "please", "consult", ".", "The", "notable", "architecture", "changes", "compared", "to", "the", "GoogLeNet", "model", "include", ":", "The", "5", "5", "convolutional", "layers", "are", "replaced", "by", "two", "consecutive", "3", "3", "convolutional", "layers", ".", "This", "increases", "the", "maximum", "depth", "of", "the", "network", "by", "9", "weight", "layers", ".", "Also", "it", "increases", "the", "number", "of", "parameters", "by", "25", "%", "and", "the", "computational", "cost", "is", "increased", "by", "about", "30", "%", ".", "The", "number", "28", "28", "inception", "modules", "is", "increased", "from", "2", "to", "3", ".", "Inside", "the", "modules", ",", "sometimes", "average", ",", "sometimes", "maximum", "-", "pooling", "is", "employed", ".", "This", "is", "indicated", "in", "the", "entries", "corresponding", "to", "the", "pooling", "layers", "of", "the", "table", ".", "There", "are", "no", "across", "the", "board", "pooling", "layers", "between", "any", "two", "Inception", "modules", ",", "but", "stride", "-", "2", "convolution", "/", "pooling", "layers", "are", "employed", "before", "the", "filter", "concatenation", "in", "the", "modules", "3c", ",", "4e", ".", "Our", "model", "employed", "separable", "convolution", "with", "depth", "multiplier", "on", "the", "first", "convolutional", "layer", ".", "This", "reduces", "the", "computational", "cost", "while", "increasing", "the", "memory", "consumption", "at", "training", "time", "."], "entity": [[2, 4, ["Batch", "Normalization"], "Task"], [6, 9, ["Deep", "Network", "Training"], "Method"], [15, 18, ["Deep", "Neural", "Networks"], "Method"], [48, 49, ["training"], "Task"], [52, 54, ["learning", "rates"], "Metric"], [56, 58, ["parameter", "initialization"], "Method"], [97, 98, ["normalization"], "Method"], [102, 104, ["model", "architecture"], "Method"], [107, 108, ["normalization"], "Method"], [115, 117, ["Batch", "Normalization"], "Method"], [123, 125, ["learning", "rates"], "Metric"], [130, 131, ["initialization"], "Task"], [137, 138, ["regularizer"], "Method"], [158, 161, ["image", "classification", "model"], "Method"], [162, 164, ["Batch", "Normalization"], "Method"], [167, 168, ["accuracy"], "Metric"], [187, 193, ["ensemble", "of", "batch", "-", "normalized", "networks"], "Method"], [202, 204, ["ImageNet", "classification"], "Task"], [208, 213, ["top", "-", "5", "validation", "error"], "Metric"], [217, 219, ["test", "error"], "Metric"], [223, 224, ["accuracy"], "Metric"], [225, 227, ["human", "raters"], "Metric"], [231, 233, ["Deep", "learning"], "Method"], [242, 243, ["vision"], "Task"], [244, 245, ["speech"], "Task"], [251, 254, ["Stochastic", "gradient", "descent"], "Method"], [255, 256, ["SGD"], "Method"], [266, 268, ["deep", "networks"], "Method"], [274, 275, ["momentum"], "Method"], [276, 277, ["Adagrad"], "Method"], [288, 289, ["SGD"], "Method"], [310, 311, ["SGD"], "Method"], [432, 434, ["computing", "platforms"], "Method"], [436, 438, ["stochastic", "gradient"], "Method"], [456, 458, ["learning", "rate"], "Metric"], [460, 461, ["optimization"], "Task"], [474, 475, ["training"], "Task"], [541, 543, ["learning", "system"], "Method"], [558, 560, ["domain", "adaptation"], "Method"], [573, 575, ["learning", "system"], "Method"], [588, 591, ["sub", "-", "network"], "Method"], [619, 620, ["Learning"], "Task"], [631, 634, ["sub", "-", "network"], "Method"], [638, 641, ["gradient", "descent", "step"], "Method"], [646, 648, ["learning", "rate"], "Metric"], [656, 660, ["stand", "-", "alone", "network"], "Method"], [692, 695, ["sub", "-", "network"], "Method"], [858, 859, ["convergence"], "Task"], [874, 876, ["saturation", "problem"], "Task"], [886, 889, ["Rectified", "Linear", "Units"], "Method"], [895, 897, ["learning", "rates"], "Metric"], [921, 922, ["optimizer"], "Method"], [953, 955, ["deep", "network"], "Method"], [974, 975, ["training"], "Task"], [985, 987, ["Batch", "Normalization"], "Method"], [994, 997, ["internal", "covariate", "shift"], "Task"], [1007, 1010, ["deep", "neural", "nets"], "Method"], [1016, 1018, ["normalization", "step"], "Method"], [1028, 1030, ["Batch", "Normalization"], "Method"], [1068, 1070, ["learning", "rates"], "Metric"], [1078, 1080, ["batch", "normalization"], "Method"], [1092, 1094, ["Batch", "Normalization"], "Method"], [1122, 1124, ["Batch", "Normalization"], "Method"], [1155, 1156, ["accuracy"], "Metric"], [1169, 1171, ["Batch", "Normalization"], "Method"], [1175, 1180, ["top", "-", "5", "error", "rate"], "Metric"], [1188, 1190, ["ImageNet", "classification"], "Task"], [1195, 1198, ["Internal", "Covariate", "Shift"], "Task"], [1225, 1226, ["training"], "Task"], [1254, 1256, ["training", "speed"], "Metric"], [1264, 1266, ["network", "training"], "Method"], [1376, 1378, ["optimization", "algorithm"], "Method"], [1395, 1397, ["optimization", "steps"], "Method"], [1400, 1403, ["gradient", "descent", "step"], "Method"], [1478, 1481, ["gradient", "descent", "step"], "Method"], [1508, 1509, ["normalization"], "Task"], [1547, 1548, ["normalization"], "Method"], [1579, 1582, ["gradient", "descent", "step"], "Method"], [1592, 1595, ["gradient", "descent", "optimization"], "Method"], [1691, 1692, ["normalization"], "Task"], [1726, 1727, ["backpropagation"], "Task"], [1786, 1787, ["backpropagation"], "Method"], [1797, 1799, ["input", "normalization"], "Task"], [1844, 1846, ["image", "networks"], "Task"], [1905, 1906, ["Normalization"], "Task"], [1907, 1911, ["Mini", "-", "Batch", "Statistics"], "Method"], [2002, 2003, ["normalization"], "Method"], [2005, 2006, ["convergence"], "Task"], [2040, 2041, ["sigmoid"], "Method"], [2069, 2071, ["identity", "transform"], "Method"], [2140, 2142, ["batch", "setting"], "Task"], [2171, 2173, ["stochastic", "optimization"], "Method"], [2189, 2192, ["stochastic", "gradient", "training"], "Method"], [2215, 2216, ["normalization"], "Task"], [2221, 2223, ["gradient", "backpropagation"], "Method"], [2251, 2252, ["regularization"], "Task"], [2290, 2291, ["normalization"], "Method"], [2330, 2332, ["linear", "transformations"], "Method"], [2341, 2344, ["Batch", "Normalizing", "Transform"], "Method"], [2371, 2373, ["numerical", "stability"], "Metric"], [2374, 2377, ["Batch", "Normalizing", "Transform"], "Method"], [2464, 2466, ["network", "layers"], "Method"], [2544, 2547, ["sub", "-", "network"], "Method"], [2550, 2552, ["linear", "transform"], "Method"], [2605, 2608, ["sub", "-", "network"], "Method"], [2619, 2620, ["training"], "Task"], [2666, 2668, ["differentiable", "transformation"], "Method"], [2708, 2710, ["affine", "transform"], "Method"], [2734, 2735, ["Inference"], "Task"], [2736, 2740, ["Batch", "-", "Normalized", "Networks"], "Method"], [2786, 2788, ["Batch", "Normalization"], "Task"], [2792, 2795, ["batch", "gradient", "descent"], "Method"], [2797, 2800, ["Stochastic", "Gradient", "Descent"], "Method"], [2815, 2816, ["Adagrad"], "Method"], [2830, 2831, ["training"], "Task"], [2839, 2840, ["inference"], "Task"], [2867, 2868, ["normalization"], "Method"], [2900, 2903, ["unbiased", "variance", "estimate"], "Method"], [2922, 2924, ["moving", "averages"], "Method"], [2930, 2931, ["accuracy"], "Metric"], [2946, 2947, ["inference"], "Task"], [2949, 2950, ["normalization"], "Method"], [2953, 2955, ["linear", "transform"], "Method"], [2977, 2979, ["linear", "transform"], "Method"], [2991, 2995, ["batch", "-", "normalized", "networks"], "Method"], [2998, 3002, ["Batch", "-", "Normalized", "Network"], "Method"], [3005, 3009, ["Batch", "-", "normalized", "network"], "Method"], [3010, 3011, ["inference"], "Task"], [3016, 3017, ["transformation"], "Task"], [3069, 3074, ["Batch", "-", "Normalized", "Convolutional", "Networks"], "Method"], [3074, 3076, ["Batch", "Normalization"], "Method"], [3123, 3124, ["sigmoid"], "Method"], [3125, 3126, ["ReLU"], "Method"], [3131, 3137, ["fully", "-", "connected", "and", "convolutional", "layers"], "Method"], [3252, 3254, ["mean", "subtraction"], "Method"], [3439, 3440, ["inference"], "Task"], [3446, 3448, ["linear", "transformation"], "Method"], [3459, 3461, ["Batch", "Normalization"], "Method"], [3463, 3465, ["learning", "rates"], "Metric"], [3467, 3469, ["deep", "networks"], "Method"], [3473, 3475, ["learning", "rate"], "Metric"], [3495, 3497, ["Batch", "Normalization"], "Method"], [3545, 3547, ["Batch", "Normalization"], "Method"], [3549, 3550, ["training"], "Task"], [3560, 3562, ["learning", "rates"], "Metric"], [3581, 3583, ["model", "explosion"], "Task"], [3587, 3589, ["Batch", "Normalization"], "Method"], [3590, 3591, ["backpropagation"], "Method"], [3627, 3629, ["gradient", "propagation"], "Method"], [3640, 3642, ["Batch", "Normalization"], "Method"], [3652, 3654, ["Batch", "Normalization"], "Method"], [3674, 3675, ["training"], "Task"], [3707, 3709, ["linear", "transformation"], "Method"], [3772, 3774, ["Batch", "Normalization"], "Method"], [3777, 3779, ["gradient", "propagation"], "Method"], [3786, 3788, ["Batch", "Normalization"], "Method"], [3789, 3791, ["gradient", "propagation"], "Task"], [3800, 3802, ["Batch", "Normalization"], "Method"], [3808, 3810, ["Batch", "Normalization"], "Method"], [3829, 3831, ["training", "network"], "Method"], [3861, 3862, ["Dropout"], "Method"], [3867, 3868, ["overfitting"], "Task"], [3871, 3875, ["batch", "-", "normalized", "network"], "Method"], [3905, 3906, ["training"], "Task"], [3911, 3913, ["Batch", "Normalization"], "Method"], [3922, 3926, ["predicting", "the", "digit", "class"], "Task"], [3948, 3953, ["fully", "-", "connected", "hidden", "layers"], "Method"], [3959, 3961, ["hidden", "layer"], "Method"], [3984, 3988, ["fully", "-", "connected", "layer"], "Method"], [3997, 4001, ["cross", "-", "entropy", "loss"], "Metric"], [4046, 4052, ["baseline", "and", "batch", "-", "normalized", "networks"], "Method"], [4063, 4064, ["MNIST"], "Task"], [4102, 4106, ["batch", "-", "normalized", "network"], "Method"], [4109, 4111, ["test", "accuracy"], "Metric"], [4126, 4128, ["network", "N"], "Method"], [4129, 4133, ["batch", "-", "normalized", "network"], "Method"], [4213, 4217, ["batch", "-", "normalized", "network"], "Method"], [4232, 4234, ["ImageNet", "classification"], "Task"], [4236, 4238, ["Batch", "Normalization"], "Method"], [4244, 4246, ["Inception", "network"], "Method"], [4261, 4265, ["convolutional", "and", "pooling", "layers"], "Method"], [4268, 4270, ["softmax", "layer"], "Method"], [4281, 4283, ["Convolutional", "layers"], "Method"], [4284, 4285, ["ReLU"], "Method"], [4300, 4302, ["convolutional", "layers"], "Method"], [4309, 4310, ["convolutions"], "Method"], [4326, 4328, ["softmax", "layer"], "Method"], [4331, 4335, ["fully", "-", "connected", "layers"], "Method"], [4350, 4351, ["Inception"], "Method"], [4366, 4369, ["Stochastic", "Gradient", "Descent"], "Method"], [4370, 4371, ["momentum"], "Method"], [4389, 4393, ["scale", ",", "distributed", "architecture"], "Method"], [4408, 4410, ["validation", "accuracy"], "Metric"], [4447, 4448, ["Inception"], "Method"], [4449, 4451, ["Batch", "Normalization"], "Method"], [4456, 4458, ["Batch", "Normalization"], "Method"], [4496, 4498, ["Batch", "Normalization"], "Method"], [4528, 4530, ["learning", "rate"], "Metric"], [4533, 4537, ["batch", "-", "normalized", "model"], "Method"], [4549, 4551, ["learning", "rates"], "Metric"], [4577, 4579, ["Batch", "Normalization"], "Method"], [4586, 4587, ["Dropout"], "Task"], [4589, 4590, ["Dropout"], "Method"], [4592, 4593, ["BN"], "Method"], [4594, 4595, ["Inception"], "Method"], [4597, 4598, ["training"], "Task"], [4605, 4608, ["L2", "weight", "regularization"], "Method"], [4611, 4612, ["Inception"], "Method"], [4623, 4624, ["BN"], "Method"], [4625, 4626, ["Inception"], "Method"], [4645, 4646, ["accuracy"], "Metric"], [4656, 4659, ["learning", "rate", "decay"], "Metric"], [4662, 4663, ["Inception"], "Method"], [4664, 4666, ["learning", "rate"], "Metric"], [4676, 4677, ["Inception"], "Method"], [4681, 4683, ["learning", "rate"], "Metric"], [4688, 4691, ["Local", "Response", "Normalization"], "Method"], [4692, 4693, ["Inception"], "Method"], [4704, 4706, ["Batch", "Normalization"], "Method"], [4752, 4754, ["validation", "accuracy"], "Metric"], [4762, 4764, ["Batch", "Normalization"], "Method"], [4766, 4767, ["regularizer"], "Method"], [4803, 4807, ["batch", "-", "normalized", "networks"], "Method"], [4835, 4839, ["Single", "-", "Network", "Classification"], "Method"], [4860, 4861, ["Inception"], "Method"], [4878, 4880, ["learning", "rate"], "Metric"], [4883, 4884, ["BN"], "Method"], [4889, 4890, ["Inception"], "Method"], [4891, 4893, ["Batch", "Normalization"], "Method"], [4901, 4902, ["Inception"], "Method"], [4903, 4905, ["Batch", "Normalization"], "Method"], [4917, 4919, ["learning", "rate"], "Metric"], [4937, 4938, ["Inception"], "Method"], [4947, 4948, ["BN"], "Method"], [4960, 4962, ["learning", "rate"], "Metric"], [4968, 4969, ["Inception"], "Method"], [4988, 4989, ["ReLU"], "Method"], [4997, 4998, ["Inception"], "Method"], [4999, 5000, ["sigmoid"], "Method"], [5007, 5008, ["accuracy"], "Metric"], [5021, 5023, ["validation", "accuracy"], "Metric"], [5037, 5038, ["Inception"], "Method"], [5040, 5041, ["accuracy"], "Metric"], [5071, 5072, ["accuracy"], "Metric"], [5077, 5080, ["maximum", "validation", "accuracy"], "Metric"], [5096, 5098, ["Batch", "Normalization"], "Method"], [5107, 5108, ["accuracy"], "Metric"], [5109, 5110, ["Inception"], "Method"], [5135, 5137, ["training", "speed"], "Metric"], [5150, 5151, ["Inception"], "Method"], [5156, 5157, ["accuracy"], "Metric"], [5162, 5164, ["learning", "rate"], "Metric"], [5166, 5167, ["BN"], "Method"], [5187, 5188, ["accuracy"], "Metric"], [5204, 5205, ["Inception"], "Method"], [5221, 5223, ["deep", "networks"], "Method"], [5224, 5226, ["Batch", "Normalization"], "Method"], [5257, 5258, ["accuracy"], "Metric"], [5263, 5265, ["Batch", "Normalization"], "Method"], [5266, 5267, ["Inception"], "Method"], [5268, 5269, ["sigmoid"], "Method"], [5273, 5274, ["accuracy"], "Metric"], [5277, 5279, ["Ensemble", "Classification"], "Method"], [5286, 5292, ["ImageNet", "Large", "Scale", "Visual", "Recognition", "Competition"], "Task"], [5296, 5302, ["Deep", "Image", "ensemble", "of", "traditional", "models"], "Method"], [5304, 5306, ["ensemble", "model"], "Method"], [5312, 5316, ["top", "-", "5", "error"], "Metric"], [5324, 5326, ["ILSVRC", "server"], "Method"], [5331, 5336, ["top", "-", "5", "validation", "error"], "Metric"], [5341, 5343, ["test", "error"], "Metric"], [5350, 5352, ["ILSVRC", "server"], "Method"], [5366, 5367, ["accuracy"], "Metric"], [5368, 5370, ["human", "raters"], "Metric"], [5406, 5407, ["Dropout"], "Method"], [5425, 5426, ["Inception"], "Method"], [5430, 5439, ["non", "-", "convolutional", ",", "per", "-", "activation", "Batch", "Normalization"], "Method"], [5452, 5453, ["accuracy"], "Metric"], [5459, 5461, ["ensemble", "prediction"], "Task"], [5473, 5475, ["constituent", "networks"], "Method"], [5479, 5483, ["ensemble", "and", "multicrop", "inference"], "Method"], [5496, 5498, ["batch", "normalization"], "Method"], [5516, 5520, ["ImageNet", "classification", "challenge", "benchmarks"], "Material"], [5534, 5538, ["training", "of", "deep", "networks"], "Task"], [5555, 5560, ["training", "of", "machine", "learning", "systems"], "Task"], [5582, 5583, ["training"], "Task"], [5598, 5599, ["normalization"], "Method"], [5601, 5603, ["network", "architecture"], "Method"], [5609, 5610, ["normalization"], "Task"], [5615, 5617, ["optimization", "method"], "Method"], [5628, 5631, ["stochastic", "optimization", "methods"], "Method"], [5634, 5637, ["deep", "network", "training"], "Task"], [5641, 5642, ["normalization"], "Method"], [5657, 5659, ["Batch", "Normalization"], "Method"], [5686, 5687, ["training"], "Task"], [5690, 5691, ["inference"], "Task"], [5692, 5696, ["batch", "-", "normalized", "networks"], "Method"], [5712, 5714, ["training", "rates"], "Metric"], [5720, 5721, ["Dropout"], "Method"], [5722, 5723, ["regularization"], "Task"], [5726, 5728, ["Batch", "Normalization"], "Method"], [5737, 5740, ["image", "classification", "model"], "Method"], [5745, 5746, ["training"], "Task"], [5751, 5753, ["learning", "rates"], "Metric"], [5755, 5756, ["Dropout"], "Method"], [5763, 5765, ["Batch", "Normalization"], "Method"], [5792, 5797, ["single", "-", "network", "image", "classification"], "Task"], [5806, 5808, ["Batch", "Normalization"], "Method"], [5818, 5819, ["ImageNet"], "Material"], [5833, 5835, ["standardization", "layer"], "Method"], [5855, 5857, ["Batch", "Normalization"], "Task"], [5905, 5907, ["standardization", "layer"], "Method"], [5922, 5927, ["large", "-", "scale", "image", "classification"], "Task"], [5944, 5946, ["Batch", "Normalization"], "Method"], [5952, 5954, ["Batch", "Normalization"], "Method"], [5970, 5972, ["standardization", "layer"], "Method"], [5983, 5985, ["linear", "transform"], "Method"], [5999, 6001, ["convolutional", "layers"], "Method"], [6002, 6004, ["deterministic", "inference"], "Method"], [6015, 6021, ["batch", "-", "normalizing", "each", "convolutional", "layer"], "Method"], [6039, 6041, ["Batch", "Normalization"], "Method"], [6053, 6056, ["Recurrent", "Neural", "Networks"], "Task"], [6085, 6086, ["normalization"], "Method"], [6087, 6089, ["gradient", "propagation"], "Task"], [6102, 6104, ["Batch", "Normalization"], "Method"], [6107, 6109, ["domain", "adaptation"], "Task"], [6118, 6119, ["normalization"], "Method"], [6225, 6227, ["GoogLeNet", "model"], "Method"], [6232, 6234, ["convolutional", "layers"], "Method"], [6239, 6243, ["3", "3", "convolutional", "layers"], "Method"], [6269, 6271, ["computational", "cost"], "Metric"], [6282, 6283, ["inception"], "Method"], [6299, 6302, ["maximum", "-", "pooling"], "Method"], [6335, 6342, ["stride", "-", "2", "convolution", "/", "pooling", "layers"], "Method"], [6346, 6348, ["filter", "concatenation"], "Method"], [6358, 6360, ["separable", "convolution"], "Method"], [6361, 6363, ["depth", "multiplier"], "Method"], [6366, 6368, ["convolutional", "layer"], "Method"], [6372, 6374, ["computational", "cost"], "Metric"], [6377, 6379, ["memory", "consumption"], "Metric"], [6380, 6382, ["training", "time"], "Metric"], [270, 271, ["SGD"], "Method"], [1129, 1131, ["ImageNet", "classification"], "Task"], [2348, 2349, ["BN"], "Method"], [2389, 2390, ["BN"], "Method"], [2424, 2425, ["BN"], "Method"], [2645, 2646, ["BN"], "Method"], [2662, 2663, ["BN"], "Method"], [2717, 2718, ["BN"], "Method"], [2756, 2757, ["BN"], "Method"], [3013, 3014, ["BN"], "Method"], [3141, 3142, ["BN"], "Method"], [3278, 3279, ["BN"], "Method"], [3441, 3442, ["BN"], "Method"], [4250, 4252, ["ImageNet", "classification"], "Task"], [4492, 4493, ["BN"], "Method"], [4897, 4898, ["BN"], "Method"], [4952, 4953, ["BN"], "Method"], [4971, 4972, ["BN"], "Method"], [4978, 4979, ["BN"], "Method"], [5099, 5100, ["BN"], "Method"], [5141, 5142, ["BN"], "Method"], [5250, 5251, ["BN"], "Method"], [5386, 5387, ["BN"], "Method"], [5963, 5964, ["BN"], "Method"], [6184, 6185, ["Inception"], "Method"], [6331, 6332, ["Inception"], "Method"]]} {"doc_id": "188123b5afe0d3b59e9e44f33729649536fe6ea6", "text": ["document", ":", "Unifying", "Count", "-", "Based", "Exploration", "and", "Intrinsic", "Motivation", "We", "consider", "an", "agent", "’s", "uncertainty", "about", "its", "environment", "and", "the", "problem", "of", "generalizing", "this", "uncertainty", "across", "states", ".", "Specifically", ",", "we", "focus", "on", "the", "problem", "of", "exploration", "in", "non", "-", "tabular", "reinforcement", "learning", ".", "Drawing", "inspiration", "from", "the", "intrinsic", "motivation", "literature", ",", "we", "use", "density", "models", "to", "measure", "uncertainty", ",", "and", "propose", "a", "novel", "algorithm", "for", "deriving", "a", "pseudo", "-", "count", "from", "an", "arbitrary", "density", "model", ".", "This", "technique", "enables", "us", "to", "generalize", "count", "-", "based", "exploration", "algorithms", "to", "the", "non", "-", "tabular", "case", ".", "We", "apply", "our", "ideas", "to", "Atari", "2600", "games", ",", "providing", "sensible", "pseudo", "-", "counts", "from", "raw", "pixels", ".", "We", "transform", "these", "pseudo", "-", "counts", "into", "exploration", "bonuses", "and", "obtain", "significantly", "improved", "exploration", "in", "a", "number", "of", "hard", "games", ",", "including", "the", "infamously", "difficult", "Montezuma", "’s", "Revenge", ".", "section", ":", "Introduction", "Exploration", "algorithms", "for", "Markov", "Decision", "Processes", "(", "MDPs", ")", "are", "typically", "concerned", "with", "reducing", "the", "agent", "’s", "uncertainty", "over", "the", "environment", "’s", "reward", "and", "transition", "functions", ".", "In", "a", "tabular", "setting", ",", "this", "uncertainty", "can", "be", "quantified", "using", "confidence", "intervals", "derived", "from", "Chernoff", "bounds", ",", "or", "inferred", "from", "a", "posterior", "over", "the", "environment", "parameters", ".", "In", "fact", ",", "both", "confidence", "intervals", "and", "posterior", "shrink", "as", "the", "inverse", "square", "root", "of", "the", "state", "-", "action", "visit", "count", ",", "making", "this", "quantity", "fundamental", "to", "most", "theoretical", "results", "on", "exploration", ".", "Count", "-", "based", "exploration", "methods", "directly", "use", "visit", "counts", "to", "guide", "an", "agent", "’s", "behaviour", "towards", "reducing", "uncertainty", ".", "For", "example", ",", "Model", "-", "based", "Interval", "Estimation", "with", "Exploration", "Bonuses", "[", "MBIE", "-", "EB;", "][]", "strehl08analysis", "solves", "the", "augmented", "Bellman", "equation", "involving", "the", "empirical", "reward", ",", "the", "empirical", "transition", "function", ",", "and", "an", "exploration", "bonus", "proportional", "to", ".", "This", "bonus", "accounts", "for", "uncertainties", "in", "both", "transition", "and", "reward", "functions", "and", "enables", "a", "finite", "-", "time", "bound", "on", "the", "agent", "’s", "suboptimality", ".", "In", "spite", "of", "their", "pleasant", "theoretical", "guarantees", ",", "count", "-", "based", "methods", "have", "not", "played", "a", "role", "in", "the", "contemporary", "successes", "of", "reinforcement", "learning", "[", "e.g.", "][]", "mnih15human", ".", "Instead", ",", "most", "practical", "methods", "still", "rely", "on", "simple", "rules", "such", "as", "-", "greedy", ".", "The", "issue", "is", "that", "visit", "counts", "are", "not", "directly", "useful", "in", "large", "domains", ",", "where", "states", "are", "rarely", "visited", "more", "than", "once", ".", "Answering", "a", "different", "scientific", "question", ",", "intrinsic", "motivation", "aims", "to", "provide", "qualitative", "guidance", "for", "exploration", "schmidhuber91possibility", ",", "oudeyer07intrinsic", ",", "barto13intrinsic", ".", "This", "guidance", "can", "be", "summarized", "as", "“", "explore", "what", "surprises", "you", "”", ".", "A", "typical", "approach", "guides", "the", "agent", "based", "on", "change", "in", "prediction", "error", ",", "or", "learning", "progress", ".", "If", "is", "the", "error", "made", "by", "the", "agent", "at", "time", "over", "some", "event", "A", ",", "and", "the", "same", "error", "after", "observing", "a", "new", "piece", "of", "information", ",", "then", "learning", "progress", "is", "Intrinsic", "motivation", "methods", "are", "attractive", "as", "they", "remain", "applicable", "in", "the", "absence", "of", "the", "Markov", "property", "or", "the", "lack", "of", "a", "tabular", "representation", ",", "both", "of", "which", "are", "required", "by", "count", "-", "based", "algorithms", ".", "Yet", "the", "theoretical", "foundations", "of", "intrinsic", "motivation", "remain", "largely", "absent", "from", "the", "literature", ",", "which", "may", "explain", "its", "slow", "rate", "of", "adoption", "as", "a", "standard", "approach", "to", "exploration", ".", "In", "this", "paper", "we", "provide", "formal", "evidence", "that", "intrinsic", "motivation", "and", "count", "-", "based", "exploration", "are", "but", "two", "sides", "of", "the", "same", "coin", ".", "Specifically", ",", "we", "consider", "a", "frequently", "used", "measure", "of", "learning", "progress", ",", "information", "gain", "cover91elements", ".", "Defined", "as", "the", "Kullback", "-", "Leibler", "divergence", "of", "a", "prior", "distribution", "from", "its", "posterior", ",", "information", "gain", "can", "be", "related", "to", "the", "confidence", "intervals", "used", "in", "count", "-", "based", "exploration", ".", "Our", "contribution", "is", "to", "propose", "a", "new", "quantity", ",", "the", "pseudo", "-", "count", ",", "which", "connects", "information", "-", "gain", "-", "as", "-", "learning", "-", "progress", "and", "count", "-", "based", "exploration", ".", "We", "derive", "our", "pseudo", "-", "count", "from", "a", "density", "model", "over", "the", "state", "space", ".", "This", "is", "in", "departure", "from", "more", "traditional", "approaches", "to", "intrinsic", "motivation", "that", "consider", "learning", "progress", "with", "respect", "to", "a", "transition", "model", ".", "We", "expose", "the", "relationship", "between", "pseudo", "-", "counts", ",", "a", "variant", "of", "schmidhuber91possibility", "’s", "compression", "progress", "we", "call", "prediction", "gain", ",", "and", "information", "gain", ".", "Combined", "to", "kolter09near", "’s", "negative", "result", "on", "the", "frequentist", "suboptimality", "of", "Bayesian", "bonuses", ",", "our", "result", "highlights", "the", "theoretical", "advantages", "of", "pseudo", "-", "counts", "compared", "to", "many", "existing", "intrinsic", "motivation", "methods", ".", "The", "pseudo", "-", "counts", "we", "introduce", "here", "are", "best", "thought", "of", "as", "“", "function", "approximation", "for", "exploration", "”", ".", "We", "bring", "them", "to", "bear", "on", "Atari", "2600", "games", "from", "the", "Arcade", "Learning", "Environment", "bellemare13arcade", ",", "focusing", "on", "games", "where", "myopic", "exploration", "fails", ".", "We", "extract", "our", "pseudo", "-", "counts", "from", "a", "simple", "density", "model", "and", "use", "them", "within", "a", "variant", "of", "MBIE", "-", "EB", ".", "We", "apply", "them", "to", "an", "experience", "replay", "setting", "and", "to", "an", "actor", "-", "critic", "setting", ",", "and", "find", "improved", "performance", "in", "both", "cases", ".", "Our", "approach", "produces", "dramatic", "progress", "on", "the", "reputedly", "most", "difficult", "Atari", "2600", "game", ",", "Montezuma", "’s", "Revenge", ":", "within", "a", "fraction", "of", "the", "training", "time", ",", "our", "agent", "explores", "a", "significant", "portion", "of", "the", "first", "level", "and", "obtains", "significantly", "higher", "scores", "than", "previously", "published", "agents", ".", "section", ":", "Notation", "We", "consider", "a", "countable", "state", "space", ".", "We", "denote", "a", "sequence", "of", "length", "from", "by", ",", "the", "set", "of", "finite", "sequences", "from", "by", ",", "write", "to", "mean", "the", "concatenation", "of", "and", "a", "state", ",", "and", "denote", "the", "empty", "sequence", "by", ".", "A", "model", "over", "is", "a", "mapping", "from", "to", "probability", "distributions", "over", ".", "That", "is", ",", "for", "each", "the", "model", "provides", "a", "probability", "distribution", "Note", "that", "we", "do", "not", "require", "to", "be", "strictly", "positive", "for", "all", "and", ".", "When", "it", "is", ",", "however", ",", "we", "may", "understand", "to", "be", "the", "usual", "conditional", "probability", "of", "given", ".", "We", "will", "take", "particular", "interest", "in", "the", "empirical", "distribution", "derived", "from", "the", "sequence", ".", "If", "is", "the", "number", "of", "occurrences", "of", "a", "state", "in", "the", "sequence", ",", "then", "We", "call", "the", "the", "empirical", "count", "function", ",", "or", "simply", "empirical", "count", ".", "The", "above", "notation", "extends", "to", "state", "-", "action", "spaces", ",", "and", "we", "write", "to", "explicitly", "refer", "to", "the", "number", "of", "occurrences", "of", "a", "state", "-", "action", "pair", "when", "the", "argument", "requires", "it", ".", "When", "is", "generated", "by", "an", "ergodic", "Markov", "chain", ",", "for", "example", "if", "we", "follow", "a", "fixed", "policy", "in", "a", "finite", "-", "state", "MDP", ",", "then", "the", "limit", "point", "of", "is", "the", "chain", "’s", "stationary", "distribution", ".", "In", "our", "setting", ",", "a", "density", "model", "is", "any", "model", "that", "assumes", "states", "are", "independently", "(", "but", "not", "necessarily", "identically", ")", "distributed", ";", "a", "density", "model", "is", "thus", "a", "particular", "kind", "of", "generative", "model", ".", "We", "emphasize", "that", "a", "density", "model", "differs", "from", "a", "forward", "model", ",", "which", "takes", "into", "account", "the", "temporal", "relationship", "between", "successive", "states", ".", "Note", "that", "is", "itself", "a", "density", "model", ".", "section", ":", "From", "Densities", "to", "Counts", "In", "the", "introduction", "we", "argued", "that", "the", "visit", "count", "(", "and", "consequently", ",", ")", "is", "not", "directly", "useful", "in", "practical", "settings", ",", "since", "states", "are", "rarely", "revisited", ".", "Specifically", ",", "is", "almost", "always", "zero", "and", "can", "not", "help", "answer", "the", "question", "“", "How", "novel", "is", "this", "state", "?", "”", "Nor", "is", "the", "problem", "solved", "by", "a", "Bayesian", "approach", ":", "even", "variable", "-", "alphabet", "models", "[", "e.g.", "][]", "hutter13sparse", "must", "assign", "a", "small", ",", "diminishing", "probability", "to", "yet", "-", "unseen", "states", ".", "To", "estimate", "the", "uncertainty", "of", "an", "agent", "’s", "knowledge", ",", "we", "must", "instead", "look", "for", "a", "quantity", "which", "generalizes", "across", "states", ".", "Guided", "by", "ideas", "from", "the", "intrinsic", "motivation", "literature", ",", "we", "now", "derive", "such", "a", "quantity", ".", "We", "call", "it", "a", "pseudo", "-", "count", "as", "it", "extends", "the", "familiar", "notion", "from", "Bayesian", "estimation", ".", "subsection", ":", "Pseudo", "-", "Counts", "and", "the", "Recoding", "Probability", "We", "are", "given", "a", "density", "model", "over", ".", "This", "density", "model", "may", "be", "approximate", ",", "biased", ",", "or", "even", "inconsistent", ".", "We", "begin", "by", "introducing", "the", "recoding", "probability", "of", "a", "state", ":", "This", "is", "the", "probability", "assigned", "to", "by", "our", "density", "model", "after", "observing", "a", "new", "occurrence", "of", ".", "The", "term", "“", "recoding", "”", "is", "inspired", "from", "the", "statistical", "compression", "literature", ",", "where", "coding", "costs", "are", "inversely", "related", "to", "probabilities", "cover91elements", ".", "When", "admits", "a", "conditional", "probability", "distribution", ",", "We", "now", "postulate", "two", "unknowns", ":", "a", "pseudo", "-", "count", "function", ",", "and", "a", "pseudo", "-", "count", "total", ".", "We", "relate", "these", "two", "unknowns", "through", "two", "constraints", ":", "In", "words", ":", "we", "require", "that", ",", "after", "observing", "one", "instance", "of", ",", "the", "density", "model", "’s", "increase", "in", "prediction", "of", "that", "same", "should", "correspond", "to", "a", "unit", "increase", "in", "pseudo", "-", "count", ".", "The", "pseudo", "-", "count", "itself", "is", "derived", "from", "solving", "the", "linear", "system", "(", "[", "reference", "]", ")", ":", "Note", "that", "the", "equations", "(", "[", "reference", "]", ")", "yield", "(", "with", ")", "when", ",", "and", "are", "inconsistent", "when", ".", "These", "cases", "may", "arise", "from", "poorly", "behaved", "density", "models", ",", "but", "are", "easily", "accounted", "for", ".", "From", "here", "onwards", "we", "will", "assume", "a", "consistent", "system", "of", "equations", ".", "theorem", ":", "(", "Learning", "-", "positive", "density", "model", ")", ".", "A", "density", "model", "ρ", "is", "learning", "-", "positive", "if", "for", "all", "∈x:1nXn", "and", "all", "∈xX", ",", "≥⁢ρ′n", "(", "x", ")", "⁢ρn", "(", "x", ")", ".", "By", "inspecting", "(", "[", "reference", "]", ")", ",", "we", "see", "that", "if", "and", "only", "if", "is", "learning", "-", "positive", ";", "if", "and", "only", "if", ";", "and", "if", "and", "only", "if", ".", "In", "many", "cases", "of", "interest", ",", "the", "pseudo", "-", "count", "matches", "our", "intuition", ".", "If", "then", ".", "Similarly", ",", "if", "is", "a", "Dirichlet", "estimator", "then", "recovers", "the", "usual", "notion", "of", "pseudo", "-", "count", ".", "More", "importantly", ",", "if", "the", "model", "generalizes", "across", "states", "then", "so", "do", "pseudo", "-", "counts", ".", "subsection", ":", "Estimating", "the", "Frequency", "of", "a", "Salient", "Event", "in", "Freeway", "As", "an", "illustrative", "example", ",", "we", "employ", "our", "method", "to", "estimate", "the", "number", "of", "occurrences", "of", "an", "infrequent", "event", "in", "the", "Atari", "2600", "video", "game", "Freeway", "(", "Figure", "[", "reference", "]", ",", "screenshot", ")", ".", "We", "use", "the", "Arcade", "Learning", "Environment", "bellemare13arcade", ".", "We", "will", "demonstrate", "the", "following", ":", "Pseudo", "-", "counts", "are", "roughly", "zero", "for", "novel", "events", ",", "they", "exhibit", "credible", "magnitudes", ",", "they", "respect", "the", "ordering", "of", "state", "frequency", ",", "they", "grow", "linearly", "(", "on", "average", ")", "with", "real", "counts", ",", "they", "are", "robust", "in", "the", "presence", "of", "nonstationary", "data", ".", "These", "properties", "suggest", "that", "pseudo", "-", "counts", "provide", "an", "appropriate", "generalized", "notion", "of", "visit", "counts", "in", "non", "-", "tabular", "settings", ".", "In", "Freeway", ",", "the", "agent", "must", "navigate", "a", "chicken", "across", "a", "busy", "road", ".", "As", "our", "example", ",", "we", "consider", "estimating", "the", "number", "of", "times", "the", "chicken", "has", "reached", "the", "very", "top", "of", "the", "screen", ".", "As", "is", "the", "case", "for", "many", "Atari", "2600", "games", ",", "this", "naturally", "salient", "event", "is", "associated", "with", "an", "increase", "in", "score", ",", "which", "ALE", "translates", "into", "a", "positive", "reward", ".", "We", "may", "reasonably", "imagine", "that", "knowing", "how", "certain", "we", "are", "about", "this", "part", "of", "the", "environment", "is", "useful", ".", "After", "crossing", ",", "the", "chicken", "is", "teleported", "back", "to", "the", "bottom", "of", "the", "screen", ".", "To", "highlight", "the", "robustness", "of", "our", "pseudo", "-", "count", ",", "we", "consider", "a", "nonstationary", "policy", "which", "waits", "for", "250", ",", "000", "frames", ",", "then", "applies", "the", "up", "action", "for", "250", ",", "000", "frames", ",", "then", "waits", ",", "then", "goes", "up", "again", ".", "The", "salient", "event", "only", "occurs", "during", "up", "periods", ".", "It", "also", "occurs", "with", "the", "cars", "in", "different", "positions", ",", "thus", "requiring", "generalization", ".", "As", "a", "point", "of", "reference", ",", "we", "record", "the", "pseudo", "-", "counts", "for", "both", "the", "salient", "event", "and", "visits", "to", "the", "chicken", "’s", "start", "position", ".", "We", "use", "a", "simplified", ",", "pixel", "-", "level", "version", "of", "the", "CTS", "model", "for", "Atari", "2600", "frames", "proposed", "by", "bellemare14skip", ",", "ignoring", "temporal", "dependencies", ".", "While", "the", "CTS", "model", "is", "rather", "impoverished", "in", "comparison", "to", "state", "-", "of", "-", "the", "-", "art", "density", "models", "for", "images", "[", "e.g.", "][]", "vandenoord16pixel", ",", "its", "count", "-", "based", "nature", "results", "in", "extremely", "fast", "learning", ",", "making", "it", "an", "appealing", "candidate", "for", "exploration", ".", "Further", "details", "on", "the", "model", "may", "be", "found", "in", "the", "appendix", ".", "Examining", "the", "pseudo", "-", "counts", "depicted", "in", "Figure", "[", "reference", "]", "confirms", "that", "they", "exhibit", "the", "desirable", "properties", "listed", "above", ".", "In", "particular", ",", "the", "pseudo", "-", "count", "is", "almost", "zero", "on", "the", "first", "occurrence", "of", "the", "salient", "event", ";", "it", "increases", "slightly", "during", "the", "3rd", "period", ",", "since", "the", "salient", "and", "reference", "events", "share", "some", "common", "structure", ";", "throughout", ",", "it", "remains", "smaller", "than", "the", "reference", "pseudo", "-", "count", ".", "The", "linearity", "on", "average", "and", "robustness", "to", "nonstationarity", "are", "immediate", "from", "the", "graph", ".", "Note", ",", "however", ",", "that", "the", "pseudo", "-", "counts", "are", "a", "fraction", "of", "the", "real", "visit", "counts", "(", "inasmuch", "as", "we", "can", "define", "“", "real", "”", ")", ":", "by", "the", "end", "of", "the", "trial", ",", "the", "start", "position", "has", "been", "visited", "about", "140", ",", "000", "times", ",", "and", "the", "topmost", "part", "of", "the", "screen", ",", "1285", "times", ".", "Furthermore", ",", "the", "ratio", "of", "recorded", "pseudo", "-", "counts", "differs", "from", "the", "ratio", "of", "real", "counts", ".", "Both", "effects", "are", "quantifiable", ",", "as", "we", "shall", "show", "in", "Section", "[", "reference", "]", ".", "section", ":", "The", "Connection", "to", "Intrinsic", "Motivation", "Having", "argued", "that", "pseudo", "-", "counts", "appropriately", "generalize", "visit", "counts", ",", "we", "will", "now", "show", "that", "they", "are", "closely", "related", "to", "information", "gain", ",", "which", "is", "commonly", "used", "to", "quantify", "novelty", "or", "curiosity", "and", "consequently", "as", "an", "intrinsic", "reward", ".", "Information", "gain", "is", "defined", "in", "relation", "to", "a", "mixture", "model", "over", "a", "class", "of", "density", "models", ".", "This", "model", "predicts", "according", "to", "a", "weighted", "combination", "from", ":", "with", "the", "posterior", "weight", "of", ".", "This", "posterior", "is", "defined", "recursively", ",", "starting", "from", "a", "prior", "distribution", "over", ":", "Information", "gain", "is", "then", "the", "Kullback", "-", "Leibler", "divergence", "from", "prior", "to", "posterior", "that", "results", "from", "observing", ":", "Computing", "the", "information", "gain", "of", "a", "complex", "density", "model", "is", "often", "impractical", ",", "if", "not", "downright", "intractable", ".", "However", ",", "a", "quantity", "which", "we", "call", "the", "prediction", "gain", "provides", "us", "with", "a", "good", "approximation", "of", "the", "information", "gain", ".", "We", "define", "the", "prediction", "gain", "of", "a", "density", "model", "(", "and", "in", "particular", ",", ")", "as", "the", "difference", "between", "the", "recoding", "log", "-", "probability", "and", "log", "-", "probability", "of", ":", "Prediction", "gain", "is", "nonnegative", "if", "and", "only", "if", "is", "learning", "-", "positive", ".", "It", "is", "related", "to", "the", "pseudo", "-", "count", ":", "with", "equality", "when", ".", "As", "the", "following", "theorem", "shows", ",", "prediction", "gain", "allows", "us", "to", "relate", "pseudo", "-", "count", "and", "information", "gain", ".", "theorem", ":", ".", "Consider", "a", "sequence", "∈x:1nXn", ".", "Let", "ξ", "be", "a", "mixture", "model", "over", "a", "class", "of", "learning", "-", "positive", "models", "M.", "Let", "^Nn", "be", "the", "pseudo", "-", "count", "derived", "from", "ξ", "(", "Equation", ")", ".", "For", "this", "model", ",", "Theorem", "1", "suggests", "that", "using", "an", "exploration", "bonus", "proportional", "to", ",", "similar", "to", "the", "MBIE", "-", "EB", "bonus", ",", "leads", "to", "a", "behaviour", "at", "least", "as", "exploratory", "as", "one", "derived", "from", "an", "information", "gain", "bonus", ".", "Since", "pseudo", "-", "counts", "correspond", "to", "empirical", "counts", "in", "the", "tabular", "setting", ",", "this", "approach", "also", "preserves", "known", "theoretical", "guarantees", ".", "In", "fact", ",", "we", "are", "confident", "pseudo", "-", "counts", "may", "be", "used", "to", "prove", "similar", "results", "in", "non", "-", "tabular", "settings", ".", "On", "the", "other", "hand", ",", "it", "may", "be", "difficult", "to", "provide", "theoretical", "guarantees", "about", "existing", "bonus", "-", "based", "intrinsic", "motivation", "approaches", ".", "kolter09near", "showed", "that", "no", "algorithm", "based", "on", "a", "bonus", "upper", "bounded", "by", "for", "any", "can", "guarantee", "PAC", "-", "MDP", "optimality", ".", "Again", "considering", "the", "tabular", "setting", "and", "combining", "their", "result", "to", "Theorem", "[", "reference", "]", ",", "we", "conclude", "that", "bonuses", "proportional", "to", "immediate", "information", "(", "or", "prediction", ")", "gain", "are", "insufficient", "for", "theoretically", "near", "-", "optimal", "exploration", ":", "to", "paraphrase", "kolter09near", ",", "these", "methods", "produce", "explore", "too", "little", "in", "comparison", "to", "pseudo", "-", "count", "bonuses", ".", "By", "inspecting", "(", "[", "reference", "]", ")", "we", "come", "to", "a", "similar", "negative", "conclusion", "for", "bonuses", "proportional", "to", "the", "L1", "or", "L2", "distance", "between", "and", ".", "Unlike", "many", "intrinsic", "motivation", "algorithms", ",", "pseudo", "-", "counts", "also", "do", "not", "rely", "on", "learning", "a", "forward", "(", "transition", "and", "/", "or", "reward", ")", "model", ".", "This", "point", "is", "especially", "important", "because", "a", "number", "of", "powerful", "density", "models", "for", "images", "exist", "vandenoord16pixel", ",", "and", "because", "optimality", "guarantees", "can", "not", "in", "general", "exist", "for", "intrinsic", "motivation", "algorithms", "based", "on", "forward", "models", ".", "section", ":", "Asymptotic", "Analysis", "In", "this", "section", "we", "analyze", "the", "limiting", "behaviour", "of", "the", "ratio", ".", "We", "use", "this", "analysis", "to", "assert", "the", "consistency", "of", "pseudo", "-", "counts", "derived", "from", "tabular", "density", "models", ",", "i.e.", "models", "which", "maintain", "per", "-", "state", "visit", "counts", ".", "In", "the", "appendix", "we", "use", "the", "same", "result", "to", "bound", "the", "approximation", "error", "of", "pseudo", "-", "counts", "derived", "from", "directed", "graphical", "models", ",", "of", "which", "our", "CTS", "model", "is", "a", "special", "case", ".", "Consider", "a", "fixed", ",", "infinite", "sequence", "from", ".", "We", "define", "the", "limit", "of", "a", "sequence", "of", "functions", "with", "respect", "to", "the", "length", "of", "the", "subsequence", ".", "We", "additionally", "assume", "that", "the", "empirical", "distribution", "converges", "pointwise", "to", "a", "distribution", ",", "and", "write", "for", "the", "recoding", "probability", "of", "under", ".", "We", "begin", "with", "two", "assumptions", "on", "our", "density", "model", ".", "theorem", ":", ".", "The", "limits", "exist", "for", "all", "x", ";", "furthermore", ",", ">", "⁢˙r", "(", "x", ")", "0", ".", "Assumption", "(", "a", ")", "states", "that", "should", "eventually", "assign", "a", "probability", "to", "proportional", "to", "the", "limiting", "empirical", "distribution", ".", "In", "particular", "there", "must", "be", "a", "state", "for", "which", ",", "unless", ".", "Assumption", "(", "b", ")", ",", "on", "the", "other", "hand", ",", "imposes", "a", "restriction", "on", "the", "learning", "rate", "of", "relative", "to", "’", "s.", "As", "both", "and", "exist", ",", "Assumption", "[", "reference", "]", "also", "implies", "that", "and", "have", "a", "common", "limit", ".", "theorem", ":", ".", "Under", "Assumption", ",", "the", "limit", "of", "the", "ratio", "of", "pseudo", "-", "counts", "⁢^Nn", "(", "x", ")", "to", "empirical", "counts", "⁢Nn", "(", "x", ")", "exists", "for", "all", "x.", "This", "limit", "is", "The", "model", "’s", "relative", "rate", "of", "change", ",", "whose", "convergence", "to", "we", "require", ",", "plays", "an", "essential", "role", "in", "the", "ratio", "of", "pseudo", "-", "to", "empirical", "counts", ".", "To", "see", "this", ",", "consider", "a", "sequence", "generated", "i.i.d", ".", "from", "a", "distribution", "over", "a", "finite", "state", "space", ",", "and", "a", "density", "model", "defined", "from", "a", "sequence", "of", "nonincreasing", "step", "-", "sizes", ":", "with", "initial", "condition", ".", "For", ",", "this", "density", "model", "is", "the", "empirical", "distribution", ".", "For", ",", "we", "may", "appeal", "to", "well", "-", "known", "results", "from", "stochastic", "approximation", "[", "e.g.", "][]", "bertsekas96neurodynamic", "and", "find", "that", "almost", "surely", "Since", ",", "we", "may", "think", "of", "Assumption", "1", "(", "b", ")", "as", "also", "requiring", "to", "converge", "at", "a", "rate", "of", "for", "a", "comparison", "with", "the", "empirical", "count", "to", "be", "meaningful", ".", "Note", ",", "however", ",", "that", "a", "density", "model", "that", "does", "not", "satisfy", "Assumption", "1", "(", "b", ")", "may", "still", "yield", "useful", "(", "but", "incommensurable", ")", "pseudo", "-", "counts", ".", "theorem", ":", ".", "Let", ">", "⁢ϕ", "(", "x", ")", "0", "with", "<", "∑∈xX⁢ϕ", "(", "x", ")", "∞", "and", "consider", "the", "count", "-", "based", "estimator", "If", "^Nn", "is", "the", "pseudo", "-", "count", "corresponding", "to", "ρn", "then", "→⁢", "/", "⁢^Nn", "(", "x", ")", "Nn", "(", "x", ")", "1", "for", "all", "x", "with", ">", "⁢μ", "(", "x", ")", "0", ".", "section", ":", "Empirical", "Evaluation", "In", "this", "section", "we", "demonstrate", "the", "use", "of", "pseudo", "-", "counts", "to", "guide", "exploration", ".", "We", "return", "to", "the", "Arcade", "Learning", "Environment", ",", "now", "using", "the", "CTS", "model", "to", "generate", "an", "exploration", "bonus", ".", "subsection", ":", "Exploration", "in", "Hard", "Atari", "2600", "Games", "From", "60", "games", "available", "through", "the", "Arcade", "Learning", "Environment", "we", "selected", "five", "“", "hard", "”", "games", ",", "in", "the", "sense", "that", "an", "-", "greedy", "policy", "is", "inefficient", "at", "exploring", "them", ".", "We", "used", "a", "bonus", "of", "the", "form", "where", "was", "selected", "from", "a", "coarse", "parameter", "sweep", ".", "We", "also", "compared", "our", "method", "to", "the", "optimistic", "initialization", "trick", "proposed", "by", "machado14domainindependent", ".", "We", "trained", "our", "agents", "’", "Q", "-", "functions", "with", "Double", "DQN", "vanhasselt16deep", ",", "with", "one", "important", "modification", ":", "we", "mixed", "the", "Double", "Q", "-", "Learning", "target", "with", "the", "Monte", "Carlo", "return", ".", "This", "modification", "led", "to", "improved", "results", "both", "with", "and", "without", "exploration", "bonuses", "(", "details", "in", "the", "appendix", ")", ".", "Figure", "[", "reference", "]", "depicts", "the", "result", "of", "our", "experiment", ",", "averaged", "across", "5", "trials", ".", "Although", "optimistic", "initialization", "helps", "in", "Freeway", ",", "it", "otherwise", "yields", "performance", "similar", "to", "DQN", ".", "By", "contrast", ",", "the", "count", "-", "based", "exploration", "bonus", "enables", "us", "to", "make", "quick", "progress", "on", "a", "number", "of", "games", ",", "most", "dramatically", "in", "Montezuma", "’s", "Revenge", "and", "Venture", ".", "Montezuma", "’s", "Revenge", "is", "perhaps", "the", "hardest", "Atari", "2600", "game", "available", "through", "the", "ALE", ".", "The", "game", "is", "infamous", "for", "its", "hostile", ",", "unforgiving", "environment", ":", "the", "agent", "must", "navigate", "a", "number", "of", "different", "rooms", ",", "each", "filled", "with", "traps", ".", "Due", "to", "its", "sparse", "reward", "function", ",", "most", "published", "agents", "achieve", "an", "average", "score", "close", "to", "zero", "and", "completely", "fail", "to", "explore", "most", "of", "the", "24", "rooms", "that", "constitute", "the", "first", "level", "(", "Figure", "[", "reference", "]", ",", "top", ")", ".", "By", "contrast", ",", "within", "50", "million", "frames", "our", "agent", "learns", "a", "policy", "which", "consistently", "navigates", "through", "15", "rooms", "(", "Figure", "[", "reference", "]", ",", "bottom", ")", ".", "Our", "agent", "also", "achieves", "a", "score", "higher", "than", "anything", "previously", "reported", ",", "with", "one", "run", "consistently", "achieving", "6600", "points", "by", "100", "million", "frames", "(", "half", "the", "training", "samples", "used", "by", "mnih15human", ")", ".", "We", "believe", "the", "success", "of", "our", "method", "in", "this", "game", "is", "a", "strong", "indicator", "of", "the", "usefulness", "of", "pseudo", "-", "counts", "for", "exploration", ".", "subsection", ":", "Exploration", "for", "Actor", "-", "Critic", "Methods", "We", "next", "used", "our", "exploration", "bonuses", "in", "conjunction", "with", "the", "A3C", "(", "Asynchronous", "Advantage", "Actor", "-", "Critic", ")", "algorithm", "of", "mnih16asynchronous", ".", "One", "appeal", "of", "actor", "-", "critic", "methods", "is", "their", "explicit", "separation", "of", "policy", "and", "Q", "-", "function", "parameters", ",", "which", "leads", "to", "a", "richer", "behaviour", "space", ".", "This", "very", "separation", ",", "however", ",", "often", "leads", "to", "deficient", "exploration", ":", "to", "produce", "any", "sensible", "results", ",", "the", "A3C", "policy", "must", "be", "regularized", "with", "an", "entropy", "cost", ".", "We", "trained", "A3C", "on", "60", "Atari", "2600", "games", ",", "with", "and", "without", "the", "exploration", "bonus", "(", "[", "reference", "]", ")", ".", "We", "refer", "to", "our", "augmented", "algorithm", "as", "A3C", "+", ".", "Full", "details", "and", "additional", "results", "may", "be", "found", "in", "the", "appendix", ".", "We", "found", "that", "A3C", "fails", "to", "learn", "in", "15", "games", ",", "in", "the", "sense", "that", "the", "agent", "does", "not", "achieve", "a", "score", "50", "%", "better", "than", "random", ".", "In", "comparison", ",", "there", "are", "only", "10", "games", "for", "which", "A3C", "+", "fails", "to", "improve", "on", "the", "random", "agent", ";", "of", "these", ",", "8", "are", "games", "where", "DQN", "fails", "in", "the", "same", "sense", ".", "We", "normalized", "the", "two", "algorithms", "’", "scores", "so", "that", "0", "and", "1", "are", "respectively", "the", "minimum", "and", "maximum", "of", "the", "random", "agent", "’s", "and", "A3C", "’s", "end", "-", "of", "-", "training", "score", "on", "a", "particular", "game", ".", "Figure", "[", "reference", "]", "depicts", "the", "in", "-", "training", "median", "score", "for", "A3C", "and", "A3C", "+", ",", "along", "with", "1st", "and", "3rd", "quartile", "intervals", ".", "Not", "only", "does", "A3C", "+", "achieve", "slightly", "superior", "median", "performance", ",", "but", "it", "also", "significantly", "outperforms", "A3C", "on", "at", "least", "a", "quarter", "of", "the", "games", ".", "This", "is", "particularly", "important", "given", "the", "large", "proportion", "of", "Atari", "2600", "games", "for", "which", "an", "-", "greedy", "policy", "is", "sufficient", "for", "exploration", ".", "section", ":", "Related", "Work", "Information", "-", "theoretic", "quantities", "have", "been", "repeatedly", "used", "to", "describe", "intrinsically", "motivated", "behaviour", ".", "Closely", "related", "to", "prediction", "gain", "is", "schmidhuber91possibility", "’s", "notion", "of", "compression", "progress", ",", "which", "equates", "novelty", "with", "an", "agent", "’s", "improvement", "in", "its", "ability", "to", "compress", "its", "past", ".", "More", "recently", ",", "lopes12exploration", "showed", "the", "relationship", "between", "time", "-", "averaged", "prediction", "gain", "and", "visit", "counts", "in", "a", "tabular", "setting", ";", "their", "result", "is", "a", "special", "case", "of", "Theorem", "[", "reference", "]", ".", "orseau13universal", "demonstrated", "that", "maximizing", "the", "sum", "of", "future", "information", "gains", "does", "lead", "to", "optimal", "behaviour", ",", "even", "though", "maximizing", "immediate", "information", "gain", "does", "not", "(", "Section", "[", "reference", "]", ")", ".", "Finally", ",", "there", "may", "be", "a", "connection", "between", "sequential", "normalized", "maximum", "likelihood", "estimators", "and", "our", "pseudo", "-", "count", "derivation", "[", "see", "e.g.", "][]", "ollivier15laplace", ".", "Intrinsic", "motivation", "has", "also", "been", "studied", "in", "reinforcement", "learning", "proper", ",", "in", "particular", "in", "the", "context", "of", "discovering", "skills", "singh04intrinsically", ",", "barto13intrinsic", ".", "Recently", ",", "stadie15incentivizing", "used", "a", "squared", "prediction", "error", "bonus", "for", "exploring", "in", "Atari", "2600", "games", ".", "Closest", "to", "our", "work", "is", "houthooft16curiosity", "’s", "variational", "approach", "to", "intrinsic", "motivation", ",", "which", "is", "equivalent", "to", "a", "second", "order", "Taylor", "approximation", "to", "prediction", "gain", ".", "mohamed15variational", "also", "considered", "a", "variational", "approach", "to", "the", "different", "problem", "of", "maximizing", "an", "agent", "’s", "ability", "to", "influence", "its", "environment", ".", "Aside", "for", "orseau13universal", "’s", "above", "-", "cited", "work", ",", "it", "is", "only", "recently", "that", "theoretical", "guarantees", "for", "exploration", "have", "emerged", "for", "non", "-", "tabular", ",", "stateful", "settings", ".", "We", "note", "pazis16efficient", "’s", "PAC", "-", "MDP", "result", "for", "metric", "spaces", "and", "leike16thompson", "’s", "asymptotic", "analysis", "of", "Thompson", "sampling", "in", "general", "environments", ".", "section", ":", "Future", "Directions", "The", "last", "few", "years", "have", "seen", "tremendous", "advances", "in", "learning", "representations", "for", "reinforcement", "learning", ".", "Surprisingly", ",", "these", "advances", "have", "yet", "to", "carry", "over", "to", "the", "problem", "of", "exploration", ".", "In", "this", "paper", ",", "we", "reconciled", "counts", ",", "the", "fundamental", "unit", "of", "uncertainty", ",", "with", "prediction", "-", "based", "heuristics", "and", "intrinsic", "motivation", ".", "Combining", "our", "work", "with", "more", "ideas", "from", "deep", "learning", "and", "better", "density", "models", "seems", "a", "plausible", "avenue", "for", "quick", "progress", "in", "practical", ",", "efficient", "exploration", ".", "We", "now", "conclude", "by", "outlining", "a", "few", "research", "directions", "we", "believe", "are", "promising", ".", "Induced", "metric", ".", "We", "did", "not", "address", "the", "question", "of", "where", "the", "generalization", "comes", "from", ".", "Clearly", ",", "the", "choice", "of", "density", "model", "induces", "a", "particular", "metric", "over", "the", "state", "space", ".", "A", "better", "understanding", "of", "this", "metric", "should", "allow", "us", "to", "tailor", "the", "density", "model", "to", "the", "problem", "of", "exploration", ".", "Compatible", "value", "function", ".", "There", "may", "be", "a", "mismatch", "in", "the", "learning", "rates", "of", "the", "density", "model", "and", "the", "value", "function", ":", "DQN", "learns", "much", "more", "slowly", "than", "our", "CTS", "model", ".", "As", "such", ",", "it", "should", "be", "beneficial", "to", "design", "value", "functions", "compatible", "with", "density", "models", "(", "or", "vice", "-", "versa", ")", ".", "The", "continuous", "case", ".", "Although", "we", "focused", "here", "on", "countable", "state", "spaces", ",", "we", "can", "as", "easily", "define", "a", "pseudo", "-", "count", "in", "terms", "of", "probability", "density", "functions", ".", "At", "present", "it", "is", "unclear", "whether", "this", "provides", "us", "with", "the", "right", "notion", "of", "counts", "for", "continuous", "spaces", ".", "subsubsection", ":", "Acknowledgments", "The", "authors", "would", "like", "to", "thank", "Laurent", "Orseau", ",", "Alex", "Graves", ",", "Joel", "Veness", ",", "Charles", "Blundell", ",", "Shakir", "Mohamed", ",", "Ivo", "Danihelka", ",", "Ian", "Osband", ",", "Matt", "Hoffman", ",", "Greg", "Wayne", ",", "Will", "Dabney", ",", "and", "Aäron", "van", "den", "Oord", "for", "their", "excellent", "feedback", "early", "and", "late", "in", "the", "writing", ",", "and", "Pierre", "-", "Yves", "Oudeyer", "and", "Yann", "Ollivier", "for", "pointing", "out", "additional", "connections", "to", "the", "literature", ".", "bibliography", ":", "References", "appendix", ":", "The", "Connection", "to", "Intrinsic", "Motivation", "The", "following", "provides", "an", "identity", "connecting", "information", "gain", "and", "prediction", "gain", ".", "theorem", ":", ".", "Consider", "a", "mixture", "model", "ξ", "over", "M", "with", "prediction", "gain", "PGn", "and", "information", "gain", "IGn", ",", "a", "fixed", "∈xX", ",", "and", "let", ":", "=", "⁢w′n", "(", "x", ")", "⁢wn", "(", "ρ", ",", "x", ")", "be", "the", "posterior", "of", "ξ", "over", "M", "after", "observing", "x.", "Let", ":", "=", "⁢w′′n", "(", "x", ")", "⁢w′n", "(", "ρ", ",", "x", ")", "be", "the", "same", "posterior", "after", "observing", "x", "a", "second", "time", ",", "and", "let", "⁢PGρn", "(", "x", ")", "denote", "the", "prediction", "gain", "of", "∈ρM.", "Then", "In", "particular", ",", "if", "M", "is", "a", "class", "of", "non", "-", "adaptive", "models", "in", "the", "sense", "that", "=", "⁢ρn", "(", "x", ")", "⁢ρ", "(", "x", ")", "for", "all", "x:1n", ",", "then", "A", "model", "which", "is", "non", "-", "adaptive", "is", "also", "learning", "-", "positive", "in", "the", "sense", "of", "Definition", "[", "reference", "]", ".", "Many", "common", "mixture", "models", ",", "for", "example", "Dirichlet", "-", "multinomial", "estimators", ",", "are", "mixtures", "over", "non", "-", "adaptive", "models", ".", "proof", ":", "Proof", ".", "We", "rewrite", "the", "posterior", "update", "rule", "(", "[", "reference", "]", ")", "to", "show", "that", "for", "any", "and", "any", ",", "Write", ".", "Now", "The", "second", "statement", "follows", "immediately", ".", "theorem", ":", ".", "The", "functions", ":", "=", "⁢f", "(", "x", ")-", "ex1x", "and", ":", "=", "⁢g", "(", "x", ")-", "ex1x2", "are", "nonnegative", "on", "∈x", "[", "0", ",", "∞", ")", ".", "proof", ":", "Proof", ".", "The", "statement", "regarding", "follows", "directly", "from", "the", "Taylor", "expansion", "for", ".", "Now", ",", "the", "first", "derivative", "of", "is", ".", "It", "is", "clearly", "positive", "for", ".", "For", ",", "Since", ",", "the", "second", "result", "follows", ".", "∎", "proof", ":", "Proof", "(", "Theorem", "[", "reference", "]", ")", ".", "The", "inequality", "follows", "directly", "from", "Lemma", "[", "reference", "]", ",", "the", "nonnegativity", "of", "the", "Kullback", "-", "Leibler", "divergence", ",", "and", "the", "fact", "that", "all", "models", "in", "are", "learning", "-", "positive", ".", "For", "the", "inequality", ",", "we", "write", "where", "(", "a", ")", "follows", "by", "definition", "of", "prediction", "gain", ",", "(", "b", ")", "from", ",", "and", "(", "c", ")", "from", "Lemma", "[", "reference", "]", ".", "Using", "the", "second", "part", "of", "Lemma", "[", "reference", "]", "in", "(", "c", ")", "yields", "the", "inequality", ".", "∎", "appendix", ":", "Asymptotic", "Analysis", "We", "begin", "with", "a", "simple", "lemma", "which", "will", "prove", "useful", "throughout", ".", "theorem", ":", ".", "The", "rate", "of", "change", "of", "the", "empirical", "distribution", ",", "-", "⁢μ′n", "(", "x", ")", "⁢μn", "(", "x", ")", ",", "is", "such", "that", "proof", ":", "Proof", ".", "We", "expand", "the", "definition", "of", "and", ":", "∎", "Using", "this", "lemma", ",", "we", "derive", "an", "asymptotic", "relationship", "between", "and", ".", "proof", ":", "Proof", "(", "Theorem", "[", "reference", "]", ")", ".", "We", "expand", "the", "definition", "of", "and", ":", "with", "the", "last", "line", "following", "from", "Lemma", "[", "reference", "]", ".", "Under", "Assumption", "[", "reference", "]", ",", "all", "terms", "of", "the", "right", "-", "hand", "side", "converge", "as", ".", "Taking", "the", "limit", "on", "both", "sides", ",", "where", "(", "a", ")", "is", "justified", "by", "the", "existence", "of", "the", "relevant", "limits", "and", ",", "and", "(", "b", ")", "follows", "from", "writing", "as", ",", "where", "all", "limits", "involved", "exist", ".", "∎", "subsection", ":", "Directed", "Graphical", "Models", "We", "say", "that", "is", "a", "factored", "state", "space", "if", "it", "is", "the", "Cartesian", "product", "of", "subspaces", ",", "i.e.", ".", "This", "factored", "structure", "allows", "us", "to", "construct", "approximate", "density", "models", "over", ",", "for", "example", "by", "modelling", "the", "joint", "density", "as", "a", "product", "of", "marginals", ".", "We", "write", "the", "factor", "of", "a", "state", "as", ",", "and", "write", "the", "sequence", "of", "the", "factor", "across", "as", ".", "We", "will", "show", "that", "directed", "graphical", "models", "wainwright08graphical", "satisfy", "Assumption", "[", "reference", "]", ".", "A", "directed", "graphical", "model", "describes", "a", "probability", "distribution", "over", "a", "factored", "state", "space", ".", "To", "the", "factor", "is", "associated", "a", "parent", "set", ".", "Let", "denote", "the", "value", "of", "the", "factors", "in", "the", "parent", "set", ".", "The", "factor", "model", "is", ",", "with", "the", "understanding", "that", "is", "allowed", "to", "make", "a", "different", "prediction", "for", "each", "value", "of", ".", "The", "state", "is", "assigned", "the", "joint", "probability", "Common", "choices", "for", "include", "the", "conditional", "empirical", "distribution", "and", "the", "Dirichlet", "estimator", ".", "theorem", ":", ".", "Suppose", "that", "each", "factor", "model", "ρin", "converges", "to", "the", "conditional", "probability", "distribution", "μ", "(", "xi|x⁢π", "(", "i", ")", ")", "and", "that", "for", "each", "xi", "with", "μ", "(", "xi|x⁢π", "(", "i", ")", ")", ",", "Then", "for", "all", "x", "with", ">", "⁢μ", "(", "x", ")", "0", ",", "the", "density", "model", "ρgm", "satisfies", "Assumption", "with", "The", "CTS", "density", "model", "used", "in", "our", "experiments", "is", "in", "fact", "a", "particular", "kind", "of", "induced", "graphical", "model", ".", "The", "result", "above", "thus", "describes", "how", "the", "pseudo", "-", "counts", "computed", "in", "Section", "[", "reference", "]", "are", "asymptotically", "related", "to", "the", "empirical", "counts", ".", "proof", ":", "Proof", ".", "By", "hypothesis", ",", ".", "Combining", "this", "with", ",", "Similarly", ",", "where", "in", "(", "a", ")", "we", "used", "the", "identity", "derived", "in", "the", "proof", "of", "Theorem", "[", "reference", "]", ".", "Now", "Let", "and", ".", "The", "difference", "of", "products", "above", "is", "and", "By", "the", "hypothesis", "on", "the", "rate", "of", "change", "of", "and", "the", "identity", ",", "we", "have", "Since", "the", "limits", "of", "and", "are", "both", ",", "we", "deduce", "that", "Now", ",", "if", "then", "also", "for", "each", "factor", ".", "Hence", ".", "∎", "subsection", ":", "Tabular", "Density", "Models", "(", "Corollary", "[", "reference", "]", ")", "We", "shall", "prove", "the", "following", ",", "which", "includes", "Corollary", "[", "reference", "]", "as", "a", "special", "case", ".", "theorem", ":", ".", "Consider", ":", "ϕ→×XX*R", "+", ".", "Suppose", "that", "for", "all", "(", ":", "xn∈nN", ")", "and", "every", "∈xX", ",", "and", ".", "Let", "⁢ρn", "(", "x", ")", "be", "the", "count", "-", "based", "estimator", "If", "^Nn", "is", "the", "pseudo", "-", "count", "corresponding", "to", "ρn", "then", "→⁢", "/", "⁢^Nn", "(", "x", ")", "Nn", "(", "x", ")", "1", "for", "all", "x", "with", ">", "⁢μ", "(", "x", ")", "0", ".", "Condition", "2", "is", "satisfied", "if", "with", "monotonically", "increasing", "in", "(", "but", "not", "too", "quickly", "!", ")", "and", "converging", "to", "some", "distribution", "for", "all", "sequences", ".", "This", "is", "the", "case", "for", "most", "tabular", "density", "models", ".", "proof", ":", "Proof", ".", "We", "will", "show", "that", "the", "condition", "on", "the", "rate", "of", "change", "required", "by", "Proposition", "[", "reference", "]", "is", "satisfied", "under", "the", "stated", "conditions", ".", "Let", ",", ",", "and", ".", "By", "hypothesis", ",", "Note", "that", "we", "do", "not", "require", ".", "Now", "Using", "Lemma", "[", "reference", "]", "we", "deduce", "that", "Since", "and", "similarly", "for", ",", "then", "pointwise", "implies", "that", "also", ".", "For", "any", ",", "where", "a", ")", "follows", "from", "and", "b", ")", "is", "justified", "by", "and", "the", "hypothesis", "that", ".", "Therefore", ".", "Hence", "Since", ",", "we", "further", "deduce", "from", "Theorem", "[", "reference", "]", "that", "The", "condition", ",", "which", "was", "also", "needed", "in", "Proposition", "[", "reference", "]", ",", "is", "necessary", "for", "the", "ratio", "to", "converge", "to", "1", ":", "for", "example", ",", "if", "grows", "as", "but", "grows", "as", "(", "with", "finite", ")", "then", "will", "grow", "as", "the", "larger", ".", "appendix", ":", "Experimental", "Methods", "subsection", ":", "CTS", "Density", "Model", "Our", "state", "space", "is", "the", "set", "of", "all", "preprocessed", "Atari", "2600", "frames", ".", "Each", "raw", "frame", "is", "composed", "of", "7", "-", "bit", "NTSC", "pixels", "bellemare13arcade", ".", "We", "preprocess", "these", "frames", "by", "first", "converting", "them", "to", "grayscale", "(", "luminance", ")", ",", "then", "downsampling", "to", "by", "averaging", "over", "pixel", "values", "(", "Figure", "[", "reference", "]", ")", ".", "Aside", "from", "this", "preprocessing", ",", "our", "model", "is", "very", "similar", "to", "the", "model", "used", "by", "bellemare14skip", "and", "veness15compress", ".", "The", "CTS", "density", "model", "treats", "as", "a", "factored", "state", ",", "where", "each", "pixel", "corresponds", "to", "a", "factor", ".", "The", "parents", "of", "this", "factor", "are", "its", "upper", "-", "left", "neighbours", ",", "i.e.", "pixels", ",", ",", "and", "(", "in", "this", "order", ")", ".", "The", "probability", "of", "is", "then", "the", "product", "of", "the", "probability", "assigned", "to", "its", "factors", ".", "Each", "factor", "is", "modelled", "using", "a", "location", "-", "dependent", "CTS", "model", ",", "which", "predicts", "the", "pixel", "’s", "colour", "value", "conditional", "on", "some", ",", "all", ",", "or", "possibly", "none", ",", "of", "the", "pixel", "’s", "parents", "(", "Figure", "[", "reference", "]", ")", ".", "subsection", ":", "A", "Taxonomy", "of", "Exploration", "We", "provide", "in", "Table", "[", "reference", "]", "a", "rough", "taxonomy", "of", "the", "Atari", "2600", "games", "available", "through", "the", "ALE", "in", "terms", "of", "the", "difficulty", "of", "exploration", ".", "We", "first", "divided", "the", "games", "into", "two", "groups", ":", "those", "for", "which", "local", "exploration", "(", "e.g.", "-", "greedy", ")", "is", "sufficient", "to", "achieve", "a", "high", "scoring", "policy", "(", "easy", ")", ",", "and", "those", "for", "which", "it", "is", "not", "(", "hard", ")", ".", "For", "example", ",", "Space", "Invaders", "versus", "Pitfall", "!", ".", "We", "further", "divided", "the", "easy", "group", "based", "on", "whether", "an", "-", "greedy", "scheme", "finds", "a", "score", "exploit", ",", "that", "is", "maximizes", "the", "score", "without", "achieving", "the", "game", "’s", "stated", "objective", ".", "For", "example", ",", "Kung", "-", "Fu", "Master", "versus", "Boxing", ".", "While", "this", "distinction", "is", "not", "directly", "used", "here", ",", "score", "exploits", "lead", "to", "behaviours", "which", "are", "optimal", "from", "an", "ALE", "perspective", "but", "uninteresting", "to", "humans", ".", "We", "divide", "the", "games", "in", "the", "hard", "category", "into", "dense", "reward", "games", "(", "Ms.", "Pac", "-", "Man", ")", "and", "sparse", "reward", "games", "(", "Montezuma", "’s", "Revenge", ")", ".", "subsection", ":", "Exploration", "in", "Montezuma", "’s", "Revenge", "Montezuma", "’s", "Revenge", "is", "divided", "into", "three", "levels", ",", "each", "composed", "of", "24", "rooms", "arranged", "in", "a", "pyramidal", "shape", "(", "Figure", "[", "reference", "]", ")", ".", "As", "discussed", "above", ",", "each", "room", "poses", "a", "number", "of", "challenges", ":", "to", "escape", "the", "very", "first", "room", ",", "the", "agent", "must", "climb", "ladders", ",", "dodge", "a", "creature", ",", "pick", "up", "a", "key", ",", "then", "backtrack", "to", "open", "one", "of", "two", "doors", ".", "The", "number", "of", "rooms", "reached", "by", "an", "agent", "is", "therefore", "a", "good", "measure", "of", "its", "ability", ".", "By", "accessing", "the", "game", "RAM", ",", "we", "recorded", "the", "location", "of", "the", "agent", "at", "each", "step", "during", "the", "course", "of", "training", ".", "We", "computed", "the", "visit", "count", "to", "each", "room", ",", "averaged", "over", "epochs", "each", "lasting", "one", "million", "frames", ".", "From", "this", "information", "we", "constructed", "a", "map", "of", "the", "agent", "’s", "“", "known", "world", "”", ",", "that", "is", ",", "all", "rooms", "visited", "at", "least", "once", ".", "The", "agent", "’s", "current", "room", "number", "ranges", "from", "0", "to", "23", "(", "Figure", "[", "reference", "]", ")", "and", "is", "stored", "at", "RAM", "location", "0x83", ".", "Figure", "[", "reference", "]", "shows", "the", "set", "of", "rooms", "explored", "by", "our", "DQN", "agents", "at", "different", "points", "during", "training", ".", "Figure", "[", "reference", "]", "paints", "a", "clear", "picture", ":", "after", "50", "million", "frames", ",", "the", "agent", "using", "exploration", "bonuses", "has", "seen", "a", "total", "of", "15", "rooms", ",", "while", "the", "no", "-", "bonus", "agent", "has", "seen", "two", ".", "At", "that", "point", "in", "time", ",", "our", "agent", "achieves", "an", "average", "score", "of", "2461", ";", "by", "100", "million", "frames", ",", "this", "figure", "stands", "at", "3439", ",", "higher", "than", "anything", "previously", "reported", ".", "We", "believe", "the", "success", "of", "our", "method", "in", "this", "game", "is", "a", "strong", "indicator", "of", "the", "usefulness", "of", "pseudo", "-", "counts", "for", "exploration", ".", "We", "remark", "that", "without", "mixing", "in", "the", "Monte", "-", "Carlo", "return", ",", "our", "bonus", "-", "based", "agent", "still", "explores", "significantly", "more", "than", "the", "no", "-", "bonus", "agent", ".", "However", ",", "the", "deep", "network", "seems", "unable", "to", "maintain", "a", "sufficiently", "good", "approximation", "to", "the", "value", "function", ",", "and", "performance", "quickly", "deteriorates", ".", "Comparable", "results", "using", "the", "A3C", "method", "provide", "another", "example", "of", "the", "practical", "importance", "of", "eligibility", "traces", "and", "return", "-", "based", "methods", "in", "reinforcement", "learning", ".", "subsection", ":", "Improving", "Exploration", "for", "Actor", "-", "Critic", "Methods", "Our", "implementation", "of", "A3C", "was", "along", "the", "lines", "mentioned", "in", "and", "uses", "16", "threads", ".", "Each", "thread", "corresponds", "to", "an", "actor", "learner", "and", "maintains", "a", "copy", "of", "the", "density", "model", ".", "All", "the", "threads", "are", "synchronized", "with", "the", "master", "thread", "at", "regular", "intervals", "of", "250", ",", "000", "steps", ".", "We", "followed", "the", "same", "training", "procedure", "as", "that", "reported", "in", "the", "A3C", "paper", "with", "the", "following", "additional", "steps", ":", "We", "update", "our", "density", "model", "with", "the", "states", "generated", "by", "following", "the", "policy", ".", "During", "the", "policy", "gradient", "step", ",", "we", "compute", "the", "intrinsic", "rewards", "by", "querying", "the", "density", "model", "and", "add", "it", "to", "the", "extrinsic", "rewards", "before", "clipping", "them", "in", "the", "range", "as", "was", "done", "in", "the", "A3C", "paper", ".", "This", "resulted", "in", "minimal", "overhead", "in", "computation", "costs", "and", "the", "memory", "footprint", "was", "manageable", "(", "32", "GB", ")", "for", "most", "of", "the", "Atari", "games", ".", "Our", "training", "times", "were", "almost", "the", "same", "as", "the", "ones", "reported", "in", "the", "A3C", "paper", ".", "We", "picked", "after", "performing", "a", "short", "parameter", "sweep", "over", "the", "training", "games", ".", "The", "choice", "of", "training", "games", "is", "the", "same", "as", "mentioned", "in", "the", "A3C", "paper", ".", "The", "games", "on", "which", "DQN", "achieves", "a", "score", "of", "150", "%", "or", "less", "of", "the", "random", "score", "are", ":", "Asteroids", ",", "Double", "Dunk", ",", "Gravitar", ",", "Ice", "Hockey", ",", "Montezuma", "’s", "Revenge", ",", "Pitfall", "!", ",", "Skiing", ",", "Surround", ",", "Tennis", ",", "Time", "Pilot", ".", "The", "games", "on", "which", "A3C", "achieves", "a", "score", "of", "150", "%", "or", "less", "of", "the", "random", "score", "are", ":", "Battle", "Zone", ",", "Bowling", ",", "Enduro", ",", "Freeway", ",", "Gravitar", ",", "Kangaroo", ",", "Pitfall", "!", ",", "Robotank", ",", "Skiing", ",", "Solaris", ",", "Surround", ",", "Tennis", ",", "Time", "Pilot", ",", "Venture", ".", "The", "games", "on", "which", "A3C", "+", "achieves", "a", "score", "of", "150", "%", "or", "less", "of", "the", "random", "score", "are", ":", "Double", "Dunk", ",", "Gravitar", ",", "Ice", "Hockey", ",", "Pitfall", "!", ",", "Skiing", ",", "Solaris", ",", "Surround", ",", "Tennis", ",", "Time", "Pilot", ",", "Venture", ".", "Our", "experiments", "involved", "the", "stochastic", "version", "of", "the", "Arcade", "Learning", "Environment", "(", "ALE", ")", "without", "a", "terminal", "signal", "for", "life", "loss", ",", "which", "is", "now", "the", "default", "ALE", "setting", ".", "Briefly", ",", "the", "stochasticity", "is", "achieved", "by", "accepting", "the", "agent’", "action", "at", "each", "frame", "with", "probability", "and", "using", "the", "agent’s", "previous", "action", "during", "rejection", ".", "We", "used", "the", "ALE", "’s", "default", "value", "of", "as", "has", "been", "previously", "used", "in", ".", "For", "comparison", ",", "Table", "[", "reference", "]", "also", "reports", "the", "deterministic", "+", "life", "loss", "setting", "also", "used", "in", "the", "literature", ".", "Anecdotally", ",", "we", "found", "that", "using", "the", "life", "loss", "signal", ",", "while", "helpful", "in", "achieving", "high", "scores", "in", "some", "games", ",", "is", "detrimental", "in", "Montezuma", "’s", "Revenge", ".", "Recall", "that", "the", "life", "loss", "signal", "was", "used", "by", "mnih15human", "to", "treat", "each", "of", "the", "agent", "’", "lives", "as", "a", "separate", "episode", ".", "For", "comparison", ",", "after", "200", "million", "frames", "A3C", "+", "achieves", "the", "following", "average", "scores", ":", "1", ")", "Stochastic", "+", "Life", "Loss", ":", "142.50", ";", "2", ")", "Deterministic", "+", "Life", "Loss", ":", "273.70", "3", ")", "Stochastic", "without", "Life", "Loss", ":", "1127.05", "4", ")", "Deterministic", "without", "Life", "Loss", ":", "273.70", ".", "The", "maximum", "score", "achieved", "by", "3", ")", "is", "3600", ",", "in", "comparison", "to", "the", "maximum", "of", "500", "achieved", "by", "1", ")", "and", "3", ")", ".", "This", "large", "discrepancy", "is", "not", "unsurprising", "when", "one", "considers", "that", "losing", "a", "life", "in", "Montezuma", "’s", "Revenge", ",", "and", "in", "fact", "in", "most", "games", ",", "is", "very", "different", "from", "restarting", "a", "new", "episode", ".", "subsection", ":", "Comparing", "Exploration", "Bonuses", "In", "this", "section", "we", "compare", "the", "effect", "of", "using", "different", "exploration", "bonuses", "derived", "from", "our", "density", "model", ".", "We", "consider", "the", "following", "variants", ":", "no", "exploration", "bonus", ",", ",", "as", "per", "MBIE", "-", "EB", "strehl08analysis", ";", ",", "as", "per", "BEB", "kolter09near", ";", "and", ",", "related", "to", "compression", "progress", "schmidhuber08driven", ".", "The", "exact", "form", "of", "these", "bonuses", "is", "analogous", "to", "(", "[", "reference", "]", ")", ".", "We", "compare", "these", "variants", "after", "10", ",", "50", ",", "100", ",", "and", "200", "million", "frames", "of", "training", ",", "again", "in", "the", "A3C", "setup", ".", "To", "compare", "scores", "across", "60", "games", ",", "we", "use", "inter", "-", "algorithm", "score", "distributions", "bellemare13arcade", ".", "Inter", "-", "algorithm", "scores", "are", "normalized", "so", "that", "0", "corresponds", "to", "the", "worst", "score", "on", "a", "game", ",", "and", "1", ",", "to", "the", "best", ".", "If", "is", "a", "game", "and", "the", "inter", "-", "algorithm", "score", "on", "for", "algorithm", ",", "then", "the", "score", "distribution", "function", "is", "The", "score", "distribution", "effectively", "depicts", "a", "kind", "of", "cumulative", "distribution", ",", "with", "a", "higher", "overall", "curve", "implying", "better", "scores", "across", "the", "gamut", "of", "Atari", "2600", "games", ".", "A", "higher", "curve", "at", "implies", "top", "performance", "on", "more", "games", ";", "a", "higher", "curve", "at", "indicates", "the", "algorithm", "does", "not", "perform", "poorly", "on", "many", "games", ".", "The", "scale", "parameter", "was", "optimized", "to", "for", "each", "variant", "separately", ".", "Figure", "[", "reference", "]", "shows", "that", ",", "while", "prediction", "gain", "initially", "achieves", "strong", "performance", ",", "by", "50", "million", "frames", "all", "three", "algorithms", "perform", "equally", "well", ".", "By", "200", "million", "frames", ",", "the", "exploration", "bonus", "outperforms", "both", "prediction", "gain", "and", "no", "bonus", ".", "The", "prediction", "gain", "achieves", "a", "decent", ",", "but", "not", "top", "-", "performing", "score", "on", "all", "games", ".", "This", "matches", "our", "earlier", "argument", "that", "using", "prediction", "gain", "results", "in", "too", "little", "exploration", ".", "We", "hypothesize", "that", "the", "poor", "performance", "of", "the", "bonus", "stems", "from", "too", "abrupt", "a", "decay", "from", "a", "large", "to", "small", "intrinsic", "reward", ",", "although", "more", "experiments", "are", "needed", ".", "As", "a", "whole", ",", "these", "results", "show", "how", "using", "PG", "offers", "an", "advantage", "over", "the", "baseline", "A3C", "algorithm", ",", "which", "is", "furthered", "by", "using", "our", "count", "-", "based", "exploration", "bonus", "."], "entity": [[3, 7, ["Count", "-", "Based", "Exploration"], "Method"], [8, 10, ["Intrinsic", "Motivation"], "Method"], [37, 44, ["exploration", "in", "non", "-", "tabular", "reinforcement", "learning"], "Task"], [55, 57, ["density", "models"], "Method"], [75, 77, ["density", "model"], "Method"], [84, 89, ["count", "-", "based", "exploration", "algorithms"], "Method"], [91, 95, ["non", "-", "tabular", "case"], "Task"], [101, 104, ["Atari", "2600", "games"], "Task"], [127, 128, ["exploration"], "Task"], [132, 134, ["hard", "games"], "Task"], [139, 142, ["Montezuma", "’s", "Revenge"], "Material"], [146, 148, ["Exploration", "algorithms"], "Method"], [149, 152, ["Markov", "Decision", "Processes"], "Task"], [153, 154, ["MDPs"], "Method"], [232, 233, ["exploration"], "Task"], [234, 239, ["Count", "-", "based", "exploration", "methods"], "Method"], [250, 252, ["reducing", "uncertainty"], "Task"], [256, 261, ["Model", "-", "based", "Interval", "Estimation"], "Method"], [262, 264, ["Exploration", "Bonuses"], "Method"], [265, 266, ["MBIE"], "Method"], [272, 275, ["augmented", "Bellman", "equation"], "Method"], [324, 328, ["count", "-", "based", "methods"], "Method"], [338, 340, ["reinforcement", "learning"], "Task"], [354, 355, ["rules"], "Method"], [397, 398, ["exploration"], "Task"], [402, 403, ["barto13intrinsic"], "Method"], [425, 429, ["change", "in", "prediction", "error"], "Metric"], [466, 468, ["motivation", "methods"], "Method"], [486, 488, ["tabular", "representation"], "Method"], [495, 499, ["count", "-", "based", "algorithms"], "Method"], [505, 507, ["intrinsic", "motivation"], "Task"], [527, 528, ["exploration"], "Task"], [537, 539, ["intrinsic", "motivation"], "Task"], [540, 544, ["count", "-", "based", "exploration"], "Method"], [562, 564, ["learning", "progress"], "Metric"], [565, 568, ["information", "gain", "cover91elements"], "Metric"], [584, 586, ["information", "gain"], "Metric"], [595, 599, ["count", "-", "based", "exploration"], "Method"], [610, 613, ["pseudo", "-", "count"], "Method"], [616, 625, ["information", "-", "gain", "-", "as", "-", "learning", "-", "progress"], "Method"], [626, 630, ["count", "-", "based", "exploration"], "Method"], [639, 641, ["density", "model"], "Method"], [655, 657, ["intrinsic", "motivation"], "Task"], [665, 667, ["transition", "model"], "Method"], [680, 684, ["schmidhuber91possibility", "’s", "compression", "progress"], "Method"], [686, 688, ["prediction", "gain"], "Metric"], [690, 692, ["information", "gain"], "Metric"], [721, 724, ["intrinsic", "motivation", "methods"], "Method"], [738, 740, ["function", "approximation"], "Method"], [741, 742, ["exploration"], "Task"], [750, 753, ["Atari", "2600", "games"], "Task"], [764, 766, ["myopic", "exploration"], "Task"], [777, 779, ["density", "model"], "Method"], [786, 789, ["MBIE", "-", "EB"], "Method"], [795, 798, ["experience", "replay", "setting"], "Task"], [801, 805, ["actor", "-", "critic", "setting"], "Method"], [828, 831, ["Montezuma", "’s", "Revenge"], "Material"], [997, 999, ["empirical", "count"], "Method"], [1038, 1041, ["ergodic", "Markov", "chain"], "Method"], [1048, 1050, ["fixed", "policy"], "Method"], [1052, 1056, ["finite", "-", "state", "MDP"], "Task"], [1064, 1068, ["chain", "’s", "stationary", "distribution"], "Method"], [1074, 1076, ["density", "model"], "Method"], [1093, 1095, ["density", "model"], "Method"], [1101, 1103, ["generative", "model"], "Method"], [1108, 1110, ["density", "model"], "Method"], [1113, 1115, ["forward", "model"], "Method"], [1132, 1134, ["density", "model"], "Method"], [1197, 1199, ["Bayesian", "approach"], "Method"], [1201, 1205, ["variable", "-", "alphabet", "models"], "Method"], [1274, 1276, ["Bayesian", "estimation"], "Task"], [1284, 1286, ["Recoding", "Probability"], "Method"], [1290, 1292, ["density", "model"], "Method"], [1295, 1297, ["density", "model"], "Method"], [1326, 1328, ["density", "model"], "Method"], [1338, 1339, ["recoding"], "Task"], [1344, 1347, ["statistical", "compression", "literature"], "Task"], [1349, 1351, ["coding", "costs"], "Metric"], [1372, 1376, ["pseudo", "-", "count", "function"], "Method"], [1379, 1383, ["pseudo", "-", "count", "total"], "Method"], [1407, 1409, ["density", "model"], "Method"], [1437, 1439, ["linear", "system"], "Method"], [1472, 1474, ["density", "models"], "Method"], [1496, 1501, ["Learning", "-", "positive", "density", "model"], "Method"], [1504, 1507, ["density", "model", "ρ"], "Method"], [1581, 1583, ["Dirichlet", "estimator"], "Method"], [1613, 1618, ["Frequency", "of", "a", "Salient", "Event"], "Task"], [1619, 1620, ["Freeway"], "Material"], [1641, 1646, ["Atari", "2600", "video", "game", "Freeway"], "Material"], [1658, 1661, ["Arcade", "Learning", "Environment"], "Method"], [1726, 1728, ["visit", "counts"], "Task"], [1729, 1733, ["non", "-", "tabular", "settings"], "Task"], [1735, 1736, ["Freeway"], "Material"], [1776, 1779, ["Atari", "2600", "games"], "Task"], [1790, 1791, ["score"], "Metric"], [1840, 1843, ["pseudo", "-", "count"], "Method"], [1847, 1849, ["nonstationary", "policy"], "Method"], [1930, 1934, ["pixel", "-", "level", "version"], "Method"], [1936, 1938, ["CTS", "model"], "Method"], [1939, 1942, ["Atari", "2600", "frames"], "Material"], [1952, 1954, ["CTS", "model"], "Method"], [1967, 1969, ["density", "models"], "Method"], [1985, 1986, ["learning"], "Task"], [1993, 1994, ["exploration"], "Task"], [2032, 2035, ["pseudo", "-", "count"], "Metric"], [2083, 2084, ["robustness"], "Metric"], [2187, 2189, ["Intrinsic", "Motivation"], "Task"], [2210, 2212, ["information", "gain"], "Metric"], [2229, 2231, ["Information", "gain"], "Task"], [2237, 2239, ["mixture", "model"], "Method"], [2243, 2245, ["density", "models"], "Method"], [2275, 2277, ["Information", "gain"], "Metric"], [2295, 2297, ["information", "gain"], "Metric"], [2299, 2302, ["complex", "density", "model"], "Method"], [2319, 2321, ["prediction", "gain"], "Metric"], [2329, 2331, ["information", "gain"], "Metric"], [2335, 2337, ["prediction", "gain"], "Metric"], [2339, 2341, ["density", "model"], "Method"], [2357, 2360, ["log", "-", "probability"], "Metric"], [2362, 2364, ["Prediction", "gain"], "Metric"], [2419, 2421, ["mixture", "model"], "Method"], [2425, 2430, ["learning", "-", "positive", "models", "M."], "Method"], [2439, 2440, ["ξ"], "Method"], [2462, 2466, ["MBIE", "-", "EB", "bonus"], "Method"], [2480, 2483, ["information", "gain", "bonus"], "Method"], [2522, 2526, ["non", "-", "tabular", "settings"], "Task"], [2542, 2548, ["bonus", "-", "based", "intrinsic", "motivation", "approaches"], "Method"], [2565, 2569, ["PAC", "-", "MDP", "optimality"], "Task"], [2601, 2606, ["theoretically", "near", "-", "optimal", "exploration"], "Task"], [2653, 2656, ["intrinsic", "motivation", "algorithms"], "Method"], [2657, 2660, ["pseudo", "-", "counts"], "Method"], [2667, 2676, ["forward", "(", "transition", "and", "/", "or", "reward", ")", "model"], "Method"], [2687, 2689, ["density", "models"], "Method"], [2704, 2707, ["intrinsic", "motivation", "algorithms"], "Method"], [2709, 2711, ["forward", "models"], "Method"], [2714, 2716, ["Asymptotic", "Analysis"], "Task"], [2742, 2745, ["tabular", "density", "models"], "Method"], [2767, 2773, ["approximation", "error", "of", "pseudo", "-", "counts"], "Metric"], [2775, 2778, ["directed", "graphical", "models"], "Method"], [2782, 2784, ["CTS", "model"], "Method"], [2844, 2846, ["density", "model"], "Method"], [2912, 2914, ["learning", "rate"], "Metric"], [3019, 3021, ["density", "model"], "Method"], [3038, 3040, ["density", "model"], "Method"], [3042, 3044, ["empirical", "distribution"], "Method"], [3056, 3058, ["stochastic", "approximation"], "Method"], [3104, 3106, ["density", "model"], "Method"], [3147, 3151, ["count", "-", "based", "estimator"], "Method"], [3152, 3153, ["^Nn"], "Method"], [3201, 3202, ["exploration"], "Task"], [3207, 3210, ["Arcade", "Learning", "Environment"], "Method"], [3214, 3216, ["CTS", "model"], "Method"], [3224, 3225, ["Exploration"], "Task"], [3226, 3230, ["Hard", "Atari", "2600", "Games"], "Task"], [3236, 3239, ["Arcade", "Learning", "Environment"], "Method"], [3252, 3255, ["-", "greedy", "policy"], "Method"], [3284, 3287, ["optimistic", "initialization", "trick"], "Method"], [3289, 3290, ["machado14domainindependent"], "Method"], [3296, 3299, ["Q", "-", "functions"], "Method"], [3300, 3303, ["Double", "DQN", "vanhasselt16deep"], "Method"], [3312, 3317, ["Double", "Q", "-", "Learning", "target"], "Method"], [3319, 3322, ["Monte", "Carlo", "return"], "Method"], [3359, 3361, ["optimistic", "initialization"], "Method"], [3363, 3364, ["Freeway"], "Material"], [3371, 3372, ["DQN"], "Method"], [3377, 3382, ["count", "-", "based", "exploration", "bonus"], "Method"], [3397, 3400, ["Montezuma", "’s", "Revenge"], "Material"], [3401, 3402, ["Venture"], "Material"], [3403, 3406, ["Montezuma", "’s", "Revenge"], "Material"], [3410, 3413, ["Atari", "2600", "game"], "Material"], [3457, 3458, ["score"], "Metric"], [3517, 3518, ["score"], "Metric"], [3567, 3568, ["exploration"], "Task"], [3573, 3577, ["Actor", "-", "Critic", "Methods"], "Method"], [3581, 3583, ["exploration", "bonuses"], "Method"], [3587, 3588, ["A3C"], "Method"], [3589, 3594, ["Asynchronous", "Advantage", "Actor", "-", "Critic"], "Method"], [3602, 3606, ["actor", "-", "critic", "methods"], "Method"], [3636, 3637, ["exploration"], "Task"], [3660, 3663, ["Atari", "2600", "games"], "Task"], [3680, 3682, ["augmented", "algorithm"], "Method"], [3719, 3720, ["score"], "Metric"], [3724, 3725, ["random"], "Method"], [3743, 3745, ["random", "agent"], "Method"], [3753, 3754, ["DQN"], "Method"], [3830, 3832, ["median", "performance"], "Metric"], [3857, 3860, ["Atari", "2600", "games"], "Task"], [3863, 3866, ["-", "greedy", "policy"], "Method"], [3869, 3870, ["exploration"], "Task"], [3885, 3888, ["intrinsically", "motivated", "behaviour"], "Task"], [3892, 3894, ["prediction", "gain"], "Metric"], [3990, 3995, ["sequential", "normalized", "maximum", "likelihood", "estimators"], "Method"], [3997, 4001, ["pseudo", "-", "count", "derivation"], "Method"], [4014, 4017, ["reinforcement", "learning", "proper"], "Task"], [4024, 4027, ["discovering", "skills", "singh04intrinsically"], "Task"], [4032, 4033, ["stadie15incentivizing"], "Method"], [4035, 4039, ["squared", "prediction", "error", "bonus"], "Method"], [4040, 4041, ["exploring"], "Task"], [4042, 4045, ["Atari", "2600", "games"], "Task"], [4051, 4055, ["houthooft16curiosity", "’s", "variational", "approach"], "Method"], [4056, 4058, ["intrinsic", "motivation"], "Task"], [4064, 4068, ["second", "order", "Taylor", "approximation"], "Method"], [4076, 4078, ["variational", "approach"], "Method"], [4110, 4111, ["exploration"], "Task"], [4114, 4120, ["non", "-", "tabular", ",", "stateful", "settings"], "Task"], [4125, 4129, ["PAC", "-", "MDP", "result"], "Method"], [4130, 4132, ["metric", "spaces"], "Task"], [4135, 4137, ["asymptotic", "analysis"], "Task"], [4138, 4140, ["Thompson", "sampling"], "Task"], [4157, 4159, ["learning", "representations"], "Task"], [4160, 4162, ["reinforcement", "learning"], "Task"], [4176, 4177, ["exploration"], "Task"], [4193, 4197, ["prediction", "-", "based", "heuristics"], "Method"], [4208, 4210, ["deep", "learning"], "Method"], [4212, 4214, ["density", "models"], "Method"], [4225, 4226, ["exploration"], "Task"], [4241, 4243, ["Induced", "metric"], "Metric"], [4253, 4254, ["generalization"], "Task"], [4262, 4264, ["density", "model"], "Method"], [4285, 4287, ["density", "model"], "Method"], [4291, 4292, ["exploration"], "Task"], [4304, 4306, ["learning", "rates"], "Metric"], [4308, 4310, ["density", "model"], "Method"], [4312, 4314, ["value", "function"], "Method"], [4315, 4316, ["DQN"], "Method"], [4322, 4324, ["CTS", "model"], "Method"], [4334, 4336, ["value", "functions"], "Method"], [4338, 4340, ["density", "models"], "Method"], [4348, 4350, ["continuous", "case"], "Task"], [4372, 4375, ["probability", "density", "functions"], "Method"], [4494, 4497, ["mixture", "model", "ξ"], "Method"], [4500, 4503, ["prediction", "gain", "PGn"], "Method"], [4504, 4507, ["information", "gain", "IGn"], "Method"], [4577, 4578, ["M"], "Method"], [4582, 4586, ["non", "-", "adaptive", "models"], "Method"], [4627, 4629, ["mixture", "models"], "Method"], [4632, 4636, ["Dirichlet", "-", "multinomial", "estimators"], "Method"], [4638, 4644, ["mixtures", "over", "non", "-", "adaptive", "models"], "Method"], [4652, 4655, ["posterior", "update", "rule"], "Method"], [4718, 4720, ["Taylor", "expansion"], "Method"], [4839, 4841, ["Asymptotic", "Analysis"], "Method"], [4987, 4990, ["Directed", "Graphical", "Models"], "Method"], [5010, 5012, ["factored", "structure"], "Method"], [5016, 5019, ["approximate", "density", "models"], "Method"], [5030, 5033, ["product", "of", "marginals"], "Method"], [5057, 5060, ["directed", "graphical", "models"], "Method"], [5068, 5071, ["directed", "graphical", "model"], "Method"], [5073, 5075, ["probability", "distribution"], "Method"], [5103, 5105, ["factor", "model"], "Method"], [5135, 5138, ["conditional", "empirical", "distribution"], "Method"], [5140, 5142, ["Dirichlet", "estimator"], "Method"], [5149, 5152, ["factor", "model", "ρin"], "Method"], [5192, 5195, ["density", "model", "ρgm"], "Method"], [5199, 5202, ["CTS", "density", "model"], "Method"], [5213, 5216, ["induced", "graphical", "model"], "Method"], [5325, 5328, ["Tabular", "Density", "Models"], "Method"], [5380, 5384, ["count", "-", "based", "estimator"], "Method"], [5385, 5386, ["^Nn"], "Method"], [5448, 5451, ["tabular", "density", "models"], "Method"], [5597, 5600, ["CTS", "Density", "Model"], "Method"], [5608, 5612, ["preprocessed", "Atari", "2600", "frames"], "Material"], [5675, 5678, ["CTS", "density", "model"], "Method"], [5736, 5741, ["location", "-", "dependent", "CTS", "model"], "Method"], [5776, 5777, ["Exploration"], "Task"], [5789, 5792, ["Atari", "2600", "games"], "Task"], [5802, 5803, ["exploration"], "Task"], [5865, 5868, ["-", "greedy", "scheme"], "Method"], [5870, 5871, ["score"], "Metric"], [5877, 5878, ["score"], "Metric"], [5905, 5906, ["score"], "Metric"], [5931, 5934, ["dense", "reward", "games"], "Method"], [5935, 5939, ["Ms.", "Pac", "-", "Man"], "Method"], [5941, 5944, ["sparse", "reward", "games"], "Method"], [5945, 5948, ["Montezuma", "’s", "Revenge"], "Material"], [5952, 5953, ["Exploration"], "Task"], [5954, 5957, ["Montezuma", "’s", "Revenge"], "Material"], [5957, 5960, ["Montezuma", "’s", "Revenge"], "Material"], [6146, 6148, ["DQN", "agents"], "Method"], [6202, 6203, ["score"], "Metric"], [6245, 6246, ["exploration"], "Task"], [6260, 6264, ["bonus", "-", "based", "agent"], "Method"], [6270, 6274, ["no", "-", "bonus", "agent"], "Method"], [6278, 6280, ["deep", "network"], "Method"], [6312, 6314, ["eligibility", "traces"], "Method"], [6315, 6319, ["return", "-", "based", "methods"], "Method"], [6320, 6322, ["reinforcement", "learning"], "Task"], [6326, 6327, ["Exploration"], "Task"], [6328, 6332, ["Actor", "-", "Critic", "Methods"], "Method"], [6335, 6336, ["A3C"], "Method"], [6352, 6354, ["actor", "learner"], "Method"], [6360, 6362, ["density", "model"], "Method"], [6403, 6405, ["density", "model"], "Method"], [6416, 6419, ["policy", "gradient", "step"], "Method"], [6428, 6430, ["density", "model"], "Method"], [6457, 6459, ["computation", "costs"], "Metric"], [6461, 6463, ["memory", "footprint"], "Metric"], [6473, 6475, ["Atari", "games"], "Material"], [6477, 6479, ["training", "times"], "Metric"], [6524, 6525, ["DQN"], "Method"], [6527, 6528, ["score"], "Metric"], [6544, 6545, ["Gravitar"], "Material"], [6549, 6552, ["Montezuma", "’s", "Revenge"], "Material"], [6572, 6573, ["score"], "Metric"], [6591, 6592, ["Freeway"], "Material"], [6593, 6594, ["Gravitar"], "Material"], [6613, 6614, ["Venture"], "Material"], [6623, 6624, ["score"], "Metric"], [6638, 6639, ["Gravitar"], "Material"], [6657, 6658, ["Venture"], "Material"], [6663, 6665, ["stochastic", "version"], "Method"], [6667, 6670, ["Arcade", "Learning", "Environment"], "Method"], [6671, 6672, ["ALE"], "Method"], [6774, 6777, ["Montezuma", "’s", "Revenge"], "Material"], [6813, 6815, ["average", "scores"], "Metric"], [6818, 6822, ["Stochastic", "+", "Life", "Loss"], "Metric"], [6827, 6831, ["Deterministic", "+", "Life", "Loss"], "Metric"], [6835, 6839, ["Stochastic", "without", "Life", "Loss"], "Method"], [6852, 6853, ["score"], "Metric"], [6889, 6892, ["Montezuma", "’s", "Revenge"], "Material"], [6912, 6914, ["Exploration", "Bonuses"], "Task"], [6929, 6931, ["density", "model"], "Method"], [6945, 6946, ["MBIE"], "Method"], [6960, 6961, ["compression"], "Method"], [7019, 7023, ["Inter", "-", "algorithm", "scores"], "Metric"], [7053, 7054, ["score"], "Metric"], [7065, 7066, ["score"], "Metric"], [7072, 7074, ["cumulative", "distribution"], "Method"], [7087, 7090, ["Atari", "2600", "games"], "Task"], [7136, 7138, ["prediction", "gain"], "Metric"], [7160, 7162, ["exploration", "bonus"], "Method"], [7164, 7166, ["prediction", "gain"], "Metric"], [7171, 7173, ["prediction", "gain"], "Metric"], [7182, 7183, ["score"], "Metric"], [7240, 7241, ["PG"], "Method"], [7256, 7261, ["count", "-", "based", "exploration", "bonus"], "Method"], [755, 758, ["Arcade", "Learning", "Environment"], "Method"], [1793, 1794, ["ALE"], "Method"], [3416, 3417, ["ALE"], "Method"], [3645, 3646, ["A3C"], "Method"], [3657, 3658, ["A3C"], "Method"], [3683, 3684, ["A3C"], "Method"], [3701, 3702, ["A3C"], "Method"], [3736, 3737, ["A3C"], "Method"], [3784, 3785, ["A3C"], "Method"], [3791, 3792, ["score"], "Metric"], [3807, 3808, ["score"], "Metric"], [3809, 3810, ["A3C"], "Method"], [3811, 3812, ["A3C"], "Method"], [3825, 3826, ["A3C"], "Method"], [3838, 3839, ["A3C"], "Method"], [5795, 5796, ["ALE"], "Method"], [5915, 5916, ["ALE"], "Method"], [6302, 6303, ["A3C"], "Method"], [6392, 6393, ["A3C"], "Method"], [6448, 6449, ["A3C"], "Method"], [6489, 6490, ["A3C"], "Method"], [6517, 6518, ["A3C"], "Method"], [6536, 6537, ["score"], "Metric"], [6569, 6570, ["A3C"], "Method"], [6581, 6582, ["score"], "Metric"], [6619, 6620, ["A3C"], "Method"], [6632, 6633, ["score"], "Metric"], [6686, 6687, ["ALE"], "Method"], [6717, 6718, ["ALE"], "Method"], [6808, 6809, ["A3C"], "Method"], [7000, 7001, ["A3C"], "Method"], [7015, 7016, ["score"], "Metric"], [7032, 7033, ["score"], "Metric"], [7060, 7061, ["score"], "Metric"], [7247, 7248, ["A3C"], "Method"]]} {"doc_id": "189e6bb7523733c4e524214b9e6ae92d4ed50dac", "text": ["document", ":", "Transfer", "Learning", "for", "Sequence", "Tagging", "with", "Hierarchical", "Recurrent", "Networks", "Recent", "papers", "have", "shown", "that", "neural", "networks", "obtain", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "several", "different", "sequence", "tagging", "tasks", ".", "One", "appealing", "property", "of", "such", "systems", "is", "their", "generality", ",", "as", "excellent", "performance", "can", "be", "achieved", "with", "a", "unified", "architecture", "and", "without", "task", "-", "specific", "feature", "engineering", ".", "However", ",", "it", "is", "unclear", "if", "such", "systems", "can", "be", "used", "for", "tasks", "without", "large", "amounts", "of", "training", "data", ".", "In", "this", "paper", "we", "explore", "the", "problem", "of", "transfer", "learning", "for", "neural", "sequence", "taggers", ",", "where", "a", "source", "task", "with", "plentiful", "annotations", "(", "e.g.", ",", "POS", "tagging", "on", "Penn", "Treebank", ")", "is", "used", "to", "improve", "performance", "on", "a", "target", "task", "with", "fewer", "available", "annotations", "(", "e.g.", ",", "POS", "tagging", "for", "microblogs", ")", ".", "We", "examine", "the", "effects", "of", "transfer", "learning", "for", "deep", "hierarchical", "recurrent", "networks", "across", "domains", ",", "applications", ",", "and", "languages", ",", "and", "show", "that", "significant", "improvement", "can", "often", "be", "obtained", ".", "These", "improvements", "lead", "to", "improvements", "over", "the", "current", "state", "-", "of", "-", "the", "-", "art", "on", "several", "well", "-", "studied", "tasks", ".", "section", ":", "Introduction", "Sequence", "tagging", "is", "an", "important", "problem", "in", "natural", "language", "processing", ",", "which", "has", "wide", "applications", "including", "part", "-", "of", "-", "speech", "(", "POS", ")", "tagging", ",", "text", "chunking", ",", "and", "named", "entity", "recognition", "(", "NER", ")", ".", "Given", "a", "sequence", "of", "words", ",", "sequence", "tagging", "aims", "to", "predict", "a", "linguistic", "tag", "for", "each", "word", "such", "as", "the", "POS", "tag", ".", "An", "important", "challenge", "for", "sequence", "tagging", "is", "how", "to", "transfer", "knowledge", "from", "one", "task", "to", "another", ",", "which", "is", "often", "referred", "to", "as", "transfer", "learning", "pan2010survey", ".", "Transfer", "learning", "can", "be", "used", "in", "several", "settings", ",", "notably", "for", "low", "-", "resource", "languages", "zirikly2cross", ",", "wang2013cross", "and", "low", "-", "resource", "domains", "such", "as", "biomedical", "corpora", "kim2003genia", "and", "Twitter", "corpora", "ritter2011named", ")", ".", "In", "these", "cases", ",", "transfer", "learning", "can", "improve", "performance", "by", "taking", "advantage", "of", "more", "plentiful", "labels", "from", "related", "tasks", ".", "Even", "on", "datasets", "with", "relatively", "abundant", "labels", ",", "multi", "-", "task", "transfer", "can", "sometimes", "achieve", "improvement", "over", "state", "-", "of", "-", "the", "-", "art", "results", "collobert2011natural", ".", "Recently", ",", "a", "number", "of", "approaches", "based", "on", "deep", "neural", "networks", "have", "addressed", "the", "problem", "of", "sequence", "tagging", "in", "an", "end", "-", "to", "-", "end", "manner", "collobert2011natural", ",", "lample2016neural", ",", "ling2015finding", ",", "ma2016end", ".", "These", "neural", "networks", "consist", "of", "multiple", "layers", "of", "neurons", "organized", "in", "a", "hierarchy", "and", "can", "transform", "the", "input", "tokens", "to", "the", "output", "labels", "without", "explicit", "hand", "-", "engineered", "feature", "extraction", ".", "The", "aforementioned", "neural", "networks", "require", "minimal", "assumptions", "about", "the", "task", "at", "hand", "and", "thus", "demonstrate", "significant", "generality", "—", "one", "single", "model", "can", "be", "applied", "to", "multiple", "applications", "in", "multiple", "languages", "without", "changing", "the", "architecture", ".", "A", "natural", "question", "is", "whether", "the", "representation", "learned", "from", "one", "task", "can", "be", "useful", "for", "another", "task", ".", "In", "other", "words", ",", "is", "there", "a", "way", "we", "can", "exploit", "the", "generality", "of", "neural", "networks", "to", "improve", "task", "performance", "by", "sharing", "model", "parameters", "and", "feature", "representations", "with", "another", "task", "?", "To", "address", "the", "above", "question", ",", "we", "study", "the", "transfer", "learning", "setting", ",", "which", "aims", "to", "improve", "the", "performance", "on", "a", "target", "task", "by", "joint", "training", "with", "a", "source", "task", ".", "We", "present", "a", "transfer", "learning", "approach", "based", "on", "a", "deep", "hierarchical", "recurrent", "neural", "network", ",", "which", "shares", "the", "hidden", "feature", "representation", "and", "part", "of", "the", "model", "parameters", "between", "the", "source", "task", "and", "the", "target", "task", ".", "Our", "approach", "combines", "the", "objectives", "of", "the", "two", "tasks", "and", "uses", "gradient", "-", "based", "methods", "for", "efficient", "training", ".", "We", "study", "cross", "-", "domain", ",", "cross", "-", "application", ",", "and", "cross", "-", "lingual", "transfer", ",", "and", "present", "a", "parameter", "-", "sharing", "architecture", "for", "each", "case", ".", "Experimental", "results", "show", "that", "our", "approach", "can", "significantly", "improve", "the", "performance", "of", "the", "target", "task", "when", "the", "the", "target", "task", "has", "few", "labels", "and", "is", "more", "related", "to", "the", "source", "task", ".", "Furthermore", ",", "we", "show", "that", "transfer", "learning", "can", "improve", "performance", "over", "state", "-", "of", "-", "the", "-", "art", "results", "even", "if", "the", "amount", "of", "labels", "is", "relatively", "abundant", ".", "We", "have", "novel", "contributions", "in", "two", "folds", ".", "First", ",", "our", "work", "is", ",", "to", "the", "best", "of", "our", "knowledge", ",", "the", "first", "that", "focuses", "on", "studying", "the", "transferability", "of", "different", "layers", "of", "representations", "for", "hierarchical", "RNNs", ".", "Second", ",", "different", "from", "previous", "transfer", "learning", "methods", "that", "usually", "focus", "on", "one", "specific", "transfer", "setting", ",", "our", "framework", "exploits", "different", "levels", "of", "representation", "sharing", "and", "provides", "a", "unified", "framework", "to", "handle", "cross", "-", "application", ",", "cross", "-", "lingual", ",", "and", "cross", "-", "domain", "transfer", ".", "section", ":", "Related", "Work", "There", "are", "two", "common", "paradigms", "for", "transfer", "learning", "for", "natural", "language", "processing", "(", "NLP", ")", "tasks", ",", "resource", "-", "based", "transfer", "and", "model", "-", "based", "transfer", ".", "Resource", "-", "based", "transfer", "utilizes", "additional", "linguistic", "annotations", "as", "weak", "supervision", "for", "transfer", "learning", ",", "such", "as", "cross", "-", "lingual", "dictionaries", "zirikly2cross", ",", "corpora", "wang2013cross", ",", "and", "word", "alignments", "yarowsky2001inducing", ".", "Resource", "-", "based", "methods", "demonstrate", "considerable", "success", "in", "cross", "-", "lingual", "transfer", ",", "but", "are", "quite", "sensitive", "to", "the", "scale", "and", "quality", "of", "the", "additional", "resources", ".", "Resource", "-", "based", "transfer", "is", "mostly", "limited", "to", "cross", "-", "lingual", "transfer", "in", "previous", "works", ",", "and", "there", "is", "not", "extensive", "research", "on", "extending", "resource", "-", "based", "methods", "to", "cross", "-", "domain", "and", "cross", "-", "application", "settings", ".", "Model", "-", "based", "transfer", ",", "on", "the", "other", "hand", ",", "does", "not", "require", "additional", "resources", ".", "Model", "-", "based", "transfer", "exploits", "the", "similarity", "and", "relatedness", "between", "the", "source", "task", "and", "the", "target", "task", "by", "adaptively", "modifying", "the", "model", "architectures", ",", "training", "algorithms", ",", "or", "feature", "representation", ".", "For", "example", ",", "proposed", "a", "transfer", "learning", "framework", "that", "shares", "structural", "parameters", "across", "multiple", "tasks", ",", "and", "improve", "the", "performance", "on", "various", "tasks", "including", "NER", ";", "presented", "a", "task", "-", "independent", "convolutional", "neural", "network", "and", "employed", "joint", "training", "to", "transfer", "knowledge", "from", "NER", "and", "POS", "tagging", "to", "chunking", ";", "studied", "transfer", "learning", "between", "named", "entity", "recognition", "and", "word", "segmentation", "in", "Chinese", "based", "on", "recurrent", "neural", "networks", ".", "Cross", "-", "domain", "transfer", ",", "or", "domain", "adaptation", ",", "is", "also", "a", "well", "-", "studied", "branch", "of", "model", "-", "based", "transfer", "in", "NLP", ".", "Techniques", "in", "cross", "-", "domain", "transfer", "include", "the", "design", "of", "robust", "feature", "representations", "schnabel2014flors", ",", "co", "-", "training", "chen2011co", ",", "hierarchical", "Bayesian", "prior", "finkel2009hierarchical", ",", "and", "canonical", "component", "analysis", "kim2015new", ".", "While", "our", "approach", "falls", "into", "the", "paradigm", "of", "model", "-", "based", "transfer", ",", "in", "contrast", "to", "the", "above", "methods", ",", "our", "method", "focuses", "on", "exploiting", "the", "generality", "of", "deep", "recurrent", "neural", "networks", "and", "is", "applicable", "to", "transfer", "between", "domains", ",", "applications", ",", "and", "languages", ".", "Our", "work", "builds", "on", "previous", "work", "on", "sequence", "tagging", "based", "on", "deep", "neural", "networks", ".", "develop", "end", "-", "to", "-", "end", "neural", "networks", "for", "sequence", "tagging", "without", "hand", "-", "engineered", "features", ".", "Later", "architectures", "based", "on", "different", "combinations", "of", "convolutional", "networks", "and", "recurrent", "networks", "have", "achieved", "state", "-", "of", "-", "the", "-", "art", "results", "on", "many", "tasks", "collobert2011natural", ",", "huang2015bidirectional", ",", "chiu2015named", ",", "lample2016neural", ",", "ma2016end", ".", "These", "models", "demonstrate", "significant", "generality", "since", "they", "can", "be", "applied", "to", "multiple", "applications", "in", "multiple", "languages", "with", "a", "unified", "network", "architecture", "and", "without", "task", "-", "specific", "feature", "extraction", ".", "section", ":", "Approach", "In", "this", "section", ",", "we", "introduce", "our", "transfer", "learning", "approach", ".", "We", "first", "introduce", "an", "abstract", "framework", "for", "neural", "sequence", "tagging", ",", "summarizing", "previous", "work", ",", "and", "then", "discuss", "three", "different", "transfer", "learning", "architectures", ".", "subsection", ":", "Base", "Model", "Though", "many", "different", "variants", "of", "neural", "networks", "have", "been", "proposed", "for", "the", "problem", "of", "sequence", "tagging", ",", "we", "find", "that", "most", "of", "the", "models", "can", "be", "described", "with", "the", "hierarchical", "framework", "illustrated", "in", "Figure", "[", "reference", "]", ".", "A", "character", "-", "level", "layer", "takes", "a", "sequence", "of", "characters", "(", "represented", "as", "embeddings", ")", "as", "input", ",", "and", "outputs", "a", "representation", "that", "encodes", "the", "morphological", "information", "at", "the", "character", "level", ".", "A", "word", "-", "level", "layer", "subsequently", "combines", "the", "character", "-", "level", "feature", "representation", "and", "a", "word", "embedding", ",", "and", "further", "incorporates", "the", "contextual", "information", "to", "output", "a", "new", "feature", "representation", ".", "After", "two", "levels", "of", "feature", "extraction", "(", "encoding", ")", ",", "the", "feature", "representation", "output", "by", "the", "word", "-", "level", "layer", "is", "fed", "to", "a", "conditional", "random", "field", "(", "CRF", ")", "layer", "that", "outputs", "the", "label", "sequence", ".", "Both", "of", "the", "word", "-", "level", "layer", "and", "the", "character", "-", "level", "layer", "can", "be", "implemented", "as", "convolutional", "neural", "networks", "(", "CNNs", ")", "or", "recurrent", "neural", "networks", "(", "RNNs", ")", "collobert2011natural", ",", "chiu2015named", ",", "lample2016neural", ",", "ma2016end", ".", "We", "discuss", "the", "details", "of", "the", "model", "we", "use", "in", "this", "work", "in", "Section", "[", "reference", "]", ".", "subsection", ":", "Transfer", "Learning", "Architectures", "We", "develop", "three", "architectures", "for", "transfer", "learning", ",", "T", "-", "A", ",", "T", "-", "B", ",", "and", "T", "-", "C", ",", "are", "illustrated", "in", "Figures", "[", "reference", "]", ",", "[", "reference", "]", ",", "and", "[", "reference", "]", "respectively", ".", "The", "three", "architectures", "are", "all", "extensions", "of", "the", "base", "model", "discussed", "in", "the", "previous", "section", "with", "different", "parameter", "sharing", "schemes", ".", "We", "now", "discuss", "the", "use", "cases", "for", "the", "different", "architectures", ".", "subsubsection", ":", "Cross", "-", "Domain", "Transfer", "Since", "different", "domains", "are", "“", "sub", "-", "languages", "”", "that", "have", "domain", "-", "specific", "regularities", ",", "sequence", "taggers", "trained", "on", "one", "domain", "might", "not", "have", "optimal", "performance", "on", "another", "domain", ".", "The", "goal", "of", "cross", "-", "domain", "transfer", "is", "to", "learn", "a", "sequence", "tagger", "that", "transfers", "knowledge", "from", "a", "source", "domain", "to", "a", "target", "domain", ".", "We", "assume", "that", "few", "labels", "are", "available", "in", "the", "target", "domain", ".", "There", "are", "two", "cases", "of", "cross", "-", "domain", "transfer", ".", "The", "two", "domains", "can", "have", "label", "sets", "that", "can", "be", "mapped", "to", "each", "other", ",", "or", "disparate", "label", "sets", ".", "For", "example", ",", "POS", "tags", "in", "the", "Genia", "biomedical", "corpus", "can", "be", "mapped", "to", "Penn", "Treebank", "tags", "barrett2014token", ",", "while", "some", "POS", "tags", "in", "Twitter", "(", "e.g.", ",", "“", "URL", "”", ")", "can", "not", "be", "mapped", "to", "Penn", "Treebank", "tags", "ritter2011named", ".", "If", "the", "two", "domains", "have", "mappable", "label", "sets", ",", "we", "share", "all", "the", "model", "parameters", "and", "feature", "representation", "in", "the", "neural", "networks", ",", "including", "the", "word", "and", "character", "embedding", ",", "the", "word", "-", "level", "layer", ",", "the", "character", "-", "level", "layer", ",", "and", "the", "CRF", "layer", ".", "We", "perform", "a", "label", "mapping", "step", "on", "top", "of", "the", "CRF", "layer", ".", "This", "becomes", "the", "model", "T", "-", "A", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "If", "the", "two", "domains", "have", "disparate", "label", "sets", ",", "we", "untie", "the", "parameter", "sharing", "in", "the", "CRF", "layer", "—", "i.e", ".", ",", "each", "task", "learns", "a", "separate", "CRF", "layer", ".", "This", "parameter", "sharing", "scheme", "reduces", "to", "model", "T", "-", "B", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "subsubsection", ":", "Cross", "-", "Application", "Transfer", "Sequence", "tagging", "has", "a", "couple", "of", "applications", "including", "POS", "tagging", ",", "chunking", ",", "and", "named", "entity", "recognition", ".", "Similar", "to", "the", "motivation", "in", "collobert2011natural", ",", "it", "is", "usually", "desirable", "to", "exploit", "the", "underlying", "similarities", "and", "regularities", "of", "different", "applications", ",", "and", "improve", "the", "performance", "of", "one", "application", "via", "joint", "training", "with", "another", ".", "Moreover", ",", "transfer", "between", "multiple", "applications", "can", "be", "helpful", "when", "the", "labels", "are", "limited", ".", "In", "the", "cross", "-", "application", "setting", ",", "we", "assume", "that", "multiple", "applications", "are", "in", "the", "same", "language", ".", "Since", "different", "applications", "share", "the", "same", "alphabet", ",", "the", "case", "is", "similar", "to", "cross", "-", "domain", "transfer", "with", "disparate", "label", "sets", ".", "We", "adopt", "the", "architecture", "of", "model", "T", "-", "B", "for", "cross", "-", "application", "transfer", "learning", "where", "only", "the", "CRF", "layers", "are", "disjoint", "for", "different", "applications", ".", "subsubsection", ":", "Cross", "-", "Lingual", "Transfer", "Though", "cross", "-", "lingual", "transfer", "is", "usually", "accomplished", "with", "additional", "multi", "-", "lingual", "resources", ",", "these", "methods", "are", "sensitive", "to", "the", "size", "and", "quality", "of", "the", "additional", "resources", "yarowsky2001inducing", ",", "wang2013cross", ".", "In", "this", "work", ",", "instead", ",", "we", "explore", "a", "complementary", "method", "that", "exploits", "the", "cross", "-", "lingual", "regularities", "purely", "on", "the", "model", "level", ".", "Our", "approach", "focuses", "on", "transfer", "learning", "between", "languages", "with", "similar", "alphabets", ",", "such", "as", "English", "and", "Spanish", ",", "since", "it", "is", "very", "difficult", "for", "transfer", "learning", "between", "languages", "with", "disparate", "alphabets", "(", "e.g.", ",", "English", "and", "Chinese", ")", "to", "work", "without", "additional", "resources", "zirikly2cross", ".", "Model", "-", "level", "transfer", "learning", "is", "achieved", "through", "exploiting", "the", "morphologies", "shared", "by", "the", "two", "languages", ".", "For", "example", ",", "“", "Canada", "”", "in", "English", "and", "“", "Canadá", "”", "in", "Spanish", "refer", "to", "the", "same", "named", "entity", ",", "and", "the", "morphological", "similarities", "can", "be", "leveraged", "for", "NER", "and", "also", "POS", "tagging", "with", "nouns", ".", "Thus", "we", "share", "the", "character", "embeddings", "and", "the", "character", "-", "level", "layer", "between", "different", "languages", "for", "transfer", "learning", ",", "which", "is", "illustrated", "as", "the", "model", "T", "-", "C", "in", "Figure", "[", "reference", "]", ".", "subsection", ":", "Training", "In", "the", "above", "sections", ",", "we", "introduced", "three", "neural", "architectures", "with", "different", "parameter", "sharing", "schemes", ",", "designed", "for", "different", "transfer", "learning", "settings", ".", "Now", "we", "describe", "how", "we", "train", "the", "neural", "networks", "jointly", "for", "two", "tasks", ".", "Suppose", "we", "are", "transferring", "from", "a", "source", "task", "to", "a", "target", "task", ",", "with", "the", "training", "instances", "being", "and", ".", "Let", "and", "denote", "the", "set", "of", "model", "parameters", "for", "the", "source", "and", "target", "tasks", "respectively", ".", "The", "model", "parameters", "are", "divided", "into", "two", "sets", ",", "task", "specific", "parameters", "and", "shared", "parameters", ",", "i.e.", ",", "where", "shared", "parameters", "are", "jointly", "optimized", "by", "the", "two", "tasks", ",", "while", "task", "specific", "parameters", "and", "are", "trained", "for", "each", "task", "separately", ".", "The", "training", "procedure", "is", "as", "follows", ".", "At", "each", "iteration", ",", "we", "sample", "a", "task", "(", "i.e.", ",", "either", "or", ")", "from", "based", "on", "a", "binomial", "distribution", "(", "the", "binomial", "probability", "is", "set", "as", "a", "hyperparameter", ")", ".", "Given", "the", "sampled", "task", ",", "we", "sample", "a", "batch", "of", "training", "instances", "from", "the", "given", "task", ",", "and", "then", "perform", "a", "gradient", "update", "according", "to", "the", "loss", "function", "of", "the", "given", "task", ".", "We", "update", "both", "the", "shared", "parameters", "and", "the", "task", "specific", "parameters", ".", "We", "repeat", "the", "above", "iterations", "until", "stopping", ".", "We", "adopt", "AdaGrad", "duchi2011adaptive", "to", "dynamically", "compute", "the", "learning", "rates", "for", "each", "iteration", ".", "Since", "the", "source", "and", "target", "tasks", "might", "have", "different", "convergence", "rates", ",", "we", "do", "early", "stopping", "on", "the", "target", "task", "performance", ".", "subsection", ":", "Model", "Implementation", "In", "this", "section", ",", "we", "describe", "our", "implementation", "of", "the", "base", "model", ".", "Both", "the", "character", "-", "level", "and", "word", "-", "level", "neural", "networks", "are", "implemented", "as", "RNNs", ".", "More", "specifically", ",", "we", "employ", "gated", "recurrent", "units", "(", "GRUs", ")", "cho2014properties", ".", "Let", "be", "a", "sequence", "of", "inputs", "that", "can", "be", "embeddings", "or", "hidden", "states", "of", "other", "layers", ".", "Let", "be", "the", "GRU", "hidden", "state", "at", "time", "step", ".", "Formally", ",", "a", "GRU", "unit", "at", "time", "step", "can", "be", "expressed", "as", "where", "’s", "are", "model", "parameters", "of", "each", "unit", ",", "is", "a", "candidate", "hidden", "state", "that", "is", "used", "to", "compute", ",", "is", "an", "element", "-", "wise", "sigmoid", "logistic", "function", "defined", "as", ",", "and", "denotes", "element", "-", "wise", "multiplication", "of", "two", "vectors", ".", "Intuitively", ",", "the", "update", "gate", "controls", "how", "much", "the", "unit", "updates", "its", "hidden", "state", ",", "and", "the", "reset", "gate", "determines", "how", "much", "information", "from", "the", "previous", "hidden", "state", "needs", "to", "be", "reset", ".", "The", "input", "to", "the", "character", "-", "level", "GRUs", "is", "character", "embeddings", ",", "while", "the", "input", "to", "the", "word", "-", "level", "GRUs", "is", "the", "concatenation", "of", "character", "-", "level", "GRU", "hidden", "states", "and", "word", "embeddings", ".", "Both", "GRUs", "are", "bi", "-", "directional", "and", "have", "two", "layers", ".", "Given", "an", "input", "sequence", "of", "words", ",", "the", "word", "-", "level", "GRUs", "and", "the", "character", "-", "level", "GRUs", "together", "learn", "a", "feature", "representation", "for", "the", "-", "th", "word", "in", "the", "sequence", ",", "which", "forms", "a", "sequence", ".", "Let", "denote", "the", "tag", "sequence", ".", "Given", "the", "feature", "representation", "and", "the", "tag", "sequence", "for", "each", "training", "instance", ",", "the", "CRF", "layer", "defines", "the", "objective", "function", "to", "maximize", "based", "on", "a", "max", "-", "margin", "principle", "gimpel2010softmax", "as", ":", "where", "is", "a", "function", "that", "assigns", "a", "score", "for", "each", "pair", "of", "and", ",", "and", "denotes", "the", "space", "of", "tag", "sequences", "for", ".", "The", "cost", "function", "is", "added", "based", "on", "the", "max", "-", "margin", "principle", "gimpel2010softmax", "that", "high", "-", "cost", "tags", "should", "be", "penalized", "more", "heavily", ".", "Our", "base", "model", "is", "similar", "to", ",", "but", "in", "contrast", "to", "their", "model", ",", "we", "employ", "GRUs", "for", "the", "character", "-", "level", "and", "word", "-", "level", "networks", "instead", "of", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "units", ",", "and", "define", "the", "objective", "function", "based", "on", "the", "max", "-", "margin", "principle", ".", "We", "note", "that", "our", "transfer", "learning", "framework", "does", "not", "make", "assumptions", "about", "specific", "model", "implementation", ",", "and", "could", "be", "applied", "to", "other", "neural", "architectures", "collobert2011natural", ",", "chiu2015named", ",", "lample2016neural", ",", "ma2016end", "as", "well", ".", "section", ":", "Experiments", "subsection", ":", "Datasets", "We", "use", "the", "following", "benchmark", "datasets", "in", "our", "experiments", ":", "Penn", "Treebank", "(", "PTB", ")", "POS", "tagging", ",", "CoNLL", "2000", "chunking", ",", "CoNLL", "2003", "English", "NER", ",", "CoNLL", "2002", "Dutch", "NER", ",", "CoNLL", "2002", "Spanish", "NER", ",", "the", "Genia", "biomedical", "corpus", "kim2003genia", ",", "and", "a", "Twitter", "corpus", "ritter2011named", ".", "The", "statistics", "of", "the", "datasets", "are", "described", "in", "Table", "[", "reference", "]", ".", "We", "construct", "the", "POS", "tagging", "dataset", "with", "the", "instructions", "described", "in", ".", "Note", "that", "as", "a", "standard", "practice", ",", "the", "POS", "tags", "are", "extracted", "from", "the", "parsed", "trees", ".", "For", "the", "CoNLL", "2003", "English", "NER", "dataset", ",", "we", "follow", "previous", "works", "collobert2011natural", "to", "append", "one", "-", "hot", "gazetteer", "features", "to", "the", "input", "of", "the", "CRF", "layer", "for", "fair", "comparison", ".", "Since", "there", "is", "no", "standard", "training", "/", "dev", "/", "test", "data", "split", "for", "the", "Genia", "and", "Twitter", "corpora", ",", "we", "randomly", "sample", "10", "%", "for", "test", ",", "10", "%", "for", "development", ",", "and", "80", "%", "for", "training", ".", "We", "follow", "previous", "work", "barrett2014token", "to", "map", "Genia", "POS", "tags", "to", "PTB", "POS", "tags", ".", "subsection", ":", "Transfer", "Learning", "Performance", "We", "evaluate", "our", "transfer", "learning", "approach", "on", "the", "above", "datasets", ".", "We", "fix", "the", "hyperparameters", "for", "all", "the", "results", "reported", "in", "this", "section", ":", "we", "set", "the", "character", "embedding", "dimension", "at", ",", "the", "word", "embedding", "dimension", "at", "for", "English", "and", "for", "Spanish", ",", "the", "dimension", "of", "hidden", "states", "of", "the", "character", "-", "level", "GRUs", "at", ",", "the", "dimension", "of", "hidden", "states", "of", "the", "word", "-", "level", "GRUs", "at", ",", "and", "the", "initial", "learning", "rate", "at", ".", "Except", "for", "the", "Twitter", "datasets", ",", "these", "datasets", "are", "fairly", "large", ".", "To", "simulate", "a", "low", "-", "resource", "setting", ",", "we", "also", "use", "random", "subsets", "of", "the", "data", ".", "We", "vary", "the", "labeling", "rate", "of", "the", "target", "task", "at", ",", ",", "and", ".", "Given", "a", "labeling", "rate", ",", "we", "randomly", "sample", "a", "ratio", "of", "the", "sentences", "from", "the", "training", "set", "and", "discard", "the", "rest", "of", "the", "training", "data", "—", "e.g", ".", ",", "a", "labeling", "rate", "of", "results", "in", "around", "900", "training", "tokens", "on", "PTB", "POS", "tagging", "(", "Cf", ".", "Table", "[", "reference", "]", ")", ".", "The", "results", "on", "transfer", "learning", "are", "plotted", "in", "Figure", "[", "reference", "]", ",", "where", "we", "compare", "the", "results", "with", "and", "without", "transfer", "learning", "under", "various", "labeling", "rates", ".", "The", "numbers", "in", "the", "y", "-", "axes", "are", "accuracies", "for", "POS", "tagging", ",", "and", "chunk", "-", "level", "F1", "scores", "for", "chunking", "and", "NER", ".", "The", "numbers", "are", "shown", "in", "Table", "[", "reference", "]", ".", "We", "can", "see", "that", "our", "transfer", "learning", "approach", "consistently", "improved", "over", "the", "non", "-", "transfer", "results", ".", "We", "also", "observe", "that", "the", "improvement", "by", "transfer", "learning", "is", "more", "substantial", "when", "the", "labeling", "rate", "is", "lower", ".", "For", "cross", "-", "domain", "transfer", ",", "we", "obtained", "substantial", "improvement", "on", "the", "Genia", "and", "Twitter", "corpora", "by", "transferring", "the", "knowledge", "from", "PTB", "POS", "tagging", "and", "CoNLL", "2003", "NER", ".", "For", "example", ",", "as", "shown", "in", "Figure", "[", "reference", "]", ",", "we", "can", "obtain", "an", "tagging", "accuracy", "of", "with", "zero", "labels", "and", "with", "only", "labels", "when", "transferring", "from", "PTB", "to", "Genia", ".", "As", "shown", "in", "Figures", "[", "reference", "]", "and", "[", "reference", "]", ",", "our", "transfer", "learning", "approach", "can", "improve", "the", "performance", "on", "Twitter", "POS", "tagging", "and", "NER", "for", "all", "labeling", "rates", ",", "and", "the", "improvements", "with", "labels", "are", "more", "than", "for", "both", "datasets", ".", "Cross", "-", "application", "transfer", "also", "leads", "to", "substantial", "improvement", "under", "low", "-", "resource", "conditions", ".", "For", "example", ",", "as", "shown", "in", "Figures", "[", "reference", "]", "and", "[", "reference", "]", ",", "the", "improvements", "with", "labels", "are", "and", "on", "CoNLL", "2000", "chunking", "and", "CoNLL", "2003", "NER", "respectively", "when", "transferring", "from", "PTB", "POS", "tagging", ".", "Figures", "[", "reference", "]", "and", "[", "reference", "]", "show", "that", "cross", "-", "lingual", "transfer", "can", "improve", "the", "performance", "when", "few", "labels", "are", "available", ".", "Figure", "[", "reference", "]", "further", "shows", "that", "the", "improvements", "by", "different", "architectures", "are", "in", "the", "following", "order", ":", "T", "-", "A", "T", "-", "B", "T", "-", "C.", "This", "phenomenon", "can", "be", "explained", "by", "the", "fact", "that", "T", "-", "A", "shares", "the", "most", "model", "parameters", "while", "T", "-", "C", "shares", "the", "least", ".", "Transfer", "settings", "like", "cross", "-", "lingual", "transfer", "can", "only", "use", "T", "-", "C", "because", "the", "underlying", "similarities", "between", "the", "source", "task", "and", "the", "target", "task", "are", "less", "prominent", "(", "i.e.", ",", "less", "transferable", ")", ",", "and", "in", "those", "cases", "the", "improvement", "by", "transfer", "learning", "is", "less", "substantial", ".", "Another", "interesting", "comparison", "is", "among", "Figures", "[", "reference", "]", ",", "[", "reference", "]", ",", "and", "[", "reference", "]", ".", "Figure", "[", "reference", "]", "is", "cross", "-", "domain", "transfer", ",", "Figure", "[", "reference", "]", "is", "transfer", "across", "domains", "and", "applications", "at", "the", "same", "time", ",", "and", "Figure", "[", "reference", "]", "combines", "all", "the", "three", "transfer", "settings", "(", "i.e.", ",", "from", "Spanish", "NER", "in", "the", "general", "domain", "to", "English", "POS", "tagging", "in", "the", "biomedical", "domain", ")", ".", "The", "results", "show", "that", "the", "improvement", "by", "transfer", "learning", "diminishes", "when", "the", "transfer", "becomes", "“", "indirect", "”", "(", "i.e.", ",", "the", "source", "task", "and", "the", "target", "task", "are", "more", "loosely", "related", ")", ".", "We", "also", "study", "using", "different", "transfer", "learning", "models", "for", "the", "same", "task", ".", "We", "study", "the", "effects", "of", "using", "T", "-", "A", ",", "T", "-", "B", ",", "and", "T", "-", "C", "when", "transferring", "from", "PTB", "to", "Genia", ",", "and", "the", "results", "are", "included", "in", "the", "lower", "part", "of", "Table", "[", "reference", "]", ".", "We", "observe", "that", "the", "performance", "gain", "decreases", "when", "less", "parameters", "are", "shared", "(", "i.e.", ",", "T", "-", "A", "T", "-", "B", "T", "-", "C", ")", ".", "subsection", ":", "Comparison", "with", "State", "-", "of", "-", "the", "-", "Art", "Results", "In", "the", "above", "section", ",", "we", "examine", "the", "effects", "of", "different", "transfer", "learning", "architectures", ".", "Now", "we", "compare", "our", "approach", "with", "state", "-", "of", "-", "the", "-", "art", "systems", "on", "these", "datasets", ".", "We", "use", "publicly", "available", "pretrained", "word", "embeddings", "as", "initialization", ".", "On", "the", "English", "datasets", ",", "following", "previous", "works", "that", "are", "based", "on", "neural", "networks", "collobert2011natural", ",", "huang2015bidirectional", ",", "chiu2015named", ",", "ma2016end", ",", "we", "experiment", "with", "both", "the", "50", "-", "dimensional", "SENNA", "embeddings", "collobert2011natural", "and", "the", "100", "-", "dimensional", "GloVe", "embeddings", "pennington2014glove", "and", "use", "the", "development", "set", "to", "choose", "the", "embeddings", "for", "different", "tasks", "and", "settings", ".", "For", "Spanish", "and", "Dutch", ",", "we", "use", "the", "64", "-", "dimensional", "Polyglot", "embeddings", "al2013polyglot", ".", "We", "set", "the", "hidden", "state", "dimensions", "to", "be", "300", "for", "the", "word", "-", "level", "GRU", ".", "The", "initial", "learning", "rate", "for", "AdaGrad", "is", "fixed", "at", "0.01", ".", "We", "use", "the", "development", "set", "to", "tune", "the", "other", "hyperparameters", "of", "our", "model", ".", "Our", "results", "are", "reported", "in", "Table", "[", "reference", "]", ".", "Since", "there", "are", "no", "standard", "data", "splits", "on", "the", "Genia", "and", "Twitter", "corpora", ",", "we", "do", "not", "include", "these", "datasets", "into", "our", "comparison", ".", "The", "results", "for", "CoNLL", "2000", "chunking", ",", "CoNLL", "2003", "NER", ",", "and", "PTB", "POS", "tagging", "are", "obtained", "by", "transfer", "learning", "between", "the", "three", "tasks", ",", "i.e.", ",", "transferring", "from", "two", "tasks", "to", "the", "other", ".", "The", "results", "for", "Spanish", "and", "Dutch", "NER", "are", "obtained", "with", "transfer", "learning", "between", "the", "NER", "datasets", "in", "three", "languages", "(", "English", ",", "Spanish", ",", "and", "Dutch", ")", ".", "From", "Table", "[", "reference", "]", ",", "we", "can", "draw", "two", "conclusions", ".", "First", ",", "our", "transfer", "learning", "approach", "achieves", "new", "state", "-", "of", "-", "the", "-", "art", "results", "on", "all", "the", "considered", "benchmark", "datasets", "except", "PTB", "POS", "tagging", ",", "which", "indicates", "that", "transfer", "learning", "can", "still", "improve", "the", "performance", "even", "on", "datasets", "with", "relatively", "abundant", "labels", ".", "Second", ",", "our", "base", "model", "(", "w", "/", "o", "transfer", ")", "performs", "competitively", "compared", "to", "the", "state", "-", "of", "-", "the", "-", "art", "systems", ",", "which", "means", "that", "the", "improvements", "shown", "in", "Section", "[", "reference", "]", "are", "obtained", "over", "a", "strong", "baseline", ".", "section", ":", "Conclusion", "In", "this", "paper", "we", "develop", "a", "transfer", "learning", "approach", "for", "sequence", "tagging", ",", "which", "exploits", "the", "generality", "demonstrated", "by", "deep", "neural", "networks", "in", "previous", "work", ".", "We", "design", "three", "neural", "network", "architectures", "for", "the", "settings", "of", "cross", "-", "domain", ",", "cross", "-", "application", ",", "and", "cross", "-", "lingual", "transfer", ".", "Our", "transfer", "learning", "approach", "achieves", "significant", "improvement", "on", "various", "datasets", "under", "low", "-", "resource", "conditions", ",", "as", "well", "as", "new", "state", "-", "of", "-", "the", "-", "art", "results", "on", "some", "of", "the", "benchmarks", ".", "With", "thorough", "experiments", ",", "we", "observe", "that", "the", "following", "factors", "are", "crucial", "for", "the", "performance", "of", "our", "transfer", "learning", "approach", ":", "a", ")", "label", "abundance", "for", "the", "target", "task", ",", "b", ")", "relatedness", "between", "the", "source", "and", "target", "tasks", ",", "and", "c", ")", "the", "number", "of", "parameters", "that", "can", "be", "shared", ".", "In", "the", "future", ",", "it", "will", "be", "interesting", "to", "combine", "model", "-", "based", "transfer", "(", "as", "in", "this", "work", ")", "with", "resource", "-", "based", "transfer", "for", "cross", "-", "lingual", "transfer", "learning", ".", "subsubsection", ":", "Acknowledgments", "This", "work", "was", "funded", "by", "NVIDIA", ",", "the", "Office", "of", "Naval", "Research", "grant", "N000141512791", ",", "the", "ADeLAIDE", "grant", "FA8750", "-", "16C", "-", "0130", "-", "001", ",", "the", "NSF", "grant", "IIS1250956", ",", "and", "Google", "Research", ".", "bibliography", ":", "References"], "entity": [[2, 4, ["Transfer", "Learning"], "Task"], [5, 7, ["Sequence", "Tagging"], "Task"], [8, 11, ["Hierarchical", "Recurrent", "Networks"], "Method"], [16, 18, ["neural", "networks"], "Method"], [30, 33, ["sequence", "tagging", "tasks"], "Task"], [52, 54, ["unified", "architecture"], "Method"], [56, 61, ["task", "-", "specific", "feature", "engineering"], "Method"], [90, 92, ["transfer", "learning"], "Method"], [93, 96, ["neural", "sequence", "taggers"], "Task"], [107, 109, ["POS", "tagging"], "Task"], [110, 112, ["Penn", "Treebank"], "Material"], [129, 131, ["POS", "tagging"], "Task"], [140, 142, ["transfer", "learning"], "Method"], [143, 147, ["deep", "hierarchical", "recurrent", "networks"], "Method"], [190, 192, ["Sequence", "tagging"], "Task"], [197, 200, ["natural", "language", "processing"], "Task"], [206, 211, ["part", "-", "of", "-", "speech"], "Task"], [212, 213, ["POS"], "Task"], [216, 218, ["text", "chunking"], "Task"], [220, 223, ["named", "entity", "recognition"], "Task"], [224, 225, ["NER"], "Task"], [233, 235, ["sequence", "tagging"], "Method"], [254, 256, ["sequence", "tagging"], "Task"], [273, 275, ["transfer", "learning"], "Method"], [277, 279, ["Transfer", "learning"], "Method"], [315, 317, ["transfer", "learning"], "Method"], [339, 343, ["multi", "-", "task", "transfer"], "Task"], [374, 376, ["sequence", "tagging"], "Task"], [393, 395, ["neural", "networks"], "Method"], [420, 422, ["feature", "extraction"], "Task"], [425, 427, ["neural", "networks"], "Method"], [490, 492, ["neural", "networks"], "Method"], [498, 500, ["model", "parameters"], "Method"], [501, 503, ["feature", "representations"], "Method"], [531, 533, ["joint", "training"], "Method"], [547, 552, ["deep", "hierarchical", "recurrent", "neural", "network"], "Method"], [556, 559, ["hidden", "feature", "representation"], "Method"], [585, 589, ["gradient", "-", "based", "methods"], "Method"], [591, 592, ["training"], "Task"], [612, 616, ["parameter", "-", "sharing", "architecture"], "Method"], [657, 659, ["transfer", "learning"], "Method"], [716, 718, ["hierarchical", "RNNs"], "Task"], [742, 744, ["representation", "sharing"], "Method"], [751, 754, ["cross", "-", "application"], "Task"], [755, 758, ["cross", "-", "lingual"], "Task"], [760, 764, ["cross", "-", "domain", "transfer"], "Task"], [775, 777, ["transfer", "learning"], "Method"], [778, 781, ["natural", "language", "processing"], "Task"], [782, 783, ["NLP"], "Task"], [786, 790, ["resource", "-", "based", "transfer"], "Task"], [791, 795, ["model", "-", "based", "transfer"], "Method"], [796, 800, ["Resource", "-", "based", "transfer"], "Method"], [808, 810, ["transfer", "learning"], "Method"], [827, 831, ["Resource", "-", "based", "methods"], "Method"], [854, 858, ["Resource", "-", "based", "transfer"], "Method"], [878, 882, ["resource", "-", "based", "methods"], "Method"], [883, 891, ["cross", "-", "domain", "and", "cross", "-", "application", "settings"], "Task"], [892, 896, ["Model", "-", "based", "transfer"], "Method"], [908, 912, ["Model", "-", "based", "transfer"], "Method"], [929, 931, ["model", "architectures"], "Method"], [932, 934, ["training", "algorithms"], "Method"], [936, 938, ["feature", "representation"], "Method"], [963, 964, ["NER"], "Task"], [967, 973, ["task", "-", "independent", "convolutional", "neural", "network"], "Method"], [975, 977, ["joint", "training"], "Method"], [981, 982, ["NER"], "Task"], [983, 985, ["POS", "tagging"], "Task"], [986, 987, ["chunking"], "Task"], [989, 991, ["transfer", "learning"], "Method"], [992, 995, ["named", "entity", "recognition"], "Task"], [996, 998, ["word", "segmentation"], "Task"], [1006, 1010, ["Cross", "-", "domain", "transfer"], "Task"], [1012, 1014, ["domain", "adaptation"], "Method"], [1023, 1027, ["model", "-", "based", "transfer"], "Method"], [1028, 1029, ["NLP"], "Task"], [1032, 1036, ["cross", "-", "domain", "transfer"], "Task"], [1040, 1043, ["robust", "feature", "representations"], "Method"], [1045, 1048, ["co", "-", "training"], "Method"], [1050, 1053, ["hierarchical", "Bayesian", "prior"], "Method"], [1056, 1059, ["canonical", "component", "analysis"], "Method"], [1069, 1073, ["model", "-", "based", "transfer"], "Method"], [1113, 1115, ["sequence", "tagging"], "Task"], [1130, 1132, ["sequence", "tagging"], "Task"], [1145, 1147, ["convolutional", "networks"], "Method"], [1148, 1150, ["recurrent", "networks"], "Method"], [1191, 1194, ["unified", "network", "architecture"], "Method"], [1196, 1201, ["task", "-", "specific", "feature", "extraction"], "Method"], [1223, 1226, ["neural", "sequence", "tagging"], "Task"], [1249, 1251, ["neural", "networks"], "Method"], [1258, 1260, ["sequence", "tagging"], "Task"], [1273, 1275, ["hierarchical", "framework"], "Method"], [1283, 1287, ["character", "-", "level", "layer"], "Method"], [1315, 1319, ["word", "-", "level", "layer"], "Method"], [1322, 1327, ["character", "-", "level", "feature", "representation"], "Method"], [1329, 1331, ["word", "embedding"], "Method"], [1342, 1344, ["feature", "representation"], "Method"], [1349, 1351, ["feature", "extraction"], "Task"], [1352, 1353, ["encoding"], "Task"], [1356, 1358, ["feature", "representation"], "Method"], [1361, 1365, ["word", "-", "level", "layer"], "Method"], [1369, 1372, ["conditional", "random", "field"], "Method"], [1373, 1374, ["CRF"], "Method"], [1385, 1389, ["word", "-", "level", "layer"], "Method"], [1391, 1395, ["character", "-", "level", "layer"], "Method"], [1403, 1404, ["CNNs"], "Method"], [1410, 1411, ["RNNs"], "Method"], [1440, 1443, ["Transfer", "Learning", "Architectures"], "Method"], [1448, 1450, ["transfer", "learning"], "Method"], [1451, 1454, ["T", "-", "A"], "Method"], [1455, 1458, ["T", "-", "B"], "Method"], [1460, 1463, ["T", "-", "C"], "Method"], [1490, 1492, ["base", "model"], "Method"], [1499, 1502, ["parameter", "sharing", "schemes"], "Method"], [1516, 1520, ["Cross", "-", "Domain", "Transfer"], "Method"], [1536, 1538, ["sequence", "taggers"], "Method"], [1554, 1558, ["cross", "-", "domain", "transfer"], "Task"], [1562, 1564, ["sequence", "tagger"], "Method"], [1593, 1597, ["cross", "-", "domain", "transfer"], "Task"], [1621, 1623, ["POS", "tags"], "Task"], [1632, 1635, ["Penn", "Treebank", "tags"], "Task"], [1639, 1641, ["POS", "tags"], "Task"], [1655, 1658, ["Penn", "Treebank", "tags"], "Task"], [1676, 1678, ["feature", "representation"], "Method"], [1680, 1682, ["neural", "networks"], "Method"], [1685, 1689, ["word", "and", "character", "embedding"], "Method"], [1691, 1695, ["word", "-", "level", "layer"], "Method"], [1697, 1701, ["character", "-", "level", "layer"], "Method"], [1710, 1713, ["label", "mapping", "step"], "Method"], [1766, 1769, ["parameter", "sharing", "scheme"], "Method"], [1785, 1789, ["Cross", "-", "Application", "Transfer"], "Method"], [1789, 1791, ["Sequence", "tagging"], "Task"], [1797, 1799, ["POS", "tagging"], "Task"], [1800, 1801, ["chunking"], "Task"], [1803, 1806, ["named", "entity", "recognition"], "Task"], [1837, 1839, ["joint", "training"], "Method"], [1859, 1863, ["cross", "-", "application", "setting"], "Task"], [1888, 1892, ["cross", "-", "domain", "transfer"], "Task"], [1925, 1929, ["Cross", "-", "Lingual", "Transfer"], "Method"], [1989, 1991, ["transfer", "learning"], "Method"], [2009, 2011, ["transfer", "learning"], "Method"], [2076, 2077, ["NER"], "Task"], [2079, 2081, ["POS", "tagging"], "Task"], [2100, 2102, ["transfer", "learning"], "Method"], [2129, 2131, ["neural", "architectures"], "Method"], [2133, 2136, ["parameter", "sharing", "schemes"], "Method"], [2151, 2153, ["neural", "networks"], "Method"], [2294, 2296, ["gradient", "update"], "Method"], [2328, 2330, ["AdaGrad", "duchi2011adaptive"], "Method"], [2349, 2351, ["convergence", "rates"], "Metric"], [2376, 2378, ["base", "model"], "Method"], [2393, 2394, ["RNNs"], "Method"], [2400, 2403, ["gated", "recurrent", "units"], "Method"], [2404, 2405, ["GRUs"], "Method"], [2438, 2440, ["GRU", "unit"], "Method"], [2469, 2475, ["element", "-", "wise", "sigmoid", "logistic", "function"], "Method"], [2553, 2555, ["word", "embeddings"], "Method"], [2557, 2558, ["GRUs"], "Method"], [2588, 2590, ["feature", "representation"], "Method"], [2612, 2614, ["feature", "representation"], "Method"], [2628, 2630, ["objective", "function"], "Metric"], [2635, 2639, ["max", "-", "margin", "principle"], "Method"], [2673, 2677, ["max", "-", "margin", "principle"], "Method"], [2705, 2706, ["GRUs"], "Method"], [2718, 2723, ["Long", "Short", "-", "Term", "Memory"], "Method"], [2724, 2725, ["LSTM"], "Method"], [2736, 2740, ["max", "-", "margin", "principle"], "Method"], [2754, 2756, ["model", "implementation"], "Method"], [2763, 2765, ["neural", "architectures"], "Method"], [2791, 2793, ["Penn", "Treebank"], "Material"], [2794, 2795, ["PTB"], "Material"], [2796, 2798, ["POS", "tagging"], "Task"], [2803, 2807, ["CoNLL", "2003", "English", "NER"], "Material"], [2808, 2812, ["CoNLL", "2002", "Dutch", "NER"], "Material"], [2813, 2817, ["CoNLL", "2002", "Spanish", "NER"], "Material"], [2819, 2822, ["Genia", "biomedical", "corpus"], "Material"], [2826, 2828, ["Twitter", "corpus"], "Material"], [2846, 2848, ["POS", "tagging"], "Task"], [2917, 2918, ["Genia"], "Material"], [2919, 2920, ["Twitter"], "Material"], [2933, 2934, ["development"], "Task"], [2948, 2949, ["Genia"], "Material"], [2949, 2951, ["POS", "tags"], "Task"], [2952, 2953, ["PTB"], "Material"], [2953, 2955, ["POS", "tags"], "Task"], [2958, 2960, ["Transfer", "Learning"], "Task"], [3033, 3035, ["learning", "rate"], "Metric"], [3052, 3056, ["low", "-", "resource", "setting"], "Task"], [3069, 3071, ["labeling", "rate"], "Metric"], [3082, 3084, ["labeling", "rate"], "Metric"], [3110, 3112, ["labeling", "rate"], "Metric"], [3120, 3123, ["PTB", "POS", "tagging"], "Task"], [3135, 3137, ["transfer", "learning"], "Method"], [3153, 3155, ["transfer", "learning"], "Method"], [3157, 3159, ["labeling", "rates"], "Metric"], [3168, 3169, ["accuracies"], "Metric"], [3170, 3172, ["POS", "tagging"], "Task"], [3174, 3179, ["chunk", "-", "level", "F1", "scores"], "Metric"], [3180, 3181, ["chunking"], "Task"], [3182, 3183, ["NER"], "Task"], [3218, 3220, ["transfer", "learning"], "Method"], [3225, 3227, ["labeling", "rate"], "Metric"], [3231, 3235, ["cross", "-", "domain", "transfer"], "Task"], [3251, 3254, ["PTB", "POS", "tagging"], "Task"], [3255, 3258, ["CoNLL", "2003", "NER"], "Task"], [3274, 3276, ["tagging", "accuracy"], "Metric"], [3312, 3315, ["Twitter", "POS", "tagging"], "Task"], [3316, 3317, ["NER"], "Task"], [3319, 3321, ["labeling", "rates"], "Metric"], [3334, 3338, ["Cross", "-", "application", "transfer"], "Method"], [3382, 3385, ["PTB", "POS", "tagging"], "Task"], [3462, 3464, ["Transfer", "settings"], "Method"], [3504, 3506, ["transfer", "learning"], "Method"], [3534, 3538, ["cross", "-", "domain", "transfer"], "Task"], [3569, 3571, ["Spanish", "NER"], "Task"], [3576, 3579, ["English", "POS", "tagging"], "Task"], [3592, 3594, ["transfer", "learning"], "Method"], [3754, 3756, ["English", "datasets"], "Material"], [3764, 3766, ["neural", "networks"], "Method"], [3779, 3784, ["50", "-", "dimensional", "SENNA", "embeddings"], "Method"], [3787, 3792, ["100", "-", "dimensional", "GloVe", "embeddings"], "Method"], [3816, 3820, ["64", "-", "dimensional", "Polyglot"], "Method"], [3834, 3838, ["word", "-", "level", "GRU"], "Method"], [3841, 3843, ["learning", "rate"], "Metric"], [3844, 3845, ["AdaGrad"], "Method"], [3905, 3908, ["CoNLL", "2003", "NER"], "Task"], [3910, 3913, ["PTB", "POS", "tagging"], "Task"], [3916, 3918, ["transfer", "learning"], "Method"], [3936, 3937, ["Spanish"], "Task"], [3938, 3940, ["Dutch", "NER"], "Task"], [3943, 3945, ["transfer", "learning"], "Method"], [3947, 3948, ["NER"], "Task"], [3953, 3954, ["English"], "Material"], [3996, 3999, ["PTB", "POS", "tagging"], "Task"], [4003, 4005, ["transfer", "learning"], "Method"], [4024, 4028, ["w", "/", "o", "transfer"], "Method"], [4074, 4076, ["sequence", "tagging"], "Task"], [4093, 4096, ["neural", "network", "architectures"], "Method"], [4100, 4107, ["cross", "-", "domain", ",", "cross", "-", "application"], "Task"], [4210, 4214, ["model", "-", "based", "transfer"], "Method"], [4221, 4225, ["resource", "-", "based", "transfer"], "Method"], [247, 248, ["POS"], "Task"], [306, 307, ["Twitter"], "Material"], [367, 369, ["neural", "networks"], "Method"], [516, 518, ["transfer", "learning"], "Method"], [541, 543, ["transfer", "learning"], "Method"], [604, 607, ["cross", "-", "lingual"], "Task"], [724, 726, ["transfer", "learning"], "Method"], [813, 816, ["cross", "-", "lingual"], "Task"], [835, 838, ["cross", "-", "lingual"], "Task"], [862, 865, ["cross", "-", "lingual"], "Task"], [944, 946, ["transfer", "learning"], "Method"], [1003, 1005, ["neural", "networks"], "Method"], [1091, 1093, ["neural", "networks"], "Method"], [1118, 1120, ["neural", "networks"], "Method"], [1127, 1129, ["neural", "networks"], "Method"], [1212, 1214, ["transfer", "learning"], "Method"], [1236, 1238, ["transfer", "learning"], "Method"], [1400, 1402, ["neural", "networks"], "Method"], [1407, 1409, ["neural", "networks"], "Method"], [1625, 1626, ["Genia"], "Material"], [1642, 1643, ["Twitter"], "Material"], [1704, 1705, ["CRF"], "Method"], [1717, 1718, ["CRF"], "Method"], [1724, 1727, ["T", "-", "A"], "Method"], [1751, 1752, ["CRF"], "Method"], [1762, 1763, ["CRF"], "Method"], [1772, 1775, ["T", "-", "B"], "Method"], [1903, 1906, ["T", "-", "B"], "Method"], [1910, 1912, ["transfer", "learning"], "Method"], [1915, 1916, ["CRF"], "Method"], [1930, 1933, ["cross", "-", "lingual"], "Task"], [1975, 1978, ["cross", "-", "lingual"], "Task"], [2001, 2002, ["Spanish"], "Task"], [2033, 2035, ["transfer", "learning"], "Method"], [2060, 2061, ["Spanish"], "Task"], [2109, 2112, ["T", "-", "C"], "Method"], [2140, 2142, ["transfer", "learning"], "Method"], [2388, 2390, ["neural", "networks"], "Method"], [2528, 2529, ["GRUs"], "Method"], [2541, 2542, ["GRUs"], "Method"], [2578, 2579, ["GRUs"], "Method"], [2584, 2585, ["GRUs"], "Method"], [2624, 2625, ["CRF"], "Method"], [2745, 2747, ["transfer", "learning"], "Method"], [2863, 2864, ["POS"], "Task"], [2877, 2878, ["NER"], "Task"], [2897, 2898, ["CRF"], "Method"], [2964, 2966, ["transfer", "learning"], "Method"], [3002, 3003, ["Spanish"], "Task"], [3014, 3015, ["GRUs"], "Method"], [3027, 3028, ["GRUs"], "Method"], [3040, 3041, ["Twitter"], "Material"], [3199, 3201, ["transfer", "learning"], "Method"], [3242, 3243, ["Genia"], "Material"], [3244, 3245, ["Twitter"], "Material"], [3287, 3288, ["PTB"], "Material"], [3289, 3290, ["Genia"], "Material"], [3304, 3306, ["transfer", "learning"], "Method"], [3377, 3378, ["NER"], "Task"], [3396, 3399, ["cross", "-", "lingual"], "Task"], [3428, 3431, ["T", "-", "A"], "Method"], [3431, 3434, ["T", "-", "B"], "Method"], [3446, 3449, ["T", "-", "A"], "Method"], [3455, 3458, ["T", "-", "C"], "Method"], [3465, 3468, ["cross", "-", "lingual"], "Task"], [3472, 3475, ["T", "-", "C"], "Method"], [3623, 3625, ["transfer", "learning"], "Method"], [3637, 3640, ["T", "-", "A"], "Method"], [3641, 3644, ["T", "-", "B"], "Method"], [3646, 3649, ["T", "-", "C"], "Method"], [3652, 3653, ["PTB"], "Material"], [3654, 3655, ["Genia"], "Material"], [3686, 3689, ["T", "-", "A"], "Method"], [3689, 3692, ["T", "-", "B"], "Method"], [3692, 3695, ["T", "-", "C"], "Method"], [3720, 3722, ["transfer", "learning"], "Method"], [3809, 3810, ["Spanish"], "Task"], [3883, 3884, ["Genia"], "Material"], [3885, 3886, ["Twitter"], "Material"], [3955, 3956, ["Spanish"], "Task"], [3976, 3978, ["transfer", "learning"], "Method"], [4070, 4072, ["transfer", "learning"], "Method"], [4084, 4086, ["neural", "networks"], "Method"], [4109, 4112, ["cross", "-", "lingual"], "Task"], [4115, 4117, ["transfer", "learning"], "Method"], [4165, 4167, ["transfer", "learning"], "Method"], [4226, 4229, ["cross", "-", "lingual"], "Task"], [4229, 4231, ["transfer", "learning"], "Method"]]} {"doc_id": "19d583bf8c5533d1261ccdc068fdc3ef53b9ffb9", "text": ["document", ":", "FaceNet", ":", "A", "Unified", "Embedding", "for", "Face", "Recognition", "and", "Clustering", "Despite", "significant", "recent", "advances", "in", "the", "field", "of", "face", "recognition", ",", "implementing", "face", "verification", "and", "recognition", "efficiently", "at", "scale", "presents", "serious", "challenges", "to", "current", "approaches", ".", "In", "this", "paper", "we", "present", "a", "system", ",", "called", "FaceNet", ",", "that", "directly", "learns", "a", "mapping", "from", "face", "images", "to", "a", "compact", "Euclidean", "space", "where", "distances", "directly", "correspond", "to", "a", "measure", "of", "face", "similarity", ".", "Once", "this", "space", "has", "been", "produced", ",", "tasks", "such", "as", "face", "recognition", ",", "verification", "and", "clustering", "can", "be", "easily", "implemented", "using", "standard", "techniques", "with", "FaceNet", "embeddings", "as", "feature", "vectors", ".", "Our", "method", "uses", "a", "deep", "convolutional", "network", "trained", "to", "directly", "optimize", "the", "embedding", "itself", ",", "rather", "than", "an", "intermediate", "bottleneck", "layer", "as", "in", "previous", "deep", "learning", "approaches", ".", "To", "train", ",", "we", "use", "triplets", "of", "roughly", "aligned", "matching", "/", "non", "-", "matching", "face", "patches", "generated", "using", "a", "novel", "online", "triplet", "mining", "method", ".", "The", "benefit", "of", "our", "approach", "is", "much", "greater", "representational", "efficiency", ":", "we", "achieve", "state", "-", "of", "-", "the", "-", "art", "face", "recognition", "performance", "using", "only", "128", "-", "bytes", "per", "face", ".", "On", "the", "widely", "used", "Labeled", "Faces", "in", "the", "Wild", "(", "LFW", ")", "dataset", ",", "our", "system", "achieves", "a", "new", "record", "accuracy", "of", "99.63", "%", ".", "On", "YouTube", "Faces", "DB", "it", "achieves", "95.12", "%", ".", "Our", "system", "cuts", "the", "error", "rate", "in", "comparison", "to", "the", "best", "published", "result", "by", "30", "%", "on", "both", "datasets", ".", "We", "also", "introduce", "the", "concept", "of", "harmonic", "embeddings", ",", "and", "a", "harmonic", "triplet", "loss", ",", "which", "describe", "different", "versions", "of", "face", "embeddings", "(", "produced", "by", "different", "networks", ")", "that", "are", "compatible", "to", "each", "other", "and", "allow", "for", "direct", "comparison", "between", "each", "other", ".", "section", ":", "Introduction", "In", "this", "paper", "we", "present", "a", "unified", "system", "for", "face", "verification", "(", "is", "this", "the", "same", "person", ")", ",", "recognition", "(", "who", "is", "this", "person", ")", "and", "clustering", "(", "find", "common", "people", "among", "these", "faces", ")", ".", "Our", "method", "is", "based", "on", "learning", "a", "Euclidean", "embedding", "per", "image", "using", "a", "deep", "convolutional", "network", ".", "The", "network", "is", "trained", "such", "that", "the", "squared", "L2", "distances", "in", "the", "embedding", "space", "directly", "correspond", "to", "face", "similarity", ":", "faces", "of", "the", "same", "person", "have", "small", "distances", "and", "faces", "of", "distinct", "people", "have", "large", "distances", ".", "Once", "this", "embedding", "has", "been", "produced", ",", "then", "the", "aforementioned", "tasks", "become", "straight", "-", "forward", ":", "face", "verification", "simply", "involves", "thresholding", "the", "distance", "between", "the", "two", "embeddings", ";", "recognition", "becomes", "a", "k", "-", "NN", "classification", "problem", ";", "and", "clustering", "can", "be", "achieved", "using", "off", "-", "the", "-", "shelf", "techniques", "such", "as", "k", "-", "means", "or", "agglomerative", "clustering", ".", "Previous", "face", "recognition", "approaches", "based", "on", "deep", "networks", "use", "a", "classification", "layer", "trained", "over", "a", "set", "of", "known", "face", "identities", "and", "then", "take", "an", "intermediate", "bottleneck", "layer", "as", "a", "representation", "used", "to", "generalize", "recognition", "beyond", "the", "set", "of", "identities", "used", "in", "training", ".", "The", "downsides", "of", "this", "approach", "are", "its", "indirectness", "and", "its", "inefficiency", ":", "one", "has", "to", "hope", "that", "the", "bottleneck", "representation", "generalizes", "well", "to", "new", "faces", ";", "and", "by", "using", "a", "bottleneck", "layer", "the", "representation", "size", "per", "face", "is", "usually", "very", "large", "(", "1000s", "of", "dimensions", ")", ".", "Some", "recent", "work", "has", "reduced", "this", "dimensionality", "using", "PCA", ",", "but", "this", "is", "a", "linear", "transformation", "that", "can", "be", "easily", "learnt", "in", "one", "layer", "of", "the", "network", ".", "In", "contrast", "to", "these", "approaches", ",", "FaceNet", "directly", "trains", "its", "output", "to", "be", "a", "compact", "128", "-", "D", "embedding", "using", "a", "triplet", "-", "based", "loss", "function", "based", "on", "LMNN", ".", "Our", "triplets", "consist", "of", "two", "matching", "face", "thumbnails", "and", "a", "non", "-", "matching", "face", "thumbnail", "and", "the", "loss", "aims", "to", "separate", "the", "positive", "pair", "from", "the", "negative", "by", "a", "distance", "margin", ".", "The", "thumbnails", "are", "tight", "crops", "of", "the", "face", "area", ",", "no", "2D", "or", "3D", "alignment", ",", "other", "than", "scale", "and", "translation", "is", "performed", ".", "Choosing", "which", "triplets", "to", "use", "turns", "out", "to", "be", "very", "important", "for", "achieving", "good", "performance", "and", ",", "inspired", "by", "curriculum", "learning", ",", "we", "present", "a", "novel", "online", "negative", "exemplar", "mining", "strategy", "which", "ensures", "consistently", "increasing", "difficulty", "of", "triplets", "as", "the", "network", "trains", ".", "To", "improve", "clustering", "accuracy", ",", "we", "also", "explore", "hard", "-", "positive", "mining", "techniques", "which", "encourage", "spherical", "clusters", "for", "the", "embeddings", "of", "a", "single", "person", ".", "As", "an", "illustration", "of", "the", "incredible", "variability", "that", "our", "method", "can", "handle", "see", "Figure", "[", "reference", "]", ".", "Shown", "are", "image", "pairs", "from", "PIE", "that", "previously", "were", "considered", "to", "be", "very", "difficult", "for", "face", "verification", "systems", ".", "An", "overview", "of", "the", "rest", "of", "the", "paper", "is", "as", "follows", ":", "in", "section", "[", "reference", "]", "we", "review", "the", "literature", "in", "this", "area", ";", "section", "[", "reference", "]", "defines", "the", "triplet", "loss", "and", "section", "[", "reference", "]", "describes", "our", "novel", "triplet", "selection", "and", "training", "procedure", ";", "in", "section", "[", "reference", "]", "we", "describe", "the", "model", "architecture", "used", ".", "Finally", "in", "section", "[", "reference", "]", "and", "[", "reference", "]", "we", "present", "some", "quantitative", "results", "of", "our", "embeddings", "and", "also", "qualitatively", "explore", "some", "clustering", "results", ".", "section", ":", "Related", "Work", "Similarly", "to", "other", "recent", "works", "which", "employ", "deep", "networks", ",", "our", "approach", "is", "a", "purely", "data", "driven", "method", "which", "learns", "its", "representation", "directly", "from", "the", "pixels", "of", "the", "face", ".", "Rather", "than", "using", "engineered", "features", ",", "we", "use", "a", "large", "dataset", "of", "labelled", "faces", "to", "attain", "the", "appropriate", "invariances", "to", "pose", ",", "illumination", ",", "and", "other", "variational", "conditions", ".", "In", "this", "paper", "we", "explore", "two", "different", "deep", "network", "architectures", "that", "have", "been", "recently", "used", "to", "great", "success", "in", "the", "computer", "vision", "community", ".", "Both", "are", "deep", "convolutional", "networks", ".", "The", "first", "architecture", "is", "based", "on", "the", "Zeiler", "&", "Fergus", "model", "which", "consists", "of", "multiple", "interleaved", "layers", "of", "convolutions", ",", "non", "-", "linear", "activations", ",", "local", "response", "normalizations", ",", "and", "max", "pooling", "layers", ".", "We", "additionally", "add", "several", "convolution", "layers", "inspired", "by", "the", "work", "of", ".", "The", "second", "architecture", "is", "based", "on", "the", "Inception", "model", "of", "Szegedy", "which", "was", "recently", "used", "as", "the", "winning", "approach", "for", "ImageNet", "2014", ".", "These", "networks", "use", "mixed", "layers", "that", "run", "several", "different", "convolutional", "and", "pooling", "layers", "in", "parallel", "and", "concatenate", "their", "responses", ".", "We", "have", "found", "that", "these", "models", "can", "reduce", "the", "number", "of", "parameters", "by", "up", "to", "20", "times", "and", "have", "the", "potential", "to", "reduce", "the", "number", "of", "FLOPS", "required", "for", "comparable", "performance", ".", "There", "is", "a", "vast", "corpus", "of", "face", "verification", "and", "recognition", "works", ".", "Reviewing", "it", "is", "out", "of", "the", "scope", "of", "this", "paper", "so", "we", "will", "only", "briefly", "discuss", "the", "most", "relevant", "recent", "work", ".", "The", "works", "of", "all", "employ", "a", "complex", "system", "of", "multiple", "stages", ",", "that", "combines", "the", "output", "of", "a", "deep", "convolutional", "network", "with", "PCA", "for", "dimensionality", "reduction", "and", "an", "SVM", "for", "classification", ".", "Zhenyao", "employ", "a", "deep", "network", "to", "‘", "‘", "warp", "’", "’", "faces", "into", "a", "canonical", "frontal", "view", "and", "then", "learn", "CNN", "that", "classifies", "each", "face", "as", "belonging", "to", "a", "known", "identity", ".", "For", "face", "verification", ",", "PCA", "on", "the", "network", "output", "in", "conjunction", "with", "an", "ensemble", "of", "SVMs", "is", "used", ".", "Taigman", "propose", "a", "multi", "-", "stage", "approach", "that", "aligns", "faces", "to", "a", "general", "3D", "shape", "model", ".", "A", "multi", "-", "class", "network", "is", "trained", "to", "perform", "the", "face", "recognition", "task", "on", "over", "four", "thousand", "identities", ".", "The", "authors", "also", "experimented", "with", "a", "so", "called", "Siamese", "network", "where", "they", "directly", "optimize", "the", "-", "distance", "between", "two", "face", "features", ".", "Their", "best", "performance", "on", "LFW", "(", ")", "stems", "from", "an", "ensemble", "of", "three", "networks", "using", "different", "alignments", "and", "color", "channels", ".", "The", "predicted", "distances", "(", "non", "-", "linear", "SVM", "predictions", "based", "on", "the", "kernel", ")", "of", "those", "networks", "are", "combined", "using", "a", "non", "-", "linear", "SVM", ".", "Sun", "propose", "a", "compact", "and", "therefore", "relatively", "cheap", "to", "compute", "network", ".", "They", "use", "an", "ensemble", "of", "25", "of", "these", "network", ",", "each", "operating", "on", "a", "different", "face", "patch", ".", "For", "their", "final", "performance", "on", "LFW", "(", ")", "the", "authors", "combine", "50", "responses", "(", "regular", "and", "flipped", ")", ".", "Both", "PCA", "and", "a", "Joint", "Bayesian", "model", "that", "effectively", "correspond", "to", "a", "linear", "transform", "in", "the", "embedding", "space", "are", "employed", ".", "Their", "method", "does", "not", "require", "explicit", "2D", "/", "3D", "alignment", ".", "The", "networks", "are", "trained", "by", "using", "a", "combination", "of", "classification", "and", "verification", "loss", ".", "The", "verification", "loss", "is", "similar", "to", "the", "triplet", "loss", "we", "employ", ",", "in", "that", "it", "minimizes", "the", "-", "distance", "between", "faces", "of", "the", "same", "identity", "and", "enforces", "a", "margin", "between", "the", "distance", "of", "faces", "of", "different", "identities", ".", "The", "main", "difference", "is", "that", "only", "pairs", "of", "images", "are", "compared", ",", "whereas", "the", "triplet", "loss", "encourages", "a", "relative", "distance", "constraint", ".", "A", "similar", "loss", "to", "the", "one", "used", "here", "was", "explored", "in", "Wang", "for", "ranking", "images", "by", "semantic", "and", "visual", "similarity", ".", "section", ":", "Method", "FaceNet", "uses", "a", "deep", "convolutional", "network", ".", "We", "discuss", "two", "different", "core", "architectures", ":", "The", "Zeiler", "&", "Fergus", "style", "networks", "and", "the", "recent", "Inception", "type", "networks", ".", "The", "details", "of", "these", "networks", "are", "described", "in", "section", "[", "reference", "]", ".", "Given", "the", "model", "details", ",", "and", "treating", "it", "as", "a", "black", "box", "(", "see", "Figure", "[", "reference", "]", ")", ",", "the", "most", "important", "part", "of", "our", "approach", "lies", "in", "the", "end", "-", "to", "-", "end", "learning", "of", "the", "whole", "system", ".", "To", "this", "end", "we", "employ", "the", "triplet", "loss", "that", "directly", "reflects", "what", "we", "want", "to", "achieve", "in", "face", "verification", ",", "recognition", "and", "clustering", ".", "Namely", ",", "we", "strive", "for", "an", "embedding", ",", "from", "an", "image", "into", "a", "feature", "space", ",", "such", "that", "the", "squared", "distance", "between", "all", "faces", ",", "independent", "of", "imaging", "conditions", ",", "of", "the", "same", "identity", "is", "small", ",", "whereas", "the", "squared", "distance", "between", "a", "pair", "of", "face", "images", "from", "different", "identities", "is", "large", ".", "Although", "we", "did", "not", "directly", "compare", "to", "other", "losses", ",", "the", "one", "using", "pairs", "of", "positives", "and", "negatives", ",", "as", "used", "in", "Eq", ".", "(", "2", ")", ",", "we", "believe", "that", "the", "triplet", "loss", "is", "more", "suitable", "for", "face", "verification", ".", "The", "motivation", "is", "that", "the", "loss", "from", "encourages", "all", "faces", "of", "one", "identity", "to", "be", "projected", "onto", "a", "single", "point", "in", "the", "embedding", "space", ".", "The", "triplet", "loss", ",", "however", ",", "tries", "to", "enforce", "a", "margin", "between", "each", "pair", "of", "faces", "from", "one", "person", "to", "all", "other", "faces", ".", "This", "allows", "the", "faces", "for", "one", "identity", "to", "live", "on", "a", "manifold", ",", "while", "still", "enforcing", "the", "distance", "and", "thus", "discriminability", "to", "other", "identities", ".", "The", "following", "section", "describes", "this", "triplet", "loss", "and", "how", "it", "can", "be", "learned", "efficiently", "at", "scale", ".", "subsection", ":", "Triplet", "Loss", "The", "embedding", "is", "represented", "by", ".", "It", "embeds", "an", "image", "into", "a", "-", "dimensional", "Euclidean", "space", ".", "Additionally", ",", "we", "constrain", "this", "embedding", "to", "live", "on", "the", "-", "dimensional", "hypersphere", ",", ".", "This", "loss", "is", "motivated", "in", "in", "the", "context", "of", "nearest", "-", "neighbor", "classification", ".", "Here", "we", "want", "to", "ensure", "that", "an", "image", "(", "anchor", ")", "of", "a", "specific", "person", "is", "closer", "to", "all", "other", "images", "(", "positive", ")", "of", "the", "same", "person", "than", "it", "is", "to", "any", "image", "(", "negative", ")", "of", "any", "other", "person", ".", "This", "is", "visualized", "in", "Figure", "[", "reference", "]", ".", "Thus", "we", "want", ",", "where", "is", "a", "margin", "that", "is", "enforced", "between", "positive", "and", "negative", "pairs", ".", "is", "the", "set", "of", "all", "possible", "triplets", "in", "the", "training", "set", "and", "has", "cardinality", ".", "The", "loss", "that", "is", "being", "minimized", "is", "then", "Generating", "all", "possible", "triplets", "would", "result", "in", "many", "triplets", "that", "are", "easily", "satisfied", "(", "fulfill", "the", "constraint", "in", "Eq", ".", "(", "[", "reference", "]", ")", ")", ".", "These", "triplets", "would", "not", "contribute", "to", "the", "training", "and", "result", "in", "slower", "convergence", ",", "as", "they", "would", "still", "be", "passed", "through", "the", "network", ".", "It", "is", "crucial", "to", "select", "hard", "triplets", ",", "that", "are", "active", "and", "can", "therefore", "contribute", "to", "improving", "the", "model", ".", "The", "following", "section", "talks", "about", "the", "different", "approaches", "we", "use", "for", "the", "triplet", "selection", ".", "subsection", ":", "Triplet", "Selection", "In", "order", "to", "ensure", "fast", "convergence", "it", "is", "crucial", "to", "select", "triplets", "that", "violate", "the", "triplet", "constraint", "in", "Eq", ".", "(", "[", "reference", "]", ")", ".", "This", "means", "that", ",", "given", ",", "we", "want", "to", "select", "an", "(", "hard", "positive", ")", "such", "that", "and", "similarly", "(", "hard", "negative", ")", "such", "that", ".", "It", "is", "infeasible", "to", "compute", "the", "and", "across", "the", "whole", "training", "set", ".", "Additionally", ",", "it", "might", "lead", "to", "poor", "training", ",", "as", "mislabelled", "and", "poorly", "imaged", "faces", "would", "dominate", "the", "hard", "positives", "and", "negatives", ".", "There", "are", "two", "obvious", "choices", "that", "avoid", "this", "issue", ":", "Generate", "triplets", "offline", "every", "n", "steps", ",", "using", "the", "most", "recent", "network", "checkpoint", "and", "computing", "the", "and", "on", "a", "subset", "of", "the", "data", ".", "Generate", "triplets", "online", ".", "This", "can", "be", "done", "by", "selecting", "the", "hard", "positive", "/", "negative", "exemplars", "from", "within", "a", "mini", "-", "batch", ".", "Here", ",", "we", "focus", "on", "the", "online", "generation", "and", "use", "large", "mini", "-", "batches", "in", "the", "order", "of", "a", "few", "thousand", "exemplars", "and", "only", "compute", "the", "and", "within", "a", "mini", "-", "batch", ".", "To", "have", "a", "meaningful", "representation", "of", "the", "anchor", "-", "positive", "distances", ",", "it", "needs", "to", "be", "ensured", "that", "a", "minimal", "number", "of", "exemplars", "of", "any", "one", "identity", "is", "present", "in", "each", "mini", "-", "batch", ".", "In", "our", "experiments", "we", "sample", "the", "training", "data", "such", "that", "around", "40", "faces", "are", "selected", "per", "identity", "per", "mini", "-", "batch", ".", "Additionally", ",", "randomly", "sampled", "negative", "faces", "are", "added", "to", "each", "mini", "-", "batch", ".", "Instead", "of", "picking", "the", "hardest", "positive", ",", "we", "use", "all", "anchor", "-", "positive", "pairs", "in", "a", "mini", "-", "batch", "while", "still", "selecting", "the", "hard", "negatives", ".", "We", "do", "n’t", "have", "a", "side", "-", "by", "-", "side", "comparison", "of", "hard", "anchor", "-", "positive", "pairs", "versus", "all", "anchor", "-", "positive", "pairs", "within", "a", "mini", "-", "batch", ",", "but", "we", "found", "in", "practice", "that", "the", "all", "anchor", "-", "positive", "method", "was", "more", "stable", "and", "converged", "slightly", "faster", "at", "the", "beginning", "of", "training", ".", "We", "also", "explored", "the", "offline", "generation", "of", "triplets", "in", "conjunction", "with", "the", "online", "generation", "and", "it", "may", "allow", "the", "use", "of", "smaller", "batch", "sizes", ",", "but", "the", "experiments", "were", "inconclusive", ".", "Selecting", "the", "hardest", "negatives", "can", "in", "practice", "lead", "to", "bad", "local", "minima", "early", "on", "in", "training", ",", "specifically", "it", "can", "result", "in", "a", "collapsed", "model", "(", ")", ".", "In", "order", "to", "mitigate", "this", ",", "it", "helps", "to", "select", "such", "that", "We", "call", "these", "negative", "exemplars", "semi", "-", "hard", ",", "as", "they", "are", "further", "away", "from", "the", "anchor", "than", "the", "positive", "exemplar", ",", "but", "still", "hard", "because", "the", "squared", "distance", "is", "close", "to", "the", "anchor", "-", "positive", "distance", ".", "Those", "negatives", "lie", "inside", "the", "margin", ".", "As", "mentioned", "before", ",", "correct", "triplet", "selection", "is", "crucial", "for", "fast", "convergence", ".", "On", "the", "one", "hand", "we", "would", "like", "to", "use", "small", "mini", "-", "batches", "as", "these", "tend", "to", "improve", "convergence", "during", "Stochastic", "Gradient", "Descent", "(", "SGD", ")", ".", "On", "the", "other", "hand", ",", "implementation", "details", "make", "batches", "of", "tens", "to", "hundreds", "of", "exemplars", "more", "efficient", ".", "The", "main", "constraint", "with", "regards", "to", "the", "batch", "size", ",", "however", ",", "is", "the", "way", "we", "select", "hard", "relevant", "triplets", "from", "within", "the", "mini", "-", "batches", ".", "In", "most", "experiments", "we", "use", "a", "batch", "size", "of", "around", "1", ",", "800", "exemplars", ".", "subsection", ":", "Deep", "Convolutional", "Networks", "In", "all", "our", "experiments", "we", "train", "the", "CNN", "using", "Stochastic", "Gradient", "Descent", "(", "SGD", ")", "with", "standard", "backprop", "and", "AdaGrad", ".", "In", "most", "experiments", "we", "start", "with", "a", "learning", "rate", "of", "which", "we", "lower", "to", "finalize", "the", "model", ".", "The", "models", "are", "initialized", "from", "random", ",", "similar", "to", ",", "and", "trained", "on", "a", "CPU", "cluster", "for", "1", ",", "000", "to", "2", ",", "000", "hours", ".", "The", "decrease", "in", "the", "loss", "(", "and", "increase", "in", "accuracy", ")", "slows", "down", "drastically", "after", "500h", "of", "training", ",", "but", "additional", "training", "can", "still", "significantly", "improve", "performance", ".", "The", "margin", "is", "set", "to", ".", "We", "used", "two", "types", "of", "architectures", "and", "explore", "their", "trade", "-", "offs", "in", "more", "detail", "in", "the", "experimental", "section", ".", "Their", "practical", "differences", "lie", "in", "the", "difference", "of", "parameters", "and", "FLOPS", ".", "The", "best", "model", "may", "be", "different", "depending", "on", "the", "application", ".", "a", "model", "running", "in", "a", "datacenter", "can", "have", "many", "parameters", "and", "require", "a", "large", "number", "of", "FLOPS", ",", "whereas", "a", "model", "running", "on", "a", "mobile", "phone", "needs", "to", "have", "few", "parameters", ",", "so", "that", "it", "can", "fit", "into", "memory", ".", "All", "our", "models", "use", "rectified", "linear", "units", "as", "the", "non", "-", "linear", "activation", "function", ".", "The", "first", "category", ",", "shown", "in", "Table", "[", "reference", "]", ",", "adds", "convolutional", "layers", ",", "as", "suggested", "in", ",", "between", "the", "standard", "convolutional", "layers", "of", "the", "Zeiler", "&", "Fergus", "architecture", "and", "results", "in", "a", "model", "22", "layers", "deep", ".", "It", "has", "a", "total", "of", "140", "million", "parameters", "and", "requires", "around", "1.6", "billion", "FLOPS", "per", "image", ".", "The", "second", "category", "we", "use", "is", "based", "on", "GoogLeNet", "style", "Inception", "models", ".", "These", "models", "have", "fewer", "parameters", "(", "around", "6.6M", "-", "7.5", "M", ")", "and", "up", "to", "fewer", "FLOPS", "(", "between", "500M", "-", "1.6B", ")", ".", "Some", "of", "these", "models", "are", "dramatically", "reduced", "in", "size", "(", "both", "depth", "and", "number", "of", "filters", ")", ",", "so", "that", "they", "can", "be", "run", "on", "a", "mobile", "phone", ".", "One", ",", "NNS1", ",", "has", "26", "M", "parameters", "and", "only", "requires", "220", "M", "FLOPS", "per", "image", ".", "The", "other", ",", "NNS2", ",", "has", "4.3", "M", "parameters", "and", "20", "M", "FLOPS", ".", "Table", "[", "reference", "]", "describes", "NN2", "our", "largest", "network", "in", "detail", ".", "NN3", "is", "identical", "in", "architecture", "but", "has", "a", "reduced", "input", "size", "of", "160x160", ".", "NN4", "has", "an", "input", "size", "of", "only", "96x96", ",", "thereby", "drastically", "reducing", "the", "CPU", "requirements", "(", "285", "M", "FLOPS", "vs", "1.6B", "for", "NN2", ")", ".", "In", "addition", "to", "the", "reduced", "input", "size", "it", "does", "not", "use", "5x5", "convolutions", "in", "the", "higher", "layers", "as", "the", "receptive", "field", "is", "already", "too", "small", "by", "then", ".", "Generally", "we", "found", "that", "the", "5x5", "convolutions", "can", "be", "removed", "throughout", "with", "only", "a", "minor", "drop", "in", "accuracy", ".", "Figure", "[", "reference", "]", "compares", "all", "our", "models", ".", "section", ":", "Datasets", "and", "Evaluation", "We", "evaluate", "our", "method", "on", "four", "datasets", "and", "with", "the", "exception", "of", "Labelled", "Faces", "in", "the", "Wild", "and", "YouTube", "Faces", "we", "evaluate", "our", "method", "on", "the", "face", "verification", "task", ".", "given", "a", "pair", "of", "two", "face", "images", "a", "squared", "distance", "threshold", "is", "used", "to", "determine", "the", "classification", "of", "same", "and", "different", ".", "All", "faces", "pairs", "of", "the", "same", "identity", "are", "denoted", "with", ",", "whereas", "all", "pairs", "of", "different", "identities", "are", "denoted", "with", ".", "We", "define", "the", "set", "of", "all", "true", "accepts", "as", "These", "are", "the", "face", "pairs", "that", "were", "correctly", "classified", "as", "same", "at", "threshold", ".", "Similarly", "is", "the", "set", "of", "all", "pairs", "that", "was", "incorrectly", "classified", "as", "same", "(", "false", "accept", ")", ".", "The", "validation", "rate", "and", "the", "false", "accept", "rate", "for", "a", "given", "face", "distance", "are", "then", "defined", "as", "subsection", ":", "Hold", "-", "out", "Test", "Set", "We", "keep", "a", "hold", "out", "set", "of", "around", "one", "million", "images", ",", "that", "has", "the", "same", "distribution", "as", "our", "training", "set", ",", "but", "disjoint", "identities", ".", "For", "evaluation", "we", "split", "it", "into", "five", "disjoint", "sets", "of", "images", "each", ".", "The", "and", "rate", "are", "then", "computed", "on", "image", "pairs", ".", "Standard", "error", "is", "reported", "across", "the", "five", "splits", ".", "subsection", ":", "Personal", "Photos", "This", "is", "a", "test", "set", "with", "similar", "distribution", "to", "our", "training", "set", ",", "but", "has", "been", "manually", "verified", "to", "have", "very", "clean", "labels", ".", "It", "consists", "of", "three", "personal", "photo", "collections", "with", "a", "total", "of", "around", "images", ".", "We", "compute", "the", "and", "rate", "across", "all", "12k", "squared", "pairs", "of", "images", ".", "subsection", ":", "Academic", "Datasets", "Labeled", "Faces", "in", "the", "Wild", "(", "LFW", ")", "is", "the", "de", "-", "facto", "academic", "test", "set", "for", "face", "verification", ".", "We", "follow", "the", "standard", "protocol", "for", "unrestricted", ",", "labeled", "outside", "data", "and", "report", "the", "mean", "classification", "accuracy", "as", "well", "as", "the", "standard", "error", "of", "the", "mean", ".", "Youtube", "Faces", "DB", "is", "a", "new", "dataset", "that", "has", "gained", "popularity", "in", "the", "face", "recognition", "community", ".", "The", "setup", "is", "similar", "to", "LFW", ",", "but", "instead", "of", "verifying", "pairs", "of", "images", ",", "pairs", "of", "videos", "are", "used", ".", "section", ":", "Experiments", "If", "not", "mentioned", "otherwise", "we", "use", "between", "100M", "-", "200", "M", "training", "face", "thumbnails", "consisting", "of", "about", "8", "M", "different", "identities", ".", "A", "face", "detector", "is", "run", "on", "each", "image", "and", "a", "tight", "bounding", "box", "around", "each", "face", "is", "generated", ".", "These", "face", "thumbnails", "are", "resized", "to", "the", "input", "size", "of", "the", "respective", "network", ".", "Input", "sizes", "range", "from", "96x96", "pixels", "to", "224x224", "pixels", "in", "our", "experiments", ".", "subsection", ":", "Computation", "Accuracy", "Trade", "-", "off", "Before", "diving", "into", "the", "details", "of", "more", "specific", "experiments", "we", "will", "discuss", "the", "trade", "-", "off", "of", "accuracy", "versus", "number", "of", "FLOPS", "that", "a", "particular", "model", "requires", ".", "Figure", "[", "reference", "]", "shows", "the", "FLOPS", "on", "the", "x", "-", "axis", "and", "the", "accuracy", "at", "0.001", "false", "accept", "rate", "(", ")", "on", "our", "user", "labelled", "test", "-", "data", "set", "from", "section", "[", "reference", "]", ".", "It", "is", "interesting", "to", "see", "the", "strong", "correlation", "between", "the", "computation", "a", "model", "requires", "and", "the", "accuracy", "it", "achieves", ".", "The", "figure", "highlights", "the", "five", "models", "(", "NN1", ",", "NN2", ",", "NN3", ",", "NNS1", ",", "NNS2", ")", "that", "we", "discuss", "in", "more", "detail", "in", "our", "experiments", ".", "We", "also", "looked", "into", "the", "accuracy", "trade", "-", "off", "with", "regards", "to", "the", "number", "of", "model", "parameters", ".", "However", ",", "the", "picture", "is", "not", "as", "clear", "in", "that", "case", ".", "For", "example", ",", "the", "Inception", "based", "model", "NN2", "achieves", "a", "comparable", "performance", "to", "NN1", ",", "but", "only", "has", "a", "20th", "of", "the", "parameters", ".", "The", "number", "of", "FLOPS", "is", "comparable", ",", "though", ".", "Obviously", "at", "some", "point", "the", "performance", "is", "expected", "to", "decrease", ",", "if", "the", "number", "of", "parameters", "is", "reduced", "further", ".", "Other", "model", "architectures", "may", "allow", "further", "reductions", "without", "loss", "of", "accuracy", ",", "just", "like", "Inception", "did", "in", "this", "case", ".", "subsection", ":", "Effect", "of", "CNN", "Model", "We", "now", "discuss", "the", "performance", "of", "our", "four", "selected", "models", "in", "more", "detail", ".", "On", "the", "one", "hand", "we", "have", "our", "traditional", "Zeiler", "&", "Fergus", "based", "architecture", "with", "convolutions", "(", "see", "Table", "[", "reference", "]", ")", ".", "On", "the", "other", "hand", "we", "have", "Inception", "based", "models", "that", "dramatically", "reduce", "the", "model", "size", ".", "Overall", ",", "in", "the", "final", "performance", "the", "top", "models", "of", "both", "architectures", "perform", "comparably", ".", "However", ",", "some", "of", "our", "Inception", "based", "models", ",", "such", "as", "NN3", ",", "still", "achieve", "good", "performance", "while", "significantly", "reducing", "both", "the", "FLOPS", "and", "the", "model", "size", ".", "The", "detailed", "evaluation", "on", "our", "personal", "photos", "test", "set", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "While", "the", "largest", "model", "achieves", "a", "dramatic", "improvement", "in", "accuracy", "compared", "to", "the", "tiny", "NNS2", ",", "the", "latter", "can", "be", "run", "30ms", "/", "image", "on", "a", "mobile", "phone", "and", "is", "still", "accurate", "enough", "to", "be", "used", "in", "face", "clustering", ".", "The", "sharp", "drop", "in", "the", "ROC", "for", "indicates", "noisy", "labels", "in", "the", "test", "data", "groundtruth", ".", "At", "extremely", "low", "false", "accept", "rates", "a", "single", "mislabeled", "image", "can", "have", "a", "significant", "impact", "on", "the", "curve", ".", "subsection", ":", "Sensitivity", "to", "Image", "Quality", "Table", "[", "reference", "]", "shows", "the", "robustness", "of", "our", "model", "across", "a", "wide", "range", "of", "image", "sizes", ".", "The", "network", "is", "surprisingly", "robust", "with", "respect", "to", "JPEG", "compression", "and", "performs", "very", "well", "down", "to", "a", "JPEG", "quality", "of", "20", ".", "The", "performance", "drop", "is", "very", "small", "for", "face", "thumbnails", "down", "to", "a", "size", "of", "120x120", "pixels", "and", "even", "at", "80x80", "pixels", "it", "shows", "acceptable", "performance", ".", "This", "is", "notable", ",", "because", "the", "network", "was", "trained", "on", "220x220", "input", "images", ".", "Training", "with", "lower", "resolution", "faces", "could", "improve", "this", "range", "further", ".", "subsection", ":", "Embedding", "Dimensionality", "We", "explored", "various", "embedding", "dimensionalities", "and", "selected", "128", "for", "all", "experiments", "other", "than", "the", "comparison", "reported", "in", "Table", "[", "reference", "]", ".", "One", "would", "expect", "the", "larger", "embeddings", "to", "perform", "at", "least", "as", "good", "as", "the", "smaller", "ones", ",", "however", ",", "it", "is", "possible", "that", "they", "require", "more", "training", "to", "achieve", "the", "same", "accuracy", ".", "That", "said", ",", "the", "differences", "in", "the", "performance", "reported", "in", "Table", "[", "reference", "]", "are", "statistically", "insignificant", ".", "It", "should", "be", "noted", ",", "that", "during", "training", "a", "128", "dimensional", "float", "vector", "is", "used", ",", "but", "it", "can", "be", "quantized", "to", "128", "-", "bytes", "without", "loss", "of", "accuracy", ".", "Thus", "each", "face", "is", "compactly", "represented", "by", "a", "128", "dimensional", "byte", "vector", ",", "which", "is", "ideal", "for", "large", "scale", "clustering", "and", "recognition", ".", "Smaller", "embeddings", "are", "possible", "at", "a", "minor", "loss", "of", "accuracy", "and", "could", "be", "employed", "on", "mobile", "devices", ".", "subsection", ":", "Amount", "of", "Training", "Data", "Table", "[", "reference", "]", "shows", "the", "impact", "of", "large", "amounts", "of", "training", "data", ".", "Due", "to", "time", "constraints", "this", "evaluation", "was", "run", "on", "a", "smaller", "model", ";", "the", "effect", "may", "be", "even", "larger", "on", "larger", "models", ".", "It", "is", "clear", "that", "using", "tens", "of", "millions", "of", "exemplars", "results", "in", "a", "clear", "boost", "of", "accuracy", "on", "our", "personal", "photo", "test", "set", "from", "section", "[", "reference", "]", ".", "Compared", "to", "only", "millions", "of", "images", "the", "relative", "reduction", "in", "error", "is", "60", "%", ".", "Using", "another", "order", "of", "magnitude", "more", "images", "(", "hundreds", "of", "millions", ")", "still", "gives", "a", "small", "boost", ",", "but", "the", "improvement", "tapers", "off", ".", "subsection", ":", "Performance", "on", "LFW", "False", "accept", "False", "reject", "We", "evaluate", "our", "model", "on", "LFW", "using", "the", "standard", "protocol", "for", "unrestricted", ",", "labeled", "outside", "data", ".", "Nine", "training", "splits", "are", "used", "to", "select", "the", "-", "distance", "threshold", ".", "Classification", "(", "same", "or", "different", ")", "is", "then", "performed", "on", "the", "tenth", "test", "split", ".", "The", "selected", "optimal", "threshold", "is", "for", "all", "test", "splits", "except", "split", "eighth", "(", ")", ".", "Our", "model", "is", "evaluated", "in", "two", "modes", ":", "Fixed", "center", "crop", "of", "the", "LFW", "provided", "thumbnail", ".", "A", "proprietary", "face", "detector", "(", "similar", "to", "Picasa", ")", "is", "run", "on", "the", "provided", "LFW", "thumbnails", ".", "If", "it", "fails", "to", "align", "the", "face", "(", "this", "happens", "for", "two", "images", ")", ",", "the", "LFW", "alignment", "is", "used", ".", "Figure", "[", "reference", "]", "gives", "an", "overview", "of", "all", "failure", "cases", ".", "It", "shows", "false", "accepts", "on", "the", "top", "as", "well", "as", "false", "rejects", "at", "the", "bottom", ".", "We", "achieve", "a", "classification", "accuracy", "of", "98.87%±0.15", "when", "using", "the", "fixed", "center", "crop", "described", "in", "(", "1", ")", "and", "the", "record", "breaking", "99.63%±", "0.09", "standard", "error", "of", "the", "mean", "when", "using", "the", "extra", "face", "alignment", "(", "2", ")", ".", "This", "reduces", "the", "error", "reported", "for", "DeepFace", "in", "by", "more", "than", "a", "factor", "of", "7", "and", "the", "previous", "state", "-", "of", "-", "the", "-", "art", "reported", "for", "DeepId2", "+", "in", "by", "30", "%", ".", "This", "is", "the", "performance", "of", "model", "NN1", ",", "but", "even", "the", "much", "smaller", "NN3", "achieves", "performance", "that", "is", "not", "statistically", "significantly", "different", ".", "subsection", ":", "Performance", "on", "Youtube", "Faces", "DB", "We", "use", "the", "average", "similarity", "of", "all", "pairs", "of", "the", "first", "one", "hundred", "frames", "that", "our", "face", "detector", "detects", "in", "each", "video", ".", "This", "gives", "us", "a", "classification", "accuracy", "of", "95.12%±0.39", ".", "Using", "the", "first", "one", "thousand", "frames", "results", "in", "95.18", "%", ".", "Compared", "to", "91.4", "%", "who", "also", "evaluate", "one", "hundred", "frames", "per", "video", "we", "reduce", "the", "error", "rate", "by", "almost", "half", ".", "DeepId2", "+", "achieved", "93.2", "%", "and", "our", "method", "reduces", "this", "error", "by", "30", "%", ",", "comparable", "to", "our", "improvement", "on", "LFW", ".", "subsection", ":", "Face", "Clustering", "Our", "compact", "embedding", "lends", "itself", "to", "be", "used", "in", "order", "to", "cluster", "a", "users", "personal", "photos", "into", "groups", "of", "people", "with", "the", "same", "identity", ".", "The", "constraints", "in", "assignment", "imposed", "by", "clustering", "faces", ",", "compared", "to", "the", "pure", "verification", "task", ",", "lead", "to", "truly", "amazing", "results", ".", "Figure", "[", "reference", "]", "shows", "one", "cluster", "in", "a", "users", "personal", "photo", "collection", ",", "generated", "using", "agglomerative", "clustering", ".", "It", "is", "a", "clear", "showcase", "of", "the", "incredible", "invariance", "to", "occlusion", ",", "lighting", ",", "pose", "and", "even", "age", ".", "section", ":", "Summary", "We", "provide", "a", "method", "to", "directly", "learn", "an", "embedding", "into", "an", "Euclidean", "space", "for", "face", "verification", ".", "This", "sets", "it", "apart", "from", "other", "methods", "who", "use", "the", "CNN", "bottleneck", "layer", ",", "or", "require", "additional", "post", "-", "processing", "such", "as", "concatenation", "of", "multiple", "models", "and", "PCA", ",", "as", "well", "as", "SVM", "classification", ".", "Our", "end", "-", "to", "-", "end", "training", "both", "simplifies", "the", "setup", "and", "shows", "that", "directly", "optimizing", "a", "loss", "relevant", "to", "the", "task", "at", "hand", "improves", "performance", ".", "Another", "strength", "of", "our", "model", "is", "that", "it", "only", "requires", "minimal", "alignment", "(", "tight", "crop", "around", "the", "face", "area", ")", ".", ",", "for", "example", ",", "performs", "a", "complex", "3D", "alignment", ".", "We", "also", "experimented", "with", "a", "similarity", "transform", "alignment", "and", "notice", "that", "this", "can", "actually", "improve", "performance", "slightly", ".", "It", "is", "not", "clear", "if", "it", "is", "worth", "the", "extra", "complexity", ".", "Future", "work", "will", "focus", "on", "better", "understanding", "of", "the", "error", "cases", ",", "further", "improving", "the", "model", ",", "and", "also", "reducing", "model", "size", "and", "reducing", "CPU", "requirements", ".", "We", "will", "also", "look", "into", "ways", "of", "improving", "the", "currently", "extremely", "long", "training", "times", ",", "variations", "of", "our", "curriculum", "learning", "with", "smaller", "batch", "sizes", "and", "offline", "as", "well", "as", "online", "positive", "and", "negative", "mining", ".", "section", ":", "Appendix", ":", "Harmonic", "Embedding", "In", "this", "section", "we", "introduce", "the", "concept", "of", "harmonic", "embeddings", ".", "By", "this", "we", "denote", "a", "set", "of", "embeddings", "that", "are", "generated", "by", "different", "models", "v1", "and", "v2", "but", "are", "compatible", "in", "the", "sense", "that", "they", "can", "be", "compared", "to", "each", "other", ".", "This", "compatibility", "greatly", "simplifies", "upgrade", "paths", ".", "in", "an", "scenario", "where", "embedding", "v1", "was", "computed", "across", "a", "large", "set", "of", "images", "and", "a", "new", "embedding", "model", "v2", "is", "being", "rolled", "out", ",", "this", "compatibility", "ensures", "a", "smooth", "transition", "without", "the", "need", "to", "worry", "about", "version", "incompatibilities", ".", "Figure", "[", "reference", "]", "shows", "results", "on", "our", "3", "G", "dataset", ".", "It", "can", "be", "seen", "that", "the", "improved", "model", "NN2", "significantly", "outperforms", "NN1", ",", "while", "the", "comparison", "of", "NN2", "embeddings", "to", "NN1", "embeddings", "performs", "at", "an", "intermediate", "level", ".", "subsection", ":", "Harmonic", "Triplet", "Loss", "In", "order", "to", "learn", "the", "harmonic", "embedding", "we", "mix", "embeddings", "of", "v1", "together", "with", "the", "embeddings", "v2", ",", "that", "are", "being", "learned", ".", "This", "is", "done", "inside", "the", "triplet", "loss", "and", "results", "in", "additionally", "generated", "triplets", "that", "encourage", "the", "compatibility", "between", "the", "different", "embedding", "versions", ".", "Figure", "[", "reference", "]", "visualizes", "the", "different", "combinations", "of", "triplets", "that", "contribute", "to", "the", "triplet", "loss", ".", "We", "initialized", "the", "v2", "embedding", "from", "an", "independently", "trained", "NN2", "and", "retrained", "the", "last", "layer", "(", "embedding", "layer", ")", "from", "random", "initialization", "with", "the", "compatibility", "encouraging", "triplet", "loss", ".", "First", "only", "the", "last", "layer", "is", "retrained", ",", "then", "we", "continue", "training", "the", "whole", "v2", "network", "with", "the", "harmonic", "loss", ".", "Figure", "[", "reference", "]", "shows", "a", "possible", "interpretation", "of", "how", "this", "compatibility", "may", "work", "in", "practice", ".", "The", "vast", "majority", "of", "v2", "embeddings", "may", "be", "embedded", "near", "the", "corresponding", "v1", "embedding", ",", "however", ",", "incorrectly", "placed", "v1", "embeddings", "can", "be", "perturbed", "slightly", "such", "that", "their", "new", "location", "in", "embedding", "space", "improves", "verification", "accuracy", ".", "subsection", ":", "Summary", "These", "are", "very", "interesting", "findings", "and", "it", "is", "somewhat", "surprising", "that", "it", "works", "so", "well", ".", "Future", "work", "can", "explore", "how", "far", "this", "idea", "can", "be", "extended", ".", "Presumably", "there", "is", "a", "limit", "as", "to", "how", "much", "the", "v2", "embedding", "can", "improve", "over", "v1", ",", "while", "still", "being", "compatible", ".", "Additionally", "it", "would", "be", "interesting", "to", "train", "small", "networks", "that", "can", "run", "on", "a", "mobile", "phone", "and", "are", "compatible", "to", "a", "larger", "server", "side", "model", ".", "section", ":", "Acknowledgments", "We", "would", "like", "to", "thank", "Johannes", "Steffens", "for", "his", "discussions", "and", "great", "insights", "on", "face", "recognition", "and", "Christian", "Szegedy", "for", "providing", "new", "network", "architectures", "like", "and", "discussing", "network", "design", "choices", ".", "Also", "we", "are", "indebted", "to", "the", "DistBelief", "team", "for", "their", "support", "especially", "to", "Rajat", "Monga", "for", "help", "in", "setting", "up", "efficient", "training", "schemes", ".", "Also", "our", "work", "would", "not", "have", "been", "possible", "without", "the", "support", "of", "Chuck", "Rosenberg", ",", "Hartwig", "Adam", ",", "and", "Simon", "Han", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["FaceNet"], "Method"], [5, 7, ["Unified", "Embedding"], "Method"], [8, 12, ["Face", "Recognition", "and", "Clustering"], "Task"], [24, 26, ["face", "verification"], "Task"], [27, 28, ["recognition"], "Task"], [47, 48, ["FaceNet"], "Method"], [70, 72, ["face", "similarity"], "Metric"], [86, 87, ["verification"], "Task"], [88, 89, ["clustering"], "Task"], [121, 124, ["intermediate", "bottleneck", "layer"], "Method"], [127, 130, ["deep", "learning", "approaches"], "Method"], [151, 155, ["online", "triplet", "mining", "method"], "Method"], [164, 166, ["representational", "efficiency"], "Metric"], [191, 196, ["Labeled", "Faces", "in", "the", "Wild"], "Material"], [197, 198, ["LFW"], "Material"], [206, 208, ["record", "accuracy"], "Metric"], [213, 216, ["YouTube", "Faces", "DB"], "Material"], [225, 227, ["error", "rate"], "Metric"], [247, 249, ["harmonic", "embeddings"], "Method"], [296, 298, ["face", "verification"], "Task"], [306, 307, ["recognition"], "Task"], [314, 315, ["clustering"], "Task"], [331, 335, ["Euclidean", "embedding", "per", "image"], "Method"], [394, 396, ["face", "verification"], "Task"], [406, 407, ["recognition"], "Task"], [409, 414, ["k", "-", "NN", "classification", "problem"], "Task"], [416, 417, ["clustering"], "Method"], [421, 427, ["off", "-", "the", "-", "shelf", "techniques"], "Method"], [429, 432, ["k", "-", "means"], "Method"], [433, 435, ["agglomerative", "clustering"], "Method"], [442, 444, ["deep", "networks"], "Method"], [446, 448, ["classification", "layer"], "Method"], [460, 463, ["intermediate", "bottleneck", "layer"], "Method"], [469, 470, ["recognition"], "Task"], [497, 499, ["bottleneck", "representation"], "Method"], [509, 511, ["bottleneck", "layer"], "Method"], [534, 535, ["PCA"], "Method"], [540, 542, ["linear", "transformation"], "Method"], [560, 561, ["FaceNet"], "Method"], [569, 573, ["128", "-", "D", "embedding"], "Method"], [575, 580, ["triplet", "-", "based", "loss", "function"], "Method"], [582, 583, ["LMNN"], "Method"], [659, 661, ["curriculum", "learning"], "Method"], [666, 671, ["online", "negative", "exemplar", "mining", "strategy"], "Method"], [685, 687, ["clustering", "accuracy"], "Metric"], [691, 696, ["hard", "-", "positive", "mining", "techniques"], "Method"], [786, 788, ["triplet", "selection"], "Method"], [789, 791, ["training", "procedure"], "Method"], [827, 828, ["clustering"], "Task"], [841, 843, ["deep", "networks"], "Method"], [849, 852, ["data", "driven", "method"], "Method"], [900, 903, ["deep", "network", "architectures"], "Method"], [913, 916, ["computer", "vision", "community"], "Task"], [930, 934, ["Zeiler", "&", "Fergus", "model"], "Method"], [938, 942, ["interleaved", "layers", "of", "convolutions"], "Method"], [948, 951, ["local", "response", "normalizations"], "Method"], [961, 963, ["convolution", "layers"], "Method"], [976, 978, ["Inception", "model"], "Method"], [979, 980, ["Szegedy"], "Method"], [989, 991, ["ImageNet", "2014"], "Task"], [995, 997, ["mixed", "layers"], "Method"], [1001, 1002, ["convolutional"], "Method"], [1003, 1005, ["pooling", "layers"], "Method"], [1050, 1052, ["face", "verification"], "Task"], [1053, 1054, ["recognition"], "Task"], [1100, 1101, ["PCA"], "Method"], [1102, 1104, ["dimensionality", "reduction"], "Task"], [1106, 1107, ["SVM"], "Method"], [1108, 1109, ["classification"], "Task"], [1113, 1115, ["deep", "network"], "Method"], [1130, 1131, ["CNN"], "Method"], [1143, 1145, ["face", "verification"], "Task"], [1146, 1147, ["PCA"], "Method"], [1155, 1158, ["ensemble", "of", "SVMs"], "Method"], [1161, 1162, ["Taigman"], "Method"], [1164, 1168, ["multi", "-", "stage", "approach"], "Method"], [1174, 1177, ["3D", "shape", "model"], "Method"], [1179, 1183, ["multi", "-", "class", "network"], "Method"], [1205, 1207, ["Siamese", "network"], "Method"], [1223, 1224, ["LFW"], "Material"], [1229, 1233, ["ensemble", "of", "three", "networks"], "Method"], [1244, 1249, ["non", "-", "linear", "SVM", "predictions"], "Method"], [1261, 1265, ["non", "-", "linear", "SVM"], "Method"], [1275, 1277, ["compute", "network"], "Method"], [1301, 1302, ["LFW"], "Material"], [1316, 1317, ["PCA"], "Method"], [1319, 1322, ["Joint", "Bayesian", "model"], "Method"], [1327, 1329, ["linear", "transform"], "Method"], [1356, 1360, ["classification", "and", "verification", "loss"], "Metric"], [1362, 1364, ["verification", "loss"], "Task"], [1434, 1436, ["ranking", "images"], "Task"], [1445, 1446, ["FaceNet"], "Method"], [1460, 1465, ["Zeiler", "&", "Fergus", "style", "networks"], "Method"], [1468, 1471, ["Inception", "type", "networks"], "Method"], [1515, 1521, ["end", "-", "to", "-", "end", "learning"], "Task"], [1543, 1545, ["face", "verification"], "Task"], [1546, 1547, ["recognition"], "Task"], [1548, 1549, ["clustering"], "Task"], [1556, 1557, ["embedding"], "Task"], [1635, 1637, ["triplet", "loss"], "Method"], [1641, 1643, ["face", "verification"], "Task"], [1737, 1739, ["Triplet", "Loss"], "Method"], [1740, 1741, ["embedding"], "Task"], [1780, 1784, ["nearest", "-", "neighbor", "classification"], "Task"], [1910, 1911, ["training"], "Task"], [1915, 1916, ["convergence"], "Metric"], [1959, 1961, ["triplet", "selection"], "Task"], [1964, 1966, ["Triplet", "Selection"], "Method"], [2038, 2039, ["training"], "Task"], [2117, 2119, ["online", "generation"], "Task"], [2278, 2282, ["anchor", "-", "positive", "method"], "Method"], [2299, 2303, ["offline", "generation", "of", "triplets"], "Task"], [2307, 2309, ["online", "generation"], "Method"], [2341, 2342, ["training"], "Task"], [2349, 2351, ["collapsed", "model"], "Method"], [2416, 2418, ["triplet", "selection"], "Task"], [2421, 2423, ["fast", "convergence"], "Task"], [2442, 2443, ["convergence"], "Metric"], [2444, 2447, ["Stochastic", "Gradient", "Descent"], "Method"], [2448, 2449, ["SGD"], "Method"], [2513, 2516, ["Deep", "Convolutional", "Networks"], "Method"], [2523, 2524, ["CNN"], "Method"], [2525, 2528, ["Stochastic", "Gradient", "Descent"], "Method"], [2529, 2530, ["SGD"], "Method"], [2533, 2534, ["backprop"], "Method"], [2535, 2536, ["AdaGrad"], "Method"], [2544, 2546, ["learning", "rate"], "Metric"], [2585, 2586, ["loss"], "Metric"], [2590, 2591, ["accuracy"], "Metric"], [2598, 2599, ["training"], "Task"], [2602, 2603, ["training"], "Method"], [2702, 2705, ["rectified", "linear", "units"], "Method"], [2707, 2712, ["non", "-", "linear", "activation", "function"], "Method"], [2739, 2743, ["Zeiler", "&", "Fergus", "architecture"], "Method"], [2777, 2781, ["GoogLeNet", "style", "Inception", "models"], "Method"], [2837, 2838, ["NNS1"], "Method"], [2855, 2856, ["NNS2"], "Method"], [2871, 2872, ["NN2"], "Method"], [2878, 2879, ["NN3"], "Method"], [2892, 2893, ["NN4"], "Method"], [2905, 2907, ["CPU", "requirements"], "Metric"], [2914, 2915, ["NN2"], "Method"], [2929, 2930, ["convolutions"], "Method"], [2962, 2963, ["accuracy"], "Metric"], [2990, 2995, ["Labelled", "Faces", "in", "the", "Wild"], "Material"], [2996, 2998, ["YouTube", "Faces"], "Material"], [3093, 3095, ["validation", "rate"], "Metric"], [3097, 3100, ["false", "accept", "rate"], "Metric"], [3156, 3158, ["and", "rate"], "Metric"], [3165, 3167, ["Standard", "error"], "Metric"], [3219, 3221, ["and", "rate"], "Metric"], [3233, 3238, ["Labeled", "Faces", "in", "the", "Wild"], "Material"], [3239, 3240, ["LFW"], "Material"], [3250, 3252, ["face", "verification"], "Task"], [3267, 3270, ["mean", "classification", "accuracy"], "Metric"], [3274, 3276, ["standard", "error"], "Metric"], [3280, 3283, ["Youtube", "Faces", "DB"], "Material"], [3302, 3303, ["LFW"], "Material"], [3344, 3346, ["face", "detector"], "Method"], [3391, 3394, ["Computation", "Accuracy", "Trade"], "Metric"], [3413, 3414, ["accuracy"], "Metric"], [3438, 3439, ["accuracy"], "Metric"], [3441, 3444, ["false", "accept", "rate"], "Metric"], [3476, 3477, ["accuracy"], "Metric"], [3487, 3488, ["NN1"], "Method"], [3489, 3490, ["NN2"], "Method"], [3491, 3492, ["NN3"], "Method"], [3493, 3494, ["NNS1"], "Method"], [3495, 3496, ["NNS2"], "Method"], [3512, 3513, ["accuracy"], "Metric"], [3541, 3545, ["Inception", "based", "model", "NN2"], "Method"], [3550, 3551, ["NN1"], "Method"], [3591, 3593, ["model", "architectures"], "Method"], [3600, 3601, ["accuracy"], "Metric"], [3614, 3616, ["CNN", "Model"], "Method"], [3638, 3643, ["Zeiler", "&", "Fergus", "based", "architecture"], "Method"], [3644, 3645, ["convolutions"], "Method"], [3659, 3662, ["Inception", "based", "models"], "Method"], [3666, 3668, ["model", "size"], "Metric"], [3689, 3692, ["Inception", "based", "models"], "Method"], [3695, 3696, ["NN3"], "Method"], [3706, 3707, ["FLOPS"], "Metric"], [3709, 3711, ["model", "size"], "Metric"], [3738, 3739, ["accuracy"], "Metric"], [3743, 3744, ["NNS2"], "Method"], [3766, 3768, ["face", "clustering"], "Task"], [3774, 3775, ["ROC"], "Metric"], [3788, 3791, ["false", "accept", "rates"], "Metric"], [3806, 3807, ["Sensitivity"], "Metric"], [3808, 3810, ["Image", "Quality"], "Metric"], [3836, 3838, ["JPEG", "compression"], "Task"], [3845, 3847, ["JPEG", "quality"], "Metric"], [3903, 3905, ["Embedding", "Dimensionality"], "Task"], [3958, 3959, ["accuracy"], "Metric"], [4006, 4007, ["accuracy"], "Metric"], [4016, 4020, ["128", "dimensional", "byte", "vector"], "Method"], [4040, 4041, ["accuracy"], "Metric"], [4108, 4109, ["accuracy"], "Metric"], [4131, 4132, ["error"], "Metric"], [4164, 4165, ["LFW"], "Material"], [4174, 4175, ["LFW"], "Material"], [4198, 4199, ["Classification"], "Task"], [4241, 4242, ["LFW"], "Material"], [4246, 4249, ["proprietary", "face", "detector"], "Method"], [4252, 4253, ["Picasa"], "Method"], [4259, 4261, ["LFW", "thumbnails"], "Material"], [4278, 4280, ["LFW", "alignment"], "Material"], [4314, 4316, ["classification", "accuracy"], "Metric"], [4353, 4354, ["error"], "Metric"], [4356, 4358, ["DeepFace", "in"], "Task"], [4377, 4380, ["DeepId2", "+", "in"], "Metric"], [4389, 4391, ["model", "NN1"], "Method"], [4397, 4398, ["NN3"], "Method"], [4411, 4414, ["Youtube", "Faces", "DB"], "Material"], [4430, 4432, ["face", "detector"], "Method"], [4441, 4443, ["classification", "accuracy"], "Metric"], [4472, 4474, ["error", "rate"], "Metric"], [4478, 4480, ["DeepId2", "+"], "Method"], [4498, 4499, ["LFW"], "Material"], [4502, 4504, ["Face", "Clustering"], "Task"], [4505, 4507, ["compact", "embedding"], "Method"], [4532, 4533, ["assignment"], "Task"], [4541, 4544, ["pure", "verification", "task"], "Task"], [4567, 4569, ["agglomerative", "clustering"], "Method"], [4600, 4601, ["embedding"], "Task"], [4606, 4608, ["face", "verification"], "Task"], [4619, 4622, ["CNN", "bottleneck", "layer"], "Method"], [4626, 4629, ["post", "-", "processing"], "Method"], [4631, 4635, ["concatenation", "of", "multiple", "models"], "Method"], [4636, 4637, ["PCA"], "Method"], [4641, 4643, ["SVM", "classification"], "Method"], [4649, 4651, ["end", "training"], "Task"], [4661, 4662, ["loss"], "Task"], [4699, 4701, ["3D", "alignment"], "Task"], [4707, 4710, ["similarity", "transform", "alignment"], "Method"], [4752, 4754, ["model", "size"], "Metric"], [4756, 4758, ["CPU", "requirements"], "Metric"], [4771, 4773, ["training", "times"], "Metric"], [4777, 4779, ["curriculum", "learning"], "Method"], [4788, 4793, ["online", "positive", "and", "negative", "mining"], "Task"], [4798, 4800, ["Harmonic", "Embedding"], "Method"], [4808, 4810, ["harmonic", "embeddings"], "Task"], [4854, 4856, ["embedding", "v1"], "Task"], [4867, 4869, ["embedding", "model"], "Method"], [4909, 4911, ["model", "NN2"], "Method"], [4913, 4914, ["NN1"], "Method"], [4919, 4921, ["NN2", "embeddings"], "Method"], [4922, 4924, ["NN1", "embeddings"], "Method"], [4940, 4942, ["harmonic", "embedding"], "Method"], [5001, 5003, ["v2", "embedding"], "Method"], [5007, 5008, ["NN2"], "Method"], [5014, 5016, ["embedding", "layer"], "Method"], [5018, 5020, ["random", "initialization"], "Method"], [5041, 5043, ["v2", "network"], "Method"], [5099, 5101, ["verification", "accuracy"], "Metric"], [5143, 5145, ["v2", "embedding"], "Method"], [5177, 5180, ["server", "side", "model"], "Method"], [5206, 5208, ["network", "architectures"], "Method"], [5211, 5214, ["network", "design", "choices"], "Method"], [5236, 5238, ["training", "schemes"], "Method"], [21, 22, ["recognition"], "Task"], [84, 85, ["recognition"], "Task"], [97, 98, ["FaceNet"], "Method"], [108, 109, ["convolutional"], "Method"], [177, 178, ["recognition"], "Task"], [338, 339, ["convolutional"], "Method"], [438, 439, ["recognition"], "Task"], [741, 743, ["face", "verification"], "Task"], [920, 921, ["convolutional"], "Method"], [954, 956, ["pooling", "layers"], "Method"], [1097, 1098, ["convolutional"], "Method"], [1189, 1190, ["recognition"], "Task"], [1449, 1450, ["convolutional"], "Method"], [2725, 2726, ["convolutional"], "Method"], [2735, 2736, ["convolutional"], "Method"], [3004, 3006, ["face", "verification"], "Task"], [3294, 3295, ["recognition"], "Task"], [4029, 4030, ["recognition"], "Task"], [5199, 5200, ["recognition"], "Task"]]} {"doc_id": "1c059493904b2244d2280b8b4c0c7d3ca115be73", "text": ["document", ":", "node2vec", ":", "Scalable", "Feature", "Learning", "for", "Networks", "Prediction", "tasks", "over", "nodes", "and", "edges", "in", "networks", "require", "careful", "effort", "in", "engineering", "features", "used", "by", "learning", "algorithms", ".", "Recent", "research", "in", "the", "broader", "field", "of", "representation", "learning", "has", "led", "to", "significant", "progress", "in", "automating", "prediction", "by", "learning", "the", "features", "themselves", ".", "However", ",", "present", "feature", "learning", "approaches", "are", "not", "expressive", "enough", "to", "capture", "the", "diversity", "of", "connectivity", "patterns", "observed", "in", "networks", ".", "Here", "we", "propose", "node2vec", ",", "an", "algorithmic", "framework", "for", "learning", "continuous", "feature", "representations", "for", "nodes", "in", "networks", ".", "In", "node2vec", ",", "we", "learn", "a", "mapping", "of", "nodes", "to", "a", "low", "-", "dimensional", "space", "of", "features", "that", "maximizes", "the", "likelihood", "of", "preserving", "network", "neighborhoods", "of", "nodes", ".", "We", "define", "a", "flexible", "notion", "of", "a", "node", "’s", "network", "neighborhood", "and", "design", "a", "biased", "random", "walk", "procedure", ",", "which", "efficiently", "explores", "diverse", "neighborhoods", ".", "Our", "algorithm", "generalizes", "prior", "work", "which", "is", "based", "on", "rigid", "notions", "of", "network", "neighborhoods", ",", "and", "we", "argue", "that", "the", "added", "flexibility", "in", "exploring", "neighborhoods", "is", "the", "key", "to", "learning", "richer", "representations", ".", "We", "demonstrate", "the", "efficacy", "of", "node2vec", "over", "existing", "state", "-", "of", "-", "the", "-", "art", "techniques", "on", "multi", "-", "label", "classification", "and", "link", "prediction", "in", "several", "real", "-", "world", "networks", "from", "diverse", "domains", ".", "Taken", "together", ",", "our", "work", "represents", "a", "new", "way", "for", "efficiently", "learning", "state", "-", "of", "-", "the", "-", "art", "task", "-", "independent", "representations", "in", "complex", "networks", ".", "2", "2016", "acmlicensed", "KDD", "’", "16", ",", "August", "13", "-", "17", ",", "2016", ",", "San", "Francisco", ",", "CA", ",", "USA", "978", "-", "1", "-", "4503", "-", "4232", "-", "2", "/", "16", "/", "08", "$", "15.00", "http:", "//", "dx.doi.org", "/", "10.1145", "/", "2939672.2939754", "Categories", "and", "Subject", "Descriptors", ":", "H.2.8", "[", "Database", "Management", "]", ":", "Database", "applications—", "Data", "mining", ";", "I.2.6", "[", "Artificial", "Intelligence", "]", ":", "Learning", "General", "Terms", ":", "Algorithms", ";", "Experimentation", ".", "Keywords", ":", "Information", "networks", ",", "Feature", "learning", ",", "Node", "embeddings", ",", "Graph", "representations", ".", "section", ":", "Introduction", "Many", "important", "tasks", "in", "network", "analysis", "involve", "predictions", "over", "nodes", "and", "edges", ".", "In", "a", "typical", "node", "classification", "task", ",", "we", "are", "interested", "in", "predicting", "the", "most", "probable", "labels", "of", "nodes", "in", "a", "network", ".", "For", "example", ",", "in", "a", "social", "network", ",", "we", "might", "be", "interested", "in", "predicting", "interests", "of", "users", ",", "or", "in", "a", "protein", "-", "protein", "interaction", "network", "we", "might", "be", "interested", "in", "predicting", "functional", "labels", "of", "proteins", ".", "Similarly", ",", "in", "link", "prediction", ",", "we", "wish", "to", "predict", "whether", "a", "pair", "of", "nodes", "in", "a", "network", "should", "have", "an", "edge", "connecting", "them", ".", "Link", "prediction", "is", "useful", "in", "a", "wide", "variety", "of", "domains", ";", "for", "instance", ",", "in", "genomics", ",", "it", "helps", "us", "discover", "novel", "interactions", "between", "genes", ",", "and", "in", "social", "networks", ",", "it", "can", "identify", "real", "-", "world", "friends", ".", "Any", "supervised", "machine", "learning", "algorithm", "requires", "a", "set", "of", "informative", ",", "discriminating", ",", "and", "independent", "features", ".", "In", "prediction", "problems", "on", "networks", "this", "means", "that", "one", "has", "to", "construct", "a", "feature", "vector", "representation", "for", "the", "nodes", "and", "edges", ".", "A", "typical", "solution", "involves", "hand", "-", "engineering", "domain", "-", "specific", "features", "based", "on", "expert", "knowledge", ".", "Even", "if", "one", "discounts", "the", "tedious", "effort", "required", "for", "feature", "engineering", ",", "such", "features", "are", "usually", "designed", "for", "specific", "tasks", "and", "do", "not", "generalize", "across", "different", "prediction", "tasks", ".", "An", "alternative", "approach", "is", "to", "learn", "feature", "representations", "by", "solving", "an", "optimization", "problem", ".", "The", "challenge", "in", "feature", "learning", "is", "defining", "an", "objective", "function", ",", "which", "involves", "a", "trade", "-", "off", "in", "balancing", "computational", "efficiency", "and", "predictive", "accuracy", ".", "On", "one", "side", "of", "the", "spectrum", ",", "one", "could", "directly", "aim", "to", "find", "a", "feature", "representation", "that", "optimizes", "performance", "of", "a", "downstream", "prediction", "task", ".", "While", "this", "supervised", "procedure", "results", "in", "good", "accuracy", ",", "it", "comes", "at", "the", "cost", "of", "high", "training", "time", "complexity", "due", "to", "a", "blowup", "in", "the", "number", "of", "parameters", "that", "need", "to", "be", "estimated", ".", "At", "the", "other", "extreme", ",", "the", "objective", "function", "can", "be", "defined", "to", "be", "independent", "of", "the", "downstream", "prediction", "task", "and", "the", "representations", "can", "be", "learned", "in", "a", "purely", "unsupervised", "way", ".", "This", "makes", "the", "optimization", "computationally", "efficient", "and", "with", "a", "carefully", "designed", "objective", ",", "it", "results", "in", "task", "-", "independent", "features", "that", "closely", "match", "task", "-", "specific", "approaches", "in", "predictive", "accuracy", ".", "However", ",", "current", "techniques", "fail", "to", "satisfactorily", "define", "and", "optimize", "a", "reasonable", "objective", "required", "for", "scalable", "unsupervised", "feature", "learning", "in", "networks", ".", "Classic", "approaches", "based", "on", "linear", "and", "non", "-", "linear", "dimensionality", "reduction", "techniques", "such", "as", "Principal", "Component", "Analysis", ",", "Multi", "-", "Dimensional", "Scaling", "and", "their", "extensions", "optimize", "an", "objective", "that", "transforms", "a", "representative", "data", "matrix", "of", "the", "network", "such", "that", "it", "maximizes", "the", "variance", "of", "the", "data", "representation", ".", "Consequently", ",", "these", "approaches", "invariably", "involve", "eigendecomposition", "of", "the", "appropriate", "data", "matrix", "which", "is", "expensive", "for", "large", "real", "-", "world", "networks", ".", "Moreover", ",", "the", "resulting", "latent", "representations", "give", "poor", "performance", "on", "various", "prediction", "tasks", "over", "networks", ".", "Alternatively", ",", "we", "can", "design", "an", "objective", "that", "seeks", "to", "preserve", "local", "neighborhoods", "of", "nodes", ".", "The", "objective", "can", "be", "efficiently", "optimized", "using", "stochastic", "gradient", "descent", "(", "SGD", ")", "akin", "to", "backpropogation", "on", "just", "single", "hidden", "-", "layer", "feedforward", "neural", "networks", ".", "Recent", "attempts", "in", "this", "direction", "propose", "efficient", "algorithms", "but", "rely", "on", "a", "rigid", "notion", "of", "a", "network", "neighborhood", ",", "which", "results", "in", "these", "approaches", "being", "largely", "insensitive", "to", "connectivity", "patterns", "unique", "to", "networks", ".", "Specifically", ",", "nodes", "in", "networks", "could", "be", "organized", "based", "on", "communities", "they", "belong", "to", "(", "i.e.", ",", "homophily", ")", ";", "in", "other", "cases", ",", "the", "organization", "could", "be", "based", "on", "the", "structural", "roles", "of", "nodes", "in", "the", "network", "(", "i.e.", ",", "structural", "equivalence", ")", ".", "For", "instance", ",", "in", "Figure", "[", "reference", "]", ",", "we", "observe", "nodes", "and", "belonging", "to", "the", "same", "tightly", "knit", "community", "of", "nodes", ",", "while", "the", "nodes", "and", "in", "the", "two", "distinct", "communities", "share", "the", "same", "structural", "role", "of", "a", "hub", "node", ".", "Real", "-", "world", "networks", "commonly", "exhibit", "a", "mixture", "of", "such", "equivalences", ".", "Thus", ",", "it", "is", "essential", "to", "allow", "for", "a", "flexible", "algorithm", "that", "can", "learn", "node", "representations", "obeying", "both", "principles", ":", "ability", "to", "learn", "representations", "that", "embed", "nodes", "from", "the", "same", "network", "community", "closely", "together", ",", "as", "well", "as", "to", "learn", "representations", "where", "nodes", "that", "share", "similar", "roles", "have", "similar", "embeddings", ".", "This", "would", "allow", "feature", "learning", "algorithms", "to", "generalize", "across", "a", "wide", "variety", "of", "domains", "and", "prediction", "tasks", ".", "Present", "work", ".", "We", "propose", "node2vec", ",", "a", "semi", "-", "supervised", "algorithm", "for", "scalable", "feature", "learning", "in", "networks", ".", "We", "optimize", "a", "custom", "graph", "-", "based", "objective", "function", "using", "SGD", "motivated", "by", "prior", "work", "on", "natural", "language", "processing", ".", "Intuitively", ",", "our", "approach", "returns", "feature", "representations", "that", "maximize", "the", "likelihood", "of", "preserving", "network", "neighborhoods", "of", "nodes", "in", "a", "-", "dimensional", "feature", "space", ".", "We", "use", "a", "2", "order", "random", "walk", "approach", "to", "generate", "(", "sample", ")", "network", "neighborhoods", "for", "nodes", ".", "Our", "key", "contribution", "is", "in", "defining", "a", "flexible", "notion", "of", "a", "node", "’s", "network", "neighborhood", ".", "By", "choosing", "an", "appropriate", "notion", "of", "a", "neighborhood", ",", "node2vec", "can", "learn", "representations", "that", "organize", "nodes", "based", "on", "their", "network", "roles", "and", "/", "or", "communities", "they", "belong", "to", ".", "We", "achieve", "this", "by", "developing", "a", "family", "of", "biased", "random", "walks", ",", "which", "efficiently", "explore", "diverse", "neighborhoods", "of", "a", "given", "node", ".", "The", "resulting", "algorithm", "is", "flexible", ",", "giving", "us", "control", "over", "the", "search", "space", "through", "tunable", "parameters", ",", "in", "contrast", "to", "rigid", "search", "procedures", "in", "prior", "work", ".", "Consequently", ",", "our", "method", "generalizes", "prior", "work", "and", "can", "model", "the", "full", "spectrum", "of", "equivalences", "observed", "in", "networks", ".", "The", "parameters", "governing", "our", "search", "strategy", "have", "an", "intuitive", "interpretation", "and", "bias", "the", "walk", "towards", "different", "network", "exploration", "strategies", ".", "These", "parameters", "can", "also", "be", "learned", "directly", "using", "a", "tiny", "fraction", "of", "labeled", "data", "in", "a", "semi", "-", "supervised", "fashion", ".", "We", "also", "show", "how", "feature", "representations", "of", "individual", "nodes", "can", "be", "extended", "to", "pairs", "of", "nodes", "(", "i.e.", ",", "edges", ")", ".", "In", "order", "to", "generate", "feature", "representations", "of", "edges", ",", "we", "compose", "the", "learned", "feature", "representations", "of", "the", "individual", "nodes", "using", "simple", "binary", "operators", ".", "This", "compositionality", "lends", "node2vec", "to", "prediction", "tasks", "involving", "nodes", "as", "well", "as", "edges", ".", "Our", "experiments", "focus", "on", "two", "common", "prediction", "tasks", "in", "networks", ":", "a", "multi", "-", "label", "classification", "task", ",", "where", "every", "node", "is", "assigned", "one", "or", "more", "class", "labels", "and", "a", "link", "prediction", "task", ",", "where", "we", "predict", "the", "existence", "of", "an", "edge", "given", "a", "pair", "of", "nodes", ".", "We", "contrast", "the", "performance", "of", "node2vec", "with", "state", "-", "of", "-", "the", "-", "art", "feature", "learning", "algorithms", ".", "We", "experiment", "with", "several", "real", "-", "world", "networks", "from", "diverse", "domains", ",", "such", "as", "social", "networks", ",", "information", "networks", ",", "as", "well", "as", "networks", "from", "systems", "biology", ".", "Experiments", "demonstrate", "that", "node2vec", "outperforms", "state", "-", "of", "-", "the", "-", "art", "methods", "by", "up", "to", "26.7", "%", "on", "multi", "-", "label", "classification", "and", "up", "to", "12.6", "%", "on", "link", "prediction", ".", "The", "algorithm", "shows", "competitive", "performance", "with", "even", "10", "%", "labeled", "data", "and", "is", "also", "robust", "to", "perturbations", "in", "the", "form", "of", "noisy", "or", "missing", "edges", ".", "Computationally", ",", "the", "major", "phases", "of", "node2vec", "are", "trivially", "parallelizable", ",", "and", "it", "can", "scale", "to", "large", "networks", "with", "millions", "of", "nodes", "in", "a", "few", "hours", ".", "Overall", "our", "paper", "makes", "the", "following", "contributions", ":", "We", "propose", "node2vec", ",", "an", "efficient", "scalable", "algorithm", "for", "feature", "learning", "in", "networks", "that", "efficiently", "optimizes", "a", "novel", "network", "-", "aware", ",", "neighborhood", "preserving", "objective", "using", "SGD", ".", "We", "show", "how", "node2vec", "is", "in", "accordance", "with", "established", "principles", "in", "network", "science", ",", "providing", "flexibility", "in", "discovering", "representations", "conforming", "to", "different", "equivalences", ".", "We", "extend", "node2vec", "and", "other", "feature", "learning", "methods", "based", "on", "neighborhood", "preserving", "objectives", ",", "from", "nodes", "to", "pairs", "of", "nodes", "for", "edge", "-", "based", "prediction", "tasks", ".", "We", "empirically", "evaluate", "node2vec", "for", "multi", "-", "label", "classification", "and", "link", "prediction", "on", "several", "real", "-", "world", "datasets", ".", "The", "rest", "of", "the", "paper", "is", "structured", "as", "follows", ".", "In", "Section", "[", "reference", "]", ",", "we", "briefly", "survey", "related", "work", "in", "feature", "learning", "for", "networks", ".", "We", "present", "the", "technical", "details", "for", "feature", "learning", "using", "node2vec", "in", "Section", "[", "reference", "]", ".", "In", "Section", "[", "reference", "]", ",", "we", "empirically", "evaluate", "node2vec", "on", "prediction", "tasks", "over", "nodes", "and", "edges", "on", "various", "real", "-", "world", "networks", "and", "assess", "the", "parameter", "sensitivity", ",", "perturbation", "analysis", ",", "and", "scalability", "aspects", "of", "our", "algorithm", ".", "We", "conclude", "with", "a", "discussion", "of", "the", "node2vec", "framework", "and", "highlight", "some", "promising", "directions", "for", "future", "work", "in", "Section", "[", "reference", "]", ".", "Datasets", "and", "a", "reference", "implementation", "of", "node2vec", "are", "available", "on", "the", "project", "page", ":", ".", "section", ":", "Related", "work", "Feature", "engineering", "has", "been", "extensively", "studied", "by", "the", "machine", "learning", "community", "under", "various", "headings", ".", "In", "networks", ",", "the", "conventional", "paradigm", "for", "generating", "features", "for", "nodes", "is", "based", "on", "feature", "extraction", "techniques", "which", "typically", "involve", "some", "seed", "hand", "-", "crafted", "features", "based", "on", "network", "properties", ".", "In", "contrast", ",", "our", "goal", "is", "to", "automate", "the", "whole", "process", "by", "casting", "feature", "extraction", "as", "a", "representation", "learning", "problem", "in", "which", "case", "we", "do", "not", "require", "any", "hand", "-", "engineered", "features", ".", "Unsupervised", "feature", "learning", "approaches", "typically", "exploit", "the", "spectral", "properties", "of", "various", "matrix", "representations", "of", "graphs", ",", "especially", "the", "Laplacian", "and", "the", "adjacency", "matrices", ".", "Under", "this", "linear", "algebra", "perspective", ",", "these", "methods", "can", "be", "viewed", "as", "dimensionality", "reduction", "techniques", ".", "Several", "linear", "(", "e.g.", ",", "PCA", ")", "and", "non", "-", "linear", "(", "e.g.", ",", "IsoMap", ")", "dimensionality", "reduction", "techniques", "have", "been", "proposed", ".", "These", "methods", "suffer", "from", "both", "computational", "and", "statistical", "performance", "drawbacks", ".", "In", "terms", "of", "computational", "efficiency", ",", "eigendecomposition", "of", "a", "data", "matrix", "is", "expensive", "unless", "the", "solution", "quality", "is", "significantly", "compromised", "with", "approximations", ",", "and", "hence", ",", "these", "methods", "are", "hard", "to", "scale", "to", "large", "networks", ".", "Secondly", ",", "these", "methods", "optimize", "for", "objectives", "that", "are", "not", "robust", "to", "the", "diverse", "patterns", "observed", "in", "networks", "(", "such", "as", "homophily", "and", "structural", "equivalence", ")", "and", "make", "assumptions", "about", "the", "relationship", "between", "the", "underlying", "network", "structure", "and", "the", "prediction", "task", ".", "For", "instance", ",", "spectral", "clustering", "makes", "a", "strong", "homophily", "assumption", "that", "graph", "cuts", "will", "be", "useful", "for", "classification", ".", "Such", "assumptions", "are", "reasonable", "in", "many", "scenarios", ",", "but", "unsatisfactory", "in", "effectively", "generalizing", "across", "diverse", "networks", ".", "Recent", "advancements", "in", "representational", "learning", "for", "natural", "language", "processing", "opened", "new", "ways", "for", "feature", "learning", "of", "discrete", "objects", "such", "as", "words", ".", "In", "particular", ",", "the", "Skip", "-", "gram", "model", "aims", "to", "learn", "continuous", "feature", "representations", "for", "words", "by", "optimizing", "a", "neighborhood", "preserving", "likelihood", "objective", ".", "The", "algorithm", "proceeds", "as", "follows", ":", "It", "scans", "over", "the", "words", "of", "a", "document", ",", "and", "for", "every", "word", "it", "aims", "to", "embed", "it", "such", "that", "the", "word", "’s", "features", "can", "predict", "nearby", "words", "(", "i.e.", ",", "words", "inside", "some", "context", "window", ")", ".", "The", "word", "feature", "representations", "are", "learned", "by", "optmizing", "the", "likelihood", "objective", "using", "SGD", "with", "negative", "sampling", ".", "The", "Skip", "-", "gram", "objective", "is", "based", "on", "the", "distributional", "hypothesis", "which", "states", "that", "words", "in", "similar", "contexts", "tend", "to", "have", "similar", "meanings", ".", "That", "is", ",", "similar", "words", "tend", "to", "appear", "in", "similar", "word", "neighborhoods", ".", "Inspired", "by", "the", "Skip", "-", "gram", "model", ",", "recent", "research", "established", "an", "analogy", "for", "networks", "by", "representing", "a", "network", "as", "a", "“", "document", "”", ".", "The", "same", "way", "as", "a", "document", "is", "an", "ordered", "sequence", "of", "words", ",", "one", "could", "sample", "sequences", "of", "nodes", "from", "the", "underlying", "network", "and", "turn", "a", "network", "into", "a", "ordered", "sequence", "of", "nodes", ".", "However", ",", "there", "are", "many", "possible", "sampling", "strategies", "for", "nodes", ",", "resulting", "in", "different", "learned", "feature", "representations", ".", "In", "fact", ",", "as", "we", "shall", "show", ",", "there", "is", "no", "clear", "winning", "sampling", "strategy", "that", "works", "across", "all", "networks", "and", "all", "prediction", "tasks", ".", "This", "is", "a", "major", "shortcoming", "of", "prior", "work", "which", "fail", "to", "offer", "any", "flexibility", "in", "sampling", "of", "nodes", "from", "a", "network", ".", "Our", "algorithm", "node2vec", "overcomes", "this", "limitation", "by", "designing", "a", "flexible", "objective", "that", "is", "not", "tied", "to", "a", "particular", "sampling", "strategy", "and", "provides", "parameters", "to", "tune", "the", "explored", "search", "space", "(", "see", "Section", "[", "reference", "]", ")", ".", "Finally", ",", "for", "both", "node", "and", "edge", "based", "prediction", "tasks", ",", "there", "is", "a", "body", "of", "recent", "work", "for", "supervised", "feature", "learning", "based", "on", "existing", "and", "novel", "graph", "-", "specific", "deep", "network", "architectures", ".", "These", "architectures", "directly", "minimize", "the", "loss", "function", "for", "a", "downstream", "prediction", "task", "using", "several", "layers", "of", "non", "-", "linear", "transformations", "which", "results", "in", "high", "accuracy", ",", "but", "at", "the", "cost", "of", "scalability", "due", "to", "high", "training", "time", "requirements", ".", "section", ":", "Feature", "learning", "framework", "We", "formulate", "feature", "learning", "in", "networks", "as", "a", "maximum", "likelihood", "optimization", "problem", ".", "Let", "be", "a", "given", "network", ".", "Our", "analysis", "is", "general", "and", "applies", "to", "any", "(", "un", ")", "directed", ",", "(", "un", ")", "weighted", "network", ".", "Let", "be", "the", "mapping", "function", "from", "nodes", "to", "feature", "representaions", "we", "aim", "to", "learn", "for", "a", "downstream", "prediction", "task", ".", "Here", "is", "a", "parameter", "specifying", "the", "number", "of", "dimensions", "of", "our", "feature", "representation", ".", "Equivalently", ",", "is", "a", "matrix", "of", "size", "parameters", ".", "For", "every", "source", "node", ",", "we", "define", "as", "a", "network", "neighborhood", "of", "node", "generated", "through", "a", "neighborhood", "sampling", "strategy", ".", "We", "proceed", "by", "extending", "the", "Skip", "-", "gram", "architecture", "to", "networks", ".", "We", "seek", "to", "optimize", "the", "following", "objective", "function", ",", "which", "maximizes", "the", "log", "-", "probability", "of", "observing", "a", "network", "neighborhood", "for", "a", "node", "conditioned", "on", "its", "feature", "representation", ",", "given", "by", ":", "In", "order", "to", "make", "the", "optimization", "problem", "tractable", ",", "we", "make", "two", "standard", "assumptions", ":", "Conditional", "independence", ".", "We", "factorize", "the", "likelihood", "by", "assuming", "that", "the", "likelihood", "of", "observing", "a", "neighborhood", "node", "is", "independent", "of", "observing", "any", "other", "neighborhood", "node", "given", "the", "feature", "representation", "of", "the", "source", ":", "Symmetry", "in", "feature", "space", ".", "A", "source", "node", "and", "neighborhood", "node", "have", "a", "symmetric", "effect", "over", "each", "other", "in", "feature", "space", ".", "Accordingly", ",", "we", "model", "the", "conditional", "likelihood", "of", "every", "source", "-", "neighborhood", "node", "pair", "as", "a", "softmax", "unit", "parametrized", "by", "a", "dot", "product", "of", "their", "features", ":", "With", "the", "above", "assumptions", ",", "the", "objective", "in", "Eq", ".", "[", "reference", "]", "simplifies", "to", ":", "The", "per", "-", "node", "partition", "function", ",", ",", "is", "expensive", "to", "compute", "for", "large", "networks", "and", "we", "approximate", "it", "using", "negative", "sampling", ".", "We", "optimize", "Eq", ".", "[", "reference", "]", "using", "stochastic", "gradient", "ascent", "over", "the", "model", "parameters", "defining", "the", "features", ".", "Feature", "learning", "methods", "based", "on", "the", "Skip", "-", "gram", "architecture", "have", "been", "originally", "developed", "in", "the", "context", "of", "natural", "language", ".", "Given", "the", "linear", "nature", "of", "text", ",", "the", "notion", "of", "a", "neighborhood", "can", "be", "naturally", "defined", "using", "a", "sliding", "window", "over", "consecutive", "words", ".", "Networks", ",", "however", ",", "are", "not", "linear", ",", "and", "thus", "a", "richer", "notion", "of", "a", "neighborhood", "is", "needed", ".", "To", "resolve", "this", "issue", ",", "we", "propose", "a", "randomized", "procedure", "that", "samples", "many", "different", "neighborhoods", "of", "a", "given", "source", "node", ".", "The", "neighborhoods", "are", "not", "restricted", "to", "just", "immediate", "neighbors", "but", "can", "have", "vastly", "different", "structures", "depending", "on", "the", "sampling", "strategy", ".", "subsection", ":", "Classic", "search", "strategies", "We", "view", "the", "problem", "of", "sampling", "neighborhoods", "of", "a", "source", "node", "as", "a", "form", "of", "local", "search", ".", "Figure", "[", "reference", "]", "shows", "a", "graph", ",", "where", "given", "a", "source", "node", "we", "aim", "to", "generate", "(", "sample", ")", "its", "neighborhood", ".", "Importantly", ",", "to", "be", "able", "to", "fairly", "compare", "different", "sampling", "strategies", ",", "we", "shall", "constrain", "the", "size", "of", "the", "neighborhood", "set", "to", "nodes", "and", "then", "sample", "multiple", "sets", "for", "a", "single", "node", ".", "Generally", ",", "there", "are", "two", "extreme", "sampling", "strategies", "for", "generating", "neighborhood", "set", "(", "s", ")", "of", "nodes", ":", "Breadth", "-", "first", "Sampling", "(", "BFS", ")", "The", "neighborhood", "is", "restricted", "to", "nodes", "which", "are", "immediate", "neighbors", "of", "the", "source", ".", "For", "example", ",", "in", "Figure", "[", "reference", "]", "for", "a", "neighborhood", "of", "size", ",", "BFS", "samples", "nodes", ",", ",", ".", "Depth", "-", "first", "Sampling", "(", "DFS", ")", "The", "neighborhood", "consists", "of", "nodes", "sequentially", "sampled", "at", "increasing", "distances", "from", "the", "source", "node", ".", "In", "Figure", "[", "reference", "]", ",", "DFS", "samples", ",", ",", ".", "The", "breadth", "-", "first", "and", "depth", "-", "first", "sampling", "represent", "extreme", "scenarios", "in", "terms", "of", "the", "search", "space", "they", "explore", "leading", "to", "interesting", "implications", "on", "the", "learned", "representations", ".", "In", "particular", ",", "prediction", "tasks", "on", "nodes", "in", "networks", "often", "shuttle", "between", "two", "kinds", "of", "similarities", ":", "homophily", "and", "structural", "equivalence", ".", "Under", "the", "homophily", "hypothesis", "nodes", "that", "are", "highly", "interconnected", "and", "belong", "to", "similar", "network", "clusters", "or", "communities", "should", "be", "embedded", "closely", "together", "(", "e.g.", ",", "nodes", "and", "in", "Figure", "[", "reference", "]", "belong", "to", "the", "same", "network", "community", ")", ".", "In", "contrast", ",", "under", "the", "structural", "equivalence", "hypothesis", "nodes", "that", "have", "similar", "structural", "roles", "in", "networks", "should", "be", "embedded", "closely", "together", "(", "e.g.", ",", "nodes", "and", "in", "Figure", "[", "reference", "]", "act", "as", "hubs", "of", "their", "corresponding", "communities", ")", ".", "Importantly", ",", "unlike", "homophily", ",", "structural", "equivalence", "does", "not", "emphasize", "connectivity", ";", "nodes", "could", "be", "far", "apart", "in", "the", "network", "and", "still", "have", "the", "same", "structural", "role", ".", "In", "real", "-", "world", ",", "these", "equivalence", "notions", "are", "not", "exclusive", ";", "networks", "commonly", "exhibit", "both", "behaviors", "where", "some", "nodes", "exhibit", "homophily", "while", "others", "reflect", "structural", "equivalence", ".", "We", "observe", "that", "BFS", "and", "DFS", "strategies", "play", "a", "key", "role", "in", "producing", "representations", "that", "reflect", "either", "of", "the", "above", "equivalences", ".", "In", "particular", ",", "the", "neighborhoods", "sampled", "by", "BFS", "lead", "to", "embeddings", "that", "correspond", "closely", "to", "structural", "equivalence", ".", "Intuitively", ",", "we", "note", "that", "in", "order", "to", "ascertain", "structural", "equivalence", ",", "it", "is", "often", "sufficient", "to", "characterize", "the", "local", "neighborhoods", "accurately", ".", "For", "example", ",", "structural", "equivalence", "based", "on", "network", "roles", "such", "as", "bridges", "and", "hubs", "can", "be", "inferred", "just", "by", "observing", "the", "immediate", "neighborhoods", "of", "each", "node", ".", "By", "restricting", "search", "to", "nearby", "nodes", ",", "BFS", "achieves", "this", "characterization", "and", "obtains", "a", "microscopic", "view", "of", "the", "neighborhood", "of", "every", "node", ".", "Additionally", ",", "in", "BFS", ",", "nodes", "in", "the", "sampled", "neighborhoods", "tend", "to", "repeat", "many", "times", ".", "This", "is", "also", "important", "as", "it", "reduces", "the", "variance", "in", "characterizing", "the", "distribution", "of", "1", "-", "hop", "nodes", "with", "respect", "the", "source", "node", ".", "However", ",", "a", "very", "small", "portion", "of", "the", "graph", "is", "explored", "for", "any", "given", ".", "The", "opposite", "is", "true", "for", "DFS", "which", "can", "explore", "larger", "parts", "of", "the", "network", "as", "it", "can", "move", "further", "away", "from", "the", "source", "node", "(", "with", "sample", "size", "being", "fixed", ")", ".", "In", "DFS", ",", "the", "sampled", "nodes", "more", "accurately", "reflect", "a", "macro", "-", "view", "of", "the", "neighborhood", "which", "is", "essential", "in", "inferring", "communities", "based", "on", "homophily", ".", "However", ",", "the", "issue", "with", "DFS", "is", "that", "it", "is", "important", "to", "not", "only", "infer", "which", "node", "-", "to", "-", "node", "dependencies", "exist", "in", "a", "network", ",", "but", "also", "to", "characterize", "the", "exact", "nature", "of", "these", "dependencies", ".", "This", "is", "hard", "given", "we", "have", "a", "constrain", "on", "the", "sample", "size", "and", "a", "large", "neighborhood", "to", "explore", ",", "resulting", "in", "high", "variance", ".", "Secondly", ",", "moving", "to", "much", "greater", "depths", "leads", "to", "complex", "dependencies", "since", "a", "sampled", "node", "may", "be", "far", "from", "the", "source", "and", "potentially", "less", "representative", ".", "subsection", ":", "node2vec", "Building", "on", "the", "above", "observations", ",", "we", "design", "a", "flexible", "neighborhood", "sampling", "strategy", "which", "allows", "us", "to", "smoothly", "interpolate", "between", "BFS", "and", "DFS", ".", "We", "achieve", "this", "by", "developing", "a", "flexible", "biased", "random", "walk", "procedure", "that", "can", "explore", "neighborhoods", "in", "a", "BFS", "as", "well", "as", "DFS", "fashion", ".", "subsubsection", ":", "Random", "Walks", "Formally", ",", "given", "a", "source", "node", ",", "we", "simulate", "a", "random", "walk", "of", "fixed", "length", ".", "Let", "denote", "the", "th", "node", "in", "the", "walk", ",", "starting", "with", ".", "Nodes", "are", "generated", "by", "the", "following", "distribution", ":", "where", "is", "the", "unnormalized", "transition", "probability", "between", "nodes", "and", ",", "and", "is", "the", "normalizing", "constant", ".", "subsubsection", ":", "Search", "bias", "The", "simplest", "way", "to", "bias", "our", "random", "walks", "would", "be", "to", "sample", "the", "next", "node", "based", "on", "the", "static", "edge", "weights", "i.e.", ",", ".", "(", "In", "case", "of", "unweighted", "graphs", ".", ")", "However", ",", "this", "does", "not", "allow", "us", "to", "account", "for", "the", "network", "structure", "and", "guide", "our", "search", "procedure", "to", "explore", "different", "types", "of", "network", "neighborhoods", ".", "Additionally", ",", "unlike", "BFS", "and", "DFS", "which", "are", "extreme", "sampling", "paradigms", "suited", "for", "structural", "equivalence", "and", "homophily", "respectively", ",", "our", "random", "walks", "should", "accommodate", "for", "the", "fact", "that", "these", "notions", "of", "equivalence", "are", "not", "competing", "or", "exclusive", ",", "and", "real", "-", "world", "networks", "commonly", "exhibit", "a", "mixture", "of", "both", ".", "We", "define", "a", "2", "order", "random", "walk", "with", "two", "parameters", "and", "which", "guide", "the", "walk", ":", "Consider", "a", "random", "walk", "that", "just", "traversed", "edge", "and", "now", "resides", "at", "node", "(", "Figure", "[", "reference", "]", ")", ".", "The", "walk", "now", "needs", "to", "decide", "on", "the", "next", "step", "so", "it", "evaluates", "the", "transition", "probabilities", "on", "edges", "leading", "from", ".", "We", "set", "the", "unnormalized", "transition", "probability", "to", ",", "where", "and", "denotes", "the", "shortest", "path", "distance", "between", "nodes", "and", ".", "Note", "that", "must", "be", "one", "of", ",", "and", "hence", ",", "the", "two", "parameters", "are", "necessary", "and", "sufficient", "to", "guide", "the", "walk", ".", "Intuitively", ",", "parameters", "and", "control", "how", "fast", "the", "walk", "explores", "and", "leaves", "the", "neighborhood", "of", "starting", "node", ".", "In", "particular", ",", "the", "parameters", "allow", "our", "search", "procedure", "to", "(", "approximately", ")", "interpolate", "between", "BFS", "and", "DFS", "and", "thereby", "reflect", "an", "affinity", "for", "different", "notions", "of", "node", "equivalences", ".", "Return", "parameter", ",", "p.", "Parameter", "controls", "the", "likelihood", "of", "immediately", "revisiting", "a", "node", "in", "the", "walk", ".", "Setting", "it", "to", "a", "high", "value", "(", ")", "ensures", "that", "we", "are", "less", "likely", "to", "sample", "an", "already", "-", "visited", "node", "in", "the", "following", "two", "steps", "(", "unless", "the", "next", "node", "in", "the", "walk", "had", "no", "other", "neighbor", ")", ".", "This", "strategy", "encourages", "moderate", "exploration", "and", "avoids", "-", "hop", "redundancy", "in", "sampling", ".", "On", "the", "other", "hand", ",", "if", "is", "low", "(", ")", ",", "it", "would", "lead", "the", "walk", "to", "backtrack", "a", "step", "(", "Figure", "[", "reference", "]", ")", "and", "this", "would", "keep", "the", "walk", "“", "local", "”", "close", "to", "the", "starting", "node", ".", "In", "-", "out", "parameter", ",", "q.", "Parameter", "allows", "the", "search", "to", "differentiate", "between", "“", "inward", "”", "and", "“", "outward", "”", "nodes", ".", "Going", "back", "to", "Figure", "[", "reference", "]", ",", "if", ",", "the", "random", "walk", "is", "biased", "towards", "nodes", "close", "to", "node", ".", "Such", "walks", "obtain", "a", "local", "view", "of", "the", "underlying", "graph", "with", "respect", "to", "the", "start", "node", "in", "the", "walk", "and", "approximate", "BFS", "behavior", "in", "the", "sense", "that", "our", "samples", "comprise", "of", "nodes", "within", "a", "small", "locality", ".", "In", "contrast", ",", "if", ",", "the", "walk", "is", "more", "inclined", "to", "visit", "nodes", "which", "are", "further", "away", "from", "the", "node", ".", "Such", "behavior", "is", "reflective", "of", "DFS", "which", "encourages", "outward", "exploration", ".", "However", ",", "an", "essential", "difference", "here", "is", "that", "we", "achieve", "DFS", "-", "like", "exploration", "within", "the", "random", "walk", "framework", ".", "Hence", ",", "the", "sampled", "nodes", "are", "not", "at", "strictly", "increasing", "distances", "from", "a", "given", "source", "node", ",", "but", "in", "turn", ",", "we", "benefit", "from", "tractable", "preprocessing", "and", "superior", "sampling", "efficiency", "of", "random", "walks", ".", "Note", "that", "by", "setting", "to", "be", "a", "function", "of", "the", "preceeding", "node", "in", "the", "walk", ",", "the", "random", "walks", "are", "2", "order", "Markovian", ".", "Benefits", "of", "random", "walks", ".", "There", "are", "several", "benefits", "of", "random", "walks", "over", "pure", "BFS", "/", "DFS", "approaches", ".", "Random", "walks", "are", "computationally", "efficient", "in", "terms", "of", "both", "space", "and", "time", "requirements", ".", "The", "space", "complexity", "to", "store", "the", "immediate", "neighbors", "of", "every", "node", "in", "the", "graph", "is", ".", "For", "2", "order", "random", "walks", ",", "it", "is", "helpful", "to", "store", "the", "interconnections", "between", "the", "neighbors", "of", "every", "node", ",", "which", "incurs", "a", "space", "complexity", "of", "where", "is", "the", "average", "degree", "of", "the", "graph", "and", "is", "usually", "small", "for", "real", "-", "world", "networks", ".", "The", "other", "key", "advantage", "of", "random", "walks", "over", "classic", "search", "-", "based", "sampling", "strategies", "is", "its", "time", "complexity", ".", "In", "particular", ",", "by", "imposing", "graph", "connectivity", "in", "the", "sample", "generation", "process", ",", "random", "walks", "provide", "a", "convenient", "mechanism", "to", "increase", "the", "effective", "sampling", "rate", "by", "reusing", "samples", "across", "different", "source", "nodes", ".", "By", "simulating", "a", "random", "walk", "of", "length", "we", "can", "generate", "samples", "for", "nodes", "at", "once", "due", "to", "the", "Markovian", "nature", "of", "the", "random", "walk", ".", "Hence", ",", "our", "effective", "complexity", "is", "per", "sample", ".", "For", "example", ",", "in", "Figure", "[", "reference", "]", "we", "sample", "a", "random", "walk", "of", "length", ",", "which", "results", "in", ",", "and", ".", "Note", "that", "sample", "reuse", "can", "introduce", "some", "bias", "in", "the", "overall", "procedure", ".", "However", ",", "we", "observe", "that", "it", "greatly", "improves", "the", "efficiency", ".", "subsubsection", ":", "The", "node2vec", "algorithm", "[", "h", "]", "LearnFeatures", "(", "Graph", ",", "Dimensions", ",", "Walks", "per", "node", ",", "Walk", "length", ",", "Context", "size", ",", "Return", ",", "In", "-", "out", ")", "PreprocessModifiedWeights", "(", ")", "Initialize", "to", "Empty", "to", "nodes", "node2vecWalk", "(", ")", "Append", "to", "StochasticGradientDescent", "(", ",", ",", ")", "return", "node2vecWalk", "(", "Graph", ",", "Start", "node", ",", "Length", ")", "Inititalize", "to", "to", "GetNeighbors", "(", ",", ")", "AliasSample", "(", ")", "Append", "to", "return", "The", "node2vec", "algorithm", ".", "The", "pseudocode", "for", "node2vec", ",", "is", "given", "in", "Algorithm", "[", "reference", "]", ".", "In", "any", "random", "walk", ",", "there", "is", "an", "implicit", "bias", "due", "to", "the", "choice", "of", "the", "start", "node", ".", "Since", "we", "learn", "representations", "for", "all", "nodes", ",", "we", "offset", "this", "bias", "by", "simulating", "random", "walks", "of", "fixed", "length", "starting", "from", "every", "node", ".", "At", "every", "step", "of", "the", "walk", ",", "sampling", "is", "done", "based", "on", "the", "transition", "probabilities", ".", "The", "transition", "probabilities", "for", "the", "2", "order", "Markov", "chain", "can", "be", "precomputed", "and", "hence", ",", "sampling", "of", "nodes", "while", "simulating", "the", "random", "walk", "can", "be", "done", "efficiently", "in", "time", "using", "alias", "sampling", ".", "The", "three", "phases", "of", "node2vec", ",", "i.e.", ",", "preprocessing", "to", "compute", "transition", "probabilities", ",", "random", "walk", "simulations", "and", "optimization", "using", "SGD", ",", "are", "executed", "sequentially", ".", "Each", "phase", "is", "parallelizable", "and", "executed", "asynchronously", ",", "contributing", "to", "the", "overall", "scalability", "of", "node2vec", ".", "node2vec", "is", "available", "at", ":", ".", "subsection", ":", "Learning", "edge", "features", "The", "node2vec", "algorithm", "provides", "a", "semi", "-", "supervised", "method", "to", "learn", "rich", "feature", "representations", "for", "nodes", "in", "a", "network", ".", "However", ",", "we", "are", "often", "interested", "in", "prediction", "tasks", "involving", "pairs", "of", "nodes", "instead", "of", "individual", "nodes", ".", "For", "instance", ",", "in", "link", "prediction", ",", "we", "predict", "whether", "a", "link", "exists", "between", "two", "nodes", "in", "a", "network", ".", "Since", "our", "random", "walks", "are", "naturally", "based", "on", "the", "connectivity", "structure", "between", "nodes", "in", "the", "underlying", "network", ",", "we", "extend", "them", "to", "pairs", "of", "nodes", "using", "a", "bootstrapping", "approach", "over", "the", "feature", "representations", "of", "the", "individual", "nodes", ".", "Given", "two", "nodes", "and", ",", "we", "define", "a", "binary", "operator", "over", "the", "corresponding", "feature", "vectors", "and", "in", "order", "to", "generate", "a", "representation", "such", "that", "where", "is", "the", "representation", "size", "for", "the", "pair", ".", "We", "want", "our", "operators", "to", "be", "generally", "defined", "for", "any", "pair", "of", "nodes", ",", "even", "if", "an", "edge", "does", "not", "exist", "between", "the", "pair", "since", "doing", "so", "makes", "the", "representations", "useful", "for", "link", "prediction", "where", "our", "test", "set", "contains", "both", "true", "and", "false", "edges", "(", "i.e.", ",", "do", "not", "exist", ")", ".", "We", "consider", "several", "choices", "for", "the", "operator", "such", "that", "which", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "section", ":", "Experiments", "The", "objective", "in", "Eq", ".", "[", "reference", "]", "is", "independent", "of", "any", "downstream", "task", "and", "the", "flexibility", "in", "exploration", "offered", "by", "node2vec", "lends", "the", "learned", "feature", "representations", "to", "a", "wide", "variety", "of", "network", "analysis", "settings", "discussed", "below", ".", "subsection", ":", "Case", "Study", ":", "Les", "Misérables", "network", "In", "Section", "[", "reference", "]", "we", "observed", "that", "BFS", "and", "DFS", "strategies", "represent", "extreme", "ends", "on", "the", "spectrum", "of", "embedding", "nodes", "based", "on", "the", "principles", "of", "homophily", "(", "i.e.", ",", "network", "communities", ")", "and", "structural", "equivalence", "(", "i.e.", ",", "structural", "roles", "of", "nodes", ")", ".", "We", "now", "aim", "to", "empirically", "demonstrate", "this", "fact", "and", "show", "that", "node2vec", "in", "fact", "can", "discover", "embeddings", "that", "obey", "both", "principles", ".", "We", "use", "a", "network", "where", "nodes", "correspond", "to", "characters", "in", "the", "novel", "Les", "Misérables", "and", "edges", "connect", "coappearing", "characters", ".", "The", "network", "has", "77", "nodes", "and", "254", "edges", ".", "We", "set", "and", "run", "node2vec", "to", "learn", "feature", "representation", "for", "every", "node", "in", "the", "network", ".", "The", "feature", "representations", "are", "clustered", "using", "-", "means", ".", "We", "then", "visualize", "the", "original", "network", "in", "two", "dimensions", "with", "nodes", "now", "assigned", "colors", "based", "on", "their", "clusters", ".", "Figure", "[", "reference", "]", "(", "top", ")", "shows", "the", "example", "when", "we", "set", ".", "Notice", "how", "regions", "of", "the", "network", "(", "i.e.", ",", "network", "communities", ")", "are", "colored", "using", "the", "same", "color", ".", "In", "this", "setting", "node2vec", "discovers", "clusters", "/", "communities", "of", "characters", "that", "frequently", "interact", "with", "each", "other", "in", "the", "major", "sub", "-", "plots", "of", "the", "novel", ".", "Since", "the", "edges", "between", "characters", "are", "based", "on", "coappearances", ",", "we", "can", "conclude", "this", "characterization", "closely", "relates", "with", "homophily", ".", "In", "order", "to", "discover", "which", "nodes", "have", "the", "same", "structural", "roles", "we", "use", "the", "same", "network", "but", "set", ",", "use", "node2vec", "to", "get", "node", "features", "and", "then", "cluster", "the", "nodes", "based", "on", "the", "obtained", "features", ".", "Here", "node2vec", "obtains", "a", "complementary", "assignment", "of", "node", "to", "clusters", "such", "that", "the", "colors", "correspond", "to", "structural", "equivalence", "as", "illustrated", "in", "Figure", "[", "reference", "]", "(", "bottom", ")", ".", "For", "instance", ",", "node2vec", "embeds", "blue", "-", "colored", "nodes", "close", "together", ".", "These", "nodes", "represent", "characters", "that", "act", "as", "bridges", "between", "different", "sub", "-", "plots", "of", "the", "novel", ".", "Similarly", ",", "the", "yellow", "nodes", "mostly", "represent", "characters", "that", "are", "at", "the", "periphery", "and", "have", "limited", "interactions", ".", "One", "could", "assign", "alternate", "semantic", "interpretations", "to", "these", "clusters", "of", "nodes", ",", "but", "the", "key", "takeaway", "is", "that", "node2vec", "is", "not", "tied", "to", "a", "particular", "notion", "of", "equivalence", ".", "As", "we", "show", "through", "our", "experiments", ",", "these", "equivalence", "notions", "are", "commonly", "exhibited", "in", "most", "real", "-", "world", "networks", "and", "have", "a", "significant", "impact", "on", "the", "performance", "of", "the", "learned", "representations", "for", "prediction", "tasks", ".", "subsection", ":", "Experimental", "setup", "Our", "experiments", "evaluate", "the", "feature", "representations", "obtained", "through", "node2vec", "on", "standard", "supervised", "learning", "tasks", ":", "multi", "-", "label", "classification", "for", "nodes", "and", "link", "prediction", "for", "edges", ".", "For", "both", "tasks", ",", "we", "evaluate", "the", "performance", "of", "node2vec", "against", "the", "following", "feature", "learning", "algorithms", ":", "Spectral", "clustering", ":", "This", "is", "a", "matrix", "factorization", "approach", "in", "which", "we", "take", "the", "top", "eigenvectors", "of", "the", "normalized", "Laplacian", "matrix", "of", "graph", "as", "the", "feature", "vector", "representations", "for", "nodes", ".", "DeepWalk", ":", "This", "approach", "learns", "-", "dimensional", "feature", "representations", "by", "simulating", "uniform", "random", "walks", ".", "The", "sampling", "strategy", "in", "DeepWalk", "can", "be", "seen", "as", "a", "special", "case", "of", "node2vec", "with", "and", ".", "LINE", ":", "This", "approach", "learns", "-", "dimensional", "feature", "representations", "in", "two", "separate", "phases", ".", "In", "the", "first", "phase", ",", "it", "learns", "dimensions", "by", "BFS", "-", "style", "simulations", "over", "immediate", "neighbors", "of", "nodes", ".", "In", "the", "second", "phase", ",", "it", "learns", "the", "next", "dimensions", "by", "sampling", "nodes", "strictly", "at", "a", "2", "-", "hop", "distance", "from", "the", "source", "nodes", ".", "We", "exclude", "other", "matrix", "factorization", "approaches", "which", "have", "already", "been", "shown", "to", "be", "inferior", "to", "DeepWalk", ".", "We", "also", "exclude", "a", "recent", "approach", ",", "GraRep", ",", "that", "generalizes", "LINE", "to", "incorporate", "information", "from", "network", "neighborhoods", "beyond", "2", "-", "hops", ",", "but", "is", "unable", "to", "efficiently", "scale", "to", "large", "networks", ".", "In", "contrast", "to", "the", "setup", "used", "in", "prior", "work", "for", "evaluating", "sampling", "-", "based", "feature", "learning", "algorithms", ",", "we", "generate", "an", "equal", "number", "of", "samples", "for", "each", "method", "and", "then", "evaluate", "the", "quality", "of", "the", "obtained", "features", "on", "the", "prediction", "task", ".", "In", "doing", "so", ",", "we", "discount", "for", "performance", "gain", "observed", "purely", "because", "of", "the", "implementation", "language", "(", "C", "/", "C", "++/", "Python", ")", "since", "it", "is", "secondary", "to", "the", "algorithm", ".", "Thus", ",", "in", "the", "sampling", "phase", ",", "the", "parameters", "for", "DeepWalk", ",", "LINE", "and", "node2vec", "are", "set", "such", "that", "they", "generate", "equal", "number", "of", "samples", "at", "runtime", ".", "As", "an", "example", ",", "if", "is", "the", "overall", "sampling", "budget", ",", "then", "the", "node2vec", "parameters", "satisfy", ".", "In", "the", "optimization", "phase", ",", "all", "these", "benchmarks", "optimize", "using", "SGD", "with", "two", "key", "differences", "that", "we", "correct", "for", ".", "First", ",", "DeepWalk", "uses", "hierarchical", "sampling", "to", "approximate", "the", "softmax", "probabilities", "with", "an", "objective", "similar", "to", "the", "one", "use", "by", "node2vec", ".", "However", ",", "hierarchical", "softmax", "is", "inefficient", "when", "compared", "with", "negative", "sampling", ".", "Hence", ",", "keeping", "everything", "else", "the", "same", ",", "we", "switch", "to", "negative", "sampling", "in", "DeepWalk", "which", "is", "also", "the", "de", "facto", "approximation", "in", "node2vec", "and", "LINE", ".", "Second", ",", "both", "node2vec", "and", "DeepWalk", "have", "a", "parameter", "for", "the", "number", "of", "context", "neighborhood", "nodes", "to", "optimize", "for", "and", "the", "greater", "the", "number", ",", "the", "more", "rounds", "of", "optimization", "are", "required", ".", "This", "parameter", "is", "set", "to", "unity", "for", "LINE", ",", "but", "since", "LINE", "completes", "a", "single", "epoch", "quicker", "than", "other", "approaches", ",", "we", "let", "it", "run", "for", "epochs", ".", "The", "parameter", "settings", "used", "for", "node2vec", "are", "in", "line", "with", "typical", "values", "used", "for", "DeepWalk", "and", "LINE", ".", "Specifically", ",", "we", "set", ",", ",", ",", ",", "and", "the", "optimization", "is", "run", "for", "a", "single", "epoch", ".", "We", "repeat", "our", "experiments", "for", "random", "seed", "initializations", ",", "and", "our", "results", "are", "statistically", "significant", "with", "a", "p", "-", "value", "of", "less", "than", "0.01", ".", "The", "best", "in", "-", "out", "and", "return", "hyperparameters", "were", "learned", "using", "10", "-", "fold", "cross", "-", "validation", "on", "10", "%", "labeled", "data", "with", "a", "grid", "search", "over", ".", "subsection", ":", "Multi", "-", "label", "classification", "In", "the", "multi", "-", "label", "classification", "setting", ",", "every", "node", "is", "assigned", "one", "or", "more", "labels", "from", "a", "finite", "set", ".", "During", "the", "training", "phase", ",", "we", "observe", "a", "certain", "fraction", "of", "nodes", "and", "all", "their", "labels", ".", "The", "task", "is", "to", "predict", "the", "labels", "for", "the", "remaining", "nodes", ".", "This", "is", "a", "challenging", "task", "especially", "if", "is", "large", ".", "We", "utilize", "the", "following", "datasets", ":", "BlogCatalog", ":", "This", "is", "a", "network", "of", "social", "relationships", "of", "the", "bloggers", "listed", "on", "the", "BlogCatalog", "website", ".", "The", "labels", "represent", "blogger", "interests", "inferred", "through", "the", "meta", "-", "data", "provided", "by", "the", "bloggers", ".", "The", "network", "has", "10", ",", "312", "nodes", ",", "333", ",", "983", "edges", ",", "and", "39", "different", "labels", ".", "Protein", "-", "Protein", "Interactions", "(", "PPI", ")", ":", "We", "use", "a", "subgraph", "of", "the", "PPI", "network", "for", "Homo", "Sapiens", ".", "The", "subgraph", "corresponds", "to", "the", "graph", "induced", "by", "nodes", "for", "which", "we", "could", "obtain", "labels", "from", "the", "hallmark", "gene", "sets", "and", "represent", "biological", "states", ".", "The", "network", "has", "3", ",", "890", "nodes", ",", "76", ",", "584", "edges", ",", "and", "50", "different", "labels", ".", "Wikipedia", ":", "This", "is", "a", "cooccurrence", "network", "of", "words", "appearing", "in", "the", "first", "million", "bytes", "of", "the", "Wikipedia", "dump", ".", "The", "labels", "represent", "the", "Part", "-", "of", "-", "Speech", "(", "POS", ")", "tags", "inferred", "using", "the", "Stanford", "POS", "-", "Tagger", ".", "The", "network", "has", "4", ",", "777", "nodes", ",", "184", ",", "812", "edges", ",", "and", "40", "different", "labels", ".", "All", "these", "networks", "exhibit", "a", "fair", "mix", "of", "homophilic", "and", "structural", "equivalences", ".", "For", "example", ",", "we", "expect", "the", "social", "network", "of", "bloggers", "to", "exhibit", "strong", "homophily", "-", "based", "relationships", ";", "however", ",", "there", "might", "also", "be", "some", "“", "familiar", "strangers", "”", ",", "i.e.", ",", "bloggers", "that", "do", "not", "interact", "but", "share", "interests", "and", "hence", "are", "structurally", "equivalent", "nodes", ".", "The", "biological", "states", "of", "proteins", "in", "a", "protein", "-", "protein", "interaction", "network", "also", "exhibit", "both", "types", "of", "equivalences", ".", "For", "example", ",", "they", "exhibit", "structural", "equivalence", "when", "proteins", "perform", "functions", "complementary", "to", "those", "of", "neighboring", "proteins", ",", "and", "at", "other", "times", ",", "they", "organize", "based", "on", "homophily", "in", "assisting", "neighboring", "proteins", "in", "performing", "similar", "functions", ".", "The", "word", "cooccurence", "network", "is", "fairly", "dense", ",", "since", "edges", "exist", "between", "words", "cooccuring", "in", "a", "2", "-", "length", "window", "in", "the", "Wikipedia", "corpus", ".", "Hence", ",", "words", "having", "the", "same", "POS", "tags", "are", "not", "hard", "to", "find", ",", "lending", "a", "high", "degree", "of", "homophily", ".", "At", "the", "same", "time", ",", "we", "expect", "some", "structural", "equivalence", "in", "the", "POS", "tags", "due", "to", "syntactic", "grammar", "patterns", "such", "as", "nouns", "following", "determiners", ",", "punctuations", "succeeding", "nouns", "etc", ".", "Experimental", "results", ".", "The", "node", "feature", "representations", "are", "input", "to", "a", "one", "-", "vs", "-", "rest", "logistic", "regression", "classifier", "with", "L2", "regularization", ".", "The", "train", "and", "test", "data", "is", "split", "equally", "over", "10", "random", "instances", ".", "We", "use", "the", "Macro", "-", "F", "scores", "for", "comparing", "performance", "in", "Table", "[", "reference", "]", "and", "the", "relative", "performance", "gain", "is", "over", "the", "closest", "benchmark", ".", "The", "trends", "are", "similar", "for", "Micro", "-", "F", "and", "accuracy", "and", "are", "not", "shown", ".", "From", "the", "results", ",", "it", "is", "evident", "we", "can", "see", "how", "the", "added", "flexibility", "in", "exploring", "neighborhoods", "allows", "node2vec", "to", "outperform", "the", "other", "benchmark", "algorithms", ".", "In", "BlogCatalog", ",", "we", "can", "discover", "the", "right", "mix", "of", "homophily", "and", "structural", "equivalence", "by", "setting", "parameters", "and", "to", "low", "values", ",", "giving", "us", "22.3", "%", "gain", "over", "DeepWalk", "and", "229.2", "%", "gain", "over", "LINE", "in", "Macro", "-", "F", "scores", ".", "LINE", "showed", "worse", "performance", "than", "expected", ",", "which", "can", "be", "explained", "by", "its", "inability", "to", "reuse", "samples", ",", "a", "feat", "that", "can", "be", "easily", "done", "using", "the", "random", "walk", "methods", ".", "Even", "in", "our", "other", "two", "networks", ",", "where", "we", "have", "a", "mix", "of", "equivalences", "present", ",", "the", "semi", "-", "supervised", "nature", "of", "node2vec", "can", "help", "us", "infer", "the", "appropriate", "degree", "of", "exploration", "necessary", "for", "feature", "learning", ".", "In", "the", "case", "of", "PPI", "network", ",", "the", "best", "exploration", "strategy", "(", ",", ")", "turns", "out", "to", "be", "virtually", "indistinguishable", "from", "DeepWalk", "’s", "uniform", "(", ",", ")", "exploration", "giving", "us", "only", "a", "slight", "edge", "over", "DeepWalk", "by", "avoiding", "redudancy", "in", "already", "visited", "nodes", "through", "a", "high", "value", ",", "but", "a", "convincing", "23.8", "%", "gain", "over", "LINE", "in", "Macro", "-", "F", "scores", ".", "However", ",", "in", "general", ",", "the", "uniform", "random", "walks", "can", "be", "much", "worse", "than", "the", "exploration", "strategy", "learned", "by", "node2vec", ".", "As", "we", "can", "see", "in", "the", "Wikipedia", "word", "cooccurrence", "network", ",", "uniform", "walks", "can", "not", "guide", "the", "search", "procedure", "towards", "the", "best", "samples", "and", "hence", ",", "we", "achieve", "a", "gain", "of", "21.8", "%", "over", "DeepWalk", "and", "33.2", "%", "over", "LINE", ".", "For", "a", "more", "fine", "-", "grained", "analysis", ",", "we", "also", "compare", "performance", "while", "varying", "the", "train", "-", "test", "split", "from", "10", "%", "to", "90", "%", ",", "while", "learning", "parameters", "and", "on", "10", "%", "of", "the", "data", "as", "before", ".", "For", "brevity", ",", "we", "summarize", "the", "results", "for", "the", "Micro", "-", "F", "and", "Macro", "-", "F", "scores", "graphically", "in", "Figure", "[", "reference", "]", ".", "Here", "we", "make", "similar", "observations", ".", "All", "methods", "significantly", "outperform", "Spectral", "clustering", ",", "DeepWalk", "outperforms", "LINE", ",", "node2vec", "consistently", "outperforms", "LINE", "and", "achieves", "large", "improvement", "over", "DeepWalk", "across", "domains", ".", "For", "example", ",", "we", "achieve", "the", "biggest", "improvement", "over", "DeepWalk", "of", "26.7", "%", "on", "BlogCatalog", "at", "70", "%", "labeled", "data", ".", "In", "the", "worst", "case", ",", "the", "search", "phase", "has", "little", "bearing", "on", "learned", "representations", "in", "which", "case", "node2vec", "is", "equivalent", "to", "DeepWalk", ".", "Similarly", ",", "the", "improvements", "are", "even", "more", "striking", "when", "compared", "to", "LINE", ",", "where", "in", "addition", "to", "drastic", "gain", "(", "over", "200", "%", ")", "on", "BlogCatalog", ",", "we", "observe", "high", "magnitude", "improvements", "upto", "41.1", "%", "on", "other", "datasets", "such", "as", "PPI", "while", "training", "on", "just", "10", "%", "labeled", "data", ".", "[", "b", "]", "0.70", "[", "b", "]", "0.28", "subsection", ":", "Parameter", "sensitivity", "The", "node2vec", "algorithm", "involves", "a", "number", "of", "parameters", "and", "in", "Figure", "[", "reference", "]", ",", "we", "examine", "how", "the", "different", "choices", "of", "parameters", "affect", "the", "performance", "of", "node2vec", "on", "the", "BlogCatalog", "dataset", "using", "a", "50", "-", "50", "split", "between", "labeled", "and", "unlabeled", "data", ".", "Except", "for", "the", "parameter", "being", "tested", ",", "all", "other", "parameters", "assume", "default", "values", ".", "The", "default", "values", "for", "and", "are", "set", "to", "unity", ".", "We", "measure", "the", "Macro", "-", "F", "score", "as", "a", "function", "of", "parameters", "and", ".", "The", "performance", "of", "node2vec", "improves", "as", "the", "in", "-", "out", "parameter", "and", "the", "return", "parameter", "decrease", ".", "This", "increase", "in", "performance", "can", "be", "based", "on", "the", "homophilic", "and", "structural", "equivalences", "we", "expect", "to", "see", "in", "BlogCatalog", ".", "While", "a", "low", "encourages", "outward", "exploration", ",", "it", "is", "balanced", "by", "a", "low", "which", "ensures", "that", "the", "walk", "does", "not", "go", "too", "far", "from", "the", "start", "node", ".", "We", "also", "examine", "how", "the", "number", "of", "features", "and", "the", "node", "’s", "neighborhood", "parameters", "(", "number", "of", "walks", ",", "walk", "length", ",", "and", "neighborhood", "size", ")", "affect", "the", "performance", ".", "We", "observe", "that", "performance", "tends", "to", "saturate", "once", "the", "dimensions", "of", "the", "representations", "reaches", "around", "100", ".", "Similarly", ",", "we", "observe", "that", "increasing", "the", "number", "and", "length", "of", "walks", "per", "source", "improves", "performance", ",", "which", "is", "not", "surprising", "since", "we", "have", "a", "greater", "overall", "sampling", "budget", "to", "learn", "representations", ".", "Both", "these", "parameters", "have", "a", "relatively", "high", "impact", "on", "the", "performance", "of", "the", "method", ".", "Interestingly", ",", "the", "context", "size", ",", "also", "improves", "performance", "at", "the", "cost", "of", "increased", "optimization", "time", ".", "However", ",", "the", "performance", "differences", "are", "not", "that", "large", "in", "this", "case", ".", "subsection", ":", "Perturbation", "Analysis", "For", "many", "real", "-", "world", "networks", ",", "we", "do", "not", "have", "access", "to", "accurate", "information", "about", "the", "network", "structure", ".", "We", "performed", "a", "perturbation", "study", "where", "we", "analyzed", "the", "performance", "of", "node2vec", "for", "two", "imperfect", "information", "scenarios", "related", "to", "the", "edge", "structure", "in", "the", "BlogCatalog", "network", ".", "In", "the", "first", "scenario", ",", "we", "measure", "performace", "as", "a", "function", "of", "the", "fraction", "of", "missing", "edges", "(", "relative", "to", "the", "full", "network", ")", ".", "The", "missing", "edges", "are", "chosen", "randomly", ",", "subject", "to", "the", "constraint", "that", "the", "number", "of", "connected", "components", "in", "the", "network", "remains", "fixed", ".", "As", "we", "can", "see", "in", "Figure", "[", "reference", "]", "(", "top", ")", ",", "the", "decrease", "in", "Macro", "-", "F", "score", "as", "the", "fraction", "of", "missing", "edges", "increases", "is", "roughly", "linear", "with", "a", "small", "slope", ".", "Robustness", "to", "missing", "edges", "in", "the", "network", "is", "especially", "important", "in", "cases", "where", "the", "graphs", "are", "evolving", "over", "time", "(", "e.g.", ",", "citation", "networks", ")", ",", "or", "where", "network", "construction", "is", "expensive", "(", "e.g.", ",", "biological", "networks", ")", ".", "In", "the", "second", "perturbation", "setting", ",", "we", "have", "noisy", "edges", "between", "randomly", "selected", "pairs", "of", "nodes", "in", "the", "network", ".", "As", "shown", "in", "Figure", "[", "reference", "]", "(", "bottom", ")", ",", "the", "performance", "of", "node2vec", "declines", "slightly", "faster", "initially", "when", "compared", "with", "the", "setting", "of", "missing", "edges", ",", "however", ",", "the", "rate", "of", "decrease", "in", "Macro", "-", "F", "score", "gradually", "slows", "down", "over", "time", ".", "Again", ",", "the", "robustness", "of", "node2vec", "to", "false", "edges", "is", "useful", "in", "several", "situations", "such", "as", "sensor", "networks", "where", "the", "measurements", "used", "for", "constructing", "the", "network", "are", "noisy", ".", "subsection", ":", "Scalability", "To", "test", "for", "scalability", ",", "we", "learn", "node", "representations", "using", "node2vec", "with", "default", "parameter", "values", "for", "Erdos", "-", "Renyi", "graphs", "with", "increasing", "sizes", "from", "100", "to", "1", ",", "000", ",", "000", "nodes", "and", "constant", "average", "degree", "of", "10", ".", "In", "Figure", "[", "reference", "]", ",", "we", "empirically", "observe", "that", "node2vec", "scales", "linearly", "with", "increase", "in", "number", "of", "nodes", "generating", "representations", "for", "one", "million", "nodes", "in", "less", "than", "four", "hours", ".", "The", "sampling", "procedure", "comprises", "of", "preprocessing", "for", "computing", "transition", "probabilities", "for", "our", "walk", "(", "negligibly", "small", ")", "and", "simulation", "of", "random", "walks", ".", "The", "optimization", "phase", "is", "made", "efficient", "using", "negative", "sampling", "and", "asynchronous", "SGD", ".", "Many", "ideas", "from", "prior", "work", "serve", "as", "useful", "pointers", "in", "making", "the", "sampling", "procedure", "computationally", "efficient", ".", "We", "showed", "how", "random", "walks", ",", "also", "used", "in", "DeepWalk", ",", "allow", "the", "sampled", "nodes", "to", "be", "reused", "as", "neighborhoods", "for", "different", "source", "nodes", "appearing", "in", "the", "walk", ".", "Alias", "sampling", "allows", "our", "walks", "to", "generalize", "to", "weighted", "networks", ",", "with", "little", "preprocessing", ".", "Though", "we", "are", "free", "to", "set", "the", "search", "parameters", "based", "on", "the", "underlying", "task", "and", "domain", "at", "no", "additional", "cost", ",", "learning", "the", "best", "settings", "of", "our", "search", "parameters", "adds", "an", "overhead", ".", "However", ",", "as", "our", "experiments", "confirm", ",", "this", "overhead", "is", "minimal", "since", "node2vec", "is", "semi", "-", "supervised", "and", "hence", ",", "can", "learn", "these", "parameters", "efficiently", "with", "very", "little", "labeled", "data", ".", "subsection", ":", "Link", "prediction", "In", "link", "prediction", ",", "we", "are", "given", "a", "network", "with", "a", "certain", "fraction", "of", "edges", "removed", ",", "and", "we", "would", "like", "to", "predict", "these", "missing", "edges", ".", "We", "generate", "the", "labeled", "dataset", "of", "edges", "as", "follows", ":", "To", "obtain", "positive", "examples", ",", "we", "remove", "50", "%", "of", "edges", "chosen", "randomly", "from", "the", "network", "while", "ensuring", "that", "the", "residual", "network", "obtained", "after", "the", "edge", "removals", "is", "connected", ",", "and", "to", "generate", "negative", "examples", ",", "we", "randomly", "sample", "an", "equal", "number", "of", "node", "pairs", "from", "the", "network", "which", "have", "no", "edge", "connecting", "them", ".", "Since", "none", "of", "feature", "learning", "algorithms", "have", "been", "previously", "used", "for", "link", "prediction", ",", "we", "additionally", "evaluate", "node2vec", "against", "some", "popular", "heuristic", "scores", "that", "achieve", "good", "performance", "in", "link", "prediction", ".", "The", "scores", "we", "consider", "are", "defined", "in", "terms", "of", "the", "neighborhood", "sets", "of", "the", "nodes", "constituting", "the", "pair", "(", "see", "Table", "[", "reference", "]", ")", ".", "We", "test", "our", "benchmarks", "on", "the", "following", "datasets", ":", "Facebook", ":", "In", "the", "Facebook", "network", ",", "nodes", "represent", "users", ",", "and", "edges", "represent", "a", "friendship", "relation", "between", "any", "two", "users", ".", "The", "network", "has", "4", ",", "039", "nodes", "and", "88", ",", "234", "edges", ".", "Protein", "-", "Protein", "Interactions", "(", "PPI", ")", ":", "In", "the", "PPI", "network", "for", "Homo", "Sapiens", ",", "nodes", "represent", "proteins", ",", "and", "an", "edge", "indicates", "a", "biological", "interaction", "between", "a", "pair", "of", "proteins", ".", "The", "network", "has", "19", ",", "706", "nodes", "and", "390", ",", "633", "edges", ".", "arXiv", "ASTRO", "-", "PH", ":", "This", "is", "a", "collaboration", "network", "generated", "from", "papers", "submitted", "to", "the", "e", "-", "print", "arXiv", "where", "nodes", "represent", "scientists", ",", "and", "an", "edge", "is", "present", "between", "two", "scientists", "if", "they", "have", "collaborated", "in", "a", "paper", ".", "The", "network", "has", "18", ",", "722", "nodes", "and", "198", ",", "110", "edges", ".", "Experimental", "results", ".", "We", "summarize", "our", "results", "for", "link", "prediction", "in", "Table", "[", "reference", "]", ".", "The", "best", "and", "parameter", "settings", "for", "each", "node2vec", "entry", "are", "omitted", "for", "ease", "of", "presentation", ".", "A", "general", "observation", "we", "can", "draw", "from", "the", "results", "is", "that", "the", "learned", "feature", "representations", "for", "node", "pairs", "significantly", "outperform", "the", "heuristic", "benchmark", "scores", "with", "node2vec", "achieving", "the", "best", "AUC", "improvement", "on", "12.6", "%", "on", "the", "arXiv", "dataset", "over", "the", "best", "performing", "baseline", "(", "Adamic", "-", "Adar", ")", ".", "Amongst", "the", "feature", "learning", "algorithms", ",", "node2vec", "outperforms", "both", "DeepWalk", "and", "LINE", "in", "all", "networks", "with", "gain", "up", "to", "3.8", "%", "and", "6.5", "%", "respectively", "in", "the", "AUC", "scores", "for", "the", "best", "possible", "choices", "of", "the", "binary", "operator", "for", "each", "algorithm", ".", "When", "we", "look", "at", "operators", "individually", "(", "Table", "[", "reference", "]", ")", ",", "node2vec", "outperforms", "DeepWalk", "and", "LINE", "barring", "a", "couple", "of", "cases", "involving", "the", "Weighted", "-", "L1", "and", "Weighted", "-", "L2", "operators", "in", "which", "LINE", "performs", "better", ".", "Overall", ",", "the", "Hadamard", "operator", "when", "used", "with", "node2vec", "is", "highly", "stable", "and", "gives", "the", "best", "performance", "on", "average", "across", "all", "networks", ".", "section", ":", "Discussion", "and", "Conclusion", "In", "this", "paper", ",", "we", "studied", "feature", "learning", "in", "networks", "as", "a", "search", "-", "based", "optimization", "problem", ".", "This", "perspective", "gives", "us", "multiple", "advantages", ".", "It", "can", "explain", "classic", "search", "strategies", "on", "the", "basis", "of", "the", "exploration", "-", "exploitation", "trade", "-", "off", ".", "Additionally", ",", "it", "provides", "a", "degree", "of", "interpretability", "to", "the", "learned", "representations", "when", "applied", "for", "a", "prediction", "task", ".", "For", "instance", ",", "we", "observed", "that", "BFS", "can", "explore", "only", "limited", "neighborhoods", ".", "This", "makes", "BFS", "suitable", "for", "characterizing", "structural", "equivalences", "in", "network", "that", "rely", "on", "the", "immediate", "local", "structure", "of", "nodes", ".", "On", "the", "other", "hand", ",", "DFS", "can", "freely", "explore", "network", "neighborhoods", "which", "is", "important", "in", "discovering", "homophilous", "communities", "at", "the", "cost", "of", "high", "variance", ".", "Both", "DeepWalk", "and", "LINE", "can", "be", "seen", "as", "rigid", "search", "strategies", "over", "networks", ".", "DeepWalk", "proposes", "search", "using", "uniform", "random", "walks", ".", "The", "obvious", "limitation", "with", "such", "a", "strategy", "is", "that", "it", "gives", "us", "no", "control", "over", "the", "explored", "neighborhoods", ".", "LINE", "proposes", "primarily", "a", "breadth", "-", "first", "strategy", ",", "sampling", "nodes", "and", "optimizing", "the", "likelihood", "independently", "over", "only", "1", "-", "hop", "and", "2", "-", "hop", "neighbors", ".", "The", "effect", "of", "such", "an", "exploration", "is", "easier", "to", "characterize", ",", "but", "it", "is", "restrictive", "and", "provides", "no", "flexibility", "in", "exploring", "nodes", "at", "further", "depths", ".", "In", "contrast", ",", "the", "search", "strategy", "in", "node2vec", "is", "both", "flexible", "and", "controllable", "exploring", "network", "neighborhoods", "through", "parameters", "and", ".", "While", "these", "search", "parameters", "have", "intuitive", "interpretations", ",", "we", "obtain", "best", "results", "on", "complex", "networks", "when", "we", "can", "learn", "them", "directly", "from", "data", ".", "From", "a", "practical", "standpoint", ",", "node2vec", "is", "scalable", "and", "robust", "to", "perturbations", ".", "We", "showed", "how", "extensions", "of", "node", "embeddings", "to", "link", "prediction", "outperform", "popular", "heuristic", "scores", "designed", "specifically", "for", "this", "task", ".", "Our", "method", "permits", "additional", "binary", "operators", "beyond", "those", "listed", "in", "Table", "[", "reference", "]", ".", "As", "a", "future", "work", ",", "we", "would", "like", "to", "explore", "the", "reasons", "behind", "the", "success", "of", "Hadamard", "operator", "over", "others", ",", "as", "well", "as", "establish", "interpretable", "equivalence", "notions", "for", "edges", "based", "on", "the", "search", "parameters", ".", "Future", "extensions", "of", "node2vec", "could", "involve", "networks", "with", "special", "structure", "such", "as", "heterogeneous", "information", "networks", ",", "networks", "with", "explicit", "domain", "features", "for", "nodes", "and", "edges", "and", "signed", "-", "edge", "networks", ".", "Continuous", "feature", "representations", "are", "the", "backbone", "of", "many", "deep", "learning", "algorithms", ",", "and", "it", "would", "be", "interesting", "to", "use", "node2vec", "representations", "as", "building", "blocks", "for", "end", "-", "to", "-", "end", "deep", "learning", "on", "graphs", ".", "Acknowledgements", ".", "We", "are", "thankful", "to", "Austin", "Benson", ",", "Will", "Hamilton", ",", "Rok", "Sosič", ",", "Marinka", "Žitnik", "as", "well", "as", "the", "anonymous", "reviewers", "for", "their", "helpful", "comments", ".", "This", "research", "has", "been", "supported", "in", "part", "by", "NSF", "CNS", "-", "1010921", ",", "IIS", "-", "1149837", ",", "NIH", "BD2", "K", ",", "ARO", "MURI", ",", "DARPA", "XDATA", ",", "DARPA", "SIMPLEX", ",", "Stanford", "Data", "Science", "Initiative", ",", "Boeing", ",", "Lightspeed", ",", "SAP", ",", "and", "Volkswagen", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["node2vec"], "Method"], [4, 7, ["Scalable", "Feature", "Learning"], "Method"], [8, 9, ["Networks"], "Task"], [9, 11, ["Prediction", "tasks"], "Task"], [25, 27, ["learning", "algorithms"], "Method"], [35, 37, ["representation", "learning"], "Task"], [43, 45, ["automating", "prediction"], "Task"], [54, 57, ["feature", "learning", "approaches"], "Method"], [75, 76, ["node2vec"], "Method"], [78, 80, ["algorithmic", "framework"], "Method"], [82, 85, ["continuous", "feature", "representations"], "Task"], [86, 89, ["nodes", "in", "networks"], "Task"], [91, 92, ["node2vec"], "Method"], [132, 136, ["biased", "random", "walk", "procedure"], "Method"], [173, 175, ["richer", "representations"], "Task"], [181, 182, ["node2vec"], "Method"], [193, 197, ["multi", "-", "label", "classification"], "Task"], [198, 200, ["link", "prediction"], "Task"], [221, 233, ["learning", "state", "-", "of", "-", "the", "-", "art", "task", "-", "independent", "representations"], "Task"], [234, 236, ["complex", "networks"], "Task"], [301, 302, ["Learning"], "Task"], [311, 313, ["Information", "networks"], "Task"], [314, 316, ["Feature", "learning"], "Task"], [317, 319, ["Node", "embeddings"], "Task"], [320, 322, ["Graph", "representations"], "Method"], [330, 332, ["network", "analysis"], "Task"], [342, 345, ["node", "classification", "task"], "Task"], [366, 368, ["social", "network"], "Task"], [374, 378, ["predicting", "interests", "of", "users"], "Task"], [382, 387, ["protein", "-", "protein", "interaction", "network"], "Task"], [401, 403, ["link", "prediction"], "Task"], [423, 425, ["Link", "prediction"], "Task"], [438, 439, ["genomics"], "Task"], [451, 453, ["social", "networks"], "Task"], [463, 467, ["supervised", "machine", "learning", "algorithm"], "Method"], [480, 482, ["prediction", "problems"], "Task"], [483, 484, ["networks"], "Task"], [492, 495, ["feature", "vector", "representation"], "Method"], [526, 528, ["feature", "engineering"], "Task"], [543, 545, ["prediction", "tasks"], "Task"], [552, 554, ["feature", "representations"], "Method"], [557, 559, ["optimization", "problem"], "Task"], [563, 565, ["feature", "learning"], "Task"], [578, 581, ["balancing", "computational", "efficiency"], "Metric"], [599, 601, ["feature", "representation"], "Method"], [606, 609, ["downstream", "prediction", "task"], "Task"], [612, 614, ["supervised", "procedure"], "Method"], [617, 618, ["accuracy"], "Metric"], [626, 629, ["training", "time", "complexity"], "Metric"], [650, 652, ["objective", "function"], "Metric"], [660, 663, ["downstream", "prediction", "task"], "Task"], [672, 674, ["unsupervised", "way"], "Method"], [678, 679, ["optimization"], "Task"], [718, 719, ["objective"], "Metric"], [721, 727, ["scalable", "unsupervised", "feature", "learning", "in", "networks"], "Task"], [732, 740, ["linear", "and", "non", "-", "linear", "dimensionality", "reduction", "techniques"], "Method"], [742, 745, ["Principal", "Component", "Analysis"], "Method"], [746, 750, ["Multi", "-", "Dimensional", "Scaling"], "Method"], [773, 775, ["data", "representation"], "Method"], [782, 783, ["eigendecomposition"], "Method"], [786, 788, ["data", "matrix"], "Method"], [792, 797, ["large", "real", "-", "world", "networks"], "Task"], [802, 804, ["latent", "representations"], "Method"], [809, 811, ["prediction", "tasks"], "Task"], [812, 813, ["networks"], "Task"], [837, 840, ["stochastic", "gradient", "descent"], "Method"], [841, 843, ["SGD", ")"], "Method"], [845, 846, ["backpropogation"], "Method"], [849, 855, ["hidden", "-", "layer", "feedforward", "neural", "networks"], "Method"], [977, 981, ["Real", "-", "world", "networks"], "Task"], [1003, 1005, ["node", "representations"], "Method"], [1043, 1046, ["feature", "learning", "algorithms"], "Method"], [1055, 1057, ["prediction", "tasks"], "Task"], [1063, 1064, ["node2vec"], "Method"], [1066, 1070, ["semi", "-", "supervised", "algorithm"], "Method"], [1071, 1076, ["scalable", "feature", "learning", "in", "networks"], "Task"], [1081, 1086, ["graph", "-", "based", "objective", "function"], "Method"], [1087, 1088, ["SGD"], "Method"], [1093, 1096, ["natural", "language", "processing"], "Task"], [1102, 1104, ["feature", "representations"], "Method"], [1124, 1129, ["2", "order", "random", "walk", "approach"], "Method"], [1164, 1165, ["node2vec"], "Method"], [1192, 1195, ["biased", "random", "walks"], "Method"], [1226, 1229, ["rigid", "search", "procedures"], "Method"], [1256, 1258, ["search", "strategy"], "Method"], [1268, 1271, ["network", "exploration", "strategies"], "Method"], [1288, 1292, ["semi", "-", "supervised", "fashion"], "Method"], [1297, 1299, ["feature", "representations"], "Method"], [1328, 1330, ["feature", "representations"], "Method"], [1336, 1338, ["binary", "operators"], "Method"], [1342, 1343, ["node2vec"], "Method"], [1344, 1346, ["prediction", "tasks"], "Task"], [1359, 1361, ["prediction", "tasks"], "Task"], [1362, 1363, ["networks"], "Task"], [1365, 1370, ["multi", "-", "label", "classification", "task"], "Task"], [1383, 1386, ["link", "prediction", "task"], "Task"], [1406, 1407, ["node2vec"], "Method"], [1415, 1418, ["feature", "learning", "algorithms"], "Method"], [1450, 1451, ["node2vec"], "Method"], [1466, 1470, ["multi", "-", "label", "classification"], "Task"], [1476, 1478, ["link", "prediction"], "Task"], [1511, 1512, ["node2vec"], "Method"], [1542, 1543, ["node2vec"], "Method"], [1546, 1548, ["scalable", "algorithm"], "Method"], [1549, 1553, ["feature", "learning", "in", "networks"], "Task"], [1558, 1565, ["network", "-", "aware", ",", "neighborhood", "preserving", "objective"], "Task"], [1566, 1567, ["SGD"], "Method"], [1571, 1572, ["node2vec"], "Method"], [1579, 1581, ["network", "science"], "Task"], [1594, 1595, ["node2vec"], "Method"], [1597, 1600, ["feature", "learning", "methods"], "Method"], [1613, 1618, ["edge", "-", "based", "prediction", "tasks"], "Task"], [1622, 1623, ["node2vec"], "Method"], [1660, 1662, ["feature", "learning"], "Task"], [1663, 1664, ["networks"], "Task"], [1671, 1673, ["feature", "learning"], "Task"], [1674, 1675, ["node2vec"], "Method"], [1690, 1691, ["node2vec"], "Method"], [1692, 1694, ["prediction", "tasks"], "Task"], [1707, 1709, ["parameter", "sensitivity"], "Metric"], [1710, 1712, ["perturbation", "analysis"], "Method"], [1749, 1750, ["node2vec"], "Method"], [1762, 1764, ["Feature", "engineering"], "Method"], [1770, 1773, ["machine", "learning", "community"], "Task"], [1778, 1779, ["networks"], "Task"], [1784, 1786, ["generating", "features"], "Task"], [1791, 1794, ["feature", "extraction", "techniques"], "Method"], [1821, 1823, ["feature", "extraction"], "Task"], [1825, 1828, ["representation", "learning", "problem"], "Task"], [1841, 1845, ["Unsupervised", "feature", "learning", "approaches"], "Method"], [1852, 1856, ["matrix", "representations", "of", "graphs"], "Method"], [1867, 1870, ["linear", "algebra", "perspective"], "Method"], [1877, 1880, ["dimensionality", "reduction", "techniques"], "Method"], [1882, 1883, ["linear"], "Method"], [1886, 1887, ["PCA"], "Method"], [1889, 1892, ["non", "-", "linear"], "Method"], [1895, 1900, ["IsoMap", ")", "dimensionality", "reduction", "techniques"], "Method"], [1909, 1913, ["computational", "and", "statistical", "performance"], "Metric"], [1918, 1920, ["computational", "efficiency"], "Metric"], [1921, 1926, ["eigendecomposition", "of", "a", "data", "matrix"], "Method"], [1930, 1932, ["solution", "quality"], "Metric"], [1990, 1992, ["prediction", "task"], "Task"], [1996, 1998, ["spectral", "clustering"], "Method"], [2010, 2011, ["classification"], "Task"], [2032, 2034, ["representational", "learning"], "Method"], [2035, 2038, ["natural", "language", "processing"], "Task"], [2042, 2047, ["feature", "learning", "of", "discrete", "objects"], "Task"], [2055, 2059, ["Skip", "-", "gram", "model"], "Method"], [2120, 2123, ["word", "feature", "representations"], "Method"], [2131, 2132, ["SGD"], "Method"], [2133, 2135, ["negative", "sampling"], "Method"], [2137, 2141, ["Skip", "-", "gram", "objective"], "Method"], [2145, 2147, ["distributional", "hypothesis"], "Method"], [2176, 2180, ["Skip", "-", "gram", "model"], "Method"], [2187, 2188, ["networks"], "Task"], [2238, 2240, ["sampling", "strategies"], "Method"], [2247, 2249, ["feature", "representations"], "Method"], [2262, 2265, ["winning", "sampling", "strategy"], "Method"], [2272, 2274, ["prediction", "tasks"], "Task"], [2290, 2293, ["sampling", "of", "nodes"], "Task"], [2299, 2300, ["node2vec"], "Method"], [2315, 2317, ["sampling", "strategy"], "Method"], [2338, 2344, ["node", "and", "edge", "based", "prediction", "tasks"], "Task"], [2353, 2356, ["supervised", "feature", "learning"], "Task"], [2361, 2367, ["graph", "-", "specific", "deep", "network", "architectures"], "Method"], [2377, 2380, ["downstream", "prediction", "task"], "Task"], [2384, 2388, ["non", "-", "linear", "transformations"], "Method"], [2392, 2393, ["accuracy"], "Metric"], [2399, 2400, ["scalability"], "Metric"], [2403, 2406, ["training", "time", "requirements"], "Metric"], [2409, 2412, ["Feature", "learning", "framework"], "Method"], [2414, 2418, ["feature", "learning", "in", "networks"], "Task"], [2420, 2424, ["maximum", "likelihood", "optimization", "problem"], "Task"], [2439, 2449, ["(", "un", ")", "directed", ",", "(", "un", ")", "weighted", "network"], "Task"], [2458, 2460, ["feature", "representaions"], "Method"], [2466, 2469, ["downstream", "prediction", "task"], "Task"], [2481, 2483, ["feature", "representation"], "Method"], [2509, 2512, ["neighborhood", "sampling", "strategy"], "Method"], [2518, 2522, ["Skip", "-", "gram", "architecture"], "Method"], [2531, 2533, ["objective", "function"], "Metric"], [2551, 2553, ["feature", "representation"], "Method"], [2562, 2564, ["optimization", "problem"], "Task"], [2599, 2601, ["feature", "representation"], "Method"], [2683, 2685, ["large", "networks"], "Task"], [2690, 2692, ["negative", "sampling"], "Method"], [2701, 2704, ["stochastic", "gradient", "ascent"], "Method"], [2712, 2715, ["Feature", "learning", "methods"], "Method"], [2718, 2722, ["Skip", "-", "gram", "architecture"], "Method"], [2757, 2758, ["Networks"], "Method"], [2784, 2786, ["randomized", "procedure"], "Method"], [2815, 2817, ["sampling", "strategy"], "Method"], [2820, 2823, ["Classic", "search", "strategies"], "Method"], [2828, 2834, ["sampling", "neighborhoods", "of", "a", "source", "node"], "Task"], [2838, 2840, ["local", "search"], "Method"], [2873, 2875, ["sampling", "strategies"], "Method"], [2903, 2905, ["sampling", "strategies"], "Method"], [2915, 2919, ["Breadth", "-", "first", "Sampling"], "Method"], [2920, 2921, ["BFS"], "Method"], [2956, 2960, ["Depth", "-", "first", "Sampling"], "Method"], [2961, 2962, ["DFS"], "Method"], [2990, 2998, ["breadth", "-", "first", "and", "depth", "-", "first", "sampling"], "Method"], [3021, 3023, ["prediction", "tasks"], "Task"], [3024, 3027, ["nodes", "in", "networks"], "Task"], [3179, 3183, ["BFS", "and", "DFS", "strategies"], "Method"], [3205, 3206, ["BFS"], "Method"], [3273, 3274, ["BFS"], "Method"], [3349, 3350, ["DFS"], "Method"], [3377, 3378, ["DFS"], "Method"], [3396, 3398, ["inferring", "communities"], "Task"], [3407, 3408, ["DFS"], "Method"], [3492, 3493, ["node2vec"], "Method"], [3503, 3506, ["neighborhood", "sampling", "strategy"], "Method"], [3513, 3514, ["BFS"], "Method"], [3515, 3516, ["DFS"], "Method"], [3524, 3528, ["biased", "random", "walk", "procedure"], "Method"], [3538, 3540, ["DFS", "fashion"], "Method"], [3543, 3545, ["Random", "Walks"], "Method"], [3555, 3560, ["random", "walk", "of", "fixed", "length"], "Method"], [3629, 3631, ["unweighted", "graphs"], "Task"], [3649, 3651, ["search", "procedure"], "Method"], [3662, 3665, ["BFS", "and", "DFS"], "Method"], [3668, 3670, ["sampling", "paradigms"], "Method"], [3672, 3674, ["structural", "equivalence"], "Task"], [3675, 3677, ["homophily", "respectively"], "Task"], [3679, 3681, ["random", "walks"], "Method"], [3712, 3716, ["2", "order", "random", "walk"], "Method"], [3727, 3729, ["random", "walk"], "Method"], [3832, 3834, ["search", "procedure"], "Method"], [3840, 3841, ["BFS"], "Method"], [3842, 3843, ["DFS"], "Method"], [3923, 3924, ["sampling"], "Task"], [4072, 4073, ["DFS"], "Method"], [4075, 4077, ["outward", "exploration"], "Task"], [4088, 4092, ["DFS", "-", "like", "exploration"], "Method"], [4094, 4097, ["random", "walk", "framework"], "Method"], [4123, 4124, ["preprocessing"], "Metric"], [4126, 4128, ["sampling", "efficiency"], "Metric"], [4129, 4131, ["random", "walks"], "Method"], [4149, 4151, ["random", "walks"], "Method"], [4152, 4155, ["2", "order", "Markovian"], "Method"], [4158, 4160, ["random", "walks"], "Method"], [4166, 4168, ["random", "walks"], "Method"], [4169, 4174, ["pure", "BFS", "/", "DFS", "approaches"], "Method"], [4175, 4177, ["Random", "walks"], "Method"], [4184, 4188, ["space", "and", "time", "requirements"], "Metric"], [4190, 4192, ["space", "complexity"], "Metric"], [4206, 4210, ["2", "order", "random", "walks"], "Task"], [4228, 4230, ["space", "complexity"], "Metric"], [4244, 4248, ["real", "-", "world", "networks"], "Task"], [4254, 4256, ["random", "walks"], "Method"], [4258, 4263, ["search", "-", "based", "sampling", "strategies"], "Method"], [4265, 4267, ["time", "complexity"], "Metric"], [4277, 4280, ["sample", "generation", "process"], "Method"], [4281, 4283, ["random", "walks"], "Method"], [4290, 4293, ["effective", "sampling", "rate"], "Metric"], [4329, 4331, ["effective", "complexity"], "Metric"], [4424, 4426, ["StochasticGradientDescent", "("], "Method"], [4459, 4460, ["node2vec"], "Method"], [4471, 4473, ["random", "walk"], "Method"], [4519, 4520, ["sampling"], "Task"], [4533, 4537, ["2", "order", "Markov", "chain"], "Method"], [4543, 4546, ["sampling", "of", "nodes"], "Task"], [4558, 4560, ["alias", "sampling"], "Method"], [4565, 4566, ["node2vec"], "Method"], [4569, 4570, ["preprocessing"], "Task"], [4575, 4578, ["random", "walk", "simulations"], "Method"], [4579, 4580, ["optimization"], "Task"], [4581, 4582, ["SGD"], "Method"], [4601, 4602, ["node2vec"], "Method"], [4603, 4604, ["node2vec"], "Method"], [4619, 4623, ["semi", "-", "supervised", "method"], "Method"], [4641, 4643, ["prediction", "tasks"], "Task"], [4656, 4658, ["link", "prediction"], "Task"], [4674, 4676, ["random", "walks"], "Method"], [4699, 4701, ["bootstrapping", "approach"], "Method"], [4703, 4705, ["feature", "representations"], "Method"], [4775, 4777, ["link", "prediction"], "Task"], [4828, 4830, ["downstream", "task"], "Task"], [4834, 4835, ["exploration"], "Task"], [4837, 4838, ["node2vec"], "Method"], [4841, 4843, ["feature", "representations"], "Method"], [4848, 4851, ["network", "analysis", "settings"], "Task"], [4870, 4874, ["BFS", "and", "DFS", "strategies"], "Method"], [4918, 4919, ["node2vec"], "Method"], [4962, 4963, ["node2vec"], "Method"], [4965, 4967, ["feature", "representation"], "Method"], [4975, 4977, ["feature", "representations"], "Method"], [5038, 5039, ["node2vec"], "Method"], [5101, 5102, ["node2vec"], "Method"], [5118, 5119, ["node2vec"], "Method"], [5149, 5150, ["node2vec"], "Method"], [5211, 5212, ["node2vec"], "Method"], [5237, 5241, ["real", "-", "world", "networks"], "Task"], [5254, 5256, ["prediction", "tasks"], "Task"], [5265, 5267, ["feature", "representations"], "Method"], [5269, 5270, ["node2vec"], "Method"], [5272, 5275, ["supervised", "learning", "tasks"], "Task"], [5276, 5280, ["multi", "-", "label", "classification"], "Task"], [5281, 5282, ["nodes"], "Task"], [5283, 5285, ["link", "prediction"], "Task"], [5297, 5298, ["node2vec"], "Method"], [5301, 5304, ["feature", "learning", "algorithms"], "Method"], [5305, 5307, ["Spectral", "clustering"], "Method"], [5311, 5314, ["matrix", "factorization", "approach"], "Method"], [5330, 5333, ["feature", "vector", "representations"], "Method"], [5336, 5337, ["DeepWalk"], "Method"], [5341, 5345, ["-", "dimensional", "feature", "representations"], "Method"], [5347, 5350, ["uniform", "random", "walks"], "Method"], [5352, 5354, ["sampling", "strategy"], "Method"], [5355, 5356, ["DeepWalk"], "Method"], [5364, 5365, ["node2vec"], "Method"], [5368, 5369, ["LINE"], "Method"], [5373, 5377, ["-", "dimensional", "feature", "representations"], "Method"], [5391, 5395, ["BFS", "-", "style", "simulations"], "Method"], [5429, 5432, ["matrix", "factorization", "approaches"], "Method"], [5441, 5442, ["DeepWalk"], "Method"], [5450, 5451, ["GraRep"], "Method"], [5454, 5455, ["LINE"], "Method"], [5487, 5493, ["sampling", "-", "based", "feature", "learning", "algorithms"], "Method"], [5515, 5517, ["prediction", "task"], "Task"], [5532, 5534, ["implementation", "language"], "Method"], [5535, 5540, ["C", "/", "C", "++/", "Python"], "Method"], [5553, 5555, ["sampling", "phase"], "Task"], [5561, 5562, ["LINE"], "Method"], [5563, 5564, ["node2vec"], "Method"], [5590, 5591, ["node2vec"], "Method"], [5596, 5598, ["optimization", "phase"], "Task"], [5604, 5605, ["SGD"], "Method"], [5616, 5617, ["DeepWalk"], "Method"], [5618, 5620, ["hierarchical", "sampling"], "Method"], [5634, 5635, ["node2vec"], "Method"], [5638, 5640, ["hierarchical", "softmax"], "Method"], [5645, 5647, ["negative", "sampling"], "Method"], [5659, 5661, ["negative", "sampling"], "Task"], [5662, 5663, ["DeepWalk"], "Method"], [5671, 5672, ["node2vec"], "Method"], [5673, 5674, ["LINE"], "Method"], [5678, 5679, ["node2vec"], "Method"], [5680, 5681, ["DeepWalk"], "Method"], [5704, 5705, ["optimization"], "Task"], [5715, 5716, ["LINE"], "Method"], [5719, 5720, ["LINE"], "Method"], [5741, 5742, ["node2vec"], "Method"], [5750, 5751, ["DeepWalk"], "Method"], [5752, 5753, ["LINE"], "Method"], [5764, 5765, ["optimization"], "Task"], [5808, 5814, ["10", "-", "fold", "cross", "-", "validation"], "Method"], [5827, 5831, ["Multi", "-", "label", "classification"], "Task"], [5833, 5838, ["multi", "-", "label", "classification", "setting"], "Task"], [5897, 5898, ["BlogCatalog"], "Material"], [5949, 5953, ["Protein", "-", "Protein", "Interactions"], "Task"], [5954, 5955, ["PPI"], "Task"], [5963, 5965, ["PPI", "network"], "Method"], [5966, 5968, ["Homo", "Sapiens"], "Material"], [6012, 6013, ["Wikipedia"], "Material"], [6017, 6019, ["cooccurrence", "network"], "Method"], [6048, 6052, ["Stanford", "POS", "-", "Tagger"], "Method"], [6138, 6143, ["protein", "-", "protein", "interaction", "network"], "Task"], [6177, 6178, ["homophily"], "Method"], [6188, 6191, ["word", "cooccurence", "network"], "Method"], [6209, 6211, ["Wikipedia", "corpus"], "Material"], [6267, 6270, ["node", "feature", "representations"], "Method"], [6274, 6282, ["one", "-", "vs", "-", "rest", "logistic", "regression", "classifier"], "Method"], [6283, 6285, ["L2", "regularization"], "Method"], [6330, 6333, ["Micro", "-", "F"], "Metric"], [6334, 6335, ["accuracy"], "Metric"], [6358, 6359, ["node2vec"], "Method"], [6367, 6368, ["BlogCatalog"], "Material"], [6394, 6395, ["DeepWalk"], "Method"], [6407, 6408, ["LINE"], "Method"], [6434, 6437, ["random", "walk", "methods"], "Method"], [6455, 6459, ["semi", "-", "supervised", "nature"], "Method"], [6460, 6461, ["node2vec"], "Method"], [6472, 6474, ["feature", "learning"], "Task"], [6479, 6481, ["PPI", "network"], "Task"], [6484, 6486, ["exploration", "strategy"], "Method"], [6530, 6531, ["LINE"], "Method"], [6552, 6554, ["exploration", "strategy"], "Method"], [6556, 6557, ["node2vec"], "Method"], [6575, 6577, ["search", "procedure"], "Method"], [6592, 6593, ["DeepWalk"], "Method"], [6597, 6598, ["LINE"], "Method"], [6602, 6606, ["fine", "-", "grained", "analysis"], "Task"], [6647, 6650, ["Micro", "-", "F"], "Metric"], [6651, 6654, ["Macro", "-", "F"], "Metric"], [6672, 6674, ["Spectral", "clustering"], "Method"], [6675, 6676, ["DeepWalk"], "Method"], [6677, 6678, ["LINE"], "Method"], [6679, 6680, ["node2vec"], "Method"], [6682, 6683, ["LINE"], "Method"], [6701, 6702, ["DeepWalk"], "Method"], [6706, 6707, ["BlogCatalog"], "Material"], [6719, 6721, ["search", "phase"], "Method"], [6730, 6731, ["node2vec"], "Method"], [6734, 6735, ["DeepWalk"], "Method"], [6747, 6748, ["LINE"], "Method"], [6761, 6762, ["BlogCatalog"], "Material"], [6776, 6777, ["PPI"], "Method"], [6796, 6798, ["Parameter", "sensitivity"], "Metric"], [6825, 6826, ["node2vec"], "Method"], [6883, 6884, ["node2vec"], "Method"], [6915, 6916, ["BlogCatalog"], "Material"], [7054, 7056, ["optimization", "time"], "Metric"], [7072, 7074, ["Perturbation", "Analysis"], "Method"], [7076, 7080, ["real", "-", "world", "networks"], "Task"], [7105, 7106, ["node2vec"], "Method"], [7108, 7111, ["imperfect", "information", "scenarios"], "Task"], [7204, 7205, ["Robustness"], "Metric"], [7232, 7234, ["network", "construction"], "Task"], [7246, 7248, ["perturbation", "setting"], "Task"], [7277, 7278, ["node2vec"], "Method"], [7311, 7312, ["robustness"], "Metric"], [7313, 7314, ["node2vec"], "Method"], [7324, 7326, ["sensor", "networks"], "Task"], [7347, 7349, ["node", "representations"], "Method"], [7350, 7351, ["node2vec"], "Method"], [7389, 7390, ["node2vec"], "Method"], [7411, 7413, ["sampling", "procedure"], "Method"], [7415, 7416, ["preprocessing"], "Task"], [7428, 7432, ["simulation", "of", "random", "walks"], "Task"], [7434, 7436, ["optimization", "phase"], "Task"], [7440, 7442, ["negative", "sampling"], "Method"], [7443, 7445, ["asynchronous", "SGD"], "Method"], [7458, 7460, ["sampling", "procedure"], "Method"], [7472, 7473, ["DeepWalk"], "Method"], [7492, 7494, ["Alias", "sampling"], "Method"], [7500, 7502, ["weighted", "networks"], "Task"], [7552, 7553, ["node2vec"], "Method"], [7554, 7557, ["semi", "-", "supervised"], "Method"], [7573, 7575, ["Link", "prediction"], "Task"], [7576, 7578, ["link", "prediction"], "Task"], [7637, 7639, ["edge", "removals"], "Method"], [7670, 7673, ["feature", "learning", "algorithms"], "Method"], [7678, 7680, ["link", "prediction"], "Task"], [7684, 7685, ["node2vec"], "Method"], [7688, 7690, ["heuristic", "scores"], "Method"], [7695, 7697, ["link", "prediction"], "Task"], [7733, 7734, ["Facebook"], "Material"], [7737, 7739, ["Facebook", "network"], "Material"], [7768, 7772, ["Protein", "-", "Protein", "Interactions"], "Task"], [7773, 7774, ["PPI"], "Task"], [7781, 7783, ["Homo", "Sapiens"], "Material"], [7815, 7818, ["ASTRO", "-", "PH"], "Task"], [7876, 7878, ["link", "prediction"], "Task"], [7891, 7892, ["node2vec"], "Method"], [7913, 7915, ["feature", "representations"], "Method"], [7921, 7924, ["heuristic", "benchmark", "scores"], "Metric"], [7925, 7926, ["node2vec"], "Method"], [7929, 7930, ["AUC"], "Metric"], [7936, 7938, ["arXiv", "dataset"], "Material"], [7944, 7947, ["Adamic", "-", "Adar"], "Method"], [7951, 7954, ["feature", "learning", "algorithms"], "Method"], [7955, 7956, ["node2vec"], "Method"], [7958, 7959, ["DeepWalk"], "Method"], [7960, 7961, ["LINE"], "Method"], [7976, 7978, ["AUC", "scores"], "Metric"], [8004, 8005, ["node2vec"], "Method"], [8006, 8007, ["DeepWalk"], "Method"], [8008, 8009, ["LINE"], "Method"], [8020, 8024, ["Weighted", "-", "L2", "operators"], "Method"], [8026, 8027, ["LINE"], "Method"], [8033, 8035, ["Hadamard", "operator"], "Method"], [8038, 8039, ["node2vec"], "Method"], [8064, 8068, ["feature", "learning", "in", "networks"], "Task"], [8070, 8075, ["search", "-", "based", "optimization", "problem"], "Task"], [8087, 8089, ["search", "strategies"], "Method"], [8094, 8100, ["exploration", "-", "exploitation", "trade", "-", "off"], "Metric"], [8117, 8119, ["prediction", "task"], "Task"], [8126, 8127, ["BFS"], "Method"], [8135, 8136, ["BFS"], "Method"], [8139, 8143, ["structural", "equivalences", "in", "network"], "Task"], [8158, 8159, ["DFS"], "Method"], [8176, 8177, ["variance"], "Metric"], [8179, 8180, ["DeepWalk"], "Method"], [8181, 8182, ["LINE"], "Method"], [8186, 8189, ["rigid", "search", "strategies"], "Method"], [8192, 8193, ["DeepWalk"], "Method"], [8194, 8195, ["search"], "Task"], [8196, 8199, ["uniform", "random", "walks"], "Method"], [8219, 8220, ["LINE"], "Method"], [8223, 8227, ["breadth", "-", "first", "strategy"], "Method"], [8276, 8278, ["search", "strategy"], "Method"], [8279, 8280, ["node2vec"], "Method"], [8305, 8307, ["complex", "networks"], "Task"], [8321, 8322, ["node2vec"], "Method"], [8334, 8336, ["node", "embeddings"], "Method"], [8337, 8339, ["link", "prediction"], "Task"], [8341, 8343, ["heuristic", "scores"], "Method"], [8403, 8404, ["node2vec"], "Method"], [8431, 8434, ["Continuous", "feature", "representations"], "Method"], [8439, 8442, ["deep", "learning", "algorithms"], "Method"], [8456, 8465, ["end", "-", "to", "-", "end", "deep", "learning", "on", "graphs"], "Task"], [583, 584, ["accuracy"], "Metric"], [704, 705, ["accuracy"], "Metric"], [1624, 1628, ["multi", "-", "label", "classification"], "Task"], [1727, 1728, ["node2vec"], "Method"], [4384, 4385, ["node2vec"], "Method"], [4453, 4454, ["node2vec"], "Method"], [4615, 4616, ["node2vec"], "Method"], [5559, 5560, ["DeepWalk"], "Method"], [5912, 5913, ["BlogCatalog"], "Material"], [6029, 6030, ["Wikipedia"], "Material"], [6302, 6305, ["Macro", "-", "F"], "Metric"], [6400, 6401, ["LINE"], "Method"], [6402, 6405, ["Macro", "-", "F"], "Metric"], [6496, 6497, ["DeepWalk"], "Method"], [6510, 6511, ["DeepWalk"], "Method"], [6532, 6535, ["Macro", "-", "F"], "Metric"], [6564, 6565, ["Wikipedia"], "Material"], [6688, 6689, ["DeepWalk"], "Method"], [6799, 6800, ["node2vec"], "Method"], [6828, 6829, ["BlogCatalog"], "Material"], [6869, 6872, ["Macro", "-", "F"], "Metric"], [7118, 7119, ["BlogCatalog"], "Material"], [7185, 7188, ["Macro", "-", "F"], "Metric"], [7298, 7301, ["Macro", "-", "F"], "Metric"], [8450, 8451, ["node2vec"], "Method"]]} {"doc_id": "1d73144204eb1604b26cd57ef75536f0f2753c69", "text": ["document", ":", "Subcategory", "-", "aware", "Convolutional", "Neural", "Networks", "for", "Object", "Proposals", "and", "Detection", "In", "Convolutional", "Neural", "Network", "(", "CNN", ")-", "based", "object", "detection", "methods", ",", "region", "proposal", "becomes", "a", "bottleneck", "when", "objects", "exhibit", "significant", "scale", "variation", ",", "occlusion", "or", "truncation", ".", "In", "addition", ",", "these", "methods", "mainly", "focus", "on", "2D", "object", "detection", "and", "can", "not", "estimate", "detailed", "properties", "of", "objects", ".", "In", "this", "paper", ",", "we", "propose", "subcategory", "-", "aware", "CNNs", "for", "object", "detection", ".", "We", "introduce", "a", "novel", "region", "proposal", "network", "that", "uses", "subcategory", "information", "to", "guide", "the", "proposal", "generating", "process", ",", "and", "a", "new", "detection", "network", "for", "joint", "detection", "and", "subcategory", "classification", ".", "By", "using", "subcategories", "related", "to", "object", "pose", ",", "we", "achieve", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "both", "detection", "and", "pose", "estimation", "on", "commonly", "used", "benchmarks", ".", "section", ":", "Introduction", "Convolutional", "Neural", "Networks", "(", "CNNs", ")", "have", "become", "dominating", "in", "solving", "different", "recognition", "problems", "recently", ".", "CNNs", "are", "powerful", "due", "to", "their", "capability", "in", "both", "representation", "and", "learning", ".", "With", "millions", "of", "weights", "in", "the", "contemporary", "CNNs", ",", "they", "are", "able", "to", "learn", "much", "richer", "representations", "from", "data", ".", "In", "object", "detection", ",", "we", "have", "witnessed", "the", "performance", "boost", "when", "CNNs", "are", "applied", "to", "commonly", "used", "benchmarks", "such", "as", "PASCAL", "VOC", "and", "ImageNet", ".", "However", ",", "there", "are", "two", "main", "limitations", "of", "the", "state", "-", "of", "-", "the", "-", "art", "CNN", "-", "based", "object", "detection", "methods", ".", "First", ",", "they", "rely", "on", "region", "proposal", "methods", "to", "generate", "object", "candidates", ",", "which", "are", "often", "based", "on", "low", "-", "level", "image", "features", "such", "as", "superpixels", "or", "edges", ".", "Although", "these", "methods", "work", "very", "well", "on", "PASCAL", "VOC", "and", "ImageNet", ",", "however", ",", "when", "it", "comes", "to", "the", "KITTI", "dataset", "for", "autonomous", "driving", "where", "objects", "have", "large", "scale", "variation", ",", "occlusion", "and", "truncation", ",", "these", "region", "proposal", "methods", "perform", "very", "poor", "as", "observed", "in", "our", "experiments", ".", "Recently", ",", "the", "Region", "Proposal", "Network", "(", "RPN", ")", "in", "is", "able", "to", "improve", "over", "the", "traditional", "region", "proposal", "methods", ".", "However", ",", "it", "still", "can", "not", "efficiently", "handle", "the", "scale", "change", "of", "object", ",", "occlusion", "and", "truncation", ".", "Second", ",", "the", "existing", "CNN", "-", "based", "object", "detection", "methods", "mainly", "focus", "on", "2D", "object", "detection", "with", "bounding", "boxes", ".", "As", "a", "result", ",", "they", "are", "not", "able", "to", "estimate", "detailed", "information", "about", "objects", "such", "as", "2D", "segmentation", "boundary", ",", "3D", "pose", "or", "occlusion", "relationship", "between", "objects", ",", "while", "these", "information", "is", "critical", "for", "various", "applications", "such", "as", "autonomous", "driving", ",", "robotics", "and", "augmented", "reality", ".", "In", "this", "work", ",", "we", "explore", "subcategory", "information", ",", "which", "is", "widely", "used", "in", "traditional", "object", "detection", ",", "to", "tackle", "the", "aforementioned", "two", "limitations", "in", "CNN", "-", "based", "object", "detection", ".", "For", "region", "proposal", "generation", ",", "we", "introduce", "a", "new", "CNN", "architecture", "that", "uses", "subcategory", "detections", "as", "object", "candidates", ".", "For", "detection", ",", "we", "modify", "the", "network", "in", "Fast", "R", "-", "CNN", "for", "joint", "detection", "and", "subcategory", "classification", ".", "Fig", ".", "[", "reference", "]", "illustrates", "our", "object", "detection", "framework", ".", "The", "concept", "of", "subcategory", "is", "general", "here", ".", "A", "subcategory", "can", "be", "objects", "with", "similar", "properties", "or", "attributes", "such", "as", "2D", "appearance", ",", "3D", "pose", "or", "3D", "shape", ".", "By", "associating", "object", "attributes", "to", "subcategories", ",", "we", "are", "able", "to", "estimate", "these", "attributes", "(", "e.g.", ",", "2D", "segmentation", "boundary", "or", "3D", "pose", ")", "by", "conducting", "subcategory", "classification", ".", "Specifically", ",", "motivated", "by", "the", "traditional", "detection", "methods", "that", "train", "a", "template", "or", "a", "detector", "for", "each", "subcategory", ",", "we", "introduce", "a", "subcategory", "convolutional", "(", "conv", ")", "layer", "in", "our", "Region", "Proposal", "Network", "(", "RPN", ")", ",", "where", "each", "filter", "in", "the", "conv", "layer", "is", "trained", "discriminatively", "for", "subcategory", "detection", ".", "The", "subcategory", "conv", "layer", "outputs", "heat", "maps", "about", "the", "presence", "of", "certain", "subcategories", "at", "a", "specific", "location", "and", "scale", ".", "Using", "these", "heat", "maps", ",", "our", "RPN", "is", "able", "to", "output", "confident", "subcategory", "detections", "as", "proposals", ".", "For", "classifying", "region", "proposals", "and", "refining", "their", "locations", ",", "we", "introduce", "a", "new", "object", "detection", "network", "by", "injecting", "subcategory", "information", "into", "the", "network", "proposed", "in", "Fast", "R", "-", "CNN", ".", "Our", "detection", "network", "is", "able", "to", "perform", "object", "detection", "and", "subcategory", "classification", "jointly", ".", "By", "using", "3D", "Voxel", "Patterns", "(", "3DVPs", ")", "as", "subcategories", ",", "our", "method", "is", "able", "to", "jointly", "detect", "the", "object", ",", "estimate", "its", "3D", "pose", ",", "segment", "its", "boundary", "and", "estimate", "its", "occluded", "or", "truncated", "regions", ".", "In", "addition", ",", "in", "both", "our", "RPN", "and", "our", "detection", "CNN", ",", "we", "use", "image", "pyramids", "as", "input", ",", "and", "we", "introduce", "a", "new", "feature", "extrapolating", "layer", "to", "efficiently", "compute", "conv", "features", "in", "multiple", "scales", ".", "In", "this", "way", ",", "our", "method", "is", "able", "to", "detect", "objects", "with", "large", "scale", "variations", ".", "We", "conduct", "experiments", "on", "the", "KITTI", "dataset", ",", "the", "PASCAL3D", "+", "dataset", "and", "the", "PASCAL", "VOC", "2007", "dataset", ".", "Comparisons", "with", "the", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "these", "benchmarks", "demonstrate", "the", "advantages", "of", "our", "subcategory", "-", "aware", "CNNs", "for", "object", "recognition", ".", "section", ":", "Related", "Work", "Subcategory", "in", "Object", "Detection", ".", "Subcategory", "has", "been", "widely", "utilized", "to", "facilitate", "object", "detection", ",", "and", "different", "methods", "of", "discovering", "object", "subcategories", "have", "been", "proposed", ".", "In", "DPM", ",", "subcategories", "are", "discovered", "by", "clustering", "objects", "according", "to", "the", "aspect", "ratio", "of", "their", "bounding", "boxes", ".", "performs", "clustering", "according", "to", "the", "viewpoint", "of", "the", "object", "to", "discover", "subcategories", ".", "Visual", "subcategories", "are", "constructed", "by", "clustering", "in", "the", "appearance", "space", "of", "object", ".", "3DVP", "performs", "clustering", "in", "the", "3D", "voxel", "space", "according", "to", "the", "visibility", "of", "the", "voxels", ".", "Unlike", "previous", "works", ",", "we", "utilize", "subcategory", "to", "improve", "CNN", "-", "based", "detection", ",", "and", "our", "framework", "is", "general", "to", "employ", "different", "types", "of", "object", "subcategories", ".", "CNN", "-", "based", "Object", "Detection", ".", "We", "can", "categorize", "the", "state", "-", "of", "-", "the", "-", "art", "CNN", "-", "based", "object", "detection", "methods", "into", "two", "classes", ":", "one", "-", "stage", "detection", "and", "two", "-", "stage", "detection", ".", "In", "one", "-", "stage", "detection", ",", "such", "as", "the", "Overfeat", "framework", ",", "a", "CNN", "directly", "processes", "an", "input", "image", ",", "and", "outputs", "object", "detections", ".", "In", "two", "-", "stage", "detection", ",", "such", "as", "R", "-", "CNNs", ",", "region", "proposals", "are", "first", "generated", "from", "an", "input", "image", ",", "where", "different", "region", "proposal", "methods", "can", "be", "employed", ".", "Then", "these", "region", "proposals", "are", "fed", "into", "a", "CNN", "for", "classification", "and", "location", "refinement", ".", "It", "is", "debatable", "which", "detection", "paradigm", "is", "better", ".", "We", "adopt", "the", "two", "-", "stage", "detection", "framework", "in", "this", "work", ",", "and", "consider", "the", "region", "proposal", "process", "to", "be", "the", "coarse", "detection", "step", "in", "coarse", "-", "to", "-", "fine", "detection", ".", "We", "propose", "a", "novel", "RPN", "motivated", "by", "and", "demonstrate", "its", "advantages", ".", "section", ":", "Subcategory", "-", "aware", "RPN", "Ideally", ",", "we", "want", "to", "have", "a", "region", "proposal", "approach", "that", "can", "cover", "objects", "in", "an", "input", "image", "with", "as", "few", "proposals", "as", "possible", ".", "Since", "objects", "in", "images", "appear", "at", "different", "locations", "and", "scales", ",", "region", "proposal", "itself", "is", "a", "challenging", "problem", ".", "Recently", ",", "proposed", "to", "tackle", "the", "region", "proposal", "problem", "with", "CNNs", ",", "demonstrating", "the", "advantages", "of", "using", "CNNs", "over", "traditional", "approaches", "for", "region", "proposal", ".", "In", "this", "section", ",", "we", "describe", "our", "subcategory", "-", "aware", "Region", "Proposal", "Network", "(", "RPN", ")", ".", "subsection", ":", "Network", "Architecture", "We", "introduce", "a", "novel", "network", "architecture", "for", "generating", "object", "proposals", "from", "images", ".", "The", "architecture", "is", "inspired", "by", "the", "traditional", "sliding", "-", "window", "-", "based", "object", "detectors", ",", "such", "as", "the", "Aggregated", "Channel", "Feature", "(", "ACF", ")", "detector", "and", "the", "Deformable", "Part", "Model", "(", "DPM", ")", ".", "Fig", ".", "[", "reference", "]", "illustrates", "the", "architecture", "of", "our", "region", "proposal", "network", ".", "i", ")", "To", "handle", "different", "scales", "of", "objects", ",", "we", "input", "into", "our", "RPN", "an", "image", "pyramid", ".", "This", "pyramid", "is", "processed", "by", "several", "convolutional", "(", "conv", ")", "and", "max", "pooling", "layers", "to", "extract", "the", "conv", "feature", "maps", ",", "with", "one", "conv", "feature", "map", "for", "each", "scale", ".", "ii", ")", "In", "order", "to", "speed", "up", "the", "computation", "of", "conv", "features", "on", "image", "pyramids", ",", "we", "introduce", "the", "feature", "extrapolating", "layer", ",", "which", "generates", "feature", "maps", "for", "scales", "that", "are", "not", "covered", "by", "the", "image", "pyramid", "via", "extrapolation", ".", "iii", ")", "After", "computing", "the", "extrapolated", "conv", "feature", "maps", ",", "we", "specifically", "design", "a", "conv", "layer", "for", "object", "subcategory", "detection", ",", "where", "each", "filter", "in", "the", "conv", "layer", "corresponds", "to", "an", "object", "subcategory", ".", "We", "train", "these", "filters", "to", "make", "sure", "they", "fire", "on", "correct", "locations", "and", "scales", "of", "objects", "in", "the", "corresponding", "subcategories", "during", "the", "network", "training", ".", "The", "subcategory", "conv", "layer", "outputs", "a", "heat", "map", "for", "each", "scale", ",", "where", "each", "value", "in", "the", "heat", "map", "indicates", "the", "confidence", "of", "an", "object", "in", "the", "corresponding", "location", ",", "scale", "and", "subcategory", ".", "v", ")", "Using", "the", "subcategory", "heat", "maps", ",", "we", "design", "a", "RoI", "generating", "layer", "that", "generates", "object", "candidates", "(", "RoIs", ")", "by", "thresholding", "the", "heat", "maps", ".", "vi", ")", "The", "RoIs", "are", "used", "in", "a", "RoI", "pooling", "layer", "to", "pool", "conv", "features", "from", "the", "extrapolated", "conv", "feature", "maps", ".", "vii", ")", "Finally", ",", "our", "RPN", "terminates", "at", "two", "sibling", "layers", ":", "one", "that", "outputs", "softmax", "probability", "estimates", "over", "object", "subcategories", ",", "and", "the", "other", "layer", "that", "refines", "the", "RoI", "location", "with", "a", "bounding", "box", "regressor", ".", "subsection", ":", "Feature", "Extrapolating", "Layer", "In", "our", "RPN", ",", "we", "use", "fixed", "-", "size", "conv", "filters", "in", "the", "subcategory", "conv", "layer", "to", "localize", "objects", "(", "e.g.", ",", "conv", "filters", ")", ".", "In", "order", "to", "handle", "different", "scales", "of", "objects", ",", "we", "resort", "to", "image", "pyramids", ".", "An", "image", "pyramid", "consists", "of", "images", "with", "different", "resolutions", "obtained", "by", "rescaling", "the", "original", "image", "according", "to", "different", "sampled", "scales", ".", "After", "constructing", "the", "image", "pyramid", "for", "an", "input", "image", ",", "multi", "-", "resolution", "conv", "feature", "maps", "can", "be", "computed", "by", "applying", "several", "conv", "layers", "and", "max", "pooling", "layers", "to", "each", "image", "in", "the", "pyramid", "(", "Fig", ".", "[", "reference", "]", ")", ".", "If", "we", "perform", "convolution", "on", "every", "scale", "explicitly", ",", "it", "is", "computationally", "expensive", ",", "especially", "when", "a", "finely", "-", "sampled", "image", "pyramid", "is", "needed", "as", "in", "the", "region", "proposal", "process", ".", "In", ",", "Dollár", "et", "al", ".", "demonstrate", "that", "multi", "-", "resolution", "image", "features", "can", "be", "approximated", "by", "extrapolation", "from", "nearby", "scales", "rather", "than", "being", "computed", "explicitly", ".", "Inspired", "by", "their", "work", ",", "we", "introduce", "a", "feature", "extrapolating", "layer", "to", "accelerate", "the", "computation", "of", "conv", "features", "on", "an", "image", "pyramid", ".", "Specifically", ",", "a", "feature", "extrapolating", "layer", "takes", "as", "input", "feature", "maps", "that", "are", "supplied", "by", "the", "last", "conv", "layer", "for", "feature", "extraction", ",", "where", "equals", "to", "the", "number", "of", "scales", "in", "the", "input", "image", "pyramid", ".", "Each", "feature", "map", "is", "a", "multi", "-", "dimensional", "array", "of", "size", ",", "with", "rows", ",", "columns", ",", "and", "channels", ".", "The", "width", "and", "height", "of", "the", "feature", "map", "corresponds", "to", "the", "largest", "scale", "in", "the", "image", "pyramid", ",", "where", "images", "in", "smaller", "scales", "are", "padded", "with", "zeros", "in", "order", "to", "generate", "feature", "maps", "with", "the", "same", "size", ".", "The", "feature", "extrapolating", "layer", "constructs", "feature", "maps", "at", "intermediate", "scales", "by", "extrapolating", "features", "from", "the", "nearest", "scales", "among", "the", "scales", "using", "bilinear", "interpolation", ".", "Suppose", "we", "add", "intermediate", "scales", "between", "every", "th", "scale", "and", "th", "scale", ",", ".", "The", "output", "of", "the", "feature", "extrapolating", "layer", "is", "feature", "maps", ",", "each", "with", "size", ".", "Since", "extrapolating", "a", "multi", "-", "dimensional", "array", "is", "much", "faster", "than", "computing", "a", "conv", "feature", "map", "explicitly", ",", "the", "feature", "extrapolating", "layer", "speeds", "up", "the", "feature", "computation", "with", "less", "memory", ".", "subsection", ":", "Subcategory", "Conv", "Layer", "After", "computing", "the", "conv", "feature", "maps", ",", "we", "design", "a", "subcategory", "conv", "layer", "for", "subcategory", "detection", ".", "Motivated", "by", "the", "traditional", "object", "detection", "methods", "that", "train", "a", "classifier", "or", "a", "template", "for", "each", "subcategory", ",", "we", "train", "a", "conv", "filter", "in", "the", "subcategory", "conv", "layer", "to", "detect", "a", "specific", "subcategory", ".", "Suppose", "there", "are", "subcategories", "to", "be", "considered", ".", "Then", ",", "the", "subcategory", "conv", "layer", "consists", "of", "conv", "filters", "with", "one", "additional", "conv", "filter", "for", "a", "special", "“", "background", "”", "category", ".", "For", "multi", "-", "class", "detection", "(", "e.g.", ",", "car", ",", "pedestrian", ",", "cyclist", ",", "etc", ".", ")", ",", "the", "subcategories", "are", "the", "aggregation", "of", "all", "the", "subcategories", "from", "all", "the", "classes", ".", "These", "conv", "filters", "operate", "on", "the", "extrapolated", "conv", "feature", "maps", "and", "output", "heat", "maps", "that", "indicate", "the", "confidences", "of", "the", "presence", "of", "objects", "in", "the", "input", "image", ".", "We", "use", "fixed", "-", "size", "conv", "filters", "in", "this", "layer", "(", "e.g.", ",", "conv", "filters", ")", ",", "which", "are", "trained", "to", "fire", "on", "specific", "scales", "in", "the", "feature", "pyramid", ".", "Sec", ".", "[", "reference", "]", "explains", "how", "we", "back", "-", "propagate", "errors", "from", "the", "loss", "layer", "to", "train", "these", "subcategory", "conv", "filters", ".", "subsection", ":", "RoI", "Generating", "Layer", "The", "RoI", "generating", "layer", "takes", "as", "input", "heat", "maps", "and", "outputs", "a", "set", "of", "region", "proposals", "(", "RoIs", ")", ",", "where", "is", "the", "number", "of", "scales", "in", "the", "feature", "pyramid", "after", "extrapolation", ".", "Each", "heat", "map", "is", "a", "multi", "-", "dimensional", "array", "of", "size", "for", "subcategories", "(", "i.e.", ",", "for", "RoI", "generating", ",", "we", "ignore", "the", "“", "background", "”", "channel", "in", "the", "heat", "map", ")", ".", "The", "RoI", "generating", "layer", "first", "converts", "each", "heat", "map", "into", "a", "2D", "array", "by", "performing", "max", "operation", "over", "the", "channels", "for", "subcategory", ".", "Then", ",", "it", "thresholds", "the", "2D", "heat", "map", "to", "generate", "RoIs", ".", "In", "this", "way", ",", "we", "measure", "the", "objectness", "of", "a", "region", "by", "aggregating", "information", "from", "subcategories", ".", "Different", "generating", "strategies", "are", "used", "in", "testing", "and", "training", ".", "In", "testing", ",", "each", "location", "in", "a", "heat", "map", "with", "a", "score", "larger", "than", "a", "predefined", "threshold", "is", "used", "to", "generate", "RoIs", ".", "First", ",", "a", "canonical", "bounding", "box", "is", "centered", "on", ".", "The", "width", "and", "height", "of", "the", "box", "are", "the", "same", "as", "those", "of", "the", "conv", "filters", "(", "e.g.", ",", ")", "in", "the", "subcategory", "conv", "layer", ",", "which", "have", "an", "aspect", "ratio", "one", ".", "Second", ",", "a", "number", "of", "boxes", "centered", "on", "with", "the", "same", "areas", "as", "the", "canonical", "box", "(", "e.g.", ",", ")", "but", "with", "different", "aspect", "ratios", "are", "generated", ".", "Finally", ",", "the", "RoI", "generating", "layer", "rescales", "the", "generated", "boxes", "according", "to", "the", "scale", "of", "the", "heat", "map", ",", "so", "as", "to", "cover", "objects", "in", "different", "scales", "and", "aspect", "ratios", ".", "In", "training", ",", "the", "RoI", "generating", "layer", "outputs", "hard", "positive", "RoIs", "and", "hard", "negative", "RoIs", "for", "training", "the", "subcategory", "conv", "filters", ",", "given", "a", "budget", "on", "batch", "size", "in", "stochastic", "gradient", "descent", ".", "First", ",", "we", "use", "the", "same", "procedure", "as", "described", "in", "testing", "to", "generate", "a", "number", "of", "bounding", "boxes", "for", "each", "location", "in", "each", "heat", "map", ".", "Second", ",", "according", "to", "the", "ground", "truth", "bounding", "boxes", "of", "objects", "in", "a", "training", "image", ",", "we", "compute", "the", "intersection", "over", "union", "(", "IoU", ")", "overlap", "between", "the", "generated", "boxes", "and", "the", "ground", "truth", "boxes", ".", "Bounding", "boxes", "with", "IoU", "overlap", "larger", "/", "smaller", "than", "some", "threshold", "(", "e.g.", ",", "0.5", ")", "are", "considered", "to", "be", "positive", "/", "negative", ".", "Finally", ",", "given", "the", "number", "of", "RoIs", "to", "be", "generated", "for", "each", "training", "image", "(", "i.e.", ",", "batch", "size", "divided", "by", "the", "number", "of", "images", "in", "a", "batch", ")", ",", "the", "RoI", "generating", "layer", "outputs", "hard", "positives", "(", "i.e.", ",", "positive", "bounding", "boxes", "with", "lowest", "scores", "in", "the", "heat", "maps", ")", "and", "hard", "negatives", "(", "i.e.", ",", "negative", "bounding", "boxes", "with", "highest", "scores", "in", "the", "heat", "maps", ")", ",", "where", "is", "the", "percentage", "of", "positive", "examples", ".", "subsection", ":", "Network", "Training", "After", "generating", "RoIs", ",", "we", "apply", "the", "RoI", "pooling", "layer", "proposed", "in", "to", "pool", "conv", "features", "for", "each", "RoI.", "Then", "the", "pooled", "conv", "features", "are", "used", "for", "two", "tasks", ":", "subcategory", "classification", "and", "bounding", "box", "regression", ".", "As", "illustrated", "in", "Fig", ".", "[", "reference", "]", ",", "our", "RPN", "has", "two", "sibling", "output", "layers", ".", "The", "first", "layer", "outputs", "a", "discrete", "probability", "distribution", ",", "over", "subcategories", ",", "which", "is", "computed", "by", "applying", "a", "softmax", "function", "over", "the", "outputs", "of", "the", "subcategory", "conv", "layer", ".", "The", "second", "layer", "outputs", "bounding", "box", "regression", "offsets", "for", "object", "classes", "(", ")", ".", "We", "parameterize", "as", "in", ",", "which", "specifies", "a", "scale", "-", "invariant", "translation", "and", "log", "-", "space", "width", "/", "height", "shift", "relative", "to", "a", "RoI.", "We", "employ", "a", "multi", "-", "task", "loss", "as", "in", "to", "train", "our", "RPN", "for", "subcategory", "classification", "and", "bounding", "box", "regression", ":", "where", "and", "are", "the", "truth", "subcategory", "label", "and", "the", "true", "class", "label", "respectively", ",", "is", "the", "standard", "cross", "-", "entropy", "loss", ",", "is", "the", "true", "bounding", "box", "regression", "targets", "for", "class", ",", "and", "is", "the", "prediction", "for", "class", ".", "We", "use", "the", "smoothed", "loss", "defined", "in", "for", "the", "bounding", "box", "regression", "loss", ".", "The", "indicator", "function", "indicates", "that", "bounding", "box", "regression", "is", "ignored", "if", "the", "RoI", "is", "background", "(", "i.e.", ",", ")", ".", "is", "a", "predefined", "weight", "to", "balance", "the", "two", "losses", ".", "In", "training", ",", "derivatives", "from", "the", "loss", "function", "are", "back", "-", "propagated", "(", "see", "red", "arrows", "in", "Fig", ".", "[", "reference", "]", ")", ".", "The", "two", "subcategory", "conv", "layers", "in", "our", "RPN", "share", "their", "weights", ".", "These", "weights", "/", "conv", "filters", "are", "updated", "according", "to", "the", "derivatives", "from", "the", "softmax", "loss", "function", "for", "subcategory", "classification", ",", "so", "we", "are", "able", "to", "train", "these", "filters", "for", "subcategory", "detection", ".", "There", "is", "no", "derivative", "flow", "in", "computing", "heat", "maps", "using", "the", "subcategory", "conv", "layer", "and", "in", "the", "RoI", "generating", "layer", ".", "Finally", ",", "our", "RPN", "generates", "confident", "subcategory", "detections", "as", "region", "proposals", ".", "section", ":", "Subcategory", "-", "aware", "Detection", "Network", "After", "the", "region", "proposal", "process", ",", "CNNs", "are", "utilized", "to", "classify", "these", "proposals", "and", "refine", "their", "locations", ".", "Since", "region", "proposal", "significantly", "reduces", "the", "search", "space", ",", "more", "powerful", "CNNs", "can", "be", "used", "in", "the", "detection", "step", ",", "which", "usually", "contain", "several", "fully", "connected", "layers", "with", "high", "dimensions", ".", "In", "this", "section", ",", "we", "introduce", "our", "subcategory", "-", "aware", "object", "detection", "network", "for", "joint", "detection", "and", "subcategory", "classification", ".", "subsection", ":", "Network", "Architecture", "Fig", ".", "[", "reference", "]", "illustrates", "the", "architecture", "of", "our", "detection", "network", ".", "The", "network", "is", "constructed", "based", "on", "the", "Fast", "R", "-", "CNN", "detection", "network", "with", "a", "number", "of", "improvements", ".", "i", ")", "We", "use", "image", "pyramids", "to", "handle", "the", "scale", "variation", "of", "objects", ".", "After", "the", "last", "conv", "layer", "for", "feature", "extraction", ",", "we", "add", "the", "feature", "extrapolating", "layer", "to", "increase", "the", "number", "of", "scales", "in", "the", "conv", "feature", "pyramid", ".", "ii", ")", "Given", "the", "region", "proposals", "generated", "from", "our", "RPN", ",", "we", "employ", "a", "RoI", "pooling", "layer", "to", "pool", "conv", "features", "for", "each", "RoI.", "Each", "RoI", "is", "mapped", "to", "a", "scale", "in", "the", "conv", "feature", "pyramid", "such", "that", "smaller", "RoIs", "pool", "features", "from", "larger", "scales", ".", "iii", ")", "The", "pooled", "conv", "features", "are", "fed", "into", "three", "fully", "connected", "(", "FC", ")", "layers", ",", "where", "the", "last", "FC", "layer", "is", "designed", "for", "subcategory", "classification", ".", "For", "subcategories", ",", "the", "“", "subcategory", "FC", "”", "layer", "outputs", "a", "dimensional", "vector", "with", "one", "additional", "dimension", "for", "the", "background", "class", ".", "We", "consider", "the", "output", ",", "named", "RoI", "feature", "vector", ",", "to", "be", "an", "embedding", "in", "the", "subcategory", "space", ".", "iv", ")", "Finally", ",", "the", "network", "terminates", "at", "three", "output", "layers", ".", "The", "first", "output", "layer", "applies", "a", "softmax", "function", "directly", "on", "the", "output", "of", "the", "“", "subcategory", "FC", "”", "layer", "for", "subcategory", "classification", ".", "The", "other", "two", "output", "layers", "operate", "on", "the", "RoI", "feature", "vector", "and", "apply", "FC", "layers", "for", "object", "class", "classification", "and", "bounding", "box", "regression", ".", "subsection", ":", "Network", "Training", "We", "train", "our", "object", "detection", "network", "with", "a", "multi", "-", "task", "loss", "for", "joint", "object", "class", "classification", ",", "subcategory", "classification", "and", "bounding", "box", "regression", ":", "where", "is", "a", "probability", "distribution", "over", "subcategories", ",", "is", "a", "probability", "distribution", "over", "object", "classes", ",", "and", "are", "the", "truth", "subcategory", "label", "and", "the", "true", "class", "label", "respectively", ",", "and", "are", "the", "predicted", "vector", "and", "the", "true", "vector", "for", "bounding", "box", "regression", "respectively", ",", "and", "and", "are", "predefined", "weights", "to", "balance", "the", "losses", "of", "different", "tasks", ".", "and", "are", "the", "standard", "cross", "-", "entropy", "loss", ",", "and", "is", "the", "smoothed", "loss", "as", "in", "our", "RPN", ".", "In", "back", "-", "propagation", "training", ",", "derivatives", "for", "the", "multi", "-", "task", "loss", "are", "back", "-", "propagated", "to", "the", "previous", "layers", ".", "Red", "arrows", "in", "Fig", ".", "[", "reference", "]", "indicate", "the", "route", "of", "the", "derivative", "flow", ".", "section", ":", "Experiments", "subsection", ":", "Experimental", "Settings", "Datasets", ".", "We", "evaluate", "our", "object", "detection", "framework", "on", "the", "KITTI", "detection", "benchmark", ",", "the", "PASCAL3D", "+", "dataset", "and", "the", "PASCAL", "VOC", "2007", "dataset", ".", "i", ")", "The", "KITTI", "dataset", "consists", "of", "video", "frames", "from", "autonomous", "driving", "scenes", ",", "with", "7", ",", "481", "images", "for", "training", "and", "7", ",", "518", "images", "for", "testing", ".", "Car", ",", "pedestrian", "and", "cyclist", "are", "evaluated", "for", "object", "detection", ".", "Since", "the", "ground", "truth", "annotations", "of", "the", "KITTI", "test", "set", "are", "not", "released", ",", "we", "split", "the", "KITTI", "training", "images", "into", "a", "train", "set", "and", "a", "validation", "set", "for", "analyses", "as", "in", ".", "ii", ")", "The", "PASCAL3D", "+", "dataset", "augments", "12", "rigid", "categories", "in", "the", "PASCAL", "VOC", "2012", "with", "3D", "annotations", ".", "Each", "object", "in", "the", "12", "categories", "is", "registered", "with", "a", "3D", "CAD", "model", ".", "The", "train", "set", "of", "PASCAL", "VOC", "2012", "is", "used", "for", "training", "(", "5", ",", "717", "images", ")", ",", "while", "the", "val", "set", "is", "used", "for", "testing", "(", "5", ",", "823", "images", ")", ".", "iii", ")", "The", "PASCAL", "VOC", "2007", "dataset", "contains", "5", ",", "011", "training", "images", "and", "4", ",", "952", "testing", "images", "on", "20", "categories", ".", "Evaluation", "Metrics", ".", "On", "KITTI", ",", "we", "evaluate", "our", "detection", "framework", "at", "three", "levels", "of", "difficulty", "as", "suggested", "by", ",", "i.e.", ",", "easy", ",", "moderate", "and", "hard", ",", "where", "the", "difficulty", "is", "measured", "by", "the", "minimal", "scale", "of", "object", "to", "be", "considered", "and", "the", "occlusion", "and", "truncation", "of", "the", "object", ".", "Average", "Precision", "(", "AP", ")", "is", "used", "to", "measure", "the", "detection", "performance", ",", "where", "70", "%", ",", "50", "%", ",", "and", "50", "%", "overlap", "thresholds", "are", "adopted", "by", "the", "KITTI", "benchmark", "for", "car", ",", "pedestrian", "and", "cyclist", "respectively", ".", "To", "evaluate", "joint", "detection", "and", "orientation", "estimation", "on", "KITTI", ",", "introduces", "Average", "Orientation", "Similarity", "(", "AOS", ")", ",", "which", "evaluates", "the", "orientation", "similarity", "between", "detections", "and", "ground", "truths", "at", "different", "detection", "recalls", ".", "introduces", "Average", "Segmentation", "Accuracy", "(", "ASA", ")", "for", "joint", "detection", "and", "segmentation", ",", "and", "Average", "Location", "Precision", "(", "ALP", ")", "for", "joint", "detection", "and", "3D", "location", "similar", "to", "AOS", ".", "We", "also", "use", "these", "metrics", "here", ".", "On", "PASCAL3D", "+", "and", "PASCAL", "VOC", "2007", ",", "the", "standard", "AP", "with", "50", "%", "overlap", "ratio", "is", "adopted", "to", "evaluate", "object", "detection", ".", "For", "joint", "detection", "and", "pose", "estimation", ",", "we", "use", "the", "Average", "Viewpoint", "Precision", "(", "AVP", ")", "suggested", "by", ",", "where", "a", "detection", "is", "considered", "to", "be", "a", "true", "positive", "if", "its", "location", "and", "viewpoint", "are", "both", "correct", ".", "Subcategories", ".", "We", "experiment", "with", "both", "2D", "subcategories", "and", "3D", "subcategories", ".", "For", "2D", "subcategories", ",", "we", "cluster", "objects", "using", "2D", "image", "features", "(", "i.e.", ",", "aggregated", "channel", "features", "from", ")", ".", "Only", "bounding", "box", "annotations", "are", "needed", "for", "2D", "subcategories", ".", "When", "additional", "annotations", "are", "available", ",", "we", "can", "obtain", "3D", "subcategories", ".", "We", "adopt", "the", "3D", "Voxel", "Pattern", "(", "3DVP", ")", "representation", "for", "rigid", "objects", "(", "i.e.", ",", "car", "in", "KITTI", "and", "the", "12", "categories", "in", "PASCAL3D", "+", ")", ",", "which", "jointly", "models", "object", "pose", ",", "occlusion", "and", "truncation", "in", "the", "clustering", "process", ".", "Each", "3DVP", "is", "considered", "to", "be", "a", "subcategory", ".", "For", "pedestrian", "and", "cyclist", "in", "KITTI", ",", "we", "perform", "clustering", "according", "to", "the", "object", "orientation", ",", "and", "each", "cluster", "is", "considered", "to", "be", "a", "subcategory", ".", "In", "this", "way", ",", "by", "subcategory", "classification", ",", "we", "can", "transfer", "the", "meta", "data", "carried", "by", "3DVPs", "(", "3D", "pose", ",", "segmentation", "boundary", "and", "occluded", "regions", ")", "to", "the", "detected", "object", ".", "For", "validation", "on", "KITTI", "(", "3", ",", "682", "images", "for", "training", ",", "3", ",", "799", "images", "for", "testing", ")", ",", "we", "use", "173", "subcategories", "(", "125", "3DVPs", "for", "car", ",", "24", "poses", "for", "pedestrian", "and", "cyclist", "each", ")", ",", "while", "for", "testing", "on", "KITTI", "(", "7", ",", "481", "images", "for", "training", ",", "7", ",", "518", "images", "for", "testing", ")", ",", "we", "use", "275", "subcategories", "(", "227", "3DVPs", "for", "car", ",", "24", "poses", "for", "pedestrian", "and", "cyclist", "each", ")", ".", "3DVPs", "are", "discovered", "with", "affinity", "propagation", "clustering", ",", "which", "automatically", "discovers", "the", "number", "of", "clusters", "from", "the", "data", ".", "For", "PASCAL3D", "+", ",", "337", "3DVPs", "are", "discovered", "among", "the", "12", "categories", ".", "For", "PASCAL", "VOC", "2007", ",", "we", "use", "240", "2D", "subcategories", ",", "with", "12", "for", "each", "class", ".", "Correspondingly", ",", "the", "output", "number", "of", "the", "subcategory", "conv", "layer", "in", "our", "RPN", "and", "that", "of", "the", "subcategory", "FC", "layer", "in", "our", "detection", "network", "equal", "to", "the", "number", "of", "subcategory", "plus", "one", ".", "Region", "Proposal", "Network", "Hyper", "-", "parameters", ".", "In", "our", "RPN", ",", "we", "use", "5", "scales", "for", "KITTI", "in", "the", "input", "image", "pyramid", "and", "4", "scales", "for", "PASCAL", "(", "both", "PASCAL3D", "+", "and", "PASCAL", "VOC", "2007", ")", ",", "where", "each", "number", "indicates", "the", "rescaling", "factor", "with", "respect", "to", "the", "original", "image", "size", ".", "Objects", "in", "PASCAL", "have", "smaller", "scale", "variation", "compared", "to", "objects", "in", "KITTI", ".", "Adding", "larger", "scales", "for", "PASCAL", "only", "results", "in", "marginal", "improvement", "but", "significantly", "increases", "the", "computation", ".", "The", "feature", "extrapolating", "layer", "extrapolates", "4", "scales", "with", "equal", "intervals", "between", "every", "two", "input", "scales", ",", "so", "the", "final", "conv", "feature", "pyramid", "has", "21", "scales", "for", "KITTI", "and", "16", "scales", "for", "PASCAL", ".", "In", "the", "RoI", "generating", "layer", ",", "each", "location", "in", "a", "heat", "map", "generates", "7", "boxes", "with", "7", "different", "aspect", "ratios", "for", "KITTI", "and", "5", "aspect", "ratios", "for", "PASCAL", ",", "where", "each", "number", "indicates", "the", "ratio", "between", "the", "height", "and", "the", "width", "of", "the", "bounding", "box", ".", "In", "training", "the", "RPN", ",", "each", "SGD", "mini", "-", "batch", "is", "constructed", "from", "a", "single", "image", ",", "chosen", "uniformly", "at", "random", ".", "A", "mini", "-", "batch", "has", "size", "128", ",", "with", "64", "positive", "RoIs", "and", "64", "negative", "RoIs", ",", "where", "the", "IoU", "threshold", "is", "70", "%", "for", "both", "KITTI", "and", "PASCAL", ".", "Detection", "Network", "Hyper", "-", "parameters", ".", "In", "our", "detection", "network", ",", "we", "use", "4", "scales", "in", "the", "input", "image", "pyramid", "for", "KITTI", "and", "2", "scales", "for", "PASCAL", ",", "both", "with", "4", "scales", "extrapolated", "between", "every", "two", "scales", ".", "Each", "SGD", "mini", "-", "batch", "is", "constructed", "from", "2", "images", ".", "A", "mini", "-", "batch", "has", "size", "128", ",", "with", "64", "RoIs", "from", "each", "image", ".", "25", "%", "of", "the", "RoIs", "are", "positive", ",", "where", "the", "IoU", "threshold", "is", "70", "%", "for", "car", "in", "KITTI", ",", "and", "50", "%", "for", "the", "other", "categories", ".", "The", "same", "SGD", "hyper", "-", "parameters", "are", "used", "as", "in", "for", "region", "proposal", "and", "detection", ".", "Fine", "-", "tuning", "Pre", "-", "trained", "Networks", ".", "Our", "framework", "is", "implemented", "in", "Caffe", ".", "We", "initialize", "the", "conv", "layers", "for", "feature", "extraction", "in", "both", "networks", "and", "the", "two", "FC", "layers", "before", "subcategory", "FC", "layer", "in", "the", "detection", "network", "with", "pre", "-", "trained", "networks", "on", "ImageNet", ".", "On", "KITTI", ",", "we", "experiment", "with", "the", "AlexNet", ",", "the", "VGG16", "network", "and", "the", "GoogleNet", ".", "On", "PASCAL", ",", "we", "fine", "-", "tune", "the", "VGG16", "network", ".", "subsection", ":", "Analysis", "on", "KITTI", "Validation", "Set", "Region", "Proposal", "Evalutaion", "on", "Recall", ".", "We", "evaluate", "the", "detection", "recall", "of", "our", "RPN", "and", "compare", "it", "with", "the", "state", "-", "of", "-", "the", "-", "art", "methods", "in", "Table", "[", "reference", "]", "on", "the", "KITTI", "validation", "set", ".", "For", "each", "image", ",", "we", "use", "2k", "proposals", "for", "all", "the", "methods", ".", "First", ",", "two", "popular", "methods", "that", "work", "well", "on", "PASCAL", "VOC", ",", "Selective", "Search", "and", "Edge", "Boxes", ",", "do", "not", "perform", "well", "on", "KITTI", ",", "mainly", "because", "objects", "in", "KITTI", "exhibit", "more", "significant", "scale", "variation", ",", "occlusion", "and", "truncation", ".", "It", "is", "challenging", "for", "a", "bottom", "-", "up", "proposal", "method", "to", "achieve", "high", "recall", "under", "a", "small", "budget", "(", "i.e", ",", "2k", "boxes", "per", "image", ")", ".", "Second", ",", "the", "RPN", "in", "Faster", "R", "-", "CNN", "performs", "much", "better", "than", "Selective", "Search", "and", "Edge", "Boxes", ",", "which", "demonstrates", "the", "ability", "of", "discriminatively", "trained", "CNNs", "for", "region", "proposal", ".", "But", "we", "have", "to", "increase", "its", "parameter", "setting", "from", "3", "scales", "and", "3", "aspect", "ratios", "in", "to", "10", "scales", "and", "7", "aspect", "ratios", "in", "order", "to", "make", "it", "work", "on", "KITTI", ".", "Finally", ",", "our", "RPN", "performs", "on", "par", "with", "Faster", "R", "-", "CNN", "on", "car", ",", "and", "outperforms", "it", "on", "pedestrian", "and", "cyclist", "using", "the", "same", "number", "of", "proposals", "per", "image", ".", "Our", "new", "architecture", "can", "better", "handle", "scale", "variation", "using", "image", "pyramid", ".", "It", "also", "benefits", "from", "data", "mining", "hard", "training", "examples", "in", "our", "RoI", "generating", "layer", ".", "Region", "Proposal", "Evalutaion", "on", "Detection", "and", "Oritentaion", "Estimation", ".", "Detection", "recall", "measures", "the", "coverage", "of", "region", "proposals", ",", "which", "can", "not", "demonstrate", "the", "quality", "of", "the", "region", "proposals", "for", "detection", ".", "In", "this", "experiment", ",", "we", "directly", "measure", "the", "detection", "and", "orientation", "estimation", "performance", "using", "different", "region", "proposals", ".", "Table", "[", "reference", "]", "presents", "the", "detection", "and", "orientation", "estimation", "results", "using", "RPN", "in", "Faster", "R", "-", "CNN", "and", "the", "RPN", "we", "propose", ",", "while", "keeping", "the", "detection", "network", "the", "same", "as", "described", "in", "Sec", ".", "[", "reference", "]", ".", "We", "compare", "our", "RPN", "with", "two", "variations", "of", "the", "RPN", "in", "Faster", "R", "-", "CNN", ".", "For", "the", "first", "model", ",", "the", "RPN", "and", "the", "detection", "network", "are", "trained", "independently", "to", "each", "other", "(", "“", "unshared", "”", ")", ".", "For", "the", "second", "model", ",", "the", "RPN", "and", "the", "detection", "network", "share", "their", "conv", "layers", "for", "feature", "extraction", "in", "order", "to", "save", "computation", "on", "convolution", "(", "“", "shared", "”", ")", ".", "The", "sharing", "is", "achieved", "by", "the", "four", "-", "step", "alternating", "optimization", "training", "algorithm", "described", "in", ".", "By", "comparing", "the", "two", "models", "in", "Table", "[", "reference", "]", ",", "we", "find", "that", "sharing", "conv", "layers", "hurts", "the", "performance", "on", "car", "and", "pedestrian", ",", "but", "improves", "the", "performance", "on", "cyclist", ".", "Car", "and", "pedestrian", "have", "much", "more", "training", "examples", "available", "than", "cyclist", ".", "With", "enough", "training", "data", ",", "the", "RPN", "and", "the", "detection", "network", "trained", "independently", "can", "develop", "conv", "features", "suitable", "for", "its", "own", "task", ".", "In", "this", "case", ",", "shared", "conv", "features", "degrade", "the", "performance", ".", "However", ",", "when", "the", "training", "data", "is", "insufficient", ",", "sharing", "conv", "features", "can", "help", ".", "In", "Table", "[", "reference", "]", ",", "by", "using", "region", "proposals", "from", "our", "RPN", ",", "we", "achieve", "better", "performance", "on", "detection", "and", "orientation", "estimation", "across", "all", "the", "three", "categories", ".", "The", "experimental", "results", "demonstrate", "the", "advantages", "of", "our", "RPN", ".", "We", "also", "tried", "to", "share", "the", "conv", "layers", "in", "our", "RPN", "and", "our", "detection", "network", ".", "However", ",", "since", "the", "architecture", "of", "our", "RPN", "after", "the", "conv", "layers", "for", "feature", "extraction", "is", "quite", "different", "from", "that", "of", "the", "detection", "network", ",", "we", "found", "that", "the", "training", "can", "not", "converge", ",", "which", "verifies", "our", "observation", "that", "the", "RPN", "and", "the", "detection", "network", "have", "developed", "their", "own", "conv", "features", "that", "are", "suitable", "for", "its", "own", "task", ".", "Detection", "Network", "Evalutaion", ".", "In", "Table", "[", "reference", "]", ",", "we", "first", "show", "that", "our", "RPN", "achieves", "significantly", "better", "performance", "than", "the", "RPN", "in", "when", "the", "two", "RPNs", "are", "used", "with", "Fast", "R", "-", "CNN", "on", "the", "KITTI", "validation", "set", "respectively", ".", "Then", ",", "we", "use", "region", "proposals", "from", "our", "RPN", "and", "compare", "different", "variations", "of", "the", "network", "architecture", "for", "detection", ".", "i", ")", "“", "Ours", "w", "/", "o", "Pose", "”", "indicates", "using", "2D", "subcategories", "from", "clustering", "on", "2D", "appearances", "of", "objects", "without", "using", "additional", "pose", "information", ".", "As", "we", "can", "see", ",", "our", "method", "still", "outperforms", "Fatser", "R", "-", "CNN", "in", "this", "case", ".", "ii", ")", "By", "using", "pose", "information", "to", "obtain", "subcategories", ",", "our", "detection", "network", "is", "also", "able", "to", "estimate", "the", "orientation", "of", "the", "object", ".", "“", "Ours", "w", "/", "o", "Extra", "”", "refers", "to", "a", "network", "without", "feature", "extrapolating", ".", "By", "augmenting", "the", "network", "with", "the", "feature", "extrapolating", "layer", ",", "our", "full", "model", "(", "“", "Ours", "Full", "”", "in", "Table", "[", "reference", "]", ")", "further", "boosts", "the", "performance", ",", "except", "for", "a", "minor", "drop", "on", "orientation", "estimation", "of", "pedestrian", ".", "Evaluation", "on", "2D", "Segmentation", "and", "3D", "Localization", ".", "3DVPs", "enable", "us", "to", "transfer", "the", "meta", "data", "to", "the", "detect", "objects", ",", "so", "our", "method", "is", "able", "to", "segment", "the", "boundary", "of", "object", ".", "In", "addition", ",", "after", "detecting", "the", "objects", "and", "estimating", "their", "3D", "poses", ",", "we", "can", "back", "-", "project", "them", "into", "3D", "using", "the", "camera", "parameters", "provided", "in", "KITTI", ",", "so", "as", "to", "evaluate", "the", "3D", "localization", "performance", ".", "In", "table", "[", "reference", "]", ",", "we", "compare", "our", "method", "on", "2D", "segmentation", "and", "3D", "localization", "of", "car", "with", "DPM", "and", "3DVP", "on", "the", "KITTI", "validation", "set", ".", "We", "have", "significantly", "improve", "the", "segmentation", "accuracy", "and", "3D", "location", "accuracy", "when", "the", "2", "-", "meter", "threshold", "is", "used", "(", "i.e.", ",", "a", "detection", "within", "2", "meters", "from", "the", "ground", "truth", "location", "is", "considered", "to", "be", "correct", ")", ".", "Surprisingly", ",", "obtains", "better", "3D", "localization", "accuracy", "with", "the", "1", "-", "meter", "threshold", ",", "which", "indicates", "that", "more", "detections", "from", "are", "within", "the", "1", "-", "meter", "distance", "from", "the", "ground", "truth", ".", "subsection", ":", "KITTI", "Test", "Set", "Evaluation", "To", "compare", "with", "the", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "the", "KITTI", "detection", "benchmark", ",", "we", "train", "our", "RPN", "and", "detection", "network", "with", "all", "the", "KITTI", "training", "data", ",", "and", "then", "test", "our", "method", "on", "the", "KITTI", "test", "set", "by", "submitting", "our", "results", "to", ".", "Table", "[", "reference", "]", "presents", "the", "detection", "and", "orientation", "estimation", "results", "on", "the", "three", "categories", ",", "where", "we", "compare", "our", "method", "(", "SubCNN", ")", "with", "different", "methods", "evaluated", "on", "KITTI", ".", "We", "have", "experimented", "fine", "-", "tuning", "both", "the", "VGG16", "network", "and", "the", "GoogleNet", "for", "the", "detection", "network", ".", "Our", "method", "ranks", "on", "top", "among", "all", "the", "published", "methods", ".", "The", "experimental", "results", "demonstrate", "the", "ability", "of", "our", "CNNs", "in", "using", "subcategory", "information", "for", "detection", "and", "orientation", "estimation", ".", "Fig", ".", "[", "reference", "]", "presents", "some", "examples", "of", "our", "detection", "and", "3D", "localization", "results", "on", "KITTI", ".", "subsection", ":", "Evaluation", "on", "PASCAL3D", "+", "and", "PASCAL", "VOC", "We", "also", "evaluate", "our", "detection", "framework", "on", "the", "12", "categories", "in", "PASCAL3D", "+", ".", "Table", "[", "reference", "]", "presents", "the", "detection", "results", "in", "AP", "and", "the", "joint", "detection", "and", "pose", "estimation", "results", "in", "AVP", ".", "After", "generating", "region", "proposals", "from", "our", "RPN", ",", "we", "experiment", "with", "our", "detection", "networks", "with", "and", "without", "feature", "extrapolation", ".", "First", ",", "in", "terms", "of", "detection", ",", "our", "method", "improves", "over", "R", "-", "CNN", "on", "all", "12", "categories", ".", "Second", ",", "in", "terms", "of", "join", "detection", "and", "pose", "estimation", ",", "our", "method", "significantly", "outperforms", "two", "state", "-", "of", "-", "the", "-", "art", "methods", ":", "VDPM", "and", "DPM", "-", "VOC", "+", "VP", ".", "Third", ",", "feature", "extrapolation", "helps", "both", "detection", "and", "pose", "estimation", "on", "PASCAL3D", "+", ".", "It", "is", "worth", "mentioning", "that", "PASCAL3D", "+", "has", "much", "fewer", "training", "examples", "in", "each", "subcategory", "compared", "to", "KITTI", ".", "Our", "pose", "estimation", "performance", "is", "limited", "by", "the", "number", "of", "training", "examples", "available", "in", "PASCAL3D", "+", ".", "We", "also", "note", "that", "the", "two", "recent", "methods", "achieve", "very", "appealing", "pose", "estimation", "results", "on", "PASCAL3D", "+", ".", "However", ",", "both", "of", "them", "utilize", "additional", "training", "images", "(", "ImageNet", "images", "in", "and", "synthetic", "images", "in", ")", "and", "conduct", "detection", "and", "pose", "estimation", "with", "separate", "CNNs", ",", "where", "a", "CNN", "is", "specifically", "designed", "for", "pose", "estimation", ".", "Our", "method", "is", "capable", "of", "simultaneous", "object", "detection", "and", "viewpoint", "estimation", "even", "in", "the", "presence", "of", "limited", "training", "examples", "per", "viewpoint", ".", "Fig", ".", "[", "reference", "]", "shows", "some", "detection", "results", "from", "our", "method", ".", "We", "again", "transfer", "segmentation", "masks", "of", "3DVPs", "to", "the", "detected", "objects", "according", "to", "the", "subcategory", "classification", "results", ".", "Please", "see", "supplementary", "material", "for", "more", "examples", ".", "To", "demonstrate", "that", "our", "method", "also", "works", "on", "datasets", "with", "bounding", "box", "annotations", "only", ",", "we", "have", "conducted", "experiments", "on", "the", "PASCAL", "VOC", "2007", "dataset", ",", "where", "subcategories", "are", "obtained", "by", "clustering", "on", "image", "features", ".", "In", "table", "[", "reference", "]", ",", "we", "compare", "with", "Fast", "R", "-", "CNN", "and", "Faster", "R", "-", "CNN", ".", "We", "have", "achieved", "comparable", "performance", "to", "the", "state", "-", "of", "-", "the", "-", "arts", ".", "Region", "proposal", "on", "PASCAL", "VOC", "is", "relatively", "easy", "compared", "to", "KITTI", ".", "So", "we", "do", "not", "see", "much", "improvement", "with", "our", "RPN", "on", "PASCAL", "VOC", "2007", ".", "section", ":", "Conclusion", "In", "this", "work", ",", "we", "explore", "how", "subcategory", "information", "can", "be", "exploited", "in", "CNN", "-", "based", "object", "detection", ".", "We", "have", "proposed", "a", "novel", "region", "proposal", "network", ",", "and", "a", "novel", "object", "detection", "network", ",", "where", "we", "explicitly", "employ", "subcategory", "information", "to", "improve", "region", "proposal", "generation", ",", "object", "detection", "and", "object", "pose", "estimation", ".", "Our", "subcategory", "-", "aware", "CNNs", "can", "also", "handle", "the", "scale", "variation", "of", "objects", "using", "image", "pyramids", "in", "an", "efficient", "way", ".", "We", "have", "conducted", "extensive", "experiments", "on", "the", "KITTI", "detection", "benchmark", ",", "the", "PASCAL3D", "+", "dataset", "and", "PASCAL", "VOC", "2007", "dataset", ".", "Our", "method", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "results", "on", "these", "benchmarks", ".", "Acknowledgments", ".", "We", "acknowledge", "the", "support", "of", "Nissan", "grant", "1188371", "-", "1", "-", "UDARQ", "and", "MURI", "grant", "1186514", "-", "1", "-", "TBCJE", ".", "bibliography", ":", "References"], "entity": [[2, 8, ["Subcategory", "-", "aware", "Convolutional", "Neural", "Networks"], "Method"], [9, 11, ["Object", "Proposals"], "Task"], [12, 13, ["Detection"], "Task"], [14, 17, ["Convolutional", "Neural", "Network"], "Method"], [18, 19, ["CNN"], "Method"], [21, 24, ["object", "detection", "methods"], "Method"], [25, 27, ["region", "proposal"], "Method"], [67, 71, ["subcategory", "-", "aware", "CNNs"], "Method"], [72, 74, ["object", "detection"], "Task"], [79, 82, ["region", "proposal", "network"], "Method"], [89, 92, ["proposal", "generating", "process"], "Method"], [96, 98, ["detection", "network"], "Method"], [99, 101, ["joint", "detection"], "Task"], [102, 104, ["subcategory", "classification"], "Task"], [137, 140, ["Convolutional", "Neural", "Networks"], "Method"], [141, 142, ["CNNs"], "Method"], [149, 151, ["recognition", "problems"], "Task"], [153, 154, ["CNNs"], "Method"], [162, 163, ["representation"], "Task"], [164, 165, ["learning"], "Task"], [173, 174, ["CNNs"], "Method"], [187, 189, ["object", "detection"], "Task"], [197, 198, ["CNNs"], "Method"], [206, 208, ["PASCAL", "VOC"], "Material"], [209, 210, ["ImageNet"], "Material"], [239, 242, ["region", "proposal", "methods"], "Method"], [270, 272, ["PASCAL", "VOC"], "Material"], [273, 274, ["ImageNet"], "Material"], [285, 287, ["autonomous", "driving"], "Task"], [299, 302, ["region", "proposal", "methods"], "Method"], [314, 317, ["Region", "Proposal", "Network"], "Method"], [328, 331, ["region", "proposal", "methods"], "Method"], [408, 410, ["autonomous", "driving"], "Task"], [411, 412, ["robotics"], "Task"], [413, 415, ["augmented", "reality"], "Task"], [431, 433, ["object", "detection"], "Task"], [448, 451, ["region", "proposal", "generation"], "Task"], [467, 468, ["detection"], "Task"], [479, 481, ["joint", "detection"], "Task"], [482, 484, ["subcategory", "classification"], "Task"], [551, 553, ["subcategory", "classification"], "Task"], [560, 562, ["detection", "methods"], "Method"], [568, 569, ["detector"], "Method"], [576, 582, ["subcategory", "convolutional", "(", "conv", ")", "layer"], "Method"], [584, 587, ["Region", "Proposal", "Network"], "Method"], [588, 589, ["RPN"], "Method"], [593, 594, ["filter"], "Method"], [596, 598, ["conv", "layer"], "Method"], [602, 604, ["subcategory", "detection"], "Task"], [606, 609, ["subcategory", "conv", "layer"], "Method"], [627, 629, ["heat", "maps"], "Method"], [631, 632, ["RPN"], "Method"], [643, 646, ["classifying", "region", "proposals"], "Task"], [673, 675, ["detection", "network"], "Method"], [679, 681, ["object", "detection"], "Task"], [682, 684, ["subcategory", "classification"], "Task"], [692, 693, ["3DVPs"], "Method"], [729, 730, ["RPN"], "Method"], [747, 750, ["feature", "extrapolating", "layer"], "Method"], [784, 787, ["PASCAL3D", "+", "dataset"], "Material"], [813, 817, ["subcategory", "-", "aware", "CNNs"], "Method"], [818, 820, ["object", "recognition"], "Task"], [825, 826, ["Subcategory"], "Method"], [830, 831, ["Subcategory"], "Method"], [837, 839, ["object", "detection"], "Task"], [844, 847, ["discovering", "object", "subcategories"], "Task"], [852, 853, ["DPM"], "Method"], [871, 872, ["clustering"], "Task"], [888, 889, ["clustering"], "Method"], [896, 897, ["3DVP"], "Method"], [898, 899, ["clustering"], "Task"], [968, 970, ["stage", "detection"], "Method"], [971, 975, ["two", "-", "stage", "detection"], "Task"], [977, 981, ["one", "-", "stage", "detection"], "Task"], [985, 987, ["Overfeat", "framework"], "Method"], [989, 990, ["CNN"], "Method"], [998, 1000, ["object", "detections"], "Task"], [1002, 1006, ["two", "-", "stage", "detection"], "Task"], [1009, 1012, ["R", "-", "CNNs"], "Method"], [1013, 1015, ["region", "proposals"], "Method"], [1025, 1028, ["region", "proposal", "methods"], "Method"], [1034, 1036, ["region", "proposals"], "Method"], [1040, 1041, ["CNN"], "Method"], [1042, 1046, ["classification", "and", "location", "refinement"], "Task"], [1051, 1053, ["detection", "paradigm"], "Method"], [1071, 1074, ["region", "proposal", "process"], "Method"], [1077, 1080, ["coarse", "detection", "step"], "Task"], [1081, 1087, ["coarse", "-", "to", "-", "fine", "detection"], "Task"], [1092, 1093, ["RPN"], "Method"], [1113, 1116, ["region", "proposal", "approach"], "Method"], [1142, 1144, ["region", "proposal"], "Method"], [1156, 1159, ["region", "proposal", "problem"], "Task"], [1160, 1161, ["CNNs"], "Method"], [1167, 1168, ["CNNs"], "Method"], [1172, 1174, ["region", "proposal"], "Task"], [1182, 1188, ["subcategory", "-", "aware", "Region", "Proposal", "Network"], "Method"], [1189, 1190, ["RPN"], "Method"], [1194, 1196, ["Network", "Architecture"], "Method"], [1200, 1202, ["network", "architecture"], "Method"], [1203, 1206, ["generating", "object", "proposals"], "Task"], [1216, 1223, ["sliding", "-", "window", "-", "based", "object", "detectors"], "Method"], [1227, 1234, ["Aggregated", "Channel", "Feature", "(", "ACF", ")", "detector"], "Method"], [1236, 1239, ["Deformable", "Part", "Model"], "Method"], [1240, 1241, ["DPM"], "Method"], [1253, 1256, ["region", "proposal", "network"], "Method"], [1270, 1271, ["RPN"], "Method"], [1272, 1274, ["image", "pyramid"], "Method"], [1281, 1289, ["convolutional", "(", "conv", ")", "and", "max", "pooling", "layers"], "Method"], [1298, 1301, ["conv", "feature", "map"], "Method"], [1313, 1317, ["computation", "of", "conv", "features"], "Task"], [1318, 1320, ["image", "pyramids"], "Task"], [1324, 1327, ["feature", "extrapolating", "layer"], "Method"], [1343, 1344, ["extrapolation"], "Method"], [1359, 1361, ["conv", "layer"], "Method"], [1362, 1365, ["object", "subcategory", "detection"], "Task"], [1368, 1369, ["filter"], "Method"], [1371, 1373, ["conv", "layer"], "Method"], [1401, 1403, ["network", "training"], "Method"], [1405, 1408, ["subcategory", "conv", "layer"], "Method"], [1449, 1452, ["RoI", "generating", "layer"], "Method"], [1462, 1464, ["heat", "maps"], "Method"], [1473, 1476, ["RoI", "pooling", "layer"], "Method"], [1483, 1486, ["conv", "feature", "maps"], "Method"], [1492, 1493, ["RPN"], "Method"], [1520, 1523, ["bounding", "box", "regressor"], "Method"], [1526, 1529, ["Feature", "Extrapolating", "Layer"], "Method"], [1531, 1532, ["RPN"], "Method"], [1535, 1540, ["fixed", "-", "size", "conv", "filters"], "Method"], [1551, 1553, ["conv", "filters"], "Method"], [1571, 1573, ["image", "pyramid"], "Task"], [1613, 1615, ["conv", "layers"], "Method"], [1616, 1619, ["max", "pooling", "layers"], "Method"], [1636, 1637, ["convolution"], "Method"], [1650, 1655, ["finely", "-", "sampled", "image", "pyramid"], "Task"], [1660, 1663, ["region", "proposal", "process"], "Task"], [1672, 1677, ["multi", "-", "resolution", "image", "features"], "Task"], [1699, 1702, ["feature", "extrapolating", "layer"], "Method"], [1705, 1709, ["computation", "of", "conv", "features"], "Task"], [1717, 1720, ["feature", "extrapolating", "layer"], "Method"], [1731, 1733, ["conv", "layer"], "Method"], [1734, 1736, ["feature", "extraction"], "Task"], [1751, 1753, ["feature", "map"], "Method"], [1809, 1812, ["feature", "extrapolating", "layer"], "Method"], [1829, 1831, ["bilinear", "interpolation"], "Method"], [1850, 1853, ["feature", "extrapolating", "layer"], "Method"], [1880, 1883, ["feature", "extrapolating", "layer"], "Method"], [1886, 1888, ["feature", "computation"], "Task"], [1894, 1897, ["Subcategory", "Conv", "Layer"], "Method"], [1907, 1910, ["subcategory", "conv", "layer"], "Method"], [1911, 1913, ["subcategory", "detection"], "Task"], [1918, 1921, ["object", "detection", "methods"], "Method"], [1924, 1925, ["classifier"], "Method"], [1935, 1937, ["conv", "filter"], "Method"], [1939, 1942, ["subcategory", "conv", "layer"], "Method"], [1959, 1962, ["subcategory", "conv", "layer"], "Method"], [1964, 1966, ["conv", "filters"], "Method"], [1969, 1971, ["conv", "filter"], "Method"], [1980, 1984, ["multi", "-", "class", "detection"], "Task"], [2012, 2014, ["conv", "filters"], "Method"], [2041, 2046, ["fixed", "-", "size", "conv", "filters"], "Method"], [2052, 2054, ["conv", "filters"], "Method"], [2083, 2085, ["loss", "layer"], "Method"], [2088, 2091, ["subcategory", "conv", "filters"], "Method"], [2094, 2097, ["RoI", "Generating", "Layer"], "Method"], [2098, 2101, ["RoI", "generating", "layer"], "Method"], [2131, 2133, ["heat", "map"], "Method"], [2147, 2149, ["RoI", "generating"], "Task"], [2164, 2167, ["RoI", "generating", "layer"], "Method"], [2178, 2180, ["max", "operation"], "Method"], [2191, 2194, ["2D", "heat", "map"], "Method"], [2216, 2218, ["generating", "strategies"], "Method"], [2223, 2224, ["training"], "Task"], [2272, 2274, ["conv", "filters"], "Method"], [2280, 2283, ["subcategory", "conv", "layer"], "Method"], [2322, 2325, ["RoI", "generating", "layer"], "Method"], [2354, 2357, ["RoI", "generating", "layer"], "Method"], [2368, 2371, ["subcategory", "conv", "filters"], "Method"], [2379, 2382, ["stochastic", "gradient", "descent"], "Method"], [2428, 2431, ["intersection", "over", "union"], "Metric"], [2432, 2433, ["IoU"], "Metric"], [2500, 2503, ["RoI", "generating", "layer"], "Method"], [2548, 2550, ["Network", "Training"], "Method"], [2557, 2560, ["RoI", "pooling", "layer"], "Method"], [2580, 2582, ["subcategory", "classification"], "Task"], [2583, 2586, ["bounding", "box", "regression"], "Task"], [2597, 2598, ["RPN"], "Method"], [2622, 2624, ["softmax", "function"], "Method"], [2629, 2632, ["subcategory", "conv", "layer"], "Method"], [2637, 2641, ["bounding", "box", "regression", "offsets"], "Method"], [2683, 2684, ["RPN"], "Method"], [2685, 2687, ["subcategory", "classification"], "Task"], [2688, 2691, ["bounding", "box", "regression"], "Task"], [2709, 2713, ["cross", "-", "entropy", "loss"], "Metric"], [2717, 2721, ["bounding", "box", "regression", "targets"], "Metric"], [2740, 2744, ["bounding", "box", "regression", "loss"], "Method"], [2750, 2753, ["bounding", "box", "regression"], "Method"], [2776, 2777, ["training"], "Task"], [2801, 2804, ["subcategory", "conv", "layers"], "Method"], [2806, 2807, ["RPN"], "Method"], [2814, 2816, ["conv", "filters"], "Method"], [2828, 2830, ["subcategory", "classification"], "Task"], [2840, 2842, ["subcategory", "detection"], "Task"], [2850, 2852, ["heat", "maps"], "Task"], [2854, 2857, ["subcategory", "conv", "layer"], "Method"], [2860, 2863, ["RoI", "generating", "layer"], "Method"], [2867, 2868, ["RPN"], "Method"], [2878, 2883, ["Subcategory", "-", "aware", "Detection", "Network"], "Method"], [2885, 2888, ["region", "proposal", "process"], "Method"], [2889, 2890, ["CNNs"], "Method"], [2902, 2904, ["region", "proposal"], "Method"], [2912, 2913, ["CNNs"], "Method"], [2918, 2920, ["detection", "step"], "Task"], [2939, 2945, ["subcategory", "-", "aware", "object", "detection", "network"], "Method"], [2954, 2956, ["Network", "Architecture"], "Method"], [2966, 2968, ["detection", "network"], "Method"], [2992, 2994, ["image", "pyramids"], "Method"], [3005, 3007, ["conv", "layer"], "Method"], [3008, 3010, ["feature", "extraction"], "Task"], [3014, 3017, ["feature", "extrapolating", "layer"], "Method"], [3038, 3039, ["RPN"], "Method"], [3043, 3046, ["RoI", "pooling", "layer"], "Method"], [3085, 3087, ["fully", "connected"], "Method"], [3088, 3089, ["FC"], "Method"], [3100, 3102, ["subcategory", "classification"], "Task"], [3176, 3178, ["subcategory", "classification"], "Task"], [3195, 3198, ["object", "class", "classification"], "Task"], [3199, 3202, ["bounding", "box", "regression"], "Task"], [3205, 3207, ["Network", "Training"], "Method"], [3215, 3219, ["multi", "-", "task", "loss"], "Task"], [3220, 3224, ["joint", "object", "class", "classification"], "Task"], [3225, 3227, ["subcategory", "classification"], "Task"], [3228, 3231, ["bounding", "box", "regression"], "Task"], [3271, 3274, ["bounding", "box", "regression"], "Method"], [3293, 3297, ["cross", "-", "entropy", "loss"], "Metric"], [3306, 3307, ["RPN"], "Method"], [3309, 3313, ["back", "-", "propagation", "training"], "Method"], [3317, 3321, ["multi", "-", "task", "loss"], "Task"], [3363, 3366, ["KITTI", "detection", "benchmark"], "Material"], [3368, 3371, ["PASCAL3D", "+", "dataset"], "Material"], [3415, 3417, ["object", "detection"], "Task"], [3454, 3457, ["PASCAL3D", "+", "dataset"], "Material"], [3480, 3483, ["3D", "CAD", "model"], "Method"], [3544, 3545, ["KITTI"], "Material"], [3549, 3551, ["detection", "framework"], "Method"], [3591, 3593, ["Average", "Precision"], "Metric"], [3594, 3595, ["AP"], "Metric"], [3601, 3603, ["detection", "performance"], "Metric"], [3614, 3616, ["overlap", "thresholds"], "Metric"], [3638, 3639, ["KITTI"], "Material"], [3641, 3644, ["Average", "Orientation", "Similarity"], "Metric"], [3645, 3647, ["AOS", ")"], "Metric"], [3651, 3653, ["orientation", "similarity"], "Metric"], [3660, 3662, ["detection", "recalls"], "Metric"], [3664, 3667, ["Average", "Segmentation", "Accuracy"], "Metric"], [3668, 3669, ["ASA"], "Metric"], [3677, 3680, ["Average", "Location", "Precision"], "Metric"], [3681, 3682, ["ALP"], "Metric"], [3684, 3686, ["joint", "detection"], "Task"], [3687, 3689, ["3D", "location"], "Task"], [3691, 3692, ["AOS"], "Method"], [3701, 3703, ["PASCAL3D", "+"], "Material"], [3704, 3707, ["PASCAL", "VOC", "2007"], "Material"], [3710, 3711, ["AP"], "Metric"], [3713, 3716, ["%", "overlap", "ratio"], "Metric"], [3720, 3722, ["object", "detection"], "Task"], [3724, 3726, ["joint", "detection"], "Task"], [3727, 3729, ["pose", "estimation"], "Task"], [3733, 3736, ["Average", "Viewpoint", "Precision"], "Metric"], [3737, 3738, ["AVP"], "Metric"], [3744, 3745, ["detection"], "Task"], [3774, 3776, ["2D", "subcategories"], "Task"], [3818, 3821, ["3D", "Voxel", "Pattern"], "Method"], [3822, 3823, ["3DVP"], "Method"], [3833, 3834, ["KITTI"], "Material"], [3839, 3841, ["PASCAL3D", "+"], "Material"], [3854, 3856, ["clustering", "process"], "Method"], [3858, 3859, ["3DVP"], "Method"], [3867, 3870, ["pedestrian", "and", "cyclist"], "Task"], [3871, 3872, ["KITTI"], "Material"], [3875, 3876, ["clustering"], "Method"], [3897, 3899, ["subcategory", "classification"], "Task"], [3925, 3926, ["validation"], "Task"], [3927, 3928, ["KITTI"], "Material"], [3967, 3968, ["KITTI"], "Material"], [4003, 4004, ["3DVPs"], "Method"], [4007, 4010, ["affinity", "propagation", "clustering"], "Method"], [4023, 4025, ["PASCAL3D", "+"], "Material"], [4036, 4039, ["PASCAL", "VOC", "2007"], "Material"], [4059, 4062, ["subcategory", "conv", "layer"], "Method"], [4064, 4065, ["RPN"], "Method"], [4074, 4076, ["detection", "network"], "Method"], [4085, 4091, ["Region", "Proposal", "Network", "Hyper", "-", "parameters"], "Method"], [4094, 4095, ["RPN"], "Method"], [4101, 4102, ["KITTI"], "Material"], [4111, 4112, ["PASCAL"], "Material"], [4114, 4116, ["PASCAL3D", "+"], "Material"], [4117, 4120, ["PASCAL", "VOC", "2007"], "Material"], [4139, 4140, ["PASCAL"], "Material"], [4148, 4149, ["KITTI"], "Material"], [4154, 4155, ["PASCAL"], "Material"], [4167, 4170, ["feature", "extrapolating", "layer"], "Method"], [4185, 4188, ["conv", "feature", "pyramid"], "Method"], [4192, 4193, ["KITTI"], "Material"], [4197, 4198, ["PASCAL"], "Material"], [4201, 4204, ["RoI", "generating", "layer"], "Method"], [4220, 4221, ["KITTI"], "Material"], [4226, 4227, ["PASCAL"], "Material"], [4248, 4249, ["RPN"], "Method"], [4251, 4255, ["SGD", "mini", "-", "batch"], "Method"], [4293, 4294, ["KITTI"], "Material"], [4295, 4296, ["PASCAL"], "Material"], [4305, 4307, ["detection", "network"], "Method"], [4318, 4319, ["KITTI"], "Material"], [4323, 4324, ["PASCAL"], "Material"], [4336, 4340, ["SGD", "mini", "-", "batch"], "Method"], [4379, 4380, ["KITTI"], "Material"], [4400, 4404, ["region", "proposal", "and", "detection"], "Task"], [4405, 4412, ["Fine", "-", "tuning", "Pre", "-", "trained", "Networks"], "Method"], [4418, 4419, ["Caffe"], "Method"], [4423, 4425, ["conv", "layers"], "Method"], [4426, 4428, ["feature", "extraction"], "Task"], [4442, 4444, ["detection", "network"], "Method"], [4447, 4449, ["trained", "networks"], "Method"], [4450, 4451, ["ImageNet"], "Material"], [4453, 4454, ["KITTI"], "Material"], [4459, 4460, ["AlexNet"], "Method"], [4462, 4464, ["VGG16", "network"], "Method"], [4466, 4467, ["GoogleNet"], "Method"], [4469, 4470, ["PASCAL"], "Material"], [4476, 4478, ["VGG16", "network"], "Method"], [4486, 4489, ["Region", "Proposal", "Evalutaion"], "Method"], [4490, 4491, ["Recall"], "Metric"], [4495, 4497, ["detection", "recall"], "Metric"], [4499, 4500, ["RPN"], "Method"], [4546, 4548, ["PASCAL", "VOC"], "Material"], [4549, 4551, ["Selective", "Search"], "Task"], [4560, 4561, ["KITTI"], "Material"], [4566, 4567, ["KITTI"], "Material"], [4582, 4587, ["bottom", "-", "up", "proposal", "method"], "Method"], [4590, 4591, ["recall"], "Metric"], [4607, 4608, ["RPN"], "Method"], [4628, 4631, ["discriminatively", "trained", "CNNs"], "Method"], [4632, 4634, ["region", "proposal"], "Task"], [4665, 4666, ["KITTI"], "Material"], [4670, 4671, ["RPN"], "Method"], [4680, 4681, ["car"], "Task"], [4686, 4687, ["pedestrian"], "Task"], [4688, 4689, ["cyclist"], "Task"], [4714, 4716, ["data", "mining"], "Task"], [4721, 4724, ["RoI", "generating", "layer"], "Method"], [4725, 4728, ["Region", "Proposal", "Evalutaion"], "Method"], [4754, 4755, ["detection"], "Task"], [4764, 4768, ["detection", "and", "orientation", "estimation"], "Task"], [4771, 4773, ["region", "proposals"], "Method"], [4780, 4784, ["detection", "and", "orientation", "estimation"], "Task"], [4786, 4787, ["RPN"], "Method"], [4794, 4795, ["RPN"], "Method"], [4801, 4803, ["detection", "network"], "Method"], [4817, 4818, ["RPN"], "Method"], [4823, 4824, ["RPN"], "Method"], [4836, 4837, ["RPN"], "Method"], [4839, 4841, ["detection", "network"], "Method"], [4859, 4860, ["RPN"], "Method"], [4862, 4864, ["detection", "network"], "Method"], [4866, 4868, ["conv", "layers"], "Method"], [4869, 4871, ["feature", "extraction"], "Task"], [4877, 4878, ["convolution"], "Task"], [4893, 4897, ["alternating", "optimization", "training", "algorithm"], "Method"], [4921, 4922, ["car"], "Task"], [4923, 4924, ["pedestrian"], "Task"], [4930, 4931, ["cyclist"], "Task"], [4950, 4951, ["RPN"], "Method"], [4953, 4955, ["detection", "network"], "Method"], [5001, 5003, ["region", "proposals"], "Method"], [5005, 5006, ["RPN"], "Method"], [5012, 5013, ["detection"], "Task"], [5014, 5016, ["orientation", "estimation"], "Task"], [5030, 5031, ["RPN"], "Method"], [5042, 5043, ["RPN"], "Method"], [5045, 5047, ["detection", "network"], "Method"], [5055, 5056, ["RPN"], "Method"], [5058, 5060, ["conv", "layers"], "Method"], [5061, 5063, ["feature", "extraction"], "Task"], [5070, 5072, ["detection", "network"], "Method"], [5088, 5089, ["RPN"], "Method"], [5091, 5093, ["detection", "network"], "Method"], [5122, 5123, ["RPN"], "Method"], [5129, 5130, ["RPN"], "Method"], [5134, 5135, ["RPNs"], "Method"], [5157, 5158, ["RPN"], "Method"], [5164, 5166, ["network", "architecture"], "Method"], [5167, 5168, ["detection"], "Task"], [5183, 5184, ["clustering"], "Method"], [5223, 5225, ["detection", "network"], "Method"], [5248, 5250, ["feature", "extrapolating"], "Method"], [5257, 5260, ["feature", "extrapolating", "layer"], "Method"], [5262, 5264, ["full", "model"], "Method"], [5286, 5290, ["orientation", "estimation", "of", "pedestrian"], "Task"], [5293, 5295, ["2D", "Segmentation"], "Task"], [5296, 5298, ["3D", "Localization"], "Task"], [5299, 5300, ["3DVPs"], "Method"], [5351, 5352, ["KITTI"], "Material"], [5358, 5360, ["3D", "localization"], "Task"], [5373, 5375, ["2D", "segmentation"], "Task"], [5376, 5380, ["3D", "localization", "of", "car"], "Task"], [5381, 5382, ["DPM"], "Method"], [5383, 5384, ["3DVP"], "Method"], [5395, 5397, ["segmentation", "accuracy"], "Metric"], [5398, 5401, ["3D", "location", "accuracy"], "Metric"], [5433, 5436, ["3D", "localization", "accuracy"], "Metric"], [5481, 5484, ["KITTI", "detection", "benchmark"], "Material"], [5506, 5507, ["KITTI"], "Material"], [5521, 5525, ["detection", "and", "orientation", "estimation"], "Task"], [5537, 5538, ["SubCNN"], "Method"], [5544, 5545, ["KITTI"], "Material"], [5554, 5556, ["VGG16", "network"], "Method"], [5558, 5559, ["GoogleNet"], "Method"], [5561, 5563, ["detection", "network"], "Task"], [5583, 5584, ["CNNs"], "Method"], [5589, 5593, ["detection", "and", "orientation", "estimation"], "Task"], [5604, 5608, ["detection", "and", "3D", "localization"], "Task"], [5610, 5611, ["KITTI"], "Material"], [5616, 5618, ["PASCAL3D", "+"], "Material"], [5619, 5621, ["PASCAL", "VOC"], "Material"], [5625, 5627, ["detection", "framework"], "Method"], [5632, 5634, ["PASCAL3D", "+"], "Material"], [5641, 5642, ["detection"], "Task"], [5644, 5645, ["AP"], "Metric"], [5654, 5655, ["AVP"], "Metric"], [5662, 5663, ["RPN"], "Method"], [5668, 5670, ["detection", "networks"], "Method"], [5673, 5675, ["feature", "extrapolation"], "Method"], [5681, 5682, ["detection"], "Task"], [5700, 5702, ["join", "detection"], "Task"], [5703, 5705, ["pose", "estimation"], "Task"], [5720, 5721, ["VDPM"], "Method"], [5722, 5727, ["DPM", "-", "VOC", "+", "VP"], "Method"], [5730, 5732, ["feature", "extrapolation"], "Task"], [5739, 5741, ["PASCAL3D", "+"], "Material"], [5747, 5749, ["PASCAL3D", "+"], "Material"], [5759, 5760, ["KITTI"], "Material"], [5762, 5764, ["pose", "estimation"], "Task"], [5775, 5777, ["PASCAL3D", "+"], "Material"], [5789, 5791, ["pose", "estimation"], "Task"], [5793, 5795, ["PASCAL3D", "+"], "Material"], [5806, 5808, ["ImageNet", "images"], "Material"], [5822, 5823, ["CNNs"], "Method"], [5826, 5827, ["CNN"], "Method"], [5831, 5833, ["pose", "estimation"], "Task"], [5863, 5864, ["detection"], "Task"], [5883, 5885, ["subcategory", "classification"], "Task"], [5926, 5927, ["clustering"], "Method"], [5965, 5967, ["Region", "proposal"], "Method"], [5968, 5970, ["PASCAL", "VOC"], "Material"], [5975, 5976, ["KITTI"], "Material"], [5986, 5987, ["RPN"], "Method"], [5988, 5991, ["PASCAL", "VOC", "2007"], "Material"], [6019, 6022, ["region", "proposal", "network"], "Method"], [6038, 6041, ["region", "proposal", "generation"], "Task"], [6042, 6044, ["object", "detection"], "Task"], [6050, 6054, ["subcategory", "-", "aware", "CNNs"], "Method"], [6077, 6080, ["KITTI", "detection", "benchmark"], "Material"], [6082, 6085, ["PASCAL3D", "+", "dataset"], "Material"], [50, 52, ["object", "detection"], "Task"], [127, 129, ["pose", "estimation"], "Task"], [227, 228, ["CNN"], "Method"], [230, 232, ["object", "detection"], "Task"], [282, 283, ["KITTI"], "Material"], [318, 319, ["RPN"], "Method"], [354, 355, ["CNN"], "Method"], [357, 359, ["object", "detection"], "Task"], [364, 366, ["object", "detection"], "Task"], [441, 442, ["CNN"], "Method"], [444, 446, ["object", "detection"], "Task"], [456, 457, ["CNN"], "Method"], [477, 478, ["CNN"], "Method"], [492, 494, ["object", "detection"], "Task"], [655, 657, ["object", "detection"], "Task"], [670, 671, ["CNN"], "Method"], [733, 734, ["CNN"], "Method"], [780, 781, ["KITTI"], "Material"], [789, 790, ["PASCAL"], "Material"], [828, 829, ["Detection"], "Task"], [921, 922, ["CNN"], "Method"], [939, 940, ["CNN"], "Method"], [943, 944, ["Detection"], "Task"], [956, 957, ["CNN"], "Method"], [959, 961, ["object", "detection"], "Task"], [1059, 1063, ["two", "-", "stage", "detection"], "Task"], [1105, 1106, ["RPN"], "Method"], [2448, 2449, ["IoU"], "Metric"], [2925, 2927, ["fully", "connected"], "Method"], [2946, 2948, ["joint", "detection"], "Task"], [2949, 2951, ["subcategory", "classification"], "Task"], [2979, 2980, ["CNN"], "Method"], [3095, 3096, ["FC"], "Method"], [3109, 3110, ["FC"], "Method"], [3172, 3173, ["FC"], "Method"], [3192, 3193, ["FC"], "Method"], [3210, 3212, ["object", "detection"], "Task"], [3358, 3360, ["object", "detection"], "Task"], [3373, 3374, ["PASCAL"], "Material"], [3381, 3382, ["KITTI"], "Material"], [3425, 3426, ["KITTI"], "Material"], [3435, 3436, ["KITTI"], "Material"], [3463, 3464, ["PASCAL"], "Material"], [3488, 3489, ["PASCAL"], "Material"], [3520, 3521, ["PASCAL"], "Material"], [3620, 3621, ["KITTI"], "Material"], [3632, 3634, ["joint", "detection"], "Task"], [3671, 3673, ["joint", "detection"], "Task"], [4070, 4071, ["FC"], "Method"], [4286, 4287, ["IoU"], "Metric"], [4297, 4298, ["Detection"], "Task"], [4371, 4372, ["IoU"], "Metric"], [4434, 4435, ["FC"], "Method"], [4438, 4439, ["FC"], "Method"], [4483, 4484, ["KITTI"], "Material"], [4520, 4521, ["KITTI"], "Material"], [4612, 4613, ["CNN"], "Method"], [4678, 4679, ["CNN"], "Method"], [4729, 4730, ["Detection"], "Task"], [4734, 4735, ["Detection"], "Task"], [4791, 4792, ["CNN"], "Method"], [4828, 4829, ["CNN"], "Method"], [5107, 5108, ["Detection"], "Task"], [5141, 5142, ["CNN"], "Method"], [5144, 5145, ["KITTI"], "Material"], [5207, 5208, ["CNN"], "Method"], [5386, 5387, ["KITTI"], "Material"], [5463, 5464, ["KITTI"], "Material"], [5488, 5489, ["RPN"], "Method"], [5495, 5496, ["KITTI"], "Material"], [5647, 5649, ["joint", "detection"], "Task"], [5650, 5652, ["pose", "estimation"], "Task"], [5689, 5690, ["CNN"], "Method"], [5736, 5738, ["pose", "estimation"], "Task"], [5818, 5820, ["pose", "estimation"], "Task"], [5840, 5842, ["object", "detection"], "Task"], [5916, 5917, ["PASCAL"], "Material"], [5943, 5944, ["CNN"], "Method"], [5948, 5949, ["CNN"], "Method"], [6008, 6009, ["CNN"], "Method"], [6011, 6013, ["object", "detection"], "Task"], [6026, 6028, ["object", "detection"], "Task"], [6046, 6048, ["pose", "estimation"], "Task"], [6086, 6087, ["PASCAL"], "Material"]]} {"doc_id": "1e5b1bf032059caddb4ef0d6bfec4706d8f3a946", "text": ["Geometric", "Matrix", "Completion", "with", "Recurrent", "Multi", "-", "Graph", "Neural", "Networks", "section", ":", "Abstract", "Matrix", "completion", "models", "are", "among", "the", "most", "common", "formulations", "of", "recommender", "systems", ".", "Recent", "works", "have", "showed", "a", "boost", "of", "performance", "of", "these", "techniques", "when", "introducing", "the", "pairwise", "relationships", "between", "users", "/", "items", "in", "the", "form", "of", "graphs", ",", "and", "imposing", "smoothness", "priors", "on", "these", "graphs", ".", "However", ",", "such", "techniques", "do", "not", "fully", "exploit", "the", "local", "stationarity", "structures", "of", "user", "/", "item", "graphs", ",", "and", "the", "number", "of", "parameters", "to", "learn", "is", "linear", "w.r.t", ".", "the", "number", "of", "users", "and", "items", ".", "We", "propose", "a", "novel", "approach", "to", "overcome", "these", "limitations", "by", "using", "geometric", "deep", "learning", "on", "graphs", ".", "Our", "matrix", "completion", "architecture", "combines", "graph", "convolutional", "neural", "networks", "and", "recurrent", "neural", "networks", "to", "learn", "meaningful", "statistical", "graph", "-", "structured", "patterns", "and", "the", "non", "-", "linear", "diffusion", "process", "that", "generates", "the", "known", "ratings", ".", "This", "neural", "network", "system", "requires", "a", "constant", "number", "of", "parameters", "independent", "of", "the", "matrix", "size", ".", "We", "apply", "our", "method", "on", "both", "synthetic", "and", "real", "datasets", ",", "showing", "that", "it", "outperforms", "state", "-", "of", "-", "the", "-", "art", "techniques", ".", "section", ":", "Introduction", "Recommender", "systems", "have", "become", "a", "central", "part", "of", "modern", "intelligent", "systems", ".", "Recommending", "movies", "on", "Netflix", ",", "friends", "on", "Facebook", ",", "furniture", "on", "Amazon", ",", "jobs", "on", "LinkedIn", "are", "a", "few", "examples", "of", "the", "main", "purpose", "of", "these", "systems", ".", "Two", "major", "approach", "to", "recommender", "systems", "are", "collaborative", "[", "reference", "]", "and", "content", "[", "reference", "]", "use", "of", "similarities", "between", "products", "and", "customers", "to", "recommend", "new", "products", ".", "Hybrid", "systems", "combine", "collaborative", "and", "content", "techniques", ".", "Matrix", "completion", ".", "Mathematically", ",", "a", "recommendation", "method", "can", "be", "posed", "as", "a", "matrix", "completion", "problem", "[", "reference", "]", ",", "where", "columns", "and", "rows", "represent", "users", "and", "items", ",", "respectively", ",", "and", "matrix", "values", "represent", "a", "score", "determining", "whether", "a", "user", "would", "like", "an", "item", "or", "not", ".", "Given", "a", "small", "subset", "of", "known", "elements", "of", "the", "matrix", ",", "the", "goal", "is", "to", "fill", "in", "the", "rest", ".", "A", "famous", "example", "is", "the", "Netflix", "challenge", "[", "reference", "]", ")", "offered", "in", "2009", "and", "carrying", "a", "1M$", "prize", "for", "the", "algorithm", "that", "can", "best", "predict", "user", "ratings", "for", "movies", "based", "on", "previous", "ratings", ".", "The", "size", "of", "the", "Netflix", "is", "480k", "movies", "×", "18k", "users", "(", "8.5B", "entries", ")", ",", "with", "only", "0.011", "%", "known", "entries", ".", "Recently", ",", "there", "have", "been", "several", "attempts", "to", "incorporate", "geometric", "structure", "into", "matrix", "completion", "problems", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ",", "e.g.", "in", "the", "form", "of", "column", "and", "row", "graphs", "representing", "similarity", "of", "users", "and", "items", ",", "respectively", ".", "Such", "additional", "information", "makes", "well", "-", "defined", "e.g.", "the", "notion", "of", "smoothness", "of", "data", "and", "was", "shown", "beneficial", "for", "the", "performance", "of", "recommender", "systems", ".", "These", "approaches", "can", "be", "generally", "related", "to", "the", "field", "of", "signal", "processing", "on", "graphs", "[", "reference", "]", ",", "extending", "classical", "harmonic", "analysis", "methods", "to", "non", "-", "Euclidean", "domains", ".", "Geometric", "deep", "learning", ".", "Of", "key", "interest", "to", "the", "design", "of", "recommender", "systems", "are", "deep", "learning", "approaches", ".", "In", "the", "recent", "years", ",", "deep", "neural", "networks", "and", ",", "in", "particular", ",", "convolutional", "neural", "networks", "(", "CNNs", ")", "[", "reference", "]", "have", "been", "applied", "with", "great", "success", "to", "numerous", "computer", "vision", "-", "related", "applications", ".", "However", ",", "original", "CNN", "models", "can", "not", "be", "directly", "applied", "to", "the", "recommendation", "problem", "to", "extract", "meaningful", "patterns", "in", "users", ",", "items", "and", "ratings", "because", "these", "data", "are", "not", "Euclidean", "structured", ",", "i.e.", "they", "do", "not", "lie", "on", "regular", "lattices", "like", "images", "but", "irregular", "domains", "like", "graphs", "or", "manifolds", ".", "This", "strongly", "motivates", "the", "development", "of", "geometric", "deep", "learning", "[", "reference", "]", "techniques", "that", "can", "mathematically", "deal", "with", "graph", "-", "structured", "data", ",", "which", "arises", "in", "numerous", "applications", ",", "ranging", "from", "computer", "graphics", "and", "vision", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "to", "chemistry", "[", "reference", "]", ".", "The", "earliest", "attempts", "to", "apply", "neural", "networks", "to", "graphs", "are", "due", "to", "[", "reference", "]", "(", "see", "more", "recent", "formulation", "[", "reference", "][", "reference", "]", ")", ".", "[", "reference", "][", "reference", "]", "formulated", "CNN", "-", "like", "deep", "neural", "architectures", "on", "graphs", "in", "the", "spectral", "domain", ",", "employing", "the", "analogy", "between", "the", "classical", "Fourier", "transforms", "and", "projections", "onto", "the", "eigenbasis", "of", "the", "graph", "Laplacian", "operator", "[", "reference", "]", ".", "In", "a", "follow", "-", "up", "work", ",", "proposed", "an", "efficient", "filtering", "scheme", "using", "recurrent", "Chebyshev", "polynomials", ",", "which", "reduces", "the", "complexity", "of", "CNNs", "on", "graphs", "to", "the", "same", "complexity", "of", "standard", "CNNs", "(", "on", "grids", ")", ".", "This", "model", "was", "later", "extended", "to", "deal", "with", "dynamic", "data", "[", "reference", "]", "Main", "contribution", ".", "In", "this", "work", ",", "we", "treat", "matrix", "completion", "problem", "as", "deep", "learning", "on", "graph", "-", "structured", "data", ".", "We", "introduce", "a", "novel", "neural", "network", "architecture", "that", "is", "able", "to", "extract", "local", "stationary", "patterns", "from", "the", "highdimensional", "spaces", "of", "users", "and", "items", ",", "and", "use", "these", "meaningful", "representations", "to", "infer", "the", "non", "-", "linear", "temporal", "diffusion", "mechanism", "of", "ratings", ".", "The", "spatial", "patterns", "are", "extracted", "by", "a", "new", "CNN", "architecture", "designed", "to", "work", "on", "multiple", "graphs", ".", "The", "temporal", "dynamics", "of", "the", "rating", "diffusion", "is", "produced", "by", "a", "Long", "-", "Short", "Term", "Memory", "(", "LSTM", ")", "recurrent", "neural", "network", "(", "RNN", ")", "[", "reference", "]", ".", "To", "our", "knowledge", ",", "our", "work", "is", "the", "first", "application", "of", "graph", "-", "based", "deep", "learning", "to", "matrix", "completion", "problem", ".", "The", "rest", "of", "the", "paper", "is", "organized", "as", "follows", ".", "Section", "2", "reviews", "the", "matrix", "completion", "models", ".", "Section", "3", "presents", "the", "proposed", "approach", ".", "Section", "4", "presents", "experimental", "results", "demonstrating", "the", "efficiency", "of", "our", "techniques", "on", "synthetic", "and", "real", "-", "world", "datasets", ",", "and", "Section", "5", "concludes", "the", "paper", ".", "section", ":", "Background", "section", ":", "Matrix", "Completion", "Matrix", "completion", "problem", ".", "Recovering", "the", "missing", "values", "of", "a", "matrix", "given", "a", "small", "fraction", "of", "its", "entries", "is", "an", "ill", "-", "posed", "problem", "without", "additional", "mathematical", "constraints", "on", "the", "space", "of", "solutions", ".", "A", "well", "-", "posed", "problem", "is", "to", "assume", "that", "the", "variables", "lie", "in", "a", "smaller", "subspace", ",", "i.e.", ",", "that", "the", "matrix", "is", "of", "low", "rank", ",", "where", "X", "denotes", "the", "matrix", "to", "recover", ",", "Ω", "is", "the", "set", "of", "the", "known", "entries", "and", "y", "ij", "are", "their", "values", ".", "To", "make", "(", "1", ")", "robust", "against", "noise", "and", "perturbation", ",", "the", "equality", "constraint", "can", "be", "replaced", "with", "a", "penalty", "where", "Ω", "is", "the", "indicator", "matrix", "of", "the", "known", "entries", "Ω", "and", "•", "denotes", "the", "Hadamard", "pointwise", "product", ".", "Unfortunately", ",", "rank", "minimization", "turns", "out", "an", "NP", "-", "hard", "combinatorial", "problem", "that", "is", "computationally", "intractable", "in", "practical", "cases", ".", "The", "tightest", "possible", "convex", "relaxation", "of", "the", "previous", "problem", "is", "where", "·", "is", "the", "nuclear", "norm", "of", "a", "matrix", "equal", "to", "the", "sum", "of", "its", "singular", "values", "[", "reference", "]", ")", ".", "[", "reference", "]", "proved", "that", "the", "1", "relaxation", "of", "the", "SVD", "lead", "to", "solutions", "that", "recover", "almost", "exactly", "the", "original", "low", "-", "rank", "matrix", ".", "Geometric", "matrix", "completion", "An", "alternative", "relaxation", "of", "the", "rank", "operator", "in", "(", "1", ")", "is", "to", "constraint", "the", "space", "of", "solutions", "to", "be", "smooth", "w.r.t", ".", "some", "geometric", "structure", "of", "the", "matrix", "rows", "and", "columns", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "The", "simplest", "model", "is", "proximity", "structure", "represented", "as", "an", "undirected", "weighted", "column", "graph", "G", "c", "=", "(", "{", "1", ",", ".", ".", ".", ",", "n", "}", ",", "E", "c", ",", "W", "c", ")", "with", "In", "our", "notation", ",", "the", "column", "graph", "could", "be", "thought", "of", "as", "a", "social", "network", "capturing", "relations", "between", "users", "and", "the", "similarity", "of", "their", "tastes", ".", "The", "row", "graph", "G", "r", "=", "(", "{", "1", ",", ".", ".", ".", ",", "m", "}", ",", "E", "r", ",", "W", "r", ")", "representing", "the", "similarities", "of", "the", "items", "is", "defined", "in", "a", "similar", "manner", ".", "On", "each", "of", "these", "graphs", "one", "can", "construct", "the", "(", "unnormalized", ")", "graph", "Laplacian", ",", "an", "n", "×", "n", "symmetric", "positivesemidefinite", "matrix", "w", "ij", "is", "the", "degree", "matrix", ".", "We", "denote", "the", "Laplacian", "associated", "with", "row", "and", "column", "graphs", "by", "∆", "r", "and", "∆", "c", ",", "respectively", ".", "Considering", "the", "columns", "(", "respectively", ",", "rows", ")", "of", "matrix", "X", "as", "vector", "-", "valued", "functions", "on", "the", "column", "graph", "G", "c", "(", "respectively", ",", "row", "graph", "G", "r", ")", ",", "their", "smoothness", "can", "be", "expressed", "as", "the", "Dirichlet", "norm", "X", "2", "Gr", "=", "trace", "(", "X", "∆", "r", "X", ")", "(", "respecitvely", ",", "X", "Gc", "=", "trace", "(", "X∆", "c", "X", ")", ")", ".", "The", "geometric", "matrix", "completion", "problem", "thus", "boils", "down", "to", "minimizing", "n", "users", "Factorized", "models", ".", "Matrix", "completion", "algorithms", "introduced", "in", "the", "previous", "section", "are", "well", "-", "posed", "as", "convex", "optimization", "problems", ",", "guaranteeing", "existence", ",", "uniqueness", "and", "robustness", "of", "solutions", ".", "Besides", ",", "fast", "algorithms", "have", "been", "developed", "in", "the", "context", "of", "compressed", "sensing", "to", "solve", "the", "non", "-", "differential", "nuclear", "norm", "problem", ".", "However", ",", "the", "variables", "in", "this", "formulation", "are", "the", "full", "m", "×", "n", "matrix", "X", ",", "making", "such", "methods", "hard", "to", "scale", "up", "to", "large", "matrices", "such", "as", "the", "notorious", "Netflix", "challenge", ".", "A", "solution", "is", "to", "use", "a", "factorized", "representation", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ")", "X", "=", "WH", ",", "where", "W", ",", "H", "are", "m", "×", "r", "and", "n", "×", "r", "matrices", ",", "respectively", ",", "with", "r", "min", "(", "m", ",", "n", ")", ".", "The", "use", "of", "factors", "W", ",", "H", "reduce", "the", "number", "of", "degrees", "of", "freedom", "from", "O", "(", "mn", ")", "to", "O", "(", "m", "+", "n", ")", ";", "this", "representation", "is", "also", "attractive", "as", "solving", "the", "matrix", "completion", "problem", "often", "assumes", "the", "original", "matrix", "to", "be", "low", "-", "rank", ",", "and", "rank", "(", "WH", ")", "≤", "r", "by", "construction", ".", "Figure", "1", "shows", "the", "full", "and", "factorized", "settings", "of", "the", "matrix", "completion", "problem", ".", "The", "nuclear", "norm", "minimization", "problem", "in", "the", "previous", "section", "can", "be", "equivalently", "rewritten", "in", "a", "factorized", "form", "as", "[", "reference", "]", ":", "and", "the", "factorized", "formulation", "of", "the", "graph", "-", "based", "minimization", "problem", "(", "4", ")", "as", "The", "limitation", "of", "model", "(", "6", ")", "is", "to", "decouple", "the", "regularization", "process", "applied", "simultaneously", "on", "the", "rows", "and", "columns", "of", "X", "in", "(", "4", ")", ",", "but", "the", "advantage", "is", "linear", "instead", "of", "quadratic", "complexity", ".", "section", ":", "Deep", "learning", "on", "graphs", "The", "key", "idea", "to", "our", "work", "is", "geometric", "deep", "learning", ",", "an", "extension", "of", "the", "popular", "CNNs", "to", "graphs", ".", "A", "graph", "Laplacian", "admits", "a", "spectral", "eigendecomposition", "of", "the", "form", "∆", "=", "ΦΛΦ", ",", "where", "Φ", "=", "(", "φ", "1", ",", ".", ".", ".", "φ", "n", ")", "denotes", "the", "matrix", "of", "orthonormal", "eigenvectors", "and", "Λ", "=", "diag", "(", "λ", "1", ",", ".", ".", ".", ",", "λ", "n", ")", "is", "the", "diagonal", "matrix", "of", "the", "corresponding", "eigenvalues", ".", "The", "eigenvectors", "play", "the", "role", "of", "Fourier", "atoms", "in", "classical", "harmonic", "analysis", "and", "the", "eigenvalues", "can", "be", "interpreted", "as", "frequencies", ".", "Given", "a", "function", "x", "=", "(", "x", "1", ",", ".", ".", ".", ",", "x", "n", ")", "on", "the", "vertices", "of", "the", "graph", ",", "its", "graph", "Fourier", "transform", "is", "given", "byx", "=", "Φ", "x.", "The", "spectral", "convolution", "of", "two", "functions", "x", ",", "y", "can", "be", "defined", "as", "the", "element", "-", "wise", "product", "of", "the", "respective", "Fourier", "transforms", ",", "Bruna", "et", "al", ".", "2013", "used", "the", "spectral", "definition", "of", "convolution", "(", "7", ")", "to", "generalize", "CNNs", "on", "graphs", ".", "A", "spectral", "convolutional", "layer", "has", "the", "form", "where", "q", ",", "q", "denote", "the", "number", "of", "input", "and", "output", "channels", ",", "respectively", ",", "Ŷ", "ll", "=", "diag", "(", "ŷ", "ll", ",", "1", ",", ".", ".", ".", ",", "ŷ", "ll", ",", "n", ")", "is", "a", "diagonal", "matrix", "of", "spectral", "multipliers", "representing", "a", "learnable", "filter", "in", "the", "spectral", "domain", ",", "and", "ξ", "is", "a", "nonlinearity", "(", "e.g.", "ReLU", ")", "applied", "on", "the", "vertex", "-", "wise", "function", "values", ".", "Unlike", "classical", "convolutions", "carried", "out", "efficiently", "in", "the", "spectral", "domain", "using", "FFT", ",", "the", "computations", "of", "the", "forward", "and", "inverse", "graph", "Fourier", "transform", "incur", "expensive", "O", "(", "n", "2", ")", "multiplication", "by", "the", "matrices", "Φ", ",", "Φ", ",", "as", "there", "are", "no", "FFTlike", "algorithms", "on", "general", "graphs", ".", "Furthermore", ",", "there", "is", "no", "guarantee", "that", "the", "filters", "represented", "in", "the", "spectral", "domain", "are", "localized", "in", "the", "spatial", "domain", ",", "which", "is", "an", "important", "property", "of", "classical", "CNNs", ".", "To", "address", "these", "issues", ",", "[", "reference", "]", "proposed", "using", "an", "explicit", "expansion", "in", "the", "Chebyshev", "polynomial", "basis", "to", "represent", "the", "spectral", "filters", "row", "+", "column", "filtering", "Figure", "2", ".", "Recurrent", "GCNN", "(", "RGCNN", ")", "architecture", "using", "the", "full", "matrix", "completion", "model", "and", "operating", "simultaneously", "on", "the", "rows", "and", "columns", "of", "the", "matrix", "X.", "The", "output", "of", "the", "Multi", "-", "Graph", "CNN", "(", "MGCNN", ")", "module", "is", "a", "q", "-", "dimensional", "feature", "vector", "for", "each", "element", "of", "the", "input", "matrix", ".", "The", "number", "of", "parameters", "to", "learn", "is", "O", "(", "1", ")", "and", "the", "learning", "complexity", "is", "O", "(", "mn", ")", ".", "n", "∆", "−", "I", "is", "the", "rescaled", "Laplacian", "such", "that", "its", "eigenvaluesΛ", "=", "2λ", "θ", "is", "the", "p", "-", "dimensional", "vector", "of", "polynomial", "coefficients", "parametrizing", "the", "filter", ",", "and", "T", "j", "(", "λ", ")", "=", "2λT", "j−1", "(", "λ", ")", "−T", "j−2", "(", "λ", ")", "denotes", "the", "Chebyshev", "polynomial", "of", "degree", "j", "defined", "in", "a", "recursive", "manner", "with", "T", "1", "(", "λ", ")", "=", "λ", "and", "T", "0", "(", "λ", ")", "=", "1", ".", "1", "This", "approach", "benefits", "from", "several", "advantages", ".", "First", ",", "it", "does", "not", "require", "an", "explicit", "computation", "of", "the", "Laplacian", "eigenvectors", ",", "and", "due", "to", "the", "recursive", "definition", "of", "the", "Chebyshev", "polynomials", ",", "the", "computation", "of", "the", "filter", "incurs", "applying", "the", "Laplacian", "p", "times", ".", "Multiplication", "by", "Laplacian", "has", "the", "cost", "of", "O", "(", "|E|", ")", ",", "and", "assuming", "the", "graph", "has", "|E|", "=", "O", "(", "n", ")", "edges", "(", "which", "is", "the", "case", "for", "k", "-", "nearest", "neighbors", "graphs", "and", "most", "real", "-", "world", "networks", ")", ",", "the", "overall", "complexity", "is", "O", "(", "n", ")", "rather", "than", "O", "(", "n", "2", ")", "operations", ",", "which", "is", "the", "same", "complexity", "than", "standard", "CNNs", ".", "Moreover", ",", "since", "the", "Laplacian", "is", "a", "local", "operator", "affecting", "only", "1", "-", "hop", "neighbors", "of", "a", "vertex", "and", "accordingly", "its", "(", "p", "−", "1", ")", "st", "power", "affects", "the", "p", "-", "hop", "neighborhood", ",", "the", "resulting", "filters", "are", "spatially", "localized", ".", "section", ":", "Our", "approach", "In", "this", "paper", ",", "we", "propose", "formulating", "matrix", "completion", "as", "a", "learnable", "diffusion", "process", "applied", "to", "the", "score", "values", ".", "The", "deep", "learning", "architecture", "considered", "for", "this", "purpose", "consists", "of", "a", "spatial", "part", "extracting", "spatial", "features", "from", "the", "matrix", "(", "we", "consider", "two", "different", "approaches", "working", "on", "the", "full", "and", "factorized", "matrix", "models", ")", ",", "and", "a", "temporal", "part", "using", "a", "recurrent", "LSTM", "network", ".", "The", "two", "architectures", "are", "1", "Tj", "(", "λ", ")", "=", "cos", "(", "j", "cos", "−1", "(", "λ", ")", ")", "is", "an", "oscillating", "function", "on", "[", "−1", ",", "1", "]", "with", "j", "roots", ",", "j", "+", "1", "equally", "spaced", "extrema", ",", "and", "a", "frequency", "linearly", "dependent", "on", "j.", "Chebyshev", "polynomials", "form", "an", "orthogonal", "basis", "for", "the", "space", "of", "smooth", "functions", "on", "[", "−1", ",", "1", "]", "and", "are", "thus", "convenient", "to", "compactly", "represent", "spectral", "filters", ".", "summarized", "in", "Figures", "2", "and", "3", "and", "described", "in", "details", "in", "the", "following", ".", "section", ":", "Multi", "-", "Graph", "CNNs", "Multi", "-", "graph", "convolution", ".", "Our", "first", "goal", "is", "to", "extend", "the", "notion", "of", "the", "aforementioned", "graph", "Fourier", "transform", "to", "matrices", "whose", "rows", "and", "columns", "are", "defined", "on", "row", "-", "and", "column", "-", "graphs", ".", "We", "recall", "that", "a", "classical", "two", "-", "dimensional", "Fourier", "transform", "of", "an", "image", "(", "matrix", ")", "can", "be", "thought", "of", "as", "applying", "a", "one", "-", "dimensional", "Fourier", "transform", "to", "its", "rows", "and", "columns", ".", "In", "our", "setting", ",", "the", "analogy", "of", "the", "twodimensional", "Fourier", "transform", "has", "the", "form", "where", "Φ", "c", ",", "Φ", "r", "and", "Λ", "c", ",", "Λ", "r", "denote", "the", "n", "×", "n", "and", "m", "×", "m", "eigenvector", "-", "and", "eigenvalue", "matrices", "of", "the", "column", "-", "and", "row", "-", "graph", "Laplacians", "∆", "c", ",", "∆", "r", ",", "respectively", ".", "The", "multigraph", "version", "of", "the", "spectral", "convolution", "(", "7", ")", "is", "given", "by", "Representing", "the", "filters", "as", "their", "spectral", "multipliersŶ", "would", "yield", "O", "(", "mn", ")", "parameters", ",", "prohibitive", "in", "any", "practical", "application", ".", "To", "overcome", "this", "limitation", ",", "we", "resort", "to", "the", "representation", "of", "the", "filters", "in", "Chebychev", "polynomial", "bases", "of", "degree", "p", ",", "where", "Θ", "=", "(", "θ", "jj", ")", "is", "the", "(", "p", "+", "1", ")", "×", "(", "p", "+", "1", ")", "matrix", "of", "coefficients", ",", "i.e.", ",", "O", "(", "1", ")", "parameters", ".", "The", "application", "of", "such", "filters", "to", "the", "matrix", "X", "A", "Multi", "-", "Graph", "CNN", "(", "MGCNN", ")", "using", "this", "parametrization", "of", "filters", "(", "13", ")", "in", "the", "convolutional", "layer", "is", "applied", "to", "the", "m×n", "matrix", "X", "(", "single", "input", "channel", ")", ",", "producing", "q", "outputs", "(", "i.e.", ",", "a", "tensor", "of", "size", "m", "×", "n", "×", "q", ")", ".", "Separable", "convolution", ".", "A", "simplification", "of", "the", "multigraph", "convolution", "is", "obtained", "considering", "the", "factorized", "form", "of", "the", "matrix", "X", "=", "WH", "and", "applying", "onedimensional", "convolution", "on", "the", "respective", "graph", "to", "each", "factor", ",", "w", "with", "2", "(", "p", "+", "1", ")", "qq", "parameters", "in", "total", ".", "section", ":", "Matrix", "diffusion", "with", "RNN", "The", "next", "step", "of", "our", "approach", "is", "to", "feed", "the", "features", "extracted", "from", "the", "matrix", "by", "the", "MGCNN", "(", "or", "alternatively", ",", "the", "row", "-", "and", "column", "-", "GCNNs", ")", "to", "a", "Recurrent", "Neural", "Network", "(", "RNN", ")", "implementing", "the", "score", "diffusion", "process", ".", "We", "use", "the", "classical", "Long", "-", "Short", "Term", "Memory", "(", "LSTM", ")", "RNN", "architecture", "[", "reference", "]", ",", "which", "has", "demonstrated", "to", "be", "highly", "efficient", "to", "learn", "the", "dynamical", "property", "of", "data", "sequences", "as", "LSTM", "is", "able", "to", "keep", "long", "-", "term", "internal", "states", "(", "in", "particular", ",", "avoiding", "the", "vanishing", "gradient", "issue", ")", ".", "The", "input", "of", "the", "LSTM", "gate", "is", "given", "by", "the", "static", "features", "extracted", "from", "the", "MGCNN", ",", "which", "can", "be", "seen", "as", "a", "projection", "or", "dimensionality", "reduction", "of", "the", "original", "matrix", "in", "the", "space", "of", "the", "most", "meaningful", "and", "representative", "information", "(", "the", "disentanglement", "effect", ")", ".", "This", "representation", "coupled", "with", "LSTM", "appears", "particularly", "well", "-", "suited", "to", "keep", "a", "long", "term", "internal", "state", ",", "which", "allows", "to", "predict", "accurate", "small", "changes", "dX", "of", "the", "matrix", "X", "(", "or", "dW", ",", "dH", "of", "the", "factors", "W", ",", "H", ")", "that", "can", "propagate", "through", "the", "full", "temporal", "steps", ".", "Figures", "2", "and", "3", "provides", "an", "illustration", "of", "the", "proposed", "matrix", "completion", "model", ".", "We", "also", "give", "a", "precise", "description", "of", "the", "two", "settings", "of", "our", "model", "in", "Algorithms", "1", "and", "2", ".", "We", "refer", "to", "the", "whole", "architecture", "combining", "the", "MGCNN", "and", "RNN", "in", "the", "full", "matrix", "completion", "setting", "as", "Recurrent", "Graph", "CNN", "(", "RGCNN", ")", ".", "The", "factorized", "version", "with", "two", "GCNNs", "and", "RNN", "is", "referred", "to", "as", "separable", "Recurrent", "Graph", "CNN", "(", "sRGCNN", ")", ".", "The", "complexity", "of", "Algorithm", "1", "scales", "quadratically", "as", "O", "(", "mn", ")", "due", "to", "the", "use", "of", "MGCNN", ".", "For", "large", "matrices", ",", "we", "can", "opt", "for", "Algorithm", "2", "that", "processes", "the", "rows", "and", "columns", "separately", "with", "standard", "GCNNs", "and", "scales", "linearly", "as", "O", "(", "m", "+", "n", ")", ".", "section", ":", "Algorithm", "1", "Full", "matrix", "completion", "model", "using", "RGCNN", "input", "m", "×", "n", "matrix", "X", "(", "0", ")", "containing", "initial", "values", "1", ":", "for", "t", "=", "0", ":", "T", "do", "2", ":", "Apply", "the", "Multi", "-", "Graph", "CNN", "(", "13", ")", "on", "X", "(", "t", ")", "producing", "an", "m×n×q", "outputX", "(", "t", ")", "containing", "a", "q", "-", "dimensional", "feature", "vector", "for", "each", "matrix", "element", ".", "section", ":", "3", ":", "for", "all", "elements", "(", "i", ",", "j", ")", "do", "section", ":", "4", ":", "Apply", "RNN", "to", "feature", "vectorx", "ijq", ")", "producing", "the", "predicted", "incremental", "value", "dx", "(", "t", ")", "ij", "5", ":", "7", ":", "end", "for", "Algorithm", "2", "Factorized", "matrix", "completion", "model", "using", "sRGCNN", "input", "m×r", "factor", "H", "(", "0", ")", "and", "n×r", "factor", "W", "(", "0", ")", "representing", "the", "matrix", "X", "Apply", "the", "Graph", "CNN", "on", "H", "(", "t", ")", "producing", "an", "n", "×", "q", "outputH", "(", "t", ")", ".", "3", ":", "Apply", "RNN", "to", "feature", "vectorh", "jq", ")", "producing", "the", "predicted", "incremental", "value", "dh", "Update", "Apply", "the", "Graph", "CNN", "on", "W", "(", "t", ")", "producing", "an", "m", "×", "q", "outputW", "(", "t", ")", ".", "8", ":", "Apply", "RNN", "to", "feature", "vectorw", "iq", ")", "producing", "the", "predicted", "incremental", "value", "dw", "12", ":", "end", "for", "section", ":", "Training", "Training", "of", "the", "networks", "is", "performed", "by", "minimizing", "the", "loss", "Here", ",", "T", "denotes", "the", "number", "of", "diffusion", "iterations", "(", "applications", "of", "the", "RNN", ")", ",", "and", "we", "use", "the", "notation", "X", "(", "T", ")", "Θ", ",", "σ", "to", "emphasize", "that", "the", "matrix", "depends", "on", "the", "parameters", "of", "the", "MGCNN", "(", "Chebyshev", "polynomial", "coefficients", "Θ", ")", "and", "those", "of", "the", "LSTM", "(", "denoted", "by", "σ", ")", ".", "In", "the", "factorized", "setting", ",", "we", "use", "the", "loss", "where", "θ", "c", ",", "θ", "r", "are", "the", "parameters", "of", "the", "two", "GCNNs", ".", "section", ":", "Results", "Experimental", "settings", ".", "We", "closely", "followed", "the", "experimental", "setup", "of", "[", "reference", "]", ",", "using", "five", "standard", "datasets", ":", "Synthetic", "dataset", "from", "[", "reference", "]", ",", "MovieLens", "[", "reference", "]", ",", "Flixster", "[", "reference", "]", ",", "Douban", "[", "reference", "]", ",", "and", "YahooMusic", "[", "reference", "]", ".", "Classical", "Matrix", "Completion", "(", "MC", ")", "[", "reference", "]", ",", "Inductive", "Matrix", "Completion", "(", "IMC", ")", "[", "reference", "][", "reference", "]", ",", "Geometric", "Matrix", "Completion", "(", "GMC", ")", "[", "reference", "]", ",", "and", "Graph", "Regularized", "Alternating", "Least", "Squares", "(", "GRALS", ")", "[", "reference", "]", "were", "used", "as", "baseline", "methods", ".", "In", "all", "the", "experiments", ",", "we", "used", "the", "following", "settings", "for", "our", "RGCNNs", ":", "Chebyshev", "polynomials", "of", "order", "p", "=", "5", ",", "outputting", "k", "=", "32", "-", "dimensional", "features", ",", "LSTM", "cells", "with", "32", "features", "and", "T", "=", "10", "diffusion", "steps", ".", "All", "the", "models", "were", "implemented", "in", "Google", "TensorFlow", "and", "trained", "using", "the", "Adam", "stochastic", "optimization", "algorithm", "[", "reference", "]", "with", "learning", "rate", "10", "−3", ".", "In", "factorized", "models", ",", "rank", "r", "=", "15", "and", "10", "was", "used", "for", "the", "synthetic", "and", "real", "datasets", ",", "respectively", ".", "For", "all", "methods", ",", "hyperparameters", "were", "chosen", "by", "cross", "-", "validation", ".", "section", ":", "Synthetic", "data", "We", "start", "our", "experimental", "evaluation", "showing", "the", "performance", "of", "our", "approach", "on", "a", "small", "synthetic", "dataset", ",", "in", "which", "the", "user", "and", "item", "graphs", "have", "strong", "communities", "structure", ".", "Though", "rather", "simple", ",", "such", "a", "dataset", "allows", "to", "study", "the", "behavior", "of", "different", "algorithms", "in", "controlled", "settings", ".", "The", "performance", "of", "different", "matrix", "completion", "methods", "is", "reported", "in", "Table", "1", ",", "along", "with", "their", "theoretical", "complexity", ".", "Our", "RGCNN", "model", "achieves", "the", "best", "accuracy", ",", "followed", "by", "the", "separable", "RGCNN", ".", "Different", "diffusion", "time", "steps", "of", "these", "two", "models", "are", "visualized", "in", "Figure", "4", ".", "Figure", "5", "shows", "the", "convergence", "rates", "of", "different", "methods", ".", "Figures", "6", "and", "7", "depict", "the", "spectral", "filters", "learnt", "by", "the", "MGCNN", "and", "row", "-", "and", "column", "-", "GCNNs", ".", "We", "repeated", "the", "same", "experiment", "considering", "only", "the", "column", "(", "users", ")", "graph", "to", "be", "given", ".", "In", "this", "setting", ",", "the", "RGCNN", "can", "not", "be", "applied", ",", "while", "the", "sRGCNN", "has", "only", "one", "GCNN", "applied", "on", "the", "factor", "H", ",", "and", "the", "other", "factor", "W", "is", "free", ".", "Figure", "4", ".", "Evolution", "of", "the", "matrix", "X", "(", "t", ")", "with", "our", "architecture", "using", "full", "matrix", "completion", "model", "RGCNN", "(", "top", ")", "and", "factorized", "matrix", "completion", "model", "sRGCNN", "(", "bottom", ")", ".", "Numbers", "indicate", "the", "RMS", "error", ".", "Figure", "5", ".", "Convergence", "rates", "of", "the", "tested", "algorithms", "over", "the", "Synthetic", "Netflix", "dataset", ".", "section", ":", "Real", "data", "Following", "[", "reference", "]", ",", "we", "evaluated", "the", "proposed", "approach", "on", "the", "MovieLens", ",", "Flixster", ",", "Douban", "and", "YahooMusic", "datasets", ".", "For", "the", "MovieLens", "dataset", "we", "constructed", "the", "user", "and", "item", "(", "movie", ")", "graphs", "as", "unweighted", "10", "-", "nearest", "neighbor", "graphs", "in", "the", "space", "of", "user", "and", "movie", "features", ",", "respectively", ".", "For", "Flixster", ",", "the", "user", "and", "item", "graphs", "were", "constructed", "from", "the", "scores", "of", "the", "original", "matrix", ".", "On", "this", "dataset", ",", "we", "also", "performed", "an", "experiment", "using", "only", "the", "users", "graph", ".", "For", "the", "Douban", "dataset", ",", "we", "used", "only", "the", "user", "graph", "(", "the", "provided", "social", "network", "of", "the", "user", ")", ".", "For", "the", "YahooMusic", "dataset", ",", "we", "used", "only", "the", "item", "graph", ",", "constructed", "with", "unweighted", "10", "-", "nearest", "neighbors", "in", "the", "space", "of", "item", "features", "(", "artists", ",", "albums", ",", "and", "genres", ")", ".", "For", "the", "latter", "three", "datasets", ",", "we", "used", "a", "sub", "-", "matrix", "of", "3000", "×", "3000", "entries", "for", "evaluating", "the", "performance", ".", "Tables", "3", "and", "4", "summarize", "the", "performance", "of", "different", "methods", ".", "RGCNN", "outperforms", "the", "competitors", "in", "all", "the", "experiments", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "presented", "a", "new", "deep", "learning", "approach", "for", "matrix", "completion", "based", "on", "a", "specially", "designed", "multigraph", "convolutional", "neural", "network", "architecture", ".", "Among", "the", "key", "advantages", "of", "our", "approach", "compared", "to", "traditional", "methods", "is", "its", "low", "computational", "complexity", "and", "constant", "number", "of", "degrees", "of", "freedom", "independent", "of", "the", "matrix", "size", ".", "We", "showed", "that", "the", "use", "of", "deep", "learning", "for", "matrix", "completion", "allows", "to", "beat", "current", "state", "-", "of", "-", "the", "-", "art", "recommender", "system", "methods", ".", "To", "our", "knowledge", ",", "our", "work", "is", "the", "first", "application", "of", "deep", "learning", "on", "graphs", "to", "this", "class", "of", "problems", ".", "We", "believe", "that", "it", "shows", "the", "potential", "of", "the", "nascent", "field", "of", "geometric", "deep", "learning", "on", "non", "-", "Euclidean", "domains", ",", "and", "will", "encourage", "future", "works", "in", "this", "direction", ".", "section", ":", "section", ":", "Acknowledgments", "section", ":"], "entity": [[0, 3, ["Geometric", "Matrix", "Completion"], "Task"], [4, 10, ["Recurrent", "Multi", "-", "Graph", "Neural", "Networks"], "Method"], [13, 16, ["Matrix", "completion", "models"], "Method"], [23, 25, ["recommender", "systems"], "Task"], [107, 112, ["geometric", "deep", "learning", "on", "graphs"], "Method"], [114, 117, ["matrix", "completion", "architecture"], "Method"], [118, 122, ["graph", "convolutional", "neural", "networks"], "Method"], [123, 126, ["recurrent", "neural", "networks"], "Method"], [136, 141, ["non", "-", "linear", "diffusion", "process"], "Method"], [148, 151, ["neural", "network", "system"], "Method"], [190, 192, ["Recommender", "systems"], "Method"], [199, 201, ["intelligent", "systems"], "Task"], [234, 236, ["recommender", "systems"], "Method"], [258, 260, ["Hybrid", "systems"], "Method"], [261, 265, ["collaborative", "and", "content", "techniques"], "Method"], [266, 268, ["Matrix", "completion"], "Task"], [272, 274, ["recommendation", "method"], "Method"], [279, 282, ["matrix", "completion", "problem"], "Task"], [373, 374, ["Netflix"], "Material"], [404, 407, ["matrix", "completion", "problems"], "Task"], [457, 459, ["recommender", "systems"], "Method"], [470, 474, ["signal", "processing", "on", "graphs"], "Task"], [480, 483, ["harmonic", "analysis", "methods"], "Method"], [489, 492, ["Geometric", "deep", "learning"], "Method"], [500, 502, ["recommender", "systems"], "Task"], [503, 506, ["deep", "learning", "approaches"], "Method"], [512, 515, ["deep", "neural", "networks"], "Method"], [520, 523, ["convolutional", "neural", "networks"], "Method"], [524, 525, ["CNNs"], "Method"], [537, 542, ["computer", "vision", "-", "related", "applications"], "Task"], [546, 548, ["CNN", "models"], "Method"], [555, 557, ["recommendation", "problem"], "Task"], [599, 602, ["geometric", "deep", "learning"], "Method"], [624, 628, ["computer", "graphics", "and", "vision"], "Task"], [648, 650, ["neural", "networks"], "Method"], [676, 682, ["CNN", "-", "like", "deep", "neural", "architectures"], "Method"], [694, 697, ["classical", "Fourier", "transforms"], "Method"], [721, 723, ["filtering", "scheme"], "Method"], [724, 727, ["recurrent", "Chebyshev", "polynomials"], "Method"], [731, 732, ["complexity"], "Metric"], [733, 736, ["CNNs", "on", "graphs"], "Method"], [739, 740, ["complexity"], "Metric"], [742, 743, ["CNNs"], "Method"], [770, 773, ["matrix", "completion", "problem"], "Task"], [774, 776, ["deep", "learning"], "Task"], [786, 789, ["neural", "network", "architecture"], "Method"], [814, 822, ["non", "-", "linear", "temporal", "diffusion", "mechanism", "of", "ratings"], "Method"], [831, 833, ["CNN", "architecture"], "Method"], [841, 847, ["temporal", "dynamics", "of", "the", "rating", "diffusion"], "Task"], [851, 856, ["Long", "-", "Short", "Term", "Memory"], "Method"], [857, 858, ["LSTM"], "Method"], [859, 862, ["recurrent", "neural", "network"], "Method"], [863, 864, ["RNN"], "Method"], [880, 885, ["graph", "-", "based", "deep", "learning"], "Method"], [886, 889, ["matrix", "completion", "problem"], "Task"], [904, 907, ["matrix", "completion", "models"], "Method"], [946, 948, ["Matrix", "Completion"], "Task"], [948, 951, ["Matrix", "completion", "problem"], "Task"], [970, 972, ["posed", "problem"], "Task"], [983, 987, ["well", "-", "posed", "problem"], "Task"], [1067, 1070, ["Hadamard", "pointwise", "product"], "Method"], [1073, 1075, ["rank", "minimization"], "Task"], [1078, 1083, ["NP", "-", "hard", "combinatorial", "problem"], "Task"], [1129, 1134, ["1", "relaxation", "of", "the", "SVD"], "Method"], [1148, 1151, ["Geometric", "matrix", "completion"], "Method"], [1202, 1206, ["undirected", "weighted", "column", "graph"], "Method"], [1254, 1256, ["row", "graph"], "Method"], [1403, 1407, ["geometric", "matrix", "completion", "problem"], "Task"], [1413, 1416, ["users", "Factorized", "models"], "Method"], [1417, 1420, ["Matrix", "completion", "algorithms"], "Method"], [1430, 1433, ["convex", "optimization", "problems"], "Task"], [1454, 1456, ["compressed", "sensing"], "Task"], [1459, 1465, ["non", "-", "differential", "nuclear", "norm", "problem"], "Task"], [1505, 1507, ["factorized", "representation"], "Method"], [1585, 1588, ["matrix", "completion", "problem"], "Task"], [1619, 1622, ["matrix", "completion", "problem"], "Task"], [1624, 1628, ["nuclear", "norm", "minimization", "problem"], "Task"], [1638, 1640, ["factorized", "form"], "Method"], [1647, 1649, ["factorized", "formulation"], "Method"], [1651, 1656, ["graph", "-", "based", "minimization", "problem"], "Task"], [1671, 1673, ["regularization", "process"], "Method"], [1694, 1696, ["quadratic", "complexity"], "Metric"], [1699, 1703, ["Deep", "learning", "on", "graphs"], "Task"], [1710, 1713, ["geometric", "deep", "learning"], "Method"], [1719, 1720, ["CNNs"], "Method"], [1728, 1730, ["spectral", "eigendecomposition"], "Method"], [1789, 1792, ["classical", "harmonic", "analysis"], "Method"], [1825, 1828, ["graph", "Fourier", "transform"], "Method"], [1855, 1857, ["Fourier", "transforms"], "Method"], [1865, 1869, ["spectral", "definition", "of", "convolution"], "Method"], [1874, 1877, ["CNNs", "on", "graphs"], "Method"], [1879, 1882, ["spectral", "convolutional", "layer"], "Method"], [1921, 1926, ["diagonal", "matrix", "of", "spectral", "multipliers"], "Method"], [1928, 1930, ["learnable", "filter"], "Method"], [1942, 1943, ["ReLU"], "Method"], [1955, 1956, ["convolutions"], "Method"], [1961, 1963, ["spectral", "domain"], "Method"], [1964, 1965, ["FFT"], "Method"], [1970, 1974, ["forward", "and", "inverse", "graph"], "Method"], [1974, 1976, ["Fourier", "transform"], "Method"], [1995, 1997, ["FFTlike", "algorithms"], "Method"], [2028, 2030, ["classical", "CNNs"], "Method"], [2042, 2044, ["explicit", "expansion"], "Method"], [2046, 2049, ["Chebyshev", "polynomial", "basis"], "Method"], [2052, 2054, ["spectral", "filters"], "Method"], [2054, 2058, ["row", "+", "column", "filtering"], "Method"], [2061, 2063, ["Recurrent", "GCNN"], "Method"], [2064, 2065, ["RGCNN"], "Method"], [2069, 2073, ["full", "matrix", "completion", "model"], "Method"], [2089, 2093, ["Multi", "-", "Graph", "CNN"], "Method"], [2094, 2095, ["MGCNN"], "Method"], [2125, 2127, ["learning", "complexity"], "Metric"], [2159, 2160, ["filter"], "Method"], [2233, 2239, ["recursive", "definition", "of", "the", "Chebyshev", "polynomials"], "Method"], [2244, 2245, ["filter"], "Method"], [2248, 2251, ["Laplacian", "p", "times"], "Method"], [2252, 2253, ["Multiplication"], "Method"], [2282, 2287, ["k", "-", "nearest", "neighbors", "graphs"], "Task"], [2289, 2293, ["real", "-", "world", "networks"], "Task"], [2297, 2298, ["complexity"], "Metric"], [2319, 2320, ["CNNs"], "Method"], [2358, 2359, ["filters"], "Method"], [2374, 2376, ["matrix", "completion"], "Method"], [2378, 2381, ["learnable", "diffusion", "process"], "Method"], [2388, 2391, ["deep", "learning", "architecture"], "Method"], [2398, 2400, ["spatial", "part"], "Method"], [2415, 2420, ["full", "and", "factorized", "matrix", "models"], "Method"], [2424, 2426, ["temporal", "part"], "Method"], [2479, 2481, ["Chebyshev", "polynomials"], "Method"], [2483, 2485, ["orthogonal", "basis"], "Method"], [2504, 2506, ["spectral", "filters"], "Method"], [2523, 2527, ["Multi", "-", "Graph", "CNNs"], "Method"], [2527, 2531, ["Multi", "-", "graph", "convolution"], "Method"], [2543, 2546, ["graph", "Fourier", "transform"], "Method"], [2567, 2577, ["two", "-", "dimensional", "Fourier", "transform", "of", "an", "image", "(", "matrix"], "Method"], [2585, 2590, ["one", "-", "dimensional", "Fourier", "transform"], "Method"], [2604, 2607, ["twodimensional", "Fourier", "transform"], "Method"], [2654, 2656, ["multigraph", "version"], "Method"], [2658, 2660, ["spectral", "convolution"], "Method"], [2671, 2673, ["spectral", "multipliersŶ"], "Method"], [2750, 2754, ["Multi", "-", "Graph", "CNN"], "Method"], [2755, 2756, ["MGCNN"], "Method"], [2767, 2769, ["convolutional", "layer"], "Method"], [2799, 2801, ["Separable", "convolution"], "Method"], [2806, 2808, ["multigraph", "convolution"], "Task"], [2822, 2824, ["onedimensional", "convolution"], "Method"], [2847, 2849, ["Matrix", "diffusion"], "Task"], [2850, 2851, ["RNN"], "Method"], [2868, 2869, ["MGCNN"], "Method"], [2883, 2886, ["Recurrent", "Neural", "Network"], "Method"], [2887, 2888, ["RNN"], "Method"], [2891, 2894, ["score", "diffusion", "process"], "Method"], [2929, 2930, ["LSTM"], "Method"], [2945, 2948, ["vanishing", "gradient", "issue"], "Task"], [2965, 2966, ["MGCNN"], "Method"], [2973, 2977, ["projection", "or", "dimensionality", "reduction"], "Method"], [3001, 3002, ["LSTM"], "Method"], [3058, 3061, ["matrix", "completion", "model"], "Method"], [3094, 3098, ["full", "matrix", "completion", "setting"], "Task"], [3099, 3102, ["Recurrent", "Graph", "CNN"], "Method"], [3103, 3104, ["RGCNN"], "Method"], [3107, 3109, ["factorized", "version"], "Method"], [3111, 3112, ["GCNNs"], "Method"], [3113, 3114, ["RNN"], "Method"], [3118, 3122, ["separable", "Recurrent", "Graph", "CNN"], "Method"], [3123, 3124, ["sRGCNN"], "Method"], [3127, 3128, ["complexity"], "Metric"], [3143, 3144, ["MGCNN"], "Method"], [3164, 3165, ["GCNNs"], "Method"], [3180, 3184, ["Full", "matrix", "completion", "model"], "Method"], [3185, 3186, ["RGCNN"], "Method"], [3211, 3215, ["Multi", "-", "Graph", "CNN"], "Method"], [3260, 3261, ["RNN"], "Method"], [3262, 3264, ["feature", "vectorx"], "Task"], [3284, 3288, ["Factorized", "matrix", "completion", "model"], "Method"], [3289, 3290, ["sRGCNN"], "Method"], [3310, 3312, ["Graph", "CNN"], "Method"], [3330, 3331, ["RNN"], "Method"], [3332, 3334, ["feature", "vectorh"], "Task"], [3345, 3347, ["Graph", "CNN"], "Method"], [3365, 3366, ["RNN"], "Method"], [3367, 3369, ["feature", "vectorw"], "Task"], [3393, 3394, ["loss"], "Metric"], [3407, 3408, ["RNN"], "Method"], [3433, 3434, ["MGCNN"], "Method"], [3444, 3445, ["LSTM"], "Method"], [3448, 3449, ["σ"], "Method"], [3453, 3455, ["factorized", "setting"], "Task"], [3472, 3473, ["GCNNs"], "Method"], [3503, 3504, ["MovieLens"], "Material"], [3508, 3509, ["Flixster"], "Material"], [3513, 3514, ["Douban"], "Material"], [3519, 3520, ["YahooMusic"], "Material"], [3525, 3527, ["Matrix", "Completion"], "Method"], [3528, 3529, ["MC"], "Method"], [3534, 3537, ["Inductive", "Matrix", "Completion"], "Method"], [3538, 3539, ["IMC"], "Method"], [3546, 3549, ["Geometric", "Matrix", "Completion"], "Method"], [3550, 3552, ["GMC", ")"], "Method"], [3557, 3562, ["Graph", "Regularized", "Alternating", "Least", "Squares"], "Method"], [3563, 3564, ["GRALS"], "Method"], [3586, 3587, ["RGCNNs"], "Method"], [3588, 3590, ["Chebyshev", "polynomials"], "Method"], [3622, 3624, ["Google", "TensorFlow"], "Method"], [3628, 3632, ["Adam", "stochastic", "optimization", "algorithm"], "Method"], [3636, 3638, ["learning", "rate"], "Metric"], [3642, 3644, ["factorized", "models"], "Method"], [3670, 3673, ["cross", "-", "validation"], "Method"], [3730, 3733, ["matrix", "completion", "methods"], "Method"], [3742, 3744, ["theoretical", "complexity"], "Metric"], [3746, 3748, ["RGCNN", "model"], "Method"], [3751, 3752, ["accuracy"], "Metric"], [3756, 3758, ["separable", "RGCNN"], "Method"], [3777, 3779, ["convergence", "rates"], "Metric"], [3789, 3791, ["spectral", "filters"], "Method"], [3794, 3795, ["MGCNN"], "Method"], [3796, 3798, ["row", "-"], "Method"], [3799, 3802, ["column", "-", "GCNNs"], "Method"], [3825, 3826, ["RGCNN"], "Method"], [3833, 3834, ["sRGCNN"], "Method"], [3876, 3881, ["factorized", "matrix", "completion", "model", "sRGCNN"], "Method"], [3888, 3890, ["RMS", "error"], "Metric"], [3894, 3896, ["Convergence", "rates"], "Metric"], [3902, 3905, ["Synthetic", "Netflix", "dataset"], "Material"], [3922, 3923, ["MovieLens"], "Material"], [3924, 3925, ["Flixster"], "Material"], [3926, 3927, ["Douban"], "Material"], [3928, 3930, ["YahooMusic", "datasets"], "Material"], [3964, 3965, ["Flixster"], "Material"], [4084, 4085, ["RGCNN"], "Method"], [4104, 4107, ["deep", "learning", "approach"], "Method"], [4108, 4110, ["matrix", "completion"], "Task"], [4115, 4120, ["multigraph", "convolutional", "neural", "network", "architecture"], "Method"], [4135, 4137, ["computational", "complexity"], "Metric"], [4156, 4158, ["deep", "learning"], "Method"], [4159, 4161, ["matrix", "completion"], "Task"], [4172, 4175, ["recommender", "system", "methods"], "Method"], [4187, 4191, ["deep", "learning", "on", "graphs"], "Task"], [4209, 4212, ["geometric", "deep", "learning"], "Task"], [4213, 4217, ["non", "-", "Euclidean", "domains"], "Task"], [2429, 2430, ["LSTM"], "Method"], [2555, 2557, ["row", "-"], "Method"], [2641, 2643, ["row", "-"], "Method"], [2874, 2876, ["row", "-"], "Method"], [2879, 2880, ["GCNNs"], "Method"], [2899, 2904, ["Long", "-", "Short", "Term", "Memory"], "Method"], [2905, 2906, ["LSTM"], "Method"], [2907, 2908, ["RNN"], "Method"], [2954, 2955, ["LSTM"], "Method"], [3089, 3090, ["MGCNN"], "Method"], [3091, 3092, ["RNN"], "Method"], [3604, 3605, ["LSTM"], "Method"], [3871, 3872, ["RGCNN"], "Method"], [3933, 3934, ["MovieLens"], "Material"], [3998, 3999, ["Douban"], "Material"], [4019, 4020, ["YahooMusic"], "Material"]]} {"doc_id": "1efdad6f91e830fd64306e4625f74191b05ef9c4", "text": ["Numerous", "important", "problems", "can", "be", "framed", "as", "learning", "from", "graph", "data", ".", "We", "propose", "a", "framework", "for", "learning", "convolutional", "neural", "networks", "for", "arbitrary", "graphs", ".", "These", "graphs", "may", "be", "undirected", ",", "directed", ",", "and", "with", "both", "discrete", "and", "continuous", "node", "and", "edge", "attributes", ".", "Analogous", "to", "image", "-", "based", "convolutional", "networks", "that", "operate", "on", "locally", "connected", "regions", "of", "the", "input", ",", "we", "present", "a", "general", "approach", "to", "extracting", "locally", "connected", "regions", "from", "graphs", ".", "Using", "established", "benchmark", "data", "sets", ",", "we", "demonstrate", "that", "the", "learned", "feature", "representations", "are", "competitive", "with", "state", "of", "the", "art", "graph", "kernels", "and", "that", "their", "computation", "is", "highly", "efficient", ".", "LearningConvolutionalNeuralNetworksforGraphs", "section", ":", "Introduction", "With", "this", "paper", "we", "aim", "to", "bring", "convolutional", "neural", "networks", "to", "bear", "on", "a", "large", "class", "of", "graph", "-", "based", "learning", "problems", ".", "We", "consider", "the", "following", "two", "problems", ".", "Given", "a", "collection", "of", "graphs", ",", "learn", "a", "function", "that", "can", "be", "used", "for", "classification", "and", "regression", "problems", "on", "unseen", "graphs", ".", "The", "nodes", "of", "any", "two", "graphs", "are", "not", "necessarily", "in", "correspondence", ".", "For", "instance", ",", "each", "graph", "of", "the", "collection", "could", "model", "a", "chemical", "compound", "and", "the", "output", "could", "be", "a", "function", "mapping", "unseen", "compounds", "to", "their", "level", "of", "activity", "against", "cancer", "cells", ".", "Given", "a", "large", "graph", ",", "learn", "graph", "representations", "that", "can", "be", "used", "to", "infer", "unseen", "graph", "properties", "such", "as", "node", "types", "and", "missing", "edges", ".", "We", "propose", "a", "framework", "for", "learning", "representations", "for", "classes", "of", "directed", "and", "undirected", "graphs", ".", "The", "graphs", "may", "have", "nodes", "and", "edges", "with", "multiple", "discrete", "and", "continuous", "attributes", "and", "may", "have", "multiple", "types", "of", "edges", ".", "Similar", "to", "convolutional", "neural", "network", "for", "images", ",", "we", "construct", "locally", "connected", "neighborhoods", "from", "the", "input", "graphs", ".", "These", "neighborhoods", "are", "generated", "efficiently", "and", "serve", "as", "the", "receptive", "fields", "of", "a", "convolutional", "architecture", ",", "allowing", "the", "framework", "to", "learn", "effective", "graph", "representations", ".", "The", "proposed", "approach", "builds", "on", "concepts", "from", "convolutional", "neural", "networks", "(", "CNNs", ")", "for", "images", "and", "extends", "them", "to", "arbitrary", "graphs", ".", "Figure", "[", "reference", "]", "illustrates", "the", "locally", "connected", "receptive", "fields", "of", "a", "CNN", "for", "images", ".", "An", "image", "can", "be", "represented", "as", "a", "square", "grid", "graph", "whose", "nodes", "represent", "pixels", ".", "Now", ",", "a", "CNN", "can", "be", "seen", "as", "traversing", "a", "node", "sequence", "(", "nodes", "-", "in", "Figure", "[", "reference", "]", "(", "a", ")", ")", "and", "generating", "fixed", "-", "size", "neighborhood", "graphs", "(", "the", "x", "grids", "in", "Figure", "[", "reference", "]", "(", "b", ")", ")", "for", "each", "of", "the", "nodes", ".", "The", "neighborhood", "graphs", "serve", "as", "the", "receptive", "fields", "to", "read", "feature", "values", "from", "the", "pixel", "nodes", ".", "Due", "to", "the", "implicit", "spatial", "order", "of", "the", "pixels", ",", "the", "sequence", "of", "nodes", "for", "which", "neighborhood", "graphs", "are", "created", ",", "from", "left", "to", "right", "and", "top", "to", "bottom", ",", "is", "uniquely", "determined", ".", "The", "same", "holds", "for", "NLP", "problems", "where", "each", "sentence", "(", "and", "its", "parse", "-", "tree", ")", "determines", "a", "sequence", "of", "words", ".", "However", ",", "for", "numerous", "graph", "collections", "a", "problem", "-", "specific", "ordering", "(", "spatial", ",", "temporal", ",", "or", "otherwise", ")", "is", "missing", "and", "the", "nodes", "of", "the", "graphs", "are", "not", "in", "correspondence", ".", "In", "these", "instances", ",", "one", "has", "to", "solve", "two", "problems", ":", "(", "i", ")", "Determining", "the", "node", "sequences", "for", "which", "neighborhood", "graphs", "are", "created", "and", "(", "ii", ")", "computing", "a", "normalization", "of", "neighborhood", "graphs", ",", "that", "is", ",", "a", "unique", "mapping", "from", "a", "graph", "representation", "into", "a", "vector", "space", "representation", ".", "The", "proposed", "approach", ",", "termed", "Patchy", "-", "san", ",", "addresses", "these", "two", "problems", "for", "arbitrary", "graphs", ".", "For", "each", "input", "graph", ",", "it", "first", "determines", "nodes", "(", "and", "their", "order", ")", "for", "which", "neighborhood", "graphs", "are", "created", ".", "For", "each", "of", "these", "nodes", ",", "a", "neighborhood", "consisting", "of", "exactly", "nodes", "is", "extracted", "and", "normalized", ",", "that", "is", ",", "it", "is", "uniquely", "mapped", "to", "a", "space", "with", "a", "fixed", "linear", "order", ".", "The", "normalized", "neighborhood", "serves", "as", "the", "receptive", "field", "for", "a", "node", "under", "consideration", ".", "Finally", ",", "feature", "learning", "components", "such", "as", "convolutional", "and", "dense", "layers", "are", "combined", "with", "the", "normalized", "neighborhood", "graphs", "as", "the", "CNN", "’s", "receptive", "fields", ".", "Figure", "[", "reference", "]", "illustrates", "the", "Patchy", "-", "san", "architecture", "which", "has", "several", "advantages", "over", "existing", "approaches", ":", "First", ",", "it", "is", "highly", "efficient", ",", "naively", "parallelizable", ",", "and", "applicable", "to", "large", "graphs", ".", "Second", ",", "for", "a", "number", "of", "applications", ",", "ranging", "from", "computational", "biology", "to", "social", "network", "analysis", ",", "it", "is", "important", "to", "visualize", "learned", "network", "motifs", ".", "Patchy", "-", "san", "supports", "feature", "visualizations", "providing", "insights", "into", "the", "structural", "properties", "of", "graphs", ".", "Third", ",", "instead", "of", "crafting", "yet", "another", "graph", "kernel", ",", "Patchy", "-", "san", "learns", "application", "dependent", "features", "without", "the", "need", "to", "feature", "engineering", ".", "Our", "theoretical", "contributions", "are", "the", "definition", "of", "the", "normalization", "problem", "on", "graphs", "and", "its", "complexity", ";", "a", "method", "for", "comparing", "graph", "labeling", "approaches", "for", "a", "collection", "of", "graphs", ";", "and", "a", "result", "that", "shows", "that", "Patchy", "-", "san", "generalizes", "CNNs", "on", "images", ".", "Using", "standard", "benchmark", "data", "sets", ",", "we", "demonstrate", "that", "the", "learned", "CNNs", "for", "graphs", "are", "both", "efficient", "and", "effective", "compared", "to", "state", "of", "the", "art", "graph", "kernels", ".", "section", ":", "Related", "Work", "Graph", "kernels", "allow", "kernel", "-", "based", "learning", "approaches", "such", "as", "SVMs", "to", "work", "directly", "on", "graphs", ".", "Kernels", "on", "graphs", "were", "originally", "defined", "as", "similarity", "functions", "on", "the", "nodes", "of", "a", "single", "graph", ".", "Two", "representative", "classes", "of", "kernels", "are", "the", "skew", "spectrum", "kernel", "and", "kernels", "based", "on", "graphlets", ".", "The", "latter", "is", "related", "to", "our", "work", ",", "as", "it", "builds", "kernels", "based", "on", "fixed", "-", "sized", "subgraphs", ".", "These", "subgraphs", ",", "which", "are", "often", "called", "motifs", "or", "graphlets", ",", "reflect", "functional", "network", "properties", ".", "However", ",", "due", "to", "the", "combinatorial", "complexity", "of", "subgraph", "enumeration", ",", "graphlet", "kernels", "are", "restricted", "to", "subgraphs", "with", "few", "nodes", ".", "An", "effective", "class", "of", "graph", "kernels", "are", "the", "Weisfeiler", "-", "Lehman", "(", "WL", ")", "kernels", ".", "WL", "kernels", ",", "however", ",", "only", "support", "discrete", "features", "and", "use", "memory", "linear", "in", "the", "number", "of", "training", "examples", "at", "test", "time", ".", "Patchy", "-", "san", "uses", "WL", "as", "one", "possible", "labeling", "procedure", "to", "compute", "receptive", "fields", ".", "Deep", "graph", "kernels", "and", "graph", "invariant", "kernels", "compare", "graphs", "based", "on", "the", "existence", "or", "count", "of", "small", "substructures", "such", "as", "shortest", "paths", ",", "graphlets", ",", "subtrees", ",", "and", "other", "graph", "invariants", ".", "In", "contrast", ",", "Patchy", "-", "san", "learns", "substructures", "from", "graph", "data", "and", "is", "not", "limited", "to", "a", "predefined", "set", "of", "motifs", ".", "Moreover", ",", "while", "all", "graph", "kernels", "have", "a", "training", "complexity", "at", "least", "quadratic", "in", "the", "number", "of", "graphs", ",", "which", "is", "prohibitive", "for", "large", "-", "scale", "problems", ",", "Patchy", "-", "san", "scales", "linearly", "with", "the", "number", "of", "graphs", ".", "Graph", "neural", "networks", "(", "GNNs", ")", "are", "a", "recurrent", "neural", "network", "architecture", "defined", "on", "graphs", ".", "GNNs", "apply", "recurrent", "neural", "networks", "for", "walks", "on", "the", "graph", "structure", ",", "propagating", "node", "representations", "until", "a", "fixed", "point", "is", "reached", ".", "The", "resulting", "node", "representations", "are", "then", "used", "as", "features", "in", "classification", "and", "regression", "problems", ".", "GNNs", "support", "only", "discrete", "labels", "and", "perform", "as", "many", "backpropagation", "operations", "as", "there", "are", "edges", "and", "nodes", "in", "the", "graph", "per", "learning", "iteration", ".", "Gated", "Graph", "Sequence", "Neural", "Networks", "modify", "GNNs", "to", "use", "gated", "recurrent", "units", "and", "to", "output", "sequences", ".", "Recent", "work", "extended", "CNNs", "to", "topologies", "that", "differ", "from", "the", "low", "-", "dimensional", "grid", "structure", ".", "All", "of", "these", "methods", ",", "however", ",", "assume", "one", "global", "graph", "structure", ",", "that", "is", ",", "a", "correspondence", "of", "the", "vertices", "across", "input", "examples", ".", "perform", "convolutional", "type", "operations", "on", "graphs", ",", "developing", "a", "differentiable", "variant", "of", "one", "specific", "graph", "feature", ".", "section", ":", "Background", "We", "provide", "a", "brief", "introduction", "to", "the", "required", "background", "in", "convolutional", "networks", "and", "graph", "theory", ".", "subsection", ":", "Convolutional", "Neural", "Networks", "CNNs", "were", "inspired", "by", "earlier", "work", "that", "showed", "that", "the", "visual", "cortex", "in", "animals", "contains", "complex", "arrangements", "of", "cells", ",", "responsible", "for", "detecting", "light", "in", "small", "local", "regions", "of", "the", "visual", "field", ".", "CNNs", "were", "developed", "in", "the", "s", "and", "have", "been", "applied", "to", "image", ",", "speech", ",", "text", ",", "and", "drug", "discovery", "problems", ".", "A", "predecessor", "to", "CNNs", "was", "the", "Neocognitron", ".", "A", "typical", "CNN", "is", "composed", "of", "convolutional", "and", "dense", "layers", ".", "The", "purpose", "of", "the", "first", "convolutional", "layer", "is", "the", "extraction", "of", "common", "patterns", "found", "within", "local", "regions", "of", "the", "input", "images", ".", "CNNs", "convolve", "learned", "filters", "over", "the", "input", "image", ",", "computing", "the", "inner", "product", "at", "every", "image", "location", "in", "the", "image", "and", "outputting", "the", "result", "as", "tensors", "whose", "depth", "is", "the", "number", "of", "filters", ".", "subsection", ":", "Graphs", "A", "graph", "is", "a", "pair", "with", "the", "set", "of", "vertices", "and", "the", "set", "of", "edges", ".", "Let", "be", "the", "number", "of", "vertices", "and", "the", "number", "of", "edges", ".", "Each", "graph", "can", "be", "represented", "by", "an", "adjacency", "matrix", "of", "size", ",", "where", "if", "there", "is", "an", "edge", "from", "vertex", "to", "vertex", ",", "and", "otherwise", ".", "In", "this", "case", ",", "we", "say", "that", "vertex", "has", "position", "in", ".", "Moreover", ",", "if", "we", "say", "and", "are", "adjacent", ".", "Node", "and", "edge", "attributes", "are", "features", "that", "attain", "one", "value", "for", "each", "node", "and", "edge", "of", "a", "graph", ".", "We", "use", "the", "term", "attribute", "value", "instead", "of", "label", "to", "avoid", "confusion", "with", "the", "graph", "-", "theoretical", "concept", "of", "a", "labeling", ".", "A", "walk", "is", "a", "sequence", "of", "nodes", "in", "a", "graph", ",", "in", "which", "consecutive", "nodes", "are", "connected", "by", "an", "edge", ".", "A", "path", "is", "a", "walk", "with", "distinct", "nodes", ".", "We", "write", "to", "denote", "the", "distance", "between", "and", ",", "that", "is", ",", "the", "length", "of", "the", "shortest", "path", "between", "and", ".", "is", "the", "-", "neighborhood", "of", "a", "node", ",", "that", "is", ",", "all", "nodes", "that", "are", "adjacent", "to", ".", "Labeling", "and", "Node", "Partitions", ".", "Patchy", "-", "san", "utilizes", "graph", "labelings", "to", "impose", "an", "order", "on", "nodes", ".", "A", "graph", "labeling", "is", "a", "function", "from", "the", "set", "of", "vertices", "to", "an", "ordered", "set", "such", "as", "the", "real", "numbers", "and", "integers", ".", "A", "graph", "labeling", "procedure", "computes", "a", "graph", "labeling", "for", "an", "input", "graph", ".", "When", "it", "is", "clear", "from", "the", "context", ",", "we", "use", "labeling", "to", "refer", "to", "both", ",", "the", "graph", "labeling", "and", "the", "procedure", "to", "compute", "it", ".", "A", "ranking", "(", "or", "coloring", ")", "is", "a", "function", ".", "Every", "labeling", "induces", "a", "ranking", "with", "if", "and", "only", "if", ".", "If", "the", "labeling", "of", "graph", "is", "injective", ",", "it", "determines", "a", "total", "order", "of", "’s", "vertices", "and", "a", "unique", "adjacency", "matrix", "of", "where", "vertex", "has", "position", "in", ".", "Moreover", ",", "every", "graph", "labeling", "induces", "a", "partition", "on", "with", "if", "and", "only", "if", ".", "Examples", "of", "graph", "labeling", "procedures", "are", "node", "degree", "and", "other", "measures", "of", "centrality", "commonly", "used", "in", "the", "analysis", "of", "networks", ".", "For", "instance", ",", "the", "betweeness", "centrality", "of", "a", "vertex", "computes", "the", "fractions", "of", "shortest", "paths", "that", "pass", "through", ".", "The", "Weisfeiler", "-", "Lehman", "algorithm", "is", "a", "procedure", "for", "partitioning", "the", "vertices", "of", "a", "graph", ".", "It", "is", "also", "known", "as", "color", "refinement", "and", "naive", "vertex", "classification", ".", "Color", "refinement", "has", "attracted", "considerable", "interest", "in", "the", "ML", "community", "since", "it", "can", "be", "applied", "to", "speed", "-", "up", "inference", "in", "graphical", "models", "and", "as", "a", "method", "to", "compute", "graph", "kernels", ".", "Patchy", "-", "san", "applies", "these", "labeling", "procedures", ",", "among", "others", "(", "degree", ",", "page", "-", "rank", ",", "eigenvector", "centrality", ",", "etc", ".", ")", ",", "to", "impose", "an", "order", "on", "the", "nodes", "of", "graphs", ",", "replacing", "application", "-", "dependent", "orders", "(", "temporal", ",", "spatial", ",", "etc", ".", ")", "where", "missing", ".", "Isomorphism", "and", "Canonicalization", ".", "The", "computational", "problem", "of", "deciding", "whether", "two", "graphs", "are", "isomorphic", "surfaces", "in", "several", "application", "domains", ".", "The", "graph", "isomorphism", "(", "GI", ")", "problem", "is", "in", "NP", "but", "not", "known", "to", "be", "in", "P", "or", "NP", "-", "hard", ".", "Under", "several", "mild", "restrictions", ",", "GI", "is", "known", "to", "be", "in", "P.", "For", "instance", ",", "GI", "is", "in", "P", "for", "graphs", "of", "bounded", "degree", ".", "A", "canonicalization", "of", "a", "graph", "is", "a", "graph", "with", "a", "fixed", "vertex", "order", "which", "is", "isomorphic", "to", "and", "which", "represents", "its", "entire", "isomorphism", "class", ".", "In", "practice", ",", "the", "graph", "canonicalization", "tool", "Nauty", "has", "shown", "remarkable", "performance", ".", "section", ":", "Learning", "CNNs", "for", "Arbitrary", "Graphs", "When", "CNNs", "are", "applied", "to", "images", ",", "a", "receptive", "field", "(", "a", "square", "grid", ")", "is", "moved", "over", "each", "image", "with", "a", "particular", "step", "size", ".", "The", "receptive", "field", "reads", "the", "pixels", "’", "feature", "values", ",", "for", "each", "channel", "once", ",", "and", "a", "patch", "of", "values", "is", "created", "for", "each", "channel", ".", "Since", "the", "pixels", "of", "an", "image", "have", "an", "implicit", "arrangement", "–", "a", "spatial", "order", "–", "the", "receptive", "fields", "are", "always", "moved", "from", "left", "to", "right", "and", "top", "to", "bottom", ".", "Moreover", ",", "the", "spatial", "order", "uniquely", "determines", "the", "nodes", "of", "each", "receptive", "field", "and", "the", "way", "these", "nodes", "are", "mapped", "to", "a", "vector", "space", "representation", "(", "see", "Figure", "[", "reference", "]", "(", "b", ")", ")", ".", "Consequently", ",", "the", "values", "read", "from", "two", "pixels", "using", "two", "different", "locations", "of", "the", "receptive", "field", "are", "assigned", "to", "the", "same", "relative", "position", "if", "and", "only", "if", "the", "pixels", "’", "structural", "roles", "(", "their", "spatial", "position", "within", "the", "receptive", "field", ")", "are", "identical", ".", "To", "show", "the", "connection", "between", "CNNs", "and", "Patchy", "-", "san", ",", "we", "frame", "CNNs", "on", "images", "as", "identifying", "a", "sequence", "of", "nodes", "in", "the", "square", "grid", "graph", "representing", "the", "image", "and", "building", "a", "normalized", "neighborhood", "graph", "–", "a", "receptive", "field", "–", "for", "each", "node", "in", "the", "identified", "sequence", ".", "For", "graph", "collections", "where", "an", "application", "-", "dependent", "node", "order", "is", "missing", "and", "where", "the", "nodes", "of", "any", "two", "graphs", "are", "not", "yet", "aligned", ",", "we", "need", "to", "determine", "for", "each", "graph", "(", "i", ")", "the", "sequences", "of", "nodes", "for", "which", "we", "create", "neighborhoods", ",", "and", "(", "ii", ")", "a", "unique", "mapping", "from", "the", "graph", "representation", "to", "a", "vector", "representation", "such", "that", "nodes", "with", "similar", "structural", "roles", "in", "the", "neighborhood", "graphs", "are", "positioned", "similarly", "in", "the", "vector", "representation", ".", "We", "address", "these", "problems", "by", "leveraging", "graph", "labeling", "procedures", "that", "assigns", "nodes", "from", "two", "different", "graphs", "to", "a", "similar", "relative", "position", "in", "their", "respective", "adjacency", "matrices", "if", "their", "structural", "roles", "within", "the", "graphs", "are", "similar", ".", "Given", "a", "collection", "of", "graphs", ",", "Patchy", "-", "san", "(", "Select", "-", "Assemble", "-", "Normalize", ")", "applies", "the", "following", "steps", "to", "each", "graph", ":", "(", "1", ")", "Select", "a", "fixed", "-", "length", "sequence", "of", "nodes", "from", "the", "graph", ";", "(", "2", ")", "assemble", "a", "fixed", "-", "size", "neighborhood", "for", "each", "node", "in", "the", "selected", "sequence", ";", "(", "3", ")", "normalize", "the", "extracted", "neighborhood", "graph", ";", "and", "(", "4", ")", "learn", "neighborhood", "representations", "with", "convolutional", "neural", "networks", "from", "the", "resulting", "sequence", "of", "patches", ".", "In", "the", "following", ",", "we", "describe", "methods", "that", "address", "the", "above", "-", "mentioned", "challenges", ".", "[", "t", "!", "]", "SelNodeSeq", ":", "Select", "Node", "Sequence", "{", "algorithmic}", "[", "1", "]", "graph", "labeling", "procedure", "ℓ", ",", "graph", "=", "G", "(", "V", ",", "E", ")", ",", "stride", "s", ",", "width", "w", ",", "receptive", "field", "size", "k", "=", "top", "w", "elements", "of", "V", "according", "to", "ℓ", "f", "to", "each", "input", "channel", "=", "j", "+", "j1", "subsection", ":", "Node", "Sequence", "Selection", "Node", "sequence", "selection", "is", "the", "process", "of", "identifying", ",", "for", "each", "input", "graph", ",", "a", "sequence", "of", "nodes", "for", "which", "receptive", "fields", "are", "created", ".", "Algorithm", "[", "reference", "]", "lists", "one", "such", "procedure", ".", "First", ",", "the", "vertices", "of", "the", "input", "graph", "are", "sorted", "with", "respect", "to", "a", "given", "graph", "labeling", ".", "Second", ",", "the", "resulting", "node", "sequence", "is", "traversed", "using", "a", "given", "stride", "and", "for", "each", "visited", "node", ",", "Algorithm", "[", "reference", "]", "is", "executed", "to", "construct", "a", "receptive", "field", ",", "until", "exactly", "receptive", "fields", "have", "been", "created", ".", "The", "stride", "determines", "the", "distance", ",", "relative", "to", "the", "selected", "node", "sequence", ",", "between", "two", "consecutive", "nodes", "for", "which", "a", "receptive", "field", "is", "created", ".", "If", "the", "number", "of", "nodes", "is", "smaller", "than", ",", "the", "algorithm", "creates", "all", "-", "zero", "receptive", "fields", "for", "padding", "purposes", ".", "Several", "alternative", "methods", "for", "vertex", "sequence", "selection", "are", "possible", ".", "For", "instance", ",", "a", "depth", "-", "first", "traversal", "of", "the", "input", "graph", "guided", "by", "the", "values", "of", "the", "graph", "labeling", ".", "We", "leave", "these", "ideas", "to", "future", "work", ".", "[", "t", "!", "]", "NeighAssemb", ":", "Neighborhood", "Assembly", "{", "algorithmic}", "[", "1", "]", "vertex", "v", ",", "receptive", "field", "size", "k", "set", "of", "neighborhood", "nodes", "N", "for", "v", "and", ">", "|L|0", "the", "set", "of", "vertices", "N", "subsection", ":", "Neighborhood", "Assembly", "For", "each", "of", "the", "nodes", "identified", "in", "the", "previous", "step", ",", "a", "receptive", "field", "has", "to", "be", "constructed", ".", "Algorithm", "[", "reference", "]", "first", "calls", "Algorithm", "[", "reference", "]", "to", "assembles", "a", "local", "neighborhood", "for", "the", "input", "node", ".", "The", "nodes", "of", "the", "neighborhood", "are", "the", "candidates", "for", "the", "receptive", "field", ".", "Algorithm", "[", "reference", "]", "lists", "the", "neighborhood", "assembly", "steps", ".", "Given", "as", "inputs", "a", "node", "and", "the", "size", "of", "the", "receptive", "field", ",", "the", "procedure", "performs", "a", "breadth", "-", "first", "search", ",", "exploring", "vertices", "with", "an", "increasing", "distance", "from", ",", "and", "adds", "these", "vertices", "to", "a", "set", ".", "If", "the", "number", "of", "collected", "nodes", "is", "smaller", "than", ",", "the", "-", "neighborhood", "of", "the", "vertices", "most", "recently", "added", "to", "are", "collected", ",", "and", "so", "on", ",", "until", "at", "least", "vertices", "are", "in", ",", "or", "until", "there", "are", "no", "more", "neighbors", "to", "add", ".", "Note", "that", "at", "this", "time", ",", "the", "size", "of", "is", "possibly", "different", "to", ".", "subsection", ":", "Graph", "Normalization", "The", "receptive", "field", "for", "a", "node", "is", "constructed", "by", "normalizing", "the", "neighborhood", "assembled", "in", "the", "previous", "step", ".", "Illustrated", "in", "Figure", "[", "reference", "]", ",", "the", "normalization", "imposes", "an", "order", "on", "the", "nodes", "of", "the", "neighborhood", "graph", "so", "as", "to", "map", "from", "the", "unordered", "graph", "space", "to", "a", "vector", "space", "with", "a", "linear", "order", ".", "The", "basic", "idea", "is", "to", "leverage", "graph", "labeling", "procedures", "that", "assigns", "nodes", "of", "two", "different", "graphs", "to", "a", "similar", "relative", "position", "in", "the", "respective", "adjacency", "matrices", "if", "and", "only", "if", "their", "structural", "roles", "within", "the", "graphs", "are", "similar", ".", "To", "formalize", "this", "intuition", ",", "we", "define", "the", "optimal", "graph", "normalization", "problem", "which", "aims", "to", "find", "a", "labeling", "that", "is", "optimal", "relative", "to", "a", "given", "collection", "of", "graphs", ".", "theorem", ":", "(", "Optimal", "graph", "normalization", ")", ".", "Let", "G", "be", "a", "collection", "of", "unlabeled", "graphs", "with", "k", "nodes", ",", "let", "ℓ", "be", "an", "injective", "graph", "labeling", "procedure", ",", "let", "dG", "be", "a", "distance", "measure", "on", "graphs", "with", "k", "nodes", ",", "and", "let", "dA", "be", "a", "distance", "measure", "on", "×kk", "matrices", ".", "Find", "^ℓ", "such", "that", "The", "problem", "amounts", "to", "finding", "a", "graph", "labeling", "procedure", ",", "such", "that", ",", "for", "any", "two", "graphs", "drawn", "uniformly", "at", "random", "from", ",", "the", "expected", "difference", "between", "the", "distance", "of", "the", "graphs", "in", "vector", "space", "(", "with", "respect", "to", "the", "adjacency", "matrices", "based", "on", ")", "and", "the", "distance", "of", "the", "graphs", "in", "graph", "space", "is", "minimized", ".", "The", "optimal", "graph", "normalization", "problem", "is", "a", "generalization", "of", "the", "classical", "graph", "canonicalization", "problem", ".", "A", "canonical", "labeling", "algorithm", ",", "however", ",", "is", "optimal", "only", "for", "isomorphic", "graphs", "and", "might", "perform", "poorly", "for", "graphs", "that", "are", "similar", "but", "not", "isomorphic", ".", "In", "contrast", ",", "the", "smaller", "the", "expectation", "of", "the", "optimal", "normalization", "problem", ",", "the", "better", "the", "labeling", "aligns", "nodes", "with", "similar", "structural", "roles", ".", "Note", "that", "the", "similarity", "is", "determined", "by", ".", "[", "t", "!", "]", "ReceptiveField", ":", "Create", "Receptive", "Field", "{", "algorithmic}", "[", "1", "]", "vertex", "v", ",", "graph", "labeling", "ℓ", ",", "receptive", "field", "size", "k", "Gnorm", "We", "have", "the", "following", "result", "concerning", "the", "complexity", "of", "the", "optimal", "normalization", "problem", ".", "theorem", ":", ".", "Optimal", "graph", "normalization", "is", "NP", "-", "hard", ".", "proof", ":", "Proof", ".", "By", "reduction", "from", "subgraph", "isomorphism", ".", "∎", "Patchy", "-", "san", "does", "not", "solve", "the", "above", "optimization", "problem", ".", "Instead", ",", "it", "may", "compare", "different", "graph", "labeling", "methods", "and", "choose", "the", "one", "that", "performs", "best", "relative", "to", "a", "given", "collection", "of", "graphs", ".", "[", "t", "!", "]", "NormalizeGraph", ":", "Graph", "Normalization", "{", "algorithmic}", "[", "1", "]", "subset", "of", "vertices", "U", "from", "original", "graph", "G", ",", "vertex", "v", ",", "graph", "labeling", "ℓ", ",", "receptive", "field", "size", "k", "receptive", "field", "for", "v", "ranking", "r", "of", "U", "using", "ℓ", ",", "subject", "to", ":", "∈∀u", ",", "wU⁢d", "(", "u", ",", "v", ")", "<⁢d", "(", "w", ",", "v", ")", "⇒⁢r", "(", "u", ")", "<⁢r", "(", "w", ")", "top", "k", "vertices", "in", "U", "according", "to", "r", "ranking", "r", "of", "N", "using", "ℓ", ",", "subject", "to", ":", "∈∀u", ",", "wN⁢d", "(", "u", ",", "v", ")", "<⁢d", "(", "w", ",", "v", ")", "⇒⁢r", "(", "u", ")", "<⁢r", "(", "w", ")", "and", "-", "k|U|", "dummy", "nodes", "the", "subgraph", "⁢G", "[", "N", "]", "for", "the", "vertices", "N", "⁢G", "[", "N", "]", ",", "respecting", "the", "prior", "coloring", "r", "⁢G", "[", "N", "]", "theorem", ":", ".", "Let", "G", "be", "a", "collection", "of", "graphs", "and", "let", "(", "G1", ",", "G1′", ")", ",", "…", ",", "(", "GN", ",", "GN′", ")", "be", "a", "sequence", "of", "pairs", "of", "graphs", "sampled", "independently", "and", "uniformly", "at", "random", "from", "G.", "Let", ":", "=", "^θℓ∑=i1N", "/", "⁢dA", "(", "⁢Aℓ", "(", "Gi", "),", "⁢Aℓ", "(", "Gi′", "))", "N", "and", ":", "=", "θℓ⁢EG", "[", "|", "-", "⁢dA", "(", "⁢Aℓ", "(", "G", "),", "⁢Aℓ", "(", "G′", "))", "⁢dG", "(", "G", ",", "G′", ")", "|", "]", ".", "If", "≥dAdG", ",", "then", "<", "⁢EG", "[", "^θℓ1", "]", "⁢EG", "[", "^θℓ2", "]", "if", "and", "only", "if", "<", "θℓ1θℓ2", ".", "Theorem", "[", "reference", "]", "enables", "us", "to", "compare", "different", "labeling", "procedures", "in", "an", "unsupervised", "manner", "via", "a", "comparison", "of", "the", "corresponding", "estimators", ".", "Under", "the", "assumption", ",", "the", "smaller", "the", "estimate", "the", "smaller", "the", "absolute", "difference", ".", "Therefore", ",", "we", "can", "simply", "choose", "the", "labeling", "for", "which", "is", "minimal", ".", "The", "assumption", "holds", ",", "for", "instance", ",", "for", "the", "edit", "distance", "on", "graphs", "and", "the", "Hamming", "distance", "on", "adjacency", "matrices", ".", "Finally", ",", "note", "that", "all", "of", "the", "above", "results", "can", "be", "extended", "to", "directed", "graphs", ".", "The", "graph", "normalization", "problem", "and", "the", "application", "of", "appropriate", "graph", "labeling", "procedures", "for", "the", "normalization", "of", "local", "graph", "structures", "is", "at", "the", "core", "of", "the", "proposed", "approach", ".", "Within", "the", "Patchy", "-", "san", "framework", ",", "we", "normalize", "the", "neighborhood", "graphs", "of", "a", "vertex", ".", "The", "labeling", "of", "the", "vertices", "is", "therefore", "constrained", "by", "the", "graph", "distance", "to", ":", "for", "any", "two", "vertices", ",", "if", "is", "closer", "to", "than", ",", "then", "is", "always", "ranked", "higher", "than", ".", "This", "definition", "ensures", "that", "has", "always", "rank", ",", "and", "that", "the", "closer", "a", "vertex", "is", "to", "in", ",", "the", "higher", "it", "is", "ranked", "in", "the", "vector", "space", "representation", ".", "Since", "most", "labeling", "methods", "are", "not", "injective", ",", "it", "is", "necessary", "to", "break", "ties", "between", "same", "-", "label", "nodes", ".", "To", "do", "so", ",", "we", "use", "Nauty", ".", "Nauty", "accepts", "prior", "node", "partitions", "as", "input", "and", "breaks", "remaining", "ties", "by", "choosing", "the", "lexicographically", "maximal", "adjacency", "matrix", ".", "It", "is", "known", "that", "graph", "isomorphism", "is", "in", "PTIME", "for", "graphs", "of", "bounded", "degree", ".", "Due", "to", "the", "constant", "size", "of", "the", "neighborhood", "graphs", ",", "the", "algorithm", "runs", "in", "time", "polynomial", "in", "the", "size", "of", "the", "original", "graph", "and", ",", "on", "average", ",", "in", "time", "linear", "in", ".", "Our", "experiments", "verify", "that", "computing", "a", "canonical", "labeling", "of", "the", "graph", "neigborhoods", "adds", "a", "negligible", "overhead", ".", "Algorithm", "[", "reference", "]", "lists", "the", "normalization", "procedure", ".", "If", "the", "size", "of", "the", "input", "set", "is", "larger", "than", ",", "it", "first", "applies", "the", "ranking", "based", "on", "to", "select", "the", "top", "nodes", "and", "recomputes", "a", "ranking", "on", "the", "smaller", "set", "of", "nodes", ".", "If", "the", "size", "of", "is", "smaller", "than", ",", "it", "adds", "disconnected", "dummy", "nodes", ".", "Finally", ",", "it", "induces", "the", "subgraph", "on", "the", "vertices", "and", "canonicalizes", "the", "graph", "taking", "the", "ranking", "as", "prior", "coloring", ".", "We", "can", "relate", "Patchy", "-", "san", "to", "CNNs", "for", "images", "as", "follows", ".", "theorem", ":", ".", "Given", "a", "sequence", "of", "pixels", "taken", "from", "an", "image", ".", "Applying", "Patchy", "-", "san", "with", "receptive", "field", "size", "(", "-", "⁢2m1", ")", "2", ",", "stride", "s", ",", "no", "zero", "padding", ",", "and", "1", "-", "WL", "normalization", "to", "the", "sequence", "is", "identical", "(", "up", "to", "a", "fixed", "permutation", "of", "the", "receptive", "field", ")", "to", "the", "first", "layer", "of", "a", "CNN", "with", "receptive", "field", "size", "-", "⁢2m1", ",", "stride", "s", ",", "and", "no", "zero", "padding", ".", "proof", ":", "Proof", ".", "It", "is", "possible", "to", "show", "that", "if", "an", "input", "graph", "is", "a", "square", "grid", ",", "then", "the", "-", "WL", "normalized", "receptive", "field", "constructed", "for", "a", "vertex", "is", "always", "a", "square", "grid", "graph", "with", "a", "unique", "vertex", "order", ".", "∎", "subsection", ":", "Convolutional", "Architecture", "Patchy", "-", "san", "is", "able", "to", "process", "both", "vertex", "and", "edge", "attributes", "(", "discrete", "and", "continuous", ")", ".", "Let", "be", "the", "number", "of", "vertex", "attributes", "and", "let", "be", "the", "number", "of", "edge", "attributes", ".", "For", "each", "input", "graph", ",", "it", "applies", "normalized", "receptive", "fields", "for", "vertices", "and", "edges", "which", "results", "in", "one", "and", "one", "tensor", ".", "These", "can", "be", "reshaped", "to", "a", "and", "a", "tensors", ".", "Note", "that", "and", "are", "the", "number", "of", "input", "channels", ".", "We", "can", "now", "apply", "a", "-", "dimensional", "convolutional", "layer", "with", "stride", "and", "receptive", "field", "size", "to", "the", "first", "and", "to", "the", "second", "tensor", ".", "The", "rest", "of", "the", "architecture", "can", "be", "chosen", "arbitrarily", ".", "We", "may", "use", "merge", "layers", "to", "combine", "convolutional", "layers", "representing", "nodes", "and", "edges", ",", "respectively", ".", "section", ":", "Complexity", "and", "Implementation", "Patchy", "-", "san", "’s", "algorithm", "for", "creating", "receptive", "fields", "is", "highly", "efficient", "and", "naively", "parallelizable", "because", "the", "fields", "are", "generated", "independently", ".", "We", "can", "show", "the", "following", "asymptotic", "worst", "-", "case", "result", ".", "theorem", ":", ".", "Let", "N", "be", "the", "number", "of", "graphs", ",", "let", "k", "be", "the", "receptive", "field", "size", ",", "w", "the", "width", ",", "and", "⁢O", "(", "⁢f", "(", "n", ",", "m", ")", ")", "the", "complexity", "of", "computing", "a", "given", "labeling", "ℓ", "for", "a", "graph", "with", "n", "vertices", "and", "m", "edges", ".", "Patchy", "-", "san", "has", "a", "worst", "-", "case", "complexity", "of", "⁢O", "(", "⁢Nw", "(+", "⁢f", "(", "n", ",", "m", ")", "⁢nlog", "(", "n", ")", "exp", "(", "k", ")", ")", ")", "for", "computing", "the", "receptive", "fields", "for", "N", "graphs", ".", "proof", ":", "Proof", ".", "Node", "sequence", "selection", "requires", "the", "labeling", "of", "each", "input", "graph", "and", "the", "retrieval", "of", "the", "highest", "ranked", "nodes", ".", "For", "the", "creation", "of", "normalized", "graph", "patches", ",", "most", "computational", "effort", "is", "spent", "applying", "the", "labeling", "procedure", "to", "a", "neighborhood", "whose", "size", "may", "be", "larger", "than", ".", "Let", "be", "the", "maximum", "degree", "of", "the", "input", "graph", ",", "and", "the", "neighborhood", "returned", "by", "Algorithm", "[", "reference", "]", ".", "We", "have", ".", "The", "term", "comes", "from", "the", "worst", "-", "case", "complexity", "of", "the", "graph", "canonicalization", "algorithm", "Nauty", "on", "a", "node", "graph", ".", "∎", "For", "instance", ",", "for", "the", "Weisfeiler", "-", "Lehman", "algorithm", ",", "which", "has", "a", "complexity", "of", ",", "and", "constants", "and", ",", "the", "complexity", "of", "Patchy", "-", "san", "is", "linear", "in", "and", "quasi", "-", "linear", "in", "and", ".", "section", ":", "Experiments", "We", "conduct", "three", "types", "of", "experiments", ":", "a", "runtime", "analysis", ",", "a", "qualitative", "analysis", "of", "the", "learned", "features", ",", "and", "a", "comparison", "to", "graph", "kernels", "on", "benchmark", "data", "sets", ".", "[", "subfigure", "]", "labelformat", "=", "empty", "subsection", ":", "Runtime", "Analysis", "We", "assess", "the", "efficiency", "of", "Patchy", "-", "san", "by", "applying", "it", "to", "real", "-", "world", "graphs", ".", "The", "objective", "is", "to", "compare", "the", "rates", "at", "which", "receptive", "fields", "are", "generated", "to", "the", "rate", "at", "which", "state", "of", "the", "art", "CNNs", "perform", "learning", ".", "All", "input", "graphs", "are", "part", "of", "the", "collection", "of", "the", "Python", "module", "graph", "-", "toolhttps:", "//", "graph", "-", "tool.skewed.de", "/", ".", "For", "a", "given", "graph", ",", "we", "used", "Patchy", "-", "san", "to", "compute", "a", "receptive", "field", "for", "all", "nodes", "using", "the", "-", "dimensional", "Weisfeiler", "-", "Lehman", "(", "1", "-", "WL", ")", "algorithm", "for", "the", "normalization", ".", "torus", "is", "a", "periodic", "lattice", "with", "nodes", ";", "random", "is", "a", "random", "undirected", "graph", "with", "nodes", "and", "a", "degree", "distribution", "and", ";", "power", "is", "a", "network", "representing", "the", "topology", "of", "a", "power", "grid", "in", "the", "US", ";", "polbooks", "is", "a", "co", "-", "purchasing", "network", "of", "books", "about", "US", "politics", "published", "during", "the", "presidential", "election", ";", "preferential", "is", "a", "preferential", "attachment", "network", "model", "where", "newly", "added", "vertices", "have", "degree", ";", "astro", "-", "ph", "is", "a", "coauthorship", "network", "between", "authors", "of", "preprints", "posted", "on", "the", "astrophysics", "arxiv", ";", "email", "-", "enron", "is", "a", "communication", "network", "generated", "from", "about", "half", "a", "million", "sent", "emails", ".", "All", "experiments", "were", "run", "on", "commodity", "hardware", "with", "64", "G", "RAM", "and", "a", "single", "2.8", "GHz", "CPU", ".", "Figure", "[", "reference", "]", "depicts", "the", "receptive", "fields", "per", "second", "rates", "for", "each", "input", "graph", ".", "For", "receptive", "field", "size", "and", "Patchy", "-", "san", "creates", "fields", "at", "a", "rate", "of", "more", "than", "except", "for", "email", "-", "enron", "with", "a", "rate", "of", "and", ",", "respectively", ".", "For", ",", "the", "largest", "tested", "size", ",", "fields", "are", "created", "at", "a", "rate", "of", "at", "least", ".", "A", "CNN", "with", "convolutional", "and", "dense", "layers", "learns", "at", "a", "rate", "of", "about", "-", "training", "examples", "per", "second", "on", "the", "same", "machine", ".", "Hence", ",", "the", "speed", "at", "which", "receptive", "fields", "are", "generated", "is", "sufficient", "to", "saturate", "a", "downstream", "CNN", ".", "subsection", ":", "Feature", "Visualization", "The", "visualization", "experiments", "’", "aim", "is", "to", "qualitatively", "investigate", "whether", "popular", "models", "such", "as", "the", "restricted", "Boltzman", "machine", "(", "RBM", ")", "can", "be", "combined", "with", "Patchy", "-", "san", "for", "unsupervised", "feature", "learning", ".", "For", "every", "input", "graph", ",", "we", "have", "generated", "receptive", "fields", "for", "all", "nodes", "and", "used", "these", "as", "input", "to", "an", "RBM", ".", "The", "RBM", "had", "hidden", "nodes", "and", "was", "trained", "for", "epochs", "with", "contrastive", "divergence", "and", "a", "learning", "rate", "of", ".", "We", "visualize", "the", "features", "learned", "by", "a", "single", "-", "layer", "RBM", "for", "-", "dimensional", "Weisfeiler", "-", "Lehman", "(", "1", "-", "WL", ")", "normalized", "receptive", "fields", "of", "size", ".", "Note", "that", "the", "features", "learned", "by", "the", "RBM", "correspond", "to", "reoccurring", "receptive", "field", "patterns", ".", "Figure", "[", "reference", "]", "depicts", "some", "of", "the", "features", "and", "samples", "drawn", "from", "it", "for", "four", "different", "graphs", ".", "subsection", ":", "Graph", "Classification", "Graph", "classification", "is", "the", "problem", "of", "assigning", "graphs", "to", "one", "of", "several", "categories", ".", "Data", "Sets", ".", "We", "use", "standard", "benchmark", "data", "sets", "to", "compare", "run", "-", "time", "and", "classification", "accuracy", "with", "state", "of", "the", "art", "graph", "kernels", ":", "MUTAG", ",", "PCT", ",", "NCI1", ",", "NCI109", ",", "PROTEIN", ",", "and", "D", "&", "D.", "MUTAG", "is", "a", "data", "set", "of", "nitro", "compounds", "where", "classes", "indicate", "whether", "the", "compound", "has", "a", "mutagenic", "effect", "on", "a", "bacterium", ".", "PTC", "consists", "of", "chemical", "compounds", "where", "classes", "indicate", "carcinogenicity", "for", "male", "and", "female", "rats", ".", "NCI1", "and", "NCI109", "are", "chemical", "compounds", "screened", "for", "activity", "against", "non", "-", "small", "cell", "lung", "cancer", "and", "ovarian", "cancer", "cell", "lines", ".", "PROTEINS", "is", "a", "graph", "collection", "where", "nodes", "are", "secondary", "structure", "elements", "and", "edges", "indicate", "neighborhood", "in", "the", "amino", "-", "acid", "sequence", "or", "in", "3D", "space", ".", "Graphs", "are", "classified", "as", "enzyme", "or", "non", "-", "enzyme", ".", "D", "&", "D", "is", "a", "data", "set", "of", "protein", "structures", "classified", "into", "enzymes", "and", "non", "-", "enzymes", ".", "Experimental", "Set", "-", "up", ".", "We", "compared", "Patchy", "-", "san", "with", "the", "shortest", "-", "path", "kernel", "(", "SP", ")", ",", "the", "random", "walk", "kernel", "(", "RW", ")", ",", "the", "graphlet", "count", "kernel", "(", "GK", ")", ",", "and", "the", "Weisfeiler", "-", "Lehman", "subtree", "kernel", "(", "WL", ")", ".", "Similar", "to", "previous", "work", ",", "we", "set", "the", "height", "parameter", "of", "WL", "to", ",", "the", "size", "of", "the", "graphlets", "for", "GK", "to", ",", "and", "chose", "the", "decay", "factor", "for", "RW", "from", ".", "We", "performed", "-", "fold", "cross", "-", "validation", "with", "LIB", "-", "SVM", ",", "using", "folds", "for", "training", "and", "for", "testing", ",", "and", "repeated", "the", "experiments", "times", ".", "We", "report", "average", "prediction", "accuracies", "and", "standard", "deviations", ".", "For", "Patchy", "-", "san", "(", "referred", "to", "as", "PSCN", ")", ",", "we", "used", "-", "dimensional", "WL", "normalization", ",", "a", "width", "equal", "to", "the", "average", "number", "of", "nodes", "(", "see", "Table", "[", "reference", "]", ")", ",", "and", "receptive", "field", "sizes", "of", "and", ".", "For", "the", "experiments", "we", "only", "used", "node", "attributes", ".", "In", "addition", ",", "we", "ran", "experiments", "for", "where", "we", "combined", "receptive", "fields", "for", "nodes", "and", "edges", "using", "a", "merge", "layer", "(", ")", ".", "To", "make", "a", "fair", "comparison", ",", "we", "used", "a", "single", "network", "architecture", "with", "two", "convolutional", "layers", ",", "one", "dense", "hidden", "layer", ",", "and", "a", "softmax", "layer", "for", "all", "experiments", ".", "The", "first", "convolutional", "layer", "had", "output", "channels", "(", "feature", "maps", ")", ".", "The", "second", "conv", "layer", "has", "output", "channels", ",", "a", "stride", "of", ",", "and", "a", "field", "size", "of", ".", "The", "convolutional", "layers", "have", "rectified", "linear", "units", ".", "The", "dense", "layer", "has", "rectified", "linear", "units", "with", "a", "dropout", "rate", "of", ".", "Dropout", "and", "the", "relatively", "small", "number", "of", "neurons", "are", "needed", "to", "avoid", "overfitting", "on", "the", "smaller", "data", "sets", ".", "The", "only", "hyperparameter", "we", "optimized", "is", "the", "number", "of", "epochs", "and", "the", "batch", "size", "for", "the", "mini", "-", "batch", "gradient", "decent", "algorithm", "rmsprop", ".", "All", "of", "the", "above", "was", "implemented", "with", "the", "Theano", "wrapper", "Keras", ".", "We", "also", "applied", "a", "logistic", "regression", "(", "PSLR", ")", "classifier", "on", "the", "patches", "for", ".", "Moreover", ",", "we", "ran", "experiments", "with", "the", "same", "set", "-", "up", "on", "larger", "social", "graph", "data", "sets", "(", "up", "to", "graphs", "each", ",", "with", "an", "average", "of", "nodes", ")", ",", "and", "compared", "Patchy", "-", "san", "with", "previously", "reported", "results", "for", "the", "graphlet", "count", "(", "GK", ")", "and", "the", "deep", "graphlet", "count", "kernel", "(", "DGK", ")", ".", "We", "used", "the", "normalized", "node", "degree", "as", "attribute", "for", "Patchy", "-", "san", ",", "highlighting", "one", "of", "its", "advantages", ":", "it", "can", "easily", "incorporate", "continuous", "features", ".", "Results", ".", "Table", "[", "reference", "]", "lists", "the", "results", "of", "the", "experiments", ".", "We", "omit", "the", "results", "for", "NCI109", "as", "they", "are", "almost", "identical", "to", "NCI1", ".", "Despite", "using", "a", "one", "-", "fits", "-", "all", "CNN", "architecture", ",", "the", "CNNs", "accuracy", "is", "highly", "competitive", "with", "existing", "graph", "kernels", ".", "In", "most", "cases", ",", "a", "receptive", "field", "size", "of", "results", "in", "the", "best", "classification", "accuracy", ".", "The", "relatively", "high", "variance", "can", "be", "explained", "with", "the", "small", "size", "of", "the", "benchmark", "data", "sets", "and", "the", "fact", "that", "the", "CNNs", "hyperparameters", "(", "with", "the", "exception", "of", "epochs", "and", "batch", "size", ")", "were", "not", "tuned", "to", "individual", "data", "sets", ".", "Similar", "to", "the", "experience", "on", "image", "and", "text", "data", ",", "we", "expect", "Patchy", "-", "san", "to", "perform", "even", "better", "for", "large", "data", "sets", ".", "Moreover", ",", "Patchy", "-", "san", "is", "between", "and", "times", "more", "efficient", "than", "the", "most", "efficient", "graph", "kernel", "(", "WL", ")", ".", "We", "expect", "the", "performance", "advantage", "to", "be", "much", "more", "pronounced", "for", "data", "sets", "with", "a", "large", "number", "of", "graphs", ".", "Results", "for", "betweeness", "centrality", "normalization", "are", "similar", "with", "the", "exception", "of", "the", "runtime", "which", "increases", "by", "about", "%", ".", "Logistic", "regression", "applied", "to", "Patchy", "-", "san", "’s", "receptive", "fields", "performs", "worse", ",", "indicating", "that", "Patchy", "-", "san", "works", "especially", "well", "in", "conjunction", "with", "CNNs", "which", "learn", "non", "-", "linear", "feature", "combinations", "and", "which", "share", "weights", "across", "receptive", "fields", ".", "Patchy", "-", "san", "is", "also", "highly", "competitive", "on", "the", "social", "graph", "data", ".", "It", "significantly", "outperforms", "the", "other", "two", "kernels", "on", "four", "of", "the", "six", "data", "sets", "and", "achieves", "ties", "on", "the", "rest", ".", "Table", "[", "reference", "]", "lists", "the", "results", "of", "the", "experiments", ".", "section", ":", "Conclusion", "and", "Future", "Work", "We", "proposed", "a", "framework", "for", "learning", "graph", "representations", "that", "are", "especially", "beneficial", "in", "conjunction", "with", "CNNs", ".", "It", "combines", "two", "complementary", "procedures", ":", "(", "a", ")", "selecting", "a", "sequence", "of", "nodes", "that", "covers", "large", "parts", "of", "the", "graph", "and", "(", "b", ")", "generating", "local", "normalized", "neighborhood", "representations", "for", "each", "of", "the", "nodes", "in", "the", "sequence", ".", "Experiments", "show", "that", "the", "approach", "is", "competitive", "with", "state", "of", "the", "art", "graph", "kernels", ".", "Directions", "for", "future", "work", "include", "the", "use", "of", "alternative", "neural", "network", "architectures", "such", "as", "RNNs", ";", "combining", "different", "receptive", "field", "sizes", ";", "pretraining", "with", "RBMs", "and", "autoencoders", ";", "and", "statistical", "relational", "models", "based", "on", "the", "ideas", "of", "the", "approach", ".", "section", ":", "Acknowledgments", "Many", "thanks", "to", "the", "anonymous", "ICML", "reviewers", "who", "provided", "tremendously", "helpful", "comments", ".", "The", "research", "leading", "to", "these", "results", "has", "received", "funding", "from", "the", "European", "Union", "’s", "Horizon", "2020", "innovation", "action", "program", "under", "grant", "agreement", "No", "653449", "-", "TYPES", ".", "bibliography", ":", "References"], "entity": [[7, 8, ["learning"], "Task"], [9, 11, ["graph", "data"], "Material"], [18, 21, ["convolutional", "neural", "networks"], "Method"], [22, 24, ["arbitrary", "graphs"], "Material"], [46, 51, ["image", "-", "based", "convolutional", "networks"], "Method"], [67, 71, ["extracting", "locally", "connected", "regions"], "Task"], [76, 79, ["benchmark", "data", "sets"], "Material"], [85, 87, ["feature", "representations"], "Method"], [94, 96, ["graph", "kernels"], "Method"], [104, 105, ["LearningConvolutionalNeuralNetworksforGraphs"], "Task"], [115, 118, ["convolutional", "neural", "networks"], "Method"], [125, 130, ["graph", "-", "based", "learning", "problems"], "Task"], [142, 143, ["graphs"], "Material"], [152, 156, ["classification", "and", "regression", "problems"], "Task"], [157, 159, ["unseen", "graphs"], "Material"], [210, 212, ["graph", "representations"], "Method"], [234, 236, ["learning", "representations"], "Task"], [237, 243, ["classes", "of", "directed", "and", "undirected", "graphs"], "Task"], [267, 270, ["convolutional", "neural", "network"], "Method"], [271, 272, ["images"], "Material"], [296, 298, ["convolutional", "architecture"], "Method"], [305, 307, ["graph", "representations"], "Method"], [315, 318, ["convolutional", "neural", "networks"], "Method"], [319, 320, ["CNNs"], "Method"], [322, 323, ["images"], "Material"], [342, 343, ["CNN"], "Method"], [344, 345, ["images"], "Material"], [347, 348, ["image"], "Material"], [353, 356, ["square", "grid", "graph"], "Method"], [364, 365, ["CNN"], "Method"], [466, 468, ["NLP", "problems"], "Task"], [488, 490, ["graph", "collections"], "Task"], [546, 550, ["normalization", "of", "neighborhood", "graphs"], "Method"], [559, 561, ["graph", "representation"], "Method"], [563, 566, ["vector", "space", "representation"], "Method"], [572, 575, ["Patchy", "-", "san"], "Method"], [581, 583, ["arbitrary", "graphs"], "Task"], [654, 657, ["feature", "learning", "components"], "Method"], [659, 663, ["convolutional", "and", "dense", "layers"], "Method"], [672, 676, ["CNN", "’s", "receptive", "fields"], "Method"], [683, 687, ["Patchy", "-", "san", "architecture"], "Method"], [708, 710, ["large", "graphs"], "Task"], [721, 723, ["computational", "biology"], "Task"], [724, 727, ["social", "network", "analysis"], "Task"], [737, 740, ["Patchy", "-", "san"], "Method"], [741, 743, ["feature", "visualizations"], "Task"], [759, 761, ["graph", "kernel"], "Method"], [762, 765, ["Patchy", "-", "san"], "Method"], [773, 775, ["feature", "engineering"], "Method"], [784, 788, ["normalization", "problem", "on", "graphs"], "Task"], [790, 791, ["complexity"], "Metric"], [796, 799, ["graph", "labeling", "approaches"], "Method"], [811, 814, ["Patchy", "-", "san"], "Method"], [815, 816, ["CNNs"], "Method"], [817, 818, ["images"], "Material"], [821, 824, ["benchmark", "data", "sets"], "Material"], [830, 831, ["CNNs"], "Method"], [832, 833, ["graphs"], "Task"], [844, 846, ["graph", "kernels"], "Method"], [851, 853, ["Graph", "kernels"], "Method"], [854, 859, ["kernel", "-", "based", "learning", "approaches"], "Method"], [861, 862, ["SVMs"], "Method"], [868, 869, ["Kernels"], "Method"], [892, 895, ["skew", "spectrum", "kernel"], "Method"], [896, 897, ["kernels"], "Method"], [899, 900, ["graphlets"], "Method"], [912, 913, ["kernels"], "Method"], [941, 943, ["combinatorial", "complexity"], "Metric"], [944, 946, ["subgraph", "enumeration"], "Task"], [947, 949, ["graphlet", "kernels"], "Method"], [961, 963, ["graph", "kernels"], "Method"], [965, 972, ["Weisfeiler", "-", "Lehman", "(", "WL", ")", "kernels"], "Method"], [973, 975, ["WL", "kernels"], "Method"], [996, 999, ["Patchy", "-", "san"], "Method"], [1000, 1001, ["WL"], "Method"], [1004, 1006, ["labeling", "procedure"], "Method"], [1011, 1014, ["Deep", "graph", "kernels"], "Method"], [1015, 1018, ["graph", "invariant", "kernels"], "Method"], [1019, 1020, ["graphs"], "Method"], [1046, 1049, ["Patchy", "-", "san"], "Method"], [1050, 1051, ["substructures"], "Method"], [1052, 1054, ["graph", "data"], "Material"], [1069, 1071, ["graph", "kernels"], "Method"], [1073, 1075, ["training", "complexity"], "Metric"], [1088, 1092, ["large", "-", "scale", "problems"], "Task"], [1093, 1096, ["Patchy", "-", "san"], "Method"], [1104, 1107, ["Graph", "neural", "networks"], "Method"], [1108, 1109, ["GNNs"], "Method"], [1112, 1116, ["recurrent", "neural", "network", "architecture"], "Method"], [1118, 1119, ["graphs"], "Material"], [1120, 1121, ["GNNs"], "Method"], [1122, 1125, ["recurrent", "neural", "networks"], "Method"], [1133, 1135, ["node", "representations"], "Method"], [1144, 1146, ["node", "representations"], "Method"], [1152, 1156, ["classification", "and", "regression", "problems"], "Task"], [1157, 1158, ["GNNs"], "Method"], [1166, 1168, ["backpropagation", "operations"], "Method"], [1181, 1186, ["Gated", "Graph", "Sequence", "Neural", "Networks"], "Method"], [1187, 1188, ["GNNs"], "Method"], [1190, 1193, ["gated", "recurrent", "units"], "Method"], [1201, 1202, ["CNNs"], "Method"], [1240, 1243, ["convolutional", "type", "operations"], "Method"], [1248, 1250, ["differentiable", "variant"], "Method"], [1269, 1271, ["convolutional", "networks"], "Method"], [1272, 1274, ["graph", "theory"], "Method"], [1277, 1280, ["Convolutional", "Neural", "Networks"], "Method"], [1280, 1281, ["CNNs"], "Method"], [1313, 1314, ["CNNs"], "Method"], [1324, 1334, ["image", ",", "speech", ",", "text", ",", "and", "drug", "discovery", "problems"], "Task"], [1338, 1339, ["CNNs"], "Method"], [1341, 1342, ["Neocognitron"], "Method"], [1345, 1346, ["CNN"], "Method"], [1349, 1353, ["convolutional", "and", "dense", "layers"], "Method"], [1359, 1361, ["convolutional", "layer"], "Method"], [1376, 1377, ["CNNs"], "Method"], [1521, 1525, ["graph", "-", "theoretical", "concept"], "Method"], [1527, 1528, ["labeling"], "Method"], [1598, 1599, ["Labeling"], "Task"], [1600, 1602, ["Node", "Partitions"], "Task"], [1603, 1606, ["Patchy", "-", "san"], "Method"], [1607, 1609, ["graph", "labelings"], "Method"], [1617, 1619, ["graph", "labeling"], "Method"], [1640, 1643, ["graph", "labeling", "procedure"], "Method"], [1645, 1647, ["graph", "labeling"], "Method"], [1662, 1663, ["labeling"], "Task"], [1669, 1671, ["graph", "labeling"], "Method"], [1679, 1680, ["ranking"], "Task"], [1682, 1683, ["coloring"], "Task"], [1701, 1704, ["labeling", "of", "graph"], "Method"], [1744, 1747, ["graph", "labeling", "procedures"], "Method"], [1752, 1755, ["measures", "of", "centrality"], "Metric"], [1759, 1762, ["analysis", "of", "networks"], "Task"], [1783, 1787, ["Weisfeiler", "-", "Lehman", "algorithm"], "Method"], [1791, 1797, ["partitioning", "the", "vertices", "of", "a", "graph"], "Task"], [1803, 1805, ["color", "refinement"], "Task"], [1806, 1809, ["naive", "vertex", "classification"], "Task"], [1810, 1812, ["Color", "refinement"], "Task"], [1818, 1820, ["ML", "community"], "Task"], [1829, 1830, ["inference"], "Task"], [1831, 1833, ["graphical", "models"], "Task"], [1842, 1845, ["Patchy", "-", "san"], "Method"], [1847, 1849, ["labeling", "procedures"], "Method"], [1892, 1893, ["Isomorphism"], "Method"], [1894, 1895, ["Canonicalization"], "Method"], [1897, 1899, ["computational", "problem"], "Task"], [1913, 1919, ["graph", "isomorphism", "(", "GI", ")", "problem"], "Task"], [1939, 1940, ["GI"], "Method"], [1960, 1964, ["canonicalization", "of", "a", "graph"], "Method"], [1988, 1992, ["graph", "canonicalization", "tool", "Nauty"], "Method"], [1999, 2001, ["Learning", "CNNs"], "Task"], [2002, 2004, ["Arbitrary", "Graphs"], "Task"], [2005, 2006, ["CNNs"], "Method"], [2108, 2111, ["vector", "space", "representation"], "Method"], [2171, 2172, ["CNNs"], "Method"], [2173, 2176, ["Patchy", "-", "san"], "Task"], [2179, 2180, ["CNNs"], "Method"], [2181, 2182, ["images"], "Material"], [2190, 2193, ["square", "grid", "graph"], "Method"], [2199, 2202, ["normalized", "neighborhood", "graph"], "Method"], [2216, 2218, ["graph", "collections"], "Task"], [2269, 2271, ["graph", "representation"], "Method"], [2273, 2275, ["vector", "representation"], "Method"], [2291, 2293, ["vector", "representation"], "Method"], [2300, 2303, ["graph", "labeling", "procedures"], "Method"], [2334, 2335, ["graphs"], "Material"], [2336, 2345, ["Patchy", "-", "san", "(", "Select", "-", "Assemble", "-", "Normalize"], "Method"], [2400, 2402, ["neighborhood", "representations"], "Method"], [2403, 2406, ["convolutional", "neural", "networks"], "Method"], [2432, 2433, ["SelNodeSeq"], "Method"], [2442, 2445, ["graph", "labeling", "procedure"], "Method"], [2486, 2489, ["Node", "Sequence", "Selection"], "Method"], [2489, 2492, ["Node", "sequence", "selection"], "Method"], [2622, 2624, ["padding", "purposes"], "Task"], [2629, 2632, ["vertex", "sequence", "selection"], "Task"], [2639, 2643, ["depth", "-", "first", "traversal"], "Method"], [2653, 2655, ["graph", "labeling"], "Method"], [2701, 2703, ["Neighborhood", "Assembly"], "Method"], [2782, 2786, ["breadth", "-", "first", "search"], "Method"], [2863, 2865, ["Graph", "Normalization"], "Method"], [2926, 2929, ["graph", "labeling", "procedures"], "Method"], [2967, 2971, ["optimal", "graph", "normalization", "problem"], "Task"], [2976, 2977, ["labeling"], "Task"], [2991, 2994, ["Optimal", "graph", "normalization"], "Method"], [3002, 3004, ["unlabeled", "graphs"], "Material"], [3012, 3016, ["injective", "graph", "labeling", "procedure"], "Method"], [3018, 3019, ["dG"], "Method"], [3021, 3023, ["distance", "measure"], "Method"], [3034, 3036, ["distance", "measure"], "Metric"], [3037, 3039, ["×kk", "matrices"], "Method"], [3050, 3053, ["graph", "labeling", "procedure"], "Method"], [3102, 3106, ["optimal", "graph", "normalization", "problem"], "Task"], [3112, 3115, ["graph", "canonicalization", "problem"], "Task"], [3117, 3120, ["canonical", "labeling", "algorithm"], "Method"], [3151, 3154, ["optimal", "normalization", "problem"], "Task"], [3178, 3179, ["ReceptiveField"], "Method"], [3207, 3208, ["complexity"], "Metric"], [3210, 3213, ["optimal", "normalization", "problem"], "Task"], [3217, 3220, ["Optimal", "graph", "normalization"], "Task"], [3232, 3234, ["subgraph", "isomorphism"], "Method"], [3236, 3239, ["Patchy", "-", "san"], "Method"], [3244, 3246, ["optimization", "problem"], "Task"], [3253, 3256, ["graph", "labeling", "methods"], "Method"], [3275, 3276, ["NormalizeGraph"], "Method"], [3277, 3279, ["Graph", "Normalization"], "Method"], [3520, 3522, ["labeling", "procedures"], "Method"], [3595, 3597, ["directed", "graphs"], "Task"], [3599, 3602, ["graph", "normalization", "problem"], "Task"], [3607, 3610, ["graph", "labeling", "procedures"], "Method"], [3612, 3617, ["normalization", "of", "local", "graph", "structures"], "Task"], [3628, 3632, ["Patchy", "-", "san", "framework"], "Method"], [3699, 3702, ["vector", "space", "representation"], "Method"], [3705, 3707, ["labeling", "methods"], "Method"], [3754, 3756, ["graph", "isomorphism"], "Task"], [3804, 3810, ["canonical", "labeling", "of", "the", "graph", "neigborhoods"], "Method"], [3821, 3823, ["normalization", "procedure"], "Method"], [3839, 3840, ["ranking"], "Method"], [3895, 3898, ["Patchy", "-", "san"], "Method"], [3899, 3900, ["CNNs"], "Method"], [3901, 3902, ["images"], "Material"], [3919, 3922, ["Patchy", "-", "san"], "Method"], [3940, 3944, ["1", "-", "WL", "normalization"], "Method"], [3962, 3964, ["first", "layer"], "Method"], [3966, 3967, ["CNN"], "Method"], [4015, 4018, ["square", "grid", "graph"], "Method"], [4027, 4029, ["Convolutional", "Architecture"], "Method"], [4029, 4032, ["Patchy", "-", "san"], "Method"], [4110, 4114, ["-", "dimensional", "convolutional", "layer"], "Method"], [4157, 4158, ["Complexity"], "Metric"], [4160, 4165, ["Patchy", "-", "san", "’s", "algorithm"], "Method"], [4166, 4169, ["creating", "receptive", "fields"], "Task"], [4244, 4247, ["Patchy", "-", "san"], "Method"], [4249, 4253, ["worst", "-", "case", "complexity"], "Metric"], [4287, 4290, ["Node", "sequence", "selection"], "Task"], [4308, 4313, ["creation", "of", "normalized", "graph", "patches"], "Task"], [4321, 4323, ["labeling", "procedure"], "Method"], [4361, 4365, ["worst", "-", "case", "complexity"], "Metric"], [4367, 4371, ["graph", "canonicalization", "algorithm", "Nauty"], "Method"], [4382, 4386, ["Weisfeiler", "-", "Lehman", "algorithm"], "Method"], [4390, 4391, ["complexity"], "Metric"], [4398, 4399, ["complexity"], "Metric"], [4400, 4403, ["Patchy", "-", "san"], "Method"], [4424, 4426, ["runtime", "analysis"], "Task"], [4439, 4441, ["graph", "kernels"], "Method"], [4442, 4445, ["benchmark", "data", "sets"], "Material"], [4454, 4456, ["Runtime", "Analysis"], "Task"], [4461, 4464, ["Patchy", "-", "san"], "Method"], [4468, 4472, ["real", "-", "world", "graphs"], "Material"], [4495, 4496, ["CNNs"], "Method"], [4497, 4498, ["learning"], "Method"], [4509, 4519, ["Python", "module", "graph", "-", "toolhttps:", "//", "graph", "-", "tool.skewed.de", "/"], "Method"], [4527, 4530, ["Patchy", "-", "san"], "Method"], [4540, 4551, ["-", "dimensional", "Weisfeiler", "-", "Lehman", "(", "1", "-", "WL", ")", "algorithm"], "Method"], [4553, 4554, ["normalization"], "Task"], [4555, 4556, ["torus"], "Method"], [4558, 4560, ["periodic", "lattice"], "Material"], [4563, 4564, ["random"], "Material"], [4566, 4569, ["random", "undirected", "graph"], "Method"], [4577, 4578, ["power"], "Material"], [4586, 4588, ["power", "grid"], "Material"], [4592, 4593, ["polbooks"], "Material"], [4597, 4601, ["purchasing", "network", "of", "books"], "Material"], [4602, 4604, ["US", "politics"], "Material"], [4607, 4609, ["presidential", "election"], "Material"], [4610, 4611, ["preferential"], "Method"], [4613, 4617, ["preferential", "attachment", "network", "model"], "Method"], [4624, 4627, ["astro", "-", "ph"], "Method"], [4629, 4631, ["coauthorship", "network"], "Method"], [4638, 4640, ["astrophysics", "arxiv"], "Material"], [4643, 4644, ["enron"], "Material"], [4646, 4648, ["communication", "network"], "Method"], [4696, 4699, ["Patchy", "-", "san"], "Method"], [4709, 4710, ["email"], "Method"], [4738, 4744, ["CNN", "with", "convolutional", "and", "dense", "layers"], "Method"], [4775, 4777, ["downstream", "CNN"], "Method"], [4780, 4782, ["Feature", "Visualization"], "Method"], [4783, 4784, ["visualization"], "Task"], [4792, 4794, ["popular", "models"], "Method"], [4797, 4803, ["restricted", "Boltzman", "machine", "(", "RBM", ")"], "Method"], [4807, 4810, ["Patchy", "-", "san"], "Method"], [4811, 4814, ["unsupervised", "feature", "learning"], "Task"], [4835, 4836, ["RBM"], "Method"], [4838, 4839, ["RBM"], "Method"], [4852, 4854, ["learning", "rate"], "Metric"], [4863, 4867, ["single", "-", "layer", "RBM"], "Method"], [4891, 4892, ["RBM"], "Method"], [4920, 4922, ["Graph", "Classification"], "Task"], [4922, 4924, ["Graph", "classification"], "Task"], [4928, 4930, ["assigning", "graphs"], "Task"], [4942, 4945, ["benchmark", "data", "sets"], "Material"], [4947, 4950, ["run", "-", "time"], "Metric"], [4951, 4953, ["classification", "accuracy"], "Metric"], [4957, 4960, ["art", "graph", "kernels"], "Method"], [4961, 4962, ["MUTAG"], "Method"], [4963, 4964, ["PCT"], "Material"], [4965, 4966, ["NCI1"], "Material"], [4967, 4968, ["NCI109"], "Material"], [4969, 4970, ["PROTEIN"], "Material"], [4972, 4976, ["D", "&", "D.", "MUTAG"], "Method"], [4997, 4998, ["PTC"], "Method"], [5012, 5013, ["NCI1"], "Method"], [5014, 5015, ["NCI109"], "Method"], [5022, 5033, ["non", "-", "small", "cell", "lung", "cancer", "and", "ovarian", "cancer", "cell", "lines"], "Task"], [5034, 5035, ["PROTEINS"], "Material"], [5060, 5061, ["Graphs"], "Method"], [5070, 5073, ["D", "&", "D"], "Method"], [5078, 5080, ["protein", "structures"], "Material"], [5095, 5098, ["Patchy", "-", "san"], "Method"], [5100, 5107, ["shortest", "-", "path", "kernel", "(", "SP", ")"], "Method"], [5109, 5112, ["random", "walk", "kernel"], "Method"], [5113, 5114, ["RW"], "Method"], [5117, 5120, ["graphlet", "count", "kernel"], "Method"], [5121, 5122, ["GK"], "Method"], [5126, 5131, ["Weisfeiler", "-", "Lehman", "subtree", "kernel"], "Method"], [5132, 5133, ["WL"], "Method"], [5169, 5174, ["-", "fold", "cross", "-", "validation"], "Method"], [5175, 5178, ["LIB", "-", "SVM"], "Method"], [5195, 5198, ["average", "prediction", "accuracies"], "Metric"], [5203, 5206, ["Patchy", "-", "san"], "Task"], [5210, 5211, ["PSCN"], "Method"], [5215, 5219, ["-", "dimensional", "WL", "normalization"], "Method"], [5271, 5273, ["merge", "layer"], "Method"], [5286, 5288, ["network", "architecture"], "Method"], [5290, 5292, ["convolutional", "layers"], "Method"], [5294, 5297, ["dense", "hidden", "layer"], "Method"], [5300, 5302, ["softmax", "layer"], "Method"], [5308, 5310, ["convolutional", "layer"], "Method"], [5320, 5322, ["conv", "layer"], "Method"], [5337, 5339, ["convolutional", "layers"], "Method"], [5340, 5343, ["rectified", "linear", "units"], "Method"], [5345, 5347, ["dense", "layer"], "Method"], [5348, 5351, ["rectified", "linear", "units"], "Method"], [5353, 5355, ["dropout", "rate"], "Metric"], [5357, 5358, ["Dropout"], "Method"], [5392, 5399, ["mini", "-", "batch", "gradient", "decent", "algorithm", "rmsprop"], "Method"], [5408, 5411, ["Theano", "wrapper", "Keras"], "Method"], [5416, 5418, ["logistic", "regression"], "Method"], [5419, 5422, ["PSLR", ")", "classifier"], "Method"], [5440, 5444, ["social", "graph", "data", "sets"], "Material"], [5459, 5462, ["Patchy", "-", "san"], "Method"], [5468, 5470, ["graphlet", "count"], "Method"], [5471, 5472, ["GK"], "Method"], [5475, 5479, ["deep", "graphlet", "count", "kernel"], "Method"], [5480, 5481, ["DGK"], "Method"], [5492, 5495, ["Patchy", "-", "san"], "Task"], [5527, 5528, ["NCI109"], "Method"], [5539, 5546, ["one", "-", "fits", "-", "all", "CNN", "architecture"], "Method"], [5548, 5550, ["CNNs", "accuracy"], "Metric"], [5555, 5557, ["graph", "kernels"], "Method"], [5571, 5573, ["classification", "accuracy"], "Metric"], [5595, 5597, ["CNNs", "hyperparameters"], "Method"], [5620, 5624, ["image", "and", "text", "data"], "Material"], [5627, 5630, ["Patchy", "-", "san"], "Method"], [5641, 5644, ["Patchy", "-", "san"], "Method"], [5654, 5656, ["graph", "kernel"], "Method"], [5657, 5658, ["WL"], "Method"], [5682, 5685, ["betweeness", "centrality", "normalization"], "Task"], [5692, 5693, ["runtime"], "Metric"], [5699, 5701, ["Logistic", "regression"], "Method"], [5703, 5709, ["Patchy", "-", "san", "’s", "receptive", "fields"], "Method"], [5714, 5717, ["Patchy", "-", "san"], "Method"], [5723, 5724, ["CNNs"], "Method"], [5739, 5742, ["Patchy", "-", "san"], "Method"], [5748, 5751, ["social", "graph", "data"], "Material"], [5795, 5798, ["learning", "graph", "representations"], "Task"], [5805, 5806, ["CNNs"], "Method"], [5833, 5837, ["local", "normalized", "neighborhood", "representations"], "Method"], [5858, 5860, ["graph", "kernels"], "Method"], [5870, 5873, ["neural", "network", "architectures"], "Method"], [5875, 5876, ["RNNs"], "Method"], [5883, 5884, ["pretraining"], "Method"], [5885, 5886, ["RBMs"], "Method"], [5887, 5888, ["autoencoders"], "Method"], [5890, 5893, ["statistical", "relational", "models"], "Method"]]} {"doc_id": "1f8f0abfe4689aa93f2f6cc7ec4fd4c6adc2c2d6", "text": ["document", ":", "Semantic", "Instance", "Segmentation", "with", "a", "Discriminative", "Loss", "Function", "Semantic", "instance", "segmentation", "remains", "a", "challenging", "task", ".", "In", "this", "work", "we", "propose", "to", "tackle", "the", "problem", "with", "a", "discriminative", "loss", "function", ",", "operating", "at", "the", "pixel", "level", ",", "that", "encourages", "a", "convolutional", "network", "to", "produce", "a", "representation", "of", "the", "image", "that", "can", "easily", "be", "clustered", "into", "instances", "with", "a", "simple", "post", "-", "processing", "step", ".", "The", "loss", "function", "encourages", "the", "network", "to", "map", "each", "pixel", "to", "a", "point", "in", "feature", "space", "so", "that", "pixels", "belonging", "to", "the", "same", "instance", "lie", "close", "together", "while", "different", "instances", "are", "separated", "by", "a", "wide", "margin", ".", "Our", "approach", "of", "combining", "an", "off", "-", "the", "-", "shelf", "network", "with", "a", "principled", "loss", "function", "inspired", "by", "a", "metric", "learning", "objective", "is", "conceptually", "simple", "and", "distinct", "from", "recent", "efforts", "in", "instance", "segmentation", ".", "In", "contrast", "to", "previous", "works", ",", "our", "method", "does", "not", "rely", "on", "object", "proposals", "or", "recurrent", "mechanisms", ".", "A", "key", "contribution", "of", "our", "work", "is", "to", "demonstrate", "that", "such", "a", "simple", "setup", "without", "bells", "and", "whistles", "is", "effective", "and", "can", "perform", "on", "-", "par", "with", "more", "complex", "methods", ".", "Moreover", ",", "we", "show", "that", "it", "does", "not", "suffer", "from", "some", "of", "the", "limitations", "of", "the", "popular", "detect", "-", "and", "-", "segment", "approaches", ".", "We", "achieve", "competitive", "performance", "on", "the", "Cityscapes", "and", "CVPPP", "leaf", "segmentation", "benchmarks", ".", "section", ":", "Introduction", "Semantic", "instance", "segmentation", "has", "recently", "gained", "in", "popularity", ".", "As", "an", "extension", "of", "regular", "semantic", "segmentation", ",", "the", "task", "is", "to", "generate", "a", "binary", "segmentation", "mask", "for", "each", "individual", "object", "along", "with", "a", "semantic", "label", ".", "It", "is", "considered", "a", "fundamentally", "harder", "problem", "than", "semantic", "segmentation", "-", "where", "overlapping", "objects", "of", "the", "same", "class", "are", "segmented", "as", "one", "-", "and", "is", "closely", "related", "to", "the", "tasks", "of", "object", "counting", "and", "object", "detection", ".", "One", "could", "also", "say", "instance", "segmentation", "is", "a", "generalization", "of", "object", "detection", ",", "with", "the", "goal", "of", "producing", "a", "segmentation", "mask", "rather", "than", "a", "bounding", "box", "for", "each", "object", ".", "Pinheiro", "obtain", "bounding", "boxes", "from", "instance", "segmentations", "by", "simply", "drawing", "the", "tightest", "bounding", "box", "around", "each", "segmentation", "mask", ",", "and", "show", "that", "their", "system", "reaches", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "an", "object", "detection", "benchmark", ".", "The", "relation", "between", "instance", "segmentation", "and", "semantic", "segmentation", "is", "less", "clear", ".", "Intuitively", ",", "the", "two", "tasks", "feel", "very", "closely", "related", ",", "but", "it", "turns", "out", "not", "to", "be", "obvious", "how", "to", "apply", "the", "network", "architectures", "and", "loss", "functions", "that", "are", "successful", "in", "semantic", "segmentation", "to", "this", "related", "instance", "task", ".", "One", "key", "factor", "that", "complicates", "the", "naive", "application", "of", "the", "popular", "softmax", "cross", "-", "entropy", "loss", "function", "to", "instance", "segmentation", ",", "is", "the", "fact", "that", "an", "image", "can", "contain", "an", "arbitrary", "number", "of", "instances", "and", "that", "the", "labeling", "is", "permutation", "-", "invariant", ":", "it", "does", "not", "matter", "which", "specific", "label", "an", "instance", "gets", ",", "as", "long", "as", "it", "is", "different", "from", "all", "other", "instance", "labels", ".", "One", "possible", "solution", "is", "to", "set", "an", "upper", "limit", "to", "the", "number", "of", "detectable", "instances", "and", "to", "impose", "extra", "constraints", "on", "the", "labeling", ",", "but", "this", "may", "unnecessarily", "limit", "the", "representational", "power", "of", "the", "network", "and", "introduce", "unwanted", "biases", ",", "leading", "to", "unsatisfying", "results", ".", "Most", "recent", "works", "on", "instance", "segmentation", "with", "deep", "networks", "go", "a", "different", "route", ".", "Two", "popular", "approaches", "introduce", "a", "multi", "-", "stage", "pipeline", "with", "object", "proposals", ",", "or", "train", "a", "recurrent", "network", "end", "-", "to", "-", "end", "with", "a", "custom", "loss", "function", "that", "outputs", "instances", "sequentially", ".", "Another", "line", "of", "research", "is", "to", "train", "a", "network", "to", "transform", "the", "image", "into", "a", "representation", "that", "is", "clustered", "into", "individual", "instances", "with", "a", "post", "-", "processing", "step", ".", "Our", "method", "belongs", "to", "this", "last", "category", ",", "but", "takes", "a", "more", "principled", "(", "less", "ad", "-", "hoc", ")", "approach", "than", "previous", "works", "and", "reduces", "the", "post", "-", "processing", "step", "to", "a", "minimum", ".", "Inspired", "by", "the", "success", "of", "siamese", "networks", "and", "the", "triplet", "loss", "in", "image", "classification", ",", "we", "introduce", "a", "discriminative", "loss", "function", "to", "replace", "the", "pixel", "-", "wise", "softmax", "loss", "that", "is", "commonly", "used", "in", "semantic", "segmentation", ".", "Our", "loss", "function", "enforces", "the", "network", "to", "map", "each", "pixel", "in", "the", "image", "to", "an", "n", "-", "dimensional", "vector", "in", "feature", "space", ",", "such", "that", "feature", "vectors", "of", "pixels", "that", "belong", "to", "the", "same", "instance", "lie", "close", "together", "while", "feature", "vectors", "of", "pixels", "that", "belong", "to", "different", "instances", "lie", "far", "apart", ".", "The", "output", "of", "the", "network", "can", "easily", "be", "clustered", "with", "a", "fast", "and", "simple", "post", "-", "processing", "operation", ".", "With", "this", "mechanism", ",", "we", "optimize", "an", "objective", "that", "avoids", "the", "aforementioned", "problems", "related", "to", "variable", "number", "of", "instances", "and", "permutation", "-", "invariance", ".", "Our", "work", "mainly", "focuses", "on", "the", "loss", "function", ",", "as", "we", "aim", "to", "be", "able", "to", "re", "-", "use", "network", "architectures", "that", "were", "designed", "for", "semantic", "segmentation", ":", "we", "plug", "in", "an", "off", "-", "the", "-", "shelf", "architecture", "and", "retrain", "the", "system", "with", "our", "discriminative", "loss", "function", ".", "In", "our", "framework", ",", "the", "tasks", "of", "semantic", "and", "instance", "segmentation", "can", "be", "treated", "in", "a", "consistent", "and", "similar", "manner", "and", "do", "not", "require", "changes", "on", "the", "architecture", "side", ".", "The", "rest", "of", "this", "paper", "is", "structured", "as", "follows", ".", "First", "we", "give", "an", "extensive", "overview", "of", "the", "related", "work", "in", "section", "[", "reference", "]", ".", "In", "section", "[", "reference", "]", "we", "discuss", "our", "proposed", "method", "in", "detail", ".", "In", "section", "[", "reference", "]", "we", "set", "up", "experiments", "on", "two", "instance", "segmentation", "benchmarks", "and", "show", "that", "we", "get", "a", "performance", "that", "is", "competitive", "with", "the", "state", "-", "of", "-", "the", "-", "art", ".", "section", ":", "Related", "Work", "In", "the", "last", "few", "years", ",", "deep", "networks", "have", "achieved", "impressive", "results", "in", "semantic", "and", "instance", "segmentation", ".", "All", "top", "-", "performing", "methods", "across", "different", "benchmarks", "use", "a", "deep", "network", "in", "their", "pipeline", ".", "Here", "we", "discuss", "these", "prior", "works", "and", "situate", "our", "model", "between", "them", ".", "Proposal", "-", "based", "Many", "instance", "segmentation", "approaches", "build", "a", "multi", "-", "stage", "pipeline", "with", "a", "separate", "object", "proposal", "and", "classification", "step", ".", "Hariharan", "and", "Chen", "use", "MCG", "to", "generate", "category", "-", "independent", "region", "proposals", ",", "followed", "by", "a", "classification", "step", ".", "Pinheiro", "use", "the", "same", "general", "approach", ",", "but", "their", "work", "focuses", "on", "generating", "segmentation", "proposals", "with", "a", "deep", "network", ".", "Dai", "won", "the", "2015", "MS", "-", "COCO", "instance", "segmentation", "challenge", "with", "a", "cascade", "of", "networks", "(", "MNC", ")", "to", "merge", "bounding", "boxes", ",", "segmentation", "masks", "and", "category", "information", ".", "Many", "works", "were", "inspired", "by", "this", "approach", "and", "also", "combine", "an", "object", "detector", "with", "a", "semantic", "segmentation", "network", "to", "produce", "instances", ".", "In", "contrast", "to", "these", "works", ",", "our", "method", "does", "not", "rely", "on", "object", "proposals", "or", "bounding", "boxes", "but", "treats", "the", "image", "holistically", ",", "which", "we", "show", "to", "be", "beneficial", "for", "handling", "certain", "tasks", "with", "complex", "occlusions", "as", "discussed", "in", "section", "[", "reference", "]", ".", "Recurrent", "methods", "Other", "recent", "works", "employ", "recurrent", "networks", "to", "generate", "the", "individual", "instances", "sequentially", ".", "Stewart", "train", "a", "network", "for", "end", "-", "to", "-", "end", "object", "detection", "using", "an", "LSTM", ".", "Their", "loss", "function", "is", "permutation", "-", "invariant", "as", "it", "incorporates", "the", "Hungarian", "algorithm", "to", "match", "candidate", "hypotheses", "to", "ground", "-", "truth", "instances", ".", "Inspired", "by", "their", "work", ",", "Romera", "propose", "an", "end", "-", "to", "-", "end", "recurrent", "network", "with", "convolutional", "LSTMs", "that", "sequentially", "outputs", "binary", "segmentation", "maps", "for", "each", "instance", ".", "Ren", "improve", "upon", "by", "adding", "a", "box", "network", "to", "confine", "segmentations", "within", "a", "local", "window", "and", "skip", "connections", "instead", "of", "graphical", "models", "to", "restore", "the", "resolution", "at", "the", "output", ".", "Their", "final", "framework", "consists", "of", "four", "major", "components", ":", "an", "external", "memory", "and", "networks", "for", "box", "proposal", ",", "segmentation", "and", "scoring", ".", "We", "argue", "that", "our", "proposed", "method", "is", "conceptually", "simpler", "and", "easier", "to", "implement", "than", "these", "methods", ".", "Our", "method", "does", "not", "involve", "recurrent", "mechanisms", "and", "can", "work", "with", "any", "off", "-", "the", "-", "shelf", "segmentation", "architecture", ".", "Moreover", ",", "our", "loss", "function", "is", "permutation", "-", "invariant", "by", "design", ",", "without", "the", "need", "to", "resort", "to", "a", "Hungarian", "algorithm", ".", "Clustering", "Another", "approach", "is", "to", "transform", "the", "image", "into", "a", "representation", "that", "is", "subsequently", "clustered", "into", "discrete", "instances", ".", "Silberman", "produce", "a", "segmentation", "tree", "and", "use", "a", "coverage", "loss", "to", "cut", "it", "into", "non", "-", "overlapping", "regions", ".", "Zhang", "impose", "an", "ordering", "on", "the", "individual", "instances", "based", "on", "their", "depth", ",", "and", "use", "a", "MRF", "to", "merge", "overlapping", "predicted", "patches", "into", "a", "coherent", "segmentation", ".", "Two", "earlier", "works", "also", "use", "depth", "information", "to", "segment", "instances", ".", "Uhrig", "train", "a", "network", "to", "predict", "each", "pixel", "’s", "direction", "towards", "its", "instance", "center", ",", "along", "with", "monocular", "depth", "and", "semantic", "labels", ".", "They", "use", "template", "matching", "and", "proposal", "fusion", "techniques", "to", "extract", "the", "individual", "instances", "from", "this", "representation", ".", "Liang", "predict", "pixel", "-", "wise", "feature", "vectors", "representing", "the", "ground", "truth", "bounding", "box", "of", "the", "instance", "it", "belongs", "to", ".", "With", "the", "help", "of", "a", "sub", "-", "network", "that", "predicts", "an", "object", "count", ",", "they", "cluster", "the", "output", "of", "the", "network", "into", "individual", "instances", ".", "Our", "work", "is", "similar", "to", "these", "works", "in", "that", "we", "have", "a", "separate", "clustering", "step", ",", "but", "our", "loss", "does", "not", "constrain", "the", "output", "of", "the", "network", "to", "a", "specific", "representation", "like", "instance", "center", "coordinates", "or", "depth", "ordering", ";", "it", "is", "less", "ad", "-", "hoc", "in", "that", "sense", ".", "Other", "Bai", "use", "deep", "networks", "to", "directly", "learn", "the", "energy", "of", "the", "watershed", "transform", ".", "A", "drawback", "of", "this", "bottom", "-", "up", "approach", "is", "that", "they", "can", "not", "handle", "occlusions", "where", "instances", "are", "separated", "into", "multiple", "pieces", ".", "Kirillov", "use", "a", "CRF", ",", "but", "with", "a", "novel", "MultiCut", "formulation", "to", "combine", "semantic", "segmentations", "with", "edge", "maps", "to", "extract", "instances", "as", "connected", "regions", ".", "A", "shortcoming", "of", "this", "method", "is", "that", ",", "although", "they", "reason", "globally", "about", "instances", ",", "they", "also", "can", "not", "handle", "occlusions", ".", "Arnab", "combine", "an", "object", "detector", "with", "a", "semantic", "segmentation", "module", "using", "a", "CRF", "model", ".", "By", "considering", "the", "image", "holistically", "it", "can", "handle", "occlusions", "and", "produce", "more", "precise", "segmentations", ".", "Loss", "function", "Our", "loss", "function", "is", "inspired", "by", "earlier", "works", "on", "distance", "metric", "learning", ",", "discriminative", "loss", "functions", "and", "siamese", "networks", ".", "Most", "similar", "to", "our", "loss", "function", ",", "Weinberger", "propose", "to", "learn", "a", "distance", "metric", "for", "large", "margin", "nearest", "neighbor", "classification", ".", "Kostinger", "further", "explore", "a", "similar", "LDA", "based", "objective", ".", "More", "recently", "Schroff", ",", "building", "on", "Sun", ",", "introduced", "the", "triplet", "loss", "for", "face", "recognition", ".", "The", "triplet", "loss", "enforces", "a", "margin", "between", "each", "pair", "of", "faces", "from", "one", "person", ",", "to", "all", "other", "faces", ".", "Xie", "propose", "a", "clustering", "objective", "for", "unsupervised", "learning", ".", "Whereas", "these", "works", "employ", "a", "discriminative", "loss", "function", "to", "optimize", "distances", "between", "images", "in", "a", "dataset", ",", "our", "method", "operates", "at", "the", "pixel", "level", ",", "optimizing", "distances", "between", "individual", "pixels", "in", "an", "image", ".", "To", "our", "knowledge", ",", "we", "are", "the", "first", "to", "successfully", "use", "a", "discriminative", "loss", "based", "on", "distance", "metric", "learning", "principles", "for", "the", "task", "of", "instance", "segmentation", "with", "deep", "networks", ".", "section", ":", "Method", "subsection", ":", "Discriminative", "loss", "function", "Consider", "a", "differentiable", "function", "that", "maps", "each", "pixel", "in", "an", "input", "image", "to", "a", "point", "in", "n", "-", "dimensional", "feature", "space", ",", "referred", "to", "as", "the", "pixel", "embedding", ".", "The", "intuition", "behind", "our", "loss", "function", "is", "that", "embeddings", "with", "the", "same", "label", "(", "same", "instance", ")", "should", "end", "up", "close", "together", ",", "while", "embeddings", "with", "a", "different", "label", "(", "different", "instance", ")", "should", "end", "up", "far", "apart", ".", "Weinberger", "propose", "a", "loss", "function", "with", "two", "competing", "terms", "to", "achieve", "this", "objective", ":", "a", "term", "to", "penalize", "large", "distances", "between", "embeddings", "with", "the", "same", "label", ",", "and", "a", "term", "to", "penalize", "small", "distances", "between", "embeddings", "with", "a", "different", "label", ".", "In", "our", "loss", "function", "we", "keep", "the", "first", "term", ",", "but", "replace", "the", "second", "term", "with", "a", "more", "tractable", "one", ":", "instead", "of", "directly", "penalizing", "small", "distances", "between", "every", "pair", "of", "differently", "-", "labeled", "embeddings", ",", "we", "only", "penalize", "small", "distances", "between", "the", "mean", "embeddings", "of", "different", "labels", ".", "If", "the", "number", "of", "different", "labels", "is", "smaller", "than", "the", "number", "of", "inputs", ",", "this", "is", "computationally", "much", "cheaper", "than", "calculating", "the", "distances", "between", "every", "pair", "of", "embeddings", ".", "This", "is", "a", "valid", "assumption", "for", "instance", "segmentation", ",", "where", "there", "are", "orders", "of", "magnitude", "fewer", "instances", "than", "pixels", "in", "an", "image", ".", "We", "now", "formulate", "our", "discriminative", "loss", "in", "terms", "of", "push", "(", "i.e.", "repelling", ")", "and", "pull", "forces", "between", "and", "within", "clusters", ".", "A", "cluster", "is", "defined", "as", "a", "group", "of", "pixel", "embeddings", "sharing", "the", "same", "label", ",", "e.g.", "pixels", "belonging", "to", "the", "same", "instance", ".", "Our", "loss", "consists", "of", "three", "terms", ":", "variance", "term", ":", "an", "intra", "-", "cluster", "pull", "-", "force", "that", "draws", "embeddings", "towards", "the", "mean", "embedding", ",", "i.e.", "the", "cluster", "center", ".", "distance", "term", ":", "an", "inter", "-", "cluster", "push", "-", "force", "that", "pushes", "clusters", "away", "from", "each", "other", ",", "increasing", "the", "distance", "between", "the", "cluster", "centers", ".", "regularization", "term", ":", "a", "small", "pull", "-", "force", "that", "draws", "all", "clusters", "towards", "the", "origin", ",", "to", "keep", "the", "activations", "bounded", ".", "The", "variance", "and", "distance", "terms", "are", "hinged", ":", "their", "forces", "are", "only", "active", "up", "to", "a", "certain", "distance", ".", "Embeddings", "within", "a", "distance", "of", "from", "their", "cluster", "centers", "are", "no", "longer", "attracted", "to", "it", ",", "which", "means", "that", "they", "can", "exist", "on", "a", "local", "manifold", "in", "feature", "space", "rather", "than", "having", "to", "converge", "to", "a", "single", "point", ".", "Analogously", ",", "cluster", "centers", "further", "apart", "than", "are", "no", "longer", "repulsed", "and", "can", "move", "freely", "in", "feature", "space", ".", "Hinging", "the", "forces", "relaxes", "the", "constraints", "on", "the", "network", ",", "giving", "it", "more", "representational", "power", "to", "achieve", "its", "goal", ".", "The", "interacting", "forces", "in", "feature", "space", "are", "illustrated", "in", "figure", "[", "reference", "]", ".", "The", "loss", "function", "can", "also", "be", "written", "down", "exactly", ".", "We", "use", "the", "following", "definitions", ":", "is", "the", "number", "of", "clusters", "in", "the", "ground", "truth", ",", "is", "the", "number", "of", "elements", "in", "cluster", ",", "is", "an", "embedding", ",", "is", "the", "mean", "embedding", "of", "cluster", "(", "the", "cluster", "center", ")", ",", "is", "the", "L1", "or", "L2", "distance", ",", "and", "denotes", "the", "hinge", ".", "and", "are", "respectively", "the", "margins", "for", "the", "variance", "and", "distance", "loss", ".", "The", "loss", "can", "then", "be", "written", "as", "follows", ":", "In", "our", "experiments", "we", "set", "and", ".", "The", "loss", "is", "minimized", "by", "stochastic", "gradient", "descent", ".", "Comparison", "with", "softmax", "loss", "We", "discuss", "the", "relation", "of", "our", "loss", "function", "with", "the", "popular", "pixel", "-", "wise", "multi", "-", "class", "cross", "-", "entropy", "loss", ",", "often", "referred", "to", "as", "the", "softmax", "loss", ".", "In", "the", "case", "of", "a", "softmax", "loss", "with", "classes", ",", "each", "pixel", "embedding", "is", "driven", "to", "a", "one", "-", "hot", "vector", ",", "i.e.", "the", "unit", "vector", "on", "one", "of", "the", "axes", "of", "an", "-", "dimensional", "feature", "space", ".", "Because", "the", "softmax", "function", "has", "the", "normalizing", "property", "that", "its", "outputs", "are", "positive", "and", "sum", "to", "one", ",", "the", "embeddings", "are", "restricted", "to", "lie", "on", "the", "unit", "simplex", ".", "When", "the", "loss", "reaches", "zero", ",", "all", "embeddings", "lie", "on", "one", "of", "the", "unit", "vectors", ".", "By", "design", ",", "the", "dimensions", "of", "the", "output", "feature", "space", "(", "which", "correspond", "to", "the", "number", "of", "feature", "maps", "in", "the", "last", "layer", "of", "the", "network", ")", "must", "be", "equal", "to", "the", "number", "of", "classes", ".", "To", "add", "a", "class", "after", "training", ",", "the", "architecture", "needs", "to", "be", "updated", "too", ".", "In", "comparison", ",", "our", "loss", "function", "does", "not", "drive", "the", "embeddings", "to", "a", "specific", "point", "in", "feature", "space", ".", "The", "network", "could", "for", "example", "place", "similar", "clusters", "(", "e.g.", "two", "small", "objects", ")", "closer", "together", "than", "dissimilar", "ones", "(", "e.g.", "a", "small", "and", "a", "large", "object", ")", ".", "When", "the", "loss", "reaches", "zero", ",", "the", "system", "of", "push", "and", "pull", "forces", "has", "minimal", "energy", "and", "the", "clusters", "have", "organized", "themselves", "in", "n", "-", "dimensional", "space", ".", "Most", "importantly", ",", "the", "dimensionality", "of", "the", "feature", "space", "is", "independent", "of", "the", "number", "of", "instances", "that", "needs", "to", "be", "segmented", ".", "Figure", "[", "reference", "]", "depicts", "the", "convergence", "of", "our", "loss", "function", "when", "overfitting", "on", "a", "single", "image", "with", "instances", ",", "in", "a", "2", "-", "dimensional", "feature", "space", ".", "subsection", ":", "Post", "-", "processing", "When", "the", "variance", "and", "distance", "terms", "of", "the", "loss", "are", "zero", ",", "the", "following", "is", "true", ":", "all", "embeddings", "are", "within", "a", "distance", "of", "from", "their", "cluster", "center", "all", "cluster", "centers", "are", "at", "least", "apart", "If", ",", "then", "each", "embedding", "is", "closer", "to", "its", "own", "cluster", "center", "than", "to", "any", "other", "cluster", "center", ".", "It", "follows", "that", "during", "inference", ",", "we", "can", "threshold", "with", "a", "bandwidth", "around", "a", "cluster", "center", "to", "select", "all", "embeddings", "belonging", "to", "that", "cluster", ".", "Thresholding", "in", "this", "case", "means", "selecting", "all", "embeddings", "that", "lie", "within", "a", "hypersphere", "with", "radius", "around", "the", "cluster", "center", ":", "For", "the", "tasks", "of", "classification", "and", "semantic", "segmentation", ",", "with", "a", "fixed", "set", "of", "classes", ",", "this", "leads", "to", "a", "simple", "strategy", "for", "post", "-", "processing", "the", "output", "of", "the", "network", "into", "discrete", "classes", ":", "after", "training", ",", "calculate", "the", "cluster", "centers", "of", "each", "class", "over", "the", "entire", "training", "set", ".", "During", "inference", ",", "threshold", "around", "each", "of", "the", "cluster", "centers", "to", "select", "all", "pixels", "belonging", "to", "the", "corresponding", "semantic", "class", ".", "This", "requires", "that", "the", "cluster", "centers", "of", "a", "specific", "class", "are", "the", "same", "in", "each", "image", ",", "which", "can", "be", "accomplished", "by", "coupling", "the", "cluster", "centers", "across", "a", "mini", "-", "batch", ".", "For", "the", "task", "of", "instance", "segmentation", "things", "are", "more", "complicated", ".", "As", "the", "labeling", "is", "permutation", "invariant", ",", "we", "can", "not", "simply", "record", "cluster", "centers", "and", "threshold", "around", "them", "during", "inference", ".", "We", "could", "follow", "a", "different", "strategy", ":", "if", "we", "set", ",", "then", "each", "embedding", "is", "closer", "to", "all", "embeddings", "of", "its", "own", "cluster", "than", "to", "any", "embedding", "of", "a", "different", "cluster", ".", "It", "follows", "that", "we", "can", "threshold", "around", "any", "embedding", "to", "select", "all", "embeddings", "belonging", "to", "the", "same", "cluster", ".", "The", "procedure", "during", "inference", "is", "to", "select", "an", "unlabeled", "pixel", ",", "threshold", "around", "its", "embedding", "to", "find", "all", "pixels", "belonging", "to", "the", "same", "instance", ",", "and", "assign", "them", "all", "the", "same", "label", ".", "Then", "select", "another", "pixel", "that", "does", "not", "yet", "belong", "to", "an", "instance", "and", "repeat", "until", "all", "pixels", "are", "labeled", ".", "Increasing", "robustness", "In", "a", "real", "-", "world", "problem", "the", "loss", "on", "the", "test", "set", "will", "not", "be", "zero", ",", "potentially", "causing", "our", "clustering", "algorithm", "for", "instance", "segmentation", "to", "make", "mistakes", ".", "If", "a", "cluster", "is", "not", "compact", "and", "we", "accidentally", "select", "an", "outlier", "to", "threshold", "around", ",", "it", "could", "happen", "that", "a", "real", "cluster", "gets", "predicted", "as", "two", "sub", "-", "clusters", ".", "To", "avoid", "this", "issue", ",", "we", "make", "the", "clustering", "more", "robust", "against", "outliers", "by", "applying", "a", "fast", "variant", "of", "the", "mean", "-", "shift", "algorithm", ".", "As", "before", ",", "we", "select", "a", "random", "unlabeled", "pixel", "and", "threshold", "around", "its", "embedding", ".", "Next", "however", ",", "we", "calculate", "the", "mean", "of", "the", "selected", "group", "of", "embeddings", "and", "use", "the", "mean", "to", "threshold", "again", ".", "We", "repeat", "this", "process", "until", "mean", "convergence", ".", "This", "has", "the", "effect", "of", "moving", "to", "a", "high", "-", "density", "area", "in", "feature", "space", ",", "likely", "corresponding", "to", "a", "true", "cluster", "center", ".", "In", "the", "experiments", "section", ",", "we", "investigate", "the", "effect", "of", "this", "clustering", "algorithm", "by", "comparing", "against", "ground", "truth", "clustering", ",", "where", "the", "thresholding", "targets", "are", "calculated", "as", "an", "average", "embedding", "over", "the", "ground", "truth", "instance", "labels", ".", "subsection", ":", "Pros", "and", "cons", "Our", "proposed", "method", "has", "some", "distinctive", "advantages", "and", "disadvantages", "compared", "to", "other", "methods", "that", "we", "now", "discuss", ".", "One", "big", "limitation", "of", "detect", "-", "and", "-", "segment", "approaches", "that", "is", "not", "immediately", "apparent", "from", "their", "excellent", "results", "on", "popular", "benchmarks", ",", "is", "that", "they", "rely", "on", "the", "assumption", "that", "an", "object", "’s", "segmentation", "mask", "can", "be", "unambiguously", "extracted", "from", "its", "bounding", "box", ".", "This", "is", "an", "implicit", "prior", "that", "is", "very", "effective", "for", "datasets", "like", "MS", "COCO", "and", "Pascal", "VOC", ",", "which", "contain", "relatively", "blobby", "objects", "that", "do", "not", "occlude", "each", "other", "in", "complex", "ways", ".", "However", ",", "the", "assumption", "is", "problematic", "for", "tasks", "where", "an", "object", "’s", "bounding", "box", "conveys", "insufficient", "information", "to", "recover", "the", "object", "’s", "segmentation", "mask", ".", "Consider", "the", "synthetic", "scattered", "sticks", "dataset", "shown", "in", "figure", "[", "reference", "]", "as", "an", "example", "to", "illustrate", "the", "issue", ".", "When", "two", "sticks", "overlap", "like", "two", "crossed", "swords", ",", "their", "bounding", "boxes", "are", "highly", "overlapping", ".", "Given", "only", "a", "detection", "in", "the", "form", "of", "a", "bounding", "box", ",", "it", "is", "exceedingly", "hard", "to", "unambigously", "extract", "a", "segmentation", "mask", "of", "the", "indicated", "object", ".", "Methods", "that", "rely", "on", "bounding", "boxes", "in", "their", "pipeline", "all", "suffer", "from", "this", "issue", ".", "In", "contrast", ",", "our", "method", "can", "handle", "such", "complex", "occlusions", "without", "problems", "as", "it", "treats", "the", "image", "holistically", "and", "learns", "to", "reason", "about", "occlusions", ",", "but", "does", "not", "employ", "a", "computationally", "expensive", "CRF", "like", ".", "Many", "real", "-", "world", "industrial", "or", "medical", "applications", "(", "conveyor", "belt", "sorting", "systems", ",", "overlapping", "cell", "and", "chromosome", "segmentation", ",", "etc", ".", ")", "exhibit", "this", "kind", "of", "occlusions", ".", "To", "the", "best", "of", "our", "knowledge", "no", "sufficiently", "large", "datasets", "for", "such", "tasks", "are", "publicly", "available", ",", "which", "unfortunately", "prevents", "us", "from", "showcasing", "this", "particular", "strength", "of", "our", "method", "to", "the", "full", ".", "On", "the", "other", "hand", ",", "our", "method", "also", "has", "some", "drawbacks", ".", "Due", "to", "the", "holistic", "treatment", "of", "the", "image", ",", "our", "method", "performs", "well", "on", "datasets", "with", "a", "lot", "of", "similarity", "across", "the", "images", "(", "traffic", "scenes", "in", "Cityscapes", "or", "leaf", "configurations", "in", "CVPPP", ")", ",", "but", "underperforms", "on", "datasets", "where", "objects", "can", "appear", "in", "random", "constellations", "and", "diverse", "settings", ",", "like", "Pascal", "VOC", "and", "MSCOCO", ".", "A", "sliding", "-", "window", "detection", "-", "based", "approach", "with", "non", "-", "max", "suppression", "is", "more", "suited", "for", "such", "datasets", ".", "For", "example", ",", "if", "our", "method", "were", "trained", "on", "images", "with", "only", "one", "object", ",", "it", "would", "perform", "badly", "on", "an", "image", "that", "unexpectedly", "contained", "many", "of", "these", "objects", ".", "A", "detection", "-", "based", "approach", "has", "no", "trouble", "with", "this", ".", "section", ":", "Experiments", "We", "test", "our", "loss", "function", "on", "two", "instance", "segmentation", "datasets", ":", "The", "CVPPP", "Leaf", "Segmentation", "dataset", "and", "the", "Cityscapes", "Instance", "-", "Level", "Semantic", "Labeling", "Task", ".", "These", "datasets", "contain", "a", "median", "number", "of", "more", "than", "15", "instances", "per", "image", ".", "We", "also", "study", "the", "influence", "of", "the", "different", "components", "of", "our", "method", "and", "point", "out", "where", "there", "is", "room", "for", "improvement", ".", "subsection", ":", "Datasets", "The", "LSC", "competition", "of", "the", "CVPPP", "workshop", "is", "a", "small", "but", "challenging", "benchmark", ".", "The", "task", "is", "to", "individually", "segment", "each", "leaf", "of", "a", "plant", ".", "The", "dataset", "was", "developed", "to", "encourage", "the", "use", "of", "computer", "vision", "methods", "to", "aid", "in", "the", "study", "of", "plant", "phenotyping", ".", "We", "use", "the", "A1", "subset", "which", "consists", "of", "128", "labeled", "images", "and", "33", "test", "images", ".", "gives", "an", "overview", "of", "results", "on", "this", "dataset", ".", "We", "compare", "our", "performance", "with", "some", "of", "these", "works", "and", "two", "other", "recent", "approaches", ".", "We", "report", "two", "metrics", "defined", "in", ":", "Symmetric", "Best", "Dice", "(", ")", ",", "which", "denotes", "the", "accuracy", "of", "the", "instance", "segmentation", "and", "Absolute", "Difference", "in", "Count", "(", ")", "which", "is", "the", "absolute", "value", "of", "the", "mean", "of", "the", "difference", "between", "the", "predicted", "number", "of", "leaves", "and", "the", "ground", "truth", "over", "all", "images", ".", "The", "large", "-", "scale", "Cityscapes", "dataset", "focuses", "on", "semantic", "understanding", "of", "urban", "street", "scenes", ".", "It", "has", "a", "benchmark", "for", "pixel", "-", "level", "and", "instance", "-", "level", "semantic", "segmentation", ".", "We", "test", "our", "method", "on", "the", "latter", ",", "using", "only", "the", "fine", "-", "grained", "annotations", ".", "The", "dataset", "is", "split", "up", "in", "2975", "training", "images", ",", "500", "validation", "images", ",", "and", "1525", "test", "images", ".", "We", "tune", "hyperparameters", "using", "the", "validation", "set", "and", "only", "use", "the", "train", "set", "to", "train", "our", "final", "model", ".", "We", "compare", "our", "results", "with", "the", "published", "works", "in", "the", "official", "leaderboard", ".", "We", "report", "accuracy", "using", "4", "metrics", "defined", "in", ":", "mean", "Average", "Precision", "(", "AP", ")", ",", "mean", "Average", "Precision", "with", "overlap", "of", "50", "%", "(", "AP0.5", ")", ",", "AP50", "m", "and", "AP100", "m", ",", "where", "evaluation", "is", "restricted", "to", "objects", "within", "50", "m", "and", "100", "m", "distance", ",", "respectively", ".", "subsection", ":", "Setup", "Model", "architecture", "and", "general", "setup", "Since", "we", "want", "to", "stress", "the", "fact", "that", "our", "loss", "can", "be", "used", "with", "an", "off", "-", "the", "-", "shelf", "network", ",", "we", "use", "the", "ResNet", "-", "38", "network", ",", "designed", "for", "semantic", "segmentation", ".", "We", "finetune", "from", "a", "model", "that", "was", "pre", "-", "trained", "on", "CityScapes", "semantic", "segmentation", ".", "In", "the", "following", "experiments", ",", "all", "models", "are", "trained", "using", "Adam", ",", "with", "a", "learning", "rate", "of", "1e", "-", "4", "on", "a", "NVidia", "Titan", "X", "GPU", ".", "Leaf", "segmentation", "Since", "this", "dataset", "only", "consists", "of", "128", "images", ",", "we", "use", "online", "data", "augmentation", "to", "prevent", "the", "model", "from", "overfitting", "and", "to", "increase", "the", "overall", "robustness", ".", "We", "apply", "random", "left", "-", "right", "flip", ",", "random", "rotation", "with", "and", "random", "scale", "deformation", "with", ".", "All", "images", "are", "rescaled", "to", "512x512", "and", "concatenated", "with", "an", "x", "-", "and", "y", "-", "coordinate", "map", "with", "values", "between", "-", "1", "and", "1", ".", "We", "train", "the", "network", "with", "margins", ",", ",", "and", "16", "output", "dimensions", ".", "Foreground", "masks", "are", "included", "with", "the", "test", "set", ",", "since", "this", "challenge", "only", "focuses", "on", "instance", "segmentation", ".", "Cityscapes", "Our", "final", "model", "is", "trained", "on", "the", "training", "images", ",", "downsampled", "to", ".", "Because", "of", "the", "size", "and", "variability", "of", "the", "dataset", ",", "there", "is", "no", "need", "for", "extra", "data", "augmentation", ".", "We", "train", "the", "network", "with", "margins", ",", ",", "and", "8", "output", "dimensions", ".", "In", "contrast", "to", "the", "CVPPP", "dataset", ",", "Cityscapes", "is", "a", "multi", "-", "class", "instance", "segmentation", "challenge", ".", "Therefore", ",", "we", "run", "our", "loss", "function", "independently", "on", "every", "semantic", "class", ",", "so", "that", "instances", "belonging", "to", "the", "same", "class", "are", "far", "apart", "in", "feature", "space", ",", "whereas", "instances", "from", "different", "classes", "can", "occupy", "the", "same", "space", ".", "For", "example", ",", "the", "cluster", "centers", "of", "a", "pedestrian", "and", "a", "car", "that", "appear", "in", "the", "same", "image", "are", "not", "pushed", "away", "from", "each", "other", ".", "We", "use", "a", "pretrained", "ResNet", "-", "38", "network", "to", "generate", "segmentation", "masks", "for", "the", "semantic", "classes", ".", "subsection", ":", "Analysis", "of", "the", "individual", "components", "The", "final", "result", "of", "the", "semantic", "instance", "segmentation", "is", "influenced", "by", "three", "main", "components", ":", "the", "performance", "of", "the", "network", "architecture", "with", "our", "loss", "function", ",", "the", "quality", "of", "the", "semantic", "labels", ",", "and", "the", "post", "-", "processing", "step", ".", "To", "disentangle", "the", "effects", "of", "the", "semantic", "segmentation", "and", "the", "post", "-", "processing", "and", "to", "point", "out", "potential", "for", "improvement", ",", "we", "run", "two", "extra", "experiments", "on", "the", "Cityscapes", "validation", "set", ":", "Semantic", "segmentation", "vs", "ground", "truth", "For", "the", "Cityscapes", "challenge", ",", "we", "rely", "on", "semantic", "segmentation", "masks", "to", "make", "a", "distinction", "between", "the", "different", "classes", ".", "Since", "our", "instance", "segmentation", "will", "discard", "regions", "not", "indicated", "in", "the", "semantic", "segmentation", "labels", ",", "the", "results", "will", "be", "influenced", "by", "the", "quality", "of", "the", "semantic", "segmentation", "masks", ".", "To", "measure", "the", "size", "of", "this", "influence", ",", "we", "also", "report", "performance", "with", "the", "ground", "truth", "semantic", "segmentation", "masks", ".", "Mean", "shift", "clustering", "vs", "ground", "truth", "clustering", "Since", "the", "output", "of", "our", "network", "needs", "to", "be", "clustered", "into", "discrete", "instances", ",", "the", "clustering", "method", "can", "potentially", "influence", "the", "accuracy", "of", "the", "overall", "instance", "segmentation", ".", "In", "this", "experiment", ",", "we", "measure", "this", "influence", "by", "clustering", "with", "our", "adapted", "mean", "shift", "clustering", "algorithm", "versus", "thresholding", "around", "the", "mean", "embeddings", "over", "the", "ground", "truth", "instance", "masks", ",", "as", "explained", "in", "section", "[", "reference", "]", ")", ".", "subsection", ":", "Results", "and", "discussion", "Figure", "[", "reference", "]", "shows", "some", "results", "of", "our", "method", "on", "the", "validation", "set", "of", "the", "CVPPP", "dataset", ".", "The", "network", "makes", "very", "few", "mistakes", ":", "only", "the", "segmentation", "of", "the", "smallest", "leafs", "and", "the", "leaf", "stalks", "sometimes", "show", "a", "small", "error", ".", "Table", "[", "reference", "]", "contains", "the", "numerical", "results", ".", "We", "achieve", "competitive", "results", "(", "SBD", "score", "of", "84.2", ")", "that", "are", "on", "-", "par", "with", "the", "state", "-", "of", "-", "the", "art", "(", "SBD", "score", "of", "84.9", ")", ".", "We", "outperform", "all", "non", "-", "deep", "learning", "methods", "and", "also", "the", "recurrent", "instance", "segmentation", "of", ",", "with", "a", "method", "that", "is", "arguably", "less", "complex", ".", "Figure", "[", "reference", "]", "shows", "some", "visual", "results", "on", "the", "Cityscapes", "validation", "set", ".", "We", "see", "that", "even", "in", "difficult", "scenarios", ",", "with", "street", "scenes", "containing", "a", "lot", "of", "cars", "or", "pedestrians", ",", "our", "method", "often", "manages", "to", "identify", "the", "individual", "objects", ".", "Failure", "cases", "mostly", "involve", "the", "splitting", "up", "of", "a", "single", "object", "into", "multiple", "instances", "or", "incorrect", "merging", "of", "neighboring", "instances", ".", "This", "happens", "in", "the", "lower", "left", "example", ",", "where", "the", "two", "rightmost", "cars", "are", "merged", ".", "Another", "failure", "mode", "is", "incorrect", "semantic", "segmentation", ":", "in", "the", "lower", "right", "example", ",", "the", "semantic", "segmentation", "network", "accidentally", "mistakes", "an", "empty", "bicycle", "storage", "for", "actual", "bikes", ".", "The", "instance", "segmentation", "network", "is", "left", "no", "choice", "but", "to", "give", "it", "a", "shot", ",", "and", "tries", "to", "split", "up", "the", "imaginary", "bikes", "into", "individual", "objects", ".", "Nevertheless", ",", "we", "achieve", "competitive", "results", "on", "the", "Cityscapes", "leaderboard", ",", "outperforming", "all", "but", "one", "unpublished", "work", ".", "Note", "that", "we", "perform", "on", "-", "par", "with", "the", "MNC", "-", "based", "method", "SAIS", "on", "this", "dataset", ".", "See", "table", "[", "reference", "]", "for", "a", "complete", "overview", ".", "A", "video", "of", "the", "results", "is", "available", "at", ".", "As", "discussed", "in", "section", "[", "reference", "]", ",", "we", "are", "interested", "to", "know", "the", "influence", "of", "the", "quality", "of", "the", "semantic", "segmentations", "and", "the", "clustering", "algorithm", "on", "the", "overall", "performance", ".", "The", "results", "of", "these", "experiments", "can", "be", "found", "in", "table", "[", "reference", "]", ".", "As", "expected", ",", "the", "performance", "increases", "when", "we", "switch", "out", "a", "component", "with", "its", "ground", "truth", "counterpart", ".", "The", "effect", "of", "the", "semantic", "segmentation", "is", "the", "largest", ":", "comparing", "the", "first", "row", "(", "our", "method", ")", "to", "the", "third", "row", ",", "we", "see", "a", "large", "performance", "increase", "when", "replacing", "the", "ResNet", "-", "38", "semantic", "segmentation", "masks", "with", "the", "ground", "truth", "masks", ".", "This", "can", "be", "explained", "by", "the", "fact", "that", "the", "average", "precision", "metric", "is", "an", "average", "over", "the", "semantic", "classes", ".", "Some", "classes", "like", "tram", ",", "train", "and", "bus", "almost", "never", "have", "more", "than", "one", "instance", "per", "image", ",", "causing", "the", "semantic", "segmentation", "to", "have", "a", "big", "influence", "on", "this", "metric", ".", "It", "is", "clear", "that", "the", "overall", "performance", "can", "be", "increased", "by", "having", "better", "semantic", "segmentations", ".", "The", "last", "two", "entries", "of", "the", "table", "show", "the", "difference", "between", "ground", "truth", "clustering", "and", "mean", "shift", "clustering", ",", "both", "using", "ground", "truth", "segmentation", "masks", ".", "Here", "also", ",", "there", "is", "a", "performance", "gap", ".", "The", "main", "reason", "is", "that", "the", "loss", "on", "the", "validation", "set", "is", "not", "zero", "which", "means", "the", "constraints", "imposed", "by", "the", "loss", "function", "are", "not", "met", ".", "Clustering", "using", "mean", "-", "shift", "will", "therefore", "not", "lead", "to", "perfect", "results", ".", "The", "effect", "is", "more", "pronounced", "for", "small", "instances", ",", "as", "also", "noticeable", "in", "the", "shown", "examples", ".", "subsection", ":", "Speed", "-", "accuracy", "trade", "-", "off", "To", "investigate", "the", "trade", "-", "off", "between", "speed", ",", "accuracy", "and", "memory", "requirements", ",", "we", "train", "four", "different", "network", "models", "on", "different", "resolutions", "and", "evaluate", "them", "on", "the", "car", "class", "of", "the", "Cityscapes", "validation", "set", ".", "This", "also", "illustrates", "the", "benefit", "that", "our", "method", "can", "be", "used", "with", "any", "off", "-", "the", "-", "shelf", "network", "designed", "for", "semantic", "segmentation", ".", "Table", "[", "reference", "]", "shows", "the", "results", ".", "We", "can", "conclude", "that", "Resnet", "-", "38", "is", "best", "for", "accuracy", ",", "but", "requires", "some", "more", "memory", ".", "If", "speed", "is", "important", ",", "ENet", "would", "favor", "over", "Segnet", "since", "it", "is", "much", "faster", "with", "almost", "the", "same", "accuracy", ".", "It", "also", "shows", "that", "running", "on", "a", "higher", "resolution", "than", "768x384", "does", "n’t", "increase", "accuracy", "much", "for", "the", "tested", "networks", ".", "Note", "that", "the", "post", "-", "processing", "step", "can", "be", "implemented", "efficiently", ",", "causing", "only", "a", "negligible", "overhead", ".", "section", ":", "Conclusion", "In", "this", "work", "we", "have", "proposed", "a", "discriminative", "loss", "function", "for", "the", "task", "of", "instance", "segmentation", ".", "After", "training", ",", "the", "output", "of", "the", "network", "can", "be", "clustered", "into", "discrete", "instances", "with", "a", "simple", "post", "-", "processing", "thresholding", "operation", "that", "is", "tailored", "to", "the", "loss", "function", ".", "Furthermore", ",", "we", "showed", "that", "our", "method", "can", "handle", "complex", "occlusions", "as", "opposed", "to", "popular", "detect", "-", "and", "-", "segment", "approaches", ".", "Our", "method", "achieves", "competitive", "performance", "on", "two", "benchmarks", ".", "In", "this", "paper", "we", "still", "used", "a", "pretrained", "network", "to", "produce", "the", "semantic", "segmentation", "masks", ".", "We", "will", "investigate", "the", "joint", "training", "of", "instance", "and", "semantic", "segmentation", "with", "our", "loss", "function", "in", "future", "work", ".", "Acknowledgement", ":", "This", "work", "was", "supported", "by", "Toyota", ",", "and", "was", "carried", "out", "at", "the", "TRACE", "Lab", "at", "KU", "Leuven", "(", "Toyota", "Research", "on", "Automated", "Cars", "in", "Europe", "-", "Leuven", ")", ".", "bibliography", ":", "References"], "entity": [[2, 5, ["Semantic", "Instance", "Segmentation"], "Task"], [7, 10, ["Discriminative", "Loss", "Function"], "Method"], [10, 13, ["Semantic", "instance", "segmentation"], "Task"], [42, 44, ["convolutional", "network"], "Method"], [61, 65, ["post", "-", "processing", "step"], "Method"], [108, 114, ["off", "-", "the", "-", "shelf", "network"], "Method"], [116, 119, ["principled", "loss", "function"], "Method"], [122, 125, ["metric", "learning", "objective"], "Method"], [134, 136, ["instance", "segmentation"], "Task"], [149, 151, ["object", "proposals"], "Method"], [152, 154, ["recurrent", "mechanisms"], "Method"], [203, 209, ["detect", "-", "and", "-", "segment", "approaches"], "Method"], [216, 217, ["Cityscapes"], "Material"], [218, 222, ["CVPPP", "leaf", "segmentation", "benchmarks"], "Material"], [226, 229, ["Semantic", "instance", "segmentation"], "Task"], [239, 242, ["regular", "semantic", "segmentation"], "Task"], [270, 272, ["semantic", "segmentation"], "Task"], [293, 295, ["object", "counting"], "Task"], [296, 298, ["object", "detection"], "Task"], [303, 305, ["instance", "segmentation"], "Task"], [307, 311, ["generalization", "of", "object", "detection"], "Task"], [334, 336, ["instance", "segmentations"], "Method"], [371, 373, ["instance", "segmentation"], "Task"], [374, 376, ["semantic", "segmentation"], "Task"], [402, 404, ["network", "architectures"], "Method"], [405, 407, ["loss", "functions"], "Method"], [411, 413, ["semantic", "segmentation"], "Task"], [416, 418, ["instance", "task"], "Task"], [437, 439, ["instance", "segmentation"], "Task"], [534, 536, ["instance", "segmentation"], "Task"], [537, 539, ["deep", "networks"], "Method"], [549, 553, ["multi", "-", "stage", "pipeline"], "Method"], [554, 556, ["object", "proposals"], "Method"], [560, 562, ["recurrent", "network"], "Method"], [570, 572, ["loss", "function"], "Method"], [601, 605, ["post", "-", "processing", "step"], "Method"], [632, 636, ["post", "-", "processing", "step"], "Task"], [645, 647, ["siamese", "networks"], "Method"], [649, 651, ["triplet", "loss"], "Method"], [652, 654, ["image", "classification"], "Task"], [664, 669, ["pixel", "-", "wise", "softmax", "loss"], "Method"], [674, 676, ["semantic", "segmentation"], "Task"], [678, 680, ["loss", "function"], "Method"], [743, 747, ["post", "-", "processing", "operation"], "Method"], [791, 793, ["network", "architectures"], "Method"], [797, 799, ["semantic", "segmentation"], "Task"], [808, 810, ["shelf", "architecture"], "Method"], [827, 831, ["semantic", "and", "instance", "segmentation"], "Task"], [900, 903, ["instance", "segmentation", "benchmarks"], "Task"], [933, 935, ["deep", "networks"], "Method"], [940, 944, ["semantic", "and", "instance", "segmentation"], "Task"], [955, 957, ["deep", "network"], "Method"], [974, 981, ["Proposal", "-", "based", "Many", "instance", "segmentation", "approaches"], "Method"], [983, 987, ["multi", "-", "stage", "pipeline"], "Method"], [990, 992, ["object", "proposal"], "Method"], [993, 995, ["classification", "step"], "Method"], [1000, 1001, ["MCG"], "Method"], [1003, 1008, ["category", "-", "independent", "region", "proposals"], "Task"], [1012, 1014, ["classification", "step"], "Method"], [1027, 1030, ["generating", "segmentation", "proposals"], "Task"], [1032, 1034, ["deep", "network"], "Method"], [1039, 1045, ["MS", "-", "COCO", "instance", "segmentation", "challenge"], "Task"], [1047, 1050, ["cascade", "of", "networks"], "Method"], [1051, 1052, ["MNC"], "Method"], [1075, 1077, ["object", "detector"], "Method"], [1079, 1082, ["semantic", "segmentation", "network"], "Method"], [1101, 1103, ["bounding", "boxes"], "Method"], [1130, 1132, ["Recurrent", "methods"], "Method"], [1136, 1138, ["recurrent", "networks"], "Method"], [1150, 1157, ["end", "-", "to", "-", "end", "object", "detection"], "Task"], [1159, 1160, ["LSTM"], "Method"], [1162, 1164, ["loss", "function"], "Method"], [1172, 1174, ["Hungarian", "algorithm"], "Method"], [1192, 1199, ["end", "-", "to", "-", "end", "recurrent", "network"], "Method"], [1200, 1202, ["convolutional", "LSTMs"], "Method"], [1218, 1220, ["box", "network"], "Method"], [1232, 1234, ["graphical", "models"], "Method"], [1252, 1256, ["external", "memory", "and", "networks"], "Method"], [1257, 1259, ["box", "proposal"], "Task"], [1260, 1261, ["segmentation"], "Task"], [1262, 1263, ["scoring"], "Task"], [1286, 1288, ["recurrent", "mechanisms"], "Method"], [1293, 1300, ["off", "-", "the", "-", "shelf", "segmentation", "architecture"], "Method"], [1304, 1306, ["loss", "function"], "Method"], [1320, 1322, ["Hungarian", "algorithm"], "Method"], [1323, 1324, ["Clustering"], "Method"], [1342, 1343, ["Silberman"], "Method"], [1345, 1347, ["segmentation", "tree"], "Method"], [1377, 1378, ["MRF"], "Method"], [1385, 1387, ["coherent", "segmentation"], "Method"], [1399, 1400, ["Uhrig"], "Method"], [1424, 1426, ["template", "matching"], "Method"], [1427, 1430, ["proposal", "fusion", "techniques"], "Method"], [1464, 1467, ["sub", "-", "network"], "Method"], [1497, 1499, ["clustering", "step"], "Method"], [1536, 1538, ["deep", "networks"], "Method"], [1552, 1556, ["bottom", "-", "up", "approach"], "Method"], [1574, 1575, ["CRF"], "Method"], [1580, 1582, ["MultiCut", "formulation"], "Method"], [1584, 1586, ["semantic", "segmentations"], "Method"], [1618, 1619, ["Arnab"], "Method"], [1621, 1623, ["object", "detector"], "Method"], [1625, 1628, ["semantic", "segmentation", "module"], "Method"], [1630, 1632, ["CRF", "model"], "Method"], [1646, 1647, ["segmentations"], "Task"], [1651, 1653, ["loss", "function"], "Method"], [1659, 1662, ["distance", "metric", "learning"], "Task"], [1667, 1669, ["siamese", "networks"], "Method"], [1674, 1676, ["loss", "function"], "Method"], [1682, 1684, ["distance", "metric"], "Method"], [1685, 1690, ["large", "margin", "nearest", "neighbor", "classification"], "Task"], [1696, 1699, ["LDA", "based", "objective"], "Method"], [1710, 1712, ["triplet", "loss"], "Method"], [1713, 1715, ["face", "recognition"], "Task"], [1739, 1741, ["clustering", "objective"], "Method"], [1742, 1744, ["unsupervised", "learning"], "Task"], [1791, 1793, ["discriminative", "loss"], "Method"], [1795, 1799, ["distance", "metric", "learning", "principles"], "Method"], [1803, 1805, ["instance", "segmentation"], "Task"], [1806, 1808, ["deep", "networks"], "Method"], [1819, 1821, ["differentiable", "function"], "Method"], [1843, 1845, ["pixel", "embedding"], "Method"], [1850, 1852, ["loss", "function"], "Method"], [1888, 1890, ["loss", "function"], "Method"], [1928, 1930, ["loss", "function"], "Method"], [2010, 2012, ["instance", "segmentation"], "Task"], [2102, 2104, ["distance", "term"], "Method"], [2128, 2130, ["regularization", "term"], "Method"], [2330, 2331, ["variance"], "Metric"], [2332, 2334, ["distance", "loss"], "Metric"], [2356, 2359, ["stochastic", "gradient", "descent"], "Method"], [2362, 2364, ["softmax", "loss"], "Method"], [2370, 2372, ["loss", "function"], "Method"], [2375, 2385, ["pixel", "-", "wise", "multi", "-", "class", "cross", "-", "entropy", "loss"], "Method"], [2399, 2401, ["softmax", "loss"], "Task"], [2405, 2407, ["pixel", "embedding"], "Method"], [2434, 2436, ["softmax", "function"], "Method"], [2532, 2534, ["loss", "function"], "Method"], [2635, 2637, ["loss", "function"], "Method"], [2656, 2659, ["Post", "-", "processing"], "Task"], [2717, 2718, ["inference"], "Task"], [2738, 2739, ["Thresholding"], "Task"], [2762, 2766, ["classification", "and", "semantic", "segmentation"], "Task"], [2781, 2784, ["post", "-", "processing"], "Task"], [2810, 2811, ["inference"], "Task"], [2854, 2856, ["cluster", "centers"], "Method"], [2866, 2868, ["instance", "segmentation"], "Task"], [2875, 2876, ["labeling"], "Task"], [2892, 2893, ["inference"], "Task"], [2948, 2949, ["inference"], "Task"], [2998, 3000, ["Increasing", "robustness"], "Task"], [3002, 3006, ["real", "-", "world", "problem"], "Task"], [3020, 3022, ["clustering", "algorithm"], "Method"], [3023, 3025, ["instance", "segmentation"], "Task"], [3068, 3069, ["clustering"], "Method"], [3080, 3084, ["mean", "-", "shift", "algorithm"], "Method"], [3164, 3166, ["clustering", "algorithm"], "Method"], [3169, 3172, ["ground", "truth", "clustering"], "Method"], [3181, 3183, ["average", "embedding"], "Method"], [3217, 3223, ["detect", "-", "and", "-", "segment", "approaches"], "Method"], [3270, 3272, ["MS", "COCO"], "Material"], [3273, 3275, ["Pascal", "VOC"], "Material"], [3355, 3356, ["detection"], "Task"], [3426, 3427, ["CRF"], "Method"], [3430, 3437, ["real", "-", "world", "industrial", "or", "medical", "applications"], "Task"], [3438, 3442, ["conveyor", "belt", "sorting", "systems"], "Task"], [3443, 3448, ["overlapping", "cell", "and", "chromosome", "segmentation"], "Task"], [3530, 3531, ["Cityscapes"], "Material"], [3535, 3536, ["CVPPP"], "Material"], [3554, 3556, ["Pascal", "VOC"], "Material"], [3557, 3558, ["MSCOCO"], "Material"], [3568, 3572, ["non", "-", "max", "suppression"], "Method"], [3626, 3628, ["loss", "function"], "Method"], [3635, 3639, ["CVPPP", "Leaf", "Segmentation", "dataset"], "Material"], [3689, 3691, ["LSC", "competition"], "Material"], [3693, 3695, ["CVPPP", "workshop"], "Material"], [3723, 3726, ["computer", "vision", "methods"], "Method"], [3732, 3734, ["plant", "phenotyping"], "Task"], [3738, 3740, ["A1", "subset"], "Material"], [3782, 3787, ["Symmetric", "Best", "Dice", "(", ")"], "Metric"], [3791, 3792, ["accuracy"], "Metric"], [3794, 3796, ["instance", "segmentation"], "Method"], [3797, 3801, ["Absolute", "Difference", "in", "Count"], "Metric"], [3836, 3842, ["semantic", "understanding", "of", "urban", "street", "scenes"], "Task"], [3848, 3857, ["pixel", "-", "level", "and", "instance", "-", "level", "semantic", "segmentation"], "Task"], [3927, 3928, ["accuracy"], "Metric"], [3934, 3937, ["mean", "Average", "Precision"], "Metric"], [3938, 3939, ["AP"], "Metric"], [3941, 3944, ["mean", "Average", "Precision"], "Metric"], [3945, 3946, ["overlap"], "Metric"], [3978, 3980, ["Model", "architecture"], "Method"], [3998, 4004, ["off", "-", "the", "-", "shelf", "network"], "Method"], [4008, 4012, ["ResNet", "-", "38", "network"], "Method"], [4015, 4017, ["semantic", "segmentation"], "Task"], [4029, 4032, ["CityScapes", "semantic", "segmentation"], "Task"], [4043, 4044, ["Adam"], "Method"], [4047, 4049, ["learning", "rate"], "Metric"], [4060, 4062, ["Leaf", "segmentation"], "Task"], [4073, 4076, ["online", "data", "augmentation"], "Task"], [4087, 4088, ["robustness"], "Metric"], [4101, 4104, ["random", "scale", "deformation"], "Method"], [4159, 4161, ["instance", "segmentation"], "Task"], [4162, 4163, ["Cityscapes"], "Material"], [4192, 4194, ["data", "augmentation"], "Task"], [4212, 4214, ["CVPPP", "dataset"], "Material"], [4215, 4216, ["Cityscapes"], "Material"], [4218, 4224, ["multi", "-", "class", "instance", "segmentation", "challenge"], "Task"], [4230, 4232, ["loss", "function"], "Method"], [4293, 4298, ["pretrained", "ResNet", "-", "38", "network"], "Method"], [4319, 4322, ["semantic", "instance", "segmentation"], "Task"], [4333, 4335, ["network", "architecture"], "Method"], [4337, 4339, ["loss", "function"], "Method"], [4349, 4353, ["post", "-", "processing", "step"], "Method"], [4360, 4362, ["semantic", "segmentation"], "Method"], [4364, 4367, ["post", "-", "processing"], "Task"], [4386, 4388, ["Semantic", "segmentation"], "Task"], [4389, 4391, ["ground", "truth"], "Task"], [4413, 4415, ["instance", "segmentation"], "Method"], [4460, 4463, ["Mean", "shift", "clustering"], "Method"], [4464, 4467, ["ground", "truth", "clustering"], "Method"], [4482, 4484, ["clustering", "method"], "Method"], [4488, 4489, ["accuracy"], "Metric"], [4492, 4494, ["instance", "segmentation"], "Task"], [4504, 4505, ["clustering"], "Method"], [4508, 4512, ["mean", "shift", "clustering", "algorithm"], "Method"], [4513, 4514, ["thresholding"], "Method"], [4555, 4557, ["CVPPP", "dataset"], "Material"], [4596, 4598, ["SBD", "score"], "Metric"], [4615, 4617, ["SBD", "score"], "Metric"], [4624, 4629, ["non", "-", "deep", "learning", "methods"], "Method"], [4632, 4635, ["recurrent", "instance", "segmentation"], "Method"], [4731, 4733, ["semantic", "segmentation"], "Method"], [4741, 4744, ["semantic", "segmentation", "network"], "Method"], [4755, 4758, ["instance", "segmentation", "network"], "Method"], [4856, 4858, ["semantic", "segmentations"], "Method"], [4860, 4862, ["clustering", "algorithm"], "Method"], [4903, 4905, ["semantic", "segmentation"], "Task"], [4952, 4955, ["average", "precision", "metric"], "Metric"], [4983, 4985, ["semantic", "segmentation"], "Task"], [5007, 5009, ["semantic", "segmentations"], "Method"], [5021, 5024, ["ground", "truth", "clustering"], "Method"], [5025, 5028, ["mean", "shift", "clustering"], "Method"], [5072, 5073, ["Clustering"], "Method"], [5074, 5077, ["mean", "-", "shift"], "Method"], [5117, 5118, ["speed"], "Metric"], [5119, 5120, ["accuracy"], "Metric"], [5121, 5123, ["memory", "requirements"], "Metric"], [5128, 5130, ["network", "models"], "Method"], [5159, 5165, ["off", "-", "the", "-", "shelf", "network"], "Method"], [5167, 5169, ["semantic", "segmentation"], "Task"], [5182, 5185, ["Resnet", "-", "38"], "Method"], [5188, 5189, ["accuracy"], "Metric"], [5197, 5198, ["speed"], "Metric"], [5201, 5202, ["ENet"], "Method"], [5205, 5206, ["Segnet"], "Method"], [5215, 5216, ["accuracy"], "Metric"], [5231, 5232, ["accuracy"], "Metric"], [5241, 5245, ["post", "-", "processing", "step"], "Method"], [5273, 5275, ["instance", "segmentation"], "Task"], [5293, 5298, ["post", "-", "processing", "thresholding", "operation"], "Method"], [5321, 5327, ["detect", "-", "and", "-", "segment", "approaches"], "Method"], [5344, 5346, ["pretrained", "network"], "Method"], [5357, 5364, ["joint", "training", "of", "instance", "and", "semantic", "segmentation"], "Task"], [5366, 5368, ["loss", "function"], "Method"], [5396, 5398, ["Automated", "Cars"], "Task"], [30, 32, ["loss", "function"], "Method"], [67, 69, ["loss", "function"], "Method"], [364, 366, ["object", "detection"], "Task"], [434, 436, ["loss", "function"], "Method"], [659, 661, ["loss", "function"], "Method"], [778, 780, ["loss", "function"], "Method"], [817, 819, ["loss", "function"], "Method"], [1664, 1666, ["loss", "functions"], "Method"], [1751, 1753, ["loss", "function"], "Method"], [1815, 1817, ["loss", "function"], "Method"], [2079, 2080, ["variance"], "Metric"], [2151, 2152, ["variance"], "Metric"], [2262, 2264, ["loss", "function"], "Method"], [2661, 2662, ["variance"], "Metric"], [3563, 3564, ["detection"], "Task"], [3610, 3611, ["detection"], "Task"], [3641, 3642, ["Cityscapes"], "Material"], [3832, 3833, ["Cityscapes"], "Material"], [4382, 4383, ["Cityscapes"], "Material"], [4393, 4394, ["Cityscapes"], "Material"], [4656, 4657, ["Cityscapes"], "Material"], [4789, 4790, ["Cityscapes"], "Material"], [4808, 4809, ["MNC"], "Method"], [5066, 5068, ["loss", "function"], "Method"], [5106, 5107, ["accuracy"], "Metric"], [5142, 5143, ["Cityscapes"], "Material"], [5267, 5269, ["loss", "function"], "Method"], [5303, 5305, ["loss", "function"], "Method"]]} {"doc_id": "22634b09c3c83f4a959f4f732b03ec3c92808094", "text": ["document", ":", "DeepMatching", ":", "Hierarchical", "Deformable", "Dense", "Matching", "We", "introduce", "a", "novel", "matching", "algorithm", ",", "called", "DeepMatching", ",", "to", "compute", "dense", "correspondences", "between", "images", ".", "DeepMatching", "relies", "on", "a", "hierarchical", ",", "multi", "-", "layer", ",", "correlational", "architecture", "designed", "for", "matching", "images", "and", "was", "inspired", "by", "deep", "convolutional", "approaches", ".", "The", "proposed", "matching", "algorithm", "can", "handle", "non", "-", "rigid", "deformations", "and", "repetitive", "textures", "and", "efficiently", "determines", "dense", "correspondences", "in", "the", "presence", "of", "significant", "changes", "between", "images", ".", "We", "evaluate", "the", "performance", "of", "DeepMatching", ",", "in", "comparison", "with", "state", "-", "of", "-", "the", "-", "art", "matching", "algorithms", ",", "on", "the", "Mikolajczyk", "Mikolajczyk2005", ",", "the", "MPI", "-", "Sintel", "sintel", "and", "the", "Kitti", "kitti", "datasets", ".", "DeepMatching", "outperforms", "the", "state", "-", "of", "-", "the", "-", "art", "algorithms", "and", "shows", "excellent", "results", "in", "particular", "for", "repetitive", "textures", ".", "We", "also", "propose", "a", "method", "for", "estimating", "optical", "flow", ",", "called", "DeepFlow", ",", "by", "integrating", "DeepMatching", "in", "the", "large", "displacement", "optical", "flow", "(", "LDOF", ")", "approach", "of", "[", "]", ".", "Compared", "to", "existing", "matching", "algorithms", ",", "additional", "robustness", "to", "large", "displacements", "and", "complex", "motion", "is", "obtained", "thanks", "to", "our", "matching", "approach", ".", "DeepFlow", "obtains", "competitive", "performance", "on", "public", "benchmarks", "for", "optical", "flow", "estimation", ".", "section", ":", "Introduction", "Computing", "correspondences", "between", "related", "images", "is", "a", "central", "issue", "in", "many", "computer", "vision", "problems", ",", "ranging", "from", "scene", "recognition", "to", "optical", "flow", "estimation", "forsyth2011computer", ",", "Szeliski2010", ".", "The", "goal", "of", "a", "matching", "algorithm", "is", "to", "discover", "shared", "visual", "content", "between", "two", "images", ",", "and", "to", "establish", "as", "many", "as", "possible", "precise", "point", "-", "wise", "correspondences", ",", "called", "matches", ".", "An", "essential", "aspect", "of", "matching", "approaches", "is", "the", "amount", "of", "rigidity", "they", "assume", "when", "computing", "the", "correspondences", ".", "In", "fact", ",", "matching", "approaches", "range", "between", "two", "extreme", "cases", ":", "stereo", "matching", ",", "where", "matching", "hinges", "upon", "strong", "geometric", "constraints", ",", "and", "matching", "‘", "‘", "in", "the", "wild", "’", "’", ",", "where", "the", "set", "of", "possible", "transformations", "from", "the", "source", "image", "to", "the", "target", "one", "is", "large", "and", "the", "problem", "is", "basically", "almost", "unconstrained", ".", "Effective", "approaches", "have", "been", "designed", "for", "matching", "rigid", "objects", "across", "images", "in", "the", "presence", "of", "large", "viewpoint", "changes", "Lowe2004", ",", "Barnes2010", ",", "HaCohen2011", ".", "However", ",", "the", "performance", "of", "current", "state", "-", "of", "-", "the", "-", "art", "matching", "algorithms", "for", "images", "‘", "‘", "in", "the", "wild", "’", "’", ",", "such", "as", "consecutive", "images", "in", "real", "-", "world", "videos", "featuring", "fast", "non", "-", "rigid", "motion", ",", "still", "calls", "for", "improvement", "mdpof", ",", "Chen2013", ".", "In", "this", "paper", ",", "we", "aim", "at", "tackling", "matching", "in", "such", "a", "general", "setting", ".", "Matching", "algorithms", "for", "images", "‘", "‘", "in", "the", "wild", "’", "’", "need", "to", "accommodate", "several", "requirements", ",", "that", "turn", "out", "to", "be", "often", "in", "contradiction", ".", "On", "one", "hand", ",", "matching", "objects", "necessarily", "requires", "rigidity", "assumptions", "to", "some", "extent", ".", "It", "is", "also", "mandatory", "that", "these", "objects", "have", "sufficiently", "discriminative", "textures", "to", "make", "the", "problem", "well", "-", "defined", ".", "On", "the", "other", "hand", ",", "many", "objects", "or", "regions", "are", "not", "rigid", "objects", ",", "like", "humans", "or", "animals", ".", "Furthermore", ",", "large", "portions", "of", "an", "image", "are", "usually", "occupied", "by", "weakly", "-", "to", "-", "no", "textured", "regions", ",", "often", "with", "repetitive", "textures", ",", "like", "sky", "or", "bucolic", "background", ".", "Descriptor", "matching", "approaches", ",", "such", "as", "SIFT", "Lowe2004", "or", "HOG", "Dalal2005", ",", "Bro11a", "matching", ",", "compute", "discriminative", "feature", "representations", "from", "rectangular", "patches", ".", "However", ",", "while", "these", "approaches", "succeed", "in", "case", "of", "rigid", "motion", ",", "they", "fail", "to", "match", "regions", "with", "weak", "or", "repetitive", "textures", ",", "as", "local", "patches", "are", "poorly", "discriminative", ".", "Furthermore", ",", "matches", "are", "usually", "poor", "and", "imprecise", "in", "case", "of", "non", "-", "rigid", "deformations", ",", "as", "these", "approaches", "rely", "on", "rigid", "patches", ".", "Discriminative", "power", "can", "be", "traded", "against", "increased", "robustness", "to", "non", "-", "rigid", "deformations", ".", "Indeed", ",", "propagation", "-", "based", "approaches", ",", "such", "as", "Generalized", "PatchMatch", "Barnes2010", "or", "Non", "-", "rigid", "Dense", "Correspondences", "HaCohen2011", ",", "compute", "simple", "feature", "representations", "from", "small", "patches", "and", "propagate", "matches", "to", "neighboring", "patches", ".", "They", "yield", "good", "performance", "in", "case", "of", "non", "-", "rigid", "deformations", ".", "However", ",", "matching", "repetitive", "textures", "remains", "beyond", "the", "reach", "of", "these", "approaches", ".", "In", "this", "paper", "we", "propose", "a", "novel", "approach", ",", "called", "DeepMatching", ",", "that", "gracefully", "combines", "the", "strengths", "of", "these", "two", "families", "of", "approaches", ".", "DeepMatching", "is", "computed", "using", "a", "multi", "-", "layer", "architecture", ",", "which", "breaks", "down", "patches", "into", "a", "hierarchy", "of", "sub", "-", "patches", ".", "This", "architecture", "allows", "to", "work", "at", "several", "scales", "and", "handle", "repetitive", "textures", ".", "Furthermore", ",", "within", "each", "layer", ",", "local", "matches", "are", "computed", "assuming", "a", "restricted", "set", "of", "feasible", "rigid", "deformations", ".", "Local", "matches", "are", "then", "propagated", "up", "the", "hierarchy", ",", "which", "progressively", "discard", "spurious", "incorrect", "matches", ".", "We", "called", "our", "approach", "DeepMatching", ",", "as", "it", "is", "inspired", "by", "deep", "convolutional", "approaches", ".", "In", "summary", ",", "we", "make", "three", "contributions", ":", "Dense", "matching", ":", "we", "propose", "a", "matching", "algorithm", ",", "DeepMatching", ",", "that", "allows", "to", "robustly", "determine", "dense", "correspondences", "between", "two", "images", ".", "It", "explicitly", "handles", "non", "-", "rigid", "deformations", ",", "with", "bounds", "on", "the", "deformation", "tolerance", ",", "and", "incorporates", "a", "multi", "-", "scale", "scoring", "of", "the", "matches", ",", "making", "it", "robust", "to", "repetitive", "or", "weak", "textures", ".", "Furthermore", ",", "our", "approach", "is", "based", "on", "gradient", "histograms", ",", "and", "thus", "robust", "to", "appearance", "changes", "caused", "by", "illumination", "and", "color", "variations", ".", "Fast", ",", "scale", "/", "rotation", "-", "invariant", "matching", ":", "we", "propose", "a", "computationally", "efficient", "version", "of", "DeepMatching", ",", "which", "performs", "almost", "as", "well", "as", "exact", "DeepMatching", ",", "but", "at", "a", "much", "lower", "memory", "cost", ".", "Furthermore", ",", "this", "fast", "version", "of", "DeepMatching", "can", "be", "extended", "to", "a", "scale", "and", "rotation", "-", "invariant", "version", ",", "making", "it", "an", "excellent", "competitor", "to", "state", "-", "of", "-", "the", "-", "art", "descriptor", "matching", "approaches", ".", "Large", "-", "displacement", "optical", "flow", ":", "we", "propose", "an", "optical", "flow", "approach", "which", "uses", "DeepMatching", "in", "the", "matching", "term", "of", "the", "large", "displacement", "variational", "energy", "minimization", "of", "[", "]", ".", "We", "show", "that", "DeepMatching", "is", "a", "better", "choice", "compared", "to", "the", "HOG", "descriptor", "used", "by", "[", "]", "and", "other", "state", "-", "of", "-", "the", "-", "art", "matching", "algorithms", ".", "The", "approach", ",", "named", "DeepFlow", ",", "obtains", "competitive", "results", "on", "public", "optical", "flow", "benchmarks", ".", "This", "paper", "is", "organized", "as", "follows", ".", "After", "a", "review", "of", "previous", "works", "(", "Section", "[", "reference", "]", ")", ",", "we", "start", "by", "presenting", "the", "proposed", "matching", "algorithm", ",", "DeepMatching", ",", "in", "Section", "[", "reference", "]", ".", "Then", ",", "Section", "[", "reference", "]", "describes", "several", "extensions", "of", "DeepMatching", ".", "In", "particular", ",", "we", "propose", "an", "optical", "flow", "approach", ",", "DeepFlow", ",", "in", "Section", "[", "reference", "]", ".", "Finally", ",", "we", "present", "experimental", "results", "in", "Section", "[", "reference", "]", ".", "A", "preliminary", "version", "of", "this", "article", "has", "appeared", "in", ".", "This", "version", "adds", "(", "1", ")", "an", "in", "-", "depth", "presentation", "of", "DeepMatching", ";", "(", "2", ")", "an", "enhanced", "version", "of", "DeepMatching", ",", "which", "improves", "the", "match", "scoring", "and", "the", "selection", "of", "entry", "points", "for", "backtracking", ";", "(", "3", ")", "proofs", "on", "time", "and", "memory", "complexity", "of", "DeepMatching", "as", "well", "as", "its", "deformation", "tolerance", ";", "(", "4", ")", "a", "discussion", "on", "the", "connection", "between", "Deep", "Convolutional", "Neural", "Networks", "and", "DeepMatching", ";", "(", "5", ")", "a", "fast", "approximate", "version", "of", "DeepMatching", ";", "(", "6", ")", "a", "scale", "and", "rotation", "invariant", "version", "of", "DeepMatching", ";", "and", "(", "7", ")", "an", "extensive", "experimental", "evaluation", "of", "DeepMatching", "on", "several", "state", "-", "of", "-", "the", "-", "art", "benchmarks", ".", "The", "code", "for", "DeepMatching", "as", "well", "as", "DeepFlow", "are", "available", "at", "and", ".", "Note", "that", "we", "provide", "a", "GPU", "implementation", "in", "addition", "to", "the", "CPU", "one", ".", "section", ":", "Related", "work", "In", "this", "section", "we", "review", "related", "work", "on", "‘", "‘", "general", "’", "’", "image", "matching", ",", "that", "is", "matching", "without", "prior", "knowledge", "and", "constraints", ",", "and", "on", "matching", "in", "the", "context", "of", "optical", "flow", "estimation", ",", "that", "is", "matching", "consecutive", "images", "in", "videos", ".", "subsection", ":", "General", "image", "matching", "Image", "matching", "based", "on", "local", "features", "has", "been", "extensively", "studied", "in", "the", "past", "decade", ".", "It", "has", "been", "applied", "successfully", "to", "various", "domains", ",", "such", "as", "wide", "baseline", "stereo", "matching", "Furukawa2010", "and", "image", "retrieval", "Philbin2010", ".", "It", "consists", "of", "two", "steps", ",", "i.e", ".", ",", "extracting", "local", "descriptors", "and", "matching", "them", ".", "Image", "descriptors", "are", "extracted", "in", "rigid", "(", "generally", "square", ")", "local", "frames", "at", "sparse", "invariant", "image", "locations", "Mikolajczyk2005", ",", "Szeliski2010", ".", "Matching", "then", "equals", "nearest", "neighbor", "search", "between", "descriptors", ",", "followed", "by", "an", "optional", "geometric", "verification", ".", "Note", "that", "a", "confidence", "value", "can", "be", "obtained", "by", "computing", "the", "uniqueness", "of", "a", "match", ",", "i.e.", ",", "by", "looking", "at", "the", "distance", "of", "its", "nearest", "neighbors", "While", "this", "class", "of", "techniques", "is", "well", "suited", "for", "well", "-", "textured", "rigid", "objects", ",", "it", "fails", "to", "match", "non", "-", "rigid", "objects", "and", "weakly", "textured", "regions", ".", "In", "contrast", ",", "the", "proposed", "matching", "algorithm", ",", "called", "DeepMatching", ",", "is", "inspired", "by", "non", "-", "rigid", "2D", "warping", "and", "deep", "convolutional", "networks", "LeCun98", ",", "Uchida1998", ",", "Keysers2007", ".", "This", "family", "of", "approaches", "explicitly", "models", "non", "-", "rigid", "deformations", ".", "We", "employ", "a", "novel", "family", "of", "feasible", "warpings", "that", "does", "not", "enforce", "monotonicity", "nor", "continuity", "constraints", ",", "in", "contrast", "to", "traditional", "2D", "warping", "Uchida1998", ",", "Keysers2007", ".", "This", "makes", "the", "problem", "computationally", "much", "less", "expensive", ".", "It", "is", "also", "worthwhile", "to", "mention", "the", "similarity", "with", "non", "-", "rigid", "matching", "approaches", "developed", "for", "a", "broad", "range", "of", "applications", ".", "proposed", "a", "similar", "pipeline", "to", "ours", "(", "albeit", "more", "complex", ")", "to", "measure", "the", "similarity", "of", "small", "images", ".", "However", ",", "their", "method", "lacks", "a", "way", "of", "merging", "correspondences", "belonging", "to", "objects", "with", "contradictory", "motions", ",", "e.g", ".", ",", "on", "different", "focal", "planes", ".", "For", "the", "purpose", "of", "establishing", "dense", "correspondences", "between", "images", ",", "estimated", "a", "non", "-", "rigid", "matching", "by", "robustly", "fitting", "smooth", "parametric", "models", "(", "homography", "and", "splines", ")", "to", "local", "descriptor", "matches", ".", "In", "contrast", ",", "our", "approach", "is", "non", "-", "parametric", "and", "model", "-", "free", ".", "Recently", ",", "fast", "algorithms", "for", "dense", "patch", "matching", "have", "taken", "advantage", "of", "the", "redundancy", "between", "overlapping", "patches", "Barnes2010", ",", "Korman2011", ",", "kpm", ",", "daisyff", ".", "The", "insight", "is", "to", "propagate", "good", "matches", "to", "their", "neighborhood", "in", "a", "loose", "fashion", ",", "yielding", "dense", "non", "-", "rigid", "matches", ".", "In", "practice", ",", "however", ",", "the", "lack", "of", "a", "smoothness", "constraint", "leads", "to", "highly", "discontinuous", "matches", ".", "Several", "works", "have", "proposed", "ways", "to", "fix", "this", ".", "reinforce", "neighboring", "matches", "using", "an", "iterative", "multiscale", "expansion", "and", "contraction", "strategy", ",", "performed", "in", "a", "coarse", "-", "to", "-", "fine", "manner", ".", "include", "a", "guided", "filtering", "stage", "on", "top", "of", "PatchMatch", ",", "which", "obtains", "smooth", "correspondence", "fields", "by", "locally", "approximating", "a", "MRF", ".", "Finally", ",", "propose", "a", "hierarchical", "matching", "to", "obtain", "dense", "correspondences", ",", "using", "a", "coarse", "-", "to", "-", "fine", "(", "top", "-", "down", ")", "strategy", ".", "Loopy", "belief", "propagation", "is", "used", "to", "perform", "inference", ".", "In", "contrast", "to", "these", "approaches", ",", "DeepMatching", "proceeds", "bottom", "-", "up", "and", ",", "then", ",", "top", "-", "down", ".", "Due", "to", "its", "hierarchical", "nature", ",", "DeepMatching", "is", "able", "to", "consider", "patches", "at", "several", "scales", ",", "thus", "overcoming", "the", "lack", "of", "distinctiveness", "that", "affects", "small", "patches", ".", "Yet", ",", "the", "multi", "-", "layer", "construction", "allows", "to", "efficiently", "perform", "matching", "allowing", "semi", "-", "rigid", "local", "deformations", ".", "In", "addition", ",", "DeepMatching", "can", "be", "computed", "efficiently", ",", "and", "can", "be", "further", "accelerated", "to", "satisfy", "low", "-", "memory", "requirements", "with", "negligible", "loss", "in", "accuracy", ".", "subsection", ":", "Matching", "for", "flow", "estimation", "Variational", "energy", "minimization", "is", "currently", "the", "most", "popular", "framework", "for", "optical", "flow", "estimation", ".", "Since", "the", "pioneering", "work", "of", ",", "research", "has", "focused", "on", "alleviating", "the", "drawbacks", "of", "this", "approach", ".", "A", "series", "of", "improvements", "were", "proposed", "over", "the", "years", "Black1996", ",", "Werlberger2009", ",", "Bruhn2005", ",", "Papenberg2006", ",", "middlebury", ",", "Sun2014", ",", "Vogel2013data", ".", "The", "variational", "approach", "of", "combines", "most", "of", "these", "improvements", "in", "a", "unified", "framework", ".", "The", "energy", "decomposes", "into", "several", "terms", ",", "resp", ".", "the", "data", "-", "fitting", "and", "the", "smoothness", "terms", ".", "Energy", "minimization", "is", "performed", "by", "solving", "the", "Euler", "-", "Lagrange", "equations", ",", "reducing", "the", "problem", "to", "solving", "a", "sequence", "of", "large", "and", "structured", "linear", "systems", ".", "More", "recently", ",", "the", "addition", "of", "a", "descriptor", "matching", "term", "in", "the", "energy", "to", "be", "minimized", "was", "proposed", "by", ".", "Following", "this", "idea", ",", "several", "papers", "Tola2008", ",", "Bro11a", ",", "siftflow", ",", "siftscales", "show", "that", "dense", "descriptor", "matching", "improves", "performance", ".", "Strategies", "such", "as", "reciprocal", "nearest", "-", "neighbor", "verification", "allow", "to", "prune", "most", "of", "the", "false", "matches", ".", "However", ",", "a", "variational", "energy", "minimization", "approach", "that", "includes", "such", "a", "descriptor", "matching", "term", "may", "fail", "at", "locations", "where", "matches", "are", "missing", "or", "wrong", ".", "Related", "approaches", "tackle", "the", "problem", "of", "dense", "scene", "correspondence", ".", "SIFT", "-", "flow", "one", "of", "the", "most", "famous", "method", "in", "this", "context", ",", "also", "formulates", "the", "matching", "problem", "in", "a", "variational", "framework", ".", "[", "]", "improve", "over", "SIFT", "-", "flow", "by", "using", "multi", "-", "scale", "patches", ".", "However", ",", "this", "decreases", "performance", "in", "cases", "where", "scale", "invariance", "is", "not", "required", ".", "integrate", "matching", "of", "SIFT", "Lowe2004", "and", "PatchMatch", "Barnes2010", "to", "refine", "the", "flow", "initialization", "at", "each", "level", ".", "Excellent", "results", "are", "obtained", "for", "optical", "flow", "estimation", ",", "yet", "at", "the", "cost", "of", "expensive", "fusion", "steps", ".", "extends", "sparse", "matches", "with", "locally", "affine", "constraints", "to", "dense", "matches", "and", ",", "then", ",", "uses", "a", "total", "variation", "algorithm", "to", "refine", "the", "flow", "estimation", ".", "We", "present", "here", "a", "computationally", "efficient", "and", "competitive", "approach", "for", "large", "displacement", "optical", "flow", "by", "integrating", "the", "proposed", "DeepMatching", "algorithm", "into", "the", "approach", "of", ".", "section", ":", "DeepMatching", "This", "section", "introduces", "our", "matching", "algorithm", "DeepMatching", ".", "DeepMatching", "is", "a", "matching", "algorithm", "based", "on", "correlations", "at", "the", "patch", "-", "level", ",", "that", "proceeds", "in", "a", "multi", "-", "layer", "fashion", ".", "The", "multi", "-", "layer", "architecture", "relies", "on", "a", "quadtree", "-", "like", "patch", "subdivision", "scheme", ",", "with", "an", "extra", "degree", "of", "freedom", "to", "locally", "re", "-", "optimize", "the", "positions", "of", "each", "quadrant", ".", "In", "order", "to", "enhance", "the", "contrast", "of", "the", "spatial", "correlation", "maps", "output", "by", "the", "local", "correlations", ",", "a", "nonlinear", "transformation", "is", "applied", "after", "each", "layer", ".", "We", "first", "give", "an", "overview", "of", "DeepMatching", "in", "Section", "[", "reference", "]", "and", "show", "that", "it", "can", "be", "decomposed", "in", "a", "bottom", "-", "up", "pass", "followed", "by", "a", "top", "-", "down", "pass", ".", "We", ",", "then", ",", "present", "the", "bottom", "-", "up", "pass", "in", "Section", "[", "reference", "]", "and", "the", "top", "-", "down", "one", "in", "Section", "[", "reference", "]", ".", "Finally", ",", "we", "analyze", "DeepMatching", "in", "Section", "[", "reference", "]", ".", "subsection", ":", "Overview", "of", "the", "approach", "A", "state", "-", "of", "-", "the", "-", "art", "approach", "for", "matching", "regions", "between", "two", "images", "is", "based", "on", "the", "SIFT", "descriptor", "Lowe2004", ".", "SIFT", "is", "a", "histogram", "of", "gradients", "with", "spatial", "and", "8", "orientation", "bins", ",", "yielding", "a", "robust", "descriptor", "that", "effectively", "encodes", "a", "square", "image", "region", ".", "Note", "that", "its", "cell", "grid", "can", "also", "be", "viewed", "as", "4", "so", "-", "called", "‘", "‘", "quadrants", "’", "’", "of", "cells", ",", "see", "Figure", "[", "reference", "]", ".", "We", "can", ",", "then", ",", "rewrite", "with", ".", "Let", "and", "be", "the", "SIFT", "descriptors", "of", "the", "corresponding", "regions", "in", "the", "source", "and", "target", "image", ".", "In", "order", "to", "remove", "the", "effect", "of", "non", "-", "rigid", "motion", ",", "we", "propose", "to", "optimize", "the", "positions", "of", "the", "quadrants", "of", "the", "target", "descriptor", "(", "rather", "than", "keeping", "them", "fixed", ")", ",", "in", "order", "to", "maximize", "where", "is", "the", "descriptor", "of", "a", "single", "quadrant", "extracted", "at", "position", "and", "a", "similarity", "function", ".", "Now", ",", "is", "able", "to", "handle", "situations", "such", "as", "the", "one", "presented", "in", "Figure", "[", "reference", "]", ",", "where", "a", "region", "contains", "multiple", "objects", "moving", "in", "different", "directions", ".", "Furthermore", ",", "if", "the", "four", "quadrants", "can", "move", "independently", "(", "of", "course", ",", "within", "some", "extent", ")", ",", "it", "can", "be", "calculated", "more", "efficiently", "as", ":", "When", "applied", "recursively", "to", "each", "quadrant", "by", "subdivided", "it", "into", "4", "sub", "-", "quadrants", "until", "a", "minimum", "patch", "size", "is", "reached", "(", "atomic", "patches", ")", ",", "this", "strategy", "allows", "for", "accurate", "non", "-", "rigid", "matching", ".", "Such", "a", "recursive", "decomposition", "can", "be", "represented", "as", "a", "quad", "-", "tree", ",", "see", "Figure", "[", "reference", "]", ".", "Given", "an", "initial", "pair", "of", "two", "matching", "regions", ",", "retrieving", "atomic", "patch", "correspondences", "is", "then", "done", "in", "a", "top", "-", "down", "fashion", "(", "i.e", ".", "by", "recursively", "applying", "Eq", ".", "(", "[", "reference", "]", ")", "to", "the", "quadrant", "’s", "positions", ")", ".", "Nevertheless", ",", "in", "order", "to", "first", "determine", "the", "set", "of", "matching", "regions", "between", "the", "two", "images", ",", "we", "need", "to", "compute", "beforehand", "the", "matching", "scores", "(", "i.e", ".", "similarity", ")", "of", "all", "large", "-", "enough", "patches", "in", "the", "two", "images", "(", "as", "in", "Figure", "[", "reference", "]", ")", ",", "and", "keep", "the", "pairs", "with", "maximum", "similarity", ".", "As", "indicated", "by", "Eq", ".", "(", "[", "reference", "]", ")", ",", "the", "score", "is", "formed", "by", "averaging", "the", "max", "-", "pooled", "scores", "of", "the", "quadrants", ".", "Hence", ",", "the", "process", "of", "computing", "the", "matching", "scores", "is", "bottom", "-", "up", ".", "In", "the", "following", ",", "we", "call", "correlation", "map", "the", "matching", "scores", "of", "a", "single", "patch", "from", "the", "first", "image", "at", "every", "position", "in", "the", "second", "image", ".", "Selecting", "matching", "patches", "then", "corresponds", "to", "finding", "local", "maxima", "in", "the", "correlation", "maps", ".", "To", "sum", "-", "up", ",", "the", "algorithm", "can", "be", "decomposed", "in", "two", "steps", ":", "(", "i", ")", "first", ",", "correlation", "maps", "are", "computed", "using", "a", "bottom", "-", "up", "algorithm", ",", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "Correlation", "maps", "of", "small", "patches", "are", "first", "computed", "and", "then", "aggregated", "to", "form", "correlation", "maps", "of", "larger", "patches", ";", "(", "ii", ")", "next", ",", "a", "top", "-", "down", "method", "estimates", "the", "motion", "of", "atomic", "patches", "starting", "from", "matches", "of", "large", "patches", ".", "In", "the", "remainder", "of", "this", "section", ",", "we", "detail", "the", "two", "steps", "described", "above", "(", "Section", "[", "reference", "]", "and", "Section", "[", "reference", "]", ")", ",", "before", "analyzing", "the", "properties", "of", "DeepMatching", "in", "Section", "[", "reference", "]", ".", "subsection", ":", "Bottom", "-", "up", "correlation", "pyramid", "computation", "Let", "and", "be", "two", "images", "of", "resolution", "and", ".", "paragraph", ":", "Bottom", "level", ".", "We", "use", "patches", "of", "size", "pixels", "as", "atomic", "patches", ".", "We", "split", "into", "non", "-", "overlapping", "atomic", "patches", ",", "and", "compute", "the", "correlation", "map", "with", "image", "for", "each", "of", "them", ",", "see", "Figure", "[", "reference", "]", ".", "The", "score", "between", "two", "atomic", "patches", "and", "is", "defined", "as", "the", "average", "pixel", "-", "wise", "similarity", ":", "where", "each", "pixel", "is", "represented", "as", "a", "histogram", "of", "oriented", "gradients", "pooled", "over", "a", "local", "neighborhood", ".", "We", "detail", "below", "how", "the", "pixel", "descriptor", "is", "computed", ".", "paragraph", ":", "Pixel", "descriptor", ":", "We", "rely", "on", "a", "robust", "pixel", "representation", "that", "is", "similar", "in", "spirit", "to", "SIFT", "and", "DAISY", "Given", "an", "input", "image", "I", ",", "we", "first", "apply", "a", "Gaussian", "smoothing", "of", "radius", "ν1", "in", "order", "to", "denoise", "I", "from", "potential", "artifacts", "caused", "for", "example", "by", "JPEG", "compression", ".", "We", "then", "extract", "the", "gradient", "(", "⁢δx", ",", "⁢δy", ")", "at", "each", "pixel", "and", "compute", "its", "non", "-", "negative", "projection", "onto", "8", "orientations", "{", "(", "cos⁢iπ4", ",", "sin⁢iπ4", ")", "}=i⁢1", "…", "8", ".", "At", "this", "point", ",", "we", "obtain", "8", "oriented", "gradient", "maps", ".", "We", "smooth", "each", "map", "with", "a", "Gaussian", "filter", "of", "radius", "ν2", ".", "Next", "we", "cap", "strong", "gradients", "using", "a", "sigmoid", "↦x", "-/", "2", "(+", "1exp", "(-", "⁢ςx", "))", "1", ",", "to", "help", "canceling", "out", "effects", "of", "varying", "illumination", ".", "We", "smooth", "gradients", "one", "more", "time", "for", "each", "orientation", "with", "a", "Gaussian", "filter", "of", "radius", "ν3", ".", "Finally", ",", "the", "descriptor", "for", "each", "pixel", "is", "obtained", "by", "the", "ℓ2", "-", "normalized", "concatenation", "of", "8", "oriented", "gradients", "and", "a", "ninth", "small", "constant", "value", "μ", ".", "Appending", "μ", "amounts", "to", "adding", "a", "regularizer", "that", "will", "reduce", "the", "importance", "of", "small", "gradients", "(", "i.e.", "noise", ")", "and", "ensures", "that", "two", "pixels", "lying", "in", "areas", "without", "gradient", "information", "will", "still", "correlate", "positively", ".", "Pixel", "descriptors", "Ri", ",", "j", "are", "compared", "using", "dot", "-", "product", "and", "the", "similarity", "function", "takes", "value", "in", "the", "interval", "[", "0", ",", "1", "]", ".", "In", "Section", ",", "we", "evaluate", "the", "impact", "of", "the", "parameters", "of", "this", "pixel", "descriptor", ".", "paragraph", ":", "Bottom", "-", "level", "correlation", "map", ":", "We", "can", "express", "the", "correlation", "map", "computation", "obtained", "from", "Eq", ".", "(", "[", "reference", "]", ")", "more", "conveniently", "in", "a", "convolutional", "framework", ".", "Let", "be", "a", "patch", "of", "size", "from", "the", "first", "image", "centered", "at", "(", "is", "a", "power", "of", "2", ")", ".", "Let", "be", "a", "grid", "with", "step", "pixels", ".", "is", "the", "set", "of", "the", "centers", "of", "the", "atomic", "patches", ".", "For", "each", ",", "we", "convolve", "the", "flipped", "patch", "over", "to", "get", "the", "correlation", "map", ",", "where", ".F", "denotes", "an", "horizontal", "and", "vertical", "flipThis", "amounts", "to", "the", "cross", "-", "correlation", "of", "the", "patch", "and", "I′.", ".", "For", "any", "pixel", "of", ",", "is", "a", "measure", "of", "similarity", "between", "and", ".", "Examples", "of", "such", "correlation", "maps", "are", "shown", "in", "Figure", "[", "reference", "]", "and", "Figure", "[", "reference", "]", ".", "Without", "surprise", "we", "can", "observe", "that", "atomic", "patches", "are", "not", "discriminative", ".", "Recursive", "aggregation", "of", "patches", "in", "subsequent", "stages", "will", "be", "the", "key", "to", "create", "discriminative", "responses", ".", "paragraph", ":", "Iteration", ".", "We", "then", "compute", "the", "correlation", "maps", "of", "larger", "patches", "by", "aggregating", "those", "of", "smaller", "patches", ".", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "a", "patch", "is", "the", "concatenation", "of", "patches", "of", "size", ":", "They", "correspond", "respectively", "to", "the", "bottom", "-", "left", ",", "top", "-", "left", ",", "bottom", "-", "right", "and", "top", "-", "right", "quadrants", ".", "The", "correlation", "map", "of", "can", "thus", "be", "computed", "using", "its", "children", "’s", "correlation", "maps", ".", "For", "the", "sake", "of", "clarity", ",", "we", "define", "the", "short", "-", "hand", "notation", "=", "sN", ",", "i⁢N4oi", "describing", "the", "positional", "shift", "of", "a", "children", "patch", "∈i", "[", "0", ",", "3", "]", "relatively", "to", "its", "parent", "patch", "(", "see", "Figure", ")", ".", "Using", "the", "above", "notations", ",", "we", "rewrite", "Eq", ".", "(", ")", "by", "replacing", "=", "⁢def⁢sim", "(", "R", ",", "R′", ")", "⁢CN", ",", "p", "(", "p′", ")", "(", "i.e.", "assuming", "here", "that", "patch", "=", "RIN", ",", "p", "and", "that", "R′", "is", "centered", "at", "∈p′I′", ")", ".", "Similarly", ",", "we", "replace", "the", "similarity", "between", "children", "patches", "⁢sim", "(", "Ri", ",", "⁢Ri′", "(", "p′i", ")", ")", "by", "⁢CN2", ",+", "psN", ",", "i", "(", "pi′", ")", ".", "For", "each", "child", ",", "we", "retain", "the", "maximum", "similarity", "over", "a", "small", "neighborhood", "Θi", "of", "width", "and", "height", "N8", "centered", "at", "+", "p′sN", ",", "i", ".", "We", "then", "obtain", ":", "We", "now", "explain", "how", "we", "can", "break", "down", "Eq", ".", "(", ")", "into", "a", "succession", "of", "simple", "operations", ".", "First", ",", "let", "us", "assume", "that", "=", "N×42ℓ", ",", "where", "≥ℓ1", "is", "the", "current", "iteration", ".", "During", "iteration", "ℓ", ",", "we", "want", "to", "compute", "the", "correlation", "maps", "CN", ",", "p", "of", "every", "patch", "IN", ",", "p", "from", "the", "first", "image", "for", "which", "correlation", "maps", "of", "its", "children", "have", "been", "computed", "in", "the", "previous", "iteration", ".", "Formally", ",", "the", "position", "GN", "of", "such", "patches", "is", "defined", "according", "to", "the", "position", "of", "children", "patches", "GN2", "according", "to", "Eq", ".", "(", ")", ":", "We", "observe", "that", "the", "larger", "a", "patch", "is", "(", "i.e", ".", "after", "several", "iterations", ")", ",", "the", "smaller", "the", "spatial", "variation", "of", "its", "correlation", "map", "(", "see", "Figure", "[", "reference", "]", ")", ".", "This", "is", "due", "to", "the", "statistics", "of", "natural", "images", ",", "in", "which", "low", "frequencies", "significantly", "dominate", "over", "high", "frequencies", ".", "As", "a", "consequence", ",", "we", "choose", "to", "subsample", "each", "map", "CN", ",", "p", "by", "a", "factor", "2", ".", "We", "express", "this", "with", "an", "operator", ":", "The", "subsampling", "reduces", "by", "the", "area", "of", "the", "correlation", "maps", "and", ",", "as", "a", "direct", "consequence", ",", "the", "computational", "requirements", ".", "Instead", "of", "computing", "the", "subsampling", "on", "top", "of", "Eq", ".", "(", ")", ",", "it", "is", "actually", "more", "efficient", "to", "propagate", "it", "towards", "the", "children", "maps", "and", "perform", "it", "jointly", "with", "max", "-", "pooling", ".", "It", "also", "makes", "the", "max", "-", "pooling", "domain", "become", "independent", "from", "in", "the", "subsampled", "maps", ",", "as", "it", "exactly", "cancels", "out", "the", "effect", "of", "doubling", "at", "each", "iteration", ".", "We", "call", "the", "max", "-", "pooling", "operator", "with", "the", "iteration", "-", "independent", "domain", ":", "For", "the", "same", "reason", ",", "the", "shift", "applied", "to", "the", "correlation", "maps", "in", "’s", "definition", "becomes", "simply", "after", "subsampling", ".", "Let", "be", "the", "shift", "(", "or", "translation", ")", "operator", "on", "the", "correlation", "map", ":", "Finally", ",", "we", "incorporate", "an", "additional", "non", "-", "linear", "mapping", "at", "each", "iteration", "on", "top", "of", "Eq", ".", "(", "[", "reference", "]", ")", "by", "applying", "a", "power", "transform", "MalikPerona90", ",", "LeCun98", ":", "This", "step", ",", "commonly", "referred", "to", "as", "rectification", ",", "is", "added", "in", "order", "to", "better", "propagate", "high", "correlations", "after", "each", "level", ",", "or", ",", "in", "other", "words", ",", "to", "counterbalance", "the", "fact", "that", "max", "-", "pooling", "tends", "to", "retain", "only", "high", "scores", ".", "Indeed", ",", "its", "effect", "is", "to", "decrease", "the", "correlation", "values", "(", "which", "are", "in", "[", "0", ",", "1", "]", ")", "as", "we", "use", ">", "λ1", ".", "Such", "post", "-", "processing", "is", "commonly", "used", "in", "deep", "convolutional", "networks", "lecun", "-", "98b", ",", "Bengio09", ".", "In", "practice", ",", "good", "performance", "is", "obtained", "with", ",", "see", "Section", "[", "reference", "]", ".", "The", "final", "expression", "of", "Eq", ".", "(", "[", "reference", "]", ")", "is", ":", "Figure", "[", "reference", "]", "illustrates", "the", "computation", "of", "correlation", "maps", "for", "different", "patch", "sizes", "and", "Algorithm", "[", "reference", "]", "summarizes", "our", "approach", ".", "The", "resulting", "set", "of", "correlation", "maps", "across", "iterations", "is", "referred", "to", "as", "multi", "-", "level", "correlation", "pyramid", ".", "paragraph", ":", "Boundary", "effects", ":", "In", "practice", ",", "a", "patch", "IN", ",", "p", "can", "overlap", "with", "the", "image", "boundary", ",", "as", "long", "as", "its", "center", "p", "remains", "inside", "the", "image", "(", "from", "Eq", ".", "(", ")", ")", ".", "For", "instance", ",", "a", "patch", "IN", ",", "p0", "with", "center", "at", "p0=", "(", "0", ",", "0", ")", "∈GN", "has", "only", "a", "single", "valid", "child", "(", "the", "one", "for", "which", "=", "i3", "as", "∈", "+", "p0sN", ",", "3I", ")", ".", "In", "such", "degenerate", "cases", ",", "the", "average", "sum", "in", "Eq", ".", "(", ")", "is", "carried", "out", "on", "valid", "children", "only", ".", "For", "IN", ",", "p0", ",", "it", "thus", "only", "comprises", "one", "term", "weighted", "by", "1", "instead", "of", "14", ".", "Note", "that", "Eq", ".", "(", "[", "reference", "]", ")", "implicitly", "defines", "the", "set", "of", "possible", "displacements", "of", "the", "approach", ",", "see", "Figures", "[", "reference", "]", "and", "[", "reference", "]", ".", "Given", "the", "position", "of", "a", "parent", "patch", ",", "each", "child", "patch", "can", "move", "only", "within", "a", "small", "extent", ",", "equal", "to", "the", "quarter", "of", "its", "own", "size", ".", "Figure", "[", "reference", "]", "shows", "the", "correlation", "maps", "for", "patches", "of", "size", ",", "and", ".", "Clearly", ",", "correlation", "maps", "for", "larger", "patch", "are", "more", "and", "more", "discriminative", ",", "while", "still", "allowing", "non", "-", "rigid", "matching", ".", "Input", ":", "Images", ",", "For", "do", "(", "convolution", ",", "Eq", ".", "(", "[", "reference", "]", ")", ")", "(", "rectification", ",", "Eq", ".", "(", "[", "reference", "]", ")", ")", "While", "do", "For", "do", "(", "max", "-", "pooling", "and", "subsampling", ")", "For", "do", "(", "shift", "and", "average", ")", "(", "rectification", ",", "Eq", ".", "(", "[", "reference", "]", ")", ")", "Return", "the", "multi", "-", "level", "correlation", "pyramid", "Computing", "the", "multi", "-", "level", "correlation", "pyramid", ".", "subsection", ":", "Top", "-", "down", "correspondence", "extraction", "A", "score", "in", "the", "multi", "-", "level", "correlation", "pyramid", "represents", "the", "deformation", "-", "tolerant", "similarity", "of", "two", "patches", "and", ".", "Since", "this", "score", "is", "built", "from", "the", "similarity", "of", "4", "matching", "sub", "-", "patches", "at", "the", "lower", "pyramid", "level", ",", "we", "can", "thus", "recursively", "backtrack", "a", "set", "of", "correspondences", "to", "the", "bottom", "level", "(", "corresponding", "to", "matches", "of", "atomic", "patches", ")", ".", "In", "this", "section", ",", "we", "first", "describe", "this", "backtracking", ".", "We", ",", "then", ",", "present", "the", "procedure", "for", "merging", "atomic", "correspondences", "backtracked", "from", "different", "entry", "points", "in", "the", "multi", "-", "level", "pyramid", ",", "which", "constitute", "the", "final", "output", "of", "DeepMatching", ".", "Compared", "to", "our", "initial", "version", "of", "DeepMatching", "DeepFlow", ",", "we", "have", "updated", "match", "scoring", "and", "entry", "point", "selection", "to", "optimize", "the", "execution", "time", "and", "the", "matching", "accuracy", ".", "A", "quantitative", "comparison", "is", "provided", "in", "Section", "[", "reference", "]", ".", "paragraph", ":", "Backtracking", "atomic", "correspondences", ".", "Given", "an", "entry", "point", "in", "the", "pyramid", "(", "i.e", ".", "a", "match", "between", "two", "patches", "and", ")", ",", "we", "retrieve", "atomic", "correspondences", "by", "successively", "undoing", "the", "steps", "used", "to", "aggregate", "correlation", "maps", "during", "the", "pyramid", "construction", ",", "see", "Figure", "[", "reference", "]", ".", "The", "entry", "patch", "is", "itself", "composed", "of", "four", "moving", "quadrants", ",", ".", "Due", "to", "the", "subsampling", ",", "the", "quadrant", "matches", "with", "where", "For", "the", "sake", "of", "clarity", ",", "we", "define", "the", "short", "-", "hand", "notations", "and", ".", "Let", "be", "the", "function", "that", "assigns", "to", "a", "tuple", ",", "representing", "a", "correspondence", "between", "pixel", "and", "for", "patch", "of", "size", "with", "a", "score", ",", "the", "set", "of", "the", "correspondences", "of", "children", "patches", ":", "Given", "a", "set", "of", "such", "tuples", ",", "let", "be", "the", "union", "of", "the", "sets", "for", "all", ".", "Note", "that", "if", "all", "candidate", "correspondences", "corresponds", "to", "atomic", "patches", ",", "then", ".", "Thus", ",", "the", "algorithm", "for", "backtracking", "correspondences", "is", "the", "following", ".", "Consider", "an", "entry", "match", ".", "We", "repeatedly", "apply", "on", ".", "After", "calls", ",", "we", "get", "one", "correspondence", "for", "each", "of", "the", "atomic", "patches", ".", "Furthermore", ",", "their", "score", "is", "equal", "to", "the", "sum", "of", "all", "patch", "similarities", "along", "their", "backtracking", "path", ".", "paragraph", ":", "Merging", "correspondences", ".", "We", "have", "shown", "how", "to", "retrieve", "atomic", "correspondences", "from", "a", "match", "between", "two", "deformable", "(", "potentially", "large", ")", "patches", ".", "Despite", "this", "flexibility", ",", "a", "single", "match", "is", "unlikely", "to", "explain", "the", "complex", "set", "of", "motions", "that", "can", "occur", ",", "for", "example", ",", "between", "two", "adjacent", "frames", "in", "a", "video", ",", "i.e", ".", ",", "two", "objects", "moving", "independently", "with", "significantly", "different", "motions", "exceeds", "the", "deformation", "range", "of", "DeepMatching", ".", "We", "quantitatively", "specify", "this", "range", "in", "the", "next", "subsection", ".", "We", "thus", "merge", "atomic", "correspondences", "gathered", "from", "different", "entry", "points", "(", "matches", ")", "in", "the", "pyramid", ".", "In", "the", "initial", "version", "of", "DeepMatching", "DeepFlow", ",", "entry", "points", "were", "local", "maxima", "over", "all", "correlation", "maps", ".", "This", "is", "now", "replaced", "by", "a", "faster", "procedure", ",", "that", "starts", "with", "all", "possible", "matches", "in", "the", "top", "pyramid", "level", "(", "i.e", ".", ")", ".", "Using", "this", "level", "only", "results", "in", "significantly", "less", "entry", "points", "than", "starting", "from", "all", "maxima", "in", "the", "entire", "pyramid", ".", "We", "did", "not", "observe", "any", "impact", "on", "the", "matching", "performance", ",", "see", "Section", "[", "reference", "]", ".", "Because", "contains", "a", "lot", "of", "overlapping", "patches", ",", "most", "of", "the", "computation", "during", "repeated", "calls", "to", "can", "be", "factorized", ".", "In", "other", "words", ",", "as", "soon", "as", "two", "tuples", "in", "are", "equal", "in", "terms", "of", ",", "and", ",", "the", "one", "with", "the", "lowest", "score", "is", "simply", "eliminated", ".", "We", "thus", "obtain", "a", "set", "of", "atomic", "correspondences", ":", "that", "we", "filter", "with", "reciprocal", "match", "verification", ".", "The", "final", "set", "of", "correspondences", "is", "obtained", "as", ":", "where", "(", "resp", ".", ")", "returns", "the", "best", "match", "in", "a", "small", "vicinity", "of", "pixels", "around", "in", "(", "resp", ".", "around", "in", ")", "from", ".", "subsection", ":", "Discussion", "and", "Analysis", "of", "DeepMatching", "paragraph", ":", "Multi", "-", "size", "patches", "and", "repetitive", "textures", ".", "During", "the", "bottom", "-", "up", "pass", "of", "the", "algorithm", ",", "we", "iteratively", "aggregate", "correlation", "maps", "of", "smaller", "patches", "to", "form", "the", "correlation", "maps", "of", "larger", "patches", ".", "Doing", "so", ",", "we", "effectively", "consider", "patches", "of", "different", "sizes", "(", ")", ",", "in", "contrast", "to", "most", "existing", "matching", "methods", ".", "This", "is", "a", "key", "feature", "of", "our", "approach", "when", "dealing", "with", "repetitive", "textures", ".", "As", "one", "moves", "up", "to", "upper", "levels", ",", "the", "matching", "problem", "gets", "less", "ambiguous", ".", "Hence", ",", "our", "method", "can", "correctly", "match", "repetitive", "patterns", ",", "see", "for", "instance", "Figure", "[", "reference", "]", ".", "paragraph", ":", "Quasi", "-", "dense", "correspondences", ".", "Our", "method", "retrieves", "dense", "correspondences", "for", "every", "single", "match", "between", "large", "regions", "(", "i.e", ".", "entry", "point", "for", "the", "backtracking", "in", "the", "top", "-", "level", "correlation", "maps", ")", ",", "even", "in", "weakly", "textured", "areas", ";", "this", "is", "in", "contrast", "to", "correspondences", "obtained", "when", "matching", "descriptors", "(", "e.g", ".", "SIFT", ")", ".", "A", "quantitative", "assessment", ",", "which", "compares", "the", "coverage", "of", "matches", "obtained", "with", "several", "matching", "schemes", ",", "is", "given", "in", "Section", "[", "reference", "]", ".", "paragraph", ":", "Non", "-", "rigid", "deformations", ".", "Our", "matching", "algorithm", "is", "able", "to", "cope", "with", "various", "sources", "of", "image", "deformations", ":", "object", "-", "induced", "or", "camera", "-", "induced", ".", "The", "set", "of", "feasible", "deformations", ",", "explicitly", "defined", "by", "Eq", ".", "(", "[", "reference", "]", ")", ",", "theoretically", "allows", "to", "deal", "with", "a", "scaling", "factor", "in", "the", "range", "and", "rotations", "approximately", "in", "the", "range", ".", "Note", "also", "that", "DeepMatching", "is", "translation", "-", "invariant", "by", "construction", ",", "thanks", "to", "the", "convolutional", "nature", "of", "the", "processing", ".", "Given", "a", "patch", "of", "size", "located", "at", "level", ",", "Eq", ".", "(", "[", "reference", "]", ")", "allows", "each", "of", "its", "children", "patches", "to", "move", "by", "at", "most", "pixels", "from", "their", "ideal", "location", "in", ".", "By", "recursively", "summing", "the", "displacements", "at", "each", "level", ",", "the", "maximal", "displacements", "for", "an", "atomic", "patch", "is", ".", "An", "example", "is", "given", "in", "Figure", "[", "reference", "]", "with", "and", ".", "Relatively", "to", ",", "we", "thus", "have", "and", ".", "For", "a", "rotation", ",", "the", "rationale", "is", "similar", ",", "see", "Figure", "[", "reference", "]", ".", "Note", "that", "the", "displacement", "tolerance", "in", "Θi", "from", "Eq", ".", "(", ")", "could", "be", "extended", "to", "/", "×xN8", "pixels", "with", "∈x{2", ",", "3", ",", "…", "}", "(", "instead", "of", "=", "x1", ")", ".", "Then", "the", "above", "formula", "for", "computing", "the", "lower", "bound", "on", "the", "scale", "factor", "of", "DeepMatching", "generalizes", "to", "=", "⁢LB", "(", "x", ")", "lim→N∞", "/(-", "N⁢2xdN", ")", "N.", "Hence", ",", "for", "≥x2", "we", "obtain", "<", "⁢LB", "(", "x", ")", "0", "instead", "of", "=", "⁢LB", "(", "1", ")", "12", ".", "This", "implies", "that", "the", "deformation", "range", "is", "extended", "to", "a", "point", "where", "any", "patch", "can", "be", "matched", "to", "a", "single", "pixel", ",", "i.e.", ",", "this", "results", "in", "unrealistic", "deformations", ".", "For", "this", "reason", ",", "we", "choose", "to", "not", "expand", "the", "deformation", "range", "of", "DeepMatching", ".", "paragraph", ":", "Built", "-", "in", "smoothing", ".", "Furthermore", ",", "correspondences", "generated", "through", "backtracking", "of", "a", "single", "entry", "point", "in", "the", "correlation", "maps", "are", "naturally", "smooth", ".", "Indeed", ",", "feasible", "deformations", "can", "not", "be", "too", "‘", "‘", "far", "’", "’", "from", "the", "identity", "deformation", ".", "To", "verify", "this", "assumption", ",", "we", "conduct", "the", "following", "experiment", ".", "We", "artificially", "generate", "two", "types", "of", "correspondences", "between", "two", "images", "of", "size", "×128128", ".", "The", "first", "one", "is", "completely", "random", ",", "i.e.", "for", "each", "atomic", "patch", "in", "the", "first", "image", "we", "assign", "randomly", "a", "match", "in", "the", "second", "image", ".", "The", "second", "one", "respects", "the", "backtracking", "constraints", ".", "Starting", "from", "a", "single", "entry", "point", "in", "the", "top", "level", "we", "simulate", "the", "backtracking", "procedure", "from", "Section", "by", "replacing", "in", "Eq", ".", "(", ")", "the", "max", "operation", "by", "a", "random", "sampling", "over", "{", "-", "1", ",", "0", ",", "1}2", ".", "By", "generating", "10", ",", "000", "sets", "of", "possible", "atomic", "correspondences", ",", "we", "simulate", "a", "set", "which", "respects", "the", "deformations", "allowed", "by", "DeepMatching", ".", "Figure", "[", "reference", "]", "compares", "the", "smoothness", "of", "these", "two", "types", "of", "artificial", "correspondences", ".", "Smoothness", "is", "measured", "by", "interpreting", "the", "correspondences", "as", "flow", "and", "measuring", "the", "gradient", "flow", "norm", ",", "see", "Eq", ".", "(", ")", ".", "Clearly", ",", "the", "two", "types", "of", "warpings", "are", "different", "by", "orders", "of", "magnitude", ".", "Furthermore", ",", "the", "one", "which", "respects", "the", "built", "-", "in", "constraints", "of", "DeepMatching", "is", "close", "to", "the", "identity", "warping", ".", "paragraph", ":", "Relation", "to", "Deep", "Convolutional", "Neural", "Networks", "(", "CNNs", ")", ".", "DeepMatching", "relies", "on", "a", "hierarchical", ",", "multi", "-", "layer", ",", "correlational", "architecture", "designed", "for", "matching", "images", "and", "was", "inspired", "by", "deep", "convolutional", "approaches", "In", "the", "following", "we", "describe", "the", "major", "similarities", "and", "differences", ".", "Deep", "networks", "learn", "from", "data", "the", "weights", "of", "the", "convolutions", ".", "In", "contrast", ",", "DeepMatching", "does", "not", "learn", "any", "feature", "representations", "and", "instead", "directly", "computes", "correlations", "at", "the", "patch", "level", ".", "It", "uses", "patches", "from", "the", "first", "image", "as", "convolution", "filters", "for", "the", "second", "one", ".", "However", ",", "the", "bottom", "-", "up", "pipeline", "of", "DeepMatching", "is", "similar", "to", "CNNs", ".", "It", "alternates", "aggregating", "channels", "from", "the", "previous", "layer", "with", "channel", "-", "wise", "max", "-", "pooling", "and", "subsampling", ".", "As", "in", "CNNs", ",", "max", "-", "pooling", "in", "DeepMatching", "allows", "for", "invariance", "w.r.t", ".", "small", "deformations", ".", "Likewise", ",", "the", "algorithm", "propagates", "pairwise", "patch", "similarity", "scores", "through", "the", "hierarchy", "using", "non", "-", "linear", "rectifying", "stages", "in", "-", "between", "layers", ".", "Finally", ",", "DeepMatching", "includes", "a", "top", "-", "down", "pass", "which", "is", "not", "present", "in", "CNNs", ".", "paragraph", ":", "Time", "and", "space", "complexity", ".", "DeepMatching", "has", "a", "complexity", "in", "memory", "and", "time", ",", "where", "and", "are", "the", "number", "of", "pixels", "per", "image", ".", "Computing", "the", "initial", "correlations", "is", "a", "operation", ".", "Then", ",", "at", "each", "level", "of", "the", "pyramid", ",", "the", "process", "is", "repeated", "while", "the", "complexity", "is", "divided", "by", "a", "factor", "due", "to", "the", "subsampling", "step", "in", "the", "target", "image", "(", "since", "the", "cardinality", "of", "remains", "approximately", "constant", ")", ".", "Thus", ",", "the", "total", "complexity", "of", "the", "correlation", "maps", "computation", "is", ",", "at", "worst", ",", ".", "During", "the", "top", "-", "down", "pass", ",", "most", "backtracking", "paths", "can", "be", "pruned", "as", "soon", "as", "they", "cross", "a", "concurrent", "path", "with", "a", "higher", "score", "(", "see", "Section", "[", "reference", "]", ")", ".", "Thus", ",", "all", "correlations", "will", "be", "examined", "at", "most", "once", ",", "and", "there", "are", "values", "in", "total", ".", "However", ",", "this", "analysis", "is", "worst", "-", "case", ".", "In", "practice", ",", "only", "correlations", "lying", "on", "maximal", "paths", "are", "actually", "examined", ".", "section", ":", "Extensions", "of", "DeepMatching", "subsection", ":", "Approximate", "DeepMatching", "As", "a", "consequence", "of", "its", "space", "complexity", ",", "DeepMatching", "requires", "an", "amount", "of", "RAM", "that", "is", "orders", "of", "magnitude", "above", "other", "state", "-", "of", "-", "the", "-", "art", "matching", "methods", ".", "This", "could", "correspond", "to", "several", "gigabytes", "for", "images", "of", "moderate", "size", "(", "800", "600", "pixels", ")", ";", "see", "Section", "[", "reference", "]", ".", "This", "section", "introduces", "an", "approximation", "of", "DeepMatching", "that", "allows", "to", "trade", "matching", "quality", "for", "reduced", "time", "and", "memory", "usage", ".", "As", "shown", "in", "Section", "[", "reference", "]", ",", "near", "-", "optimal", "results", "can", "be", "obtained", "at", "a", "fraction", "of", "the", "original", "cost", ".", "Our", "approximation", "proposes", "to", "compress", "the", "representation", "of", "atomic", "patches", ".", "Atomic", "patches", "carry", "little", "information", ",", "and", "thus", "are", "highly", "redundant", ".", "For", "instance", ",", "in", "uniform", "regions", ",", "all", "patches", "are", "nearly", "identical", "(", "i.e", ".", ",", "gradient", "-", "wise", ")", ".", "To", "exploit", "this", "property", ",", "we", "index", "atomic", "patches", "with", "a", "small", "set", "of", "patch", "prototypes", ".", "We", "substitute", "each", "patch", "with", "its", "closest", "neighbor", "in", "a", "fixed", "dictionary", "of", "prototypes", ".", "Hence", ",", "we", "need", "to", "perform", "and", "store", "only", "convolutions", "at", "the", "first", "level", ",", "instead", "of", "(", "with", ")", ".", "This", "significantly", "reduces", "both", "memory", "and", "time", "complexity", ".", "Note", "that", "higher", "pyramid", "levels", "also", "benefit", "from", "this", "optimization", ".", "Indeed", ",", "two", "parent", "patches", "at", "the", "second", "level", "have", "the", "exact", "same", "correlation", "map", "in", "case", "their", "children", "are", "assigned", "the", "same", "prototypes", ".", "The", "same", "reasoning", "also", "holds", "for", "all", "subsequent", "levels", ",", "but", "the", "gains", "rapidly", "diminish", "due", "to", "statistical", "unlikeliness", "of", "the", "required", "condition", ".", "This", "is", "not", "really", "an", "issue", ",", "since", "the", "memory", "and", "computational", "cost", "mostly", "rests", "on", "the", "initial", "levels", ";", "see", "Section", "[", "reference", "]", ".", "In", "practice", ",", "we", "build", "the", "prototype", "dictionary", "using", "k", "-", "means", ",", "as", "it", "is", "designed", "to", "minimize", "the", "approximation", "error", "between", "input", "descriptors", "and", "resulting", "centroids", "(", "i.e", ".", "prototypes", ")", ".", "Given", "a", "pair", "of", "images", "to", "match", ",", "we", "perform", "on", "-", "line", "clustering", "of", "all", "descriptors", "of", "atomic", "patches", "in", "the", "first", "image", ".", "Since", "the", "original", "descriptors", "lie", "on", "an", "hypersphere", "(", "each", "pixel", "descriptor", "has", "norm", "1", ")", ",", "we", "modify", "the", "k", "-", "means", "approach", "so", "as", "to", "project", "the", "estimated", "centroids", "on", "the", "hypersphere", "at", "each", "iteration", ".", "We", "find", "experimentally", "that", "this", "is", "important", "to", "obtain", "good", "results", ".", "subsection", ":", "Scale", "and", "rotation", "invariant", "DeepMatching", "For", "a", "variety", "of", "tasks", ",", "objects", "to", "be", "matched", "can", "appear", "under", "image", "rotations", "or", "at", "different", "scales", "Lowe2004", ",", "Mikolajczyk2005", ",", "Szeliski2010", ",", "HaCohen2011", ".", "As", "discussed", "above", ",", "DeepMatching", "(", "DM", ")", "is", "only", "robust", "to", "moderate", "scale", "changes", "and", "rotations", ".", "We", "now", "present", "a", "scale", "and", "rotation", "invariant", "version", ".", "The", "approach", "is", "straightforward", ":", "we", "apply", "DM", "to", "several", "rotated", "and", "scaled", "versions", "of", "the", "second", "image", ".", "According", "to", "the", "invariance", "range", "of", "DM", ",", "we", "use", "steps", "of", "for", "image", "rotation", "and", "power", "of", "for", "scale", "changes", ".", "While", "iterating", "over", "all", "combinations", "of", "scale", "changes", "and", "rotations", ",", "we", "maintain", "a", "list", "of", "all", "atomic", "correspondences", "obtained", "so", "far", ",", "i.e", ".", "corresponding", "positions", "and", "scores", ".", "As", "before", ",", "the", "final", "output", "correspondences", "consists", "of", "the", "reciprocal", "matches", "in", ".", "Storing", "all", "matches", "and", "finally", "choosing", "the", "best", "ones", "based", "on", "reciprocal", "verification", "permits", "to", "capture", "distinct", "motions", "possibly", "occurring", "together", "in", "the", "same", "scene", "(", "e.g", ".", "one", "object", "could", "have", "undergone", "a", "rotation", ",", "while", "the", "rest", "of", "the", "scene", "did", "not", "move", ")", ".", "The", "steps", "of", "the", "approach", "are", "described", "in", "Algorithm", "[", "reference", "]", ".", "Since", "we", "iterate", "sequentially", "over", "a", "fixed", "list", "of", "rotations", "and", "scale", "changes", ",", "the", "space", "and", "time", "complexity", "of", "the", "algorithm", "remains", "unchanged", "(", "i.e", ".", ")", ".", "In", "practice", ",", "the", "run", "-", "time", "compared", "to", "DM", "is", "multiplied", "by", "a", "constant", "approximately", "equal", "to", "25", ",", "see", "Section", "[", "reference", "]", ".", "Note", "that", "the", "algorithm", "permits", "a", "straightforward", "parallelization", ".", "Input", ":", ",", "are", "the", "images", "to", "be", "matched", "Initialize", "an", "empty", "set", "of", "correspondences", "For", "do", "#", "either", "downsize", "image", "1", "#", "or", "downsize", "image", "2", "For", "∈θ{0", ",", "⁢π4", "…", ",", "⁢7π4", "}", "do", "#", "get", "raw", "atomic", "correspondences", "(", "Eq", ".", "(", ")", ")", "#", "Geometric", "rectification", "to", "the", "input", "image", "space", ":", "#", "Concatenate", "results", ":", "#", "keep", "reciprocal", "matches", "(", "Eq", ".", "(", ")", ")", "Return", "Scale", "and", "rotation", "invariant", "version", "of", "DeepMatching", "(", "DM", ")", ".", "Iσ", "denotes", "the", "image", "I", "downsized", "by", "a", "factor", "σ", ",", "and", "Rθ", "denotes", "rotation", "by", "an", "angle", "θ", ".", "subsection", ":", "DeepFlow", "We", "now", "present", "our", "approach", "for", "optical", "flow", "estimation", ",", "DeepFlow", ".", "We", "adopt", "the", "method", "introduced", "by", "[", "]", ",", "where", "a", "matching", "term", "penalizes", "the", "differences", "between", "optical", "flow", "and", "input", "matches", ",", "and", "replace", "their", "matching", "approach", "by", "DeepMatching", ".", "In", "addition", ",", "we", "make", "a", "few", "minor", "modifications", "introduced", "recently", "in", "the", "state", "of", "the", "art", ":", "(", "i", ")", "we", "add", "a", "normalization", "in", "the", "data", "term", "to", "downweight", "the", "impact", "of", "locations", "with", "high", "spatial", "image", "derivatives", "(", "ii", ")", "we", "use", "a", "different", "weight", "at", "each", "level", "to", "downweight", "the", "matching", "term", "at", "finer", "scales", "and", "(", "iii", ")", "the", "smoothness", "term", "is", "locally", "weighted", "Let", "be", "two", "consecutive", "images", "defined", "on", "with", "channels", ".", "The", "goal", "is", "to", "estimate", "the", "flow", ".", "We", "assume", "that", "the", "images", "are", "already", "smoothed", "using", "a", "Gaussian", "filter", "of", "standard", "deviation", ".", "The", "energy", "we", "optimize", "is", "a", "weighted", "sum", "of", "a", "data", "term", ",", "a", "smoothness", "term", "and", "a", "matching", "term", ":", "For", "the", "three", "terms", ",", "we", "use", "a", "robust", "penalizer", "with", "which", "has", "shown", "excellent", "results", "Sun2014", ".", "paragraph", ":", "Data", "term", ".", "The", "data", "term", "is", "a", "separate", "penalization", "of", "the", "color", "and", "gradient", "constancy", "assumptions", "with", "a", "normalization", "factor", "as", "proposed", "by", ".", "We", "start", "from", "the", "optical", "flow", "constraint", "assuming", "brightness", "constancy", ":", "the", "spatio", "-", "temporal", "gradient", ".", "A", "basic", "way", "to", "build", "a", "data", "term", "is", "to", "penalize", "it", ",", "i.e", ".", "with", "the", "tensor", "defined", "by", ".", "As", "highlighted", "by", ",", "such", "a", "data", "term", "adds", "a", "higher", "weight", "in", "locations", "corresponding", "to", "high", "spatial", "image", "derivatives", ".", "We", "normalize", "it", "by", "the", "norm", "of", "the", "spatial", "derivatives", "plus", "a", "small", "factor", "to", "avoid", "division", "by", "zero", ",", "and", "to", "reduce", "a", "bit", "the", "influence", "in", "tiny", "gradient", "locations", "Zimmer2011", ".", "Let", "be", "the", "normalized", "tensor", "with", ".", "We", "set", "in", "the", "following", ".", "To", "deal", "with", "color", "images", ",", "we", "consider", "the", "tensor", "defined", "for", "a", "channel", "denoted", "by", "upper", "indices", "and", "we", "penalize", "the", "sum", "over", "channels", ":", ".", "We", "consider", "images", "in", "the", "RGB", "color", "space", ".", "We", "separately", "penalize", "the", "gradient", "constancy", "assumption", "Bruhn2005", ".", "Let", "and", "be", "the", "derivatives", "of", "the", "images", "with", "respect", "to", "the", "and", "axis", "respectively", ".", "Let", "be", "the", "tensor", "for", "the", "channel", "including", "the", "normalization", "The", "data", "term", "is", "the", "sum", "of", "two", "terms", ",", "balanced", "by", "two", "weights", "and", ":", "paragraph", ":", "Smoothness", "term", ".", "The", "smoothness", "term", "is", "a", "robust", "penalization", "of", "the", "gradient", "flow", "norm", ":", "The", "smoothness", "weight", "α", "is", "locally", "set", "according", "to", "image", "derivatives", "with", "=", "⁢α", "(", "x", ")", "exp", "(-", "⁢κ∇2I", "(", "x", ")", ")", "where", "κ", "is", "experimentally", "set", "to", "=", "κ5", ".", "paragraph", ":", "Matching", "term", ".", "The", "matching", "term", "encourages", "the", "flow", "estimation", "to", "be", "similar", "to", "a", "precomputed", "vector", "field", ".", "To", "this", "end", ",", "we", "penalize", "the", "difference", "between", "and", "using", "the", "robust", "penalizer", ".", "Since", "the", "matching", "is", "not", "totally", "dense", ",", "we", "add", "a", "binary", "term", "which", "is", "equal", "to", "if", "and", "only", "if", "a", "match", "is", "available", "at", ".", "We", "also", "multiply", "each", "matching", "penalization", "by", "a", "weight", ",", "which", "is", "low", "in", "uniform", "regions", "where", "matching", "is", "ambiguous", "and", "when", "matched", "patches", "are", "dissimilar", ".", "To", "that", "aim", ",", "we", "rely", "on", ",", "the", "minimum", "eigenvalue", "of", "the", "autocorrelation", "matrix", "multiplied", "by", ".", "We", "also", "compute", "the", "visual", "similarity", "between", "matches", "as", ".", "We", "then", "compute", "the", "score", "as", "a", "Gaussian", "kernel", "on", "weighted", "by", "with", "a", "parameter", ",", "experimentally", "set", "to", ".", "More", "precisely", ",", "we", "define", "at", "each", "point", "with", "a", "match", "as", ":", "The", "matching", "term", "is", "then", ".", "paragraph", ":", "Minimization", ".", "This", "energy", "objective", "is", "non", "-", "convex", "and", "non", "-", "linear", ".", "To", "solve", "it", ",", "we", "use", "a", "numerical", "optimization", "algorithm", "similar", "as", ".", "An", "incremental", "coarse", "-", "to", "-", "fine", "warping", "strategy", "is", "used", "with", "a", "downsampling", "factor", ".", "The", "remaining", "equations", "are", "still", "non", "-", "linear", "due", "to", "the", "robust", "penalizers", ".", "We", "apply", "5", "inner", "fixed", "point", "iterations", "where", "the", "non", "-", "linear", "weights", "and", "the", "flow", "increments", "are", "iteratively", "updated", "while", "fixing", "the", "other", ".", "To", "approximate", "the", "solution", "of", "the", "linear", "system", ",", "we", "use", "25", "iterations", "of", "the", "Successive", "Over", "Relaxation", "(", "SOR", ")", "method", "Young", ":", "sor", ".", "To", "downweight", "the", "matching", "term", "on", "fine", "scales", ",", "we", "use", "a", "different", "weight", "at", "each", "level", "as", "proposed", "by", ".", "We", "set", "where", "k", "is", "the", "current", "level", "of", "computation", ",", "kmax", "the", "coarsest", "level", "and", "b", "a", "parameter", "which", "is", "optimized", "together", "with", "the", "other", "parameters", ",", "see", "Section", ".", "section", ":", "Experiments", "This", "section", "presents", "an", "experimental", "evaluation", "of", "DeepMatching", "and", "DeepFlow", ".", "The", "datasets", "and", "metrics", "used", "to", "evaluate", "DeepMatching", "and", "DeepFlow", "are", "introduced", "in", "Section", "[", "reference", "]", ".", "Experimental", "results", "are", "given", "in", "Sections", "[", "reference", "]", "and", "[", "reference", "]", "respectively", ".", "subsection", ":", "Datasets", "and", "metrics", "In", "this", "section", "we", "briefly", "introduce", "the", "matching", "and", "flow", "datasets", "used", "in", "our", "experiments", ".", "Since", "consecutive", "frames", "of", "a", "video", "are", "well", "-", "suited", "to", "evaluate", "a", "matching", "approach", ",", "we", "use", "several", "optical", "flow", "datasets", "for", "evaluating", "both", "the", "quality", "of", "matching", "and", "flow", ",", "but", "we", "rely", "on", "different", "metrics", ".", "paragraph", ":", "The", "Mikolajczyk", "dataset", "was", "originally", "proposed", "by", "to", "evaluate", "and", "compare", "the", "performance", "of", "keypoint", "detectors", "and", "descriptors", ".", "It", "is", "one", "of", "the", "standard", "benchmarks", "for", "evaluating", "matching", "approaches", ".", "The", "dataset", "consists", "of", "8", "sequences", "of", "6", "images", "each", "viewing", "a", "scene", "under", "different", "conditions", ",", "such", "as", "illumination", "changes", "or", "viewpoint", "changes", ".", "The", "images", "of", "a", "sequence", "are", "related", "by", "homographies", ".", "During", "the", "evaluation", ",", "we", "comply", "to", "the", "standard", "procedure", "in", "which", "the", "first", "image", "of", "each", "scene", "is", "matched", "to", "the", "5", "remaining", "ones", ".", "Since", "our", "goal", "is", "to", "study", "robustness", "of", "DeepMatching", "to", "geometric", "distortions", ",", "we", "follow", "and", "restrict", "our", "evaluation", "to", "the", "4", "most", "difficult", "sequences", "with", "viewpoint", "changes", ":", "bark", ",", "boat", ",", "graf", "and", "wall", ".", "paragraph", ":", "The", "MPI", "-", "Sintel", "dataset", "sintel", "is", "a", "challenging", "evaluation", "benchmark", "for", "optical", "flow", "estimation", ",", "constructed", "from", "realistic", "computer", "-", "animated", "films", ".", "The", "dataset", "contains", "sequences", "with", "large", "motions", "and", "specular", "reflections", ".", "In", "the", "training", "set", ",", "more", "than", "of", "the", "pixels", "have", "a", "motion", "over", "pixels", ",", "approximately", "over", "pixels", ".", "We", "use", "the", "‘", "‘", "final", "’", "’", "version", ",", "featuring", "rendering", "effects", "such", "as", "motion", "blur", ",", "defocus", "blur", "and", "atmospheric", "effects", ".", "Note", "that", "ground", "-", "truth", "optical", "flows", "for", "the", "test", "set", "are", "not", "publicly", "available", ".", "paragraph", ":", "The", "Middlebury", "dataset", "middlebury", "has", "been", "extensively", "used", "for", "evaluating", "optical", "flow", "methods", ".", "The", "dataset", "contains", "complex", "motions", ",", "but", "most", "of", "the", "motions", "are", "small", ".", "Less", "than", "of", "the", "pixels", "have", "a", "motion", "over", "pixels", ",", "and", "no", "motion", "exceeds", "pixels", "(", "training", "set", ")", ".", "Ground", "-", "truth", "optical", "flows", "for", "the", "test", "set", "are", "not", "publicly", "available", ".", "paragraph", ":", "The", "Kitti", "dataset", "contains", "real", "-", "world", "sequences", "taken", "from", "a", "driving", "platform", ".", "The", "dataset", "includes", "non", "-", "Lambertian", "surfaces", ",", "different", "lighting", "conditions", ",", "a", "large", "variety", "of", "materials", "and", "large", "displacements", ".", "More", "than", "16", "%", "of", "the", "pixels", "have", "motion", "over", "20", "pixels", ".", "Again", ",", "ground", "-", "truth", "optical", "flows", "for", "the", "test", "set", "are", "not", "publicly", "available", ".", "paragraph", ":", "Performance", "metric", "for", "matching", ".", "Choosing", "a", "performance", "measure", "for", "matching", "approaches", "is", "delicate", ".", "Matching", "approaches", "typically", "do", "not", "return", "dense", "correspondences", ",", "but", "output", "varying", "numbers", "of", "matches", ".", "Furthermore", ",", "correspondences", "might", "be", "concentrated", "in", "different", "areas", "of", "the", "image", ".", "Most", "matching", "approaches", ",", "including", "DeepMatching", ",", "are", "based", "on", "establishing", "correspondences", "between", "patches", ".", "Given", "a", "pair", "of", "matching", "patches", ",", "it", "is", "possible", "to", "obtain", "a", "list", "of", "pixel", "correspondences", "for", "all", "pixels", "within", "the", "patches", ".", "We", "introduce", "a", "measure", "based", "on", "the", "number", "of", "correctly", "matched", "pixels", "compared", "to", "the", "overall", "number", "of", "pixels", ".", "We", "define", "‘", "‘", "accuracy@", "’’", "as", "the", "proportion", "of", "‘", "‘", "correct", "’", "’", "pixels", "from", "the", "first", "image", "with", "respect", "to", "the", "total", "number", "of", "pixels", ".", "A", "pixel", "is", "considered", "correct", "if", "its", "pixel", "match", "in", "the", "second", "image", "is", "closer", "than", "pixels", "to", "ground", "-", "truth", ".", "In", "practice", ",", "we", "use", "a", "threshold", "of", "pixels", ",", "as", "this", "represents", "a", "sufficiently", "precise", "estimation", "(", "about", "1", "%", "of", "image", "diagonal", "for", "all", "datasets", ")", ",", "while", "allowing", "some", "tolerance", "in", "blurred", "areas", "that", "are", "difficult", "to", "match", "exactly", ".", "If", "a", "pixel", "belongs", "to", "several", "matches", ",", "we", "choose", "the", "one", "with", "the", "highest", "score", "to", "predict", "its", "correspondence", ".", "Pixels", "which", "do", "not", "belong", "to", "any", "patch", "have", "an", "infinite", "error", ".", "paragraph", ":", "Performance", "metric", "for", "optical", "flow", ".", "To", "evaluate", "optical", "flow", ",", "we", "follow", "the", "standard", "protocol", "and", "measure", "the", "average", "endpoint", "error", "over", "all", "pixels", ",", "denoted", "as", "‘", "‘", "EPE", "’", "’", ".", "The", "‘", "‘", "s10", "-", "40", "’", "’", "variant", "measures", "the", "EPE", "only", "for", "pixels", "with", "a", "ground", "-", "truth", "displacement", "between", "10", "and", "40", "pixels", ",", "and", "likewise", "for", "‘", "‘", "s0", "-", "10", "’", "’", "and", "‘", "‘", "s40", "+", "’", "’", ".", "In", "all", "cases", ",", "scores", "are", "averaged", "over", "all", "image", "pairs", "to", "yield", "the", "final", "result", "for", "a", "given", "dataset", ".", "subsection", ":", "Matching", "Experiments", "In", "this", "section", ",", "we", "evaluate", "DeepMatching", "(", "DM", ")", ".", "We", "present", "results", "for", "all", "datasets", "presented", "above", "but", "Middlebury", ",", "which", "does", "not", "feature", "long", "-", "range", "motions", ",", "the", "main", "difficulty", "in", "image", "matching", ".", "When", "evaluating", "on", "the", "Mikolajczyk", "dataset", ",", "we", "employ", "the", "scale", "and", "rotation", "invariant", "version", "of", "DM", "presented", "in", "Section", "[", "reference", "]", ".", "For", "all", "the", "matching", "experiments", "reported", "in", "this", "section", ",", "we", "use", "the", "Mikolajczyk", "dataset", "and", "the", "training", "sets", "of", "MPI", "-", "Sintel", "and", "Kitti", ".", "subsubsection", ":", "Impact", "of", "the", "parameters", "We", "optimize", "the", "different", "parameters", "of", "DM", "jointly", "on", "all", "datasets", ".", "To", "prevent", "overfitting", ",", "we", "use", "the", "same", "parameters", "across", "all", "datasets", ".", "paragraph", ":", "Pixel", "descriptor", "parameters", ":", "We", "first", "optimize", "the", "parameters", "of", "the", "pixel", "representation", "(", "Section", ")", ":", "ν1", ",", "ν2", ",", "ν3", "(", "different", "smoothing", "stages", ")", ",", "ς", "(", "sigmoid", "slope", ")", "and", "μ", "(", "regularization", "constant", ")", ".", "After", "performing", "a", "grid", "search", ",", "we", "find", "that", "good", "results", "are", "obtained", "at", "ν1=ν2=ν3=1", ",", "=", "ς0.2", "and", "=", "μ0.3", "across", "all", "datasets", ".", "Figure", "shows", "the", "accuracy@10", "in", "the", "neighborhood", "of", "these", "values", "for", "all", "parameters", ".", "Image", "pre", "-", "smoothing", "seems", "to", "be", "crucial", "for", "JPEG", "images", "(", "Mikolajczyk", "dataset", ")", ",", "as", "it", "smooths", "out", "compression", "artifacts", ",", "whereas", "it", "slightly", "degrades", "performance", "for", "uncompressed", "PNG", "images", "(", "MPI", "-", "Sintel", "and", "Kitti", ")", ".", "As", "expected", ",", "similar", "findings", "are", "observed", "for", "the", "regularization", "constant", "μ", "since", "it", "acts", "as", "a", "regularizer", "that", "reduces", "the", "impact", "of", "small", "gradients", "(", "i.e.", "noise", ")", ".", "In", "the", "following", ",", "we", "thus", "use", "low", "values", "of", "ν1", "and", "μ", "when", "dealing", "with", "PNG", "images", "(", "we", "set", "=", "ν10", "and", "=", "μ0.1", ",", "other", "parameters", "are", "unchanged", ")", ".", "paragraph", ":", "Non", "-", "linear", "rectification", ":", "We", "also", "evaluate", "the", "impact", "of", "the", "parameter", "of", "the", "non", "-", "linear", "rectification", "obtained", "by", "applying", "power", "normalization", ",", "see", "Eq", ".", "(", "[", "reference", "]", ")", ".", "Figure", "[", "reference", "]", "displays", "the", "accuracy@10", "for", "various", "values", "of", ".", "We", "can", "observe", "that", "the", "optimal", "performance", "is", "achieved", "at", "for", "all", "datasets", ".", "We", "use", "this", "value", "in", "the", "remainder", "of", "our", "experiments", ".", "subsubsection", ":", "Evaluation", "of", "the", "backtracking", "and", "scoring", "schemes", "We", "now", "evaluate", "two", "improvements", "of", "DM", "with", "respect", "to", "the", "previous", "version", "published", "in", "[", "]", ",", "referred", "to", "as", "DM", "*", ":", "Backtracking", "(", "BT", ")", "entry", "points", ":", "in", "DM", "*", "we", "select", "as", "entry", "points", "local", "maxima", "in", "the", "correlation", "maps", "from", "all", "pyramid", "levels", ".", "The", "new", "alternative", "is", "to", "start", "from", "all", "possible", "points", "in", "the", "top", "pyramid", "level", ".", "Scoring", "scheme", ":", "In", "DM", "*", "we", "scored", "atomic", "correspondences", "based", "on", "the", "correlation", "values", "of", "start", "and", "end", "point", "of", "the", "backtracking", "path", ".", "The", "new", "scoring", "scheme", "is", "the", "sum", "of", "correlation", "values", "along", "the", "full", "backtracking", "path", ".", "We", "report", "results", "for", "the", "different", "variants", "in", "Table", "on", "each", "dataset", ".", "The", "first", "two", "rows", "for", "each", "dataset", "correspond", "to", "the", "exact", "settings", "used", "for", "DM", "*", "(", "i.e.", "with", "an", "image", "resolution", "of", "1", "/", "4", "and", "1", "/", "2", ")", ".", "We", "observe", "a", "steady", "increase", "in", "performance", "on", "all", "datasets", "when", "we", "add", "the", "new", "scoring", "and", "backtracking", "approach", ".", "We", "can", "observe", "that", "starting", "from", "all", "possible", "entry", "points", "in", "the", "top", "pyramid", "level", "(", "i.e.", "considering", "all", "possible", "translations", ")", "yields", "slightly", "better", "results", "than", "starting", "from", "local", "maxima", ".", "This", "demonstrates", "that", "some", "ground", "-", "truth", "matches", "are", "not", "covered", "by", "any", "local", "maximum", ".", "By", "enumerating", "all", "possible", "patch", "translations", "from", "the", "top", "-", "level", ",", "we", "instead", "ensure", "to", "fully", "explore", "the", "space", "of", "all", "possible", "matches", ".", "Furthermore", ",", "it", "is", "interesting", "to", "note", "that", "memory", "usage", "and", "run", "-", "time", "significantly", "decreases", "when", "using", "the", "new", "options", ".", "This", "is", "because", "(", "1", ")", "searching", "and", "storing", "local", "maxima", "(", "which", "are", "exponentially", "more", "numerous", "in", "lower", "pyramid", "levels", ")", "is", "not", "necessary", "anymore", ",", "and", "(", "2", ")", "the", "new", "scoring", "scheme", "allows", "for", "further", "optimization", ",", "i.e.", "early", "pruning", "of", "backtracking", "paths", "(", "Section", ")", ".", "subsubsection", ":", "Approximate", "DeepMatching", "We", "now", "evaluate", "the", "performance", "of", "approximate", "DeepMatching", "(", "Section", "[", "reference", "]", ")", "and", "report", "its", "run", "-", "time", "and", "memory", "usage", ".", "We", "evaluate", "and", "compare", "two", "different", "ways", "of", "reducing", "the", "computational", "load", ".", "The", "first", "one", "simply", "consists", "in", "downsizing", "the", "input", "images", ",", "and", "upscaling", "the", "resulting", "matches", "accordingly", ".", "The", "second", "option", "is", "the", "compression", "scheme", "proposed", "in", "Section", "[", "reference", "]", ".", "We", "evaluate", "both", "schemes", "jointly", "by", "varying", "the", "input", "image", "size", "(", "expressed", "as", "a", "fraction", "of", "the", "original", "resolution", ")", "and", "the", "size", "of", "the", "prototype", "dictionary", "(", "i.e", ".", "parameter", "of", "k", "-", "means", "in", "Section", "[", "reference", "]", ")", ".", "corresponds", "to", "the", "original", "dataset", "image", "size", "(", "no", "downsizing", ")", ".", "We", "display", "the", "results", "in", "terms", "of", "matching", "accuracy", "(", "accuracy@10", ")", "against", "memory", "consumption", "in", "Figure", "[", "reference", "]", "and", "as", "a", "function", "of", "in", "Figure", "[", "reference", "]", ".", "Figure", "[", "reference", "]", "shows", "that", "DeepMatching", "can", "be", "computed", "in", "an", "approximate", "manner", "for", "any", "given", "memory", "budget", ".", "Unsurprisingly", ",", "too", "low", "settings", "(", "e.g", ".", ",", ")", "result", "in", "a", "strong", "loss", "of", "performance", ".", "It", "should", "be", "noted", "that", "that", "we", "were", "unable", "to", "compute", "DeepMatching", "at", "full", "resolution", "(", "for", ",", "as", "the", "memory", "consumption", "explodes", ".", "As", "a", "consequence", ",", "all", "subsequent", "experiments", "in", "the", "paper", "are", "done", "at", ".", "In", "Figure", "[", "reference", "]", ",", "we", "observe", "that", "good", "trades", "-", "off", "are", "achieved", "for", "dictionary", "sizes", "comprised", "in", ".", "For", "instance", ",", "on", "MPI", "-", "Sintel", ",", "at", ",", "94", "%", "of", "the", "performance", "of", "the", "uncompressed", "case", "(", ")", "is", "reached", "for", "half", "the", "computation", "time", "and", "one", "third", "of", "the", "memory", "usage", ".", "Detailed", "timings", "of", "the", "different", "stages", "of", "DeepMatching", "are", "given", "in", "Table", ".", "As", "expected", ",", "only", "the", "bottom", "-", "up", "pass", "is", "affected", "by", "the", "approximation", ",", "with", "a", "run", "-", "time", "of", "the", "different", "operations", "involved", "(", "patch", "correlations", ",", "max", "-", "pooling", ",", "subsampling", ",", "aggregation", "and", "non", "-", "linear", "rectification", ")", "roughly", "proportional", "to", "D", "(", "or", "to", "|G4|", ",", "the", "actual", "number", "of", "atomic", "patches", ",", "if", "=", "D∞", ")", ".", "The", "overhead", "of", "clustering", "the", "dictionary", "prototypes", "with", "k", "-", "means", "appears", "negligible", ",", "with", "the", "exception", "of", "the", "largest", "dictionary", "size", "(", ")", "for", "which", "it", "induces", "a", "slightly", "longer", "run", "-", "time", "than", "in", "the", "uncompressed", "case", ".", "Overall", ",", "the", "proposed", "method", "for", "approximating", "DeepMatching", "is", "highly", "effective", ".", "paragraph", ":", "GPU", "Implementation", ".", "We", "have", "implemented", "DM", "on", "GPU", "in", "the", "Caffe", "framework", "caffe", ".", "Using", "existing", "Caffe", "layers", "like", "ConvolutionLayer", "and", "PoolingLayer", ",", "the", "implementation", "is", "straightforward", "for", "most", "layers", ".", "We", "had", "to", "specifically", "code", "a", "few", "layers", "which", "are", "not", "available", "in", "Caffe", "(", "e.g", ".", "the", "backtracking", "pass", ")", ".", "For", "the", "aggregation", "layer", "which", "consists", "in", "selecting", "and", "averaging", "4", "children", "channels", "out", "of", "many", "channels", ",", "we", "relied", "on", "the", "sparse", "matrix", "multiplication", "in", "the", "cuSPARSE", "toolbox", ".", "Detailed", "timings", "are", "given", "in", "Table", "[", "reference", "]", "on", "a", "GeForce", "Titan", "X.", "Our", "code", "runs", "in", "about", "0.2s", "for", "a", "pair", "of", "MPI", "-", "Sintel", "image", ".", "As", "expected", ",", "the", "computation", "bottleneck", "essentially", "lies", "in", "the", "computation", "of", "bottom", "-", "level", "patch", "correlations", "and", "the", "backtracking", "pass", ".", "Note", "that", "computing", "patch", "descriptors", "takes", "significantly", "more", "time", ",", "in", "proportion", ",", "than", "on", "CPU", ":", "it", "takes", "about", "0.024s", "=", "11", "%", "of", "total", "time", "(", "not", "shown", "in", "table", ")", ".", "This", "is", "because", "it", "involves", "a", "succession", "of", "many", "small", "layers", "(", "image", "smoothing", ",", "gradient", "extraction", "and", "projection", ",", "etc", ".", ")", ",", "which", "causes", "overhead", "and", "is", "rather", "inefficient", ".", "subsubsection", ":", "Comparison", "to", "the", "state", "of", "the", "art", "We", "compare", "DM", "with", "several", "baselines", "and", "state", "-", "of", "-", "the", "-", "art", "matching", "algorithms", ",", "namely", ":", "SIFT", "keypoints", "extracted", "with", "DoG", "detector", "Lowe2004", "and", "matched", "with", "FLANN", "flann", ",", "referred", "to", "as", "SIFT", "-", "NN", ",", "dense", "HOG", "matching", ",", "followed", "by", "nearest", "-", "neighbor", "matching", "with", "reciprocal", "verification", "as", "done", "in", "LDOF", "Bro11a", ",", "referred", "to", "as", "HOG", "-", "NN", ",", "Generalized", "PatchMatch", "(", "GPM", ")", "Barnes2010", ",", "with", "default", "parameters", ",", "32x32", "patches", "and", "20", "iterations", "(", "best", "settings", "in", "our", "experiments", ")", ",", "Kd", "-", "tree", "PatchMatch", "(", "KPM", ")", "kpm", ",", "an", "improved", "version", "of", "PatchMatch", "based", "on", "better", "patch", "descriptors", "and", "kd", "-", "trees", "optimized", "for", "correspondence", "propagation", ",", "Non", "-", "Rigid", "Dense", "Correspondences", "(", "NRDC", ")", "HaCohen2011", ",", "an", "improved", "version", "of", "GPM", "based", "on", "a", "multiscale", "iterative", "expansion", "/", "contraction", "strategy", ",", "SIFT", "-", "flow", "a", "dense", "matching", "algorithm", "based", "on", "an", "energy", "minimization", "where", "pixels", "are", "represented", "as", "SIFT", "features", "and", "a", "smoothness", "term", "is", "incorporated", "to", "explicitly", "preserve", "spatial", "discontinuities", "Scale", "-", "less", "SIFT", "(", "SLS", ")", "an", "improvement", "of", "SIFT", "-", "flow", "to", "handle", "scale", "changes", "(", "multiple", "sized", "SIFTs", "are", "extracted", "and", "combined", "to", "form", "a", "scale", "-", "invariant", "pixel", "representation", ")", "DaisyFilterFlow", "(", "DaisyFF", ")", "a", "dense", "matching", "approach", "that", "combines", "filter", "-", "based", "efficient", "flow", "inference", "and", "the", "Patch", "-", "Match", "fast", "search", "algorithm", "to", "match", "pixels", "described", "using", "the", "DAISY", "representation", "Deformable", "Pyramid", "Matching", "(", "DSP", ")", "a", "dense", "matching", "approach", "based", "on", "a", "coarse", "-", "to", "-", "fine", "(", "top", "-", "down", ")", "strategy", "where", "inference", "is", "performed", "with", "(", "inexact", ")", "loopy", "belief", "propagation", "SIFT", "-", "NN", ",", "HOG", "-", "NN", "and", "DM", "output", "sparse", "matches", ",", "whereas", "the", "other", "methods", "output", "fully", "dense", "correspondence", "fields", ".", "SIFT", "keypoints", ",", "GPM", ",", "NRDC", "and", "DaisyFF", "are", "scale", "and", "rotation", "invariant", ",", "whereas", "HOG", "-", "NN", ",", "KPM", ",", "SIFT", "-", "flow", ",", "SLS", "and", "DSP", "are", "not", ".", "We", ",", "therefore", ",", "do", "not", "report", "results", "for", "these", "latter", "methods", "on", "the", "Mikolajczyk", "dataset", "which", "includes", "image", "rotations", "and", "scale", "changes", ".", "Statistics", "about", "each", "method", "(", "average", "number", "of", "matches", "per", "image", "and", "their", "coverage", ")", "are", "reported", "in", "Table", "[", "reference", "]", ".", "Coverage", "is", "computed", "as", "the", "proportion", "of", "points", "on", "a", "regular", "grid", "with", "10", "pixel", "spacing", "for", "which", "there", "exists", "a", "correspondence", "(", "in", "the", "raw", "output", "of", "the", "considered", "method", ")", "within", "a", "10", "pixel", "neighborhood", ".", "Thus", ",", "it", "measures", "how", "well", "matches", "‘", "‘", "cover", "’", "’", "the", "image", ".", "Table", "[", "reference", "]", "shows", "that", "DeepMatching", "outputs", "2", "to", "7", "times", "more", "matches", "than", "SIFT", "-", "NN", "and", "a", "comparable", "number", "to", "HOG", "-", "NN", ".", "Yet", ",", "the", "coverage", "for", "DM", "matches", "is", "much", "higher", "than", "for", "HOG", "-", "NN", "and", "SIFT", "-", "NN", ".", "This", "shows", "that", "DM", "matches", "are", "well", "distributed", "over", "the", "entire", "image", ",", "which", "is", "not", "the", "case", "for", "HOG", "-", "NN", "and", "SIFT", "-", "NN", ",", "as", "they", "have", "difficulties", "estimating", "matches", "in", "regions", "with", "weak", "or", "repetitive", "textures", ".", "Quantitative", "results", "are", "listed", "in", "Table", "[", "reference", "]", ",", "and", "qualitative", "results", "in", "Figures", "[", "reference", "]", ",", "[", "reference", "]", "and", "[", "reference", "]", ".", "Overall", ",", "DM", "significantly", "outperforms", "all", "other", "methods", ",", "even", "when", "reduced", "settings", "are", "used", "(", "e.g", ".", "for", "image", "resolution", "and", "prototypes", ")", ".", "As", "expected", ",", "SIFT", "-", "NN", "performs", "rather", "well", "in", "presence", "of", "global", "image", "transformation", "(", "Mikolajczyk", "dataset", ")", ",", "but", "yields", "the", "worst", "result", "for", "the", "case", "of", "more", "complex", "motions", "(", "flow", "datasets", ")", ".", "Figures", "[", "reference", "]", "and", "[", "reference", "]", "illustrate", "the", "reason", ":", "SIFT", "’s", "large", "patches", "are", "way", "too", "coarse", "to", "follow", "motion", "boundaries", "precisely", ".", "The", "same", "issue", "also", "holds", "for", "HOG", "-", "NN", ".", "Methods", "predicting", "dense", "correspondence", "fields", "return", "a", "more", "precise", "estimate", ",", "yet", "most", "of", "them", "(", "KPM", ",", "GPM", ",", "SIFT", "-", "flow", ",", "DSP", ")", "are", "not", "robust", "to", "repetitive", "textures", "in", "the", "Kitti", "dataset", "(", "Figure", ")", "as", "they", "rely", "on", "weakly", "discriminative", "small", "patches", ".", "Despite", "this", "limitation", ",", "SIFT", "-", "flow", "and", "DSP", "are", "still", "able", "to", "perform", "well", "on", "MPI", "-", "Sintel", "as", "this", "dataset", "contains", "little", "scale", "changes", ".", "Other", "dense", "methods", ",", "NRDC", ",", "SLS", "and", "DaisyFF", ",", "can", "handle", "patches", "of", "different", "sizes", "and", "thus", "perform", "better", "on", "Kitti", ".", "But", "in", "turn", "this", "is", "at", "the", "cost", "of", "reduced", "performance", "on", "the", "MPI", "-", "Sintel", "or", "Mikolajczyk", "datasets", "(", "qualitative", "results", "are", "in", "Figure", ")", ".", "In", "conclusion", ",", "DM", "outperforms", "all", "other", "methods", "on", "the", "3", "datasets", ",", "including", "DSP", "which", "also", "relies", "on", "a", "hierarchical", "matching", ".", "In", "terms", "of", "computing", "resources", ",", "DeepMatching", "with", "full", "settings", "(", ",", ")", "is", "one", "of", "the", "most", "costly", "method", "(", "only", "SLS", "and", "DaisyFF", "require", "the", "same", "order", "of", "memory", "and", "longer", "run", "-", "time", ")", ".", "The", "scale", "and", "rotation", "invariant", "version", "of", "DM", ",", "used", "for", "the", "Mikolajczyk", "dataset", ",", "is", "slow", "compared", "to", "most", "other", "approaches", ",", "due", "to", "its", "sequential", "processing", "(", "i.e", ".", "treating", "each", "combination", "of", "rotation", "and", "scaling", "sequentially", ")", ",", "yet", "yields", "near", "perfect", "results", ".", "However", ",", "running", "DM", "with", "reduced", "settings", "is", "very", "competitive", "to", "the", "other", "approaches", ".", "On", "MPI", "-", "Sintel", "and", "Kitti", ",", "for", "instance", ",", "DM", "with", "a", "quarter", "resolution", "has", "a", "run", "-", "time", "comparable", "to", "the", "fastest", "method", ",", "SIFT", "-", "NN", ",", "with", "a", "reasonable", "memory", "usage", ",", "while", "still", "outperforming", "nearly", "all", "methods", "in", "terms", "of", "the", "accuracy@10", "measure", ".", "Ground", "-", "truth", "SIFT", "-", "NN", "GPM", "NRDC", "DaisyFF", "DeepMatching", "GT", "SIFT", "-", "NN", "HOG", "-", "NN", "KPM", "GPM", "SIFT", "-", "flow", "SLS", "DaisyFF", "DSP", "DM", "GT", "SIFT", "-", "NN", "HOG", "-", "NN", "KPM", "GPM", "SIFT", "-", "flow", "SLS", "DaisyFF", "DSP", "DM", "subsection", ":", "Optical", "Flow", "Experiments", "We", "now", "present", "experimental", "results", "for", "the", "optical", "flow", "estimation", ".", "Optical", "flow", "is", "predicted", "using", "the", "variational", "framework", "presented", "in", "Section", "[", "reference", "]", "that", "takes", "as", "input", "a", "set", "of", "matches", ".", "In", "the", "following", ",", "we", "evaluate", "the", "impact", "of", "DeepMatching", "against", "other", "matching", "methods", ",", "and", "compare", "to", "the", "state", "of", "the", "art", ".", "subsubsection", ":", "Optimization", "of", "the", "parameters", "We", "optimize", "the", "parameters", "of", "DeepFlow", "on", "a", "subset", "of", "the", "MPI", "-", "Sintel", "training", "set", "(", "20", "%", ")", ",", "called", "‘", "‘", "small", "’", "’", "set", ",", "and", "report", "results", "on", "the", "remaining", "image", "pairs", "(", "80", "%", ",", "called", "‘", "‘", "validation", "set", "’", "’", ")", "and", "on", "the", "training", "sets", "of", "Kitti", "and", "Middlebury", ".", "Ground", "-", "truth", "optical", "flows", "for", "the", "three", "test", "sets", "are", "not", "publicly", "available", ",", "in", "order", "to", "prevent", "parameter", "tuning", "on", "the", "test", "set", ".", "We", "first", "optimize", "the", "different", "flow", "parameters", "(", ",", ",", ",", "and", ")", "by", "employing", "a", "gradient", "descent", "strategy", "with", "multiple", "initializations", "followed", "by", "a", "local", "grid", "search", ".", "For", "the", "data", "term", ",", "we", "find", "an", "optimum", "at", ",", "which", "is", "equivalent", "to", "removing", "the", "color", "constancy", "assumption", ".", "This", "can", "be", "explained", "by", "the", "fact", "that", "the", "‘", "‘", "final", "’", "’", "version", "contains", "atmospheric", "effects", ",", "reflections", ",", "blurs", ",", "etc", ".", "The", "remaining", "parameters", "are", "optimal", "at", ",", ",", ",", ".", "These", "parameters", "are", "used", "in", "the", "remaining", "of", "the", "experiments", "for", "DeepFlow", ",", "i.e.", "using", "matches", "obtained", "with", "DeepMatching", ",", "except", "when", "reporting", "results", "on", "Kitti", "and", "Middlebury", "test", "sets", "in", "Section", ".", "In", "this", "case", "the", "parameters", "are", "optimized", "on", "their", "respective", "training", "set", ".", "subsubsection", ":", "Impact", "of", "the", "matches", "on", "the", "flow", "We", "examine", "the", "impact", "of", "different", "matching", "methods", "on", "the", "flow", ",", "i.e", ".", ",", "different", "matches", "are", "used", "in", "DeepFlow", ",", "see", "Section", "[", "reference", "]", ".", "For", "all", "matching", "approaches", "evaluated", "in", "the", "previous", "section", ",", "we", "use", "their", "output", "as", "matching", "term", "in", "Eq", ".", "(", "[", "reference", "]", ")", ".", "Because", "these", "approaches", "may", "output", "matches", "with", "statistics", "different", "from", "DM", ",", "we", "separately", "optimize", "the", "flow", "parameters", "for", "each", "matching", "approach", "on", "the", "small", "training", "set", "of", "MPI", "-", "Sintel", ".", "Table", "[", "reference", "]", "shows", "the", "endpoint", "error", ",", "averaged", "over", "all", "pixels", ".", "Clearly", ",", "a", "sufficiently", "dense", "and", "accurate", "matching", "like", "DM", "allows", "to", "considerably", "improve", "the", "flow", "estimation", "on", "datasets", "with", "large", "displacements", "(", "MPI", "-", "Sintel", ",", "Kitti", ")", ".", "In", "contrast", ",", "none", "of", "the", "methods", "presented", "have", "a", "tangible", "effect", "on", "the", "Middlebury", "dataset", ",", "where", "the", "displacements", "are", "small", ".", "The", "relatively", "small", "gains", "achieved", "by", "SIFT", "-", "NN", "and", "HOG", "-", "NN", "on", "MPI", "-", "Sintel", "and", "Kitti", "are", "due", "to", "the", "fact", "that", "a", "lot", "of", "regions", "with", "large", "displacements", "are", "not", "covered", "by", "any", "matches", ",", "such", "as", "the", "sky", "or", "the", "blurred", "character", "in", "the", "first", "and", "second", "column", "of", "Figure", "[", "reference", "]", ".", "Hence", ",", "SIFT", "-", "NN", "and", "HOG", "-", "NN", "have", "only", "a", "limited", "impact", "on", "the", "variational", "approach", ".", "On", "the", "other", "hand", ",", "the", "gains", "are", "also", "small", "(", "or", "even", "negative", ")", "for", "the", "dense", "methods", "despite", "the", "fact", "that", "they", "output", "significantly", "more", "correspondences", ".", "We", "observe", "for", "these", "methods", "that", "the", "weight", "of", "the", "matching", "term", "tends", "to", "be", "small", "after", "optimizing", "the", "parameters", ",", "thus", "indicating", "that", "the", "matches", "are", "found", "unreliable", "and", "noisy", "during", "training", ".", "The", "cause", "is", "clearly", "visible", "in", "Figure", "[", "reference", "]", ",", "where", "large", "portions", "containing", "repetitive", "textures", "(", "e.g", ".", "road", ",", "trees", ")", "are", "incorrectly", "matched", ".", "The", "poor", "quality", "of", "these", "matches", "even", "leads", "to", "a", "significant", "drop", "in", "performance", "on", "the", "Kitti", "dataset", ".", "In", "contrast", ",", "DeepMatching", "generates", "accurate", "matches", "well", "covering", "the", "image", "that", "enable", "to", "boost", "the", "optical", "flow", "accuracy", "in", "case", "of", "large", "displacements", ".", "Namely", ",", "we", "observe", "a", "relative", "improvement", "of", "30", "%", "on", "MPI", "-", "Sintel", "and", "of", "50", "%", "on", "Kitti", ".", "It", "is", "interesting", "to", "observe", "that", "DM", "is", "able", "to", "effectively", "prune", "false", "matches", "arising", "in", "occluded", "areas", "(", "black", "areas", "in", "Figures", "and", ")", ".", "This", "is", "due", "to", "the", "reciprocal", "verification", "filtering", "incorporated", "in", "DM", "(", "Eq", ".", "(", ")", ")", ".", "When", "using", "the", "approximation", "with", "1024", "prototypes", ",", "however", ",", "a", "significant", "drop", "is", "observed", "on", "the", "Kitti", "dataset", ",", "while", "the", "performance", "remains", "good", "on", "MPI", "-", "Sintel", ".", "This", "indicates", "that", "approximating", "DeepMatching", "can", "result", "in", "a", "significant", "loss", "of", "robustness", "when", "matching", "repetitive", "textures", ",", "that", "are", "more", "frequent", "in", "Kitti", "than", "in", "MPI", "-", "Sintel", ".", "subsubsection", ":", "Comparison", "to", "the", "state", "of", "the", "art", "In", "this", "section", ",", "we", "compare", "DeepFlow", "to", "the", "state", "of", "the", "art", "on", "the", "test", "sets", "of", "MPI", "-", "Sintel", ",", "Kitti", "and", "Middlebury", "datasets", ".", "For", "theses", "datasets", ",", "the", "results", "are", "submitted", "to", "a", "dedicated", "server", "which", "performs", "the", "evaluation", ".", "Prior", "to", "submitting", "our", "results", "for", "Kitti", "and", "Middlebury", "test", "sets", ",", "we", "have", "optimized", "the", "parameters", "on", "the", "respective", "training", "set", ".", "paragraph", ":", "Results", "on", "MPI", "-", "Sintel", ".", "Images", "Ground", "-", "Truth", "DeepMatching", "DeepFlow", "MDP", "-", "Flow2", "LDOF", "Table", "[", "reference", "]", "compares", "our", "method", "to", "state", "-", "of", "-", "the", "-", "art", "algorithms", "on", "the", "MPI", "-", "Sintel", "test", "set", ".", "A", "comparison", "with", "the", "preliminary", "version", "of", "DeepFlow", "DeepFlow", ",", "referred", "to", "as", "DeepFlow", "*", ",", "is", "also", "provided", ".", "In", "this", "early", "version", ",", "we", "used", "a", "constant", "smoothness", "weight", "instead", "of", "a", "local", "one", "here", "(", "see", "Section", "[", "reference", "]", ")", "and", "used", "DM", "*", "as", "input", "matches", ".", "We", "can", "see", "that", "DeepFlow", "is", "among", "the", "best", "performing", "methods", "on", "MPI", "-", "Sintel", ",", "particularly", "for", "large", "displacements", ".", "This", "is", "due", "to", "the", "use", "of", "a", "reliable", "matching", "term", "in", "the", "variational", "approach", ",", "and", "this", "property", "is", "shared", "by", "all", "top", "performing", "approaches", ",", "e.g", ".", "epicflow", ",", "Leordeanu2013", ".", "Furthermore", ",", "it", "is", "interesting", "to", "note", "that", "among", "the", "top", "performers", "on", "MPI", "-", "Sintel", ",", "3", "methods", "out", "of", "6", "actually", "employ", "DeepMatching", ".", "In", "particular", ",", "the", "top", "-", "3", "method", "EpicFlow", "epicflow", "relies", "on", "the", "output", "of", "DeepMatching", "to", "produce", "a", "piece", "-", "wise", "affine", "flow", ",", "and", "SparseFlowFused", "sparseflowfused", "combines", "matches", "obtained", "with", "DeepMatching", "and", "another", "algorithm", ".", "We", "refer", "to", "the", "webpage", "of", "the", "MPI", "-", "Sintel", "dataset", "for", "complete", "results", "including", "the", "‘", "‘", "clean", "’", "’", "version", ".", "paragraph", ":", "Timings", ".", "As", "mentioned", "before", ",", "DeepMatching", "at", "half", "the", "resolution", "takes", "15", "seconds", "to", "compute", "on", "CPU", "and", "0.2", "second", "on", "GPU", ".", "The", "variational", "part", "requires", "10", "additional", "seconds", "on", "CPU", ".", "Note", "that", "by", "implementing", "it", "on", "GPU", ",", "we", "could", "obtain", "a", "significant", "speed", "-", "up", "as", "well", ".", "DeepFlow", "consequently", "takes", "25", "seconds", "in", "total", "on", "a", "single", "CPU", "core", "@", "3.6", "GHz", "or", "10.2s", "with", "GPU", "+", "CPU", ".", "This", "is", "in", "the", "same", "order", "of", "magnitude", "as", "the", "fastest", "among", "the", "best", "competitors", ",", "EpicFlow", "epicflow", ".", "paragraph", ":", "Results", "on", "Kitti", ".", "Table", "[", "reference", "]", "summarizes", "the", "main", "results", "on", "the", "Kitti", "benchmark", "(", "see", "official", "website", "for", "complete", "results", ")", ",", "when", "optimizing", "the", "parameters", "on", "the", "Kitti", "training", "set", ".", "EPE", "-", "Noc", "is", "the", "EPE", "computed", "only", "in", "non", "-", "occluded", "areas", ".", "‘", "‘", "Out", "3", "’", "’", "corresponds", "to", "the", "proportion", "of", "incorrect", "pixel", "correspondences", "for", "an", "error", "threshold", "of", "3", "pixels", ",", "i.e", ".", "it", "corresponds", "to", ",", "and", "likewise", "for", "‘", "‘", "Out", "-", "Noc", "3", "’", "’", "for", "non", "-", "occluded", "areas", ".", "In", "terms", "of", "EPE", "-", "noc", ",", "DeepFlow", "is", "on", "par", "with", "the", "best", "approaches", ",", "but", "performs", "somewhat", "worse", "in", "the", "occluded", "areas", ".", "This", "is", "due", "to", "a", "specificity", "of", "the", "Kitti", "dataset", ",", "in", "which", "motion", "is", "mostly", "homographic", "(", "especially", "on", "the", "image", "borders", ",", "where", "most", "surfaces", "like", "roads", "and", "walls", "are", "planar", ")", ".", "In", "such", "cases", ",", "flow", "is", "better", "predicted", "using", "an", "affine", "motion", "prior", ",", "which", "locally", "well", "approximates", "homographies", "(", "a", "constant", "motion", "prior", "is", "used", "in", "DeepFlow", ")", ".", "As", "a", "matter", "of", "facts", ",", "all", "top", "performing", "methods", "in", "terms", "of", "total", "EPE", "output", "piece", "-", "wise", "affine", "optical", "flow", ",", "either", "due", "to", "affine", "regularizers", "(", "BTF", "-", "ILLUM", "demetz2014learning", ",", "NLTGB", "-", "SC", "ranftl2014non", ",", "TGV2ADCSIFT", "Braux", "-", "Zin_2013_ICCV", ")", "or", "due", "to", "local", "affine", "estimators", "(", "EpicFlow", "epicflow", ")", ".", "Note", "that", "the", "learned", "parameters", "on", "Kitti", "and", "MPI", "-", "Sintel", "are", "close", ".", "In", "particular", ",", "running", "the", "experiments", "with", "the", "same", "parameters", "as", "MPI", "-", "Sintel", "decreases", "EPE", "-", "Noc", "by", "only", "pixels", "on", "the", "training", "set", ".", "This", "shows", "that", "our", "method", "does", "not", "suffer", "from", "overfitting", ".", "paragraph", ":", "Results", "on", "Middlebury", ".", "We", "optimize", "the", "parameters", "on", "the", "Middlebury", "training", "set", "by", "minimizing", "the", "average", "angular", "error", "with", "the", "same", "strategy", "as", "for", "MPI", "-", "Sintel", ".", "We", "find", "weights", "quasi", "-", "zero", "for", "the", "matching", "term", "due", "to", "the", "absence", "of", "large", "displacements", ".", "DeepFlow", "obtained", "an", "average", "endpoint", "error", "of", "on", "the", "test", "which", "is", "competitive", "with", "the", "state", "of", "the", "art", ".", "section", ":", "Conclusion", "We", "have", "introduced", "a", "dense", "matching", "algorithm", ",", "termed", "DeepMatching", ".", "The", "proposed", "algorithm", "gracefully", "handles", "complex", "non", "-", "rigid", "object", "deformations", "and", "repetitive", "textured", "regions", ".", "DeepMatching", "yields", "state", "-", "of", "-", "the", "-", "art", "performance", "for", "image", "matching", ",", "on", "the", "Mikolajczyk", "Mikolajczyk2005", ",", "the", "MPI", "-", "Sintel", "sintel", "and", "the", "Kitti", "kitti", "datasets", ".", "Integrated", "in", "a", "variational", "energy", "minimization", "approach", "Bro11a", ",", "the", "resulting", "approach", "for", "optical", "flow", "estimation", ",", "termed", "DeepFlow", ",", "shows", "competitive", "performance", "on", "optical", "flow", "benchmarks", ".", "Future", "work", "includes", "incorporating", "a", "weighting", "of", "the", "patches", "in", "Eq", ".", "(", "[", "reference", "]", ")", "instead", "of", "weighting", "all", "patches", "equally", "to", "take", "into", "account", "that", "different", "parts", "of", "a", "large", "patch", "may", "belong", "to", "different", "objects", ".", "This", "could", "improve", "the", "performance", "of", "DeepMatching", "for", "thin", "objects", ",", "such", "as", "human", "limbs", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["DeepMatching"], "Method"], [4, 8, ["Hierarchical", "Deformable", "Dense", "Matching"], "Task"], [12, 14, ["matching", "algorithm"], "Method"], [16, 17, ["DeepMatching"], "Method"], [20, 24, ["dense", "correspondences", "between", "images"], "Task"], [25, 26, ["DeepMatching"], "Method"], [29, 37, ["hierarchical", ",", "multi", "-", "layer", ",", "correlational", "architecture"], "Method"], [45, 48, ["deep", "convolutional", "approaches"], "Method"], [51, 53, ["matching", "algorithm"], "Method"], [81, 82, ["DeepMatching"], "Method"], [93, 95, ["matching", "algorithms"], "Method"], [112, 113, ["DeepMatching"], "Method"], [139, 142, ["estimating", "optical", "flow"], "Task"], [144, 145, ["DeepFlow"], "Method"], [148, 149, ["DeepMatching"], "Method"], [151, 155, ["large", "displacement", "optical", "flow"], "Method"], [156, 157, ["LDOF"], "Method"], [166, 168, ["matching", "algorithms"], "Method"], [170, 171, ["robustness"], "Metric"], [182, 184, ["matching", "approach"], "Method"], [185, 186, ["DeepFlow"], "Method"], [193, 196, ["optical", "flow", "estimation"], "Task"], [200, 202, ["Computing", "correspondences"], "Task"], [211, 214, ["computer", "vision", "problems"], "Task"], [217, 219, ["scene", "recognition"], "Task"], [220, 223, ["optical", "flow", "estimation"], "Task"], [231, 233, ["matching", "algorithm"], "Method"], [263, 265, ["matching", "approaches"], "Method"], [280, 282, ["matching", "approaches"], "Method"], [292, 293, ["matching"], "Method"], [370, 372, ["matching", "algorithms"], "Method"], [373, 374, ["images"], "Task"], [414, 415, ["matching"], "Method"], [421, 423, ["Matching", "algorithms"], "Method"], [424, 425, ["images"], "Task"], [535, 536, ["SIFT"], "Method"], [538, 540, ["HOG", "Dalal2005"], "Method"], [545, 548, ["discriminative", "feature", "representations"], "Method"], [561, 563, ["rigid", "motion"], "Task"], [613, 614, ["robustness"], "Metric"], [622, 626, ["propagation", "-", "based", "approaches"], "Method"], [629, 631, ["Generalized", "PatchMatch"], "Method"], [633, 638, ["Non", "-", "rigid", "Dense", "Correspondences"], "Method"], [642, 644, ["feature", "representations"], "Method"], [661, 665, ["non", "-", "rigid", "deformations"], "Task"], [689, 690, ["DeepMatching"], "Method"], [703, 704, ["DeepMatching"], "Method"], [708, 712, ["multi", "-", "layer", "architecture"], "Method"], [777, 778, ["DeepMatching"], "Method"], [784, 787, ["deep", "convolutional", "approaches"], "Method"], [802, 804, ["matching", "algorithm"], "Method"], [805, 806, ["DeepMatching"], "Method"], [821, 825, ["non", "-", "rigid", "deformations"], "Task"], [860, 862, ["gradient", "histograms"], "Method"], [892, 893, ["DeepMatching"], "Method"], [908, 910, ["memory", "cost"], "Metric"], [917, 918, ["DeepMatching"], "Method"], [923, 929, ["scale", "and", "rotation", "-", "invariant", "version"], "Method"], [947, 952, ["Large", "-", "displacement", "optical", "flow"], "Task"], [956, 959, ["optical", "flow", "approach"], "Method"], [961, 962, ["DeepMatching"], "Method"], [964, 966, ["matching", "term"], "Method"], [968, 973, ["large", "displacement", "variational", "energy", "minimization"], "Method"], [980, 981, ["DeepMatching"], "Method"], [988, 990, ["HOG", "descriptor"], "Method"], [1003, 1005, ["matching", "algorithms"], "Method"], [1010, 1011, ["DeepFlow"], "Method"], [1047, 1049, ["matching", "algorithm"], "Method"], [1050, 1051, ["DeepMatching"], "Method"], [1068, 1069, ["DeepMatching"], "Method"], [1076, 1079, ["optical", "flow", "approach"], "Method"], [1080, 1081, ["DeepFlow"], "Method"], [1122, 1123, ["DeepMatching"], "Method"], [1131, 1132, ["DeepMatching"], "Method"], [1136, 1138, ["match", "scoring"], "Task"], [1152, 1156, ["time", "and", "memory", "complexity"], "Metric"], [1157, 1158, ["DeepMatching"], "Method"], [1162, 1164, ["deformation", "tolerance"], "Metric"], [1174, 1178, ["Deep", "Convolutional", "Neural", "Networks"], "Method"], [1179, 1180, ["DeepMatching"], "Method"], [1186, 1188, ["approximate", "version"], "Method"], [1189, 1190, ["DeepMatching"], "Method"], [1195, 1200, ["scale", "and", "rotation", "invariant", "version"], "Method"], [1201, 1202, ["DeepMatching"], "Method"], [1212, 1213, ["DeepMatching"], "Method"], [1227, 1228, ["DeepMatching"], "Method"], [1231, 1232, ["DeepFlow"], "Method"], [1242, 1244, ["GPU", "implementation"], "Method"], [1248, 1250, ["CPU", "one"], "Method"], [1273, 1274, ["matching"], "Method"], [1282, 1283, ["matching"], "Method"], [1287, 1290, ["optical", "flow", "estimation"], "Task"], [1330, 1334, ["wide", "baseline", "stereo", "matching"], "Task"], [1336, 1338, ["image", "retrieval"], "Task"], [1350, 1352, ["local", "descriptors"], "Method"], [1356, 1358, ["Image", "descriptors"], "Method"], [1377, 1378, ["Matching"], "Task"], [1380, 1383, ["nearest", "neighbor", "search"], "Method"], [1390, 1392, ["geometric", "verification"], "Method"], [1396, 1398, ["confidence", "value"], "Metric"], [1453, 1455, ["matching", "algorithm"], "Method"], [1457, 1458, ["DeepMatching"], "Method"], [1462, 1467, ["non", "-", "rigid", "2D", "warping"], "Method"], [1468, 1471, ["deep", "convolutional", "networks"], "Method"], [1492, 1496, ["family", "of", "feasible", "warpings"], "Method"], [1509, 1511, ["2D", "warping"], "Method"], [1560, 1564, ["similarity", "of", "small", "images"], "Task"], [1595, 1599, ["dense", "correspondences", "between", "images"], "Task"], [1602, 1606, ["non", "-", "rigid", "matching"], "Method"], [1609, 1612, ["smooth", "parametric", "models"], "Method"], [1630, 1635, ["parametric", "and", "model", "-", "free"], "Method"], [1733, 1736, ["guided", "filtering", "stage"], "Method"], [1739, 1740, ["PatchMatch"], "Method"], [1750, 1751, ["MRF"], "Method"], [1756, 1758, ["hierarchical", "matching"], "Method"], [1760, 1762, ["dense", "correspondences"], "Task"], [1771, 1776, ["top", "-", "down", ")", "strategy"], "Method"], [1777, 1780, ["Loopy", "belief", "propagation"], "Method"], [1784, 1785, ["inference"], "Task"], [1792, 1793, ["DeepMatching"], "Method"], [1811, 1812, ["DeepMatching"], "Method"], [1835, 1839, ["multi", "-", "layer", "construction"], "Method"], [1843, 1844, ["matching"], "Method"], [1845, 1850, ["semi", "-", "rigid", "local", "deformations"], "Task"], [1854, 1855, ["DeepMatching"], "Method"], [1867, 1871, ["low", "-", "memory", "requirements"], "Metric"], [1875, 1876, ["accuracy"], "Metric"], [1879, 1880, ["Matching"], "Task"], [1881, 1883, ["flow", "estimation"], "Task"], [1883, 1886, ["Variational", "energy", "minimization"], "Method"], [1893, 1896, ["optical", "flow", "estimation"], "Task"], [1938, 1940, ["variational", "approach"], "Method"], [1961, 1964, ["data", "-", "fitting"], "Method"], [1966, 1968, ["smoothness", "terms"], "Method"], [1969, 1971, ["Energy", "minimization"], "Task"], [1976, 1980, ["Euler", "-", "Lagrange", "equations"], "Method"], [1989, 1994, ["large", "and", "structured", "linear", "systems"], "Task"], [2025, 2026, ["siftflow"], "Method"], [2027, 2028, ["siftscales"], "Method"], [2030, 2033, ["dense", "descriptor", "matching"], "Method"], [2039, 2044, ["reciprocal", "nearest", "-", "neighbor", "verification"], "Method"], [2056, 2060, ["variational", "energy", "minimization", "approach"], "Method"], [2084, 2087, ["dense", "scene", "correspondence"], "Task"], [2108, 2110, ["variational", "framework"], "Method"], [2142, 2143, ["SIFT"], "Method"], [2145, 2146, ["PatchMatch"], "Method"], [2150, 2152, ["flow", "initialization"], "Task"], [2161, 2164, ["optical", "flow", "estimation"], "Task"], [2171, 2173, ["fusion", "steps"], "Method"], [2190, 2193, ["total", "variation", "algorithm"], "Method"], [2196, 2198, ["flow", "estimation"], "Task"], [2209, 2213, ["large", "displacement", "optical", "flow"], "Method"], [2217, 2219, ["DeepMatching", "algorithm"], "Method"], [2226, 2227, ["DeepMatching"], "Method"], [2233, 2234, ["DeepMatching"], "Method"], [2235, 2236, ["DeepMatching"], "Method"], [2238, 2240, ["matching", "algorithm"], "Method"], [2253, 2257, ["multi", "-", "layer", "fashion"], "Method"], [2259, 2263, ["multi", "-", "layer", "architecture"], "Method"], [2266, 2272, ["quadtree", "-", "like", "patch", "subdivision", "scheme"], "Method"], [2308, 2310, ["nonlinear", "transformation"], "Method"], [2322, 2323, ["DeepMatching"], "Method"], [2380, 2381, ["DeepMatching"], "Method"], [2412, 2414, ["SIFT", "descriptor"], "Method"], [2416, 2417, ["SIFT"], "Method"], [2431, 2433, ["robust", "descriptor"], "Method"], [2640, 2642, ["recursive", "decomposition"], "Method"], [2667, 2670, ["atomic", "patch", "correspondences"], "Task"], [2862, 2866, ["bottom", "-", "up", "algorithm"], "Method"], [2900, 2904, ["top", "-", "down", "method"], "Method"], [2906, 2910, ["motion", "of", "atomic", "patches"], "Task"], [2948, 2949, ["DeepMatching"], "Method"], [2960, 2963, ["correlation", "pyramid", "computation"], "Method"], [3025, 3030, ["average", "pixel", "-", "wise", "similarity"], "Metric"], [3053, 3055, ["pixel", "descriptor"], "Method"], [3060, 3062, ["Pixel", "descriptor"], "Method"], [3067, 3070, ["robust", "pixel", "representation"], "Method"], [3076, 3077, ["SIFT"], "Method"], [3078, 3079, ["DAISY"], "Method"], [3089, 3094, ["Gaussian", "smoothing", "of", "radius", "ν1"], "Method"], [3106, 3108, ["JPEG", "compression"], "Task"], [3149, 3152, ["oriented", "gradient", "maps"], "Method"], [3159, 3164, ["Gaussian", "filter", "of", "radius", "ν2"], "Method"], [3203, 3208, ["Gaussian", "filter", "of", "radius", "ν3"], "Method"], [3237, 3238, ["μ"], "Method"], [3242, 3243, ["regularizer"], "Method"], [3279, 3282, ["dot", "-", "product"], "Method"], [3284, 3286, ["similarity", "function"], "Method"], [3309, 3311, ["pixel", "descriptor"], "Method"], [3317, 3319, ["correlation", "map"], "Metric"], [3340, 3342, ["convolutional", "framework"], "Method"], [3424, 3427, ["measure", "of", "similarity"], "Metric"], [3460, 3464, ["Recursive", "aggregation", "of", "patches"], "Method"], [3565, 3566, ["sN"], "Method"], [3817, 3819, ["correlation", "map"], "Metric"], [3873, 3874, ["subsampling"], "Method"], [3890, 3892, ["computational", "requirements"], "Metric"], [3923, 3926, ["max", "-", "pooling"], "Method"], [3931, 3935, ["max", "-", "pooling", "domain"], "Method"], [3959, 3963, ["max", "-", "pooling", "operator"], "Method"], [4010, 4014, ["non", "-", "linear", "mapping"], "Method"], [4030, 4032, ["power", "transform"], "Method"], [4069, 4072, ["max", "-", "pooling"], "Method"], [4087, 4089, ["correlation", "values"], "Metric"], [4106, 4109, ["post", "-", "processing"], "Method"], [4113, 4116, ["deep", "convolutional", "networks"], "Method"], [4396, 4400, ["non", "-", "rigid", "matching"], "Method"], [4406, 4409, ["do", "(", "convolution"], "Task"], [4434, 4437, ["max", "-", "pooling"], "Method"], [4438, 4439, ["subsampling"], "Method"], [4464, 4466, ["pyramid", "Computing"], "Method"], [4475, 4480, ["Top", "-", "down", "correspondence", "extraction"], "Task"], [4550, 4551, ["backtracking"], "Method"], [4581, 4582, ["DeepMatching"], "Method"], [4595, 4597, ["match", "scoring"], "Method"], [4598, 4601, ["entry", "point", "selection"], "Method"], [4604, 4606, ["execution", "time"], "Metric"], [4624, 4627, ["Backtracking", "atomic", "correspondences"], "Task"], [4662, 4664, ["pyramid", "construction"], "Method"], [4776, 4778, ["backtracking", "correspondences"], "Task"], [4826, 4828, ["Merging", "correspondences"], "Task"], [4996, 4997, ["matching"], "Method"], [5066, 5069, ["reciprocal", "match", "verification"], "Task"], [5110, 5111, ["DeepMatching"], "Method"], [5166, 5168, ["matching", "methods"], "Method"], [5266, 5268, ["matching", "descriptors"], "Method"], [5271, 5272, ["SIFT"], "Method"], [5287, 5289, ["matching", "schemes"], "Method"], [5306, 5308, ["matching", "algorithm"], "Method"], [5365, 5366, ["DeepMatching"], "Method"], [5376, 5378, ["convolutional", "nature"], "Method"], [5509, 5511, ["lower", "bound"], "Metric"], [5516, 5517, ["DeepMatching"], "Method"], [5597, 5601, ["Built", "-", "in", "smoothing"], "Method"], [5615, 5617, ["correlation", "maps"], "Method"], [5711, 5713, ["backtracking", "procedure"], "Method"], [5723, 5725, ["max", "operation"], "Method"], [5727, 5729, ["random", "sampling"], "Method"], [5759, 5760, ["DeepMatching"], "Method"], [5776, 5777, ["Smoothness"], "Metric"], [5829, 5831, ["identity", "warping"], "Method"], [5836, 5840, ["Deep", "Convolutional", "Neural", "Networks"], "Method"], [5841, 5842, ["CNNs"], "Method"], [5844, 5845, ["DeepMatching"], "Method"], [5848, 5856, ["hierarchical", ",", "multi", "-", "layer", ",", "correlational", "architecture"], "Method"], [5864, 5867, ["deep", "convolutional", "approaches"], "Method"], [5878, 5880, ["Deep", "networks"], "Method"], [5887, 5888, ["convolutions"], "Method"], [5892, 5893, ["DeepMatching"], "Method"], [5897, 5899, ["feature", "representations"], "Method"], [5917, 5919, ["convolution", "filters"], "Method"], [5927, 5931, ["bottom", "-", "up", "pipeline"], "Method"], [5932, 5933, ["DeepMatching"], "Method"], [5936, 5937, ["CNNs"], "Method"], [5947, 5953, ["channel", "-", "wise", "max", "-", "pooling"], "Method"], [5954, 5955, ["subsampling"], "Method"], [5958, 5959, ["CNNs"], "Method"], [5960, 5963, ["max", "-", "pooling"], "Method"], [5964, 5965, ["DeepMatching"], "Method"], [5967, 5968, ["invariance"], "Task"], [5988, 5991, ["linear", "rectifying", "stages"], "Method"], [5998, 5999, ["DeepMatching"], "Method"], [6001, 6005, ["top", "-", "down", "pass"], "Method"], [6010, 6011, ["CNNs"], "Method"], [6014, 6018, ["Time", "and", "space", "complexity"], "Metric"], [6019, 6020, ["DeepMatching"], "Method"], [6024, 6025, ["memory"], "Metric"], [6061, 6062, ["complexity"], "Metric"], [6090, 6091, ["complexity"], "Metric"], [6093, 6096, ["correlation", "maps", "computation"], "Method"], [6179, 6180, ["DeepMatching"], "Method"], [6189, 6191, ["space", "complexity"], "Metric"], [6192, 6193, ["DeepMatching"], "Method"], [6212, 6214, ["matching", "methods"], "Method"], [6252, 6257, ["reduced", "time", "and", "memory", "usage"], "Metric"], [6287, 6291, ["representation", "of", "atomic", "patches"], "Method"], [6382, 6386, ["memory", "and", "time", "complexity"], "Metric"], [6456, 6460, ["memory", "and", "computational", "cost"], "Metric"], [6479, 6481, ["prototype", "dictionary"], "Method"], [6482, 6485, ["k", "-", "means"], "Method"], [6519, 6521, ["line", "clustering"], "Method"], [6523, 6527, ["descriptors", "of", "atomic", "patches"], "Method"], [6542, 6544, ["pixel", "descriptor"], "Method"], [6552, 6556, ["k", "-", "means", "approach"], "Method"], [6620, 6621, ["DeepMatching"], "Method"], [6622, 6623, ["DM"], "Method"], [6638, 6643, ["scale", "and", "rotation", "invariant", "version"], "Method"], [6651, 6652, ["DM"], "Method"], [6676, 6678, ["image", "rotation"], "Task"], [6740, 6742, ["reciprocal", "verification"], "Task"], [6804, 6808, ["space", "and", "time", "complexity"], "Metric"], [6822, 6825, ["run", "-", "time"], "Metric"], [6827, 6828, ["DM"], "Method"], [6901, 6903, ["Geometric", "rectification"], "Method"], [6930, 6931, ["DeepMatching"], "Method"], [6932, 6933, ["DM"], "Method"], [6957, 6958, ["DeepFlow"], "Method"], [6964, 6967, ["optical", "flow", "estimation"], "Task"], [6968, 6969, ["DeepFlow"], "Task"], [6981, 6983, ["matching", "term"], "Method"], [6996, 6998, ["matching", "approach"], "Method"], [6999, 7000, ["DeepMatching"], "Method"], [7025, 7026, ["normalization"], "Method"], [7055, 7057, ["matching", "term"], "Method"], [7098, 7103, ["Gaussian", "filter", "of", "standard", "deviation"], "Method"], [7110, 7112, ["weighted", "sum"], "Method"], [7122, 7124, ["matching", "term"], "Method"], [7133, 7135, ["robust", "penalizer"], "Method"], [7149, 7151, ["data", "term"], "Method"], [7214, 7216, ["data", "term"], "Method"], [7372, 7374, ["robust", "penalization"], "Method"], [7376, 7379, ["gradient", "flow", "norm"], "Method"], [7419, 7421, ["matching", "term"], "Method"], [7423, 7425, ["flow", "estimation"], "Task"], [7430, 7433, ["precomputed", "vector", "field"], "Method"], [7446, 7448, ["robust", "penalizer"], "Method"], [7451, 7452, ["matching"], "Method"], [7538, 7540, ["Gaussian", "kernel"], "Method"], [7572, 7573, ["Minimization"], "Task"], [7593, 7596, ["numerical", "optimization", "algorithm"], "Method"], [7600, 7608, ["incremental", "coarse", "-", "to", "-", "fine", "warping", "strategy"], "Method"], [7626, 7628, ["robust", "penalizers"], "Method"], [7632, 7636, ["inner", "fixed", "point", "iterations"], "Method"], [7660, 7662, ["linear", "system"], "Task"], [7669, 7676, ["Successive", "Over", "Relaxation", "(", "SOR", ")", "method"], "Method"], [7678, 7679, ["sor"], "Method"], [7742, 7743, ["DeepMatching"], "Method"], [7744, 7745, ["DeepFlow"], "Method"], [7753, 7754, ["DeepMatching"], "Method"], [7755, 7756, ["DeepFlow"], "Method"], [7813, 7815, ["matching", "approach"], "Method"], [7830, 7831, ["flow"], "Task"], [7855, 7857, ["keypoint", "detectors"], "Method"], [7858, 7859, ["descriptors"], "Method"], [7869, 7871, ["matching", "approaches"], "Method"], [7939, 7940, ["robustness"], "Metric"], [7981, 7983, ["evaluation", "benchmark"], "Metric"], [7984, 7987, ["optical", "flow", "estimation"], "Task"], [8079, 8082, ["optical", "flow", "methods"], "Method"], [8145, 8147, ["driving", "platform"], "Method"], [8200, 8202, ["Performance", "metric"], "Metric"], [8203, 8204, ["matching"], "Method"], [8207, 8209, ["performance", "measure"], "Metric"], [8210, 8212, ["matching", "approaches"], "Method"], [8215, 8217, ["Matching", "approaches"], "Method"], [8245, 8247, ["matching", "approaches"], "Method"], [8249, 8250, ["DeepMatching"], "Method"], [8307, 8308, ["accuracy@"], "Metric"], [8433, 8435, ["Performance", "metric"], "Metric"], [8436, 8438, ["optical", "flow"], "Task"], [8441, 8443, ["optical", "flow"], "Task"], [8452, 8455, ["average", "endpoint", "error"], "Metric"], [8463, 8464, ["EPE"], "Metric"], [8478, 8479, ["EPE"], "Method"], [8535, 8536, ["Matching"], "Task"], [8543, 8544, ["DeepMatching"], "Method"], [8545, 8546, ["DM"], "Method"], [8602, 8603, ["matching"], "Method"], [8619, 8622, ["MPI", "-", "Sintel"], "Material"], [8623, 8624, ["Kitti"], "Material"], [8637, 8638, ["DM"], "Method"], [8669, 8671, ["pixel", "representation"], "Method"], [8701, 8703, ["grid", "search"], "Task"], [8726, 8727, ["accuracy@10"], "Metric"], [8737, 8741, ["Image", "pre", "-", "smoothing"], "Task"], [8746, 8748, ["JPEG", "images"], "Material"], [8766, 8769, ["uncompressed", "PNG", "images"], "Material"], [8770, 8773, ["MPI", "-", "Sintel"], "Material"], [8774, 8775, ["Kitti"], "Material"], [8794, 8795, ["regularizer"], "Method"], [8823, 8825, ["PNG", "images"], "Material"], [8842, 8846, ["Non", "-", "linear", "rectification"], "Task"], [8864, 8866, ["power", "normalization"], "Method"], [8882, 8883, ["accuracy@10"], "Metric"], [8918, 8922, ["backtracking", "and", "scoring", "schemes"], "Method"], [8928, 8929, ["DM"], "Method"], [8943, 8944, ["DM"], "Method"], [8988, 8990, ["Scoring", "scheme"], "Method"], [9015, 9017, ["scoring", "scheme"], "Method"], [9056, 9057, ["DM"], "Method"], [9089, 9093, ["scoring", "and", "backtracking", "approach"], "Method"], [9175, 9177, ["memory", "usage"], "Metric"], [9178, 9181, ["run", "-", "time"], "Metric"], [9222, 9224, ["scoring", "scheme"], "Method"], [9227, 9228, ["optimization"], "Task"], [9230, 9235, ["early", "pruning", "of", "backtracking", "paths"], "Task"], [9260, 9266, ["run", "-", "time", "and", "memory", "usage"], "Metric"], [9277, 9279, ["computational", "load"], "Metric"], [9303, 9305, ["compression", "scheme"], "Method"], [9345, 9348, ["k", "-", "means"], "Method"], [9377, 9378, ["accuracy@10"], "Metric"], [9380, 9382, ["memory", "consumption"], "Metric"], [9404, 9405, ["DeepMatching"], "Method"], [9490, 9492, ["dictionary", "sizes"], "Metric"], [9499, 9502, ["MPI", "-", "Sintel"], "Material"], [9521, 9523, ["computation", "time"], "Metric"], [9538, 9539, ["DeepMatching"], "Method"], [9549, 9553, ["bottom", "-", "up", "pass"], "Method"], [9557, 9558, ["approximation"], "Method"], [9561, 9564, ["run", "-", "time"], "Metric"], [9573, 9576, ["max", "-", "pooling"], "Method"], [9577, 9578, ["subsampling"], "Method"], [9579, 9580, ["aggregation"], "Method"], [9581, 9585, ["non", "-", "linear", "rectification"], "Method"], [9615, 9618, ["k", "-", "means"], "Method"], [9638, 9641, ["run", "-", "time"], "Metric"], [9661, 9663, ["GPU", "Implementation"], "Method"], [9667, 9668, ["DM"], "Method"], [9669, 9670, ["GPU"], "Method"], [9672, 9675, ["Caffe", "framework", "caffe"], "Method"], [9678, 9680, ["Caffe", "layers"], "Method"], [9681, 9682, ["ConvolutionLayer"], "Method"], [9683, 9684, ["PoolingLayer"], "Method"], [9717, 9719, ["aggregation", "layer"], "Method"], [9737, 9740, ["sparse", "matrix", "multiplication"], "Method"], [9742, 9744, ["cuSPARSE", "toolbox"], "Method"], [9756, 9759, ["GeForce", "Titan", "X."], "Material"], [9778, 9780, ["computation", "bottleneck"], "Metric"], [9799, 9801, ["patch", "descriptors"], "Method"], [9842, 9844, ["image", "smoothing"], "Method"], [9845, 9847, ["gradient", "extraction"], "Method"], [9873, 9874, ["DM"], "Method"], [9885, 9887, ["matching", "algorithms"], "Method"], [9894, 9896, ["DoG", "detector"], "Method"], [9900, 9902, ["FLANN", "flann"], "Method"], [9906, 9909, ["SIFT", "-", "NN"], "Method"], [9910, 9913, ["dense", "HOG", "matching"], "Method"], [9921, 9923, ["reciprocal", "verification"], "Method"], [9932, 9935, ["HOG", "-", "NN"], "Method"], [9936, 9938, ["Generalized", "PatchMatch"], "Method"], [9939, 9940, ["GPM"], "Method"], [9960, 9964, ["Kd", "-", "tree", "PatchMatch"], "Method"], [9965, 9966, ["KPM"], "Method"], [9967, 9968, ["kpm"], "Method"], [9973, 9974, ["PatchMatch"], "Method"], [9977, 9979, ["patch", "descriptors"], "Method"], [9980, 9983, ["kd", "-", "trees"], "Method"], [9985, 9987, ["correspondence", "propagation"], "Task"], [10002, 10003, ["GPM"], "Method"], [10006, 10009, ["multiscale", "iterative", "expansion"], "Method"], [10010, 10012, ["contraction", "strategy"], "Method"], [10023, 10025, ["energy", "minimization"], "Method"], [10048, 10049, ["SLS"], "Method"], [10071, 10076, ["scale", "-", "invariant", "pixel", "representation"], "Method"], [10077, 10078, ["DaisyFilterFlow"], "Method"], [10079, 10080, ["DaisyFF"], "Method"], [10082, 10085, ["dense", "matching", "approach"], "Method"], [10087, 10093, ["filter", "-", "based", "efficient", "flow", "inference"], "Method"], [10095, 10101, ["Patch", "-", "Match", "fast", "search", "algorithm"], "Method"], [10107, 10109, ["DAISY", "representation"], "Method"], [10109, 10112, ["Deformable", "Pyramid", "Matching"], "Method"], [10113, 10114, ["DSP"], "Method"], [10116, 10119, ["dense", "matching", "approach"], "Method"], [10134, 10135, ["inference"], "Task"], [10148, 10151, ["HOG", "-", "NN"], "Method"], [10152, 10153, ["DM"], "Method"], [10170, 10171, ["GPM"], "Method"], [10172, 10173, ["NRDC"], "Method"], [10174, 10175, ["DaisyFF"], "Method"], [10176, 10180, ["scale", "and", "rotation", "invariant"], "Method"], [10182, 10185, ["HOG", "-", "NN"], "Method"], [10186, 10187, ["KPM"], "Method"], [10192, 10193, ["SLS"], "Method"], [10194, 10195, ["DSP"], "Method"], [10245, 10246, ["Coverage"], "Task"], [10304, 10305, ["DeepMatching"], "Method"], [10313, 10316, ["SIFT", "-", "NN"], "Method"], [10321, 10324, ["HOG", "-", "NN"], "Method"], [10330, 10332, ["DM", "matches"], "Task"], [10337, 10340, ["HOG", "-", "NN"], "Method"], [10341, 10344, ["SIFT", "-", "NN"], "Method"], [10348, 10350, ["DM", "matches"], "Task"], [10364, 10367, ["HOG", "-", "NN"], "Method"], [10368, 10371, ["SIFT", "-", "NN"], "Method"], [10415, 10416, ["DM"], "Method"], [10441, 10444, ["SIFT", "-", "NN"], "Method"], [10450, 10453, ["global", "image", "transformation"], "Task"], [10507, 10510, ["HOG", "-", "NN"], "Method"], [10512, 10516, ["predicting", "dense", "correspondence", "fields"], "Task"], [10527, 10528, ["KPM"], "Method"], [10529, 10530, ["GPM"], "Method"], [10535, 10536, ["DSP"], "Method"], [10567, 10568, ["DSP"], "Method"], [10575, 10578, ["MPI", "-", "Sintel"], "Material"], [10587, 10589, ["dense", "methods"], "Method"], [10590, 10591, ["NRDC"], "Method"], [10592, 10593, ["SLS"], "Method"], [10594, 10595, ["DaisyFF"], "Method"], [10607, 10608, ["Kitti"], "Material"], [10639, 10640, ["DM"], "Method"], [10650, 10651, ["DSP"], "Method"], [10656, 10658, ["hierarchical", "matching"], "Method"], [10681, 10682, ["SLS"], "Method"], [10683, 10684, ["DaisyFF"], "Method"], [10692, 10695, ["run", "-", "time"], "Metric"], [10723, 10725, ["sequential", "processing"], "Method"], [10747, 10748, ["DM"], "Method"], [10760, 10763, ["MPI", "-", "Sintel"], "Material"], [10764, 10765, ["Kitti"], "Material"], [10769, 10770, ["DM"], "Method"], [10776, 10779, ["run", "-", "time"], "Metric"], [10785, 10788, ["SIFT", "-", "NN"], "Method"], [10792, 10794, ["memory", "usage"], "Metric"], [10805, 10807, ["accuracy@10", "measure"], "Metric"], [10852, 10855, ["Optical", "Flow", "Experiments"], "Task"], [10862, 10865, ["optical", "flow", "estimation"], "Task"], [10866, 10868, ["Optical", "flow"], "Task"], [10872, 10874, ["variational", "framework"], "Method"], [10898, 10899, ["DeepMatching"], "Method"], [10901, 10903, ["matching", "methods"], "Method"], [10924, 10925, ["DeepFlow"], "Method"], [10974, 10975, ["Kitti"], "Material"], [10976, 10977, ["Middlebury"], "Material"], [10997, 10999, ["parameter", "tuning"], "Method"], [11020, 11023, ["gradient", "descent", "strategy"], "Method"], [11029, 11032, ["local", "grid", "search"], "Method"], [11100, 11101, ["DeepFlow"], "Method"], [11107, 11108, ["DeepMatching"], "Method"], [11114, 11115, ["Kitti"], "Material"], [11116, 11117, ["Middlebury"], "Material"], [11150, 11152, ["matching", "methods"], "Method"], [11164, 11165, ["DeepFlow"], "Method"], [11174, 11176, ["matching", "approaches"], "Method"], [11208, 11209, ["DM"], "Method"], [11218, 11220, ["matching", "approach"], "Method"], [11226, 11229, ["MPI", "-", "Sintel"], "Material"], [11236, 11238, ["endpoint", "error"], "Metric"], [11251, 11252, ["matching"], "Method"], [11253, 11254, ["DM"], "Method"], [11259, 11261, ["flow", "estimation"], "Task"], [11267, 11270, ["MPI", "-", "Sintel"], "Material"], [11271, 11272, ["Kitti"], "Material"], [11303, 11306, ["SIFT", "-", "NN"], "Method"], [11307, 11310, ["HOG", "-", "NN"], "Method"], [11311, 11314, ["MPI", "-", "Sintel"], "Material"], [11315, 11316, ["Kitti"], "Material"], [11358, 11361, ["SIFT", "-", "NN"], "Method"], [11362, 11365, ["HOG", "-", "NN"], "Method"], [11372, 11374, ["variational", "approach"], "Method"], [11392, 11394, ["dense", "methods"], "Method"], [11488, 11489, ["DeepMatching"], "Method"], [11501, 11504, ["optical", "flow", "accuracy"], "Metric"], [11521, 11524, ["MPI", "-", "Sintel"], "Material"], [11529, 11530, ["Kitti"], "Material"], [11537, 11538, ["DM"], "Method"], [11562, 11565, ["reciprocal", "verification", "filtering"], "Method"], [11567, 11568, ["DM"], "Method"], [11578, 11579, ["approximation"], "Method"], [11601, 11604, ["MPI", "-", "Sintel"], "Material"], [11609, 11610, ["DeepMatching"], "Method"], [11617, 11618, ["robustness"], "Metric"], [11631, 11634, ["MPI", "-", "Sintel"], "Material"], [11650, 11651, ["DeepFlow"], "Method"], [11662, 11665, ["MPI", "-", "Sintel"], "Material"], [11666, 11667, ["Kitti"], "Material"], [11668, 11669, ["Middlebury"], "Material"], [11694, 11695, ["Kitti"], "Material"], [11696, 11697, ["Middlebury"], "Material"], [11715, 11718, ["MPI", "-", "Sintel"], "Material"], [11760, 11762, ["DeepFlow", "DeepFlow"], "Method"], [11766, 11768, ["DeepFlow", "*"], "Method"], [11809, 11810, ["DeepFlow"], "Method"], [11817, 11820, ["MPI", "-", "Sintel"], "Material"], [11839, 11841, ["variational", "approach"], "Method"], [11855, 11856, ["epicflow"], "Method"], [11883, 11884, ["DeepMatching"], "Method"], [11900, 11901, ["DeepMatching"], "Method"], [11911, 11913, ["SparseFlowFused", "sparseflowfused"], "Method"], [11917, 11918, ["DeepMatching"], "Method"], [11953, 11954, ["DeepMatching"], "Method"], [11972, 11974, ["variational", "part"], "Method"], [11987, 11988, ["GPU"], "Method"], [12000, 12001, ["DeepFlow"], "Method"], [12038, 12039, ["EpicFlow"], "Method"], [12078, 12081, ["EPE", "-", "Noc"], "Method"], [12083, 12084, ["EPE"], "Method"], [12140, 12143, ["EPE", "-", "noc"], "Task"], [12144, 12145, ["DeepFlow"], "Method"], [12224, 12225, ["DeepFlow"], "Method"], [12241, 12242, ["EPE"], "Metric"], [12253, 12255, ["affine", "regularizers"], "Method"], [12256, 12259, ["BTF", "-", "ILLUM"], "Method"], [12261, 12264, ["NLTGB", "-", "SC"], "Method"], [12266, 12267, ["TGV2ADCSIFT"], "Method"], [12274, 12277, ["local", "affine", "estimators"], "Method"], [12278, 12279, ["EpicFlow"], "Method"], [12288, 12289, ["Kitti"], "Material"], [12351, 12354, ["average", "angular", "error"], "Metric"], [12360, 12363, ["MPI", "-", "Sintel"], "Material"], [12382, 12383, ["DeepFlow"], "Method"], [12385, 12388, ["average", "endpoint", "error"], "Metric"], [12414, 12415, ["DeepMatching"], "Method"], [12432, 12433, ["DeepMatching"], "Method"], [12448, 12449, ["Mikolajczyk"], "Material"], [12452, 12455, ["MPI", "-", "Sintel"], "Material"], [12458, 12459, ["Kitti"], "Material"], [12465, 12469, ["variational", "energy", "minimization", "approach"], "Method"], [12475, 12478, ["optical", "flow", "estimation"], "Task"], [12480, 12481, ["DeepFlow"], "Method"], [12486, 12489, ["optical", "flow", "benchmarks"], "Task"], [12536, 12537, ["DeepMatching"], "Method"], [12538, 12540, ["thin", "objects"], "Task"], [39, 40, ["matching"], "Method"], [98, 99, ["Mikolajczyk"], "Material"], [102, 105, ["MPI", "-", "Sintel"], "Material"], [108, 109, ["Kitti"], "Material"], [289, 290, ["matching"], "Method"], [300, 301, ["matching"], "Method"], [339, 340, ["matching"], "Method"], [451, 452, ["matching"], "Method"], [530, 531, ["matching"], "Method"], [542, 543, ["matching"], "Method"], [668, 669, ["matching"], "Method"], [797, 798, ["matching"], "Method"], [883, 884, ["matching"], "Method"], [901, 902, ["DeepMatching"], "Method"], [944, 945, ["matching"], "Method"], [1269, 1270, ["matching"], "Method"], [1293, 1294, ["matching"], "Method"], [1303, 1304, ["matching"], "Method"], [1305, 1306, ["matching"], "Method"], [1353, 1354, ["matching"], "Method"], [1533, 1537, ["non", "-", "rigid", "matching"], "Method"], [1643, 1644, ["matching"], "Method"], [1657, 1658, ["kpm"], "Method"], [1718, 1720, ["contraction", "strategy"], "Method"], [2003, 2004, ["matching"], "Method"], [2065, 2066, ["matching"], "Method"], [2088, 2089, ["SIFT"], "Method"], [2104, 2105, ["matching"], "Method"], [2115, 2116, ["SIFT"], "Method"], [2140, 2141, ["matching"], "Method"], [2231, 2232, ["matching"], "Method"], [2403, 2404, ["matching"], "Method"], [2481, 2482, ["SIFT"], "Method"], [2633, 2637, ["non", "-", "rigid", "matching"], "Method"], [2663, 2664, ["matching"], "Method"], [2709, 2710, ["matching"], "Method"], [2722, 2723, ["matching"], "Method"], [2789, 2790, ["matching"], "Method"], [2805, 2806, ["matching"], "Method"], [2824, 2825, ["matching"], "Method"], [4510, 4511, ["matching"], "Method"], [4589, 4590, ["DeepMatching"], "Method"], [4608, 4609, ["matching"], "Method"], [4896, 4897, ["DeepMatching"], "Method"], [4930, 4931, ["DeepMatching"], "Method"], [5192, 5193, ["matching"], "Method"], [5593, 5594, ["DeepMatching"], "Method"], [5824, 5825, ["DeepMatching"], "Method"], [5858, 5859, ["matching"], "Method"], [6183, 6184, ["DeepMatching"], "Method"], [6244, 6245, ["DeepMatching"], "Method"], [6249, 6250, ["matching"], "Method"], [6588, 6589, ["DeepMatching"], "Method"], [6669, 6670, ["DM"], "Method"], [7480, 7481, ["matching"], "Method"], [7493, 7494, ["matching"], "Method"], [7565, 7566, ["matching"], "Method"], [7683, 7684, ["matching"], "Method"], [7791, 7792, ["matching"], "Method"], [7828, 7829, ["matching"], "Method"], [7842, 7843, ["Mikolajczyk"], "Material"], [7941, 7942, ["DeepMatching"], "Method"], [7973, 7976, ["MPI", "-", "Sintel"], "Material"], [8070, 8071, ["Middlebury"], "Material"], [8135, 8136, ["Kitti"], "Material"], [8263, 8264, ["matching"], "Method"], [8557, 8558, ["Middlebury"], "Material"], [8573, 8574, ["matching"], "Method"], [8579, 8580, ["Mikolajczyk"], "Material"], [8591, 8592, ["DM"], "Method"], [8612, 8613, ["Mikolajczyk"], "Material"], [8749, 8750, ["Mikolajczyk"], "Material"], [8954, 8955, ["DM"], "Method"], [8992, 8993, ["DM"], "Method"], [9242, 9243, ["DeepMatching"], "Method"], [9250, 9251, ["DeepMatching"], "Method"], [9374, 9375, ["matching"], "Method"], [9447, 9448, ["DeepMatching"], "Method"], [9654, 9655, ["DeepMatching"], "Method"], [9769, 9772, ["MPI", "-", "Sintel"], "Material"], [9890, 9891, ["SIFT"], "Method"], [9919, 9920, ["matching"], "Method"], [9926, 9927, ["LDOF"], "Method"], [10013, 10014, ["SIFT"], "Method"], [10018, 10019, ["matching"], "Method"], [10030, 10031, ["SIFT"], "Method"], [10046, 10047, ["SIFT"], "Method"], [10053, 10054, ["SIFT"], "Method"], [10144, 10145, ["SIFT"], "Method"], [10167, 10168, ["SIFT"], "Method"], [10188, 10189, ["SIFT"], "Method"], [10212, 10213, ["Mikolajczyk"], "Material"], [10454, 10455, ["Mikolajczyk"], "Material"], [10487, 10488, ["SIFT"], "Method"], [10531, 10532, ["SIFT"], "Method"], [10545, 10546, ["Kitti"], "Material"], [10563, 10564, ["SIFT"], "Method"], [10622, 10625, ["MPI", "-", "Sintel"], "Material"], [10626, 10627, ["Mikolajczyk"], "Material"], [10665, 10666, ["DeepMatching"], "Method"], [10704, 10705, ["DM"], "Method"], [10709, 10710, ["Mikolajczyk"], "Material"], [10811, 10812, ["SIFT"], "Method"], [10814, 10815, ["GPM"], "Method"], [10817, 10818, ["DeepMatching"], "Method"], [10819, 10820, ["SIFT"], "Method"], [10822, 10825, ["HOG", "-", "NN"], "Method"], [10825, 10826, ["KPM"], "Method"], [10826, 10827, ["GPM"], "Method"], [10827, 10828, ["SIFT"], "Method"], [10833, 10834, ["DM"], "Method"], [10835, 10836, ["SIFT"], "Method"], [10838, 10841, ["HOG", "-", "NN"], "Method"], [10841, 10842, ["KPM"], "Method"], [10842, 10843, ["GPM"], "Method"], [10843, 10844, ["SIFT"], "Method"], [10849, 10850, ["DM"], "Method"], [10930, 10933, ["MPI", "-", "Sintel"], "Material"], [11187, 11188, ["matching"], "Method"], [11288, 11289, ["Middlebury"], "Material"], [11414, 11415, ["matching"], "Method"], [11482, 11483, ["Kitti"], "Material"], [11592, 11593, ["Kitti"], "Material"], [11619, 11620, ["matching"], "Method"], [11628, 11629, ["Kitti"], "Material"], [11723, 11724, ["DeepMatching"], "Method"], [11728, 11729, ["LDOF"], "Method"], [11747, 11750, ["MPI", "-", "Sintel"], "Material"], [11799, 11800, ["DM"], "Method"], [11835, 11836, ["matching"], "Method"], [11872, 11875, ["MPI", "-", "Sintel"], "Material"], [11893, 11894, ["EpicFlow"], "Method"], [11894, 11895, ["epicflow"], "Method"], [11929, 11932, ["MPI", "-", "Sintel"], "Material"], [12039, 12040, ["epicflow"], "Method"], [12045, 12046, ["Kitti"], "Material"], [12057, 12058, ["Kitti"], "Material"], [12074, 12075, ["Kitti"], "Material"], [12170, 12171, ["Kitti"], "Material"], [12279, 12280, ["epicflow"], "Method"], [12290, 12293, ["MPI", "-", "Sintel"], "Material"], [12307, 12310, ["MPI", "-", "Sintel"], "Material"], [12337, 12338, ["Middlebury"], "Material"], [12345, 12346, ["Middlebury"], "Material"], [12372, 12373, ["matching"], "Method"], [12410, 12411, ["matching"], "Method"], [12444, 12445, ["matching"], "Method"]]} {"doc_id": "2315fc6c2c0c4abd2443e26a26e7bb86df8e24cc", "text": ["ImageNet", "Classification", "with", "Deep", "Convolutional", "Neural", "Networks", "We", "trained", "a", "large", ",", "deep", "convolutional", "neural", "network", "to", "classify", "the", "1.2", "million", "high", "-", "resolution", "images", "in", "the", "ImageNet", "LSVRC", "-", "2010", "contest", "into", "the", "1000", "different", "classes", ".", "On", "the", "test", "data", ",", "we", "achieved", "top", "-", "1", "and", "top", "-", "5", "error", "rates", "of", "37.5", "%", "and", "17.0", "%", ",", "respectively", ",", "which", "is", "considerably", "better", "than", "the", "previous", "state", "-", "of", "-", "the", "-", "art", ".", "The", "neural", "network", ",", "which", "has", "60", "million", "parameters", "and", "650", ",", "000", "neurons", ",", "consists", "of", "five", "convolutional", "layers", ",", "some", "of", "which", "are", "followed", "by", "max", "-", "pooling", "layers", ",", "and", "three", "fully", "connected", "layers", "with", "a", "final", "1000", "-", "way", "softmax", ".", "To", "make", "training", "faster", ",", "we", "used", "non", "-", "saturating", "neurons", "and", "a", "very", "efficient", "GPU", "implementation", "of", "the", "convolution", "operation", ".", "To", "reduce", "overfitting", "in", "the", "fully", "connected", "layers", "we", "employed", "a", "recently", "developed", "regularization", "method", "called", "\"", "dropout", "\"", "that", "proved", "to", "be", "very", "effective", ".", "We", "also", "entered", "a", "variant", "of", "this", "model", "in", "the", "ILSVRC", "-", "2012", "competition", "and", "achieved", "a", "winning", "top", "-", "5", "test", "error", "rate", "of", "15.3", "%", ",", "compared", "to", "26.2", "%", "achieved", "by", "the", "second", "-", "best", "entry", ".", "We", "trained", "a", "large", ",", "deep", "convolutional", "neural", "network", "to", "classify", "the", "1.2", "million", "high", "-", "resolution", "images", "in", "the", "ImageNet", "LSVRC", "-", "2010", "contest", "into", "the", "1000", "different", "classes", ".", "On", "the", "test", "data", ",", "we", "achieved", "top", "-", "1", "and", "top", "-", "5", "error", "rates", "of", "37.5", "%", "and", "17.0", "%", "which", "is", "considerably", "better", "than", "the", "previous", "state", "-", "of", "-", "the", "-", "art", ".", "The", "neural", "network", ",", "which", "has", "60", "million", "parameters", "and", "650", ",", "000", "neurons", ",", "consists", "of", "five", "convolutional", "layers", ",", "some", "of", "which", "are", "followed", "by", "max", "-", "pooling", "layers", ",", "and", "three", "fully", "-", "connected", "layers", "with", "a", "final", "1000", "-", "way", "softmax", ".", "To", "make", "training", "faster", ",", "we", "used", "non", "-", "saturating", "neurons", "and", "a", "very", "efficient", "GPU", "implementation", "of", "the", "convolution", "operation", ".", "To", "reduce", "overfitting", "in", "the", "fully", "-", "connected", "layers", "we", "employed", "a", "recently", "-", "developed", "regularization", "method", "called", "“", "dropout", "”", "that", "proved", "to", "be", "very", "effective", ".", "We", "also", "entered", "a", "variant", "of", "this", "model", "in", "the", "ILSVRC", "-", "2012", "competition", "and", "achieved", "a", "winning", "top", "-", "5", "test", "error", "rate", "of", "15.3", "%", ",", "compared", "to", "26.2", "%", "achieved", "by", "the", "second", "-", "best", "entry", ".", "1", "Introduction", "Current", "approaches", "to", "object", "recognition", "make", "essential", "use", "of", "machine", "learning", "methods", ".", "To", "improve", "their", "performance", ",", "we", "can", "collect", "larger", "datasets", ",", "learn", "more", "powerful", "models", ",", "and", "use", "better", "techniques", "for", "preventing", "overfitting", ".", "Until", "recently", ",", "datasets", "of", "labeled", "images", "were", "relatively", "small", "—", "on", "the", "order", "of", "tens", "of", "thousands", "of", "images", "(", "e.g.", ",", "NORB", "[", "16", "]", ",", "Caltech", "-", "101", "/", "256", "[", "8", ",", "9", "]", ",", "and", "CIFAR", "-", "10", "/", "100", "[", "12", "]", ")", ".", "Simple", "recognition", "tasks", "can", "be", "solved", "quite", "well", "with", "datasets", "of", "this", "size", ",", "especially", "if", "they", "are", "augmented", "with", "label", "-", "preserving", "transformations", ".", "For", "example", ",", "the", "currentbest", "error", "rate", "on", "the", "MNIST", "digit", "-", "recognition", "task", "(", "<", "0.3", "%", ")", "approaches", "human", "performance", "[", "4", "]", ".", "But", "objects", "in", "realistic", "settings", "exhibit", "considerable", "variability", ",", "so", "to", "learn", "to", "recognize", "them", "it", "is", "necessary", "to", "use", "much", "larger", "training", "sets", ".", "And", "indeed", ",", "the", "shortcomings", "of", "small", "image", "datasets", "have", "been", "widely", "recognized", "(", "e.g.", ",", "Pinto", "et", "al", ".", "[", "21", "]", ")", ",", "but", "it", "has", "only", "recently", "become", "possible", "to", "collect", "labeled", "datasets", "with", "millions", "of", "images", ".", "The", "new", "larger", "datasets", "include", "LabelMe", "[", "23", "]", ",", "which", "consists", "of", "hundreds", "of", "thousands", "of", "fully", "-", "segmented", "images", ",", "and", "ImageNet", "[", "6", "]", ",", "which", "consists", "of", "over", "15", "million", "labeled", "high", "-", "resolution", "images", "in", "over", "22", ",", "000", "categories", ".", "To", "learn", "about", "thousands", "of", "objects", "from", "millions", "of", "images", ",", "we", "need", "a", "model", "with", "a", "large", "learning", "capacity", ".", "However", ",", "the", "immense", "complexity", "of", "the", "object", "recognition", "task", "means", "that", "this", "problem", "can", "not", "be", "specified", "even", "by", "a", "dataset", "as", "large", "as", "ImageNet", ",", "so", "our", "model", "should", "also", "have", "lots", "of", "prior", "knowledge", "to", "compensate", "for", "all", "the", "data", "we", "do", "n’t", "have", ".", "Convolutional", "neural", "networks", "(", "CNNs", ")", "constitute", "one", "such", "class", "of", "models", "[", "16", ",", "11", ",", "13", ",", "18", ",", "15", ",", "22", ",", "26", "]", ".", "Their", "capacity", "can", "be", "controlled", "by", "varying", "their", "depth", "and", "breadth", ",", "and", "they", "also", "make", "strong", "and", "mostly", "correct", "assumptions", "about", "the", "nature", "of", "images", "(", "namely", ",", "stationarity", "of", "statistics", "and", "locality", "of", "pixel", "dependencies", ")", ".", "Thus", ",", "compared", "to", "standard", "feedforward", "neural", "networks", "with", "similarly", "-", "sized", "layers", ",", "CNNs", "have", "much", "fewer", "connections", "and", "parameters", "and", "so", "they", "are", "easier", "to", "train", ",", "while", "their", "theoretically", "-", "best", "performance", "is", "likely", "to", "be", "only", "slightly", "worse", ".", "Despite", "the", "attractive", "qualities", "of", "CNNs", ",", "and", "despite", "the", "relative", "efficiency", "of", "their", "local", "architecture", ",", "they", "have", "still", "been", "prohibitively", "expensive", "to", "apply", "in", "large", "scale", "to", "high", "-", "resolution", "images", ".", "Luckily", ",", "current", "GPUs", ",", "paired", "with", "a", "highly", "-", "optimized", "implementation", "of", "2D", "convolution", ",", "are", "powerful", "enough", "to", "facilitate", "the", "training", "of", "interestingly", "-", "large", "CNNs", ",", "and", "recent", "datasets", "such", "as", "ImageNet", "contain", "enough", "labeled", "examples", "to", "train", "such", "models", "without", "severe", "overfitting", ".", "The", "specific", "contributions", "of", "this", "paper", "are", "as", "follows", ":", "we", "trained", "one", "of", "the", "largest", "convolutional", "neural", "networks", "to", "date", "on", "the", "subsets", "of", "ImageNet", "used", "in", "the", "ILSVRC", "-", "2010", "and", "ILSVRC", "-", "2012", "competitions", "[", "2", "]", "and", "achieved", "by", "far", "the", "best", "results", "ever", "reported", "on", "these", "datasets", ".", "We", "wrote", "a", "highly", "-", "optimized", "GPU", "implementation", "of", "2D", "convolution", "and", "all", "the", "other", "operations", "inherent", "in", "training", "convolutional", "neural", "networks", ",", "which", "we", "make", "available", "publicly1", ".", "Our", "network", "contains", "a", "number", "of", "new", "and", "unusual", "features", "which", "improve", "its", "performance", "and", "reduce", "its", "training", "time", ",", "which", "are", "detailed", "in", "Section", "3", ".", "The", "size", "of", "our", "network", "made", "overfitting", "a", "significant", "problem", ",", "even", "with", "1.2", "million", "labeled", "training", "examples", ",", "so", "we", "used", "several", "effective", "techniques", "for", "preventing", "overfitting", ",", "which", "are", "described", "in", "Section", "4", ".", "Our", "final", "network", "contains", "five", "convolutional", "and", "three", "fully", "-", "connected", "layers", ",", "and", "this", "depth", "seems", "to", "be", "important", ":", "we", "found", "that", "removing", "any", "convolutional", "layer", "(", "each", "of", "which", "contains", "no", "more", "than", "1", "%", "of", "the", "model", "’s", "parameters", ")", "resulted", "in", "inferior", "performance", ".", "In", "the", "end", ",", "the", "network", "’s", "size", "is", "limited", "mainly", "by", "the", "amount", "of", "memory", "available", "on", "current", "GPUs", "and", "by", "the", "amount", "of", "training", "time", "that", "we", "are", "willing", "to", "tolerate", ".", "Our", "network", "takes", "between", "five", "and", "six", "days", "to", "train", "on", "two", "GTX", "580", "3", "GB", "GPUs", ".", "All", "of", "our", "experiments", "suggest", "that", "our", "results", "can", "be", "improved", "simply", "by", "waiting", "for", "faster", "GPUs", "and", "bigger", "datasets", "to", "become", "available", ".", "2", "The", "Dataset", "ImageNet", "is", "a", "dataset", "of", "over", "15", "million", "labeled", "high", "-", "resolution", "images", "belonging", "to", "roughly", "22", ",", "000", "categories", ".", "The", "images", "were", "collected", "from", "the", "web", "and", "labeled", "by", "human", "labelers", "using", "Amazon", "’s", "Mechanical", "Turk", "crowd", "-", "sourcing", "tool", ".", "Starting", "in", "2010", ",", "as", "part", "of", "the", "Pascal", "Visual", "Object", "Challenge", ",", "an", "annual", "competition", "called", "the", "ImageNet", "Large", "-", "Scale", "Visual", "Recognition", "Challenge", "(", "ILSVRC", ")", "has", "been", "held", ".", "ILSVRC", "uses", "a", "subset", "of", "ImageNet", "with", "roughly", "1000", "images", "in", "each", "of", "1000", "categories", ".", "In", "all", ",", "there", "are", "roughly", "1.2", "million", "training", "images", ",", "50", ",", "000", "validation", "images", ",", "and", "150", ",", "000", "testing", "images", ".", "ILSVRC", "-", "2010", "is", "the", "only", "version", "of", "ILSVRC", "for", "which", "the", "test", "set", "labels", "are", "available", ",", "so", "this", "is", "the", "version", "on", "which", "we", "performed", "most", "of", "our", "experiments", ".", "Since", "we", "also", "entered", "our", "model", "in", "the", "ILSVRC", "-", "2012", "competition", ",", "in", "Section", "6", "we", "report", "our", "results", "on", "this", "version", "of", "the", "dataset", "as", "well", ",", "for", "which", "test", "set", "labels", "are", "unavailable", ".", "On", "ImageNet", ",", "it", "is", "customary", "to", "report", "two", "error", "rates", ":", "top", "-", "1", "and", "top", "-", "5", ",", "where", "the", "top", "-", "5", "error", "rate", "is", "the", "fraction", "of", "test", "images", "for", "which", "the", "correct", "label", "is", "not", "among", "the", "five", "labels", "considered", "most", "probable", "by", "the", "model", ".", "ImageNet", "consists", "of", "variable", "-", "resolution", "images", ",", "while", "our", "system", "requires", "a", "constant", "input", "dimensionality", ".", "Therefore", ",", "we", "down", "-", "sampled", "the", "images", "to", "a", "fixed", "resolution", "of", "256", "⇥", "256", ".", "Given", "a", "rectangular", "image", ",", "we", "first", "rescaled", "the", "image", "such", "that", "the", "shorter", "side", "was", "of", "length", "256", ",", "and", "then", "cropped", "out", "the", "central", "256", "⇥", "256", "patch", "from", "the", "resulting", "image", ".", "We", "did", "not", "pre", "-", "process", "the", "images", "in", "any", "other", "way", ",", "except", "for", "subtracting", "the", "mean", "activity", "over", "the", "training", "set", "from", "each", "pixel", ".", "So", "we", "trained", "our", "network", "on", "the", "(", "centered", ")", "raw", "RGB", "values", "of", "the", "pixels", ".", "3", "The", "Architecture", "The", "architecture", "of", "our", "network", "is", "summarized", "in", "Figure", "2", ".", "It", "contains", "eight", "learned", "layers", "—", "five", "convolutional", "and", "three", "fully", "-", "connected", ".", "Below", ",", "we", "describe", "some", "of", "the", "novel", "or", "unusual", "features", "of", "our", "network", "’s", "architecture", ".", "Sections", "3.1", "-", "3.4", "are", "sorted", "according", "to", "our", "estimation", "of", "their", "importance", ",", "with", "the", "most", "important", "first", ".", "1http:", "//", "code.google.com", "/", "p", "/", "cuda", "-", "convnet", "/", "3.1", "ReLU", "Nonlinearity", "The", "standard", "way", "to", "model", "a", "neuron", "’s", "output", "f", "as", "a", "function", "of", "its", "input", "x", "is", "with", "f", "(", "x", ")", "=", "tanh", "(", "x", ")", "or", "f", "(", "x", ")", "=", "(", "1", "+", "e", "x", ")", "1", ".", "In", "terms", "of", "training", "time", "with", "gradient", "descent", ",", "these", "saturating", "nonlinearities", "are", "much", "slower", "than", "the", "non", "-", "saturating", "nonlinearity", "f", "(", "x", ")", "=", "max", "(", "0", ",", "x", ")", ".", "Following", "Nair", "and", "Hinton", "[", "20", "]", ",", "we", "refer", "to", "neurons", "with", "this", "nonlinearity", "as", "Rectified", "Linear", "Units", "(", "ReLUs", ")", ".", "Deep", "convolutional", "neural", "networks", "with", "ReLUs", "train", "several", "times", "faster", "than", "their", "equivalents", "with", "tanh", "units", ".", "This", "is", "demonstrated", "in", "Figure", "1", ",", "which", "shows", "the", "number", "of", "iterations", "required", "to", "reach", "25", "%", "training", "error", "on", "the", "CIFAR", "-", "10", "dataset", "for", "a", "particular", "four", "-", "layer", "convolutional", "network", ".", "This", "plot", "shows", "that", "we", "would", "not", "have", "been", "able", "to", "experiment", "with", "such", "large", "neural", "networks", "for", "this", "work", "if", "we", "had", "used", "traditional", "saturating", "neuron", "models", ".", "We", "are", "not", "the", "first", "to", "consider", "alternatives", "to", "traditional", "neuron", "models", "in", "CNNs", ".", "For", "example", ",", "Jarrett", "et", "al", ".", "[", "11", "]", "claim", "that", "the", "nonlinearity", "f", "(", "x", ")", "=", "|tanh", "(", "x", ")", "|", "works", "particularly", "well", "with", "their", "type", "of", "contrast", "normalization", "followed", "by", "local", "average", "pooling", "on", "the", "Caltech", "-", "101", "dataset", ".", "However", ",", "on", "this", "dataset", "the", "primary", "concern", "is", "preventing", "overfitting", ",", "so", "the", "effect", "they", "are", "observing", "is", "different", "from", "the", "accelerated", "ability", "to", "fit", "the", "training", "set", "which", "we", "report", "when", "using", "ReLUs", ".", "Faster", "learning", "has", "a", "great", "influence", "on", "the", "performance", "of", "large", "models", "trained", "on", "large", "datasets", ".", "3.2", "Training", "on", "Multiple", "GPUs", "A", "single", "GTX", "580", "GPU", "has", "only", "3", "GB", "of", "memory", ",", "which", "limits", "the", "maximum", "size", "of", "the", "networks", "that", "can", "be", "trained", "on", "it", ".", "It", "turns", "out", "that", "1.2", "million", "training", "examples", "are", "enough", "to", "train", "networks", "which", "are", "too", "big", "to", "fit", "on", "one", "GPU", ".", "Therefore", "we", "spread", "the", "net", "across", "two", "GPUs", ".", "Current", "GPUs", "are", "particularly", "well", "-", "suited", "to", "cross", "-", "GPU", "parallelization", ",", "as", "they", "are", "able", "to", "read", "from", "and", "write", "to", "one", "another", "’s", "memory", "directly", ",", "without", "going", "through", "host", "machine", "memory", ".", "The", "parallelization", "scheme", "that", "we", "employ", "essentially", "puts", "half", "of", "the", "kernels", "(", "or", "neurons", ")", "on", "each", "GPU", ",", "with", "one", "additional", "trick", ":", "the", "GPUs", "communicate", "only", "in", "certain", "layers", ".", "This", "means", "that", ",", "for", "example", ",", "the", "kernels", "of", "layer", "3", "take", "input", "from", "all", "kernel", "maps", "in", "layer", "2", ".", "However", ",", "kernels", "in", "layer", "4", "take", "input", "only", "from", "those", "kernel", "maps", "in", "layer", "3", "which", "reside", "on", "the", "same", "GPU", ".", "Choosing", "the", "pattern", "of", "connectivity", "is", "a", "problem", "for", "cross", "-", "validation", ",", "but", "this", "allows", "us", "to", "precisely", "tune", "the", "amount", "of", "communication", "until", "it", "is", "an", "acceptable", "fraction", "of", "the", "amount", "of", "computation", ".", "The", "resultant", "architecture", "is", "somewhat", "similar", "to", "that", "of", "the", "“", "columnar", "”", "CNN", "employed", "by", "Cireşan", "et", "al", ".", "[", "5", "]", ",", "except", "that", "our", "columns", "are", "not", "independent", "(", "see", "Figure", "2", ")", ".", "This", "scheme", "reduces", "our", "top", "-", "1", "and", "top", "-", "5", "error", "rates", "by", "1.7", "%", "and", "1.2", "%", ",", "respectively", ",", "as", "compared", "with", "a", "net", "with", "half", "as", "many", "kernels", "in", "each", "convolutional", "layer", "trained", "on", "one", "GPU", ".", "The", "two", "-", "GPU", "net", "takes", "slightly", "less", "time", "to", "train", "than", "the", "one", "-", "GPU", "net2", ".", "2The", "one", "-", "GPU", "net", "actually", "has", "the", "same", "number", "of", "kernels", "as", "the", "two", "-", "GPU", "net", "in", "the", "final", "convolutional", "layer", ".", "This", "is", "because", "most", "of", "the", "net", "’s", "parameters", "are", "in", "the", "first", "fully", "-", "connected", "layer", ",", "which", "takes", "the", "last", "convolutional", "layer", "as", "input", ".", "So", "to", "make", "the", "two", "nets", "have", "approximately", "the", "same", "number", "of", "parameters", ",", "we", "did", "not", "halve", "the", "size", "of", "the", "final", "convolutional", "layer", "(", "nor", "the", "fully", "-", "conneced", "layers", "which", "follow", ")", ".", "Therefore", "this", "comparison", "is", "biased", "in", "favor", "of", "the", "one", "-", "GPU", "net", ",", "since", "it", "is", "bigger", "than", "“", "half", "the", "size", "”", "of", "the", "two", "-", "GPU", "net", ".", "3.3", "Local", "Response", "Normalization", "ReLUs", "have", "the", "desirable", "property", "that", "they", "do", "not", "require", "input", "normalization", "to", "prevent", "them", "from", "saturating", ".", "If", "at", "least", "some", "training", "examples", "produce", "a", "positive", "input", "to", "a", "ReLU", ",", "learning", "will", "happen", "in", "that", "neuron", ".", "However", ",", "we", "still", "find", "that", "the", "following", "local", "normalization", "scheme", "aids", "generalization", ".", "Denoting", "by", "aix", ",", "y", "the", "activity", "of", "a", "neuron", "computed", "by", "applying", "kernel", "i", "at", "position", "(", "x", ",", "y", ")", "and", "then", "applying", "the", "ReLU", "nonlinearity", ",", "the", "response", "-", "normalized", "activity", "bix", ",", "y", "is", "given", "by", "the", "expression", "bix", ",", "y", "=", "a", "i", "x", ",", "y", "/", "0", "@k", "+", "↵", "min", "(", "N", "1", ",", "i", "+", "n", "/", "2", ")", "X", "j", "=", "max", "(", "0", ",", "i", "n", "/", "2", ")", "(", "ajx", ",", "y", ")", "2", "1", "A", "where", "the", "sum", "runs", "over", "n", "“", "adjacent", "”", "kernel", "maps", "at", "the", "same", "spatial", "position", ",", "and", "N", "is", "the", "total", "number", "of", "kernels", "in", "the", "layer", ".", "The", "ordering", "of", "the", "kernel", "maps", "is", "of", "course", "arbitrary", "and", "determined", "before", "training", "begins", ".", "This", "sort", "of", "response", "normalization", "implements", "a", "form", "of", "lateral", "inhibition", "inspired", "by", "the", "type", "found", "in", "real", "neurons", ",", "creating", "competition", "for", "big", "activities", "amongst", "neuron", "outputs", "computed", "using", "different", "kernels", ".", "The", "constants", "k", ",", "n", ",", "↵", ",", "and", "are", "hyper", "-", "parameters", "whose", "values", "are", "determined", "using", "a", "validation", "set", ";", "we", "used", "k", "=", "2", ",", "n", "=", "5", ",", "↵", "=", "10", "4", ",", "and", "=", "0.75", ".", "We", "applied", "this", "normalization", "after", "applying", "the", "ReLU", "nonlinearity", "in", "certain", "layers", "(", "see", "Section", "3.5", ")", ".", "This", "scheme", "bears", "some", "resemblance", "to", "the", "local", "contrast", "normalization", "scheme", "of", "Jarrett", "et", "al", ".", "[", "11", "]", ",", "but", "ours", "would", "be", "more", "correctly", "termed", "“", "brightness", "normalization", "”", ",", "since", "we", "do", "not", "subtract", "the", "mean", "activity", ".", "Response", "normalization", "reduces", "our", "top", "-", "1", "and", "top", "-", "5", "error", "rates", "by", "1.4", "%", "and", "1.2", "%", ",", "respectively", ".", "We", "also", "verified", "the", "effectiveness", "of", "this", "scheme", "on", "the", "CIFAR", "-", "10", "dataset", ":", "a", "four", "-", "layer", "CNN", "achieved", "a", "13", "%", "test", "error", "rate", "without", "normalization", "and", "11", "%", "with", "normalization3", ".", "3.4", "Overlapping", "Pooling", "Pooling", "layers", "in", "CNNs", "summarize", "the", "outputs", "of", "neighboring", "groups", "of", "neurons", "in", "the", "same", "kernel", "map", ".", "Traditionally", ",", "the", "neighborhoods", "summarized", "by", "adjacent", "pooling", "units", "do", "not", "overlap", "(", "e.g.", ",", "[", "17", ",", "11", ",", "4", "]", ")", ".", "To", "be", "more", "precise", ",", "a", "pooling", "layer", "can", "be", "thought", "of", "as", "consisting", "of", "a", "grid", "of", "pooling", "units", "spaced", "s", "pixels", "apart", ",", "each", "summarizing", "a", "neighborhood", "of", "size", "z", "⇥", "z", "centered", "at", "the", "location", "of", "the", "pooling", "unit", ".", "If", "we", "set", "s", "=", "z", ",", "we", "obtain", "traditional", "local", "pooling", "as", "commonly", "employed", "in", "CNNs", ".", "If", "we", "set", "s", "<", "z", ",", "we", "obtain", "overlapping", "pooling", ".", "This", "is", "what", "we", "use", "throughout", "our", "network", ",", "with", "s", "=", "2", "and", "z", "=", "3", ".", "This", "scheme", "reduces", "the", "top", "-", "1", "and", "top", "-", "5", "error", "rates", "by", "0.4", "%", "and", "0.3", "%", ",", "respectively", ",", "as", "compared", "with", "the", "non", "-", "overlapping", "scheme", "s", "=", "2", ",", "z", "=", "2", ",", "which", "produces", "output", "of", "equivalent", "dimensions", ".", "We", "generally", "observe", "during", "training", "that", "models", "with", "overlapping", "pooling", "find", "it", "slightly", "more", "difficult", "to", "overfit", ".", "3.5", "Overall", "Architecture", "Now", "we", "are", "ready", "to", "describe", "the", "overall", "architecture", "of", "our", "CNN", ".", "As", "depicted", "in", "Figure", "2", ",", "the", "net", "contains", "eight", "layers", "with", "weights", ";", "the", "first", "five", "are", "convolutional", "and", "the", "remaining", "three", "are", "fullyconnected", ".", "The", "output", "of", "the", "last", "fully", "-", "connected", "layer", "is", "fed", "to", "a", "1000", "-", "way", "softmax", "which", "produces", "a", "distribution", "over", "the", "1000", "class", "labels", ".", "Our", "network", "maximizes", "the", "multinomial", "logistic", "regression", "objective", ",", "which", "is", "equivalent", "to", "maximizing", "the", "average", "across", "training", "cases", "of", "the", "log", "-", "probability", "of", "the", "correct", "label", "under", "the", "prediction", "distribution", ".", "The", "kernels", "of", "the", "second", ",", "fourth", ",", "and", "fifth", "convolutional", "layers", "are", "connected", "only", "to", "those", "kernel", "maps", "in", "the", "previous", "layer", "which", "reside", "on", "the", "same", "GPU", "(", "see", "Figure", "2", ")", ".", "The", "kernels", "of", "the", "third", "convolutional", "layer", "are", "connected", "to", "all", "kernel", "maps", "in", "the", "second", "layer", ".", "The", "neurons", "in", "the", "fullyconnected", "layers", "are", "connected", "to", "all", "neurons", "in", "the", "previous", "layer", ".", "Response", "-", "normalization", "layers", "follow", "the", "first", "and", "second", "convolutional", "layers", ".", "Max", "-", "pooling", "layers", ",", "of", "the", "kind", "described", "in", "Section", "3.4", ",", "follow", "both", "response", "-", "normalization", "layers", "as", "well", "as", "the", "fifth", "convolutional", "layer", ".", "The", "ReLU", "non", "-", "linearity", "is", "applied", "to", "the", "output", "of", "every", "convolutional", "and", "fully", "-", "connected", "layer", ".", "The", "first", "convolutional", "layer", "filters", "the", "224", "⇥", "224", "⇥", "3", "input", "image", "with", "96", "kernels", "of", "size", "11", "⇥", "11", "⇥", "3", "with", "a", "stride", "of", "4", "pixels", "(", "this", "is", "the", "distance", "between", "the", "receptive", "field", "centers", "of", "neighboring", "3We", "can", "not", "describe", "this", "network", "in", "detail", "due", "to", "space", "constraints", ",", "but", "it", "is", "specified", "precisely", "by", "the", "code", "and", "parameter", "files", "provided", "here", ":", "http:", "//", "code.google.com", "/", "p", "/", "cuda", "-", "convnet", "/", ".", "neurons", "in", "a", "kernel", "map", ")", ".", "The", "second", "convolutional", "layer", "takes", "as", "input", "the", "(", "response", "-", "normalized", "and", "pooled", ")", "output", "of", "the", "first", "convolutional", "layer", "and", "filters", "it", "with", "256", "kernels", "of", "size", "5", "⇥", "5", "⇥", "48", ".", "The", "third", ",", "fourth", ",", "and", "fifth", "convolutional", "layers", "are", "connected", "to", "one", "another", "without", "any", "intervening", "pooling", "or", "normalization", "layers", ".", "The", "third", "convolutional", "layer", "has", "384", "kernels", "of", "size", "3", "⇥", "3", "⇥", "256", "connected", "to", "the", "(", "normalized", ",", "pooled", ")", "outputs", "of", "the", "second", "convolutional", "layer", ".", "The", "fourth", "convolutional", "layer", "has", "384", "kernels", "of", "size", "3", "⇥", "3", "⇥", "192", ",", "and", "the", "fifth", "convolutional", "layer", "has", "256", "kernels", "of", "size", "3", "⇥", "3", "⇥", "192", ".", "The", "fully", "-", "connected", "layers", "have", "4096", "neurons", "each", ".", "4", "Reducing", "Overfitting", "Our", "neural", "network", "architecture", "has", "60", "million", "parameters", ".", "Although", "the", "1000", "classes", "of", "ILSVRC", "make", "each", "training", "example", "impose", "10", "bits", "of", "constraint", "on", "the", "mapping", "from", "image", "to", "label", ",", "this", "turns", "out", "to", "be", "insufficient", "to", "learn", "so", "many", "parameters", "without", "considerable", "overfitting", ".", "Below", ",", "we", "describe", "the", "two", "primary", "ways", "in", "which", "we", "combat", "overfitting", ".", "4.1", "Data", "Augmentation", "The", "easiest", "and", "most", "common", "method", "to", "reduce", "overfitting", "on", "image", "data", "is", "to", "artificially", "enlarge", "the", "dataset", "using", "label", "-", "preserving", "transformations", "(", "e.g.", ",", "[", "25", ",", "4", ",", "5", "]", ")", ".", "We", "employ", "two", "distinct", "forms", "of", "data", "augmentation", ",", "both", "of", "which", "allow", "transformed", "images", "to", "be", "produced", "from", "the", "original", "images", "with", "very", "little", "computation", ",", "so", "the", "transformed", "images", "do", "not", "need", "to", "be", "stored", "on", "disk", ".", "In", "our", "implementation", ",", "the", "transformed", "images", "are", "generated", "in", "Python", "code", "on", "the", "CPU", "while", "the", "GPU", "is", "training", "on", "the", "previous", "batch", "of", "images", ".", "So", "these", "data", "augmentation", "schemes", "are", ",", "in", "effect", ",", "computationally", "free", ".", "The", "first", "form", "of", "data", "augmentation", "consists", "of", "generating", "image", "translations", "and", "horizontal", "reflections", ".", "We", "do", "this", "by", "extracting", "random", "224", "⇥", "224", "patches", "(", "and", "their", "horizontal", "reflections", ")", "from", "the", "256", "⇥", "256", "images", "and", "training", "our", "network", "on", "these", "extracted", "patches4", ".", "This", "increases", "the", "size", "of", "our", "training", "set", "by", "a", "factor", "of", "2048", ",", "though", "the", "resulting", "training", "examples", "are", ",", "of", "course", ",", "highly", "interdependent", ".", "Without", "this", "scheme", ",", "our", "network", "suffers", "from", "substantial", "overfitting", ",", "which", "would", "have", "forced", "us", "to", "use", "much", "smaller", "networks", ".", "At", "test", "time", ",", "the", "network", "makes", "a", "prediction", "by", "extracting", "five", "224", "⇥", "224", "patches", "(", "the", "four", "corner", "patches", "and", "the", "center", "patch", ")", "as", "well", "as", "their", "horizontal", "reflections", "(", "hence", "ten", "patches", "in", "all", ")", ",", "and", "averaging", "the", "predictions", "made", "by", "the", "network", "’s", "softmax", "layer", "on", "the", "ten", "patches", ".", "The", "second", "form", "of", "data", "augmentation", "consists", "of", "altering", "the", "intensities", "of", "the", "RGB", "channels", "in", "training", "images", ".", "Specifically", ",", "we", "perform", "PCA", "on", "the", "set", "of", "RGB", "pixel", "values", "throughout", "the", "ImageNet", "training", "set", ".", "To", "each", "training", "image", ",", "we", "add", "multiples", "of", "the", "found", "principal", "components", ",", "4This", "is", "the", "reason", "why", "the", "input", "images", "in", "Figure", "2", "are", "224", "⇥", "224", "⇥", "3", "-", "dimensional", ".", "with", "magnitudes", "proportional", "to", "the", "corresponding", "eigenvalues", "times", "a", "random", "variable", "drawn", "from", "a", "Gaussian", "with", "mean", "zero", "and", "standard", "deviation", "0.1", ".", "Therefore", "to", "each", "RGB", "image", "pixel", "Ixy", "=", "[", "IRxy", ",", "I", "G", "xy", ",", "I", "B", "xy", "]", "T", "we", "add", "the", "following", "quantity", ":", "[", "p1", ",", "p2", ",", "p3", "]", "[", "↵", "1", "1", ",", "↵", "2", "2", ",", "↵", "3", "3", "]", "T", "where", "pi", "and", "i", "are", "ith", "eigenvector", "and", "eigenvalue", "of", "the", "3", "⇥", "3", "covariance", "matrix", "of", "RGB", "pixel", "values", ",", "respectively", ",", "and", "↵", "i", "is", "the", "aforementioned", "random", "variable", ".", "Each", "↵", "i", "is", "drawn", "only", "once", "for", "all", "the", "pixels", "of", "a", "particular", "training", "image", "until", "that", "image", "is", "used", "for", "training", "again", ",", "at", "which", "point", "it", "is", "re", "-", "drawn", ".", "This", "scheme", "approximately", "captures", "an", "important", "property", "of", "natural", "images", ",", "namely", ",", "that", "object", "identity", "is", "invariant", "to", "changes", "in", "the", "intensity", "and", "color", "of", "the", "illumination", ".", "This", "scheme", "reduces", "the", "top", "-", "1", "error", "rate", "by", "over", "1", "%", ".", "4.2", "Dropout", "Combining", "the", "predictions", "of", "many", "different", "models", "is", "a", "very", "successful", "way", "to", "reduce", "test", "errors", "[", "1", ",", "3", "]", ",", "but", "it", "appears", "to", "be", "too", "expensive", "for", "big", "neural", "networks", "that", "already", "take", "several", "days", "to", "train", ".", "There", "is", ",", "however", ",", "a", "very", "efficient", "version", "of", "model", "combination", "that", "only", "costs", "about", "a", "factor", "of", "two", "during", "training", ".", "The", "recently", "-", "introduced", "technique", ",", "called", "“", "dropout", "”", "[", "10", "]", ",", "consists", "of", "setting", "to", "zero", "the", "output", "of", "each", "hidden", "neuron", "with", "probability", "0.5", ".", "The", "neurons", "which", "are", "“", "dropped", "out", "”", "in", "this", "way", "do", "not", "contribute", "to", "the", "forward", "pass", "and", "do", "not", "participate", "in", "backpropagation", ".", "So", "every", "time", "an", "input", "is", "presented", ",", "the", "neural", "network", "samples", "a", "different", "architecture", ",", "but", "all", "these", "architectures", "share", "weights", ".", "This", "technique", "reduces", "complex", "co", "-", "adaptations", "of", "neurons", ",", "since", "a", "neuron", "can", "not", "rely", "on", "the", "presence", "of", "particular", "other", "neurons", ".", "It", "is", ",", "therefore", ",", "forced", "to", "learn", "more", "robust", "features", "that", "are", "useful", "in", "conjunction", "with", "many", "different", "random", "subsets", "of", "the", "other", "neurons", ".", "At", "test", "time", ",", "we", "use", "all", "the", "neurons", "but", "multiply", "their", "outputs", "by", "0.5", ",", "which", "is", "a", "reasonable", "approximation", "to", "taking", "the", "geometric", "mean", "of", "the", "predictive", "distributions", "produced", "by", "the", "exponentially", "-", "many", "dropout", "networks", ".", "We", "use", "dropout", "in", "the", "first", "two", "fully", "-", "connected", "layers", "of", "Figure", "2", ".", "Without", "dropout", ",", "our", "network", "exhibits", "substantial", "overfitting", ".", "Dropout", "roughly", "doubles", "the", "number", "of", "iterations", "required", "to", "converge", ".", "5", "Details", "of", "learning", "We", "trained", "our", "models", "using", "stochastic", "gradient", "descent", "with", "a", "batch", "size", "of", "128", "examples", ",", "momentum", "of", "0.9", ",", "and", "weight", "decay", "of", "0.0005", ".", "We", "found", "that", "this", "small", "amount", "of", "weight", "decay", "was", "important", "for", "the", "model", "to", "learn", ".", "In", "other", "words", ",", "weight", "decay", "here", "is", "not", "merely", "a", "regularizer", ":", "it", "reduces", "the", "model", "’s", "training", "error", ".", "The", "update", "rule", "for", "weight", "w", "was", "vi", "+", "1", ":", "=", "0.9", "·", "vi", "0.0005", "·", "✏", "·", "wi", "✏", "·", "⌧", "@L", "@w", "wi", "Di", "wi", "+", "1", ":", "=", "wi", "+", "vi", "+", "1", "where", "i", "is", "the", "iteration", "index", ",", "v", "is", "the", "momentum", "variable", ",", "✏", "is", "the", "learning", "rate", ",", "and", "D", "@L", "@w", "wi", "E", "Di", "is", "the", "average", "over", "the", "ith", "batch", "Di", "of", "the", "derivative", "of", "the", "objective", "with", "respect", "to", "w", ",", "evaluated", "at", "wi", ".", "We", "initialized", "the", "weights", "in", "each", "layer", "from", "a", "zero", "-", "mean", "Gaussian", "distribution", "with", "standard", "deviation", "0.01", ".", "We", "initialized", "the", "neuron", "biases", "in", "the", "second", ",", "fourth", ",", "and", "fifth", "convolutional", "layers", ",", "as", "well", "as", "in", "the", "fully", "-", "connected", "hidden", "layers", ",", "with", "the", "constant", "1", ".", "This", "initialization", "accelerates", "the", "early", "stages", "of", "learning", "by", "providing", "the", "ReLUs", "with", "positive", "inputs", ".", "We", "initialized", "the", "neuron", "biases", "in", "the", "remaining", "layers", "with", "the", "constant", "0", ".", "We", "used", "an", "equal", "learning", "rate", "for", "all", "layers", ",", "which", "we", "adjusted", "manually", "throughout", "training", ".", "The", "heuristic", "which", "we", "followed", "was", "to", "divide", "the", "learning", "rate", "by", "10", "when", "the", "validation", "error", "rate", "stopped", "improving", "with", "the", "current", "learning", "rate", ".", "The", "learning", "rate", "was", "initialized", "at", "0.01", "and", "reduced", "three", "times", "prior", "to", "termination", ".", "We", "trained", "the", "network", "for", "roughly", "90", "cycles", "through", "the", "training", "set", "of", "1.2", "million", "images", ",", "which", "took", "five", "to", "six", "days", "on", "two", "NVIDIA", "GTX", "580", "3", "GB", "GPUs", ".", "6", "Results", "Our", "results", "on", "ILSVRC", "-", "2010", "are", "summarized", "in", "Table", "1", ".", "Our", "network", "achieves", "top", "-", "1", "and", "top", "-", "5", "test", "set", "error", "rates", "of", "37.5", "%", "and", "17.0%5", ".", "The", "best", "performance", "achieved", "during", "the", "ILSVRC2010", "competition", "was", "47.1", "%", "and", "28.2", "%", "with", "an", "approach", "that", "averages", "the", "predictions", "produced", "from", "six", "sparse", "-", "coding", "models", "trained", "on", "different", "features", "[", "2", "]", ",", "and", "since", "then", "the", "best", "published", "results", "are", "45.7", "%", "and", "25.7", "%", "with", "an", "approach", "that", "averages", "the", "predictions", "of", "two", "classifiers", "trained", "on", "Fisher", "Vectors", "(", "FVs", ")", "computed", "from", "two", "types", "of", "densely", "-", "sampled", "features", "[", "24", "]", ".", "We", "also", "entered", "our", "model", "in", "the", "ILSVRC", "-", "2012", "competition", "and", "report", "our", "results", "in", "Table", "2", ".", "Since", "the", "ILSVRC", "-", "2012", "test", "set", "labels", "are", "not", "publicly", "available", ",", "we", "can", "not", "report", "test", "error", "rates", "for", "all", "the", "models", "that", "we", "tried", ".", "In", "the", "remainder", "of", "this", "paragraph", ",", "we", "use", "validation", "and", "test", "error", "rates", "interchangeably", "because", "in", "our", "experience", "they", "do", "not", "differ", "by", "more", "than", "0.1", "%", "(", "see", "Table", "2", ")", ".", "The", "CNN", "described", "in", "this", "paper", "achieves", "a", "top", "-", "5", "error", "rate", "of", "18.2", "%", ".", "Averaging", "the", "predictions", "of", "five", "similar", "CNNs", "gives", "an", "error", "rate", "of", "16.4", "%", ".", "Training", "one", "CNN", ",", "with", "an", "extra", "sixth", "convolutional", "layer", "over", "the", "last", "pooling", "layer", ",", "to", "classify", "the", "entire", "ImageNet", "Fall", "2011", "release", "(", "15", "M", "images", ",", "22", "K", "categories", ")", ",", "and", "then", "“", "fine", "-", "tuning", "”", "it", "on", "ILSVRC", "-", "2012", "gives", "an", "error", "rate", "of", "16.6", "%", ".", "Averaging", "the", "predictions", "of", "two", "CNNs", "that", "were", "pre", "-", "trained", "on", "the", "entire", "Fall", "2011", "release", "with", "the", "aforementioned", "five", "CNNs", "gives", "an", "error", "rate", "of", "15.3", "%", ".", "The", "second", "-", "best", "contest", "entry", "achieved", "an", "error", "rate", "of", "26.2", "%", "with", "an", "approach", "that", "averages", "the", "predictions", "of", "several", "classifiers", "trained", "on", "FVs", "computed", "from", "different", "types", "of", "densely", "-", "sampled", "features", "[", "7", "]", ".", "Finally", ",", "we", "also", "report", "our", "error", "rates", "on", "the", "Fall", "2009", "version", "of", "ImageNet", "with", "10", ",", "184", "categories", "and", "8.9", "million", "images", ".", "On", "this", "dataset", "we", "follow", "the", "convention", "in", "the", "literature", "of", "using", "half", "of", "the", "images", "for", "training", "and", "half", "for", "testing", ".", "Since", "there", "is", "no", "established", "test", "set", ",", "our", "split", "necessarily", "differs", "from", "the", "splits", "used", "by", "previous", "authors", ",", "but", "this", "does", "not", "affect", "the", "results", "appreciably", ".", "Our", "top", "-", "1", "and", "top", "-", "5", "error", "rates", "on", "this", "dataset", "are", "67.4", "%", "and", "40.9", "%", ",", "attained", "by", "the", "net", "described", "above", "but", "with", "an", "additional", ",", "sixth", "convolutional", "layer", "over", "the", "last", "pooling", "layer", ".", "The", "best", "published", "results", "on", "this", "dataset", "are", "78.1", "%", "and", "60.9", "%", "[", "19", "]", ".", "6.1", "Qualitative", "Evaluations", "Figure", "3", "shows", "the", "convolutional", "kernels", "learned", "by", "the", "network", "’s", "two", "data", "-", "connected", "layers", ".", "The", "network", "has", "learned", "a", "variety", "of", "frequency", "-", "and", "orientation", "-", "selective", "kernels", ",", "as", "well", "as", "various", "colored", "blobs", ".", "Notice", "the", "specialization", "exhibited", "by", "the", "two", "GPUs", ",", "a", "result", "of", "the", "restricted", "connectivity", "described", "in", "Section", "3.5", ".", "The", "kernels", "on", "GPU", "1", "are", "largely", "color", "-", "agnostic", ",", "while", "the", "kernels", "on", "on", "GPU", "2", "are", "largely", "color", "-", "specific", ".", "This", "kind", "of", "specialization", "occurs", "during", "every", "run", "and", "is", "independent", "of", "any", "particular", "random", "weight", "initialization", "(", "modulo", "a", "renumbering", "of", "the", "GPUs", ")", ".", "5The", "error", "rates", "without", "averaging", "predictions", "over", "ten", "patches", "as", "described", "in", "Section", "4.1", "are", "39.0", "%", "and", "18.3", "%", ".", "In", "the", "left", "panel", "of", "Figure", "4", "we", "qualitatively", "assess", "what", "the", "network", "has", "learned", "by", "computing", "its", "top", "-", "5", "predictions", "on", "eight", "test", "images", ".", "Notice", "that", "even", "off", "-", "center", "objects", ",", "such", "as", "the", "mite", "in", "the", "top", "-", "left", ",", "can", "be", "recognized", "by", "the", "net", ".", "Most", "of", "the", "top", "-", "5", "labels", "appear", "reasonable", ".", "For", "example", ",", "only", "other", "types", "of", "cat", "are", "considered", "plausible", "labels", "for", "the", "leopard", ".", "In", "some", "cases", "(", "grille", ",", "cherry", ")", "there", "is", "genuine", "ambiguity", "about", "the", "intended", "focus", "of", "the", "photograph", ".", "Another", "way", "to", "probe", "the", "network", "’s", "visual", "knowledge", "is", "to", "consider", "the", "feature", "activations", "induced", "by", "an", "image", "at", "the", "last", ",", "4096", "-", "dimensional", "hidden", "layer", ".", "If", "two", "images", "produce", "feature", "activation", "vectors", "with", "a", "small", "Euclidean", "separation", ",", "we", "can", "say", "that", "the", "higher", "levels", "of", "the", "neural", "network", "consider", "them", "to", "be", "similar", ".", "Figure", "4", "shows", "five", "images", "from", "the", "test", "set", "and", "the", "six", "images", "from", "the", "training", "set", "that", "are", "most", "similar", "to", "each", "of", "them", "according", "to", "this", "measure", ".", "Notice", "that", "at", "the", "pixel", "level", ",", "the", "retrieved", "training", "images", "are", "generally", "not", "close", "in", "L2", "to", "the", "query", "images", "in", "the", "first", "column", ".", "For", "example", ",", "the", "retrieved", "dogs", "and", "elephants", "appear", "in", "a", "variety", "of", "poses", ".", "We", "present", "the", "results", "for", "many", "more", "test", "images", "in", "the", "supplementary", "material", ".", "Computing", "similarity", "by", "using", "Euclidean", "distance", "between", "two", "4096", "-", "dimensional", ",", "real", "-", "valued", "vectors", "is", "inefficient", ",", "but", "it", "could", "be", "made", "efficient", "by", "training", "an", "auto", "-", "encoder", "to", "compress", "these", "vectors", "to", "short", "binary", "codes", ".", "This", "should", "produce", "a", "much", "better", "image", "retrieval", "method", "than", "applying", "autoencoders", "to", "the", "raw", "pixels", "[", "14", "]", ",", "which", "does", "not", "make", "use", "of", "image", "labels", "and", "hence", "has", "a", "tendency", "to", "retrieve", "images", "with", "similar", "patterns", "of", "edges", ",", "whether", "or", "not", "they", "are", "semantically", "similar", ".", "7", "Discussion", "Our", "results", "show", "that", "a", "large", ",", "deep", "convolutional", "neural", "network", "is", "capable", "of", "achieving", "recordbreaking", "results", "on", "a", "highly", "challenging", "dataset", "using", "purely", "supervised", "learning", ".", "It", "is", "notable", "that", "our", "network", "’s", "performance", "degrades", "if", "a", "single", "convolutional", "layer", "is", "removed", ".", "For", "example", ",", "removing", "any", "of", "the", "middle", "layers", "results", "in", "a", "loss", "of", "about", "2", "%", "for", "the", "top", "-", "1", "performance", "of", "the", "network", ".", "So", "the", "depth", "really", "is", "important", "for", "achieving", "our", "results", ".", "To", "simplify", "our", "experiments", ",", "we", "did", "not", "use", "any", "unsupervised", "pre", "-", "training", "even", "though", "we", "expect", "that", "it", "will", "help", ",", "especially", "if", "we", "obtain", "enough", "computational", "power", "to", "significantly", "increase", "the", "size", "of", "the", "network", "without", "obtaining", "a", "corresponding", "increase", "in", "the", "amount", "of", "labeled", "data", ".", "Thus", "far", ",", "our", "results", "have", "improved", "as", "we", "have", "made", "our", "network", "larger", "and", "trained", "it", "longer", "but", "we", "still", "have", "many", "orders", "of", "magnitude", "to", "go", "in", "order", "to", "match", "the", "infero", "-", "temporal", "pathway", "of", "the", "human", "visual", "system", ".", "Ultimately", "we", "would", "like", "to", "use", "very", "large", "and", "deep", "convolutional", "nets", "on", "video", "sequences", "where", "the", "temporal", "structure", "provides", "very", "helpful", "information", "that", "is", "missing", "or", "far", "less", "obvious", "in", "static", "images", "."], "entity": [[0, 2, ["ImageNet", "Classification"], "Task"], [3, 7, ["Deep", "Convolutional", "Neural", "Networks"], "Method"], [12, 16, ["deep", "convolutional", "neural", "network"], "Method"], [27, 32, ["ImageNet", "LSVRC", "-", "2010", "contest"], "Material"], [45, 48, ["top", "-", "1"], "Metric"], [49, 54, ["top", "-", "5", "error", "rates"], "Metric"], [79, 81, ["neural", "network"], "Method"], [96, 98, ["convolutional", "layers"], "Method"], [105, 109, ["max", "-", "pooling", "layers"], "Method"], [112, 115, ["fully", "connected", "layers"], "Method"], [118, 122, ["1000", "-", "way", "softmax"], "Method"], [125, 126, ["training"], "Task"], [130, 134, ["non", "-", "saturating", "neurons"], "Method"], [138, 140, ["GPU", "implementation"], "Method"], [142, 144, ["convolution", "operation"], "Method"], [158, 160, ["regularization", "method"], "Method"], [162, 163, ["dropout"], "Method"], [189, 195, ["top", "-", "5", "test", "error", "rate"], "Metric"], [216, 220, ["deep", "convolutional", "neural", "network"], "Method"], [231, 236, ["ImageNet", "LSVRC", "-", "2010", "contest"], "Material"], [249, 252, ["top", "-", "1"], "Metric"], [253, 258, ["top", "-", "5", "error", "rates"], "Metric"], [280, 282, ["neural", "network"], "Method"], [297, 299, ["convolutional", "layers"], "Method"], [306, 310, ["max", "-", "pooling", "layers"], "Method"], [313, 317, ["fully", "-", "connected", "layers"], "Method"], [320, 324, ["1000", "-", "way", "softmax"], "Method"], [327, 328, ["training"], "Task"], [332, 336, ["non", "-", "saturating", "neurons"], "Method"], [340, 342, ["GPU", "implementation"], "Method"], [344, 346, ["convolution", "operation"], "Method"], [362, 364, ["regularization", "method"], "Method"], [366, 367, ["dropout"], "Method"], [393, 399, ["top", "-", "5", "test", "error", "rate"], "Metric"], [420, 422, ["object", "recognition"], "Task"], [426, 429, ["machine", "learning", "methods"], "Method"], [452, 453, ["overfitting"], "Task"], [477, 478, ["NORB"], "Material"], [482, 485, ["Caltech", "-", "101"], "Material"], [494, 497, ["CIFAR", "-", "10"], "Material"], [505, 507, ["recognition", "tasks"], "Task"], [534, 536, ["error", "rate"], "Metric"], [538, 543, ["MNIST", "digit", "-", "recognition", "task"], "Task"], [626, 627, ["LabelMe"], "Material"], [644, 645, ["ImageNet"], "Material"], [685, 687, ["learning", "capacity"], "Metric"], [695, 698, ["object", "recognition", "task"], "Task"], [713, 714, ["ImageNet"], "Material"], [736, 739, ["Convolutional", "neural", "networks"], "Method"], [740, 741, ["CNNs"], "Method"], [808, 811, ["feedforward", "neural", "networks"], "Method"], [812, 816, ["similarly", "-", "sized", "layers"], "Method"], [817, 818, ["CNNs"], "Method"], [851, 852, ["CNNs"], "Method"], [860, 862, ["local", "architecture"], "Method"], [875, 879, ["high", "-", "resolution", "images"], "Task"], [883, 884, ["GPUs"], "Method"], [888, 895, ["highly", "-", "optimized", "implementation", "of", "2D", "convolution"], "Method"], [914, 915, ["ImageNet"], "Material"], [943, 946, ["convolutional", "neural", "networks"], "Method"], [952, 953, ["ImageNet"], "Material"], [956, 959, ["ILSVRC", "-", "2010"], "Material"], [986, 991, ["GPU", "implementation", "of", "2D", "convolution"], "Method"], [999, 1002, ["convolutional", "neural", "networks"], "Method"], [1026, 1028, ["training", "time"], "Metric"], [1063, 1064, ["overfitting"], "Task"], [1080, 1084, ["fully", "-", "connected", "layers"], "Method"], [1098, 1100, ["convolutional", "layer"], "Method"], [1140, 1141, ["GPUs"], "Method"], [1146, 1148, ["training", "time"], "Metric"], [1189, 1190, ["GPUs"], "Method"], [1199, 1201, ["Dataset", "ImageNet"], "Material"], [1231, 1233, ["human", "labelers"], "Method"], [1234, 1242, ["Amazon", "’s", "Mechanical", "Turk", "crowd", "-", "sourcing", "tool"], "Method"], [1251, 1255, ["Pascal", "Visual", "Object", "Challenge"], "Task"], [1261, 1268, ["ImageNet", "Large", "-", "Scale", "Visual", "Recognition", "Challenge"], "Task"], [1269, 1270, ["ILSVRC"], "Material"], [1275, 1276, ["ILSVRC"], "Material"], [1280, 1281, ["ImageNet"], "Material"], [1315, 1318, ["ILSVRC", "-", "2010"], "Material"], [1323, 1324, ["ILSVRC"], "Material"], [1385, 1386, ["ImageNet"], "Material"], [1393, 1395, ["error", "rates"], "Metric"], [1396, 1399, ["top", "-", "1"], "Metric"], [1400, 1403, ["top", "-", "5"], "Metric"], [1406, 1411, ["top", "-", "5", "error", "rate"], "Metric"], [1435, 1436, ["ImageNet"], "Material"], [1569, 1570, ["convolutional"], "Method"], [1572, 1575, ["fully", "-", "connected"], "Method"], [1589, 1592, ["network", "’s", "architecture"], "Method"], [1671, 1673, ["training", "time"], "Metric"], [1674, 1676, ["gradient", "descent"], "Method"], [1678, 1680, ["saturating", "nonlinearities"], "Method"], [1717, 1720, ["Rectified", "Linear", "Units"], "Method"], [1721, 1722, ["ReLUs"], "Method"], [1724, 1728, ["Deep", "convolutional", "neural", "networks"], "Method"], [1729, 1730, ["ReLUs"], "Method"], [1738, 1740, ["tanh", "units"], "Method"], [1759, 1761, ["training", "error"], "Metric"], [1763, 1767, ["CIFAR", "-", "10", "dataset"], "Material"], [1770, 1775, ["four", "-", "layer", "convolutional", "network"], "Method"], [1791, 1793, ["neural", "networks"], "Method"], [1801, 1804, ["saturating", "neuron", "models"], "Method"], [1815, 1817, ["neuron", "models"], "Method"], [1818, 1819, ["CNNs"], "Method"], [1851, 1853, ["contrast", "normalization"], "Method"], [1855, 1858, ["local", "average", "pooling"], "Method"], [1875, 1876, ["overfitting"], "Task"], [1899, 1900, ["ReLUs"], "Method"], [1901, 1903, ["Faster", "learning"], "Method"], [1911, 1913, ["large", "models"], "Method"], [1919, 1920, ["Training"], "Task"], [1927, 1928, ["GPU"], "Method"], [1983, 1984, ["GPUs"], "Method"], [1990, 1994, ["cross", "-", "GPU", "parallelization"], "Task"], [2019, 2021, ["parallelization", "scheme"], "Method"], [2044, 2045, ["GPUs"], "Method"], [2067, 2069, ["kernel", "maps"], "Method"], [2084, 2086, ["kernel", "maps"], "Method"], [2105, 2108, ["cross", "-", "validation"], "Task"], [2203, 2205, ["convolutional", "layer"], "Method"], [2213, 2215, ["GPU", "net"], "Method"], [2223, 2227, ["one", "-", "GPU", "net2"], "Method"], [2229, 2233, ["one", "-", "GPU", "net"], "Method"], [2244, 2246, ["GPU", "net"], "Method"], [2249, 2251, ["convolutional", "layer"], "Method"], [2265, 2269, ["fully", "-", "connected", "layer"], "Method"], [2274, 2276, ["convolutional", "layer"], "Method"], [2302, 2304, ["convolutional", "layer"], "Method"], [2307, 2311, ["fully", "-", "conneced", "layers"], "Method"], [2326, 2328, ["GPU", "net"], "Method"], [2343, 2345, ["GPU", "net"], "Method"], [2347, 2351, ["Local", "Response", "Normalization", "ReLUs"], "Method"], [2397, 2400, ["local", "normalization", "scheme"], "Method"], [2401, 2402, ["generalization"], "Task"], [2416, 2418, ["kernel", "i"], "Method"], [2429, 2431, ["ReLU", "nonlinearity"], "Method"], [2523, 2525, ["kernel", "maps"], "Method"], [2538, 2540, ["response", "normalization"], "Method"], [2616, 2618, ["ReLU", "nonlinearity"], "Method"], [2634, 2638, ["local", "contrast", "normalization", "scheme"], "Method"], [2655, 2657, ["brightness", "normalization"], "Method"], [2668, 2670, ["Response", "normalization"], "Method"], [2700, 2704, ["CIFAR", "-", "10", "dataset"], "Material"], [2714, 2717, ["test", "error", "rate"], "Metric"], [2718, 2719, ["normalization"], "Metric"], [2723, 2724, ["normalization3"], "Task"], [2726, 2728, ["Overlapping", "Pooling"], "Method"], [2728, 2730, ["Pooling", "layers"], "Method"], [2731, 2732, ["CNNs"], "Method"], [2776, 2778, ["pooling", "layer"], "Method"], [2788, 2790, ["pooling", "units"], "Method"], [2823, 2825, ["local", "pooling"], "Method"], [2829, 2830, ["CNNs"], "Method"], [2887, 2891, ["non", "-", "overlapping", "scheme"], "Method"], [2938, 2939, ["CNN"], "Method"], [2958, 2959, ["convolutional"], "Method"], [2971, 2975, ["fully", "-", "connected", "layer"], "Method"], [2979, 2983, ["1000", "-", "way", "softmax"], "Method"], [2997, 3001, ["multinomial", "logistic", "regression", "objective"], "Task"], [3036, 3038, ["convolutional", "layers"], "Method"], [3066, 3068, ["convolutional", "layer"], "Method"], [3072, 3074, ["kernel", "maps"], "Method"], [3095, 3099, ["Response", "-", "normalization", "layers"], "Method"], [3104, 3106, ["convolutional", "layers"], "Method"], [3107, 3111, ["Max", "-", "pooling", "layers"], "Method"], [3122, 3126, ["response", "-", "normalization", "layers"], "Method"], [3130, 3133, ["fifth", "convolutional", "layer"], "Method"], [3146, 3152, ["convolutional", "and", "fully", "-", "connected", "layer"], "Method"], [3155, 3157, ["convolutional", "layer"], "Method"], [3235, 3237, ["kernel", "map"], "Method"], [3241, 3243, ["convolutional", "layer"], "Method"], [3258, 3260, ["convolutional", "layer"], "Method"], [3281, 3283, ["convolutional", "layers"], "Method"], [3290, 3295, ["intervening", "pooling", "or", "normalization", "layers"], "Method"], [3298, 3300, ["convolutional", "layer"], "Method"], [3322, 3324, ["convolutional", "layer"], "Method"], [3327, 3329, ["convolutional", "layer"], "Method"], [3343, 3345, ["convolutional", "layer"], "Method"], [3357, 3361, ["fully", "-", "connected", "layers"], "Method"], [3368, 3369, ["Overfitting"], "Task"], [3370, 3373, ["neural", "network", "architecture"], "Method"], [3383, 3384, ["ILSVRC"], "Material"], [3431, 3433, ["Data", "Augmentation"], "Task"], [3441, 3442, ["overfitting"], "Task"], [3474, 3476, ["data", "augmentation"], "Task"], [3525, 3526, ["GPU"], "Method"], [3537, 3540, ["data", "augmentation", "schemes"], "Method"], [3552, 3554, ["data", "augmentation"], "Task"], [3556, 3559, ["generating", "image", "translations"], "Task"], [3692, 3694, ["softmax", "layer"], "Method"], [3703, 3705, ["data", "augmentation"], "Task"], [3722, 3723, ["PCA"], "Method"], [3727, 3730, ["RGB", "pixel", "values"], "Material"], [3732, 3735, ["ImageNet", "training", "set"], "Material"], [3796, 3798, ["RGB", "image"], "Material"], [3981, 3984, ["big", "neural", "networks"], "Method"], [4002, 4004, ["model", "combination"], "Method"], [4023, 4025, ["dropout", "”"], "Method"], [4067, 4068, ["backpropagation"], "Method"], [4078, 4080, ["neural", "network"], "Method"], [4175, 4180, ["exponentially", "-", "many", "dropout", "networks"], "Method"], [4183, 4184, ["dropout"], "Method"], [4197, 4198, ["dropout"], "Method"], [4205, 4206, ["Dropout"], "Method"], [4219, 4220, ["learning"], "Task"], [4225, 4228, ["stochastic", "gradient", "descent"], "Method"], [4267, 4269, ["weight", "decay"], "Method"], [4274, 4275, ["regularizer"], "Method"], [4281, 4283, ["training", "error"], "Metric"], [4285, 4287, ["update", "rule"], "Method"], [4337, 4339, ["learning", "rate"], "Metric"], [4379, 4384, ["zero", "-", "mean", "Gaussian", "distribution"], "Method"], [4428, 4429, ["learning"], "Task"], [4477, 4479, ["learning", "rate"], "Metric"], [4483, 4486, ["validation", "error", "rate"], "Metric"], [4491, 4493, ["learning", "rate"], "Metric"], [4495, 4497, ["learning", "rate"], "Metric"], [4546, 4549, ["ILSVRC", "-", "2010"], "Material"], [4581, 4583, ["ILSVRC2010", "competition"], "Metric"], [4599, 4603, ["sparse", "-", "coding", "models"], "Method"], [4633, 4634, ["classifiers"], "Method"], [4639, 4640, ["FVs"], "Method"], [4690, 4693, ["test", "error", "rates"], "Metric"], [4710, 4715, ["validation", "and", "test", "error", "rates"], "Metric"], [4736, 4737, ["CNN"], "Method"], [4743, 4748, ["top", "-", "5", "error", "rate"], "Metric"], [4758, 4759, ["CNNs"], "Method"], [4761, 4763, ["error", "rate"], "Metric"], [4769, 4770, ["CNN"], "Method"], [4774, 4777, ["sixth", "convolutional", "layer"], "Method"], [4780, 4782, ["pooling", "layer"], "Method"], [4787, 4791, ["ImageNet", "Fall", "2011", "release"], "Material"], [4810, 4813, ["ILSVRC", "-", "2012"], "Material"], [4815, 4817, ["error", "rate"], "Metric"], [4826, 4827, ["CNNs"], "Method"], [4842, 4843, ["CNNs"], "Method"], [4845, 4847, ["error", "rate"], "Metric"], [4859, 4861, ["error", "rate"], "Metric"], [4873, 4874, ["classifiers"], "Method"], [4876, 4877, ["FVs"], "Method"], [4896, 4898, ["error", "rates"], "Metric"], [4904, 4905, ["ImageNet"], "Material"], [4999, 5001, ["convolutional", "layer"], "Method"], [5004, 5006, ["pooling", "layer"], "Method"], [5031, 5033, ["convolutional", "kernels"], "Method"], [5039, 5043, ["data", "-", "connected", "layers"], "Method"], [5073, 5074, ["GPUs"], "Method"], [5089, 5091, ["GPU", "1"], "Method"], [5124, 5127, ["random", "weight", "initialization"], "Method"], [5137, 5139, ["error", "rates"], "Metric"], [5306, 5308, ["neural", "network"], "Method"], [5399, 5401, ["Computing", "similarity"], "Task"], [5427, 5430, ["auto", "-", "encoder"], "Method"], [5435, 5438, ["short", "binary", "codes"], "Method"], [5445, 5448, ["image", "retrieval", "method"], "Method"], [5450, 5451, ["autoencoders"], "Method"], [5498, 5502, ["deep", "convolutional", "neural", "network"], "Method"], [5515, 5517, ["supervised", "learning"], "Method"], [5530, 5532, ["convolutional", "layer"], "Method"], [5583, 5587, ["unsupervised", "pre", "-", "training"], "Task"], [5662, 5665, ["human", "visual", "system"], "Method"], [5675, 5678, ["deep", "convolutional", "nets"], "Method"], [181, 182, ["ILSVRC"], "Material"], [385, 386, ["ILSVRC"], "Material"], [907, 908, ["CNNs"], "Method"], [960, 961, ["ILSVRC"], "Material"], [1171, 1172, ["GPUs"], "Method"], [1355, 1356, ["ILSVRC"], "Material"], [1860, 1863, ["Caltech", "-", "101"], "Material"], [1922, 1923, ["GPUs"], "Method"], [1980, 1981, ["GPUs"], "Method"], [2145, 2146, ["CNN"], "Method"], [2173, 2176, ["top", "-", "1"], "Metric"], [2177, 2182, ["top", "-", "5", "error", "rates"], "Metric"], [2672, 2675, ["top", "-", "1"], "Metric"], [2676, 2681, ["top", "-", "5", "error", "rates"], "Metric"], [2709, 2710, ["CNN"], "Method"], [2865, 2868, ["top", "-", "1"], "Metric"], [2869, 2874, ["top", "-", "5", "error", "rates"], "Metric"], [3939, 3942, ["top", "-", "1"], "Metric"], [4539, 4540, ["GPUs"], "Method"], [4558, 4561, ["top", "-", "1"], "Metric"], [4661, 4662, ["ILSVRC"], "Material"], [4675, 4676, ["ILSVRC"], "Material"], [4767, 4768, ["Training"], "Task"], [4968, 4971, ["top", "-", "1"], "Metric"], [4972, 4977, ["top", "-", "5", "error", "rates"], "Metric"], [5133, 5134, ["GPUs"], "Method"], [5554, 5557, ["top", "-", "1"], "Metric"]]} {"doc_id": "2347efd31b723a636902ccaa7b5bc351788e793f", "text": ["document", ":", "Attending", "to", "Characters", "in", "Neural", "Sequence", "Labeling", "Models", "Sequence", "labeling", "architectures", "use", "word", "embeddings", "for", "capturing", "similarity", ",", "but", "suffer", "when", "handling", "previously", "unseen", "or", "rare", "words", ".", "We", "investigate", "character", "-", "level", "extensions", "to", "such", "models", "and", "propose", "a", "novel", "architecture", "for", "combining", "alternative", "word", "representations", ".", "By", "using", "an", "attention", "mechanism", ",", "the", "model", "is", "able", "to", "dynamically", "decide", "how", "much", "information", "to", "use", "from", "a", "word", "-", "or", "character", "-", "level", "component", ".", "We", "evaluated", "different", "architectures", "on", "a", "range", "of", "sequence", "labeling", "datasets", ",", "and", "character", "-", "level", "extensions", "were", "found", "to", "improve", "performance", "on", "every", "benchmark", ".", "In", "addition", ",", "the", "proposed", "attention", "-", "based", "architecture", "delivered", "the", "best", "results", "even", "with", "a", "smaller", "number", "of", "trainable", "parameters", ".", "section", ":", "Introduction", "This", "work", "is", "licenced", "under", "a", "Creative", "Commons", "Attribution", "4.0", "International", "Licence", ".", "Licence", "details", ":", "Many", "NLP", "tasks", ",", "including", "named", "entity", "recognition", "(", "NER", ")", ",", "part", "-", "of", "-", "speech", "(", "POS", ")", "tagging", "and", "shallow", "parsing", "can", "be", "framed", "as", "types", "of", "sequence", "labeling", ".", "The", "development", "of", "accurate", "and", "efficient", "sequence", "labeling", "models", "is", "thereby", "useful", "for", "a", "wide", "range", "of", "downstream", "applications", ".", "Work", "in", "this", "area", "has", "traditionally", "involved", "task", "-", "specific", "feature", "engineering", "–", "for", "example", ",", "integrating", "gazetteers", "for", "named", "entity", "recognition", ",", "or", "using", "features", "from", "a", "morphological", "analyser", "in", "POS", "-", "tagging", ".", "Recent", "developments", "in", "neural", "architectures", "and", "representation", "learning", "have", "opened", "the", "door", "to", "models", "that", "can", "discover", "useful", "features", "automatically", "from", "the", "data", ".", "Such", "sequence", "labeling", "systems", "are", "applicable", "to", "many", "tasks", ",", "using", "only", "the", "surface", "text", "as", "input", ",", "yet", "are", "able", "to", "achieve", "competitive", "results", ".", "Current", "neural", "models", "generally", "make", "use", "of", "word", "embeddings", ",", "which", "allow", "them", "to", "learn", "similar", "representations", "for", "semantically", "or", "functionally", "similar", "words", ".", "While", "this", "is", "an", "important", "improvement", "over", "count", "-", "based", "models", ",", "they", "still", "have", "weaknesses", "that", "should", "be", "addressed", ".", "The", "most", "obvious", "problem", "arises", "when", "dealing", "with", "out", "-", "of", "-", "vocabulary", "(", "OOV", ")", "words", "–", "if", "a", "token", "has", "never", "been", "seen", "before", ",", "then", "it", "does", "not", "have", "an", "embedding", "and", "the", "model", "needs", "to", "back", "-", "off", "to", "a", "generic", "OOV", "representation", ".", "Words", "that", "have", "been", "seen", "very", "infrequently", "have", "embeddings", ",", "but", "they", "will", "likely", "have", "low", "quality", "due", "to", "lack", "of", "training", "data", ".", "The", "approach", "can", "also", "be", "sub", "-", "optimal", "in", "terms", "of", "parameter", "usage", "–", "for", "example", ",", "certain", "suffixes", "indicate", "more", "likely", "POS", "tags", "for", "these", "words", ",", "but", "this", "information", "gets", "encoded", "into", "each", "individual", "embedding", "as", "opposed", "to", "being", "shared", "between", "the", "whole", "vocabulary", ".", "In", "this", "paper", ",", "we", "construct", "a", "task", "-", "independent", "neural", "network", "architecture", "for", "sequence", "labeling", ",", "and", "then", "extend", "it", "with", "two", "different", "approaches", "for", "integrating", "character", "-", "level", "information", ".", "By", "operating", "on", "individual", "characters", ",", "the", "model", "is", "able", "to", "infer", "representations", "for", "previously", "unseen", "words", "and", "share", "information", "about", "morpheme", "-", "level", "regularities", ".", "We", "propose", "a", "novel", "architecture", "for", "combining", "character", "-", "level", "representations", "with", "word", "embeddings", "using", "a", "gating", "mechanism", ",", "also", "referred", "to", "as", "attention", ",", "which", "allows", "the", "model", "to", "dynamically", "decide", "which", "source", "of", "information", "to", "use", "for", "each", "word", ".", "In", "addition", ",", "we", "describe", "a", "new", "objective", "for", "model", "training", "where", "the", "character", "-", "level", "representations", "are", "optimised", "to", "mimic", "the", "current", "state", "of", "word", "embeddings", ".", "We", "evaluate", "the", "neural", "models", "on", "8", "datasets", "from", "the", "fields", "of", "NER", ",", "POS", "-", "tagging", ",", "chunking", "and", "error", "detection", "in", "learner", "texts", ".", "Our", "experiments", "show", "that", "including", "a", "character", "-", "based", "component", "in", "the", "sequence", "labeling", "model", "provides", "substantial", "performance", "improvements", "on", "all", "the", "benchmarks", ".", "In", "addition", ",", "the", "attention", "-", "based", "architecture", "achieves", "the", "best", "results", "on", "all", "evaluations", ",", "while", "requiring", "a", "smaller", "number", "of", "parameters", ".", "section", ":", "Bidirectional", "LSTM", "for", "sequence", "labeling", "We", "first", "describe", "a", "basic", "word", "-", "level", "neural", "network", "for", "sequence", "labeling", ",", "following", "the", "models", "described", "by", "Lample2016", "and", "Rei2016", ",", "and", "then", "propose", "two", "alternative", "methods", "for", "incorporating", "character", "-", "level", "information", ".", "Figure", "[", "reference", "]", "shows", "the", "general", "architecture", "of", "the", "sequence", "labeling", "network", ".", "The", "model", "receives", "a", "sequence", "of", "tokens", "as", "input", ",", "and", "predicts", "a", "label", "corresponding", "to", "each", "of", "the", "input", "tokens", ".", "The", "tokens", "are", "first", "mapped", "to", "a", "distributed", "vector", "space", ",", "resulting", "in", "a", "sequence", "of", "word", "embeddings", ".", "Next", ",", "the", "embeddings", "are", "given", "as", "input", "to", "two", "LSTM", "components", "moving", "in", "opposite", "directions", "through", "the", "text", ",", "creating", "context", "-", "specific", "representations", ".", "The", "respective", "forward", "-", "and", "backward", "-", "conditioned", "representations", "are", "concatenated", "for", "each", "word", "position", ",", "resulting", "in", "representations", "that", "are", "conditioned", "on", "the", "whole", "sequence", ":", "We", "include", "an", "extra", "narrow", "hidden", "layer", "on", "top", "of", "the", "LSTM", ",", "which", "proved", "to", "be", "a", "useful", "modification", "based", "on", "development", "experiments", ".", "An", "additional", "hidden", "layer", "allows", "the", "model", "to", "detect", "higher", "-", "level", "feature", "combinations", ",", "while", "constraining", "it", "to", "be", "small", "forces", "it", "to", "focus", "on", "more", "generalisable", "patterns", ":", "where", "is", "a", "weight", "matrix", "between", "the", "layers", ",", "and", "the", "size", "of", "is", "intentionally", "kept", "small", ".", "Finally", ",", "to", "produce", "label", "predictions", ",", "we", "use", "either", "a", "softmax", "layer", "or", "a", "conditional", "random", "field", "(", "CRF", ",", "Lafferty2001", ")", ".", "The", "softmax", "calculates", "a", "normalised", "probability", "distribution", "over", "all", "the", "possible", "labels", "for", "each", "word", ":", "where", "is", "the", "probability", "of", "the", "label", "of", "the", "-", "th", "word", "(", ")", "being", ",", "is", "the", "set", "of", "all", "possible", "labels", ",", "and", "is", "the", "-", "th", "row", "of", "output", "weight", "matrix", ".", "To", "optimise", "this", "model", ",", "we", "minimise", "categorical", "crossentropy", ",", "which", "is", "equivalent", "to", "minimising", "the", "negative", "log", "-", "probability", "of", "the", "correct", "labels", ":", "Following", "Huang2015", ",", "we", "can", "also", "use", "a", "CRF", "as", "the", "output", "layer", ",", "which", "conditions", "each", "prediction", "on", "the", "previously", "predicted", "label", ".", "In", "this", "architecture", ",", "the", "last", "hidden", "layer", "is", "used", "to", "predict", "confidence", "scores", "for", "the", "word", "having", "each", "of", "the", "possible", "labels", ".", "A", "separate", "weight", "matrix", "is", "used", "to", "learn", "transition", "probabilities", "between", "different", "labels", ",", "and", "the", "Viterbi", "algorithm", "is", "used", "to", "find", "an", "optimal", "sequence", "of", "weights", ".", "Given", "that", "is", "a", "sequence", "of", "labels", ",", "then", "the", "CRF", "score", "for", "this", "sequence", "can", "be", "calculated", "as", ":", "where", "shows", "how", "confident", "the", "network", "is", "that", "the", "label", "on", "the", "-", "th", "word", "is", ".", "shows", "the", "likelihood", "of", "transitioning", "from", "label", "to", "label", ",", "and", "these", "values", "are", "optimised", "during", "training", ".", "The", "output", "from", "the", "model", "is", "the", "sequence", "of", "labels", "with", "the", "largest", "score", ",", "which", "can", "be", "found", "efficiently", "using", "the", "Viterbi", "algorithm", ".", "In", "order", "to", "optimise", "the", "CRF", "model", ",", "the", "loss", "function", "maximises", "the", "score", "for", "the", "correct", "label", "sequence", ",", "while", "minimising", "the", "scores", "for", "all", "other", "sequences", ":", "where", "is", "the", "set", "of", "all", "possible", "label", "sequences", ".", "section", ":", "Character", "-", "level", "sequence", "labeling", "Distributed", "embeddings", "map", "words", "into", "a", "space", "where", "semantically", "similar", "words", "have", "similar", "vector", "representations", ",", "allowing", "the", "models", "to", "generalise", "better", ".", "However", ",", "they", "still", "treat", "words", "as", "atomic", "units", "and", "ignore", "any", "surface", "-", "or", "morphological", "similarities", "between", "different", "words", ".", "By", "constructing", "models", "that", "operate", "over", "individual", "characters", "in", "each", "word", ",", "we", "can", "take", "advantage", "of", "these", "regularities", ".", "This", "can", "be", "particularly", "useful", "for", "handling", "unseen", "words", "–", "for", "example", ",", "if", "we", "have", "never", "seen", "the", "word", "cabinets", "before", ",", "a", "character", "-", "level", "model", "could", "still", "infer", "a", "representation", "for", "this", "word", "if", "it", "has", "previously", "seen", "the", "word", "cabinet", "and", "other", "words", "with", "the", "suffix", "-", "s", ".", "In", "contrast", ",", "a", "word", "-", "level", "model", "can", "only", "represent", "this", "word", "with", "a", "generic", "out", "-", "of", "-", "vocabulary", "representation", ",", "which", "is", "shared", "between", "all", "other", "unseen", "words", ".", "Research", "into", "character", "-", "level", "models", "is", "still", "in", "fairly", "early", "stages", ",", "and", "models", "that", "operate", "exclusively", "on", "characters", "are", "not", "yet", "competitive", "to", "word", "-", "level", "models", "on", "most", "tasks", ".", "However", ",", "instead", "of", "fully", "replacing", "word", "embeddings", ",", "we", "are", "interested", "in", "combining", "the", "two", "approaches", ",", "thereby", "allowing", "the", "model", "to", "take", "advantage", "of", "information", "at", "both", "granularity", "levels", ".", "The", "general", "outline", "of", "our", "approach", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "Each", "word", "is", "broken", "down", "into", "individual", "characters", ",", "these", "are", "then", "mapped", "to", "a", "sequence", "of", "character", "embeddings", ",", "which", "are", "passed", "through", "a", "bidirectional", "LSTM", ":", "We", "then", "use", "the", "last", "hidden", "vectors", "from", "each", "of", "the", "LSTM", "components", ",", "concatenate", "them", "together", ",", "and", "pass", "the", "result", "through", "a", "separate", "non", "-", "linear", "layer", ".", "where", "is", "a", "weight", "matrix", "mapping", "the", "concatenated", "hidden", "vectors", "from", "both", "LSTMs", "into", "a", "joint", "word", "representation", ",", "built", "from", "individual", "characters", ".", "We", "now", "have", "two", "alternative", "feature", "representations", "for", "each", "word", "–", "from", "Section", "[", "reference", "]", "is", "an", "embedding", "learned", "on", "the", "word", "level", ",", "and", "is", "a", "representation", "dynamically", "built", "from", "individual", "characters", "in", "the", "-", "th", "word", "of", "the", "input", "text", ".", "Following", "Lample2016", ",", "one", "possible", "approach", "is", "to", "concatenate", "the", "two", "vectors", "and", "use", "this", "as", "the", "new", "word", "-", "level", "representation", "for", "the", "sequence", "labeling", "model", ":", "This", "approach", ",", "also", "illustrated", "in", "Figure", "[", "reference", "]", ",", "assumes", "that", "the", "word", "-", "level", "and", "character", "-", "level", "components", "learn", "somewhat", "disjoint", "information", ",", "and", "it", "is", "beneficial", "to", "give", "them", "separately", "as", "input", "to", "the", "sequence", "labeler", ".", "section", ":", "Attention", "over", "character", "features", "Alternatively", ",", "we", "can", "have", "the", "word", "embedding", "and", "the", "character", "-", "level", "component", "learn", "the", "same", "semantic", "features", "for", "each", "word", ".", "Instead", "of", "concatenating", "them", "as", "alternative", "feature", "sets", ",", "we", "specifically", "construct", "the", "network", "so", "that", "they", "would", "learn", "the", "same", "representations", ",", "and", "then", "allow", "the", "model", "to", "decide", "how", "to", "combine", "the", "information", "for", "each", "specific", "word", ".", "We", "first", "construct", "the", "word", "representation", "from", "characters", "using", "the", "same", "architecture", "–", "a", "bidirectional", "LSTM", "operates", "over", "characters", ",", "and", "the", "last", "hidden", "states", "are", "used", "to", "create", "vector", "for", "the", "input", "word", ".", "Instead", "of", "concatenating", "this", "with", "the", "word", "embedding", ",", "the", "two", "vectors", "are", "added", "together", "using", "a", "weighted", "sum", ",", "where", "the", "weights", "are", "predicted", "by", "a", "two", "-", "layer", "network", ":", "where", ",", "and", "are", "weight", "matrices", "for", "calculating", ",", "and", "is", "the", "logistic", "function", "with", "values", "in", "the", "range", ".", "The", "vector", "has", "the", "same", "dimensions", "as", "or", ",", "acting", "as", "the", "weight", "between", "the", "two", "vectors", ".", "It", "allows", "the", "model", "to", "dynamically", "decide", "how", "much", "information", "to", "use", "from", "the", "character", "-", "level", "component", "or", "from", "the", "word", "embedding", ".", "This", "decision", "is", "done", "for", "each", "feature", "separately", ",", "which", "adds", "extra", "flexiblity", "–", "for", "example", ",", "words", "with", "regular", "suffixes", "can", "share", "some", "character", "-", "level", "features", ",", "whereas", "irregular", "words", "can", "store", "exceptions", "into", "word", "embeddings", ".", "Furthermore", ",", "previously", "unknown", "words", "are", "able", "to", "use", "character", "-", "level", "regularities", "whenever", "possible", ",", "and", "are", "still", "able", "to", "revert", "to", "using", "the", "generic", "OOV", "token", "when", "necessary", ".", "The", "main", "benefits", "of", "character", "-", "level", "modeling", "are", "expected", "to", "come", "from", "improved", "handling", "of", "rare", "and", "unseen", "words", ",", "whereas", "frequent", "words", "are", "likely", "able", "to", "learn", "high", "-", "quality", "word", "-", "level", "embeddings", "directly", ".", "We", "would", "like", "to", "take", "advantage", "of", "this", ",", "and", "train", "the", "character", "component", "to", "predict", "these", "word", "embeddings", ".", "Our", "attention", "-", "based", "architecture", "requires", "the", "learned", "features", "in", "both", "word", "representations", "to", "align", ",", "and", "we", "can", "add", "in", "an", "extra", "constraint", "to", "encourage", "this", ".", "During", "training", ",", "we", "add", "a", "term", "to", "the", "loss", "function", "that", "optimises", "the", "vector", "to", "be", "similar", "to", "the", "word", "embedding", ":", "Equation", "[", "reference", "]", "maximises", "the", "cosine", "similarity", "between", "and", ".", "Importantly", ",", "this", "is", "done", "only", "for", "words", "that", "are", "not", "out", "-", "of", "-", "vocabulary", "–", "we", "want", "the", "character", "-", "level", "component", "to", "learn", "from", "the", "word", "embeddings", ",", "but", "this", "should", "exclude", "the", "OOV", "embedding", ",", "as", "it", "is", "shared", "between", "many", "words", ".", "We", "use", "to", "set", "this", "cost", "component", "to", "for", "any", "OOV", "tokens", ".", "While", "the", "character", "component", "learns", "general", "regularities", "that", "are", "shared", "between", "all", "the", "words", ",", "individual", "word", "embeddings", "provide", "a", "way", "for", "the", "model", "to", "store", "word", "-", "specific", "information", "and", "any", "exceptions", ".", "Therefore", ",", "while", "we", "want", "the", "character", "-", "based", "model", "to", "shift", "towards", "predicting", "high", "-", "quality", "word", "embeddings", ",", "it", "is", "not", "desireable", "to", "optimise", "the", "word", "embeddings", "towards", "the", "character", "-", "level", "representations", ".", "This", "can", "be", "achieved", "by", "making", "sure", "that", "the", "optimisation", "is", "performed", "only", "in", "one", "direction", ";", "in", "Theano", ",", "the", "disconnected_grad", "function", "gives", "the", "desired", "effect", ".", "section", ":", "Datasets", "We", "evaluate", "the", "sequence", "labeling", "models", "and", "character", "architectures", "on", "8", "different", "datasets", ".", "Table", "[", "reference", "]", "contains", "information", "about", "the", "number", "of", "labels", "and", "dataset", "sizes", "for", "each", "of", "them", ".", "CoNLL00", ":", "The", "CoNLL", "-", "2000", "dataset", "is", "a", "frequently", "used", "benchmark", "for", "the", "task", "of", "chunking", ".", "Wall", "Street", "Journal", "Sections", "15", "-", "18", "from", "the", "Penn", "Treebank", "are", "used", "for", "training", ",", "and", "Section", "20", "as", "the", "test", "data", ".", "As", "there", "is", "no", "official", "development", "set", ",", "we", "separated", "some", "of", "the", "training", "set", "for", "this", "purpose", ".", "CoNLL03", ":", "The", "CoNLL", "-", "2003", "corpus", "was", "created", "for", "the", "shared", "task", "on", "language", "-", "independent", "NER", ".", "We", "use", "the", "English", "section", "of", "the", "dataset", ",", "containing", "news", "stories", "from", "the", "Reuters", "Corpus", ".", "PTB", "-", "POS", ":", "The", "Penn", "Treebank", "POS", "-", "tag", "corpus", "contains", "texts", "from", "the", "Wall", "Street", "Journal", ",", "annotated", "for", "part", "-", "of", "-", "speech", "tags", ".", "The", "PTB", "label", "set", "includes", "36", "main", "tags", "and", "an", "additional", "12", "tags", "covering", "items", "such", "as", "punctuation", ".", "FCEPUBLIC", ":", "The", "publicly", "released", "subset", "of", "the", "First", "Certificate", "in", "English", "(", "FCE", ")", "dataset", "contains", "short", "essays", "written", "by", "language", "learners", "and", "manual", "corrections", "by", "examiners", ".", "We", "use", "a", "version", "of", "this", "corpus", "converted", "into", "a", "binary", "error", "detection", "task", ",", "where", "each", "token", "is", "labeled", "as", "being", "correct", "or", "incorrect", "in", "the", "given", "context", ".", "BC2GM", ":", "The", "BioCreative", "II", "Gene", "Mention", "corpus", "consists", "of", "20", ",", "000", "sentences", "from", "biomedical", "publication", "abstracts", "and", "is", "annotated", "for", "mentions", "of", "the", "names", "of", "genes", ",", "proteins", "and", "related", "entities", "using", "a", "single", "NE", "class", ".", "CHEMDNER", ":", "The", "BioCreative", "IV", "Chemical", "and", "Drug", "NER", "corpus", "consists", "of", "10", ",", "000", "abstracts", "annotated", "for", "mentions", "of", "chemical", "and", "drug", "names", "using", "a", "single", "class", ".", "We", "make", "use", "of", "the", "official", "splits", "provided", "by", "the", "shared", "task", "organizers", ".", "JNLPBA", ":", "The", "JNLPBA", "corpus", "consists", "of", "2", ",", "404", "biomedical", "abstracts", "and", "is", "annotated", "for", "mentions", "of", "five", "entity", "types", ":", "cell", "line", ",", "cell", "type", ",", "dna", ",", "rna", ",", "and", "protein", ".", "The", "corpus", "was", "derived", "from", "GENIA", "corpus", "entity", "annotations", "for", "use", "in", "the", "shared", "task", "organized", "in", "conjuction", "with", "the", "BioNLP", "2004", "workshop", ".", "GENIA", "-", "POS", ":", "The", "GENIA", "corpus", "is", "one", "of", "the", "most", "widely", "used", "resources", "for", "biomedical", "NLP", "and", "has", "a", "rich", "set", "of", "annotations", "including", "parts", "of", "speech", ",", "phrase", "structure", "syntax", ",", "entity", "mentions", ",", "and", "events", ".", "Here", ",", "we", "make", "use", "of", "the", "GENIA", "POS", "annotations", ",", "which", "cover", "2", ",", "000", "PubMed", "abstracts", "(", "approx", ".", "20", ",", "000", "sentences", ")", ".", "We", "use", "the", "same", "210", "-", "document", "test", "set", "as", "Tsuruoka2005", ",", "and", "additionally", "split", "off", "a", "sample", "of", "210", "from", "the", "remaining", "documents", "as", "a", "development", "set", ".", "section", ":", "Experiment", "settings", "For", "data", "prepocessing", ",", "all", "digits", "were", "replaced", "with", "the", "character", "’", "0", "’", ".", "Any", "words", "that", "occurred", "only", "once", "in", "the", "training", "data", "were", "replaced", "by", "the", "generic", "OOV", "token", "for", "word", "embeddings", ",", "but", "were", "still", "used", "in", "the", "character", "-", "level", "components", ".", "The", "word", "embeddings", "were", "initialised", "with", "publicly", "available", "pretrained", "vectors", ",", "created", "using", "word2vec", ",", "and", "then", "fine", "-", "tuned", "during", "model", "training", ".", "For", "the", "general", "-", "domain", "datasets", "we", "used", "300", "-", "dimensional", "vectors", "trained", "on", "Google", "News", ";", "for", "the", "biomedical", "datasets", "we", "used", "200", "-", "dimensional", "vectors", "trained", "on", "PubMed", "and", "PMC", ".", "The", "embeddings", "for", "characters", "were", "set", "to", "length", "and", "initialised", "randomly", ".", "The", "LSTM", "layer", "size", "was", "set", "to", "in", "each", "direction", "for", "both", "word", "-", "and", "character", "-", "level", "components", ".", "The", "hidden", "layer", "has", "size", ",", "and", "the", "combined", "representation", "has", "the", "same", "length", "as", "the", "word", "embeddings", ".", "CRF", "was", "used", "as", "the", "output", "layer", "for", "all", "the", "experiments", "–", "we", "found", "that", "this", "gave", "most", "benefits", "to", "tasks", "with", "larger", "numbers", "of", "possible", "labels", ".", "Parameters", "were", "optimised", "using", "AdaDelta", "with", "default", "learning", "rate", "and", "sentences", "were", "grouped", "into", "batches", "of", "size", ".", "Performance", "on", "the", "development", "set", "was", "measured", "at", "every", "epoch", "and", "training", "was", "stopped", "if", "performance", "had", "not", "improved", "for", "7", "epochs", ";", "the", "best", "-", "performing", "model", "on", "the", "development", "set", "was", "then", "used", "for", "evaluation", "on", "the", "test", "set", ".", "In", "order", "to", "avoid", "any", "outlier", "results", "due", "to", "randomness", "in", "the", "model", "initialisation", ",", "we", "trained", "each", "configuration", "with", "10", "different", "random", "seeds", "and", "present", "here", "the", "averaged", "results", ".", "When", "evaluating", "on", "each", "dataset", ",", "we", "report", "the", "measures", "established", "in", "previous", "work", ".", "Token", "-", "level", "accuracy", "is", "used", "for", "PTB", "-", "POS", "and", "GENIA", "-", "POS", ";", "score", "over", "the", "erroneous", "words", "for", "FCEPUBLIC", ";", "the", "official", "evaluation", "script", "for", "BC2GM", "which", "allows", "for", "alternative", "correct", "entity", "spans", ";", "and", "microaveraged", "mention", "-", "level", "score", "for", "the", "remaining", "datasets", ".", "section", ":", "Results", "While", "optimising", "the", "hyperparameters", "for", "each", "dataset", "separately", "would", "likely", "improve", "individual", "performance", ",", "we", "conduct", "more", "controlled", "experiments", "on", "a", "task", "-", "independent", "model", ".", "Therefore", ",", "we", "use", "the", "same", "hyperparameters", "from", "Section", "[", "reference", "]", "on", "all", "datasets", ",", "and", "the", "development", "set", "is", "only", "used", "for", "the", "stopping", "condition", ".", "With", "these", "experiments", ",", "we", "wish", "to", "determine", "1", ")", "on", "which", "sequence", "labeling", "tasks", "do", "character", "-", "based", "models", "offer", "an", "advantange", ",", "and", "2", ")", "which", "character", "-", "based", "architecture", "performs", "better", ".", "Results", "for", "the", "different", "model", "architectures", "on", "all", "8", "datasets", "are", "shown", "in", "Table", "[", "reference", "]", ".", "As", "can", "be", "seen", ",", "including", "a", "character", "-", "based", "component", "in", "the", "sequence", "labeling", "architecture", "improves", "performance", "on", "every", "benchmark", ".", "The", "NER", "datasets", "have", "the", "largest", "absolute", "improvement", "–", "the", "model", "is", "able", "to", "learn", "character", "-", "level", "patterns", "for", "names", ",", "and", "also", "improve", "the", "handling", "of", "any", "previously", "unseen", "tokens", ".", "Compared", "to", "concatenating", "the", "word", "-", "and", "character", "-", "level", "representations", ",", "the", "attention", "-", "based", "character", "model", "outperforms", "the", "former", "on", "all", "evaluations", ".", "The", "mechanism", "for", "dynamically", "deciding", "how", "much", "character", "-", "level", "information", "to", "use", "allows", "the", "model", "to", "better", "handle", "individual", "word", "representations", ",", "giving", "it", "an", "advantage", "in", "the", "experiments", ".", "Visualisation", "of", "the", "attention", "values", "in", "Figure", "[", "reference", "]", "shows", "that", "the", "model", "is", "actively", "using", "character", "-", "based", "features", ",", "and", "the", "attention", "areas", "vary", "between", "different", "words", ".", "The", "results", "of", "this", "general", "tagging", "architecture", "are", "competitive", ",", "even", "when", "compared", "to", "previous", "work", "using", "hand", "-", "crafted", "features", ".", "The", "network", "achieves", "97.27", "%", "on", "PTB", "-", "POS", "compared", "to", "97.55", "%", "by", "Huang2015", ",", "and", "72.70", "%", "on", "JNLPBA", "compared", "to", "72.55", "%", "by", "Zhou2004", ".", "In", "some", "cases", ",", "we", "are", "also", "able", "to", "beat", "the", "previous", "best", "results", "–", "87.99", "%", "on", "BC2GM", "compared", "to", "87.48", "%", "by", "Campos2015", ",", "and", "41.88", "%", "on", "FCEPUBLIC", "compared", "to", "41.1", "%", "by", "Rei2016", ".", "Lample2016", "report", "a", "considerably", "higher", "result", "of", "90.94", "%", "on", "CoNLL03", ",", "indicating", "that", "the", "chosen", "hyperparameters", "for", "the", "baseline", "system", "are", "suboptimal", "for", "this", "specific", "task", ".", "Compared", "to", "the", "experiments", "presented", "here", ",", "their", "model", "used", "the", "IOBES", "tagging", "scheme", "instead", "of", "the", "original", "IOB", ",", "and", "embeddings", "pretrained", "with", "a", "more", "specialised", "method", "that", "accounts", "for", "word", "order", ".", "It", "is", "important", "to", "also", "compare", "the", "parameter", "counts", "of", "alternative", "neural", "architectures", ",", "as", "this", "shows", "their", "learning", "capacity", "and", "indicates", "their", "time", "requirements", "in", "practice", ".", "Table", "[", "reference", "]", "contains", "the", "parameter", "counts", "on", "three", "representative", "datasets", ".", "While", "keeping", "the", "model", "hyperparameters", "constant", ",", "the", "character", "-", "level", "models", "require", "additional", "parameters", "for", "the", "character", "composition", "and", "character", "embeddings", ".", "However", ",", "the", "attention", "-", "based", "model", "uses", "fewer", "parameters", "compared", "to", "the", "concatenation", "approach", ".", "When", "the", "two", "representations", "are", "concatenated", ",", "the", "overall", "word", "representation", "size", "is", "increased", ",", "which", "in", "turn", "increases", "the", "number", "of", "parameters", "required", "for", "the", "word", "-", "level", "bidirectional", "LSTM", ".", "Therefore", ",", "the", "attention", "-", "based", "character", "architecture", "achieves", "improved", "results", "even", "with", "a", "smaller", "parameter", "footprint", ".", "section", ":", "Related", "work", "There", "is", "a", "wide", "range", "of", "previous", "work", "on", "constructing", "and", "optimising", "neural", "architectures", "applicable", "to", "sequence", "labeling", ".", "Collobert2011", "described", "one", "of", "the", "first", "task", "-", "independent", "neural", "tagging", "models", "using", "convolutional", "neural", "networks", ".", "They", "were", "able", "to", "achieve", "good", "results", "on", "POS", "tagging", ",", "chunking", ",", "NER", "and", "semantic", "role", "labeling", ",", "without", "relying", "on", "hand", "-", "engineered", "features", ".", "Irsoy2014a", "experimented", "with", "multi", "-", "layer", "bidirectional", "Elman", "-", "style", "recurrent", "networks", ",", "and", "found", "that", "the", "deep", "models", "outperformed", "conditional", "random", "fields", "on", "the", "task", "of", "opinion", "mining", ".", "Huang2015", "described", "a", "bidirectional", "LSTM", "model", "with", "a", "CRF", "layer", ",", "which", "included", "hand", "-", "crafted", "features", "specialised", "for", "the", "task", "of", "named", "entity", "recognition", ".", "Rei2016", "evaluated", "a", "range", "of", "neural", "architectures", ",", "including", "convolutional", "and", "recurrent", "networks", ",", "on", "the", "task", "of", "error", "detection", "in", "learner", "writing", ".", "The", "word", "-", "level", "sequence", "labeling", "model", "described", "in", "this", "paper", "follows", "the", "previous", "work", ",", "combining", "useful", "design", "choices", "from", "each", "of", "them", ".", "In", "addition", ",", "we", "extended", "the", "model", "with", "two", "alternative", "character", "-", "level", "architectures", ",", "and", "evaluated", "its", "performance", "on", "8", "different", "datasets", ".", "Character", "-", "level", "models", "have", "the", "potential", "of", "capturing", "morpheme", "patterns", ",", "thereby", "improving", "generalisation", "on", "both", "frequent", "and", "unseen", "words", ".", "In", "recent", "years", ",", "there", "has", "been", "an", "increase", "in", "research", "into", "these", "models", ",", "resulting", "in", "several", "interesting", "applications", ".", "Ling2015b", "described", "a", "character", "-", "level", "neural", "model", "for", "machine", "translation", ",", "performing", "both", "encoding", "and", "decoding", "on", "individual", "characters", ".", "Kim2016", "implemented", "a", "language", "model", "where", "encoding", "is", "performed", "by", "a", "convolutional", "network", "and", "LSTM", "over", "characters", ",", "whereas", "predictions", "are", "given", "on", "the", "word", "-", "level", ".", "Cao2016", "proposed", "a", "method", "for", "learning", "both", "word", "embeddings", "and", "morphological", "segmentation", "with", "a", "bidirectional", "recurrent", "network", "over", "characters", ".", "There", "is", "also", "research", "on", "performing", "parsing", "and", "text", "classification", "with", "character", "-", "level", "neural", "models", ".", "Ling2015a", "proposed", "a", "neural", "architecture", "that", "replaces", "word", "embeddings", "with", "dynamically", "-", "constructed", "character", "-", "based", "representations", ".", "We", "applied", "a", "similar", "method", "for", "operating", "over", "characters", ",", "but", "combined", "them", "with", "word", "embeddings", "instead", "of", "replacing", "them", ",", "as", "this", "allows", "the", "model", "to", "benefit", "from", "both", "approaches", ".", "Lample2016", "described", "a", "model", "where", "the", "character", "-", "level", "representation", "is", "combined", "with", "word", "embeddings", "through", "concatenation", ".", "In", "this", "work", ",", "we", "proposed", "an", "alternative", "architecture", ",", "where", "the", "representations", "are", "combined", "using", "an", "attention", "mechanism", ",", "and", "evaluated", "both", "approaches", "on", "a", "range", "of", "tasks", "and", "datasets", ".", "Recently", ",", "Miyamoto2016", "have", "also", "described", "a", "related", "method", "for", "the", "task", "of", "language", "modelling", ",", "combining", "characters", "and", "word", "embeddings", "using", "gating", ".", "section", ":", "Conclusion", "Developments", "in", "neural", "network", "research", "allow", "for", "model", "architectures", "that", "work", "well", "on", "a", "wide", "range", "of", "sequence", "labeling", "datasets", "without", "requiring", "hand", "-", "crafted", "data", ".", "While", "word", "-", "level", "representation", "learning", "is", "a", "powerful", "tool", "for", "automatically", "discovering", "useful", "features", ",", "these", "models", "still", "come", "with", "certain", "weaknesses", "–", "rare", "words", "have", "low", "-", "quality", "representations", ",", "previously", "unseen", "words", "can", "not", "be", "modeled", "at", "all", ",", "and", "morpheme", "-", "level", "information", "is", "not", "shared", "with", "the", "whole", "vocabulary", ".", "In", "this", "paper", ",", "we", "investigated", "character", "-", "level", "model", "components", "for", "a", "sequence", "labeling", "architecture", ",", "which", "allow", "the", "system", "to", "learn", "useful", "patterns", "from", "sub", "-", "word", "units", ".", "In", "addition", "to", "a", "bidirectional", "LSTM", "operating", "over", "words", ",", "a", "separate", "bidirectional", "LSTM", "is", "used", "to", "construct", "word", "representations", "from", "individual", "characters", ".", "We", "proposed", "a", "novel", "architecture", "for", "combining", "the", "character", "-", "based", "representation", "with", "the", "word", "embedding", "by", "using", "an", "attention", "mechanism", ",", "allowing", "the", "model", "to", "dynamically", "choose", "which", "information", "to", "use", "from", "each", "information", "source", ".", "In", "addition", ",", "the", "character", "-", "level", "composition", "function", "is", "augmented", "with", "a", "novel", "training", "objective", ",", "optimising", "it", "to", "predict", "representations", "that", "are", "similar", "to", "the", "word", "embeddings", "in", "the", "model", ".", "The", "evaluation", "was", "performed", "on", "8", "different", "sequence", "labeling", "datasets", ",", "covering", "a", "range", "of", "tasks", "and", "domains", ".", "We", "found", "that", "incorporating", "character", "-", "level", "information", "into", "the", "model", "improved", "performance", "on", "every", "benchmark", ",", "indicating", "that", "capturing", "features", "regarding", "characters", "and", "morphmes", "is", "indeed", "useful", "in", "a", "general", "-", "purpose", "tagging", "system", ".", "In", "addition", ",", "the", "attention", "-", "based", "model", "for", "combining", "character", "representations", "outperformed", "the", "concatenation", "method", "used", "in", "previous", "work", "in", "all", "evaluations", ".", "Even", "though", "the", "proposed", "method", "requires", "fewer", "parameters", ",", "the", "added", "ability", "of", "controlling", "how", "much", "character", "-", "level", "information", "is", "used", "for", "each", "word", "has", "led", "to", "improved", "performance", "on", "a", "range", "of", "different", "tasks", ".", "bibliography", ":", "References"], "entity": [[6, 10, ["Neural", "Sequence", "Labeling", "Models"], "Method"], [10, 13, ["Sequence", "labeling", "architectures"], "Method"], [18, 19, ["similarity"], "Task"], [32, 36, ["character", "-", "level", "extensions"], "Method"], [47, 49, ["word", "representations"], "Method"], [53, 55, ["attention", "mechanism"], "Method"], [109, 113, ["attention", "-", "based", "architecture"], "Method"], [146, 148, ["NLP", "tasks"], "Task"], [150, 153, ["named", "entity", "recognition"], "Task"], [154, 155, ["NER"], "Task"], [157, 162, ["part", "-", "of", "-", "speech"], "Task"], [163, 164, ["POS"], "Task"], [175, 177, ["sequence", "labeling"], "Task"], [184, 187, ["sequence", "labeling", "models"], "Method"], [195, 197, ["downstream", "applications"], "Task"], [208, 210, ["feature", "engineering"], "Task"], [215, 216, ["gazetteers"], "Task"], [217, 220, ["named", "entity", "recognition"], "Task"], [226, 228, ["morphological", "analyser"], "Method"], [229, 232, ["POS", "-", "tagging"], "Task"], [236, 238, ["neural", "architectures"], "Method"], [239, 241, ["representation", "learning"], "Method"], [258, 261, ["sequence", "labeling", "systems"], "Method"], [284, 286, ["neural", "models"], "Method"], [290, 292, ["word", "embeddings"], "Method"], [314, 318, ["count", "-", "based", "models"], "Method"], [336, 341, ["out", "-", "of", "-", "vocabulary"], "Task"], [342, 343, ["OOV"], "Task"], [454, 460, ["task", "-", "independent", "neural", "network", "architecture"], "Method"], [461, 463, ["sequence", "labeling"], "Task"], [474, 478, ["character", "-", "level", "information"], "Task"], [512, 516, ["character", "-", "level", "representations"], "Method"], [517, 519, ["word", "embeddings"], "Method"], [521, 523, ["gating", "mechanism"], "Method"], [556, 558, ["model", "training"], "Task"], [560, 564, ["character", "-", "level", "representations"], "Method"], [578, 580, ["neural", "models"], "Method"], [587, 588, ["NER"], "Task"], [589, 592, ["POS", "-", "tagging"], "Task"], [593, 594, ["chunking"], "Task"], [595, 597, ["error", "detection"], "Task"], [607, 611, ["character", "-", "based", "component"], "Method"], [613, 616, ["sequence", "labeling", "model"], "Method"], [629, 633, ["attention", "-", "based", "architecture"], "Method"], [654, 656, ["sequence", "labeling"], "Task"], [661, 666, ["word", "-", "level", "neural", "network"], "Method"], [667, 669, ["sequence", "labeling"], "Task"], [702, 705, ["sequence", "labeling", "network"], "Method"], [744, 746, ["word", "embeddings"], "Method"], [768, 772, ["context", "-", "specific", "representations"], "Method"], [775, 782, ["forward", "-", "and", "backward", "-", "conditioned", "representations"], "Method"], [811, 812, ["LSTM"], "Method"], [877, 879, ["label", "predictions"], "Task"], [884, 886, ["softmax", "layer"], "Method"], [888, 891, ["conditional", "random", "field"], "Method"], [892, 893, ["CRF"], "Method"], [898, 899, ["softmax"], "Method"], [955, 957, ["categorical", "crossentropy"], "Method"], [981, 982, ["CRF"], "Method"], [984, 986, ["output", "layer"], "Method"], [1037, 1039, ["Viterbi", "algorithm"], "Method"], [1059, 1061, ["CRF", "score"], "Metric"], [1126, 1128, ["Viterbi", "algorithm"], "Method"], [1134, 1136, ["CRF", "model"], "Method"], [1138, 1140, ["loss", "function"], "Method"], [1170, 1175, ["Character", "-", "level", "sequence", "labeling"], "Task"], [1175, 1177, ["Distributed", "embeddings"], "Method"], [1188, 1190, ["vector", "representations"], "Method"], [1263, 1267, ["character", "-", "level", "model"], "Method"], [1296, 1300, ["word", "-", "level", "model"], "Method"], [1326, 1330, ["character", "-", "level", "models"], "Method"], [1349, 1353, ["word", "-", "level", "models"], "Method"], [1456, 1460, ["non", "-", "linear", "layer"], "Method"], [1473, 1474, ["LSTMs"], "Method"], [1476, 1479, ["joint", "word", "representation"], "Method"], [1490, 1492, ["feature", "representations"], "Method"], [1503, 1504, ["embedding"], "Method"], [1547, 1551, ["word", "-", "level", "representation"], "Method"], [1553, 1556, ["sequence", "labeling", "model"], "Method"], [1596, 1598, ["sequence", "labeler"], "Task"], [1611, 1613, ["word", "embedding"], "Method"], [1615, 1619, ["character", "-", "level", "component"], "Method"], [1672, 1674, ["word", "representation"], "Method"], [1709, 1711, ["word", "embedding"], "Method"], [1720, 1722, ["weighted", "sum"], "Method"], [1732, 1734, ["layer", "network"], "Method"], [1747, 1749, ["logistic", "function"], "Method"], [1787, 1791, ["character", "-", "level", "component"], "Method"], [1794, 1796, ["word", "embedding"], "Method"], [1871, 1875, ["character", "-", "level", "modeling"], "Method"], [1917, 1919, ["character", "component"], "Method"], [1926, 1930, ["attention", "-", "based", "architecture"], "Method"], [1936, 1938, ["word", "representations"], "Method"], [1954, 1955, ["training"], "Task"], [2007, 2011, ["character", "-", "level", "component"], "Method"], [2039, 2041, ["cost", "component"], "Method"], [2049, 2051, ["character", "component"], "Method"], [2087, 2091, ["character", "-", "based", "model"], "Method"], [2095, 2100, ["high", "-", "quality", "word", "embeddings"], "Task"], [2112, 2116, ["character", "-", "level", "representations"], "Method"], [2126, 2127, ["optimisation"], "Task"], [2151, 2154, ["sequence", "labeling", "models"], "Method"], [2155, 2157, ["character", "architectures"], "Method"], [2181, 2182, ["CoNLL00"], "Material"], [2184, 2188, ["CoNLL", "-", "2000", "dataset"], "Material"], [2197, 2198, ["chunking"], "Task"], [2199, 2203, ["Wall", "Street", "Journal", "Sections"], "Material"], [2208, 2210, ["Penn", "Treebank"], "Material"], [2242, 2243, ["CoNLL03"], "Material"], [2245, 2249, ["CoNLL", "-", "2003", "corpus"], "Material"], [2253, 2255, ["shared", "task"], "Task"], [2264, 2266, ["English", "section"], "Material"], [2275, 2277, ["Reuters", "Corpus"], "Material"], [2278, 2281, ["PTB", "-", "POS"], "Material"], [2293, 2296, ["Wall", "Street", "Journal"], "Material"], [2307, 2310, ["PTB", "label", "set"], "Material"], [2325, 2326, ["FCEPUBLIC"], "Material"], [2333, 2337, ["First", "Certificate", "in", "English"], "Material"], [2338, 2339, ["FCE"], "Material"], [2364, 2368, ["binary", "error", "detection", "task"], "Task"], [2384, 2385, ["BC2GM"], "Material"], [2387, 2392, ["BioCreative", "II", "Gene", "Mention", "corpus"], "Material"], [2423, 2424, ["CHEMDNER"], "Method"], [2466, 2467, ["JNLPBA"], "Material"], [2506, 2510, ["GENIA", "corpus", "entity", "annotations"], "Material"], [2514, 2516, ["shared", "task"], "Task"], [2521, 2524, ["BioNLP", "2004", "workshop"], "Material"], [2530, 2532, ["GENIA", "corpus"], "Material"], [2541, 2543, ["biomedical", "NLP"], "Task"], [2626, 2628, ["data", "prepocessing"], "Task"], [2667, 2671, ["character", "-", "level", "components"], "Method"], [2693, 2695, ["model", "training"], "Method"], [2710, 2712, ["Google", "News"], "Material"], [2725, 2726, ["PubMed"], "Material"], [2727, 2728, ["PMC"], "Material"], [2762, 2764, ["hidden", "layer"], "Method"], [2769, 2771, ["combined", "representation"], "Method"], [2780, 2781, ["CRF"], "Method"], [2812, 2813, ["AdaDelta"], "Method"], [2815, 2817, ["learning", "rate"], "Metric"], [2880, 2882, ["model", "initialisation"], "Method"], [2914, 2918, ["Token", "-", "level", "accuracy"], "Metric"], [2921, 2924, ["PTB", "-", "POS"], "Material"], [2935, 2936, ["FCEPUBLIC"], "Material"], [2938, 2941, ["official", "evaluation", "script"], "Metric"], [2942, 2943, ["BC2GM"], "Material"], [2952, 2957, ["microaveraged", "mention", "-", "level", "score"], "Metric"], [2986, 2990, ["task", "-", "independent", "model"], "Method"], [3031, 3034, ["sequence", "labeling", "tasks"], "Task"], [3035, 3039, ["character", "-", "based", "models"], "Method"], [3047, 3051, ["character", "-", "based", "architecture"], "Method"], [3079, 3083, ["character", "-", "based", "component"], "Method"], [3085, 3088, ["sequence", "labeling", "architecture"], "Method"], [3131, 3138, ["word", "-", "and", "character", "-", "level", "representations"], "Method"], [3140, 3145, ["attention", "-", "based", "character", "model"], "Method"], [3172, 3174, ["word", "representations"], "Method"], [3219, 3221, ["tagging", "architecture"], "Method"], [3242, 3245, ["PTB", "-", "POS"], "Material"], [3256, 3257, ["JNLPBA"], "Material"], [3282, 3283, ["BC2GM"], "Material"], [3294, 3295, ["FCEPUBLIC"], "Material"], [3312, 3313, ["CoNLL03"], "Material"], [3318, 3319, ["hyperparameters"], "Method"], [3341, 3344, ["IOBES", "tagging", "scheme"], "Method"], [3348, 3349, ["IOB"], "Method"], [3351, 3353, ["embeddings", "pretrained"], "Method"], [3375, 3377, ["neural", "architectures"], "Method"], [3382, 3384, ["learning", "capacity"], "Metric"], [3413, 3417, ["character", "-", "level", "models"], "Method"], [3431, 3435, ["attention", "-", "based", "model"], "Method"], [3441, 3443, ["concatenation", "approach"], "Method"], [3453, 3456, ["word", "representation", "size"], "Metric"], [3470, 3475, ["word", "-", "level", "bidirectional", "LSTM"], "Method"], [3479, 3484, ["attention", "-", "based", "character", "architecture"], "Method"], [3510, 3512, ["neural", "architectures"], "Method"], [3514, 3516, ["sequence", "labeling"], "Task"], [3523, 3529, ["task", "-", "independent", "neural", "tagging", "models"], "Method"], [3530, 3533, ["convolutional", "neural", "networks"], "Method"], [3545, 3546, ["chunking"], "Task"], [3547, 3548, ["NER"], "Task"], [3549, 3552, ["semantic", "role", "labeling"], "Task"], [3564, 3573, ["multi", "-", "layer", "bidirectional", "Elman", "-", "style", "recurrent", "networks"], "Method"], [3578, 3580, ["deep", "models"], "Method"], [3581, 3584, ["conditional", "random", "fields"], "Method"], [3588, 3590, ["opinion", "mining"], "Task"], [3594, 3597, ["bidirectional", "LSTM", "model"], "Method"], [3599, 3601, ["CRF", "layer"], "Method"], [3613, 3616, ["named", "entity", "recognition"], "Task"], [3622, 3624, ["neural", "architectures"], "Method"], [3626, 3630, ["convolutional", "and", "recurrent", "networks"], "Method"], [3635, 3637, ["error", "detection"], "Task"], [3638, 3640, ["learner", "writing"], "Task"], [3642, 3648, ["word", "-", "level", "sequence", "labeling", "model"], "Method"], [3676, 3680, ["character", "-", "level", "architectures"], "Method"], [3690, 3694, ["Character", "-", "level", "models"], "Method"], [3704, 3705, ["generalisation"], "Task"], [3736, 3741, ["character", "-", "level", "neural", "model"], "Method"], [3742, 3744, ["machine", "translation"], "Task"], [3747, 3748, ["encoding"], "Task"], [3749, 3750, ["decoding"], "Task"], [3757, 3759, ["language", "model"], "Method"], [3760, 3761, ["encoding"], "Method"], [3765, 3767, ["convolutional", "network"], "Method"], [3768, 3769, ["LSTM"], "Method"], [3789, 3791, ["word", "embeddings"], "Task"], [3792, 3794, ["morphological", "segmentation"], "Task"], [3796, 3799, ["bidirectional", "recurrent", "network"], "Method"], [3808, 3809, ["parsing"], "Task"], [3810, 3812, ["text", "classification"], "Task"], [3813, 3818, ["character", "-", "level", "neural", "models"], "Method"], [3822, 3824, ["neural", "architecture"], "Method"], [3832, 3836, ["character", "-", "based", "representations"], "Method"], [3875, 3879, ["character", "-", "level", "representation"], "Method"], [3882, 3884, ["word", "embeddings"], "Method"], [3885, 3886, ["concatenation"], "Method"], [3904, 3906, ["attention", "mechanism"], "Method"], [3932, 3934, ["language", "modelling"], "Task"], [3938, 3940, ["word", "embeddings"], "Method"], [3941, 3942, ["gating"], "Method"], [3948, 3951, ["neural", "network", "research"], "Method"], [3953, 3955, ["model", "architectures"], "Method"], [3974, 3979, ["word", "-", "level", "representation", "learning"], "Method"], [3984, 3988, ["automatically", "discovering", "useful", "features"], "Task"], [4034, 4039, ["character", "-", "level", "model", "components"], "Method"], [4041, 4044, ["sequence", "labeling", "architecture"], "Method"], [4077, 4079, ["word", "representations"], "Method"], [4091, 4095, ["character", "-", "based", "representation"], "Method"], [4097, 4099, ["word", "embedding"], "Method"], [4102, 4104, ["attention", "mechanism"], "Method"], [4134, 4136, ["training", "objective"], "Metric"], [4205, 4207, ["tagging", "system"], "Task"], [4212, 4216, ["attention", "-", "based", "model"], "Method"], [4218, 4220, ["character", "representations"], "Method"], [4222, 4224, ["concatenation", "method"], "Method"], [91, 95, ["character", "-", "level", "extensions"], "Method"], [168, 169, ["parsing"], "Task"], [373, 374, ["OOV"], "Task"], [422, 423, ["POS"], "Task"], [652, 653, ["LSTM"], "Method"], [757, 758, ["LSTM"], "Method"], [1308, 1313, ["out", "-", "of", "-", "vocabulary"], "Task"], [1429, 1430, ["LSTM"], "Method"], [1442, 1443, ["LSTM"], "Method"], [1683, 1684, ["LSTM"], "Method"], [1862, 1863, ["OOV"], "Task"], [1998, 2003, ["out", "-", "of", "-", "vocabulary"], "Task"], [2023, 2024, ["OOV"], "Task"], [2044, 2045, ["OOV"], "Task"], [2259, 2260, ["NER"], "Task"], [2283, 2285, ["Penn", "Treebank"], "Material"], [2285, 2286, ["POS"], "Task"], [2299, 2304, ["part", "-", "of", "-", "speech"], "Task"], [2431, 2432, ["NER"], "Task"], [2469, 2470, ["JNLPBA"], "Material"], [2527, 2528, ["POS"], "Task"], [2573, 2574, ["POS"], "Task"], [2581, 2582, ["PubMed"], "Material"], [2655, 2656, ["OOV"], "Task"], [2742, 2743, ["LSTM"], "Method"], [2927, 2928, ["POS"], "Task"], [3095, 3096, ["NER"], "Task"], [3542, 3543, ["POS"], "Task"], [4064, 4065, ["LSTM"], "Method"], [4072, 4073, ["LSTM"], "Method"]]} {"doc_id": "25ace72fc3e483d5d52c9209f76b04f0fdd08f9b", "text": ["document", ":", "Fused", "Text", "Segmentation", "Networks", "for", "Multi", "-", "oriented", "Scene", "Text", "Detection", "In", "this", "paper", ",", "we", "introduce", "a", "novel", "end", "-", "end", "framework", "for", "multi", "-", "oriented", "scene", "text", "detection", "from", "an", "instance", "-", "aware", "semantic", "segmentation", "perspective", ".", "We", "present", "Fused", "Text", "Segmentation", "Networks", ",", "which", "combine", "multi", "-", "level", "features", "during", "the", "feature", "extracting", "as", "text", "instance", "may", "rely", "on", "finer", "feature", "expression", "compared", "to", "general", "objects", ".", "It", "detects", "and", "segments", "the", "text", "instance", "jointly", "and", "simultaneously", ",", "leveraging", "merits", "from", "both", "semantic", "segmentation", "task", "and", "region", "proposal", "based", "object", "detection", "task", ".", "Not", "involving", "any", "extra", "pipelines", ",", "our", "approach", "surpasses", "the", "current", "state", "of", "the", "art", "on", "multi", "-", "oriented", "scene", "text", "detection", "benchmarks", ":", "ICDAR2015", "Incidental", "Scene", "Text", "and", "MSRA", "-", "TD500", "reaching", "Hmean", "84.1", "%", "and", "82.0", "%", "respectively", ".", "Morever", ",", "we", "report", "a", "baseline", "on", "total", "-", "text", "containing", "curved", "text", "which", "suggests", "effectiveness", "of", "the", "proposed", "approach", ".", "section", ":", "Introduction", "Recently", ",", "scene", "text", "detection", "has", "drawn", "great", "attention", "from", "computer", "vision", "and", "machine", "learning", "community", ".", "Driven", "by", "many", "content", "-", "based", "image", "applications", "such", "as", "photo", "translation", "and", "receipt", "content", "recognition", ",", "it", "has", "become", "a", "promising", "and", "challenging", "research", "area", "both", "in", "academia", "and", "industry", ".", "Detecting", "text", "in", "natural", "images", "is", "difficult", ",", "because", "both", "text", "and", "background", "may", "be", "complex", "in", "the", "wild", "and", "it", "often", "suffers", "from", "disturbance", "such", "as", "occlusion", "and", "uncontrollable", "lighting", "conditions", ".", "Previous", "text", "detection", "methods", "have", "achieved", "promising", "results", "on", "several", "benchmarks", ".", "The", "essential", "problem", "in", "text", "detection", "is", "to", "represent", "text", "region", "using", "discriminative", "features", ".", "Conventionally", ",", "hand", "-", "crafted", "features", "are", "designed", "to", "capture", "the", "properties", "of", "text", "region", "such", "as", "texture", "and", "shape", ",", "while", "in", "the", "past", "few", "years", ",", "deep", "learning", "based", "approaches", "directly", "learn", "hierarchical", "features", "from", "training", "data", ",", "demonstrating", "more", "accurate", "and", "efficient", "performance", "in", "various", "benchmarks", "such", "as", "ICDAR", "series", "contests", ".", "Existing", "methods", "have", "obtained", "decent", "performance", "for", "detecting", "horizontal", "or", "near", "-", "horizontal", "text", ".", "While", "horizontal", "text", "detection", "has", "constraints", "of", "axis", "-", "aligned", "bounding", "-", "box", "ground", "truth", ",", "the", "multi", "-", "oriented", "text", "is", "not", "restrictive", "to", "a", "particular", "orientation", "and", "usually", "uses", "quadrilaterals", "for", "annotations", ".", "Therefore", ",", "it", "reports", "relatively", "lower", "accuracies", "in", "ICDAR", "2015", "Competition", "Challenge", "4", "“Incidental", "scene", "text", "localization”", "compared", "to", "horizontal", "scene", "text", "detection", "benchmarks", ".", "Recently", ",", "a", "few", "approaches", "have", "been", "proposed", "to", "address", "the", "multi", "-", "oriented", "text", "detection", ".", "In", "general", ",", "there", "are", "currently", "four", "different", "types", "of", "methods", ".", "Region", "based", "methods", "leverage", "advanced", "object", "detection", "techniques", "such", "as", "Faster", "RCNN", "and", "SSD", ".", "Segmentation", "-", "based", "methods", "mainly", "utilize", "fully", "convolutional", "neural", "networks", "(", "FCN", ")", "for", "generating", "text", "score", "maps", ",", "which", "often", "need", "several", "stages", "and", "components", "to", "achieve", "final", "detections", ".", "Direct", "regression", "based", "method", "regresses", "the", "position", "and", "size", "of", "an", "object", "from", "a", "given", "point", ".", "Finally", ",", "hybrid", "method", "combines", "text", "scores", "map", "and", "rotated", "/", "quadrangle", "bounding", "boxes", "generation", "to", "collaboratively", "obtain", "the", "efficient", "and", "accurate", "performance", "in", "multi", "-", "oriented", "text", "detection", ".", "Inspired", "by", "recent", "advance", "of", "instance", "-", "aware", "semantic", "segmentation", ",", "we", "present", "a", "novel", "perspective", "to", "handle", "the", "task", "of", "multi", "-", "oriented", "text", "detection", ".", "In", "this", "work", ",", "we", "leverage", "the", "merits", "from", "accurate", "region", "proposal", "based", "methods", ",", "and", "flexible", "segmentation", "based", "methods", "which", "can", "easily", "generate", "arbitrary", "-", "shaped", "text", "mask", ".", "It", "is", "an", "end", "-", "to", "-", "end", "trainable", "framework", "excluding", "redundant", "and", "low", "-", "efficient", "pipelines", "such", "as", "the", "use", "of", "text", "/", "nontext", "salient", "map", "and", "text", "-", "line", "generation", ".", "Based", "on", "region", "proposal", "network", "(", "RPN", ")", ",", "our", "approach", "detects", "and", "segments", "text", "instance", "simultaneously", ",", "followed", "by", "non", "-", "maximum", "suppression", "(", "NMS", ")", "to", "suppress", "overlapping", "instances", ".", "Finally", ",", "a", "minimum", "quadrangle", "bounding", "box", "to", "fit", "each", "instance", "area", "is", "generated", "as", "the", "result", "of", "the", "whole", "detection", "process", ".", "Our", "main", "contributions", "are", "summarized", "as", "follows", ":", "We", "present", "an", "end", "-", "end", "efficient", "and", "trainable", "solution", "for", "multi", "-", "oriented", "text", "detection", "from", "an", "instance", "aware", "segmentation", "perspective", ",", "excluding", "any", "redundant", "pipelines", ".", "During", "feature", "extraction", ",", "feature", "maps", "are", "composed", "in", "a", "fused", "fashion", "to", "adaptively", "satisfy", "the", "finer", "representation", "of", "text", "instance", ".", "Mask", "-", "NMS", "is", "introduced", "to", "improve", "the", "standard", "NMS", "when", "facing", "heavily", "inclined", "or", "line", "-", "level", "text", "instances", ".", "Without", "many", "bells", "and", "whistles", ",", "our", "approach", "outperforms", "state", "of", "the", "art", "on", "current", "multi", "-", "oriented", "text", "detection", "benchmarks", ".", "section", ":", "Related", "work", "Detecting", "text", "in", "natural", "images", "has", "been", "widely", "studied", "in", "past", "few", "years", ",", "motivated", "by", "many", "text", "-", "related", "real", "-", "world", "applications", "such", "as", "photo", "OCR", "and", "blind", "navigation", ".", "One", "of", "the", "mainstream", "traditional", "methods", "for", "scene", "text", "detection", "are", "Connected", "Components", "(", "CCs", ")", "based", "methods", ",", "which", "consider", "text", "as", "a", "group", "of", "individual", "components", "such", "as", "characters", ".", "Within", "these", "methods", ",", "stroke", "width", "transform", "(", "SWT", ")", "and", "maximally", "stable", "extremal", "region", "(", "MSER", ")", "are", "usually", "used", "to", "seek", "character", "candidates", ".", "Finally", ",", "these", "candidates", "are", "combined", "to", "obtain", "text", "objects", ".", "Although", "these", "bottom", "up", "approaches", "may", "be", "accurate", "on", "some", "benchmarks", ",", "they", "often", "suffer", "from", "too", "many", "pipelines", ",", "which", "may", "cause", "inefficiency", ".", "Another", "mainstream", "traditional", "methods", "are", "sliding", "window", "based", ".", "These", "methods", "often", "use", "a", "fixed", "-", "size", "or", "multi", "-", "scale", "window", "to", "slide", "through", "the", "image", "searching", "the", "region", "which", "most", "likely", "contains", "text", ".", "However", ",", "the", "process", "of", "sliding", "window", "may", "involve", "large", "computational", "cost", "which", "results", "in", "inefficiency", ".", "Generally", ",", "traditional", "methods", "often", "require", "several", "steps", "to", "obtain", "final", "detections", ",", "and", "hand", "-", "designed", "features", "are", "usually", "used", "to", "represent", "properties", "of", "text", ".", "Therefore", ",", "they", "may", "suffer", "from", "inefficiency", "and", "low", "generalization", "ability", "against", "complex", "situations", "such", "as", "non", "-", "uniform", "illumination", ".", "Recent", "progress", "on", "deep", "learning", "based", "approaches", "for", "object", "detection", "and", "semantic", "segmentation", "has", "provided", "new", "techniques", "for", "reading", "text", "in", "the", "wild", ",", "which", "can", "be", "also", "seen", "as", "an", "instance", "of", "general", "object", "detection", ".", "Driven", "by", "the", "advance", "of", "object", "detection", "frameworks", "such", "as", "Faster", "RCNN", "and", "SSD", ",", "these", "methods", "achieved", "state", "of", "the", "art", "by", "either", "using", "a", "region", "proposal", "network", "to", "first", "classify", "some", "text", "region", "proposals", ",", "or", "directly", "regress", "text", "bounding", "boxes", "coordinates", "from", "a", "set", "of", "default", "boxes", ".", "These", "methods", "are", "able", "to", "achieve", "leading", "performance", "on", "horizontal", "or", "multi", "-", "oriented", "scene", "text", "detection", "benchmarks", ".", "However", ",", "they", "may", "also", "be", "restricted", "to", "rectangular", "bounding", "box", "constraints", "even", "with", "appropriate", "rotation", ".", "Different", "from", "these", "methods", ",", "FCN", "based", "approaches", "generate", "text", "/", "non", "-", "text", "map", "which", "classifies", "text", "at", "the", "pixel", "level", ".", "Though", "it", "may", "be", "suited", "well", "for", "arbitrary", "shape", "of", "text", "in", "natural", "images", ",", "it", "often", "involves", "several", "pipelines", "which", "leads", "to", "inefficiency", ".", "Inspired", "by", "recent", "advance", "on", "instance", "-", "aware", "semantic", "segmentation", ",", "we", "present", "an", "end", "-", "end", "trainable", "framework", "called", "Fused", "Text", "Segmentation", "Networks", "(", "FTSN", ")", "to", "handle", "arbitrary", "-", "shape", "text", "detection", "with", "no", "extra", "pipelines", "involved", ".", "It", "inherits", "merits", "from", "both", "object", "detection", "and", "semantic", "segmentation", "architecture", "which", "efficiently", "detects", "and", "segments", "an", "text", "instance", "simultaneously", "and", "accurately", "gives", "predictions", "in", "the", "pixel", "level", ".", "As", "text", "may", "rely", "on", "finer", "feature", "representation", ",", "a", "fused", "structure", "formed", "by", "multi", "-", "level", "feature", "maps", "is", "set", "to", "fit", "this", "property", ".", "section", ":", "Methods", "The", "proposed", "framework", "for", "multi", "-", "oriented", "scene", "text", "detection", "is", "diagrammed", "in", "Fig.2", ".", "It", "is", "a", "deep", "CNN", "model", "which", "mainly", "consists", "of", "three", "parts", ".", "Feature", "representations", "of", "each", "image", "are", "extracted", "through", "resnet", "-", "101", "backbone", ",", "then", "multi", "-", "level", "feature", "maps", "are", "fused", "as", "FusedMapA", "which", "is", "fed", "to", "the", "region", "proposed", "network", "(", "RPN", ")", "for", "text", "region", "of", "interest", "(", "ROI", ")", "generation", "and", "FusedMapB", "for", "later", "rois", "’", "PSROIPooling", ".", "Finally", "the", "rois", "are", "sent", "to", "the", "detection", ",", "segmentation", "and", "box", "regression", "branches", "to", "output", "text", "instances", "in", "pixel", "level", "along", "with", "their", "corresponding", "bounding", "boxes", ".", "The", "post", "-", "processing", "part", "includes", "NMS", "and", "minimal", "quadrilateral", "generation", ".", "subsection", ":", "Network", "Architecture", "The", "convolutional", "feature", "representation", "is", "designed", "in", "a", "fusion", "fashion", ".", "The", "text", "instance", "is", "not", "like", "the", "general", "object", "such", "as", "people", "and", "cars", "which", "have", "relatively", "strong", "semantics", ".", "On", "the", "contrary", ",", "texts", "often", "vary", "tremendously", "in", "intra", "-", "class", "geometries", ".", "Consequently", ",", "low", "-", "level", "features", "should", "be", "taken", "into", "consideration", ".", "Basically", ",", "resnet", "-", "101", "consists", "of", "five", "stages", ".", "Before", "region", "proposing", ",", "stage3", "and", "upsampled", "stage4", "feature", "maps", "are", "combined", "to", "form", "FusedMapA", "through", "element", "-", "wise", "adding", ",", "then", "upsampled", "feature", "maps", "from", "stage5", "are", "fused", "with", "FusedMapA", "to", "form", "FusedMapB.", "It", "is", "noted", "that", "downsampling", "is", "not", "involved", "during", "stage5", ".", "Instead", ",", "we", "use", "the", "“hole", "algorithm”", "to", "keep", "the", "feature", "stride", "and", "maintain", "the", "receptive", "field", ".", "The", "reason", "for", "this", "is", "that", "both", "text", "properties", "and", "the", "segmentation", "task", "may", "require", "finer", "features", "and", "involving", "final", "downsampling", "may", "lose", "some", "useful", "information", ".", "Because", "using", "feature", "stride", "of", "stage3", "may", "cause", "millions", "of", "anchors", "in", "original", "RPN", "which", "makes", "model", "training", "hard", ",", "so", "we", "add", "a", "with", "stride", "2", "convolution", "to", "reduce", "such", "huge", "number", "of", "anchors", ".", "Followed", "FCIS", ",", "we", "use", "Joint", "Mask", "Prediction", "and", "Classification", "to", "simultaneously", "classify", "and", "mask", "the", "text", "instance", "on", "inside", "/", "outside", "score", "maps", "generated", "through", "PSROIPooling", "on", "conv", "-", "cls", "-", "seg", "feature", "maps", ",", "and", "box", "regression", "branch", "utilizes", "feature", "maps", "from", "conv", "-", "box", "after", "PSROIPooling", "(", "”", "”", "means", "one", "class", "is", "for", "text", "and", "the", "other", "for", "background", ")", ".", "We", "use", "shown", "in", "Fig.2", "in", "our", "experiments", "by", "default", ".", "It", "is", "noted", "that", "after", "PSROIPooling", ",", "the", "resolution", "of", "feature", "maps", "becomes", ".", "Therefore", ",", "we", "use", "global", "average", "pooling", "for", "classification", "(", "after", "pixel", "-", "wise", "max", ")", "and", "box", "regression", "branches", ",", "and", "pixel", "-", "wise", "softmax", "on", "mask", "branch", ".", "subsection", ":", "Ground", "Truth", "and", "Loss", "Function", "The", "whole", "multi", "-", "task", "loss", "can", "be", "interpreted", "as", "The", "full", "loss", "consists", "of", "two", "sub", "stage", "losses", ":", "RPN", "loss", "where", "is", "for", "region", "proposal", "classification", "and", "is", "for", "box", "regression", ",", "and", "text", "instance", "loss", "based", "on", "each", "ROI", ",", "where", "represent", "losses", "for", "instance", "classification", ",", "mask", "and", "box", "regression", "task", "respectively", ".", "is", "the", "hyper", "-", "parameter", "to", "control", "the", "balance", "among", "each", "loss", "term", ".", "They", "are", "set", "as", "in", "our", "experiments", ".", "Classification", "and", "mask", "task", "both", "use", "cross", "-", "entropy", "as", "loss", "function", ",", "whereas", "we", "use", "smooth", "-", "L1", "for", "box", "regression", "task", "formulated", "as", "is", "set", "to", "3", "in", "our", "experiments", "which", "makes", "the", "box", "regression", "loss", "less", "sensitive", "to", "outliers", ".", "Ground", "truth", "of", "each", "text", "instance", "is", "presented", "by", "bounding", "boxes", "and", "masks", "shown", "in", "Fig.3", ".", "In", "most", "multi", "-", "oriented", "text", "detection", "dataset", ",", "annotations", "are", "given", "in", "quadrilaterals", "such", "as", "IC15", "or", "can", "be", "converted", "to", "quadrilaterals", "such", "as", "TD500", ".", "For", "each", "instance", ",", "we", "directly", "generate", "mask", "from", "quadrilateral", "coordinates", "and", "use", "the", "minimal", "rectangle", "containing", "the", "mask", "as", "the", "bounding", "box", ".", "subsection", ":", "Post", "Processing", "Mask", "-", "NMS", "To", "obtain", "final", "detection", "results", ",", "we", "use", "Non", "-", "Maximum", "Suppression", "mechanism", "(", "NMS", ")", "to", "filter", "overlapped", "text", "instances", "and", "preserve", "those", "with", "highest", "scores", ".", "After", "NMS", ",", "we", "generate", "a", "minimum", "quadrilateral", "for", "each", "text", "instance", "covering", "the", "mask", "as", "shown", "in", "Fig.1", ".", "Standard", "NMS", "computes", "IOU", "among", "bounding", "boxes", ",", "which", "may", "be", "fine", "for", "word", "-", "level", "and", "near", "-", "horizontal", "results", "’", "filtering", ".", "However", ",", "it", "may", "filter", "some", "correct", "line", "-", "level", "detections", "when", "they", "are", "close", "and", "heavily", "inclined", "as", "shown", "in", "Fig.4", "or", "when", "words", "stay", "close", "in", "the", "same", "line", "as", "shown", "in", "Fig.5", ".", "Consequently", ",", "we", "propose", "a", "modified", "NMS", "called", "Mask", "-", "NMS", "to", "handle", "such", "situations", ".", "Mask", "-", "NMS", "mainly", "changes", "bounding", "box", "IOU", "computation", "to", "so", "-", "called", "mask", "-", "maximum", "-", "intersection", "(", "MMI", ")", "as", "formulated", ":", "are", "mask", "areas", "of", "two", "text", "instances", "to", "be", "computed", ",", "is", "the", "intersection", "area", "between", "the", "masks", ".", "Maximum", "intersection", "over", "the", "mask", "areas", "are", "used", "to", "replace", "original", "IOU", "for", "the", "reason", "that", "detections", "may", "easily", "involve", "line", "-", "level", "and", "word", "-", "level", "text", "instances", "simultaneously", "at", "the", "same", "line", "as", "shown", "in", "Fig.5", ".", "The", "proposed", "Mask", "-", "NMS", "has", "significantly", "improved", "performance", "for", "multi", "-", "oriented", "scene", "text", "detection", "as", "shown", "in", "section.5", ".", "section", ":", "Experiments", "To", "evaluate", "the", "proposed", "framework", ",", "we", "conduct", "quantitative", "experiments", "on", "three", "public", "benchmarks", ":", "ICDAR2015", ",", "MSRA", "-", "TD500", "and", "Total", "-", "Text", ".", "subsection", ":", "Datasets", "ICDAR", "2015", "Incidental", "Text", "(", "IC15", ")", "the", "Challenge", "4", "of", "ICDAR", "2015", "Robust", "Reading", "Competition", ".", "IC15", "contains", "1000", "training", "and", "500", "testing", "incidental", "images", "taken", "by", "Google", "Glasses", "without", "paying", "attention", "to", "viewpoint", "and", "image", "quality", ".", "Therefore", ",", "large", "variations", "in", "text", "scale", ",", "orientation", "and", "resolution", "lead", "to", "difficulty", "for", "text", "detection", ".", "Annotations", "of", "the", "dataset", "are", "given", "in", "word", "-", "level", "quadrilaterals", ".", "MSRA", "-", "TD500", "(", "TD500", ")", "is", "early", "presented", "in", ".", "The", "dataset", "is", "multi", "-", "oriented", "and", "multi", "-", "lingual", "including", "both", "Chinese", "and", "English", "text,", "which", "consists", "of", "300", "training", "and", "200", "testing", "images", ".", "Different", "from", "IC15", ",", "annotations", "of", "TD500", "are", "at", "line", "level", "which", "are", "rotated", "rectangles", ".", "Total", "-", "Text", "is", "presented", "in", "ICDAR2017", ".", "It", "consists", "of", "1555", "images", "with", "more", "than", "3", "different", "text", "orientations", ":", "Horizontal", ",", "Multi", "-", "Oriented", ",", "and", "Curved", ".", "SynthText", "in", "the", "Wild", "(", "SynthText", ")", "The", "dataset", "contains", "800", ",", "000", "synthetic", "images", ",", "text", "with", "random", "color", ",", "fonts", ",", "scale", "and", "orientation", "are", "rendered", "on", "natural", "images", "carefully", "to", "have", "a", "realistic", "look", ".", "Annotations", "are", "given", "in", "character", ",", "word", "and", "line", "level", ".", "subsection", ":", "Implementation", "Details", "Training", "We", "pretrain", "the", "proposed", "FTSN", "on", "a", "subset", "of", "SynthText", "containing", "160", ",", "000", "images", ",", "then", "finetune", "on", "IC15", ",", "TD500", "and", "Total", "-", "Text", ".", "For", "optimization", ",", "standard", "SGD", "is", "used", "during", "training", "with", "learning", "rate", "for", "first", "5", "epochs", "and", "for", "the", "last", "epoch", ",", "and", "we", "also", "apply", "online", "hard", "example", "mining", "(", "OHEM", ")", "for", "balancing", "the", "positive", "and", "negative", "samples", ".", "Different", "from", "original", "RPN", "anchor", "ratios", "and", "scales", "setting", "for", "object", "detection", ",", "anchor", "scales", "of", "[", "]", "and", "ratios", "of", "[", "1", "/", "3", ",", "1", "/", "2", ",", "1", ",", "2", ",", "3", ",", "5", ",", "7", "]", "are", "set", "because", "text", "often", "has", "a", "large", "aspect", "ratio", "and", "a", "small", "scale", ".", "Data", "augmentation", "Multi", "-", "scale", "training", ",", "rotation", "and", "color", "jittering", "are", "applied", "during", "training", ".", "Scales", "are", "randomly", "chosen", "from", "[", "600", ",", "720", ",", "960", ",", "1100", "]", "and", "each", "number", "represents", "the", "short", "edge", "of", "input", "images", ".", "Rotation", "with", ",", "and", "are", "applied", "with", "horizontal", "flip", ".", "Consequently", ",", "it", "enlarges", "8x", "dataset", "size", "than", "the", "original", "one", ".", "Random", "brightness", ",", "contrast", "and", "saturation", "jittering", "are", "applied", "for", "input", "images", ".", "Testing", "Input", "images", "are", "resized", "to", "when", "testing", ".", "After", "NMS", ",", "mask", "voting", "is", "used", "to", "obtain", "an", "ensemble", "text", "instance", "mask", "by", "averaging", "all", "reasonable", "detections", ".", "Experiments", "are", "conducted", "on", "MXNet", "and", "run", "on", "a", "server", "with", "Intel", "i7", "6700", "K", "CPU", ",", "64", "GB", "RAM", ",", "GTX", "1080", "and", "Ubuntu", "14.04", "OS", ".", "subsection", ":", "Results", "Tabel.1", "shows", "results", "of", "the", "proposed", "FTSN", "on", "IC15", "compared", "with", "previous", "state", "of", "art", "published", "methods", ".", "SNMS", "and", "MNMS", "represent", "standard", "NMS", "and", "Mask", "-", "NMS", "respectively", ".", "Our", "FTSN", "with", "Mask", "-", "NMS", "outperforms", "former", "best", "result", "by", "5.3", "%", "in", "Precision", "and", "3.1", "%", "in", "Hmean", ".", "It", "is", "evaluated", "by", "the", "official", "submission", "server", ".", "Results", "on", "TD500", "are", "shown", "in", "Table.2", "along", "with", "other", "state", "of", "art", "methods", ".", "It", "is", "shown", "that", "our", "methods", "outperform", "the", "current", "state", "of", "art", "approaches", "by", "a", "large", "margin", "in", "Hmean", "and", "Recall", ",", "without", "adding", "extra", "real", "-", "world", "training", "images", ".", "Our", "method", "also", "shows", "great", "flexibility", "on", "the", "total", "-", "text", "dataset", "containing", "curved", "text", ".", "As", "the", "dataset", "is", "new", "to", "the", "community", ",", "experiments", "are", "seldom", "conducted", "on", "it", "which", "makes", "our", "results", "as", "a", "baseline", "shown", "in", "Table3", ".", "The", "evaluation", "metric", "uses", "IoU", "of", "0.5", "between", "each", "instance", "masks", ".", "Outperforming", "the", "current", "state", "of", "the", "art", ",", "our", "approach", "runs", "about", "4", "FPS", "on", "images", "and", "2.5", "FPS", "when", "using", "Mask", "-", "NMS", ",", "which", "presents", "efficiency", "and", "accuracy", ".", "It", "is", "noted", "that", "the", "proposed", "Mask", "-", "NMS", "significantly", "improved", "Hmean", "by", "0.7", "and", "0.3", "percent", "on", "IC15", "and", "TD500", ",", "which", "mainly", "target", "the", "situations", "in", "Fig.4", "and", "Fig.5", ".", "Fig.6", "shows", "example", "results", "of", "FTSN", ".", "From", "left", "to", "right", ",", "it", "illustrates", "results", "on", "IC15", ",", "TD500", "and", "Total", "-", "Text", "dataset", ".", "The", "decent", "performance", "for", "word", "-", "level", ",", "line", "-", "level", "and", "curved", "text", "detection", "with", "large", "variation", "in", "resolution", ",", "view", "point", ",", "scale", "and", "linguistics", "suggests", "excellent", "generalization", "ability", ".", "section", ":", "Conclusion", "We", "present", "FTSN", ",", "an", "end", "-", "end", "efficient", "and", "accurate", "multi", "-", "oriented", "scene", "text", "detection", "framework", ".", "It", "has", "outperformed", "previous", "state", "of", "the", "art", "approaches", "on", "word", "-", "level", "line", "-", "level", "annotated", "benchmarks", "and", "report", "a", "baseline", "on", "total", "-", "text", "demonstrating", "decent", "generalization", "ability", "and", "flexibility", ".", "section", ":", "Acknowlegements", "The", "research", "is", "supported", "by", "The", "National", "Key", "Research", "and", "Development", "Program", "of", "China", "under", "grant", "2017YFB1002401", ".", "bibliography", ":", "References"], "entity": [[2, 6, ["Fused", "Text", "Segmentation", "Networks"], "Method"], [7, 13, ["Multi", "-", "oriented", "Scene", "Text", "Detection"], "Task"], [21, 25, ["end", "-", "end", "framework"], "Method"], [34, 40, ["instance", "-", "aware", "semantic", "segmentation", "perspective"], "Task"], [43, 47, ["Fused", "Text", "Segmentation", "Networks"], "Method"], [56, 58, ["feature", "extracting"], "Method"], [87, 90, ["semantic", "segmentation", "task"], "Task"], [91, 97, ["region", "proposal", "based", "object", "detection", "task"], "Task"], [122, 123, ["ICDAR2015"], "Material"], [123, 126, ["Incidental", "Scene", "Text"], "Material"], [127, 130, ["MSRA", "-", "TD500"], "Material"], [131, 132, ["Hmean"], "Method"], [165, 168, ["scene", "text", "detection"], "Task"], [173, 175, ["computer", "vision"], "Task"], [176, 179, ["machine", "learning", "community"], "Task"], [183, 188, ["content", "-", "based", "image", "applications"], "Task"], [190, 192, ["photo", "translation"], "Task"], [193, 196, ["receipt", "content", "recognition"], "Task"], [212, 214, ["Detecting", "text"], "Task"], [246, 249, ["text", "detection", "methods"], "Method"], [261, 263, ["text", "detection"], "Task"], [300, 304, ["deep", "learning", "based", "approaches"], "Method"], [323, 326, ["ICDAR", "series", "contests"], "Material"], [334, 341, ["detecting", "horizontal", "or", "near", "-", "horizontal", "text"], "Task"], [343, 346, ["horizontal", "text", "detection"], "Task"], [383, 384, ["accuracies"], "Metric"], [385, 389, ["ICDAR", "2015", "Competition", "Challenge"], "Task"], [390, 394, ["“Incidental", "scene", "text", "localization”"], "Task"], [413, 418, ["multi", "-", "oriented", "text", "detection"], "Task"], [431, 434, ["Region", "based", "methods"], "Method"], [436, 439, ["object", "detection", "techniques"], "Method"], [442, 443, ["RCNN"], "Method"], [444, 445, ["SSD"], "Method"], [446, 450, ["Segmentation", "-", "based", "methods"], "Method"], [452, 456, ["fully", "convolutional", "neural", "networks"], "Method"], [457, 458, ["FCN"], "Method"], [461, 464, ["text", "score", "maps"], "Task"], [475, 476, ["detections"], "Task"], [477, 481, ["Direct", "regression", "based", "method"], "Method"], [499, 502, ["text", "scores", "map"], "Method"], [503, 509, ["rotated", "/", "quadrangle", "bounding", "boxes", "generation"], "Method"], [518, 523, ["multi", "-", "oriented", "text", "detection"], "Task"], [529, 534, ["instance", "-", "aware", "semantic", "segmentation"], "Task"], [545, 550, ["multi", "-", "oriented", "text", "detection"], "Task"], [561, 565, ["region", "proposal", "based", "methods"], "Method"], [567, 571, ["flexible", "segmentation", "based", "methods"], "Method"], [588, 591, ["end", "trainable", "framework"], "Method"], [603, 608, ["text", "/", "nontext", "salient", "map"], "Task"], [609, 613, ["text", "-", "line", "generation"], "Task"], [616, 619, ["region", "proposal", "network"], "Method"], [620, 621, ["RPN"], "Method"], [634, 638, ["non", "-", "maximum", "suppression"], "Method"], [639, 640, ["NMS"], "Method"], [666, 668, ["detection", "process"], "Task"], [685, 687, ["trainable", "solution"], "Method"], [688, 693, ["multi", "-", "oriented", "text", "detection"], "Task"], [695, 699, ["instance", "aware", "segmentation", "perspective"], "Task"], [702, 704, ["redundant", "pipelines"], "Method"], [706, 708, ["feature", "extraction"], "Task"], [727, 730, ["Mask", "-", "NMS"], "Method"], [736, 737, ["NMS"], "Method"], [763, 769, ["multi", "-", "oriented", "text", "detection", "benchmarks"], "Task"], [774, 776, ["Detecting", "text"], "Task"], [791, 798, ["text", "-", "related", "real", "-", "world", "applications"], "Task"], [800, 802, ["photo", "OCR"], "Task"], [803, 805, ["blind", "navigation"], "Task"], [813, 816, ["scene", "text", "detection"], "Task"], [817, 819, ["Connected", "Components"], "Method"], [820, 821, ["CCs"], "Method"], [842, 845, ["stroke", "width", "transform"], "Method"], [846, 847, ["SWT"], "Method"], [854, 855, ["MSER"], "Method"], [877, 880, ["bottom", "up", "approaches"], "Method"], [905, 908, ["sliding", "window", "based"], "Method"], [941, 943, ["sliding", "window"], "Method"], [946, 948, ["computational", "cost"], "Metric"], [989, 991, ["generalization", "ability"], "Metric"], [1004, 1008, ["deep", "learning", "based", "approaches"], "Method"], [1009, 1011, ["object", "detection"], "Task"], [1012, 1014, ["semantic", "segmentation"], "Task"], [1019, 1021, ["reading", "text"], "Task"], [1034, 1037, ["general", "object", "detection"], "Task"], [1043, 1046, ["object", "detection", "frameworks"], "Method"], [1048, 1050, ["Faster", "RCNN"], "Method"], [1051, 1052, ["SSD"], "Method"], [1064, 1067, ["region", "proposal", "network"], "Method"], [1130, 1133, ["FCN", "based", "approaches"], "Method"], [1134, 1140, ["text", "/", "non", "-", "text", "map"], "Task"], [1167, 1168, ["pipelines"], "Method"], [1178, 1183, ["instance", "-", "aware", "semantic", "segmentation"], "Task"], [1187, 1192, ["end", "-", "end", "trainable", "framework"], "Method"], [1193, 1197, ["Fused", "Text", "Segmentation", "Networks"], "Method"], [1198, 1199, ["FTSN"], "Method"], [1202, 1207, ["arbitrary", "-", "shape", "text", "detection"], "Task"], [1218, 1220, ["object", "detection"], "Method"], [1221, 1224, ["semantic", "segmentation", "architecture"], "Method"], [1248, 1250, ["feature", "representation"], "Method"], [1256, 1261, ["multi", "-", "level", "feature", "maps"], "Method"], [1289, 1292, ["deep", "CNN", "model"], "Method"], [1299, 1301, ["Feature", "representations"], "Method"], [1307, 1311, ["resnet", "-", "101", "backbone"], "Method"], [1321, 1322, ["FusedMapA"], "Method"], [1327, 1330, ["region", "proposed", "network"], "Method"], [1331, 1332, ["RPN"], "Method"], [1334, 1338, ["text", "region", "of", "interest"], "Task"], [1339, 1342, ["ROI", ")", "generation"], "Task"], [1346, 1349, ["rois", "’", "PSROIPooling"], "Task"], [1357, 1364, ["detection", ",", "segmentation", "and", "box", "regression", "branches"], "Task"], [1379, 1383, ["post", "-", "processing", "part"], "Task"], [1384, 1385, ["NMS"], "Method"], [1386, 1389, ["minimal", "quadrilateral", "generation"], "Task"], [1392, 1394, ["Network", "Architecture"], "Method"], [1395, 1398, ["convolutional", "feature", "representation"], "Method"], [1402, 1404, ["fusion", "fashion"], "Method"], [1453, 1456, ["resnet", "-", "101"], "Method"], [1462, 1464, ["region", "proposing"], "Task"], [1477, 1481, ["element", "-", "wise", "adding"], "Method"], [1491, 1492, ["FusedMapA"], "Method"], [1494, 1495, ["FusedMapB."], "Method"], [1499, 1500, ["downsampling"], "Task"], [1535, 1537, ["segmentation", "task"], "Task"], [1564, 1565, ["RPN"], "Method"], [1567, 1569, ["model", "training"], "Task"], [1576, 1579, ["stride", "2", "convolution"], "Method"], [1588, 1589, ["FCIS"], "Method"], [1592, 1595, ["Joint", "Mask", "Prediction"], "Method"], [1596, 1597, ["Classification"], "Method"], [1613, 1614, ["PSROIPooling"], "Method"], [1615, 1622, ["conv", "-", "cls", "-", "seg", "feature", "maps"], "Method"], [1624, 1627, ["box", "regression", "branch"], "Method"], [1635, 1636, ["PSROIPooling"], "Method"], [1668, 1669, ["PSROIPooling"], "Method"], [1681, 1684, ["global", "average", "pooling"], "Method"], [1685, 1686, ["classification"], "Task"], [1709, 1711, ["Ground", "Truth"], "Metric"], [1712, 1714, ["Loss", "Function"], "Metric"], [1716, 1720, ["multi", "-", "task", "loss"], "Task"], [1730, 1733, ["sub", "stage", "losses"], "Method"], [1734, 1736, ["RPN", "loss"], "Method"], [1739, 1742, ["region", "proposal", "classification"], "Task"], [1745, 1747, ["box", "regression"], "Method"], [1749, 1752, ["text", "instance", "loss"], "Method"], [1761, 1763, ["instance", "classification"], "Task"], [1764, 1769, ["mask", "and", "box", "regression", "task"], "Task"], [1793, 1794, ["Classification"], "Task"], [1795, 1797, ["mask", "task"], "Task"], [1799, 1802, ["cross", "-", "entropy"], "Metric"], [1803, 1805, ["loss", "function"], "Metric"], [1809, 1812, ["smooth", "-", "L1"], "Method"], [1813, 1816, ["box", "regression", "task"], "Task"], [1828, 1831, ["box", "regression", "loss"], "Method"], [1855, 1861, ["multi", "-", "oriented", "text", "detection", "dataset"], "Task"], [1869, 1870, ["IC15"], "Material"], [1878, 1879, ["TD500"], "Material"], [1906, 1908, ["Post", "Processing"], "Task"], [1908, 1911, ["Mask", "-", "NMS"], "Method"], [1914, 1915, ["detection"], "Task"], [1919, 1924, ["Non", "-", "Maximum", "Suppression", "mechanism"], "Method"], [1925, 1926, ["NMS"], "Method"], [1940, 1941, ["NMS"], "Method"], [1960, 1961, ["NMS"], "Method"], [1962, 1963, ["IOU"], "Method"], [1972, 1982, ["word", "-", "level", "and", "near", "-", "horizontal", "results", "’", "filtering"], "Task"], [2025, 2026, ["NMS"], "Method"], [2027, 2030, ["Mask", "-", "NMS"], "Method"], [2035, 2038, ["Mask", "-", "NMS"], "Method"], [2040, 2044, ["bounding", "box", "IOU", "computation"], "Method"], [2048, 2053, ["mask", "-", "maximum", "-", "intersection"], "Method"], [2054, 2055, ["MMI"], "Method"], [2119, 2122, ["Mask", "-", "NMS"], "Method"], [2156, 2157, ["ICDAR2015"], "Material"], [2158, 2161, ["MSRA", "-", "TD500"], "Material"], [2162, 2165, ["Total", "-", "Text"], "Material"], [2169, 2171, ["ICDAR", "2015"], "Material"], [2171, 2173, ["Incidental", "Text"], "Material"], [2174, 2175, ["IC15"], "Material"], [2186, 2187, ["IC15"], "Material"], [2197, 2199, ["Google", "Glasses"], "Material"], [2205, 2207, ["image", "quality"], "Metric"], [2223, 2225, ["text", "detection"], "Task"], [2238, 2241, ["MSRA", "-", "TD500"], "Material"], [2242, 2243, ["TD500"], "Material"], [2261, 2264, ["Chinese", "and", "English"], "Material"], [2277, 2278, ["IC15"], "Material"], [2291, 2294, ["Total", "-", "Text"], "Material"], [2297, 2298, ["ICDAR2017"], "Material"], [2321, 2322, ["SynthText"], "Material"], [2326, 2327, ["SynthText"], "Material"], [2379, 2380, ["FTSN"], "Method"], [2384, 2385, ["SynthText"], "Material"], [2394, 2395, ["IC15"], "Material"], [2396, 2397, ["TD500"], "Material"], [2398, 2401, ["Total", "-", "Text"], "Material"], [2403, 2404, ["optimization"], "Task"], [2406, 2407, ["SGD"], "Method"], [2412, 2414, ["learning", "rate"], "Metric"], [2428, 2432, ["online", "hard", "example", "mining"], "Method"], [2433, 2435, ["OHEM", ")"], "Method"], [2446, 2449, ["RPN", "anchor", "ratios"], "Method"], [2453, 2455, ["object", "detection"], "Task"], [2498, 2504, ["Data", "augmentation", "Multi", "-", "scale", "training"], "Task"], [2584, 2585, ["NMS"], "Method"], [2586, 2588, ["mask", "voting"], "Method"], [2607, 2608, ["MXNet"], "Method"], [2640, 2641, ["FTSN"], "Method"], [2642, 2643, ["IC15"], "Material"], [2652, 2653, ["SNMS"], "Method"], [2654, 2655, ["MNMS"], "Method"], [2657, 2662, ["NMS", "and", "Mask", "-", "NMS"], "Method"], [2665, 2666, ["FTSN"], "Method"], [2667, 2670, ["Mask", "-", "NMS"], "Method"], [2678, 2679, ["Precision"], "Metric"], [2683, 2684, ["Hmean"], "Method"], [2696, 2697, ["TD500"], "Material"], [2725, 2726, ["margin"], "Metric"], [2727, 2730, ["Hmean", "and", "Recall"], "Metric"], [2783, 2785, ["evaluation", "metric"], "Metric"], [2807, 2808, ["FPS"], "Metric"], [2812, 2813, ["FPS"], "Metric"], [2815, 2818, ["Mask", "-", "NMS"], "Method"], [2823, 2824, ["accuracy"], "Metric"], [2831, 2834, ["Mask", "-", "NMS"], "Method"], [2836, 2837, ["Hmean"], "Method"], [2843, 2844, ["IC15"], "Material"], [2845, 2846, ["TD500"], "Material"], [2862, 2863, ["FTSN"], "Method"], [2873, 2874, ["IC15"], "Material"], [2875, 2876, ["TD500"], "Material"], [2877, 2881, ["Total", "-", "Text", "dataset"], "Material"], [2886, 2897, ["word", "-", "level", ",", "line", "-", "level", "and", "curved", "text", "detection"], "Task"], [2901, 2902, ["resolution"], "Metric"], [2911, 2913, ["generalization", "ability"], "Metric"], [2919, 2920, ["FTSN"], "Method"], [2964, 2966, ["generalization", "ability"], "Metric"], [29, 32, ["scene", "text", "detection"], "Task"], [117, 120, ["scene", "text", "detection"], "Task"], [397, 400, ["scene", "text", "detection"], "Task"], [1103, 1106, ["scene", "text", "detection"], "Task"], [1278, 1281, ["scene", "text", "detection"], "Task"], [2130, 2133, ["scene", "text", "detection"], "Task"], [2180, 2182, ["ICDAR", "2015"], "Material"], [2281, 2282, ["TD500"], "Material"], [2931, 2934, ["scene", "text", "detection"], "Task"]]} {"doc_id": "25b9b1ede6519eeef3e67657a97e2f551940b0d6", "text": ["document", ":", "Segmental", "Recurrent", "Neural", "Networks", "for", "End", "-", "to", "-", "end", "Speech", "Recognition", "We", "study", "the", "segmental", "recurrent", "neural", "network", "for", "end", "-", "to", "-", "end", "acoustic", "modelling", ".", "This", "model", "connects", "the", "segmental", "conditional", "random", "field", "(", "CRF", ")", "with", "a", "recurrent", "neural", "network", "(", "RNN", ")", "used", "for", "feature", "extraction", ".", "Compared", "to", "most", "previous", "CRF", "-", "based", "acoustic", "models", ",", "it", "does", "not", "rely", "on", "an", "external", "system", "to", "provide", "features", "or", "segmentation", "boundaries", ".", "Instead", ",", "this", "model", "marginalises", "out", "all", "the", "possible", "segmentations", ",", "and", "features", "are", "extracted", "from", "the", "RNN", "trained", "together", "with", "the", "segmental", "CRF", ".", "Essentially", ",", "this", "model", "is", "self", "-", "contained", "and", "can", "be", "trained", "end", "-", "to", "-", "end", ".", "In", "this", "paper", ",", "we", "discuss", "practical", "training", "and", "decoding", "issues", "as", "well", "as", "the", "method", "to", "speed", "up", "the", "training", "in", "the", "context", "of", "speech", "recognition", ".", "We", "performed", "experiments", "on", "the", "TIMIT", "dataset", ".", "We", "achieved", "17.3", "%", "phone", "error", "rate", "(", "PER", ")", "from", "the", "first", "-", "pass", "decoding", "—", "the", "best", "reported", "result", "using", "CRFs", ",", "despite", "the", "fact", "that", "we", "only", "used", "a", "zeroth", "-", "order", "CRF", "and", "without", "using", "any", "language", "model", ".", "CentreforSpeechTechnologyResearch", ",", "TheUniversityofEdinburgh", ",", "Edinburgh", ",", "UK", "SchoolofComputerScience", ",", "CarnegieMellonUniversity", ",", "Pittsburgh", ",", "USA", "ComputerScience", "&", "Engineering", ",", "TheUniversityofWashington", ",", "Seattle", ",", "USA", "{", "liang.lu", ",", "s.renals}@ed.ac.uk", ",", "{lingpenk", ",", "cdyer}@cs.cmu.edu", ",", "nasmith@cs.washington.edu", "Index", "Terms", ":", "end", "-", "to", "-", "end", "speech", "recognition", ",", "segmental", "CRF", ",", "recurrent", "neural", "networks", ".", "section", ":", "Introduction", "Speech", "recognition", "is", "a", "typical", "sequence", "to", "sequence", "transduction", "problem", ",", "i.e.", ",", "given", "a", "sequence", "of", "acoustic", "observations", ",", "the", "speech", "recognition", "engine", "decodes", "the", "corresponding", "sequence", "of", "words", "or", "phonemes", ".", "A", "key", "component", "in", "a", "speech", "recognition", "system", "is", "the", "acoustic", "model", ",", "which", "computes", "the", "conditional", "probability", "of", "the", "output", "sequence", "given", "the", "input", "sequence", ".", "However", ",", "directly", "computing", "this", "conditional", "probability", "is", "challenging", "due", "to", "many", "factors", "including", "the", "variable", "lengths", "of", "the", "input", "and", "output", "sequences", ".", "The", "hidden", "Markov", "model", "(", "HMM", ")", "converts", "this", "sequence", "-", "level", "classification", "task", "into", "a", "frame", "-", "level", "classification", "problem", ",", "where", "each", "acoustic", "frame", "is", "classified", "into", "one", "of", "the", "hidden", "states", ",", "and", "each", "output", "sequence", "corresponds", "to", "a", "sequence", "of", "hidden", "states", ".", "To", "make", "it", "computationally", "tractable", ",", "HMMs", "usually", "rely", "on", "the", "conditional", "independence", "assumption", "and", "the", "first", "-", "order", "Markov", "rule", "—", "the", "well", "-", "known", "weaknesses", "of", "HMMs", ".", "Furthermore", ",", "the", "HMM", "-", "based", "pipeline", "is", "composed", "of", "a", "few", "relatively", "independent", "modules", ",", "which", "makes", "the", "joint", "optimisation", "nontrivial", ".", "There", "has", "been", "a", "consistent", "research", "effort", "to", "seek", "architectures", "to", "replace", "HMMs", "and", "overcome", "their", "limitation", "for", "acoustic", "modelling", ",", "e.g.", ",", ";", "however", "these", "approaches", "have", "not", "yet", "improved", "speech", "recognition", "accuracy", "over", "HMMs", ".", "In", "the", "past", "few", "years", ",", "several", "neural", "network", "based", "approaches", "have", "been", "proposed", "and", "demonstrated", "promising", "results", ".", "In", "particular", ",", "the", "connectionist", "temporal", "classification", "(", "CTC", ")", "approach", "defines", "the", "loss", "function", "directly", "to", "maximise", "the", "conditional", "probability", "of", "the", "output", "sequence", "given", "the", "input", "sequence", ",", "and", "it", "usually", "uses", "a", "recurrent", "neural", "network", "to", "extract", "features", ".", "However", ",", "CTC", "simplifies", "the", "sequence", "-", "level", "error", "function", "by", "a", "product", "of", "the", "frame", "-", "level", "error", "functions", "(", "i.e.", ",", "independence", "assumption", ")", ",", "which", "means", "it", "essentially", "still", "does", "frame", "-", "level", "classification", ".", "It", "also", "requires", "the", "lengths", "of", "the", "input", "and", "output", "sequence", "to", "be", "the", "same", ",", "which", "is", "inappropriate", "for", "speech", "recognition", ".", "CTC", "deals", "with", "this", "problem", "by", "replicating", "the", "output", "labels", "so", "that", "a", "consecutive", "frames", "may", "correspond", "to", "the", "same", "output", "label", "or", "a", "blank", "token", ".", "Attention", "-", "based", "RNNs", "have", "been", "demonstrated", "to", "be", "a", "powerful", "alternative", "sequence", "-", "to", "-", "sequence", "transducer", ",", "e.g.", ",", "in", "machine", "translation", ",", "and", "speech", "recognition", ".", "A", "key", "difference", "of", "this", "model", "from", "HMMs", "and", "CTCs", "is", "that", "the", "attention", "-", "based", "approach", "does", "not", "apply", "the", "conditional", "independence", "assumption", "to", "the", "input", "sequence", ".", "Instead", ",", "it", "maps", "the", "variable", "-", "length", "input", "sequence", "into", "a", "fixed", "-", "size", "vector", "representation", "at", "each", "decoding", "step", "by", "an", "attention", "-", "based", "scheme", "(", "see", "for", "further", "explanation", ")", ".", "It", "then", "generates", "the", "output", "sequence", "using", "an", "RNN", "conditioned", "on", "the", "vector", "representation", "from", "the", "source", "sequence", ".", "The", "attentive", "scheme", "suits", "the", "machine", "translation", "task", "well", ",", "because", "there", "may", "be", "no", "clear", "alignment", "between", "the", "source", "and", "target", "sequence", "for", "many", "language", "pairs", ".", "However", ",", "this", "approach", "does", "not", "naturally", "apply", "to", "the", "speech", "recognition", "task", ",", "as", "each", "output", "token", "only", "corresponds", "to", "a", "small", "size", "window", "of", "acoustic", "spectrum", ".", "In", "this", "paper", ",", "we", "study", "segmental", "RNNs", "for", "acoustic", "modelling", ".", "This", "model", "is", "similar", "to", "CTC", "and", "attention", "-", "based", "RNN", "in", "the", "sense", "that", "an", "RNN", "encoder", "is", "also", "used", "for", "feature", "extraction", ",", "but", "it", "differs", "in", "the", "sense", "that", "the", "sequence", "-", "level", "conditional", "probability", "is", "defined", "using", "an", "segmental", "(", "semi", "-", "Markov", ")", "CRF", ",", "which", "is", "an", "extension", "on", "the", "standard", "CRF", ".", "There", "have", "been", "numerous", "works", "on", "CRFs", "and", "their", "variants", "for", "speech", "recognition", ",", "e.g", ",", "(", "see", "for", "an", "overview", ")", ".", "In", "particular", ",", "feed", "-", "forward", "neural", "networks", "have", "been", "used", "with", "segmental", "CRFs", "for", "speech", "recognition", ".", "However", ",", "segmental", "RNNs", "are", "different", "in", "that", "they", "are", "end", "-", "to", "-", "end", "models", "—", "they", "do", "not", "depend", "on", "external", "systems", "to", "provide", "segmentation", "boundaries", "and", "features", ",", "instead", ",", "they", "are", "trained", "by", "marginalising", "out", "all", "possible", "segmentations", ",", "while", "the", "features", "are", "derived", "from", "the", "encoder", "RNNs", ",", "which", "are", "trained", "jointly", "with", "the", "segmental", "CRFs", ".", "Our", "experiments", "were", "performed", "on", "the", "TIMIT", "dataset", ",", "and", "we", "achieved", "17.3", "%", "PER", "from", "first", "-", "pass", "decoding", "with", "zeroth", "-", "order", "CRF", "and", "without", "using", "any", "language", "model", "—", "the", "best", "reported", "result", "using", "CRFs", ".", "section", ":", "Segmental", "Recurrent", "Neural", "Networks", "subsection", ":", "Segmental", "Conditional", "Random", "Fields", "Given", "a", "sequence", "of", "acoustic", "frames", "and", "its", "corresponding", "sequence", "of", "output", "labels", ",", "where", ",", "segmental", "(", "or", "semi", "-", "Markov", ")", "conditional", "random", "field", "defines", "the", "sequence", "-", "level", "conditional", "probability", "with", "the", "auxiliary", "segment", "labels", "as", "where", "is", "a", "tuple", "of", "the", "beginning", "(", ")", "and", "the", "end", "(", ")", "time", "tag", "for", "the", "segment", "of", ",", "and", "while", ";", "and", "denotes", "the", "vocabulary", "set", ";", "is", "the", "normaliser", "that", "that", "sums", "over", "all", "the", "possible", "pairs", ",", "i.e.", ",", "Here", ",", "we", "only", "consider", "the", "zeroth", "-", "order", "CRF", ",", "while", "the", "extension", "to", "higher", "order", "models", "is", "straightforward", ".", "Similar", "to", "other", "CRF", "-", "based", "models", ",", "the", "function", "is", "defined", "as", "where", "denotes", "the", "feature", "function", ",", "and", "is", "the", "weight", "vector", ".", "Previous", "works", "on", "CRF", "-", "based", "acoustic", "models", "mainly", "use", "heuristically", "handcrafted", "feature", "function", ".", "They", "also", "usually", "rely", "on", "an", "external", "system", "to", "provide", "the", "segment", "labels", ".", "In", "this", "paper", ",", "we", "define", "using", "neural", "networks", ",", "and", "the", "segmentation", "is", "marginalised", "out", "during", "training", ",", "which", "makes", "our", "model", "self", "-", "contained", ".", "subsection", ":", "Feature", "Representations", "We", "use", "neural", "networks", "to", "define", "the", "feature", "function", ",", "which", "maps", "the", "acoustic", "segment", "and", "its", "corresponding", "label", "into", "a", "joint", "feature", "space", ".", "More", "specifically", ",", "is", "firstly", "represented", "as", "a", "one", "-", "hot", "vector", ",", "and", "it", "is", "then", "mapped", "into", "a", "continuous", "space", "by", "a", "linear", "embedding", "matrix", "as", "Given", "the", "segment", "label", ",", "we", "use", "an", "RNN", "to", "map", "the", "acoustic", "segment", "to", "a", "fixed", "-", "dimensional", "vector", "representation", ",", "i.e.", ",", "where", "denotes", "the", "initial", "hidden", "state", ",", "denotes", "the", "duration", "of", "the", "segment", "and", "is", "a", "non", "-", "linear", "function", ".", "We", "take", "the", "final", "hidden", "state", "as", "the", "segment", "embedding", "vector", ",", "then", "can", "be", "represented", "as", "where", "corresponds", "to", "one", "layer", "or", "multiple", "layers", "of", "linear", "or", "non", "-", "linear", "transformation", ".", "In", "fact", ",", "it", "is", "flexible", "to", "include", "other", "relevant", "features", "as", "additional", "inputs", "to", "the", "function", ",", "e.g.", ",", "the", "duration", "feature", "which", "can", "be", "obtained", "by", "converting", "into", "another", "embedding", "vector", ".", "In", "practice", ",", "multiple", "RNN", "layers", "can", "be", "used", "transform", "the", "acoustic", "signal", "before", "extracting", "the", "segment", "embedding", "vector", "as", "Figure", "[", "reference", "]", ".", "subsection", ":", "Conditional", "Maximum", "Likelihood", "Training", "For", "speech", "recognition", ",", "the", "segmentation", "labels", "are", "usually", "unknown", ",", "training", "the", "model", "by", "maximising", "the", "conditional", "probability", "as", "Eq", ".", "(", "[", "reference", "]", ")", "is", "therefore", "not", "practical", ".", "The", "problem", "can", "be", "addressed", "by", "defining", "the", "loss", "function", "as", "the", "negative", "marginal", "log", "-", "likelihood", "as", "where", "denotes", "the", "set", "of", "model", "parameters", ",", "and", "denotes", "the", "summation", "over", "all", "the", "possible", "segmentations", "when", "only", "is", "observed", ".", "To", "simplify", "notations", ",", "the", "objective", "function", "is", "define", "with", "only", "one", "training", "utterance", ".", "However", ",", "the", "number", "of", "possible", "segmentations", "is", "exponential", "with", "the", "length", "of", ",", "which", "makes", "the", "naive", "computation", "of", "both", "and", "impractical", ".", "Fortunately", ",", "this", "can", "be", "addressed", "by", "using", "the", "following", "dynamic", "programming", "algorithm", "as", "proposed", "in", ":", "In", "Eq", ".", "(", "[", "reference", "]", ")", ",", "the", "first", "summation", "is", "over", "all", "the", "possible", "segmentation", "up", "to", "timestep", ",", "and", "the", "second", "summation", "is", "over", "all", "the", "possible", "labels", "from", "the", "vocabulary", ".", "The", "computation", "cost", "of", "this", "algorithm", "is", ",", "where", "is", "the", "size", "of", "the", "vocabulary", ".", "The", "cost", "can", "be", "further", "reduced", "by", "introducing", "an", "upper", "bound", "of", "the", "segment", "length", ",", "in", "which", "case", "Eq", ".", "(", "[", "reference", "]", ")", "can", "be", "rewritten", "as", "where", "denotes", "the", "maximum", "value", "of", "the", "segment", "length", ".", "The", "cost", "is", "then", "reduced", "to", ",", "and", "for", "long", "sequences", "like", "speech", "signals", "where", ",", "the", "computational", "savings", "are", "substantial", ".", "The", "term", "can", "be", "computed", "similarly", ".", "In", "this", "case", ",", "since", "the", "label", "is", "now", "observed", ",", "the", "summation", "over", "all", "the", "possible", "labels", "in", "Eq", ".", "(", "[", "reference", "]", ")", "is", "not", "necessary", ",", "i.e.", ",", "Again", ",", "we", "can", "limit", "the", "length", "of", "the", "possible", "segments", "as", "Eq", ".", "(", "[", "reference", "]", ")", ".", "Given", "and", ",", "the", "loss", "function", "can", "be", "minimised", "using", "the", "stochastic", "gradient", "decent", "(", "SGD", ")", "algorithm", "similar", "to", "training", "other", "neural", "network", "models", ".", "Other", "losses", ",", "for", "example", ",", "hinge", ",", "can", "be", "considered", "in", "future", "work", ".", "subsection", ":", "Viterbi", "Decoding", "During", "decoding", ",", "we", "need", "to", "search", "the", "target", "label", "sequence", "that", "yields", "the", "highest", "posterior", "probability", "given", "by", "marginalising", "out", "all", "the", "possible", "segmentations", ":", "This", "involves", "minor", "modification", "of", "the", "recursive", "algorithm", "in", "Eq", ".", "(", "[", "reference", "]", ")", "that", "instead", "of", "summing", "over", "all", "the", "possible", "labels", ",", "the", "Viterbi", "path", "up", "to", "the", "timestep", "is", "However", ",", "marginalising", "out", "all", "the", "possible", "segmentations", "is", "still", "expensive", ".", "The", "computational", "cost", "can", "be", "further", "reduced", "by", "greedy", "searching", "the", "most", "likely", "segmentation", ",", "i.e.", ",", "which", "corresponds", "to", "the", "decoding", "objective", "as", "This", "joint", "maximization", "algorithm", "may", "yield", "high", "search", "error", ",", "because", "it", "only", "considers", "one", "segmentation", ".", "In", "the", "future", ",", "we", "shall", "investigate", "the", "beam", "search", "algorithm", "which", "may", "yield", "a", "lower", "search", "error", ".", "subsection", ":", "Further", "Speedup", "It", "is", "computationally", "expensive", "for", "RNNs", "to", "model", "long", "sequences", ",", "and", "the", "number", "of", "possible", "segmentations", "is", "exponential", "with", "the", "length", "of", "the", "input", "sequence", "as", "mentioned", "before", ".", "The", "computational", "cost", "can", "be", "significantly", "reduced", "by", "using", "the", "hierarchical", "subsampling", "RNN", "to", "shorten", "the", "input", "sequences", ",", "where", "the", "subsampling", "layer", "takes", "a", "window", "of", "hidden", "states", "from", "the", "lower", "layer", "as", "input", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "In", "this", "work", ",", "we", "consider", "three", "variants", ":", "a", ")", "concatenate", "–", "the", "hidden", "states", "in", "the", "subsampling", "window", "are", "concatenated", "before", "been", "fed", "into", "the", "next", "layer", ";", "b", ")", "add", "–", "the", "hidden", "states", "are", "added", "into", "one", "vector", "for", "the", "next", "layer", ";", "c", ")", "skip", "–", "only", "the", "last", "hidden", "state", "in", "the", "window", "is", "kept", "and", "all", "the", "others", "are", "skipped", ".", "The", "last", "two", "schemes", "are", "computationally", "cheaper", "as", "they", "do", "not", "introduce", "extra", "model", "parameters", ".", "section", ":", "Experiments", "subsection", ":", "System", "Setup", "We", "used", "the", "TIMIT", "dataset", "to", "evaluate", "the", "segmental", "RNN", "acoustic", "models", ".", "This", "dataset", "was", "preferred", "for", "the", "rapid", "evaluation", "of", "different", "system", "settings", ",", "and", "for", "the", "comparison", "to", "other", "CRF", "and", "end", "-", "to", "-", "end", "systems", ".", "We", "followed", "the", "standard", "protocol", "of", "the", "TIMIT", "dataset", ",", "and", "our", "experiments", "were", "based", "on", "the", "Kaldi", "recipe", ".", "We", "used", "the", "core", "test", "set", "as", "our", "evaluation", "set", ",", "which", "has", "192", "utterances", ".", "We", "used", "24", "dimensional", "log", "fiterbanks", "(", "FBANKs", ")", "with", "delta", "and", "double", "-", "delta", "coefficients", ",", "yielding", "72", "dimensional", "feature", "vectors", ".", "Our", "models", "were", "trained", "with", "48", "phonemes", ",", "and", "their", "predictions", "were", "converted", "to", "39", "phonemes", "before", "scoring", ".", "The", "dimension", "of", "was", "fixed", "to", "be", "32", ".", "For", "all", "our", "experiments", ",", "we", "used", "the", "long", "short", "-", "term", "memory", "(", "LSTM", ")", "networks", "as", "the", "implementation", "of", "RNNs", ",", "and", "the", "networks", "were", "always", "bi", "-", "directional", ".", "We", "set", "the", "initial", "SGD", "learning", "rate", "to", "be", "0.1", ",", "and", "we", "exponentially", "decay", "the", "learning", "rate", "by", "a", "factor", "of", "2", "when", "the", "validation", "error", "stopped", "decreasing", ".", "Our", "models", "were", "trained", "with", "dropout", "regularisation", ",", "using", "an", "specific", "implementation", "for", "recurrent", "networks", ".", "The", "dropout", "rate", "was", "0.2", "unless", "specified", "otherwise", ".", "Our", "models", "were", "randomly", "initialised", "with", "the", "same", "random", "seed", ".", "subsection", ":", "Results", "of", "Hierarchical", "Subsampling", "We", "first", "demonstrate", "the", "results", "of", "the", "hierarchical", "subsampling", "recurrent", "network", ",", "which", "is", "the", "key", "to", "speed", "up", "our", "experiments", ".", "We", "set", "the", "size", "of", "the", "subsampling", "window", "to", "be", "2", ",", "therefore", "each", "subsampling", "layer", "reduced", "the", "time", "resolution", "by", "a", "factor", "of", "2", ".", "We", "set", "the", "maximum", "segment", "length", "in", "Eq", ".", "(", "[", "reference", "]", ")", "to", "be", "300", "milliseconds", ",", "which", "corresponded", "to", "30", "frames", "of", "FBANKs", "(", "sampled", "at", "the", "rate", "of", "10", "milliseconds", ")", ".", "With", "two", "layers", "of", "subsampling", "recurrent", "networks", ",", "the", "time", "resolution", "was", "reduced", "by", "a", "factor", "of", "4", ",", "and", "the", "value", "of", "was", "reduced", "to", "be", "8", ",", "yielding", "around", "10", "times", "speedup", "as", "shown", "in", "Table", "[", "reference", "]", ".", "Table", "[", "reference", "]", "compares", "the", "three", "implementations", "of", "the", "recurrent", "subsampling", "network", "detailed", "in", "section", "[", "reference", "]", ".", "We", "observed", "that", "concatenating", "all", "the", "hidden", "states", "in", "the", "subsampling", "window", "did", "not", "yield", "lower", "phone", "error", "rate", "(", "PER", ")", "than", "using", "the", "simple", "skipping", "approach", ",", "which", "may", "be", "due", "to", "the", "fact", "that", "the", "TIMIT", "dataset", "is", "small", "and", "it", "prefers", "a", "smaller", "model", ".", "On", "the", "other", "hand", ",", "adding", "the", "hidden", "states", "in", "the", "subsampling", "window", "together", "worked", "even", "worse", ",", "possibly", "due", "to", "that", "the", "sequential", "information", "in", "the", "subsampling", "window", "was", "flattened", ".", "In", "the", "following", "experiments", ",", "we", "sticked", "to", "the", "skipping", "method", ",", "and", "using", "two", "subsampling", "layers", ".", "subsection", ":", "Hyperparameters", "and", "Different", "Features", "We", "then", "evaluated", "the", "model", "by", "tuning", "the", "hyperparameters", ",", "and", "the", "results", "are", "given", "in", "Table", "[", "reference", "]", ".", "We", "tuned", "the", "number", "of", "LSTM", "layers", ",", "and", "the", "dimension", "of", "LSTM", "cells", ",", "as", "well", "as", "the", "dimensions", "of", "and", "the", "segment", "vector", ".", "In", "general", ",", "larger", "models", "with", "dropout", "regularisation", "yielded", "higher", "recognition", "accuracy", ".", "Our", "best", "result", "was", "obtained", "using", "6", "layers", "of", "250", "-", "dimensional", "LSTMs", ".", "However", ",", "without", "the", "dropout", "regularisation", ",", "the", "model", "can", "be", "easily", "overfit", "due", "to", "the", "small", "size", "of", "training", "set", ".", "In", "the", "future", ",", "we", "shall", "evaluate", "this", "model", "with", "a", "large", "dataset", ".", "We", "then", "evaluated", "another", "two", "types", "of", "features", "using", "the", "same", "system", "configuration", "that", "achieved", "the", "best", "result", "in", "Table", "[", "reference", "]", ".", "We", "increased", "the", "number", "of", "FBANKs", "from", "24", "to", "40", ",", "which", "yielded", "slightly", "lower", "PER", ".", "We", "also", "evaluated", "the", "standard", "Kaldi", "features", "—", "39", "dimensional", "MFCCs", "spliced", "by", "a", "context", "window", "of", "7", ",", "followed", "by", "LDA", "and", "MLLT", "transform", "and", "with", "feature", "-", "space", "speaker", "-", "dependent", "MLLR", ",", "which", "were", "the", "same", "features", "used", "in", "the", "HMM", "-", "DNN", "baseline", "in", "Table", "[", "reference", "]", ".", "The", "well", "-", "engineered", "features", "improved", "the", "accuracy", "of", "our", "system", "by", "more", "than", "1", "%", "absolute", ".", "subsection", ":", "Comparison", "to", "Related", "Works", "In", "Table", "[", "reference", "]", ",", "we", "compare", "our", "result", "to", "other", "reported", "results", "using", "segmental", "CRFs", "as", "well", "as", "recent", "end", "-", "to", "-", "end", "systems", ".", "Previous", "state", "-", "of", "-", "the", "-", "art", "result", "using", "segmental", "CRFs", "on", "the", "TIMIT", "dataset", "is", "reported", "in", ",", "where", "the", "first", "-", "pass", "decoding", "was", "used", "to", "prune", "the", "search", "space", ",", "and", "the", "second", "-", "pass", "was", "used", "to", "re", "-", "score", "the", "hypothesis", "using", "various", "features", "including", "neural", "network", "features", ".", "Besides", ",", "the", "ground", "-", "truth", "segmentation", "was", "used", "in", ".", "We", "achieved", "considerably", "lower", "PER", "with", "first", "-", "pass", "decoding", ",", "despite", "the", "fact", "that", "our", "CRF", "was", "zeroth", "-", "order", ",", "and", "we", "did", "not", "use", "any", "language", "model", ".", "Furthermore", ",", "our", "results", "are", "also", "comparable", "to", "that", "from", "the", "CTC", "and", "attention", "-", "based", "RNN", "end", "-", "to", "-", "end", "systems", ".", "The", "accuracy", "of", "segmental", "RNNs", "may", "be", "further", "improved", "by", "using", "higher", "-", "order", "CRFs", "or", "incorporating", "a", "language", "model", "into", "the", "decode", "step", ",", "and", "using", "beam", "search", "to", "reduce", "the", "search", "error", ".", "section", ":", "Conclusions", "In", "this", "paper", ",", "we", "present", "the", "segmental", "RNN", "—", "a", "novel", "acoustic", "model", "that", "combines", "the", "segmental", "CRF", "with", "an", "encoder", "RNN", "for", "end", "-", "to", "-", "end", "speech", "recognition", ".", "We", "discuss", "the", "practical", "training", "and", "decoding", "algorithms", "of", "this", "model", "for", "speech", "recognition", ",", "and", "the", "subsampling", "network", "to", "reduce", "the", "computational", "cost", ".", "Our", "experiments", "were", "performed", "on", "the", "TIMIT", "dataset", ",", "and", "we", "achieved", "strong", "recognition", "accuracy", "using", "zeroth", "-", "order", "CRF", ",", "and", "without", "using", "any", "language", "model", ".", "In", "the", "future", ",", "we", "shall", "investigate", "discriminative", "training", "criteria", ",", "and", "incorporating", "a", "language", "model", "into", "the", "decoding", "step", ".", "Future", "works", "also", "include", "implementing", "a", "weighted", "finite", "sate", "transducer", "(", "WFST", ")", "based", "decoder", "and", "scaling", "this", "model", "to", "large", "vocabulary", "datasets", ".", "bibliography", ":", "References"], "entity": [[2, 6, ["Segmental", "Recurrent", "Neural", "Networks"], "Method"], [7, 14, ["End", "-", "to", "-", "end", "Speech", "Recognition"], "Task"], [17, 21, ["segmental", "recurrent", "neural", "network"], "Method"], [22, 29, ["end", "-", "to", "-", "end", "acoustic", "modelling"], "Task"], [35, 38, ["conditional", "random", "field"], "Method"], [39, 40, ["CRF"], "Method"], [43, 46, ["recurrent", "neural", "network"], "Method"], [47, 49, ["RNN", ")"], "Method"], [51, 53, ["feature", "extraction"], "Task"], [70, 72, ["external", "system"], "Method"], [96, 97, ["RNN"], "Method"], [128, 133, ["practical", "training", "and", "decoding", "issues"], "Task"], [142, 143, ["training"], "Task"], [147, 149, ["speech", "recognition"], "Task"], [155, 157, ["TIMIT", "dataset"], "Material"], [162, 165, ["phone", "error", "rate"], "Metric"], [166, 167, ["PER"], "Metric"], [170, 174, ["first", "-", "pass", "decoding"], "Method"], [180, 181, ["CRFs"], "Method"], [198, 200, ["language", "model"], "Method"], [248, 251, ["recurrent", "neural", "networks"], "Method"], [255, 257, ["Speech", "recognition"], "Task"], [260, 265, ["sequence", "to", "sequence", "transduction", "problem"], "Task"], [298, 300, ["acoustic", "model"], "Method"], [340, 343, ["hidden", "Markov", "model"], "Method"], [344, 345, ["HMM"], "Method"], [348, 353, ["sequence", "-", "level", "classification", "task"], "Task"], [355, 360, ["frame", "-", "level", "classification", "problem"], "Task"], [392, 393, ["HMMs"], "Method"], [402, 407, ["first", "-", "order", "Markov", "rule"], "Method"], [414, 415, ["HMMs"], "Method"], [419, 423, ["HMM", "-", "based", "pipeline"], "Method"], [435, 437, ["joint", "optimisation"], "Task"], [451, 452, ["HMMs"], "Method"], [457, 459, ["acoustic", "modelling"], "Task"], [474, 475, ["HMMs"], "Method"], [483, 487, ["neural", "network", "based", "approaches"], "Method"], [530, 533, ["recurrent", "neural", "network"], "Method"], [539, 540, ["CTC"], "Method"], [570, 574, ["frame", "-", "level", "classification"], "Method"], [595, 597, ["speech", "recognition"], "Task"], [598, 599, ["CTC"], "Method"], [637, 643, ["sequence", "-", "to", "-", "sequence", "transducer"], "Method"], [647, 649, ["machine", "translation"], "Task"], [651, 653, ["speech", "recognition"], "Task"], [661, 662, ["HMMs"], "Method"], [663, 664, ["CTCs"], "Method"], [667, 671, ["attention", "-", "based", "approach"], "Method"], [698, 700, ["vector", "representation"], "Method"], [706, 710, ["attention", "-", "based", "scheme"], "Method"], [725, 726, ["RNN"], "Method"], [729, 731, ["vector", "representation"], "Method"], [737, 739, ["attentive", "scheme"], "Method"], [741, 744, ["machine", "translation", "task"], "Task"], [802, 804, ["acoustic", "modelling"], "Task"], [810, 811, ["CTC"], "Method"], [812, 816, ["attention", "-", "based", "RNN"], "Method"], [827, 829, ["feature", "extraction"], "Task"], [862, 863, ["CRF"], "Method"], [870, 871, ["CRFs"], "Method"], [875, 877, ["speech", "recognition"], "Task"], [890, 895, ["feed", "-", "forward", "neural", "networks"], "Method"], [902, 904, ["speech", "recognition"], "Task"], [927, 929, ["external", "systems"], "Method"], [973, 975, ["TIMIT", "dataset"], "Material"], [981, 982, ["PER"], "Metric"], [983, 987, ["first", "-", "pass", "decoding"], "Method"], [996, 998, ["language", "model"], "Method"], [1004, 1005, ["CRFs"], "Method"], [1008, 1012, ["Segmental", "Recurrent", "Neural", "Networks"], "Method"], [1014, 1018, ["Segmental", "Conditional", "Random", "Fields"], "Method"], [1117, 1119, ["order", "models"], "Method"], [1183, 1185, ["neural", "networks"], "Method"], [1188, 1189, ["segmentation"], "Method"], [1205, 1207, ["Feature", "Representations"], "Method"], [1209, 1211, ["neural", "networks"], "Method"], [1256, 1259, ["linear", "embedding", "matrix"], "Method"], [1268, 1269, ["RNN"], "Method"], [1278, 1281, ["dimensional", "vector", "representation"], "Method"], [1331, 1337, ["linear", "or", "non", "-", "linear", "transformation"], "Method"], [1399, 1403, ["Conditional", "Maximum", "Likelihood", "Training"], "Method"], [1404, 1406, ["speech", "recognition"], "Task"], [1480, 1482, ["objective", "function"], "Metric"], [1524, 1527, ["dynamic", "programming", "algorithm"], "Method"], [1568, 1570, ["computation", "cost"], "Metric"], [1640, 1642, ["computational", "savings"], "Metric"], [1715, 1718, ["stochastic", "gradient", "decent"], "Method"], [1719, 1720, ["SGD"], "Method"], [1726, 1729, ["neural", "network", "models"], "Method"], [1747, 1749, ["Viterbi", "Decoding"], "Method"], [1750, 1751, ["decoding"], "Task"], [1781, 1783, ["recursive", "algorithm"], "Method"], [1822, 1824, ["computational", "cost"], "Metric"], [1842, 1844, ["decoding", "objective"], "Task"], [1846, 1849, ["joint", "maximization", "algorithm"], "Method"], [1852, 1854, ["search", "error"], "Metric"], [1870, 1873, ["beam", "search", "algorithm"], "Method"], [1878, 1880, ["search", "error"], "Metric"], [1890, 1891, ["RNNs"], "Method"], [1916, 1918, ["computational", "cost"], "Metric"], [1936, 1938, ["subsampling", "layer"], "Method"], [1969, 1970, ["concatenate"], "Method"], [2052, 2054, ["TIMIT", "dataset"], "Material"], [2072, 2074, ["system", "settings"], "Task"], [2097, 2099, ["TIMIT", "dataset"], "Material"], [2107, 2109, ["Kaldi", "recipe"], "Method"], [2131, 2132, ["fiterbanks"], "Method"], [2133, 2134, ["FBANKs"], "Method"], [2185, 2190, ["long", "short", "-", "term", "memory"], "Method"], [2191, 2192, ["LSTM"], "Method"], [2198, 2199, ["RNNs"], "Method"], [2225, 2227, ["learning", "rate"], "Metric"], [2234, 2236, ["validation", "error"], "Metric"], [2244, 2246, ["dropout", "regularisation"], "Method"], [2252, 2254, ["recurrent", "networks"], "Method"], [2256, 2258, ["dropout", "rate"], "Metric"], [2279, 2281, ["Hierarchical", "Subsampling"], "Method"], [2288, 2292, ["hierarchical", "subsampling", "recurrent", "network"], "Method"], [2317, 2319, ["subsampling", "layer"], "Method"], [2369, 2372, ["subsampling", "recurrent", "networks"], "Method"], [2374, 2376, ["time", "resolution"], "Metric"], [2417, 2420, ["recurrent", "subsampling", "network"], "Method"], [2443, 2446, ["phone", "error", "rate"], "Metric"], [2447, 2448, ["PER"], "Metric"], [2453, 2455, ["skipping", "approach"], "Method"], [2465, 2467, ["TIMIT", "dataset"], "Material"], [2517, 2519, ["skipping", "method"], "Method"], [2523, 2525, ["subsampling", "layers"], "Method"], [2585, 2587, ["dropout", "regularisation"], "Method"], [2589, 2591, ["recognition", "accuracy"], "Metric"], [2601, 2605, ["250", "-", "dimensional", "LSTMs"], "Method"], [2610, 2612, ["dropout", "regularisation"], "Method"], [2681, 2682, ["PER"], "Metric"], [2688, 2690, ["Kaldi", "features"], "Method"], [2692, 2694, ["dimensional", "MFCCs"], "Method"], [2704, 2708, ["LDA", "and", "MLLT", "transform"], "Method"], [2710, 2717, ["feature", "-", "space", "speaker", "-", "dependent", "MLLR"], "Method"], [2726, 2730, ["HMM", "-", "DNN", "baseline"], "Method"], [2743, 2744, ["accuracy"], "Metric"], [2785, 2787, ["end", "systems"], "Method"], [2802, 2804, ["TIMIT", "dataset"], "Material"], [2812, 2814, ["pass", "decoding"], "Method"], [2839, 2842, ["neural", "network", "features"], "Method"], [2846, 2850, ["ground", "-", "truth", "segmentation"], "Metric"], [2858, 2859, ["PER"], "Metric"], [2860, 2864, ["first", "-", "pass", "decoding"], "Method"], [2870, 2871, ["CRF"], "Method"], [2882, 2884, ["language", "model"], "Method"], [2910, 2911, ["accuracy"], "Metric"], [2927, 2929, ["language", "model"], "Method"], [2931, 2933, ["decode", "step"], "Method"], [2936, 2938, ["beam", "search"], "Method"], [2941, 2943, ["search", "error"], "Metric"], [2959, 2961, ["acoustic", "model"], "Method"], [2983, 2987, ["training", "and", "decoding", "algorithms"], "Method"], [2991, 2993, ["speech", "recognition"], "Task"], [2996, 2998, ["subsampling", "network"], "Method"], [3001, 3003, ["computational", "cost"], "Metric"], [3010, 3012, ["TIMIT", "dataset"], "Material"], [3017, 3019, ["recognition", "accuracy"], "Metric"], [3029, 3031, ["language", "model"], "Method"], [3039, 3042, ["discriminative", "training", "criteria"], "Metric"], [3046, 3048, ["language", "model"], "Method"], [3050, 3052, ["decoding", "step"], "Task"], [3059, 3068, ["weighted", "finite", "sate", "transducer", "(", "WFST", ")", "based", "decoder"], "Method"], [58, 59, ["CRF"], "Method"], [102, 103, ["CRF"], "Method"], [193, 194, ["CRF"], "Method"], [242, 244, ["speech", "recognition"], "Task"], [246, 247, ["CRF"], "Method"], [276, 278, ["speech", "recognition"], "Task"], [293, 295, ["speech", "recognition"], "Task"], [470, 472, ["speech", "recognition"], "Task"], [503, 504, ["CTC"], "Method"], [628, 629, ["RNNs"], "Method"], [774, 776, ["speech", "recognition"], "Task"], [800, 801, ["RNNs"], "Method"], [821, 822, ["RNN"], "Method"], [853, 854, ["CRF"], "Method"], [900, 901, ["CRFs"], "Method"], [908, 909, ["RNNs"], "Method"], [956, 957, ["RNNs"], "Method"], [965, 966, ["CRFs"], "Method"], [991, 992, ["CRF"], "Method"], [1041, 1044, ["conditional", "random", "field"], "Method"], [1110, 1111, ["CRF"], "Method"], [1125, 1126, ["CRF"], "Method"], [1150, 1151, ["CRF"], "Method"], [1376, 1377, ["RNN"], "Method"], [1927, 1928, ["RNN"], "Method"], [2058, 2059, ["RNN"], "Method"], [2081, 2082, ["CRF"], "Method"], [2213, 2214, ["SGD"], "Method"], [2354, 2355, ["FBANKs"], "Method"], [2558, 2559, ["LSTM"], "Method"], [2565, 2566, ["LSTM"], "Method"], [2671, 2672, ["FBANKs"], "Method"], [2776, 2777, ["CRFs"], "Method"], [2799, 2800, ["CRFs"], "Method"], [2896, 2897, ["CTC"], "Method"], [2901, 2902, ["RNN"], "Method"], [2913, 2914, ["RNNs"], "Method"], [2923, 2924, ["CRFs"], "Method"], [2955, 2956, ["RNN"], "Method"], [2965, 2966, ["CRF"], "Method"], [2969, 2970, ["RNN"], "Method"], [2976, 2978, ["speech", "recognition"], "Task"], [3023, 3024, ["CRF"], "Method"]]} {"doc_id": "26d4ab9b60b91bb610202b58fa1766951fedb9e9", "text": ["This", "paper", "introduces", "the", "Deep", "Recurrent", "Attentive", "Writer", "(", "DRAW", ")", "neural", "network", "architecture", "for", "image", "generation", ".", "DRAW", "networks", "combine", "a", "novel", "spatial", "attention", "mechanism", "that", "mimics", "the", "foveation", "of", "the", "human", "eye", ",", "with", "a", "sequential", "variational", "auto", "-", "encoding", "framework", "that", "allows", "for", "the", "iterative", "construction", "of", "complex", "images", ".", "The", "system", "substantially", "improves", "on", "the", "state", "of", "the", "art", "for", "generative", "models", "on", "MNIST", ",", "and", ",", "when", "trained", "on", "the", "Street", "View", "House", "Numbers", "dataset", ",", "it", "generates", "images", "that", "can", "not", "be", "distinguished", "from", "real", "data", "with", "the", "naked", "eye", ".", "DRAW", ":", "ARecurrentNeuralNetworkForImageGeneration", "section", ":", "Introduction", "A", "person", "asked", "to", "draw", ",", "paint", "or", "otherwise", "recreate", "a", "visual", "scene", "will", "naturally", "do", "so", "in", "a", "sequential", ",", "iterative", "fashion", ",", "reassessing", "their", "handiwork", "after", "each", "modification", ".", "Rough", "outlines", "are", "gradually", "replaced", "by", "precise", "forms", ",", "lines", "are", "sharpened", ",", "darkened", "or", "erased", ",", "shapes", "are", "altered", ",", "and", "the", "final", "picture", "emerges", ".", "Most", "approaches", "to", "automatic", "image", "generation", ",", "however", ",", "aim", "to", "generate", "entire", "scenes", "at", "once", ".", "In", "the", "context", "of", "generative", "neural", "networks", ",", "this", "typically", "means", "that", "all", "the", "pixels", "are", "conditioned", "on", "a", "single", "latent", "distribution", ".", "As", "well", "as", "precluding", "the", "possibility", "of", "iterative", "self", "-", "correction", ",", "the", "“", "one", "shot", "”", "approach", "is", "fundamentally", "difficult", "to", "scale", "to", "large", "images", ".", "The", "Deep", "Recurrent", "Attentive", "Writer", "(", "DRAW", ")", "architecture", "represents", "a", "shift", "towards", "a", "more", "natural", "form", "of", "image", "construction", ",", "in", "which", "parts", "of", "a", "scene", "are", "created", "independently", "from", "others", ",", "and", "approximate", "sketches", "are", "successively", "refined", ".", "The", "core", "of", "the", "DRAW", "architecture", "is", "a", "pair", "of", "recurrent", "neural", "networks", ":", "an", "encoder", "network", "that", "compresses", "the", "real", "images", "presented", "during", "training", ",", "and", "a", "decoder", "that", "reconstitutes", "images", "after", "receiving", "codes", ".", "The", "combined", "system", "is", "trained", "end", "-", "to", "-", "end", "with", "stochastic", "gradient", "descent", ",", "where", "the", "loss", "function", "is", "a", "variational", "upper", "bound", "on", "the", "log", "-", "likelihood", "of", "the", "data", ".", "It", "therefore", "belongs", "to", "the", "family", "of", "variational", "auto", "-", "encoders", ",", "a", "recently", "emerged", "hybrid", "of", "deep", "learning", "and", "variational", "inference", "that", "has", "led", "to", "significant", "advances", "in", "generative", "modelling", ".", "Where", "DRAW", "differs", "from", "its", "siblings", "is", "that", ",", "rather", "than", "generating", "images", "in", "a", "single", "pass", ",", "it", "iteratively", "constructs", "scenes", "through", "an", "accumulation", "of", "modifications", "emitted", "by", "the", "decoder", ",", "each", "of", "which", "is", "observed", "by", "the", "encoder", ".", "An", "obvious", "correlate", "of", "generating", "images", "step", "by", "step", "is", "the", "ability", "to", "selectively", "attend", "to", "parts", "of", "the", "scene", "while", "ignoring", "others", ".", "A", "wealth", "of", "results", "in", "the", "past", "few", "years", "suggest", "that", "visual", "structure", "can", "be", "better", "captured", "by", "a", "sequence", "of", "partial", "glimpses", ",", "or", "foveations", ",", "than", "by", "a", "single", "sweep", "through", "the", "entire", "image", ".", "The", "main", "challenge", "faced", "by", "sequential", "attention", "models", "is", "learning", "where", "to", "look", ",", "which", "can", "be", "addressed", "with", "reinforcement", "learning", "techniques", "such", "as", "policy", "gradients", ".", "The", "attention", "model", "in", "DRAW", ",", "however", ",", "is", "fully", "differentiable", ",", "making", "it", "possible", "to", "train", "with", "standard", "backpropagation", ".", "In", "this", "sense", "it", "resembles", "the", "selective", "read", "and", "write", "operations", "developed", "for", "the", "Neural", "Turing", "Machine", ".", "The", "following", "section", "defines", "the", "DRAW", "architecture", ",", "along", "with", "the", "loss", "function", "used", "for", "training", "and", "the", "procedure", "for", "image", "generation", ".", "Section", "[", "reference", "]", "presents", "the", "selective", "attention", "model", "and", "shows", "how", "it", "is", "applied", "to", "reading", "and", "modifying", "images", ".", "Section", "[", "reference", "]", "provides", "experimental", "results", "on", "the", "MNIST", ",", "Street", "View", "House", "Numbers", "and", "CIFAR", "-", "10", "datasets", ",", "with", "examples", "of", "generated", "images", ";", "and", "concluding", "remarks", "are", "given", "in", "Section", "[", "reference", "]", ".", "Lastly", ",", "we", "would", "like", "to", "direct", "the", "reader", "to", "the", "video", "accompanying", "this", "paper", "(", ")", "which", "contains", "examples", "of", "DRAW", "networks", "reading", "and", "generating", "images", ".", "section", ":", "The", "DRAW", "Network", "The", "basic", "structure", "of", "a", "DRAW", "network", "is", "similar", "to", "that", "of", "other", "variational", "auto", "-", "encoders", ":", "an", "encoder", "network", "determines", "a", "distribution", "over", "latent", "codes", "that", "capture", "salient", "information", "about", "the", "input", "data", ";", "a", "decoder", "network", "receives", "samples", "from", "the", "code", "distribuion", "and", "uses", "them", "to", "condition", "its", "own", "distribution", "over", "images", ".", "However", "there", "are", "three", "key", "differences", ".", "Firstly", ",", "both", "the", "encoder", "and", "decoder", "are", "recurrent", "networks", "in", "DRAW", ",", "so", "that", "a", "sequence", "of", "code", "samples", "is", "exchanged", "between", "them", ";", "moreover", "the", "encoder", "is", "privy", "to", "the", "decoder", "’s", "previous", "outputs", ",", "allowing", "it", "to", "tailor", "the", "codes", "it", "sends", "according", "to", "the", "decoder", "’s", "behaviour", "so", "far", ".", "Secondly", ",", "the", "decoder", "’s", "outputs", "are", "successively", "added", "to", "the", "distribution", "that", "will", "ultimately", "generate", "the", "data", ",", "as", "opposed", "to", "emitting", "this", "distribution", "in", "a", "single", "step", ".", "And", "thirdly", ",", "a", "dynamically", "updated", "attention", "mechanism", "is", "used", "to", "restrict", "both", "the", "input", "region", "observed", "by", "the", "encoder", ",", "and", "the", "output", "region", "modified", "by", "the", "decoder", ".", "In", "simple", "terms", ",", "the", "network", "decides", "at", "each", "time", "-", "step", "“", "where", "to", "read", "”", "and", "“", "where", "to", "write", "”", "as", "well", "as", "“", "what", "to", "write", "”", ".", "The", "architecture", "is", "sketched", "in", "Fig", ".", "[", "reference", "]", ",", "alongside", "a", "feedforward", "variational", "auto", "-", "encoder", ".", "subsection", ":", "Network", "Architecture", "Let", "be", "the", "function", "enacted", "by", "the", "encoder", "network", "at", "a", "single", "time", "-", "step", ".", "The", "output", "of", "at", "time", "is", "the", "encoder", "hidden", "vector", ".", "Similarly", "the", "output", "of", "the", "decoder", "at", "is", "the", "hidden", "vector", ".", "In", "general", "the", "encoder", "and", "decoder", "may", "be", "implemented", "by", "any", "recurrent", "neural", "network", ".", "In", "our", "experiments", "we", "use", "the", "Long", "Short", "-", "Term", "Memory", "architecture", "(", "LSTM", ";", ")", "for", "both", ",", "in", "the", "extended", "form", "with", "forget", "gates", ".", "We", "favour", "LSTM", "due", "to", "its", "proven", "track", "record", "for", "handling", "long", "-", "range", "dependencies", "in", "real", "sequential", "data", ".", "Throughout", "the", "paper", ",", "we", "use", "the", "notation", "to", "denote", "a", "linear", "weight", "matrix", "with", "bias", "from", "the", "vector", "to", "the", "vector", ".", "At", "each", "time", "-", "step", ",", "the", "encoder", "receives", "input", "from", "both", "the", "image", "and", "from", "the", "previous", "decoder", "hidden", "vector", ".", "The", "precise", "form", "of", "the", "encoder", "input", "depends", "on", "a", "operation", ",", "which", "will", "be", "defined", "in", "the", "next", "section", ".", "The", "output", "of", "the", "encoder", "is", "used", "to", "parameterise", "a", "distribution", "over", "the", "latent", "vector", ".", "In", "our", "experiments", "the", "latent", "distribution", "is", "a", "diagonal", "Gaussian", ":", "Bernoulli", "distributions", "are", "more", "common", "than", "Gaussians", "for", "latent", "variables", "in", "auto", "-", "encoders", ";", "however", "a", "great", "advantage", "of", "Gaussian", "latents", "is", "that", "the", "gradient", "of", "a", "function", "of", "the", "samples", "with", "respect", "to", "the", "distribution", "parameters", "can", "be", "easily", "obtained", "using", "the", "so", "-", "called", "reparameterization", "trick", ".", "This", "makes", "it", "straightforward", "to", "back", "-", "propagate", "unbiased", ",", "low", "variance", "stochastic", "gradients", "of", "the", "loss", "function", "through", "the", "latent", "distribution", ".", "At", "each", "time", "-", "step", "a", "sample", "drawn", "from", "the", "latent", "distribution", "is", "passed", "as", "input", "to", "the", "decoder", ".", "The", "output", "of", "the", "decoder", "is", "added", "(", "via", "a", "operation", ",", "defined", "in", "the", "sequel", ")", "to", "a", "cumulative", "canvas", "matrix", ",", "which", "is", "ultimately", "used", "to", "reconstruct", "the", "image", ".", "The", "total", "number", "of", "time", "-", "steps", "consumed", "by", "the", "network", "before", "performing", "the", "reconstruction", "is", "a", "free", "parameter", "that", "must", "be", "specified", "in", "advance", ".", "For", "each", "image", "presented", "to", "the", "network", ",", "are", "initialised", "to", "learned", "biases", ",", "and", "the", "DRAW", "network", "iteratively", "computes", "the", "following", "equations", "for", ":", "where", "is", "the", "error", "image", ",", "is", "the", "concatenation", "of", "vectors", "and", "into", "a", "single", "vector", ",", "and", "denotes", "the", "logistic", "sigmoid", "function", ":", ".", "Note", "that", ",", "and", "hence", ",", "depends", "on", "both", "and", "the", "history", "of", "previous", "latent", "samples", ".", "We", "will", "sometimes", "make", "this", "dependency", "explicit", "by", "writing", ",", "as", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "can", "also", "be", "passed", "as", "input", "to", "the", "operation", ";", "however", "we", "did", "not", "find", "that", "this", "helped", "performance", "and", "therefore", "omitted", "it", ".", "subsection", ":", "Loss", "Function", "The", "final", "canvas", "matrix", "is", "used", "to", "parameterise", "a", "model", "of", "the", "input", "data", ".", "If", "the", "input", "is", "binary", ",", "the", "natural", "choice", "for", "is", "a", "Bernoulli", "distribution", "with", "means", "given", "by", ".", "The", "reconstruction", "loss", "is", "defined", "as", "the", "negative", "log", "probability", "of", "under", ":", "The", "latent", "loss", "for", "a", "sequence", "of", "latent", "distributions", "is", "defined", "as", "the", "summed", "Kullback", "-", "Leibler", "divergence", "of", "some", "latent", "prior", "from", ":", "Note", "that", "this", "loss", "depends", "upon", "the", "latent", "samples", "drawn", "from", ",", "which", "depend", "in", "turn", "on", "the", "input", ".", "If", "the", "latent", "distribution", "is", "a", "diagonal", "Gaussian", "with", ",", "as", "defined", "in", "Eqs", "[", "reference", "]", "and", "[", "reference", "]", ",", "a", "simple", "choice", "for", "is", "a", "standard", "Gaussian", "with", "mean", "zero", "and", "standard", "deviation", "one", ",", "in", "which", "case", "Eq", ".", "[", "reference", "]", "becomes", "The", "total", "loss", "for", "the", "network", "is", "the", "expectation", "of", "the", "sum", "of", "the", "reconstruction", "and", "latent", "losses", ":", "which", "we", "optimise", "using", "a", "single", "sample", "of", "for", "each", "stochastic", "gradient", "descent", "step", ".", "can", "be", "interpreted", "as", "the", "number", "of", "nats", "required", "to", "transmit", "the", "latent", "sample", "sequence", "to", "the", "decoder", "from", "the", "prior", ",", "and", "(", "if", "is", "discrete", ")", "is", "the", "number", "of", "nats", "required", "for", "the", "decoder", "to", "reconstruct", "given", ".", "The", "total", "loss", "is", "therefore", "equivalent", "to", "the", "expected", "compression", "of", "the", "data", "by", "the", "decoder", "and", "prior", ".", "subsection", ":", "Stochastic", "Data", "Generation", "An", "image", "can", "be", "generated", "by", "a", "DRAW", "network", "by", "iteratively", "picking", "latent", "samples", "from", "the", "prior", ",", "then", "running", "the", "decoder", "to", "update", "the", "canvas", "matrix", ".", "After", "repetitions", "of", "this", "process", "the", "generated", "image", "is", "a", "sample", "from", ":", "Note", "that", "the", "encoder", "is", "not", "involved", "in", "image", "generation", ".", "section", ":", "Read", "and", "Write", "Operations", "The", "DRAW", "network", "described", "in", "the", "previous", "section", "is", "not", "complete", "until", "the", "and", "operations", "in", "Eqs", ".", "[", "reference", "]", "and", "[", "reference", "]", "have", "been", "defined", ".", "This", "section", "describes", "two", "ways", "to", "do", "so", ",", "one", "with", "selective", "attention", "and", "one", "without", ".", "subsection", ":", "Reading", "and", "Writing", "Without", "Attention", "In", "the", "simplest", "instantiation", "of", "DRAW", "the", "entire", "input", "image", "is", "passed", "to", "the", "encoder", "at", "every", "time", "-", "step", ",", "and", "the", "decoder", "modifies", "the", "entire", "canvas", "matrix", "at", "every", "time", "-", "step", ".", "In", "this", "case", "the", "and", "operations", "reduce", "to", "However", "this", "approach", "does", "not", "allow", "the", "encoder", "to", "focus", "on", "only", "part", "of", "the", "input", "when", "creating", "the", "latent", "distribution", ";", "nor", "does", "it", "allow", "the", "decoder", "to", "modify", "only", "a", "part", "of", "the", "canvas", "vector", ".", "In", "other", "words", "it", "does", "not", "provide", "the", "network", "with", "an", "explicit", "selective", "attention", "mechanism", ",", "which", "we", "believe", "to", "be", "crucial", "to", "large", "scale", "image", "generation", ".", "We", "refer", "to", "the", "above", "configuration", "as", "“", "DRAW", "without", "attention", "”", ".", "subsection", ":", "Selective", "Attention", "Model", "To", "endow", "the", "network", "with", "selective", "attention", "without", "sacrificing", "the", "benefits", "of", "gradient", "descent", "training", ",", "we", "take", "inspiration", "from", "the", "differentiable", "attention", "mechanisms", "recently", "used", "in", "handwriting", "synthesis", "and", "Neural", "Turing", "Machines", ".", "Unlike", "the", "aforementioned", "works", ",", "we", "consider", "an", "explicitly", "two", "-", "dimensional", "form", "of", "attention", ",", "where", "an", "array", "of", "2D", "Gaussian", "filters", "is", "applied", "to", "the", "image", ",", "yielding", "an", "image", "‘", "patch", "’", "of", "smoothly", "varying", "location", "and", "zoom", ".", "This", "configuration", ",", "which", "we", "refer", "to", "simply", "as", "“", "DRAW", "”", ",", "somewhat", "resembles", "the", "affine", "transformations", "used", "in", "computer", "graphics", "-", "based", "autoencoders", ".", "As", "illustrated", "in", "Fig", ".", "[", "reference", "]", ",", "the", "grid", "of", "Gaussian", "filters", "is", "positioned", "on", "the", "image", "by", "specifying", "the", "co", "-", "ordinates", "of", "the", "grid", "centre", "and", "the", "stride", "distance", "between", "adjacent", "filters", ".", "The", "stride", "controls", "the", "‘", "zoom", "’", "of", "the", "patch", ";", "that", "is", ",", "the", "larger", "the", "stride", ",", "the", "larger", "an", "area", "of", "the", "original", "image", "will", "be", "visible", "in", "the", "attention", "patch", ",", "but", "the", "lower", "the", "effective", "resolution", "of", "the", "patch", "will", "be", ".", "The", "grid", "centre", "and", "stride", "(", "both", "of", "which", "are", "real", "-", "valued", ")", "determine", "the", "mean", "location", "of", "the", "filter", "at", "row", ",", "column", "in", "the", "patch", "as", "follows", ":", "Two", "more", "parameters", "are", "required", "to", "fully", "specify", "the", "attention", "model", ":", "the", "isotropic", "variance", "of", "the", "Gaussian", "filters", ",", "and", "a", "scalar", "intensity", "that", "multiplies", "the", "filter", "response", ".", "Given", "an", "input", "image", ",", "all", "five", "attention", "parameters", "are", "dynamically", "determined", "at", "each", "time", "step", "via", "a", "linear", "transformation", "of", "the", "decoder", "output", ":", "where", "the", "variance", ",", "stride", "and", "intensity", "are", "emitted", "in", "the", "log", "-", "scale", "to", "ensure", "positivity", ".", "The", "scaling", "of", ",", "and", "is", "chosen", "to", "ensure", "that", "the", "initial", "patch", "(", "with", "a", "randomly", "initialised", "network", ")", "roughly", "covers", "the", "whole", "input", "image", ".", "Given", "the", "attention", "parameters", "emitted", "by", "the", "decoder", ",", "the", "horizontal", "and", "vertical", "filterbank", "matrices", "and", "(", "dimensions", "and", "respectively", ")", "are", "defined", "as", "follows", ":", "where", "is", "a", "point", "in", "the", "attention", "patch", ",", "is", "a", "point", "in", "the", "input", "image", ",", "and", "are", "normalisation", "constants", "that", "ensure", "that", "and", ".", "subsection", ":", "Reading", "and", "Writing", "With", "Attention", "Given", ",", "and", "intensity", "determined", "by", ",", "along", "with", "an", "input", "image", "and", "error", "image", ",", "the", "read", "operation", "returns", "the", "concatenation", "of", "two", "patches", "from", "the", "image", "and", "error", "image", ":", "Note", "that", "the", "same", "filterbanks", "are", "used", "for", "both", "the", "image", "and", "error", "image", ".", "For", "the", "write", "operation", ",", "a", "distinct", "set", "of", "attention", "parameters", ",", "and", "are", "extracted", "from", ",", "the", "order", "of", "transposition", "is", "reversed", ",", "and", "the", "intensity", "is", "inverted", ":", "where", "is", "the", "writing", "patch", "emitted", "by", ".", "For", "colour", "images", "each", "point", "in", "the", "input", "and", "error", "image", "(", "and", "hence", "in", "the", "reading", "and", "writing", "patches", ")", "is", "an", "RGB", "triple", ".", "In", "this", "case", "the", "same", "reading", "and", "writing", "filters", "are", "used", "for", "all", "three", "channels", ".", "section", ":", "Experimental", "Results", "We", "assess", "the", "ability", "of", "DRAW", "to", "generate", "realistic", "-", "looking", "images", "by", "training", "on", "three", "datasets", "of", "progressively", "increasing", "visual", "complexity", ":", "MNIST", ",", "Street", "View", "House", "Numbers", "(", "SVHN", ")", "and", "CIFAR", "-", "10", ".", "The", "images", "generated", "by", "the", "network", "are", "always", "novel", "(", "not", "simply", "copies", "of", "training", "examples", ")", ",", "and", "are", "virtually", "indistinguishable", "from", "real", "data", "for", "MNIST", "and", "SVHN", ";", "the", "generated", "CIFAR", "images", "are", "somewhat", "blurry", ",", "but", "still", "contain", "recognisable", "structure", "from", "natural", "scenes", ".", "The", "binarized", "MNIST", "results", "substantially", "improve", "on", "the", "state", "of", "the", "art", ".", "As", "a", "preliminary", "exercise", ",", "we", "also", "evaluate", "the", "2D", "attention", "module", "of", "the", "DRAW", "network", "on", "cluttered", "MNIST", "classification", ".", "For", "all", "experiments", ",", "the", "model", "of", "the", "input", "data", "was", "a", "Bernoulli", "distribution", "with", "means", "given", "by", ".", "For", "the", "MNIST", "experiments", ",", "the", "reconstruction", "loss", "from", "Eq", "[", "reference", "]", "was", "the", "usual", "binary", "cross", "-", "entropy", "term", ".", "For", "the", "SVHN", "and", "CIFAR", "-", "10", "experiments", ",", "the", "red", ",", "green", "and", "blue", "pixel", "intensities", "were", "represented", "as", "numbers", "between", "0", "and", "1", ",", "which", "were", "then", "interpreted", "as", "independent", "colour", "emission", "probabilities", ".", "The", "reconstruction", "loss", "was", "therefore", "the", "cross", "-", "entropy", "between", "the", "pixel", "intensities", "and", "the", "model", "probabilities", ".", "Although", "this", "approach", "worked", "well", "in", "practice", ",", "it", "means", "that", "the", "training", "loss", "did", "not", "correspond", "to", "the", "true", "compression", "cost", "of", "RGB", "images", ".", "Network", "hyper", "-", "parameters", "for", "all", "the", "experiments", "are", "presented", "in", "Table", "[", "reference", "]", ".", "The", "Adam", "optimisation", "algorithm", "was", "used", "throughout", ".", "Examples", "of", "generation", "sequences", "for", "MNIST", "and", "SVHN", "are", "provided", "in", "the", "accompanying", "video", "(", ")", ".", "subsection", ":", "Cluttered", "MNIST", "Classification", "To", "test", "the", "classification", "efficacy", "of", "the", "DRAW", "attention", "mechanism", "(", "as", "opposed", "to", "its", "ability", "to", "aid", "in", "image", "generation", ")", ",", "we", "evaluate", "its", "performance", "on", "the", "cluttered", "translated", "MNIST", "task", ".", "Each", "image", "in", "cluttered", "MNIST", "contains", "many", "digit", "-", "like", "fragments", "of", "visual", "clutter", "that", "the", "network", "must", "distinguish", "from", "the", "true", "digit", "to", "be", "classified", ".", "As", "illustrated", "in", "Fig", ".", "[", "reference", "]", ",", "having", "an", "iterative", "attention", "model", "allows", "the", "network", "to", "progressively", "zoom", "in", "on", "the", "relevant", "region", "of", "the", "image", ",", "and", "ignore", "the", "clutter", "outside", "it", ".", "Our", "model", "consists", "of", "an", "LSTM", "recurrent", "network", "that", "receives", "a", "‘", "glimpse", "’", "from", "the", "input", "image", "at", "each", "time", "-", "step", ",", "using", "the", "selective", "read", "operation", "defined", "in", "Section", "[", "reference", "]", ".", "After", "a", "fixed", "number", "of", "glimpses", "the", "network", "uses", "a", "softmax", "layer", "to", "classify", "the", "MNIST", "digit", ".", "The", "network", "is", "similar", "to", "the", "recently", "introduced", "Recurrent", "Attention", "Model", "(", "RAM", ")", ",", "except", "that", "our", "attention", "method", "is", "differentiable", ";", "we", "therefore", "refer", "to", "it", "as", "“", "Differentiable", "RAM", "”", ".", "The", "results", "in", "Table", "[", "reference", "]", "demonstrate", "a", "significant", "improvement", "in", "test", "error", "over", "the", "original", "RAM", "network", ".", "Moreover", "our", "model", "had", "only", "a", "single", "attention", "patch", "at", "each", "time", "-", "step", ",", "whereas", "RAM", "used", "four", ",", "at", "different", "zooms", ".", "subsection", ":", "MNIST", "Generation", "We", "trained", "the", "full", "DRAW", "network", "as", "a", "generative", "model", "on", "the", "binarized", "MNIST", "dataset", ".", "This", "dataset", "has", "been", "widely", "studied", "in", "the", "literature", ",", "allowing", "us", "to", "compare", "the", "numerical", "performance", "(", "measured", "in", "average", "nats", "per", "image", "on", "the", "test", "set", ")", "of", "DRAW", "with", "existing", "methods", ".", "Table", "[", "reference", "]", "shows", "that", "DRAW", "without", "selective", "attention", "performs", "comparably", "to", "other", "recent", "generative", "models", "such", "as", "DARN", ",", "NADE", "and", "DBMs", ",", "and", "that", "DRAW", "with", "attention", "considerably", "improves", "on", "the", "state", "of", "the", "art", ".", "Once", "the", "DRAW", "network", "was", "trained", ",", "we", "generated", "MNIST", "digits", "following", "the", "method", "in", "Section", "[", "reference", "]", ",", "examples", "of", "which", "are", "presented", "in", "Fig", ".", "[", "reference", "]", ".", "Fig", ".", "[", "reference", "]", "illustrates", "the", "image", "generation", "sequence", "for", "a", "DRAW", "network", "without", "selective", "attention", "(", "see", "Section", "[", "reference", "]", ")", ".", "It", "is", "interesting", "to", "compare", "this", "with", "the", "generation", "sequence", "for", "DRAW", "with", "attention", ",", "as", "depicted", "in", "Fig", ".", "[", "reference", "]", ".", "Whereas", "without", "attention", "it", "progressively", "sharpens", "a", "blurred", "image", "in", "a", "global", "way", ",", "with", "attention", "it", "constructs", "the", "digit", "by", "tracing", "the", "lines", "—", "much", "like", "a", "person", "with", "a", "pen", ".", "subsection", ":", "MNIST", "Generation", "with", "Two", "Digits", "The", "main", "motivation", "for", "using", "an", "attention", "-", "based", "generative", "model", "is", "that", "large", "images", "can", "be", "built", "up", "iteratively", ",", "by", "adding", "to", "a", "small", "part", "of", "the", "image", "at", "a", "time", ".", "To", "test", "this", "capability", "in", "a", "controlled", "fashion", ",", "we", "trained", "DRAW", "to", "generate", "images", "with", "two", "MNIST", "images", "chosen", "at", "random", "and", "placed", "at", "random", "locations", "in", "a", "black", "background", ".", "In", "cases", "where", "the", "two", "digits", "overlap", ",", "the", "pixel", "intensities", "were", "added", "together", "at", "each", "point", "and", "clipped", "to", "be", "no", "greater", "than", "one", ".", "Examples", "of", "generated", "data", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "The", "network", "typically", "generates", "one", "digit", "and", "then", "the", "other", ",", "suggesting", "an", "ability", "to", "recreate", "composite", "scenes", "from", "simple", "pieces", ".", "subsection", ":", "Street", "View", "House", "Number", "Generation", "MNIST", "digits", "are", "very", "simplistic", "in", "terms", "of", "visual", "structure", ",", "and", "we", "were", "keen", "to", "see", "how", "well", "DRAW", "performed", "on", "natural", "images", ".", "Our", "first", "natural", "image", "generation", "experiment", "used", "the", "multi", "-", "digit", "Street", "View", "House", "Numbers", "dataset", ".", "We", "used", "the", "same", "preprocessing", "as", ",", "yielding", "a", "house", "number", "image", "for", "each", "training", "example", ".", "The", "network", "was", "then", "trained", "using", "patches", "extracted", "at", "random", "locations", "from", "the", "preprocessed", "images", ".", "The", "SVHN", "training", "set", "contains", "231", ",", "053", "images", ",", "and", "the", "validation", "set", "contains", "4", ",", "701", "images", ".", "The", "house", "number", "images", "generated", "by", "the", "network", "are", "highly", "realistic", ",", "as", "shown", "in", "Figs", ".", "[", "reference", "]", "and", "[", "reference", "]", ".", "Fig", ".", "[", "reference", "]", "reveals", "that", ",", "despite", "the", "long", "training", "time", ",", "the", "DRAW", "network", "underfit", "the", "SVHN", "training", "data", ".", "s", "subsection", ":", "Generating", "CIFAR", "Images", "The", "most", "challenging", "dataset", "we", "applied", "DRAW", "to", "was", "the", "CIFAR", "-", "10", "collection", "of", "natural", "images", ".", "CIFAR", "-", "10", "is", "very", "diverse", ",", "and", "with", "only", "50", ",", "000", "training", "examples", "it", "is", "very", "difficult", "to", "generate", "realistic", "-", "looking", "objects", "without", "overfitting", "(", "in", "other", "words", ",", "without", "copying", "from", "the", "training", "set", ")", ".", "Nonetheless", "the", "images", "in", "Fig", ".", "[", "reference", "]", "demonstrate", "that", "DRAW", "is", "able", "to", "capture", "much", "of", "the", "shape", ",", "colour", "and", "composition", "of", "real", "photographs", ".", "section", ":", "Conclusion", "This", "paper", "introduced", "the", "Deep", "Recurrent", "Attentive", "Writer", "(", "DRAW", ")", "neural", "network", "architecture", ",", "and", "demonstrated", "its", "ability", "to", "generate", "highly", "realistic", "natural", "images", "such", "as", "photographs", "of", "house", "numbers", ",", "as", "well", "as", "improving", "on", "the", "best", "known", "results", "for", "binarized", "MNIST", "generation", ".", "We", "also", "established", "that", "the", "two", "-", "dimensional", "differentiable", "attention", "mechanism", "embedded", "in", "DRAW", "is", "beneficial", "not", "only", "to", "image", "generation", ",", "but", "also", "to", "image", "classification", ".", "section", ":", "Acknowledgments", "Of", "the", "many", "who", "assisted", "in", "creating", "this", "paper", ",", "we", "are", "especially", "thankful", "to", "Koray", "Kavukcuoglu", ",", "Volodymyr", "Mnih", ",", "Jimmy", "Ba", ",", "Yaroslav", "Bulatov", ",", "Greg", "Wayne", ",", "Andrei", "Rusu", "and", "Shakir", "Mohamed", ".", "bibliography", ":", "References"], "entity": [[4, 8, ["Deep", "Recurrent", "Attentive", "Writer"], "Method"], [9, 10, ["DRAW"], "Method"], [11, 14, ["neural", "network", "architecture"], "Method"], [15, 17, ["image", "generation"], "Task"], [23, 26, ["spatial", "attention", "mechanism"], "Method"], [37, 43, ["sequential", "variational", "auto", "-", "encoding", "framework"], "Method"], [47, 52, ["iterative", "construction", "of", "complex", "images"], "Task"], [64, 66, ["generative", "models"], "Method"], [67, 68, ["MNIST"], "Material"], [182, 185, ["generative", "neural", "networks"], "Method"], [208, 212, ["iterative", "self", "-", "correction"], "Task"], [215, 219, ["one", "shot", "”", "approach"], "Method"], [229, 233, ["Deep", "Recurrent", "Attentive", "Writer"], "Method"], [234, 235, ["DRAW"], "Method"], [246, 248, ["image", "construction"], "Task"], [278, 281, ["recurrent", "neural", "networks"], "Method"], [283, 285, ["encoder", "network"], "Method"], [296, 297, ["decoder"], "Method"], [315, 318, ["stochastic", "gradient", "descent"], "Method"], [344, 348, ["variational", "auto", "-", "encoders"], "Method"], [352, 356, ["hybrid", "of", "deep", "learning"], "Method"], [357, 359, ["variational", "inference"], "Method"], [366, 368, ["generative", "modelling"], "Task"], [399, 400, ["decoder"], "Method"], [408, 409, ["encoder"], "Method"], [476, 479, ["sequential", "attention", "models"], "Method"], [490, 493, ["reinforcement", "learning", "techniques"], "Method"], [495, 497, ["policy", "gradients"], "Method"], [499, 501, ["attention", "model"], "Method"], [517, 518, ["backpropagation"], "Method"], [533, 536, ["Neural", "Turing", "Machine"], "Method"], [548, 550, ["loss", "function"], "Method"], [552, 553, ["training"], "Task"], [557, 559, ["image", "generation"], "Task"], [566, 569, ["selective", "attention", "model"], "Method"], [576, 580, ["reading", "and", "modifying", "images"], "Task"], [590, 591, ["MNIST"], "Material"], [592, 596, ["Street", "View", "House", "Numbers"], "Material"], [597, 601, ["CIFAR", "-", "10", "datasets"], "Material"], [665, 669, ["variational", "auto", "-", "encoders"], "Method"], [671, 673, ["encoder", "network"], "Method"], [675, 679, ["distribution", "over", "latent", "codes"], "Method"], [689, 691, ["decoder", "network"], "Method"], [695, 697, ["code", "distribuion"], "Method"], [719, 722, ["encoder", "and", "decoder"], "Method"], [723, 725, ["recurrent", "networks"], "Method"], [803, 807, ["dynamically", "updated", "attention", "mechanism"], "Method"], [827, 828, ["decoder"], "Method"], [874, 879, ["feedforward", "variational", "auto", "-", "encoder"], "Method"], [882, 884, ["Network", "Architecture"], "Method"], [891, 893, ["encoder", "network"], "Method"], [916, 917, ["decoder"], "Method"], [926, 929, ["encoder", "and", "decoder"], "Method"], [934, 937, ["recurrent", "neural", "network"], "Method"], [944, 950, ["Long", "Short", "-", "Term", "Memory", "architecture"], "Method"], [951, 953, ["LSTM", ";"], "Method"], [967, 968, ["LSTM"], "Method"], [976, 980, ["long", "-", "range", "dependencies"], "Task"], [1015, 1016, ["encoder"], "Method"], [1055, 1056, ["encoder"], "Method"], [1075, 1077, ["diagonal", "Gaussian"], "Method"], [1078, 1080, ["Bernoulli", "distributions"], "Method"], [1084, 1085, ["Gaussians"], "Method"], [1089, 1092, ["auto", "-", "encoders"], "Method"], [1098, 1100, ["Gaussian", "latents"], "Method"], [1125, 1127, ["reparameterization", "trick"], "Method"], [1169, 1170, ["decoder"], "Method"], [1217, 1218, ["reconstruction"], "Task"], [1274, 1277, ["logistic", "sigmoid", "function"], "Method"], [1341, 1343, ["Loss", "Function"], "Method"], [1370, 1372, ["Bernoulli", "distribution"], "Method"], [1378, 1380, ["reconstruction", "loss"], "Metric"], [1436, 1438, ["latent", "distribution"], "Method"], [1440, 1442, ["diagonal", "Gaussian"], "Method"], [1462, 1464, ["standard", "Gaussian"], "Method"], [1510, 1514, ["stochastic", "gradient", "descent", "step"], "Method"], [1551, 1552, ["decoder"], "Method"], [1557, 1559, ["total", "loss"], "Metric"], [1577, 1580, ["Stochastic", "Data", "Generation"], "Task"], [1601, 1602, ["decoder"], "Method"], [1624, 1625, ["encoder"], "Method"], [1629, 1631, ["image", "generation"], "Task"], [1634, 1638, ["Read", "and", "Write", "Operations"], "Task"], [1741, 1742, ["encoder"], "Method"], [1761, 1762, ["decoder"], "Method"], [1784, 1787, ["selective", "attention", "mechanism"], "Method"], [1815, 1818, ["Selective", "Attention", "Model"], "Method"], [1830, 1833, ["gradient", "descent", "training"], "Method"], [1839, 1842, ["differentiable", "attention", "mechanisms"], "Method"], [1845, 1847, ["handwriting", "synthesis"], "Task"], [1848, 1851, ["Neural", "Turing", "Machines"], "Task"], [1863, 1867, ["dimensional", "form", "of", "attention"], "Method"], [1870, 1875, ["array", "of", "2D", "Gaussian", "filters"], "Method"], [1914, 1919, ["computer", "graphics", "-", "based", "autoencoders"], "Method"], [1930, 1934, ["grid", "of", "Gaussian", "filters"], "Method"], [2044, 2046, ["attention", "model"], "Method"], [2052, 2054, ["Gaussian", "filters"], "Method"], [2083, 2085, ["linear", "transformation"], "Method"], [2124, 2127, ["randomly", "initialised", "network"], "Method"], [2142, 2143, ["decoder"], "Method"], [2145, 2150, ["horizontal", "and", "vertical", "filterbank", "matrices"], "Method"], [2189, 2192, ["Reading", "and", "Writing"], "Task"], [2211, 2213, ["read", "operation"], "Method"], [2230, 2231, ["filterbanks"], "Method"], [2243, 2245, ["write", "operation"], "Task"], [2330, 2331, ["DRAW"], "Method"], [2333, 2337, ["realistic", "-", "looking", "images"], "Task"], [2348, 2349, ["MNIST"], "Material"], [2350, 2354, ["Street", "View", "House", "Numbers"], "Material"], [2355, 2356, ["SVHN"], "Material"], [2358, 2361, ["CIFAR", "-", "10"], "Material"], [2388, 2389, ["MNIST"], "Material"], [2390, 2391, ["SVHN"], "Material"], [2394, 2396, ["CIFAR", "images"], "Material"], [2410, 2412, ["binarized", "MNIST"], "Material"], [2431, 2434, ["2D", "attention", "module"], "Method"], [2455, 2457, ["Bernoulli", "distribution"], "Method"], [2468, 2470, ["reconstruction", "loss"], "Metric"], [2478, 2483, ["binary", "cross", "-", "entropy", "term"], "Metric"], [2486, 2487, ["SVHN"], "Material"], [2488, 2492, ["CIFAR", "-", "10", "experiments"], "Material"], [2550, 2552, ["training", "loss"], "Metric"], [2558, 2560, ["compression", "cost"], "Metric"], [2561, 2563, ["RGB", "images"], "Material"], [2581, 2584, ["Adam", "optimisation", "algorithm"], "Method"], [2593, 2594, ["MNIST"], "Material"], [2595, 2596, ["SVHN"], "Material"], [2613, 2614, ["classification"], "Task"], [2629, 2631, ["image", "generation"], "Task"], [2682, 2685, ["iterative", "attention", "model"], "Method"], [2712, 2715, ["LSTM", "recurrent", "network"], "Method"], [2753, 2755, ["softmax", "layer"], "Method"], [2769, 2772, ["Recurrent", "Attention", "Model"], "Method"], [2773, 2775, ["RAM", ")"], "Method"], [2779, 2781, ["attention", "method"], "Method"], [2791, 2793, ["Differentiable", "RAM"], "Method"], [2807, 2809, ["test", "error"], "Metric"], [2812, 2814, ["RAM", "network"], "Method"], [2851, 2853, ["generative", "model"], "Method"], [2879, 2881, ["average", "nats"], "Metric"], [2909, 2911, ["generative", "models"], "Method"], [2913, 2914, ["DARN"], "Method"], [2915, 2916, ["NADE"], "Method"], [2917, 2918, ["DBMs"], "Method"], [2923, 2924, ["attention"], "Method"], [3060, 3065, ["attention", "-", "based", "generative", "model"], "Method"], [3183, 3188, ["Street", "View", "House", "Number", "Generation"], "Task"], [3334, 3337, ["Generating", "CIFAR", "Images"], "Task"], [3355, 3358, ["CIFAR", "-", "10"], "Material"], [3406, 3407, ["DRAW"], "Method"], [3430, 3434, ["Deep", "Recurrent", "Attentive", "Writer"], "Method"], [3435, 3436, ["DRAW"], "Method"], [3479, 3483, ["dimensional", "differentiable", "attention", "mechanism"], "Method"], [3485, 3486, ["DRAW"], "Method"], [3491, 3493, ["image", "generation"], "Task"], [3497, 3499, ["image", "classification"], "Task"], [18, 19, ["DRAW"], "Method"], [75, 79, ["Street", "View", "House", "Numbers"], "Material"], [97, 98, ["DRAW"], "Method"], [165, 167, ["image", "generation"], "Task"], [272, 273, ["DRAW"], "Method"], [370, 371, ["DRAW"], "Method"], [502, 503, ["DRAW"], "Method"], [542, 543, ["DRAW"], "Method"], [640, 641, ["DRAW"], "Method"], [650, 651, ["DRAW"], "Method"], [657, 658, ["DRAW"], "Method"], [726, 727, ["DRAW"], "Method"], [1245, 1246, ["DRAW"], "Method"], [1587, 1588, ["DRAW"], "Method"], [1639, 1640, ["DRAW"], "Method"], [1696, 1697, ["DRAW"], "Method"], [1797, 1799, ["image", "generation"], "Task"], [1808, 1809, ["DRAW"], "Method"], [1904, 1905, ["DRAW"], "Method"], [2436, 2437, ["DRAW"], "Method"], [2440, 2441, ["MNIST"], "Material"], [2464, 2465, ["MNIST"], "Material"], [2608, 2609, ["MNIST"], "Material"], [2617, 2618, ["DRAW"], "Method"], [2641, 2642, ["MNIST"], "Material"], [2648, 2649, ["MNIST"], "Material"], [2758, 2759, ["MNIST"], "Material"], [2841, 2842, ["MNIST"], "Material"], [2847, 2848, ["DRAW"], "Method"], [2856, 2857, ["MNIST"], "Material"], [2889, 2890, ["DRAW"], "Method"], [2900, 2901, ["DRAW"], "Method"], [2921, 2922, ["DRAW"], "Method"], [2935, 2936, ["DRAW"], "Method"], [2942, 2943, ["MNIST"], "Material"], [2972, 2974, ["image", "generation"], "Task"], [2977, 2978, ["DRAW"], "Method"], [3001, 3002, ["DRAW"], "Method"], [3049, 3050, ["MNIST"], "Material"], [3099, 3100, ["DRAW"], "Method"], [3105, 3106, ["MNIST"], "Material"], [3188, 3189, ["MNIST"], "Material"], [3207, 3208, ["DRAW"], "Method"], [3216, 3218, ["image", "generation"], "Task"], [3224, 3228, ["Street", "View", "House", "Numbers"], "Material"], [3264, 3265, ["SVHN"], "Material"], [3323, 3324, ["DRAW"], "Method"], [3327, 3328, ["SVHN"], "Material"], [3343, 3344, ["DRAW"], "Method"], [3347, 3350, ["CIFAR", "-", "10"], "Material"], [3469, 3470, ["MNIST"], "Material"]]} {"doc_id": "27e98e09cf09bc13c913d01676e5f32624011050", "text": ["document", ":", "U", "-", "Net", ":", "Machine", "Reading", "Comprehension", "with", "Unanswerable", "Questions", "Machine", "reading", "comprehension", "with", "unanswerable", "questions", "is", "a", "new", "challenging", "task", "for", "natural", "language", "processing", ".", "A", "key", "subtask", "is", "to", "reliably", "predict", "whether", "the", "question", "is", "unanswerable", ".", "In", "this", "paper", ",", "we", "propose", "a", "unified", "model", ",", "called", "U", "-", "Net", ",", "with", "three", "important", "components", ":", "answer", "pointer", ",", "no", "-", "answer", "pointer", ",", "and", "answer", "verifier", ".", "We", "introduce", "a", "universal", "node", "and", "thus", "process", "the", "question", "and", "its", "context", "passage", "as", "a", "single", "contiguous", "sequence", "of", "tokens", ".", "The", "universal", "node", "encodes", "the", "fused", "information", "from", "both", "the", "question", "and", "passage", ",", "and", "plays", "an", "important", "role", "to", "predict", "whether", "the", "question", "is", "answerable", "and", "also", "greatly", "improves", "the", "conciseness", "of", "the", "U", "-", "Net", ".", "Different", "from", "the", "state", "-", "of", "-", "art", "pipeline", "models", ",", "U", "-", "Net", "can", "be", "learned", "in", "an", "end", "-", "to", "-", "end", "fashion", ".", "The", "experimental", "results", "on", "the", "SQuAD", "2.0", "dataset", "show", "that", "U", "-", "Net", "can", "effectively", "predict", "the", "unanswerability", "of", "questions", "and", "achieves", "an", "F1", "score", "of", "71.7", "on", "SQuAD", "2.0", ".", "section", ":", "Introduction", "Machine", "reading", "comprehension", "(", "MRC", ")", "is", "a", "challenging", "task", "in", "natural", "language", "processing", ",", "which", "requires", "that", "machine", "can", "read", ",", "understand", ",", "and", "answer", "questions", "about", "a", "text", ".", "Benefiting", "from", "the", "rapid", "development", "of", "deep", "learning", "techniques", "and", "large", "-", "scale", "benchmarks", ",", "the", "end", "-", "to", "-", "end", "neural", "methods", "have", "achieved", "promising", "results", "on", "MRC", "task", ".", "The", "best", "systems", "have", "even", "surpassed", "human", "performance", "on", "the", "Stanford", "Question", "Answering", "Dataset", "(", "SQuAD", ")", ",", "one", "of", "the", "most", "widely", "used", "MRC", "benchmarks", ".", "However", ",", "one", "of", "the", "limitations", "of", "the", "SQuAD", "task", "is", "that", "each", "question", "has", "a", "correct", "answer", "in", "the", "context", "passage", ",", "therefore", "most", "models", "just", "need", "to", "select", "the", "most", "relevant", "text", "span", "as", "the", "answer", ",", "without", "necessarily", "checking", "whether", "it", "is", "indeed", "the", "answer", "to", "the", "question", ".", "To", "remedy", "the", "deficiency", "of", "SQuAD", ",", "squad2.0", "squad2.0", "developed", "SQuAD", "2.0", "that", "combines", "SQuAD", "with", "new", "unanswerable", "questions", ".", "Table", "[", "reference", "]", "shows", "two", "examples", "of", "unanswerable", "questions", ".", "The", "new", "dataset", "requires", "the", "MRC", "systems", "to", "know", "what", "they", "do", "n’t", "know", ".", "Paragraph", ":", "“", "…", "Other", "legislation", "followed", ",", "including", "the", "Migratory", "Bird", "Conservation", "Act", "of", "1929", ",", "a", "1937", "treaty", "prohibiting", "the", "hunting", "of", "right", "and", "gray", "whales", ",", "and", "the", "Bald", "Eagle", "Protection", "Act", "of", "1940", ".", "These", "later", "laws", "had", "a", "low", "cost", "to", "society—the", "species", "were", "relatively", "rare—and", "little", "opposition", "was", "raised.”", "To", "do", "well", "on", "MRC", "with", "unanswerable", "questions", ",", "the", "model", "needs", "to", "comprehend", "the", "question", ",", "reason", "among", "the", "passage", ",", "judge", "the", "unanswerability", "and", "then", "identify", "the", "answer", "span", ".", "Since", "extensive", "work", "has", "been", "done", "on", "how", "to", "correctly", "predict", "the", "answer", "span", "when", "the", "question", "is", "answerable", "(", "e.g.", ",", "SQuAD", "1.1", ")", ",", "the", "main", "challenge", "of", "this", "task", "lies", "in", "how", "to", "reliably", "determine", "whether", "a", "question", "is", "not", "answerable", "from", "the", "passage", ".", "There", "are", "two", "kinds", "of", "approaches", "to", "model", "the", "answerability", "of", "a", "question", ".", "One", "approach", "is", "to", "directly", "extend", "previous", "MRC", "models", "by", "introducing", "a", "no", "-", "answer", "score", "to", "the", "score", "vector", "of", "the", "answer", "span", ".", "But", "this", "kind", "of", "approaches", "is", "relatively", "simple", "and", "can", "not", "effectively", "model", "the", "answerability", "of", "a", "question", ".", "Another", "approach", "introduces", "an", "answer", "verifier", "to", "determine", "whether", "the", "question", "is", "unanswerable", ".", "However", ",", "this", "kind", "of", "approaches", "usually", "has", "a", "pipeline", "structure", ".", "The", "answer", "pointer", "and", "answer", "verifier", "have", "their", "respective", "models", ",", "which", "are", "trained", "separately", ".", "Intuitively", ",", "it", "is", "unnecessary", "since", "the", "underlying", "comprehension", "and", "reasoning", "of", "language", "for", "these", "components", "is", "the", "same", ".", "In", "this", "paper", ",", "we", "decompose", "the", "problem", "of", "MRC", "with", "unanswerable", "questions", "into", "three", "sub", "-", "tasks", ":", "answer", "pointer", ",", "no", "-", "answer", "pointer", ",", "and", "answer", "verifier", ".", "Since", "these", "three", "sub", "-", "tasks", "are", "highly", "related", ",", "we", "regard", "the", "MRC", "with", "unanswerable", "questions", "as", "a", "multi", "-", "task", "learning", "problem", "by", "sharing", "some", "meta", "-", "knowledge", ".", "We", "propose", "the", "U", "-", "Net", "to", "incorporate", "these", "three", "sub", "-", "tasks", "into", "a", "unified", "model", ":", "1", ")", "an", "answer", "pointer", "to", "predict", "a", "candidate", "answer", "span", "for", "a", "question", ";", "2", ")", "a", "no", "-", "answer", "pointer", "to", "avoid", "selecting", "any", "text", "span", "when", "a", "question", "has", "no", "answer", ";", "and", "3", ")", "an", "answer", "verifier", "to", "determine", "the", "probability", "of", "the", "“", "unanswerability", "”", "of", "a", "question", "with", "candidate", "answer", "information", ".", "Additionally", ",", "we", "also", "introduce", "a", "universal", "node", "and", "process", "the", "question", "and", "its", "context", "passage", "as", "a", "single", "contiguous", "sequence", "of", "tokens", ",", "which", "greatly", "improves", "the", "conciseness", "of", "U", "-", "Net", ".", "The", "universal", "node", "acts", "on", "both", "question", "and", "passage", "to", "learn", "whether", "the", "question", "is", "answerable", ".", "Different", "from", "the", "previous", "pipeline", "models", ",", "U", "-", "Net", "can", "be", "learned", "in", "an", "end", "-", "to", "-", "end", "fashion", ".", "Our", "experimental", "results", "on", "the", "SQuAD", "2.0", "dataset", "show", "that", "U", "-", "Net", "effectively", "predicts", "the", "unanswerability", "of", "questions", "and", "achieves", "an", "F1", "score", "of", "72.6", ".", "The", "contributions", "of", "this", "paper", "can", "be", "summarized", "as", "follows", ".", "We", "decompose", "the", "problem", "of", "MRC", "with", "unanswerable", "questions", "into", "three", "sub", "-", "tasks", "and", "combine", "them", "into", "a", "unified", "model", ",", "which", "uses", "the", "shared", "encoding", "and", "interaction", "layers", ".", "Thus", ",", "the", "three", "-", "tasks", "can", "be", "trained", "simultaneously", "in", "an", "end", "-", "to", "-", "end", "fashion", ".", "We", "introduce", "a", "universal", "node", "to", "encode", "the", "common", "information", "of", "the", "question", "and", "passage", ".", "Thus", ",", "we", "can", "use", "a", "unified", "representation", "to", "model", "the", "question", "and", "passage", ",", "which", "makes", "our", "model", "more", "condensed", ".", "U", "-", "Net", "is", "very", "easy", "to", "implement", "yet", "effective", ".", "section", ":", "Proposed", "Model", "Formally", ",", "we", "can", "represent", "the", "MRC", "problem", "as", ":", "given", "a", "set", "of", "tuples", ",", "where", "is", "the", "question", "with", "words", ",", "is", "the", "context", "passage", "with", "words", ",", "and", "is", "the", "answer", "with", "and", "indicating", "the", "start", "and", "end", "points", ",", "the", "task", "is", "to", "estimate", "the", "conditional", "probability", ".", "The", "architecture", "of", "our", "proposed", "U", "-", "Net", "is", "illustrated", "in", "Figure", "[", "reference", "]", ".", "U", "-", "Net", "consists", "of", "four", "major", "blocks", ":", "Unified", "Encoding", ",", "Multi", "-", "Level", "Attention", ",", "Final", "Fusion", ",", "and", "Prediction", ".", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "we", "first", "combine", "the", "embedded", "representation", "of", "the", "question", "and", "passage", "with", "a", "universal", "node", "and", "pass", "them", "through", "a", "BiLSTM", "to", "encode", "the", "whole", "text", ".", "We", "then", "use", "the", "encoded", "representation", "to", "do", "the", "information", "interaction", ".", "Then", "we", "use", "the", "encoded", "and", "interacted", "representation", "to", "fuse", "the", "full", "representation", "and", "feed", "them", "into", "the", "final", "prediction", "layers", "to", "do", "the", "multi", "-", "task", "training", ".", "We", "will", "describe", "our", "model", "in", "details", "in", "the", "following", ".", "subsection", ":", "(", "A", ")", "Unified", "Encoding", "paragraph", ":", "Embedding", "Following", "the", "successful", "models", "on", "SQuAD", "1.1", ",", "we", "first", "embed", "both", "the", "question", "and", "the", "passage", "with", "the", "following", "features", ".", "Glove", "embedding", "and", "Elmo", "embedding", "are", "used", "as", "basic", "embeddings", ".", "Besides", ",", "we", "use", "POS", "embedding", ",", "NER", "embedding", ",", "and", "a", "feature", "embedding", "that", "includes", "the", "exact", "match", ",", "lower", "-", "case", "match", ",", "lemma", "match", ",", "and", "a", "TF", "-", "IDF", "feature", ".", "Now", "we", "get", "the", "question", "representation", "and", "the", "passage", "representation", ",", "where", "each", "word", "is", "represented", "as", "a", "-", "dim", "embedding", "by", "combining", "the", "features", "/", "embedding", "described", "above", ".", "paragraph", ":", "Universal", "Node", "We", "create", "a", "universal", "node", ",", "which", "is", "a", "key", "factor", "in", "our", "model", "and", "has", "several", "roles", "in", "predicting", "the", "unanswerability", "of", "question", ".", "We", "expect", "this", "node", "to", "learn", "universal", "information", "from", "both", "passage", "and", "question", ".", "This", "universal", "node", "is", "added", "and", "connects", "the", "passage", "and", "question", "at", "the", "phase", "of", "embedding", ",", "and", "then", "goes", "along", "with", "the", "whole", "representation", ",", "so", "it", "is", "a", "key", "factor", "in", "information", "representation", ".", "Since", "the", "universal", "node", "is", "in", "between", "and", "later", "shared", "between", "passage", "and", "question", ",", "it", "has", "an", "abstract", "semantic", "meaning", "rather", "than", "just", "a", "word", "embedding", ".", "Also", ",", "the", "universal", "node", "is", "later", "shared", "in", "the", "attention", "interaction", "mechanism", "and", "used", "in", "both", "the", "answer", "boundary", "detection", "and", "classification", "tasks", ",", "so", "this", "node", "carries", "massive", "information", "and", "has", "several", "important", "roles", "in", "our", "whole", "model", "construction", ".", "The", "universal", "node", "is", "first", "represented", "by", "a", "-", "dim", "randomly", "-", "initialized", "vector", ".", "We", "concatenated", "question", "representation", ",", "universal", "node", "representation", ",", "passage", "representation", "together", "as", ":", "is", "a", "joint", "representation", "of", "question", "and", "passage", ".", "paragraph", ":", "Word", "-", "level", "Fusion", "Then", "we", "first", "use", "two", "-", "layer", "bidirectional", "LSTM", "(", "BiLSTM", ")", "to", "fuse", "the", "joint", "representation", "of", "question", ",", "universal", "node", ",", "and", "passage", ".", "where", "is", "the", "hidden", "states", "of", "the", "first", "BiLSTM", ",", "representing", "the", "low", "-", "level", "semantic", "information", ",", "and", "is", "the", "hidden", "states", "of", "the", "second", "BiLSTM", ",", "representing", "the", "high", "-", "level", "semantic", "information", ".", "Finally", ",", "we", "concatenate", "and", "together", "and", "pass", "them", "through", "the", "third", "BiLSTM", "and", "obtain", "a", "full", "representation", ".", "Thus", ",", "represents", "the", "deep", "fusion", "information", "of", "the", "question", "and", "passage", "on", "word", "-", "level", ".", "When", "a", "BiLSTM", "is", "applied", "to", "encode", "representations", ",", "it", "learns", "the", "semantic", "information", "bi", "-", "directionally", ".", "Since", "the", "universal", "node", "is", "between", "the", "question", "and", "passage", ",", "its", "hidden", "states", "can", "learn", "both", "question", "and", "passage", "information", ".", "When", "the", "passage", "-", "question", "pair", "was", "encoded", "as", "a", "unified", "representation", "and", "information", "flows", "via", "the", "BiLSTM", ",", "the", "universal", "node", "has", "an", "important", "role", "in", "information", "representation", ".", "subsection", ":", "(", "B", ")", "Multi", "-", "Level", "Attention", "To", "fully", "fuse", "the", "semantic", "representation", "of", "the", "question", "and", "passage", ",", "we", "use", "the", "attention", "mechanism", "to", "capture", "their", "interactions", "on", "different", "levels", ".", "We", "expected", "that", "we", "could", "simply", "use", "self", "-", "attention", "on", "the", "encoded", "representation", "for", "interaction", "between", "question", "and", "passage", ",", "which", "contains", "both", "bi", "-", "attention", "and", "self", "-", "attention", "of", "the", "question", "and", "passage", ".", "But", "we", "found", "that", "it", "performed", "slightly", "worse", "than", "the", "traditional", "bi", "-", "directional", "attention", "with", "the", "universal", "node", "included", ".", "Therefore", ",", "we", "use", "a", "bi", "-", "directional", "attention", "between", "the", "question", "and", "passage", ".", "We", "first", "divide", "into", "two", "representations", ":", "attached", "passage", "and", "attached", "question", ",", "and", "let", "the", "universal", "node", "representation", "attached", "to", "both", "the", "passage", "and", "question", ",", "i.e.", ",", "Note", "is", "shared", "by", "and", ".", "Here", "the", "universal", "node", "works", "as", "a", "special", "information", "carrier", ",", "and", "both", "passage", "and", "question", "can", "focus", "attention", "information", "on", "this", "node", "so", "that", "the", "connection", "between", "them", "is", "closer", "than", "a", "traditional", "bi", "-", "attention", "interaction", ".", "Since", "both", "and", "are", "concatenated", "by", "three", "-", "level", "representations", ",", "we", "followed", "previous", "work", "FusionNet", "to", "construct", "their", "iterations", "on", "three", "levels", ".", "Take", "the", "first", "level", "as", "an", "example", ".", "We", "first", "compute", "the", "affine", "matrix", "of", "and", "by", "where", ";", "and", "are", "learnable", "parameters", ".", "Next", ",", "a", "bi", "-", "directional", "attention", "is", "used", "to", "compute", "the", "interacted", "representation", "and", ".", "where", "is", "column", "-", "wise", "normalized", "function", ".", "We", "use", "the", "same", "attention", "layer", "to", "model", "the", "interactions", "for", "all", "the", "three", "levels", ",", "and", "get", "the", "final", "fused", "representation", "for", "the", "question", "and", "passage", "respectively", ".", "Note", "that", "while", "dealing", "with", "the", "attention", "output", "of", "the", "universal", "node", ",", "we", "added", "two", "outputs", "from", "passage", "-", "to", "-", "question", "attention", "and", "question", "-", "to", "-", "passage", "attention", ".", "So", "after", "the", "interaction", ",", "the", "fused", "representation", "still", "have", "the", "same", "length", "as", "the", "encoded", "representation", ",", "and", ".", "subsection", ":", "(", "C", ")", "Final", "Fusion", "After", "the", "three", "-", "level", "attentive", "interaction", ",", "we", "generate", "the", "final", "fused", "information", "for", "the", "question", "and", "passage", ".", "We", "concatenate", "all", "the", "history", "information", ":", "we", "first", "concatenate", "the", "encoded", "representation", "and", "the", "representation", "after", "attention", "(", "again", ",", "we", "use", ",", "and", "to", "represent", "3", "different", "levels", "of", "representation", "for", "the", "two", "previous", "steps", "respectively", ")", ".", "Following", "the", "success", "of", "DenseNet", ",", "we", "concatenate", "the", "input", "and", "output", "of", "each", "layer", "as", "the", "input", "of", "the", "next", "layer", ".", "First", ",", "we", "pass", "the", "concatenated", "representation", "through", "a", "BiLSTM", "to", "get", ".", "where", "the", "representation", "is", "a", "fusion", "of", "information", "from", "different", "levels", ".", "Then", "we", "concatenate", "the", "original", "embedded", "representation", "and", "for", "better", "representation", "of", "the", "fused", "information", "of", "passage", ",", "universal", "node", ",", "and", "question", ".", "Finally", ",", "we", "use", "a", "self", "-", "attention", "layer", "to", "get", "the", "attention", "information", "within", "the", "fused", "information", ".", "The", "self", "-", "attention", "layer", "is", "constructed", "the", "same", "way", "as", ":", "where", "is", "the", "representation", "after", "self", "-", "attention", "of", "the", "fused", "information", ".", "Next", "we", "concatenated", "representation", "and", "and", "pass", "them", "through", "another", "BiLSTM", "layer", ":", "Now", "is", "the", "final", "fused", "representation", "of", "all", "the", "information", ".", "At", "this", "point", ",", "we", "divide", "into", "two", "parts", ":", ",", ",", "representing", "the", "fused", "information", "of", "the", "question", "and", "passage", "respectively", ".", "Note", "for", "the", "final", "representation", ",", "we", "attach", "the", "universal", "node", "only", "in", "the", "passage", "representation", ".", "This", "is", "because", "we", "need", "the", "universal", "node", "as", "a", "focus", "for", "the", "pointer", "when", "the", "question", "is", "unanswerable", ".", "These", "will", "be", "fed", "into", "the", "next", "decoder", "prediction", "layer", ".", "subsection", ":", "(", "D", ")", "Prediction", "The", "prediction", "layer", "receives", "fused", "information", "of", "passage", "and", "question", ",", "and", "tackles", "three", "prediction", "tasks", ":", "(", "1", ")", "answer", "pointer", ",", "(", "2", ")", "no", "-", "answer", "pointer", "and", "(", "3", ")", "answer", "verifier", ".", "First", ",", "we", "use", "a", "function", "shown", "below", "to", "summarize", "the", "question", "information", "into", "a", "fixed", "-", "dim", "representation", ".", "where", "is", "a", "learnable", "weight", "matrix", "and", "represents", "the", "word", "in", "the", "question", "representation", ".", "Then", "we", "feed", "into", "the", "answer", "pointer", "to", "find", "boundaries", "of", "answers", ",", "and", "the", "classification", "layer", "to", "distinguish", "whether", "the", "question", "is", "answerable", ".", "paragraph", ":", "(", "i", ")", "Answer", "Pointer", "We", "use", "this", "answer", "pointer", "to", "detect", "the", "answer", "boundaries", "from", "the", "passage", "when", "the", "question", "is", "answerable", "(", "i.e.", ",", "the", "answer", "is", "a", "span", "in", "the", "passage", ")", ".", "This", "layer", "is", "a", "classic", "pointer", "net", "structure", ".", "We", "use", "two", "trainable", "matrices", "and", "to", "estimate", "the", "probability", "of", "the", "answer", "start", "and", "end", "boundaries", "of", "the", "word", "in", "the", "passage", ",", "and", ".", "Note", "here", "when", "the", "question", "is", "answerable", ",", "we", "do", "not", "consider", "the", "universal", "node", "in", "answer", "boundary", "detection", ",", "so", "we", "have", "(", "is", "the", "universal", "node", "in", "the", "passage", "representation", ")", ".", "The", "loss", "function", "for", "the", "answerable", "question", "pairs", "is", ":", "where", "and", "are", "the", "ground", "-", "truth", "of", "the", "start", "and", "end", "boundary", "of", "the", "answer", ".", "paragraph", ":", "(", "ii", ")", "No", "-", "Answer", "Pointer", "Then", "we", "use", "the", "same", "pointer", "for", "questions", "that", "are", "not", "answerable", ".", "Here", "the", "loss", "is", ":", "and", "correspond", "to", "the", "position", "of", "the", "universal", "node", ",", "which", "is", "at", "the", "front", "of", "the", "passage", "representation", ".", "For", "this", "scenario", ",", "the", "loss", "is", "calculated", "for", "the", "universal", "node", ".", "Additionally", ",", "since", "there", "exits", "a", "plausible", "answer", "for", "each", "unanswerable", "question", "in", "SQuAD", "2.0", ",", "we", "introduce", "an", "auxiliary", "plausible", "answer", "pointer", "to", "predict", "the", "boundaries", "of", "the", "plausible", "answers", ".", "The", "plausible", "answer", "pointer", "has", "the", "same", "structure", "as", "the", "answer", "pointer", ",", "but", "with", "different", "parameters", ".", "Thus", ",", "the", "total", "loss", "function", "is", ":", "where", "and", "are", "the", "output", "of", "the", "plausible", "answer", "pointer", ";", "and", "are", "the", "start", "and", "end", "boundary", "of", "the", "unanswerable", "answer", ".", "The", "no", "-", "answer", "pointer", "and", "plausible", "answer", "pointer", "are", "removed", "at", "test", "phase", ".", "paragraph", ":", "(", "iii", ")", "Answer", "Verifier", "We", "use", "the", "answer", "verifier", "to", "distinguish", "whether", "the", "question", "is", "answerable", ".", "Answer", "verifier", "applies", "a", "weighted", "summary", "layer", "to", "summarize", "the", "passage", "information", "into", "a", "fixed", "-", "dim", "representation", "(", "as", "shown", "in", "Eq", ".", "(", "[", "reference", "]", ")", ")", ".", "And", "we", "use", "the", "weight", "matrix", "obtained", "from", "the", "answer", "pointer", "to", "get", "two", "representations", "of", "the", "passage", ".", "Then", "we", "use", "the", "universal", "node", "and", "concatenate", "it", "with", "the", "summary", "of", "question", "and", "passage", "to", "make", "a", "fixed", "vector", "This", "fixed", "includes", "the", "representation", "representing", "the", "question", "information", ",", "and", "and", "representing", "the", "passage", "information", ".", "Since", "these", "representations", "are", "highly", "summarized", "specially", "for", "classification", ",", "we", "believe", "that", "this", "passage", "-", "question", "pair", "contains", "information", "to", "distinguish", "whether", "this", "question", "is", "answerable", ".", "In", "addition", ",", "we", "include", "the", "universal", "node", "as", "a", "supplement", ".", "Since", "the", "universal", "node", "is", "pointed", "at", "when", "the", "question", "is", "unanswerable", "and", "this", "node", "itself", "already", "contains", "information", "collected", "from", "both", "the", "passage", "and", "question", "during", "encoding", "and", "information", "interaction", ",", "we", "believe", "that", "this", "node", "is", "important", "in", "distinguishing", "whether", "the", "question", "is", "answerable", ".", "Finally", ",", "we", "pass", "this", "fixed", "vector", "through", "a", "linear", "layer", "to", "obtain", "the", "prediction", "whether", "the", "question", "is", "answerable", ".", "where", "is", "a", "sigmoid", "function", ",", "is", "a", "learnable", "weight", "matrix", ".", "Here", "we", "use", "the", "cross", "-", "entropy", "loss", "in", "training", ".", "where", "indicates", "whether", "the", "question", "has", "an", "answer", "in", "the", "passage", ".", "Compared", "with", "other", "relatively", "complex", "structures", "developped", "for", "this", "MRC", "task", ",", "our", "U", "-", "Net", "model", "passes", "the", "original", "question", "and", "passage", "pair", "through", "embedding", "and", "encoding", ",", "which", "then", "interacts", "with", "each", "other", ",", "yielding", "fused", "information", "merged", "from", "all", "the", "levels", ".", "The", "entire", "architecture", "is", "very", "easy", "to", "construct", ".", "After", "we", "have", "the", "fused", "representation", "of", "the", "question", "and", "passage", ",", "we", "pass", "them", "through", "the", "pointer", "layer", "and", "a", "fused", "information", "classification", "layer", "in", "a", "multi", "-", "task", "setup", ".", "section", ":", "Training", "We", "jointly", "train", "the", "three", "tasks", "by", "combining", "the", "three", "loss", "functions", ".", "The", "final", "loss", "function", "is", ":", "where", "indicates", "whether", "the", "question", "has", "an", "answer", "in", "the", "passage", ",", ",", "and", "are", "the", "three", "loss", "functions", "of", "the", "answer", "pointer", ",", "no", "-", "answer", "pointer", ",", "and", "answer", "verifier", ".", "Although", "the", "three", "tasks", "could", "have", "different", "weights", "in", "the", "final", "loss", "function", "and", "be", "further", "fine", "-", "tuned", "after", "joint", "training", ",", "here", "we", "just", "consider", "them", "in", "the", "same", "weight", "and", "do", "not", "fine", "-", "tune", "them", "individually", ".", "At", "the", "test", "phase", ",", "we", "first", "use", "the", "answer", "pointer", "to", "find", "a", "potential", "answer", "to", "the", "question", ",", "while", "the", "verifier", "layer", "judges", "whether", "the", "question", "is", "answerable", ".", "If", "the", "classifier", "predicts", "the", "question", "is", "unanswerable", ",", "we", "consider", "the", "answer", "extracted", "by", "the", "answer", "pointer", "as", "plausible", ".", "In", "this", "way", ",", "we", "get", "the", "system", "result", ".", "section", ":", "Experiment", "subsection", ":", "Datasets", "Recently", ",", "machine", "reading", "comprehension", "and", "question", "answering", "have", "progressed", "rapidly", ",", "owing", "to", "the", "computation", "ability", "and", "publicly", "available", "high", "-", "quality", "datasets", "such", "as", "SQuAD", ".", "Now", "new", "research", "efforts", "have", "been", "devoted", "to", "the", "newly", "released", "answer", "extraction", "test", "with", "unanswerable", "questions", ",", "SQuAD", "2.0", ".", "It", "is", "constructed", "by", "combining", "question", "-", "answer", "pairs", "selected", "from", "SQuAD", "1.0", "and", "newly", "crafted", "unanswerable", "questions", ".", "These", "unanswerable", "questions", "are", "created", "by", "workers", "that", "were", "asked", "to", "pose", "questions", "that", "can", "not", "be", "answered", "based", "on", "the", "paragraph", "alone", "but", "are", "similar", "to", "the", "answerable", "questions", ".", "It", "is", "very", "difficult", "to", "distinguish", "these", "questions", "from", "the", "answerable", "ones", ".", "We", "evaluate", "our", "model", "using", "this", "data", "set", ".", "It", "contains", "over", "100", ",", "000", "+", "questions", "on", "500", "+", "wikipedia", "articles", ".", "subsection", ":", "Implementation", "Details", "We", "use", "Spacy", "to", "process", "each", "question", "and", "passage", "to", "obtain", "tokens", ",", "POS", "tags", ",", "NER", "tags", "and", "lemmas", "tags", "of", "each", "text", ".", "We", "use", "12", "dimensions", "to", "embed", "POS", "tags", ",", "8", "for", "NER", "tags", ".", "We", "use", "3", "binary", "features", ":", "exact", "match", ",", "lower", "-", "case", "match", "and", "lemma", "match", "between", "the", "question", "and", "passage", ".", "We", "use", "100", "-", "dim", "Glove", "pretrained", "word", "embeddings", "and", "1024", "-", "dim", "Elmo", "embeddings", ".", "All", "the", "LSTM", "blocks", "are", "bi", "-", "directional", "with", "one", "single", "layer", ".", "We", "set", "the", "hidden", "layer", "dimension", "as", "125", ",", "attention", "layer", "dimension", "as", "250", ".", "We", "added", "a", "dropout", "layer", "over", "all", "the", "modeling", "layers", ",", "including", "the", "embedding", "layer", ",", "at", "a", "dropout", "rate", "of", "0.3", ".", "We", "use", "Adam", "optimizer", "with", "a", "learning", "rate", "of", "0.002", ".", "During", "training", ",", "we", "omit", "passage", "with", "over", "400", "words", "and", "question", "with", "more", "than", "50", "words", ".", "For", "testing", ",", "when", "the", "passage", "has", "over", "600", "words", "and", "the", "question", "is", "over", "100", "words", ",", "we", "simply", "label", "these", "questions", "as", "unanswerable", ".", "subsection", ":", "Main", "Results", "Our", "model", "achieves", "an", "F1", "score", "of", "74.0", "and", "an", "EM", "score", "of", "70.3", "on", "the", "development", "set", ",", "and", "an", "F1", "score", "of", "72.6", "and", "an", "EM", "score", "of", "69.2", "on", "Test", "set", ",", "as", "shown", "in", "Table", "[", "reference", "]", ".", "Our", "model", "outperforms", "most", "of", "the", "previous", "approaches", ".", "Comparing", "to", "the", "best", "-", "performing", "systems", ",", "our", "model", "has", "a", "simple", "architecture", "and", "is", "an", "end", "-", "to", "-", "end", "model", ".", "In", "fact", ",", "among", "all", "the", "end", "-", "to", "-", "end", "models", ",", "we", "achieve", "the", "best", "F1", "scores", ".", "We", "believe", "that", "the", "performance", "of", "the", "U", "-", "Net", "can", "be", "boosted", "with", "an", "additional", "post", "-", "processing", "step", "to", "verify", "answers", "using", "approaches", "such", "as", ".", "subsection", ":", "Ablation", "Study", "We", "also", "do", "an", "ablation", "study", "on", "the", "SQuAD", "2.0", "development", "set", "to", "further", "test", "the", "effectiveness", "of", "different", "components", "in", "our", "model", ".", "In", "Table", "[", "reference", "]", ",", "we", "show", "four", "different", "configurations", ".", "First", ",", "we", "remove", "the", "universal", "node", ".", "We", "let", "the", "negative", "examples", "focus", "on", "the", "plausible", "answer", "spans", "instead", "of", "focusing", "on", "the", "universal", "node", ".", "This", "results", "in", "a", "loss", "of", "2.6", "%", "F1", "score", "on", "the", "development", "set", ",", "showing", "that", "the", "universal", "node", "indeed", "learns", "information", "about", "whether", "the", "question", "is", "answerable", ".", "We", "also", "tried", "to", "make", "the", "universal", "node", "only", "attached", "to", "the", "passage", "representation", "when", "passing", "the", "attention", "layer", ".", "Our", "results", "showed", "that", "when", "node", "is", "shared", ",", "as", "it", "is", "called", "‘", "universal", "’", ",", "it", "learns", "information", "interaction", "between", "the", "question", "and", "passage", ",", "and", "when", "it", "is", "not", "shared", ",", "the", "performance", "slightly", "degraded", ".", "As", "for", "the", "approaches", "to", "encode", "the", "representations", ",", "we", "pass", "both", "the", "question", "and", "passage", "through", "a", "shared", "BiLSTM", ".", "To", "test", "the", "effectiveness", "of", "this", ",", "we", "ran", "the", "experiment", "using", "separate", "BiLSTMs", "on", "embedded", "question", "and", "passage", "representations", ".", "Results", "show", "that", "the", "performance", "dropped", "slightly", ",", "suggesting", "sharing", "BiLSTM", "is", "an", "effective", "method", "to", "improve", "the", "quality", "of", "the", "encoder", ".", "After", "removing", "the", "plausible", "answer", "pointer", ",", "the", "performance", "also", "dropped", ",", "indicating", "the", "plausible", "answers", "are", "useful", "to", "improve", "the", "model", "even", "though", "they", "are", "incorrect", ".", "After", "removing", "the", "answer", "verifier", ",", "the", "performance", "dropped", "greatly", ",", "indicating", "it", "is", "vital", "for", "our", "model", ".", "Lastly", ",", "we", "run", "a", "test", "using", "a", "more", "concise", "configuration", ".", "In", "the", "second", "block", "(", "multi", "-", "level", "attention", ")", "of", "the", "U", "-", "Net", ",", "we", "do", "not", "split", "the", "output", "of", "the", "encoded", "presentation", "and", "let", "it", "pass", "through", "a", "self", "-", "attention", "layer", ".", "The", "bidirectional", "attention", "is", "removed", ".", "In", "this", "way", ",", "our", "model", "uses", "only", "one", "unified", "representation", "of", "the", "question", "and", "passage", "at", "all", "time", ".", "We", "simply", "pass", "this", "representation", "layer", "by", "layer", "to", "get", "the", "final", "result", ".", "Compared", "to", "the", "bi", "-", "attention", "model", ",", "the", "F1", "-", "score", "decreases", "0.5", "%", ".", "subsection", ":", "Multi", "-", "task", "Study", "We", "also", "run", "an", "experiment", "to", "test", "the", "performance", "of", "our", "multi", "-", "task", "model", ".", "We", "select", "different", "losses", "that", "participate", "in", "the", "training", "procedure", "to", "observe", "the", "performance", "affected", "by", "answer", "boundary", "detect", "or", "classification", ".", "Table", "[", "reference", "]", "shows", "the", "performance", ".", "Here", "we", "use", "and", "to", "represent", "the", "EM", "and", "F1", "score", "when", "the", "classification", "is", "not", "part", "of", "the", "task", ",", "which", "makes", "it", "very", "much", "like", "the", "task", "in", "SQuAD", "1.1", ".", "To", "test", "our", "classifier", "performance", ",", "we", "do", "not", "use", "backward", "propagation", "over", "the", "loss", "of", "answer", "boundary", "detection", "and", "simply", "run", "a", "classification", "task", ".", "Results", "(", "the", "first", "two", "rows", "in", "Table", "[", "reference", "]", ")", "show", "that", "there", "is", "a", "large", "gain", "when", "using", "the", "multi", "-", "task", "model", ".", "The", "answer", "boundary", "detection", "task", "helps", "the", "encoder", "learn", "information", "between", "the", "passage", "and", "question", "and", "also", "feed", "information", "into", "the", "universal", "node", ",", "therefore", "we", "can", "use", "a", "summarized", "representation", "of", "the", "passage", "and", "question", "as", "well", "as", "the", "universal", "node", "to", "distinguish", "whether", "the", "question", "is", "answerable", ",", "i.e.", ",", "help", "improve", "classification", ".", "For", "the", "answer", "boundary", "detection", "task", ",", "we", "find", "that", "the", "multi", "-", "task", "setup", "(", "i.e.", ",", "the", "classification", "layer", "participates", "in", "the", "training", "process", ")", "does", "not", "help", "its", "performance", ".", "Since", "the", "classifier", "and", "pointer", "layer", "shared", "the", "encoding", "process", ",", "we", "originally", "expected", "that", "classification", "information", "can", "help", "detect", "answer", "boundaries", ".", "But", "this", "is", "not", "the", "case", ".", "We", "think", "this", "is", "also", "reasonable", "since", "distinguishing", "whether", "the", "question", "is", "answerable", "is", "mainly", "focusing", "on", "the", "interactions", "between", "the", "passage", "-", "question", "pair", ",", "so", "once", "the", "question", "is", "predicted", "as", "answerable", "or", "not", ",", "it", "has", "nothing", "to", "do", "with", "the", "answer", "boundaries", ".", "This", "is", "consistent", "with", "how", "human", "-", "beings", "do", "this", "classification", "task", ".", "We", "also", "run", "the", "test", "over", "SQuAD", "1.1", "development", "test", "to", "evaluate", "the", "performance", ".", "Due", "to", "a", "condensed", "structure", ",", "our", "model", "achieves", "an", "score", "of", "less", "than", "86", "%", ",", "which", "is", "not", "a", "very", "competitive", "score", "on", "SQuAD", "1.1", "test", ".", "But", "as", "shown", "above", ",", "our", "model", "achieves", "a", "good", "score", "in", "SQuAD", "2.0", "test", ",", "which", "shows", "this", "model", "has", "the", "potential", "to", "achieve", "higher", "performance", "by", "making", "progress", "on", "both", "the", "answer", "detection", "and", "classification", "tasks", ".", "Overall", ",", "we", "can", "conclude", "that", "our", "multi", "-", "task", "model", "works", "well", "since", "the", "performance", "of", "unanswerability", "classification", "improves", "significantly", "when", "the", "answer", "pointer", "and", "answer", "verifier", "work", "simultaneously", ".", "subsection", ":", "Study", "on", "the", "Different", "Thresholds", "of", "Unanswerability", "Classification", "The", "output", "of", "the", "answer", "verifier", "is", "the", "probability", "of", "a", "question", "being", "unanswerable", ".", "The", "smaller", "the", "output", ",", "the", "lower", "the", "probability", "of", "unanswerability", "is", ".", "In", "SQuAD", "2.0", ",", "the", "proportions", "of", "unanswerable", "questions", "are", "different", "in", "the", "training", "and", "test", "sets", ".", "The", "default", "threshold", "is", "optimized", "on", "the", "training", "set", ",", "but", "not", "suitable", "for", "the", "test", "set", ".", "Therefore", ",", "it", "is", "reasonable", "to", "set", "a", "proper", "threshold", "to", "manually", "adapt", "to", "the", "test", "set", ".", "As", "mentioned", "in", "SQuAD", "2.0", "paper", ",", "different", "thresholds", "for", "answerability", "prediction", "result", "in", "fluctuated", "scores", "between", "answerable", "and", "unanswerable", "questions", ".", "Here", "we", "show", "the", "variation", "of", "the", "F1", "score", "with", "different", "thresholds", "in", "Figure", ".", "The", "threshold", "between", "is", "used", "to", "decide", "whether", "a", "question", "can", "be", "answered", ".", "When", "the", "threshold", "is", "set", "to", ",", "all", "questions", "are", "considered", "as", "answerable", ".", "As", "we", "can", "see", ",", "when", "the", "threshold", "is", "set", "to", "0.5", ",", "F1", "score", "of", "answerable", "questions", "is", "similar", "to", "that", "of", "unanswerable", "questions", ".", "When", "we", "increase", "the", "threshold", "(", "i.e.", ",", "more", "likely", "to", "predict", "the", "question", "as", "unanswerable", ")", ",", "performance", "for", "answerable", "questions", "degrades", ",", "and", "improves", "for", "unanswerable", "questions", ".", "This", "is", "as", "expected", ".", "We", "can", "see", "that", "the", "overall", "score", "is", "slightly", "better", ",", "which", "is", "consistent", "with", "the", "idea", "from", "SQuAD", "2.0", ".", "In", "addition", ",", "we", "find", "that", "for", "larger", "thresholds", ",", "the", "variance", "between", "and", "is", "narrowed", "since", "and", "scores", "for", "unanswerable", "questions", "are", "the", "same", ".", "Finally", ",", "we", "set", "the", "threshold", "to", "be", "for", "the", "submission", "system", "to", "SQuAD", "evaluation", ".", "section", ":", "Related", "Work", "subsection", ":", "End", "-", "to", "-", "end", "Models", "for", "MRC", "Currently", ",", "end", "-", "to", "-", "end", "neural", "network", "models", "have", "achieved", "great", "successes", "for", "machine", "reading", "comprehension", ".", "Most", "of", "these", "models", "consist", "of", "three", "components", ":", "encoder", ",", "interaction", ",", "and", "pointer", ".", "The", "BiLSTM", "is", "widely", "used", "for", "encoding", "the", "embedded", "representation", ".", "For", "the", "interaction", ",", "bidirectional", "attention", "mechanism", "is", "very", "effective", "to", "fuse", "information", "of", "the", "question", "and", "passage", ".", "Finally", ",", "a", "pointer", "network", "is", "used", "to", "predict", "the", "span", "boundaries", "of", "the", "answer", ".", "Specifically", ",", "in", "SQuAD", "test", ",", "there", "are", "approaches", "to", "combine", "match", "-", "LSTM", "and", "pointer", "networks", "to", "produce", "boundaries", "of", "the", "answer", "and", "employ", "variant", "bidirectional", "attention", "mechanism", "to", "match", "the", "question", "and", "passage", "mutually", ".", "In", "our", "model", ",", "we", "learn", "from", "previous", "work", "and", "develop", "a", "condensed", "end", "-", "to", "-", "end", "model", "for", "the", "SQuAD", "2.0", "task", ".", "Different", "from", "the", "previous", "models", ",", "we", "use", "a", "unified", "representation", "to", "encode", "the", "question", "and", "passage", "simultaneously", ",", "and", "introduce", "a", "universal", "node", "to", "encode", "the", "fused", "information", "of", "the", "question", "and", "passage", ",", "which", "also", "plays", "an", "important", "role", "to", "predict", "the", "unanswerability", "of", "a", "question", ".", "subsection", ":", "MRC", "with", "Unanswerable", "Questions", "MRC", "with", "unanswerable", "questions", "is", "a", "more", "challenging", "task", ".", "Previous", "work", "levy2017zero", ",", "clark2017simple", "levy2017zero", ",", "clark2017simple", "has", "attempted", "to", "normalize", "a", "no", "-", "answer", "score", "depending", "on", "the", "probability", "of", "all", "answer", "spans", "and", "still", "detect", "boundaries", "at", "the", "same", "time", ".", "But", "the", "scores", "of", "the", "answer", "span", "predictions", "are", "not", "very", "discriminative", "in", "distinguishing", "whether", "the", "question", "is", "answerable", ".", "Therefore", ",", "this", "kind", "of", "approaches", ",", "though", "relatively", "simple", ",", "can", "not", "effectively", "deal", "with", "the", "answerability", "of", "a", "question", ".", "hu2018read", ",", "tan2018know", "hu2018read", ",", "tan2018know", "introduced", "an", "answer", "verifier", "idea", "to", "construct", "a", "classification", "layer", ".", "However", ",", "this", "kind", "of", "approaches", "usually", "has", "a", "pipeline", "structure", ".", "The", "answer", "pointer", "and", "answer", "verifier", "have", "their", "respective", "models", "that", "are", "trained", "separately", ".", "paragraph", ":", "Multi", "-", "task", "models", "Different", "from", "existing", "work", ",", "we", "regard", "the", "MRC", "with", "unanswerable", "questions", "as", "a", "multi", "-", "task", "learning", "problem", "by", "sharing", "some", "meta", "-", "knowledge", ".", "Intuitively", ",", "answer", "prediction", "and", "answer", "verification", "are", "related", "tasks", "since", "the", "underlying", "comprehension", "and", "reasoning", "of", "language", "for", "these", "components", "is", "the", "same", ".", "Therefore", ",", "we", "construct", "a", "multi", "-", "task", "model", "to", "solve", "three", "sub", "-", "tasks", ":", "answer", "pointer", ",", "no", "-", "answer", "pointer", ",", "and", "answer", "verifier", ".", "section", ":", "Conclusion", "and", "Future", "Work", "In", "this", "paper", ",", "we", "regard", "the", "MRC", "with", "unanswerable", "questions", "as", "multi", "-", "task", "learning", "problems", "and", "propose", "the", "U", "-", "Net", ",", "a", "simple", "end", "-", "to", "-", "end", "model", "for", "MRC", "challenges", ".", "U", "-", "Net", "has", "good", "performance", "on", "SQuAD", "2.0", ".", "We", "first", "add", "a", "universal", "node", "to", "learn", "a", "fused", "representation", "from", "both", "the", "question", "and", "passage", ",", "then", "use", "a", "concatenated", "representation", "to", "pass", "through", "encoding", "layers", ".", "We", "only", "treat", "question", "and", "passage", "differently", "during", "attention", "interactions", ".", "In", "the", "rest", "blocks", "of", "U", "-", "Net", ",", "we", "still", "use", "the", "unified", "representation", "containing", "both", "the", "question", "and", "passage", "representation", ".", "Finally", ",", "we", "train", "the", "U", "-", "Net", "as", "a", "multi", "-", "task", "framework", "to", "determine", "the", "final", "answer", "boundaries", "as", "well", "as", "whether", "the", "question", "is", "answerable", ".", "Our", "model", "has", "very", "simple", "structure", "yet", "achieves", "good", "results", "on", "SQuAD", "2.0", "test", ".", "Our", "future", "work", "is", "to", "reconstruct", "the", "structure", "of", "U", "-", "Net", "by", "replacing", "the", "current", "multi", "-", "level", "attention", "block", "with", "a", "simpler", "self", "-", "attention", "mechanism", ",", "which", "we", "believe", "can", "capture", "the", "question", "and", "passage", "information", ",", "and", "intuitively", "is", "also", "coherent", "with", "the", "rest", "of", "our", "U", "-", "Net", "model", ".", "In", "addition", ",", "we", "will", "improve", "the", "answer", "boundary", "detection", "performance", "based", "on", "some", "of", "the", "previous", "successful", "models", ".", "Since", "our", "model", "actually", "does", "not", "achieve", "very", "competitive", "performance", "in", "the", "boundary", "detection", "task", "yet", "still", "has", "a", "good", "overall", "performance", "on", "SQuAD", "2.0", "test", ",", "we", "are", "optimistic", "that", "our", "U", "-", "Net", "model", "is", "potentially", "capable", "of", "achieving", "better", "performance", ".", "Furthermore", ",", "our", "model", "has", "a", "simple", "structure", "and", "is", "easy", "to", "implement", ",", "therefore", "we", "believe", "that", "our", "model", "can", "be", "easily", "modified", "for", "various", "datasets", ".", "section", ":", "Acknowledgement", "We", "would", "like", "to", "thank", "Robin", "Jia", ",", "Pranav", "Rajpurkar", "for", "their", "help", "with", "SQuAD", "2.0", "submissions", ".", "bibliography", ":", "References"], "entity": [[2, 5, ["U", "-", "Net"], "Method"], [6, 9, ["Machine", "Reading", "Comprehension"], "Task"], [12, 15, ["Machine", "reading", "comprehension"], "Task"], [24, 27, ["natural", "language", "processing"], "Task"], [48, 50, ["unified", "model"], "Method"], [52, 55, ["U", "-", "Net"], "Method"], [70, 72, ["answer", "verifier"], "Method"], [126, 127, ["conciseness"], "Metric"], [129, 132, ["U", "-", "Net"], "Method"], [141, 143, ["pipeline", "models"], "Method"], [144, 147, ["U", "-", "Net"], "Method"], [164, 167, ["SQuAD", "2.0", "dataset"], "Material"], [169, 172, ["U", "-", "Net"], "Method"], [182, 184, ["F1", "score"], "Metric"], [187, 189, ["SQuAD", "2.0"], "Material"], [193, 196, ["Machine", "reading", "comprehension"], "Task"], [197, 198, ["MRC"], "Task"], [204, 207, ["natural", "language", "processing"], "Task"], [230, 233, ["deep", "learning", "techniques"], "Method"], [240, 247, ["end", "-", "to", "-", "end", "neural", "methods"], "Method"], [252, 254, ["MRC", "task"], "Task"], [265, 269, ["Stanford", "Question", "Answering", "Dataset"], "Material"], [270, 272, ["SQuAD", ")"], "Material"], [279, 281, ["MRC", "benchmarks"], "Material"], [341, 342, ["squad2.0"], "Material"], [344, 346, ["SQuAD", "2.0"], "Material"], [348, 349, ["SQuAD"], "Material"], [370, 372, ["MRC", "systems"], "Method"], [439, 440, ["MRC"], "Task"], [536, 538, ["MRC", "models"], "Method"], [577, 579, ["answer", "verifier"], "Method"], [596, 598, ["pipeline", "structure"], "Method"], [603, 605, ["answer", "verifier"], "Method"], [644, 648, ["MRC", "with", "unanswerable", "questions"], "Task"], [650, 653, ["sub", "-", "tasks"], "Task"], [663, 665, ["answer", "verifier"], "Method"], [679, 680, ["MRC"], "Task"], [685, 690, ["multi", "-", "task", "learning", "problem"], "Task"], [700, 703, ["U", "-", "Net"], "Method"], [718, 720, ["answer", "pointer"], "Method"], [754, 756, ["answer", "verifier"], "Method"], [803, 806, ["U", "-", "Net"], "Method"], [828, 830, ["pipeline", "models"], "Method"], [831, 834, ["U", "-", "Net"], "Method"], [851, 854, ["SQuAD", "2.0", "dataset"], "Material"], [856, 859, ["U", "-", "Net"], "Method"], [862, 865, ["unanswerability", "of", "questions"], "Task"], [868, 870, ["F1", "score"], "Metric"], [889, 890, ["MRC"], "Task"], [909, 914, ["shared", "encoding", "and", "interaction", "layers"], "Method"], [956, 958, ["unified", "representation"], "Method"], [972, 975, ["U", "-", "Net"], "Method"], [993, 995, ["MRC", "problem"], "Task"], [1044, 1047, ["U", "-", "Net"], "Method"], [1055, 1058, ["U", "-", "Net"], "Method"], [1064, 1066, ["Unified", "Encoding"], "Method"], [1067, 1071, ["Multi", "-", "Level", "Attention"], "Task"], [1072, 1074, ["Final", "Fusion"], "Task"], [1076, 1077, ["Prediction"], "Task"], [1090, 1092, ["embedded", "representation"], "Method"], [1106, 1107, ["BiLSTM"], "Method"], [1117, 1119, ["encoded", "representation"], "Method"], [1122, 1124, ["information", "interaction"], "Task"], [1129, 1133, ["encoded", "and", "interacted", "representation"], "Method"], [1136, 1138, ["full", "representation"], "Method"], [1144, 1146, ["prediction", "layers"], "Method"], [1149, 1153, ["multi", "-", "task", "training"], "Task"], [1170, 1172, ["Unified", "Encoding"], "Method"], [1174, 1175, ["Embedding"], "Task"], [1180, 1182, ["SQuAD", "1.1"], "Material"], [1197, 1199, ["Glove", "embedding"], "Method"], [1200, 1202, ["Elmo", "embedding"], "Method"], [1205, 1207, ["basic", "embeddings"], "Method"], [1215, 1217, ["NER", "embedding"], "Method"], [1220, 1222, ["feature", "embedding"], "Method"], [1238, 1242, ["TF", "-", "IDF", "feature"], "Method"], [1247, 1249, ["question", "representation"], "Method"], [1251, 1253, ["passage", "representation"], "Method"], [1261, 1264, ["-", "dim", "embedding"], "Method"], [1267, 1270, ["features", "/", "embedding"], "Method"], [1275, 1277, ["Universal", "Node"], "Method"], [1298, 1301, ["unanswerability", "of", "question"], "Task"], [1331, 1332, ["embedding"], "Task"], [1349, 1351, ["information", "representation"], "Task"], [1377, 1379, ["word", "embedding"], "Method"], [1390, 1393, ["attention", "interaction", "mechanism"], "Method"], [1398, 1404, ["answer", "boundary", "detection", "and", "classification", "tasks"], "Task"], [1419, 1421, ["model", "construction"], "Method"], [1439, 1441, ["question", "representation"], "Method"], [1442, 1445, ["universal", "node", "representation"], "Method"], [1446, 1448, ["passage", "representation"], "Method"], [1462, 1466, ["Word", "-", "level", "Fusion"], "Method"], [1473, 1475, ["bidirectional", "LSTM"], "Method"], [1476, 1477, ["BiLSTM"], "Method"], [1500, 1501, ["BiLSTM"], "Method"], [1518, 1519, ["BiLSTM"], "Method"], [1540, 1541, ["BiLSTM"], "Method"], [1544, 1546, ["full", "representation"], "Method"], [1566, 1567, ["BiLSTM"], "Method"], [1614, 1616, ["unified", "representation"], "Method"], [1621, 1622, ["BiLSTM"], "Method"], [1631, 1633, ["information", "representation"], "Task"], [1639, 1643, ["Multi", "-", "Level", "Attention"], "Task"], [1658, 1660, ["attention", "mechanism"], "Method"], [1675, 1678, ["self", "-", "attention"], "Method"], [1680, 1682, ["encoded", "representation"], "Method"], [1716, 1720, ["bi", "-", "directional", "attention"], "Method"], [1757, 1760, ["universal", "node", "representation"], "Method"], [1810, 1814, ["bi", "-", "attention", "interaction"], "Method"], [1823, 1825, ["level", "representations"], "Method"], [1830, 1831, ["FusionNet"], "Method"], [1866, 1870, ["bi", "-", "directional", "attention"], "Method"], [1875, 1877, ["interacted", "representation"], "Method"], [1891, 1893, ["attention", "layer"], "Method"], [1907, 1909, ["fused", "representation"], "Method"], [1954, 1956, ["fused", "representation"], "Method"], [1963, 1965, ["encoded", "representation"], "Method"], [2006, 2008, ["encoded", "representation"], "Method"], [2039, 2040, ["DenseNet"], "Method"], [2063, 2065, ["concatenated", "representation"], "Method"], [2067, 2068, ["BiLSTM"], "Method"], [2088, 2090, ["embedded", "representation"], "Method"], [2112, 2116, ["self", "-", "attention", "layer"], "Method"], [2127, 2131, ["self", "-", "attention", "layer"], "Method"], [2168, 2170, ["fused", "representation"], "Method"], [2212, 2214, ["passage", "representation"], "Method"], [2242, 2245, ["decoder", "prediction", "layer"], "Method"], [2251, 2252, ["Prediction"], "Task"], [2253, 2255, ["prediction", "layer"], "Method"], [2266, 2268, ["prediction", "tasks"], "Task"], [2286, 2288, ["answer", "verifier"], "Method"], [2304, 2308, ["fixed", "-", "dim", "representation"], "Method"], [2321, 2323, ["question", "representation"], "Method"], [2339, 2341, ["classification", "layer"], "Method"], [2392, 2395, ["pointer", "net", "structure"], "Method"], [2399, 2401, ["trainable", "matrices"], "Method"], [2438, 2441, ["answer", "boundary", "detection"], "Task"], [2452, 2454, ["passage", "representation"], "Method"], [2457, 2459, ["loss", "function"], "Method"], [2527, 2529, ["passage", "representation"], "Method"], [2556, 2558, ["SQuAD", "2.0"], "Material"], [2596, 2599, ["total", "loss", "function"], "Metric"], [2649, 2651, ["answer", "verifier"], "Method"], [2659, 2661, ["Answer", "verifier"], "Method"], [2663, 2666, ["weighted", "summary", "layer"], "Method"], [2673, 2677, ["fixed", "-", "dim", "representation"], "Method"], [2755, 2756, ["classification"], "Task"], [2816, 2818, ["information", "interaction"], "Task"], [2843, 2845, ["linear", "layer"], "Method"], [2871, 2875, ["cross", "-", "entropy", "loss"], "Metric"], [2876, 2877, ["training"], "Task"], [2899, 2901, ["MRC", "task"], "Task"], [2948, 2950, ["fused", "representation"], "Method"], [2961, 2963, ["pointer", "layer"], "Method"], [2965, 2969, ["fused", "information", "classification", "layer"], "Method"], [2971, 2975, ["multi", "-", "task", "setup"], "Task"], [2989, 2991, ["loss", "functions"], "Method"], [2994, 2996, ["loss", "function"], "Metric"], [3028, 3030, ["answer", "verifier"], "Method"], [3042, 3044, ["loss", "function"], "Metric"], [3051, 3053, ["joint", "training"], "Task"], [3105, 3106, ["classifier"], "Method"], [3142, 3145, ["machine", "reading", "comprehension"], "Task"], [3155, 3157, ["computation", "ability"], "Metric"], [3166, 3167, ["SQuAD"], "Material"], [3179, 3182, ["answer", "extraction", "test"], "Task"], [3186, 3188, ["SQuAD", "2.0"], "Material"], [3281, 3282, ["Spacy"], "Method"], [3342, 3349, ["100", "-", "dim", "Glove", "pretrained", "word", "embeddings"], "Method"], [3350, 3355, ["1024", "-", "dim", "Elmo", "embeddings"], "Method"], [3358, 3360, ["LSTM", "blocks"], "Method"], [3392, 3394, ["modeling", "layers"], "Method"], [3397, 3399, ["embedding", "layer"], "Method"], [3402, 3404, ["dropout", "rate"], "Metric"], [3409, 3411, ["Adam", "optimizer"], "Method"], [3413, 3417, ["learning", "rate", "of", "0.002"], "Metric"], [3470, 3472, ["F1", "score"], "Metric"], [3476, 3478, ["EM", "score"], "Metric"], [3487, 3489, ["F1", "score"], "Metric"], [3493, 3495, ["EM", "score"], "Metric"], [3539, 3541, ["end", "model"], "Method"], [3552, 3554, ["end", "models"], "Method"], [3559, 3561, ["F1", "scores"], "Metric"], [3569, 3572, ["U", "-", "Net"], "Method"], [3578, 3582, ["post", "-", "processing", "step"], "Method"], [3592, 3594, ["Ablation", "Study"], "Task"], [3598, 3600, ["ablation", "study"], "Task"], [3665, 3667, ["F1", "score"], "Metric"], [3675, 3677, ["universal", "node"], "Method"], [3699, 3701, ["passage", "representation"], "Method"], [3704, 3706, ["attention", "layer"], "Method"], [3780, 3781, ["BiLSTMs"], "Method"], [3798, 3799, ["BiLSTM"], "Method"], [3806, 3807, ["quality"], "Metric"], [3882, 3885, ["U", "-", "Net"], "Method"], [3902, 3906, ["self", "-", "attention", "layer"], "Method"], [3922, 3924, ["unified", "representation"], "Method"], [3950, 3954, ["bi", "-", "attention", "model"], "Method"], [3956, 3959, ["F1", "-", "score"], "Metric"], [3965, 3969, ["Multi", "-", "task", "Study"], "Task"], [3980, 3984, ["multi", "-", "task", "model"], "Method"], [4001, 4004, ["answer", "boundary", "detect"], "Task"], [4005, 4006, ["classification"], "Task"], [4022, 4023, ["EM"], "Metric"], [4024, 4026, ["F1", "score"], "Metric"], [4028, 4029, ["classification"], "Task"], [4045, 4047, ["SQuAD", "1.1"], "Material"], [4051, 4052, ["classifier"], "Method"], [4058, 4060, ["backward", "propagation"], "Method"], [4062, 4067, ["loss", "of", "answer", "boundary", "detection"], "Task"], [4071, 4073, ["classification", "task"], "Task"], [4096, 4100, ["multi", "-", "task", "model"], "Method"], [4102, 4106, ["answer", "boundary", "detection", "task"], "Task"], [4130, 4132, ["summarized", "representation"], "Method"], [4155, 4156, ["classification"], "Task"], [4159, 4163, ["answer", "boundary", "detection", "task"], "Task"], [4168, 4172, ["multi", "-", "task", "setup"], "Task"], [4176, 4178, ["classification", "layer"], "Method"], [4192, 4193, ["classifier"], "Method"], [4194, 4196, ["pointer", "layer"], "Method"], [4198, 4200, ["encoding", "process"], "Method"], [4277, 4279, ["classification", "task"], "Task"], [4320, 4323, ["SQuAD", "1.1", "test"], "Material"], [4336, 4339, ["SQuAD", "2.0", "test"], "Material"], [4357, 4362, ["answer", "detection", "and", "classification", "tasks"], "Task"], [4370, 4374, ["multi", "-", "task", "model"], "Method"], [4380, 4382, ["unanswerability", "classification"], "Task"], [4389, 4391, ["answer", "verifier"], "Task"], [4402, 4404, ["Unanswerability", "Classification"], "Task"], [4408, 4410, ["answer", "verifier"], "Method"], [4433, 4435, ["SQuAD", "2.0"], "Material"], [4489, 4491, ["SQuAD", "2.0"], "Material"], [4496, 4498, ["answerability", "prediction"], "Task"], [4515, 4517, ["F1", "score"], "Metric"], [4564, 4566, ["F1", "score"], "Metric"], [4597, 4599, ["answerable", "questions"], "Task"], [4630, 4632, ["SQuAD", "2.0"], "Material"], [4669, 4671, ["submission", "system"], "Method"], [4681, 4687, ["End", "-", "to", "-", "end", "Models"], "Method"], [4688, 4689, ["MRC"], "Task"], [4696, 4699, ["neural", "network", "models"], "Method"], [4704, 4707, ["machine", "reading", "comprehension"], "Task"], [4717, 4718, ["encoder"], "Method"], [4725, 4726, ["BiLSTM"], "Method"], [4732, 4734, ["embedded", "representation"], "Task"], [4739, 4742, ["bidirectional", "attention", "mechanism"], "Method"], [4757, 4759, ["pointer", "network"], "Method"], [4781, 4787, ["match", "-", "LSTM", "and", "pointer", "networks"], "Method"], [4796, 4799, ["bidirectional", "attention", "mechanism"], "Method"], [4820, 4826, ["end", "-", "to", "-", "end", "model"], "Method"], [4828, 4831, ["SQuAD", "2.0", "task"], "Material"], [4841, 4843, ["unified", "representation"], "Method"], [4876, 4880, ["unanswerability", "of", "a", "question"], "Task"], [4883, 4884, ["MRC"], "Method"], [4887, 4888, ["MRC"], "Task"], [4936, 4939, ["answer", "span", "predictions"], "Method"], [4973, 4974, ["hu2018read"], "Method"], [4981, 4984, ["answer", "verifier", "idea"], "Method"], [4987, 4989, ["classification", "layer"], "Method"], [4999, 5001, ["pipeline", "structure"], "Method"], [5006, 5008, ["answer", "verifier"], "Method"], [5019, 5023, ["Multi", "-", "task", "models"], "Method"], [5031, 5032, ["MRC"], "Task"], [5037, 5042, ["multi", "-", "task", "learning", "problem"], "Task"], [5051, 5053, ["answer", "prediction"], "Task"], [5054, 5056, ["answer", "verification"], "Task"], [5079, 5083, ["multi", "-", "task", "model"], "Method"], [5086, 5089, ["sub", "-", "tasks"], "Task"], [5099, 5101, ["answer", "verifier"], "Task"], [5115, 5116, ["MRC"], "Task"], [5120, 5125, ["multi", "-", "task", "learning", "problems"], "Task"], [5128, 5131, ["U", "-", "Net"], "Method"], [5134, 5140, ["end", "-", "to", "-", "end", "model"], "Method"], [5141, 5143, ["MRC", "challenges"], "Task"], [5144, 5147, ["U", "-", "Net"], "Method"], [5151, 5153, ["SQuAD", "2.0"], "Material"], [5163, 5165, ["fused", "representation"], "Method"], [5175, 5177, ["concatenated", "representation"], "Method"], [5180, 5182, ["encoding", "layers"], "Method"], [5199, 5202, ["U", "-", "Net"], "Method"], [5207, 5209, ["unified", "representation"], "Method"], [5212, 5213, ["question"], "Method"], [5214, 5216, ["passage", "representation"], "Method"], [5222, 5225, ["U", "-", "Net"], "Method"], [5227, 5231, ["multi", "-", "task", "framework"], "Method"], [5257, 5260, ["SQuAD", "2.0", "test"], "Material"], [5270, 5273, ["U", "-", "Net"], "Method"], [5277, 5282, ["multi", "-", "level", "attention", "block"], "Method"], [5285, 5289, ["self", "-", "attention", "mechanism"], "Method"], [5323, 5326, ["answer", "boundary", "detection"], "Task"], [5348, 5351, ["boundary", "detection", "task"], "Task"], [5359, 5362, ["SQuAD", "2.0", "test"], "Material"], [5425, 5428, ["SQuAD", "2.0", "submissions"], "Material"], [37, 38, ["question"], "Method"], [82, 83, ["question"], "Method"], [105, 106, ["question"], "Method"], [118, 119, ["question"], "Method"], [290, 291, ["SQuAD"], "Material"], [295, 296, ["question"], "Method"], [332, 333, ["question"], "Method"], [339, 340, ["SQuAD"], "Material"], [342, 343, ["squad2.0"], "Material"], [450, 451, ["question"], "Method"], [483, 484, ["question"], "Method"], [489, 490, ["SQuAD"], "Material"], [507, 508, ["question"], "Method"], [527, 528, ["question"], "Method"], [571, 572, ["question"], "Method"], [583, 584, ["question"], "Method"], [728, 729, ["question"], "Method"], [745, 746, ["question"], "Method"], [767, 768, ["question"], "Method"], [784, 785, ["question"], "Method"], [813, 814, ["question"], "Method"], [820, 821, ["question"], "Method"], [946, 947, ["question"], "Method"], [961, 962, ["question"], "Method"], [1006, 1007, ["question"], "Method"], [1094, 1095, ["question"], "Method"], [1188, 1189, ["question"], "Method"], [1314, 1315, ["question"], "Method"], [1326, 1327, ["question"], "Method"], [1365, 1366, ["question"], "Method"], [1456, 1457, ["question"], "Method"], [1484, 1485, ["question"], "Method"], [1556, 1557, ["question"], "Method"], [1589, 1590, ["question"], "Method"], [1599, 1600, ["question"], "Method"], [1608, 1609, ["question"], "Method"], [1651, 1652, ["question"], "Method"], [1685, 1686, ["question"], "Method"], [1701, 1702, ["question"], "Method"], [1737, 1738, ["question"], "Method"], [1752, 1753, ["question"], "Method"], [1766, 1767, ["question"], "Method"], [1791, 1792, ["question"], "Method"], [1911, 1912, ["question"], "Method"], [1938, 1939, ["question"], "Method"], [1941, 1942, ["question"], "Method"], [1991, 1992, ["question"], "Method"], [2105, 2106, ["question"], "Method"], [2161, 2162, ["BiLSTM"], "Method"], [2193, 2194, ["question"], "Method"], [2231, 2232, ["question"], "Method"], [2261, 2262, ["question"], "Method"], [2300, 2301, ["question"], "Method"], [2345, 2346, ["question"], "Method"], [2371, 2372, ["question"], "Method"], [2426, 2427, ["question"], "Method"], [2462, 2463, ["question"], "Method"], [2554, 2555, ["question"], "Method"], [2655, 2656, ["question"], "Method"], [2722, 2723, ["question"], "Method"], [2737, 2738, ["question"], "Method"], [2763, 2764, ["question"], "Method"], [2771, 2772, ["question"], "Method"], [2796, 2797, ["question"], "Method"], [2812, 2813, ["question"], "Method"], [2830, 2831, ["question"], "Method"], [2851, 2852, ["question"], "Method"], [2882, 2883, ["question"], "Method"], [2903, 2906, ["U", "-", "Net"], "Method"], [2910, 2911, ["question"], "Method"], [2952, 2953, ["question"], "Method"], [3002, 3003, ["question"], "Method"], [3090, 3091, ["question"], "Method"], [3099, 3100, ["question"], "Method"], [3108, 3109, ["question"], "Method"], [3146, 3147, ["question"], "Method"], [3194, 3195, ["question"], "Method"], [3200, 3201, ["SQuAD"], "Material"], [3285, 3286, ["question"], "Method"], [3336, 3337, ["question"], "Method"], [3429, 3430, ["question"], "Method"], [3448, 3449, ["question"], "Method"], [3602, 3603, ["SQuAD"], "Material"], [3683, 3684, ["question"], "Method"], [3730, 3731, ["question"], "Method"], [3759, 3760, ["question"], "Method"], [3765, 3766, ["BiLSTM"], "Method"], [3783, 3784, ["question"], "Method"], [3926, 3927, ["question"], "Method"], [4115, 4116, ["question"], "Method"], [4136, 4137, ["question"], "Method"], [4147, 4148, ["question"], "Method"], [4230, 4231, ["question"], "Method"], [4243, 4244, ["question"], "Method"], [4249, 4250, ["question"], "Method"], [4286, 4287, ["SQuAD"], "Material"], [4415, 4416, ["question"], "Method"], [4532, 4533, ["question"], "Method"], [4590, 4591, ["question"], "Method"], [4672, 4673, ["SQuAD"], "Material"], [4750, 4751, ["question"], "Method"], [4773, 4774, ["SQuAD"], "Material"], [4802, 4803, ["question"], "Method"], [4846, 4847, ["question"], "Method"], [4863, 4864, ["question"], "Method"], [4947, 4948, ["question"], "Method"], [4971, 4972, ["question"], "Method"], [5168, 5169, ["question"], "Method"], [5186, 5187, ["question"], "Method"], [5242, 5243, ["question"], "Method"], [5296, 5297, ["question"], "Method"], [5311, 5314, ["U", "-", "Net"], "Method"], [5368, 5371, ["U", "-", "Net"], "Method"]]} {"doc_id": "289e3e6b84982eb65aea8e3a64f2f6916c98e87e", "text": ["document", ":", "Grammar", "as", "a", "Foreign", "Language", "Syntactic", "constituency", "parsing", "is", "a", "fundamental", "problem", "in", "natural", "language", "processing", "and", "has", "been", "the", "subject", "of", "intensive", "research", "and", "engineering", "for", "decades", ".", "As", "a", "result", ",", "the", "most", "accurate", "parsers", "are", "domain", "specific", ",", "complex", ",", "and", "inefficient", ".", "In", "this", "paper", "we", "show", "that", "the", "domain", "agnostic", "attention", "-", "enhanced", "sequence", "-", "to", "-", "sequence", "model", "achieves", "state", "-", "of", "-", "the", "-", "art", "results", "on", "the", "most", "widely", "used", "syntactic", "constituency", "parsing", "dataset", ",", "when", "trained", "on", "a", "large", "synthetic", "corpus", "that", "was", "annotated", "using", "existing", "parsers", ".", "It", "also", "matches", "the", "performance", "of", "standard", "parsers", "when", "trained", "only", "on", "a", "small", "human", "-", "annotated", "dataset", ",", "which", "shows", "that", "this", "model", "is", "highly", "data", "-", "efficient", ",", "in", "contrast", "to", "sequence", "-", "to", "-", "sequence", "models", "without", "the", "attention", "mechanism", ".", "Our", "parser", "is", "also", "fast", ",", "processing", "over", "a", "hundred", "sentences", "per", "second", "with", "an", "unoptimized", "CPU", "implementation", ".", "section", ":", "Introduction", "Syntactic", "constituency", "parsing", "is", "a", "fundamental", "problem", "in", "linguistics", "and", "natural", "language", "processing", "that", "has", "a", "wide", "range", "of", "applications", ".", "This", "problem", "has", "been", "the", "subject", "of", "intense", "research", "for", "decades", ",", "and", "as", "a", "result", ",", "there", "exist", "highly", "accurate", "domain", "-", "specific", "parsers", ".", "The", "computational", "requirements", "of", "traditional", "parsers", "are", "cubic", "in", "sentence", "length", ",", "and", "while", "linear", "-", "time", "shift", "-", "reduce", "constituency", "parsers", "improved", "in", "accuracy", "in", "recent", "years", ",", "they", "never", "matched", "state", "-", "of", "-", "the", "-", "art", ".", "Furthermore", ",", "standard", "parsers", "have", "been", "designed", "with", "parsing", "in", "mind", ";", "the", "concept", "of", "a", "parse", "tree", "is", "deeply", "ingrained", "into", "these", "systems", ",", "which", "makes", "these", "methods", "inapplicable", "to", "other", "problems", ".", "Recently", ",", "Sutskever", "et", "al", ".", "introduced", "a", "neural", "network", "model", "for", "solving", "the", "general", "sequence", "-", "to", "-", "sequence", "problem", ",", "and", "Bahdanau", "et", "al", ".", "proposed", "a", "related", "model", "with", "an", "attention", "mechanism", "that", "makes", "it", "capable", "of", "handling", "long", "sequences", "well", ".", "Both", "models", "achieve", "state", "-", "of", "-", "the", "-", "art", "results", "on", "large", "scale", "machine", "translation", "tasks", "(", "e.g.", ",", ")", ".", "Syntactic", "constituency", "parsing", "can", "be", "formulated", "as", "a", "sequence", "-", "to", "-", "sequence", "problem", "if", "we", "linearize", "the", "parse", "tree", "(", "cf", ".", "Figure", "[", "reference", "]", ")", ",", "so", "we", "can", "apply", "these", "models", "to", "parsing", "as", "well", ".", "Our", "early", "experiments", "focused", "on", "the", "sequence", "-", "to", "-", "sequence", "model", "of", "Sutskever", "et", "al", ".", ".", "We", "found", "this", "model", "to", "work", "poorly", "when", "we", "trained", "it", "on", "standard", "human", "-", "annotated", "parsing", "datasets", "(", "1", "M", "tokens", ")", ",", "so", "we", "constructed", "an", "artificial", "dataset", "by", "labelling", "a", "large", "corpus", "with", "the", "BerkeleyParser", ".", "To", "our", "surprise", ",", "the", "sequence", "-", "to", "-", "sequence", "model", "matched", "the", "BerkeleyParser", "that", "produced", "the", "annotation", ",", "having", "achieved", "an", "F1", "score", "of", "90.5", "on", "the", "test", "set", "(", "section", "23", "of", "the", "WSJ", ")", ".", "We", "suspected", "that", "the", "attention", "model", "of", "Bahdanau", "et", "al", ".", "might", "be", "more", "data", "efficient", "and", "we", "found", "that", "it", "is", "indeed", "the", "case", ".", "We", "trained", "a", "sequence", "-", "to", "-", "sequence", "model", "with", "attention", "on", "the", "small", "human", "-", "annotated", "parsing", "dataset", "and", "were", "able", "to", "achieve", "an", "F1", "score", "of", "88.3", "on", "section", "23", "of", "the", "WSJ", "without", "the", "use", "of", "an", "ensemble", "and", "90.5", "with", "an", "ensemble", ",", "which", "matches", "the", "performance", "of", "the", "BerkeleyParser", "(", "90.4", ")", "when", "trained", "on", "the", "same", "data", ".", "Finally", ",", "we", "constructed", "a", "second", "artificial", "dataset", "consisting", "of", "only", "high", "-", "confidence", "parse", "trees", ",", "as", "measured", "by", "the", "agreement", "of", "two", "parsers", ".", "We", "trained", "a", "sequence", "-", "to", "-", "sequence", "model", "with", "attention", "on", "this", "data", "and", "achieved", "an", "F1", "score", "of", "92.5", "on", "section", "23", "of", "the", "WSJ", "–", "a", "new", "state", "-", "of", "-", "the", "-", "art", ".", "This", "result", "did", "not", "require", "an", "ensemble", ",", "and", "as", "a", "result", ",", "the", "parser", "is", "also", "very", "fast", ".", "An", "ensemble", "further", "improves", "the", "score", "to", "92.8", ".", "section", ":", "LSTM", "+", "A", "Parsing", "Model", "[", "xscale=0.64", ",", "yscale=0.75", "]", "(", "A", ")", "at", "(", "0", ",", "-", "1.7", ")", ".", ";", "(", "B", ")", "at", "(", "2", ",", "-", "1.7", ")", "Go", ";", "[", "thick", ",-", "¿", "]", "(", "A", ")", "–", "(", "0", ",", "-", "1.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "B", ")", "–", "(", "2", ",", "-", "1.05", ")", ";", "(", "-", "0.8", ",", "-", "1.0", ")", "rectangle", "(", "2.8", ",", "0.0", ")", ";", "(", "lstm", ")", "at", "(", "1", ",", "-", "0.5", ")", "LSTM", ";", "[", "thick", ",-", "¿", "]", "(", "0", ",", "0.05", ")", "–", "(", "0", ",", "0.45", ")", ";", "[", "thick", ",-", "¿", "]", "(", "2", ",", "0.05", ")", "–", "(", "2", ",", "0.45", ")", ";", "(", "-", "0.8", ",", "0.5", ")", "rectangle", "(", "2.8", ",", "1.5", ")", ";", "(", "lstm", ")", "at", "(", "1", ",", "1", ")", "LSTM", ";", "[", "thick", ",-", "¿", "]", "(", "0", ",", "1.55", ")", "–", "(", "0", ",", "1.95", ")", ";", "[", "thick", ",-", "¿", "]", "(", "2", ",", "1.55", ")", "–", "(", "2", ",", "1.95", ")", ";", "(", "-", "0.8", ",", "2", ")", "rectangle", "(", "2.8", ",", "3", ")", ";", "(", "lstm", ")", "at", "(", "1", ",", "2.5", ")", "LSTM", ";", "[", "thick", ",-", "¿", "]", "(", "2.85", ",", "-", "0.5", ")", "–", "(", "4.15", ",", "-", "0.5", ")", ";", "[", "thick", ",-", "¿", "]", "(", "2.85", ",", "1", ")", "–", "(", "4.15", ",", "1", ")", ";", "[", "thick", ",-", "¿", "]", "(", "2.85", ",", "2.5", ")", "–", "(", "4.15", ",", "2.5", ")", ";", "[", "xshift=1", "cm", "]", "(", "end", ")", "at", "(", "4", ",", "-", "1.7", ")", "end", ";", "(", "S", ")", "at", "(", "6", ",", "-", "1.7", ")", "(", "S", ";", "(", "VP", ")", "at", "(", "8", ",", "-", "1.7", ")", "(", "VP", ";", "(", "VB", ")", "at", "(", "10", ",", "-", "1.7", ")", "XX", ";", "(", "cVP", ")", "at", "(", "12", ",", "-", "1.7", ")", ")", ";", "(", "dot", ")", "at", "(", "14", ",", "-", "1.7", ")", ".", ";", "(", "cS", ")", "at", "(", "16", ",", "-", "1.7", ")", ")", ";", "[", "thick", ",-", "¿", "]", "(", "end", ")", "–", "(", "4", ",", "-", "1.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "S", ")", "–", "(", "6", ",", "-", "1.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "VP", ")", "–", "(", "8", ",", "-", "1.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "VB", ")", "–", "(", "10", ",", "-", "1.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "cVP", ")", "–", "(", "12", ",", "-", "1.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "dot", ")", "–", "(", "14", ",", "-", "1.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "cS", ")", "–", "(", "16", ",", "-", "1.05", ")", ";", "(", "3.2", ",", "-", "1.0", ")", "rectangle", "(", "16.8", ",", "0.0", ")", ";", "(", "lstm", ")", "at", "(", "10", ",", "-", "0.5", ")", "LSTM", ";", "[", "thick", ",-", "¿", "]", "(", "4", ",", "0.05", ")", "–", "(", "4", ",", "0.45", ")", ";", "[", "thick", ",-", "¿", "]", "(", "6", ",", "0.05", ")", "–", "(", "6", ",", "0.45", ")", ";", "[", "thick", ",-", "¿", "]", "(", "8", ",", "0.05", ")", "–", "(", "8", ",", "0.45", ")", ";", "[", "thick", ",-", "¿", "]", "(", "10", ",", "0.05", ")", "–", "(", "10", ",", "0.45", ")", ";", "[", "thick", ",-", "¿", "]", "(", "12", ",", "0.05", ")", "–", "(", "12", ",", "0.45", ")", ";", "[", "thick", ",-", "¿", "]", "(", "14", ",", "0.05", ")", "–", "(", "14", ",", "0.45", ")", ";", "[", "thick", ",-", "¿", "]", "(", "16", ",", "0.05", ")", "–", "(", "16", ",", "0.45", ")", ";", "(", "3.2", ",", "0.5", ")", "rectangle", "(", "16.8", ",", "1.5", ")", ";", "(", "lstm", ")", "at", "(", "10", ",", "1", ")", "LSTM", ";", "[", "thick", ",-", "¿", "]", "(", "4", ",", "1.55", ")", "–", "(", "4", ",", "1.95", ")", ";", "[", "thick", ",-", "¿", "]", "(", "6", ",", "1.55", ")", "–", "(", "6", ",", "1.95", ")", ";", "[", "thick", ",-", "¿", "]", "(", "8", ",", "1.55", ")", "–", "(", "8", ",", "1.95", ")", ";", "[", "thick", ",-", "¿", "]", "(", "10", ",", "1.55", ")", "–", "(", "10", ",", "1.95", ")", ";", "[", "thick", ",-", "¿", "]", "(", "12", ",", "1.55", ")", "–", "(", "12", ",", "1.95", ")", ";", "[", "thick", ",-", "¿", "]", "(", "14", ",", "1.55", ")", "–", "(", "14", ",", "1.95", ")", ";", "[", "thick", ",-", "¿", "]", "(", "16", ",", "1.55", ")", "–", "(", "16", ",", "1.95", ")", ";", "(", "3.2", ",", "2", ")", "rectangle", "(", "16.8", ",", "3", ")", ";", "(", "lstm", ")", "at", "(", "10", ",", "2.5", ")", "LSTM", ";", "(", "S", ")", "at", "(", "4", ",", "3.7", ")", "(", "S", ";", "(", "VP", ")", "at", "(", "6", ",", "3.7", ")", "(", "VP", ";", "(", "VB", ")", "at", "(", "8", ",", "3.7", ")", "XX", ";", "(", "cVP", ")", "at", "(", "10", ",", "3.7", ")", ")", ";", "(", "dot", ")", "at", "(", "12", ",", "3.7", ")", ".", ";", "(", "cS", ")", "at", "(", "14", ",", "3.7", ")", ")", ";", "(", "end", ")", "at", "(", "16", ",", "3.7", ")", "end", ";", "[", "thick", ",-", "¿", "]", "(", "4", ",", "3.05", ")", "–", "(", "S", ")", ";", "0", ",", "3.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "6", ",", "3.05", ")", "–", "(", "VP", ")", ";", "0", ",", "3.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "8", ",", "3.05", ")", "–", "(", "VB", ")", ";", "0", ",", "3.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "10", ",", "3.05", ")", "–", "(", "cVP", ")", ";", "0", ",", "3.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "12", ",", "3.05", ")", "–", "(", "dot", ")", ";", "0", ",", "3.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "14", ",", "3.05", ")", "–", "(", "cS", ")", ";", "0", ",", "3.05", ")", ";", "[", "thick", ",-", "¿", "]", "(", "16", ",", "3.05", ")", "–", "(", "end", ")", ";", "0", ",", "3.05", ")", ";", "Let", "us", "first", "recall", "the", "sequence", "-", "to", "-", "sequence", "LSTM", "model", ".", "The", "Long", "Short", "-", "Term", "Memory", "model", "of", "is", "defined", "as", "follows", ".", "Let", ",", ",", "and", "be", "the", "input", ",", "control", "state", ",", "and", "memory", "state", "at", "timestep", ".", "Given", "a", "sequence", "of", "inputs", ",", "the", "LSTM", "computes", "the", "-", "sequence", "and", "the", "-", "sequence", "as", "follows", ".", "The", "operator", "denotes", "element", "-", "wise", "multiplication", ",", "the", "matrices", "and", "the", "vector", "are", "the", "parameters", "of", "the", "model", ",", "and", "all", "the", "nonlinearities", "are", "computed", "element", "-", "wise", ".", "In", "a", "deep", "LSTM", ",", "each", "subsequent", "layer", "uses", "the", "-", "sequence", "of", "the", "previous", "layer", "for", "its", "input", "sequence", ".", "The", "deep", "LSTM", "defines", "a", "distribution", "over", "output", "sequences", "given", "an", "input", "sequence", ":", "The", "above", "equation", "assumes", "a", "deep", "LSTM", "whose", "input", "sequence", "is", ",", "so", "denotes", "-", "th", "element", "of", "the", "-", "sequence", "of", "topmost", "LSTM", ".", "The", "matrix", "consists", "of", "the", "vector", "representations", "of", "each", "output", "symbol", "and", "the", "symbol", "is", "a", "Kronecker", "delta", "with", "a", "dimension", "for", "each", "output", "symbol", ",", "so", "is", "precisely", "the", "’", "th", "element", "of", "the", "distribution", "defined", "by", "the", "softmax", ".", "Every", "output", "sequence", "terminates", "with", "a", "special", "end", "-", "of", "-", "sequence", "token", "which", "is", "necessary", "in", "order", "to", "define", "a", "distribution", "over", "sequences", "of", "variable", "lengths", ".", "We", "use", "two", "different", "sets", "of", "LSTM", "parameters", ",", "one", "for", "the", "input", "sequence", "and", "one", "for", "the", "output", "sequence", ",", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "Stochastic", "gradient", "descent", "is", "used", "to", "maximize", "the", "training", "objective", "which", "is", "the", "average", "over", "the", "training", "set", "of", "the", "log", "probability", "of", "the", "correct", "output", "sequence", "given", "the", "input", "sequence", ".", "subsection", ":", "Attention", "Mechanism", "An", "important", "extension", "of", "the", "sequence", "-", "to", "-", "sequence", "model", "is", "by", "adding", "an", "attention", "mechanism", ".", "We", "adapted", "the", "attention", "model", "from", "which", ",", "to", "produce", "each", "output", "symbol", ",", "uses", "an", "attention", "mechanism", "over", "the", "encoder", "LSTM", "states", ".", "Similar", "to", "our", "sequence", "-", "to", "-", "sequence", "model", "described", "in", "the", "previous", "section", ",", "we", "use", "two", "separate", "LSTMs", "(", "one", "to", "encode", "the", "sequence", "of", "input", "words", ",", "and", "another", "one", "to", "produce", "or", "decode", "the", "output", "symbols", ")", ".", "Recall", "that", "the", "encoder", "hidden", "states", "are", "denoted", "and", "we", "denote", "the", "hidden", "states", "of", "the", "decoder", "by", ".", "To", "compute", "the", "attention", "vector", "at", "each", "output", "time", "over", "the", "input", "words", "we", "define", ":", "The", "vector", "and", "matrices", "are", "learnable", "parameters", "of", "the", "model", ".", "The", "vector", "has", "length", "and", "its", "-", "th", "item", "contains", "a", "score", "of", "how", "much", "attention", "should", "be", "put", "on", "the", "-", "th", "hidden", "encoder", "state", ".", "These", "scores", "are", "normalized", "by", "softmax", "to", "create", "the", "attention", "mask", "over", "encoder", "hidden", "states", ".", "In", "all", "our", "experiments", ",", "we", "use", "the", "same", "hidden", "dimensionality", "(", "256", ")", "at", "the", "encoder", "and", "the", "decoder", ",", "so", "is", "a", "vector", "and", "and", "are", "square", "matrices", ".", "Lastly", ",", "we", "concatenate", "with", ",", "which", "becomes", "the", "new", "hidden", "state", "from", "which", "we", "make", "predictions", ",", "and", "which", "is", "fed", "to", "the", "next", "time", "step", "in", "our", "recurrent", "model", ".", "In", "Section", "[", "reference", "]", "we", "provide", "an", "analysis", "of", "what", "the", "attention", "mechanism", "learned", ",", "and", "we", "visualize", "the", "normalized", "attention", "vector", "for", "all", "in", "Figure", "[", "reference", "]", ".", "subsection", ":", "Linearizing", "Parsing", "Trees", "To", "apply", "the", "model", "described", "above", "to", "parsing", ",", "we", "need", "to", "design", "an", "invertible", "way", "of", "converting", "the", "parse", "tree", "into", "a", "sequence", "(", "linearization", ")", ".", "We", "do", "this", "in", "a", "very", "simple", "way", "following", "a", "depth", "-", "first", "traversal", "order", ",", "as", "depicted", "in", "Figure", "[", "reference", "]", ".", "[", "growth", "parent", "anchor", "=", "north", ",", "level", "distance=2em", ",", "xscale=1.5", ",", "yscale=1.3", "]", "(", "tokens", ")", "at", "(", "-", "2", ",", "0", ")", "John", "has", "a", "dog", ".", ";", "(", "arrow", ")", "at", "(", "-", "0.7", ",", "0", ")", ";", "(", "S", ")", "at", "(", "2", ",", "0.5", ")", "S", "child", "node", "NP", "child", "node", "NNP", "child", "node", "VP", "child", "node", "VBZ", "child", "node", "NP", "child", "node", "DT", "child", "node", "NN", "child", "node", ".", ";", "(", "tokens", ")", "at", "(", "-", "2", ",", "-", "1.7", ")", "John", "has", "a", "dog", ".", ";", "(", "arrow", ")", "at", "(", "-", "0.7", ",", "-", "1.7", ")", ";", "[", "anchor", "=", "west", "]", "(", "result", ")", "at", "(", "0", ",", "-", "1.7", ")", "(", "S", "(", "NP", "NNP", ")", "(", "VP", "VBZ", "(", "NP", "DT", "NN", ")", ")", ".", ")", ";", "We", "use", "the", "above", "model", "for", "parsing", "in", "the", "following", "way", ".", "First", ",", "the", "network", "consumes", "the", "sentence", "in", "a", "left", "-", "to", "-", "right", "sweep", ",", "creating", "vectors", "in", "memory", ".", "Then", ",", "it", "outputs", "the", "linearized", "parse", "tree", "using", "information", "in", "these", "vectors", ".", "As", "described", "below", ",", "we", "use", "3", "LSTM", "layers", ",", "reverse", "the", "input", "sentence", "and", "normalize", "part", "-", "of", "-", "speech", "tags", ".", "An", "example", "run", "of", "our", "LSTM", "+", "A", "model", "on", "the", "sentence", "“", "Go", ".", "”", "is", "depicted", "in", "Figure", "[", "reference", "]", "(", "top", "gray", "edges", "illustrate", "attention", ")", ".", "subsection", ":", "Parameters", "and", "Initialization", "paragraph", ":", "Sizes", ".", "In", "our", "experiments", "we", "used", "a", "model", "with", "3", "LSTM", "layers", "and", "256", "units", "in", "each", "layer", ",", "which", "we", "call", "LSTM", "+", "A.", "Our", "input", "vocabulary", "size", "was", "90", "K", "and", "we", "output", "128", "symbols", ".", "paragraph", ":", "Dropout", ".", "Training", "on", "a", "small", "dataset", "we", "additionally", "used", "2", "dropout", "layers", ",", "one", "between", "LSTM", "and", "LSTM", ",", "and", "one", "between", "LSTM", "and", "LSTM", ".", "We", "call", "this", "model", "LSTM", "+", "A", "+", "D.", "paragraph", ":", "POS", "-", "tag", "normalization", ".", "Since", "part", "-", "of", "-", "speech", "(", "POS", ")", "tags", "are", "not", "evaluated", "in", "the", "syntactic", "parsing", "F1", "score", ",", "we", "replaced", "all", "of", "them", "by", "“", "XX", "”", "in", "the", "training", "data", ".", "This", "improved", "our", "F1", "score", "by", "about", "1", "point", ",", "which", "is", "surprising", ":", "For", "standard", "parsers", ",", "including", "POS", "tags", "in", "training", "data", "helps", "significantly", ".", "All", "experiments", "reported", "below", "are", "performed", "with", "normalized", "POS", "tags", ".", "paragraph", ":", "Input", "reversing", ".", "We", "also", "found", "it", "useful", "to", "reverse", "the", "input", "sentences", "but", "not", "their", "parse", "trees", ",", "similarly", "to", ".", "Not", "reversing", "the", "input", "had", "a", "small", "negative", "impact", "on", "the", "F1", "score", "on", "our", "development", "set", "(", "about", "absolute", ")", ".", "All", "experiments", "reported", "below", "are", "performed", "with", "input", "reversing", ".", "paragraph", ":", "Pre", "-", "training", "word", "vectors", ".", "The", "embedding", "layer", "for", "our", "90", "K", "vocabulary", "can", "be", "initialized", "randomly", "or", "using", "pre", "-", "trained", "word", "-", "vector", "embeddings", ".", "We", "pre", "-", "trained", "skip", "-", "gram", "embeddings", "of", "size", "512", "using", "word2vec", "on", "a", "10B", "-", "word", "corpus", ".", "These", "embeddings", "were", "used", "to", "initialize", "our", "network", "but", "not", "fixed", ",", "they", "were", "later", "modified", "during", "training", ".", "We", "discuss", "the", "impact", "of", "pre", "-", "training", "in", "the", "experimental", "section", ".", "We", "do", "not", "apply", "any", "other", "special", "preprocessing", "to", "the", "data", ".", "In", "particular", ",", "we", "do", "not", "binarize", "the", "parse", "trees", "or", "handle", "unaries", "in", "any", "specific", "way", ".", "We", "also", "treat", "unknown", "words", "in", "a", "naive", "way", ":", "we", "map", "all", "words", "beyond", "our", "90", "K", "vocabulary", "to", "a", "single", "UNK", "token", ".", "This", "potentially", "underestimates", "our", "final", "results", ",", "but", "keeps", "our", "framework", "task", "-", "independent", ".", "section", ":", "Experiments", "subsection", ":", "Training", "Data", "We", "trained", "the", "model", "described", "above", "on", "2", "different", "datasets", ".", "For", "one", ",", "we", "trained", "on", "the", "standard", "WSJ", "training", "dataset", ".", "This", "is", "a", "very", "small", "training", "set", "by", "neural", "network", "standards", ",", "as", "it", "contains", "only", "40", "K", "sentences", "(", "compared", "to", "60", "K", "examples", "even", "in", "MNIST", ")", ".", "Still", ",", "even", "training", "on", "this", "set", ",", "we", "managed", "to", "get", "results", "that", "match", "those", "obtained", "by", "domain", "-", "specific", "parsers", ".", "To", "exceed", "the", "previous", "state", "-", "of", "-", "the", "-", "art", ",", "we", "created", "another", ",", "larger", "training", "set", "of", "11", "M", "parsed", "sentences", "(", "250", "M", "tokens", ")", ".", "First", ",", "we", "collected", "all", "publicly", "available", "treebanks", ".", "We", "used", "the", "OntoNotes", "corpus", "version", "5", ",", "the", "English", "Web", "Treebank", "and", "the", "updated", "and", "corrected", "Question", "Treebank", ".", "Note", "that", "the", "popular", "Wall", "Street", "Journal", "section", "of", "the", "Penn", "Treebank", "is", "part", "of", "the", "OntoNotes", "corpus", ".", "In", "total", ",", "these", "corpora", "give", "us", "90", "K", "training", "sentences", "(", "we", "held", "out", "certain", "sections", "for", "evaluation", ",", "as", "described", "below", ")", ".", "In", "addition", "to", "this", "gold", "standard", "data", ",", "we", "use", "a", "corpus", "parsed", "with", "existing", "parsers", "using", "the", "“", "tri", "-", "training", "”", "approach", "of", ".", "In", "this", "approach", ",", "two", "parsers", ",", "our", "reimplementation", "of", "BerkeleyParser", "and", "a", "reimplementation", "of", "ZPar", ",", "are", "used", "to", "process", "unlabeled", "sentences", "sampled", "from", "news", "appearing", "on", "the", "web", ".", "We", "select", "only", "sentences", "for", "which", "both", "parsers", "produced", "the", "same", "parse", "tree", "and", "re", "-", "sample", "to", "match", "the", "distribution", "of", "sentence", "lengths", "of", "the", "WSJ", "training", "corpus", ".", "Re", "-", "sampling", "is", "useful", "because", "parsers", "agree", "much", "more", "often", "on", "short", "sentences", ".", "We", "call", "the", "set", "of", "11", "million", "sentences", "selected", "in", "this", "way", ",", "together", "with", "the", "90", "K", "golden", "sentences", "described", "above", ",", "the", "high", "-", "confidence", "corpus", ".", "In", "earlier", "experiments", ",", "we", "only", "used", "one", "parser", ",", "our", "reimplementation", "of", "BerkeleyParser", ",", "to", "create", "a", "corpus", "of", "parsed", "sentences", ".", "In", "that", "case", "we", "just", "parsed", "7", "million", "senteces", "from", "news", "appearing", "on", "the", "web", "and", "combined", "these", "parsed", "sentences", "with", "the", "90", "K", "golden", "corpus", "described", "above", ".", "We", "call", "this", "the", "BerkeleyParser", "corpus", ".", "subsection", ":", "Evaluation", "We", "use", "the", "standard", "EVALB", "tool", "for", "evaluation", "and", "report", "F1", "scores", "on", "our", "developments", "set", "(", "section", "22", "of", "the", "Penn", "Treebank", ")", "and", "the", "final", "test", "set", "(", "section", "23", ")", "in", "Table", "[", "reference", "]", ".", "First", ",", "let", "us", "remark", "that", "our", "training", "setup", "differs", "from", "those", "reported", "in", "previous", "works", ".", "To", "the", "best", "of", "our", "knowledge", ",", "no", "standard", "parsers", "have", "ever", "been", "trained", "on", "datasets", "numbering", "in", "the", "hundreds", "of", "millions", "of", "tokens", ",", "and", "it", "would", "be", "hard", "to", "do", "due", "to", "efficiency", "problems", ".", "We", "therefore", "cite", "the", "semi", "-", "supervised", "results", ",", "which", "are", "analogous", "in", "spirit", "but", "use", "less", "data", ".", "Table", "[", "reference", "]", "shows", "performance", "of", "our", "models", "on", "the", "top", "and", "results", "from", "other", "papers", "at", "the", "bottom", ".", "We", "compare", "to", "variants", "of", "the", "BerkeleyParser", "that", "use", "self", "-", "training", "on", "unlabeled", "data", ",", "or", "built", "an", "ensemble", "of", "multiple", "parsers", ",", "or", "combine", "both", "techniques", ".", "We", "also", "include", "the", "best", "linear", "-", "time", "parser", "in", "the", "literature", ",", "the", "transition", "-", "based", "parser", "of", ".", "It", "can", "be", "seen", "that", ",", "when", "training", "on", "WSJ", "only", ",", "a", "baseline", "LSTM", "does", "not", "achieve", "any", "reasonable", "score", ",", "even", "with", "dropout", "and", "early", "stopping", ".", "But", "a", "single", "attention", "model", "gets", "to", "and", "an", "ensemble", "of", "5", "LSTM", "+", "A", "+", "D", "models", "achieves", "matching", "a", "single", "-", "model", "BerkeleyParser", "on", "WSJ", "23", ".", "When", "trained", "on", "the", "large", "high", "-", "confidence", "corpus", ",", "a", "single", "LSTM", "+", "A", "model", "achieves", "and", "so", "outperforms", "not", "only", "the", "best", "single", "model", ",", "but", "also", "the", "best", "ensemble", "result", "reported", "previously", ".", "An", "ensemble", "of", "5", "LSTM", "+", "A", "models", "further", "improves", "this", "score", "to", ".", "paragraph", ":", "Generating", "well", "-", "formed", "trees", ".", "The", "LSTM", "+", "A", "model", "trained", "on", "WSJ", "dataset", "only", "produced", "malformed", "trees", "for", "25", "of", "the", "1700", "sentences", "in", "our", "development", "set", "(", "%", "of", "all", "cases", ")", ",", "and", "the", "model", "trained", "on", "full", "high", "-", "confidence", "dataset", "did", "this", "for", "14", "sentences", "(", "%", ")", ".", "In", "these", "few", "cases", "where", "LSTM", "+", "A", "outputs", "a", "malformed", "tree", ",", "we", "simply", "add", "brackets", "to", "either", "the", "beginning", "or", "the", "end", "of", "the", "tree", "in", "order", "to", "make", "it", "balanced", ".", "It", "is", "worth", "noting", "that", "all", "14", "cases", "where", "LSTM", "+", "A", "produced", "unbalanced", "trees", "were", "sentences", "or", "sentence", "fragments", "that", "did", "not", "end", "with", "proper", "punctuation", ".", "There", "were", "very", "few", "such", "sentences", "in", "the", "training", "data", ",", "so", "it", "is", "not", "a", "surprise", "that", "our", "model", "can", "not", "deal", "with", "them", "very", "well", ".", "paragraph", ":", "Score", "by", "sentence", "length", ".", "An", "important", "concern", "with", "the", "sequence", "-", "to", "-", "sequence", "LSTM", "was", "that", "it", "may", "not", "be", "able", "to", "handle", "long", "sentences", "well", ".", "We", "determine", "the", "extent", "of", "this", "problem", "by", "partitioning", "the", "development", "set", "by", "length", ",", "and", "evaluating", "BerkeleyParser", ",", "a", "baseline", "LSTM", "model", "without", "attention", ",", "and", "LSTM", "+", "A", "on", "sentences", "of", "each", "length", ".", "The", "results", ",", "presented", "in", "Figure", "[", "reference", "]", ",", "are", "surprising", ".", "The", "difference", "between", "the", "F1", "score", "on", "sentences", "of", "length", "upto", "30", "and", "that", "upto", "70", "is", "for", "the", "BerkeleyParser", ",", "for", "the", "baseline", "LSTM", ",", "and", "for", "LSTM", "+", "A.", "So", "already", "the", "baseline", "LSTM", "has", "similar", "performance", "to", "the", "BerkeleyParser", ",", "it", "degrades", "with", "length", "only", "slightly", ".", "Surprisingly", ",", "LSTM", "+", "A", "shows", "less", "degradation", "with", "length", "than", "BerkeleyParser", "–", "a", "full", "chart", "parser", "that", "uses", "a", "lot", "more", "memory", ".", "[", "xscale=0.18", ",", "yscale=0.5", "]", "[", "xstep=10", "cm", ",", "ystep=1", "cm", ",", "color", "=", "lightgray", ",", "thin", "]", "(", "10", ",", "90", ")", "grid", "(", "71", ",", "96", ")", ";", "in", "10", ",", "20", ",", "…", ",", "70", "(", "x", ")", "at", "(", ",", "89.8", ")", ";", "in", "90", ",", "…", ",", "96", "(", "x", ")", "at", "(", "8", ",", ")", ";", "(", "bottom", ")", "at", "(", "40", ",", "89.3", ")", "Sentence", "length", ";", "[", "rotate=90", "]", "(", "left", ")", "at", "(", "5", ",", "92.5", ")", "F1", "score", ";", "[", "color", "=", "blue", ",", "fill", "=", "blue", "]", "(", "72", ",", "93.85", ")", "rectangle", "(", "73", ",", "94.15", ")", ";", "[", "anchor", "=", "west", "]", "(", "l", ")", "at", "(", "73", ",", "94", ")", "BerkeleyParser", ";", "[", "color", "=", "red", ",", "fill", "=", "red", "]", "(", "72", ",", "93.35", ")", "rectangle", "(", "73", ",", "93.65", ")", ";", "[", "anchor", "=", "west", "]", "(", "l", ")", "at", "(", "73", ",", "93.5", ")", "baseline", "LSTM", ";", "[", "color", "=", "orange", ",", "fill", "=", "orange", "]", "(", "72", ",", "92.85", ")", "rectangle", "(", "73", ",", "93.15", ")", ";", "[", "anchor", "=", "west", "]", "(", "l", ")", "at", "(", "73", ",", "93", ")", "LSTM", "+", "A", ";", "[", "thick", ",", "color", "=", "blue", "]", "plot", "coordinates", "(", "10", ",", "92.7", ")", "(", "11", ",", "92.94", ")", "(", "12", ",", "92.95", ")", "(", "13", ",", "93.49", ")", "(", "14", ",", "93.69", ")", "(", "15", ",", "93.73", ")", "(", "16", ",", "93.75", ")", "(", "17", ",", "93.83", ")", "(", "18", ",", "93.69", ")", "(", "19", ",", "93.57", ")", "(", "20", ",", "93.33", ")", "(", "21", ",", "93.17", ")", "(", "22", ",", "93.12", ")", "(", "23", ",", "93.34", ")", "(", "24", ",", "93.2", ")", "(", "25", ",", "93.05", ")", "(", "26", ",", "92.78", ")", "(", "27", ",", "92.74", ")", "(", "28", ",", "92.64", ")", "(", "29", ",", "92.54", ")", "(", "30", ",", "92.37", ")", "(", "31", ",", "92.34", ")", "(", "32", ",", "92.3", ")", "(", "33", ",", "92.29", ")", "(", "34", ",", "92.15", ")", "(", "35", ",", "92.04", ")", "(", "36", ",", "91.99", ")", "(", "37", ",", "91.97", ")", "(", "38", ",", "91.91", ")", "(", "39", ",", "91.89", ")", "(", "40", ",", "91.8", ")", "(", "41", ",", "91.72", ")", "(", "42", ",", "91.69", ")", "(", "43", ",", "91.67", ")", "(", "44", ",", "91.55", ")", "(", "45", ",", "91.55", ")", "(", "46", ",", "91.47", ")", "(", "47", ",", "91.48", ")", "(", "48", ",", "91.45", ")", "(", "49", ",", "91.47", ")", "(", "50", ",", "91.43", ")", "(", "51", ",", "91.37", ")", "(", "52", ",", "91.34", ")", "(", "53", ",", "91.3", ")", "(", "54", ",", "91.3", ")", "(", "55", ",", "91.3", ")", "(", "56", ",", "91.26", ")", "(", "57", ",", "91.14", ")", "(", "58", ",", "91.15", ")", "(", "59", ",", "91.15", ")", "(", "60", ",", "91.15", ")", "(", "61", ",", "91.14", ")", "(", "62", ",", "91.14", ")", "(", "63", ",", "91.14", ")", "(", "64", ",", "91.09", ")", "(", "65", ",", "91.1", ")", "(", "66", ",", "91.1", ")", "(", "67", ",", "91.1", ")", "(", "68", ",", "91.1", ")", "(", "69", ",", "91.1", ")", "(", "70", ",", "91.1", ")", "(", "71", ",", "91.1", ")", ";", "[", "thick", ",", "color", "=", "red", "]", "plot", "coordinates", "(", "10", ",", "92.36", ")", "(", "11", ",", "93.05", ")", "(", "12", ",", "93.33", ")", "(", "13", ",", "93.47", ")", "(", "14", ",", "93.88", ")", "(", "15", ",", "94.12", ")", "(", "16", ",", "93.77", ")", "(", "17", ",", "93.83", ")", "(", "18", ",", "93.76", ")", "(", "19", ",", "93.84", ")", "(", "20", ",", "93.67", ")", "(", "21", ",", "93.53", ")", "(", "22", ",", "93.63", ")", "(", "23", ",", "93.76", ")", "(", "24", ",", "93.38", ")", "(", "25", ",", "93.3", ")", "(", "26", ",", "93.21", ")", "(", "27", ",", "92.98", ")", "(", "28", ",", "92.81", ")", "(", "29", ",", "92.72", ")", "(", "30", ",", "92.69", ")", "(", "31", ",", "92.54", ")", "(", "32", ",", "92.57", ")", "(", "33", ",", "92.61", ")", "(", "34", ",", "92.57", ")", "(", "35", ",", "92.44", ")", "(", "36", ",", "92.31", ")", "(", "37", ",", "92.33", ")", "(", "38", ",", "92.3", ")", "(", "39", ",", "92.18", ")", "(", "40", ",", "92.09", ")", "(", "41", ",", "92.04", ")", "(", "42", ",", "92.01", ")", "(", "43", ",", "91.99", ")", "(", "44", ",", "91.92", ")", "(", "45", ",", "91.9", ")", "(", "46", ",", "91.75", ")", "(", "47", ",", "91.74", ")", "(", "48", ",", "91.71", ")", "(", "49", ",", "91.7", ")", "(", "50", ",", "91.67", ")", "(", "51", ",", "91.57", ")", "(", "52", ",", "91.53", ")", "(", "53", ",", "91.52", ")", "(", "54", ",", "91.52", ")", "(", "55", ",", "91.52", ")", "(", "56", ",", "91.46", ")", "(", "57", ",", "91.23", ")", "(", "58", ",", "91.14", ")", "(", "59", ",", "91.11", ")", "(", "60", ",", "91.11", ")", "(", "61", ",", "91.06", ")", "(", "62", ",", "91.06", ")", "(", "63", ",", "90.99", ")", "(", "64", ",", "90.99", ")", "(", "65", ",", "90.99", ")", "(", "66", ",", "90.99", ")", "(", "67", ",", "90.99", ")", "(", "68", ",", "90.99", ")", "(", "69", ",", "90.99", ")", "(", "70", ",", "90.99", ")", "(", "71", ",", "90.99", ")", ";", "[", "thick", ",", "color", "=", "orange", "]", "plot", "coordinates", "(", "10", ",", "95.41", ")", "(", "11", ",", "95.07", ")", "(", "12", ",", "95.57", ")", "(", "13", ",", "95.16", ")", "(", "14", ",", "95.39", ")", "(", "15", ",", "95.24", ")", "(", "16", ",", "95.44", ")", "(", "17", ",", "95.5", ")", "(", "18", ",", "95.5", ")", "(", "19", ",", "95.39", ")", "(", "20", ",", "95.42", ")", "(", "21", ",", "95.1", ")", "(", "22", ",", "94.85", ")", "(", "23", ",", "94.76", ")", "(", "24", ",", "94.83", ")", "(", "25", ",", "94.58", ")", "(", "26", ",", "94.54", ")", "(", "27", ",", "94.37", ")", "(", "28", ",", "94.18", ")", "(", "29", ",", "94.06", ")", "(", "30", ",", "94.03", ")", "(", "31", ",", "94.03", ")", "(", "32", ",", "93.91", ")", "(", "33", ",", "93.9", ")", "(", "34", ",", "93.99", ")", "(", "35", ",", "93.94", ")", "(", "36", ",", "93.85", ")", "(", "37", ",", "93.86", ")", "(", "38", ",", "93.9", ")", "(", "39", ",", "93.87", ")", "(", "40", ",", "93.87", ")", "(", "41", ",", "93.84", ")", "(", "42", ",", "93.77", ")", "(", "43", ",", "93.77", ")", "(", "44", ",", "93.77", ")", "(", "45", ",", "93.75", ")", "(", "46", ",", "93.72", ")", "(", "47", ",", "93.62", ")", "(", "48", ",", "93.61", ")", "(", "49", ",", "93.59", ")", "(", "50", ",", "93.6", ")", "(", "51", ",", "93.6", ")", "(", "52", ",", "93.54", ")", "(", "53", ",", "93.49", ")", "(", "54", ",", "93.49", ")", "(", "55", ",", "93.49", ")", "(", "56", ",", "93.5", ")", "(", "57", ",", "93.43", ")", "(", "58", ",", "93.42", ")", "(", "59", ",", "93.42", ")", "(", "60", ",", "93.43", ")", "(", "61", ",", "93.43", ")", "(", "62", ",", "93.42", ")", "(", "63", ",", "93.42", ")", "(", "64", ",", "93.42", ")", "(", "65", ",", "93.36", ")", "(", "66", ",", "93.37", ")", "(", "67", ",", "93.37", ")", "(", "68", ",", "93.37", ")", "(", "69", ",", "93.37", ")", "(", "70", ",", "93.37", ")", "(", "71", ",", "93.37", ")", ";", "paragraph", ":", "Beam", "size", "influence", ".", "Our", "decoder", "uses", "a", "beam", "of", "a", "fixed", "size", "to", "calculate", "the", "output", "sequence", "of", "labels", ".", "We", "experimented", "with", "different", "settings", "for", "the", "beam", "size", ".", "It", "turns", "out", "that", "it", "is", "almost", "irrelevant", ".", "We", "report", "report", "results", "that", "use", "beam", "size", "10", ",", "but", "using", "beam", "size", "2", "only", "lowers", "the", "F1", "score", "of", "LSTM", "+", "A", "on", "the", "development", "set", "by", ",", "and", "using", "beam", "size", "1", "lowers", "it", "by", "(", "to", ")", ".", "Beam", "sizes", "above", "10", "do", "not", "give", "any", "additional", "improvements", ".", "paragraph", ":", "Dropout", "influence", ".", "We", "only", "used", "dropout", "when", "training", "on", "the", "small", "WSJ", "dataset", "and", "its", "influence", "was", "significant", ".", "A", "single", "LSTM", "+", "A", "model", "only", "achieved", "an", "F1", "score", "of", "on", "our", "development", "set", ",", "that", "is", "over", "points", "lower", "than", "the", "of", "a", "LSTM", "+", "A", "+", "D", "model", ".", "paragraph", ":", "Pre", "-", "training", "influence", ".", "As", "described", "in", "the", "previous", "section", ",", "we", "initialized", "the", "word", "-", "vector", "embedding", "with", "pre", "-", "trained", "word", "vectors", "obtained", "from", "word2vec", ".", "To", "test", "the", "influence", "of", "this", "initialization", ",", "we", "trained", "a", "LSTM", "+", "A", "model", "on", "the", "high", "-", "confidence", "corpus", ",", "and", "a", "LSTM", "+", "A", "+", "D", "model", "on", "the", "WSJ", "corpus", ",", "starting", "with", "randomly", "initialized", "word", "-", "vector", "embeddings", ".", "The", "F1", "score", "on", "our", "development", "set", "was", "lower", "for", "the", "LSTM", "+", "A", "model", "(", "vs", ")", "and", "lower", "for", "the", "LSTM", "+", "A", "+", "D", "model", "(", "vs", ")", ".", "So", "the", "effect", "of", "pre", "-", "training", "is", "consistent", "but", "small", ".", "paragraph", ":", "Performance", "on", "other", "datasets", ".", "The", "WSJ", "evaluation", "set", "has", "been", "in", "use", "for", "20", "years", "and", "is", "commonly", "used", "to", "compare", "syntactic", "parsers", ".", "But", "it", "is", "not", "representative", "for", "text", "encountered", "on", "the", "web", ".", "Even", "though", "our", "model", "was", "trained", "on", "a", "news", "corpus", ",", "we", "wanted", "to", "check", "how", "well", "it", "generalizes", "to", "other", "forms", "of", "text", ".", "To", "this", "end", ",", "we", "evaluated", "it", "on", "two", "additional", "datasets", ":", "1000", "held", "-", "out", "sentences", "from", "the", "Question", "Treebank", ";", "the", "first", "half", "of", "each", "domain", "from", "the", "English", "Web", "Treebank", "(", "8310", "sentences", ")", ".", "LSTM", "+", "A", "trained", "on", "the", "high", "-", "confidence", "corpus", "(", "which", "only", "includes", "text", "from", "news", ")", "achieved", "an", "F1", "score", "of", "on", "QTB", "and", "on", "WEB", ".", "Our", "score", "on", "WEB", "is", "higher", "both", "than", "the", "best", "score", "reported", "in", "(", ")", "and", "the", "best", "score", "we", "achieved", "with", "an", "in", "-", "house", "reimplementation", "of", "BerkeleyParser", "trained", "on", "human", "-", "annotated", "data", "(", ")", ".", "We", "managed", "to", "achieve", "a", "slightly", "higher", "score", "(", ")", "with", "the", "in", "-", "house", "BerkeleyParser", "trained", "on", "a", "large", "corpus", ".", "On", "QTB", ",", "the", "score", "of", "LSTM", "+", "A", "is", "also", "lower", "than", "the", "best", "score", "of", "our", "in", "-", "house", "BerkeleyParser", "(", ")", ".", "Still", ",", "taking", "into", "account", "that", "there", "were", "only", "few", "questions", "in", "the", "training", "data", ",", "these", "scores", "show", "that", "LSTM", "+", "A", "managed", "to", "generalize", "well", "beyond", "the", "news", "language", "it", "was", "trained", "on", ".", "paragraph", ":", "Parsing", "speed", ".", "Our", "LSTM", "+", "A", "model", ",", "running", "on", "a", "multi", "-", "core", "CPU", "using", "batches", "of", "128", "sentences", "on", "a", "generic", "unoptimized", "decoder", ",", "can", "parse", "over", "120", "sentences", "from", "WSJ", "per", "second", "for", "sentences", "of", "all", "lengths", "(", "using", "beam", "-", "size", "1", ")", ".", "This", "is", "better", "than", "the", "speed", "reported", "for", "this", "batch", "size", "in", "Figure", "4", "of", "at", "100", "sentences", "per", "second", ",", "even", "though", "they", "run", "on", "a", "GPU", "and", "only", "on", "sentences", "of", "under", "40", "words", ".", "Note", "that", "they", "achieve", "F1", "score", "on", "this", "subset", "of", "sentences", "of", "section", "22", ",", "while", "our", "model", "at", "beam", "-", "size", "1", "achieves", "a", "score", "of", "on", "this", "subset", ".", "section", ":", "Analysis", "As", "shown", "in", "this", "paper", ",", "the", "attention", "mechanism", "was", "a", "key", "component", "especially", "when", "learning", "from", "a", "relatively", "small", "dataset", ".", "We", "found", "that", "the", "model", "did", "not", "overfit", "and", "learned", "the", "parsing", "function", "from", "scratch", "much", "faster", ",", "which", "resulted", "in", "a", "model", "which", "generalized", "much", "better", "than", "the", "plain", "LSTM", "without", "attention", ".", "One", "of", "the", "most", "interesting", "aspects", "of", "attention", "is", "that", "it", "allows", "us", "to", "visualize", "to", "interpret", "what", "the", "model", "has", "learned", "from", "the", "data", ".", "For", "example", ",", "in", "it", "is", "shown", "that", "for", "translation", ",", "attention", "learns", "an", "alignment", "function", ",", "which", "certainly", "should", "help", "translating", "from", "English", "to", "French", ".", "Figure", "[", "reference", "]", "shows", "an", "example", "of", "the", "attention", "model", "trained", "only", "on", "the", "WSJ", "dataset", ".", "From", "the", "attention", "matrix", ",", "where", "each", "column", "is", "the", "attention", "vector", "over", "the", "inputs", ",", "it", "is", "clear", "that", "the", "model", "focuses", "quite", "sharply", "on", "one", "word", "as", "it", "produces", "the", "parse", "tree", ".", "It", "is", "also", "clear", "that", "the", "focus", "moves", "from", "the", "first", "word", "to", "the", "last", "monotonically", ",", "and", "steps", "to", "the", "right", "deterministically", "when", "a", "word", "is", "consumed", ".", "On", "the", "bottom", "of", "Figure", "[", "reference", "]", "we", "see", "where", "the", "model", "attends", "(", "black", "arrow", ")", ",", "and", "the", "current", "output", "being", "decoded", "in", "the", "tree", "(", "black", "circle", ")", ".", "This", "stack", "procedure", "is", "learned", "from", "data", "(", "as", "all", "the", "parameters", "are", "randomly", "initialized", ")", ",", "but", "is", "not", "quite", "a", "simple", "stack", "decoding", ".", "Indeed", ",", "at", "the", "input", "side", ",", "if", "the", "model", "focuses", "on", "position", ",", "that", "state", "has", "information", "for", "all", "words", "after", "(", "since", "we", "also", "reverse", "the", "inputs", ")", ".", "It", "is", "worth", "noting", "that", ",", "in", "some", "examples", "(", "not", "shown", "here", ")", ",", "the", "model", "does", "skip", "words", ".", "[", "width=1.1", "]", "attention_figure", "section", ":", "Related", "Work", "The", "task", "of", "syntactic", "constituency", "parsing", "has", "received", "a", "tremendous", "amount", "of", "attention", "in", "the", "last", "20", "years", ".", "Traditional", "approaches", "to", "constituency", "parsing", "rely", "on", "probabilistic", "context", "-", "free", "grammars", "(", "CFGs", ")", ".", "The", "focus", "in", "these", "approaches", "is", "on", "devising", "appropriate", "smoothing", "techniques", "for", "highly", "lexicalized", "and", "thus", "rare", "events", "or", "carefully", "crafting", "the", "model", "structure", ".", "partially", "alleviate", "the", "heavy", "reliance", "on", "manual", "modeling", "of", "linguistic", "structure", "by", "using", "latent", "variables", "to", "learn", "a", "more", "articulated", "model", ".", "However", ",", "their", "model", "still", "depends", "on", "a", "CFG", "backbone", "and", "is", "thereby", "potentially", "restricted", "in", "its", "capacity", ".", "Early", "neural", "network", "approaches", "to", "parsing", ",", "for", "example", "by", "also", "relied", "on", "strong", "linguistic", "insights", ".", "introduced", "Incremental", "Sigmoid", "Belief", "Networks", "for", "syntactic", "parsing", ".", "By", "constructing", "the", "model", "structure", "incrementally", ",", "they", "are", "able", "to", "avoid", "making", "strong", "independence", "assumptions", "but", "inference", "becomes", "intractable", ".", "To", "avoid", "complex", "inference", "methods", ",", "propose", "a", "recurrent", "neural", "network", "where", "parse", "trees", "are", "decomposed", "into", "a", "stack", "of", "independent", "levels", ".", "Unfortunately", ",", "this", "decomposition", "breaks", "for", "long", "sentences", "and", "their", "accuracy", "on", "longer", "sentences", "falls", "quite", "significantly", "behind", "the", "state", "-", "of", "-", "the", "-", "art", ".", "used", "a", "tree", "-", "structured", "neural", "network", "to", "score", "candidate", "parse", "trees", ".", "Their", "model", "however", "relies", "again", "on", "the", "CFG", "assumption", "and", "furthermore", "can", "only", "be", "used", "to", "score", "candidate", "trees", "rather", "than", "for", "full", "inference", ".", "Our", "LSTM", "model", "significantly", "differs", "from", "all", "these", "models", ",", "as", "it", "makes", "no", "assumptions", "about", "the", "task", ".", "As", "a", "sequence", "-", "to", "-", "sequence", "prediction", "model", "it", "is", "somewhat", "related", "to", "the", "incremental", "parsing", "models", ",", "pioneered", "by", "and", "extended", "by", ".", "Such", "linear", "time", "parsers", "however", "typically", "need", "some", "task", "-", "specific", "constraints", "and", "might", "build", "up", "the", "parse", "in", "multiple", "passes", ".", "Relatedly", ",", "present", "excellent", "parsing", "results", "with", "a", "single", "left", "-", "to", "-", "right", "pass", ",", "but", "require", "a", "stack", "to", "explicitly", "delay", "making", "decisions", "and", "a", "parsing", "-", "specific", "transition", "strategy", "in", "order", "to", "achieve", "good", "parsing", "accuracies", ".", "The", "LSTM", "in", "contrast", "uses", "its", "short", "term", "memory", "to", "model", "the", "complex", "underlying", "structure", "that", "connects", "the", "input", "-", "output", "pairs", ".", "Recently", ",", "researchers", "have", "developed", "a", "number", "of", "neural", "network", "models", "that", "can", "be", "applied", "to", "general", "sequence", "-", "to", "-", "sequence", "problems", ".", "was", "the", "first", "to", "propose", "a", "differentiable", "attention", "mechanism", "for", "the", "general", "problem", "of", "handwritten", "text", "synthesis", ",", "although", "his", "approach", "assumed", "a", "monotonic", "alignment", "between", "the", "input", "and", "output", "sequences", ".", "Later", ",", "introduced", "a", "more", "general", "attention", "model", "that", "does", "not", "assume", "a", "monotonic", "alignment", ",", "and", "applied", "it", "to", "machine", "translation", ",", "and", "applied", "the", "same", "model", "to", "speech", "recognition", ".", "used", "a", "convolutional", "neural", "network", "to", "encode", "a", "variable", "-", "sized", "input", "sentence", "into", "a", "vector", "of", "a", "fixed", "dimension", "and", "used", "a", "RNN", "to", "produce", "the", "output", "sentence", ".", "Essentially", "the", "same", "model", "has", "been", "used", "by", "to", "successfully", "learn", "to", "generate", "image", "captions", ".", "Finally", ",", "already", "in", "1990", "experimented", "with", "applying", "recurrent", "neural", "networks", "to", "the", "problem", "of", "syntactic", "parsing", ".", "section", ":", "Conclusions", "In", "this", "work", ",", "we", "have", "shown", "that", "generic", "sequence", "-", "to", "-", "sequence", "approaches", "can", "achieve", "excellent", "results", "on", "syntactic", "constituency", "parsing", "with", "relatively", "little", "effort", "or", "tuning", ".", "In", "addition", ",", "while", "we", "found", "the", "model", "of", "Sutskever", "et", "al", ".", "to", "not", "be", "particularly", "data", "efficient", ",", "the", "attention", "model", "of", "Bahdanau", "et", "al", ".", "was", "found", "to", "be", "highly", "data", "efficient", ",", "as", "it", "has", "matched", "the", "performance", "of", "the", "BerkeleyParser", "when", "trained", "on", "a", "small", "human", "-", "annotated", "parsing", "dataset", ".", "Finally", ",", "we", "showed", "that", "synthetic", "datasets", "with", "imperfect", "labels", "can", "be", "highly", "useful", ",", "as", "our", "models", "have", "substantially", "outperformed", "the", "models", "that", "have", "been", "used", "to", "create", "their", "training", "data", ".", "We", "suspect", "it", "is", "the", "case", "due", "to", "the", "different", "natures", "of", "the", "teacher", "model", "and", "the", "student", "model", ":", "the", "student", "model", "has", "likely", "viewed", "the", "teacher", "’s", "errors", "as", "noise", "which", "it", "has", "been", "able", "to", "ignore", ".", "This", "approach", "was", "so", "successful", "that", "we", "obtained", "a", "new", "state", "-", "of", "-", "the", "-", "art", "result", "in", "syntactic", "constituency", "parsing", "with", "a", "single", "attention", "model", ",", "which", "also", "means", "that", "the", "model", "is", "exceedingly", "fast", ".", "This", "work", "shows", "that", "domain", "independent", "models", "with", "excellent", "learning", "algorithms", "can", "match", "and", "even", "outperform", "domain", "specific", "models", ".", "Acknowledgement", ".", "We", "would", "like", "to", "thank", "Amin", "Ahmad", ",", "Dan", "Bikel", "and", "Jonni", "Kanerva", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["Grammar"], "Method"], [7, 10, ["Syntactic", "constituency", "parsing"], "Task"], [15, 18, ["natural", "language", "processing"], "Task"], [38, 39, ["parsers"], "Method"], [55, 66, ["domain", "agnostic", "attention", "-", "enhanced", "sequence", "-", "to", "-", "sequence", "model"], "Method"], [97, 98, ["parsers"], "Method"], [106, 107, ["parsers"], "Method"], [132, 138, ["sequence", "-", "to", "-", "sequence", "models"], "Method"], [140, 142, ["attention", "mechanism"], "Method"], [144, 145, ["parser"], "Method"], [158, 161, ["unoptimized", "CPU", "implementation"], "Method"], [165, 168, ["Syntactic", "constituency", "parsing"], "Task"], [173, 178, ["linguistics", "and", "natural", "language", "processing"], "Task"], [207, 211, ["domain", "-", "specific", "parsers"], "Method"], [213, 215, ["computational", "requirements"], "Metric"], [217, 218, ["parsers"], "Method"], [226, 234, ["linear", "-", "time", "shift", "-", "reduce", "constituency", "parsers"], "Method"], [236, 237, ["accuracy"], "Metric"], [255, 256, ["parsers"], "Method"], [260, 261, ["parsing"], "Task"], [294, 297, ["neural", "network", "model"], "Method"], [300, 307, ["general", "sequence", "-", "to", "-", "sequence", "problem"], "Task"], [319, 321, ["attention", "mechanism"], "Method"], [343, 348, ["large", "scale", "machine", "translation", "tasks"], "Task"], [353, 356, ["Syntactic", "constituency", "parsing"], "Task"], [361, 367, ["sequence", "-", "to", "-", "sequence", "problem"], "Task"], [389, 390, ["parsing"], "Task"], [399, 405, ["sequence", "-", "to", "-", "sequence", "model"], "Method"], [448, 449, ["BerkeleyParser"], "Method"], [455, 461, ["sequence", "-", "to", "-", "sequence", "model"], "Method"], [463, 464, ["BerkeleyParser"], "Method"], [472, 474, ["F1", "score"], "Metric"], [485, 486, ["WSJ"], "Material"], [492, 494, ["attention", "model"], "Method"], [517, 523, ["sequence", "-", "to", "-", "sequence", "model"], "Method"], [524, 525, ["attention"], "Method"], [539, 541, ["F1", "score"], "Metric"], [548, 549, ["WSJ"], "Material"], [554, 555, ["ensemble"], "Method"], [559, 560, ["ensemble"], "Method"], [567, 568, ["BerkeleyParser"], "Method"], [602, 603, ["parsers"], "Method"], [607, 613, ["sequence", "-", "to", "-", "sequence", "model"], "Method"], [614, 615, ["attention"], "Method"], [621, 623, ["F1", "score"], "Metric"], [630, 631, ["WSJ"], "Material"], [656, 657, ["parser"], "Method"], [754, 755, ["lstm"], "Method"], [763, 764, ["LSTM"], "Method"], [813, 814, ["lstm"], "Method"], [879, 880, ["LSTM"], "Method"], [1094, 1095, ["cVP"], "Method"], [1150, 1151, ["lstm"], "Method"], [1159, 1160, ["LSTM"], "Method"], [1293, 1294, ["lstm"], "Method"], [1301, 1302, ["LSTM"], "Method"], [1435, 1436, ["lstm"], "Method"], [1443, 1444, ["LSTM"], "Method"], [1678, 1684, ["Long", "Short", "-", "Term", "Memory", "model"], "Method"], [1714, 1715, ["LSTM"], "Method"], [1821, 1823, ["vector", "representations"], "Method"], [1855, 1856, ["softmax"], "Method"], [1914, 1917, ["Stochastic", "gradient", "descent"], "Method"], [1922, 1924, ["training", "objective"], "Metric"], [1948, 1950, ["Attention", "Mechanism"], "Method"], [1955, 1961, ["sequence", "-", "to", "-", "sequence", "model"], "Method"], [1965, 1967, ["attention", "mechanism"], "Method"], [1971, 1973, ["attention", "model"], "Method"], [1984, 1986, ["attention", "mechanism"], "Method"], [1995, 2001, ["sequence", "-", "to", "-", "sequence", "model"], "Method"], [2011, 2012, ["LSTMs"], "Method"], [2139, 2140, ["encoder"], "Method"], [2183, 2185, ["recurrent", "model"], "Method"], [2198, 2200, ["attention", "mechanism"], "Method"], [2219, 2222, ["Linearizing", "Parsing", "Trees"], "Task"], [2229, 2230, ["parsing"], "Task"], [2260, 2265, ["depth", "-", "first", "traversal", "order"], "Method"], [2418, 2419, ["parsing"], "Task"], [2572, 2574, ["dropout", "layers"], "Method"], [2599, 2603, ["POS", "-", "tag", "normalization"], "Method"], [2641, 2643, ["F1", "score"], "Metric"], [2654, 2655, ["parsers"], "Method"], [2711, 2713, ["F1", "score"], "Metric"], [2741, 2743, ["embedding", "layer"], "Method"], [2757, 2761, ["word", "-", "vector", "embeddings"], "Method"], [2766, 2770, ["skip", "-", "gram", "embeddings"], "Method"], [2774, 2775, ["word2vec"], "Method"], [2806, 2809, ["pre", "-", "training"], "Task"], [2922, 2925, ["neural", "network", "standards"], "Method"], [2941, 2942, ["MNIST"], "Material"], [2962, 2966, ["domain", "-", "specific", "parsers"], "Method"], [3009, 3011, ["OntoNotes", "corpus"], "Material"], [3015, 3018, ["English", "Web", "Treebank"], "Material"], [3022, 3025, ["corrected", "Question", "Treebank"], "Material"], [3030, 3034, ["Wall", "Street", "Journal", "section"], "Material"], [3036, 3038, ["Penn", "Treebank"], "Material"], [3042, 3044, ["OntoNotes", "corpus"], "Material"], [3085, 3086, ["parsers"], "Method"], [3089, 3094, ["tri", "-", "training", "”", "approach"], "Method"], [3101, 3102, ["parsers"], "Method"], [3111, 3112, ["ZPar"], "Method"], [3134, 3135, ["parsers"], "Method"], [3157, 3160, ["Re", "-", "sampling"], "Task"], [3163, 3164, ["parsers"], "Method"], [3209, 3210, ["parser"], "Method"], [3214, 3215, ["BerkeleyParser"], "Method"], [3267, 3269, ["EVALB", "tool"], "Method"], [3273, 3275, ["F1", "scores"], "Metric"], [3284, 3286, ["Penn", "Treebank"], "Material"], [3328, 3329, ["parsers"], "Method"], [3353, 3355, ["efficiency", "problems"], "Task"], [3360, 3364, ["semi", "-", "supervised", "results"], "Task"], [3402, 3403, ["BerkeleyParser"], "Method"], [3405, 3408, ["self", "-", "training"], "Method"], [3415, 3419, ["ensemble", "of", "multiple", "parsers"], "Method"], [3430, 3434, ["linear", "-", "time", "parser"], "Method"], [3439, 3443, ["transition", "-", "based", "parser"], "Method"], [3454, 3455, ["WSJ"], "Material"], [3469, 3470, ["dropout"], "Method"], [3471, 3473, ["early", "stopping"], "Method"], [3477, 3479, ["attention", "model"], "Method"], [3498, 3499, ["BerkeleyParser"], "Method"], [3555, 3560, ["Generating", "well", "-", "formed", "trees"], "Task"], [3748, 3749, ["BerkeleyParser"], "Method"], [3784, 3786, ["F1", "score"], "Metric"], [3799, 3800, ["BerkeleyParser"], "Method"], [3815, 3816, ["LSTM"], "Method"], [3821, 3822, ["BerkeleyParser"], "Method"], [3826, 3827, ["length"], "Metric"], [3839, 3840, ["length"], "Metric"], [3841, 3842, ["BerkeleyParser"], "Method"], [3844, 3847, ["full", "chart", "parser"], "Method"], [3940, 3942, ["F1", "score"], "Metric"], [4016, 4017, ["LSTM"], "Method"], [5077, 5079, ["F1", "score"], "Metric"], [5120, 5121, ["dropout"], "Method"], [5143, 5145, ["F1", "score"], "Metric"], [5184, 5188, ["word", "-", "vector", "embedding"], "Method"], [5196, 5197, ["word2vec"], "Material"], [5237, 5241, ["word", "-", "vector", "embeddings"], "Method"], [5243, 5245, ["F1", "score"], "Metric"], [5310, 5312, ["syntactic", "parsers"], "Method"], [5369, 5371, ["Question", "Treebank"], "Material"], [5380, 5383, ["English", "Web", "Treebank"], "Material"], [5408, 5410, ["F1", "score"], "Metric"], [5412, 5413, ["QTB"], "Material"], [5415, 5416, ["WEB"], "Material"], [5420, 5421, ["WEB"], "Material"], [5445, 5446, ["BerkeleyParser"], "Method"], [5470, 5471, ["BerkeleyParser"], "Method"], [5478, 5479, ["QTB"], "Material"], [5498, 5499, ["BerkeleyParser"], "Method"], [5540, 5542, ["Parsing", "speed"], "Metric"], [5552, 5556, ["multi", "-", "core", "CPU"], "Method"], [5564, 5566, ["unoptimized", "decoder"], "Method"], [5573, 5574, ["WSJ"], "Material"], [5616, 5617, ["GPU"], "Method"], [5630, 5632, ["F1", "score"], "Metric"], [5667, 5669, ["attention", "mechanism"], "Method"], [5751, 5752, ["translation"], "Task"], [5753, 5754, ["attention"], "Task"], [5763, 5764, ["translating"], "Task"], [5765, 5766, ["English"], "Material"], [5767, 5768, ["French"], "Material"], [5778, 5780, ["attention", "model"], "Method"], [5885, 5887, ["stack", "procedure"], "Method"], [5907, 5909, ["stack", "decoding"], "Method"], [5973, 5976, ["syntactic", "constituency", "parsing"], "Task"], [5992, 5994, ["constituency", "parsing"], "Task"], [5996, 6001, ["probabilistic", "context", "-", "free", "grammars"], "Method"], [6002, 6003, ["CFGs"], "Method"], [6014, 6016, ["smoothing", "techniques"], "Method"], [6036, 6041, ["manual", "modeling", "of", "linguistic", "structure"], "Task"], [6049, 6051, ["articulated", "model"], "Method"], [6060, 6062, ["CFG", "backbone"], "Method"], [6072, 6075, ["neural", "network", "approaches"], "Method"], [6076, 6077, ["parsing"], "Task"], [6089, 6093, ["Incremental", "Sigmoid", "Belief", "Networks"], "Method"], [6094, 6096, ["syntactic", "parsing"], "Task"], [6100, 6102, ["model", "structure"], "Method"], [6114, 6115, ["inference"], "Task"], [6121, 6123, ["inference", "methods"], "Method"], [6126, 6129, ["recurrent", "neural", "network"], "Method"], [6151, 6152, ["accuracy"], "Metric"], [6170, 6175, ["tree", "-", "structured", "neural", "network"], "Method"], [6203, 6205, ["full", "inference"], "Task"], [6227, 6234, ["sequence", "-", "to", "-", "sequence", "prediction", "model"], "Method"], [6240, 6243, ["incremental", "parsing", "models"], "Method"], [6251, 6254, ["linear", "time", "parsers"], "Method"], [6276, 6277, ["parsing"], "Task"], [6302, 6304, ["transition", "strategy"], "Method"], [6309, 6311, ["parsing", "accuracies"], "Metric"], [6313, 6314, ["LSTM"], "Method"], [6318, 6321, ["short", "term", "memory"], "Method"], [6343, 6346, ["neural", "network", "models"], "Method"], [6351, 6358, ["general", "sequence", "-", "to", "-", "sequence", "problems"], "Task"], [6365, 6368, ["differentiable", "attention", "mechanism"], "Method"], [6373, 6376, ["handwritten", "text", "synthesis"], "Task"], [6397, 6399, ["attention", "model"], "Method"], [6411, 6413, ["machine", "translation"], "Task"], [6420, 6422, ["speech", "recognition"], "Task"], [6425, 6428, ["convolutional", "neural", "network"], "Method"], [6446, 6447, ["RNN"], "Method"], [6466, 6468, ["image", "captions"], "Task"], [6477, 6480, ["recurrent", "neural", "networks"], "Method"], [6484, 6486, ["syntactic", "parsing"], "Task"], [6499, 6505, ["sequence", "-", "to", "-", "sequence", "approaches"], "Method"], [6510, 6513, ["syntactic", "constituency", "parsing"], "Task"], [6541, 6543, ["attention", "model"], "Method"], [6564, 6565, ["BerkeleyParser"], "Method"], [6622, 6624, ["teacher", "model"], "Method"], [6626, 6628, ["student", "model"], "Method"], [6630, 6632, ["student", "model"], "Method"], [6668, 6671, ["syntactic", "constituency", "parsing"], "Task"], [6674, 6676, ["attention", "model"], "Method"], [6691, 6694, ["domain", "independent", "models"], "Method"], [6696, 6698, ["learning", "algorithms"], "Method"], [6703, 6706, ["domain", "specific", "models"], "Method"], [673, 674, ["LSTM"], "Method"], [821, 822, ["LSTM"], "Method"], [871, 872, ["lstm"], "Method"], [1674, 1675, ["LSTM"], "Method"], [1759, 1760, ["LSTM"], "Method"], [1779, 1780, ["LSTM"], "Method"], [1797, 1798, ["LSTM"], "Method"], [1814, 1815, ["LSTM"], "Method"], [1891, 1892, ["LSTM"], "Method"], [1989, 1990, ["LSTM"], "Method"], [2466, 2467, ["LSTM"], "Method"], [2487, 2488, ["LSTM"], "Method"], [2531, 2532, ["LSTM"], "Method"], [2543, 2544, ["LSTM"], "Method"], [2577, 2578, ["LSTM"], "Method"], [2579, 2580, ["LSTM"], "Method"], [2584, 2585, ["LSTM"], "Method"], [2586, 2587, ["LSTM"], "Method"], [2592, 2593, ["LSTM"], "Method"], [2621, 2623, ["F1", "score"], "Metric"], [2910, 2911, ["WSJ"], "Material"], [3106, 3107, ["BerkeleyParser"], "Method"], [3153, 3154, ["WSJ"], "Material"], [3257, 3258, ["BerkeleyParser"], "Method"], [3459, 3460, ["LSTM"], "Method"], [3486, 3487, ["LSTM"], "Method"], [3500, 3501, ["WSJ"], "Material"], [3515, 3516, ["LSTM"], "Method"], [3543, 3544, ["LSTM"], "Method"], [3562, 3563, ["LSTM"], "Method"], [3568, 3569, ["WSJ"], "Material"], [3615, 3616, ["LSTM"], "Method"], [3653, 3654, ["LSTM"], "Method"], [3717, 3718, ["LSTM"], "Method"], [3752, 3753, ["LSTM"], "Method"], [3758, 3759, ["LSTM"], "Method"], [3804, 3805, ["LSTM"], "Method"], [3808, 3809, ["LSTM"], "Method"], [3832, 3833, ["LSTM"], "Method"], [3978, 3979, ["BerkeleyParser"], "Method"], [4053, 4054, ["LSTM"], "Method"], [5080, 5081, ["LSTM"], "Method"], [5126, 5127, ["WSJ"], "Material"], [5136, 5137, ["LSTM"], "Method"], [5160, 5161, ["LSTM"], "Method"], [5209, 5210, ["LSTM"], "Method"], [5222, 5223, ["LSTM"], "Method"], [5230, 5231, ["WSJ"], "Material"], [5253, 5254, ["LSTM"], "Method"], [5264, 5265, ["LSTM"], "Method"], [5294, 5295, ["WSJ"], "Material"], [5388, 5389, ["LSTM"], "Method"], [5483, 5484, ["LSTM"], "Method"], [5522, 5523, ["LSTM"], "Method"], [5544, 5545, ["LSTM"], "Method"], [5712, 5713, ["LSTM"], "Method"], [5784, 5785, ["WSJ"], "Material"], [6207, 6208, ["LSTM"], "Method"]]} {"doc_id": "28bf3aee9eecc2f7a7b4ac71bfe89534d3fe5f19", "text": ["document", ":", "Occlusion", "-", "aware", "R", "-", "CNN", ":", "Detecting", "Pedestrians", "in", "a", "Crowd", "Pedestrian", "detection", "in", "crowded", "scenes", "is", "a", "challenging", "problem", "since", "the", "pedestrians", "often", "gather", "together", "and", "occlude", "each", "other", ".", "In", "this", "paper", ",", "we", "propose", "a", "new", "occlusion", "-", "aware", "R", "-", "CNN", "(", "OR", "-", "CNN", ")", "to", "improve", "the", "detection", "accuracy", "in", "the", "crowd", ".", "Specifically", ",", "we", "design", "a", "new", "aggregation", "loss", "to", "enforce", "proposals", "to", "be", "close", "and", "locate", "compactly", "to", "the", "corresponding", "objects", ".", "Meanwhile", ",", "we", "use", "a", "new", "part", "occlusion", "-", "aware", "region", "of", "interest", "(", "PORoI", ")", "pooling", "unit", "to", "replace", "the", "RoI", "pooling", "layer", "in", "order", "to", "integrate", "the", "prior", "structure", "information", "of", "human", "body", "with", "visibility", "prediction", "into", "the", "network", "to", "handle", "occlusion", ".", "Our", "detector", "is", "trained", "in", "an", "end", "-", "to", "-", "end", "fashion", ",", "which", "achieves", "state", "-", "of", "-", "the", "-", "art", "results", "on", "three", "pedestrian", "detection", "datasets", ",", "i.e.", ",", "CityPersons", ",", "ETH", ",", "and", "INRIA", ",", "and", "performs", "on", "-", "pair", "with", "the", "state", "-", "of", "-", "the", "-", "arts", "on", "Caltech", ".", "section", ":", "Introduction", "Pedestrian", "detection", "is", "an", "important", "research", "topic", "in", "computer", "vision", "field", "with", "various", "applications", ",", "such", "as", "autonomous", "driving", ",", "video", "surveillance", ",", "and", "robotics", ",", "which", "aims", "to", "predict", "a", "series", "of", "bounding", "boxes", "enclosing", "pedestrian", "instances", "in", "an", "image", ".", "Recent", "advances", "in", "object", "detection", "are", "driven", "by", "the", "success", "of", "deep", "convolutional", "neural", "networks", "(", "CNNs", ")", ",", "which", "uses", "the", "bounding", "box", "regression", "techniques", "to", "accurately", "localize", "the", "objects", "based", "on", "the", "deep", "features", ".", "Actually", ",", "in", "real", "life", "complex", "scenarios", ",", "occlusion", "is", "one", "of", "the", "most", "significant", "challenges", "in", "detecting", "pedestrian", ",", "especially", "in", "the", "crowded", "scenes", ".", "For", "example", ",", "as", "pointed", "out", "in", ",", "annotated", "pedestrians", "are", "occluded", "by", "other", "pedestrians", "in", "the", "CityPersons", "dataset", ".", "Previous", "methods", "only", "require", "each", "predicted", "bounding", "box", "to", "be", "close", "to", "its", "designated", "ground", "truth", ",", "without", "considering", "the", "relations", "among", "them", ".", "Thus", ",", "they", "make", "the", "detectors", "sensitive", "to", "the", "threshold", "of", "non", "-", "maximum", "suppression", "(", "NMS", ")", "in", "the", "crowded", "scenes", ",", "wherein", "filling", "with", "occlusions", ".", "To", "that", "end", ",", "Wang", "et", "al", ".", "[", "]", "design", "a", "repulsion", "loss", ",", "which", "not", "only", "pushes", "each", "proposal", "to", "approach", "its", "designated", "target", ",", "but", "also", "to", "keep", "it", "away", "from", "the", "other", "ground", "truth", "objects", "and", "their", "corresponding", "designated", "proposals", ".", "However", ",", "it", "is", "difficult", "to", "control", "the", "balance", "between", "the", "repulsion", "and", "attraction", "terms", "in", "the", "loss", "function", "to", "handle", "the", "overlapping", "pedestrians", ".", "In", "this", "paper", ",", "we", "propose", "a", "new", "occlusion", "-", "aware", "R", "-", "CNN", "(", "OR", "-", "CNN", ")", "based", "on", "the", "Faster", "R", "-", "CNN", "detection", "framework", "to", "mitigate", "the", "impact", "of", "occlusion", "challenge", ".", "Specifically", ",", "to", "reduce", "the", "false", "detections", "of", "the", "adjacent", "overlapping", "pedestrians", ",", "we", "expect", "the", "proposals", "to", "be", "close", "and", "locate", "compactly", "to", "the", "corresponding", "objects", ".", "Thus", ",", "inspired", "by", "the", "herd", "behavior", "in", "psychology", ",", "we", "design", "a", "new", "loss", "function", ",", "called", "aggregation", "loss", "(", "AggLoss", ")", ",", "not", "only", "to", "enforce", "proposals", "to", "be", "close", "to", "the", "corresponding", "objects", ",", "but", "also", "to", "minimize", "the", "internal", "region", "distances", "of", "proposals", "associated", "with", "the", "same", "objects", ".", "Meanwhile", ",", "to", "effectively", "handle", "partial", "occlusion", ",", "we", "propose", "a", "new", "part", "occlusion", "-", "aware", "region", "of", "interest", "(", "PORoI", ")", "pooling", "unit", "to", "replace", "the", "original", "RoI", "pooling", "layer", "in", "the", "second", "stage", "Fast", "R", "-", "CNN", "module", "of", "the", "detector", ",", "which", "integrates", "the", "prior", "structure", "information", "of", "human", "body", "with", "visibility", "prediction", "into", "the", "network", ".", "That", "is", ",", "we", "first", "partition", "the", "pedestrian", "region", "into", "five", "parts", ",", "and", "pool", "the", "features", "under", "each", "part", "’s", "projection", "as", "well", "as", "the", "whole", "proposal", "’s", "projection", "onto", "the", "feature", "map", "into", "fixed", "-", "length", "feature", "vectors", "by", "adaptively", "-", "sized", "pooling", "bins", ".", "After", "that", ",", "we", "use", "the", "learned", "sub", "-", "network", "to", "predict", "the", "visibility", "score", "of", "each", "part", "to", "combine", "the", "extracted", "features", "for", "pedestrian", "detection", ".", "Several", "experiments", "are", "carried", "out", "on", "four", "pedestrian", "detection", "datasets", ",", "i.e.", ",", "CityPersons", ",", "Caltech", ",", "ETH", "and", "INRIA", ",", "to", "demonstrate", "the", "superiority", "of", "the", "proposed", "method", ",", "especially", "for", "the", "crowded", "scenes", ".", "Notably", ",", "the", "proposed", "OR", "-", "CNN", "method", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "results", "with", "MR", "on", "the", "CityPersons", "dataset", ",", "MR", "on", "the", "ETH", "dataset", ",", "and", "MR", "on", "the", "INRIA", "dataset", ".", "The", "main", "contributions", "of", "this", "work", "are", "summarized", "as", "follows", ".", "We", "propose", "a", "new", "occlusion", "-", "aware", "R", "-", "CNN", "method", ",", "which", "uses", "a", "new", "designed", "AggLoss", "to", "enforce", "proposals", "to", "be", "close", "to", "the", "corresponding", "objects", ",", "as", "well", "as", "minimize", "the", "internal", "region", "distances", "of", "proposals", "associated", "with", "the", "same", "objects", ".", "We", "design", "a", "new", "PORoI", "pooling", "unit", "to", "replace", "the", "RoI", "pooling", "layer", "in", "the", "second", "Fast", "R", "-", "CNN", "module", "to", "integrate", "the", "prior", "structure", "information", "of", "human", "body", "with", "visibility", "prediction", "into", "the", "network", ".", "Several", "experiments", "are", "carried", "out", "on", "four", "challenging", "pedestrian", "detection", "datasets", ",", "i.e.", ",", "CityPersons", ",", "Caltech", ",", "ETH", ",", "and", "INRIA", ",", "to", "demonstrate", "the", "superiority", "of", "the", "proposed", "method", ".", "section", ":", "Related", "Work", "Generic", "Object", "Detection", ".", "Early", "generic", "object", "detectors", "rely", "on", "the", "sliding", "window", "paradigm", "based", "on", "the", "hand", "-", "crafted", "features", "and", "classifiers", "to", "find", "the", "objects", "of", "interest", ".", "In", "recent", "years", ",", "with", "the", "advent", "of", "deep", "convolutional", "neural", "network", "(", "CNN", ")", ",", "a", "new", "generation", "of", "more", "effective", "object", "detection", "methods", "based", "on", "CNN", "significantly", "improve", "the", "state", "-", "of", "-", "the", "-", "art", "performances", ",", "which", "can", "be", "roughly", "divided", "into", "two", "categories", ",", "i.e.", ",", "the", "one", "-", "stage", "approach", "and", "the", "two", "-", "stage", "approach", ".", "The", "one", "-", "stage", "approach", "directly", "predicts", "object", "class", "label", "and", "regresses", "object", "bounding", "box", "based", "on", "the", "pre", "-", "tiled", "anchor", "boxes", "using", "deep", "CNNs", ".", "The", "main", "advantage", "of", "the", "one", "-", "stage", "approach", "is", "its", "high", "computational", "efficiency", ".", "In", "contrast", "to", "the", "one", "-", "stage", "approach", ",", "the", "two", "-", "stage", "approach", "always", "achieves", "top", "accuracy", "on", "several", "benchmarks", ",", "which", "first", "generates", "a", "pool", "of", "object", "proposals", "by", "a", "separated", "proposal", "generator", "(", "e.g.", ",", "Selective", "Search", ",", "EdgeBoxes", ",", "and", "RPN", ")", ",", "and", "then", "predicts", "the", "class", "label", "and", "accurate", "location", "and", "size", "of", "each", "proposal", ".", "Pedestrian", "Detection", ".", "Even", "as", "one", "of", "the", "long", "-", "standing", "problems", "in", "computer", "vision", "field", "with", "an", "extensive", "literature", ",", "pedestrian", "detection", "still", "receives", "considerable", "interests", "with", "a", "wide", "range", "of", "applications", ".", "A", "common", "paradigm", "to", "address", "this", "problem", "is", "to", "train", "a", "pedestrian", "detector", "that", "exhaustively", "operates", "on", "the", "sub", "-", "images", "across", "all", "locations", "and", "scales", ".", "Dalal", "and", "Triggs", "design", "the", "histograms", "of", "oriented", "gradient", "(", "HOG", ")", "descriptors", "and", "support", "vector", "machine", "(", "SVM", ")", "classifier", "for", "human", "detection", ".", "Dollár", "et", "al", ".", "[", "]", "demonstrate", "that", "using", "features", "from", "multiple", "channels", "can", "significantly", "improve", "the", "performance", ".", "Zhang", "et", "al", ".", "[", "]", "provide", "a", "systematic", "analysis", "for", "the", "filtered", "channel", "features", ",", "and", "find", "that", "with", "the", "proper", "filter", "bank", ",", "filtered", "channel", "features", "can", "reach", "top", "detection", "quality", ".", "Paisitkriangkrai", "et", "al", ".", "[", "]", "design", "a", "new", "features", "built", "on", "the", "basis", "of", "low", "-", "level", "visual", "features", "and", "spatial", "pooling", ",", "and", "directly", "optimize", "the", "partial", "area", "under", "the", "ROC", "curve", "for", "better", "performance", ".", "Recently", ",", "pedestrian", "detection", "is", "dominated", "by", "the", "CNN", "-", "based", "methods", "(", "e.g.", ",", ")", ".", "Sermanet", "et", "al", ".", "[", "]", "present", "an", "unsupervised", "method", "using", "the", "convolutional", "sparse", "coding", "to", "pre", "-", "train", "CNN", "for", "pedestrian", "detection", ".", "In", ",", "a", "complexity", "-", "aware", "cascaded", "detector", "is", "proposed", "for", "an", "optimal", "trade", "-", "off", "between", "accuracy", "and", "speed", ".", "Angelova", "et", "al", ".", "[", "]", "combine", "the", "ideas", "of", "fast", "cascade", "and", "a", "deep", "network", "to", "detect", "pedestrian", ".", "Yang", "et", "al", ".", "[", "]", "use", "scale", "-", "dependent", "pooling", "and", "layer", "-", "wise", "cascaded", "rejection", "classifiers", "to", "detect", "objects", "efficiently", ".", "Zhang", "et", "al", ".", "[", "]", "present", "an", "effective", "pipeline", "for", "pedestrian", "detection", "via", "using", "RPN", "followed", "by", "boosted", "forests", ".", "To", "jointly", "learn", "pedestrian", "detection", "with", "the", "given", "extra", "features", ",", "a", "novel", "network", "architecture", "is", "presented", "in", ".", "Li", "et", "al", ".", "[", "]", "use", "multiple", "built", "-", "in", "sub", "-", "networks", "to", "adaptively", "detect", "pedestrians", "across", "scales", ".", "Brazil", "et", "al", ".", "[", "]", "exploit", "weakly", "annotated", "boxes", "via", "a", "segmentation", "infusion", "network", "to", "achieve", "considerable", "performance", "gains", ".", "However", ",", "occlusion", "still", "remains", "one", "of", "the", "most", "significant", "challenges", "in", "pedestrian", "detection", ",", "which", "increases", "the", "difficulty", "in", "pedestrian", "localization", ".", "Several", "methods", "use", "part", "-", "based", "model", "to", "describe", "the", "pedestrian", "in", "occlusion", "handling", ",", "which", "learn", "a", "series", "of", "part", "detectors", "and", "design", "some", "mechanisms", "to", "fuse", "the", "part", "detection", "results", "to", "localize", "partially", "occluded", "pedestrians", ".", "Besides", "the", "part", "-", "based", "model", ",", "Leibe", "et", "al", ".", "[", "]", "propose", "an", "implicit", "shape", "model", "to", "generate", "a", "set", "of", "pedestrian", "hypotheses", "that", "are", "further", "refined", "to", "obtain", "the", "visible", "regions", ".", "Wang", "et", "al", ".", "[", "]", "divide", "the", "template", "of", "pedestrian", "into", "a", "set", "of", "blocks", "and", "conduct", "occlusion", "reasoning", "by", "estimating", "the", "visibility", "status", "of", "each", "block", ".", "Ouyang", "et", "al", ".", "[", "]", "exploit", "multi", "-", "pedestrian", "detectors", "to", "aid", "single", "-", "pedestrian", "detectors", "to", "handle", "partial", "occlusions", ",", "especially", "when", "the", "pedestrians", "gather", "together", "and", "occlude", "each", "other", "in", "real", "-", "world", "scenarios", ".", "In", ",", "a", "set", "of", "occlusion", "patterns", "of", "pedestrians", "are", "discovered", "to", "learn", "a", "mixture", "of", "occlusion", "-", "specific", "detectors", ".", "Zhou", "et", "al", ".", "[", "]", "propose", "to", "jointly", "learn", "part", "detectors", "so", "as", "to", "exploit", "part", "correlations", "and", "reduce", "the", "computational", "cost", ".", "Wang", "et", "al", ".", "[", "]", "introduce", "a", "novel", "bounding", "box", "regression", "loss", "to", "detect", "pedestrians", "in", "the", "crowd", "scenes", ".", "Although", "numerous", "pedestrian", "detection", "methods", "are", "presented", "in", "literature", ",", "how", "to", "robustly", "detect", "each", "individual", "pedestrian", "in", "crowded", "scenarios", "is", "still", "one", "of", "the", "most", "critical", "issues", "for", "pedestrian", "detectors", ".", "section", ":", "Occlusion", "-", "aware", "R", "-", "CNN", "Our", "occlusion", "-", "aware", "R", "-", "CNN", "detector", "follows", "the", "adaptive", "Faster", "R", "-", "CNN", "detection", "framework", "for", "pedestrian", "detection", ",", "with", "the", "new", "designed", "aggregation", "loss", "(", "Section", "[", "reference", "]", ")", ",", "and", "the", "PORoI", "pooling", "unit", "(", "Section", "[", "reference", "]", ")", ".", "Specifically", ",", "Faster", "R", "-", "CNN", "consists", "of", "two", "modules", ",", "i.e.", ",", "the", "first", "region", "proposal", "network", "(", "RPN", ")", "module", "and", "the", "second", "Fast", "R", "-", "CNN", "module", ".", "The", "RPN", "module", "is", "designed", "to", "generate", "high", "-", "quality", "region", "proposals", ",", "and", "the", "Fast", "R", "-", "CNN", "module", "is", "used", "to", "classify", "and", "regress", "the", "accurate", "locations", "and", "sizes", "of", "objects", ",", "based", "on", "the", "generated", "proposals", ".", "To", "effectively", "generate", "accurate", "region", "proposals", "in", "the", "first", "RPN", "module", ",", "we", "design", "the", "AggLoss", "term", "to", "enforce", "the", "proposals", "locate", "closely", "and", "compactly", "to", "the", "ground", "-", "truth", "object", ",", "which", "is", "defined", "as", "where", "is", "the", "index", "of", "anchor", "in", "a", "mini", "-", "batch", ",", "and", "are", "the", "predicted", "confidence", "of", "the", "-", "th", "anchor", "being", "a", "pedestrian", "and", "the", "predicted", "coordinates", "of", "the", "pedestrian", ",", "and", "are", "the", "associated", "ground", "truth", "class", "label", "and", "coordinates", "of", "the", "-", "th", "anchor", ",", "is", "the", "hyperparameters", "used", "to", "balance", "the", "two", "loss", "terms", ",", "is", "the", "classification", "loss", ",", "and", "is", "the", "AggLoss", "(", "see", "Section", "[", "reference", "]", ")", ".", "We", "use", "the", "log", "loss", "to", "calculate", "the", "classification", "loss", "over", "two", "classes", "(", "pedestrian", "vs.", "background", ")", ",", "i.e.", ",", "where", "is", "the", "total", "number", "of", "anchors", "in", "classification", ".", "subsection", ":", "Aggregation", "Loss", "To", "reduce", "the", "false", "detections", "of", "the", "adjacent", "overlapping", "pedestrians", ",", "we", "enforce", "proposals", "to", "be", "close", "and", "locate", "compactly", "to", "the", "corresponding", "ground", "truth", "objects", ".", "To", "that", "end", ",", "we", "design", "a", "new", "aggregation", "loss", "(", "AggLoss", ")", "for", "both", "the", "region", "proposal", "network", "(", "RPN", ")", "and", "Fast", "R", "-", "CNN", "modules", "in", "the", "Faster", "R", "-", "CNN", "algorithm", ",", "which", "is", "a", "multi", "-", "task", "loss", "pushing", "proposals", "to", "be", "close", "to", "the", "corresponding", "ground", "truth", "object", ",", "while", "minimizing", "the", "internal", "region", "distances", "of", "proposals", "associated", "with", "the", "same", "objects", ",", "i.e.", ",", "where", "is", "the", "regression", "loss", "which", "requires", "each", "proposal", "to", "approach", "the", "designated", "ground", "truth", ",", "and", "is", "the", "compactness", "loss", "which", "enforces", "proposals", "locate", "compactly", "to", "the", "designated", "ground", "truth", "object", ",", "and", "is", "the", "hyper", "-", "parameters", "used", "to", "balance", "the", "two", "loss", "terms", ".", "Similar", "to", "Fast", "R", "-", "CNN", ",", "we", "use", "the", "smooth", "L1", "loss", "as", "the", "regression", "loss", "to", "measure", "the", "accuracy", "of", "predicted", "bounding", "boxes", ",", "i.e.", ",", "where", "is", "the", "total", "number", "of", "anchors", "in", "regression", ",", "and", "is", "the", "smooth", "L1", "loss", "of", "the", "predicted", "bounding", "box", ".", "The", "compactness", "term", "is", "designed", "to", "consider", "the", "attractiveness", "among", "proposals", "associated", "with", "the", "same", "ground", "truth", "object", ".", "In", "this", "way", ",", "we", "can", "make", "the", "proposals", "to", "locate", "compactly", "around", "the", "ground", "truth", "to", "reduce", "the", "false", "detections", "of", "adjacent", "overlapping", "objects", ".", "Specifically", ",", "we", "set", "to", "be", "the", "ground", "truth", "set", "associated", "with", "more", "than", "one", "anchor", ",", "and", "to", "be", "the", "index", "sets", "of", "the", "associated", "anchors", "corresponding", "to", "the", "ground", "truth", "objects", ",", "i.e.", ",", "the", "anchors", "indexed", "by", "are", "associated", "to", "the", "ground", "truth", ",", "where", "is", "the", "total", "number", "of", "ground", "-", "truth", "object", "associated", "with", "more", "than", "one", "anchor", ".", "Thus", ",", "we", "have", ",", "for", ",", "and", ".", "We", "use", "the", "smooth", "L1", "loss", "to", "measure", "the", "difference", "between", "the", "average", "predictions", "of", "the", "anchors", "indexed", "by", "each", "set", "in", "and", "the", "corresponding", "ground", "truth", "object", ",", "describing", "the", "compactness", "of", "predicted", "bounding", "boxes", "with", "respect", "to", "the", "ground", "truth", "object", ",", "i.e.", ",", "where", "is", "the", "total", "number", "of", "ground", "truth", "object", "associated", "with", "more", "than", "one", "anchor", "(", "i.e.", ",", ")", ",", "and", "is", "the", "number", "of", "anchors", "associated", "with", "the", "-", "th", "ground", "truth", "object", ".", "subsection", ":", "Part", "Occlusion", "-", "aware", "RoI", "Pooling", "Unit", "In", "real", "life", "complex", "scenarios", ",", "occlusion", "is", "ubiquitous", "challenging", "the", "accuracy", "of", "detectors", ",", "especially", "in", "crowded", "scenes", ".", "As", "indicated", "in", ",", "the", "part", "-", "based", "model", "is", "effective", "in", "handling", "occluded", "pedestrians", ".", "In", "contrast", "to", "the", "aforementioned", "methods", ",", "we", "design", "a", "new", "part", "occlusion", "-", "aware", "RoI", "pooling", "unit", "to", "integrate", "the", "prior", "structure", "information", "of", "human", "body", "with", "visibility", "prediction", "into", "the", "Fast", "R", "-", "CNN", "module", "of", "the", "detector", ",", "which", "assembles", "a", "micro", "neural", "network", "to", "estimate", "the", "part", "occlusion", "status", ".", "As", "shown", "in", "Figure", "[", "reference", "]", "(", "a", ")", ",", "we", "first", "divide", "the", "pedestrian", "region", "into", "five", "parts", "with", "the", "empirical", "ratio", "in", ".", "For", "each", "part", ",", "we", "use", "the", "RoI", "pooling", "layer", "to", "pool", "the", "features", "into", "a", "small", "feature", "map", "with", "a", "fixed", "spatial", "extent", "of", "(", "e.g.", ",", ")", ".", "We", "introduce", "an", "occlusion", "process", "unit", ",", "shown", "in", "Figure", "[", "reference", "]", "(", "b", ")", ",", "to", "predict", "the", "visibility", "score", "of", "the", "corresponding", "part", "based", "on", "the", "pooled", "features", ".", "Specifically", ",", "the", "occlusion", "process", "unit", "is", "constructed", "by", "three", "convolutional", "layers", "followed", "by", "a", "softmax", "layer", "with", "the", "log", "loss", "in", "training", ".", "Symbolically", ",", "indicates", "the", "-", "th", "part", "of", "the", "-", "th", "proposal", ",", "represents", "its", "predicted", "visibility", "score", ",", "and", "is", "the", "corresponding", "ground", "truth", "visibility", "score", ".", "If", "half", "of", "the", "part", "is", "visible", ",", ",", "otherwise", ".", "Mathematically", ",", "if", "the", "intersection", "between", "and", "the", "visible", "region", "of", "ground", "truth", "object", "divided", "by", "the", "area", "of", "is", "larger", "than", "the", "threshold", ",", ",", "otherwise", ".", "That", "is", "where", "is", "the", "area", "computing", "function", ",", "is", "the", "region", "of", ",", "is", "the", "visible", "region", "of", "the", "ground", "truth", "object", ",", "and", "is", "the", "intersection", "operation", "between", "two", "regions", ".", "Then", ",", "the", "loss", "function", "of", "the", "occlusion", "process", "unit", "is", "calculated", "as", ".", "After", "that", ",", "we", "apply", "the", "element", "-", "wise", "multiplication", "operator", "to", "multiply", "the", "pooled", "features", "of", "each", "part", "and", "the", "corresponding", "predicted", "visibility", "score", "to", "generate", "the", "final", "features", "with", "the", "dimensions", ".", "The", "element", "-", "wise", "summation", "operation", "is", "further", "used", "to", "combine", "the", "extracted", "features", "of", "the", "five", "parts", "and", "the", "whole", "proposal", "for", "classification", "and", "regression", "in", "the", "Fast", "R", "-", "CNN", "module", "(", "see", "Figure", "[", "reference", "]", ")", ".", "To", "further", "improve", "the", "regression", "accuracy", ",", "we", "also", "use", "AggLoss", "in", "the", "Fast", "R", "-", "CNN", "module", ",", "which", "is", "defined", "as", ":", "where", "and", "are", "used", "to", "balance", "the", "three", "loss", "terms", ",", "and", "are", "the", "classification", "and", "aggregation", "losses", ",", "defined", "the", "same", "as", "that", "in", "the", "RPN", "module", ",", "and", "is", "the", "occlusion", "process", "loss", ".", "section", ":", "Experiments", "Several", "experiments", "are", "conducted", "on", "four", "datasets", ":", "CityPersons", ",", "Caltech", "-", "USA", ",", "ETH", ",", "and", "INRIA", ",", "to", "demonstrate", "the", "performance", "of", "the", "proposed", "OR", "-", "CNN", "method", ".", "subsection", ":", "Experimental", "Setup", "Our", "OR", "-", "CNN", "detector", "follows", "the", "adaptive", "Faster", "R", "-", "CNN", "framework", "and", "uses", "VGG", "-", "16", "as", "the", "backbone", "network", ",", "pre", "-", "trained", "on", "the", "ILSVRC", "CLS", "-", "LOC", "dataset", ".", "To", "improve", "the", "detection", "accuracy", "of", "pedestrians", "with", "small", "scale", ",", "we", "use", "the", "method", "presented", "in", "to", "dense", "the", "anchor", "boxes", "with", "the", "height", "less", "than", "pixels", "two", "times", ",", "and", "use", "the", "matching", "strategy", "in", "to", "associate", "the", "anchors", "and", "the", "ground", "truth", "objects", ".", "All", "the", "parameters", "in", "the", "newly", "added", "convolutional", "layers", "are", "randomly", "initialized", "by", "the", "“", "xavier", "”", "method", ".", "We", "optimize", "the", "OR", "-", "CNN", "detector", "using", "the", "Stochastic", "Gradient", "Descent", "(", "SGD", ")", "algorithm", "with", "momentum", "and", "weight", "decay", ",", "which", "is", "trained", "on", "Titan", "X", "GPUs", "with", "the", "mini", "-", "batch", "involving", "image", "per", "GPU", ".", "For", "the", "Citypersons", "dataset", ",", "we", "set", "the", "learning", "rate", "to", "for", "the", "first", "iterations", ",", "and", "decay", "it", "to", "for", "another", "iterations", ".", "For", "the", "Caltech", "-", "USA", "dataset", ",", "we", "train", "the", "network", "for", "iterations", "with", "the", "initial", "learning", "rate", "and", "decrease", "it", "by", "a", "factor", "of", "after", "the", "first", "iterations", ".", "All", "the", "hyperparameters", ",", "and", "are", "empirically", "set", "to", ".", "subsection", ":", "CityPersons", "Dataset", "The", "CityPersons", "dataset", "is", "built", "upon", "the", "semantic", "segmentation", "dataset", "Cityscapes", "to", "provide", "a", "new", "dataset", "of", "interest", "for", "pedestrian", "detection", ".", "It", "is", "recorded", "across", "different", "cities", "in", "Germany", "with", "different", "seasons", "and", "various", "weather", "conditions", ".", "The", "dataset", "includes", "images", "(", "for", "training", ",", "for", "validation", ",", "and", "for", "testing", ")", "with", "manually", "annotated", "persons", "plus", "ignore", "region", "annotations", ".", "Both", "the", "bounding", "boxes", "and", "visible", "parts", "of", "pedestrians", "are", "provided", "and", "there", "are", "approximately", "pedestrians", "in", "average", "per", "image", ".", "Following", "the", "evaluation", "protocol", "in", "CityPersons", ",", "we", "train", "our", "OR", "-", "CNN", "detector", "on", "the", "training", "set", ",", "and", "evaluate", "it", "on", "both", "the", "validation", "and", "the", "testing", "sets", ".", "The", "log", "miss", "rate", "averaged", "over", "the", "false", "positive", "per", "image", "(", "FPPI", ")", "range", "of", "(", ")", "is", "used", "to", "measure", "the", "detection", "performance", "(", "lower", "score", "indicates", "better", "performance", ")", ".", "We", "use", "the", "adaptive", "Faster", "R", "-", "CNN", "method", "trained", "by", "ourselves", "as", "the", "baseline", "detector", ",", "which", "achieves", "on", "the", "validation", "set", "with", "scale", ",", "sightly", "better", "than", "the", "reported", "result", "(", ")", "in", ".", "subsubsection", ":", "Ablation", "Study", "on", "AggLoss", "To", "demonstrate", "the", "effectiveness", "of", "AggLoss", ",", "we", "construct", "a", "detector", ",", "denoted", "as", "OR", "-", "CNN", "-", "A", ",", "that", "use", "AggLoss", "instead", "of", "the", "original", "regression", "loss", "in", "the", "baseline", "detector", ",", "and", "evaluate", "it", "on", "the", "validation", "set", "of", "CityPersons", "in", "Table", "[", "reference", "]", ".", "For", "a", "fair", "comparison", ",", "we", "use", "the", "same", "setting", "of", "parameters", "of", "OR", "-", "CNN", "-", "A", "and", "our", "OR", "-", "CNN", "detector", "in", "both", "training", "and", "testing", ".", "All", "of", "the", "experiments", "are", "conducted", "on", "the", "reasonable", "train", "/", "validation", "sets", "for", "training", "and", "testing", ".", "Comparing", "the", "detection", "results", "between", "the", "baseline", "and", "OR", "-", "CNN", "-", "A", "in", "Table", "[", "reference", "]", ",", "we", "find", "that", "using", "the", "newly", "proposed", "AggLoss", "can", "reduce", "the", "by", "(", "i.e.", ",", "vs.", ")", "with", "scale", ".", "It", "is", "worth", "noting", "that", "the", "OR", "-", "CNN", "-", "A", "detector", "achieves", "with", "scale", ",", "surpassing", "the", "state", "-", "of", "-", "the", "-", "art", "method", "using", "Repulsion", "Loss", "(", ")", ",", "which", "demonstrates", "that", "AggLoss", "is", "more", "effective", "than", "Repulsion", "Loss", "for", "detecting", "the", "pedestrians", "in", "a", "crowd", ".", "In", "addition", ",", "we", "also", "show", "some", "visual", "comparison", "results", "of", "the", "predicted", "bounding", "boxes", "before", "NMS", "of", "the", "baseline", "and", "OR", "-", "CNN", "-", "A", "detectors", "in", "Figure", "[", "reference", "]", "(", "a", ")", ".", "As", "shown", "in", "Figure", "[", "reference", "]", "(", "a", ")", ",", "the", "predictions", "of", "OR", "-", "CNN", "-", "A", "locate", "more", "compactly", "than", "that", "of", "the", "baseline", "detector", ",", "and", "there", "are", "fewer", "predictions", "of", "OR", "-", "CNN", "-", "A", "lying", "in", "between", "two", "adjacent", "ground", "-", "truth", "objects", "than", "the", "baseline", "detector", ".", "This", "phenomenon", "demonstrates", "that", "AggLoss", "can", "push", "the", "predictions", "lying", "compactly", "to", "the", "ground", "-", "truth", "objects", ",", "making", "the", "detector", "less", "sensitive", "to", "the", "NMS", "threshold", "with", "better", "performance", "in", "the", "crowd", "scene", ".", "To", "further", "validate", "this", "point", ",", "we", "also", "present", "the", "results", "with", "AggLoss", "across", "various", "NMS", "threshold", "at", "in", "Figure", "[", "reference", "]", "(", "b", ")", ".", "A", "high", "NMS", "threshold", "may", "lead", "to", "more", "false", "positives", ",", "while", "a", "low", "NMS", "threshold", "may", "lead", "to", "more", "false", "negatives", ".", "As", "shown", "in", "Figure", "[", "reference", "]", "(", "b", ")", ",", "we", "find", "that", "the", "curve", "of", "OR", "-", "CNN", "-", "A", "is", "smoother", "than", "that", "of", "baseline", "(", "i.e.", ",", "the", "variances", "of", "the", "miss", "rates", "are", "vs.", ")", ",", "which", "indicates", "that", "the", "former", "is", "less", "sensitive", "to", "the", "NMS", "threshold", ".", "It", "is", "worth", "noting", "that", "across", "various", "NMS", "thresholds", "at", ",", "the", "OR", "-", "CNN", "-", "A", "method", "always", "produces", "lower", "miss", "rate", ",", "which", "is", "due", "to", "the", "NMS", "operation", "filtering", "out", "more", "false", "positives", "in", "the", "predictions", "of", "OR", "-", "CNN", "-", "A", "than", "that", "of", "baseline", ",", "implying", "that", "the", "predicted", "bounding", "boxes", "of", "OR", "-", "CNN", "-", "A", "locate", "compactly", "than", "baseline", ".", "subsubsection", ":", "Ablation", "Study", "on", "PORoI", "Pooling", "To", "validate", "the", "effectiveness", "of", "the", "PORoI", "pooling", "unit", ",", "we", "construct", "a", "detector", ",", "denoted", "as", "OR", "-", "CNN", "-", "P", ",", "that", "use", "the", "PORoI", "pooling", "unit", "instead", "of", "the", "RoI", "pooling", "layer", "in", "baseline", ",", "and", "evaluate", "it", "on", "the", "validation", "set", "of", "CityPersons", "in", "Table", "[", "reference", "]", ".", "For", "a", "fair", "comparison", ",", "we", "use", "the", "same", "parameter", "settings", "of", "OR", "-", "CNN", "-", "P", "and", "our", "OR", "-", "CNN", "detector", "in", "both", "training", "and", "testing", ".", "All", "of", "the", "ablation", "experiments", "involved", "CityPersons", "are", "conducted", "on", "the", "reasonable", "train", "/", "validation", "sets", "for", "training", "and", "testing", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "comparing", "to", "baseline", ",", "OR", "-", "CNN", "-", "P", "reduces", "with", "scale", "(", "i.e.", ",", "vs.", ")", ",", "which", "demonstrates", "the", "effectiveness", "of", "the", "PORoI", "pooling", "unit", "in", "pedestrian", "detection", ".", "Meanwhile", ",", "we", "also", "present", "some", "qualitative", "results", "of", "the", "predictions", "with", "the", "visibility", "scores", "of", "the", "corresponding", "parts", "in", "Figure", "[", "reference", "]", ".", "Notably", ",", "we", "find", "that", "the", "visibility", "scores", "predicted", "by", "the", "PORoI", "pooling", "unit", "are", "in", "accordance", "with", "the", "human", "visual", "system", ".", "As", "shown", "in", "Figure", "[", "reference", "]", "(", "a", ")", "and", "(", "b", ")", ",", "if", "the", "pedestrian", "is", "not", "occluded", ",", "the", "visibility", "score", "of", "each", "part", "of", "the", "pedestrian", "approaches", ".", "However", ",", "if", "some", "parts", "of", "the", "pedestrians", "are", "occluded", "by", "the", "background", "obstacles", "or", "other", "pedestrians", ",", "the", "scores", "of", "the", "corresponding", "parts", "decrease", ",", "such", "as", "the", "occluded", "thigh", "and", "calf", "in", "Figure", "[", "reference", "]", "(", "c", ")-(", "f", ")", ".", "Besides", ",", "if", "two", "pedestrians", "gather", "together", "and", "occlude", "each", "other", ",", "our", "PORoI", "pooling", "unit", "successfully", "detects", "the", "occluded", "human", "parts", "that", "can", "help", "lower", "the", "contributions", "of", "the", "occluded", "parts", "in", "pedestrian", "detection", ",", "see", "Figure", "[", "reference", "]", "(", "g", ")", "and", "(", "h", ")", ".", "Notably", ",", "the", "detection", "accuracy", "of", "the", "OR", "-", "CNN", "detector", "can", "not", "be", "improved", "if", "we", "fix", "the", "visibility", "score", "of", "each", "part", "to", "instead", "of", "using", "the", "predictions", "of", "the", "occlusion", "process", "unit", "(", "see", "Figure", "[", "reference", "]", ")", ".", "Thus", ",", "the", "occlusion", "process", "unit", "is", "the", "key", "component", "to", "detection", "accuracy", ",", "since", "it", "enables", "our", "PORoI", "pooling", "unit", "to", "detect", "the", "occluded", "parts", "of", "pedestrians", ",", "which", "is", "useful", "to", "help", "extract", "effective", "features", "for", "detection", ".", "subsubsection", ":", "Evaluation", "Results", "We", "compare", "the", "proposed", "OR", "-", "CNN", "method", "with", "the", "state", "-", "of", "-", "the", "-", "art", "detectors", "on", "both", "the", "validation", "and", "testing", "sets", "of", "CityPersons", "in", "Table", "[", "reference", "]", "and", "Table", "[", "reference", "]", ",", "respectively", ".", "Our", "OR", "-", "CNN", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "results", "on", "the", "validation", "set", "of", "CityPersons", "by", "reducing", "(", "i.e.", ",", "vs.", "of", ")", "with", "scale", "and", "(", "i.e.", ",", "vs.", "of", ")", "with", "scale", ",", "surpassing", "all", "published", "approaches", ",", "which", "demonstrates", "the", "superiority", "of", "the", "proposed", "method", "in", "pedestrian", "detection", ".", "To", "demonstrate", "the", "effectiveness", "of", "OR", "-", "CNN", "under", "various", "occlusion", "levels", ",", "we", "follow", "the", "strategy", "in", "to", "divide", "the", "Reasonable", "subset", "in", "the", "validation", "set", "(", "occlusion", ")", "into", "the", "Reasonable", "-", "Partial", "subset", "(", "occlusion", ")", ",", "denoted", "as", "Partial", "subset", ",", "and", "the", "Reasonable", "-", "Bare", "subset", "(", "occlusion", ")", ",", "denoted", "as", "Bare", "subset", ".", "Meanwhile", ",", "we", "denote", "the", "annotated", "pedestrians", "with", "the", "occlusion", "ratio", "larger", "than", "(", "that", "are", "not", "included", "in", "the", "Reasonable", "set", ")", "as", "Heavy", "subset", ".", "We", "report", "the", "results", "of", "the", "proposed", "OR", "-", "CNN", "method", "and", "other", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "these", "three", "subsets", "in", "Table", "[", "reference", "]", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "OR", "-", "CNN", "outperforms", "the", "state", "-", "of", "-", "the", "-", "art", "methods", "consistently", "across", "all", "three", "subsets", ",", "i.e.", ",", "reduces", "on", "the", "Bare", "subset", ",", "on", "the", "Partial", "subset", ",", "and", "on", "the", "Heavy", "subset", ".", "Notably", ",", "when", "the", "occlusion", "becomes", "severely", "(", "i.e.", ",", "from", "Bare", "subset", "to", "Heavy", "subset", ")", ",", "the", "performance", "improvement", "of", "our", "OR", "-", "CNN", "is", "more", "obvious", "compared", "to", "the", "state", "-", "of", "-", "the", "-", "art", "methods", ",", "which", "demonstrates", "that", "the", "AggLoss", "and", "PORoI", "pooling", "unit", "are", "extremely", "effective", "to", "address", "the", "occlusion", "challenge", ".", "In", "addition", ",", "we", "also", "evaluate", "the", "proposed", "OR", "-", "CNN", "method", "on", "the", "testing", "set", "of", "CityPersons", ".", "Following", "its", "evaluation", "protocol", ",", "we", "submit", "the", "detection", "results", "of", "OR", "-", "CNN", "to", "the", "authors", "for", "evaluation", "and", "report", "the", "results", "in", "Table", "[", "reference", "]", ".", "The", "proposed", "OR", "-", "CNN", "method", "achieves", "the", "top", "accuracy", "with", "only", "scale", ".", "Although", "the", "second", "best", "detector", "Repulsion", "Loss", "uses", "much", "bigger", "input", "images", "(", "i.e.", ",", "scale", "of", "vs.", "scale", "of", "OR", "-", "CNN", ")", "and", "stronger", "backbone", "network", "(", "i.e.", ",", "ResNet", "-", "50", "of", "vs.", "VGG", "-", "16", "of", "OR", "-", "CNN", ")", ",", "it", "still", "produces", "higher", "on", "the", "Reasonable", "subset", "and", "higher", "on", "the", "Reasonable", "-", "Small", "subset", ".", "We", "believe", "the", "performance", "of", "OR", "-", "CNN", "can", "be", "further", "improved", "by", "using", "bigger", "input", "images", "and", "stronger", "backbone", "network", ".", "subsection", ":", "Caltech", "-", "USA", "Dataset", "The", "Caltech", "-", "USA", "dataset", "is", "one", "of", "the", "most", "popular", "and", "challenging", "datasets", "for", "pedestrian", "detection", ",", "which", "comes", "from", "approximately", "hours", "Hz", "VGA", "video", "recorded", "by", "a", "car", "traversing", "the", "streets", "in", "the", "greater", "Los", "Angeles", "metropolitan", "area", ".", "We", "use", "the", "new", "high", "quality", "annotations", "provided", "by", "to", "evaluate", "the", "proposed", "OR", "-", "CNN", "method", ".", "The", "training", "and", "testing", "sets", "contains", "and", "frames", ",", "respectively", ".", "Following", ",", "the", "log", "-", "average", "miss", "rate", "over", "points", "ranging", "from", "to", "FPPI", "is", "used", "to", "evaluate", "the", "performance", "of", "the", "detectors", ".", "We", "directly", "fine", "-", "tune", "the", "detection", "models", "pre", "-", "trained", "on", "CityPersons", "of", "the", "proposed", "OR", "-", "CNN", "method", "on", "the", "training", "set", "in", "Caltech", "-", "USA", ".", "Similar", "to", ",", "we", "evaluate", "the", "OR", "-", "CNN", "method", "on", "the", "Reasonable", "subset", "of", "the", "Caltech", "-", "USA", "dataset", ",", "and", "compare", "it", "to", "other", "state", "-", "of", "-", "the", "-", "art", "methods", "(", "e.g.", ",", ")", "in", "Figure", "[", "reference", "]", ".", "Notably", ",", "the", "Reasonable", "subset", "(", "occlusion", ")", "only", "includes", "the", "pedestrians", "with", "at", "least", "pixels", "tall", ",", "which", "is", "widely", "used", "to", "evaluate", "the", "pedestrian", "detectors", ".", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "the", "OR", "-", "CNN", "method", "performs", "competitively", "with", "the", "state", "-", "of", "-", "the", "-", "art", "method", "by", "producing", ".", "subsection", ":", "ETH", "Dataset", "To", "verify", "the", "generalization", "capacity", "of", "the", "proposed", "OR", "-", "CNN", "detector", ",", "we", "directly", "use", "the", "model", "trained", "on", "the", "CityPersons", "dataset", "to", "detect", "the", "pedestrians", "in", "the", "ETH", "dataset", "without", "fine", "-", "tuning", ".", "That", "is", ",", "all", "frames", "in", "three", "video", "clips", "of", "the", "ETH", "dataset", "are", "used", "to", "evaluate", "the", "performance", "of", "the", "OR", "-", "CNN", "detector", ".", "We", "use", "to", "evaluate", "the", "performance", "of", "the", "detectors", ",", "and", "compare", "the", "proposed", "OR", "-", "CNN", "method", "with", "other", "state", "-", "of", "-", "the", "-", "art", "methods", "(", "i.e.", ",", ")", "in", "Figure", "[", "reference", "]", ".", "Our", "OR", "-", "CNN", "detector", "achieves", "the", "top", "accuracy", "by", "reducing", "comparing", "to", "the", "state", "-", "of", "-", "the", "-", "art", "results", "(", "i.e.", ",", "of", "OR", "-", "CNN", "vs.", "RFN", "-", "BF", ")", ".", "The", "results", "on", "the", "ETH", "dataset", "not", "only", "demonstrates", "the", "superiority", "of", "the", "proposed", "OR", "-", "CNN", "method", "in", "pedestrian", "detection", ",", "but", "also", "verifies", "its", "generalization", "capacity", "to", "other", "scenarios", ".", "subsection", ":", "INRIA", "Dataset", "The", "INRIA", "dataset", "contains", "images", "of", "high", "resolution", "pedestrians", "collected", "mostly", "from", "holiday", "photos", ",", "which", "consists", "of", "images", ",", "including", "images", "for", "training", "and", "images", ".", "Specifically", ",", "there", "are", "positive", "images", "and", "negative", "images", "in", "the", "training", "set", ".", "We", "use", "the", "positive", "images", "in", "the", "training", "set", "to", "fine", "-", "tune", "our", "model", "pre", "-", "trained", "on", "CityPersons", "for", "iterations", ",", "and", "test", "it", "on", "the", "testing", "images", ".", "Figure", "[", "reference", "]", "shows", "that", "our", "OR", "-", "CNN", "method", "achieves", "an", "of", ",", "better", "than", "the", "other", "available", "competitors", "(", "i.e.", ",", ")", ",", "which", "demonstrates", "the", "effectiveness", "of", "the", "proposed", "method", "in", "pedestrian", "detection", ".", "section", ":", "Conclusions", "In", "this", "paper", ",", "we", "present", "a", "new", "occlusion", "-", "aware", "R", "-", "CNN", "method", "to", "improve", "the", "pedestrian", "detection", "accuracy", "in", "crowded", "scenes", ".", "Specifically", ",", "we", "design", "a", "new", "aggregation", "loss", "to", "reduce", "the", "false", "detections", "of", "the", "adjacent", "overlapping", "pedestrians", ",", "by", "simultaneously", "enforcing", "the", "proposals", "to", "be", "close", "to", "the", "associated", "objects", ",", "and", "locate", "compactly", ".", "Meanwhile", ",", "to", "effectively", "handle", "partial", "occlusion", ",", "we", "propose", "a", "new", "part", "occlusion", "-", "aware", "RoI", "pooling", "unit", "to", "replace", "the", "RoI", "pooling", "layer", "in", "the", "Fast", "R", "-", "CNN", "module", "of", "the", "detector", ",", "which", "integrates", "the", "prior", "structure", "information", "of", "human", "body", "with", "visibility", "prediction", "into", "the", "network", "to", "handle", "occlusion", ".", "Our", "method", "is", "trained", "in", "an", "end", "-", "to", "-", "end", "fashion", "and", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "accuracy", "on", "three", "pedestrian", "detection", "datasets", ",", "i.e.", ",", "CityPersons", ",", "ETH", ",", "and", "INRIA", ",", "and", "performs", "on", "-", "pair", "with", "the", "state", "-", "of", "-", "the", "-", "arts", "on", "Caltech", ".", "In", "the", "future", ",", "we", "plan", "to", "improve", "the", "method", "in", "two", "aspects", ".", "First", ",", "we", "would", "like", "to", "redesign", "the", "PORoI", "pooling", "unit", "to", "jointly", "estimate", "the", "location", ",", "size", ",", "and", "occlusion", "status", "of", "the", "object", "parts", "in", "the", "network", ",", "instead", "of", "using", "the", "empirical", "ratio", ".", "And", "then", ",", "we", "plan", "to", "extend", "the", "proposed", "method", "to", "detect", "other", "kinds", "of", "objects", ",", "e.g.", ",", "car", ",", "bicycle", ",", "tricycle", ",", "etc", ".", "section", ":", "Acknowledgments", "This", "work", "was", "supported", "by", "the", "National", "Key", "Research", "and", "Development", "Plan", "(", "Grant", "No.2016YFC0801002", ")", ",", "the", "Chinese", "National", "Natural", "Science", "Foundation", "Projects", ",", ",", ",", ",", "the", "Science", "and", "Technology", "Development", "Fund", "of", "Macau", "(", "No", ".", "0025", "/", "2018", "/", "A1", ",", "151", "/", "2017", "/", "A", ",", "152", "/", "2017", "/", "A", ")", ",", "JDGrapevine", "Plan", "and", "AuthenMetric", "R", "&", "D", "Funds", ".", "We", "also", "thank", "NVIDIA", "for", "GPU", "donations", "through", "their", "academic", "program", ".", "bibliography", ":", "References"], "entity": [[2, 8, ["Occlusion", "-", "aware", "R", "-", "CNN"], "Method"], [9, 11, ["Detecting", "Pedestrians"], "Task"], [14, 16, ["Pedestrian", "detection"], "Task"], [17, 19, ["crowded", "scenes"], "Task"], [42, 48, ["occlusion", "-", "aware", "R", "-", "CNN"], "Method"], [49, 52, ["OR", "-", "CNN"], "Method"], [56, 58, ["detection", "accuracy"], "Metric"], [90, 97, ["part", "occlusion", "-", "aware", "region", "of", "interest"], "Method"], [98, 99, ["PORoI"], "Method"], [105, 108, ["RoI", "pooling", "layer"], "Method"], [160, 161, ["CityPersons"], "Material"], [162, 163, ["ETH"], "Material"], [165, 166, ["INRIA"], "Material"], [182, 183, ["Caltech"], "Material"], [187, 189, ["Pedestrian", "detection"], "Task"], [195, 198, ["computer", "vision", "field"], "Task"], [204, 206, ["autonomous", "driving"], "Task"], [207, 209, ["video", "surveillance"], "Task"], [211, 212, ["robotics"], "Task"], [232, 234, ["object", "detection"], "Task"], [240, 244, ["deep", "convolutional", "neural", "networks"], "Method"], [245, 246, ["CNNs"], "Method"], [251, 255, ["bounding", "box", "regression", "techniques"], "Method"], [269, 273, ["real", "life", "complex", "scenarios"], "Task"], [274, 275, ["occlusion"], "Task"], [283, 285, ["detecting", "pedestrian"], "Task"], [309, 311, ["CityPersons", "dataset"], "Material"], [341, 342, ["detectors"], "Method"], [347, 351, ["non", "-", "maximum", "suppression"], "Task"], [352, 353, ["NMS"], "Task"], [376, 378, ["repulsion", "loss"], "Method"], [442, 448, ["occlusion", "-", "aware", "R", "-", "CNN"], "Method"], [449, 452, ["OR", "-", "CNN"], "Method"], [467, 469, ["occlusion", "challenge"], "Task"], [512, 514, ["loss", "function"], "Method"], [516, 518, ["aggregation", "loss"], "Method"], [519, 520, ["AggLoss"], "Method"], [563, 570, ["part", "occlusion", "-", "aware", "region", "of", "interest"], "Method"], [571, 572, ["PORoI"], "Method"], [579, 582, ["RoI", "pooling", "layer"], "Method"], [593, 594, ["detector"], "Method"], [665, 668, ["sub", "-", "network"], "Method"], [682, 684, ["pedestrian", "detection"], "Task"], [698, 699, ["CityPersons"], "Material"], [700, 701, ["Caltech"], "Material"], [702, 703, ["ETH"], "Material"], [704, 705, ["INRIA"], "Material"], [725, 729, ["OR", "-", "CNN", "method"], "Method"], [740, 741, ["MR"], "Method"], [743, 745, ["CityPersons", "dataset"], "Material"], [746, 747, ["MR"], "Method"], [749, 751, ["ETH", "dataset"], "Material"], [753, 754, ["MR"], "Method"], [756, 758, ["INRIA", "dataset"], "Material"], [787, 788, ["AggLoss"], "Method"], [825, 828, ["RoI", "pooling", "layer"], "Method"], [866, 867, ["CityPersons"], "Material"], [868, 869, ["Caltech"], "Material"], [870, 871, ["ETH"], "Material"], [873, 874, ["INRIA"], "Material"], [888, 891, ["Generic", "Object", "Detection"], "Task"], [893, 896, ["generic", "object", "detectors"], "Method"], [899, 902, ["sliding", "window", "paradigm"], "Method"], [910, 911, ["classifiers"], "Method"], [926, 930, ["deep", "convolutional", "neural", "network"], "Method"], [931, 932, ["CNN"], "Method"], [940, 943, ["object", "detection", "methods"], "Method"], [945, 946, ["CNN"], "Method"], [970, 974, ["one", "-", "stage", "approach"], "Method"], [978, 980, ["stage", "approach"], "Method"], [1015, 1017, ["stage", "approach"], "Method"], [1020, 1022, ["computational", "efficiency"], "Metric"], [1029, 1031, ["stage", "approach"], "Method"], [1040, 1041, ["accuracy"], "Metric"], [1055, 1058, ["separated", "proposal", "generator"], "Method"], [1061, 1063, ["Selective", "Search"], "Method"], [1064, 1065, ["EdgeBoxes"], "Method"], [1067, 1068, ["RPN"], "Method"], [1085, 1087, ["Pedestrian", "Detection"], "Task"], [1098, 1101, ["computer", "vision", "field"], "Task"], [1106, 1108, ["pedestrian", "detection"], "Task"], [1130, 1132, ["pedestrian", "detector"], "Method"], [1151, 1159, ["histograms", "of", "oriented", "gradient", "(", "HOG", ")", "descriptors"], "Method"], [1160, 1163, ["support", "vector", "machine"], "Method"], [1164, 1165, ["SVM"], "Method"], [1166, 1167, ["classifier"], "Method"], [1168, 1170, ["human", "detection"], "Task"], [1212, 1214, ["filter", "bank"], "Method"], [1221, 1223, ["detection", "quality"], "Metric"], [1245, 1247, ["spatial", "pooling"], "Method"], [1252, 1254, ["partial", "area"], "Metric"], [1256, 1258, ["ROC", "curve"], "Metric"], [1264, 1266, ["pedestrian", "detection"], "Task"], [1287, 1289, ["unsupervised", "method"], "Method"], [1291, 1294, ["convolutional", "sparse", "coding"], "Method"], [1298, 1299, ["CNN"], "Method"], [1300, 1302, ["pedestrian", "detection"], "Task"], [1306, 1311, ["complexity", "-", "aware", "cascaded", "detector"], "Method"], [1320, 1321, ["accuracy"], "Metric"], [1322, 1323, ["speed"], "Metric"], [1334, 1336, ["fast", "cascade"], "Method"], [1338, 1340, ["deep", "network"], "Method"], [1342, 1343, ["pedestrian"], "Task"], [1351, 1355, ["scale", "-", "dependent", "pooling"], "Method"], [1356, 1362, ["layer", "-", "wise", "cascaded", "rejection", "classifiers"], "Method"], [1378, 1380, ["pedestrian", "detection"], "Task"], [1382, 1383, ["RPN"], "Method"], [1385, 1387, ["boosted", "forests"], "Method"], [1391, 1393, ["pedestrian", "detection"], "Task"], [1401, 1403, ["network", "architecture"], "Method"], [1415, 1421, ["built", "-", "in", "sub", "-", "networks"], "Method"], [1440, 1443, ["segmentation", "infusion", "network"], "Method"], [1461, 1463, ["pedestrian", "detection"], "Task"], [1469, 1471, ["pedestrian", "localization"], "Task"], [1475, 1479, ["part", "-", "based", "model"], "Method"], [1482, 1486, ["pedestrian", "in", "occlusion", "handling"], "Task"], [1492, 1494, ["part", "detectors"], "Method"], [1501, 1503, ["part", "detection"], "Method"], [1512, 1516, ["part", "-", "based", "model"], "Method"], [1525, 1528, ["implicit", "shape", "model"], "Method"], [1563, 1565, ["occlusion", "reasoning"], "Method"], [1581, 1585, ["multi", "-", "pedestrian", "detectors"], "Method"], [1589, 1591, ["pedestrian", "detectors"], "Method"], [1626, 1632, ["mixture", "of", "occlusion", "-", "specific", "detectors"], "Method"], [1643, 1645, ["part", "detectors"], "Method"], [1654, 1656, ["computational", "cost"], "Metric"], [1666, 1670, ["bounding", "box", "regression", "loss"], "Method"], [1672, 1673, ["pedestrians"], "Task"], [1680, 1683, ["pedestrian", "detection", "methods"], "Method"], [1707, 1709, ["pedestrian", "detectors"], "Task"], [1712, 1718, ["Occlusion", "-", "aware", "R", "-", "CNN"], "Method"], [1736, 1738, ["pedestrian", "detection"], "Task"], [1743, 1745, ["aggregation", "loss"], "Metric"], [1779, 1782, ["region", "proposal", "network"], "Method"], [1783, 1784, ["RPN"], "Method"], [1838, 1841, ["accurate", "region", "proposals"], "Task"], [1850, 1852, ["AggLoss", "term"], "Method"], [1933, 1935, ["classification", "loss"], "Metric"], [1939, 1940, ["AggLoss"], "Method"], [1977, 1978, ["classification"], "Task"], [1981, 1983, ["Aggregation", "Loss"], "Task"], [2018, 2020, ["aggregation", "loss"], "Method"], [2021, 2023, ["AggLoss", ")"], "Method"], [2026, 2029, ["region", "proposal", "network"], "Method"], [2030, 2031, ["RPN"], "Method"], [2049, 2053, ["multi", "-", "task", "loss"], "Task"], [2148, 2149, ["accuracy"], "Metric"], [2164, 2165, ["regression"], "Task"], [2379, 2386, ["Part", "Occlusion", "-", "aware", "RoI", "Pooling", "Unit"], "Method"], [2387, 2391, ["real", "life", "complex", "scenarios"], "Task"], [2397, 2398, ["accuracy"], "Metric"], [2399, 2400, ["detectors"], "Task"], [2411, 2415, ["part", "-", "based", "model"], "Method"], [2419, 2421, ["occluded", "pedestrians"], "Task"], [2433, 2440, ["part", "occlusion", "-", "aware", "RoI", "pooling", "unit"], "Method"], [2450, 2452, ["visibility", "prediction"], "Method"], [2461, 2462, ["detector"], "Method"], [2466, 2469, ["micro", "neural", "network"], "Method"], [2509, 2512, ["RoI", "pooling", "layer"], "Method"], [2535, 2538, ["occlusion", "process", "unit"], "Method"], [2567, 2570, ["occlusion", "process", "unit"], "Method"], [2574, 2576, ["convolutional", "layers"], "Method"], [2579, 2581, ["softmax", "layer"], "Method"], [2603, 2606, ["predicted", "visibility", "score"], "Metric"], [2611, 2615, ["ground", "truth", "visibility", "score"], "Metric"], [2695, 2698, ["occlusion", "process", "unit"], "Method"], [2708, 2713, ["element", "-", "wise", "multiplication", "operator"], "Method"], [2737, 2742, ["element", "-", "wise", "summation", "operation"], "Method"], [2759, 2762, ["classification", "and", "regression"], "Task"], [2781, 2783, ["regression", "accuracy"], "Metric"], [2787, 2788, ["AggLoss"], "Method"], [2815, 2819, ["classification", "and", "aggregation", "losses"], "Metric"], [2848, 2849, ["CityPersons"], "Material"], [2854, 2855, ["ETH"], "Material"], [2857, 2858, ["INRIA"], "Material"], [2866, 2870, ["OR", "-", "CNN", "method"], "Method"], [2890, 2893, ["VGG", "-", "16"], "Method"], [2895, 2897, ["backbone", "network"], "Method"], [2903, 2908, ["ILSVRC", "CLS", "-", "LOC", "dataset"], "Material"], [2912, 2914, ["detection", "accuracy"], "Metric"], [2915, 2919, ["pedestrians", "with", "small", "scale"], "Task"], [2943, 2945, ["matching", "strategy"], "Method"], [2971, 2974, ["xavier", "”", "method"], "Method"], [2984, 2987, ["Stochastic", "Gradient", "Descent"], "Method"], [2988, 2989, ["SGD"], "Method"], [2992, 2996, ["momentum", "and", "weight", "decay"], "Method"], [3016, 3018, ["Citypersons", "dataset"], "Material"], [3022, 3024, ["learning", "rate"], "Metric"], [3054, 3056, ["learning", "rate"], "Metric"], [3083, 3085, ["CityPersons", "dataset"], "Material"], [3089, 3093, ["semantic", "segmentation", "dataset", "Cityscapes"], "Material"], [3101, 3103, ["pedestrian", "detection"], "Task"], [3129, 3130, ["validation"], "Task"], [3170, 3171, ["CityPersons"], "Material"], [3203, 3207, ["false", "positive", "per", "image"], "Metric"], [3208, 3209, ["FPPI"], "Metric"], [3219, 3220, ["detection"], "Task"], [3265, 3266, ["subsubsection"], "Method"], [3270, 3271, ["AggLoss"], "Task"], [3276, 3277, ["AggLoss"], "Method"], [3281, 3282, ["detector"], "Method"], [3293, 3294, ["AggLoss"], "Method"], [3313, 3314, ["CityPersons"], "Material"], [3370, 3371, ["detection"], "Task"], [3394, 3395, ["AggLoss"], "Method"], [3413, 3416, ["OR", "-", "CNN"], "Method"], [3418, 3419, ["detector"], "Method"], [3421, 3422, ["scale"], "Metric"], [3434, 3436, ["Repulsion", "Loss"], "Method"], [3442, 3443, ["AggLoss"], "Method"], [3447, 3449, ["Repulsion", "Loss"], "Method"], [3473, 3474, ["NMS"], "Method"], [3476, 3477, ["baseline"], "Method"], [3478, 3481, ["OR", "-", "CNN"], "Method"], [3507, 3510, ["OR", "-", "CNN"], "Method"], [3528, 3531, ["OR", "-", "CNN"], "Method"], [3551, 3552, ["AggLoss"], "Method"], [3667, 3669, ["miss", "rates"], "Metric"], [3698, 3701, ["OR", "-", "CNN"], "Method"], [3707, 3709, ["miss", "rate"], "Metric"], [3715, 3717, ["NMS", "operation"], "Method"], [3726, 3729, ["OR", "-", "CNN"], "Method"], [3743, 3746, ["OR", "-", "CNN"], "Method"], [3755, 3757, ["Ablation", "Study"], "Task"], [3773, 3774, ["detector"], "Method"], [3792, 3795, ["RoI", "pooling", "layer"], "Method"], [3806, 3807, ["CityPersons"], "Material"], [3848, 3849, ["CityPersons"], "Material"], [3899, 3901, ["pedestrian", "detection"], "Task"], [3946, 3949, ["human", "visual", "system"], "Method"], [3973, 3975, ["visibility", "score"], "Metric"], [4060, 4062, ["pedestrian", "detection"], "Task"], [4079, 4081, ["detection", "accuracy"], "Metric"], [4108, 4111, ["occlusion", "process", "unit"], "Method"], [4122, 4125, ["occlusion", "process", "unit"], "Method"], [4130, 4132, ["detection", "accuracy"], "Metric"], [4157, 4158, ["detection"], "Task"], [4167, 4171, ["OR", "-", "CNN", "method"], "Method"], [4189, 4190, ["CityPersons"], "Material"], [4204, 4207, ["OR", "-", "CNN"], "Method"], [4222, 4223, ["CityPersons"], "Material"], [4257, 4259, ["pedestrian", "detection"], "Task"], [4265, 4268, ["OR", "-", "CNN"], "Method"], [4354, 4358, ["OR", "-", "CNN", "method"], "Method"], [4386, 4389, ["OR", "-", "CNN"], "Method"], [4415, 4417, ["Partial", "subset"], "Material"], [4447, 4450, ["OR", "-", "CNN"], "Method"], [4480, 4482, ["occlusion", "challenge"], "Task"], [4491, 4495, ["OR", "-", "CNN", "method"], "Method"], [4500, 4501, ["CityPersons"], "Material"], [4510, 4511, ["detection"], "Task"], [4513, 4516, ["OR", "-", "CNN"], "Method"], [4533, 4537, ["OR", "-", "CNN", "method"], "Method"], [4540, 4541, ["accuracy"], "Metric"], [4549, 4552, ["detector", "Repulsion", "Loss"], "Metric"], [4565, 4568, ["OR", "-", "CNN"], "Method"], [4571, 4573, ["backbone", "network"], "Method"], [4576, 4579, ["ResNet", "-", "50"], "Method"], [4581, 4584, ["VGG", "-", "16"], "Method"], [4585, 4588, ["OR", "-", "CNN"], "Method"], [4612, 4615, ["OR", "-", "CNN"], "Method"], [4626, 4628, ["backbone", "network"], "Method"], [4650, 4652, ["pedestrian", "detection"], "Task"], [4659, 4661, ["VGA", "video"], "Material"], [4689, 4693, ["OR", "-", "CNN", "method"], "Method"], [4718, 4719, ["FPPI"], "Method"], [4735, 4737, ["detection", "models"], "Method"], [4741, 4742, ["CityPersons"], "Material"], [4745, 4749, ["OR", "-", "CNN", "method"], "Method"], [4764, 4768, ["OR", "-", "CNN", "method"], "Method"], [4827, 4829, ["pedestrian", "detectors"], "Method"], [4839, 4843, ["OR", "-", "CNN", "method"], "Method"], [4860, 4862, ["ETH", "Dataset"], "Material"], [4865, 4867, ["generalization", "capacity"], "Metric"], [4883, 4885, ["CityPersons", "dataset"], "Material"], [4891, 4893, ["ETH", "dataset"], "Material"], [4909, 4911, ["ETH", "dataset"], "Material"], [4938, 4942, ["OR", "-", "CNN", "method"], "Method"], [4970, 4971, ["accuracy"], "Metric"], [4988, 4991, ["OR", "-", "CNN"], "Method"], [4992, 4995, ["RFN", "-", "BF"], "Method"], [5001, 5003, ["ETH", "dataset"], "Material"], [5011, 5015, ["OR", "-", "CNN", "method"], "Method"], [5016, 5018, ["pedestrian", "detection"], "Task"], [5023, 5025, ["generalization", "capacity"], "Metric"], [5031, 5033, ["INRIA", "Dataset"], "Material"], [5034, 5036, ["INRIA", "dataset"], "Material"], [5093, 5094, ["CityPersons"], "Material"], [5112, 5116, ["OR", "-", "CNN", "method"], "Method"], [5140, 5142, ["pedestrian", "detection"], "Task"], [5164, 5167, ["pedestrian", "detection", "accuracy"], "Task"], [5168, 5170, ["crowded", "scenes"], "Task"], [5177, 5179, ["aggregation", "loss"], "Method"], [5219, 5226, ["part", "occlusion", "-", "aware", "RoI", "pooling", "unit"], "Method"], [5229, 5232, ["RoI", "pooling", "layer"], "Method"], [5241, 5242, ["detector"], "Method"], [5284, 5285, ["accuracy"], "Metric"], [5287, 5290, ["pedestrian", "detection", "datasets"], "Task"], [5293, 5294, ["CityPersons"], "Material"], [5295, 5296, ["ETH"], "Material"], [5298, 5299, ["INRIA"], "Material"], [5315, 5316, ["Caltech"], "Material"], [459, 460, ["CNN"], "Method"], [589, 590, ["CNN"], "Method"], [774, 780, ["occlusion", "-", "aware", "R", "-", "CNN"], "Method"], [819, 820, ["PORoI"], "Method"], [834, 835, ["CNN"], "Method"], [1006, 1007, ["CNNs"], "Method"], [1270, 1271, ["CNN"], "Method"], [1719, 1725, ["occlusion", "-", "aware", "R", "-", "CNN"], "Method"], [1732, 1733, ["CNN"], "Method"], [1754, 1755, ["PORoI"], "Method"], [1769, 1770, ["CNN"], "Method"], [1792, 1793, ["CNN"], "Method"], [1796, 1797, ["RPN"], "Method"], [1813, 1814, ["CNN"], "Method"], [1844, 1845, ["RPN"], "Method"], [2036, 2037, ["CNN"], "Method"], [2043, 2044, ["CNN"], "Method"], [2133, 2134, ["CNN"], "Method"], [2457, 2458, ["CNN"], "Method"], [2767, 2768, ["CNN"], "Method"], [2793, 2794, ["CNN"], "Method"], [2827, 2828, ["RPN"], "Method"], [2850, 2851, ["Caltech"], "Material"], [2876, 2879, ["OR", "-", "CNN"], "Method"], [2886, 2887, ["CNN"], "Method"], [2978, 2981, ["OR", "-", "CNN"], "Method"], [3040, 3041, ["Caltech"], "Material"], [3080, 3081, ["CityPersons"], "Material"], [3175, 3178, ["OR", "-", "CNN"], "Method"], [3198, 3200, ["miss", "rate"], "Metric"], [3236, 3237, ["CNN"], "Method"], [3243, 3244, ["baseline"], "Method"], [3285, 3288, ["OR", "-", "CNN"], "Method"], [3302, 3303, ["baseline"], "Method"], [3333, 3336, ["OR", "-", "CNN"], "Method"], [3340, 3343, ["OR", "-", "CNN"], "Method"], [3374, 3375, ["baseline"], "Method"], [3376, 3379, ["OR", "-", "CNN"], "Method"], [3519, 3520, ["baseline"], "Method"], [3544, 3545, ["baseline"], "Method"], [3649, 3652, ["OR", "-", "CNN"], "Method"], [3659, 3660, ["baseline"], "Method"], [3734, 3735, ["baseline"], "Method"], [3751, 3752, ["baseline"], "Method"], [3758, 3759, ["PORoI"], "Method"], [3766, 3767, ["PORoI"], "Method"], [3777, 3780, ["OR", "-", "CNN"], "Method"], [3786, 3787, ["PORoI"], "Method"], [3796, 3797, ["baseline"], "Method"], [3825, 3828, ["OR", "-", "CNN"], "Method"], [3832, 3835, ["OR", "-", "CNN"], "Method"], [3873, 3874, ["baseline"], "Method"], [3875, 3878, ["OR", "-", "CNN"], "Method"], [3895, 3896, ["PORoI"], "Method"], [3938, 3939, ["PORoI"], "Method"], [4040, 4041, ["PORoI"], "Method"], [4083, 4086, ["OR", "-", "CNN"], "Method"], [4137, 4138, ["PORoI"], "Method"], [4471, 4472, ["PORoI"], "Method"], [4631, 4632, ["Caltech"], "Material"], [4636, 4637, ["Caltech"], "Material"], [4711, 4713, ["miss", "rate"], "Metric"], [4754, 4755, ["Caltech"], "Material"], [4774, 4775, ["Caltech"], "Material"], [4870, 4873, ["OR", "-", "CNN"], "Method"], [4919, 4922, ["OR", "-", "CNN"], "Method"], [4963, 4966, ["OR", "-", "CNN"], "Method"], [5154, 5160, ["occlusion", "-", "aware", "R", "-", "CNN"], "Method"], [5237, 5238, ["CNN"], "Method"], [5339, 5340, ["PORoI"], "Method"]]} {"doc_id": "2af864bc38fd1116b4af32f16b44f5b3b156f5a3", "text": ["FINDING", "REMO", "(", "RELATED", "MEMORY", "OBJECT", ")", ":", "A", "SIMPLE", "NEURAL", "ARCHITECTURE", "FOR", "TEXT", "BASED", "REASONING", "section", ":", "ABSTRACT", "To", "solve", "the", "text", "-", "based", "question", "and", "answering", "task", "that", "requires", "relational", "reasoning", ",", "it", "is", "necessary", "to", "memorize", "a", "large", "amount", "of", "information", "and", "find", "out", "the", "question", "relevant", "information", "from", "the", "memory", ".", "Most", "approaches", "were", "based", "on", "external", "memory", "and", "four", "components", "proposed", "by", "Memory", "Network", ".", "The", "distinctive", "component", "among", "them", "was", "the", "way", "of", "finding", "the", "necessary", "information", "and", "it", "contributes", "to", "the", "performance", ".", "Recently", ",", "a", "simple", "but", "powerful", "neural", "network", "module", "for", "reasoning", "called", "Relation", "Network", "(", "RN", ")", "has", "been", "introduced", ".", "We", "analyzed", "RN", "from", "the", "view", "of", "Memory", "Network", ",", "and", "realized", "that", "its", "MLP", "component", "is", "able", "to", "reveal", "the", "complicate", "relation", "between", "question", "and", "object", "pair", ".", "Motivated", "from", "it", ",", "we", "introduce", "Relation", "Memory", "Network", "(", "RMN", ")", "which", "uses", "MLP", "to", "find", "out", "relevant", "information", "on", "Memory", "Network", "architecture", ".", "It", "shows", "new", "state", "-", "of", "-", "the", "-", "art", "results", "in", "jointly", "trained", "bAbI", "-", "10k", "story", "-", "based", "question", "answering", "tasks", "and", "bAbI", "dialog", "-", "based", "question", "answering", "tasks", ".", "section", ":", "INTRODUCTION", "Neural", "network", "has", "made", "an", "enormous", "progress", "on", "the", "two", "major", "challenges", "in", "artificial", "intelligence", ":", "seeing", "and", "reading", ".", "In", "both", "areas", ",", "embedding", "methods", "have", "served", "as", "the", "main", "vehicle", "to", "process", "and", "analyze", "text", "and", "image", "data", "for", "solving", "classification", "problems", ".", "As", "for", "the", "task", "of", "logical", "reasoning", ",", "however", ",", "more", "complex", "and", "careful", "handling", "of", "features", "is", "called", "for", ".", "A", "reasoning", "task", "requires", "the", "machine", "to", "answer", "a", "simple", "question", "upon", "the", "delivery", "of", "a", "series", "of", "sequential", "information", ".", "For", "example", ",", "imagine", "that", "the", "machine", "is", "given", "the", "following", "three", "sentences", ":", "\"", "Mary", "got", "the", "milk", "there", ".", "\"", ",", "\"", "John", "moved", "to", "the", "bedroom", ".", "\"", ",", "and", "\"", "Mary", "traveled", "to", "the", "hallway", ".", "\"", "Once", "prompted", "with", "the", "question", ",", "\"", "Where", "is", "the", "milk", "?", "\"", ",", "the", "machine", "then", "needs", "to", "sequentially", "focus", "on", "the", "two", "supporting", "sentences", ",", "\"", "Mary", "got", "the", "milk", "there", ".", "\"", "and", "\"", "Mary", "traveled", "to", "the", "hallway", ".", "\"", "in", "order", "to", "successfully", "determine", "that", "the", "milk", "is", "located", "in", "the", "hallway", ".", "Inspired", "by", "this", "reasoning", "mechanism", ",", "J.", "has", "introduced", "the", "memory", "network", "(", "MemNN", ")", ",", "which", "consists", "of", "an", "external", "memory", "and", "four", "components", ":", "input", "feature", "map", "(", "I", ")", ",", "generalization", "(", "G", ")", ",", "output", "feature", "map", "(", "O", ")", ",", "and", "response", "(", "R", ")", ".", "The", "external", "memory", "enables", "the", "model", "to", "deal", "with", "a", "knowledge", "base", "without", "loss", "of", "information", ".", "Input", "feature", "map", "embeds", "the", "incoming", "sentences", ".", "Generalization", "updates", "old", "memories", "given", "the", "new", "input", "and", "output", "feature", "map", "finds", "relevant", "information", "from", "the", "memory", ".", "Finally", ",", "response", "produces", "the", "final", "output", ".", "Based", "on", "the", "memory", "network", "architecture", ",", "neural", "network", "based", "models", "like", "end", "-", "to", "-", "end", "memory", "network", "(", "MemN2N", ")", "[", "reference", "]", ",", "gated", "end", "-", "to", "-", "end", "memory", "network", "(", "GMemN2N", ")", "[", "reference", "]", ",", "dynamic", "memory", "network", "(", "DMN", ")", "[", "reference", "]", ",", "and", "dynamic", "memory", "network", "+", "(", "DMN", "+", ")", "[", "reference", "]", "are", "proposed", ".", "Since", "strong", "reasoning", "ability", "depends", "on", "whether", "the", "model", "is", "able", "to", "sequentially", "catching", "the", "right", "supporting", "sentences", "that", "lead", "to", "the", "answer", ",", "the", "most", "important", "thing", "that", "discriminates", "those", "models", "is", "the", "way", "of", "constructing", "the", "output", "feature", "map", ".", "As", "the", "output", "feature", "map", "becomes", "more", "complex", ",", "it", "is", "able", "to", "learn", "patterns", "for", "more", "complicate", "relations", ".", "For", "example", ",", "MemN2N", ",", "which", "has", "the", "lowest", "performance", "among", "the", "four", "models", ",", "measures", "the", "relatedness", "between", "question", "and", "sentence", "by", "the", "inner", "product", ",", "while", "the", "best", "performing", "DMN", "+", "uses", "inner", "product", "and", "absolute", "difference", "with", "two", "embedding", "matrices", ".", "Recently", ",", "a", "new", "architecture", "called", "Relation", "Network", "(", "RN", ")", "[", "reference", "]", "has", "been", "proposed", "as", "a", "general", "solution", "to", "relational", "reasoning", ".", "The", "design", "philosophy", "behind", "it", "is", "to", "directly", "capture", "the", "supporting", "relation", "between", "the", "sentences", "through", "the", "multi", "-", "layer", "perceptron", "(", "MLP", ")", ".", "Despite", "its", "simplicity", ",", "RN", "achieves", "better", "performance", "than", "previous", "models", "without", "any", "catastrophic", "failure", ".", "The", "interesting", "thing", "we", "found", "is", "that", "RN", "can", "also", "be", "interpreted", "in", "terms", "of", "MemNN", ".", "It", "is", "composed", "of", "O", "and", "R", "where", "each", "corresponds", "to", "MLP", "which", "focuses", "on", "the", "related", "pair", "and", "another", "MLP", "which", "infers", "the", "answer", ".", "RN", "does", "not", "need", "to", "have", "G", "because", "it", "directly", "finds", "all", "the", "supporting", "sentences", "at", "once", ".", "In", "this", "point", "of", "view", ",", "the", "significant", "component", "would", "be", "MLP", "-", "based", "output", "feature", "map", ".", "As", "MLP", "is", "enough", "to", "recognize", "highly", "non", "-", "linear", "pattern", ",", "RN", "could", "find", "the", "proper", "relation", "better", "than", "previous", "models", "to", "answer", "the", "given", "question", ".", "However", ",", "as", "RN", "considers", "a", "pair", "at", "a", "time", "unlike", "MemNN", ",", "the", "number", "of", "relations", "that", "RN", "learns", "is", "n", "2", "when", "the", "number", "of", "input", "sentence", "is", "n.", "When", "n", "is", "small", ",", "the", "cost", "of", "learning", "relation", "is", "reduced", "by", "n", "times", "compared", "to", "MemNN", "based", "models", ",", "which", "enables", "more", "data", "-", "efficient", "learning", "[", "reference", "]", ".", "However", ",", "when", "n", "increases", ",", "the", "performance", "becomes", "worse", "than", "the", "previous", "models", ".", "In", "this", "case", ",", "the", "pair", "-", "wise", "operation", "increases", "the", "number", "of", "non", "-", "related", "sentence", "pairs", "more", "than", "the", "related", "sentence", "pair", ",", "thereby", "confuses", "RN", "'s", "learning", ".", "[", "reference", "]", "has", "suggested", "attention", "mechanisms", "as", "a", "solution", "to", "filter", "out", "unimportant", "relations", ";", "however", ",", "since", "it", "interrupts", "the", "reasoning", "operation", ",", "it", "may", "not", "be", "the", "most", "optimal", "solution", "to", "the", "problem", ".", "Our", "proposed", "model", ",", "\"", "Relation", "Memory", "Network", "\"", "(", "RMN", ")", ",", "is", "able", "to", "find", "complex", "relation", "even", "when", "a", "lot", "of", "information", "is", "given", ".", "It", "uses", "MLP", "to", "find", "out", "relevant", "information", "with", "a", "new", "generalization", "which", "simply", "erase", "the", "information", "already", "used", ".", "In", "other", "words", ",", "RMN", "inherits", "RN", "'s", "MLP", "-", "based", "output", "feature", "map", "on", "Memory", "Network", "architecture", ".", "Experiments", "show", "its", "state", "-", "ofthe", "-", "art", "result", "on", "the", "text", "-", "based", "question", "answering", "tasks", ".", "Relation", "Memory", "Network", "(", "RMN", ")", "is", "composed", "of", "four", "components", "-", "embedding", ",", "attention", ",", "updating", ",", "and", "reasoning", ".", "It", "takes", "as", "the", "inputs", "a", "set", "of", "sentences", "x", "1", ",", "x", "2", ",", "...", ",", "x", "n", "and", "its", "related", "question", "u", ",", "and", "outputs", "an", "answer", "a.", "Each", "of", "the", "x", "i", ",", "u", ",", "and", "a", "is", "made", "up", "of", "one", "-", "hot", "representation", "of", "words", ",", "for", "example", ",", "section", ":", "RELATION", "MEMORY", "NETWORK", "..", ",", "n", "i", ")", ",", "V", "=", "vocabulary", "size", ",", "n", "i", "=", "number", "of", "words", "in", "sentence", "i", ")", ".", "section", ":", "EMBEDDING", "COMPONENT", "We", "first", "embed", "words", "in", "each", "x", "i", "=", "{", "x", "i1", ",", "x", "i2", ",", "x", "i3", ",", "...", ",", "x", "ini", "}", "and", "u", "to", "a", "continuous", "space", "multiplying", "an", "embedding", "matrix", "A", "∈", "R", "d×V", ".", "Then", ",", "the", "embedded", "sentence", "is", "stored", "and", "represented", "as", "a", "memory", "object", "m", "i", "while", "question", "is", "represented", "as", "q.", "Any", "of", "the", "following", "methods", "are", "available", "for", "embedding", "component", ":", "simple", "sum", "(", "equation", "1", ")", ",", "position", "encoding", "(", "J.", "(", "equation", "2", ")", ",", "concatenation", "(", "equation", "3", ")", ",", "LSTM", ",", "and", "GRU", ".", "In", "case", "of", "LSTM", "or", "GRU", ",", "m", "i", "is", "the", "final", "hidden", "state", "of", "it", ".", "As", "the", "following", "attention", "component", "takes", "the", "concatenation", "of", "m", "i", "and", "q", ",", "it", "is", "not", "necessarily", "the", "case", "that", "sentence", "and", "question", "have", "the", "same", "dimensional", "embedding", "vectors", "unlike", "previous", "memory", "-", "augmented", "neural", "networks", ".", "section", ":", "ATTENTION", "COMPONENT", "Attention", "component", "can", "be", "applied", "more", "than", "once", "depending", "on", "the", "problem", ";", "Figure", "1", "illustrates", "2", "hop", "version", "of", "RMN", ".", "We", "refer", "to", "the", "i", "th", "embedded", "sentence", "on", "the", "t", "th", "hop", "as", "m", "section", ":", "UPDATING", "COMPONENT", "To", "forget", "the", "information", "already", "used", ",", "we", "use", "intuitive", "updating", "component", "to", "renew", "the", "memory", ".", "It", "is", "replaced", "by", "the", "amount", "of", "unconsumed", "from", "the", "old", "one", ":", "Contrary", "to", "other", "components", ",", "updating", "is", "not", "a", "mandatory", "component", ".", "When", "it", "is", "considered", "to", "have", "1", "hop", ",", "there", "is", "no", "need", "to", "use", "this", ".", "section", ":", "REASONING", "COMPONENT", "Similar", "to", "attention", "component", ",", "reasoning", "component", "is", "also", "made", "up", "of", "MLP", ",", "represented", "as", "f", "φ", ".", "It", "receives", "both", "q", "and", "the", "final", "result", "of", "attention", "component", "r", "f", "and", "then", "takes", "a", "softmax", "to", "produce", "the", "model", "answerâ", ":", "To", "answer", "the", "question", "from", "a", "given", "set", "of", "facts", ",", "the", "model", "needs", "to", "memorize", "these", "facts", "from", "the", "past", ".", "Long", "short", "term", "memory", "(", "LSTM", ")", "[", "reference", "]", ",", "one", "of", "the", "variants", "of", "recurrent", "neural", "network", "(", "RNN", ")", ",", "is", "inept", "at", "remembering", "past", "stories", "because", "of", "their", "small", "internal", "memory", "[", "reference", "]", ".", "To", "cope", "with", "this", "problem", ",", "J.", "has", "proposed", "a", "new", "class", "of", "memory", "-", "augmented", "model", "called", "Memory", "Network", "(", "MemNN", ")", ".", "MemNN", "comprises", "an", "external", "memory", "m", "and", "four", "components", ":", "input", "feature", "map", "(", "I", ")", ",", "generalization", "(", "G", ")", ",", "output", "feature", "map", "(", "O", ")", ",", "and", "response", "(", "R", ")", ".", "I", "encodes", "the", "sentences", "which", "are", "stored", "in", "memory", "m.", "G", "updates", "the", "memory", ",", "whereas", "O", "reads", "output", "feature", "o", "from", "the", "memory", ".", "Finally", ",", "R", "infers", "an", "answer", "from", "o.", "MemN2N", ",", "GMemN2N", ",", "DMN", ",", "and", "DMN", "+", "all", "follow", "the", "same", "structure", "of", "MemNN", "from", "a", "broad", "perspective", ",", "however", ",", "output", "feature", "map", "is", "composed", "in", "slightly", "different", "way", ".", "The", "relation", "between", "question", "and", "supporting", "sentences", "is", "realized", "from", "its", "cooperation", ".", "MemN2N", "first", "calculates", "the", "relatedness", "of", "sentences", "in", "the", "question", "and", "memory", "by", "taking", "the", "inner", "product", ",", "and", "the", "sentence", "with", "the", "highest", "relatedness", "is", "selected", "as", "the", "first", "supporting", "sentence", "for", "the", "given", "question", ".", "The", "first", "supporting", "sentence", "is", "then", "added", "with", "the", "question", "and", "repeat", "the", "same", "operation", "with", "the", "updated", "memory", "to", "find", "the", "second", "supporting", "sentence", ".", "GMemN2N", "selects", "the", "supporting", "sentence", "in", "the", "same", "way", "as", "MemN2N", ",", "but", "uses", "the", "gate", "to", "selectively", "add", "the", "the", "question", "to", "control", "the", "influence", "of", "the", "question", "information", "in", "finding", "the", "supporting", "sentence", "in", "the", "next", "step", ".", "DMN", "and", "DMN", "+", "use", "output", "feature", "map", "based", "on", "various", "relatedness", "such", "as", "absolute", "difference", ",", "as", "well", "as", "inner", "product", ",", "to", "understand", "the", "relation", "between", "sentence", "and", "question", "at", "various", "points", ".", "The", "more", "difficult", "the", "task", ",", "the", "more", "complex", "the", "output", "feature", "map", "and", "the", "generalization", "component", "to", "get", "the", "correct", "answer", ".", "For", "a", "dataset", "experimenting", "the", "text", "-", "based", "reasoning", "ability", "of", "the", "model", ",", "the", "overall", "accuracy", "could", "be", "increased", "in", "order", "of", "MemN2N", ",", "GMemN2N", ",", "DMN", ",", "and", "DMN", "+", ",", "where", "the", "complexity", "of", "the", "component", "increases", ".", "section", ":", "RELATION", "NETWORK", "Relation", "Network", "(", "RN", ")", "has", "emerged", "as", "a", "new", "and", "simpler", "framework", "for", "solving", "the", "general", "reasoning", "problem", ".", "RN", "takes", "in", "a", "pair", "of", "objects", "as", "its", "input", "and", "simply", "learns", "from", "the", "compositions", "of", "two", "MLPs", "represented", "as", "g", "θ", "and", "f", "φ", ".", "The", "role", "of", "each", "MLP", "is", "not", "clearly", "defined", "in", "the", "original", "paper", ",", "but", "from", "the", "view", "of", "MemNN", ",", "it", "can", "be", "understood", "that", "g", "θ", "corresponds", "to", "O", "and", "f", "φ", "corresponds", "to", "R.", "Table", "1", "summarizes", "the", "interpretation", "of", "RN", "compared", "to", "MemN2N", "and", "our", "model", ",", "RMN", ".", "To", "verify", "the", "role", "of", "g", "θ", ",", "we", "compare", "the", "output", "when", "pairs", "are", "made", "with", "supporting", "sentences", "and", "when", "made", "with", "unrelated", "sentences", ".", "Figure", "2", "shows", "the", "visualization", "result", "of", "each", "output", ".", "When", "we", "focus", "on", "whether", "the", "value", "is", "activated", "or", "not", ",", "we", "can", "see", "that", "g", "θ", "distinguishes", "supporting", "sentence", "pair", "from", "non", "-", "supporting", "sentence", "pair", "as", "output", "feature", "map", "examines", "how", "relevant", "the", "sentence", "is", "to", "the", "question", ".", "Therefore", ",", "we", "can", "comprehend", "the", "output", "of", "g", "θ", "reveals", "the", "relation", "between", "the", "object", "pair", "and", "the", "question", "and", "f", "φ", "aggregates", "all", "these", "outputs", "to", "infer", "the", "answer", ".", "bAbI", "story", "-", "based", "QA", "dataset", "bAbI", "story", "-", "based", "QA", "dataset", "is", "composed", "of", "20", "different", "types", "of", "tasks", "for", "testing", "natural", "language", "reasoning", "ability", ".", "Each", "task", "requires", "different", "methods", "to", "infer", "the", "answer", ".", "The", "dataset", "includes", "a", "set", "of", "statements", "comprised", "of", "multiple", "sentences", ",", "a", "question", "and", "answer", ".", "A", "statement", "can", "be", "as", "short", "as", "two", "sentences", "and", "as", "long", "as", "320", "sentences", ".", "To", "answer", "the", "question", ",", "it", "is", "necessary", "to", "find", "relevant", "one", "or", "more", "sentences", "to", "a", "given", "question", "and", "derive", "answer", "from", "them", ".", "Answer", "is", "typically", "a", "single", "word", "but", "in", "a", "few", "tasks", ",", "answers", "are", "a", "set", "of", "words", ".", "Each", "task", "is", "regarded", "as", "success", "when", "the", "accuracy", "is", "greater", "than", "95", "%", ".", "There", "are", "two", "versions", "of", "this", "dataset", ",", "one", "that", "has", "1k", "training", "examples", "and", "the", "other", "with", "10k", "examples", ".", "Most", "of", "the", "previous", "models", "test", "their", "accuracy", "on", "10k", "dataset", "with", "trained", "jointly", ".", "bAbI", "dialog", "dataset", "bAbI", "dialog", "dataset", "[", "reference", "]", ")", "is", "a", "set", "of", "5", "tasks", "within", "the", "goal", "-", "oriented", "context", "of", "restaurant", "reservation", ".", "It", "is", "designed", "to", "test", "if", "model", "can", "learn", "various", "abilities", "such", "as", "performing", "dialog", "management", ",", "querying", "knowledge", "bases", "(", "KBs", ")", ",", "and", "interpreting", "the", "output", "of", "such", "queries", ".", "The", "KB", "can", "be", "queried", "using", "API", "calls", "and", "4", "fields", "(", "a", "type", "of", "cuisine", ",", "a", "location", ",", "a", "price", "range", ",", "and", "a", "party", "size", ")", ".", "They", "should", "be", "filled", "to", "issue", "an", "API", "call", ".", "Task", "1", "tests", "the", "capacity", "of", "interpreting", "a", "request", "and", "asking", "the", "right", "questions", "to", "issue", "an", "API", "call", ".", "Task", "2", "checks", "the", "ability", "to", "modify", "an", "API", "call", ".", "Task", "3", "and", "4", "test", "the", "capacity", "of", "using", "outputs", "from", "an", "API", "call", "to", "propose", "options", "in", "the", "order", "of", "rating", "and", "to", "provide", "extra", "-", "information", "of", "what", "user", "asks", "for", ".", "Task", "5", "combines", "everything", ".", "The", "maximum", "length", "of", "the", "dialog", "for", "each", "task", "is", "different", ":", "14", "for", "task", "1", ",", "20", "for", "task", "2", ",", "78", "for", "task", "3", ",", "13", "for", "task", "4", ",", "and", "96", "for", "task", "5", ".", "As", "restaurant", "name", ",", "locations", ",", "and", "cuisine", "types", "always", "face", "new", "entities", ",", "there", "are", "normal", "and", "OOV", "test", "sets", "to", "assess", "model", "'s", "generalization", "ability", ".", "Training", "sets", "consist", "fo", "1k", "examples", ",", "which", "is", "not", "a", "large", "amount", "of", "creating", "realistic", "learning", "conditions", ".", "section", ":", "TRAINING", "DETAILS", "bAbI", "story", "-", "based", "QA", "dataset", "We", "trained", "2", "hop", "RMN", "jointly", "on", "all", "tasks", "using", "10k", "dataset", "for", "model", "to", "infer", "the", "solution", "suited", "to", "each", "type", "of", "tasks", ".", "We", "limited", "the", "input", "to", "the", "last", "70", "stories", "for", "all", "tasks", "except", "task", "3", "for", "which", "we", "limited", "input", "to", "the", "last", "130", "stories", ",", "similar", "to", "[", "reference", "]", "which", "is", "the", "hardest", "condition", "among", "previous", "models", ".", "Then", ",", "we", "labeled", "each", "sentence", "with", "its", "relative", "position", ".", "Embedding", "component", "is", "similar", "to", "[", "reference", "]", ",", "where", "story", "and", "question", "are", "embedded", "through", "different", "LSTMs", ";", "32", "unit", "word", "-", "lookup", "embeddings", ";", "32", "unit", "LSTM", "for", "story", "and", "question", ".", "For", "attention", "component", ",", "as", "we", "use", "2", "hop", "RMN", ",", "there", "are", "g", "1", "θ", "and", "g", "2", "θ", ";", "both", "are", "three", "-", "layer", "MLP", "consisting", "of", "256", ",", "128", ",", "1", "unit", "with", "ReLU", "activation", "function", "[", "reference", "]", ".", "f", "φ", "is", "composed", "of", "512", ",", "512", ",", "and", "159", "units", "(", "the", "number", "of", "words", "appearing", "in", "bAbI", "dataset", "is", "159", ")", "of", "three", "-", "layer", "MLP", "with", "ReLU", "non", "-", "linearities", "where", "the", "final", "layer", "was", "a", "linear", "that", "produced", "logits", "for", "a", "softmax", "over", "the", "answer", "vocabulary", ".", "For", "regularization", ",", "we", "use", "batch", "normalization", "[", "reference", "]", "for", "all", "MLPs", ".", "The", "softmax", "output", "was", "optimized", "with", "a", "cross", "-", "entropy", "loss", "function", "using", "the", "Adam", "optimizer", "[", "reference", "]", ")", "with", "a", "learning", "rate", "of", "2e", "−4", ".", "bAbI", "dialog", "dataset", "We", "trained", "on", "full", "dialog", "scripts", "with", "every", "model", "response", "as", "answer", ",", "all", "previous", "dialog", "history", "as", "sentences", "to", "be", "memorized", ",", "and", "the", "last", "user", "utterance", "as", "question", ".", "Model", "selects", "the", "most", "probable", "response", "from", "4", ",", "212", "candidates", "which", "are", "ranked", "from", "a", "set", "of", "all", "bot", "utterances", "appearing", "in", "training", ",", "validation", "and", "test", "sets", "(", "plain", "and", "OOV", ")", "for", "all", "tasks", "combined", ".", "We", "also", "report", "results", "when", "we", "use", "match", "type", "features", "for", "dialog", ".", "Match", "type", "feature", "is", "an", "additional", "label", "on", "the", "candidates", "indicating", "if", "word", "is", "found", "on", "the", "dialog", "history", ".", "For", "example", ",", "if", "the", "concentrate", "on", "the", "same", "sentences", "which", "are", "all", "critical", "to", "answer", "the", "question", ",", "and", "sometimes", "g", "1", "θ", "finds", "a", "fact", "related", "to", "the", "given", "question", "and", "with", "this", "information", "g", "2", "θ", "chooses", "the", "key", "fact", "to", "answer", ".", "While", "trained", "jointly", ",", "RMN", "learns", "these", "different", "solutions", "for", "each", "task", ".", "For", "the", "task", "3", ",", "the", "only", "failed", "task", ",", "attention", "component", "still", "functions", "well", ";", "it", "focuses", "sequentially", "on", "the", "supporting", "sentences", ".", "However", ",", "the", "reasoning", "component", ",", "f", "φ", ",", "had", "difficulty", "catching", "the", "word", "'", "before", "'", ".", "We", "could", "easily", "figure", "out", "'", "before", "'", "implies", "'", "just", "before", "'", "the", "certain", "situation", ",", "whereas", "RMN", "confused", "its", "meaning", ".", "As", "shown", "in", "table", "3c", ",", "our", "model", "found", "all", "previous", "locations", "before", "the", "garden", ".", "Still", ",", "it", "is", "remarkable", "that", "the", "simple", "MLP", "carried", "out", "all", "of", "these", "various", "roles", ".", "section", ":", "BABI", "DIALOG", "The", "results", "in", "the", "Table", "4", "show", "that", "the", "RMN", "has", "the", "best", "results", "in", "any", "conditions", ".", "Without", "any", "match", "type", ",", "RN", "and", "RMN", "outperform", "previous", "memory", "-", "augmented", "models", "on", "both", "normal", "and", "OOV", "tasks", ".", "This", "is", "mainly", "attributed", "to", "the", "impressive", "result", "on", "task", "4", "which", "can", "be", "interpreted", "as", "an", "effect", "of", "MLP", "based", "output", "feature", "map", ".", "To", "solve", "task", "4", ",", "it", "is", "critical", "to", "understand", "the", "relation", "between", "'", "phone", "number", "'", "of", "user", "input", "and", "'", "r", "phone", "'", "of", "previous", "dialog", "as", "shown", "in", "Table", "8c", ".", "We", "assumed", "that", "inner", "product", "was", "not", "sufficient", "to", "capture", "their", "implicit", "similarity", "and", "performed", "an", "supporting", "experiment", ".", "We", "converted", "RMN", "'s", "attention", "component", "to", "inner", "product", "based", "attention", ",", "and", "the", "results", "revealed", "the", "error", "rate", "increased", "to", "11.3", "%", ".", "For", "the", "task", "3", "and", "task", "5", "where", "the", "maximum", "length", "is", "especially", "longer", "than", "the", "others", ",", "RN", "performs", "worse", "than", "MemN2N", ",", "GMemN2N", "and", "RMN", ".", "The", "number", "of", "unnecessary", "object", "pairs", "created", "by", "the", "RN", "not", "only", "increases", "the", "processing", "time", "but", "also", "decreases", "the", "accuracy", ".", "With", "the", "match", "type", "feature", ",", "all", "models", "other", "than", "RMN", "have", "significantly", "improved", "their", "performance", "except", "for", "task", "3", "compared", "to", "the", "plain", "condition", ".", "RMN", "was", "helped", "by", "the", "match", "type", "only", "on", "the", "OOV", "tasks", "and", "this", "implies", "RMN", "is", "able", "to", "find", "relation", "in", "the", "With", "Match", "condition", "for", "the", "normal", "tasks", ".", "When", "we", "look", "at", "the", "OOV", "tasks", "more", "precisely", ",", "RMN", "failed", "to", "perform", "well", "on", "the", "OOV", "task", "1", "and", "2", "even", "though", "g", "1", "θ", "properly", "focused", "on", "the", "related", "object", "as", "shown", "in", "Table", "8a", ".", "We", "state", "that", "this", "originated", "from", "the", "fact", "that", "the", "number", "of", "keywords", "in", "task", "1", "and", "2", "is", "bigger", "than", "that", "in", "task", "4", ".", "In", "task", "1", "and", "2", ",", "all", "four", "keywords", "(", "cuisine", ",", "location", ",", "number", "and", "price", ")", "must", "be", "correctly", "aligned", "from", "the", "supporting", "sentence", "in", "order", "to", "make", "the", "correct", "API", "call", "which", "is", "harder", "than", "task", "4", ".", "Consider", "the", "example", "in", "Table", "8a", "and", "Table", "8c", ".", "Supporting", "sentence", "of", "task", "4", "have", "one", "keyword", "out", "of", "three", "words", ",", "whereas", "supporting", "sentences", "of", "task", "1", "and", "2", "consist", "of", "four", "keywords", "(", "cuisine", ",", "location", ",", "number", "and", "price", ")", "out", "of", "sixteen", "words", ".", "Different", "from", "other", "tasks", ",", "RMN", "yields", "the", "same", "error", "rate", "25.1", "%", "with", "MemN2N", "and", "GMemN2N", "on", "the", "task", "3", ".", "The", "main", "goal", "of", "task", "3", "is", "to", "recommend", "restaurant", "from", "knowledge", "base", "in", "the", "order", "of", "rating", ".", "All", "failed", "cases", "are", "displaying", "restaurant", "where", "the", "user", "input", "is", "<", "silence", ">", "which", "is", "somewhat", "an", "ambiguous", "trigger", "to", "find", "the", "input", "relevant", "previous", "utterance", ".", "As", "shown", "in", "Table", "8b", ",", "there", "are", "two", "different", "types", "of", "response", "to", "the", "same", "user", "input", ".", "One", "is", "to", "check", "whether", "all", "the", "required", "fields", "are", "given", "from", "the", "previous", "utterances", "and", "then", "ask", "user", "for", "the", "missing", "fields", "or", "send", "a", "\"", "Ok", "let", "me", "look", "into", "some", "options", "for", "you", ".", "\"", "message", ".", "The", "other", "type", "is", "to", "recommend", "restaurant", "starting", "from", "the", "highest", "rating", ".", "All", "models", "show", "lack", "of", "ability", "to", "discriminate", "these", "two", "types", "of", "silences", "so", "that", "concluded", "to", "the", "same", "results", ".", "To", "verify", "our", "statement", ",", "we", "performed", "an", "additional", "experiment", "on", "task", "3", "and", "checked", "the", "performance", "gain", "(", "extra", "result", "is", "given", "in", "Table", "10", "of", "Appendix", "B", ")", ".", "section", ":", "MODEL", "ANALYSIS", "Effectiveness", "of", "the", "MLP", "-", "based", "output", "feature", "map", "The", "most", "important", "feature", "that", "distinguishes", "MemNN", "based", "models", "is", "the", "output", "feature", "map", ".", "Table", "5", "summarizes", "the", "experimental", "results", "for", "the", "bAbI", "story", "-", "based", "QA", "dataset", "when", "replacing", "the", "RMN", "'s", "MLP", "-", "based", "output", "feature", "map", "with", "the", "idea", "of", "the", "previous", "models", ".", "inner", "product", "was", "used", "in", "MemN2N", ",", "inner", "product", "with", "gate", "was", "used", "in", "GMemN2N", ",", "and", "inner", "product", "and", "absolute", "difference", "with", "two", "embedding", "matrices", "was", "used", "in", "DMN", "and", "DMN", "+", ".", "From", "the", "Table", "5", ",", "the", "more", "complex", "the", "output", "feature", "map", ",", "the", "better", "the", "overall", "performance", ".", "In", "this", "point", "of", "view", ",", "MLP", "is", "the", "effective", "output", "feature", "map", ".", "Performance", "of", "RN", "and", "RMN", "according", "to", "memory", "size", "Additional", "experiments", "were", "conducted", "with", "the", "bAbI", "story", "-", "based", "QA", "dataset", "to", "see", "how", "memory", "size", "affects", "both", "performance", "and", "training", "time", "of", "RN", "and", "RMN", ".", "Test", "errors", "with", "training", "time", "written", "in", "parentheses", "are", "summarized", "in", "Table", "6", ".", "When", "memory", "size", "is", "small", ",", "we", "could", "observe", "the", "data", "-", "effeciency", "of", "RN", ".", "It", "shows", "similar", "performance", "to", "RMN", "in", "less", "time", ".", "However", ",", "when", "the", "memory", "size", "increases", ",", "performance", "is", "significantly", "reduced", "compared", "to", "RMN", ",", "even", "though", "it", "has", "been", "learned", "for", "a", "longer", "time", ".", "It", "is", "even", "lower", "than", "itself", "when", "the", "memory", "size", "is", "20", ".", "On", "the", "other", "hand", ",", "RMN", "maintains", "high", "performance", "even", "when", "the", "memory", "size", "increases", ".", "Effectiveness", "of", "the", "number", "of", "hops", "bAbI", "story", "based", "QA", "dataset", "differs", "in", "the", "number", "of", "supporting", "sentences", "by", "each", "task", "that", "need", "to", "be", "referenced", "to", "solve", "problems", ".", "For", "example", ",", "task", "1", ",", "2", ",", "and", "3", "require", "single", ",", "two", ",", "and", "three", "supporting", "facts", ",", "respectively", ".", "The", "result", "of", "the", "mean", "error", "rate", "for", "each", "task", "according", "to", "the", "number", "of", "hops", "is", "in", "Table", "7", ".", "Overall", ",", "the", "number", "of", "hops", "is", "correlated", "with", "the", "number", "of", "supporting", "sentences", ".", "In", "this", "respect", ",", "when", "the", "number", "of", "relations", "increases", ",", "RMN", "could", "reason", "across", "increasing", "the", "number", "of", "hops", "to", "3", ",", "4", "or", "more", ".", "section", ":", "CONCLUSION", "Our", "work", ",", "RMN", ",", "is", "a", "simple", "and", "powerful", "architecture", "that", "effectively", "handles", "text", "-", "based", "question", "answering", "tasks", "when", "large", "size", "of", "memory", "and", "high", "reasoning", "ability", "is", "required", ".", "Multiple", "access", "to", "the", "external", "memory", "to", "find", "out", "necessary", "information", "through", "a", "multi", "-", "hop", "approach", "is", "similar", "to", "most", "existing", "approaches", ".", "However", ",", "by", "using", "a", "MLP", "that", "can", "effectively", "deal", "with", "complex", "relatedness", "when", "searching", "for", "the", "right", "supporting", "sentences", "among", "a", "lot", "of", "sentences", ",", "RMN", "raised", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "the", "story", "-", "based", "QA", "and", "goal", "-", "oriented", "dialog", "dataset", ".", "When", "comparing", "RN", "which", "also", "used", "MLP", "to", "understand", "relations", ",", "RMN", "was", "more", "effective", "in", "the", "case", "of", "large", "memory", ".", "Future", "work", "will", "apply", "RMN", "to", "image", "based", "reasoning", "task", "(", "e.g.", ",", "CLEVR", ",", "DAQUAR", ",", "VQA", "etc", ".", ")", ".", "To", "extract", "features", "from", "the", "image", ",", "VGG", "net", "[", "reference", "]", ")", "is", "used", "in", "convention", "and", "outputs", "196", "objects", "of", "512", "dimensional", "vectors", "which", "also", "require", "large", "sized", "memory", ".", "An", "important", "direction", "will", "be", "to", "find", "an", "appropriate", "way", "to", "focus", "sequentially", "on", "related", "object", "which", "was", "rather", "easy", "in", "text", "-", "based", "reasoning", ".", "5", "Number", "in", "parentheses", "indicates", "the", "number", "of", "supporting", "sentences", "to", "solve", "the", "task", "A", "MODEL", "DETAILS", "We", "modify", "the", "user", "input", "from", "<", "silence", ">", "to", "<", "silence>", "when", "looking", "for", "restaurant", "recommendations", ".", "This", "makes", "model", "to", "distinguish", "two", "different", "situations", "whether", "to", "ask", "for", "additional", "fields", "or", "to", "recommend", "restaurant", ".", "section", ":"], "entity": [[1, 2, ["REMO"], "Method"], [3, 6, ["RELATED", "MEMORY", "OBJECT"], "Method"], [10, 12, ["NEURAL", "ARCHITECTURE"], "Method"], [13, 16, ["TEXT", "BASED", "REASONING"], "Task"], [22, 29, ["text", "-", "based", "question", "and", "answering", "task"], "Task"], [31, 33, ["relational", "reasoning"], "Task"], [67, 69, ["Memory", "Network"], "Method"], [96, 99, ["neural", "network", "module"], "Method"], [100, 101, ["reasoning"], "Task"], [102, 104, ["Relation", "Network"], "Method"], [105, 106, ["RN"], "Method"], [113, 114, ["RN"], "Method"], [118, 120, ["Memory", "Network"], "Method"], [146, 149, ["Relation", "Memory", "Network"], "Method"], [150, 151, ["RMN"], "Method"], [154, 155, ["MLP"], "Method"], [161, 164, ["Memory", "Network", "architecture"], "Method"], [179, 188, ["bAbI", "-", "10k", "story", "-", "based", "question", "answering", "tasks"], "Task"], [189, 196, ["bAbI", "dialog", "-", "based", "question", "answering", "tasks"], "Task"], [200, 202, ["Neural", "network"], "Method"], [213, 215, ["artificial", "intelligence"], "Task"], [216, 219, ["seeing", "and", "reading"], "Task"], [224, 226, ["embedding", "methods"], "Method"], [242, 244, ["classification", "problems"], "Task"], [250, 252, ["logical", "reasoning"], "Task"], [267, 269, ["reasoning", "task"], "Task"], [389, 391, ["reasoning", "mechanism"], "Method"], [396, 398, ["memory", "network"], "Method"], [399, 401, ["MemNN", ")"], "Method"], [455, 457, ["feature", "map"], "Method"], [462, 463, ["Generalization"], "Method"], [492, 495, ["memory", "network", "architecture"], "Method"], [496, 500, ["neural", "network", "based", "models"], "Method"], [501, 508, ["end", "-", "to", "-", "end", "memory", "network"], "Method"], [509, 510, ["MemN2N"], "Method"], [515, 523, ["gated", "end", "-", "to", "-", "end", "memory", "network"], "Method"], [524, 525, ["GMemN2N"], "Method"], [530, 533, ["dynamic", "memory", "network"], "Method"], [534, 535, ["DMN"], "Method"], [541, 548, ["dynamic", "memory", "network", "+", "(", "DMN", "+"], "Method"], [620, 621, ["MemN2N"], "Method"], [641, 643, ["inner", "product"], "Method"], [648, 650, ["DMN", "+"], "Method"], [651, 653, ["inner", "product"], "Method"], [658, 660, ["embedding", "matrices"], "Method"], [667, 669, ["Relation", "Network"], "Method"], [670, 671, ["RN"], "Method"], [683, 685, ["relational", "reasoning"], "Task"], [703, 707, ["multi", "-", "layer", "perceptron"], "Method"], [708, 709, ["MLP"], "Method"], [715, 716, ["RN"], "Method"], [734, 735, ["RN"], "Method"], [742, 743, ["MemNN"], "Method"], [755, 756, ["MLP"], "Method"], [764, 765, ["MLP"], "Method"], [770, 771, ["RN"], "Method"], [807, 808, ["MLP"], "Method"], [818, 819, ["RN"], "Method"], [837, 838, ["RN"], "Method"], [845, 846, ["MemNN"], "Method"], [852, 853, ["RN"], "Method"], [873, 875, ["learning", "relation"], "Task"], [882, 885, ["MemNN", "based", "models"], "Method"], [892, 893, ["learning"], "Task"], [917, 921, ["pair", "-", "wise", "operation"], "Method"], [948, 950, ["attention", "mechanisms"], "Method"], [985, 988, ["Relation", "Memory", "Network"], "Method"], [1010, 1011, ["MLP"], "Method"], [1019, 1020, ["generalization"], "Method"], [1032, 1033, ["RMN"], "Method"], [1043, 1046, ["Memory", "Network", "architecture"], "Method"], [1058, 1064, ["text", "-", "based", "question", "answering", "tasks"], "Task"], [1065, 1068, ["Relation", "Memory", "Network"], "Method"], [1069, 1070, ["RMN"], "Method"], [1081, 1082, ["updating"], "Task"], [1084, 1085, ["reasoning"], "Task"], [1132, 1134, ["hot", "representation"], "Method"], [1142, 1145, ["RELATION", "MEMORY", "NETWORK"], "Method"], [1169, 1171, ["EMBEDDING", "COMPONENT"], "Method"], [1203, 1205, ["embedding", "matrix"], "Method"], [1239, 1241, ["embedding", "component"], "Method"], [1242, 1244, ["simple", "sum"], "Method"], [1249, 1251, ["position", "encoding"], "Method"], [1258, 1259, ["concatenation"], "Method"], [1264, 1265, ["LSTM"], "Method"], [1267, 1268, ["GRU"], "Method"], [1272, 1273, ["LSTM"], "Method"], [1274, 1275, ["GRU"], "Method"], [1289, 1291, ["attention", "component"], "Method"], [1318, 1323, ["memory", "-", "augmented", "neural", "networks"], "Method"], [1326, 1328, ["ATTENTION", "COMPONENT"], "Method"], [1328, 1330, ["Attention", "component"], "Method"], [1345, 1347, ["hop", "version"], "Method"], [1348, 1349, ["RMN"], "Method"], [1367, 1369, ["UPDATING", "COMPONENT"], "Method"], [1378, 1381, ["intuitive", "updating", "component"], "Method"], [1404, 1405, ["updating"], "Task"], [1430, 1432, ["REASONING", "COMPONENT"], "Method"], [1434, 1436, ["attention", "component"], "Method"], [1437, 1439, ["reasoning", "component"], "Method"], [1444, 1445, ["MLP"], "Method"], [1448, 1450, ["f", "φ"], "Method"], [1468, 1469, ["softmax"], "Method"], [1497, 1501, ["Long", "short", "term", "memory"], "Method"], [1502, 1503, ["LSTM"], "Method"], [1513, 1519, ["recurrent", "neural", "network", "(", "RNN", ")"], "Method"], [1549, 1553, ["memory", "-", "augmented", "model"], "Method"], [1554, 1556, ["Memory", "Network"], "Method"], [1557, 1558, ["MemNN"], "Method"], [1560, 1561, ["MemNN"], "Method"], [1577, 1578, ["generalization"], "Method"], [1628, 1629, ["MemN2N"], "Method"], [1630, 1631, ["GMemN2N"], "Method"], [1632, 1633, ["DMN"], "Method"], [1635, 1637, ["DMN", "+"], "Method"], [1643, 1644, ["MemNN"], "Method"], [1674, 1675, ["MemN2N"], "Method"], [1737, 1738, ["GMemN2N"], "Method"], [1747, 1748, ["MemN2N"], "Method"], [1777, 1781, ["DMN", "and", "DMN", "+"], "Method"], [1782, 1785, ["output", "feature", "map"], "Method"], [1827, 1829, ["generalization", "component"], "Method"], [1840, 1845, ["text", "-", "based", "reasoning", "ability"], "Task"], [1851, 1852, ["accuracy"], "Metric"], [1858, 1859, ["MemN2N"], "Method"], [1860, 1861, ["GMemN2N"], "Method"], [1862, 1863, ["DMN"], "Method"], [1865, 1867, ["DMN", "+"], "Method"], [1870, 1871, ["complexity"], "Metric"], [1878, 1880, ["RELATION", "NETWORK"], "Method"], [1896, 1899, ["general", "reasoning", "problem"], "Task"], [1900, 1901, ["RN"], "Method"], [1918, 1919, ["MLPs"], "Method"], [1931, 1932, ["MLP"], "Method"], [1946, 1947, ["MemNN"], "Method"], [1970, 1971, ["RN"], "Method"], [1973, 1974, ["MemN2N"], "Method"], [1978, 1979, ["RMN"], "Method"], [2090, 2096, ["bAbI", "story", "-", "based", "QA", "dataset"], "Material"], [2096, 2102, ["bAbI", "story", "-", "based", "QA", "dataset"], "Material"], [2112, 2116, ["natural", "language", "reasoning", "ability"], "Task"], [2212, 2213, ["accuracy"], "Metric"], [2247, 2248, ["accuracy"], "Metric"], [2255, 2258, ["bAbI", "dialog", "dataset"], "Material"], [2258, 2261, ["bAbI", "dialog", "dataset"], "Material"], [2278, 2280, ["restaurant", "reservation"], "Task"], [2295, 2297, ["dialog", "management"], "Task"], [2396, 2398, ["API", "call"], "Method"], [2486, 2488, ["generalization", "ability"], "Metric"], [2512, 2518, ["bAbI", "story", "-", "based", "QA", "dataset"], "Material"], [2522, 2523, ["RMN"], "Method"], [2594, 2596, ["Embedding", "component"], "Method"], [2611, 2612, ["LSTMs"], "Method"], [2614, 2619, ["unit", "word", "-", "lookup", "embeddings"], "Method"], [2624, 2627, ["story", "and", "question"], "Task"], [2629, 2631, ["attention", "component"], "Task"], [2690, 2692, ["bAbI", "dataset"], "Material"], [2701, 2705, ["ReLU", "non", "-", "linearities"], "Method"], [2724, 2725, ["regularization"], "Task"], [2728, 2730, ["batch", "normalization"], "Method"], [2735, 2736, ["MLPs"], "Method"], [2744, 2749, ["cross", "-", "entropy", "loss", "function"], "Metric"], [2751, 2753, ["Adam", "optimizer"], "Method"], [2759, 2761, ["learning", "rate"], "Metric"], [2765, 2768, ["bAbI", "dialog", "dataset"], "Material"], [2849, 2850, ["dialog"], "Task"], [2921, 2922, ["RMN"], "Method"], [2940, 2942, ["attention", "component"], "Method"], [2957, 2959, ["reasoning", "component"], "Method"], [2990, 2991, ["RMN"], "Method"], [3019, 3020, ["MLP"], "Method"], [3030, 3032, ["BABI", "DIALOG"], "Task"], [3041, 3042, ["RMN"], "Method"], [3055, 3056, ["RN"], "Method"], [3057, 3058, ["RMN"], "Method"], [3060, 3064, ["memory", "-", "augmented", "models"], "Method"], [3066, 3067, ["normal"], "Task"], [3068, 3070, ["OOV", "tasks"], "Task"], [3151, 3155, ["RMN", "'s", "attention", "component"], "Method"], [3156, 3160, ["inner", "product", "based", "attention"], "Method"], [3166, 3168, ["error", "rate"], "Metric"], [3191, 3192, ["RN"], "Method"], [3195, 3196, ["MemN2N"], "Method"], [3197, 3198, ["GMemN2N"], "Method"], [3199, 3200, ["RMN"], "Method"], [3210, 3211, ["RN"], "Method"], [3215, 3217, ["processing", "time"], "Metric"], [3221, 3222, ["accuracy"], "Metric"], [3233, 3234, ["RMN"], "Method"], [3249, 3250, ["RMN"], "Method"], [3259, 3261, ["OOV", "tasks"], "Task"], [3264, 3265, ["RMN"], "Method"], [3285, 3287, ["OOV", "tasks"], "Task"], [3290, 3291, ["RMN"], "Method"], [3297, 3299, ["OOV", "task"], "Task"], [3440, 3441, ["RMN"], "Method"], [3444, 3446, ["error", "rate"], "Metric"], [3449, 3450, ["MemN2N"], "Method"], [3451, 3452, ["GMemN2N"], "Method"], [3468, 3470, ["knowledge", "base"], "Material"], [3630, 3632, ["MODEL", "ANALYSIS"], "Method"], [3647, 3650, ["MemNN", "based", "models"], "Method"], [3664, 3670, ["bAbI", "story", "-", "based", "QA", "dataset"], "Material"], [3689, 3691, ["inner", "product"], "Method"], [3694, 3695, ["MemN2N"], "Method"], [3696, 3698, ["inner", "product"], "Method"], [3699, 3700, ["gate"], "Method"], [3703, 3704, ["GMemN2N"], "Method"], [3718, 3719, ["DMN"], "Method"], [3720, 3722, ["DMN", "+"], "Method"], [3748, 3749, ["MLP"], "Method"], [3758, 3759, ["RN"], "Method"], [3760, 3761, ["RMN"], "Method"], [3763, 3765, ["memory", "size"], "Metric"], [3771, 3777, ["bAbI", "story", "-", "based", "QA", "dataset"], "Material"], [3786, 3788, ["training", "time"], "Metric"], [3789, 3790, ["RN"], "Method"], [3791, 3792, ["RMN"], "Method"], [3793, 3795, ["Test", "errors"], "Metric"], [3796, 3798, ["training", "time"], "Metric"], [3828, 3829, ["RMN"], "Method"], [3847, 3848, ["RMN"], "Method"], [3878, 3879, ["RMN"], "Method"], [3894, 3900, ["hops", "bAbI", "story", "based", "QA", "dataset"], "Material"], [3945, 3948, ["mean", "error", "rate"], "Metric"], [3988, 3989, ["RMN"], "Method"], [4010, 4011, ["RMN"], "Method"], [4021, 4027, ["text", "-", "based", "question", "answering", "tasks"], "Task"], [4034, 4036, ["reasoning", "ability"], "Metric"], [4052, 4056, ["multi", "-", "hop", "approach"], "Method"], [4068, 4069, ["MLP"], "Method"], [4089, 4090, ["RMN"], "Method"], [4102, 4106, ["story", "-", "based", "QA"], "Task"], [4107, 4112, ["goal", "-", "oriented", "dialog", "dataset"], "Task"], [4115, 4116, ["RN"], "Method"], [4119, 4120, ["MLP"], "Method"], [4124, 4125, ["RMN"], "Method"], [4139, 4140, ["RMN"], "Method"], [4141, 4145, ["image", "based", "reasoning", "task"], "Task"], [4148, 4149, ["CLEVR"], "Task"], [4150, 4151, ["DAQUAR"], "Task"], [4152, 4153, ["VQA"], "Task"], [4164, 4166, ["VGG", "net"], "Method"], [4210, 4214, ["text", "-", "based", "reasoning"], "Task"], [4248, 4250, ["restaurant", "recommendations"], "Task"], [125, 126, ["MLP"], "Method"], [799, 800, ["MLP"], "Method"], [939, 940, ["RN"], "Method"], [990, 991, ["RMN"], "Method"], [1034, 1035, ["RN"], "Method"], [1036, 1037, ["MLP"], "Method"], [1880, 1882, ["Relation", "Network"], "Method"], [1883, 1884, ["RN"], "Method"], [2477, 2478, ["normal"], "Task"], [2622, 2623, ["LSTM"], "Method"], [2637, 2638, ["RMN"], "Method"], [2654, 2655, ["MLP"], "Method"], [2699, 2700, ["MLP"], "Method"], [3090, 3091, ["MLP"], "Method"], [3277, 3278, ["normal"], "Task"], [3635, 3636, ["MLP"], "Method"], [3673, 3674, ["RMN"], "Method"], [3675, 3676, ["MLP"], "Method"], [3821, 3822, ["RN"], "Method"]]} {"doc_id": "2b507f659b341ed0f23106446de8e4322f4a3f7e", "text": ["document", ":", "Deep", "Identity", "-", "Aware", "Transfer", "of", "Facial", "Attributes", "This", "paper", "presents", "a", "Deep", "convolutional", "network", "model", "for", "Identity", "-", "Aware", "Transfer", "(", "DIAT", ")", "of", "facial", "attributes", ".", "Given", "the", "source", "input", "image", "and", "the", "reference", "attribute", ",", "DIAT", "aims", "to", "generate", "a", "facial", "image", "that", "owns", "the", "reference", "attribute", "as", "well", "as", "keeps", "the", "same", "or", "similar", "identity", "to", "the", "input", "image", ".", "In", "general", ",", "our", "model", "consists", "of", "a", "mask", "network", "and", "an", "attribute", "transform", "network", "which", "work", "in", "synergy", "to", "generate", "photo", "-", "realistic", "facial", "image", "with", "the", "reference", "attribute", ".", "Considering", "that", "the", "reference", "attribute", "may", "be", "only", "related", "to", "some", "parts", "of", "the", "image", ",", "the", "mask", "network", "is", "introduced", "to", "avoid", "the", "incorrect", "editing", "on", "attribute", "irrelevant", "region", ".", "Then", "the", "estimated", "mask", "is", "adopted", "to", "combine", "the", "input", "and", "transformed", "image", "for", "producing", "the", "transfer", "result", ".", "For", "joint", "training", "of", "transform", "network", "and", "mask", "network", ",", "we", "incorporate", "the", "adversarial", "attribute", "loss", ",", "identity", "aware", "adaptive", "perceptual", "loss", "and", "VGG", "-", "FACE", "based", "identity", "loss", ".", "Furthermore", ",", "a", "denoising", "network", "is", "presented", "to", "serve", "for", "perceptual", "regularization", "to", "suppress", "the", "artifacts", "in", "transfer", "result", ",", "while", "an", "attribute", "ratio", "regularization", "is", "introduced", "to", "constrain", "the", "size", "of", "attribute", "relevant", "region", ".", "Our", "DIAT", "can", "provide", "a", "unified", "solution", "for", "several", "representative", "facial", "attribute", "transfer", "tasks", ",", "e", ".", "g", ".", ",", "expression", "transfer", ",", "accessory", "removal", ",", "age", "progression", "and", "gender", "transfer", ",", "and", "can", "be", "extended", "for", "other", "face", "enhancement", "tasks", "such", "as", "face", "hallucination", ".", "The", "experimental", "results", "validate", "the", "effectiveness", "of", "the", "proposed", "method", ".", "Even", "for", "the", "identity", "-", "related", "attribute", "(", "e", ".", "g", ".", ",", "gender", ")", ",", "our", "DIAT", "can", "obtain", "visually", "impressive", "results", "by", "changing", "the", "attribute", "while", "retaining", "most", "identity", "-", "aware", "features", ".", "Facial", "attribute", "transfer", ",", "generative", "adversarial", "nets", ",", "convolutional", "networks", ",", "perceptual", "loss", ".", "section", ":", "Introduction", "Face", "attributes", ",", "e", ".", "g", ".", ",", "gender", "and", "expression", ",", "can", "not", "only", "provide", "a", "natural", "description", "of", "facial", "images", ",", "but", "also", "offer", "a", "unified", "viewpoint", "for", "understanding", "many", "facial", "animation", "and", "manipulation", "tasks", ".", "For", "example", ",", "the", "goal", "of", "facial", "avatar", "and", "reenactment", "is", "to", "transfer", "the", "facial", "expression", "attributes", "of", "a", "source", "actor", "to", "a", "target", "actor", ".", "In", "most", "applications", "such", "as", "expression", "transfer", ",", "accessory", "removal", "and", "age", "progression", ",", "the", "animation", "only", "modifies", "the", "related", "attribute", "without", "changing", "the", "identity", ".", "But", "for", "some", "other", "tasks", ",", "the", "change", "of", "some", "attributes", ",", "e", ".", "g", ".", ",", "gender", "and", "ethnicity", ",", "will", "inevitably", "alter", "the", "identity", "of", "the", "source", "image", ".", "In", "recent", "years", ",", "a", "variety", "of", "methods", "have", "been", "developed", "for", "specific", "facial", "attribute", "transfer", "tasks", ",", "and", "have", "achieved", "impressive", "results", ".", "For", "expression", "transfer", ",", "approaches", "have", "been", "suggested", "to", "create", "3D", "or", "image", "-", "based", "avatars", "from", "hand", "-", "held", "video", ",", "while", "face", "trackers", "and", "expression", "modeling", "have", "been", "investigated", "for", "offline", "and", "online", "facial", "reenactment", ".", "For", "age", "progression", ",", "explicit", "and", "implicit", "synthesis", "methods", "have", "been", "proposed", "for", "different", "image", "models", ".", "Hair", "style", "generation", "and", "replacement", "have", "also", "been", "studied", "in", "literatures", ".", "Convolutional", "neural", "network", "(", "CNN", ")-", "based", "models", "have", "also", "been", "investigated", "for", "human", "face", "generation", "with", "attributes", ".", "Kulkarni", "et", "al", ".", "propose", "deep", "convolution", "inverse", "graphic", "network", "(", "DG", "-", "IGN", ")", ".", "This", "method", "requires", "a", "large", "number", "of", "faces", "of", "a", "single", "person", "for", "training", ",", "and", "can", "only", "generate", "faces", "with", "different", "pose", "and", "light", ".", "Gauthier", "developes", "a", "conditional", "generative", "adversarial", "network", "(", "cGAN", ")", "to", "generate", "facial", "image", "from", "a", "noise", "distribution", "and", "conditional", "attributes", ".", "Yan", "et", "al", ".", "suggest", "an", "attribute", "-", "conditioned", "deep", "variational", "auto", "-", "encoder", "which", "extracts", "the", "latent", "variables", "from", "a", "reference", "image", "and", "combines", "them", "with", "attributes", "to", "produce", "the", "generated", "image", "with", "a", "generative", "model", ".", "Oord", "et", "al", ".", "propose", "a", "conditional", "image", "generation", "model", "based", "on", "PixelCNN", "decoder", "for", "image", "generation", "conditioned", "on", "an", "arbitrary", "feature", "vector", ".", "However", ",", "the", "identity", "of", "the", "generated", "face", "is", "not", "emphasized", "in", ",", "making", "them", "not", "directly", "applicable", "to", "attribute", "transfer", ".", "Motivated", "by", "the", "strong", "capability", "of", "CNN", "in", "modeling", "complex", "transformation", "and", "capturing", "perceptual", "similarity", ",", "several", "approaches", "have", "also", "been", "suggested", "for", "facial", "attribute", "transfer", ".", "Li", "et", "al", ".", "suggest", "a", "CNN", "-", "based", "attribute", "transfer", "model", "from", "the", "optimization", "perspective", ",", "but", "both", "run", "time", "and", "transfer", "quality", "is", "far", "from", "satisfying", ".", "Considering", "that", "it", "is", "impracticable", "to", "collect", "labeled", "data", "for", "supervised", "learning", ",", "the", "generative", "adversarial", "net", "(", "GAN", ")", "framework", "usually", "is", "adopted", "for", "handling", "this", "task", ".", "However", ",", "visible", "artifacts", "and", "over", "-", "smoothing", "usually", "are", "inevitable", "in", "the", "transfer", "result", "for", "these", "methods", ".", "In", "this", "paper", ",", "we", "present", "a", "novel", "Deep", "CNN", "model", "for", "Identity", "-", "Aware", "Transfer", "(", "DIAT", ")", "of", "facial", "attributes", "which", "can", "provide", "a", "unified", "solution", "to", "several", "facial", "animation", "and", "manipulation", "tasks", ",", "e", ".", "g", ".", ",", "expression", "transfer", ",", "accessory", "removal", ",", "age", "progression", ",", "and", "gender", "transfer", ".", "For", "each", "reference", "attribute", "label", ",", "we", "train", "a", "CNN", "model", "for", "the", "transfer", "of", "the", "input", "image", "to", "the", "desired", "attribute", ".", "Note", "that", "the", "reference", "attribute", "may", "be", "only", "related", "to", "some", "parts", "of", "the", "image", ".", "To", "avoid", "the", "the", "incorrect", "editing", "on", "attribute", "irrelevant", "region", ",", "our", "model", "consists", "of", "a", "mask", "network", "and", "an", "attribute", "transform", "network", ".", "The", "attribute", "transform", "network", "is", "presented", "to", "edit", "the", "input", "image", "for", "generating", "the", "desired", "attribute", ".", "While", "the", "mask", "network", "is", "adopted", "to", "estimate", "a", "mask", "of", "the", "attribute", "relevant", "region", "for", "guiding", "the", "combination", "of", "the", "input", "image", "and", "transformed", "image", ".", "Then", "attribute", "transform", "network", "and", "mask", "network", "work", "collaboratively", "to", "generate", "final", "photo", "-", "realistic", "transfer", "result", ".", "For", "attribute", "transfer", ",", "the", "ground", "truth", "transfer", "results", "generally", "are", "very", "difficult", "or", "even", "impossible", "to", "obtain", ".", "Therefore", ",", "we", "follow", "the", "GAN", "framework", "to", "train", "the", "model", ".", "As", "for", "training", "data", ",", "we", "only", "consider", "the", "binary", "attribute", "labels", "presented", "in", "the", "large", "-", "scale", "CelebFaces", "Attributes", "(", "CelebA", ")", "dataset", ".", "To", "capture", "the", "convolutional", "feature", "distribution", "of", "each", "attribute", ",", "we", "construct", "an", "attribute", "guided", "set", "using", "all", "the", "images", "with", "the", "desired", "attribute", "in", "CelebA.", "Then", ",", "the", "input", "set", "is", "defined", "as", "a", "set", "of", "input", "images", "without", "the", "reference", "attribute", ".", "Due", "to", "the", "infeasibility", "of", "ground", "truth", "transfer", "results", ",", "two", "alternative", "losses", ",", "i", ".", "e", ".", ",", "adversarial", "attribute", "loss", "and", "identity", "-", "aware", "perceptual", "loss", ",", "are", "incorporated", "for", "unsupervised", "training", "of", "our", "DIAT", "model", ".", "Furthermore", ",", "two", "regularizers", ",", "i", ".", "e", ".", ",", "perceptual", "regularization", "and", "attribute", "ratio", "regularization", ",", "are", "also", "introduced", "in", "the", "learning", "of", "DIAT", ".", "In", "terms", "of", "attribute", "transfer", ",", "the", "generated", "image", "should", "have", "the", "desired", "attribute", "label", ".", "Following", "the", "GAN", "framework", ",", "we", "define", "the", "adversarial", "attribute", "loss", "on", "the", "attribute", "discriminator", "to", "require", "the", "generated", "image", "to", "have", "the", "desired", "attribute", ".", "As", "for", "identity", "-", "aware", "transfer", ",", "our", "DIAT", "requires", "that", "the", "generated", "image", "should", "keep", "the", "same", "or", "similar", "identity", "to", "the", "input", "image", ".", "To", "this", "end", ",", "the", "identity", "-", "aware", "perceptual", "loss", "is", "introduced", "on", "the", "convolutional", "feature", "map", "of", "a", "CNN", "model", "to", "model", "the", "content", "similarity", "between", "the", "reference", "face", "and", "the", "generated", "face", ".", "Instead", "of", "adopting", "any", "pre", "-", "trained", "CNNs", ",", "we", "suggest", "to", "define", "the", "perceptual", "loss", "on", "the", "attribute", "discriminator", ",", "which", "can", "be", "adaptively", "trained", "along", "with", "the", "learning", "procedure", ",", "and", "is", "named", "as", "adaptive", "perceptual", "loss", ".", "Compared", "with", "conventional", "perceptual", "loss", ",", "ours", "is", "more", "effective", "in", "computation", ",", "tailored", "to", "our", "specific", "attribute", "transfer", "task", ",", "and", "can", "serve", "as", "a", "kind", "of", "hidden", "-", "layer", "supervision", "or", "regularization", "to", "ease", "the", "training", "of", "the", "DIAT", "model", ".", "To", "further", "encourage", "the", "identity", "keeping", "property", ",", "we", "add", "an", "identity", "loss", "by", "minimizing", "the", "distance", "between", "feature", "representations", "of", "the", "generated", "face", "and", "the", "reference", ".", "Pre", "-", "trained", "VGG", "-", "Face", "is", "used", "to", "extract", "the", "identity", "related", "features", "for", "face", "verification", ".", "The", "model", "objective", "of", "DIAT", "also", "consider", "two", "regularizers", ",", "i", ".", "e", ".", ",", "perceptual", "regularization", "and", "attribute", "ratio", "regularization", ".", "To", "suppress", "the", "artifacts", ",", "we", "propose", "a", "denoising", "network", "to", "serve", "for", "a", "perceptual", "regularization", "on", "the", "generated", "image", ".", "To", "guide", "the", "learning", "of", "mask", "network", ",", "an", "attribute", "ratio", "regularization", "is", "introduced", "to", "constrain", "the", "size", "of", "attribute", "relevant", "region", ".", "Finally", ",", "our", "DIAT", "model", "can", "be", "learned", "from", "training", "data", "by", "incorporating", "adversarial", "attribute", "loss", ",", "adaptive", "perceptual", "loss", "with", "perceptual", "regularization", "and", "attribute", "ratio", "regularization", ".", "Extensive", "experiments", "are", "conducted", "on", "CelebA", "and", "real", "images", "from", "the", "website", "iStock", ".", "As", "illustrated", "in", "Fig", ".", "[", "reference", "]", ",", "our", "DIAT", "performs", "favorably", "in", "attribute", "transfer", "with", "minor", "or", "no", "modification", "on", "the", "identity", "of", "the", "input", "faces", ".", "Even", "for", "some", "identity", "-", "related", "attributes", "(", "e", ".", "g", ".", ",", "gender", ")", ",", "our", "DIAT", "can", "obtain", "visually", "impressive", "transfer", "result", "while", "retaining", "most", "identity", "-", "relevant", "features", ".", "Computational", "efficiency", "is", "also", "a", "prominent", "merit", "of", "our", "method", ".", "In", "the", "testing", "stage", ",", "our", "DIAT", "can", "process", "more", "than", "one", "hundred", "of", "images", "within", "one", "second", ".", "Furthermore", ",", "our", "model", "can", "be", "extended", "to", "face", "hallucination", ",", "and", "is", "effective", "in", "generating", "photo", "-", "realistic", "high", "resolution", "images", ".", "A", "preliminary", "report", "of", "this", "work", "is", "given", "in", "2016", ".", "To", "sum", "up", ",", "our", "contribution", "is", "three", "-", "fold", ":", "A", "novel", "DIAT", "model", "is", "developed", "for", "facial", "attribute", "transfer", ".", "For", "better", "preserving", "of", "attribute", "irrelevant", "feature", ",", "our", "model", "comprises", "a", "mask", "network", "and", "an", "attribute", "transform", "network", ",", "which", "collaborate", "to", "generate", "the", "transfer", "result", "and", "can", "be", "jointly", "learned", "from", "training", "data", ".", "Adversarial", "attribute", "loss", ",", "adaptive", "perceptual", "loss", ",", "identity", "loss", ",", "perceptual", "regularization", ",", "and", "attribute", "ratio", "regularization", "are", "incorporated", "for", "training", "our", "DIAT", "model", ".", "The", "adversarial", "attribute", "loss", "is", "adopted", "to", "make", "the", "transfer", "result", "exhibit", "the", "desired", "attribute", ",", "and", "the", "adaptive", "perceptual", "loss", "is", "defined", "on", "the", "discriminator", "for", "identity", "-", "aware", "transfer", "while", "improving", "training", "efficiency", ".", "Moreover", ",", "perceptual", "regularization", "and", "attribute", "ratio", "regularization", "are", "further", "introduced", "for", "suppressing", "the", "artifacts", "and", "constraining", "the", "mask", "network", ".", "Experimental", "results", "validate", "the", "effectiveness", "and", "efficiency", "of", "our", "method", "for", "identity", "-", "aware", "attribute", "transfer", ".", "Our", "DIAT", "can", "be", "used", "for", "the", "transfer", "of", "either", "local", "(", "e", ".", "g", ".", ",", "mouth", ")", ",", "global", "(", "e", ".", "g", ".", ",", "age", "progression", ")", "or", "identity", "-", "related", "(", "e", ".", "g", ".", ",", "gender", ")", "attributes", ",", "and", "can", "be", "extended", "to", "face", "hallucination", ".", "The", "remainder", "of", "the", "paper", "is", "organized", "as", "follows", ".", "Section", "[", "reference", "]", "gives", "a", "brief", "survey", "on", "relevant", "work", ".", "Section", "[", "reference", "]", "describes", "the", "model", "and", "learning", "of", "our", "DIAT", "method", ".", "Section", "[", "reference", "]", "reports", "the", "experimental", "results", "on", "facial", "attribute", "transfer", "and", "face", "hallucination", ".", "Finally", ",", "Section", "[", "reference", "]", "ends", "this", "work", "with", "several", "concluding", "remarks", ".", "section", ":", "Related", "work", "Deep", "convolutional", "neural", "networks", "(", "CNNs", ")", "not", "only", "have", "achieved", "unprecedented", "success", "in", "versatile", "high", "level", "vision", "problems", ",", "but", "also", "exhibited", "their", "remarkable", "power", "in", "understanding", ",", "generating", ",", "and", "recovering", "images", ".", "In", "this", "section", ",", "we", "focus", "on", "the", "task", "of", "facial", "attribute", "transfer", ",", "and", "briefly", "survey", "the", "CNN", "models", "for", "image", "generation", "and", "face", "generation", ".", "subsection", ":", "CNN", "for", "image", "generation", "Generative", "image", "modeling", "is", "a", "critical", "issue", "for", "image", "generation", "and", "many", "low", "level", "vision", "problems", ".", "Conventional", "sparse", ",", "low", "rank", ",", "FRAME", "and", "non", "-", "local", "similarity", "based", "models", "usually", "are", "limited", "in", "capturing", "highly", "complex", "and", "long", "-", "range", "dependence", "between", "pixels", ".", "For", "better", "image", "modeling", ",", "a", "number", "of", "CNN", "-", "based", "methods", "have", "been", "proposed", ",", "including", "convolutional", "auto", "-", "encoder", ",", "PixelCNN", "and", "PixelRNN", ",", "and", "they", "have", "been", "applied", "to", "image", "completion", "and", "generation", ".", "Several", "CNN", "architectures", "have", "been", "developed", "for", "image", "generation", ".", "Fully", "convolutional", "networks", "can", "be", "trained", "in", "the", "supervised", "learning", "manner", "to", "generate", "an", "image", "from", "an", "input", "image", ".", "The", "generative", "CNN", "model", "stacks", "four", "convolution", "layers", "upon", "five", "fully", "connected", "layers", "to", "generate", "images", "from", "object", "description", ".", "Kulkarni", "et", "al", ".", "suggest", "the", "Deep", "Convolution", "Inverse", "Graphics", "Network", "(", "DC", "-", "IGN", ")", ",", "which", "follows", "the", "variational", "autoencoder", "architecture", "to", "transform", "the", "input", "image", "into", "different", "pose", "and", "lighting", "condition", ".", "However", ",", "both", "generative", "CNN", "and", "DC", "-", "IGN", "require", "many", "labeled", "images", "in", "training", ".", "To", "visualize", "and", "understand", "CNN", "features", ",", "several", "methods", "have", "been", "proposed", "to", "reconstruct", "images", "by", "inverting", "deep", "representation", "or", "maximizing", "class", "score", ".", "Subsequently", ",", "Gatys", "et", "al", ".", "suggest", "to", "combine", "content", "and", "style", "losses", "defined", "on", "deep", "representation", "on", "the", "off", "-", "the", "-", "shelf", "CNNs", "for", "artistic", "style", "transfer", ".", "To", "improve", "the", "efficiency", ",", "alternative", "approaches", "have", "been", "proposed", "by", "substituting", "the", "iterative", "optimization", "procedure", "with", "pre", "-", "trained", "feed", "-", "forward", "CNN", ".", "And", "perceptual", "loss", "has", "also", "been", "adopted", "for", "style", "transfer", "and", "other", "generation", "tasks", ".", "Motivated", "by", "these", "works", ",", "both", "identity", "-", "aware", "adaptive", "perceptual", "loss", "and", "perceptual", "regularization", "are", "exploited", "in", "our", "DIAT", "model", "to", "meet", "the", "requirement", "of", "facial", "attribute", "transfer", ".", "Another", "representative", "approach", "is", "generative", "adversarial", "network", "(", "GAN", ")", ",", "where", "a", "discriminator", "and", "a", "generator", "are", "alternatingly", "trained", "as", "an", "adversarial", "game", ".", "The", "generator", "aims", "to", "generate", "images", "to", "match", "the", "data", "distribution", ",", "while", "the", "discriminator", "attempts", "to", "distinguish", "between", "the", "generated", "images", "and", "the", "training", "data", ".", "Laplacian", "Pyramid", "of", "GANs", "is", "further", "suggested", "to", "generate", "high", "quality", "image", "in", "a", "coarse", "-", "to", "-", "fine", "manner", ".", "Radford", "et", "al", ".", "extend", "GAN", "with", "the", "fully", "deep", "convolutional", "networks", "(", "i", ".", "e", ".", ",", "DCGAN", ")", "for", "image", "generation", ".", "To", "learn", "disentangled", "representations", ",", "information", "-", "theoretic", "extension", "of", "GAN", "is", "proposed", "by", "maximizing", "the", "mutual", "information", "between", "a", "subset", "of", "noise", "variables", "and", "the", "generated", "results", ".", "In", ",", "WGAN", "and", "WGAN", "-", "GP", "minimize", "the", "Wasserstein", "-", "1", "distance", "between", "the", "generated", "distribution", "and", "the", "real", "distribution", "to", "improve", "the", "stability", "of", "learning", "generator", ".", "In", "this", "work", ",", "we", "adopt", "the", "WGAN", "framework", "to", "learn", "our", "DIAT", "model", ",", "and", "further", "suggest", "adaptive", "perceptual", "loss", "for", "identity", "-", "aware", "transfer", "and", "perceptual", "regularization", "to", "suppress", "visual", "artifacts", ".", "subsection", ":", "CNN", "for", "face", "generation", "Facial", "attribute", "transfer", "has", "received", "considerable", "recent", "attention", ".", "Larsen", "et", "al", ".", "present", "to", "combine", "variational", "autoencode", "with", "GAN", "(", "VAE", "/", "GAN", ")", "for", "image", "generation", ".", "By", "modeling", "the", "attribute", "vector", "as", "the", "difference", "between", "the", "mean", "latent", "representations", "of", "the", "images", "with", "and", "without", "the", "reference", "attribute", ",", "VAE", "/", "GAN", "can", "provide", "a", "flexible", "solution", "to", "arbitrary", "facial", "attribute", "transfer", ",", "but", "is", "limited", "in", "transfer", "performance", ".", "Li", "et", "al", ".", "suggest", "an", "attribute", "driven", "and", "identity", "-", "preserving", "face", "generation", "model", "by", "solving", "an", "optimization", "problems", "with", "perceptual", "loss", ",", "which", "is", "computationally", "expensive", "and", "can", "not", "obtain", "high", "quality", "results", ".", "Perarnau", "et", "al", ".", "adopt", "an", "encoder", "-", "decoder", "architecture", ",", "where", "attribute", "transfer", "can", "be", "conducted", "by", "editing", "the", "latent", "representation", ".", "Shen", "et", "al", ".", "learn", "the", "residual", "image", "in", "the", "GAN", "framework", ",", "and", "adopt", "dual", "learning", "to", "learn", "two", "reverse", "attribute", "transfer", "models", "simultaneously", ".", "Zhou", "et", "al", ".", "propose", "a", "model", "to", "learn", "object", "transfiguration", "from", "two", "sets", "of", "unpaired", "images", "that", "have", "the", "opposite", "attribute", ".", "However", ",", "most", "existing", "methods", "can", "not", "achieve", "high", "quality", "transfer", "results", ",", "and", "visible", "artifacts", "and", "over", "-", "smoothing", "usually", "are", "inevitable", ".", "In", "comparison", ",", "our", "DIAT", "model", "can", "achieve", "much", "better", "transfer", "results", "than", "the", "competing", "methods", ".", "Besides", ",", "CNNs", "have", "also", "been", "developed", "for", "other", "face", "generation", "tasks", ".", "For", "painting", "style", "transfer", "of", "head", "portrait", ",", "Selim", "et", "al", ".", "modify", "the", "perceptual", "loss", "to", "balance", "the", "contribution", "of", "the", "input", "photograph", "and", "the", "aligned", "exemplar", "painting", ".", "Gucluturk", "et", "al", ".", "train", "a", "feed", "-", "forward", "CNN", "with", "perceptual", "loss", "for", "sketch", "inversion", ".", "Yeh", "et", "al", ".", "apply", "DCGAN", "to", "semantic", "face", "inpainting", "in", "an", "optimization", "manner", ".", "section", ":", "Deep", "CNNs", "for", "Identity", "-", "aware", "Attribute", "Transfer", "In", "this", "section", ",", "we", "present", "our", "DIAT", "model", "for", "identity", "-", "aware", "transfer", "of", "facial", "attribute", ".", "As", "illustrated", "in", "Fig", ".", "[", "reference", "]", ",", "our", "model", "involves", "a", "mask", "network", "and", "an", "attribute", "transform", "network", "which", "collaborate", "to", "produce", "the", "transfer", "result", ".", "To", "train", "our", "model", ",", "we", "incorporate", "the", "adversarial", "attribute", "loss", ",", "adaptive", "perceptual", "loss", ",", "perceptual", "regularization", "and", "attribute", "ratio", "regularization", ".", "subsection", ":", "Network", "architecture", "Most", "facial", "attributes", ",", "e", ".", "g", ".", ",", "expression", "and", "accessory", ",", "are", "local", "-", "based", "and", "only", "related", "to", "part", "of", "facial", "image", ".", "Even", "for", "global", "attributes", "such", "as", "age", "and", "gender", ",", "some", "parts", ",", "e", ".", "g", ".", ",", "the", "background", ",", "should", "also", "keep", "the", "same", "with", "the", "source", "image", ".", "For", "the", "sake", "of", "preserving", "attribute", "irrelevant", "feature", ",", "it", "is", "natural", "to", "only", "perform", "attribute", "transfer", "in", "image", "region", "related", "to", "specific", "attribute", ".", "However", ",", "it", "is", "not", "a", "trivial", "issue", "to", "find", "the", "attribute", "relevant", "region", ".", "One", "possible", "solution", "is", "to", "manually", "specify", "the", "relevant", "region", "for", "each", "attribute", "given", "a", "new", "transfer", "task", ",", "but", "it", "undoubtedly", "restricts", "the", "universality", "and", "adaptivity", "of", "the", "solution", ".", "In", "this", "work", ",", "we", "aim", "to", "provide", "a", "unified", "solution", "to", "attribute", "transfer", ",", "which", "indicates", "that", "we", "only", "require", "to", "prepare", "training", "data", "and", "retrain", "the", "model", "when", "a", "new", "transfer", "task", "comes", ".", "To", "this", "end", ",", "our", "whole", "attribute", "transfer", "network", "is", "comprised", "of", "two", "sub", "-", "networks", ",", "i", ".", "e", ".", ",", "mask", "network", "and", "attribute", "transform", "network", ".", "Both", "mask", "network", "and", "attribute", "transform", "network", "take", "the", "source", "image", "as", "input", ".", "The", "mask", "network", "is", "utilized", "to", "predict", "a", "mask", "to", "indicate", "the", "attribute", "relevant", "region", ",", "while", "the", "attribute", "transform", "network", "is", "used", "to", "produce", "the", "transformed", "image", ".", "Given", "and", ",", "the", "final", "transfer", "result", "can", "be", "obtained", "by", ",", "where", "denotes", "the", "element", "-", "wise", "product", "operator", ".", "We", "also", "note", "that", "both", "attribute", "transform", "network", "and", "mask", "network", "can", "be", "learned", "from", "training", "data", "in", "an", "end", "-", "to", "-", "end", "manner", ".", "In", "the", "following", ",", "we", "describe", "the", "architecture", "of", "attribute", "transform", "network", "and", "mask", "network", ",", "respectively", ".", "Attribute", "transform", "network", ".", "We", "adopt", "the", "Unet", "for", "attribute", "transform", "due", "to", "its", "good", "tradeoff", "between", "efficiency", "and", "reconstruction", "ability", ".", "In", "general", ",", "the", "Unet", "architecture", "involves", "an", "encoder", "subnetwork", "and", "a", "decoder", "subnetwork", ",", "then", "skip", "connection", "and", "pooling", "operation", "are", "further", "introduced", "to", "exploit", "multi", "-", "scale", "information", ".", "As", "for", "attribute", "transform", ",", "we", "design", "a", "10", "-", "layer", "Unet", ",", "which", "includes", "5", "convolution", "layers", "for", "encoding", "and", "another", "5", "convolution", "layers", "for", "decoding", ".", "In", "the", "encoder", ",", "we", "use", "convolution", "with", "stride", "2", "for", "downsampling", ".", "In", "the", "decoder", ",", "a", "depth", "to", "width", "(", "DTOW", ")", "layer", "is", "deployed", "for", "upsampling", ",", "and", "the", "element", "-", "wise", "summation", "operation", "is", "adopted", "to", "fuse", "the", "feature", "maps", "from", "the", "encoder", "and", "decoder", "subnetworks", ".", "The", "detailed", "parameters", "of", "the", "attribute", "transform", "network", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "Mask", "network", ".", "As", "for", "mask", "network", ",", "we", "first", "adopt", "a", "5", "-", "layer", "fully", "convolutional", "network", "to", "generate", "a", "binary", "mask", "for", "indicating", "the", "attribute", "relevant", "region", ".", "A", "batch", "normalization", "layer", "is", "added", "after", "each", "convolution", "layer", ".", "Then", ",", "upsampling", "is", "deployed", "by", "simply", "replicating", "each", "element", "in", "the", "binary", "mask", "times", ".", "In", "order", "to", "make", "the", "generated", "image", "smooth", ",", "we", "further", "utilize", "a", "Gaussian", "filter", "with", "the", "standard", "deviation", "of", "to", "produce", "the", "final", "mask", ".", "To", "sum", "up", ",", "the", "details", "of", "the", "mask", "network", "are", "given", "in", "Table", "[", "reference", "]", ".", "In", "our", "mask", "network", ",", "ReLU", "is", "adopted", "for", "nonlinearity", "for", "the", "first", "4", "convolution", "layers", ".", "As", "for", "the", "fifth", "convolution", "layer", ",", "we", "adopt", "the", "sigmoid", "nonlinearity", ",", "and", "the", "binarization", "operation", "is", "then", "used", "to", "obtain", "the", "binary", "mask", ",", "where", "denotes", "an", "element", "of", "the", "feature", "map", ".", "However", ",", "the", "gradient", "of", "the", "binarizer", "is", "zero", "almost", "everywhere", "except", "that", "it", "is", "infinite", "when", ",", "making", "any", "layer", "before", "the", "binarizer", "never", "be", "updated", "during", "training", ".", "As", "a", "remedy", ",", "we", "follow", "the", "straight", "-", "through", "estimator", "on", "gradient", ",", "and", "introduce", "a", "piecewise", "linear", "proxy", "function", "to", "approximate", ",", "During", "training", ",", "is", "still", "used", "in", "forward", "-", "propagation", "calculation", ",", "while", "is", "used", "in", "back", "-", "propagation", ",", "with", "its", "gradient", "computed", "by", ",", "subsection", ":", "Model", "objective", "By", "enforcing", "proper", "constraints", "on", "transfer", "result", ",", "both", "the", "attribute", "transform", "network", "and", "mask", "network", "can", "be", "learned", "from", "training", "data", ".", "However", ",", "the", "ground", "truth", "of", "attribute", "transfer", "usually", "is", "unavailable", "and", "not", "unique", ".", "For", "example", ",", "it", "is", "generally", "impossible", "to", "obtain", "the", "ground", "truth", "of", "gender", "transfer", "in", "reality", ".", "Instead", ",", "the", "training", "data", "used", "in", "this", "work", "includes", "a", "guided", "set", "of", "images", "with", "the", "desired", "reference", "attribute", "and", "a", "source", "set", "of", "input", "images", "not", "with", "the", "reference", "attribute", ".", "And", "we", "do", "not", "require", "the", "images", "from", "guided", "set", "to", "have", "the", "same", "identity", "with", "theose", "from", "source", "set", ".", "For", "the", "sake", "of", "identity", "-", "aware", "attribute", "transfer", ",", "we", "define", "two", "alternative", "losses", ":", "(", "i", ")", "adversarial", "attribute", "loss", "to", "make", "the", "transfer", "result", "exhibit", "the", "desired", "attribute", ",", "and", "(", "ii", ")", "adaptive", "perceptual", "loss", "and", "identity", "loss", "to", "make", "the", "generated", "image", "keep", "the", "same", "or", "similar", "identity", "to", "input", "image", ".", "To", "suppress", "the", "visual", "artifacts", "of", "the", "transfer", "result", ",", "we", "further", "include", "(", "iii", ")", "a", "perceptual", "regularization", "defined", "on", "a", "denoising", "network", ".", "Finally", ",", "(", "iv", ")", "an", "attribute", "ratio", "regularization", "is", "deployed", "on", "to", "guide", "the", "learning", "of", "mask", "network", ".", "In", "the", "following", ",", "we", "provide", "more", "details", "on", "these", "losses", "and", "regularization", "terms", ".", "Adversarial", "attribute", "loss", ".", "Adversarial", "strategy", "is", "a", "common", "stratey", "that", "is", "widely", "used", "in", "security", "problems", ".", "For", "computer", "vision", ",", "an", "adversarial", "learning", "framework", "called", "generative", "adversarial", "networks", "(", "GANs", ")", "also", "shows", "powerful", "ability", "on", "generating", "images", "that", "fulfil", "the", "distribution", "of", "a", "set", "of", "images", "without", "any", "ground", "truth", "targets", ".", "Considering", "the", "infeasibility", "of", "obtaining", "the", "ground", "truth", "transfer", "result", ",", "we", "define", "the", "adversarial", "attribute", "loss", "based", "on", "the", "guided", "set", "and", "the", "source", "set", ".", "If", "the", "guided", "set", "is", "of", "large", "scale", ",", "it", "can", "provide", "a", "natural", "representation", "of", "the", "attribute", "distribution", ".", "Therefore", ",", "the", "goal", "of", "adversarial", "attribute", "loss", "is", "to", "make", "that", "the", "distribution", "of", "generated", "images", "matches", "the", "real", "attribute", "distribution", ".", "To", "this", "end", ",", "we", "adopt", "the", "generative", "adversarial", "network", "framework", ",", "where", "the", "generator", "is", "the", "attribute", "transfer", "network", ",", "and", "the", "discriminator", "is", "used", "to", "define", "the", "adversarial", "attribute", "loss", ".", "The", "details", "of", "the", "discriminator", "are", "provided", "in", "Table", "[", "reference", "]", ",", "which", "contains", "6", "convolution", "layers", "followed", "by", "another", "two", "fully", "-", "connected", "layers", ".", "Denote", "by", "an", "input", "image", "from", ",", "and", "an", "image", "from", ".", "Let", "be", "the", "distribution", "of", "the", "input", "images", ",", "be", "the", "distribution", "of", "the", "images", "with", "the", "reference", "attribute", ".", "The", "discriminator", "is", "defined", "as", "to", "output", "the", "probability", "that", "the", "image", "comes", "from", "the", "set", ".", "To", "train", "the", "generator", "and", "the", "discriminator", ",", "we", "take", "use", "of", "the", "following", "adversarial", "attribute", "loss", ",", "In", "order", "to", "improve", "the", "training", "stability", ",", "the", "improved", "Wasserstein", "GAN", "is", "adopted", "by", "defining", "the", "loss", "as", ",", "For", "simplicity", ",", "we", "respectively", "define", "the", "adversarial", "attribute", "losses", "for", "the", "generator", "and", "discriminator", "as", "follows", ",", "Adaptive", "perceptual", "loss", ".", "The", "adaptive", "perceptual", "loss", "is", "introduced", "to", "guarantee", "that", "the", "transfer", "result", "keeps", "the", "same", "or", "similar", "identity", "with", "the", "input", "image", ".", "Due", "to", "identity", "is", "a", "high", "level", "semantic", "concept", ",", "it", "is", "not", "proper", "to", "define", "identity", "-", "aware", "loss", "by", "forcing", "two", "images", "to", "be", "exactly", "the", "same", "in", "pixel", "domain", ".", "Instead", ",", "we", "define", "the", "squared", "-", "error", "loss", "on", "the", "feature", "representations", "of", "the", "discriminator", ",", "resulting", "in", "our", "adaptive", "perceptual", "loss", ".", "Denote", "by", "the", "discriminator", ",", "and", "the", "feature", "map", "of", "the", "-", "th", "convolution", "layer", ".", ",", "and", "represent", "the", "channel", "number", ",", "height", ",", "and", "width", "of", "the", "feature", "map", ",", "respectively", ".", "We", "then", "define", "the", "perceptual", "loss", "between", "and", "on", "the", "-", "th", "convolution", "layer", "as", ",", "And", "the", "identity", "-", "aware", "adaptive", "perceptual", "loss", "is", "further", "defined", "as", ",", "We", "note", "that", "the", "discriminator", "is", "learned", "from", "training", "data", ".", "Thus", ",", "the", "network", "parameters", "of", "will", "be", "changed", "along", "with", "the", "updating", "of", "discriminator", ",", "and", "thus", "we", "name", "as", "adaptive", "perceptual", "loss", ".", "In", "contrast", ",", "conventional", "perceptual", "loss", "is", "defined", "on", "the", "off", "-", "the", "-", "shelf", "CNNs", "(", "e", ".", "g", ".", ",", "VGG", "-", "Face", ")", ".", "Compared", "with", "conventional", "perceptual", "loss", ",", "our", "adaptive", "perceptual", "loss", "generally", "is", "more", "effective", "in", "improving", "the", "training", "efficiency", "and", "attribute", "transfer", "performance", ":", "The", "training", "efficiency", "of", "adaptive", "perceptual", "loss", "can", "be", "further", "explained", "from", "two", "aspects", ".", "(", "i", ")", "For", "conventional", "perceptual", "loss", ",", "the", "forward", "and", "backward", "calculations", "are", "required", "for", "both", "the", "off", "-", "the", "-", "shelf", "CNN", "and", "the", "discriminator", "during", "training", ".", "Due", "to", "that", "the", "adaptive", "perceptual", "loss", "is", "defined", "on", "the", "discriminator", ",", "it", "is", "sufficient", "to", "only", "conduct", "forward", "and", "backward", "calculation", "on", "the", "discriminator", ",", "making", "our", "DIAT", "more", "efficient", "in", "training", ".", "(", "ii", ")", "For", "conventional", "GAN", ",", "the", "generator", "usually", "is", "difficult", "to", "be", "trained", ".", "As", "for", "our", "DIAT", ",", "it", "can", "be", "trained", "by", "both", "the", "adversarial", "attribute", "loss", "and", "adaptive", "perceptual", "loss", ",", "greatly", "accelerating", "the", "training", "speed", ".", "Actually", ",", "the", "adaptive", "perceptual", "loss", "is", "defined", "on", "the", "third", "and", "fourth", "convolution", "layers", "of", "the", "discriminator", ",", "which", "can", "serve", "as", "some", "kind", "of", "hidden", "-", "layer", "supervision", "and", "benefit", "the", "convergence", "of", "network", "training", ".", "For", "conventional", "perceptual", "loss", ",", "the", "off", "-", "the", "-", "shelf", "CNNs", "generally", "are", "pre", "-", "trained", "using", "other", "training", "data", "and", "are", "not", "tailored", "to", "attribute", "transfer", ".", "One", "plausible", "choice", "is", "the", "VGG", "-", "Face", ",", "which", ",", "however", ",", "is", "trained", "for", "face", "recognition", "and", "may", "not", "be", "suitable", "for", "identity", "-", "aware", "attribute", "transfer", ".", "In", "comparison", ",", "our", "adaptive", "perceptual", "loss", "is", "defined", "on", "the", "discriminator", "which", "is", "trained", "for", "modeling", ".", "Such", "loss", "can", "thus", "provide", "natural", "balance", "between", "identity", "similarity", "and", "attribute", "transfer", "and", "benefit", "transfer", "performance", ".", "For", "example", ",", "in", "terms", "of", "gender", "transfer", ",", "the", "introduction", "of", "adaptive", "perceptual", "loss", "will", "allow", "the", "adaptive", "adjustment", "on", "the", "length", "of", "hair", ".", "Identity", "Loss", ".", "The", "proposed", "adaptive", "perceptual", "loss", "does", "help", "keep", "the", "content", "similarity", "between", "the", "generated", "face", "and", "the", "reference", ".", "However", ",", "it", "can", "not", "guarantee", "the", "identity", "by", "itself", ".", "To", "further", "enhance", "the", "identity", "keeping", "property", ",", "we", "add", "constrains", "on", "the", "feature", "representation", "extracted", "for", "face", "recognition", "or", "verification", ".", "In", "face", "verification", "task", ",", "two", "faces", "are", "from", "the", "same", "person", "when", "the", "distance", "between", "two", "features", "are", "smaller", "than", "ceratain", "threshold", ".", "Here", ",", "we", "adopt", "VGG", "-", "Face", "and", "model", "the", "distance", "between", "the", "features", "of", "the", "generated", "face", "and", "the", "reference", "as", "the", "identity", "loss", ",", "Perceptual", "regularization", ".", "Despite", "the", "use", "of", "adversarial", "attribute", "loss", "and", "adaptive", "perceptual", "loss", ",", "visual", "artifacts", "are", "still", "inevitable", "in", "the", "transfer", "result", ".", "Image", "regularization", "is", "thus", "required", "to", "encourage", "the", "spatial", "smoothness", "while", "preserving", "small", "scale", "details", "of", "the", "generated", "face", ".", "One", "choice", "is", "the", "Total", "Variation", "(", "TV", ")", "regularizer", "which", "has", "been", "adopted", "in", "CNN", "feature", "visualization", "and", "artistic", "style", "transfer", ".", "However", ",", "the", "TV", "regularizer", "is", "limited", "in", "recovering", "small", "-", "scale", "texture", "details", "and", "suppressing", "complex", "artifacts", ".", "Moreover", ",", "it", "is", "a", "generic", "model", "that", "does", "not", "consider", "the", "characteristics", "of", "facial", "images", ".", "In", "this", "work", ",", "we", "take", "the", "facial", "characteristics", "into", "account", "and", "train", "a", "denoising", "network", "for", "perceptual", "regularization", ".", "To", "train", "the", "denoising", "network", ",", "we", "generate", "the", "noisy", "image", "by", "adding", "Gaussian", "noise", "with", "the", "standard", "deviation", "of", "to", "the", "clean", "facial", "image", "from", "CelebA.", "Inspired", "by", "residual", "learning", ",", "we", "train", "the", "denoising", "network", "through", "learning", "the", "residual", "between", "the", "noise", "image", "and", "the", "clean", "image", ".", "Taking", "the", "noise", "image", "as", "input", ",", "the", "denoising", "network", "utilizes", "a", "fully", "convolutional", "network", "of", "6", "layers", "to", "predict", "the", "residual", ".", "The", "denoising", "result", "can", "then", "be", "obtained", "by", ".", "The", "architecture", "of", "is", "listed", "in", "Table", "[", "reference", "]", ".", "Denote", "by", "a", "training", "set", ",", "where", "denotes", "the", "-", "th", "noisy", "image", "and", "the", "corresponding", "clean", "image", ".", "The", "objective", "for", "learning", "is", "given", "as", ",", "Given", "the", "denoising", "network", "and", "the", "transfer", "result", ",", "we", "define", "the", "perceptual", "regularization", "as", ",", "where", "denotes", "the", "Frobenius", "norm", ".", "Note", "that", "predicts", "the", "residual", "between", "the", "latent", "clean", "image", "and", ".", "Minimizing", "makes", "be", "close", "to", "the", "clean", "image", ",", "and", "can", "be", "used", "to", "suppress", "the", "noise", "and", "artifacts", "in", ".", "Furthermore", ",", "the", "threshold", "is", "introduced", "for", "better", "preserving", "of", "small", "scale", "details", ",", "and", "we", "empirically", "set", "be", "a", "value", "in", "the", "range", "of", ".", "Note", "that", "the", "regularizer", "in", "Eqn", ".", "(", "[", "reference", "]", ")", "is", "defined", "on", "the", "denoising", "network", ",", "and", "thus", "is", "named", "as", "perceptual", "regularization", ".", "Attribute", "ratio", "regularization", ".", "The", "size", "of", "attribute", "relevant", "region", "varies", "for", "different", "attributes", ".", "For", "example", ",", "the", "region", "related", "to", "mouth", "open", "/", "close", "mainly", "includes", "the", "mouth", "and", "should", "be", "small", ".", "For", "glasses", "removal", ",", "the", "attribute", "relevant", "region", "includes", "the", "two", "eyes", "and", "is", "relatively", "large", ".", "As", "for", "gender", "transfer", ",", "all", "the", "face", "region", "and", "the", "hair", "should", "be", "attribute", "relevant", ".", "Therefore", ",", "we", "introduce", "an", "attribute", "ratio", "regularization", "term", "to", "constrain", "the", "size", "of", "attribute", "relevant", "region", ".", "Specifically", ",", "such", "regularization", "is", "defined", "on", "the", "binary", "mask", "in", "Eqn", ".", "(", "[", "reference", "]", ")", ".", "Denote", "by", "the", "image", "size", ",", "and", "the", "expected", "ratio", "of", "the", "region", "for", "a", "specific", "attribute", ".", "The", "attribute", "ratio", "regularization", "is", "then", "defined", "as", ",", "In", "our", "experiments", ",", "we", "set", "smaller", "value", "for", "local", "attribute", "and", "larger", "value", "for", "global", "attribute", ".", "Objective", "function", ".", "We", "define", "the", "objective", "function", "for", "learning", "the", "transfer", "model", "and", "the", "discriminator", "by", "combining", "the", "adversarial", "attribute", "loss", ",", "adaptive", "perceptual", "loss", ",", "perceptual", "regularization", ",", "and", "attribute", "ratio", "regularization", ".", "The", "transfer", "model", "is", "learned", "by", "minimizing", "the", "following", "objective", ",", "where", ".", ",", ",", "and", "are", "the", "tradeoff", "parameters", "for", "the", "adaptive", "perceptual", "loss", ",", "attribute", "ratio", "regularization", ",", "and", "perceptual", "regularization", ",", "respectively", ".", "However", ",", "it", "is", "difficult", "to", "set", "the", "tradeoff", "parameter", ".", "Instead", ",", "we", "empirically", "find", "that", "the", "transfer", "model", "can", "be", "stably", "learned", "by", "alternatingly", "minimizing", "and", "during", "training", ".", "Finally", ",", "the", "discriminator", "is", "learned", "by", "minimizing", "the", "following", "objective", ",", "subsection", ":", "Learning", "algorithm", "Generally", ",", "both", "the", "generator", "and", "the", "discriminator", "are", "difficult", "to", "converge", "in", "GAN", ".", "Therefore", ",", "we", "adopt", "a", "two", "-", "stage", "strategy", "for", "learning", "the", "transfer", "model", "and", "the", "discriminator", ":", "(", "i", ")", "we", "first", "combine", "the", "source", "set", "and", "the", "guided", "set", "to", "pre", "-", "train", "for", "initialization", ",", "and", "(", "ii", ")", "alternate", "between", "updating", "and", ".", "The", "procedure", "for", "training", "the", "transfer", "model", "is", "summarized", "in", "Algorithm", "[", "reference", "]", ".", "Initialization", ".", "For", "the", "initialization", "of", "the", ",", "we", "only", "consider", "the", "attribute", "transform", "network", ",", "and", "leave", "the", "mask", "network", "be", "learned", "in", "the", "second", "stage", ".", "Note", "that", "the", "transform", "network", "has", "the", "architecture", "of", "auto", "-", "encoder", ".", "Thus", ",", "it", "can", "be", "pre", "-", "trained", "by", "minimizing", "the", "following", "reconstruction", "objective", "on", "and", ",", "As", "for", "the", "initialization", "of", "the", "discriminator", ",", "we", "use", "the", "images", "in", "as", "negative", "samples", "and", "the", "images", "in", "as", "positive", "samples", ".", "Then", "the", "discriminator", "can", "be", "pre", "-", "trained", "by", "minimizing", "the", "following", "objective", ",", "where", "is", "for", "positive", "image", "and", "for", "negative", "image", ".", "By", "this", "way", ",", "the", "initialization", "can", "provide", "a", "good", "start", "point", "and", "benefit", "the", "convergence", "and", "stability", "of", "DIAT", "training", ".", "Network", "training", ".", "After", "the", "initialization", "of", "and", ",", "network", "training", "is", "further", "performed", "by", "updating", "the", "whole", "(", "including", "both", "and", ")", "and", "alternatingly", ".", "Moreover", ",", "is", "updated", "by", "first", "iterations", "for", "minimizing", "and", "then", "iterations", "for", "minimizing", ".", "We", "apply", "the", "RMSProp", "solver", "to", "train", "the", "transfer", "network", "and", "the", "discriminator", "with", "a", "learning", "rate", "of", ".", "[", "!", "tbp", "]", "Learning", "the", "attribute", "transfer", "network", "[", "1", "]", "Source", "set", ",", "guided", "set", ",", ",", ",", ".", "mini", "-", "batch", "size", "is", ".", "The", "attribute", "transfer", "network", "Pre", "-", "train", "the", "transform", "model", "by", "minimizing", "the", "objective", "in", "Eqn", ".", "(", "[", "reference", "]", ")", ".", "Pre", "-", "train", "the", "discriminator", "by", "minimizing", "the", "objective", "in", "Eqn", ".", "(", "[", "reference", "]", ")", ".", "not", "converged", "Select", "a", "mini", "-", "bath", "from", "to", "generate", "the", "transfer", "results", ",", "which", "is", "further", "combined", "with", "another", "mini", "-", "batch", "from", "to", "form", "the", "set", "for", "training", "the", "discriminator", ".", "Here", "we", "set", ".", "Use", "the", "RMSProp", "solver", "to", "update", "the", "discriminator", "with", "Eqn", ".", "(", "[", "reference", "]", ")", "using", "the", "mini", "-", "batch", ".", "Clip", "the", "parameters", "of", "the", "discriminator", ".", "Use", "the", "RMSProp", "solver", "to", "update", "by", "minimizing", "in", "Eqn", ".", "(", "[", "reference", "]", ")", ".", "Use", "the", "RMSProp", "solver", "to", "update", "by", "minimizing", "in", "Eqn", ".", "(", "[", "reference", "]", ")", ".", "Return", "the", "transfer", "network", "subsection", ":", "Extension", "to", "face", "hallucination", "Besides", "facial", "attribute", "transfer", ",", "our", "DIAT", "can", "also", "be", "extended", "to", "other", "face", "editing", "tasks", ".", "Here", "we", "use", "the", "face", "hallucination", "as", "an", "example", ".", "Face", "hallucination", "is", "undoubtedly", "a", "global", "transfer", "task", ",", "and", "thus", "we", "remove", "the", "mask", "network", "as", "well", "as", "the", "attribute", "ratio", "regularization", ",", "making", ".", "Moreover", ",", "the", "input", "image", "in", "face", "hallucination", "is", "of", "low", "resolution", "(", "LR", ")", "while", "the", "output", "image", "is", "of", "high", "resolution", "(", "HR", ")", ".", "To", "be", "consistent", "with", "attribute", "transfer", ",", "we", "super", "-", "resolve", "LR", "image", "to", "the", "size", "of", "HR", "image", "with", "the", "bicubic", "interpolator", ",", "which", "is", "taken", "as", "input", "to", ".", "Furthermore", ",", "the", "ground", "truth", "HR", "images", "can", "be", "available", "to", "guide", "the", "network", "training", "for", "face", "hallucination", ".", "Denote", "by", "the", "super", "-", "resolved", "image", "by", "bicubic", "interpolator", ",", "and", "the", "ground", "truth", "HR", "image", ".", "Then", ",", "the", "pixel", "-", "wise", "reconstruction", "loss", "is", "defined", "as", ",", "We", "further", "modify", "the", "definition", "of", "by", "removing", "the", "attribute", "ratio", "regularization", "and", "adding", "reconstruction", "loss", ",", "where", "is", "the", "tradeoff", "parameter", "for", "pixel", "-", "wise", "reconstruction", "loss", ",", "and", "we", "set", "in", "our", "experiment", ".", "Given", "the", "training", "data", ",", "the", "models", "can", "then", "be", "learned", "by", "updating", "and", "alternatingly", ".", "section", ":", "Experimental", "Results", "In", "this", "section", ",", "we", "first", "describe", "the", "experimental", "settings", ",", "including", "the", "training", "and", "testing", "data", ",", "competing", "methods", ",", "model", "and", "learning", "parameters", ".", "Experiments", "are", "then", "performed", "for", "local", "and", "global", "attribute", "transfer", ".", "Quantitative", "metrics", "and", "the", "results", "on", "real", "images", "are", "also", "reported", ".", "Moreover", ",", "we", "analyze", "the", "effect", "of", "adaptive", "perceptual", "loss", "and", "perceptual", "regularization", ".", "Finally", ",", "the", "results", "are", "reported", "to", "further", "assess", "the", "performance", "of", "our", "DIAT", "for", "face", "hallucination", ".", "The", "source", "code", "will", "be", "given", "after", "the", "publication", "of", "this", "work", ".", "subsection", ":", "Experimental", "settings", "Our", "DIAT", "models", "are", "trained", "using", "a", "subset", "of", "the", "aligned", "CelebA", "dataset", "by", "removing", "the", "images", "with", "poor", "quality", ".", "The", "size", "of", "the", "aligned", "images", "is", ".", "Due", "to", "the", "limitation", "of", "the", "GPU", "memory", ",", "we", "sample", "the", "central", "part", "of", "each", "image", "and", "resize", "it", "to", ".", "For", "each", "attribute", "transfer", "task", ",", "we", "use", "all", "the", "images", "with", "the", "reference", "attribute", "from", "training", "set", "to", "form", "the", "guided", "set", ",", "and", "randomly", "select", "10", ",", "000", "training", "images", "not", "with", "the", "reference", "attribute", "as", "the", "source", "set", ".", "After", "training", ",", "2", ",", "000", "images", "apart", "from", "the", "images", "for", "training", "are", "adopted", "to", "assess", "the", "attribute", "transfer", "performance", ".", "And", "we", "also", "test", "the", "models", "on", "other", "real", "images", "from", "the", "website", "iStock", ".", "Only", "a", "few", "methods", "have", "been", "proposed", "for", "facial", "attribute", "transfer", ".", "In", "our", "experiments", ",", "we", "compare", "our", "DIAT", "with", "the", "convolutional", "attribute", "-", "driven", "and", "identity", "-", "preserving", "model", "(", "CNIA", ")", ",", "IcGAN", "and", "VAE", "/", "GAN", "due", "to", "that", "their", "codes", "are", "available", ".", "As", "for", "IcGAN", "and", "VAE", "/", "GAN", ",", "the", "original", "image", "size", "is", "not", "the", "same", "with", "our", "DIAT", ",", "so", "we", "resize", "the", "result", "to", "the", "same", "size", "with", "DIAT", "for", "comparison", ".", "For", "the", "task", "of", "glasses", "removal", ",", "we", "can", "first", "manually", "detect", "the", "region", "of", "glasses", ",", "and", "then", "use", "some", "face", "inpainting", "methods", "(", "e", ".", "g", ".", ",", "semantic", "inpainting", ")", "to", "recover", "the", "missing", "pixels", ".", "Thus", "we", "also", "compare", "our", "DIAT", "with", "semantic", "inpainting", "for", "glasses", "removal", ".", "All", "the", "experiments", "are", "conducted", "on", "a", "computer", "with", "the", "GTX", "TitanX", "GPU", "of", "12", "GB", "memory", ".", "We", "set", "the", "parameters", "and", "for", "DIAT", ".", "For", "the", "threshold", "in", "the", "perceptual", "regularization", ",", "we", "set", "it", "to", "be", "a", "value", "in", "the", "range", "of", ".", "As", "for", "in", "the", "attribute", "ration", "regularization", ",", "we", "set", "it", "to", "be", "(", "i", ")", "for", "small", "local", "attributes", "(", "e", ".", "g", ".", ",", "mouth", ")", ",", "(", "ii", ")", "for", "large", "local", "attributes", "(", "e", ".", "g", ".", ",", "eyes", ")", ",", "and", "(", "iii", ")", "for", "global", "attributes", "(", "e", ".", "g", ".", ",", "gender", "and", "age", ")", ".", "subsection", ":", "Local", "attribute", "transfer", "We", "assess", "the", "local", "attribute", "transfer", "models", "on", "three", "tasks", ",", "i", ".", "e", ".", ",", "mouth", "open", ",", "mouth", "close", ",", "and", "eyeglasses", "removal", ".", "Fig", ".", "[", "reference", "]", "illustrates", "the", "transfer", "results", "by", "our", "DIAT", ".", "It", "can", "be", "seen", "that", "our", "DIAT", "performs", "favorably", "for", "transferring", "the", "input", "images", "to", "the", "desired", "attribute", "with", "satisfying", "visual", "quality", ".", "Benefited", "from", "the", "mask", "network", ",", "the", "results", "by", "DIAT", "can", "preserve", "more", "identity", "-", "aware", "and", "attribute", "irrelevant", "details", ".", "Moreover", ",", "when", "the", "training", "data", "are", "sufficient", ",", "it", "is", "feasible", "to", "separately", "train", "two", "DIAT", "models", "for", "reverse", "tasks", ",", "e", ".", "g", ".", ",", "one", "for", "mouth", "open", "and", "another", "for", "mouth", "close", ".", "We", "further", "compare", "our", "DIAT", "with", "three", "competing", "methods", ",", "i", ".", "e", ".", ",", "CNIA", ",", "IcGAN", "and", "VAE", "/", "GAN", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "the", "results", "by", "our", "DIAT", "are", "visually", "more", "pleasing", "than", "those", "by", "CNIA", "for", "all", "the", "three", "local", "attribute", "transfer", "tasks", ".", "In", "terms", "of", "run", "time", ",", "CNIA", "takes", "about", "seconds", "(", ")", "to", "deal", "with", "an", "image", ",", "while", "our", "DIAT", "only", "needs", ".", "Fig", ".", "[", "reference", "]", "further", "provides", "the", "results", "by", "IcGAN", ",", "VAE", "/", "GAN", ",", "and", "our", "DIAT", ".", "In", "comparison", "with", "the", "competing", "methods", ",", "our", "DIAT", "can", "well", "address", "the", "attribute", "transfer", "tasks", "while", "recovering", "more", "visual", "details", "in", "both", "attribute", "relevant", "and", "attribute", "irrelevant", "regions", ".", "Finally", ",", "for", "glasses", "removal", ",", "we", "compare", "our", "DIAT", "with", "semantic", "inpainting", ",", "and", "the", "results", "in", "Fig", ".", "[", "reference", "]", "clearly", "demonstrate", "the", "superiority", "of", "DIAT", ".", "subsection", ":", "Global", "attribute", "transfer", "We", "consider", "two", "global", "attribute", "transfer", "tasks", ",", "i", ".", "e", ".", ",", "gender", "transfer", "and", "age", "transfer", ".", "For", "gender", "transfer", ",", "we", "only", "evaluate", "the", "model", "for", "male", "-", "to", "-", "female", ".", "For", "age", "transfer", ",", "we", "only", "test", "the", "model", "for", "older", "-", "to", "-", "younger", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "transfer", "results", ",", "and", "our", "DIAT", "is", "also", "effective", "for", "global", "attribute", "transfer", ".", "Even", "gender", "transfer", "certainly", "causes", "the", "change", "of", "the", "identity", ",", "as", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "a", ")", ",", "our", "DIAT", "can", "still", "retain", "most", "identity", "-", "aware", "features", ",", "making", "the", "transfer", "result", "similar", "to", "the", "input", "image", "in", "appearance", ".", "Figs", ".", "[", "reference", "]", "and", "[", "reference", "]", "show", "the", "results", "by", "our", "DIAT", ",", "CNAI", ",", "IcGAN", "and", "VAE", "/", "GAN", ".", "Compared", "with", "the", "competing", "methods", ",", "the", "results", "by", "our", "DIAT", "well", "exhibit", "the", "desired", "attribute", ",", "and", "are", "of", "high", "visual", "quality", "with", "photo", "-", "realistic", "details", ".", "Finally", ",", "we", "also", "note", "that", "for", "gender", "transfer", "our", "DIAT", "is", "able", "of", "adjusting", "the", "hair", "length", "due", "to", "the", "introduction", "of", "adaptive", "perceptual", "loss", ".", "subsection", ":", "Quantitative", "evaluation", "Given", "each", "attribute", "transfer", "task", ",", "we", "randomly", "select", "images", "without", "the", "reference", "attribute", "from", "the", "testing", "partition", "of", "CelebA", "to", "form", "our", "testing", "set", ".", "Then", ",", "three", "groups", "of", "experiments", "are", "conducted", "to", "evaluate", "the", "transfer", "performance", "quantitatively", ":", "Attribute", "classification", ".", "For", "attribute", "transfer", ",", "it", "is", "natural", "to", "require", "the", "transfer", "result", "to", "exhibit", "the", "desired", "attribute", ".", "Thus", ",", "we", "first", "train", "a", "CNN", "-", "based", "attribute", "classifier", "(", "including", "two", "convolution", "layers", ",", "three", "residual", "blocks", "and", "two", "fully", "-", "connected", "layers", ")", "using", "the", "training", "set", "of", "CelebA.", "Given", "an", "attribute", "transfer", "task", ",", "we", "test", "the", "classification", "accuracy", "of", "the", "desired", "attribute", "for", "the", "transfer", "results", "of", "2", ",", "000", "testing", "images", ".", "Tabel", "[", "reference", "]", "lists", "the", "classification", "accuracy", "for", "five", "attribute", "transfer", "tasks", ",", "i", ".", "e", ".", ",", "mouth", "open", ",", "mouth", "close", ",", "glasses", "removal", ",", "gender", "transfer", ",", "and", "age", "transfer", ".", "It", "can", "be", "observed", "that", "our", "DIAT", "achieves", "satisfying", "accuracy", "(", "i", ".", "e", ".", ",", ")", "for", "all", "the", "tasks", ",", "indicating", "that", "the", "results", "by", "our", "DIAT", "generally", "are", "with", "the", "desired", "attribute", ".", "Identity", "verification", ".", "As", "for", "local", "attribute", "transfer", ",", "we", "also", "require", "the", "transfer", "result", "to", "preserve", "the", "identity", "of", "input", "image", ".", "Here", "we", "use", "the", "open", "source", "face", "recognition", "platform", "Openface", "for", "matching", "the", "input", "image", "with", "the", "transfer", "result", ".", "By", "setting", "the", "threshold", "be", "0.99", ",", "Table", "[", "reference", "]", "lists", "the", "identity", "verification", "accuracy", "for", "mouth", "open", ",", "mouth", "close", ",", "and", "glasses", "removal", ".", "The", "results", "demonstrate", "that", "our", "DIAT", "can", "well", "preserve", "the", "identity", "-", "aware", "feature", "for", "local", "attribute", "transfer", ".", "Image", "quality", ".", "Image", "quality", "is", "another", "crucial", "metric", "to", "assess", "attribute", "transfer", ".", "However", ",", "the", "ground", "truth", "of", "transfer", "result", "is", "unavailable", ",", "making", "it", "difficult", "to", "perform", "quantitative", "evaluation", ".", "Here", "we", "use", "a", "pair", "of", "reverse", "attribute", "transfer", "tasks", "(", "i", ".", "e", ".", ",", "mouth", "close", "and", "mouth", "open", ")", "as", "an", "example", ",", "and", "adopt", "an", "indirect", "scheme", "to", "compute", "average", "PSNR", "on", "the", "2000", "testing", "images", ".", "Specifically", ",", "we", "first", "perform", "mouth", "open", "to", "the", "images", "with", "mouth", "close", ",", "and", "then", "perform", "reverse", "mouth", "close", "to", "the", "transfer", "results", ".", "Finally", ",", "the", "input", "images", "are", "taken", "as", "ground", "truth", "and", "the", "images", "after", "two", "steps", "of", "transfer", "can", "be", "viewed", "as", "the", "generated", "images", ".", "By", "this", "way", ",", "we", "obtain", "the", "average", "PSNR", "of", "33.27dB", ",", "indicating", "the", "effectiveness", "of", "our", "DIAT", "for", "local", "attribute", "transfer", ".", "subsection", ":", "Results", "on", "other", "real", "facial", "images", "To", "assess", "the", "generalization", "ability", ",", "we", "use", "the", "DIAT", "models", "learned", "on", "CelebA", "to", "other", "real", "facial", "images", "from", "the", "website", "iStock", ".", "Each", "test", "image", "is", "first", "aligned", "with", "the", "5", "facial", "landmarks", ",", "and", "then", "input", "to", "the", "DIAT", "models", ".", "Taking", "mouth", "open", "and", "gender", "transfer", "as", "examples", ",", "Fig", ".", "[", "reference", "]", "gives", "the", "transfer", "results", "on", "15", "images", "for", "each", "task", ",", "clearly", "demonstrating", "the", "generalization", "ability", "of", "our", "models", "to", "other", "real", "facial", "images", ".", "subsection", ":", "Evaluation", "on", "adaptive", "perceptual", "loss", "and", "perceptual", "regularization", "We", "also", "implement", "a", "variant", "of", "DIAT", "(", "i", ".", "e", ".", ",", "DIAT", "-", "1", ")", "by", "replacing", "the", "adaptive", "perceptual", "loss", "with", "the", "conventional", "perceptual", "loss", "defined", "on", "VGG", "-", "Face", ".", "Taking", "gender", "transfer", "as", "an", "example", ",", "Fig", ".", "[", "reference", "]", "compares", "DIAT", "with", "DIAT", "-", "1", ".", "It", "can", "be", "observed", "that", "DIAT", "converges", "very", "fast", "and", "can", "generate", "satisfying", "results", "after", "4", "epochs", "of", "training", ".", "In", "comparison", ",", "DIAT", "-", "1", "requires", "much", "more", "epochs", "in", "training", ",", "and", "the", "gender", "just", "begins", "to", "be", "modified", "after", "18", "epochs", ".", "Moreover", ",", "the", "adoption", "of", "adaptive", "perceptual", "loss", "also", "benefits", "the", "transfer", "performance", ",", "and", "adaptive", "adjustment", "on", "the", "hair", "length", "can", "be", "observed", "on", "the", "transfer", "results", "by", "DIAT", ".", "Furthermore", ",", "Fig", ".", "[", "reference", "]", "shows", "the", "transfer", "results", "by", "DIAT", "with", "the", "perceptual", "regularization", "and", "the", "TV", "regularization", ".", "It", "can", "be", "clearly", "seen", "that", "the", "perceptual", "regularization", "is", "more", "effective", "on", "suppressing", "noise", "and", "artifacts", "while", "preserving", "sharp", "edges", "and", "fine", "details", ".", "subsection", ":", "Results", "of", "the", "learnt", "mask", "Fig", ".", "[", "reference", "]", "gives", "the", "masks", "generated", "by", "the", "mask", "network", "for", "different", "task", ".", "For", "the", "local", "attribute", "transformation", "tasks", "such", "as", "glasses", "removal", "and", "closing", "mouth", ",", "the", "generated", "masks", "accurately", "cover", "the", "local", "facial", "part", "which", "is", "related", "to", "the", "attribute", ".", "For", "global", "transformation", "like", "gender", "transformation", ",", "the", "mask", "covers", "most", "of", "the", "face", "and", "keep", "the", "background", "out", ".", "subsection", ":", "Experiments", "on", "face", "hallucination", "Finally", ",", "we", "evaluate", "the", "performance", "of", "DIAT", "for", "face", "hallucination", ".", "Table", "[", "reference", "]", "lists", "the", "average", "PSNR", "and", "SSIM", "values", "on", "the", "2", ",", "000", "testing", "images", "by", "DIAT", ",", "bicubic", "interpolator", ",", "and", "Unet", ",", "while", "Fig", ".", "[", "reference", "]", "shows", "the", "super", "-", "resolved", "images", ".", "Even", "our", "DIAT", "achieves", "lower", "PSNR", "/", "SSIM", "than", "the", "baseline", "Unet", ",", "it", "is", "much", "better", "in", "terms", "of", "visual", "quality", ",", "and", "can", "generate", "hallucinated", "image", "with", "rich", "textures", "and", "sharp", "edges", ".", "section", ":", "Conclusion", "A", "deep", "identity", "-", "aware", "transfer", "(", "i", ".", "e", ".", ",", "DIAT", ")", "model", "is", "presented", "for", "facial", "attribute", "transfer", ".", "Considering", "that", "some", "attributes", "may", "be", "only", "related", "with", "parts", "of", "facial", "image", ",", "the", "whole", "transfer", "model", "consists", "of", "two", "subnetworks", ",", "i", ".", "e", ".", ",", "mask", "network", "and", "attribute", "transform", "network", ",", "which", "work", "collaboratively", "to", "produce", "the", "transfer", "result", ".", "In", "order", "to", "train", "the", "model", ",", "we", "further", "incorporate", "adversarial", "attribute", "loss", ",", "adaptive", "perceptual", "loss", "with", "perceptual", "regularization", "and", "attribute", "ratio", "regularization", ".", "Experiments", "show", "that", "our", "model", "can", "obtain", "satisfying", "results", "for", "both", "local", "and", "global", "attribute", "transfer", ".", "Even", "for", "some", "identity", "-", "related", "attributes", "(", "e.g.", ",", "gender", "transfer", ")", ",", "our", "DIAT", "can", "obtain", "visually", "impressive", "results", "with", "minor", "modification", "on", "identity", "-", "related", "features", ".", "Our", "DIAT", "can", "also", "be", "extended", "to", "face", "hallucination", "and", "performs", "favorably", "in", "recovering", "facial", "details", ".", "In", "future", "work", ",", "we", "will", "further", "improve", "the", "visual", "quality", "and", "diversity", "of", "the", "transfer", "results", ",", "and", "extend", "our", "model", "to", "arbitrary", "attribute", "transfer", ".", "bibliography", ":", "References"], "entity": [[2, 10, ["Deep", "Identity", "-", "Aware", "Transfer", "of", "Facial", "Attributes"], "Task"], [14, 18, ["Deep", "convolutional", "network", "model"], "Method"], [19, 23, ["Identity", "-", "Aware", "Transfer"], "Task"], [24, 25, ["DIAT"], "Task"], [40, 41, ["DIAT"], "Task"], [74, 76, ["mask", "network"], "Method"], [78, 81, ["attribute", "transform", "network"], "Method"], [114, 116, ["mask", "network"], "Method"], [148, 153, ["joint", "training", "of", "transform", "network"], "Task"], [154, 156, ["mask", "network"], "Method"], [164, 169, ["identity", "aware", "adaptive", "perceptual", "loss"], "Method"], [170, 176, ["VGG", "-", "FACE", "based", "identity", "loss"], "Method"], [180, 182, ["denoising", "network"], "Method"], [187, 189, ["perceptual", "regularization"], "Task"], [199, 202, ["attribute", "ratio", "regularization"], "Method"], [214, 215, ["DIAT"], "Task"], [223, 227, ["facial", "attribute", "transfer", "tasks"], "Task"], [233, 235, ["expression", "transfer"], "Task"], [236, 238, ["accessory", "removal"], "Task"], [239, 241, ["age", "progression"], "Task"], [242, 244, ["gender", "transfer"], "Task"], [251, 254, ["face", "enhancement", "tasks"], "Task"], [256, 258, ["face", "hallucination"], "Task"], [287, 288, ["DIAT"], "Task"], [305, 308, ["Facial", "attribute", "transfer"], "Task"], [309, 312, ["generative", "adversarial", "nets"], "Method"], [313, 315, ["convolutional", "networks"], "Method"], [316, 318, ["perceptual", "loss"], "Task"], [354, 359, ["facial", "animation", "and", "manipulation", "tasks"], "Task"], [366, 370, ["facial", "avatar", "and", "reenactment"], "Task"], [391, 393, ["expression", "transfer"], "Task"], [394, 396, ["accessory", "removal"], "Task"], [397, 399, ["age", "progression"], "Task"], [401, 402, ["animation"], "Method"], [456, 460, ["facial", "attribute", "transfer", "tasks"], "Task"], [468, 470, ["expression", "transfer"], "Task"], [477, 483, ["3D", "or", "image", "-", "based", "avatars"], "Task"], [490, 492, ["face", "trackers"], "Method"], [493, 495, ["expression", "modeling"], "Method"], [499, 504, ["offline", "and", "online", "facial", "reenactment"], "Task"], [506, 508, ["age", "progression"], "Task"], [509, 514, ["explicit", "and", "implicit", "synthesis", "methods"], "Method"], [519, 521, ["image", "models"], "Method"], [522, 525, ["Hair", "style", "generation"], "Method"], [526, 527, ["replacement"], "Task"], [534, 537, ["Convolutional", "neural", "network"], "Method"], [538, 539, ["CNN"], "Method"], [547, 552, ["human", "face", "generation", "with", "attributes"], "Task"], [558, 563, ["deep", "convolution", "inverse", "graphic", "network"], "Method"], [564, 567, ["DG", "-", "IGN"], "Method"], [595, 596, ["Gauthier"], "Method"], [598, 602, ["conditional", "generative", "adversarial", "network"], "Method"], [603, 604, ["cGAN"], "Method"], [623, 631, ["attribute", "-", "conditioned", "deep", "variational", "auto", "-", "encoder"], "Method"], [652, 654, ["generative", "model"], "Method"], [661, 665, ["conditional", "image", "generation", "model"], "Method"], [667, 669, ["PixelCNN", "decoder"], "Method"], [670, 672, ["image", "generation"], "Task"], [698, 700, ["attribute", "transfer"], "Task"], [707, 708, ["CNN"], "Method"], [710, 712, ["complex", "transformation"], "Task"], [724, 727, ["facial", "attribute", "transfer"], "Task"], [734, 740, ["CNN", "-", "based", "attribute", "transfer", "model"], "Method"], [742, 744, ["optimization", "perspective"], "Task"], [747, 749, ["run", "time"], "Metric"], [750, 752, ["transfer", "quality"], "Metric"], [767, 769, ["supervised", "learning"], "Task"], [771, 774, ["generative", "adversarial", "net"], "Method"], [775, 776, ["GAN"], "Method"], [817, 821, ["Identity", "-", "Aware", "Transfer"], "Task"], [822, 823, ["DIAT"], "Task"], [835, 840, ["facial", "animation", "and", "manipulation", "tasks"], "Task"], [846, 848, ["expression", "transfer"], "Task"], [849, 851, ["accessory", "removal"], "Task"], [852, 854, ["age", "progression"], "Task"], [856, 858, ["gender", "transfer"], "Task"], [914, 916, ["mask", "network"], "Method"], [918, 921, ["attribute", "transform", "network"], "Method"], [923, 926, ["attribute", "transform", "network"], "Method"], [941, 943, ["mask", "network"], "Method"], [967, 970, ["attribute", "transform", "network"], "Method"], [971, 973, ["mask", "network"], "Method"], [978, 982, ["photo", "-", "realistic", "transfer"], "Task"], [985, 987, ["attribute", "transfer"], "Task"], [989, 992, ["ground", "truth", "transfer"], "Task"], [1030, 1035, ["large", "-", "scale", "CelebFaces", "Attributes"], "Material"], [1036, 1037, ["CelebA"], "Material"], [1065, 1066, ["CelebA."], "Material"], [1116, 1118, ["unsupervised", "training"], "Task"], [1136, 1139, ["attribute", "ratio", "regularization"], "Method"], [1152, 1154, ["attribute", "transfer"], "Task"], [1178, 1180, ["attribute", "discriminator"], "Method"], [1193, 1197, ["identity", "-", "aware", "transfer"], "Task"], [1199, 1200, ["DIAT"], "Task"], [1222, 1227, ["identity", "-", "aware", "perceptual", "loss"], "Task"], [1231, 1234, ["convolutional", "feature", "map"], "Method"], [1256, 1260, ["pre", "-", "trained", "CNNs"], "Method"], [1270, 1272, ["attribute", "discriminator"], "Method"], [1281, 1283, ["learning", "procedure"], "Method"], [1295, 1297, ["perceptual", "loss"], "Method"], [1303, 1304, ["computation"], "Task"], [1309, 1312, ["attribute", "transfer", "task"], "Task"], [1320, 1324, ["hidden", "-", "layer", "supervision"], "Method"], [1325, 1326, ["regularization"], "Method"], [1353, 1355, ["feature", "representations"], "Method"], [1366, 1369, ["VGG", "-", "Face"], "Method"], [1378, 1380, ["face", "verification"], "Task"], [1385, 1386, ["DIAT"], "Task"], [1389, 1390, ["regularizers"], "Method"], [1396, 1398, ["perceptual", "regularization"], "Method"], [1399, 1402, ["attribute", "ratio", "regularization"], "Method"], [1411, 1413, ["denoising", "network"], "Method"], [1417, 1419, ["perceptual", "regularization"], "Method"], [1429, 1431, ["mask", "network"], "Method"], [1433, 1436, ["attribute", "ratio", "regularization"], "Method"], [1460, 1463, ["adversarial", "attribute", "loss"], "Method"], [1464, 1467, ["adaptive", "perceptual", "loss"], "Method"], [1468, 1470, ["perceptual", "regularization"], "Method"], [1471, 1474, ["attribute", "ratio", "regularization"], "Method"], [1480, 1481, ["CelebA"], "Material"], [1482, 1484, ["real", "images"], "Material"], [1486, 1488, ["website", "iStock"], "Material"], [1499, 1500, ["DIAT"], "Task"], [1503, 1505, ["attribute", "transfer"], "Task"], [1535, 1536, ["DIAT"], "Task"], [1540, 1541, ["transfer"], "Task"], [1550, 1552, ["Computational", "efficiency"], "Metric"], [1567, 1568, ["DIAT"], "Task"], [1588, 1590, ["face", "hallucination"], "Task"], [1596, 1602, ["photo", "-", "realistic", "high", "resolution", "images"], "Task"], [1632, 1635, ["facial", "attribute", "transfer"], "Task"], [1648, 1650, ["mask", "network"], "Method"], [1652, 1655, ["attribute", "transform", "network"], "Method"], [1680, 1682, ["identity", "loss"], "Method"], [1683, 1685, ["perceptual", "regularization"], "Method"], [1687, 1690, ["attribute", "ratio", "regularization"], "Method"], [1723, 1724, ["discriminator"], "Method"], [1725, 1729, ["identity", "-", "aware", "transfer"], "Task"], [1731, 1733, ["training", "efficiency"], "Metric"], [1736, 1738, ["perceptual", "regularization"], "Method"], [1739, 1742, ["attribute", "ratio", "regularization"], "Method"], [1752, 1754, ["mask", "network"], "Method"], [1766, 1771, ["identity", "-", "aware", "attribute", "transfer"], "Task"], [1773, 1774, ["DIAT"], "Task"], [1779, 1780, ["transfer"], "Task"], [1799, 1801, ["age", "progression"], "Task"], [1821, 1823, ["face", "hallucination"], "Task"], [1869, 1872, ["facial", "attribute", "transfer"], "Task"], [1873, 1875, ["face", "hallucination"], "Task"], [1894, 1898, ["Deep", "convolutional", "neural", "networks"], "Method"], [1899, 1900, ["CNNs"], "Method"], [1909, 1913, ["high", "level", "vision", "problems"], "Task"], [1923, 1928, ["generating", ",", "and", "recovering", "images"], "Task"], [1939, 1942, ["facial", "attribute", "transfer"], "Task"], [1950, 1952, ["image", "generation"], "Task"], [1953, 1955, ["face", "generation"], "Task"], [1958, 1959, ["CNN"], "Method"], [1960, 1962, ["image", "generation"], "Task"], [1962, 1965, ["Generative", "image", "modeling"], "Method"], [1970, 1972, ["image", "generation"], "Task"], [1974, 1978, ["low", "level", "vision", "problems"], "Task"], [1980, 1993, ["sparse", ",", "low", "rank", ",", "FRAME", "and", "non", "-", "local", "similarity", "based", "models"], "Method"], [2010, 2012, ["image", "modeling"], "Task"], [2025, 2029, ["convolutional", "auto", "-", "encoder"], "Method"], [2030, 2031, ["PixelCNN"], "Method"], [2032, 2033, ["PixelRNN"], "Method"], [2040, 2044, ["image", "completion", "and", "generation"], "Task"], [2052, 2054, ["image", "generation"], "Task"], [2055, 2058, ["Fully", "convolutional", "networks"], "Method"], [2063, 2066, ["supervised", "learning", "manner"], "Method"], [2081, 2083, ["convolution", "layers"], "Method"], [2085, 2088, ["fully", "connected", "layers"], "Method"], [2101, 2106, ["Deep", "Convolution", "Inverse", "Graphics", "Network"], "Method"], [2107, 2110, ["DC", "-", "IGN"], "Method"], [2115, 2118, ["variational", "autoencoder", "architecture"], "Method"], [2136, 2139, ["DC", "-", "IGN"], "Method"], [2163, 2165, ["deep", "representation"], "Method"], [2167, 2169, ["class", "score"], "Metric"], [2185, 2187, ["deep", "representation"], "Method"], [2194, 2195, ["CNNs"], "Method"], [2196, 2199, ["artistic", "style", "transfer"], "Task"], [2213, 2216, ["iterative", "optimization", "procedure"], "Method"], [2233, 2235, ["style", "transfer"], "Task"], [2237, 2239, ["generation", "tasks"], "Task"], [2246, 2252, ["identity", "-", "aware", "adaptive", "perceptual", "loss"], "Method"], [2253, 2255, ["perceptual", "regularization"], "Method"], [2266, 2269, ["facial", "attribute", "transfer"], "Task"], [2274, 2277, ["generative", "adversarial", "network"], "Method"], [2278, 2279, ["GAN"], "Method"], [2283, 2284, ["discriminator"], "Method"], [2286, 2287, ["generator"], "Method"], [2292, 2294, ["adversarial", "game"], "Method"], [2296, 2297, ["generator"], "Method"], [2309, 2310, ["discriminator"], "Method"], [2348, 2349, ["GAN"], "Method"], [2351, 2355, ["fully", "deep", "convolutional", "networks"], "Method"], [2364, 2366, ["image", "generation"], "Task"], [2369, 2371, ["disentangled", "representations"], "Method"], [2398, 2399, ["WGAN"], "Method"], [2400, 2403, ["WGAN", "-", "GP"], "Method"], [2432, 2434, ["WGAN", "framework"], "Method"], [2443, 2446, ["adaptive", "perceptual", "loss"], "Method"], [2447, 2451, ["identity", "-", "aware", "transfer"], "Task"], [2452, 2454, ["perceptual", "regularization"], "Task"], [2461, 2462, ["CNN"], "Method"], [2463, 2465, ["face", "generation"], "Task"], [2465, 2468, ["Facial", "attribute", "transfer"], "Task"], [2481, 2485, ["variational", "autoencode", "with", "GAN"], "Method"], [2486, 2487, ["VAE"], "Method"], [2488, 2489, ["GAN"], "Method"], [2491, 2493, ["image", "generation"], "Task"], [2504, 2507, ["mean", "latent", "representations"], "Method"], [2517, 2518, ["VAE"], "Method"], [2519, 2520, ["GAN"], "Method"], [2526, 2530, ["arbitrary", "facial", "attribute", "transfer"], "Task"], [2535, 2536, ["transfer"], "Task"], [2544, 2553, ["attribute", "driven", "and", "identity", "-", "preserving", "face", "generation", "model"], "Method"], [2556, 2558, ["optimization", "problems"], "Task"], [2580, 2584, ["encoder", "-", "decoder", "architecture"], "Method"], [2586, 2588, ["attribute", "transfer"], "Task"], [2594, 2596, ["latent", "representation"], "Method"], [2612, 2614, ["dual", "learning"], "Method"], [2617, 2621, ["reverse", "attribute", "transfer", "models"], "Method"], [2632, 2634, ["object", "transfiguration"], "Task"], [2656, 2658, ["transfer", "results"], "Metric"], [2689, 2690, ["CNNs"], "Method"], [2696, 2699, ["face", "generation", "tasks"], "Task"], [2701, 2707, ["painting", "style", "transfer", "of", "head", "portrait"], "Task"], [2741, 2743, ["perceptual", "loss"], "Method"], [2744, 2746, ["sketch", "inversion"], "Task"], [2752, 2753, ["DCGAN"], "Method"], [2754, 2757, ["semantic", "face", "inpainting"], "Task"], [2759, 2761, ["optimization", "manner"], "Method"], [2764, 2766, ["Deep", "CNNs"], "Method"], [2767, 2772, ["Identity", "-", "aware", "Attribute", "Transfer"], "Task"], [2782, 2789, ["identity", "-", "aware", "transfer", "of", "facial", "attribute"], "Task"], [2803, 2805, ["mask", "network"], "Method"], [2807, 2810, ["attribute", "transform", "network"], "Method"], [2834, 2836, ["perceptual", "regularization"], "Method"], [2837, 2840, ["attribute", "ratio", "regularization"], "Method"], [2843, 2845, ["Network", "architecture"], "Method"], [2917, 2919, ["attribute", "transfer"], "Task"], [2958, 2960, ["transfer", "task"], "Task"], [2985, 2987, ["attribute", "transfer"], "Task"], [3005, 3007, ["transfer", "task"], "Task"], [3015, 3018, ["attribute", "transfer", "network"], "Method"], [3031, 3033, ["mask", "network"], "Method"], [3034, 3037, ["attribute", "transform", "network"], "Method"], [3039, 3041, ["mask", "network"], "Method"], [3042, 3045, ["attribute", "transform", "network"], "Method"], [3053, 3055, ["mask", "network"], "Method"], [3070, 3073, ["attribute", "transform", "network"], "Method"], [3107, 3110, ["attribute", "transform", "network"], "Method"], [3111, 3113, ["mask", "network"], "Method"], [3137, 3140, ["attribute", "transform", "network"], "Method"], [3141, 3143, ["mask", "network"], "Method"], [3146, 3149, ["Attribute", "transform", "network"], "Method"], [3153, 3154, ["Unet"], "Method"], [3155, 3157, ["attribute", "transform"], "Task"], [3163, 3164, ["efficiency"], "Metric"], [3165, 3167, ["reconstruction", "ability"], "Metric"], [3172, 3174, ["Unet", "architecture"], "Method"], [3176, 3178, ["encoder", "subnetwork"], "Method"], [3180, 3182, ["decoder", "subnetwork"], "Method"], [3184, 3186, ["skip", "connection"], "Method"], [3187, 3189, ["pooling", "operation"], "Method"], [3201, 3203, ["attribute", "transform"], "Task"], [3207, 3211, ["10", "-", "layer", "Unet"], "Method"], [3215, 3217, ["convolution", "layers"], "Method"], [3218, 3219, ["encoding"], "Task"], [3222, 3224, ["convolution", "layers"], "Method"], [3225, 3226, ["decoding"], "Task"], [3229, 3230, ["encoder"], "Method"], [3233, 3237, ["convolution", "with", "stride", "2"], "Method"], [3238, 3239, ["downsampling"], "Task"], [3242, 3243, ["decoder"], "Method"], [3255, 3256, ["upsampling"], "Method"], [3259, 3264, ["element", "-", "wise", "summation", "operation"], "Method"], [3273, 3277, ["encoder", "and", "decoder", "subnetworks"], "Method"], [3283, 3286, ["attribute", "transform", "network"], "Method"], [3294, 3296, ["Mask", "network"], "Method"], [3299, 3301, ["mask", "network"], "Method"], [3306, 3312, ["5", "-", "layer", "fully", "convolutional", "network"], "Method"], [3325, 3328, ["batch", "normalization", "layer"], "Method"], [3332, 3334, ["convolution", "layer"], "Method"], [3337, 3338, ["upsampling"], "Method"], [3364, 3366, ["Gaussian", "filter"], "Method"], [3385, 3387, ["mask", "network"], "Method"], [3397, 3399, ["mask", "network"], "Method"], [3400, 3401, ["ReLU"], "Method"], [3416, 3418, ["convolution", "layer"], "Method"], [3422, 3424, ["sigmoid", "nonlinearity"], "Method"], [3427, 3429, ["binarization", "operation"], "Method"], [3494, 3498, ["piecewise", "linear", "proxy", "function"], "Method"], [3502, 3503, ["training"], "Task"], [3508, 3512, ["forward", "-", "propagation", "calculation"], "Task"], [3517, 3520, ["back", "-", "propagation"], "Task"], [3541, 3544, ["attribute", "transform", "network"], "Method"], [3545, 3547, ["mask", "network"], "Method"], [3560, 3562, ["attribute", "transfer"], "Task"], [3582, 3584, ["gender", "transfer"], "Task"], [3645, 3650, ["identity", "-", "aware", "attribute", "transfer"], "Task"], [3715, 3717, ["perceptual", "regularization"], "Method"], [3720, 3722, ["denoising", "network"], "Method"], [3729, 3732, ["attribute", "ratio", "regularization"], "Method"], [3738, 3742, ["learning", "of", "mask", "network"], "Method"], [3758, 3761, ["Adversarial", "attribute", "loss"], "Task"], [3762, 3764, ["Adversarial", "strategy"], "Method"], [3773, 3775, ["security", "problems"], "Task"], [3777, 3779, ["computer", "vision"], "Task"], [3781, 3784, ["adversarial", "learning", "framework"], "Method"], [3785, 3788, ["generative", "adversarial", "networks"], "Method"], [3789, 3790, ["GANs"], "Method"], [3857, 3859, ["attribute", "distribution"], "Task"], [3865, 3868, ["adversarial", "attribute", "loss"], "Method"], [3890, 3894, ["generative", "adversarial", "network", "framework"], "Method"], [3897, 3898, ["generator"], "Method"], [3900, 3903, ["attribute", "transfer", "network"], "Method"], [3906, 3907, ["discriminator"], "Method"], [3938, 3942, ["fully", "-", "connected", "layers"], "Method"], [3976, 3977, ["discriminator"], "Method"], [3995, 3996, ["generator"], "Method"], [3998, 3999, ["discriminator"], "Method"], [4015, 4017, ["training", "stability"], "Metric"], [4042, 4043, ["generator"], "Method"], [4044, 4045, ["discriminator"], "Method"], [4091, 4095, ["identity", "-", "aware", "loss"], "Task"], [4113, 4117, ["squared", "-", "error", "loss"], "Metric"], [4119, 4121, ["feature", "representations"], "Method"], [4123, 4124, ["discriminator"], "Method"], [4128, 4131, ["adaptive", "perceptual", "loss"], "Method"], [4135, 4136, ["discriminator"], "Method"], [4143, 4147, ["-", "th", "convolution", "layer"], "Method"], [4184, 4190, ["identity", "-", "aware", "adaptive", "perceptual", "loss"], "Method"], [4199, 4200, ["discriminator"], "Method"], [4241, 4247, ["off", "-", "the", "-", "shelf", "CNNs"], "Method"], [4253, 4256, ["VGG", "-", "Face"], "Material"], [4261, 4263, ["perceptual", "loss"], "Method"], [4265, 4268, ["adaptive", "perceptual", "loss"], "Method"], [4275, 4277, ["training", "efficiency"], "Metric"], [4278, 4281, ["attribute", "transfer", "performance"], "Metric"], [4283, 4285, ["training", "efficiency"], "Metric"], [4286, 4289, ["adaptive", "perceptual", "loss"], "Method"], [4302, 4304, ["perceptual", "loss"], "Task"], [4306, 4310, ["forward", "and", "backward", "calculations"], "Task"], [4323, 4324, ["discriminator"], "Method"], [4325, 4326, ["training"], "Task"], [4338, 4339, ["discriminator"], "Method"], [4356, 4357, ["DIAT"], "Task"], [4360, 4361, ["training"], "Task"], [4367, 4368, ["GAN"], "Method"], [4370, 4371, ["generator"], "Method"], [4381, 4382, ["DIAT"], "Task"], [4390, 4393, ["adversarial", "attribute", "loss"], "Method"], [4394, 4397, ["adaptive", "perceptual", "loss"], "Method"], [4401, 4403, ["training", "speed"], "Metric"], [4421, 4422, ["discriminator"], "Method"], [4437, 4441, ["convergence", "of", "network", "training"], "Task"], [4444, 4446, ["perceptual", "loss"], "Task"], [4448, 4454, ["off", "-", "the", "-", "shelf", "CNNs"], "Method"], [4468, 4470, ["attribute", "transfer"], "Task"], [4476, 4479, ["VGG", "-", "Face"], "Method"], [4487, 4489, ["face", "recognition"], "Task"], [4495, 4500, ["identity", "-", "aware", "attribute", "transfer"], "Task"], [4505, 4508, ["adaptive", "perceptual", "loss"], "Method"], [4512, 4513, ["discriminator"], "Method"], [4517, 4518, ["modeling"], "Task"], [4530, 4532, ["attribute", "transfer"], "Task"], [4534, 4535, ["transfer"], "Task"], [4543, 4545, ["gender", "transfer"], "Task"], [4563, 4565, ["Identity", "Loss"], "Task"], [4568, 4571, ["adaptive", "perceptual", "loss"], "Method"], [4609, 4611, ["feature", "representation"], "Method"], [4613, 4617, ["face", "recognition", "or", "verification"], "Task"], [4619, 4622, ["face", "verification", "task"], "Task"], [4646, 4649, ["VGG", "-", "Face"], "Method"], [4668, 4670, ["Perceptual", "regularization"], "Task"], [4675, 4678, ["adversarial", "attribute", "loss"], "Method"], [4693, 4695, ["Image", "regularization"], "Task"], [4717, 4719, ["Total", "Variation"], "Method"], [4720, 4721, ["TV"], "Method"], [4732, 4735, ["artistic", "style", "transfer"], "Task"], [4744, 4750, ["recovering", "small", "-", "scale", "texture", "details"], "Task"], [4786, 4788, ["denoising", "network"], "Method"], [4789, 4791, ["perceptual", "regularization"], "Task"], [4795, 4797, ["denoising", "network"], "Method"], [4818, 4819, ["CelebA."], "Material"], [4821, 4823, ["residual", "learning"], "Method"], [4827, 4829, ["denoising", "network"], "Method"], [4850, 4852, ["denoising", "network"], "Method"], [4854, 4860, ["fully", "convolutional", "network", "of", "6", "layers"], "Method"], [4866, 4867, ["denoising"], "Task"], [4907, 4908, ["learning"], "Task"], [4914, 4916, ["denoising", "network"], "Method"], [4946, 4947, ["Minimizing"], "Method"], [5009, 5011, ["denoising", "network"], "Method"], [5017, 5019, ["perceptual", "regularization"], "Method"], [5020, 5023, ["Attribute", "ratio", "regularization"], "Method"], [5056, 5058, ["glasses", "removal"], "Task"], [5074, 5076, ["gender", "transfer"], "Task"], [5094, 5098, ["attribute", "ratio", "regularization", "term"], "Method"], [5110, 5111, ["regularization"], "Method"], [5171, 5173, ["Objective", "function"], "Metric"], [5177, 5179, ["objective", "function"], "Metric"], [5182, 5184, ["transfer", "model"], "Method"], [5186, 5187, ["discriminator"], "Method"], [5194, 5197, ["adaptive", "perceptual", "loss"], "Method"], [5198, 5200, ["perceptual", "regularization"], "Method"], [5202, 5205, ["attribute", "ratio", "regularization"], "Method"], [5207, 5209, ["transfer", "model"], "Method"], [5215, 5216, ["objective"], "Metric"], [5232, 5235, ["attribute", "ratio", "regularization"], "Method"], [5260, 5262, ["transfer", "model"], "Method"], [5267, 5269, ["alternatingly", "minimizing"], "Method"], [5276, 5277, ["discriminator"], "Method"], [5283, 5284, ["objective"], "Metric"], [5287, 5289, ["Learning", "algorithm"], "Method"], [5293, 5294, ["generator"], "Method"], [5296, 5297, ["discriminator"], "Method"], [5302, 5303, ["GAN"], "Method"], [5311, 5313, ["stage", "strategy"], "Method"], [5316, 5318, ["transfer", "model"], "Method"], [5320, 5321, ["discriminator"], "Method"], [5340, 5341, ["initialization"], "Task"], [5356, 5358, ["transfer", "model"], "Method"], [5366, 5367, ["Initialization"], "Task"], [5378, 5381, ["attribute", "transform", "network"], "Method"], [5385, 5387, ["mask", "network"], "Method"], [5397, 5399, ["transform", "network"], "Method"], [5403, 5406, ["auto", "-", "encoder"], "Method"], [5419, 5421, ["reconstruction", "objective"], "Metric"], [5430, 5431, ["discriminator"], "Method"], [5450, 5451, ["discriminator"], "Method"], [5487, 5490, ["convergence", "and", "stability"], "Metric"], [5494, 5496, ["Network", "training"], "Method"], [5503, 5505, ["network", "training"], "Method"], [5528, 5529, ["minimizing"], "Task"], [5533, 5534, ["minimizing"], "Task"], [5538, 5540, ["RMSProp", "solver"], "Method"], [5543, 5545, ["transfer", "network"], "Method"], [5547, 5548, ["discriminator"], "Method"], [5550, 5552, ["learning", "rate"], "Metric"], [5560, 5563, ["attribute", "transfer", "network"], "Method"], [5582, 5585, ["attribute", "transfer", "network"], "Method"], [5589, 5591, ["transform", "model"], "Method"], [5594, 5595, ["objective"], "Metric"], [5608, 5609, ["discriminator"], "Method"], [5612, 5613, ["objective"], "Metric"], [5614, 5615, ["Eqn"], "Metric"], [5653, 5654, ["discriminator"], "Method"], [5661, 5663, ["RMSProp", "solver"], "Method"], [5666, 5667, ["discriminator"], "Method"], [5677, 5680, ["mini", "-", "batch"], "Method"], [5686, 5687, ["discriminator"], "Method"], [5690, 5692, ["RMSProp", "solver"], "Method"], [5707, 5709, ["RMSProp", "solver"], "Method"], [5724, 5726, ["transfer", "network"], "Method"], [5730, 5732, ["face", "hallucination"], "Task"], [5733, 5736, ["facial", "attribute", "transfer"], "Task"], [5738, 5739, ["DIAT"], "Task"], [5745, 5748, ["face", "editing", "tasks"], "Task"], [5753, 5755, ["face", "hallucination"], "Task"], [5759, 5761, ["Face", "hallucination"], "Task"], [5764, 5767, ["global", "transfer", "task"], "Task"], [5773, 5775, ["mask", "network"], "Method"], [5779, 5782, ["attribute", "ratio", "regularization"], "Method"], [5791, 5793, ["face", "hallucination"], "Task"], [5798, 5799, ["LR"], "Method"], [5816, 5818, ["attribute", "transfer"], "Task"], [5833, 5835, ["bicubic", "interpolator"], "Method"], [5856, 5858, ["network", "training"], "Method"], [5859, 5861, ["face", "hallucination"], "Task"], [5870, 5872, ["bicubic", "interpolator"], "Method"], [5883, 5888, ["pixel", "-", "wise", "reconstruction", "loss"], "Metric"], [5901, 5904, ["attribute", "ratio", "regularization"], "Method"], [5915, 5920, ["pixel", "-", "wise", "reconstruction", "loss"], "Metric"], [5961, 5965, ["training", "and", "testing", "data"], "Material"], [5971, 5973, ["learning", "parameters"], "Method"], [5979, 5984, ["local", "and", "global", "attribute", "transfer"], "Task"], [5985, 5987, ["Quantitative", "metrics"], "Metric"], [5991, 5993, ["real", "images"], "Material"], [6008, 6010, ["perceptual", "regularization"], "Method"], [6024, 6025, ["DIAT"], "Task"], [6026, 6028, ["face", "hallucination"], "Task"], [6099, 6102, ["attribute", "transfer", "task"], "Task"], [6157, 6160, ["attribute", "transfer", "performance"], "Metric"], [6169, 6171, ["real", "images"], "Material"], [6173, 6175, ["website", "iStock"], "Material"], [6184, 6187, ["facial", "attribute", "transfer"], "Task"], [6195, 6196, ["DIAT"], "Task"], [6198, 6207, ["convolutional", "attribute", "-", "driven", "and", "identity", "-", "preserving", "model"], "Method"], [6208, 6209, ["CNIA"], "Method"], [6211, 6212, ["IcGAN"], "Method"], [6213, 6214, ["VAE"], "Method"], [6215, 6216, ["GAN"], "Method"], [6226, 6227, ["IcGAN"], "Method"], [6228, 6229, ["VAE"], "Method"], [6230, 6231, ["GAN"], "Method"], [6242, 6243, ["DIAT"], "Task"], [6262, 6264, ["glasses", "removal"], "Task"], [6279, 6282, ["face", "inpainting", "methods"], "Method"], [6288, 6290, ["semantic", "inpainting"], "Task"], [6302, 6303, ["DIAT"], "Task"], [6304, 6306, ["semantic", "inpainting"], "Task"], [6307, 6309, ["glasses", "removal"], "Task"], [6320, 6323, ["GTX", "TitanX", "GPU"], "Method"], [6341, 6343, ["perceptual", "regularization"], "Task"], [6360, 6363, ["attribute", "ration", "regularization"], "Task"], [6421, 6424, ["Local", "attribute", "transfer"], "Method"], [6427, 6431, ["local", "attribute", "transfer", "models"], "Method"], [6447, 6449, ["eyeglasses", "removal"], "Task"], [6461, 6462, ["DIAT"], "Task"], [6469, 6470, ["DIAT"], "Task"], [6483, 6485, ["visual", "quality"], "Metric"], [6489, 6491, ["mask", "network"], "Method"], [6495, 6496, ["DIAT"], "Task"], [6526, 6528, ["reverse", "tasks"], "Task"], [6548, 6549, ["DIAT"], "Task"], [6563, 6564, ["VAE"], "Method"], [6565, 6566, ["GAN"], "Method"], [6580, 6581, ["DIAT"], "Task"], [6588, 6589, ["CNIA"], "Method"], [6593, 6597, ["local", "attribute", "transfer", "tasks"], "Task"], [6601, 6603, ["run", "time"], "Metric"], [6604, 6605, ["CNIA"], "Method"], [6618, 6619, ["DIAT"], "Task"], [6632, 6633, ["IcGAN"], "Method"], [6634, 6635, ["VAE"], "Method"], [6636, 6637, ["GAN"], "Method"], [6640, 6641, ["DIAT"], "Task"], [6650, 6651, ["DIAT"], "Task"], [6655, 6658, ["attribute", "transfer", "tasks"], "Task"], [6675, 6677, ["glasses", "removal"], "Task"], [6681, 6682, ["DIAT"], "Task"], [6683, 6685, ["semantic", "inpainting"], "Task"], [6700, 6701, ["DIAT"], "Task"], [6704, 6707, ["Global", "attribute", "transfer"], "Method"], [6710, 6714, ["global", "attribute", "transfer", "tasks"], "Task"], [6720, 6722, ["gender", "transfer"], "Task"], [6723, 6725, ["age", "transfer"], "Task"], [6727, 6729, ["gender", "transfer"], "Task"], [6736, 6741, ["male", "-", "to", "-", "female"], "Task"], [6743, 6745, ["age", "transfer"], "Task"], [6770, 6771, ["DIAT"], "Task"], [6775, 6778, ["global", "attribute", "transfer"], "Task"], [6780, 6782, ["gender", "transfer"], "Task"], [6803, 6804, ["DIAT"], "Task"], [6839, 6840, ["DIAT"], "Task"], [6841, 6842, ["CNAI"], "Method"], [6843, 6844, ["IcGAN"], "Method"], [6845, 6846, ["VAE"], "Method"], [6847, 6848, ["GAN"], "Method"], [6859, 6860, ["DIAT"], "Task"], [6870, 6872, ["visual", "quality"], "Metric"], [6885, 6887, ["gender", "transfer"], "Task"], [6888, 6889, ["DIAT"], "Task"], [6907, 6909, ["Quantitative", "evaluation"], "Metric"], [6911, 6914, ["attribute", "transfer", "task"], "Task"], [6928, 6929, ["CelebA"], "Material"], [6946, 6948, ["transfer", "performance"], "Metric"], [6950, 6952, ["Attribute", "classification"], "Task"], [6954, 6956, ["attribute", "transfer"], "Task"], [6985, 6987, ["convolution", "layers"], "Method"], [6993, 6997, ["fully", "-", "connected", "layers"], "Method"], [7003, 7004, ["CelebA."], "Material"], [7006, 7009, ["attribute", "transfer", "task"], "Task"], [7013, 7015, ["classification", "accuracy"], "Metric"], [7036, 7038, ["classification", "accuracy"], "Metric"], [7040, 7043, ["attribute", "transfer", "tasks"], "Task"], [7058, 7060, ["gender", "transfer"], "Task"], [7062, 7064, ["age", "transfer"], "Task"], [7071, 7072, ["DIAT"], "Task"], [7074, 7075, ["accuracy"], "Metric"], [7093, 7094, ["DIAT"], "Task"], [7101, 7103, ["Identity", "verification"], "Task"], [7106, 7109, ["local", "attribute", "transfer"], "Task"], [7128, 7134, ["open", "source", "face", "recognition", "platform", "Openface"], "Method"], [7157, 7160, ["identity", "verification", "accuracy"], "Metric"], [7168, 7170, ["glasses", "removal"], "Task"], [7176, 7177, ["DIAT"], "Task"], [7186, 7189, ["local", "attribute", "transfer"], "Task"], [7190, 7192, ["Image", "quality"], "Metric"], [7193, 7195, ["Image", "quality"], "Metric"], [7201, 7203, ["attribute", "transfer"], "Task"], [7220, 7222, ["quantitative", "evaluation"], "Task"], [7229, 7233, ["reverse", "attribute", "transfer", "tasks"], "Task"], [7252, 7254, ["indirect", "scheme"], "Method"], [7256, 7258, ["average", "PSNR"], "Metric"], [7306, 7307, ["transfer"], "Task"], [7332, 7333, ["DIAT"], "Task"], [7334, 7337, ["local", "attribute", "transfer"], "Task"], [7349, 7351, ["generalization", "ability"], "Metric"], [7359, 7360, ["CelebA"], "Material"], [7367, 7369, ["website", "iStock"], "Material"], [7391, 7393, ["mouth", "open"], "Task"], [7433, 7436, ["adaptive", "perceptual", "loss"], "Method"], [7437, 7439, ["perceptual", "regularization"], "Task"], [7445, 7446, ["DIAT"], "Task"], [7459, 7462, ["adaptive", "perceptual", "loss"], "Method"], [7469, 7472, ["VGG", "-", "Face"], "Material"], [7474, 7476, ["gender", "transfer"], "Task"], [7486, 7487, ["DIAT"], "Task"], [7497, 7498, ["DIAT"], "Task"], [7548, 7549, ["transfer"], "Task"], [7552, 7554, ["adaptive", "adjustment"], "Method"], [7566, 7567, ["DIAT"], "Task"], [7580, 7581, ["DIAT"], "Task"], [7583, 7585, ["perceptual", "regularization"], "Method"], [7597, 7599, ["perceptual", "regularization"], "Method"], [7633, 7635, ["mask", "network"], "Method"], [7641, 7645, ["local", "attribute", "transformation", "tasks"], "Task"], [7647, 7649, ["glasses", "removal"], "Task"], [7650, 7652, ["closing", "mouth"], "Task"], [7670, 7672, ["global", "transformation"], "Task"], [7673, 7675, ["gender", "transformation"], "Task"], [7693, 7695, ["face", "hallucination"], "Task"], [7702, 7703, ["DIAT"], "Task"], [7704, 7706, ["face", "hallucination"], "Task"], [7713, 7715, ["average", "PSNR"], "Metric"], [7716, 7718, ["SSIM", "values"], "Metric"], [7726, 7727, ["DIAT"], "Task"], [7728, 7730, ["bicubic", "interpolator"], "Method"], [7732, 7733, ["Unet"], "Method"], [7749, 7750, ["DIAT"], "Task"], [7752, 7753, ["PSNR"], "Metric"], [7754, 7755, ["SSIM"], "Metric"], [7758, 7759, ["Unet"], "Method"], [7767, 7769, ["visual", "quality"], "Metric"], [7786, 7791, ["deep", "identity", "-", "aware", "transfer"], "Method"], [7797, 7798, ["DIAT"], "Task"], [7803, 7806, ["facial", "attribute", "transfer"], "Task"], [7823, 7825, ["transfer", "model"], "Method"], [7835, 7837, ["mask", "network"], "Method"], [7838, 7841, ["attribute", "transform", "network"], "Method"], [7865, 7868, ["adaptive", "perceptual", "loss"], "Method"], [7869, 7871, ["perceptual", "regularization"], "Method"], [7872, 7875, ["attribute", "ratio", "regularization"], "Method"], [7887, 7892, ["local", "and", "global", "attribute", "transfer"], "Task"], [7903, 7905, ["gender", "transfer"], "Task"], [7908, 7909, ["DIAT"], "Task"], [7924, 7925, ["DIAT"], "Task"], [7930, 7932, ["face", "hallucination"], "Task"], [7936, 7939, ["recovering", "facial", "details"], "Task"], [7949, 7951, ["visual", "quality"], "Metric"], [7963, 7966, ["arbitrary", "attribute", "transfer"], "Task"], [160, 163, ["adversarial", "attribute", "loss"], "Method"], [814, 815, ["CNN"], "Method"], [868, 869, ["CNN"], "Method"], [1008, 1009, ["GAN"], "Method"], [1103, 1106, ["adversarial", "attribute", "loss"], "Method"], [1120, 1121, ["DIAT"], "Task"], [1147, 1148, ["DIAT"], "Task"], [1167, 1168, ["GAN"], "Method"], [1173, 1176, ["adversarial", "attribute", "loss"], "Method"], [1236, 1237, ["CNN"], "Method"], [1288, 1291, ["adaptive", "perceptual", "loss"], "Method"], [1332, 1333, ["DIAT"], "Task"], [1450, 1451, ["DIAT"], "Task"], [1627, 1628, ["DIAT"], "Task"], [1676, 1679, ["adaptive", "perceptual", "loss"], "Method"], [1695, 1696, ["DIAT"], "Task"], [1699, 1702, ["adversarial", "attribute", "loss"], "Method"], [1716, 1719, ["adaptive", "perceptual", "loss"], "Method"], [1857, 1858, ["DIAT"], "Task"], [1947, 1948, ["CNN"], "Method"], [2016, 2017, ["CNN"], "Method"], [2046, 2047, ["CNN"], "Method"], [2077, 2078, ["CNN"], "Method"], [2134, 2135, ["CNN"], "Method"], [2150, 2151, ["CNN"], "Method"], [2223, 2224, ["CNN"], "Method"], [2259, 2260, ["DIAT"], "Task"], [2325, 2326, ["GANs"], "Method"], [2377, 2378, ["GAN"], "Method"], [2423, 2424, ["generator"], "Method"], [2437, 2438, ["DIAT"], "Task"], [2607, 2608, ["GAN"], "Method"], [2674, 2675, ["DIAT"], "Task"], [2739, 2740, ["CNN"], "Method"], [2779, 2780, ["DIAT"], "Task"], [2826, 2829, ["adversarial", "attribute", "loss"], "Method"], [2830, 2833, ["adaptive", "perceptual", "loss"], "Method"], [3660, 3663, ["adversarial", "attribute", "loss"], "Method"], [3677, 3680, ["adaptive", "perceptual", "loss"], "Method"], [3827, 3830, ["adversarial", "attribute", "loss"], "Method"], [3912, 3915, ["adversarial", "attribute", "loss"], "Method"], [3920, 3921, ["discriminator"], "Method"], [4006, 4009, ["adversarial", "attribute", "loss"], "Method"], [4021, 4022, ["GAN"], "Method"], [4053, 4056, ["adaptive", "perceptual", "loss"], "Method"], [4220, 4221, ["discriminator"], "Method"], [4227, 4230, ["adaptive", "perceptual", "loss"], "Method"], [4320, 4321, ["CNN"], "Method"], [4331, 4334, ["adaptive", "perceptual", "loss"], "Method"], [4352, 4353, ["discriminator"], "Method"], [4407, 4410, ["adaptive", "perceptual", "loss"], "Method"], [4549, 4552, ["adaptive", "perceptual", "loss"], "Method"], [4555, 4557, ["adaptive", "adjustment"], "Method"], [4679, 4682, ["adaptive", "perceptual", "loss"], "Method"], [4728, 4729, ["CNN"], "Method"], [4739, 4740, ["TV"], "Method"], [5190, 5193, ["adversarial", "attribute", "loss"], "Method"], [5228, 5231, ["adaptive", "perceptual", "loss"], "Method"], [5491, 5492, ["DIAT"], "Task"], [5823, 5824, ["LR"], "Method"], [6004, 6007, ["adaptive", "perceptual", "loss"], "Method"], [6047, 6048, ["DIAT"], "Task"], [6057, 6058, ["CelebA"], "Material"], [6254, 6255, ["DIAT"], "Task"], [6334, 6335, ["DIAT"], "Task"], [6523, 6524, ["DIAT"], "Task"], [6559, 6560, ["CNIA"], "Method"], [6561, 6562, ["IcGAN"], "Method"], [6901, 6904, ["adaptive", "perceptual", "loss"], "Method"], [6977, 6978, ["CNN"], "Method"], [7322, 7324, ["average", "PSNR"], "Metric"], [7355, 7356, ["DIAT"], "Task"], [7387, 7388, ["DIAT"], "Task"], [7452, 7453, ["DIAT"], "Task"], [7488, 7489, ["DIAT"], "Task"], [7515, 7516, ["DIAT"], "Task"], [7542, 7545, ["adaptive", "perceptual", "loss"], "Method"], [7587, 7588, ["TV"], "Method"], [7861, 7864, ["adversarial", "attribute", "loss"], "Method"]]} {"doc_id": "2c1d608e3cd43b3104d76296a14f1aef0e6b2c64", "text": ["document", ":", "Dynamic", "Evaluation", "of", "Neural", "Sequence", "Models", "We", "present", "methodology", "for", "using", "dynamic", "evaluation", "to", "improve", "neural", "sequence", "models", ".", "Models", "are", "adapted", "to", "recent", "history", "via", "a", "gradient", "descent", "based", "mechanism", ",", "causing", "them", "to", "assign", "higher", "probabilities", "to", "re", "-", "occurring", "sequential", "patterns", ".", "Dynamic", "evaluation", "outperforms", "existing", "adaptation", "approaches", "in", "our", "comparisons", ".", "Dynamic", "evaluation", "improves", "the", "state", "-", "of", "-", "the", "-", "art", "word", "-", "level", "perplexities", "on", "the", "Penn", "Treebank", "and", "WikiText", "-", "2", "datasets", "to", "51.1", "and", "44.3", "respectively", ",", "and", "the", "state", "-", "of", "-", "the", "-", "art", "character", "-", "level", "cross", "-", "entropies", "on", "the", "text8", "and", "Hutter", "Prize", "datasets", "to", "1.19", "bits", "/", "char", "and", "1.08", "bits", "/", "char", "respectively", ".", "section", ":", "Introduction", "Sequence", "generation", "and", "prediction", "tasks", "span", "many", "modes", "of", "data", ",", "ranging", "from", "audio", "and", "language", "modelling", ",", "to", "more", "general", "timeseries", "prediction", "tasks", ".", "Applications", "of", "such", "models", "include", "speech", "recognition", ",", "machine", "translation", ",", "dialogue", "generation", ",", "speech", "synthesis", ",", "forecasting", ",", "and", "music", "generation", ",", "among", "others", ".", "Neural", "networks", "can", "be", "applied", "to", "these", "tasks", "by", "predicting", "sequence", "elements", "one", "-", "by", "-", "one", ",", "conditioning", "on", "the", "history", "of", "sequence", "elements", ",", "forming", "an", "autoregressive", "model", ".", "Convolutional", "neural", "networks", "(", "CNNs", ")", "and", "recurrent", "neural", "networks", "(", "RNNs", ")", ",", "including", "long", "-", "short", "term", "memory", "(", "LSTM", ")", "networks", "in", "particular", ",", "have", "achieved", "many", "successes", "at", "these", "tasks", ".", "However", ",", "in", "their", "basic", "form", ",", "these", "models", "have", "a", "limited", "ability", "to", "adapt", "to", "recently", "observed", "parts", "of", "a", "sequence", ".", "Many", "sequences", "contain", "repetition", ";", "a", "pattern", "that", "occurs", "once", "is", "more", "likely", "to", "occur", "again", ".", "For", "instance", ",", "a", "word", "that", "occurs", "once", "in", "a", "document", "is", "much", "more", "likely", "to", "occur", "again", ".", "A", "sequence", "of", "handwriting", "will", "generally", "stay", "in", "the", "same", "handwriting", "style", ".", "A", "sequence", "of", "speech", "will", "generally", "stay", "in", "the", "same", "voice", ".", "Although", "RNNs", "have", "a", "hidden", "state", "that", "can", "summarize", "the", "recent", "past", ",", "they", "are", "often", "unable", "to", "exploit", "new", "patterns", "that", "occur", "repeatedly", "in", "a", "test", "sequence", ".", "This", "paper", "concerns", "dynamic", "evaluation", ",", "which", "we", "investigate", "as", "a", "candidate", "solution", "to", "this", "problem", ".", "Our", "approach", "adapts", "models", "to", "recent", "sequences", "using", "gradient", "descent", "based", "mechanisms", ".", "We", "show", "several", "ways", "to", "improve", "on", "past", "dynamic", "evaluation", "approaches", "in", "Section", "[", "reference", "]", ",", "and", "use", "our", "improved", "methodology", "to", "achieve", "state", "-", "of", "-", "the", "-", "art", "results", "in", "Section", "[", "reference", "]", "and", "Section", "[", "reference", "]", ".", "In", "Section", "[", "reference", "]", "we", "design", "a", "method", "to", "dramatically", "to", "reduce", "the", "number", "of", "adaptation", "parameters", "in", "dynamic", "evaluation", ",", "making", "it", "practical", "in", "a", "wider", "range", "of", "situations", ".", "In", "Section", "[", "reference", "]", "we", "analyse", "dynamic", "evaluation", "’s", "performance", "over", "varying", "time", "-", "scales", "and", "distribution", "shifts", ",", "and", "demonstrate", "that", "dynamically", "evaluated", "models", "can", "generate", "conditional", "samples", "that", "repeat", "many", "patterns", "from", "the", "conditioning", "data", ".", "section", ":", "Motivation", "Generative", "models", "can", "assign", "probabilities", "to", "sequences", "by", "modelling", "each", "term", "in", "the", "factorization", "given", "by", "the", "product", "rule", ".", "The", "probability", "of", "a", "sequence", "factorizes", "as", "Methods", "that", "apply", "this", "factorization", "either", "use", "a", "fixed", "context", "when", "predicting", ",", "for", "instance", "as", "in", "N", "-", "grams", "or", "CNNs", ",", "or", "use", "a", "recurrent", "hidden", "state", "to", "summarize", "the", "context", ",", "as", "in", "an", "RNN", ".", "However", ",", "for", "longer", "sequences", ",", "the", "history", "often", "contains", "re", "-", "occurring", "patterns", "that", "are", "difficult", "to", "capture", "using", "models", "with", "fixed", "parameters", "(", "static", "models", ")", ".", "In", "many", "domains", ",", "in", "a", "dataset", "of", "sequences", ",", "each", "sequence", "is", "generated", "from", "a", "slightly", "different", "distribution", ".", "At", "any", "point", "in", "time", ",", "the", "history", "of", "a", "sequence", "contains", "useful", "information", "about", "the", "generating", "distribution", "for", "that", "specific", "sequence", ".", "Therefore", "adapting", "the", "model", "parameters", "learned", "during", "training", "is", "justified", ".", "We", "aim", "to", "infer", "a", "set", "of", "model", "parameters", "from", "that", "will", "better", "approximate", "within", "sequence", ".", "Many", "sequence", "modelling", "tasks", "are", "characterised", "by", "sequences", "generated", "from", "slightly", "different", "distributions", "as", "in", "the", "scenario", "described", "above", ".", "The", "generating", "distribution", "may", "also", "change", "continuously", "across", "a", "single", "sequence", ";", "for", "instance", ",", "a", "text", "excerpt", "may", "change", "topic", ".", "Furthermore", ",", "many", "machine", "learning", "benchmarks", "do", "not", "distinguish", "between", "sequence", "boundaries", ",", "and", "concatenate", "all", "sequences", "into", "one", "continuous", "sequence", ".", "Thus", ",", "many", "sequence", "modelling", "tasks", "could", "be", "seen", "as", "having", "a", "local", "distribution", "as", "well", "as", "a", "global", "distribution", ".", "During", "training", "time", ",", "the", "goal", "is", "to", "find", "the", "best", "fixed", "model", "possible", "for", ".", "However", ",", "during", "evaluation", "time", ",", "a", "model", "that", "can", "infer", "the", "current", "from", "the", "recent", "history", "has", "an", "advantage", ".", "section", ":", "Dynamic", "evaluation", "Dynamic", "evaluation", "methods", "continuously", "adapt", "the", "model", "parameters", ",", "learned", "at", "training", "time", ",", "to", "parts", "of", "a", "sequence", "during", "evaluation", ".", "The", "goal", "is", "to", "learn", "adapted", "parameters", "that", "provide", "a", "better", "model", "of", "the", "local", "sequence", "distribution", ",", ".", "When", "dynamic", "evaluation", "is", "applied", "in", "the", "present", "work", ",", "a", "long", "test", "sequence", "is", "divided", "up", "into", "shorter", "sequences", "of", "length", ".", "We", "define", "to", "be", "a", "sequence", "of", "shorter", "sequence", "segments", "The", "initial", "adapted", "parameters", "are", "set", "to", ",", "and", "used", "to", "compute", "the", "probability", "of", "the", "first", "segment", ",", ".", "This", "probability", "gives", "a", "cross", "entropy", "loss", ",", "with", "gradient", ",", "which", "is", "computed", "using", "truncated", "back", "-", "propagation", "through", "time", ".", "The", "gradient", "is", "used", "to", "update", "the", "model", ",", "resulting", "in", "adapted", "parameters", ",", "before", "evaluating", ".", "The", "same", "procedure", "is", "then", "repeated", "for", ",", "and", "for", "each", "in", "the", "sequence", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "Gradients", "for", "each", "loss", "are", "only", "backpropagated", "to", "the", "beginning", "of", ",", "so", "computation", "is", "linear", "in", "the", "sequence", "length", ".", "Each", "update", "applies", "one", "maximum", "likelihood", "training", "step", "to", "approximate", "the", "current", "local", "distribution", ".", "The", "computational", "cost", "of", "dynamic", "evaluation", "is", "one", "forward", "pass", "and", "one", "gradient", "computation", "through", "the", "data", ",", "with", "some", "slight", "overhead", "to", "apply", "the", "update", "rule", "for", "every", "sequence", "segment", ".", "As", "in", "all", "autoregressive", "models", ",", "dynamic", "evaluation", "only", "conditions", "on", "sequence", "elements", "that", "it", "has", "already", "predicted", ",", "and", "so", "evaluates", "a", "valid", "log", "-", "probability", "for", "each", "sequence", ".", "Dynamic", "evaluation", "can", "also", "be", "used", "while", "generating", "sequences", ".", "In", "this", "case", ",", "the", "model", "generates", "each", "sequence", "segment", "using", "fixed", "weights", ",", "and", "performs", "a", "gradient", "descent", "based", "update", "step", "on", ".", "Applying", "dynamic", "evaluation", "for", "sequence", "generation", "could", "result", "in", "generated", "sequences", "with", "more", "consistent", "regularities", ",", "meaning", "that", "patterns", "that", "occur", "in", "the", "generated", "sequence", "are", "more", "likely", "to", "occur", "again", ".", "section", ":", "Background", "subsection", ":", "Related", "approaches", "Adaptive", "language", "modelling", "was", "first", "considered", "for", "n", "-", "grams", ",", "adapting", "to", "recent", "history", "via", "caching", ",", "and", "other", "methods", ".", "More", "recently", ",", "the", "neural", "cache", "approach", "and", "the", "closely", "related", "pointer", "sentinel", "-", "LSTM", "have", "been", "used", "to", "for", "adaptive", "neural", "language", "modelling", ".", "Neural", "caching", "has", "recently", "been", "used", "to", "improve", "the", "state", "-", "of", "-", "the", "-", "art", "at", "word", "-", "level", "language", "modelling", ".", "The", "neural", "cache", "model", "learns", "a", "type", "of", "non", "-", "parametric", "output", "layer", "on", "the", "fly", "at", "test", "time", ",", "which", "allows", "the", "network", "to", "adapt", "to", "recent", "observations", ".", "Each", "past", "hidden", "state", "is", "paired", "with", "the", "next", "input", ",", "and", "is", "stored", "as", "a", "tuple", ".", "When", "a", "new", "hidden", "state", "is", "observed", ",", "the", "output", "probabilities", "are", "adjusted", "to", "give", "a", "higher", "weight", "to", "output", "words", "that", "coincided", "with", "past", "hidden", "states", "with", "a", "large", "inner", "product", ".", "where", "is", "a", "one", "hot", "encoding", "of", ",", "and", "is", "a", "scaling", "parameter", ".", "The", "cache", "probabilities", "are", "interpolated", "with", "the", "base", "network", "probabilities", "to", "adapt", "the", "base", "network", "at", "test", "time", ".", "The", "neural", "cache", "closely", "relates", "to", "dynamic", "evaluation", ",", "as", "both", "methods", "can", "be", "added", "on", "top", "of", "a", "base", "model", "for", "adaptation", "at", "test", "time", ".", "The", "main", "difference", "is", "the", "mechanism", "used", "to", "fit", "to", "recent", "history", ":", "the", "neural", "cache", "approach", "uses", "a", "non", "-", "parametric", ",", "nearest", "neighbours", "-", "like", "method", ",", "whereas", "dynamic", "evaluation", "uses", "a", "gradient", "descent", "based", "method", "to", "change", "model", "parameters", "dynamically", ".", "Both", "methods", "rely", "on", "an", "autoregressive", "factorisation", ",", "as", "they", "depend", "on", "observing", "sequence", "elements", "after", "they", "are", "predicted", "in", "order", "to", "perform", "adaptation", ".", "Dynamic", "evaluation", "and", "neural", "caching", "methods", "are", "therefore", "both", "applicable", "to", "sequence", "prediction", "and", "generation", "tasks", ",", "but", "not", "directly", "to", "more", "general", "supervised", "learning", "tasks", ".", "One", "drawback", "of", "the", "neural", "cache", "method", "is", "that", "it", "can", "not", "adjust", "the", "recurrent", "hidden", "state", "dynamics", ".", "As", "a", "result", ",", "the", "neural", "cache", "’s", "ability", "to", "capture", "information", "that", "occurs", "jointly", "between", "successive", "sequence", "elements", "is", "limited", ".", "This", "capability", "is", "critical", "for", "adapting", "to", "sequences", "where", "each", "element", "has", "very", "little", "independent", "meaning", ",", "e.g.", "character", "level", "language", "modelling", ".", "subsection", ":", "Dynamic", "evaluation", "in", "neural", "networks", "Dynamic", "evaluation", "of", "neural", "language", "models", "was", "proposed", "by", ".", "Their", "approach", "simply", "used", "stochastic", "gradient", "descent", "(", "SGD", ")", "updates", "at", "every", "time", "step", ",", "computing", "the", "gradient", "with", "fully", "truncated", "backpropagation", "through", "time", ",", "which", "is", "equivalent", "to", "setting", "in", "equation", "(", "[", "reference", "]", ")", ".", "Dynamic", "evaluation", "has", "since", "been", "applied", "to", "character", "and", "word", "-", "level", "language", "models", ".", "Previous", "work", "using", "dynamic", "evaluation", "considered", "it", "as", "an", "aside", ",", "and", "did", "not", "explore", "it", "in", "depth", ".", "section", ":", "Update", "Rule", "methodology", "for", "Dynamic", "evaluation", "We", "propose", "several", "changes", "to", "’s", "dynamic", "evaluation", "method", "with", "SGD", "and", "fully", "truncated", "backpropagation", ",", "which", "we", "refer", "to", "as", "traditional", "dynamic", "evaluation", ".", "The", "first", "modification", "reduces", "the", "update", "frequency", ",", "so", "that", "gradients", "are", "backpropagated", "over", "more", "timesteps", ".", "This", "change", "provides", "more", "accurate", "gradient", "information", ",", "and", "also", "improves", "the", "computational", "efficiency", "of", "dynamic", "evaluation", ",", "since", "the", "update", "rule", "is", "applied", "much", "less", "often", ".", "We", "use", "sequence", "segments", "of", "length", "5", "for", "word", "-", "level", "tasks", "and", "20", "for", "character", "-", "level", "tasks", ".", "Next", ",", "we", "add", "a", "global", "decay", "prior", "to", "bias", "the", "model", "towards", "the", "parameters", "learned", "during", "training", ".", "Our", "motivation", "for", "dynamic", "evaluation", "assumes", "that", "the", "local", "generating", "distribution", "is", "constantly", "changing", ",", "so", "it", "is", "potentially", "desirable", "to", "weight", "recent", "sequence", "history", "higher", "in", "adaptation", ".", "Adding", "a", "global", "decay", "prior", "accomplishes", "this", "by", "causing", "previous", "adaptation", "updates", "to", "decay", "exponentially", "over", "time", ".", "For", "SGD", "with", "a", "global", "prior", ",", "learning", "rate", "and", "decay", "rate", ";", "we", "form", "the", "update", "rule", "We", "then", "consider", "using", "an", "RMSprop", "derived", "update", "rule", "for", "the", "learning", "rule", "in", "place", "of", "SGD", ".", "RMSprop", "uses", "a", "moving", "average", "of", "recent", "squared", "gradients", "to", "scale", "learning", "rates", "for", "each", "weight", ".", "In", "dynamic", "evaluation", ",", "near", "the", "start", "of", "a", "test", "sequence", ",", "RMSprop", "has", "had", "very", "few", "gradients", "to", "average", ",", "and", "therefore", "may", "not", "be", "able", "to", "leverage", "its", "updates", "as", "effectively", ".", "For", "this", "reason", ",", "we", "collect", "mean", "squared", "gradients", ",", ",", "on", "the", "training", "data", "rather", "than", "on", "recent", "test", "data", "(", "which", "is", "what", "RMSprop", "would", "do", ")", ".", "is", "given", "by", "where", "is", "the", "number", "of", "training", "batches", "and", "is", "the", "gradient", "on", "the", "th", "training", "batch", ".", "The", "mini", "-", "batch", "size", "for", "this", "computation", "becomes", "a", "hyper", "-", "parameter", ",", "as", "larger", "mini", "-", "batches", "will", "result", "in", "smaller", "mean", "squared", "gradients", ".", "The", "update", "rule", ",", "which", "we", "call", "RMS", "with", "a", "global", "prior", "in", "our", "experiments", ",", "is", "then", "where", "is", "a", "stabilization", "parameter", ".", "For", "the", "decay", "step", "of", "our", "update", "rule", ",", "we", "also", "consider", "scaling", "the", "decay", "rate", "for", "each", "parameter", "proportionally", "to", ".", "Parameters", "with", "a", "high", "RMS", "gradient", "affect", "the", "dynamics", "of", "the", "network", "more", ",", "so", "it", "makes", "sense", "to", "decay", "them", "faster", ".", "is", "divided", "by", "its", "mean", ",", "resulting", "in", "a", "normalized", "version", "of", "with", "a", "mean", "of", "1", ":", "We", "clip", "the", "values", "of", "to", "be", "no", "greater", "than", "to", "be", "sure", "that", "the", "decay", "rate", "does", "not", "exceed", "for", "any", "parameter", ".", "Combining", "the", "learning", "component", "and", "the", "regularization", "component", "results", "in", "the", "final", "update", "equation", ",", "which", "we", "refer", "to", "as", "RMS", "with", "an", "RMS", "global", "prior", "section", ":", "Sparse", "dynamic", "evaluation", "Mini", "-", "batching", "over", "sequences", "is", "desirable", "for", "some", "test", "-", "time", "sequence", "modelling", "applications", "because", "it", "allows", "faster", "processing", "of", "multiple", "sequences", "in", "parallel", ".", "Dynamic", "evaluation", "has", "a", "high", "memory", "cost", "for", "mini", "-", "batching", "because", "it", "is", "necessary", "to", "store", "a", "different", "set", "of", "parameters", "for", "each", "sequence", "in", "the", "mini", "-", "batch", ".", "Therefore", ",", "we", "consider", "a", "sparse", "dynamic", "evaluation", "variant", "that", "updates", "a", "smaller", "number", "of", "parameters", ".", "We", "introduce", "a", "new", "adaptation", "matrix", "which", "is", "initialized", "to", "zeros", ".", "multiplies", "hidden", "state", "vector", "of", "an", "RNN", "at", "every", "time", "-", "step", "to", "get", "a", "new", "hidden", "state", ",", "via", "then", "replaces", "and", "is", "propagated", "throughout", "the", "network", "via", "both", "recurrent", "and", "feed", "-", "forward", "connections", ".", "Applying", "dynamic", "evaluation", "to", "avoids", "the", "need", "to", "apply", "dynamic", "evaluation", "to", "the", "original", "parameters", "of", "the", "network", ",", "reduces", "the", "number", "of", "adaptation", "parameters", ",", "and", "makes", "mini", "-", "batching", "less", "memory", "intensive", ".", "We", "reduce", "the", "number", "of", "adaptation", "parameters", "further", "by", "only", "using", "to", "transform", "an", "arbitrary", "subset", "of", "hidden", "units", ".", "This", "results", "in", "being", "an", "matrix", "with", "adaptation", "parameters", ".", "If", "is", "chosen", "to", "be", "much", "less", "than", "the", "number", "of", "hidden", "units", ",", "this", "reduces", "the", "number", "of", "adaptation", "parameters", "dramatically", ".", "In", "Section", "[", "reference", "]", "we", "experiment", "with", "sparse", "dynamic", "evaluation", "for", "character", "-", "level", "language", "models", ".", "section", ":", "Experiments", "We", "applied", "dynamic", "evaluation", "to", "word", "-", "level", "and", "character", "-", "level", "language", "modelling", ".", "In", "all", "tasks", ",", "we", "evaluate", "dynamic", "evaluation", "on", "top", "of", "a", "base", "model", ".", "After", "training", "the", "base", "model", ",", "we", "tune", "hyper", "-", "parameters", "for", "dynamic", "evaluation", "on", "the", "validation", "set", ",", "and", "evaluate", "both", "the", "static", "and", "dynamic", "versions", "of", "the", "model", "on", "the", "test", "set", ".", "We", "also", "consider", "follow", "up", "experiments", "that", "analyse", "the", "sequence", "lengths", "for", "which", "dynamic", "evaluation", "is", "useful", ".", "Code", "for", "our", "dynamic", "evaluation", "methodology", "is", "available", ".", "subsection", ":", "Word", "-", "level", "language", "modelling", "We", "train", "base", "models", "on", "the", "Penn", "Treebank", "and", "WikiText", "-", "2", "datasets", ",", "and", "compare", "the", "performance", "of", "static", "and", "dynamic", "evaluation", ".", "These", "experiments", "compare", "dynamic", "evaluation", "against", "past", "approaches", "such", "as", "the", "neural", "cache", "and", "measure", "dynamic", "evaluation", "’s", "general", "performance", "across", "different", "models", "and", "datasets", ".", "PTB", "is", "derived", "from", "articles", "of", "the", "Wall", "Street", "Journal", ".", "It", "contains", "929k", "training", "tokens", "and", "a", "vocab", "size", "limited", "to", "10k", "words", ".", "It", "is", "one", "of", "the", "most", "commonly", "used", "benchmarks", "in", "language", "modelling", ".", "We", "consider", "two", "baseline", "models", "on", "PTB", ",", "a", "standard", "LSTM", "implementation", "with", "recurrent", "dropout", ",", "and", "the", "recent", "state", "-", "of", "-", "the", "-", "art", "AWD", "-", "LSTM", ".", "Our", "standard", "LSTM", "was", "taken", "from", "the", "Chainer", "tutorial", "on", "language", "modelling", ",", "and", "used", "two", "LSTM", "layers", "with", "650", "units", "each", ",", "trained", "with", "SGD", "and", "regularized", "with", "recurrent", "dropout", ".", "On", "our", "standard", "LSTM", ",", "we", "experiment", "with", "traditional", "dynamic", "evaluation", "as", "applied", "by", ",", "as", "well", "as", "each", "modification", "we", "make", "building", "up", "to", "our", "final", "update", "rule", "as", "described", "in", "Section", "[", "reference", "]", ".", "As", "our", "final", "update", "rule", "(", "RMS", "+", "RMS", "global", "prior", ")", "worked", "best", ",", "we", "use", "this", "for", "all", "other", "experiments", "and", "use", "‘", "‘", "dynamic", "eval", "’", "’", "by", "default", "to", "refer", "to", "this", "update", "rule", "in", "tables", ".", "We", "applied", "dynamic", "evaluation", "on", "a", "more", "powerful", "model", ",", "the", "ASGD", "weight", "-", "dropped", "LSTM", ".", "The", "AWD", "-", "LSTM", "is", "a", "vanilla", "LSTM", "that", "combines", "the", "use", "of", "drop", "-", "connect", "on", "recurrent", "weights", "for", "regularization", ",", "and", "a", "variant", "of", "averaged", "stochastic", "gradient", "descent", "for", "optimisation", ".", "Our", "model", ",", "which", "used", "3", "layers", "and", "tied", "input", "and", "output", "embeddings", ",", "was", "intended", "to", "be", "a", "direct", "replication", "of", "AWD", "-", "LSTM", ",", "using", "code", "from", "their", "implementation", ".", "Results", "are", "given", "in", "Table", "[", "reference", "]", ".", "Dynamic", "evaluation", "gives", "significant", "overall", "improvements", "to", "both", "models", "on", "this", "dataset", ".", "Dynamic", "evaluation", "also", "achieves", "better", "final", "results", "than", "the", "neural", "cache", "on", "both", "a", "standard", "LSTM", "and", "the", "AWD", "-", "LSTM", "reimplementation", ",", "and", "improves", "the", "state", "-", "of", "-", "the", "-", "art", "on", "PTB", ".", "WikiText", "-", "2", "is", "roughly", "twice", "the", "size", "of", "PTB", ",", "with", "2", "million", "training", "tokens", "and", "a", "vocab", "size", "of", "33k", ".", "It", "features", "articles", "in", "a", "non", "-", "shuffled", "order", ",", "with", "dependencies", "across", "articles", "that", "adaptive", "methods", "should", "be", "able", "to", "exploit", ".", "For", "this", "dataset", ",", "we", "use", "the", "same", "baseline", "LSTM", "implementation", "and", "AWD", "-", "LSTM", "re", "-", "implementation", "as", "on", "PTB", ".", "Results", "are", "given", "in", "Table", "[", "reference", "]", ".", "Dynamic", "evaluation", "improves", "the", "state", "-", "of", "-", "the", "-", "art", "perplexity", "on", "WikiText", "-", "2", ",", "and", "provides", "a", "significantly", "greater", "improvement", "than", "neural", "caching", "to", "both", "base", "models", ".", "This", "suggests", "that", "dynamic", "evaluation", "is", "effective", "at", "exploiting", "regularities", "that", "co", "-", "occur", "across", "non", "-", "shuffled", "documents", ".", "subsection", ":", "Character", "-", "level", "language", "modelling", "We", "consider", "dynamic", "evaluation", "on", "the", "text8", ",", "and", "Hutter", "Prize", "datasets", ".", "The", "Hutter", "Prize", "dataset", "is", "comprised", "of", "Wikipedia", "text", ",", "and", "includes", "XML", "and", "characters", "from", "non", "-", "Latin", "languages", ".", "It", "is", "100", "million", "UTF", "-", "8", "bytes", "long", "and", "contains", "205", "unique", "bytes", ".", "Similarly", "to", "other", "reported", "results", ",", "we", "use", "a", "90", "-", "5", "-", "5", "split", "for", "training", ",", "validation", ",", "and", "testing", ".", "The", "text8", "dataset", "is", "also", "derived", "from", "Wikipedia", "text", ",", "but", "has", "all", "XML", "removed", ",", "and", "is", "lower", "cased", "to", "only", "have", "26", "characters", "of", "English", "text", "plus", "spaces", ".", "As", "with", "Hutter", "Prize", ",", "we", "use", "the", "standard", "90", "-", "5", "-", "5", "split", "for", "training", ",", "validation", ",", "and", "testing", "for", "text8", ".", "We", "used", "a", "multiplicative", "LSTM", "(", "mLSTM", ")", "as", "our", "base", "model", "for", "both", "datasets", ".", "The", "mLSTMs", "for", "both", "tasks", "used", "2800", "hidden", "units", ",", "an", "embedding", "layer", "of", "400", "units", ",", "weight", "normalization", ",", "variational", "dropout", ",", "and", "ADAM", "for", "training", ".", "We", "also", "consider", "sparse", "dynamic", "evaluation", ",", "as", "described", "in", "Section", "[", "reference", "]", ",", "on", "the", "Hutter", "Prize", "dataset", ".", "For", "sparse", "dynamic", "evaluation", ",", "we", "adapted", "a", "subset", "of", "hidden", "units", ",", "resulting", "in", "a", "adaptation", "matrix", "and", "250k", "adaptation", "parameters", ".", "All", "of", "our", "dynamic", "evaluation", "results", "in", "this", "section", "use", "the", "final", "update", "rule", "given", "in", "Section", "[", "reference", "]", ".", "Results", "for", "Hutter", "Prize", "are", "given", "in", "Table", "[", "reference", "]", ",", "and", "results", "for", "text8", "are", "given", "in", "Table", "[", "reference", "]", ".", "Dynamic", "evaluation", "achieves", "large", "improvements", "to", "our", "base", "models", "and", "state", "-", "of", "-", "the", "-", "art", "results", "on", "both", "datasets", ".", "Sparse", "dynamic", "evaluation", "also", "achieves", "significant", "improvements", "on", "Hutter", "Prize", "using", "only", "0.5", "%", "of", "the", "adaptation", "parameters", "of", "regular", "dynamic", "evaluation", ".", "subsection", ":", "Time", "-", "scales", "of", "dynamic", "evaluation", "We", "measure", "time", "-", "scales", "at", "which", "dynamic", "evaluation", "gains", "an", "advantage", "over", "static", "evaluation", ".", "Starting", "from", "the", "model", "trained", "on", "Hutter", "Prize", ",", "we", "plot", "the", "performance", "of", "static", "and", "dynamic", "evaluation", "against", "the", "number", "of", "characters", "processed", "on", "sequences", "from", "the", "Hutter", "Prize", "test", "set", ",", "and", "sequences", "in", "Spanish", "from", "the", "European", "Parliament", "dataset", ".", "The", "Hutter", "Prize", "data", "experiments", "show", "the", "timescales", "at", "which", "dynamic", "evaluation", "gained", "the", "advantage", "observed", "in", "Table", "[", "reference", "]", ".", "We", "divided", "the", "Hutter", "Prize", "test", "set", "into", "500", "sequences", "of", "length", "10000", ",", "and", "applied", "static", "and", "dynamic", "evaluation", "to", "these", "sequences", "using", "the", "same", "model", "and", "methodology", "used", "to", "obtain", "results", "in", "Table", "[", "reference", "]", ".", "Losses", "were", "averaged", "across", "these", "500", "sequences", "to", "obtain", "average", "losses", "at", "each", "time", "step", ".", "Plots", "of", "the", "average", "cross", "-", "entropy", "errors", "against", "the", "number", "of", "Hutter", "characters", "sequenced", "are", "given", "in", "Figure", "[", "reference", "]", ".", "The", "Spanish", "experiments", "measure", "how", "dynamic", "evaluation", "handles", "large", "distribution", "shifts", "between", "training", "and", "test", "time", ",", "as", "Hutter", "Prize", "contains", "very", "little", "Spanish", ".", "We", "used", "the", "first", "5", "million", "characters", "of", "the", "Spanish", "European", "Parliament", "data", "in", "place", "of", "the", "Hutter", "Prize", "test", "set", ".", "The", "Spanish", "experiments", "used", "the", "same", "base", "model", "and", "dynamic", "evaluation", "settings", "as", "Hutter", "Prize", ".", "Plots", "of", "the", "average", "cross", "-", "entropy", "errors", "against", "the", "number", "of", "Spanish", "characters", "sequenced", "are", "given", "in", "Figure", "[", "reference", "]", ".", ".4", ".4", "On", "both", "datasets", ",", "dynamic", "evaluation", "gave", "a", "very", "noticeable", "advantage", "after", "a", "few", "hundred", "characters", ".", "For", "Spanish", "this", "advantage", "continued", "to", "grow", "as", "more", "of", "the", "sequence", "was", "processed", ",", "whereas", "for", "Hutter", ",", "this", "advantage", "was", "maximized", "after", "viewing", "around", "2", "-", "3k", "characters", ".", "The", "advantage", "of", "dynamic", "evaluation", "was", "also", "much", "greater", "on", "Spanish", "sequences", "than", "Hutter", "sequences", ".", "We", "also", "drew", "300", "character", "conditional", "samples", "from", "the", "static", "and", "dynamic", "versions", "of", "our", "model", "after", "viewing", "10k", "characters", "of", "Spanish", ".", "For", "the", "dynamic", "model", ",", "we", "continued", "to", "apply", "dynamic", "evaluation", "during", "sampling", "as", "well", ",", "by", "the", "process", "described", "in", "Section", "[", "reference", "]", ".", "The", "conditional", "samples", "are", "given", "in", "the", "appendix", ".", "The", "static", "samples", "quickly", "switched", "to", "English", "that", "resembled", "Hutter", "Prize", "data", ".", "The", "dynamic", "model", "generated", "data", "with", "some", "Spanish", "words", "and", "a", "number", "of", "made", "up", "words", "with", "characteristics", "of", "Spanish", "words", "for", "the", "entirety", "of", "the", "sample", ".", "This", "is", "an", "example", "of", "the", "kinds", "of", "features", "that", "dynamic", "evaluation", "was", "able", "to", "learn", "to", "model", "on", "the", "fly", ".", "section", ":", "Conclusion", "This", "work", "explores", "and", "develops", "methodology", "for", "applying", "dynamic", "evaluation", "to", "sequence", "modelling", "tasks", ".", "Experiments", "show", "that", "the", "proposed", "dynamic", "evaluation", "methodology", "gives", "large", "test", "time", "improvements", "across", "character", "and", "word", "level", "language", "modelling", ".", "Our", "improvements", "to", "language", "modelling", "have", "applications", "to", "speech", "recognition", "and", "machine", "translation", "over", "longer", "contexts", ",", "including", "broadcast", "speech", "recognition", "and", "paragraph", "level", "machine", "translation", ".", "Overall", ",", "dynamic", "evaluation", "is", "shown", "to", "be", "an", "effective", "method", "for", "exploiting", "pattern", "re", "-", "occurrence", "in", "sequences", ".", "plus", "0.3ex", "bibliography", ":", "References", "appendix", ":", "Appendix", "subsection", ":", "Dynamic", "samples", "conditioned", "on", "Spanish", "300", "character", "samples", "generated", "from", "the", "dynamic", "version", "of", "the", "model", "trained", "on", "Hutter", "Prize", ",", "conditioned", "on", "10k", "of", "Spanish", "characters", ".", "The", "final", "sentence", "fragment", "of", "the", "10k", "conditioning", "characters", "is", "given", "to", "the", "reader", ",", "with", "the", "generated", "text", "given", "in", "bold", ":", "Tiene", "importancia", "este", "compromiso", "en", "la", "medida", "en", "que", "la", "Comisión", "es", "un", "organismo", "que", "tiene", "el", "montembre", "tas", "procedíns", "la", "conscriptione", "se", "ha", "Tesalo", "del", "Pómienda", "que", "et", "hanemos", "que", "Pe", "la", "Siemina", ".", "De", "la", "Pedrera", "Orden", "es", "Señora", "Presidente", "civil", ",", "Orden", "de", "siemin", "presente", "relevante", "frónmida", "que", "esculdad", "pludiore", "e", "formidad", "President", "de", "la", "Presidenta", "Antidorne", "Adamirmidad", "i", "ciemano", "de", "el", "200", "’", ".", "Fo", "subsection", ":", "Static", "samples", "conditioned", "on", "Spanish", "300", "character", "samples", "generated", "from", "the", "static", "version", "of", "the", "model", "trained", "on", "Hutter", "Prize", ",", "conditioned", "on", "10k", "of", "Spanish", "characters", ".", "The", "final", "sentence", "fragment", "of", "the", "10k", "conditioning", "characters", "is", "given", "to", "the", "reader", ",", "with", "the", "generated", "text", "given", "in", "bold", ":", "Tiene", "importancia", "este", "compromiso", "en", "la", "medida", "en", "que", "la", "Comisión", "es", "un", "organismo", "que", "tiene", "el", "monde", ",", "&", "lt;br", "&", "gt;There", "is", "a", "secret", "act", "in", "the", "world", "except", "Cape", "Town", ",", "seen", "in", "now", "flat", "comalo", "and", "ball", "market", "and", "has", "seen", "the", "closure", "of", "the", "eagle", "as", "imprints", "in", "a", "dallas", "within", "the", "country.", "&", "quot", ";", "Is", "a", "topic", "for", "an", "increasingly", "small", "contract", "saying", "Allan", "Roth", "acquired", "the", "government", "in", "[", "[", "1916", "]]", ".", "=", "=", "="], "entity": [[2, 4, ["Dynamic", "Evaluation"], "Task"], [5, 8, ["Neural", "Sequence", "Models"], "Method"], [13, 15, ["dynamic", "evaluation"], "Metric"], [17, 20, ["neural", "sequence", "models"], "Method"], [29, 33, ["gradient", "descent", "based", "mechanism"], "Method"], [47, 49, ["Dynamic", "evaluation"], "Task"], [51, 53, ["adaptation", "approaches"], "Method"], [57, 59, ["Dynamic", "evaluation"], "Task"], [68, 72, ["word", "-", "level", "perplexities"], "Metric"], [74, 76, ["Penn", "Treebank"], "Material"], [77, 81, ["WikiText", "-", "2", "datasets"], "Material"], [96, 102, ["character", "-", "level", "cross", "-", "entropies"], "Metric"], [104, 105, ["text8"], "Material"], [106, 108, ["Hutter", "Prize"], "Material"], [124, 126, ["Sequence", "generation"], "Task"], [127, 129, ["prediction", "tasks"], "Task"], [137, 138, ["audio"], "Task"], [139, 141, ["language", "modelling"], "Task"], [145, 148, ["timeseries", "prediction", "tasks"], "Task"], [154, 156, ["speech", "recognition"], "Task"], [157, 159, ["machine", "translation"], "Task"], [160, 162, ["dialogue", "generation"], "Task"], [163, 165, ["speech", "synthesis"], "Task"], [166, 167, ["forecasting"], "Task"], [169, 171, ["music", "generation"], "Task"], [175, 177, ["Neural", "networks"], "Method"], [203, 205, ["autoregressive", "model"], "Method"], [206, 209, ["Convolutional", "neural", "networks"], "Method"], [210, 211, ["CNNs"], "Method"], [213, 216, ["recurrent", "neural", "networks"], "Method"], [217, 218, ["RNNs"], "Method"], [221, 226, ["long", "-", "short", "term", "memory"], "Method"], [227, 228, ["LSTM"], "Method"], [326, 327, ["RNNs"], "Method"], [357, 359, ["dynamic", "evaluation"], "Metric"], [379, 383, ["gradient", "descent", "based", "mechanisms"], "Method"], [446, 448, ["dynamic", "evaluation"], "Metric"], [466, 468, ["dynamic", "evaluation"], "Metric"], [482, 485, ["dynamically", "evaluated", "models"], "Method"], [501, 503, ["Generative", "models"], "Method"], [518, 520, ["product", "rule"], "Method"], [539, 540, ["predicting"], "Task"], [545, 548, ["N", "-", "grams"], "Task"], [549, 550, ["CNNs"], "Method"], [565, 566, ["RNN"], "Method"], [589, 591, ["fixed", "parameters"], "Method"], [632, 634, ["generating", "distribution"], "Method"], [668, 671, ["sequence", "modelling", "tasks"], "Task"], [688, 690, ["generating", "distribution"], "Method"], [712, 715, ["machine", "learning", "benchmarks"], "Method"], [734, 737, ["sequence", "modelling", "tasks"], "Task"], [763, 765, ["fixed", "model"], "Method"], [791, 793, ["Dynamic", "evaluation"], "Task"], [793, 796, ["Dynamic", "evaluation", "methods"], "Method"], [835, 837, ["dynamic", "evaluation"], "Metric"], [902, 906, ["truncated", "back", "-", "propagation"], "Method"], [973, 977, ["maximum", "likelihood", "training", "step"], "Method"], [985, 987, ["computational", "cost"], "Metric"], [988, 990, ["dynamic", "evaluation"], "Metric"], [992, 994, ["forward", "pass"], "Method"], [996, 998, ["gradient", "computation"], "Method"], [1009, 1011, ["update", "rule"], "Method"], [1019, 1021, ["autoregressive", "models"], "Method"], [1022, 1024, ["dynamic", "evaluation"], "Metric"], [1047, 1049, ["Dynamic", "evaluation"], "Task"], [1054, 1056, ["generating", "sequences"], "Task"], [1074, 1079, ["gradient", "descent", "based", "update", "step"], "Method"], [1082, 1084, ["dynamic", "evaluation"], "Metric"], [1085, 1087, ["sequence", "generation"], "Task"], [1120, 1123, ["Adaptive", "language", "modelling"], "Task"], [1127, 1130, ["n", "-", "grams"], "Task"], [1146, 1149, ["neural", "cache", "approach"], "Method"], [1167, 1169, ["Neural", "caching"], "Method"], [1191, 1194, ["neural", "cache", "model"], "Method"], [1198, 1203, ["non", "-", "parametric", "output", "layer"], "Method"], [1274, 1277, ["one", "hot", "encoding"], "Method"], [1305, 1307, ["neural", "cache"], "Method"], [1310, 1312, ["dynamic", "evaluation"], "Metric"], [1326, 1327, ["adaptation"], "Task"], [1345, 1348, ["neural", "cache", "approach"], "Method"], [1350, 1359, ["non", "-", "parametric", ",", "nearest", "neighbours", "-", "like", "method"], "Method"], [1361, 1363, ["dynamic", "evaluation"], "Metric"], [1365, 1369, ["gradient", "descent", "based", "method"], "Method"], [1380, 1382, ["autoregressive", "factorisation"], "Method"], [1398, 1399, ["adaptation"], "Task"], [1400, 1402, ["Dynamic", "evaluation"], "Method"], [1403, 1406, ["neural", "caching", "methods"], "Method"], [1411, 1416, ["sequence", "prediction", "and", "generation", "tasks"], "Task"], [1423, 1426, ["supervised", "learning", "tasks"], "Task"], [1431, 1434, ["neural", "cache", "method"], "Method"], [1451, 1453, ["neural", "cache"], "Method"], [1486, 1490, ["character", "level", "language", "modelling"], "Task"], [1493, 1495, ["Dynamic", "evaluation"], "Task"], [1496, 1498, ["neural", "networks"], "Task"], [1498, 1500, ["Dynamic", "evaluation"], "Task"], [1501, 1504, ["neural", "language", "models"], "Method"], [1512, 1515, ["stochastic", "gradient", "descent"], "Method"], [1516, 1517, ["SGD"], "Method"], [1547, 1549, ["Dynamic", "evaluation"], "Task"], [1565, 1567, ["dynamic", "evaluation"], "Metric"], [1583, 1586, ["Update", "Rule", "methodology"], "Method"], [1587, 1589, ["Dynamic", "evaluation"], "Task"], [1595, 1598, ["dynamic", "evaluation", "method"], "Method"], [1599, 1600, ["SGD"], "Method"], [1601, 1604, ["fully", "truncated", "backpropagation"], "Method"], [1611, 1613, ["dynamic", "evaluation"], "Metric"], [1643, 1645, ["computational", "efficiency"], "Metric"], [1646, 1648, ["dynamic", "evaluation"], "Metric"], [1651, 1653, ["update", "rule"], "Method"], [1667, 1671, ["word", "-", "level", "tasks"], "Task"], [1701, 1703, ["dynamic", "evaluation"], "Metric"], [1706, 1709, ["local", "generating", "distribution"], "Method"], [1725, 1726, ["adaptation"], "Task"], [1746, 1747, ["SGD"], "Method"], [1752, 1754, ["learning", "rate"], "Metric"], [1755, 1757, ["decay", "rate"], "Metric"], [1761, 1763, ["update", "rule"], "Method"], [1768, 1772, ["RMSprop", "derived", "update", "rule"], "Method"], [1774, 1776, ["learning", "rule"], "Method"], [1779, 1780, ["SGD"], "Method"], [1781, 1782, ["RMSprop"], "Method"], [1784, 1790, ["moving", "average", "of", "recent", "squared", "gradients"], "Method"], [1799, 1801, ["dynamic", "evaluation"], "Metric"], [1810, 1811, ["RMSprop"], "Method"], [1910, 1912, ["update", "rule"], "Method"], [1916, 1917, ["RMS"], "Method"], [1939, 1941, ["update", "rule"], "Method"], [1947, 1949, ["decay", "rate"], "Metric"], [2011, 2013, ["decay", "rate"], "Metric"], [2022, 2024, ["learning", "component"], "Method"], [2026, 2028, ["regularization", "component"], "Method"], [2032, 2034, ["update", "equation"], "Method"], [2040, 2041, ["RMS"], "Method"], [2043, 2046, ["RMS", "global", "prior"], "Method"], [2051, 2054, ["Mini", "-", "batching"], "Method"], [2060, 2066, ["test", "-", "time", "sequence", "modelling", "applications"], "Task"], [2069, 2074, ["faster", "processing", "of", "multiple", "sequences"], "Task"], [2077, 2079, ["Dynamic", "evaluation"], "Task"], [2082, 2084, ["memory", "cost"], "Metric"], [2085, 2088, ["mini", "-", "batching"], "Task"], [2129, 2131, ["adaptation", "matrix"], "Method"], [2143, 2144, ["RNN"], "Method"], [2167, 2173, ["recurrent", "and", "feed", "-", "forward", "connections"], "Method"], [2175, 2177, ["dynamic", "evaluation"], "Metric"], [2183, 2185, ["dynamic", "evaluation"], "Metric"], [2202, 2205, ["mini", "-", "batching"], "Method"], [2285, 2287, ["dynamic", "evaluation"], "Metric"], [2304, 2306, ["dynamic", "evaluation"], "Metric"], [2325, 2327, ["dynamic", "evaluation"], "Metric"], [2361, 2363, ["dynamic", "evaluation"], "Metric"], [2377, 2382, ["Word", "-", "level", "language", "modelling"], "Task"], [2384, 2386, ["base", "models"], "Method"], [2388, 2390, ["Penn", "Treebank"], "Material"], [2391, 2395, ["WikiText", "-", "2", "datasets"], "Material"], [2401, 2402, ["static"], "Metric"], [2403, 2405, ["dynamic", "evaluation"], "Metric"], [2409, 2411, ["dynamic", "evaluation"], "Metric"], [2417, 2419, ["neural", "cache"], "Method"], [2421, 2423, ["dynamic", "evaluation"], "Metric"], [2432, 2433, ["PTB"], "Material"], [2439, 2442, ["Wall", "Street", "Journal"], "Material"], [2467, 2469, ["language", "modelling"], "Task"], [2476, 2477, ["PTB"], "Material"], [2483, 2485, ["recurrent", "dropout"], "Method"], [2496, 2499, ["AWD", "-", "LSTM"], "Method"], [2502, 2503, ["LSTM"], "Method"], [2507, 2509, ["Chainer", "tutorial"], "Method"], [2510, 2512, ["language", "modelling"], "Task"], [2525, 2526, ["SGD"], "Method"], [2527, 2528, ["regularized"], "Method"], [2529, 2531, ["recurrent", "dropout"], "Method"], [2535, 2536, ["LSTM"], "Method"], [2541, 2543, ["dynamic", "evaluation"], "Metric"], [2572, 2574, ["update", "rule"], "Method"], [2605, 2607, ["update", "rule"], "Method"], [2612, 2614, ["dynamic", "evaluation"], "Metric"], [2621, 2626, ["ASGD", "weight", "-", "dropped", "LSTM"], "Method"], [2628, 2631, ["AWD", "-", "LSTM"], "Method"], [2640, 2643, ["drop", "-", "connect"], "Method"], [2647, 2648, ["regularization"], "Task"], [2653, 2657, ["averaged", "stochastic", "gradient", "descent"], "Method"], [2658, 2659, ["optimisation"], "Task"], [2682, 2685, ["AWD", "-", "LSTM"], "Method"], [2701, 2703, ["Dynamic", "evaluation"], "Task"], [2714, 2716, ["Dynamic", "evaluation"], "Task"], [2723, 2725, ["neural", "cache"], "Method"], [2729, 2730, ["LSTM"], "Method"], [2748, 2749, ["PTB"], "Material"], [2750, 2753, ["WikiText", "-", "2"], "Material"], [2759, 2760, ["PTB"], "Material"], [2788, 2790, ["adaptive", "methods"], "Method"], [2816, 2817, ["PTB"], "Material"], [2827, 2829, ["Dynamic", "evaluation"], "Task"], [2838, 2839, ["perplexity"], "Metric"], [2840, 2843, ["WikiText", "-", "2"], "Material"], [2851, 2853, ["neural", "caching"], "Method"], [2861, 2863, ["dynamic", "evaluation"], "Metric"], [2880, 2885, ["Character", "-", "level", "language", "modelling"], "Task"], [2887, 2889, ["dynamic", "evaluation"], "Metric"], [2891, 2892, ["text8"], "Material"], [2894, 2897, ["Hutter", "Prize", "datasets"], "Material"], [2905, 2907, ["Wikipedia", "text"], "Material"], [2910, 2911, ["XML"], "Material"], [2914, 2918, ["non", "-", "Latin", "languages"], "Material"], [2950, 2951, ["training"], "Task"], [2952, 2953, ["validation"], "Task"], [2955, 2956, ["testing"], "Task"], [2964, 2966, ["Wikipedia", "text"], "Material"], [2983, 2985, ["English", "text"], "Material"], [3006, 3007, ["validation"], "Task"], [3011, 3012, ["text8"], "Material"], [3016, 3018, ["multiplicative", "LSTM"], "Method"], [3019, 3020, ["mLSTM"], "Method"], [3030, 3031, ["mLSTMs"], "Method"], [3040, 3042, ["embedding", "layer"], "Method"], [3046, 3048, ["weight", "normalization"], "Method"], [3049, 3051, ["variational", "dropout"], "Method"], [3053, 3054, ["ADAM"], "Method"], [3055, 3056, ["training"], "Task"], [3104, 3106, ["dynamic", "evaluation"], "Metric"], [3124, 3126, ["Hutter", "Prize"], "Material"], [3137, 3138, ["text8"], "Material"], [3146, 3148, ["Dynamic", "evaluation"], "Task"], [3176, 3178, ["Hutter", "Prize"], "Material"], [3197, 3199, ["dynamic", "evaluation"], "Metric"], [3206, 3208, ["dynamic", "evaluation"], "Metric"], [3221, 3223, ["Hutter", "Prize"], "Material"], [3229, 3230, ["static"], "Metric"], [3231, 3233, ["dynamic", "evaluation"], "Metric"], [3251, 3252, ["Spanish"], "Material"], [3254, 3257, ["European", "Parliament", "dataset"], "Material"], [3268, 3270, ["dynamic", "evaluation"], "Metric"], [3338, 3343, ["average", "cross", "-", "entropy", "errors"], "Metric"], [3347, 3348, ["Hutter"], "Material"], [3363, 3365, ["dynamic", "evaluation"], "Metric"], [3376, 3378, ["Hutter", "Prize"], "Material"], [3381, 3382, ["Spanish"], "Material"], [3392, 3396, ["Spanish", "European", "Parliament", "data"], "Material"], [3418, 3420, ["Hutter", "Prize"], "Material"], [3424, 3429, ["average", "cross", "-", "entropy", "errors"], "Metric"], [3450, 3452, ["dynamic", "evaluation"], "Metric"], [3480, 3481, ["Hutter"], "Material"], [3497, 3499, ["dynamic", "evaluation"], "Metric"], [3504, 3506, ["Spanish", "sequences"], "Material"], [3531, 3532, ["Spanish"], "Material"], [3535, 3537, ["dynamic", "model"], "Method"], [3542, 3544, ["dynamic", "evaluation"], "Metric"], [3545, 3546, ["sampling"], "Task"], [3574, 3575, ["English"], "Material"], [3582, 3584, ["dynamic", "model"], "Method"], [3588, 3590, ["Spanish", "words"], "Material"], [3619, 3621, ["dynamic", "evaluation"], "Metric"], [3642, 3644, ["dynamic", "evaluation"], "Metric"], [3645, 3648, ["sequence", "modelling", "tasks"], "Task"], [3659, 3661, ["test", "time"], "Metric"], [3663, 3664, ["character"], "Task"], [3667, 3669, ["language", "modelling"], "Task"], [3673, 3675, ["language", "modelling"], "Task"], [3678, 3680, ["speech", "recognition"], "Task"], [3681, 3683, ["machine", "translation"], "Task"], [3688, 3691, ["broadcast", "speech", "recognition"], "Task"], [3692, 3696, ["paragraph", "level", "machine", "translation"], "Task"], [3699, 3701, ["dynamic", "evaluation"], "Metric"], [3710, 3716, ["pattern", "re", "-", "occurrence", "in", "sequences"], "Task"], [3731, 3732, ["Spanish"], "Material"], [3738, 3740, ["dynamic", "version"], "Method"], [3745, 3747, ["Hutter", "Prize"], "Material"], [3752, 3754, ["Spanish", "characters"], "Material"], [3849, 3851, ["Static", "samples"], "Material"], [3853, 3854, ["Spanish"], "Material"], [3867, 3869, ["Hutter", "Prize"], "Material"], [3874, 3876, ["Spanish", "characters"], "Material"], [392, 394, ["dynamic", "evaluation"], "Metric"], [592, 593, ["static"], "Metric"], [1156, 1157, ["LSTM"], "Method"], [1164, 1166, ["language", "modelling"], "Task"], [1187, 1189, ["language", "modelling"], "Task"], [1207, 1209, ["test", "time"], "Metric"], [1301, 1303, ["test", "time"], "Metric"], [1328, 1330, ["test", "time"], "Metric"], [1554, 1555, ["character"], "Task"], [1674, 1675, ["character"], "Task"], [2049, 2051, ["dynamic", "evaluation"], "Metric"], [2114, 2116, ["dynamic", "evaluation"], "Metric"], [2271, 2273, ["dynamic", "evaluation"], "Metric"], [2274, 2275, ["character"], "Task"], [2292, 2293, ["character"], "Task"], [2295, 2297, ["language", "modelling"], "Task"], [2336, 2337, ["static"], "Metric"], [2369, 2371, ["dynamic", "evaluation"], "Metric"], [2480, 2481, ["LSTM"], "Method"], [2516, 2517, ["LSTM"], "Method"], [2634, 2635, ["LSTM"], "Method"], [2734, 2735, ["LSTM"], "Method"], [2805, 2806, ["LSTM"], "Method"], [2810, 2811, ["LSTM"], "Method"], [2899, 2900, ["Hutter"], "Material"], [2958, 2959, ["text8"], "Material"], [2990, 2991, ["Hutter"], "Material"], [3061, 3063, ["dynamic", "evaluation"], "Metric"], [3074, 3075, ["Hutter"], "Material"], [3080, 3082, ["dynamic", "evaluation"], "Metric"], [3169, 3171, ["dynamic", "evaluation"], "Metric"], [3188, 3190, ["dynamic", "evaluation"], "Metric"], [3212, 3213, ["static"], "Metric"], [3243, 3244, ["Hutter"], "Material"], [3259, 3260, ["Hutter"], "Material"], [3283, 3284, ["Hutter"], "Material"], [3296, 3297, ["static"], "Metric"], [3298, 3300, ["dynamic", "evaluation"], "Metric"], [3372, 3374, ["test", "time"], "Metric"], [3400, 3401, ["Hutter"], "Material"], [3414, 3416, ["dynamic", "evaluation"], "Metric"], [3507, 3508, ["Hutter"], "Material"], [3514, 3515, ["character"], "Task"], [3519, 3520, ["static"], "Metric"], [3569, 3570, ["static"], "Metric"], [3577, 3578, ["Hutter"], "Material"], [3654, 3656, ["dynamic", "evaluation"], "Metric"], [3733, 3734, ["character"], "Task"], [3855, 3856, ["character"], "Task"], [3860, 3861, ["static"], "Metric"]]} {"doc_id": "2ead783089fc757052abb908287a2fb743a4ebef", "text": ["document", ":", "Squeezed", "Very", "Deep", "Convolutional", "Neural", "Networks", "for", "Text", "Classification", "Most", "of", "the", "research", "in", "convolutional", "neural", "networks", "has", "focused", "on", "increasing", "network", "depth", "to", "improve", "accuracy", ",", "resulting", "in", "a", "massive", "number", "of", "parameters", "which", "restricts", "the", "trained", "network", "to", "platforms", "with", "memory", "and", "processing", "constraints", ".", "We", "propose", "to", "modify", "the", "structure", "of", "the", "Very", "Deep", "Convolutional", "Neural", "Networks", "(", "VDCNN", ")", "model", "to", "fit", "mobile", "platforms", "constraints", "and", "keep", "performance", ".", "In", "this", "paper", ",", "we", "evaluate", "the", "impact", "of", "Temporal", "Depthwise", "Separable", "Convolutions", "and", "Global", "Average", "Pooling", "in", "the", "network", "parameters", ",", "storage", "size", ",", "and", "latency", ".", "The", "squeezed", "model", "(", "SVDCNN", ")", "is", "between", "10x", "and", "20x", "smaller", ",", "depending", "on", "the", "network", "depth", ",", "maintaining", "a", "maximum", "size", "of", "6", "MB", ".", "Regarding", "accuracy", ",", "the", "network", "experiences", "a", "loss", "between", "0.4", "%", "and", "1.3", "%", "and", "obtains", "lower", "latencies", "compared", "to", "the", "baseline", "model", ".", "section", ":", "Introduction", "The", "general", "trend", "in", "deep", "learning", "approaches", "has", "been", "developing", "models", "with", "increasing", "layers", ".", "Deeper", "neural", "networks", "have", "achieved", "high", "-", "quality", "results", "in", "different", "tasks", "such", "as", "image", "classification", ",", "detection", ",", "and", "segmentation", ".", "Deep", "models", "can", "also", "learn", "hierarchical", "feature", "representations", "from", "images", ".", "In", "the", "Natural", "Language", "Processing", "(", "NLP", ")", "field", ",", "the", "belief", "that", "compositional", "models", "can", "also", "be", "used", "to", "text", "-", "related", "tasks", "is", "more", "recent", ".", "The", "increasing", "availability", "of", "text", "data", "motivates", "research", "for", "models", "able", "to", "improve", "accuracy", "in", "different", "language", "tasks", ".", "Following", "the", "image", "classification", "Convolutional", "Neural", "Network", "(", "CNN", ")", "tendency", ",", "the", "research", "in", "text", "classification", "has", "placed", "effort", "into", "developing", "deeper", "networks", ".", "The", "first", "CNN", "based", "approach", "for", "text", "was", "a", "shallow", "network", "with", "one", "layer", ".", "Following", "this", "work", ",", "deeper", "architectures", "were", "proposed", ".", "Conneau", "et", "al", ".", "were", "the", "first", "to", "propose", "Very", "Deep", "Convolutional", "Neural", "Networks", "(", "VDCNN", ")", "applied", "to", "text", "classification", ".", "VDCNN", "accuracy", "increases", "with", "depth", ".", "The", "approach", "with", "29", "layers", "is", "the", "state", "-", "of", "-", "the", "-", "art", "accuracy", "of", "CNNs", "for", "text", "classification", ".", "However", ",", "regardless", "of", "making", "networks", "deeper", "to", "improve", "accuracy", ",", "little", "effort", "has", "been", "made", "to", "build", "text", "classification", "models", "to", "constrained", "resources", ".", "It", "is", "a", "very", "different", "scenario", "compared", "to", "image", "approaches", ",", "where", "size", "and", "speed", "constrained", "models", "have", "been", "proposed", ".", "In", "real", "-", "world", "applications", ",", "size", "and", "speed", "are", "constraints", "to", "an", "efficient", "mobile", "and", "embedded", "deployment", "of", "deep", "models", ".", "Several", "relevant", "real", "-", "world", "applications", "depend", "on", "text", "classification", "tasks", "such", "as", "sentiment", "analysis", ",", "recommendation", "and", "opinion", "mining", ".", "The", "appeal", "for", "these", "applications", "combined", "with", "the", "boost", "in", "mobile", "devices", "usage", "motivates", "the", "need", "for", "research", "in", "restrained", "text", "classification", "models", ".", "Concerning", "mobile", "development", ",", "there", "are", "numerous", "benefits", "to", "developing", "smaller", "models", ".", "Some", "of", "the", "most", "relevant", "are", "requiring", "fewer", "data", "transferring", "while", "updating", "the", "client", "model", "and", "increasing", "usability", "by", "diminishing", "the", "inference", "time", ".", "Such", "advantages", "would", "boost", "the", "usage", "of", "deep", "neural", "models", "in", "text", "-", "based", "applications", "for", "embedded", "platforms", ".", "In", "this", "paper", ",", "we", "investigate", "modifications", "on", "the", "network", "proposed", "by", "Conneau", "et", "al", ".", "with", "the", "aim", "of", "reducing", "its", "number", "of", "parameters", ",", "storage", "size", "and", "latency", "with", "minimal", "performance", "degradation", ".", "To", "achieve", "these", "improvements", "we", "used", "Temporal", "Depthwise", "Separable", "Convolution", "and", "Global", "Average", "Pooling", "techniques", ".", "Therefore", ",", "our", "main", "contribution", "is", "to", "propose", "the", "Squeezed", "Very", "Deep", "Convolutional", "Neural", "Networks", "(", "SVDCNN", ")", ",", "a", "text", "classification", "model", "which", "requires", "significantly", "fewer", "parameters", "compared", "to", "the", "state", "-", "of", "-", "the", "-", "art", "CNNs", ".", "Section", "II", "provides", "an", "overview", "of", "the", "state", "-", "of", "-", "the", "-", "art", "in", "CNNs", "for", "text", "classification", ".", "Section", "III", "presents", "the", "VDCNN", "model", ".", "Section", "IV", "explains", "the", "proposed", "model", "SVDCNN", "and", "the", "subsequent", "impact", "in", "the", "total", "number", "of", "parameters", "of", "the", "network", ".", "Section", "V", "details", "how", "we", "perform", "experiments", ".", "Section", "VI", "analyses", "the", "results", "and", "lastly", ",", "Section", "VII", ",", "presents", "conclusions", "and", "direction", "for", "future", "works", ".", "section", ":", "Related", "Work", "CNNs", "were", "originally", "designed", "for", "Computer", "Vision", "with", "the", "aim", "of", "considering", "feature", "extraction", "and", "classification", "as", "one", "task", ".", "Although", "CNNs", "are", "very", "successful", "in", "image", "classification", "tasks", ",", "its", "use", "in", "text", "classification", "is", "relatively", "new", "and", "has", "some", "peculiarities", ".", "Contrasting", "with", "traditional", "image", "bi", "-", "dimensional", "representations", ",", "texts", "are", "one", "-", "dimensionally", "represented", ".", "Due", "to", "this", "property", ",", "the", "convolutions", "are", "designed", "as", "temporal", "convolutions", ".", "Furthermore", ",", "it", "is", "necessary", "to", "generate", "a", "numerical", "representation", "from", "the", "text", "so", "the", "network", "can", "be", "trained", "using", "this", "representation", ".", "This", "representation", ",", "namely", "embeddings", ",", "is", "usually", "obtained", "through", "the", "application", "of", "a", "lookup", "table", ",", "generated", "from", "a", "given", "dictionary", ".", "An", "early", "approach", "for", "text", "classification", "tasks", "consisted", "of", "a", "shallow", "neural", "network", "working", "on", "the", "word", "level", "and", "using", "only", "one", "convolutional", "layer", ".", "The", "author", "reported", "results", "in", "smaller", "datasets", ".", "Later", ",", "Zhang", "et", "al", ".", "proposed", "the", "first", "CNN", "performing", "on", "a", "character", "level", "(", "Char", "-", "CNN", ")", ",", "which", "allowed", "them", "to", "train", "up", "to", "6", "convolutional", "layers", ",", "followed", "by", "three", "fully", "connected", "classification", "layers", ".", "Char", "-", "CNN", "uses", "convolutional", "kernels", "of", "size", "3", "and", "7", ",", "as", "well", "as", "simple", "max", "-", "pooling", "layers", ".", "Conneau", "et", "al", ".", "(", "2016", ")", "proposed", "the", "Very", "Deep", "CNN", "(", "VDCNN", ")", "also", "on", "a", "character", "level", ",", "presenting", "improvements", "compared", "to", "Char", "-", "CNN", ".", "Conneau", "et", "al", ".", "(", "2016", ")", "have", "shown", "that", "text", "classification", "accuracy", "increases", "when", "the", "proposed", "model", "becomes", "deeper", ".", "VDCNN", "uses", "only", "small", "kernel", "convolutions", "and", "pooling", "operations", ".", "The", "proposed", "architecture", "relies", "on", "the", "VGG", "and", "ResNet", "philosophy", ":", "The", "number", "of", "feature", "maps", "and", "the", "temporal", "resolution", "is", "modeled", "so", "that", "their", "product", "is", "constant", ".", "This", "approach", "makes", "it", "easier", "to", "control", "the", "memory", "footprint", "of", "the", "network", ".", "Both", "Zhang", "and", "Conneau", "et", "al", ".", "CNNs", "utilized", "standard", "convolutional", "blocks", "and", "fully", "connected", "layers", "to", "combine", "convolution", "information", ".", "This", "architecture", "choice", "increases", "the", "number", "of", "parameters", "and", "storage", "size", "of", "the", "models", ".", "However", ",", "size", "and", "speed", "was", "not", "the", "focus", "of", "those", "works", ".", "The", "idea", "of", "developing", "smaller", "and", "more", "efficient", "CNNs", "without", "losing", "representative", "accuracy", "is", "a", "less", "explored", "research", "direction", "in", "NLP", ",", "but", "it", "has", "already", "been", "a", "trend", "for", "computer", "vision", "applications", ".", "Most", "approaches", "consist", "in", "compressing", "pre", "-", "trained", "networks", "or", "training", "small", "networks", "directly", ".", "A", "recent", "tendency", "in", "deep", "models", "is", "replacing", "standard", "convolutional", "blocks", "with", "Depthwise", "Separable", "Convolutions", "(", "DSCs", ")", ".", "The", "purpose", "is", "to", "reduce", "the", "number", "of", "parameters", "and", "consequently", "the", "model", "size", ".", "DSCs", "were", "initially", "introduced", "in", "and", "since", "then", "have", "been", "successfully", "applied", "to", "image", "classification", "and", "machine", "translation", "to", "reduce", "the", "computation", "in", "convolutional", "blocks", ".", "Another", "approach", "is", "the", "use", "of", "a", "Global", "Average", "Pooling", "(", "GAP", ")", "layer", "at", "the", "output", "of", "the", "network", "to", "replace", "fully", "connected", "layers", ".", "This", "approach", "has", "become", "a", "standard", "architectural", "decision", "for", "newer", "CNNs", ".", "section", ":", "VDCNN", "Model", "for", "Text", "Classification", "The", "VDCNN", "is", "a", "modular", "architecture", "for", "text", "classification", "tasks", "developed", "to", "offer", "different", "depth", "levels", "(", "9", ",", "17", ",", "29", "and", "49", ")", ".", "Fig", ".", "[", "reference", "]", "presents", "the", "architecture", "for", "depth", "9", ".", "The", "network", "begins", "with", "a", "lookup", "table", ",", "which", "generates", "the", "embeddings", "for", "the", "input", "text", "and", "stores", "them", "in", "a", "2D", "tensor", "of", "size", ".", "The", "number", "of", "input", "characters", "is", "fixed", "to", "1", ",", "024", "while", "the", "embedding", "dimension", "is", "16", ".", "The", "embedding", "dimension", "can", "be", "seen", "as", "the", "number", "of", "RGB", "channels", "of", "an", "image", ".", "The", "following", "layer", "(", "3", ",", "Temp", "Convolution", ",", "64", ")", "applies", "64", "temporal", "convolutions", "of", "kernel", "size", "3", ",", "so", "the", "output", "tensor", "has", "size", ".", "Its", "primary", "function", "is", "to", "fit", "the", "lookup", "table", "output", "with", "the", "modular", "network", "segment", "input", "composed", "by", "convolutional", "blocks", ".", "Each", "aforenamed", "block", "is", "a", "sequence", "of", "two", "temporal", "convolutional", "layers", ",", "each", "one", "accompanied", "by", "a", "temporal", "batch", "normalization", "layer", "and", "a", "ReLU", "activation", ".", "Besides", ",", "the", "different", "network", "depths", "are", "obtained", "varying", "the", "number", "of", "convolutional", "blocks", ".", "As", "a", "convention", ",", "the", "depth", "of", "a", "network", "is", "given", "as", "its", "total", "number", "of", "convolutions", ".", "For", "instance", ",", "the", "architecture", "of", "depth", "17", "has", "two", "convolutional", "blocks", "of", "each", "level", "of", "feature", "maps", ",", "which", "results", "in", "4", "convolutional", "layers", "for", "each", "level", "(", "see", "Table", "[", "reference", "]", ")", ".", "Considering", "the", "first", "convolutional", "layer", "of", "the", "network", ",", "we", "obtain", "the", "depth", ".", "The", "different", "depth", "architectures", "provided", "by", "VDCNN", "model", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "The", "following", "rule", "is", "employed", "to", "minimize", "the", "network", "’s", "memory", "footprint", ":", "Before", "each", "convolutional", "block", "doubling", "the", "number", "of", "feature", "maps", ",", "a", "pooling", "layer", "halves", "the", "temporal", "dimension", ".", "This", "strategy", "is", "inspired", "by", "the", "VGG", "and", "ResNets", "philosophy", "and", "results", "in", "three", "levels", "of", "feature", "maps", ":", "128", ",", "256", "and", "512", "(", "see", "Fig", ".", "[", "reference", "]", ")", ".", "Additionally", ",", "the", "VDCNN", "network", "also", "contains", "shortcut", "connections", "for", "each", "convolutional", "blocks", "implemented", "through", "the", "usage", "of", "convolutions", ".", "Lastly", ",", "for", "the", "classification", "task", ",", "the", "most", "valuable", "features", "are", "extracted", "using", "-", "max", "pooling", ",", "generating", "a", "one", "-", "dimensional", "vector", "which", "supplies", "three", "fully", "connected", "layers", "with", "ReLU", "hidden", "units", "and", "softmax", "outputs", ".", "The", "number", "of", "hidden", "units", "is", "2", ",", "048", ",", "and", "they", "do", "not", "use", "dropout", "but", "rather", "batch", "normalization", "after", "convolutional", "layers", "perform", "the", "network", "regularization", ".", "section", ":", "SVDCNN", "Model", "for", "Text", "Classification", "The", "primary", "objective", "is", "reducing", "the", "number", "of", "parameters", "so", "that", "the", "resulting", "network", "has", "a", "significative", "lower", "storage", "size", ".", "We", "first", "propose", "to", "modify", "the", "convolutional", "blocks", "of", "VDCNN", "model", "by", "the", "usage", "of", "Temporal", "Depthwise", "Separable", "Convolutions", "(", "TDSCs", ")", ".", "Next", ",", "we", "reduce", "the", "number", "of", "fully", "connected", "layers", "using", "the", "Global", "Average", "Pooling", "(", "GAP", ")", "technique", ".", "The", "resulting", "proposed", "architecture", "is", "called", "Squeezed", "Very", "Deep", "Convolutional", "Neural", "Networks", "(", "SVDCNN", ")", ".", "paragraph", ":", "Temporal", "Depthwise", "Separable", "Convolutions", "(", "TDSCs", ")", "The", "use", "of", "TDSCs", "over", "standard", "convolutions", "allowed", "reducing", "the", "number", "of", "parameters", "without", "relevant", "accuracy", "loss", ".", "TDSCs", "work", "decompounding", "the", "standard", "convolution", "into", "two", "parts", ":", "Depthwise", "and", "Pointwise", ".", "The", "first", "one", "is", "responsible", "for", "applying", "a", "convolutional", "filter", "to", "each", "channel", "of", "the", "input", "at", "a", "time", ".", "For", "an", "image", "input", ",", "one", "possibility", "of", "channels", "are", "the", "RGB", "components", ",", "whereas", "in", "a", "text", "input", "the", "dimensions", "of", "the", "embedding", "can", "be", "used", "instead", ".", "For", "both", "cases", "mentioned", "above", ",", "the", "result", "is", "one", "feature", "map", "by", "channel", ".", "The", "second", "convolution", "unifies", "the", "generated", "feature", "maps", "successively", "applying", "1x1", "convolutions", "so", "that", "the", "target", "amount", "of", "feature", "maps", "can", "be", "achieved", ".", "[", "]", "[", "]", "TDSCs", "are", "DSCs", "which", "work", "with", "one", "-", "dimensional", "convolutions", ".", "Although", "DSCs", "hold", "verified", "results", "in", "image", "classification", "networks", ",", "the", "usage", "of", "its", "temporal", "version", "for", "text", "related", "tasks", "is", "less", "explored", ".", "Fig", ".", "2a", "presents", "the", "architecture", "of", "a", "temporal", "standard", "convolution", "while", "Fig", ".", "2b", "presents", "the", "TDSC", ".", "For", "a", "more", "formal", "definition", ",", "let", "be", "the", "number", "of", "parameters", "of", "a", "temporal", "standard", "convolution", ",", "where", "In", "and", "Out", "are", "the", "numbers", "of", "Input", "and", "Output", "channels", "respectively", ",", "and", "is", "the", "kernel", "size", ":", "Alternatively", ",", "a", "TDSC", "achieves", "fewer", "parameters", "(", ")", ":", "In", "the", "VDCNN", "model", ",", "one", "convolutional", "block", "is", "composed", "of", "two", "temporal", "standard", "convolutional", "layers", ".", "The", "first", "one", "doubles", "the", "number", "of", "feature", "maps", "while", "the", "second", "keeps", "the", "same", "value", "received", "as", "input", ".", "Besides", ",", "each", "convolutional", "layer", "is", "followed", "by", "a", "Batch", "Normalization", "and", "a", "ReLU", "layers", ".", "In", "our", "model", ",", "we", "proposed", "changing", "the", "temporal", "standard", "convolutions", "by", "TDSCs", ".", "[", "]", "[", "]", "Fig", ".", "3", "presents", "the", "standard", "convolutional", "block", "on", "the", "left", "and", "the", "proposed", "convolutional", "block", "using", "TDSC", "on", "the", "right", ".", "The", "pattern", "used", "in", "the", "figure", "for", "the", "convolutional", "layers", "is", "the", "following", ":", "”", "Kernel", "Size", ",", "Conv", "type", ",", "Output", "Feature", "Maps", "”", ";", "as", "a", "brief", "example", "consider", "”", "3x1", ",", "Temporal", "Conv", ",", "256", "”", ",", "which", "means", "a", "Temporal", "Convolution", "with", "kernel", "size", "3", "and", "256", "feature", "maps", "as", "output", ".", "From", "Equation", "1", ",", "we", "have", "the", "number", "of", "parameters", "of", "the", "original", "convolutional", "block", "(", ")", "as", "follows", ":", "Moreover", ",", "from", "equation", "2", ",", "the", "number", "of", "parameters", "of", "the", "proposed", "convolutional", "block", "(", ")", "that", "uses", "TDSC", "being", ":", "For", "illustration", ",", "following", "the", "same", "characteristics", "of", "Fig", ".", "3", ",", "consider", "that", "the", "number", "of", "input", "channels", "In", "is", "equal", "to", "128", "and", "the", "number", "of", "output", "channels", "Out", "is", "equal", "to", "256", ".", "Our", "proposed", "approach", "accumulates", "a", "total", "of", "99", ",", "456", "parameters", ".", "In", "contrast", ",", "there", "are", "294", ",", "912", "parameters", "in", "the", "original", "convolutional", "block", ".", "The", "use", "of", "TDSC", "yields", "a", "reduction", "of", "66.28", "%", "in", "the", "network", "size", ".", "Lastly", ",", "since", "each", "standard", "temporal", "convolution", "turns", "into", "two", "(", "Depthwise", "and", "Pointwise", ")", ",", "the", "number", "of", "convolutions", "per", "convolutional", "block", "has", "doubled", ".", "Nevertheless", ",", "these", "two", "convolutions", "work", "as", "one", "because", "it", "is", "not", "possible", "to", "use", "them", "separately", "keeping", "the", "same", "propose", ".", "In", "this", "way", ",", "we", "count", "them", "as", "one", "layer", "in", "the", "network", "depth", ".", "This", "decision", "holds", "the", "provided", "depth", "architectures", "the", "same", "as", "the", "VDCNN", "model", "summarized", "in", "Table", "[", "reference", "]", ",", "contributing", "to", "a", "proper", "comparison", "between", "the", "models", ".", "paragraph", ":", "Global", "Average", "Pooling", "(", "GAP", ")", "The", "VDCNN", "model", "uses", "a", "-", "max", "pooling", "layer", "followed", "by", "three", "fully", "connected", "(", "FC", ")", "layers", "to", "perform", "the", "classification", "task", "(", "Fig", ".", "4a", ")", ".", "Although", "this", "approach", "is", "the", "traditional", "architecture", "choice", "for", "text", "classification", "CNNs", ",", "it", "introduces", "a", "significant", "number", "of", "parameter", "in", "the", "network", ".", "The", "resulting", "number", "of", "the", "FC", "layers", "parameters", "(", ")", "aforementioned", "is", "presented", "below", ",", "for", "a", "problem", "with", "four", "target", "classes", ":", "[", "]", "[", "]", "Instead", "of", "maintaining", "these", "fully", "connected", "layers", ",", "we", "directly", "aggregate", "the", "output", "of", "the", "last", "convolutional", "block", "through", "the", "usage", "of", "an", "average", "pooling", "layer", ".", "This", "method", ",", "known", "as", "Global", "Average", "Pooling", ",", "contributes", "substantially", "to", "the", "parameters", "reduction", "without", "degrading", "the", "network", "accuracy", "significantly", ".", "The", "number", "of", "resulting", "feature", "maps", "given", "by", "the", "average", "pooling", "layer", "was", "the", "same", "as", "the", "original", "-", "max", "pooling", "layer", ".", "Fig", ".", "4b", "presents", "this", "proposed", "modification", ".", "The", "number", "of", "parameters", "obtained", "by", "the", "usage", "of", "GAP", "(", ")", "is", "revealed", "as", "follows", ":", "Our", "proposed", "approach", "accumulates", "a", "total", "of", "16", ",", "384", "parameters", ".", "In", "contrast", ",", "there", "are", "12", ",", "591", ",", "104", "parameters", "in", "the", "original", "classification", "method", ".", "The", "use", "of", "GAP", "yields", "a", "reduction", "of", "99.86", "%", ".", "section", ":", "Experiments", "The", "experiment", "goal", "is", "to", "investigate", "the", "impact", "of", "modifying", "the", "convolutional", "block", "of", "VDCNN", "to", "TDSCs", "and", "using", "GAP", "instead", "of", "the", "original", "fully", "connected", "layers", ".", "We", "evaluate", "Char", "-", "CNN", ",", "VDCNN", ",", "and", "SVDCNN", "according", "to", "the", "number", "of", "parameters", ",", "storage", "size", ",", "inference", "time", "and", "accuracy", ".", "The", "source", "code", "of", "the", "proposed", "model", "is", "available", "in", "the", "GitHub", "repository", "SVDCNN", "The", "original", "VDCNN", "paper", "reported", "the", "number", "of", "parameters", "of", "the", "convolutional", "layers", ",", "in", "which", "we", "reproduce", "in", "this", "article", ".", "For", "SVDCNN", "and", "Char", "-", "CNN", ",", "we", "calculated", "the", "abovementioned", "number", "from", "the", "network", "architecture", "implemented", "in", "PyTorch", ".", "As", "for", "the", "FC", "layer", "’s", "parameters", ",", "the", "number", "is", "obtained", "as", "the", "summation", "of", "the", "product", "of", "the", "input", "and", "output", "size", "of", "each", "FC", "layer", "for", "each", "CNN", ".", "Considering", "the", "network", "parameters", "and", "assuming", "that", "one", "float", "number", "on", "Cuda", "environment", "takes", "4", "bytes", ",", "we", "can", "calculate", "the", "network", "storage", "in", "megabytes", ",", "for", "all", "the", "models", ",", "as", "follows", ":", "Regarding", "the", "inference", "time", ",", "its", "average", "and", "standard", "deviation", "were", "calculated", "as", "the", "time", "to", "predict", "one", "instance", "of", "the", "AG", "’s", "News", "dataset", "throughout", "1", ",", "000", "repetitions", ".", "The", "SVDCNN", "experimental", "settings", "are", "similar", "to", "the", "original", "VDCNN", "paper", ",", "using", "the", "same", "dictionary", "and", "the", "same", "embedding", "size", "of", "16", ".", "The", "training", "is", "also", "performed", "with", "SGD", ",", "utilizing", "size", "batch", "of", "64", ",", "with", "a", "maximum", "of", "100", "epochs", ".", "We", "use", "an", "initial", "learning", "rate", "of", "0.01", ",", "a", "momentum", "of", "0.9", "and", "a", "weight", "decay", "of", "0.001", ".", "All", "the", "experiments", "were", "performed", "on", "an", "NVIDIA", "GTX", "1060", "GPU", "+", "Intel", "Core", "i7", "4770s", "CPU", ".", "The", "model", "’s", "performance", "is", "evaluated", "on", "three", "large", "-", "scale", "public", "datasets", "also", "used", "by", "Zhang", "et", "al", ".", "in", "the", "introduction", "of", "Char", "-", "CNN", "and", "VDCNN", "models", ".", "Table", "[", "reference", "]", "presents", "the", "details", "of", "the", "utilized", "datasets", ":", "AG", "’s", "News", ",", "Yelp", "Polarity", "and", "Yelp", "Full", ".", "section", ":", "Results", "Table", "[", "reference", "]", "presents", "the", "number", "of", "parameters", ",", "storage", "size", ",", "and", "accuracy", "for", "the", "SVDCNN", ",", "VDCNN", ",", "and", "Char", "-", "CNN", "in", "all", "datasets", ".", "The", "use", "of", "TDSCs", "promoted", "a", "significant", "reduction", "in", "convolutional", "parameters", "compared", "to", "VDCNN", ".", "For", "the", "most", "in", "-", "depth", "network", "evaluated", ",", "which", "contains", "29", "convolutional", "layers", "(", "depth", "29", ")", ",", "the", "number", "of", "parameters", "of", "these", "convolutional", "layers", "had", "a", "reduction", "of", "66.08", "%", ",", "from", "4.6", "to", "1.56", "million", "parameters", ".", "This", "quantity", "is", "slightly", "larger", "than", "the", "one", "obtained", "from", "the", "Char", "-", "CNN", ",", "1.40", "million", "parameters", ",", "but", "this", "network", "has", "only", "six", "convolutional", "layers", "(", "depth", "6", ")", ".", "The", "network", "reduction", "obtained", "by", "the", "GAP", "is", "even", "more", "representative", "since", "both", "compared", "models", "use", "three", "FC", "layers", "for", "their", "classification", "tasks", ".", "Considering", "a", "dataset", "with", "four", "target", "classes", ",", "and", "comparing", "SVDCNN", "with", "VDCNN", ",", "the", "number", "of", "parameters", "of", "the", "FC", "layers", "has", "passed", "from", "12.59", "to", "0.02", "million", "parameters", ",", "representing", "a", "reduction", "of", "99.84", "%", ".", "Following", "with", "the", "same", "comparison", ",", "but", "to", "Char", "-", "CNN", ",", "the", "proposed", "model", "is", "99.82", "%", "smaller", ",", "0.02", "against", "11.36", "million", "of", "FC", "parameters", ".", "The", "reduction", "of", "the", "total", "parameters", "impacts", "directly", "on", "the", "storage", "size", "of", "the", "networks", ".", "While", "our", "most", "in", "-", "depth", "model", "(", "29", ")", "occupies", "only", "6", "MB", ",", "VDCNN", "with", "the", "same", "depth", "occupies", "64.16", "MB", "of", "storage", ".", "Likewise", ",", "Char", "-", "CNN", "(", "which", "has", "depth", "6", ")", "occupies", "43.25", "MB", ".", "This", "reduction", "is", "a", "significant", "result", "because", "many", "embedded", "platforms", "have", "several", "memory", "constraints", ".", "For", "example", ",", "FPGAs", "often", "have", "less", "than", "10", "MB", "of", "on", "-", "chip", "memory", "and", "no", "off", "-", "chip", "memory", "or", "storage", ".", "Regarding", "accuracy", "results", ",", "usually", ",", "a", "model", "with", "such", "parameter", "reduction", "should", "present", "some", "loss", "of", "accuracy", "in", "comparison", "to", "the", "original", "model", ".", "Nevertheless", ",", "the", "performance", "difference", "between", "VDCNN", "and", "SVDCNN", "models", "varies", "between", "0.4", "and", "1.3", "%", ",", "which", "is", "pretty", "modest", "considering", "the", "parameters", "and", "storage", "size", "reduction", "aforementioned", ".", "In", "Table", "[", "reference", "]", ",", "it", "is", "possible", "to", "see", "the", "accuracy", "scores", "obtained", "by", "the", "compared", "models", ".", "Another", "two", "fundamental", "results", "obtained", "are", "a", ")", "The", "base", "property", "of", "VDCNN", "model", "is", "preserved", "on", "its", "squeezed", "model", ":", "the", "performance", "still", "increasing", "up", "with", "the", "depth", "and", "b", ")", "The", "performance", "evaluated", "for", "the", "most", "extensive", "dataset", ",", "i.e.", ",", "Yelp", "Review", "(", "62.30", "%", ")", ",", "still", "overcomes", "the", "accuracy", "of", "the", "Char", "-", "CNN", "model", "(", "62.05", "%", ")", ".", "Deep", "learning", "processing", "architecture", "has", "the", "property", "of", "being", "high", "parallelizable", ";", "it", "is", "expected", "smaller", "latencies", "when", "performing", "inferences", "in", "hardware", "with", "high", "parallelization", "power", ".", "Despite", "this", "property", ",", "the", "model", "ability", "to", "use", "all", "hardware", "parallel", "potential", "available", "also", "depends", "on", "the", "network", "architecture", ".", "The", "more", "parameters", "per", "layers", ",", "the", "more", "parallelizable", "a", "model", "tends", "to", "be", ",", "while", "the", "increase", "of", "the", "depth", "gets", "the", "opposite", "result", ".", "Another", "natural", "comprehension", "fact", "is", "if", "a", "model", "has", "few", "parameters", ",", "there", "exists", "less", "content", "to", "be", "processed", ",", "and", "then", "we", "have", "a", "faster", "inference", "time", ".", "Concerning", "mobile", "devices", ",", "the", "presence", "of", "dedicated", "hardware", "for", "deep", "learning", "is", "not", "entirely", "feasible", ".", "This", "hardware", "usually", "requires", "more", "energy", "and", "dissipates", "more", "heat", ",", "two", "undesirable", "features", "for", "a", "mobile", "platform", ".", "Therefore", ",", "obtaining", "fewer", "inference", "times", ",", "even", "out", "of", "environments", "with", "high", "parallelization", "capabilities", ",", "is", "a", "pretty", "desirable", "characteristic", "for", "a", "model", "designed", "to", "work", "on", "mobile", "platforms", ".", "The", "latency", "ratio", "between", "CPU", "and", "GPU", "inference", "times", "indicates", "how", "undependable", "of", "dedicated", "hardware", "a", "model", "is", ",", "with", "higher", "values", "meaning", "more", "independence", ".", "The", "inference", "times", "obtained", "for", "the", "three", "models", "compared", "are", "available", "in", "Table", "[", "reference", "]", ".", "As", "explained", "in", "Section", "IV", "a", ")", ",", "each", "convolutional", "layer", "of", "the", "convolutional", "blocks", "was", "substituted", "by", "two", "convolutions", ".", "This", "change", "could", "impact", "the", "inference", "time", "negatively", ",", "but", "the", "significant", "parameter", "reduction", "allows", "the", "SVDCNN", "to", "obtain", "better", "results", "than", "the", "VDCNN", "model", ".", "The", "CPU", "inference", "time", "obtained", "by", "the", "proposed", "model", "was", "smaller", "than", "the", "base", "model", "for", "the", "depth", "9", "(", "25.88ms", "against", "29", ",", "13ms", ")", "and", "depth", "17", "(", "47.80ms", "against", "48.05ms", ")", ",", "while", "the", "Ratio", "was", "higher", "for", "all", "depths", "(", "0.20", "against", "0.15", "in", "average", ")", ".", "These", "results", ",", "as", "explained", "above", ",", "are", "pretty", "significant", "for", "mobile", "platforms", ".", "Looking", "to", "Char", "-", "CNN", ",", "this", "model", "got", "notably", "inferior", "results", "compared", "to", "the", "proposed", "method", ",", "with", "313.53ms", "of", "CPU", "inference", "time", "and", "Ratio", "of", "0.03", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "presented", "a", "squeezed", "version", "of", "the", "VDCNN", "model", "considering", "the", "number", "of", "parameters", "and", "size", ".", "The", "new", "model", "proprieties", "became", "it", "feasible", "for", "mobile", "platforms", ".", "To", "achieve", "this", "goal", ",", "we", "analyzed", "the", "impact", "of", "including", "Temporal", "Depthwise", "Separable", "Convolutions", "and", "a", "Global", "Average", "Pooling", "layer", "in", "a", "very", "deep", "convolutional", "neural", "network", "for", "text", "classification", ".", "The", "SVDCNN", "model", "reduces", "about", "92.45", "%", "the", "number", "of", "parameters", "and", "storage", "size", "while", "presents", "an", "inference", "time", "ratio", "(", "CPU", "/", "GPU", ")", ",", "31.94", "%", "higher", ".", "For", "future", "works", ",", "we", "plan", "to", "evaluate", "other", "techniques", "able", "to", "reduce", "storage", "size", ",", "such", "as", "model", "compression", ".", "Moreover", ",", "the", "model", "accuracy", "over", "even", "more", "massive", "datasets", "will", "be", "evaluated", "as", "well", "as", "the", "efficiency", "of", "its", "depth", "49", "configuration", ".", "section", ":", "Acknowledgment", "We", "would", "like", "to", "thank", "FACEPE", "and", "CNPq", "(", "Brazilian", "research", "agencies", ")", "for", "financial", "support", ".", "bibliography", ":", "References"], "entity": [[2, 8, ["Squeezed", "Very", "Deep", "Convolutional", "Neural", "Networks"], "Method"], [9, 11, ["Text", "Classification"], "Task"], [16, 19, ["convolutional", "neural", "networks"], "Method"], [27, 28, ["accuracy"], "Metric"], [57, 62, ["Very", "Deep", "Convolutional", "Neural", "Networks"], "Method"], [63, 64, ["VDCNN"], "Method"], [84, 88, ["Temporal", "Depthwise", "Separable", "Convolutions"], "Method"], [89, 92, ["Global", "Average", "Pooling"], "Method"], [97, 99, ["storage", "size"], "Metric"], [101, 102, ["latency"], "Metric"], [104, 106, ["squeezed", "model"], "Method"], [107, 108, ["SVDCNN"], "Method"], [131, 132, ["accuracy"], "Metric"], [147, 148, ["latencies"], "Metric"], [161, 164, ["deep", "learning", "approaches"], "Method"], [172, 175, ["Deeper", "neural", "networks"], "Method"], [186, 188, ["image", "classification"], "Task"], [189, 190, ["detection"], "Task"], [192, 193, ["segmentation"], "Task"], [194, 196, ["Deep", "models"], "Method"], [199, 202, ["hierarchical", "feature", "representations"], "Method"], [207, 210, ["Natural", "Language", "Processing"], "Task"], [211, 212, ["NLP"], "Task"], [218, 220, ["compositional", "models"], "Method"], [225, 229, ["text", "-", "related", "tasks"], "Task"], [246, 247, ["accuracy"], "Metric"], [249, 251, ["language", "tasks"], "Task"], [256, 259, ["Convolutional", "Neural", "Network"], "Method"], [260, 261, ["CNN"], "Method"], [267, 269, ["text", "classification"], "Task"], [274, 276, ["deeper", "networks"], "Method"], [286, 288, ["shallow", "network"], "Method"], [310, 315, ["Very", "Deep", "Convolutional", "Neural", "Networks"], "Method"], [316, 318, ["VDCNN", ")"], "Method"], [320, 322, ["text", "classification"], "Task"], [343, 344, ["accuracy"], "Metric"], [345, 346, ["CNNs"], "Method"], [347, 349, ["text", "classification"], "Task"], [359, 360, ["accuracy"], "Metric"], [368, 371, ["text", "classification", "models"], "Method"], [383, 385, ["image", "approaches"], "Method"], [387, 392, ["size", "and", "speed", "constrained", "models"], "Method"], [397, 401, ["real", "-", "world", "applications"], "Task"], [404, 405, ["speed"], "Metric"], [410, 417, ["mobile", "and", "embedded", "deployment", "of", "deep", "models"], "Task"], [420, 424, ["real", "-", "world", "applications"], "Task"], [426, 429, ["text", "classification", "tasks"], "Task"], [431, 433, ["sentiment", "analysis"], "Task"], [434, 435, ["recommendation"], "Task"], [436, 438, ["opinion", "mining"], "Task"], [449, 452, ["mobile", "devices", "usage"], "Task"], [458, 462, ["restrained", "text", "classification", "models"], "Method"], [464, 466, ["mobile", "development"], "Task"], [473, 475, ["smaller", "models"], "Method"], [484, 486, ["data", "transferring"], "Task"], [489, 491, ["client", "model"], "Method"], [497, 499, ["inference", "time"], "Metric"], [507, 510, ["deep", "neural", "models"], "Method"], [511, 515, ["text", "-", "based", "applications"], "Task"], [516, 518, ["embedded", "platforms"], "Task"], [545, 547, ["storage", "size"], "Metric"], [548, 549, ["latency"], "Metric"], [560, 564, ["Temporal", "Depthwise", "Separable", "Convolution"], "Method"], [565, 569, ["Global", "Average", "Pooling", "techniques"], "Method"], [579, 585, ["Squeezed", "Very", "Deep", "Convolutional", "Neural", "Networks"], "Method"], [586, 587, ["SVDCNN"], "Method"], [590, 593, ["text", "classification", "model"], "Method"], [608, 609, ["CNNs"], "Method"], [625, 626, ["CNNs"], "Method"], [627, 629, ["text", "classification"], "Task"], [634, 636, ["VDCNN", "model"], "Method"], [643, 644, ["SVDCNN"], "Method"], [689, 690, ["CNNs"], "Method"], [694, 696, ["Computer", "Vision"], "Task"], [701, 703, ["feature", "extraction"], "Task"], [704, 705, ["classification"], "Task"], [710, 711, ["CNNs"], "Method"], [715, 718, ["image", "classification", "tasks"], "Task"], [722, 724, ["text", "classification"], "Task"], [735, 740, ["image", "bi", "-", "dimensional", "representations"], "Method"], [754, 755, ["convolutions"], "Method"], [758, 760, ["temporal", "convolutions"], "Method"], [769, 771, ["numerical", "representation"], "Method"], [811, 814, ["text", "classification", "tasks"], "Task"], [817, 820, ["shallow", "neural", "network"], "Method"], [829, 831, ["convolutional", "layer"], "Method"], [849, 850, ["CNN"], "Method"], [856, 859, ["Char", "-", "CNN"], "Method"], [869, 871, ["convolutional", "layers"], "Method"], [880, 883, ["Char", "-", "CNN"], "Method"], [884, 886, ["convolutional", "kernels"], "Method"], [896, 900, ["max", "-", "pooling", "layers"], "Method"], [914, 915, ["VDCNN"], "Method"], [926, 929, ["Char", "-", "CNN"], "Method"], [940, 943, ["text", "classification", "accuracy"], "Metric"], [951, 952, ["VDCNN"], "Method"], [955, 957, ["kernel", "convolutions"], "Method"], [958, 960, ["pooling", "operations"], "Method"], [967, 971, ["VGG", "and", "ResNet", "philosophy"], "Method"], [1011, 1012, ["CNNs"], "Method"], [1014, 1016, ["convolutional", "blocks"], "Method"], [1034, 1036, ["storage", "size"], "Metric"], [1061, 1062, ["CNNs"], "Method"], [1064, 1066, ["representative", "accuracy"], "Metric"], [1073, 1074, ["NLP"], "Task"], [1083, 1086, ["computer", "vision", "applications"], "Task"], [1092, 1096, ["pre", "-", "trained", "networks"], "Method"], [1098, 1100, ["small", "networks"], "Method"], [1106, 1108, ["deep", "models"], "Method"], [1111, 1113, ["convolutional", "blocks"], "Method"], [1114, 1117, ["Depthwise", "Separable", "Convolutions"], "Method"], [1118, 1119, ["DSCs"], "Method"], [1136, 1137, ["DSCs"], "Method"], [1149, 1151, ["image", "classification"], "Task"], [1152, 1154, ["machine", "translation"], "Task"], [1157, 1158, ["computation"], "Task"], [1169, 1176, ["Global", "Average", "Pooling", "(", "GAP", ")", "layer"], "Method"], [1198, 1199, ["CNNs"], "Method"], [1202, 1204, ["VDCNN", "Model"], "Method"], [1205, 1207, ["Text", "Classification"], "Task"], [1208, 1209, ["VDCNN"], "Method"], [1211, 1213, ["modular", "architecture"], "Method"], [1214, 1217, ["text", "classification", "tasks"], "Task"], [1242, 1244, ["depth", "9"], "Task"], [1290, 1292, ["embedding", "dimension"], "Metric"], [1318, 1320, ["temporal", "convolutions"], "Method"], [1344, 1348, ["modular", "network", "segment", "input"], "Method"], [1350, 1352, ["convolutional", "blocks"], "Method"], [1361, 1364, ["temporal", "convolutional", "layers"], "Method"], [1370, 1374, ["temporal", "batch", "normalization", "layer"], "Method"], [1376, 1378, ["ReLU", "activation"], "Method"], [1451, 1453, ["convolutional", "layer"], "Method"], [1464, 1466, ["depth", "architectures"], "Method"], [1468, 1470, ["VDCNN", "model"], "Method"], [1493, 1495, ["convolutional", "block"], "Method"], [1503, 1505, ["pooling", "layer"], "Method"], [1516, 1520, ["VGG", "and", "ResNets", "philosophy"], "Method"], [1546, 1548, ["VDCNN", "network"], "Method"], [1554, 1556, ["convolutional", "blocks"], "Method"], [1561, 1562, ["convolutions"], "Method"], [1567, 1569, ["classification", "task"], "Task"], [1577, 1580, ["-", "max", "pooling"], "Method"], [1616, 1617, ["dropout"], "Method"], [1619, 1621, ["batch", "normalization"], "Method"], [1622, 1624, ["convolutional", "layers"], "Method"], [1626, 1628, ["network", "regularization"], "Method"], [1631, 1633, ["SVDCNN", "Model"], "Method"], [1634, 1636, ["Text", "Classification"], "Task"], [1654, 1656, ["storage", "size"], "Metric"], [1663, 1668, ["convolutional", "blocks", "of", "VDCNN", "model"], "Method"], [1672, 1676, ["Temporal", "Depthwise", "Separable", "Convolutions"], "Method"], [1677, 1678, ["TDSCs"], "Method"], [1692, 1699, ["Global", "Average", "Pooling", "(", "GAP", ")", "technique"], "Method"], [1706, 1712, ["Squeezed", "Very", "Deep", "Convolutional", "Neural", "Networks"], "Method"], [1713, 1714, ["SVDCNN"], "Method"], [1718, 1722, ["Temporal", "Depthwise", "Separable", "Convolutions"], "Method"], [1723, 1724, ["TDSCs"], "Method"], [1728, 1729, ["TDSCs"], "Method"], [1731, 1732, ["convolutions"], "Method"], [1740, 1742, ["accuracy", "loss"], "Metric"], [1743, 1744, ["TDSCs"], "Method"], [1748, 1749, ["convolution"], "Method"], [1765, 1767, ["convolutional", "filter"], "Method"], [1832, 1833, ["convolutions"], "Method"], [1849, 1850, ["TDSCs"], "Method"], [1851, 1852, ["DSCs"], "Method"], [1855, 1859, ["one", "-", "dimensional", "convolutions"], "Method"], [1861, 1862, ["DSCs"], "Method"], [1866, 1869, ["image", "classification", "networks"], "Task"], [1874, 1876, ["temporal", "version"], "Method"], [1877, 1880, ["text", "related", "tasks"], "Task"], [1892, 1895, ["temporal", "standard", "convolution"], "Method"], [1901, 1902, ["TDSC"], "Method"], [1917, 1920, ["temporal", "standard", "convolution"], "Method"], [1944, 1945, ["TDSC"], "Method"], [1953, 1955, ["VDCNN", "model"], "Method"], [1957, 1959, ["convolutional", "block"], "Method"], [1963, 1967, ["temporal", "standard", "convolutional", "layers"], "Method"], [1991, 1993, ["convolutional", "layer"], "Method"], [1997, 1999, ["Batch", "Normalization"], "Method"], [2001, 2003, ["ReLU", "layers"], "Method"], [2012, 2015, ["temporal", "standard", "convolutions"], "Method"], [2016, 2017, ["TDSCs"], "Method"], [2028, 2030, ["convolutional", "block"], "Method"], [2036, 2038, ["convolutional", "block"], "Method"], [2039, 2040, ["TDSC"], "Method"], [2052, 2054, ["convolutional", "layers"], "Method"], [2087, 2089, ["Temporal", "Convolution"], "Method"], [2133, 2135, ["convolutional", "block"], "Method"], [2139, 2140, ["TDSC"], "Method"], [2208, 2209, ["TDSC"], "Method"], [2217, 2219, ["network", "size"], "Metric"], [2225, 2227, ["temporal", "convolution"], "Method"], [2239, 2240, ["convolutions"], "Method"], [2241, 2243, ["convolutional", "block"], "Method"], [2294, 2296, ["VDCNN", "model"], "Method"], [2314, 2317, ["Global", "Average", "Pooling"], "Method"], [2318, 2319, ["GAP"], "Method"], [2321, 2323, ["VDCNN", "model"], "Method"], [2325, 2329, ["-", "max", "pooling", "layer"], "Method"], [2332, 2334, ["fully", "connected"], "Method"], [2335, 2336, ["FC"], "Method"], [2341, 2343, ["classification", "task"], "Task"], [2358, 2361, ["text", "classification", "CNNs"], "Task"], [2423, 2426, ["average", "pooling", "layer"], "Method"], [2432, 2435, ["Global", "Average", "Pooling"], "Method"], [2440, 2442, ["parameters", "reduction"], "Task"], [2445, 2447, ["network", "accuracy"], "Metric"], [2458, 2461, ["average", "pooling", "layer"], "Method"], [2467, 2471, ["-", "max", "pooling", "layer"], "Method"], [2489, 2491, ["GAP", "("], "Method"], [2523, 2525, ["classification", "method"], "Method"], [2529, 2530, ["GAP"], "Method"], [2556, 2557, ["TDSCs"], "Method"], [2559, 2560, ["GAP"], "Method"], [2570, 2573, ["Char", "-", "CNN"], "Method"], [2574, 2575, ["VDCNN"], "Method"], [2577, 2578, ["SVDCNN"], "Method"], [2585, 2587, ["storage", "size"], "Metric"], [2588, 2590, ["inference", "time"], "Metric"], [2591, 2592, ["accuracy"], "Metric"], [2609, 2610, ["VDCNN"], "Method"], [2618, 2620, ["convolutional", "layers"], "Method"], [2630, 2631, ["SVDCNN"], "Method"], [2632, 2635, ["Char", "-", "CNN"], "Method"], [2643, 2645, ["network", "architecture"], "Method"], [2647, 2648, ["PyTorch"], "Method"], [2679, 2680, ["CNN"], "Method"], [2717, 2719, ["inference", "time"], "Metric"], [2721, 2722, ["average"], "Metric"], [2723, 2725, ["standard", "deviation"], "Metric"], [2747, 2748, ["SVDCNN"], "Method"], [2776, 2777, ["SGD"], "Method"], [2795, 2797, ["learning", "rate"], "Metric"], [2818, 2820, ["NVIDIA", "GTX"], "Material"], [2821, 2822, ["GPU"], "Method"], [2872, 2875, ["AG", "’s", "News"], "Material"], [2876, 2878, ["Yelp", "Polarity"], "Material"], [2879, 2881, ["Yelp", "Full"], "Material"], [2895, 2897, ["storage", "size"], "Metric"], [2899, 2900, ["accuracy"], "Metric"], [2902, 2903, ["SVDCNN"], "Method"], [2904, 2905, ["VDCNN"], "Method"], [2907, 2910, ["Char", "-", "CNN"], "Method"], [2917, 2918, ["TDSCs"], "Method"], [2923, 2925, ["convolutional", "parameters"], "Metric"], [2927, 2928, ["VDCNN"], "Method"], [2932, 2936, ["in", "-", "depth", "network"], "Task"], [2954, 2956, ["convolutional", "layers"], "Method"], [2981, 2984, ["Char", "-", "CNN"], "Method"], [2995, 2997, ["convolutional", "layers"], "Method"], [3003, 3005, ["network", "reduction"], "Task"], [3008, 3009, ["GAP"], "Method"], [3023, 3025, ["classification", "tasks"], "Task"], [3036, 3037, ["SVDCNN"], "Method"], [3038, 3039, ["VDCNN"], "Method"], [3072, 3075, ["Char", "-", "CNN"], "Method"], [3102, 3107, ["storage", "size", "of", "the", "networks"], "Task"], [3111, 3115, ["in", "-", "depth", "model"], "Method"], [3123, 3124, ["VDCNN"], "Method"], [3136, 3139, ["Char", "-", "CNN"], "Method"], [3157, 3159, ["embedded", "platforms"], "Method"], [3189, 3190, ["accuracy"], "Metric"], [3198, 3200, ["parameter", "reduction"], "Method"], [3205, 3206, ["accuracy"], "Metric"], [3219, 3220, ["VDCNN"], "Method"], [3221, 3222, ["SVDCNN"], "Method"], [3238, 3241, ["storage", "size", "reduction"], "Task"], [3255, 3257, ["accuracy", "scores"], "Metric"], [3275, 3277, ["VDCNN", "model"], "Method"], [3281, 3283, ["squeezed", "model"], "Method"], [3306, 3308, ["Yelp", "Review"], "Material"], [3316, 3317, ["accuracy"], "Metric"], [3328, 3332, ["Deep", "learning", "processing", "architecture"], "Method"], [3373, 3375, ["network", "architecture"], "Method"], [3428, 3430, ["inference", "time"], "Metric"], [3438, 3440, ["dedicated", "hardware"], "Method"], [3441, 3443, ["deep", "learning"], "Task"], [3464, 3466, ["mobile", "platform"], "Task"], [3471, 3473, ["inference", "times"], "Metric"], [3495, 3497, ["mobile", "platforms"], "Task"], [3499, 3501, ["latency", "ratio"], "Metric"], [3504, 3507, ["GPU", "inference", "times"], "Metric"], [3525, 3527, ["inference", "times"], "Metric"], [3550, 3552, ["convolutional", "layer"], "Method"], [3554, 3556, ["convolutional", "blocks"], "Method"], [3560, 3561, ["convolutions"], "Method"], [3567, 3569, ["inference", "time"], "Metric"], [3574, 3576, ["parameter", "reduction"], "Method"], [3578, 3579, ["SVDCNN"], "Method"], [3585, 3587, ["VDCNN", "model"], "Method"], [3589, 3592, ["CPU", "inference", "time"], "Metric"], [3650, 3652, ["mobile", "platforms"], "Task"], [3655, 3658, ["Char", "-", "CNN"], "Method"], [3674, 3677, ["CPU", "inference", "time"], "Metric"], [3678, 3681, ["Ratio", "of", "0.03"], "Metric"], [3692, 3694, ["squeezed", "version"], "Method"], [3696, 3698, ["VDCNN", "model"], "Method"], [3714, 3716, ["mobile", "platforms"], "Task"], [3728, 3732, ["Temporal", "Depthwise", "Separable", "Convolutions"], "Method"], [3734, 3738, ["Global", "Average", "Pooling", "layer"], "Method"], [3741, 3745, ["deep", "convolutional", "neural", "network"], "Method"], [3746, 3748, ["text", "classification"], "Task"], [3761, 3763, ["storage", "size"], "Metric"], [3766, 3769, ["inference", "time", "ratio"], "Metric"], [3770, 3773, ["CPU", "/", "GPU"], "Metric"], [3797, 3799, ["model", "compression"], "Method"], [3803, 3805, ["model", "accuracy"], "Metric"], [279, 280, ["CNN"], "Method"], [323, 324, ["VDCNN"], "Method"], [875, 877, ["fully", "connected"], "Method"], [912, 913, ["CNN"], "Method"], [1017, 1019, ["fully", "connected"], "Method"], [1184, 1186, ["fully", "connected"], "Method"], [1590, 1592, ["fully", "connected"], "Method"], [1687, 1689, ["fully", "connected"], "Method"], [2378, 2379, ["FC"], "Method"], [2404, 2406, ["fully", "connected"], "Method"], [2554, 2555, ["VDCNN"], "Method"], [2564, 2566, ["fully", "connected"], "Method"], [2606, 2607, ["SVDCNN"], "Method"], [2652, 2653, ["FC"], "Method"], [2675, 2676, ["FC"], "Method"], [2736, 2739, ["AG", "’s", "News"], "Material"], [2755, 2756, ["VDCNN"], "Method"], [2855, 2856, ["CNN"], "Method"], [2857, 2858, ["VDCNN"], "Method"], [3019, 3020, ["FC"], "Method"], [3046, 3047, ["FC"], "Method"], [3089, 3090, ["FC"], "Method"], [3321, 3322, ["CNN"], "Method"], [3750, 3751, ["SVDCNN"], "Method"]]} {"doc_id": "3070a1bd503c3767def898bbd50c7eea2bbf29c9", "text": ["document", ":", "Wider", "or", "Deeper", ":", "Revisiting", "the", "ResNet", "Model", "for", "Visual", "Recognition", "The", "trend", "towards", "increasingly", "deep", "neural", "networks", "has", "been", "driven", "by", "a", "general", "observation", "that", "increasing", "depth", "increases", "the", "performance", "of", "a", "network", ".", "Recently", ",", "however", ",", "evidence", "has", "been", "amassing", "that", "simply", "increasing", "depth", "may", "not", "be", "the", "best", "way", "to", "increase", "performance", ",", "particularly", "given", "other", "limitations", ".", "Investigations", "into", "deep", "residual", "networks", "have", "also", "suggested", "that", "they", "may", "not", "in", "fact", "be", "operating", "as", "a", "single", "deep", "network", ",", "but", "rather", "as", "an", "ensemble", "of", "many", "relatively", "shallow", "networks", ".", "We", "examine", "these", "issues", ",", "and", "in", "doing", "so", "arrive", "at", "a", "new", "interpretation", "of", "the", "unravelled", "view", "of", "deep", "residual", "networks", "which", "explains", "some", "of", "the", "behaviours", "that", "have", "been", "observed", "experimentally", ".", "As", "a", "result", ",", "we", "are", "able", "to", "derive", "a", "new", ",", "shallower", ",", "architecture", "of", "residual", "networks", "which", "significantly", "outperforms", "much", "deeper", "models", "such", "as", "ResNet", "-", "200", "on", "the", "ImageNet", "classification", "dataset", ".", "We", "also", "show", "that", "this", "performance", "is", "transferable", "to", "other", "problem", "domains", "by", "developing", "a", "semantic", "segmentation", "approach", "which", "outperforms", "the", "state", "-", "of", "-", "the", "-", "art", "by", "a", "remarkable", "margin", "on", "datasets", "including", "PASCAL", "VOC", ",", "PASCAL", "Context", ",", "and", "Cityscapes", ".", "The", "architecture", "that", "we", "propose", "thus", "outperforms", "its", "comparators", ",", "including", "very", "deep", "ResNets", ",", "and", "yet", "is", "more", "efficient", "in", "memory", "use", "and", "sometimes", "also", "in", "training", "time", ".", "The", "code", "and", "models", "are", "available", "at", ".", "section", ":", "Introduction", "The", "convolutional", "networks", "used", "by", "the", "computer", "vision", "community", "have", "been", "growing", "deeper", "and", "deeper", "each", "year", "since", "Krizhevsky", "et", "al", ".", "proposed", "AlexNet", "in", "2012", ".", "The", "deepest", "network", "in", "the", "literature", "is", "a", "residual", "network", "(", "ResNet", ")", "with", "1", ",", "202", "trainable", "layers", ",", "which", "was", "trained", "using", "the", "tiny", "images", "in", "the", "CIFAR", "-", "10", "dataset", ".", "The", "image", "size", "here", "is", "important", ",", "because", "it", "means", "that", "the", "size", "of", "corresponding", "feature", "maps", "is", "relatively", "small", ",", "which", "is", "critical", "in", "training", "extremely", "deep", "models", ".", "Most", "networks", "operating", "on", "more", "practically", "interesting", "image", "sizes", "tend", "to", "have", "the", "order", "of", "one", ",", "to", "two", ",", "hundred", "layers", ",", "the", "200", "-", "layer", "ResNet", "and", "96", "-", "layer", "Inception", "-", "ResNet", ".", "The", "progression", "to", "deeper", "networks", "continues", ",", "however", ",", "with", "Zhao", "et", "al", ".", "having", "trained", "a", "269", "-", "layer", "network", "for", "semantic", "image", "segmentation", ".", "These", "networks", "were", "trained", "using", "the", "ImageNet", "classification", "dataset", ",", "where", "the", "images", "are", "of", "much", "higher", "resolution", ".", "Each", "additional", "layer", "requires", "not", "only", "additional", "memory", ",", "but", "also", "additional", "training", ".", "The", "marginal", "gains", "achieved", "by", "each", "additional", "layer", "diminish", "with", "depth", ",", "however", ",", "to", "the", "point", "where", "Zhao", "et", "al", ".", "achieved", "only", "an", "improvement", "of", "1.1", "%", "(", "from", "42.2", "%", "to", "43.3", "%", "by", "mean", "intersection", "-", "over", "-", "union", "scores", ")", "after", "almost", "doubling", "the", "number", "of", "layers", "(", "from", "152", "to", "269", ")", ".", "On", "the", "other", "hand", ",", "Zagoruyko", "and", "Komodakis", "showed", "that", "it", "is", "possible", "to", "train", "much", "shallower", "but", "wider", "networks", "on", "CIFAR", "-", "10", ",", "which", "outperform", "a", "ResNet", "with", "its", "more", "than", "one", "thousand", "layers", ".", "The", "question", "thus", "naturally", "arises", "as", "to", "whether", "deep", ",", "or", "wide", ",", "is", "the", "right", "strategy", ".", "In", "order", "to", "examine", "the", "issue", "we", "first", "need", "to", "understand", "the", "mechanism", "behind", "ResNets", ".", "Veit", "et", "al", ".", "have", "claimed", "that", "they", "actually", "behave", "as", "exponential", "ensembles", "of", "relatively", "shallow", "networks", ".", "However", ",", "there", "is", "a", "gap", "between", "their", "proposed", "unravelled", "view", "of", "a", "ResNet", ",", "and", "a", "real", "exponential", "ensemble", "of", "sub", "-", "networks", ",", "as", "illustrated", "in", "the", "top", "row", "of", "Fig", ".", "[", "reference", "]", ".", "Since", "the", "residual", "units", "are", "non", "-", "linear", ",", "we", "can", "not", "further", "split", "the", "bottom", "path", "into", "two", "sub", "-", "networks", ",", "i.e.", ",", "and", ".", "It", "turns", "out", "that", "ResNets", "are", "only", "assembling", "linearly", "growing", "numbers", "of", "sub", "-", "networks", ".", "Besides", ",", "the", "key", "characteristic", "of", "our", "introduced", "view", "is", "that", "it", "depends", "on", "the", "effective", "depth", "of", "a", "network", ".", "This", "amounts", "to", "the", "number", "of", "residual", "units", "which", "backward", "gradients", "during", "training", "can", "go", "through", ".", "When", ",", "the", "two", "-", "unit", "ResNet", "in", "Fig", ".", "[", "reference", "]", "can", "be", "seen", "as", "an", "ensemble", "of", "three", "sub", "-", "networks", ",", "i.e.", ",", ",", ",", "and", ",", "as", "shown", "in", "the", "bottom", "left", ".", "When", ",", "nothing", "changes", "except", "that", "we", "replace", "the", "third", "sub", "-", "network", "with", "a", "shallower", "one", ",", "as", "shown", "in", "the", "bottom", "right", "example", ".", "The", "superscripts", "in", "and", "denote", "their", "actual", "depths", ".", "About", "the", "unravelled", "view", ",", "the", "effective", "depth", "of", "a", "ResNet", ",", "and", "the", "actual", "depth", "of", "a", "sub", "-", "network", ",", "more", "details", "will", "be", "provided", "in", "the", "sequence", ".", "It", "is", "also", "worth", "noting", "that", "Veit", "et", "al", ".", "empirically", "found", "that", "most", "gradients", "in", "a", "110", "-", "layer", "ResNet", "can", "only", "go", "through", "up", "to", "seventeen", "residual", "units", ",", "which", "supports", "our", "above", "hypothesis", "that", "the", "effective", "depth", "exists", "for", "a", "specific", "network", ".", "In", "this", "paper", ",", "our", "contributions", "include", ":", "We", "introduce", "a", "further", "developed", "intuitive", "view", "of", "ResNets", ",", "which", "helps", "us", "understand", "their", "behaviours", ",", "and", "find", "possible", "directions", "to", "further", "improvements", ".", "We", "propose", "a", "group", "of", "relatively", "shallow", "convolutional", "networks", "based", "on", "our", "new", "understanding", ".", "Some", "of", "them", "achieve", "the", "state", "-", "of", "-", "the", "-", "art", "results", "on", "the", "ImageNet", "classification", "dataset", ".", "We", "evaluate", "the", "impact", "of", "using", "different", "networks", "on", "the", "performance", "of", "semantic", "image", "segmentation", ",", "and", "show", "these", "networks", ",", "as", "pre", "-", "trained", "features", ",", "can", "boost", "existing", "algorithms", "a", "lot", ".", "We", "achieve", "the", "best", "results", "on", "PASCAL", "VOC", ",", "PASCAL", "Context", ",", "and", "Cityscapes", ".", "section", ":", "Related", "work", "Our", "work", "here", "is", "closely", "related", "to", "two", "topics", ",", "residual", "network", "(", "ResNet", ")", "based", "image", "classification", "and", "semantic", "image", "segmentation", "using", "fully", "convolutional", "networks", ".", "As", "we", "have", "noted", "above", ",", "He", "et", "al", ".", "recently", "proposed", "the", "ResNets", "to", "combat", "the", "vanishing", "gradient", "problem", "during", "training", "very", "deep", "convolutional", "networks", ".", "ResNets", "have", "outperformed", "previous", "models", "at", "a", "variety", "of", "tasks", ",", "such", "as", "object", "detection", "and", "semantic", "image", "segmentation", ".", "They", "are", "gradually", "replacing", "VGGNets", "in", "the", "computer", "vision", "community", ",", "as", "the", "standard", "feature", "extractors", ".", "Nevertheless", ",", "the", "real", "mechanism", "underpinning", "the", "effectiveness", "of", "ResNets", "is", "not", "yet", "clear", ".", "Veit", "et", "al", ".", "claimed", "that", "they", "behave", "like", "exponential", "ensembles", "of", "relatively", "shallow", "networks", ",", "yet", "the", "‘", "exponential", "’", "nature", "of", "the", "ensembles", "has", "yet", "to", "be", "theoretically", "verified", ".", "Residual", "units", "are", "usually", "non", "-", "linear", ",", "which", "prevents", "a", "ResNet", "from", "exponentially", "expanding", "into", "separated", "sub", "-", "networks", ",", "as", "illustrated", "in", "Fig", ".", "[", "reference", "]", ".", "It", "is", "also", "unclear", "as", "to", "whether", "a", "residual", "structure", "is", "required", "to", "train", "very", "deep", "networks", ".", "For", "example", ",", "Szegedy", "et", "al", ".", "showed", "that", "it", "is", "‘", "not", "very", "difficult", "’", "to", "train", "competitively", "deep", "networks", ",", "even", "without", "residual", "shortcuts", ".", "Currently", ",", "the", "most", "clear", "advantage", "of", "ResNets", "is", "in", "their", "fast", "convergence", ".", "Szegedy", "et", "al", ".", "observed", "similar", "empirically", "results", "to", "support", "that", ".", "On", "the", "other", "hand", ",", "Zagoruyko", "and", "Komodakis", "found", "that", "a", "wide", "sixteen", "-", "layer", "ResNet", "outperformed", "the", "original", "thin", "thousand", "-", "layer", "ResNet", "on", "datasets", "composed", "of", "tiny", "images", "such", "as", "CIFAR", "-", "10", ".", "The", "analysis", "we", "present", "here", "is", "motivated", "by", "their", "empirical", "testing", ",", "but", "aims", "at", "a", "more", "theoretical", "approach", ",", "and", "the", "observation", "that", "a", "grid", "search", "of", "configuration", "space", "is", "impractical", "on", "large", "scale", "datasets", "such", "as", "the", "ImageNet", "classification", "dataset", ".", "Semantic", "image", "segmentation", "amounts", "to", "predicting", "the", "categories", "for", "each", "pixel", "in", "an", "image", ".", "Long", "et", "al", ".", "proposed", "the", "fully", "convolutional", "networks", "(", "FCN", ")", "to", "this", "end", ".", "FCNs", "soon", "became", "the", "mainstream", "approach", "to", "dense", "prediction", "based", "tasks", ",", "especially", "due", "to", "its", "efficiency", ".", "Besides", ",", "empirical", "results", "in", "the", "literature", "showed", "that", "stronger", "pre", "-", "trained", "features", "can", "yet", "further", "improve", "their", "performance", ".", "We", "thus", "here", "base", "our", "semantic", "image", "segmentation", "approach", "on", "fully", "convolutional", "networks", ",", "and", "will", "show", "the", "impact", "of", "different", "pre", "-", "trained", "features", "on", "final", "segmentation", "results", ".", "section", ":", "Residual", "networks", "revisited", "We", "are", "concerned", "here", "with", "the", "full", "pre", "-", "activation", "version", "of", "residual", "networks", "(", "ResNet", ")", ".", "For", "shortcut", "connections", ",", "we", "consider", "identity", "mappings", "only", ".", "We", "omit", "the", "raw", "input", "and", "the", "top", "-", "most", "linear", "classifier", "for", "clarity", ".", "Usually", ",", "there", "may", "be", "a", "stem", "block", "or", "several", "traditional", "convolution", "layers", "directly", "after", "the", "raw", "input", ".", "We", "omit", "these", "also", ",", "for", "the", "purpose", "of", "simplicity", ".", "For", "the", "residual", "Unit", ",", "let", "be", "the", "input", ",", "and", "let", "be", "its", "trainable", "non", "-", "linear", "mappings", ",", "also", "named", "Block", ".", "The", "output", "of", "Unit", "is", "recursively", "defined", "as", ":", "where", "denotes", "the", "trainable", "parameters", ",", "and", "is", "often", "two", "or", "three", "stacked", "convolution", "stages", ".", "In", "the", "full", "pre", "-", "activation", "version", ",", "the", "components", "of", "a", "stage", "are", "in", "turn", "a", "batch", "normalization", ",", "a", "rectified", "linear", "unit", "(", "ReLU", ")", "non", "-", "linearity", ",", "and", "a", "convolution", "layer", ".", "subsection", ":", "Residual", "networks", "unravelled", "online", "Applying", "Eqn", ".", "(", "[", "reference", "]", ")", "in", "one", "substitution", "step", ",", "we", "expand", "the", "forward", "pass", "into", ":", "which", "describes", "the", "unravelled", "view", "by", "Veit", "et", "al", ".", ",", "as", "shown", "in", "the", "top", "row", "of", "Fig", ".", "[", "reference", "]", ".", "Since", "is", "non", "-", "linear", ",", "we", "can", "not", "derive", "Eqn", ".", "(", "[", "reference", "]", ")", "from", "Eqn", ".", "(", "[", "reference", "]", ")", ".", "So", "the", "whole", "network", "is", "not", "equivalent", "to", "an", "exponentially", "growing", "ensemble", "of", "sub", "-", "networks", ".", "It", "is", "rather", ",", "more", "accurately", ",", "described", "as", "a", "linearly", "growing", "ensemble", "of", "sub", "-", "networks", ".", "For", "the", "two", "-", "unit", "ResNet", "as", "illustrated", "in", "Fig", ".", "[", "reference", "]", ",", "there", "are", "three", ",", "e.g.", ",", ",", ",", "and", ",", "sub", "-", "networks", "respectively", "corresponding", "to", "the", "three", "terms", "in", "Eqn", ".", "(", "[", "reference", "]", ")", ",", "i.e.", ",", ",", ",", "and", ".", "Veit", "et", "al", ".", "in", "showed", "that", "the", "paths", "which", "gradients", "take", "through", "a", "ResNet", "are", "typically", "far", "shorter", "than", "the", "total", "depth", "of", "that", "network", ".", "They", "thus", "introduced", "the", "idea", "of", "effective", "depth", "as", "a", "measure", "for", "the", "true", "length", "of", "these", "paths", ".", "By", "characterising", "the", "units", "of", "a", "ResNet", "given", "its", "effective", "depth", ",", "we", "illuminate", "the", "impact", "of", "varying", "paths", "that", "gradients", "actually", "take", ",", "as", "in", "Fig", ".", "[", "reference", "]", ".", "We", "here", "illustrate", "this", "impact", "in", "terms", "of", "small", "effective", "depths", ",", "because", "to", "do", "so", "for", "larger", "ones", "would", "require", "diagrams", "of", "enormous", "networks", ".", "The", "impact", "is", "the", "same", ",", "however", ".", "Take", "the", "ResNet", "in", "Fig", ".", "[", "reference", "]", "for", "example", "again", ".", "In", "an", "SGD", "iteration", ",", "the", "backward", "gradients", "are", ":", "where", "denotes", "the", "derivative", "of", "to", "its", "input", ".", "When", "effective", "depth", ",", "both", "terms", "in", "Eqn", ".", "(", "[", "reference", "]", ")", "are", "non", "-", "zeros", ",", "which", "corresponds", "to", "the", "bottom", "-", "left", "case", "in", "Fig", ".", "[", "reference", "]", ".", "Namely", ",", "Block", "1", "receives", "gradients", "from", "both", "and", ".", "However", ",", "when", "effective", "depth", ",", "the", "gradient", "vanishes", "after", "passing", "through", "Block", "2", ".", "Namely", ",", ".", "So", ",", "the", "second", "term", "in", "Eqn", ".", "(", "[", "reference", "]", ")", "also", "goes", "to", "zeros", ",", "which", "is", "illustrated", "by", "the", "bottom", "-", "right", "case", "in", "Fig", ".", "[", "reference", "]", ".", "The", "weights", "in", "Block", "1", "indeed", "vary", "across", "different", "iterations", ",", "but", "they", "are", "updated", "only", "by", ".", "To", ",", "Block", "1", "is", "no", "more", "than", "an", "additional", "input", "providing", "preprocessed", "representations", ",", "because", "Block", "1", "is", "not", "end", "-", "to", "-", "end", "trained", ",", "from", "the", "point", "of", "view", "of", ".", "In", "this", "case", ",", "we", "name", "to", "have", "an", "actual", "depth", "of", "one", ".", "We", "say", "that", "the", "ResNet", "is", "over", "-", "deepened", ",", "and", "that", "it", "can", "not", "be", "trained", "in", "a", "fully", "end", "-", "to", "-", "end", "manner", ",", "even", "with", "those", "shortcut", "connections", ".", "Let", "be", "the", "total", "number", "of", "residual", "units", ".", "We", "can", "see", "a", "ResNet", "as", "an", "ensemble", "of", "different", "sub", "-", "networks", ",", "i.e.", ",", ".", "The", "actual", "depth", "of", "is", ".", "We", "show", "an", "unravelled", "three", "-", "unit", "ResNet", "with", "different", "effective", "depths", "in", "Fig", ".", "[", "reference", "]", ".", "By", "way", "of", "example", ",", "note", "that", "in", "Fig", ".", "[", "reference", "]", "contains", "only", "Block", "1", ",", "whereas", "contains", "both", "Block", "1", "and", "Block", "2", ".", "Among", "the", "cases", "illustrated", ",", "the", "bottom", "left", "example", "is", "more", "complicated", ",", "where", "and", ".", "From", "the", "point", "of", "view", "of", ",", "the", "gradient", "of", "Block", "1", "is", ",", "where", "the", "first", "term", "is", "non", "-", "zero", ".", "will", "thus", "update", "Block", "1", "at", "each", "iteration", ".", "Considering", "the", "non", "-", "linearity", "in", "Block", "3", ",", "it", "is", "non", "-", "trivial", "to", "tell", "if", "this", "is", "as", "good", "as", "the", "fully", "end", "-", "to", "-", "end", "training", "case", ",", "as", "illustrated", "by", "in", "the", "top", "right", "example", ".", "An", "investigation", "of", "this", "issue", "remains", "future", "work", ".", "subsection", ":", "Residual", "networks", "behaviours", "revisited", "Very", "deep", "ResNets", ".", "Conventionally", ",", "it", "is", "not", "easy", "to", "train", "very", "deep", "networks", "due", "to", "the", "vanishing", "gradient", "problem", ".", "To", "understand", "how", "a", "very", "deep", "ResNet", "is", "trained", ",", "the", "observation", "by", "Veit", "et", "al", ".", "is", "important", ",", "i.e.", ",", "gradients", "vanish", "exponentially", "as", "the", "length", "of", "paths", "increases", ".", "Now", "refer", "to", "the", "top", "-", "right", "example", "in", "Fig", ".", "[", "reference", "]", ".", "This", "is", "somewhat", "similar", "to", "the", "case", "of", "a", "shallow", "or", "reasonably", "deep", "ResNet", ",", "when", ".", "At", "the", "beginning", ",", "the", "shallowest", "sub", "-", "network", ",", "i.e.", ",", ",", "converges", "fast", ",", "because", "it", "gives", "Block", "1", "the", "largest", "gradients", ".", "From", "the", "point", "of", "view", "of", ",", "Block", "2", "may", "also", "receive", "large", "gradients", "due", "to", "the", "path", "with", "a", "length", "of", "one", ".", "However", ",", "the", "input", "of", "Block", "2", "partly", "depends", "on", "Block", "1", ".", "It", "would", "not", "be", "easy", "for", "Block", "2", "to", "converge", "before", "the", "output", "of", "Block", "1", "stabilises", ".", "Similarly", ",", "Block", "3", "will", "need", "to", "wait", "for", "Blocks", "1", "and", "2", ",", "and", "so", "forth", ".", "In", "this", "way", ",", "a", "ResNet", "seems", "like", "an", "ensemble", "with", "a", "growing", "number", "of", "sub", "-", "networks", ".", "Besides", ",", "each", "newly", "added", "sub", "-", "network", "will", "have", "a", "larger", "actual", "depth", "than", "all", "the", "previous", "ones", ".", "Note", "that", "Littwin", "and", "Wolf", ",", "in", "a", "concurrent", "work", ",", "have", "theoretically", "showed", "that", "ResNets", "are", "virtual", "ensembles", "whose", "depth", "grows", "as", "training", "progresses", ".", "Their", "result", "to", "some", "extent", "coincides", "with", "the", "above", "described", "process", ".", "The", "story", "will", "however", "be", "different", "when", "the", "actual", "depth", "becomes", "as", "large", "as", "the", "effective", "depth", ".", "Refer", "to", "the", "bottom", "-", "right", "example", "in", "Fig", ".", "[", "reference", "]", ".", "This", "is", "somewhat", "similar", "to", "the", "case", "of", "an", "over", "-", "deepened", "ResNet", ",", "when", "is", "much", "larger", "than", ".", "Again", ",", "Block", "1", "in", "gets", "trained", "and", "stabilises", "first", ".", "However", ",", "this", "time", "is", "not", "fully", "end", "-", "to", "-", "end", "trained", "any", "more", ".", "Since", "gives", "no", "gradients", "to", "Block", "1", ",", "it", "becomes", "a", "one", "-", "block", "sub", "-", "network", "trained", "on", "top", "of", "some", "preprocessed", "representations", ",", "which", "are", "obtained", "by", "adding", "the", "output", "of", "Block", "1", "up", "to", "the", "original", "input", ".", "In", "this", "way", ",", "the", "newly", "added", "sub", "-", "network", "still", "has", "an", "actual", "depth", "of", "one", ",", "which", "is", "no", "deeper", "than", "the", "previous", "one", ",", "i.e.", ",", ",", "and", "so", "forth", "for", ".", "ResNets", "thus", "avoid", "the", "vanishing", "gradient", "problem", "by", "reshaping", "themselves", "into", "multiple", "shallower", "sub", "-", "networks", ".", "This", "is", "just", "another", "view", "of", "delivering", "gradients", "to", "bottom", "layers", "through", "shortcut", "connections", ".", "Researchers", "have", "claimed", "that", "the", "residual", "shortcut", "connections", "are", "not", "necessary", "even", "in", "very", "deep", "networks", ".", "However", ",", "there", "are", "usually", "short", "paths", "in", "their", "proposed", "networks", "as", "well", ".", "For", "example", ",", "the", "76", "-", "layer", "Inception", "-", "v4", "network", "has", "a", "much", "shorter", "twenty", "-", "layer", "route", "from", "its", "input", "to", "the", "output", ".", "There", "might", "be", "differences", "in", "the", "details", "between", "fusion", "by", "concatenation", "(", "Inception", "-", "v4", ")", "and", "fusion", "by", "summation", "(", "ResNets", ")", ".", "However", ",", "the", "manner", "of", "avoiding", "the", "vanishing", "gradient", "problem", "is", "probably", "similar", ",", "i.e.", ",", "using", "shortcuts", ",", "either", "with", "trainable", "weights", "or", "not", ".", "We", "are", "thus", "not", "yet", "in", "a", "position", "to", "be", "able", "to", "claim", "that", "the", "vanishing", "gradient", "problem", "has", "been", "solved", ".", "Wide", "ResNets", ".", "Conventionally", ",", "wide", "layers", "are", "more", "prone", "to", "over", "-", "fitting", ",", "and", "sometimes", "require", "extra", "regularization", "such", "as", "dropout", ".", "However", ",", "Zagoruyko", "and", "Komodakis", "showed", "the", "possibility", "to", "effectively", "train", "times", "wider", "ResNets", ",", "even", "without", "any", "extra", "regularization", ".", "To", "understand", "how", "a", "wide", "ResNet", "is", "trained", ",", "refer", "to", "the", "top", "right", "example", "in", "Fig", ".", "[", "reference", "]", ".", "This", "simulates", "the", "case", "of", "a", "rather", "shallow", "network", ",", "when", "is", "smaller", "than", ".", "We", "reuse", "the", "weights", "of", "Block", "1", "for", "four", "times", ".", "Among", "these", ",", "Block", "1", "is", "located", "in", "three", "different", "kinds", "of", "circumstances", ".", "In", "the", "bottom", "-", "most", "path", "of", "the", "sub", "-", "network", ",", "it", "is", "supposed", "to", "learn", "some", "low", "-", "level", "features", ";", "in", ",", "it", "should", "learn", "both", "low", "-", "level", "and", "mid", "-", "level", "features", ";", "and", "in", ",", "it", "has", "to", "learn", "everything", ".", "This", "format", "of", "weight", "sharing", "may", "suppress", "over", "-", "fitting", ",", "especially", "for", "those", "units", "far", "from", "the", "top", "-", "most", "linear", "classifier", ".", "Hence", "ResNets", "inherently", "introduce", "regularization", "by", "weight", "sharing", "among", "multiple", "very", "different", "sub", "-", "networks", ".", "Residual", "unit", "choices", ".", "For", "better", "performance", ",", "we", "hope", "that", "a", "ResNet", "should", "expand", "into", "a", "sufficiently", "large", "number", "of", "sub", "-", "networks", ",", "some", "of", "which", "should", "have", "large", "model", "capacity", ".", "So", ",", "given", "our", "previous", "observations", ",", "the", "requirements", "for", "an", "ideal", "mapping", "function", "in", "a", "residual", "unit", "are", ",", "1", ")", "being", "strong", "enough", "to", "converge", "even", "if", "it", "is", "reused", "in", "many", "sub", "-", "networks", ",", "and", "2", ")", "being", "shallow", "enough", "to", "enable", "an", "large", "effective", "depth", ".", "Since", "it", "is", "very", "hard", "to", "build", "a", "model", "with", "large", "capacity", "using", "a", "single", "trainable", "layer", ",", "the", "most", "natural", "choice", "would", "be", "a", "residual", "unit", "with", "two", "wide", "convolution", "stages", ".", "This", "coincides", "with", "empirical", "results", "reported", "by", "Zagoruyko", "and", "Komodakis", ".", "They", "found", "that", ",", "among", "the", "most", "trivial", "structure", "choices", ",", "the", "best", "one", "is", "to", "stack", "two", "convolution", "stages", ".", "subsection", ":", "Wider", "or", "deeper", "?", "To", "summarize", "the", "previous", "subsections", ",", "shortcut", "connections", "enable", "us", "to", "train", "wider", "and", "deeper", "networks", ".", "As", "they", "growing", "to", "some", "point", ",", "we", "will", "face", "the", "dilemma", "between", "width", "and", "depth", ".", "From", "that", "point", ",", "going", "deep", ",", "we", "will", "actually", "get", "a", "wider", "network", ",", "with", "extra", "features", "which", "are", "not", "completely", "end", "-", "to", "-", "end", "trained", ";", "going", "wider", ",", "we", "will", "literally", "get", "a", "wider", "network", ",", "without", "changing", "its", "end", "-", "to", "-", "end", "characteristic", ".", "We", "have", "learned", "the", "strength", "of", "depth", "from", "the", "previous", "plain", "deep", "networks", "without", "any", "shortcuts", ",", "e.g.", ",", "the", "AlexNet", "and", "VGGNets", ".", "However", ",", "it", "is", "not", "clear", "whether", "those", "extra", "features", "in", "very", "deep", "residual", "networks", "can", "perform", "as", "well", "as", "conventional", "fully", "end", "-", "to", "-", "end", "trained", "features", ".", "So", "in", "this", "paper", ",", "we", "only", "favour", "a", "deeper", "model", ",", "when", "it", "can", "be", "completely", "end", "-", "to", "-", "end", "trained", ".", "In", "practice", ",", "algorithms", "are", "often", "limited", "by", "their", "spatial", "costs", ".", "One", "way", "is", "to", "use", "more", "devices", ",", "which", "will", "however", "increase", "communication", "costs", "among", "them", ".", "With", "similar", "memory", "costs", ",", "a", "shallower", "but", "wider", "network", "can", "have", "times", "more", "number", "of", "trainable", "parameters", ".", "Therefore", ",", "given", "the", "following", "observations", "in", "the", "literature", ",", "Zagoruyko", "and", "Komodakis", "found", "that", "the", "performance", "of", "a", "ResNet", "was", "related", "to", "the", "number", "of", "trainable", "parameters", ".", "Szegedy", "et", "al", ".", "came", "to", "a", "similar", "conclusion", ",", "according", "to", "the", "comparison", "between", "their", "proposed", "Inception", "networks", ".", "Veit", "et", "al", ".", "found", "that", "there", "is", "a", "relatively", "small", "effective", "depth", "for", "a", "very", "deep", "ResNet", ",", "e.g.", ",", "seventeen", "residual", "units", "for", "a", "110", "-", "layer", "ResNet", ".", "most", "of", "the", "current", "state", "-", "of", "-", "the", "-", "art", "models", "on", "the", "ImageNet", "classification", "dataset", "seem", "over", "-", "deepened", ",", "e.g.", ",", "the", "200", "-", "layer", "ResNet", "and", "96", "-", "layer", "Inception", "-", "ResNet", ".", "The", "reason", "is", "that", ",", "to", "effectively", "utilize", "GPU", "memories", ",", "we", "should", "make", "a", "model", "shallow", ".", "According", "to", "our", "previous", "analysis", ",", "paths", "longer", "than", "the", "effective", "depth", "in", "ResNets", "are", "not", "trained", "in", "a", "fully", "end", "-", "to", "-", "end", "manner", ".", "Thus", ",", "we", "can", "remove", "most", "of", "these", "paths", "by", "directly", "reducing", "the", "number", "of", "residual", "units", ".", "For", "example", ",", "in", "our", "best", "performing", "network", ",", "there", "are", "exactly", "seventeen", "residual", "units", ".", "With", "empirical", "results", ",", "we", "will", "show", "that", "our", "fully", "end", "-", "to", "-", "end", "networks", "can", "perform", "much", "better", "than", "the", "previous", "much", "deeper", "ResNets", ",", "especially", "as", "feature", "extractors", ".", "However", ",", "even", "if", "a", "rather", "shallow", "network", "(", "eight", "-", "unit", ",", "or", "twenty", "-", "layer", ")", "can", "outperform", "ResNet", "-", "152", "on", "the", "ImageNet", "classification", "dataset", ",", "we", "will", "not", "go", "that", "shallow", ",", "because", "an", "appropriate", "depth", "is", "vital", "to", "train", "good", "features", ".", "section", ":", "Approach", "to", "image", "classification", "We", "show", "the", "proposed", "networks", "in", "Fig", ".", "[", "reference", "]", ".", "There", "are", "three", "architectures", ",", "with", "different", "input", "sizes", ".", "Dashed", "blue", "rectangles", "to", "denote", "convolution", "stages", ",", "which", "are", "respectively", "composed", "of", "a", "batch", "normalization", ",", "an", "ReLU", "non", "-", "linearity", "and", "a", "convolution", "layer", ",", "following", "the", "second", "version", "of", "ResNets", ".", "The", "closely", "stacked", "two", "or", "three", "convolution", "stages", "denote", "different", "kinds", "of", "residual", "units", "(", "B1–B7", ")", ",", "with", "inner", "shortcut", "connections", ".", "Each", "kind", "corresponds", "to", "a", "level", ",", "where", "all", "units", "share", "the", "same", "kernel", "sizes", "and", "numbers", "of", "channels", ",", "as", "given", "in", "the", "dashed", "black", "rectangles", "in", "the", "left", "-", "most", "column", "of", "Fig", ".", "[", "reference", "]", ".", "As", "mentioned", "before", ",", "there", "are", "two", "3", "3", "convolution", "layers", "in", "most", "residual", "units", "(", "B1–B5", ")", ".", "However", ",", "in", "B6", "and", "B7", ",", "we", "use", "bottleneck", "structures", "as", "in", "ResNets", ",", "except", "that", "we", "adjust", "the", "numbers", "of", "channels", "to", "avoid", "drastic", "changes", "in", "width", ".", "Each", "of", "our", "networks", "usually", "consists", "of", "one", "B6", ",", "one", "B7", ",", "and", "different", "numbers", "of", "B1–B5", ".", "For", "those", "with", "a", "224", "224", "input", ",", "we", "do", "not", "use", "B1", "due", "to", "limited", "GPU", "memories", ".", "Each", "of", "the", "green", "triangles", "denotes", "a", "down", "-", "sampling", "operation", "with", "a", "rate", "of", "two", ",", "which", "is", "clear", "given", "the", "feature", "map", "sizes", "of", "different", "convolution", "stages", "(", "in", "dashed", "blue", "rectangles", ")", ".", "To", "this", "end", ",", "we", "can", "let", "the", "first", "convolution", "layer", "at", "according", "levels", "have", "a", "stride", "of", "two", ".", "Or", ",", "we", "can", "use", "an", "extra", "spatial", "pooling", "layer", ",", "whose", "kernel", "size", "is", "three", "and", "stride", "is", "two", ".", "In", "a", "network", "whose", "classification", "results", "are", "reported", "in", "this", "paper", ",", "we", "always", "use", "pooling", "layers", "for", "down", "-", "sampling", ".", "We", "average", "the", "top", "-", "most", "feature", "maps", "into", "4", ",", "096", "-", "dimensional", "final", "features", ",", "which", "matches", "the", "cases", "of", "AlexNet", "and", "VGGNets", ".", "We", "will", "show", "more", "details", "about", "network", "structures", "in", "Subsection", "[", "reference", "]", ".", "Implementation", "details", ".", "We", "run", "all", "experiments", "using", "the", "MXNet", "framework", ",", "with", "four", "devices", "(", "two", "K80", "or", "four", "Maxwell", "Titan", "X", "cards", ")", "on", "a", "single", "node", ".", "We", "follow", "settings", "in", "the", "re", "-", "implementation", "of", "ResNets", "by", "Gross", "and", "Wilber", "as", "possible", ".", "But", ",", "we", "use", "a", "linear", "learning", "rate", "schedule", ",", "which", "was", "reported", "as", "a", "better", "choice", "by", "Mishkin", "et", "al", ".", ".", "Take", "Model", "A", "in", "Table", "[", "reference", "]", "for", "example", ".", "We", "start", "from", "0.1", ",", "and", "linearly", "reduce", "the", "learning", "rate", "to", "within", "450k", "iterations", ".", "section", ":", "Approach", "to", "semantic", "image", "segmentation", "Our", "approach", "is", "similar", "to", "the", "fully", "convolutional", "networks", "(", "FCN", ")", "implemented", "in", "the", "first", "version", "of", "DeepLab", ".", "However", ",", "without", "getting", "too", "many", "factors", "entangled", ",", "we", "in", "this", "paper", "do", "not", "introduce", "any", "multi", "-", "scale", "structures", ",", "deep", "supervision", "signals", ",", "or", "global", "context", "features", ".", "Besides", ",", "we", "do", "not", "apply", "any", "multi", "-", "scale", "testing", ",", "model", "averaging", "or", "CRF", "based", "post", "-", "processing", ",", "except", "for", "the", "test", "set", "of", "ADE20", "K", ".", "Given", "a", "pre", "-", "trained", "network", ",", "there", "are", "three", "steps", "to", "reshape", "it", "into", "a", "network", "suitable", "for", "semantic", "image", "segmentation", ",", "as", "stated", "below", ".", "1", ")", "Resolution", ".", "To", "generate", "score", "maps", "at", "1", "/", "8", "resolution", ",", "we", "remove", "down", "-", "sampling", "operations", "and", "increase", "dilation", "rates", "accordingly", "in", "some", "convolution", "layers", ".", "For", "clarity", ",", "first", "suppose", "that", "we", "always", "down", "-", "sample", "features", "maps", "using", "a", "convolution", "layer", "with", "a", "stride", "of", "two", ".", "Take", "networks", "with", "224", "224", "inputs", "for", "example", ".", "We", "set", "stride", "of", "the", "first", "convolution", "layer", "in", "B5", "to", "one", ",", "and", "increase", "the", "dilation", "rate", "from", "one", "to", "two", "for", "the", "following", "layers", ";", "We", "do", "the", "same", "thing", "to", "the", "first", "convolution", "layer", "in", "B6", "too", ",", "and", "increase", "the", "dilation", "rate", "from", "two", "to", "four", "for", "the", "following", "layers", ".", "In", "the", "case", "of", "down", "-", "sampling", "using", "a", "pooling", "layer", ",", "everything", "is", "the", "same", "except", "that", "we", "set", "stride", "of", "that", "pooling", "layer", "to", "one", ".", "Sometimes", ",", "we", "will", "have", "to", "apply", "a", "pooling", "layer", "with", "dilation", ".", "On", "the", "other", "hand", ",", "we", "do", "not", "make", "any", "change", "for", "networks", "with", "56", "56", "inputs", ",", "since", "there", "are", "only", "three", "down", "-", "sampling", "operations", "in", "each", "of", "them", ".", "It", "is", "notable", "that", "all", "down", "-", "sampling", "operations", "are", "implemented", "using", "spatial", "pooling", "layers", "in", "our", "originally", "pre", "-", "trained", "networks", ".", "We", "find", "it", "harmful", "for", "FCNs", "in", "our", "preliminary", "experiments", ",", "probably", "due", "to", "too", "strong", "spatial", "invariance", ".", "To", "this", "end", ",", "we", "replace", "several", "top", "-", "most", "down", "-", "sampling", "operations", "in", "a", "network", ",", "and", "then", "tune", "it", "for", "some", "additional", "iterations", ".", "Take", "Model", "A", "in", "Table", "[", "reference", "]", "for", "example", "again", ".", "We", "remove", "the", "top", "-", "most", "three", "pooling", "layers", "(", "before", "B4", ",", "B5", "and", "B6", ")", ",", "increase", "the", "strides", "of", "according", "convolution", "layers", "up", "to", "two", ",", "and", "tune", "it", "for", "45k", "iterations", "using", "the", "ImageNet", "dataset", ",", "starting", "from", "a", "learning", "rate", "of", "0.01", ".", "2", ")", "Classifier", ".", "We", "remove", "the", "top", "-", "most", "linear", "classifier", "and", "the", "global", "pooling", "layer", ",", "and", "then", "consider", "two", "cases", ".", "For", "one", "thing", ",", "we", "follow", "a", "basic", "large", "field", "-", "of", "-", "view", "setting", "in", "DeepLab", "-", "v2", ",", "called", "‘", "1", "convolution", "’", ".", "Namely", ",", "we", "just", "add", "back", "a", "single", "linear", "layer", "as", "the", "new", "classifier", ".", "For", "anther", ",", "we", "insert", "an", "additional", "non", "-", "linear", "convolution", "stage", "(", "without", "batch", "normalization", ")", "below", "the", "linear", "classifier", ".", "This", "case", "is", "called", "‘", "2", "convolutions", "’", ".", "Both", "of", "the", "added", "layers", "have", "3", "3", "kernels", ",", "with", "a", "dilation", "rate", "of", "twelve", ".", "The", "top", "-", "most", "two", "-", "layer", "classifier", "thus", "has", "a", "receptive", "field", "of", "392", "392", "on", "the", "final", "feature", "maps", ".", "By", "default", ",", "we", "let", "the", "number", "of", "channels", "in", "the", "hidden", "layer", "be", "512", ".", "3", ")", "Dropout", ".", "To", "alleviate", "over", "-", "fitting", ",", "we", "also", "apply", "the", "traditional", "dropout", "to", "very", "wide", "residual", "units", ".", "The", "dropout", "rate", "is", "0.3", "for", "those", "with", "2", ",", "048", "channels", ",", "e.g.", ",", "the", "last", "three", "units", "in", "ResNets", "and", "the", "second", "last", "units", "(", "B6", ")", "in", "our", "networks", ";", "while", "0.5", "for", "those", "with", "4", ",", "096", "channels", ",", "e.g.", ",", "the", "top", "-", "most", "units", "(", "B7", ")", "in", "our", "networks", ".", "Implementation", "details", ".", "We", "fix", "the", "moving", "means", "and", "variations", "in", "batch", "normalization", "layers", "during", "fine", "-", "tuning", ".", "We", "use", "four", "devices", "on", "a", "single", "node", ".", "The", "batch", "size", "is", "sixteen", ",", "so", "there", "are", "four", "examples", "per", "device", ".", "We", "first", "tune", "each", "network", "for", "a", "number", "of", "iterations", ",", "keeping", "the", "learning", "rate", "unchanged", "at", "0.0016", ".", "And", "then", ",", "we", "reduce", "the", "learning", "rate", "gradually", "during", "another", "number", "of", "iterations", ",", "following", "a", "linear", "schedule", ".", "For", "datasets", "with", "available", "testing", "sets", ",", "we", "evaluate", "these", "numbers", "of", "iterations", "on", "validation", "sets", ".", "During", "training", ",", "we", "first", "resize", "an", "image", "by", "a", "ratio", "randomly", "sampled", "from", ",", "and", "then", "generate", "a", "sample", "by", "cropping", "one", "500", "500", "sub", "-", "window", "at", "a", "randomly", "selected", "location", ".", "section", ":", "Experimental", "results", "subsection", ":", "Image", "classification", "results", "We", "evaluate", "our", "proposed", "networks", "on", "the", "ILSVRC", "2012", "classification", "dataset", ",", "with", "1.28", "million", "images", "for", "training", ",", "respectively", "belonging", "to", "1", ",", "000", "categories", ".", "We", "report", "top", "-", "1", "and", "top", "-", "5", "error", "rates", "on", "the", "validation", "set", ".", "We", "compare", "various", "networks", "in", "Table", "[", "reference", "]", ",", "where", "we", "obtain", "all", "the", "results", "by", "testing", "on", "a", "single", "crop", ".", "However", ",", "we", "list", "the", "ten", "-", "crop", "result", "for", "VGG16", "since", "it", "is", "not", "inherently", "a", "fully", "convolutional", "network", ".", "For", "networks", "trained", "with", "224", "224", "inputs", ",", "the", "testing", "crop", "size", "is", "320", "320", ",", "following", "the", "setting", "used", "by", "He", "et", "al", ".", ".", "For", "those", "with", "112", "112", "and", "56", "56", "inputs", ",", "we", "use", "160", "160", "and", "80", "80", "crops", "respectively", ".", "For", "Inception", "networks", ",", "the", "testing", "crop", "size", "is", "299", "299", ".", "The", "names", "of", "our", "proposed", "networks", "are", "composed", "of", "training", "crop", "sizes", "and", "the", "numbers", "of", "residual", "units", "on", "different", "levels", ".", "Take", "56", "-", "1", "-", "1", "-", "1", "-", "1", "-", "9", "-", "1", "-", "1", "for", "example", ".", "Its", "input", "size", "is", "56", ",", "and", "there", "are", "only", "one", "unit", "on", "all", "levels", "except", "for", "Level", "5", "(", "B5", "in", "Fig", ".", "[", "reference", "]", ")", ".", "Notable", "points", "about", "the", "results", "are", "as", "follows", ".", "1", ")", "Relatively", "shallow", "networks", "can", "outperform", "very", "deep", "ones", ",", "which", "is", "probably", "due", "to", "large", "model", "capacity", ",", "coinciding", "with", "the", "results", "reported", "by", "Zagoruyko", "and", "Komodakis", ".", "For", "example", ",", "the", "much", "shallower", "Model", "B", "achieves", "similar", "error", "rates", "as", "ResNet", "-", "152", ",", "and", "even", "runs", "slightly", "faster", ".", "And", "particularly", ",", "Model", "A", "performs", "the", "best", "among", "all", "the", "networks", ".", "2", ")", "We", "can", "trade", "performance", "for", "efficiency", "by", "using", "a", "small", "input", "size", ".", "For", "example", ",", "Model", "D", "performs", "slightly", "worse", "than", "ResNet", "-", "152", ",", "but", "is", "almost", "two", "times", "faster", ".", "This", "may", "be", "useful", "when", "efficiency", "is", "strictly", "required", ".", "Mishkin", "et", "al", ".", "also", "reduced", "the", "input", "size", "for", "efficiency", ".", "However", ",", "they", "did", "not", "remove", "down", "-", "sampling", "operations", "accordingly", "to", "preserve", "the", "size", "of", "final", "feature", "maps", ",", "which", "resulted", "in", "much", "degraded", "performance", ".", "3", ")", "Models", "C", ",", "D", "and", "E", "perform", "comparably", ",", "even", "though", "Model", "C", "has", "larger", "depth", "and", "more", "parameters", ".", "This", "comparison", "shows", "the", "importance", "of", "designing", "a", "network", "properly", ".", "In", "these", "models", ",", "we", "put", "too", "many", "layers", "on", "low", "resolution", "levels", "(", "7", "7", ",", "B5", "in", "Fig", ".", "[", "reference", "]", ")", ".", "subsection", ":", "Semantic", "image", "segmentation", "results", "We", "evaluate", "our", "proposed", "networks", "on", "four", "widely", "used", "datasets", ".", "When", "available", ",", "we", "report", ",", "1", ")", "the", "pixel", "accuracy", ",", "which", "is", "the", "percentage", "of", "correctly", "labelled", "pixels", "on", "a", "whole", "test", "set", ",", "2", ")", "the", "mean", "pixel", "accuracy", ",", "which", "is", "the", "mean", "of", "class", "-", "wise", "pixel", "accuracies", ",", "and", "3", ")", "the", "mean", "IoU", "score", ",", "which", "is", "the", "mean", "of", "class", "-", "wise", "intersection", "-", "over", "-", "union", "scores", ".", "PASCAL", "VOC", "2012", ".", "This", "dataset", "consists", "of", "daily", "life", "photos", ".", "There", "are", "1", ",", "464", "labelled", "images", "for", "training", "and", "another", "1", ",", "449", "for", "validation", ".", "Pixels", "either", "belong", "to", "the", "background", "or", "twenty", "object", "categories", ",", "including", "bus", ",", "car", ",", "cat", ",", "sofa", ",", "monitor", ",", "etc", ".", "Following", "the", "common", "criteria", "in", "the", "literature", ",", "we", "augment", "the", "dataset", "with", "extra", "labelled", "images", "from", "the", "semantic", "boundaries", "dataset", ".", "So", "in", "total", ",", "there", "are", "10", ",", "582", "images", "for", "training", ".", "We", "first", "compare", "different", "networks", "in", "Table", "[", "reference", "]", ".", "Notable", "points", "about", "the", "results", "are", "as", "follows", ".", "1", ")", "We", "can", "not", "make", "statistically", "significant", "improvement", "by", "using", "ResNet", "-", "152", "instead", "of", "ResNet", "-", "101", ".", "However", ",", "Model", "A", "performs", "better", "than", "ResNet", "-", "152", "by", "3.4", "%", ".", "Using", "one", "hidden", "layer", "leads", "to", "a", "further", "improvement", "by", "2.1", "%", ".", "2", ")", "The", "very", "deep", "ResNet", "-", "152", "uses", "too", "many", "memories", "due", "to", "intentionally", "enlarged", "depth", ".", "With", "our", "settings", ",", "it", "even", "can", "not", "be", "tuned", "using", "many", "mainstream", "GPUs", "with", "only", "12", "GB", "memories", ".", "3", ")", "Model", "B", "performs", "worse", "than", "ResNet", "-", "101", ",", "even", "if", "it", "performs", "better", "on", "the", "classification", "task", "as", "shown", "in", "Table", "[", "reference", "]", ".", "This", "shows", "that", "it", "is", "not", "reliable", "to", "tell", "a", "good", "feature", "extractor", "only", "depending", "on", "its", "classification", "performance", ".", "And", "it", "again", "shows", "why", "we", "should", "favour", "deeper", "models", ".", "4", ")", "Model", "A2", "performs", "worse", "than", "Model", "A", "on", "this", "dataset", ".", "We", "initialize", "it", "using", "weights", "from", "Model", "A", ",", "and", "tune", "it", "with", "the", "Places", "365", "data", "for", "45k", "iterations", ".", "This", "is", "reasonable", "since", "there", "are", "only", "object", "categories", "in", "this", "dataset", ",", "while", "Places", "365", "is", "for", "scene", "classification", "tasks", ".", "We", "then", "compare", "our", "method", "with", "previous", "ones", "on", "the", "test", "set", "in", "Table", "[", "reference", "]", ".", "Only", "using", "the", "augmented", "PASCAL", "VOC", "data", "for", "training", ",", "we", "achieve", "a", "mean", "IoU", "score", "of", "82.5", "%", ",", "which", "is", "better", "than", "the", "previous", "best", "one", "by", "3.4", "%", ".", "This", "is", "a", "significant", "margin", ",", "considering", "that", "the", "gap", "between", "ResNet", "-", "based", "and", "VGGNet", "-", "based", "methods", "is", "3.8", "%", ".", "Our", "method", "wins", "for", "seventeen", "out", "of", "the", "twenty", "object", "categories", ",", "which", "was", "the", "official", "criteria", "used", "in", "the", "PASCAL", "VOC", "challenges", ".", "In", "some", "works", ",", "models", "were", "further", "pre", "-", "trained", "using", "the", "Microsoft", "COCO", "data", ",", "which", "consists", "of", "120k", "labelled", "images", ".", "In", "this", "case", ",", "the", "current", "best", "mean", "IoU", "is", "79.7", "%", "reported", "by", "Chen", "et", "al", ".", ".", "They", "also", "used", "multi", "-", "scale", "structure", "and", "CRF", "-", "based", "post", "-", "processing", "in", "their", "submission", ",", "which", "we", "do", "not", "consider", "here", ".", "Nevertheless", ",", "our", "method", "outperforms", "theirs", "by", "2.8", "%", ",", "which", "further", "shows", "the", "effectiveness", "of", "our", "features", "pre", "-", "trained", "only", "using", "the", "ImageNet", "classification", "data", ".", "aero", ".", "bicy", ".", "bird", "boat", "bott", ".", "bus", "car", "cat", "chai", ".", "cow", "dini", ".", "dog", "hors", ".", "moto", ".", "pers", ".", "pott", ".", "shee", ".", "sofa", "trai", ".", "tvmo", ".", "mean", "Cityscapes", ".", "This", "dataset", "consists", "of", "street", "scene", "photos", "taken", "by", "car", "-", "carried", "cameras", ".", "There", "are", "2975", "labelled", "images", "for", "training", "and", "another", "500", "for", "validation", ".", "Besides", ",", "there", "is", "also", "an", "extended", "set", "with", "19", ",", "998", "coarsely", "labelled", "images", ".", "Pixels", "belong", "to", "nineteen", "semantic", "classes", ",", "including", "road", ",", "car", ",", "pedestrian", ",", "bicycle", ",", "etc", ".", "These", "classes", "further", "belong", "to", "seven", "categories", ",", "i.e.", ",", "flat", ",", "nature", ",", "object", ",", "sky", ",", "construction", ",", "human", ",", "and", "vehicle", ".", "We", "first", "compare", "different", "networks", "in", "Table", "[", "reference", "]", ".", "On", "this", "dataset", ",", "ResNet", "-", "152", "again", "shows", "no", "advantage", "against", "ResNet", "-", "101", ".", "However", ",", "Model", "A1", "outperforms", "ResNet", "-", "101", "by", "4.2", "%", "in", "terms", "of", "mean", "IoU", "scores", ",", "which", "again", "is", "a", "significant", "margin", ".", "Because", "there", "are", "many", "scene", "classes", ",", "models", "pre", "-", "trained", "using", "Places", "365", "are", "supposed", "to", "perform", "better", ",", "which", "coincides", "with", "our", "results", ".", "We", "then", "compare", "our", "method", "with", "previous", "ones", "on", "the", "test", "set", "in", "Table", "[", "reference", "]", ".", "The", "official", "criteria", "on", "this", "dataset", "includes", "two", "levels", ",", "i.e.", ",", "class", "and", "category", ".", "Besides", ",", "there", "is", "also", "an", "instance", "-", "weighted", "IoU", "score", "for", "each", "of", "the", "two", ",", "which", "assigns", "high", "scores", "to", "those", "pixels", "of", "small", "instances", ".", "Namely", ",", "this", "score", "penalizes", "methods", "ignoring", "small", "instances", ",", "which", "may", "cause", "fatal", "problems", "in", "vehicle", "-", "centric", "scenarios", ".", "Our", "method", "achieves", "a", "class", "-", "level", "IoU", "score", "of", "78.4", "%", ",", "and", "outperforms", "the", "previous", "best", "one", "by", "6.6", "%", ".", "Furthermore", ",", "in", "the", "case", "of", "instance", "-", "weighted", "IoU", "score", ",", "our", "method", "also", "performs", "better", "than", "the", "previous", "best", "one", "by", "6.4", "%", ".", "It", "is", "notable", "that", "these", "significant", "improvements", "show", "the", "strength", "of", "our", "pre", "-", "trained", "features", ",", "considering", "that", "DeepLab", "-", "v2", "uses", "ResNet", "-", "101", ",", "and", "LRR", "uses", "much", "more", "data", "for", "training", ".", "ADE20", "K", ".", "This", "dataset", "consists", "of", "both", "indoor", "and", "outdoor", "images", "with", "large", "variations", ".", "There", "are", "20", ",", "210", "labelled", "images", "for", "training", "and", "another", "2k", "for", "validation", ".", "Pixels", "belong", "to", "150", "semantic", "categories", ",", "including", "sky", ",", "house", ",", "bottle", ",", "food", ",", "toy", ",", "etc", ".", "We", "first", "compare", "different", "networks", "in", "Table", "[", "reference", "]", ".", "On", "this", "dataset", ",", "ResNet", "-", "152", "performs", "slightly", "better", "than", "ResNet", "-", "101", ".", "However", ",", "Model", "A2", "outperforms", "ResNet", "-", "152", "by", "4.0", "%", "in", "terms", "of", "mean", "IoU", "scores", ".", "Being", "similar", "with", "Cityscapes", ",", "this", "dataset", "has", "many", "scene", "categories", ".", "So", ",", "Model", "A2", "performs", "slightly", "better", "than", "Model", "A.", "Another", "notable", "point", "is", "that", ",", "Model", "C", "takes", "the", "second", "place", "on", "this", "dataset", ",", "even", "if", "it", "performs", "worse", "than", "Model", "A", "in", "the", "image", "classification", "task", "on", "the", "ImageNet", "dataset", ".", "This", "shows", "that", "large", "model", "capacity", "may", "become", "more", "critical", "in", "complicated", "tasks", ",", "since", "there", "are", "more", "parameters", "in", "Model", "C.", "We", "then", "compare", "our", "method", "with", "others", "on", "the", "test", "set", "in", "Table", "[", "reference", "]", ".", "The", "official", "criteria", "on", "this", "dataset", "is", "the", "average", "of", "pixel", "accuracies", "and", "mean", "IoU", "scores", ".", "For", "better", "performance", ",", "we", "apply", "multi", "-", "scale", "testing", ",", "model", "averaging", "and", "post", "-", "processing", "with", "CRFs", ".", "Our", "Model", "A2", "performs", "the", "best", "among", "all", "methods", "using", "only", "a", "single", "pre", "-", "trained", "model", ".", "However", ",", "in", "this", "submission", ",", "we", "only", "managed", "to", "include", "two", "kinds", "of", "pre", "-", "trained", "features", ",", "i.e.", ",", "Models", "A", "and", "C.", "Nevertheless", ",", "our", "method", "only", "performs", "slightly", "worse", "than", "the", "winner", "by", "a", "margin", "of", "0.47", "%", ".", "PASCAL", "Context", ".", "This", "dataset", "consists", "of", "images", "from", "PASCAL", "VOC", "2010", "with", "extra", "object", "and", "stuff", "labels", ".", "There", "are", "4", ",", "998", "images", "for", "training", "and", "another", "5", ",", "105", "for", "validation", ".", "Pixels", "either", "belong", "to", "the", "background", "category", "or", "59", "semantic", "categories", ",", "including", "bag", ",", "food", ",", "sign", ",", "ceiling", ",", "ground", "and", "snow", ".", "All", "images", "in", "this", "dataset", "are", "no", "larger", "than", "500", "500", ".", "Since", "the", "test", "set", "is", "not", "available", ",", "here", "we", "directly", "apply", "the", "hyper", "-", "parameters", "which", "are", "used", "on", "the", "PASCAL", "VOC", "dataset", ".", "Our", "method", "again", "performs", "the", "best", "with", "a", "clear", "margin", "by", "all", "the", "three", "kinds", "of", "scores", ",", "as", "shown", "in", "Table", "[", "reference", "]", ".", "In", "particular", ",", "we", "improve", "the", "IoU", "score", "by", "2.4", "%", "compared", "to", "the", "previous", "best", "method", ".", "section", ":", "Conclusion", "We", "have", "analysed", "the", "ResNet", "architecture", ",", "in", "terms", "of", "the", "ensemble", "classifiers", "therein", "and", "the", "effective", "depths", "of", "the", "residual", "units", ".", "On", "the", "basis", "of", "that", "analysis", "we", "calculated", "a", "new", ",", "more", "spatially", "efficient", ",", "and", "better", "performing", "architecture", "which", "actually", "achieves", "fully", "end", "-", "to", "-", "end", "training", "for", "large", "networks", ".", "Using", "this", "new", "architecture", "we", "designed", "a", "group", "of", "correspondingly", "shallow", "networks", ",", "and", "showed", "that", "they", "outperform", "the", "previous", "very", "deep", "residual", "networks", "not", "only", "on", "the", "ImageNet", "classification", "dataset", ",", "but", "also", "when", "applied", "to", "semantic", "image", "segmentation", ".", "These", "results", "show", "that", "the", "proposed", "architecture", "delivers", "better", "feature", "extraction", "performance", "than", "the", "current", "state", "-", "of", "-", "the", "-", "art", ".", "appendix", ":", "Appendix", "subsection", ":", "Network", "structures", "The", "graph", "structures", "of", "Model", "A", "for", "the", "ImageNet", "(", "ILSVRC", "2012", ")", "classification", "can", "be", "accessed", "at", ":", "Model", "A2", "for", "the", "PASCAL", "VOC", "2012", "segmentation", "can", "be", "accessed", "at", ":", ".", "subsection", ":", "Gradients", "in", "residual", "networks", "We", "show", "results", "of", "the", "experiment", "on", "gradients", "proposed", "by", "Veit", "et", "al", ".", ",", "with", "various", "residual", "networks", ".", "Namely", ",", "for", "a", "trained", "network", "with", "units", ",", "we", "sample", "individual", "paths", "of", "a", "certain", "length", ",", "and", "measure", "the", "norm", "of", "gradients", "that", "arrive", "at", "the", "input", ".", "Each", "time", ",", "we", "first", "feed", "a", "batch", "forward", "through", "the", "whole", "network", ";", "then", "during", "the", "backward", "pass", ",", "we", "randomly", "sample", "units", ".", "For", "them", ",", "we", "only", "propagate", "gradients", "through", "their", "trainable", "mapping", "functions", ",", "but", "without", "their", "shortcut", "connections", ".", "For", "the", "remaining", "units", ",", "we", "do", "the", "opposite", ",", "namely", ",", "only", "propagating", "gradients", "through", "their", "shortcut", "connections", ".", "We", "record", "the", "norm", "of", "those", "gradients", "that", "reach", "the", "input", "for", "varying", "path", "length", ",", "and", "show", "the", "results", "in", "Fig", ".", "[", "reference", "]", ".", "Note", "the", "varying", "magnitude", "and", "maximum", "path", "length", "in", "individual", "figures", ".", "These", "are", "compared", "to", "the", "middle", "part", "of", "Fig", ".", "6", "in", ".", "However", ",", "differently", "we", "further", "divide", "the", "computed", "norm", "of", "a", "batch", "by", "its", "number", "of", "examples", ".", "According", "to", "the", "results", "in", "Fig", ".", "[", "reference", "]", ",", "ResNet", "-", "110", "trained", "on", "CIFAR", "-", "10", ",", "as", "well", "as", "ResNet", "-", "101", "and", "ResNet", "-", "152", "trained", "on", "ILSVRC", "2012", ",", "generate", "much", "smaller", "gradients", "from", "their", "long", "paths", "than", "from", "their", "short", "paths", ".", "In", "contrast", ",", "our", "Model", "A", "trained", "on", "ILSVRC", "2012", ",", "generates", "more", "comparable", "gradients", "from", "its", "paths", "with", "different", "lengths", ".", "subsection", ":", "Qualitative", "results", "We", "show", "qualitative", "results", "of", "semantic", "image", "segmentation", "on", "PASCAL", "VOC", ",", "Cityscapes", ",", "ADE20", "K", ",", "and", "PASCAL", "Context", ",", "respectively", "in", "Figs", ".", "[", "reference", "]", ",", "[", "reference", "]", ",", "[", "reference", "]", ",", "[", "reference", "]", "and", "[", "reference", "]", ",", "and", "show", "some", "failure", "cases", "in", "Figs", "[", "reference", "]", "and", "[", "reference", "]", ".", "In", "a", "difference", "map", ",", "grey", "and", "black", "respectively", "denotes", "correctly", "and", "wrongly", "labelled", "pixels", ",", "while", "white", "denotes", "the", "officially", "ignored", "pixels", "during", "evaluation", ".", "Note", "that", "we", "do", "not", "apply", "post", "-", "processing", "with", "CRFs", ",", "which", "can", "smooth", "the", "output", "but", "is", "too", "slow", "in", "practice", ",", "especially", "for", "large", "images", ".", "bibliography", ":", "References"], "entity": [[8, 10, ["ResNet", "Model"], "Method"], [11, 13, ["Visual", "Recognition"], "Task"], [17, 20, ["deep", "neural", "networks"], "Method"], [66, 69, ["deep", "residual", "networks"], "Method"], [83, 85, ["deep", "network"], "Method"], [94, 96, ["shallow", "networks"], "Method"], [116, 119, ["deep", "residual", "networks"], "Method"], [143, 149, ["shallower", ",", "architecture", "of", "residual", "networks"], "Method"], [153, 155, ["deeper", "models"], "Method"], [157, 160, ["ResNet", "-", "200"], "Method"], [181, 184, ["semantic", "segmentation", "approach"], "Method"], [201, 203, ["PASCAL", "VOC"], "Material"], [204, 206, ["PASCAL", "Context"], "Material"], [208, 209, ["Cityscapes"], "Material"], [218, 219, ["comparators"], "Method"], [222, 224, ["deep", "ResNets"], "Method"], [231, 233, ["memory", "use"], "Metric"], [237, 239, ["training", "time"], "Metric"], [252, 254, ["convolutional", "networks"], "Method"], [257, 260, ["computer", "vision", "community"], "Task"], [274, 275, ["AlexNet"], "Method"], [279, 281, ["deepest", "network"], "Method"], [286, 288, ["residual", "network"], "Method"], [289, 290, ["ResNet"], "Method"], [295, 297, ["trainable", "layers"], "Method"], [339, 341, ["deep", "models"], "Method"], [366, 370, ["200", "-", "layer", "ResNet"], "Method"], [371, 377, ["96", "-", "layer", "Inception", "-", "ResNet"], "Method"], [395, 399, ["269", "-", "layer", "network"], "Method"], [400, 403, ["semantic", "image", "segmentation"], "Task"], [475, 481, ["intersection", "-", "over", "-", "union", "scores"], "Metric"], [517, 520, ["CIFAR", "-", "10"], "Material"], [524, 525, ["ResNet"], "Method"], [565, 566, ["ResNets"], "Method"], [578, 584, ["exponential", "ensembles", "of", "relatively", "shallow", "networks"], "Method"], [598, 599, ["ResNet"], "Method"], [603, 609, ["exponential", "ensemble", "of", "sub", "-", "networks"], "Method"], [654, 655, ["ResNets"], "Method"], [725, 728, ["sub", "-", "networks"], "Method"], [752, 755, ["sub", "-", "network"], "Method"], [867, 869, ["intuitive", "view"], "Method"], [870, 871, ["ResNets"], "Method"], [893, 896, ["shallow", "convolutional", "networks"], "Method"], [933, 936, ["semantic", "image", "segmentation"], "Task"], [961, 963, ["PASCAL", "VOC"], "Material"], [964, 966, ["PASCAL", "Context"], "Material"], [968, 969, ["Cityscapes"], "Material"], [984, 986, ["residual", "network"], "Method"], [987, 988, ["ResNet"], "Method"], [990, 992, ["image", "classification"], "Task"], [993, 996, ["semantic", "image", "segmentation"], "Task"], [997, 1000, ["fully", "convolutional", "networks"], "Method"], [1014, 1015, ["ResNets"], "Method"], [1018, 1021, ["vanishing", "gradient", "problem"], "Task"], [1024, 1027, ["deep", "convolutional", "networks"], "Method"], [1028, 1029, ["ResNets"], "Method"], [1041, 1043, ["object", "detection"], "Task"], [1044, 1047, ["semantic", "image", "segmentation"], "Task"], [1052, 1053, ["VGGNets"], "Method"], [1055, 1058, ["computer", "vision", "community"], "Task"], [1062, 1064, ["feature", "extractors"], "Method"], [1074, 1075, ["ResNets"], "Method"], [1089, 1095, ["exponential", "ensembles", "of", "relatively", "shallow", "networks"], "Method"], [1157, 1159, ["deep", "networks"], "Method"], [1178, 1181, ["competitively", "deep", "networks"], "Method"], [1194, 1195, ["ResNets"], "Method"], [1199, 1200, ["convergence"], "Metric"], [1245, 1248, ["CIFAR", "-", "10"], "Material"], [1274, 1279, ["grid", "search", "of", "configuration", "space"], "Method"], [1292, 1295, ["Semantic", "image", "segmentation"], "Task"], [1313, 1316, ["fully", "convolutional", "networks"], "Method"], [1317, 1318, ["FCN"], "Method"], [1323, 1324, ["FCNs"], "Method"], [1330, 1334, ["dense", "prediction", "based", "tasks"], "Task"], [1372, 1375, ["fully", "convolutional", "networks"], "Method"], [1389, 1390, ["segmentation"], "Task"], [1394, 1396, ["Residual", "networks"], "Method"], [1403, 1411, ["full", "pre", "-", "activation", "version", "of", "residual", "networks"], "Method"], [1412, 1413, ["ResNet"], "Method"], [1435, 1437, ["linear", "classifier"], "Method"], [1451, 1453, ["convolution", "layers"], "Method"], [1472, 1474, ["residual", "Unit"], "Method"], [1485, 1489, ["non", "-", "linear", "mappings"], "Method"], [1515, 1518, ["stacked", "convolution", "stages"], "Method"], [1522, 1526, ["pre", "-", "activation", "version"], "Method"], [1536, 1538, ["batch", "normalization"], "Method"], [1540, 1543, ["rectified", "linear", "unit"], "Method"], [1544, 1545, ["ReLU"], "Method"], [1552, 1554, ["convolution", "layer"], "Method"], [1557, 1559, ["Residual", "networks"], "Method"], [1644, 1647, ["sub", "-", "networks"], "Method"], [1658, 1665, ["linearly", "growing", "ensemble", "of", "sub", "-", "networks"], "Method"], [1729, 1730, ["ResNet"], "Method"], [1829, 1830, ["ResNet"], "Method"], [1842, 1844, ["SGD", "iteration"], "Method"], [2025, 2026, ["ResNet"], "Method"], [2067, 2068, ["ResNet"], "Method"], [2204, 2211, ["end", "-", "to", "-", "end", "training", "case"], "Task"], [2237, 2239, ["deep", "ResNets"], "Method"], [2249, 2251, ["deep", "networks"], "Method"], [2254, 2257, ["vanishing", "gradient", "problem"], "Task"], [2327, 2331, ["shallowest", "sub", "-", "network"], "Method"], [2425, 2426, ["ResNet"], "Method"], [2429, 2430, ["ensemble"], "Method"], [2474, 2475, ["ResNets"], "Method"], [2476, 2478, ["virtual", "ensembles"], "Method"], [2587, 2593, ["one", "-", "block", "sub", "-", "network"], "Method"], [2598, 2600, ["preprocessed", "representations"], "Method"], [2652, 2653, ["ResNets"], "Method"], [2656, 2659, ["vanishing", "gradient", "problem"], "Task"], [2664, 2668, ["shallower", "sub", "-", "networks"], "Method"], [2698, 2700, ["deep", "networks"], "Method"], [2749, 2750, ["fusion"], "Task"], [2751, 2752, ["concatenation"], "Method"], [2753, 2756, ["Inception", "-", "v4"], "Method"], [2758, 2759, ["fusion"], "Method"], [2760, 2761, ["summation"], "Method"], [2762, 2763, ["ResNets"], "Method"], [2772, 2775, ["vanishing", "gradient", "problem"], "Task"], [2782, 2783, ["shortcuts"], "Method"], [2806, 2809, ["vanishing", "gradient", "problem"], "Task"], [2818, 2820, ["wide", "layers"], "Method"], [2835, 2836, ["dropout"], "Method"], [2850, 2851, ["ResNets"], "Method"], [2863, 2864, ["ResNet"], "Method"], [2887, 2889, ["shallow", "network"], "Method"], [2928, 2931, ["sub", "-", "network"], "Method"], [2970, 2972, ["weight", "sharing"], "Method"], [2974, 2977, ["over", "-", "fitting"], "Task"], [2988, 2990, ["linear", "classifier"], "Method"], [2992, 2993, ["ResNets"], "Method"], [2997, 2999, ["weight", "sharing"], "Method"], [3019, 3020, ["ResNet"], "Method"], [3107, 3109, ["trainable", "layer"], "Method"], [3117, 3119, ["residual", "unit"], "Method"], [3122, 3124, ["convolution", "stages"], "Method"], [3154, 3156, ["convolution", "stages"], "Method"], [3177, 3179, ["deeper", "networks"], "Method"], [3257, 3260, ["plain", "deep", "networks"], "Method"], [3267, 3268, ["AlexNet"], "Method"], [3269, 3270, ["VGGNets"], "Method"], [3283, 3286, ["deep", "residual", "networks"], "Method"], [3310, 3312, ["deeper", "model"], "Method"], [3334, 3336, ["spatial", "costs"], "Metric"], [3349, 3351, ["communication", "costs"], "Metric"], [3360, 3364, ["shallower", "but", "wider", "network"], "Method"], [3392, 3393, ["ResNet"], "Method"], [3419, 3421, ["Inception", "networks"], "Method"], [3478, 3482, ["200", "-", "layer", "ResNet"], "Method"], [3483, 3489, ["96", "-", "layer", "Inception", "-", "ResNet"], "Method"], [3505, 3507, ["model", "shallow"], "Method"], [3578, 3585, ["fully", "end", "-", "to", "-", "end", "networks"], "Method"], [3594, 3595, ["ResNets"], "Method"], [3598, 3600, ["feature", "extractors"], "Method"], [3607, 3609, ["shallow", "network"], "Method"], [3615, 3618, ["twenty", "-", "layer"], "Method"], [3621, 3624, ["ResNet", "-", "152"], "Method"], [3652, 3654, ["image", "classification"], "Task"], [3690, 3692, ["batch", "normalization"], "Method"], [3700, 3702, ["convolution", "layer"], "Method"], [3708, 3709, ["ResNets"], "Method"], [3716, 3718, ["convolution", "stages"], "Method"], [3782, 3784, ["convolution", "layers"], "Method"], [3867, 3871, ["down", "-", "sampling", "operation"], "Method"], [3887, 3889, ["convolution", "stages"], "Method"], [3905, 3907, ["convolution", "layer"], "Method"], [3923, 3926, ["spatial", "pooling", "layer"], "Method"], [3941, 3942, ["classification"], "Task"], [3952, 3954, ["pooling", "layers"], "Method"], [3955, 3958, ["down", "-", "sampling"], "Task"], [3981, 3982, ["AlexNet"], "Method"], [3983, 3984, ["VGGNets"], "Method"], [4008, 4010, ["MXNet", "framework"], "Method"], [4038, 4039, ["ResNets"], "Method"], [4051, 4055, ["linear", "learning", "rate", "schedule"], "Method"], [4089, 4091, ["learning", "rate"], "Metric"], [4100, 4103, ["semantic", "image", "segmentation"], "Task"], [4109, 4112, ["fully", "convolutional", "networks"], "Method"], [4113, 4114, ["FCN"], "Method"], [4121, 4122, ["DeepLab"], "Method"], [4161, 4165, ["multi", "-", "scale", "testing"], "Method"], [4166, 4168, ["model", "averaging"], "Method"], [4169, 4174, ["CRF", "based", "post", "-", "processing"], "Method"], [4181, 4183, ["ADE20", "K"], "Material"], [4188, 4190, ["trained", "network"], "Method"], [4203, 4206, ["semantic", "image", "segmentation"], "Task"], [4213, 4214, ["Resolution"], "Metric"], [4233, 4235, ["dilation", "rates"], "Metric"], [4238, 4240, ["convolution", "layers"], "Method"], [4256, 4258, ["convolution", "layer"], "Method"], [4279, 4281, ["convolution", "layer"], "Method"], [4289, 4291, ["dilation", "rate"], "Metric"], [4308, 4310, ["convolution", "layer"], "Method"], [4317, 4319, ["dilation", "rate"], "Metric"], [4332, 4335, ["down", "-", "sampling"], "Task"], [4337, 4339, ["pooling", "layer"], "Method"], [4351, 4353, ["pooling", "layer"], "Method"], [4364, 4366, ["pooling", "layer"], "Method"], [4367, 4368, ["dilation"], "Method"], [4392, 4396, ["down", "-", "sampling", "operations"], "Method"], [4406, 4410, ["down", "-", "sampling", "operations"], "Method"], [4413, 4416, ["spatial", "pooling", "layers"], "Method"], [4419, 4423, ["pre", "-", "trained", "networks"], "Method"], [4429, 4430, ["FCNs"], "Method"], [4450, 4457, ["top", "-", "most", "down", "-", "sampling", "operations"], "Method"], [4505, 4507, ["convolution", "layers"], "Method"], [4525, 4527, ["learning", "rate"], "Metric"], [4532, 4533, ["Classifier"], "Method"], [4540, 4542, ["linear", "classifier"], "Method"], [4544, 4547, ["global", "pooling", "layer"], "Method"], [4570, 4573, ["DeepLab", "-", "v2"], "Method"], [4576, 4578, ["1", "convolution"], "Method"], [4588, 4590, ["linear", "layer"], "Method"], [4593, 4594, ["classifier"], "Method"], [4596, 4597, ["anther"], "Task"], [4602, 4607, ["non", "-", "linear", "convolution", "stage"], "Method"], [4609, 4611, ["batch", "normalization"], "Method"], [4614, 4616, ["linear", "classifier"], "Method"], [4622, 4624, ["2", "convolutions"], "Method"], [4638, 4640, ["dilation", "rate"], "Metric"], [4649, 4651, ["layer", "classifier"], "Method"], [4687, 4690, ["over", "-", "fitting"], "Task"], [4696, 4697, ["dropout"], "Method"], [4704, 4706, ["dropout", "rate"], "Metric"], [4771, 4774, ["batch", "normalization", "layers"], "Method"], [4775, 4778, ["fine", "-", "tuning"], "Task"], [4815, 4817, ["learning", "rate"], "Metric"], [4827, 4829, ["learning", "rate"], "Metric"], [4838, 4840, ["linear", "schedule"], "Method"], [4859, 4860, ["training"], "Task"], [4898, 4900, ["Image", "classification"], "Task"], [4930, 4939, ["top", "-", "1", "and", "top", "-", "5", "error", "rates"], "Metric"], [4977, 4978, ["VGG16"], "Method"], [4984, 4987, ["fully", "convolutional", "network"], "Method"], [5035, 5037, ["Inception", "networks"], "Task"], [5040, 5042, ["crop", "size"], "Metric"], [5128, 5130, ["shallow", "networks"], "Method"], [5160, 5163, ["shallower", "Model", "B"], "Method"], [5165, 5167, ["error", "rates"], "Metric"], [5168, 5171, ["ResNet", "-", "152"], "Method"], [5215, 5218, ["ResNet", "-", "152"], "Method"], [5336, 5339, ["Semantic", "image", "segmentation"], "Task"], [5360, 5362, ["pixel", "accuracy"], "Metric"], [5380, 5383, ["mean", "pixel", "accuracy"], "Metric"], [5389, 5394, ["class", "-", "wise", "pixel", "accuracies"], "Metric"], [5399, 5402, ["mean", "IoU", "score"], "Metric"], [5408, 5417, ["class", "-", "wise", "intersection", "-", "over", "-", "union", "scores"], "Metric"], [5418, 5421, ["PASCAL", "VOC", "2012"], "Material"], [5445, 5446, ["validation"], "Task"], [5537, 5540, ["ResNet", "-", "152"], "Method"], [5542, 5545, ["ResNet", "-", "101"], "Method"], [5553, 5556, ["ResNet", "-", "152"], "Method"], [5562, 5564, ["hidden", "layer"], "Method"], [5578, 5581, ["ResNet", "-", "152"], "Method"], [5604, 5605, ["GPUs"], "Method"], [5618, 5621, ["ResNet", "-", "101"], "Method"], [5629, 5631, ["classification", "task"], "Task"], [5650, 5652, ["feature", "extractor"], "Method"], [5656, 5657, ["classification"], "Task"], [5697, 5700, ["Places", "365", "data"], "Material"], [5722, 5725, ["scene", "classification", "tasks"], "Task"], [5747, 5751, ["augmented", "PASCAL", "VOC", "data"], "Material"], [5757, 5760, ["mean", "IoU", "score"], "Metric"], [5787, 5790, ["ResNet", "-", "based"], "Method"], [5791, 5795, ["VGGNet", "-", "based", "methods"], "Method"], [5819, 5822, ["PASCAL", "VOC", "challenges"], "Material"], [5835, 5838, ["Microsoft", "COCO", "data"], "Material"], [5853, 5855, ["mean", "IoU"], "Metric"], [5868, 5872, ["multi", "-", "scale", "structure"], "Method"], [5873, 5879, ["CRF", "-", "based", "post", "-", "processing"], "Method"], [5978, 5979, ["validation"], "Task"], [6054, 6057, ["ResNet", "-", "152"], "Method"], [6062, 6065, ["ResNet", "-", "101"], "Method"], [6071, 6074, ["ResNet", "-", "101"], "Method"], [6136, 6138, ["official", "criteria"], "Metric"], [6195, 6199, ["vehicle", "-", "centric", "scenarios"], "Task"], [6268, 6271, ["DeepLab", "-", "v2"], "Method"], [6272, 6275, ["ResNet", "-", "101"], "Method"], [6277, 6278, ["LRR"], "Method"], [6285, 6286, ["ADE20"], "Method"], [6314, 6315, ["validation"], "Task"], [6351, 6354, ["ResNet", "-", "152"], "Method"], [6358, 6361, ["ResNet", "-", "101"], "Method"], [6367, 6370, ["ResNet", "-", "152"], "Method"], [6383, 6384, ["Cityscapes"], "Material"], [6476, 6478, ["official", "criteria"], "Metric"], [6483, 6487, ["average", "of", "pixel", "accuracies"], "Metric"], [6498, 6502, ["multi", "-", "scale", "testing"], "Method"], [6503, 6505, ["model", "averaging"], "Method"], [6506, 6509, ["post", "-", "processing"], "Method"], [6510, 6511, ["CRFs"], "Method"], [6525, 6529, ["pre", "-", "trained", "model"], "Method"], [6573, 6575, ["PASCAL", "Context"], "Material"], [6582, 6585, ["PASCAL", "VOC", "2010"], "Material"], [6606, 6607, ["validation"], "Task"], [6666, 6669, ["PASCAL", "VOC", "dataset"], "Material"], [6702, 6704, ["IoU", "score"], "Metric"], [6728, 6730, ["ensemble", "classifiers"], "Method"], [6762, 6769, ["fully", "end", "-", "to", "-", "end", "training"], "Task"], [6770, 6772, ["large", "networks"], "Task"], [6782, 6785, ["correspondingly", "shallow", "networks"], "Method"], [6794, 6797, ["deep", "residual", "networks"], "Method"], [6810, 6813, ["semantic", "image", "segmentation"], "Task"], [6823, 6825, ["feature", "extraction"], "Task"], [6852, 6853, ["ImageNet"], "Task"], [6854, 6856, ["ILSVRC", "2012"], "Task"], [6867, 6871, ["PASCAL", "VOC", "2012", "segmentation"], "Task"], [6881, 6883, ["residual", "networks"], "Method"], [6890, 6891, ["gradients"], "Method"], [6900, 6902, ["residual", "networks"], "Method"], [6907, 6909, ["trained", "network"], "Method"], [6967, 6970, ["trainable", "mapping", "functions"], "Method"], [7078, 7081, ["ResNet", "-", "110"], "Method"], [7083, 7086, ["CIFAR", "-", "10"], "Material"], [7090, 7093, ["ResNet", "-", "101"], "Method"], [7094, 7097, ["ResNet", "-", "152"], "Method"], [7099, 7101, ["ILSVRC", "2012"], "Task"], [7124, 7126, ["ILSVRC", "2012"], "Task"], [7147, 7150, ["semantic", "image", "segmentation"], "Task"], [7151, 7153, ["PASCAL", "VOC"], "Material"], [7154, 7155, ["Cityscapes"], "Material"], [7156, 7158, ["ADE20", "K"], "Material"], [7160, 7162, ["PASCAL", "Context"], "Material"], [7234, 7237, ["post", "-", "processing"], "Method"], [7238, 7239, ["CRFs"], "Method"], [162, 163, ["ImageNet"], "Task"], [307, 310, ["CIFAR", "-", "10"], "Material"], [410, 411, ["ImageNet"], "Task"], [710, 711, ["ResNet"], "Method"], [787, 788, ["ResNet"], "Method"], [828, 829, ["ResNet"], "Method"], [917, 918, ["ImageNet"], "Task"], [1123, 1124, ["ResNet"], "Method"], [1228, 1229, ["ResNet"], "Method"], [1236, 1237, ["ResNet"], "Method"], [1288, 1289, ["ImageNet"], "Task"], [1367, 1370, ["semantic", "image", "segmentation"], "Task"], [1671, 1672, ["ResNet"], "Method"], [1767, 1768, ["ResNet"], "Method"], [2093, 2094, ["ResNet"], "Method"], [2264, 2265, ["ResNet"], "Method"], [2318, 2319, ["ResNet"], "Method"], [2541, 2542, ["ResNet"], "Method"], [2722, 2725, ["Inception", "-", "v4"], "Method"], [2814, 2815, ["ResNets"], "Method"], [3439, 3440, ["ResNet"], "Method"], [3451, 3452, ["ResNet"], "Method"], [3467, 3468, ["ImageNet"], "Task"], [3521, 3522, ["ResNets"], "Method"], [3626, 3627, ["ImageNet"], "Task"], [3694, 3695, ["ReLU"], "Method"], [3805, 3806, ["ResNets"], "Method"], [4519, 4520, ["ImageNet"], "Task"], [4723, 4724, ["ResNets"], "Method"], [4908, 4910, ["ILSVRC", "2012"], "Task"], [5914, 5915, ["ImageNet"], "Task"], [5951, 5952, ["Cityscapes"], "Material"], [6080, 6082, ["mean", "IoU"], "Metric"], [6160, 6162, ["IoU", "score"], "Metric"], [6207, 6209, ["IoU", "score"], "Metric"], [6232, 6234, ["IoU", "score"], "Metric"], [6376, 6378, ["mean", "IoU"], "Metric"], [6428, 6430, ["image", "classification"], "Task"], [6433, 6434, ["ImageNet"], "Task"], [6488, 6490, ["mean", "IoU"], "Metric"], [6721, 6722, ["ResNet"], "Method"], [6801, 6802, ["ImageNet"], "Task"]]} {"doc_id": "310fed695da4f36139b70b23065c34fe9fb29ca6", "text": ["document", ":", "Shorten", "Spatial", "-", "spectral", "RNN", "with", "Parallel", "-", "GRU", "for", "Hyperspectral", "Image", "Classification", "Convolutional", "neural", "networks", "(", "CNNs", ")", "attained", "a", "good", "performance", "in", "hyperspectral", "sensing", "image", "(", "HSI", ")", "classification", ",", "but", "CNNs", "consider", "spectra", "as", "orderless", "vectors", ".", "Therefore", ",", "considering", "the", "spectra", "as", "sequences", ",", "recurrent", "neural", "networks", "(", "RNNs", ")", "have", "been", "applied", "in", "HSI", "classification", ",", "for", "RNNs", "is", "skilled", "at", "dealing", "with", "sequential", "data", ".", "However", ",", "for", "a", "long", "-", "sequence", "task", ",", "RNNs", "is", "difficult", "for", "training", "and", "not", "as", "effective", "as", "we", "expected", ".", "Besides", ",", "spatial", "contextual", "features", "are", "not", "considered", "in", "RNNs", ".", "In", "this", "study", ",", "we", "propose", "a", "Shorten", "Spatial", "-", "spectral", "RNN", "with", "Parallel", "-", "GRU", "(", "St", "-", "SS", "-", "pGRU", ")", "for", "HSI", "classification", ".", "A", "shorten", "RNN", "is", "more", "efficient", "and", "easier", "for", "training", "than", "band", "-", "by", "-", "band", "RNN", ".", "By", "combining", "converlusion", "layer", ",", "the", "St", "-", "SSpGRU", "model", "considers", "not", "only", "spectral", "but", "also", "spatial", "feature", ",", "which", "results", "in", "a", "better", "performance", ".", "An", "architecture", "named", "parallel", "-", "GRU", "is", "also", "proposed", "and", "applied", "in", "St", "-", "SS", "-", "pGRU", ".", "With", "this", "architecture", ",", "the", "model", "gets", "a", "better", "performance", "and", "is", "more", "robust", ".", "deep", "learning", ",", "gated", "recurrent", "unit", "(", "GRU", ")", ",", "long", "short", "-", "term", "memory", "(", "LSTM", ")", ",", "recurrent", "neural", "networks", "(", "RNN", ")", ",", "hyperspectral", "image", "classification", "section", ":", "Introduction", "Hyperspectral", "image", "(", "HSI", ")", "has", "attracted", "considerable", "attention", "in", "the", "remote", "sensing", "community", "and", "been", "widely", "used", "in", "various", "areas", ".", "With", "the", "rich", "spectral", "information", "in", "HSI", ",", "different", "land", "cover", "categories", "can", "potentially", "be", "differentiated", "precisely", ".", "In", "recent", "years", ",", "deep", "learning", "has", "been", "widely", "used", "in", "various", "fields", ",", "including", "HSI", "classification", ".", "Convolutional", "neural", "networks", "(", "CNNs", ")", "and", "residual", "networks", "(", "ResNets", ")", "have", "obtained", "a", "successful", "result", "for", "HSI", "classification", ".", "Recurrent", "neural", "networks", "(", "RNNs", ")", "are", "also", "applied", "in", "HSI", "classification", ".", "Because", "of", "the", "ability", "to", "extract", "the", "spatial", "contextual", "information", ",", "CNNs", "and", "ResNets", "can", "achieve", "a", "high", "accuracy", "in", "the", "classification", "task", ".", "However", ",", "CNNs", "and", "ResNets", "consider", "spectra", "as", "orderless", "vectors", "in", "-", "dimensional", "feature", "space", "where", "represents", "the", "number", "of", "bands", ".", "However", ",", "spectra", "can", "be", "seen", "as", "orderly", "and", "continuing", "sequences", "in", "the", "spectral", "space", ".", "In", "other", "words", ",", "CNNs", "and", "ResNets", "ignore", "the", "continuity", "of", "spectra", ".", "RNNs", "have", "proved", "effective", "in", "solving", "many", "challenging", "problems", "involving", "sequential", "data", ",", "such", "as", "Natural", "Language", "Processing", "(", "NLP", ")", "and", "prediction", "of", "time", "series", ".", "Considering", "the", "spectrum", "as", "a", "sequential", "sequence", ",", "the", "application", "of", "RNNs", "is", "reasonable", "as", "it", "can", "take", "full", "advantage", "of", "the", "high", "spectral", "resolution", "characteristics", "of", "HSI", ".", "However", ",", "for", "a", "long", "-", "sequence", "task", ",", "RNNs", "is", "not", "as", "effective", "as", "we", "expected", ".", "Long", "distance", "dependence", ",", "gradient", "vanish", "and", "overfitting", "are", "prone", "to", "occur", ".", "Even", "if", "the", "long", "short", "-", "term", "memory", "network", "(", "LSTM", ")", "is", "used", "to", "solve", "the", "long", "-", "distance", "dependence", "problem", ",", "RNNs", "is", "still", "hard", "for", "training", "and", "easily", "overfitting", "in", "a", "long", "-", "sequence", "task", ".", "In", "previous", "work", ",", "3D", "-", "CNN", "is", "applied", "in", "HSI", "classification", "and", "obtained", "a", "good", "behavior", ".", "For", "RNNs", ",", "Convolutional", "-", "LSTM", "(", "CLSTM", ")", "also", "achieved", "a", "good", "performance", "in", "HSI", "classification", ".", "3D", "-", "CNNs", "and", "CLSTM", "consider", "both", "spatial", "contextual", "information", "and", "spectral", "continuity", ",", "which", "result", "in", "a", "high", "accuracy", ".", "Nevertheless", ",", "it", "takes", "a", "long", "time", "to", "train", "these", "two", "models", ".", "In", ",", "LSTM", "and", "its", "variant", ",", "GRU", ",", "are", "applied", "in", "HIS", "classification", ",", "and", "it", "is", "proved", "that", "GRU", "has", "a", "better", "performance", "in", "HIS", "classification", ".", "To", "solve", "the", "problem", "that", "RNNs", "are", "easily", "over", "-", "fitting", "and", "difficult", "for", "training", ",", "proposed", "band", "-", "group", "LSTM", ",", "which", "can", "effectively", "make", "training", "easier", "by", "reducing", "the", "number", "of", "timestep", "in", "LSTM", ".", "In", "this", "study", ",", "a", "Shorten", "Spatial", "-", "spectral", "RNN", "with", "Parallel", "-", "GRU", "(", "St", "-", "SS", "-", "pGRU", ")", "is", "proposed", ".", "This", "study", "contributes", "to", "the", "literature", "in", "2", "major", "respects", ":", "A", "shorten", "RNN", "with", "GRU", "is", "applied", "in", "HIS", "classification", ".", "The", "model", "is", "more", "efficient", "and", "easier", "for", "training", "than", "band", "-", "by", "-", "band", "RNN", ".", "By", "combining", "converlusion", "layer", ",", "an", "advanced", "model", "Shorten", "Spatial", "-", "spectral", "RNN", "with", "GRU", "is", "proposed", ".", "The", "model", "considers", "not", "only", "spectral", "but", "also", "spatial", "feature", ",", "which", "leads", "to", "a", "better", "performance", ".", "An", "architecture", "named", "parallel", "-", "GRU", "is", "proposed", "and", "the", "model", "with", "this", "architecture", "has", "a", "better", "performance", "and", "is", "more", "robust", ".", "The", "remainder", "of", "this", "paper", "is", "organized", "as", "follows", ".", "In", "the", "methodology", "section", ",", "firstly", "the", "structure", "of", "traditional", "RNN", ",", "LSTM", "and", "GRU", "are", "introduced", "and", "then", "the", "architecture", "of", "the", "proposed", "models", "are", "described", ".", "In", "the", "experimental", "section", ",", "the", "network", "setup", ",", "the", "experimental", "results", ",", "and", "the", "comparison", "of", "different", "models", "are", "provided", ".", "Finally", ",", "the", "conclusion", "section", "concludes", "the", "paper", ".", "section", ":", "Methodology", "subsection", ":", "Recurrent", "neural", "networks", "(", "RNN", ")", "Different", "from", "Artificial", "neural", "network", "(", "ANN", ")", ",", "RNN", ",", "a", "neural", "network", "with", "recurrent", "unit", ",", "has", "a", "better", "performance", "in", "solving", "many", "challenging", "problems", "involving", "sequential", "data", "analysis", ".", "The", "state", "of", "each", "time", "step", "of", "the", "recurrent", "unit", "is", "not", "only", "related", "to", "the", "input", "of", "the", "current", "step", ",", "but", "also", "related", "to", "the", "state", "of", "the", "previous", "step", ".", "Thus", ",", "the", "state", "of", "the", "preceding", "step", "can", "effectively", "influence", "the", "next", "step", ".", "Given", "a", "sequence", "sample", ",", "in", "which", "is", "the", "data", "at", "th", "timestep", ".", "For", "the", "th", "recurrent", "unit", ",", "its", "hidden", "state", "can", "be", "described", "as", ":", "where", "is", "the", "initial", "state", "of", "the", "recurrent", "unit", ",", "is", "a", "nonlinear", "function", ".", "Normally", ",", "is", "set", "as", "a", "zero", "vector", ".", "Optionally", ",", "in", "th", "timestep", ",", "the", "recurrent", "unit", "may", "have", "an", "output", ".", "For", "some", "task", ",", "the", "RNN", "model", "will", "finally", "have", "an", "output", "vector", ",", "while", "for", "classification", "tasks", ",", "only", "one", "output", "is", "needed", ".", "Generally", ",", "The", "last", "output", "is", "adopted", ":", "The", "recurrent", "unit", "in", "a", "traditional", "RNN", "is", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "In", "the", "traditional", "RNN", "model", ",", "the", "update", "rule", "of", "the", "recurrent", "hidden", "state", "and", "output", "in", "Eq", ".", "(", "[", "reference", "]", ")", "and", "(", "[", "reference", "]", ")", "is", "usually", "implemented", "as", "follows", ":", "where", ",", "and", "are", "the", "weight", "matrices", ".", "and", "are", "the", "bias", "vectors", ",", "and", "is", "an", "activation", "function", ",", "such", "as", "the", "sigmoid", "function", "or", "the", "hyperbolic", "tangent", "function", ".", "subsection", ":", "Long", "short", "-", "term", "memory", "(", "LSTM", ")", "The", "traditional", "RNN", "has", "the", "problem", "of", "long", "-", "distance", "dependence", ".", "The", "RNN", "has", "the", "capability", "to", "connect", "different", "timesteps", "related", "information", ".", "However", ",", "when", "the", "sequence", "is", "too", "long", ",", "the", "RNN", "becomes", "unable", "to", "connect", "related", "information", "as", "the", "distance", "increases", ",", "because", "the", "information", "losses", "when", "propagating", "through", "multi", "-", "time", "-", "step", "recurrent", "units", ".", "By", "using", "long", "short", "-", "term", "memory", "(", "LSTM", ")", ",", "the", "problems", "have", "been", "solved", ".", "As", "Fig", ".", "[", "reference", "]", "shows", ",", "LSTM", "contains", "a", "forget", "gate", ",", "an", "input", "gate", "and", "an", "output", "gate", ".", "’", "Gate", "’", "structure", "is", "actually", "a", "logistic", "regression", "model", "so", "that", "part", "of", "the", "information", "is", "filtered", "selectively", ",", "while", "the", "rest", "is", "reserved", "and", "passes", "through", "the", "gate", ".", "LSTM", "can", "simulate", "the", "process", "of", "forgetting", "and", "memory", "and", "calculate", "the", "probability", "of", "forgetting", "and", "memory", ",", "so", "information", "flow", "could", "be", "preserved", "in", "long", "-", "distance", "propagation", ".", "The", "structure", "of", "LSTM", "can", "be", "described", "as", ":", "where", "Eq", ".", "(", "[", "reference", "]", ")", ",", "(", "[", "reference", "]", ")", "and", "(", "[", "reference", "]", ")", "represent", "forget", "gate", ",", "input", "gate", "and", "output", "gate", ".", ",", ",", ",", ",", ",", ",", "and", "are", "the", "weight", "matrices", ".", ",", ",", "and", "are", "the", "bias", "vectors", ".", "refers", "to", "sigmoid", "function", "and", "tanh", "refers", "to", "the", "hyperbolic", "tangent", "function", ":", "subsection", ":", "Gated", "recurrent", "unit", "(", "GRU", ")", "Over", "the", "years", ",", "there", "have", "been", "many", "variants", "of", "LSTM", ",", "but", "there", "is", "no", "evidence", "to", "show", "that", "there", "is", "not", "a", "superior", "variant", ".", "Any", "variant", "may", "have", "advantages", "in", "a", "particular", "problem", ".", "GRU", "is", "a", "variant", "of", "LSTM", ".", "With", "fewer", "parameters", ",", "it", "is", "much", "easier", "for", "training", "than", "LSTM", ",", "and", "usually", "achieves", "the", "same", "performance", "as", "LSTM", "in", "some", "tasks", ".", "It", "is", "considered", "that", "using", "GRU", "in", "a", "HSI", "classification", "task", "is", "more", "appropriate", "than", "using", "LSTM", ".", "The", "main", "difference", "between", "LSTM", "and", "GRU", "is", "that", "an", "update", "gate", "and", "a", "reset", "gate", "are", "adopted", "in", "GRU", ",", "instead", "of", "using", "a", "forget", "gate", ",", "an", "input", "gate", "and", "an", "output", "gate", ".", "The", "structure", "of", "the", "GRU", "is", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "which", "can", "be", "defined", "as", "follows", ":", "where", "Eq", ".", "(", "[", "reference", "]", ")", "and", "(", "[", "reference", "]", ")", "represent", "update", "gate", "and", "reset", "gate", ".", ",", ",", ",", ",", "and", "are", "the", "weight", "matrices", ".", ",", "and", "are", "the", "bias", "vectors", ".", "subsection", ":", "The", "proposed", "model", "subsubsection", ":", "Shorten", "Spatial", "-", "spectral", "RNN", "with", "GRU", "(", "St", "-", "SS", "-", "GRU", ")", "A", "Shorten", "Spatial", "-", "spectral", "RNN", "with", "GRU", "(", "St", "-", "SS", "-", "GRU", ")", "model", "for", "HSI", "classification", "is", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "For", "each", "pixel", ",", "a", "square", "subgraph", "composed", "of", "5", "5", "pixels", "centered", "on", "it", "is", "used", "as", "a", "training", "sample", ".", "The", "first", "part", "of", "St", "-", "SS", "-", "GRU", "is", "actually", "a", "3D", "-", "Convolutional", "layer", "but", "both", "the", "depth", "and", "stride", "of", "the", "kernels", "are", "1", ".", "Three", "different", "convolution", "kernels", "(", "1×1", ",", "3×3", ",", "5×5", ")", "were", "used", "to", "convolve", "different", "bands", ".", "The", "output", "of", "this", "part", "is", "a", "sequence", "with", "the", "same", "length", "as", "the", "original", "input", ".", "The", "output", "sequence", "is", "a", "’", "spectra", "’", "with", "the", "spatial", "contextual", "feature", ".", "Every", "timestep", "of", "the", "sequence", "is", "a", "feature", "vector", ".", "The", "second", "part", "is", "a", "Shorten", "RNN", "with", "GRU", "(", "St", "-", "GRU", ")", ".", "The", "structure", "of", "St", "-", "GRU", "is", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "The", "1D", "converlusion", "layer", "before", "GRU", "is", "used", "to", "reduce", "the", "number", "of", "timesteps", "so", "that", "the", "network", "is", "easier", "for", "training", ".", "subsubsection", ":", "Parallel", "-", "GRU", "Architecture", "In", "order", "to", "make", "the", "model", "more", "robust", ",", "a", "Parallel", "-", "GRU", "(", "pGRU", ")", "architecture", "is", "proposed", ".", "The", "architecture", "of", "Shorten", "Parallel", "-", "GRU", "(", "St", "-", "pGRU", ")", "is", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "The", "architecture", "is", "actually", "a", "combination", "of", "several", "GRU", "units", ".", "The", "output", "of", "the", "architecture", "is", "the", "summation", "of", "every", "unit", ".", "The", "Shorten", "Spatial", "-", "spectral", "RNN", "with", "parallel", "-", "GRU", "(", "St", "-", "SS", "-", "pGRU", ")", "is", "similar", "to", "St", "-", "SS", "-", "GRU", ",", "except", "that", "St", "-", "GRU", "is", "replaced", "by", "St", "-", "SS", "-", "pGRU", ".", "section", ":", "Experiment", "subsection", ":", "Data", "In", "the", "experiment", ",", "two", "HSI", "datasets", ",", "including", "the", "Pavia", "University", "and", "Indian", "Pines", ",", "are", "used", "to", "evaluate", "the", "performance", "of", "the", "proposed", "model", ".", "The", "Pavia", "University", "dataset", "was", "acquired", "by", "the", "Reflective", "Optics", "System", "Imaging", "Spectrometer", "(", "ROSIS", ")", "sensor", "over", "Pavia", ",", "northern", "Italy", "in", "2001", ".", "The", "corrected", "data", ",", "with", "a", "spatial", "resolution", "of", "1.3", "m", "per", "pixel", ",", "contains", "103", "spectral", "bands", "ranging", "from", "0.43", "to", "0.86", ".", "The", "image", ",", "with", "610", "340", "pixels", ",", "is", "differentiated", "into", "9", "ground", "truth", "classes", ".", "Table", "[", "reference", "]", "provides", "information", "about", "all", "classes", "of", "the", "dataset", "with", "their", "corresponding", "training", "and", "test", "sample", ".", "The", "Indian", "Pines", "dataset", "was", "acquired", "by", "the", "TAirborne", "Visible", "/", "Infrared", "Imaging", "Spectrometer", "(", "AVIRIS", ")", "sensor", "over", "the", "Indian", "Pines", "test", "site", "in", "north", "-", "western", "Indiana", "in", "1992", ".", "The", "corrected", "data", "with", "a", "moderate", "spatial", "resolution", "of", "20", "m", "contains", "200", "spectral", "bands", "ranging", "from", "0.4", "to", "2.5", ".", "The", "image", "consists", "of", "145", "145", "pixels", ",", "which", "are", "differentiated", "into", "16", "ground", "truth", "classes", ".", "Table", "[", "reference", "]", "provides", "information", "about", "all", "classes", "of", "the", "dataset", "with", "their", "corresponding", "training", "and", "test", "sample", ".", "subsection", ":", "Result", "Table", "[", "reference", "]", "and", "[", "reference", "]", "list", "the", "results", "obtained", "by", "the", "experiment", ",", "and", "Fig", ".", "[", "reference", "]", "and", "[", "reference", "]", "show", "the", "classification", "maps", "on", "the", "Pavia", "University", "dataset", "and", "the", "Indian", "Pines", "dataset", ".", "Note", "that", "the", "accuracies", "list", "in", "Table", "[", "reference", "]", "and", "[", "reference", "]", "are", "overall", "accuracies", "(", "OA", ")", "along", "with", "the", "standard", "deviation", ",", "from", "10", "independent", "runs", "on", "each", "dataset", ".", "The", "experiment", "is", "implemented", "with", "an", "Intel", "i7", "-", "7700", "K", "4.20GHz", "processor", "with", "16", "GB", "of", "RAM", "and", "an", "NVIDIA", "GTX1050Ti", "graphic", "card", "under", "Python3.6", "with", "tensorflow1.8.0", ".", "First", "of", "all", ",", "for", "all", "the", "datasets", ",", "GRU", "outperforms", "LSTM", ".", "In", "addition", ",", "it", "is", "observed", "that", "LSTM", "is", "difficult", "to", "converge", "in", "the", "experiment", ",", "while", "GRU", "is", "not", ".", "Thus", ",", "it", "is", "reasonable", "to", "indicate", "that", "GRU", "is", "a", "better", "choice", "for", "a", "HSI", "classification", "task", ".", "Furthermore", ",", "it", "is", "apparent", "that", "St", "-", "GRU", "increases", "the", "accuracy", "significantly", "by", "5.33", "%", "and", "3.52", "%", "in", "the", "Pavia", "University", "dataset", "and", "the", "Indian", "Pines", "correspondingly", ".", "With", "converlusion", "layers", ",", "St", "-", "SS", "-", "GRU", "has", "a", "better", "than", "St", "-", "GRU", ".", "The", "accuracy", "of", "St", "-", "SS", "-", "GRU", "is", "4.55", "%", "and", "6.63", "%", "higher", "than", "that", "in", "St", "-", "GRU", ".", "After", "parallel", "-", "GRU", "is", "adopted", ",", "the", "model", "gains", "the", "best", "performance", "in", "this", "experiment", ".", "The", "accuracy", "of", "St", "-", "SS", "-", "pGRU", "is", "1.64", "%", "and", "3.19", "%", "higher", "than", "St", "-", "SS", "-", "GRU", ".", "What", "is", "more", ",", "the", "standard", "deviation", "of", "St", "-", "SS", "-", "pGRU", "is", "smaller", "than", "other", "models", ",", "which", "indicate", "that", "St", "-", "SS", "-", "pGRU", "is", "more", "robust", ".", "Comparing", "the", "processing", "time", "of", "different", "methods", ",", "st", "-", "GRU", "is", "significantly", "faster", "in", "training", "than", "band", "-", "by", "-", "band", "GRU", ".", "St", "-", "SS", "-", "GRU", "and", "St", "-", "SS", "-", "pGRU", "are", "as", "slow", "as", "LSTM", "and", "GRU", "in", "training", ",", "but", "they", "have", "higher", "accuracies", "than", "LSTM", "and", "GRU", ".", "section", ":", "Conclusion", "In", "the", "study", ",", "a", "St", "-", "SS", "-", "pGRU", "model", "is", "proposed", "for", "HSI", "classification", ".", "What", "is", "more", ",", "an", "architecture", "named", "parallel", "-", "GRU", "is", "proposed", "to", "promote", "the", "performance", "and", "robustness", ".", "Then", "an", "experiment", "is", "conducted", "to", "compare", "the", "performance", "of", "different", "models", ".", "From", "the", "experiment", ",", "it", "is", "confirmed", "that", "GRU", "performs", "better", "than", "LSTM", "in", "HSI", "classification", "task", ".", "Moreover", ",", "it", "is", "apparent", "that", "the", "proposed", "models", "are", "a", "lot", "more", "accurate", ",", "more", "robust", "and", "faster", "than", "the", "traditional", "GRU", "network", ".", "Specifically", ",", "St", "-", "GRU", "effectively", "reduced", "the", "training", "time", "and", "promoted", "the", "accuracy", ".", "St", "-", "SS", "-", "GRU", "needs", "more", "time", "for", "training", "but", "gains", "a", "better", "performance", "than", "St", "-", "GRU", ".", "The", "proposed", "architecture", "parallel", "-", "GRU", "also", "provided", "a", "satisfactory", "result", "in", "the", "experiment", ".", "bibliography", ":", "References"], "entity": [[8, 11, ["Parallel", "-", "GRU"], "Method"], [12, 15, ["Hyperspectral", "Image", "Classification"], "Task"], [15, 18, ["Convolutional", "neural", "networks"], "Method"], [19, 20, ["CNNs"], "Method"], [26, 29, ["hyperspectral", "sensing", "image"], "Task"], [30, 31, ["HSI"], "Task"], [32, 33, ["classification"], "Task"], [35, 36, ["CNNs"], "Method"], [50, 53, ["recurrent", "neural", "networks"], "Method"], [54, 55, ["RNNs"], "Method"], [64, 65, ["RNNs"], "Method"], [77, 81, ["long", "-", "sequence", "task"], "Task"], [82, 83, ["RNNs"], "Method"], [104, 105, ["RNNs"], "Method"], [113, 122, ["Shorten", "Spatial", "-", "spectral", "RNN", "with", "Parallel", "-", "GRU"], "Method"], [123, 128, ["St", "-", "SS", "-", "pGRU"], "Method"], [134, 136, ["shorten", "RNN"], "Method"], [144, 150, ["band", "-", "by", "-", "band", "RNN"], "Method"], [153, 155, ["converlusion", "layer"], "Method"], [157, 161, ["St", "-", "SSpGRU", "model"], "Method"], [189, 194, ["St", "-", "SS", "-", "pGRU"], "Method"], [210, 212, ["deep", "learning"], "Method"], [213, 216, ["gated", "recurrent", "unit"], "Method"], [217, 218, ["GRU"], "Method"], [220, 225, ["long", "short", "-", "term", "memory"], "Method"], [226, 227, ["LSTM"], "Method"], [229, 232, ["recurrent", "neural", "networks"], "Method"], [233, 234, ["RNN"], "Method"], [242, 244, ["Hyperspectral", "image"], "Method"], [245, 246, ["HSI"], "Task"], [253, 256, ["remote", "sensing", "community"], "Task"], [286, 288, ["deep", "learning"], "Method"], [300, 303, ["Convolutional", "neural", "networks"], "Method"], [304, 305, ["CNNs"], "Method"], [307, 309, ["residual", "networks"], "Method"], [310, 311, ["ResNets"], "Method"], [321, 324, ["Recurrent", "neural", "networks"], "Method"], [325, 326, ["RNNs"], "Method"], [345, 346, ["CNNs"], "Method"], [347, 348, ["ResNets"], "Method"], [352, 353, ["accuracy"], "Metric"], [360, 361, ["CNNs"], "Method"], [362, 363, ["ResNets"], "Method"], [400, 401, ["CNNs"], "Method"], [402, 403, ["ResNets"], "Method"], [409, 410, ["RNNs"], "Method"], [424, 427, ["Natural", "Language", "Processing"], "Task"], [428, 429, ["NLP"], "Task"], [431, 435, ["prediction", "of", "time", "series"], "Task"], [447, 448, ["RNNs"], "Method"], [463, 464, ["HSI"], "Task"], [469, 473, ["long", "-", "sequence", "task"], "Task"], [474, 475, ["RNNs"], "Method"], [499, 505, ["long", "short", "-", "term", "memory", "network"], "Method"], [506, 507, ["LSTM"], "Method"], [513, 518, ["long", "-", "distance", "dependence", "problem"], "Task"], [519, 520, ["RNNs"], "Method"], [524, 525, ["training"], "Task"], [530, 534, ["long", "-", "sequence", "task"], "Task"], [539, 542, ["3D", "-", "CNN"], "Method"], [554, 555, ["RNNs"], "Method"], [556, 559, ["Convolutional", "-", "LSTM"], "Method"], [560, 561, ["CLSTM"], "Method"], [571, 574, ["3D", "-", "CNNs"], "Method"], [575, 576, ["CLSTM"], "Method"], [590, 591, ["accuracy"], "Metric"], [607, 608, ["LSTM"], "Method"], [612, 613, ["GRU"], "Method"], [618, 619, ["classification"], "Task"], [625, 626, ["GRU"], "Method"], [632, 633, ["classification"], "Task"], [639, 640, ["RNNs"], "Method"], [660, 661, ["training"], "Task"], [669, 670, ["LSTM"], "Method"], [676, 685, ["Shorten", "Spatial", "-", "spectral", "RNN", "with", "Parallel", "-", "GRU"], "Method"], [686, 691, ["St", "-", "SS", "-", "pGRU"], "Method"], [708, 709, ["RNN"], "Method"], [710, 711, ["GRU"], "Method"], [715, 716, ["classification"], "Task"], [725, 726, ["training"], "Task"], [727, 733, ["band", "-", "by", "-", "band", "RNN"], "Method"], [736, 738, ["converlusion", "layer"], "Method"], [743, 747, ["Spatial", "-", "spectral", "RNN"], "Method"], [748, 749, ["GRU"], "Method"], [813, 814, ["RNN"], "Method"], [815, 816, ["LSTM"], "Method"], [817, 818, ["GRU"], "Method"], [867, 870, ["Recurrent", "neural", "networks"], "Method"], [871, 872, ["RNN"], "Method"], [875, 881, ["Artificial", "neural", "network", "(", "ANN", ")"], "Method"], [882, 883, ["RNN"], "Method"], [885, 887, ["neural", "network"], "Method"], [888, 890, ["recurrent", "unit"], "Method"], [901, 904, ["sequential", "data", "analysis"], "Task"], [913, 915, ["recurrent", "unit"], "Method"], [970, 972, ["recurrent", "unit"], "Method"], [988, 990, ["recurrent", "unit"], "Method"], [1012, 1014, ["recurrent", "unit"], "Method"], [1024, 1026, ["RNN", "model"], "Method"], [1053, 1055, ["recurrent", "unit"], "Method"], [1058, 1059, ["RNN"], "Method"], [1071, 1073, ["RNN", "model"], "Method"], [1127, 1129, ["sigmoid", "function"], "Method"], [1131, 1134, ["hyperbolic", "tangent", "function"], "Method"], [1137, 1142, ["Long", "short", "-", "term", "memory"], "Method"], [1143, 1144, ["LSTM"], "Method"], [1147, 1148, ["RNN"], "Method"], [1158, 1159, ["RNN"], "Method"], [1179, 1180, ["RNN"], "Method"], [1198, 1205, ["multi", "-", "time", "-", "step", "recurrent", "units"], "Method"], [1208, 1213, ["long", "short", "-", "term", "memory"], "Method"], [1214, 1215, ["LSTM"], "Method"], [1231, 1232, ["LSTM"], "Method"], [1246, 1249, ["Gate", "’", "structure"], "Method"], [1252, 1255, ["logistic", "regression", "model"], "Method"], [1276, 1277, ["LSTM"], "Method"], [1301, 1305, ["long", "-", "distance", "propagation"], "Task"], [1309, 1310, ["LSTM"], "Method"], [1370, 1371, ["tanh"], "Method"], [1380, 1383, ["Gated", "recurrent", "unit"], "Method"], [1384, 1385, ["GRU"], "Method"], [1396, 1397, ["LSTM"], "Method"], [1423, 1424, ["GRU"], "Method"], [1428, 1429, ["LSTM"], "Method"], [1441, 1442, ["LSTM"], "Method"], [1450, 1451, ["LSTM"], "Method"], [1460, 1461, ["GRU"], "Method"], [1471, 1472, ["LSTM"], "Method"], [1477, 1478, ["LSTM"], "Method"], [1479, 1480, ["GRU"], "Method"], [1492, 1493, ["GRU"], "Method"], [1513, 1514, ["GRU"], "Method"], [1575, 1582, ["Shorten", "Spatial", "-", "spectral", "RNN", "with", "GRU"], "Method"], [1583, 1588, ["St", "-", "SS", "-", "GRU"], "Method"], [1590, 1597, ["Shorten", "Spatial", "-", "spectral", "RNN", "with", "GRU"], "Method"], [1598, 1603, ["St", "-", "SS", "-", "GRU"], "Method"], [1643, 1648, ["St", "-", "SS", "-", "GRU"], "Method"], [1651, 1655, ["3D", "-", "Convolutional", "layer"], "Method"], [1669, 1671, ["convolution", "kernels"], "Method"], [1731, 1735, ["Shorten", "RNN", "with", "GRU"], "Method"], [1736, 1739, ["St", "-", "GRU"], "Method"], [1744, 1747, ["St", "-", "GRU"], "Method"], [1757, 1760, ["1D", "converlusion", "layer"], "Method"], [1761, 1762, ["GRU"], "Method"], [1795, 1798, ["Parallel", "-", "GRU"], "Method"], [1799, 1800, ["pGRU"], "Method"], [1850, 1859, ["Shorten", "Spatial", "-", "spectral", "RNN", "with", "parallel", "-", "GRU"], "Method"], [1860, 1865, ["St", "-", "SS", "-", "pGRU"], "Method"], [1869, 1874, ["St", "-", "SS", "-", "GRU"], "Method"], [1877, 1880, ["St", "-", "GRU"], "Method"], [1883, 1888, ["St", "-", "SS", "-", "pGRU"], "Method"], [1905, 1907, ["Pavia", "University"], "Material"], [1908, 1910, ["Indian", "Pines"], "Material"], [1923, 1926, ["Pavia", "University", "dataset"], "Material"], [1930, 1935, ["Reflective", "Optics", "System", "Imaging", "Spectrometer"], "Method"], [1936, 1937, ["ROSIS"], "Method"], [2008, 2011, ["Indian", "Pines", "dataset"], "Material"], [2015, 2017, ["TAirborne", "Visible"], "Method"], [2018, 2021, ["Infrared", "Imaging", "Spectrometer"], "Method"], [2022, 2023, ["AVIRIS"], "Method"], [2027, 2029, ["Indian", "Pines"], "Material"], [2132, 2135, ["Pavia", "University", "dataset"], "Material"], [2137, 2140, ["Indian", "Pines", "dataset"], "Material"], [2144, 2145, ["accuracies"], "Metric"], [2157, 2158, ["accuracies"], "Metric"], [2159, 2160, ["OA"], "Metric"], [2195, 2199, ["NVIDIA", "GTX1050Ti", "graphic", "card"], "Material"], [2200, 2201, ["Python3.6"], "Method"], [2202, 2203, ["tensorflow1.8.0"], "Method"], [2213, 2214, ["GRU"], "Method"], [2215, 2216, ["LSTM"], "Method"], [2224, 2225, ["LSTM"], "Method"], [2234, 2235, ["GRU"], "Method"], [2246, 2247, ["GRU"], "Method"], [2263, 2266, ["St", "-", "GRU"], "Method"], [2268, 2269, ["accuracy"], "Metric"], [2278, 2281, ["Pavia", "University", "dataset"], "Material"], [2283, 2285, ["Indian", "Pines"], "Material"], [2288, 2290, ["converlusion", "layers"], "Method"], [2291, 2296, ["St", "-", "SS", "-", "GRU"], "Method"], [2300, 2303, ["St", "-", "GRU"], "Method"], [2305, 2306, ["accuracy"], "Metric"], [2307, 2312, ["St", "-", "SS", "-", "GRU"], "Method"], [2322, 2325, ["St", "-", "GRU"], "Method"], [2344, 2345, ["accuracy"], "Metric"], [2346, 2351, ["St", "-", "SS", "-", "pGRU"], "Method"], [2359, 2364, ["St", "-", "SS", "-", "GRU"], "Method"], [2373, 2378, ["St", "-", "SS", "-", "pGRU"], "Method"], [2387, 2392, ["St", "-", "SS", "-", "pGRU"], "Method"], [2398, 2400, ["processing", "time"], "Metric"], [2411, 2412, ["training"], "Task"], [2435, 2436, ["LSTM"], "Method"], [2437, 2438, ["GRU"], "Method"], [2439, 2440, ["training"], "Task"], [2445, 2446, ["accuracies"], "Metric"], [2447, 2448, ["LSTM"], "Method"], [2449, 2450, ["GRU"], "Method"], [2459, 2465, ["St", "-", "SS", "-", "pGRU", "model"], "Method"], [2488, 2489, ["robustness"], "Metric"], [2511, 2512, ["GRU"], "Method"], [2515, 2516, ["LSTM"], "Method"], [2548, 2551, ["St", "-", "GRU"], "Method"], [2554, 2556, ["training", "time"], "Metric"], [2559, 2560, ["accuracy"], "Metric"], [2561, 2566, ["St", "-", "SS", "-", "GRU"], "Method"], [2570, 2571, ["training"], "Task"], [2577, 2580, ["St", "-", "GRU"], "Method"], [60, 61, ["HSI"], "Task"], [61, 62, ["classification"], "Task"], [130, 131, ["HSI"], "Task"], [131, 132, ["classification"], "Task"], [182, 183, ["GRU"], "Method"], [238, 239, ["classification"], "Task"], [270, 271, ["HSI"], "Task"], [297, 298, ["HSI"], "Task"], [298, 299, ["classification"], "Task"], [318, 319, ["HSI"], "Task"], [319, 320, ["classification"], "Task"], [331, 332, ["HSI"], "Task"], [332, 333, ["classification"], "Task"], [355, 356, ["classification"], "Task"], [545, 546, ["HSI"], "Task"], [546, 547, ["classification"], "Task"], [568, 569, ["HSI"], "Task"], [569, 570, ["classification"], "Task"], [654, 655, ["LSTM"], "Method"], [775, 776, ["GRU"], "Method"], [1035, 1036, ["classification"], "Task"], [1463, 1464, ["HSI"], "Task"], [1464, 1465, ["classification"], "Task"], [1606, 1607, ["HSI"], "Task"], [1607, 1608, ["classification"], "Task"], [1783, 1784, ["GRU"], "Method"], [1811, 1812, ["GRU"], "Method"], [1815, 1816, ["pGRU"], "Method"], [1834, 1835, ["GRU"], "Method"], [1900, 1901, ["HSI"], "Task"], [2128, 2129, ["classification"], "Task"], [2253, 2254, ["HSI"], "Task"], [2254, 2255, ["classification"], "Task"], [2329, 2330, ["GRU"], "Method"], [2406, 2407, ["GRU"], "Method"], [2418, 2419, ["GRU"], "Method"], [2424, 2425, ["GRU"], "Method"], [2426, 2431, ["St", "-", "SS", "-", "pGRU"], "Method"], [2468, 2469, ["HSI"], "Task"], [2469, 2470, ["classification"], "Task"], [2480, 2481, ["GRU"], "Method"], [2517, 2518, ["HSI"], "Task"], [2518, 2519, ["classification"], "Task"], [2543, 2544, ["GRU"], "Method"], [2586, 2587, ["GRU"], "Method"]]} {"doc_id": "3208b5abafa1146e7df1073090523c822909275e", "text": ["document", ":", "Text", "Classification", "Improved", "by", "Integrating", "Bidirectional", "LSTM", "with", "Two", "-", "dimensional", "Max", "Pooling", "Recurrent", "Neural", "Network", "(", "RNN", ")", "is", "one", "of", "the", "most", "popular", "architectures", "used", "in", "Natural", "Language", "Processsing", "(", "NLP", ")", "tasks", "because", "its", "recurrent", "structure", "is", "very", "suitable", "to", "process", "variable", "-", "length", "text", ".", "RNN", "can", "utilize", "distributed", "representations", "of", "words", "by", "first", "converting", "the", "tokens", "comprising", "each", "text", "into", "vectors", ",", "which", "form", "a", "matrix", ".", "And", "this", "matrix", "includes", "two", "dimensions", ":", "the", "time", "-", "step", "dimension", "and", "the", "feature", "vector", "dimension", ".", "Then", "most", "existing", "models", "usually", "utilize", "one", "-", "dimensional", "(", "1D", ")", "max", "pooling", "operation", "or", "attention", "-", "based", "operation", "only", "on", "the", "time", "-", "step", "dimension", "to", "obtain", "a", "fixed", "-", "length", "vector", ".", "However", ",", "the", "features", "on", "the", "feature", "vector", "dimension", "are", "not", "mutually", "independent", ",", "and", "simply", "applying", "1D", "pooling", "operation", "over", "the", "time", "-", "step", "dimension", "independently", "may", "destroy", "the", "structure", "of", "the", "feature", "representation", ".", "On", "the", "other", "hand", ",", "applying", "two", "-", "dimensional", "(", "2D", ")", "pooling", "operation", "over", "the", "two", "dimensions", "may", "sample", "more", "meaningful", "features", "for", "sequence", "modeling", "tasks", ".", "To", "integrate", "the", "features", "on", "both", "dimensions", "of", "the", "matrix", ",", "this", "paper", "explores", "applying", "2D", "max", "pooling", "operation", "to", "obtain", "a", "fixed", "-", "length", "representation", "of", "the", "text", ".", "This", "paper", "also", "utilizes", "2D", "convolution", "to", "sample", "more", "meaningful", "information", "of", "the", "matrix", ".", "Experiments", "are", "conducted", "on", "six", "text", "classification", "tasks", ",", "including", "sentiment", "analysis", ",", "question", "classification", ",", "subjectivity", "classification", "and", "newsgroup", "classification", ".", "Compared", "with", "the", "state", "-", "of", "-", "the", "-", "art", "models", ",", "the", "proposed", "models", "achieve", "excellent", "performance", "on", "4", "out", "of", "6", "tasks", ".", "Specifically", ",", "one", "of", "the", "proposed", "models", "achieves", "highest", "accuracy", "on", "Stanford", "Sentiment", "Treebank", "binary", "classification", "and", "fine", "-", "grained", "classification", "tasks", ".", "section", ":", "Introduction", "This", "work", "is", "licenced", "under", "a", "Creative", "Commons", "Attribution", "4.0", "International", "Licence", ".", "Licence", "details", ":", "Text", "classification", "is", "an", "essential", "component", "in", "many", "NLP", "applications", ",", "such", "as", "sentiment", "analysis", ",", "relation", "extraction", "and", "spam", "detection", ".", "Therefore", ",", "it", "has", "attracted", "considerable", "attention", "from", "many", "researchers", ",", "and", "various", "types", "of", "models", "have", "been", "proposed", ".", "As", "a", "traditional", "method", ",", "the", "bag", "-", "of", "-", "words", "(", "BoW", ")", "model", "treats", "texts", "as", "unordered", "sets", "of", "words", ".", "In", "this", "way", ",", "however", ",", "it", "fails", "to", "encode", "word", "order", "and", "syntactic", "feature", ".", "Recently", ",", "order", "-", "sensitive", "models", "based", "on", "neural", "networks", "have", "achieved", "tremendous", "success", "for", "text", "classification", ",", "and", "shown", "more", "significant", "progress", "compared", "with", "BoW", "models", ".", "The", "challenge", "for", "textual", "modeling", "is", "how", "to", "capture", "features", "for", "different", "text", "units", ",", "such", "as", "phrases", ",", "sentences", "and", "documents", ".", "Benefiting", "from", "its", "recurrent", "structure", ",", "RNN", ",", "as", "an", "alternative", "type", "of", "neural", "networks", ",", "is", "very", "suitable", "to", "process", "the", "variable", "-", "length", "text", ".", "RNN", "can", "capitalize", "on", "distributed", "representations", "of", "words", "by", "first", "converting", "the", "tokens", "comprising", "each", "text", "into", "vectors", ",", "which", "form", "a", "matrix", ".", "This", "matrix", "includes", "two", "dimensions", ":", "the", "time", "-", "step", "dimension", "and", "the", "feature", "vector", "dimension", ",", "and", "it", "will", "be", "updated", "in", "the", "process", "of", "learning", "feature", "representation", ".", "Then", "RNN", "utilizes", "1D", "max", "pooling", "operation", "or", "attention", "-", "based", "operation", ",", "which", "extracts", "maximum", "values", "or", "generates", "a", "weighted", "representation", "over", "the", "time", "-", "step", "dimension", "of", "the", "matrix", ",", "to", "obtain", "a", "fixed", "-", "length", "vector", ".", "Both", "of", "the", "two", "operators", "ignore", "features", "on", "the", "feature", "vector", "dimension", ",", "which", "maybe", "important", "for", "sentence", "representation", ",", "therefore", "the", "use", "of", "1D", "max", "pooling", "and", "attention", "-", "based", "operators", "may", "pose", "a", "serious", "limitation", ".", "Convolutional", "Neural", "Networks", "(", "CNN", ")", "utilizes", "1D", "convolution", "to", "perform", "the", "feature", "mapping", ",", "and", "then", "applies", "1D", "max", "pooling", "operation", "over", "the", "time", "-", "step", "dimension", "to", "obtain", "a", "fixed", "-", "length", "output", ".", "However", "the", "elements", "in", "the", "matrix", "learned", "by", "RNN", "are", "not", "independent", ",", "as", "RNN", "reads", "a", "sentence", "word", "by", "word", ",", "one", "can", "effectively", "treat", "the", "matrix", "as", "an", "’", "image", "’", ".", "Unlike", "in", "NLP", ",", "CNN", "in", "image", "processing", "tasks", "applies", "2D", "convolution", "and", "2D", "pooling", "operation", "to", "get", "a", "representation", "of", "the", "input", ".", "It", "is", "a", "good", "choice", "to", "utilize", "2D", "convolution", "and", "2D", "pooling", "to", "sample", "more", "meaningful", "features", "on", "both", "the", "time", "-", "step", "dimension", "and", "the", "feature", "vector", "dimension", "for", "text", "classification", ".", "Above", "all", ",", "this", "paper", "proposes", "Bidirectional", "Long", "Short", "-", "Term", "Memory", "Networks", "with", "Two", "-", "Dimensional", "Max", "Pooling", "(", "BLSTM", "-", "2DPooling", ")", "to", "capture", "features", "on", "both", "the", "time", "-", "step", "dimension", "and", "the", "feature", "vector", "dimension", ".", "It", "first", "utilizes", "Bidirectional", "Long", "Short", "-", "Term", "Memory", "Networks", "(", "BLSTM", ")", "to", "transform", "the", "text", "into", "vectors", ".", "And", "then", "2D", "max", "pooling", "operation", "is", "utilized", "to", "obtain", "a", "fixed", "-", "length", "vector", ".", "This", "paper", "also", "applies", "2D", "convolution", "(", "BLSTM", "-", "2DCNN", ")", "to", "capture", "more", "meaningful", "features", "to", "represent", "the", "input", "text", ".", "The", "contributions", "of", "this", "paper", "can", "be", "summarized", "as", "follows", ":", "This", "paper", "proposes", "a", "combined", "framework", ",", "which", "utilizes", "BLSTM", "to", "capture", "long", "-", "term", "sentence", "dependencies", ",", "and", "extracts", "features", "by", "2D", "convolution", "and", "2D", "max", "pooling", "operation", "for", "sequence", "modeling", "tasks", ".", "To", "the", "best", "of", "our", "knowledge", ",", "this", "work", "is", "the", "first", "example", "of", "using", "2D", "convolution", "and", "2D", "max", "pooling", "operation", "in", "NLP", "tasks", ".", "This", "work", "introduces", "two", "combined", "models", "BLSTM", "-", "2DPooling", "and", "BLSTM", "-", "2DCNN", ",", "and", "verifies", "them", "on", "six", "text", "classification", "tasks", ",", "including", "sentiment", "analysis", ",", "question", "classification", ",", "subjectivity", "classification", ",", "and", "newsgroups", "classification", ".", "Compared", "with", "the", "state", "-", "of", "-", "the", "-", "art", "models", ",", "BLSTM", "-", "2DCNN", "achieves", "excellent", "performance", "on", "out", "of", "tasks", ".", "Specifically", ",", "it", "achieves", "highest", "accuracy", "on", "Stanford", "Sentiment", "Treebank", "binary", "classification", "and", "fine", "-", "grained", "classification", "tasks", ".", "To", "better", "understand", "the", "effect", "of", "2D", "convolution", "and", "2D", "max", "pooling", "operation", ",", "this", "paper", "conducts", "experiments", "on", "Stanford", "Sentiment", "Treebank", "fine", "-", "grained", "task", ".", "It", "first", "depicts", "the", "performance", "of", "the", "proposed", "models", "on", "different", "length", "of", "sentences", ",", "and", "then", "conducts", "a", "sensitivity", "analysis", "of", "2D", "filter", "and", "max", "pooling", "size", ".", "The", "remainder", "of", "the", "paper", "is", "organized", "as", "follows", ".", "In", "Section", "2", ",", "the", "related", "work", "about", "text", "classification", "is", "reviewed", ".", "Section", "3", "presents", "the", "BLSTM", "-", "2DCNN", "architectures", "for", "text", "classification", "in", "detail", ".", "Section", "4", "describes", "details", "about", "the", "setup", "of", "the", "experiments", ".", "Section", "5", "presents", "the", "experimental", "results", ".", "The", "conclusion", "is", "drawn", "in", "the", "section", "6", ".", "section", ":", "Related", "Work", "Deep", "learning", "based", "neural", "network", "models", "have", "achieved", "great", "improvement", "on", "text", "classification", "tasks", ".", "These", "models", "generally", "consist", "of", "a", "projection", "layer", "that", "maps", "words", "of", "text", "to", "vectors", ".", "And", "then", "combine", "the", "vectors", "with", "different", "neural", "networks", "to", "make", "a", "fixed", "-", "length", "representation", ".", "According", "to", "the", "structure", ",", "they", "may", "divide", "into", "four", "categories", ":", "Recursive", "Neural", "Networks", "(", "RecNN", ")", ",", "RNN", ",", "CNN", "and", "other", "neural", "networks", ".", "Recursive", "Neural", "Networks", ":", "RecNN", "is", "defined", "over", "recursive", "tree", "structures", ".", "In", "the", "type", "of", "recursive", "models", ",", "information", "from", "the", "leaf", "nodes", "of", "a", "tree", "and", "its", "internal", "nodes", "are", "combined", "in", "a", "bottom", "-", "up", "manner", ".", "socher2013recursive", "introduced", "recursive", "neural", "tensor", "network", "to", "build", "representations", "of", "phrases", "and", "sentences", "by", "combining", "neighbour", "constituents", "based", "on", "the", "parsing", "tree", ".", "irsoy2014deep", "proposed", "deep", "recursive", "neural", "network", ",", "which", "is", "constructed", "by", "stacking", "multiple", "recursive", "layers", "on", "top", "of", "each", "other", ",", "to", "modeling", "sentence", ".", "Recurrent", "Neural", "Networks", ":", "RNN", "has", "obtained", "much", "attention", "because", "of", "their", "superior", "ability", "to", "preserve", "sequence", "information", "over", "time", ".", "tang2015target", "developed", "target", "dependent", "Long", "Short", "-", "Term", "Memory", "Networks", "(", "LSTM", ")", ",", "where", "target", "information", "is", "automatically", "taken", "into", "account", ".", "tai2015improved", "generalized", "LSTM", "to", "Tree", "-", "LSTM", "where", "each", "LSTM", "unit", "gains", "information", "from", "its", "children", "units", ".", "zhou2016attention", "introduced", "BLSTM", "with", "attention", "mechanism", "to", "automatically", "select", "features", "that", "have", "a", "decisive", "effect", "on", "classification", ".", "yang2016hierarchical", "introduced", "a", "hierarchical", "network", "with", "two", "levels", "of", "attention", "mechanisms", ",", "which", "are", "word", "attention", "and", "sentence", "attention", ",", "for", "document", "classification", ".", "This", "paper", "also", "implements", "an", "attention", "-", "based", "model", "BLSTM", "-", "Att", "like", "the", "model", "in", "zhou2016attention", ".", "Convolution", "Neural", "Networks", ":", "CNN", "is", "a", "feedforward", "neural", "network", "with", "2D", "convolution", "layers", "and", "2D", "pooling", "layers", ",", "originally", "developed", "for", "image", "processing", ".", "Then", "CNN", "is", "applied", "to", "NLP", "tasks", ",", "such", "as", "sentence", "classification", ",", "and", "relation", "classification", ".", "The", "difference", "is", "that", "the", "common", "CNN", "in", "NLP", "tasks", "is", "made", "up", "of", "1D", "convolution", "layers", "and", "1D", "pooling", "layers", ".", "kim2014convolutional", "defined", "a", "CNN", "architecture", "with", "two", "channels", ".", "kalchbrenner2014convolutional", "proposed", "a", "dynamic", "-", "max", "pooling", "mechanism", "for", "sentence", "modeling", ".", "conducted", "a", "sensitivity", "analysis", "of", "one", "-", "layer", "CNN", "to", "explore", "the", "effect", "of", "architecture", "components", "on", "model", "performance", ".", "yin2016multichannel", "introduced", "multichannel", "embeddings", "and", "unsupervised", "pretraining", "to", "improve", "classification", "accuracy", ".", "conducted", "a", "sensitivity", "analysis", "of", "one", "-", "layer", "CNN", "to", "explore", "the", "effect", "of", "architecture", "components", "on", "model", "performance", ".", "Usually", "there", "is", "a", "misunderstanding", "that", "1D", "convolutional", "filter", "in", "NLP", "tasks", "has", "one", "dimension", ".", "Actually", "it", "has", "two", "dimensions", ",", "where", ",", ".", "As", "is", "equal", "to", "the", "word", "embeddings", "size", ",", "the", "window", "slides", "only", "on", "the", "time", "-", "step", "dimension", ",", "so", "the", "convolution", "is", "usually", "called", "1D", "convolution", ".", "While", "in", "this", "paper", "varies", "from", "2", "to", ",", "to", "avoid", "confusion", "with", "common", "CNN", ",", "the", "convolution", "in", "this", "work", "is", "named", "as", "2D", "convolution", ".", "The", "details", "will", "be", "described", "in", "Section", "[", "reference", "]", ".", "Other", "Neural", "Networks", ":", "In", "addition", "to", "the", "models", "described", "above", ",", "lots", "of", "other", "neural", "networks", "have", "been", "proposed", "for", "text", "classification", ".", "iyyer2015deep", "introduced", "a", "deep", "averaging", "network", ",", "which", "fed", "an", "unweighted", "average", "of", "word", "embeddings", "through", "multiple", "hidden", "layers", "before", "classification", ".", "zhou2015c", "used", "CNN", "to", "extract", "a", "sequence", "of", "higher", "-", "level", "phrase", "representations", ",", "then", "the", "representations", "were", "fed", "into", "a", "LSTM", "to", "obtain", "the", "sentence", "representation", ".", "The", "proposed", "model", "BLSTM", "-", "2DCNN", "is", "most", "relevant", "to", "DSCNN", "and", "RCNN", ".", "The", "difference", "is", "that", "the", "former", "two", "utilize", "LSTM", ",", "bidirectional", "RNN", "respectively", ",", "while", "this", "work", "applies", "BLSTM", ",", "to", "capture", "long", "-", "term", "sentence", "dependencies", ".", "After", "that", "the", "former", "two", "both", "apply", "1D", "convolution", "and", "1D", "max", "pooling", "operation", ",", "while", "this", "paper", "uses", "2D", "convolution", "and", "2D", "max", "pooling", "operation", ",", "to", "obtain", "the", "whole", "sentence", "representation", ".", "section", ":", "Model", "As", "shown", "in", "Figure", "1", ",", "the", "overall", "model", "consists", "of", "four", "parts", ":", "BLSTM", "Layer", ",", "Two", "-", "dimensional", "Convolution", "Layer", ",", "Two", "dimensional", "max", "pooling", "Layer", ",", "and", "Output", "Layer", ".", "The", "details", "of", "different", "components", "are", "described", "in", "the", "following", "sections", ".", "subsection", ":", "BLSTM", "Layer", "LSTM", "was", "firstly", "proposed", "by", "hochreiter1997long", "to", "overcome", "the", "gradient", "vanishing", "problem", "of", "RNN", ".", "The", "main", "idea", "is", "to", "introduce", "an", "adaptive", "gating", "mechanism", ",", "which", "decides", "the", "degree", "to", "keep", "the", "previous", "state", "and", "memorize", "the", "extracted", "features", "of", "the", "current", "data", "input", ".", "Given", "a", "sequence", ",", "where", "is", "the", "length", "of", "input", "text", ",", "LSTM", "processes", "it", "word", "by", "word", ".", "At", "time", "-", "step", ",", "the", "memory", "and", "the", "hidden", "state", "are", "updated", "with", "the", "following", "equations", ":", "where", "is", "the", "input", "at", "the", "current", "time", "-", "step", ",", ",", "and", "is", "the", "input", "gate", "activation", ",", "forget", "gate", "activation", "and", "output", "gate", "activation", "respectively", ",", "is", "the", "current", "cell", "state", ",", "denotes", "the", "logistic", "sigmoid", "function", "and", "denotes", "element", "-", "wise", "multiplication", ".", "For", "the", "sequence", "modeling", "tasks", ",", "it", "is", "beneficial", "to", "have", "access", "to", "the", "past", "context", "as", "well", "as", "the", "future", "context", ".", "schuster1997bidirectional", "proposed", "BLSTM", "to", "extend", "the", "unidirectional", "LSTM", "by", "introducing", "a", "second", "hidden", "layer", ",", "where", "the", "hidden", "to", "hidden", "connections", "flow", "in", "opposite", "temporal", "order", ".", "Therefore", ",", "the", "model", "is", "able", "to", "exploit", "information", "from", "both", "the", "past", "and", "the", "future", ".", "In", "this", "paper", ",", "BLSTM", "is", "utilized", "to", "capture", "the", "past", "and", "the", "future", "information", ".", "As", "shown", "in", "Figure", "1", ",", "the", "network", "contains", "two", "sub", "-", "networks", "for", "the", "forward", "and", "backward", "sequence", "context", "respectively", ".", "The", "output", "of", "the", "word", "is", "shown", "in", "the", "following", "equation", ":", "Here", ",", "element", "-", "wise", "sum", "is", "used", "to", "combine", "the", "forward", "and", "backward", "pass", "outputs", ".", "subsection", ":", "Convolutional", "Neural", "Networks", "Since", "BLSTM", "has", "access", "to", "the", "future", "context", "as", "well", "as", "the", "past", "context", ",", "is", "related", "to", "all", "the", "other", "words", "in", "the", "text", ".", "One", "can", "effectively", "treat", "the", "matrix", ",", "which", "consists", "of", "feature", "vectors", ",", "as", "an", "’", "image", "’", ",", "so", "2D", "convolution", "and", "2D", "max", "pooling", "operation", "can", "be", "utilized", "to", "capture", "more", "meaningful", "information", ".", "subsubsection", ":", "Two", "-", "dimensional", "Convolution", "Layer", "A", "matrix", ",", ",", "is", "obtained", "from", "BLSTM", "Layer", ",", "where", "is", "the", "size", "of", "word", "embeddings", ".", "Then", "narrow", "convolution", "is", "utilized", "to", "extract", "local", "features", "over", ".", "A", "convolution", "operation", "involves", "a", "2D", "filter", ",", "which", "is", "applied", "to", "a", "window", "of", "k", "words", "and", "d", "feature", "vectors", ".", "For", "example", ",", "a", "feature", "is", "generated", "from", "a", "window", "of", "vectors", "by", "where", "ranges", "from", "1", "to", ",", "ranges", "from", "1", "to", ",", "represents", "dot", "product", ",", "is", "a", "bias", "and", "an", "is", "a", "non", "-", "linear", "function", "such", "as", "the", "hyperbolic", "tangent", ".", "This", "filter", "is", "applied", "to", "each", "possible", "window", "of", "the", "matrix", "to", "produce", "a", "feature", "map", ":", "with", ".", "It", "has", "described", "the", "process", "of", "one", "convolution", "filter", ".", "The", "convolution", "layer", "may", "have", "multiple", "filters", "for", "the", "same", "size", "filter", "to", "learn", "complementary", "features", ",", "or", "multiple", "kinds", "of", "filter", "with", "different", "size", ".", "subsubsection", ":", "Two", "-", "dimensional", "Max", "Pooling", "Layer", "Then", "2D", "max", "pooling", "operation", "is", "utilized", "to", "obtain", "a", "fixed", "length", "vector", ".", "For", "a", "2D", "max", "pooling", ",", "it", "is", "applied", "to", "each", "possible", "window", "of", "matrix", "O", "to", "extract", "the", "maximum", "value", ":", "where", "represents", "the", "2D", "max", "pooling", "function", ",", ",", "and", ".", "Then", "the", "pooling", "results", "are", "combined", "as", "follows", ":", "where", ",", "and", "the", "length", "of", "is", ".", "subsection", ":", "Output", "Layer", "For", "text", "classification", ",", "the", "output", "of", "2D", "Max", "Pooling", "Layer", "is", "the", "whole", "representation", "of", "the", "input", "text", ".", "And", "then", "it", "is", "passed", "to", "a", "softmax", "classifier", "layer", "to", "predict", "the", "semantic", "relation", "label", "from", "a", "discrete", "set", "of", "classes", ".", "The", "classifier", "takes", "the", "hidden", "state", "as", "input", ":", "A", "reasonable", "training", "objective", "to", "be", "minimized", "is", "the", "categorical", "cross", "-", "entropy", "loss", ".", "The", "loss", "is", "calculated", "as", "a", "regularized", "sum", ":", "where", "is", "the", "one", "-", "hot", "represented", "ground", "truth", ",", "is", "the", "estimated", "probability", "for", "each", "class", "by", "softmax", ",", "is", "the", "number", "of", "target", "classes", ",", "and", "is", "an", "L2", "regularization", "hyper", "-", "parameter", ".", "Training", "is", "done", "through", "stochastic", "gradient", "descent", "over", "shuffled", "mini", "-", "batches", "with", "the", "AdaDelta", "update", "rule", ".", "Training", "details", "are", "further", "introduced", "in", "Section", "[", "reference", "]", ".", "section", ":", "Experimental", "Setup", "subsection", ":", "Datasets", "The", "proposed", "models", "are", "tested", "on", "six", "datasets", ".", "Summary", "statistics", "of", "the", "datasets", "are", "in", "Table", "1", ".", "MRhttps:", "//", "www.cs.cornell.edu", "/", "people", "/", "pabo", "/", "movie", "-", "review", "-", "data", "/", ":", "Sentence", "polarity", "dataset", "from", "pang2005seeing", ".", "The", "task", "is", "to", "detect", "positive", "/", "negative", "reviews", ".", "SST", "-", "1http:", "//", "nlp.stanford.edu", "/", "sentiment", "/", ":", "Stanford", "Sentiment", "Treebank", "is", "an", "extension", "of", "MR", "from", "socher2013recursive", ".", "The", "aim", "is", "to", "classify", "a", "review", "as", "fine", "-", "grained", "labels", "(", "very", "negative", ",", "negative", ",", "neutral", ",", "positive", ",", "very", "positive", ")", ".", "SST", "-", "2", ":", "Same", "as", "SST", "-", "1", "but", "with", "neutral", "reviews", "removed", "and", "binary", "labels", "(", "negative", ",", "positive", ")", ".", "For", "both", "experiments", ",", "phrases", "and", "sentences", "are", "used", "to", "train", "the", "model", ",", "but", "only", "sentences", "are", "scored", "at", "test", "time", ".", "Thus", "the", "training", "set", "is", "an", "order", "of", "magnitude", "larger", "than", "listed", "in", "table", "1", ".", "Subjhttp:", "//", "www.cs.cornell.edu", "/", "people", "/", "pabo", "/", "movie", "-", "review", "-", "data", "/", ":", "Subjectivity", "dataset", ".", "The", "task", "is", "to", "classify", "a", "sentence", "as", "being", "subjective", "or", "objective", ".", "TREChttp:", "//", "cogcomp.cs.illinois.edu", "/", "Data", "/", "QA", "/", "QC", "/", ":", "Question", "classification", "dataset", ".", "The", "task", "involves", "classifying", "a", "question", "into", "6", "question", "types", "(", "abbreviation", ",", "description", ",", "entity", ",", "human", ",", "location", ",", "numeric", "value", ")", ".", "20Newsgroupshttp:", "//", "web.ist.utl.pt", "/", "acardoso", "/", "datasets", "/", ":", "The", "20Ng", "dataset", "contains", "messages", "from", "twenty", "newsgroups", ".", "We", "use", "the", "bydate", "version", "preprocessed", "by", "cachopo2007improving", ".", "We", "select", "four", "major", "categories", "(", "comp", ",", "politics", ",", "rec", "and", "religion", ")", "followed", "by", "hingmire2013document", ".", "subsection", ":", "Word", "Embeddings", "The", "word", "embeddings", "are", "pre", "-", "trained", "on", "much", "larger", "unannotated", "corpora", "to", "achieve", "better", "generalization", "given", "limited", "amount", "of", "training", "data", ".", "In", "particular", ",", "our", "experiments", "utilize", "the", "GloVe", "embeddings", "trained", "by", "pennington2014glove", "on", "6", "billion", "tokens", "of", "Wikipedia", "2014", "and", "Gigaword", "5", ".", "Words", "not", "present", "in", "the", "set", "of", "pre", "-", "trained", "words", "are", "initialized", "by", "randomly", "sampling", "from", "uniform", "distribution", "in", ".", "The", "word", "embeddings", "are", "fine", "-", "tuned", "during", "training", "to", "improve", "the", "performance", "of", "classification", ".", "subsection", ":", "Hyper", "-", "parameter", "Settings", "For", "datasets", "without", "a", "standard", "development", "set", "we", "randomly", "select", "of", "the", "training", "data", "as", "the", "development", "set", ".", "The", "evaluation", "metric", "of", "the", "20Ng", "is", "the", "Macro", "-", "F1", "measure", "followed", "by", "the", "state", "-", "of", "-", "the", "-", "art", "work", "and", "the", "other", "five", "datasets", "use", "accuracy", "as", "the", "metric", ".", "The", "final", "hyper", "-", "parameters", "are", "as", "follows", ".", "The", "dimension", "of", "word", "embeddings", "is", "300", ",", "the", "hidden", "units", "of", "LSTM", "is", "300", ".", "We", "use", "100", "convolutional", "filters", "each", "for", "window", "sizes", "of", "(", "3", ",", "3", ")", ",", "2D", "pooling", "size", "of", "(", "2", ",", "2", ")", ".", "We", "set", "the", "mini", "-", "batch", "size", "as", "10", "and", "the", "learning", "rate", "of", "AdaDelta", "as", "the", "default", "value", "1.0", ".", "For", "regularization", ",", "we", "employ", "Dropout", "operation", "with", "dropout", "rate", "of", "0.5", "for", "the", "word", "embeddings", ",", "0.2", "for", "the", "BLSTM", "layer", "and", "0.4", "for", "the", "penultimate", "layer", ",", "we", "also", "use", "l2", "penalty", "with", "coefficient", "over", "the", "parameters", ".", "These", "values", "are", "chosen", "via", "a", "grid", "search", "on", "the", "SST", "-", "1", "development", "set", ".", "We", "only", "tune", "these", "hyper", "-", "parameters", ",", "and", "more", "finer", "tuning", ",", "such", "as", "using", "different", "numbers", "of", "hidden", "units", "of", "LSTM", "layer", ",", "or", "using", "wide", "convolution", ",", "may", "further", "improve", "the", "performance", ".", "section", ":", "Results", "subsection", ":", "Overall", "Performance", "This", "work", "implements", "four", "models", ",", "BLSTM", ",", "BLSTM", "-", "Att", ",", "BLSTM", "-", "2DPooling", ",", "and", "BLSTM", "-", "2DCNN", ".", "Table", "2", "presents", "the", "performance", "of", "the", "four", "models", "and", "other", "state", "-", "of", "-", "the", "-", "art", "models", "on", "six", "classification", "tasks", ".", "The", "BLSTM", "-", "2DCNN", "model", "achieves", "excellent", "performance", "on", "4", "out", "of", "6", "tasks", ".", "Especially", ",", "it", "achieves", "and", "test", "accuracies", "on", "SST", "-", "1", "and", "SST", "-", "2", "respectively", ".", "BLSTM", "-", "2DPooling", "performs", "worse", "than", "the", "state", "-", "of", "-", "the", "-", "art", "models", ".", "While", "we", "expect", "performance", "gains", "through", "the", "use", "of", "2D", "convolution", ",", "we", "are", "surprised", "at", "the", "magnitude", "of", "the", "gains", ".", "BLSTM", "-", "CNN", "beats", "all", "baselines", "on", "SST", "-", "1", ",", "SST", "-", "2", ",", "and", "TREC", "datasets", ".", "As", "for", "Subj", "and", "MR", "datasets", ",", "BLSTM", "-", "2DCNN", "gets", "a", "second", "higher", "accuracies", ".", "Some", "of", "the", "previous", "techniques", "only", "work", "on", "sentences", ",", "but", "not", "paragraphs", "/", "documents", "with", "several", "sentences", ".", "Our", "question", "becomes", "whether", "it", "is", "possible", "to", "use", "our", "models", "for", "datasets", "that", "have", "a", "substantial", "number", "of", "words", ",", "such", "as", "20Ng", "and", "where", "the", "content", "consists", "of", "many", "different", "topics", ".", "For", "that", "purpose", ",", "this", "paper", "tests", "the", "four", "models", "on", "document", "-", "level", "dataset", "20Ng", ",", "by", "treating", "the", "document", "as", "a", "long", "sentence", ".", "Compared", "with", "RCNN", ",", "BLSTM", "-", "2DCNN", "achieves", "a", "comparable", "result", ".", "Besides", ",", "this", "paper", "also", "compares", "with", "ReNN", ",", "RNN", ",", "CNN", "and", "other", "neural", "networks", ":", "Compared", "with", "ReNN", ",", "the", "proposed", "two", "models", "do", "not", "depend", "on", "external", "language", "-", "specific", "features", "such", "as", "dependency", "parse", "trees", ".", "CNN", "extracts", "features", "from", "word", "embeddings", "of", "the", "input", "text", ",", "while", "BLSTM", "-", "2DPooling", "and", "BLSTM", "-", "2DCNN", "captures", "features", "from", "the", "output", "of", "BLSTM", "layer", ",", "which", "has", "already", "extracted", "features", "from", "the", "original", "input", "text", ".", "BLSTM", "-", "2DCNN", "is", "an", "extension", "of", "BLSTM", "-", "2DPooling", ",", "and", "the", "results", "show", "that", "BLSTM", "-", "2DCNN", "can", "capture", "more", "dependencies", "in", "text", ".", "AdaSent", "utilizes", "a", "more", "complicated", "model", "to", "form", "a", "hierarchy", "of", "representations", ",", "and", "it", "outperforms", "BLSTM", "-", "2DCNN", "on", "Subj", "and", "MR", "datasets", ".", "Compared", "with", "DSCNN", ",", "BLSTM", "-", "2DCNN", "outperforms", "it", "on", "five", "datasets", ".", "Compared", "with", "these", "results", ",", "2D", "convolution", "and", "2D", "max", "pooling", "operation", "are", "more", "effective", "for", "modeling", "sentence", ",", "even", "document", ".", "To", "better", "understand", "the", "effect", "of", "2D", "operations", ",", "this", "work", "conducts", "a", "sensitivity", "analysis", "on", "SST", "-", "1", "dataset", ".", "subsection", ":", "Effect", "of", "Sentence", "Length", "Figure", "2", "depicts", "the", "performance", "of", "the", "four", "models", "on", "different", "length", "of", "sentences", ".", "In", "the", "figure", ",", "the", "x", "-", "axis", "represents", "sentence", "lengths", "and", "y", "-", "axis", "is", "accuracy", ".", "The", "sentences", "collected", "in", "test", "set", "are", "no", "longer", "than", "45", "words", ".", "The", "accuracy", "here", "is", "the", "average", "value", "of", "the", "sentences", "with", "length", "in", "the", "window", ".", "Each", "data", "point", "is", "a", "mean", "score", "over", "5", "runs", ",", "and", "error", "bars", "have", "been", "omitted", "for", "clarity", ".", "It", "is", "found", "that", "both", "BLSTM", "-", "2DPooling", "and", "BLSTM", "-", "2DCNN", "outperform", "the", "other", "two", "models", ".", "This", "suggests", "that", "both", "2D", "convolution", "and", "2D", "max", "pooling", "operation", "are", "able", "to", "encode", "semantically", "-", "useful", "structural", "information", ".", "At", "the", "same", "time", ",", "it", "shows", "that", "the", "accuracies", "decline", "with", "the", "length", "of", "sentences", "increasing", ".", "In", "future", "work", ",", "we", "would", "like", "to", "investigate", "neural", "mechanisms", "to", "preserve", "long", "-", "term", "dependencies", "of", "text", ".", "subsection", ":", "Effect", "of", "2D", "Convolutional", "Filter", "and", "2D", "Max", "Pooling", "Size", "We", "are", "interested", "in", "what", "is", "the", "best", "2D", "filter", "and", "max", "pooling", "size", "to", "get", "better", "performance", ".", "We", "conduct", "experiments", "on", "SST", "-", "1", "dataset", "with", "BLSTM", "-", "2DCNN", "and", "set", "the", "number", "of", "feature", "maps", "to", "100", ".", "To", "make", "it", "simple", ",", "we", "set", "these", "two", "dimensions", "to", "the", "same", "values", ",", "thus", "both", "the", "filter", "and", "the", "pooling", "are", "square", "matrices", ".", "For", "the", "horizontal", "axis", ",", "c", "means", "2D", "convolutional", "filter", "size", ",", "and", "the", "five", "different", "color", "bar", "charts", "on", "each", "c", "represent", "different", "2D", "max", "pooling", "size", "from", "2", "to", "6", ".", "Figure", "3", "shows", "that", "different", "size", "of", "filter", "and", "pooling", "can", "get", "different", "accuracies", ".", "The", "best", "accuracy", "is", "52.6", "with", "2D", "filter", "size", "(", "5", ",", "5", ")", "and", "2D", "max", "pooling", "size", "(", "5", ",", "5", ")", ",", "this", "shows", "that", "finer", "tuning", "can", "further", "improve", "the", "performance", "reported", "here", ".", "And", "if", "a", "larger", "filter", "is", "used", ",", "the", "convolution", "can", "detector", "more", "features", ",", "and", "the", "performance", "may", "be", "improved", ",", "too", ".", "However", ",", "the", "networks", "will", "take", "up", "more", "storage", "space", ",", "and", "consume", "more", "time", ".", "section", ":", "Conclusion", "This", "paper", "introduces", "two", "combination", "models", ",", "one", "is", "BLSTM", "-", "2DPooling", ",", "the", "other", "is", "BLSTM", "-", "2DCNN", ",", "which", "can", "be", "seen", "as", "an", "extension", "of", "BLSTM", "-", "2DPooling", ".", "Both", "models", "can", "hold", "not", "only", "the", "time", "-", "step", "dimension", "but", "also", "the", "feature", "vector", "dimension", "information", ".", "The", "experiments", "are", "conducted", "on", "six", "text", "classificaion", "tasks", ".", "The", "experiments", "results", "demonstrate", "that", "BLSTM", "-", "2DCNN", "not", "only", "outperforms", "RecNN", ",", "RNN", "and", "CNN", "models", ",", "but", "also", "works", "better", "than", "the", "BLSTM", "-", "2DPooling", "and", "DSCNN", ".", "Especially", ",", "BLSTM", "-", "2DCNN", "achieves", "highest", "accuracy", "on", "SST", "-", "1", "and", "SST", "-", "2", "datasets", ".", "To", "better", "understand", "the", "effective", "of", "the", "proposed", "two", "models", ",", "this", "work", "also", "conducts", "a", "sensitivity", "analysis", "on", "SST", "-", "1", "dataset", ".", "It", "is", "found", "that", "large", "filter", "can", "detector", "more", "features", ",", "and", "this", "may", "lead", "to", "performance", "improvement", ".", "section", ":", "Acknowledgements", "We", "thank", "anonymous", "reviewers", "for", "their", "constructive", "comments", ".", "This", "research", "was", "funded", "by", "the", "National", "High", "Technology", "Research", "and", "Development", "Program", "of", "China", "(", "No.2015AA015402", ")", ",", "and", "the", "National", "Natural", "Science", "Foundation", "of", "China", "(", "No", ".", "61602479", ")", ",", "and", "the", "Strategic", "Priority", "Research", "Program", "of", "the", "Chinese", "Academy", "of", "Sciences", "(", "Grant", "No", ".", "XDB02070005", ")", ".", "bibliography", ":", "References"], "entity": [[2, 4, ["Text", "Classification"], "Task"], [10, 15, ["Two", "-", "dimensional", "Max", "Pooling"], "Method"], [15, 18, ["Recurrent", "Neural", "Network"], "Method"], [19, 20, ["RNN"], "Method"], [30, 33, ["Natural", "Language", "Processsing"], "Task"], [34, 35, ["NLP"], "Task"], [39, 41, ["recurrent", "structure"], "Method"], [51, 52, ["RNN"], "Method"], [54, 58, ["distributed", "representations", "of", "words"], "Method"], [98, 107, ["one", "-", "dimensional", "(", "1D", ")", "max", "pooling", "operation"], "Method"], [108, 112, ["attention", "-", "based", "operation"], "Method"], [144, 147, ["1D", "pooling", "operation"], "Method"], [160, 162, ["feature", "representation"], "Method"], [175, 177, ["pooling", "operation"], "Method"], [187, 190, ["sequence", "modeling", "tasks"], "Task"], [206, 210, ["2D", "max", "pooling", "operation"], "Method"], [213, 217, ["fixed", "-", "length", "representation"], "Method"], [225, 227, ["2D", "convolution"], "Method"], [241, 244, ["text", "classification", "tasks"], "Task"], [246, 248, ["sentiment", "analysis"], "Task"], [249, 251, ["question", "classification"], "Task"], [252, 254, ["subjectivity", "classification"], "Task"], [255, 257, ["newsgroup", "classification"], "Task"], [292, 293, ["accuracy"], "Metric"], [300, 305, ["fine", "-", "grained", "classification", "tasks"], "Task"], [325, 327, ["Text", "classification"], "Task"], [338, 340, ["sentiment", "analysis"], "Task"], [341, 343, ["relation", "extraction"], "Task"], [344, 346, ["spam", "detection"], "Task"], [373, 382, ["bag", "-", "of", "-", "words", "(", "BoW", ")", "model"], "Method"], [408, 412, ["order", "-", "sensitive", "models"], "Method"], [414, 416, ["neural", "networks"], "Method"], [421, 423, ["text", "classification"], "Task"], [431, 433, ["BoW", "models"], "Method"], [437, 439, ["textual", "modeling"], "Task"], [460, 462, ["recurrent", "structure"], "Method"], [463, 464, ["RNN"], "Method"], [470, 472, ["neural", "networks"], "Method"], [484, 485, ["RNN"], "Method"], [534, 537, ["learning", "feature", "representation"], "Task"], [539, 540, ["RNN"], "Method"], [541, 545, ["1D", "max", "pooling", "operation"], "Method"], [546, 550, ["attention", "-", "based", "operation"], "Method"], [558, 560, ["weighted", "representation"], "Method"], [595, 597, ["sentence", "representation"], "Task"], [602, 605, ["1D", "max", "pooling"], "Method"], [606, 610, ["attention", "-", "based", "operators"], "Method"], [623, 625, ["1D", "convolution"], "Method"], [628, 630, ["feature", "mapping"], "Task"], [634, 638, ["1D", "max", "pooling", "operation"], "Method"], [660, 661, ["RNN"], "Method"], [666, 667, ["RNN"], "Method"], [688, 689, ["NLP"], "Task"], [690, 691, ["CNN"], "Method"], [692, 695, ["image", "processing", "tasks"], "Task"], [696, 698, ["2D", "convolution"], "Method"], [699, 702, ["2D", "pooling", "operation"], "Method"], [717, 719, ["2D", "convolution"], "Method"], [720, 722, ["2D", "pooling"], "Method"], [740, 742, ["text", "classification"], "Task"], [749, 756, ["Bidirectional", "Long", "Short", "-", "Term", "Memory", "Networks"], "Method"], [757, 762, ["Two", "-", "Dimensional", "Max", "Pooling"], "Method"], [763, 766, ["BLSTM", "-", "2DPooling"], "Method"], [786, 793, ["Bidirectional", "Long", "Short", "-", "Term", "Memory", "Networks"], "Method"], [794, 795, ["BLSTM"], "Method"], [805, 809, ["2D", "max", "pooling", "operation"], "Method"], [823, 825, ["2D", "convolution"], "Method"], [826, 829, ["BLSTM", "-", "2DCNN"], "Method"], [861, 862, ["BLSTM"], "Method"], [874, 876, ["2D", "convolution"], "Method"], [877, 881, ["2D", "max", "pooling", "operation"], "Method"], [882, 885, ["sequence", "modeling", "tasks"], "Task"], [901, 903, ["2D", "convolution"], "Method"], [904, 908, ["2D", "max", "pooling", "operation"], "Method"], [918, 921, ["BLSTM", "-", "2DPooling"], "Method"], [922, 925, ["BLSTM", "-", "2DCNN"], "Method"], [931, 934, ["text", "classification", "tasks"], "Task"], [936, 938, ["sentiment", "analysis"], "Task"], [939, 941, ["question", "classification"], "Task"], [942, 944, ["subjectivity", "classification"], "Task"], [946, 948, ["newsgroups", "classification"], "Task"], [961, 964, ["BLSTM", "-", "2DCNN"], "Method"], [977, 978, ["accuracy"], "Metric"], [985, 990, ["fine", "-", "grained", "classification", "tasks"], "Task"], [997, 999, ["2D", "convolution"], "Method"], [1000, 1004, ["2D", "max", "pooling", "operation"], "Method"], [1037, 1039, ["sensitivity", "analysis"], "Method"], [1040, 1042, ["2D", "filter"], "Method"], [1043, 1046, ["max", "pooling", "size"], "Metric"], [1065, 1067, ["text", "classification"], "Task"], [1079, 1081, ["text", "classification"], "Task"], [1115, 1121, ["Deep", "learning", "based", "neural", "network", "models"], "Method"], [1126, 1129, ["text", "classification", "tasks"], "Task"], [1136, 1138, ["projection", "layer"], "Method"], [1153, 1155, ["neural", "networks"], "Method"], [1158, 1162, ["fixed", "-", "length", "representation"], "Method"], [1175, 1178, ["Recursive", "Neural", "Networks"], "Method"], [1182, 1183, ["RNN"], "Method"], [1184, 1185, ["CNN"], "Method"], [1187, 1189, ["neural", "networks"], "Method"], [1190, 1193, ["Recursive", "Neural", "Networks"], "Method"], [1194, 1195, ["RecNN"], "Method"], [1206, 1208, ["recursive", "models"], "Method"], [1232, 1236, ["recursive", "neural", "tensor", "network"], "Method"], [1255, 1259, ["deep", "recursive", "neural", "network"], "Method"], [1266, 1268, ["recursive", "layers"], "Method"], [1278, 1281, ["Recurrent", "Neural", "Networks"], "Method"], [1282, 1283, ["RNN"], "Method"], [1301, 1309, ["target", "dependent", "Long", "Short", "-", "Term", "Memory", "Networks"], "Method"], [1310, 1312, ["LSTM", ")"], "Method"], [1323, 1325, ["generalized", "LSTM"], "Method"], [1326, 1329, ["Tree", "-", "LSTM"], "Method"], [1331, 1333, ["LSTM", "unit"], "Method"], [1342, 1343, ["BLSTM"], "Method"], [1344, 1346, ["attention", "mechanism"], "Method"], [1356, 1357, ["classification"], "Task"], [1361, 1363, ["hierarchical", "network"], "Method"], [1367, 1369, ["attention", "mechanisms"], "Method"], [1379, 1381, ["document", "classification"], "Task"], [1398, 1399, ["zhou2016attention"], "Method"], [1400, 1403, ["Convolution", "Neural", "Networks"], "Method"], [1404, 1405, ["CNN"], "Method"], [1407, 1410, ["feedforward", "neural", "network"], "Method"], [1411, 1414, ["2D", "convolution", "layers"], "Method"], [1415, 1418, ["2D", "pooling", "layers"], "Method"], [1422, 1424, ["image", "processing"], "Task"], [1426, 1427, ["CNN"], "Method"], [1435, 1437, ["sentence", "classification"], "Task"], [1439, 1441, ["relation", "classification"], "Task"], [1448, 1449, ["CNN"], "Method"], [1456, 1459, ["1D", "convolution", "layers"], "Method"], [1460, 1463, ["1D", "pooling", "layers"], "Method"], [1476, 1481, ["dynamic", "-", "max", "pooling", "mechanism"], "Method"], [1482, 1484, ["sentence", "modeling"], "Task"], [1487, 1489, ["sensitivity", "analysis"], "Method"], [1499, 1501, ["architecture", "components"], "Method"], [1507, 1509, ["multichannel", "embeddings"], "Method"], [1510, 1512, ["unsupervised", "pretraining"], "Method"], [1514, 1516, ["classification", "accuracy"], "Metric"], [1519, 1521, ["sensitivity", "analysis"], "Method"], [1531, 1533, ["architecture", "components"], "Method"], [1543, 1546, ["1D", "convolutional", "filter"], "Method"], [1584, 1585, ["convolution"], "Method"], [1588, 1590, ["1D", "convolution"], "Method"], [1605, 1606, ["CNN"], "Method"], [1608, 1609, ["convolution"], "Method"], [1615, 1617, ["2D", "convolution"], "Method"], [1630, 1632, ["Neural", "Networks"], "Method"], [1644, 1646, ["neural", "networks"], "Method"], [1650, 1652, ["text", "classification"], "Task"], [1656, 1659, ["deep", "averaging", "network"], "Method"], [1663, 1668, ["unweighted", "average", "of", "word", "embeddings"], "Method"], [1670, 1672, ["hidden", "layers"], "Method"], [1673, 1674, ["classification"], "Task"], [1677, 1678, ["CNN"], "Method"], [1696, 1697, ["LSTM"], "Method"], [1700, 1702, ["sentence", "representation"], "Method"], [1706, 1709, ["BLSTM", "-", "2DCNN"], "Method"], [1713, 1714, ["DSCNN"], "Method"], [1715, 1716, ["RCNN"], "Method"], [1725, 1726, ["LSTM"], "Method"], [1735, 1736, ["BLSTM"], "Method"], [1752, 1754, ["1D", "convolution"], "Method"], [1755, 1759, ["1D", "max", "pooling", "operation"], "Method"], [1764, 1766, ["2D", "convolution"], "Method"], [1767, 1771, ["2D", "max", "pooling", "operation"], "Method"], [1776, 1778, ["sentence", "representation"], "Method"], [1796, 1798, ["BLSTM", "Layer"], "Method"], [1799, 1804, ["Two", "-", "dimensional", "Convolution", "Layer"], "Method"], [1806, 1810, ["dimensional", "max", "pooling", "Layer"], "Method"], [1812, 1814, ["Output", "Layer"], "Method"], [1829, 1831, ["BLSTM", "Layer"], "Method"], [1831, 1832, ["LSTM"], "Method"], [1853, 1856, ["adaptive", "gating", "mechanism"], "Method"], [1889, 1890, ["LSTM"], "Method"], [1950, 1953, ["logistic", "sigmoid", "function"], "Method"], [1955, 1959, ["element", "-", "wise", "multiplication"], "Method"], [1962, 1965, ["sequence", "modeling", "tasks"], "Task"], [1985, 1986, ["BLSTM"], "Method"], [2031, 2032, ["BLSTM"], "Method"], [2096, 2099, ["Convolutional", "Neural", "Networks"], "Method"], [2100, 2101, ["BLSTM"], "Method"], [2145, 2147, ["2D", "convolution"], "Method"], [2148, 2152, ["2D", "max", "pooling", "operation"], "Method"], [2163, 2168, ["Two", "-", "dimensional", "Convolution", "Layer"], "Method"], [2175, 2177, ["BLSTM", "Layer"], "Method"], [2187, 2189, ["narrow", "convolution"], "Method"], [2198, 2200, ["convolution", "operation"], "Method"], [2202, 2204, ["2D", "filter"], "Method"], [2290, 2292, ["convolution", "filter"], "Method"], [2294, 2296, ["convolution", "layer"], "Method"], [2321, 2327, ["Two", "-", "dimensional", "Max", "Pooling", "Layer"], "Method"], [2328, 2332, ["2D", "max", "pooling", "operation"], "Method"], [2343, 2346, ["2D", "max", "pooling"], "Method"], [2366, 2370, ["2D", "max", "pooling", "function"], "Method"], [2396, 2398, ["text", "classification"], "Task"], [2402, 2406, ["2D", "Max", "Pooling", "Layer"], "Method"], [2422, 2425, ["softmax", "classifier", "layer"], "Method"], [2439, 2440, ["classifier"], "Method"], [2449, 2451, ["training", "objective"], "Metric"], [2456, 2461, ["categorical", "cross", "-", "entropy", "loss"], "Metric"], [2463, 2464, ["loss"], "Metric"], [2468, 2470, ["regularized", "sum"], "Method"], [2507, 2508, ["Training"], "Task"], [2511, 2514, ["stochastic", "gradient", "descent"], "Method"], [2521, 2524, ["AdaDelta", "update", "rule"], "Method"], [2602, 2605, ["Stanford", "Sentiment", "Treebank"], "Material"], [2609, 2610, ["MR"], "Material"], [2639, 2642, ["SST", "-", "2"], "Material"], [2645, 2648, ["SST", "-", "1"], "Material"], [2716, 2718, ["Subjectivity", "dataset"], "Material"], [2743, 2746, ["Question", "classification", "dataset"], "Material"], [2782, 2784, ["20Ng", "dataset"], "Material"], [2793, 2796, ["bydate", "version", "preprocessed"], "Method"], [2819, 2821, ["Word", "Embeddings"], "Method"], [2822, 2824, ["word", "embeddings"], "Method"], [2836, 2837, ["generalization"], "Task"], [2851, 2853, ["GloVe", "embeddings"], "Method"], [2861, 2863, ["Wikipedia", "2014"], "Material"], [2864, 2866, ["Gigaword", "5"], "Material"], [2902, 2903, ["classification"], "Task"], [2930, 2932, ["evaluation", "metric"], "Metric"], [2937, 2941, ["Macro", "-", "F1", "measure"], "Metric"], [2958, 2959, ["accuracy"], "Metric"], [2984, 2985, ["LSTM"], "Method"], [2991, 2993, ["convolutional", "filters"], "Method"], [3025, 3027, ["learning", "rate"], "Metric"], [3028, 3029, ["AdaDelta"], "Method"], [3036, 3037, ["regularization"], "Task"], [3040, 3042, ["Dropout", "operation"], "Method"], [3043, 3045, ["dropout", "rate"], "Metric"], [3055, 3057, ["BLSTM", "layer"], "Method"], [3081, 3083, ["grid", "search"], "Method"], [3118, 3120, ["wide", "convolution"], "Method"], [3140, 3141, ["BLSTM"], "Method"], [3142, 3145, ["BLSTM", "-", "Att"], "Method"], [3146, 3149, ["BLSTM", "-", "2DPooling"], "Method"], [3151, 3154, ["BLSTM", "-", "2DCNN"], "Method"], [3176, 3178, ["classification", "tasks"], "Task"], [3200, 3201, ["accuracies"], "Metric"], [3202, 3205, ["SST", "-", "1"], "Material"], [3206, 3209, ["SST", "-", "2"], "Material"], [3211, 3214, ["BLSTM", "-", "2DPooling"], "Method"], [3236, 3238, ["2D", "convolution"], "Method"], [3256, 3259, ["SST", "-", "1"], "Material"], [3260, 3263, ["SST", "-", "2"], "Material"], [3265, 3267, ["TREC", "datasets"], "Material"], [3270, 3271, ["Subj"], "Material"], [3272, 3274, ["MR", "datasets"], "Material"], [3275, 3278, ["BLSTM", "-", "2DCNN"], "Method"], [3282, 3283, ["accuracies"], "Metric"], [3348, 3353, ["document", "-", "level", "dataset", "20Ng"], "Material"], [3365, 3366, ["RCNN"], "Method"], [3367, 3370, ["BLSTM", "-", "2DCNN"], "Method"], [3382, 3383, ["ReNN"], "Method"], [3384, 3385, ["RNN"], "Method"], [3386, 3387, ["CNN"], "Method"], [3389, 3391, ["neural", "networks"], "Method"], [3394, 3395, ["ReNN"], "Method"], [3415, 3416, ["CNN"], "Method"], [3427, 3430, ["BLSTM", "-", "2DPooling"], "Method"], [3431, 3434, ["BLSTM", "-", "2DCNN"], "Method"], [3440, 3442, ["BLSTM", "layer"], "Method"], [3454, 3457, ["BLSTM", "-", "2DCNN"], "Method"], [3461, 3464, ["BLSTM", "-", "2DPooling"], "Method"], [3470, 3473, ["BLSTM", "-", "2DCNN"], "Method"], [3480, 3481, ["AdaSent"], "Method"], [3489, 3492, ["hierarchy", "of", "representations"], "Method"], [3496, 3499, ["BLSTM", "-", "2DCNN"], "Method"], [3500, 3501, ["Subj"], "Material"], [3502, 3504, ["MR", "datasets"], "Material"], [3507, 3508, ["DSCNN"], "Method"], [3509, 3512, ["BLSTM", "-", "2DCNN"], "Method"], [3523, 3525, ["2D", "convolution"], "Method"], [3526, 3530, ["2D", "max", "pooling", "operation"], "Method"], [3553, 3555, ["sensitivity", "analysis"], "Method"], [3598, 3599, ["accuracy"], "Metric"], [3614, 3615, ["accuracy"], "Metric"], [3654, 3657, ["BLSTM", "-", "2DPooling"], "Method"], [3658, 3661, ["BLSTM", "-", "2DCNN"], "Method"], [3671, 3673, ["2D", "convolution"], "Method"], [3674, 3678, ["2D", "max", "pooling", "operation"], "Method"], [3697, 3698, ["accuracies"], "Metric"], [3715, 3717, ["neural", "mechanisms"], "Method"], [3730, 3733, ["2D", "Convolutional", "Filter"], "Method"], [3734, 3738, ["2D", "Max", "Pooling", "Size"], "Method"], [3746, 3748, ["2D", "filter"], "Method"], [3766, 3769, ["BLSTM", "-", "2DCNN"], "Method"], [3797, 3798, ["filter"], "Method"], [3800, 3801, ["pooling"], "Method"], [3802, 3804, ["square", "matrices"], "Method"], [3851, 3852, ["accuracies"], "Metric"], [3855, 3856, ["accuracy"], "Metric"], [3895, 3896, ["filter"], "Method"], [3900, 3901, ["convolution"], "Method"], [3938, 3940, ["combination", "models"], "Method"], [3943, 3946, ["BLSTM", "-", "2DPooling"], "Method"], [3950, 3953, ["BLSTM", "-", "2DCNN"], "Method"], [3962, 3965, ["BLSTM", "-", "2DPooling"], "Method"], [3991, 3994, ["text", "classificaion", "tasks"], "Task"], [4000, 4003, ["BLSTM", "-", "2DCNN"], "Method"], [4006, 4007, ["RecNN"], "Method"], [4008, 4009, ["RNN"], "Method"], [4010, 4011, ["CNN"], "Method"], [4019, 4022, ["BLSTM", "-", "2DPooling"], "Method"], [4023, 4024, ["DSCNN"], "Method"], [4027, 4030, ["BLSTM", "-", "2DCNN"], "Method"], [4032, 4033, ["accuracy"], "Metric"], [4034, 4037, ["SST", "-", "1"], "Material"], [4038, 4042, ["SST", "-", "2", "datasets"], "Material"], [4059, 4061, ["sensitivity", "analysis"], "Task"], [4071, 4073, ["large", "filter"], "Method"], [8, 9, ["LSTM"], "Method"], [294, 297, ["Stanford", "Sentiment", "Treebank"], "Material"], [333, 334, ["NLP"], "Task"], [620, 621, ["CNN"], "Method"], [909, 910, ["NLP"], "Task"], [979, 982, ["Stanford", "Sentiment", "Treebank"], "Material"], [1010, 1013, ["Stanford", "Sentiment", "Treebank"], "Material"], [1074, 1075, ["BLSTM"], "Method"], [1179, 1180, ["RecNN"], "Method"], [1391, 1392, ["BLSTM"], "Method"], [1430, 1431, ["NLP"], "Task"], [1450, 1451, ["NLP"], "Task"], [1467, 1468, ["CNN"], "Method"], [1493, 1494, ["CNN"], "Method"], [1525, 1526, ["CNN"], "Method"], [1547, 1548, ["NLP"], "Task"], [1728, 1729, ["RNN"], "Method"], [1844, 1845, ["RNN"], "Method"], [1990, 1991, ["LSTM"], "Method"], [3085, 3088, ["SST", "-", "1"], "Material"], [3113, 3114, ["LSTM"], "Method"], [3180, 3181, ["BLSTM"], "Method"], [3249, 3250, ["BLSTM"], "Method"], [3251, 3252, ["CNN"], "Method"], [3556, 3559, ["SST", "-", "1"], "Material"], [3761, 3764, ["SST", "-", "1"], "Material"], [4062, 4065, ["SST", "-", "1"], "Material"]]} {"doc_id": "33bcc97b605f00145098d095be2841a1fa6b9a95", "text": ["document", ":", "Order", "-", "Embeddings", "of", "Images", "and", "Language", "Hypernymy", ",", "textual", "entailment", ",", "and", "image", "captioning", "can", "be", "seen", "as", "special", "cases", "of", "a", "single", "visual", "-", "semantic", "hierarchy", "over", "words", ",", "sentences", ",", "and", "images", ".", "In", "this", "paper", "we", "advocate", "for", "explicitly", "modeling", "the", "partial", "order", "structure", "of", "this", "hierarchy", ".", "Towards", "this", "goal", ",", "we", "introduce", "a", "general", "method", "for", "learning", "ordered", "representations", ",", "and", "show", "how", "it", "can", "be", "applied", "to", "a", "variety", "of", "tasks", "involving", "images", "and", "language", ".", "We", "show", "that", "the", "resulting", "representations", "improve", "performance", "over", "current", "approaches", "for", "hypernym", "prediction", "and", "image", "-", "caption", "retrieval", ".", "section", ":", "Introduction", "Computer", "vision", "and", "natural", "language", "processing", "are", "becoming", "increasingly", "intertwined", ".", "Recent", "work", "in", "vision", "has", "moved", "beyond", "discriminating", "between", "a", "fixed", "set", "of", "object", "classes", ",", "to", "automatically", "generating", "open", "-", "ended", "lingual", "descriptions", "of", "images", "vinyals2015show", ".", "Recent", "methods", "for", "natural", "language", "processing", "such", "as", "flickr30k", "learn", "the", "semantics", "of", "language", "by", "grounding", "it", "in", "the", "visual", "world", ".", "Looking", "to", "the", "future", ",", "autonomous", "artificial", "agents", "will", "need", "to", "jointly", "model", "vision", "and", "language", "in", "order", "to", "parse", "the", "visual", "world", "and", "communicate", "with", "people", ".", "But", "what", ",", "precisely", ",", "is", "the", "relationship", "between", "images", "and", "the", "words", "or", "captions", "we", "use", "to", "describe", "them", "?", "It", "is", "akin", "to", "the", "hypernym", "relation", "between", "words", ",", "and", "textual", "entailment", "among", "phrases", ":", "captions", "are", "simply", "abstractions", "of", "images", ".", "In", "fact", ",", "all", "three", "relations", "can", "be", "seen", "as", "special", "cases", "of", "a", "partial", "order", "over", "images", "and", "language", ",", "illustrated", "in", "Figure", "[", "reference", "]", ",", "which", "we", "refer", "to", "as", "the", "visual", "-", "semantic", "hierarchy", ".", "As", "a", "partial", "order", ",", "this", "relation", "is", "transitive", ":", "“", "woman", "walking", "her", "dog", "”", ",", "“", "woman", "walking", "”", ",", "“", "person", "walking", "”", ",", "“", "person", "”", ",", "and", "“", "entity", "”", "are", "all", "valid", "abstractions", "of", "the", "rightmost", "image", ".", "Our", "goal", "in", "this", "work", "is", "to", "learn", "representations", "that", "respect", "this", "partial", "order", "structure", ".", "Most", "recent", "approaches", "to", "modeling", "the", "hypernym", ",", "entailment", ",", "and", "image", "-", "caption", "relations", "involve", "learning", "distributed", "representations", "or", "embeddings", ".", "This", "is", "a", "very", "powerful", "and", "general", "approach", "which", "maps", "the", "objects", "of", "interest", "—", "words", ",", "phrases", ",", "images—", "to", "points", "in", "a", "high", "-", "dimensional", "vector", "space", ".", "One", "line", "of", "work", ",", "exemplified", "by", "chopra2005learning", "and", "first", "applied", "to", "the", "caption", "-", "image", "relationship", "by", "socher2014grounded", ",", "requires", "the", "mapping", "to", "be", "distance", "-", "preserving", ":", "semantically", "similar", "objects", "are", "mapped", "to", "points", "that", "are", "nearby", "in", "the", "embedding", "space", ".", "A", "symmetric", "distance", "measure", "such", "as", "Euclidean", "or", "cosine", "distance", "is", "typically", "used", ".", "Since", "the", "visual", "-", "semantic", "hierarchy", "is", "an", "anti", "symmetric", "relation", ",", "we", "expect", "this", "approach", "to", "introduce", "systematic", "model", "error", ".", "Other", "approaches", "do", "not", "have", "such", "explicit", "constraints", ",", "learning", "a", "more", "-", "or", "-", "less", "general", "binary", "relation", "between", "the", "objects", "of", "interest", ",", "e.g.", "bordes2011learning", ",", "socher2013reasoning", ",", "ma2015multimodal", ".", "Notably", ",", "no", "existing", "approach", "directly", "imposes", "the", "transitivity", "and", "antisymmetry", "of", "the", "partial", "order", ",", "leaving", "the", "model", "to", "induce", "these", "properties", "from", "data", ".", "In", "contrast", ",", "we", "propose", "to", "exploit", "the", "partial", "order", "structure", "of", "the", "visual", "-", "semantic", "hierarchy", "by", "learning", "a", "mapping", "which", "is", "not", "distance", "-", "preserving", "but", "order", "-", "preserving", "between", "the", "visual", "-", "semantic", "hierarchy", "and", "a", "partial", "order", "over", "the", "embedding", "space", ".", "We", "call", "embeddings", "learned", "in", "this", "way", "order", "-", "embeddings", ".", "This", "idea", "can", "be", "integrated", "into", "existing", "relational", "learning", "methods", "simply", "by", "replacing", "their", "comparison", "operation", "with", "ours", ".", "By", "modifying", "existing", "methods", "in", "this", "way", ",", "we", "find", "that", "order", "-", "embeddings", "provide", "a", "marked", "improvement", "over", "the", "state", "-", "of", "-", "art", "for", "hypernymy", "prediction", "and", "caption", "-", "image", "retrieval", ",", "and", "near", "state", "-", "of", "-", "the", "-", "art", "performance", "for", "natural", "language", "inference", ".", "This", "paper", "is", "structured", "as", "follows", ".", "We", "begin", ",", "in", "Section", "[", "reference", "]", ",", "by", "giving", "a", "unified", "mathematical", "treatment", "of", "our", "tasks", ",", "and", "describing", "the", "general", "approach", "of", "learning", "order", "-", "embeddings", ".", "In", "the", "next", "three", "sections", "we", "describe", "in", "detail", "the", "tasks", "we", "tackle", ",", "how", "we", "apply", "the", "order", "-", "embeddings", "idea", "to", "each", "of", "them", ",", "and", "the", "results", "we", "obtain", ".", "The", "tasks", "are", "hypernym", "prediction", "(", "Section", "[", "reference", "]", ")", ",", "caption", "-", "image", "retrieval", "(", "Section", "[", "reference", "]", ")", ",", "and", "textual", "entailment", "(", "Section", "[", "reference", "]", ")", ".", "In", "the", "supplementary", "material", ",", "we", "visualize", "novel", "vector", "regularities", "that", "emerge", "in", "our", "learned", "embeddings", "of", "images", "and", "language", ".", "section", ":", "Learning", "Order", "-", "Embeddings", "To", "unify", "our", "treatment", "of", "various", "tasks", ",", "we", "introduce", "the", "problem", "of", "partial", "order", "completion", ".", "In", "partial", "order", "completion", ",", "we", "are", "given", "a", "set", "of", "positive", "examples", "of", "ordered", "pairs", "drawn", "from", "a", "partially", "ordered", "set", ",", "and", "a", "set", "of", "negative", "examples", "which", "we", "know", "to", "be", "unordered", ".", "Our", "goal", "is", "to", "predict", "whether", "an", "unseen", "pair", "is", "ordered", ".", "Note", "that", "hypernym", "prediction", ",", "caption", "-", "image", "retrieval", ",", "and", "textual", "entailment", "are", "all", "special", "cases", "of", "this", "task", ",", "since", "they", "all", "involve", "classifying", "pairs", "of", "concepts", "in", "the", "(", "partially", "ordered", ")", "visual", "-", "semantic", "hierarchy", ".", "We", "tackle", "this", "problem", "by", "learning", "a", "mapping", "from", "into", "a", "partially", "ordered", "embedding", "space", ".", "The", "idea", "is", "to", "predict", "the", "ordering", "of", "an", "unseen", "pair", "in", "based", "on", "its", "ordering", "in", "the", "embedding", "space", ".", "This", "is", "possible", "only", "if", "the", "mapping", "satisfies", "the", "following", "crucial", "property", ":", "theorem", ":", ".", "A", "function", ":", "f→", "(", "X", ",", "⪯X", ")(", "Y", ",", "⪯Y", ")", "is", "an", "order", "-", "embedding", "if", "for", "all", "∈u", ",", "vX", ",", "This", "definition", "implies", "that", "each", "combination", "of", "embedding", "space", ",", "order", ",", "and", "order", "-", "embedding", "determines", "a", "unique", "completion", "of", "our", "data", "as", "a", "partial", "order", ".", "In", "the", "following", ",", "we", "first", "consider", "the", "choice", "of", "and", ",", "and", "then", "discuss", "how", "to", "find", "an", "appropriate", ".", "subsection", ":", "The", "Reversed", "Product", "Order", "on", "The", "choice", "of", "and", "is", "somewhat", "application", "-", "dependent", ".", "For", "the", "purpose", "of", "modeling", "the", "semantic", "hierarchy", ",", "our", "choices", "are", "narrowed", "by", "the", "following", "considerations", ".", "Much", "of", "the", "expressive", "power", "of", "human", "language", "comes", "from", "abstraction", "and", "composition", ".", "For", "any", "two", "concepts", ",", "say", "“", "dog", "”", "and", "“", "cat", "”", ",", "we", "can", "name", "a", "concept", "that", "is", "an", "abstraction", "of", "the", "two", ",", "such", "as", "“", "mammal", "”", ",", "as", "well", "as", "a", "concept", "that", "composes", "the", "two", ",", "such", "as", "“", "dog", "chasing", "cat", "”", ".", "So", ",", "in", "order", "to", "represent", "the", "visual", "-", "semantic", "hierarchy", ",", "we", "need", "to", "choose", "an", "order", "that", "is", "rich", "enough", "to", "embed", "these", "two", "relations", ".", "We", "also", "restrict", "ourselves", "to", "orders", "with", "a", "top", "element", ",", "which", "is", "above", "every", "other", "element", "in", "the", "order", ".", "In", "the", "visual", "-", "semantic", "hierarchy", ",", "this", "element", "represents", "the", "most", "general", "possible", "concept", ";", "practically", ",", "it", "provides", "an", "anchor", "for", "the", "embedding", ".", "Finally", ",", "we", "choose", "the", "embedding", "space", "to", "be", "continuous", "in", "order", "to", "allow", "optimization", "with", "gradient", "-", "based", "methods", ".", "A", "natural", "choice", "that", "satisfies", "all", "three", "properties", "is", "the", "reversed", "product", "order", "on", ",", "defined", "by", "the", "conjunction", "of", "total", "orders", "on", "each", "coordinate", ":", "for", "all", "vectors", "with", "nonnegative", "coordinates", ".", "Note", "the", "reversal", "of", "direction", ":", "smaller", "coordinates", "imply", "higher", "position", "in", "the", "partial", "order", ".", "The", "origin", "is", "then", "the", "top", "element", "of", "the", "order", ",", "representing", "the", "most", "general", "concept", ".", "Instead", "of", "viewing", "our", "embeddings", "as", "single", "points", ",", "we", "can", "also", "view", "them", "as", "sets", ".", "The", "meaning", "of", "a", "word", "is", "then", "the", "union", "of", "all", "concepts", "of", "which", "it", "is", "a", "hypernym", ",", "and", "the", "meaning", "of", "a", "sentence", "is", "the", "union", "of", "all", "sentences", "that", "entail", "it", ".", "The", "visual", "-", "semantic", "hierarchy", "can", "then", "be", "seen", "as", "a", "special", "case", "of", "the", "subset", "relation", ",", "a", "connection", "also", "used", "by", "flickr30k", ".", "subsection", ":", "Penalizing", "Order", "Violations", "Having", "fixed", "the", "embedding", "space", "and", "order", ",", "we", "now", "consider", "the", "problem", "of", "finding", "an", "order", "-", "embedding", "into", "this", "space", ".", "In", "practice", ",", "the", "order", "embedding", "condition", "(", "Definition", "[", "reference", "]", ")", "is", "too", "restrictive", "to", "impose", "as", "a", "hard", "constraint", ".", "Instead", ",", "we", "aim", "to", "find", "an", "approximate", "order", "-", "embedding", ":", "a", "mapping", "which", "violates", "the", "order", "-", "embedding", "condition", ",", "imposed", "as", "a", "soft", "constraint", ",", "as", "little", "as", "possible", ".", "More", "precisely", ",", "we", "define", "a", "penalty", "that", "measures", "the", "degree", "to", "which", "a", "pair", "of", "points", "violates", "the", "product", "order", ".", "In", "particular", ",", "we", "define", "the", "penalty", "for", "an", "ordered", "pair", "of", "points", "in", "as", "Crucially", ",", "according", "to", "the", "reversed", "product", "order", ";", "if", "the", "order", "is", "not", "satisfied", ",", "is", "positive", ".", "This", "effectively", "imposes", "a", "strong", "prior", "on", "the", "space", "of", "relations", ",", "encouraging", "our", "learned", "relation", "to", "satisfy", "the", "partial", "order", "properties", "of", "transitivity", "and", "antisymmetry", ".", "This", "penalty", "is", "key", "to", "our", "method", ".", "Throughout", "the", "remainder", "of", "the", "paper", ",", "we", "will", "use", "it", "where", "previous", "work", "has", "used", "symmetric", "distances", "or", "learned", "comparison", "operators", ".", "Recall", "that", "and", "are", "our", "positive", "and", "negative", "examples", ",", "respectively", ".", "Then", ",", "to", "learn", "an", "approximate", "order", "-", "embedding", ",", "we", "could", "use", "a", "max", "-", "margin", "loss", "which", "encourages", "positive", "examples", "to", "have", "zero", "penalty", ",", "and", "negative", "examples", "to", "have", "penalty", "greater", "than", "a", "margin", ":", "In", "practice", "we", "are", "often", "not", "given", "negative", "examples", ",", "in", "which", "case", "this", "loss", "admits", "the", "trivial", "solution", "of", "mapping", "all", "objects", "to", "the", "same", "point", ".", "The", "best", "way", "of", "dealing", "with", "this", "problem", "depends", "on", "the", "application", ",", "so", "we", "will", "describe", "task", "-", "specific", "variations", "of", "the", "loss", "in", "the", "next", "several", "sections", ".", "section", ":", "Hypernym", "Prediction", "To", "test", "the", "ability", "of", "our", "model", "to", "learn", "partial", "orders", "from", "incomplete", "data", ",", "our", "first", "task", "is", "to", "predict", "withheld", "hypernym", "pairs", "in", "WordNet", "miller1995wordnet", ".", "A", "hypernym", "pair", "is", "a", "pair", "of", "concepts", "where", "the", "first", "concept", "is", "a", "specialization", "or", "an", "instance", "of", "the", "second", ",", "e.g.", ",", "(", "woman", ",", "person", ")", "or", "(", "New", "York", ",", "city", ")", ".", "Our", "setup", "differs", "significantly", "from", "previous", "work", "in", "that", "we", "use", "only", "the", "WordNet", "hierarchy", "as", "training", "data", ".", "The", "most", "similar", "evaluation", "has", "been", "that", "of", "baroni2012entailment", ",", "who", "use", "external", "linguistic", "data", "in", "the", "form", "of", "distributional", "semantic", "vectors", ".", "bordes2011learning", "and", "socher2013reasoning", "also", "evaluate", "on", "the", "WordNet", "hierarchy", ",", "but", "they", "use", "other", "relations", "in", "WordNet", "as", "training", "data", "(", "and", "external", "linguistic", "data", ",", "in", "Socher", "’s", "case", ")", ".", "Additionally", ",", "the", "latter", "two", "consider", "only", "direct", "hypernyms", ",", "rather", "than", "the", "full", ",", "transitive", "hypernymy", "relation", ".", "But", "predicting", "the", "transitive", "hypernym", "relation", "is", "a", "better", "-", "defined", "problem", "because", "individual", "hypernym", "edges", "in", "WordNet", "vary", "dramatically", "in", "the", "degree", "of", "abstraction", "they", "require", ".", "For", "instance", ",", "(", "person", ",", "organism", ")", "is", "a", "direct", "hypernym", "pair", ",", "but", "it", "takes", "eight", "hypernym", "edges", "to", "get", "from", "cat", "to", "organism", ".", "subsection", ":", "Loss", "Function", "To", "apply", "order", "-", "embeddings", "to", "hypernymy", ",", "we", "follow", "the", "setup", "of", "socher2013reasoning", "in", "learning", "an", "N", "-", "dimensional", "vector", "for", "each", "concept", "in", "WordNet", ",", "but", "we", "replace", "their", "neural", "tensor", "network", "with", "our", "order", "-", "violation", "penalty", "defined", "in", "Eq", ".", "(", "[", "reference", "]", ")", ".", "Just", "like", "them", ",", "we", "corrupt", "each", "hypernym", "pair", "by", "replacing", "one", "of", "the", "two", "concepts", "with", "a", "randomly", "chosen", "concept", ",", "and", "use", "these", "corrupted", "pairs", "as", "negative", "examples", "for", "both", "training", "and", "evaluation", ".", "We", "use", "their", "max", "-", "margin", "loss", ",", "which", "encourages", "the", "order", "-", "violation", "penalty", "to", "be", "zero", "for", "positive", "examples", ",", "and", "greater", "than", "a", "margin", "for", "negative", "examples", ":", "where", "is", "our", "order", "-", "violation", "penalty", ",", "and", "is", "a", "corrupted", "version", "of", ".", "Since", "we", "learn", "an", "independent", "embedding", "for", "each", "concept", ",", "the", "mapping", "is", "simply", "a", "lookup", "table", ".", "subsection", ":", "Dataset", "The", "transitive", "closure", "of", "the", "WordNet", "hierarchy", "gives", "us", "edges", "between", "concepts", "in", "WordNet", ".", "Like", "bordes2011learning", ",", "we", "randomly", "select", "edges", "for", "the", "test", "split", ",", "and", "another", "for", "the", "development", "set", ".", "Note", "that", "the", "majority", "of", "test", "set", "edges", "can", "be", "inferred", "simply", "by", "applying", "transitivity", ",", "giving", "us", "a", "strong", "baseline", ".", "subsection", ":", "Details", "of", "Training", "We", "learn", "a", "50", "-", "dimensional", "nonnegative", "vector", "for", "each", "concept", "in", "WordNet", "using", "the", "max", "-", "margin", "objective", "(", "[", "reference", "]", ")", "with", "margin", ",", "sampling", "500", "true", "and", "500", "false", "hypernym", "pairs", "in", "each", "batch", ".", "We", "train", "for", "30", "-", "50", "epochs", "using", "the", "Adam", "optimizer", "adam", "with", "learning", "rate", "and", "early", "stopping", "on", "the", "validation", "set", ".", "During", "evaluation", ",", "we", "find", "the", "optimal", "classification", "threshold", "on", "the", "validation", "set", ",", "then", "apply", "it", "to", "the", "test", "set", ".", "subsection", ":", "Results", "Since", "our", "setup", "is", "novel", ",", "there", "are", "no", "published", "numbers", "to", "compare", "to", ".", "We", "therefore", "compare", "three", "variants", "of", "our", "model", "to", "two", "baselines", ",", "with", "results", "shown", "in", "Table", "[", "reference", "]", ".", "The", "transitive", "closure", "baseline", "involves", "no", "learning", ";", "it", "simply", "classifies", "hypernyms", "pairs", "as", "positive", "if", "they", "are", "in", "the", "transitive", "closure", "of", "the", "union", "of", "edges", "in", "the", "training", "and", "validation", "sets", ".", "The", "word2gauss", "baseline", "evaluates", "the", "approach", "of", "word2gauss", "to", "represent", "words", "as", "Gaussian", "densities", "rather", "than", "points", "in", "the", "embedding", "space", ".", "This", "allows", "a", "natural", "representation", "of", "hierarchies", "using", "the", "KL", "divergence", ".", "We", "used", "50", "-", "dimensional", "diagonal", "Gaussian", "embeddings", ",", "trained", "for", "200", "epochs", "on", "a", "max", "-", "margin", "objective", "with", "margin", ",", "chosen", "by", "grid", "search", ".", "order", "-", "embeddings", "(", "symmetric", ")", "is", "our", "full", "model", ",", "but", "using", "symmetric", "cosine", "distance", "instead", "of", "our", "asymmetric", "penalty", ".", "order", "-", "embeddings", "(", "bilinear", ")", "replaces", "our", "penalty", "with", "the", "bilinear", "model", "used", "by", "socher2013reasoning", ".", "order", "-", "embeddings", "is", "our", "full", "model", ".", "Only", "our", "full", "model", "can", "do", "better", "than", "the", "transitive", "baseline", ",", "showing", "the", "value", "of", "exploiting", "partial", "order", "structure", "in", "contrast", "to", "using", "symmetric", "similarity", "or", "learning", "a", "general", "binary", "relation", "as", "most", "previous", "work", "and", "our", "bilinear", "baseline", "do", ".", "The", "resulting", "50", "-", "dimensional", "embeddings", "are", "difficult", "to", "visualize", ".", "To", "give", "some", "intuition", "for", "the", "structure", "being", "learned", ",", "Figure", "[", "reference", "]", "shows", "the", "results", "of", "a", "toy", "2D", "experiment", ".", "section", ":", "Caption", "-", "Image", "Retrieval", "The", "caption", "-", "image", "retrieval", "task", "has", "become", "a", "standard", "evaluation", "of", "joint", "models", "of", "vision", "and", "language", "hodosh2013framing", ",", "LinCVPR14", ".", "The", "task", "involves", "ranking", "a", "large", "dataset", "of", "images", "by", "relevance", "for", "a", "query", "caption", "(", "Image", "Retrieval", ")", ",", "and", "ranking", "captions", "by", "relevance", "for", "a", "query", "image", "(", "Caption", "Retrieval", ")", ".", "Given", "a", "set", "of", "aligned", "image", "-", "caption", "pairs", "as", "training", "data", ",", "the", "goal", "is", "then", "to", "learn", "a", "caption", "-", "image", "compatibility", "score", "to", "be", "used", "at", "test", "time", ".", "Many", "modern", "approaches", "model", "the", "caption", "-", "image", "relationship", "symmetrically", ",", "either", "by", "embedding", "into", "a", "common", "“", "visual", "-", "semantic", "”", "space", "with", "inner", "-", "product", "similarity", "socher2014grounded", ",", "kiros2014", ",", "or", "by", "using", "Canonical", "Correlations", "Analysis", "between", "distributed", "representations", "of", "images", "and", "captions", "klein2015fisher", ".", "While", "karpathydeep", "and", "plummer2015flickr30k", "model", "a", "finer", "-", "grained", "alignment", "between", "regions", "in", "the", "image", "and", "segments", "of", "the", "caption", ",", "the", "similarity", "they", "use", "is", "still", "symmetric", ".", "An", "alternative", "is", "to", "learn", "an", "unconstrained", "binary", "relation", ",", "either", "with", "a", "neural", "language", "model", "conditioned", "on", "the", "image", "vinyals2015show", ",", "mao2015", "or", "using", "a", "multimodal", "CNN", "ma2015multimodal", ".", "In", "contrast", "to", "these", "lines", "of", "work", ",", "we", "propose", "to", "treat", "the", "caption", "-", "image", "pairs", "as", "a", "two", "-", "level", "partial", "order", "with", "captions", "above", "the", "images", "they", "describe", ",", "and", "let", "with", "our", "order", "-", "violation", "penalty", "defined", "in", "Eq", "(", "[", "reference", "]", ")", ",", "and", "are", "embedding", "functions", "from", "captions", "and", "images", "into", ".", "subsection", ":", "Loss", "Function", "To", "facilitate", "comparison", ",", "we", "use", "the", "same", "pairwise", "ranking", "loss", "that", "socher2014grounded", ",", "kiros2014", "and", "karpathydeep", "have", "used", "on", "this", "task", "—", "simply", "replacing", "their", "symmetric", "similarity", "measure", "with", "our", "asymmetric", "order", "-", "violation", "penalty", ".", "This", "loss", "function", "encourages", "for", "ground", "truth", "caption", "-", "image", "pairs", "to", "be", "greater", "than", "that", "for", "all", "other", "pairs", ",", "by", "a", "margin", ":", "where", "is", "a", "ground", "truth", "caption", "-", "image", "pair", ",", "goes", "over", "captions", "that", "no", "describe", ",", "and", "goes", "over", "image", "not", "described", "by", ".", "subsection", ":", "Image", "and", "Caption", "Embeddings", "To", "learn", "and", ",", "we", "use", "the", "approach", "of", "kiros2014", "except", ",", "since", "we", "are", "embedding", "into", ",", "we", "constrain", "the", "embedding", "vectors", "to", "have", "nonnegative", "entries", "by", "taking", "their", "absolute", "value", ".", "Thus", ",", "to", "embed", "images", ",", "we", "use", "where", "is", "a", "learned", "matrix", ",", "being", "the", "dimensionality", "of", "the", "embedding", "space", ".", "is", "the", "same", "image", "feature", "used", "by", "klein2015fisher", ":", "we", "rescale", "images", "to", "have", "smallest", "side", "pixels", ",", "we", "take", "crops", "from", "the", "corners", ",", "center", ",", "and", "their", "horizontal", "reflections", ",", "run", "the", "10", "crops", "through", "the", "19", "-", "layer", "VGG", "network", "of", "vgg", "(", "weights", "pre", "-", "trained", "on", "ImageNet", "and", "fixed", "during", "training", ")", ",", "and", "average", "their", "fc7", "features", ".", "To", "embed", "the", "captions", ",", "we", "use", "a", "recurrent", "neural", "net", "encoder", "with", "GRU", "activations", "gru", ",", "so", ",", "the", "absolute", "value", "of", "hidden", "state", "after", "processing", "the", "last", "word", ".", "subsection", ":", "Dataset", "We", "evaluate", "on", "the", "Microsoft", "COCO", "dataset", "coco", ",", "which", "has", "over", "120", ",", "000", "images", ",", "each", "with", "at", "least", "five", "human", "-", "annotated", "captions", "per", "image", ".", "This", "is", "by", "far", "the", "largest", "dataset", "commonly", "used", "for", "caption", "-", "image", "retrieval", ".", "We", "use", "the", "data", "splits", "of", "karpathydeep", "for", "training", "(", "113", ",", "287", "images", ")", ",", "validation", "(", "5000", "images", ")", ",", "and", "test", "(", "5000", "images", ")", ".", "subsection", ":", "Details", "of", "Training", "To", "train", "the", "model", ",", "we", "use", "the", "standard", "pairwise", "ranking", "objective", "from", "Eq", ".", "(", "[", "reference", "]", ")", ".", "We", "sample", "minibatches", "of", "128", "random", "image", "-", "caption", "pairs", ",", "and", "draw", "all", "contrastive", "terms", "from", "the", "minibatch", ",", "giving", "us", "127", "contrastive", "images", "for", "each", "caption", "and", "captions", "for", "each", "image", ".", "We", "train", "for", "15", "-", "30", "epochs", "using", "the", "Adam", "optimizer", "with", "learning", "rate", ",", "and", "early", "stopping", "on", "the", "validation", "set", ".", "We", "set", "the", "dimension", "of", "the", "embedding", "space", "and", "the", "GRU", "hidden", "state", "to", ",", "the", "dimension", "of", "the", "learned", "word", "embeddings", "to", ",", "and", "the", "margin", "to", ".", "All", "these", "hyperparameters", ",", "as", "well", "as", "the", "learning", "rate", "and", "batchsize", ",", "were", "selected", "using", "the", "validation", "set", ".", "For", "consistency", "with", "kiros2014", "and", "to", "mitigate", "overfitting", ",", "we", "constrain", "the", "caption", "and", "image", "embeddings", "to", "have", "unit", "L2", "norm", ".", "This", "constraint", "implies", "that", "no", "two", "points", "can", "be", "exactly", "ordered", "with", "zero", "order", "-", "violation", "penalty", ",", "but", "since", "we", "use", "a", "ranking", "loss", ",", "only", "the", "relative", "size", "of", "the", "penalties", "matters", ".", "subsection", ":", "Results", "L", "—", "CCCC", "—", "CCCC", "Align", "Caption", "Retrieval", "Align", "Image", "Retrieval", "Model", "Align", "R@1", "Align", "R@10", "Align", "Med", "r", "Align", "Mean", "rAlign", "R@1", "Align", "R@10", "Align", "Med", "r", "Align", "Mean", "rAlign", "Align", "1k", "Test", "Images", "Align", "Align", "Align", "Align", "Align", "Align", "Align", "MNLM", "Align", "43.4", "Align", "85.8", "Align", "2", "Align", "*", "Align", "31.0", "Align", "79.9", "Align", "3", "Align", "*", "m", "-", "RNN", "Align41.0", "Align", "83.5", "Align", "2", "Align", "*", "Align", "29.0", "Align", "77.0", "Align", "3", "Align", "*", "DVSA", "Align", "38.4", "Align", "80.5", "Align", "1", "Align", "*", "Align", "27.4", "Align", "74.8", "Align", "3", "Align", "*", "STV", "Align", "33.8", "Align", "82.1", "Align", "3", "Align", "*", "Align", "25.9", "Align", "74.6", "Align", "4", "Align", "*", "FV", "Align", "39.4Align", "80.9", "Align", "2", "Align", "10.4", "Align", "25.1", "Align", "76.6", "Align", "4", "Align", "11.1", "m", "-", "CNN", "Align", "38.3", "Align", "81.0", "Align", "2", "Align", "*", "Align", "27.4", "Align", "79.5", "Align", "3", "Align", "*", "m", "-", "CNN⁢ENS", "Align", "42.8", "Align", "84.1", "Align", "2", "Align", "*", "Align", "32.6", "Align", "82.8", "Align", "3", "Align", "*", "Align", "Align", "Align", "Align", "Align", "Align", "Align", "order", "-", "embeddings", "(", "reversed", ")", "Align", "11.2", "Align", "44.0", "Align", "14.2", "Align", "86.6", "Align", "12.3", "Align", "53.5", "Align", "9.0", "Align", "30.1order", "-", "embeddings", "(", "1", "-", "crop", ")", "Align", "41.4", "Align", "84.2", "Align", "2.0", "Align", "8.7", "Align", "33.5", "Align", "82.2", "Align", "2.6Align", "10.0", "order", "-", "embeddings", "(", "symm", ".", ")", "Align", "45.4", "Align", "88.7", "Align", "2.0", "Align", "5.8", "Align", "36.3", "Align", "85.8", "Align", "2.0", "Align", "9.0", "order", "-", "embeddings", "Align", "46.7", "Align", "88.9", "Align", "2.0", "Align", "5.7", "Align", "37.9", "Align", "85.9", "Align", "2.0", "Align", "8.1", "5k", "Test", "Images", "Align", "Align", "Align", "Align", "Align", "Align", "Align", "DVSA", "Align", "11.8", "Align", "45.4", "Align", "12.2", "Align", "*", "Align", "8.9", "Align", "36.3", "Align", "19.5", "Align", "*", "FV", "Align", "17.3", "Align", "50.2", "Align", "10.0", "Align", "46.4", "Align", "10.8", "Align", "40.1", "Align", "17.0", "Align", "49.3", "Align", "Align", "Align", "Align", "Align", "Align", "Align", "order", "-", "embeddings", "(", "symm", ".", ")", "Align", "21.5", "Align", "62.9", "Align", "6.0", "Align", "24.4", "Align", "16.8", "Align", "56.3", "Align", "8.0", "Align", "40.4", "order", "-", "embeddings", "Align", "23.3", "Align", "65.0", "Align", "5.0", "Align", "24.4", "Align", "18.0", "Align", "57.6", "Align", "7.0", "Align", "35.9", "Given", "a", "query", "caption", "or", "image", ",", "we", "sort", "all", "the", "images", "or", "captions", "of", "the", "test", "set", "in", "order", "of", "increasing", "penalty", ".", "We", "use", "standard", "ranking", "metrics", "for", "evaluation", ".", "We", "measure", "Recall@K", ",", "the", "percent", "of", "queries", "for", "which", "the", "GT", "term", "is", "one", "of", "the", "first", "K", "retrieved", ";", "and", "median", "and", "mean", "rank", ",", "which", "are", "statistics", "over", "the", "position", "of", "the", "GT", "term", "in", "the", "retrieval", "order", ".", "Table", "shows", "a", "comparison", "between", "all", "state", "-", "of", "-", "the", "-", "art", "and", "some", "older", "methodsNote", "that", "the", "numbers", "for", "MNLM", "come", "not", "from", "the", "published", "paper", "but", "from", "the", "recently", "released", "code", "at", "http:", "//", "github.com", "/", "ryankiros", "/", "visual", "-", "semantic", "-", "embedding", ".", "along", "with", "our", "own", ";", "see", "for", "a", "more", "complete", "listing", ".", "The", "best", "results", "overall", "are", "in", "bold", ",", "and", "the", "best", "results", "using", "1", "-", "crop", "VGG", "image", "features", "are", "underlined", ".", "Note", "that", "the", "comparison", "is", "additionally", "complicated", "by", "the", "following", ":", "-", "CNN", "is", "an", "ensemble", "of", "four", "different", "models", ",", "whereas", "the", "other", "entries", "are", "all", "single", "models", ".", "STV", "and", "FV", "use", "external", "text", "corpora", "to", "learn", "their", "language", "features", ",", "whereas", "the", "other", "methods", "learn", "them", "from", "scratch", ".", "To", "facilitate", "the", "comparison", "and", "to", "evaluate", "the", "contributions", "of", "various", "components", "of", "our", "model", ",", "we", "evaluate", "four", "variations", "of", "order", "-", "embeddings", ":", "order", "-", "embeddings", "is", "our", "full", "model", "as", "described", "above", ".", "order", "-", "embeddings", "(", "reversed", ")", "reverses", "the", "order", "of", "captions", "and", "image", "embeddings", "in", "our", "order", "-", "violation", "penalty", "—", "placing", "images", "above", "captions", "in", "the", "partial", "order", "learned", "by", "our", "model", ".", "This", "seemingly", "slight", "variation", "performs", "atrociously", ",", "confirming", "our", "prior", "that", "captions", "are", "much", "more", "abstract", "than", "images", ",", "and", "should", "be", "placed", "higher", "in", "the", "semantic", "hierarchy", ".", "order", "-", "embeddings", "(", "1", "-", "crop", ")", "computes", "the", "image", "feature", "using", "just", "the", "center", "crop", ",", "instead", "of", "averaging", "over", "10", "crops", ".", "order", "-", "embeddings", "(", "symm", ".", ")", "replaces", "our", "asymmetric", "penalty", "with", "the", "symmetric", "cosine", "distance", ",", "and", "allows", "embedding", "coordinates", "to", "be", "negative", "—", "essentially", "replicating", "MNLM", ",", "but", "with", "better", "image", "features", ".", "Here", "we", "find", "that", "a", "different", "margin", "(", "=", "α0.2", ")", "works", "best", ".", "Between", "these", "four", "models", ",", "the", "only", "previous", "work", "whose", "results", "are", "incommensurable", "with", "ours", "is", "DVSA", ",", "since", "it", "uses", "the", "less", "discriminative", "CNN", "of", "but", "20", "region", "features", "instead", "of", "a", "single", "whole", "-", "image", "feature", ".", "Aside", "from", "this", "limitation", ",", "and", "if", "only", "single", "models", "are", "considered", ",", "order", "-", "embeddings", "significantly", "outperform", "the", "state", "-", "of", "-", "art", "approaches", "for", "image", "retrieval", "even", "when", "we", "control", "for", "image", "features", ".", "subsection", ":", "Exploration", "Why", "would", "order", "-", "embeddings", "do", "well", "on", "such", "a", "shallow", "partial", "order", "?", "Why", "are", "they", "much", "more", "helpful", "for", "image", "retrieval", "than", "for", "caption", "retrieval", "?", "Intuitively", ",", "symmetric", "similarity", "should", "fail", "when", "an", "image", "has", "captions", "with", "very", "different", "levels", "of", "detail", ",", "because", "the", "captions", "are", "so", "dissimilar", "that", "it", "is", "impossible", "to", "map", "both", "their", "embeddings", "close", "to", "the", "same", "image", "embedding", ".", "Order", "-", "embeddings", "do", "n’t", "have", "this", "problem", ":", "the", "less", "detailed", "caption", "can", "be", "embedded", "very", "far", "away", "from", "the", "image", "while", "remaining", "above", "it", "in", "the", "partial", "order", ".", "To", "evaluate", "this", "intuition", ",", "we", "use", "caption", "length", "as", "a", "proxy", "for", "level", "of", "detail", "and", "select", ",", "among", "pairs", "of", "co", "-", "referring", "captions", "in", "our", "validation", "set", ",", "the", "100", "pairs", "with", "the", "biggest", "length", "difference", ".", "For", "image", "retrieval", "with", "1000", "target", "images", ",", "the", "mean", "rank", "over", "captions", "in", "this", "set", "is", "for", "order", "-", "embeddings", "and", "for", "cosine", "similarity", ",", "a", "much", "bigger", "difference", "than", "over", "the", "entire", "dataset", ".", "Some", "particularly", "dramatic", "examples", "of", "this", "are", "shown", "in", "Figure", "[", "reference", "]", ".", "Moreover", ",", "if", "we", "use", "the", "shorter", "caption", "as", "a", "query", ",", "and", "retrieve", "captions", "in", "order", "of", "increasing", "error", ",", "the", "mean", "rank", "of", "the", "longer", "caption", "is", "for", "order", "-", "embeddings", "and", "for", "cosine", "similarity", ",", "showing", "that", "order", "-", "embeddings", "are", "able", "to", "capture", "the", "relatedness", "of", "co", "-", "referring", "captions", "with", "very", "different", "lengths", ".", "This", "also", "explains", "why", "order", "-", "embeddings", "provide", "a", "much", "smaller", "improvement", "for", "caption", "retrieval", "than", "for", "image", "retrieval", ":", "all", "the", "caption", "retrieval", "metrics", "are", "based", "on", "the", "position", "of", "the", "first", "ground", "truth", "caption", "in", "the", "retrieval", "order", ",", "so", "the", "embeddings", "need", "only", "learn", "to", "retrieve", "one", "of", "each", "image", "’s", "five", "captions", "well", ",", "which", "symmetric", "similarity", "is", "well", "suited", "for", ".", "section", ":", "Textual", "Entailment", "/", "Natural", "Language", "Inference", "Natural", "language", "inference", "can", "be", "seen", "as", "a", "generalization", "of", "hypernymy", "from", "words", "to", "sentences", ".", "For", "example", ",", "from", "“", "woman", "walking", "her", "dog", "in", "a", "park", "”", "we", "can", "infer", "both", "“", "woman", "walking", "her", "dog", "”", "and", "“", "dog", "in", "a", "park", "”", ",", "but", "not", "”", "old", "woman", "”", "or", "”", "black", "dog", "”", ".", "Given", "a", "pair", "of", "sentences", ",", "our", "task", "is", "to", "predict", "whether", "we", "can", "infer", "the", "second", "sentence", "(", "the", "hypothesis", ")", "from", "the", "first", "(", "the", "premise", ")", ".", "subsection", ":", "Loss", "Function", "To", "apply", "order", "-", "embeddings", "to", "this", "task", ",", "we", "again", "view", "it", "as", "partial", "order", "completion", "—", "we", "can", "infer", "a", "hypothesis", "from", "a", "premise", "exactly", "when", "the", "hypothesis", "is", "above", "the", "premise", "in", "the", "visual", "-", "semantic", "hierarchy", ".", "Unlike", "our", "other", "tasks", ",", "for", "which", "we", "had", "to", "generate", "contrastive", "negatives", ",", "datasets", "for", "natural", "language", "inference", "include", "labeled", "negative", "examples", ".", "So", ",", "we", "can", "simply", "use", "a", "max", "-", "margin", "loss", ":", "where", "are", "positive", "and", "negative", "pairs", "of", "premise", "and", "hypothesis", ".", "To", "embed", "sentences", ",", "we", "use", "the", "same", "GRU", "encoder", "as", "in", "the", "caption", "-", "image", "retrieval", "task", ".", "subsection", ":", "Dataset", "To", "evaluate", "order", "-", "embeddings", "on", "the", "natural", "language", "inference", "task", ",", "we", "use", "the", "recently", "proposed", "SNLI", "corpus", "snli", ",", "which", "contains", "570", ",", "000", "pairs", "of", "sentences", ",", "each", "labeled", "with", "“", "entailment", "”", "if", "the", "inference", "is", "valid", ",", "“", "contradiction", "”", "if", "the", "two", "sentences", "contradict", ",", "or", "“", "neutral", "”", "if", "the", "inference", "is", "invalid", "but", "there", "is", "no", "contradiction", ".", "Our", "method", "only", "allows", "us", "to", "discriminate", "between", "entailment", "and", "non", "-", "entailment", ",", "so", "we", "merge", "the", "“", "contradiction", "”", "and", "“", "neutral", "”", "classes", "together", "to", "serve", "as", "our", "negative", "examples", ".", "subsection", ":", "Implementation", "Details", "Just", "as", "for", "caption", "-", "image", "ranking", ",", "we", "set", "the", "dimensions", "of", "the", "embedding", "space", "and", "GRU", "hidden", "state", "to", "be", ",", "the", "dimension", "of", "the", "word", "embeddings", "to", "be", ",", "and", "constrain", "the", "embeddings", "to", "have", "unit", "L2", "norm", ".", "We", "train", "for", "10", "epochs", "with", "batches", "of", "sentence", "pairs", ".", "We", "use", "the", "Adam", "optimizer", "with", "learning", "rate", "and", "early", "stopping", "on", "the", "validation", "set", ".", "During", "evaluation", ",", "we", "find", "the", "optimal", "classification", "threshold", "on", "validation", ",", "then", "use", "the", "threshold", "to", "classify", "the", "test", "set", ".", "subsection", ":", "Results", "The", "state", "-", "of", "-", "the", "-", "art", "method", "for", "3", "-", "class", "classification", "on", "SNLI", "is", "that", "of", "rocktaschel2015reasoning", ".", "Unfortunately", ",", "they", "do", "not", "compute", "2", "-", "class", "accuracy", ",", "so", "we", "can", "not", "compare", "to", "them", "directly", ".", "As", "a", "bridge", "to", "facilitate", "comparison", ",", "we", "use", "a", "challenging", "baseline", "which", "can", "be", "evaluated", "on", "both", "the", "2", "-", "class", "and", "3", "-", "class", "problems", ".", "The", "baseline", ",", "referred", "to", "as", "skip", "-", "thoughts", ",", "involves", "a", "feedforward", "neural", "network", "on", "top", "of", "skip", "-", "thought", "vectors", "kiros2015skip", ",", "a", "state", "-", "of", "-", "the", "-", "art", "semantic", "representation", "of", "sentences", ".", "Given", "pairs", "of", "sentence", "vectors", "and", ",", "the", "input", "to", "the", "network", "is", "the", "concatenation", "of", ",", "and", "the", "absolute", "difference", ".", "We", "tuned", "the", "number", "of", "layers", ",", "layer", "dimensionality", "and", "dropout", "rates", "to", "optimize", "performance", "on", "the", "development", "set", ",", "using", "the", "Adam", "optimizer", ".", "Batch", "normalization", "ioffe2015batch", "and", "PReLU", "units", "he2015delving", "were", "used", ".", "Our", "best", "network", "used", "2", "hidden", "layers", "of", "1000", "units", "each", ",", "with", "dropout", "rate", "of", "0.5", "across", "both", "the", "input", "and", "hidden", "layers", ".", "We", "did", "not", "backpropagate", "through", "the", "skip", "-", "thought", "encoder", ".", "We", "also", "evaluate", "against", "EOP", "classifier", ",", "a", "2", "-", "class", "baseline", "introduced", "by", "snli", ",", "and", "against", "a", "version", "of", "our", "model", "where", "our", "order", "-", "violation", "penalty", "is", "replaced", "with", "the", "symmetric", "cosine", "distance", ",", "order", "-", "embeddings", "(", "symmetric", ")", ".", "The", "results", "for", "all", "models", "are", "shown", "in", "Table", "[", "reference", "]", ".", "We", "see", "that", "order", "-", "embeddings", "outperform", "the", "skip", "-", "thought", "baseline", "despite", "not", "using", "external", "text", "corpora", ".", "While", "our", "method", "is", "almost", "certainly", "worse", "than", "the", "state", "-", "of", "-", "the", "-", "art", "method", "of", "rocktaschel2015reasoning", ",", "which", "uses", "a", "word", "-", "by", "-", "word", "attention", "mechanism", ",", "it", "is", "also", "much", "simpler", ".", "section", ":", "Conclusion", "and", "Future", "Work", "We", "introduced", "a", "simple", "method", "to", "encode", "order", "into", "learned", "distributed", "representations", ",", "which", "allows", "us", "to", "explicitly", "model", "the", "partial", "order", "structure", "of", "the", "visual", "-", "semantic", "hierarchy", ".", "Our", "method", "can", "be", "easily", "integrated", "into", "existing", "relational", "learning", "methods", ",", "as", "we", "demonstrated", "on", "three", "challenging", "tasks", "involving", "computer", "vision", "and", "natural", "language", "processing", ".", "On", "two", "of", "these", "tasks", ",", "hypernym", "prediction", "and", "caption", "-", "image", "retrieval", ",", "our", "methods", "outperform", "all", "previous", "work", ".", "A", "promising", "direction", "of", "future", "work", "is", "to", "learn", "better", "classifiers", "on", "ImageNet", "imagenet", ",", "which", "has", "over", "21k", "image", "classes", "arranged", "by", "the", "WordNet", "hierarchy", ".", "Previous", "approaches", ",", "including", "frome2013devise", "and", "norouzi2013zero", "have", "embedded", "words", "and", "images", "into", "a", "shared", "semantic", "space", "with", "symmetric", "similarity", "—", "which", "our", "experiments", "suggest", "to", "be", "a", "poor", "fit", "with", "the", "partial", "order", "structure", "of", "WordNet", ".", "We", "expect", "significant", "progress", "on", "ImageNet", "classification", ",", "and", "the", "related", "problems", "of", "one", "-", "shot", "and", "zero", "-", "shot", "learning", ",", "to", "be", "possible", "using", "order", "-", "embeddings", ".", "Going", "further", ",", "order", "-", "embeddings", "may", "enable", "learning", "the", "entire", "semantic", "hierarchy", "in", "a", "single", "model", "which", "jointly", "reasons", "about", "hypernymy", ",", "entailment", ",", "and", "the", "relationship", "between", "perception", "and", "language", ",", "unifying", "what", "have", "been", "until", "now", "almost", "independent", "lines", "of", "work", ".", "subsubsection", ":", "Acknowledgments", "We", "thank", "Kaustav", "Kundu", "for", "many", "fruitful", "discussions", "throughout", "the", "development", "of", "this", "paper", ".", "The", "work", "was", "supported", "in", "part", "by", "an", "NSERC", "Graduate", "Scholarship", ".", "bibliography", ":", "References", "section", ":", "Supplementary", "Material", "mikolov2013linguistic", "showed", "that", "word", "representations", "learned", "using", "word2vec", "exhibit", "semantic", "regularities", ",", "such", "as", ".", "kiros2014", "showed", "that", "similar", "regularities", "hold", "for", "joint", "image", "-", "language", "models", ".", "We", "find", "that", "order", "-", "embeddings", "exhibit", "a", "novel", "form", "of", "regularity", ",", "shown", "in", "Figure", "[", "reference", "]", ".", "The", "elementwise", "and", "operations", "in", "the", "embedding", "space", "roughly", "correspond", "to", "composition", "and", "abstraction", ",", "respectively", "."], "entity": [[2, 9, ["Order", "-", "Embeddings", "of", "Images", "and", "Language"], "Task"], [9, 10, ["Hypernymy"], "Method"], [11, 13, ["textual", "entailment"], "Task"], [15, 17, ["image", "captioning"], "Task"], [64, 67, ["learning", "ordered", "representations"], "Task"], [97, 99, ["hypernym", "prediction"], "Task"], [100, 104, ["image", "-", "caption", "retrieval"], "Task"], [111, 114, ["natural", "language", "processing"], "Task"], [122, 123, ["vision"], "Task"], [138, 145, ["open", "-", "ended", "lingual", "descriptions", "of", "images"], "Method"], [150, 153, ["natural", "language", "processing"], "Task"], [155, 156, ["flickr30k"], "Method"], [174, 177, ["autonomous", "artificial", "agents"], "Task"], [346, 355, ["hypernym", ",", "entailment", ",", "and", "image", "-", "caption", "relations"], "Task"], [357, 359, ["distributed", "representations"], "Method"], [360, 361, ["embeddings"], "Method"], [405, 409, ["caption", "-", "image", "relationship"], "Task"], [414, 415, ["mapping"], "Task"], [437, 440, ["symmetric", "distance", "measure"], "Metric"], [442, 446, ["Euclidean", "or", "cosine", "distance"], "Method"], [469, 471, ["model", "error"], "Metric"], [550, 551, ["mapping"], "Method"], [583, 586, ["order", "-", "embeddings"], "Method"], [594, 597, ["relational", "learning", "methods"], "Method"], [601, 603, ["comparison", "operation"], "Method"], [617, 620, ["order", "-", "embeddings"], "Method"], [632, 634, ["hypernymy", "prediction"], "Task"], [635, 639, ["caption", "-", "image", "retrieval"], "Task"], [651, 654, ["natural", "language", "inference"], "Task"], [687, 691, ["learning", "order", "-", "embeddings"], "Task"], [710, 714, ["order", "-", "embeddings", "idea"], "Method"], [728, 730, ["hypernym", "prediction"], "Task"], [737, 741, ["caption", "-", "image", "retrieval"], "Task"], [749, 751, ["textual", "entailment"], "Task"], [781, 785, ["Learning", "Order", "-", "Embeddings"], "Task"], [798, 801, ["partial", "order", "completion"], "Task"], [803, 806, ["partial", "order", "completion"], "Task"], [852, 854, ["hypernym", "prediction"], "Task"], [855, 859, ["caption", "-", "image", "retrieval"], "Task"], [861, 863, ["textual", "entailment"], "Task"], [958, 961, ["order", "-", "embedding"], "Method"], [981, 984, ["order", "-", "embedding"], "Method"], [1040, 1042, ["semantic", "hierarchy"], "Task"], [1062, 1063, ["abstraction"], "Method"], [1064, 1065, ["composition"], "Task"], [1190, 1191, ["embedding"], "Task"], [1206, 1207, ["optimization"], "Task"], [1208, 1212, ["gradient", "-", "based", "methods"], "Method"], [1354, 1355, ["flickr30k"], "Method"], [1358, 1361, ["Penalizing", "Order", "Violations"], "Task"], [1377, 1380, ["order", "-", "embedding"], "Method"], [1414, 1418, ["approximate", "order", "-", "embedding"], "Method"], [1420, 1421, ["mapping"], "Method"], [1571, 1575, ["approximate", "order", "-", "embedding"], "Task"], [1580, 1584, ["max", "-", "margin", "loss"], "Method"], [1664, 1666, ["Hypernym", "Prediction"], "Task"], [1691, 1692, ["WordNet"], "Material"], [1789, 1790, ["WordNet"], "Material"], [1841, 1842, ["WordNet"], "Material"], [1881, 1883, ["Loss", "Function"], "Method"], [1908, 1909, ["WordNet"], "Material"], [1914, 1917, ["neural", "tensor", "network"], "Method"], [1967, 1968, ["evaluation"], "Task"], [1972, 1976, ["max", "-", "margin", "loss"], "Method"], [2084, 2085, ["transitivity"], "Method"], [2109, 2110, ["WordNet"], "Material"], [2112, 2116, ["max", "-", "margin", "objective"], "Method"], [2145, 2148, ["Adam", "optimizer", "adam"], "Method"], [2149, 2151, ["learning", "rate"], "Method"], [2152, 2154, ["early", "stopping"], "Method"], [2166, 2168, ["classification", "threshold"], "Metric"], [2221, 2224, ["transitive", "closure", "baseline"], "Method"], [2226, 2227, ["learning"], "Method"], [2255, 2257, ["word2gauss", "baseline"], "Method"], [2261, 2262, ["word2gauss"], "Method"], [2285, 2287, ["KL", "divergence"], "Method"], [2290, 2296, ["50", "-", "dimensional", "diagonal", "Gaussian", "embeddings"], "Method"], [2303, 2307, ["max", "-", "margin", "objective"], "Metric"], [2312, 2314, ["grid", "search"], "Method"], [2315, 2318, ["order", "-", "embeddings"], "Method"], [2323, 2325, ["full", "model"], "Method"], [2337, 2340, ["order", "-", "embeddings"], "Method"], [2341, 2343, ["bilinear", ")"], "Method"], [2348, 2350, ["bilinear", "model"], "Method"], [2354, 2357, ["order", "-", "embeddings"], "Method"], [2371, 2373, ["transitive", "baseline"], "Method"], [2400, 2402, ["bilinear", "baseline"], "Method"], [2440, 2441, ["Caption"], "Task"], [2442, 2444, ["Image", "Retrieval"], "Task"], [2445, 2450, ["caption", "-", "image", "retrieval", "task"], "Task"], [2456, 2458, ["joint", "models"], "Method"], [2459, 2460, ["vision"], "Task"], [2461, 2462, ["language"], "Task"], [2464, 2465, ["LinCVPR14"], "Task"], [2469, 2470, ["ranking"], "Task"], [2482, 2484, ["Image", "Retrieval"], "Task"], [2487, 2489, ["ranking", "captions"], "Task"], [2490, 2491, ["relevance"], "Task"], [2496, 2498, ["Caption", "Retrieval"], "Task"], [2520, 2525, ["caption", "-", "image", "compatibility", "score"], "Metric"], [2556, 2560, ["inner", "-", "product", "similarity"], "Method"], [2567, 2570, ["Canonical", "Correlations", "Analysis"], "Method"], [2571, 2575, ["distributed", "representations", "of", "images"], "Method"], [2576, 2577, ["captions"], "Method"], [2634, 2636, ["multimodal", "CNN"], "Method"], [2699, 2701, ["Loss", "Function"], "Method"], [2709, 2712, ["pairwise", "ranking", "loss"], "Metric"], [2727, 2730, ["symmetric", "similarity", "measure"], "Metric"], [2739, 2741, ["loss", "function"], "Method"], [2790, 2794, ["Image", "and", "Caption", "Embeddings"], "Task"], [2803, 2804, ["kiros2014"], "Method"], [2890, 2892, ["VGG", "network"], "Method"], [2893, 2894, ["vgg"], "Method"], [2900, 2901, ["ImageNet"], "Material"], [2921, 2925, ["recurrent", "neural", "net", "encoder"], "Method"], [2926, 2929, ["GRU", "activations", "gru"], "Method"], [2951, 2955, ["Microsoft", "COCO", "dataset", "coco"], "Material"], [2986, 2990, ["caption", "-", "image", "retrieval"], "Task"], [3034, 3037, ["pairwise", "ranking", "objective"], "Metric"], [3089, 3091, ["Adam", "optimizer"], "Method"], [3092, 3094, ["learning", "rate"], "Method"], [3096, 3098, ["early", "stopping"], "Method"], [3140, 3142, ["learning", "rate"], "Metric"], [3217, 3220, ["Align", "Caption", "Retrieval"], "Task"], [3221, 3223, ["Image", "Retrieval"], "Task"], [3231, 3234, ["Align", "Mean", "rAlign"], "Metric"], [3254, 3256, ["MNLM", "Align"], "Task"], [3273, 3275, ["RNN", "Align41.0"], "Method"], [3289, 3290, ["DVSA"], "Method"], [3306, 3308, ["STV", "Align"], "Method"], [3339, 3343, ["m", "-", "CNN", "Align"], "Method"], [3404, 3408, ["Align", "30.1order", "-", "embeddings"], "Task"], [3425, 3427, ["Align", "2.6Align"], "Task"], [3428, 3431, ["order", "-", "embeddings"], "Method"], [3432, 3433, ["symm"], "Method"], [3449, 3454, ["Align", "9.0", "order", "-", "embeddings"], "Task"], [3521, 3524, ["order", "-", "embeddings"], "Method"], [3525, 3526, ["symm"], "Method"], [3542, 3547, ["Align", "40.4", "order", "-", "embeddings"], "Task"], [3559, 3560, ["Align"], "Task"], [3590, 3592, ["ranking", "metrics"], "Metric"], [3593, 3594, ["evaluation"], "Task"], [3597, 3598, ["Recall@K"], "Metric"], [3617, 3621, ["median", "and", "mean", "rank"], "Metric"], [3658, 3659, ["MNLM"], "Method"], [3748, 3749, ["STV"], "Method"], [3750, 3751, ["FV"], "Method"], [3791, 3794, ["order", "-", "embeddings"], "Method"], [3795, 3798, ["order", "-", "embeddings"], "Method"], [3806, 3809, ["order", "-", "embeddings"], "Method"], [3869, 3872, ["order", "-", "embeddings"], "Method"], [3873, 3876, ["1", "-", "crop"], "Method"], [3894, 3897, ["order", "-", "embeddings"], "Method"], [3898, 3899, ["symm"], "Method"], [3921, 3922, ["MNLM"], "Method"], [3959, 3960, ["DVSA"], "Method"], [3966, 3968, ["discriminative", "CNN"], "Method"], [3995, 3998, ["order", "-", "embeddings"], "Method"], [4008, 4010, ["image", "retrieval"], "Task"], [4023, 4026, ["order", "-", "embeddings"], "Method"], [4042, 4044, ["image", "retrieval"], "Task"], [4046, 4048, ["caption", "retrieval"], "Task"], [4089, 4092, ["Order", "-", "embeddings"], "Method"], [4161, 4163, ["image", "retrieval"], "Task"], [4169, 4171, ["mean", "rank"], "Metric"], [4183, 4185, ["cosine", "similarity"], "Metric"], [4229, 4230, ["error"], "Metric"], [4232, 4234, ["mean", "rank"], "Metric"], [4240, 4243, ["order", "-", "embeddings"], "Method"], [4245, 4247, ["cosine", "similarity"], "Metric"], [4250, 4253, ["order", "-", "embeddings"], "Method"], [4273, 4276, ["order", "-", "embeddings"], "Method"], [4282, 4284, ["caption", "retrieval"], "Task"], [4286, 4288, ["image", "retrieval"], "Task"], [4291, 4294, ["caption", "retrieval", "metrics"], "Metric"], [4328, 4330, ["symmetric", "similarity"], "Metric"], [4337, 4339, ["Textual", "Entailment"], "Task"], [4340, 4343, ["Natural", "Language", "Inference"], "Task"], [4351, 4354, ["generalization", "of", "hypernymy"], "Method"], [4434, 4436, ["Loss", "Function"], "Method"], [4438, 4441, ["order", "-", "embeddings"], "Method"], [4450, 4453, ["partial", "order", "completion"], "Task"], [4493, 4496, ["natural", "language", "inference"], "Task"], [4508, 4512, ["max", "-", "margin", "loss"], "Method"], [4532, 4534, ["GRU", "encoder"], "Method"], [4537, 4542, ["caption", "-", "image", "retrieval", "task"], "Task"], [4548, 4551, ["order", "-", "embeddings"], "Task"], [4563, 4564, ["SNLI"], "Material"], [4565, 4566, ["snli"], "Material"], [4653, 4657, ["caption", "-", "image", "ranking"], "Task"], [4706, 4708, ["Adam", "optimizer"], "Method"], [4709, 4711, ["learning", "rate"], "Method"], [4712, 4714, ["early", "stopping"], "Method"], [4726, 4728, ["classification", "threshold"], "Metric"], [4754, 4758, ["3", "-", "class", "classification"], "Task"], [4759, 4760, ["SNLI"], "Material"], [4771, 4775, ["2", "-", "class", "accuracy"], "Metric"], [4804, 4807, ["2", "-", "class"], "Task"], [4808, 4812, ["3", "-", "class", "problems"], "Task"], [4825, 4828, ["feedforward", "neural", "network"], "Method"], [4831, 4835, ["skip", "-", "thought", "vectors"], "Method"], [4845, 4849, ["semantic", "representation", "of", "sentences"], "Method"], [4882, 4884, ["dropout", "rates"], "Metric"], [4894, 4896, ["Adam", "optimizer"], "Method"], [4897, 4899, ["Batch", "normalization"], "Method"], [4901, 4903, ["PReLU", "units"], "Method"], [4938, 4942, ["skip", "-", "thought", "encoder"], "Method"], [4947, 4949, ["EOP", "classifier"], "Method"], [4957, 4958, ["snli"], "Material"], [4980, 4983, ["order", "-", "embeddings"], "Method"], [4984, 4985, ["symmetric"], "Method"], [5003, 5006, ["order", "-", "embeddings"], "Method"], [5008, 5012, ["skip", "-", "thought", "baseline"], "Method"], [5042, 5049, ["word", "-", "by", "-", "word", "attention", "mechanism"], "Method"], [5072, 5074, ["distributed", "representations"], "Method"], [5100, 5103, ["relational", "learning", "methods"], "Method"], [5115, 5118, ["natural", "language", "processing"], "Task"], [5125, 5127, ["hypernym", "prediction"], "Task"], [5128, 5132, ["caption", "-", "image", "retrieval"], "Task"], [5150, 5151, ["classifiers"], "Method"], [5152, 5154, ["ImageNet", "imagenet"], "Material"], [5210, 5212, ["ImageNet", "classification"], "Task"], [5218, 5221, ["one", "-", "shot"], "Task"], [5222, 5226, ["zero", "-", "shot", "learning"], "Task"], [5231, 5234, ["order", "-", "embeddings"], "Method"], [5238, 5241, ["order", "-", "embeddings"], "Method"], [5320, 5322, ["word", "representations"], "Method"], [5348, 5351, ["order", "-", "embeddings"], "Method"], [83, 84, ["language"], "Task"], [109, 110, ["vision"], "Task"], [160, 161, ["language"], "Task"], [182, 183, ["vision"], "Task"], [184, 185, ["language"], "Task"], [260, 261, ["language"], "Task"], [777, 778, ["language"], "Task"], [1059, 1060, ["language"], "Task"], [1744, 1745, ["WordNet"], "Material"], [1780, 1781, ["WordNet"], "Material"], [2041, 2042, ["WordNet"], "Material"], [2049, 2050, ["WordNet"], "Material"], [2622, 2623, ["language"], "Task"], [3409, 3412, ["1", "-", "crop"], "Method"], [3480, 3481, ["DVSA"], "Method"], [3709, 3712, ["1", "-", "crop"], "Method"], [3758, 3759, ["language"], "Task"], [4344, 4345, ["language"], "Task"], [4554, 4555, ["language"], "Task"], [4951, 4954, ["2", "-", "class"], "Task"], [5113, 5114, ["vision"], "Task"], [5164, 5165, ["WordNet"], "Material"], [5203, 5204, ["WordNet"], "Material"], [5266, 5267, ["language"], "Task"], [5342, 5343, ["language"], "Task"]]} {"doc_id": "34f92431bb8cb9fa6a1210c5d6e4642079ffac7a", "text": ["document", ":", "Sliding", "Line", "Point", "Regression", "for", "Shape", "Robust", "Scene", "Text", "Detection", "Traditional", "text", "detection", "methods", "mostly", "focus", "on", "quadrangle", "text", ".", "In", "this", "study", "we", "propose", "a", "novel", "method", "named", "sliding", "line", "point", "regression", "(", "SLPR", ")", "in", "order", "to", "detect", "arbitrary", "-", "shape", "text", "in", "natural", "scene", ".", "SLPR", "regresses", "multiple", "points", "on", "the", "edge", "of", "text", "line", "and", "then", "utilizes", "these", "points", "to", "sketch", "the", "outlines", "of", "the", "text", ".", "The", "proposed", "SLPR", "can", "be", "adapted", "to", "many", "object", "detection", "architectures", "such", "as", "Faster", "R", "-", "CNN", "and", "R", "-", "FCN", ".", "Specifically", ",", "we", "first", "generate", "the", "smallest", "rectangular", "box", "including", "the", "text", "with", "region", "proposal", "network", "(", "RPN", ")", ",", "then", "isometrically", "regress", "the", "points", "on", "the", "edge", "of", "text", "by", "using", "the", "vertically", "and", "horizontally", "sliding", "lines", ".", "To", "make", "full", "use", "of", "information", "and", "reduce", "redundancy", ",", "we", "calculate", "x", "-", "coordinate", "or", "y", "-", "coordinate", "of", "target", "point", "by", "the", "rectangular", "box", "position", ",", "and", "just", "regress", "the", "remaining", "y", "-", "coordinate", "or", "x", "-", "coordinate", ".", "Accordingly", "we", "can", "not", "only", "reduce", "the", "parameters", "of", "system", ",", "but", "also", "restrain", "the", "points", "which", "will", "generate", "more", "regular", "polygon", ".", "Our", "approach", "achieved", "competitive", "results", "on", "traditional", "ICDAR2015", "Incidental", "Scene", "Text", "benchmark", "and", "curve", "text", "detection", "dataset", "CTW1500", ".", "section", ":", "Introduction", "Text", "detection", "is", "important", "in", "our", "daily", "life", "as", "it", "can", "be", "applied", "in", "many", "areas", ",", "such", "as", "digitization", "of", "text", ",", "text", "translation", ",", "etc", ".", "In", "this", "study", ",", "we", "focus", "on", "scene", "text", "detection", ".", "Some", "of", "the", "previous", "methods", "have", "obtained", "good", "results", "on", "many", "horizontal", "scene", "texts", "dataset", "based", "on", "Faster", "R", "-", "CNN", "or", "SSD", ".", "Some", "methods", "also", "tried", "to", "solve", "arbitrary", "-", "oriented", "text", "detection", "problem", ".", "and", "regressed", "first", "a", "horizontal", "rectangle", "and", "then", "a", "quadrilateral", ".", "aimed", "to", "generate", "an", "irregular", "polygon", "after", "regressing", "a", "rectangle", ".", "The", "methods", "mentioned", "above", "mostly", "treated", "a", "text", "line", "as", "a", "quadrilateral", "which", "can", "be", "completely", "represented", "by", "four", "points", ".", "However", ",", "besides", "the", "quadrilateral", "shape", ",", "there", "are", "many", "other", "various", "shapes", "of", "text", "line", "in", "natural", "scene", ".", "Therefore", ",", "recent", "research", "have", "begun", "to", "explore", "curve", "text", "line", "detection", ".", "In", "this", "paper", "we", "explore", "both", "arbitrary", "-", "oriented", "and", "curve", "text", "detection", ".", "Our", "method", "named", "sliding", "line", "point", "regression", "(", "SLPR", ")", "is", "based", "on", "2", "-", "step", "object", "detection", "methods", "using", "Faster", "R", "-", "CNN", "or", "R", "-", "FCN", ".", "Firstly", "we", "propose", "some", "interesting", "rectangular", "regions", "with", "region", "proposal", "network", "(", "RPN", ")", ",", "then", "regress", "the", "points", "on", "the", "edge", "of", "text", ".", "We", "generate", "some", "rules", "to", "determine", "which", "points", "should", "be", "regressed", "so", "that", "there", "will", "be", "relevance", "between", "points", ".", "Different", "from", "which", "directly", "regressed", "both", "x", "-", "coordinate", "and", "y", "-", "coordinate", "of", "fixed", "annotated", "points", "and", "employed", "RNN", "to", "learn", "their", "relevance", ",", "we", "introduce", "some", "rules", "to", "vertically", "and", "horizontally", "slide", "lines", "along", "text", "and", "then", "regress", "the", "intersection", "points", "of", "sliding", "lines", "and", "text", "lines", "as", "illustrated", "in", "Fig", ".", "[", "reference", "]", ".", "In", "this", "way", ",", "we", "can", "only", "regress", "x", "-", "coordinate", "or", "y", "-", "coordinate", "of", "these", "points", ",", "then", "calculate", "other", "coordinates", "with", "the", "position", "of", "rectangle", ",", "yielding", "reduction", "of", "unnecessary", "computation", "and", "improvement", "of", "performance", ".", "The", "contributions", "of", "this", "paper", "are", "as", "follows", ":", "1", ".", "We", "explore", "regressing", "multiple", "points", "on", "the", "border", "of", "text", "line", ",", "try", "to", "handle", "arbitrary", "-", "oriented", "and", "curve", "text", "detection", "based", "on", "Faster", "R", "-", "CNN", "and", "R", "-", "FCN", ".", "2", ".", "We", "introduce", "a", "sliding", "line", "method", "to", "determine", "the", "ground", "truth", "points", "for", "the", "regression", ",", "and", "we", "make", "full", "use", "of", "the", "relevance", "of", "these", "points", "to", "generate", "more", "regular", "polygon", ".", "section", ":", "Related", "Work", "In", "recent", "years", ",", "scene", "text", "detection", "and", "recognition", "has", "drawn", "more", "and", "more", "attention", ".", "But", "scene", "text", "detection", "remains", "a", "difficult", "problem", "due", "to", "its", "complicated", "orientation", "and", "background", ".", "All", "the", "methods", "can", "be", "divided", "into", "three", "categories", ":", "character", "based", "methods", ",", "word", "based", "methods", "and", "segmentation", "based", "methods", ".", "Character", "based", "methods", "often", "need", "synthetic", "datasets", "because", "labeling", "characters", "in", "text", "lines", "requires", "additional", "efforts", ".", "However", ",", "the", "generated", "data", "is", "greatly", "deviated", "from", "the", "real", "data", ",", "which", "can", "not", "make", "the", "trained", "model", "to", "achieve", "the", "state", "-", "of", "-", "the", "-", "art", "results", "on", "real", "dataset", "such", "as", "the", "popular", "ICDAR2015", "Incidental", "Scene", "Text", "benchmark", ".", "In", "order", "to", "solve", "this", "problem", ",", "used", "semi", "-", "supervised", "method", "to", "finetune", "model", "on", "real", "data", "and", "obtained", "good", "results", ".", "The", "segmentation", "based", "methods", "have", "also", "been", "used", "in", "text", "detection", "recently", ".", "trained", "a", "fully", "convolutional", "network", "(", "FCN", ")", "to", "predict", "the", "salient", "map", "of", "text", "regions", ",", "then", "traced", "the", "text", "line", "by", "combining", "the", "salient", "map", "and", "character", "components", ".", "added", "the", "border", "class", "to", "separate", "text", "from", "their", "neighbor", ".", "and", "generated", "text", "maps", "and", "regress", "the", "size", "and", "angle", "of", "the", "corresponding", "quadrilateral", ",", "or", "coordinates", "of", "four", "vertexes", "at", "the", "same", "time", ".", "Compared", "with", "the", "traditional", "segmentation", "methods", ",", "they", "made", "a", "huge", "breakthrough", "on", "ICDAR2015", "Incidental", "Scene", "Text", "benchmark", ".", "Many", "methods", "of", "object", "detection", "can", "be", "applied", "to", "text", "detection", ",", "e.g.", ",", "Faster", "R", "-", "CNN", ",", "SSD", ",", "R", "-", "FCN", "and", "YOLO", ".", "used", "irregular", "convolutional", "filters", "instead", "of", "the", "standard", "convolutional", "filters", "to", "make", "the", "network", "more", "suitable", "for", "long", "text", "detection", ".", "used", "the", "attention", "map", "to", "remove", "background", "noise", ".", "Recently", "more", "and", "more", "researchers", "proposed", "2", "-", "step", "methods", "based", "on", "Faster", "R", "-", "CNN", "or", "R", "-", "FCN", ".", "firstly", "generated", "axis", "-", "aligned", "bounding", "boxes", "and", "then", "regressed", "the", "text", "quadrangle", ".", "They", "used", "multi", "-", "scale", "pool", "operations", "on", "the", "roipool", "layer", ".", "tried", "to", "segment", "and", "detect", "text", "simultaneously", ".", "Considering", "the", "particularity", "of", "text", "line", ",", "appended", "different", "angle", "anchors", "which", "are", "suitable", "for", "arbitrary", "-", "oriented", "text", "line", ".", "More", "recently", ",", "considered", "the", "polygon", "case", "and", "labeled", "a", "new", "dataset", "of", "curve", "text", ".", "also", "constructed", "a", "curve", "text", "dataset", "named", "CTW1500", ",", "and", "they", "proposed", "a", "new", "structure", "named", "curve", "text", "detector", "(", "CTD", ")", "to", "solve", "curve", "text", "detection", "problem", ".", "section", ":", "Method", "Our", "model", "can", "be", "applied", "to", "any", "2", "-", "step", "object", "detection", "framework", "such", "as", "Faster", "R", "-", "CNN", "and", "R", "-", "FCN", ".", "Our", "system", "simultaneously", "regresses", "the", "minimum", "rectangle", "including", "text", "line", "and", "the", "coordinates", "of", "some", "specific", "points", "on", "the", "boundary", "of", "text", "line", ".", "More", "specifically", ",", "take", "Faster", "R", "-", "CNN", "as", "an", "example", ",", "we", "first", "get", "some", "interesting", "regions", "using", "the", "RPN", ",", "then", "we", "not", "only", "regress", "the", "position", "of", "the", "rectangle", ",", "but", "also", "regress", "the", "coordinates", "of", "the", "points", "on", "the", "edge", "of", "the", "text", "line", ",", "finally", "we", "can", "get", "arbitrary", "shape", "text", "area", ".", "subsection", ":", "Which", "points", "should", "be", "regressed", "?", "Obviously", ",", "how", "to", "determine", "the", "point", "set", "for", "restoring", "the", "polygon", "is", "quite", "important", ".", "We", "believe", "the", "simpler", "the", "rules", ",", "the", "easier", "the", "neural", "net", "learns", ".", "We", "do", "not", "regress", "the", "fixed", "points", "such", "as", "vertexes", "on", "the", "polygon", "because", "there", "are", "a", "large", "variety", "of", "shapes", "and", "angles", "in", "natural", "scene", "and", "it", "is", "difficult", "to", "define", "the", "order", "of", "fixed", "feature", "points", "for", "all", "shapes", ".", "Although", "for", "quadrilateral", ",", "we", "can", "perfectly", "restore", "it", "by", "regressing", "the", "corresponding", "four", "vertices", ",", "the", "determination", "of", "the", "order", "of", "four", "vertices", "requires", "a", "complicated", "rule", "which", "is", "difficult", "for", "the", "neural", "net", "to", "learn", ".", "Alternatively", ",", "as", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "we", "introduce", "some", "rules", "to", "vertically", "and", "horizontally", "slide", "the", "lines", "(", "we", "use", "equidistant", "sliding", "in", "our", "experiment", ")", "on", "text", "line", "and", "then", "regress", "the", "intersection", "of", "sliding", "lines", "and", "text", "line", "border", ".", "On", "the", "other", "hand", ",", "the", "correlation", "exists", "among", "the", "coordinates", "of", "different", "intersection", "points", "due", "to", "the", "constraints", "of", "the", "sliding", "lines", ".", "It", "is", "not", "necessary", "to", "regress", "both", "x", "-", "coordinate", "and", "y", "-", "coordinate", "of", "all", "points", "simultaneously", ".", "If", "it", "is", "horizontal", "sliding", ",", "the", "x", "-", "coordinate", "of", "the", "point", "on", "the", "text", "boundary", "can", "be", "calculated", "by", "the", "coordinates", "of", "the", "rectangle", ",", "so", "we", "only", "need", "to", "regress", "the", "y", "-", "coordinate", "of", "these", "points", ".", "Similarly", ",", "if", "it", "is", "vertical", "sliding", ",", "we", "only", "need", "to", "regress", "the", "x", "-", "coordinate", "of", "these", "points", ".", "This", "method", "not", "only", "reduces", "the", "computational", "complexity", "of", "the", "network", ",", "but", "also", "adds", "restraints", "to", "the", "regressed", "points", "as", "the", "prior", "knowledge", "which", "can", "prevent", "generating", "polygons", "with", "weird", "shapes", "and", "further", "improve", "the", "accuracy", ".", "As", "for", "the", "number", "of", "sliding", "lines", ",", "we", "observe", "that", "this", "parameter", "is", "not", "sensitive", "to", "quadrangle", "text", "line", ".", "But", "in", "order", "to", "restore", "other", "shape", "text", "line", "well", ",", "after", "balancing", "the", "performance", "and", "network", "complexity", ",", "seven", "sliding", "lines", "are", "used", "for", "we", "decided", "to", "for", "vertical", "and", "horizontal", "directions", ",", "respectively", ".", "Accordingly", "a", "total", "of", "14", "lines", "with", "28", "intersection", "points", "are", "generated", ".", "subsection", ":", "The", "multi", "-", "task", "learning", "To", "optimize", "the", "neural", "network", "parameters", ",", "as", "illustrated", "in", "Fig", ".", "[", "reference", "]", ",", "we", "adopt", "the", "multi", "-", "task", "learning", "to", "define", "the", "loss", "function", "as", ":", "where", "is", "the", "region", "proposal", "loss", ",", "is", "region", "proposal", "classification", "loss", ",", "is", "box", "regression", "loss", ".", "is", "the", "loss", "for", "the", "second", "step", "after", "RPN", ".", "Similarly", ",", "the", "first", "two", "items", "and", "are", "respectively", "classification", "loss", "and", "box", "regression", "loss", ".", ",", "and", "are", "the", "related", "weighting", "factors", ",", "which", "are", "all", "set", "to", "1", "in", "this", "study", ".", "is", "the", "proposed", "new", "loss", "item", "for", "SLPR", ":", "is", "the", "smooth", "L1", "loss", "for", "the", "box", "regression", "task", ":", "In", "Eq", ".", "(", "[", "reference", "]", ")", ",", "represents", "the", "number", "of", "sliding", "lines", "in", "one", "direction", "and", "we", "set", "in", "our", "experiments", ".", "In", "general", ",", "each", "line", "has", "two", "intersection", "points", "with", "the", "text", "line", "border", ".", "If", "there", "are", "more", "than", "two", "intersection", "points", ",", "we", "take", "the", "smallest", "and", "the", "largest", "coordinates", ".", "is", "x", "-", "coordinate", "of", "the", "intersection", "point", "of", "vertically", "sliding", "lines", "and", "text", "line", "border", "while", "is", "y", "-", "coordinate", "of", "the", "intersection", "point", "of", "horizontally", "sliding", "lines", "and", "text", "line", "border", ".", "and", "are", "the", "corresponding", "estimated", "points", "from", "neural", "net", "outputs", ".", "For", "horizontally", "sliding", "lines", ",", "we", "only", "regress", "the", "y", "-", "coordinate", "of", "its", "intersection", "point", ".", "For", "vertically", "sliding", "lines", ",", "we", "only", "regress", "the", "x", "-", "coordinate", "of", "its", "intersection", "point", ".", "The", "other", "coordinates", "can", "be", "restored", "through", "the", "coordinates", "of", "the", "rectangle", ":", "and", "represent", "the", "minimum", "x", "-", "coordinate", "and", "y", "-", "coordinate", "of", "the", "rectangular", "border", "while", "and", "represent", "the", "maximum", "x", "-", "coordinate", "and", "y", "-", "coordinate", "of", "the", "rectangular", "border", ".", "is", "the", "floor", "function", ".", "In", "a", "word", ",", "in", "order", "to", "regress", "the", "coordinates", "of", "polygon", ",", "32", "parameters", "should", "be", "considered", "including", "4", "parameters", "for", "the", "rectangle", "and", "28", "parameters", "to", "represent", "x", "and", "y", "coordinates", "of", "intersection", "points", "on", "text", "line", "border", ".", "subsection", ":", "Restoration", "of", "polygon", "Through", "the", "above", "SLPR", "method", ",", "we", "can", "obtain", "multiple", "points", "from", "the", "output", "of", "neural", "nets", ".", "To", "restore", "the", "final", "quadrilateral", "or", "polygon", ",", "the", "following", "two", "approaches", "are", "adopted", "and", "compared", ":", "subsubsection", ":", "Only", "Using", "Points", "in", "Long", "Side", "(", "PLS", ")", "The", "text", "line", "always", "extends", "to", "the", "long", "side", ",", "and", "the", "lines", "that", "slide", "along", "the", "long", "side", "can", "better", "reflect", "the", "shape", "of", "the", "text", ".", "In", "fact", "we", "can", "restore", "the", "polygon", "by", "only", "scanning", "the", "long", "side", ",", "as", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "Specifically", ",", "we", "firstly", "judge", "whether", "the", "text", "line", "is", "horizontal", "or", "vertical", "through", "the", "regressed", "rectangle", ",", "and", "then", "restore", "the", "polygon", "through", "points", "in", "the", "corresponding", "direction", ".", "Taking", "the", "vertical", "direction", "as", "an", "example", "in", "Fig", ".", "[", "reference", "]", ",", "since", "we", "do", "not", "regress", "the", "intersection", "point", "on", "the", "rectangular", "border", ",", "we", "firstly", "extend", "the", "four", "lines", "near", "the", "border", "to", "find", "four", "intersection", "points", "with", "the", "rectangle", ",", "then", "we", "connect", "four", "new", "points", "and", "other", "intersection", "points", "to", "generate", "polygon", ".", "subsubsection", ":", "Using", "Both", "of", "Horizontal", "and", "Vertical", "Points", "(", "BHVP", ")", "In", "fact", ",", "if", "we", "use", "both", "horizontal", "and", "vertical", "points", "to", "restore", "polygon", ",", "we", "can", "calculate", "a", "polygon", "or", "quadrangle", "that", "passes", "through", "these", "points", "roughly", "as", "shown", "in", "Fig", ".", "[", "reference", "]", "by", "using", "the", "method", "in", ".", "In", "this", "way", "we", "can", "obtain", "dense", "enough", "points", "in", "both", "horizontal", "and", "vertical", "direction", "and", "we", "do", "not", "need", "to", "calculate", "the", "intersection", "with", "the", "rectangle", "as", "in", "PLS", "method", ".", "However", ",", "we", "observe", "BHVP", "is", "not", "as", "effective", "as", "PLS", "for", "the", "polygon", "case", ".", "So", "we", "use", "this", "method", "only", "on", "the", "quadrilateral", "dataset", "(", "ICDAR2015", "Incidental", "Scene", "Text", ")", ".", "subsection", ":", "Polygonal", "non", "-", "maximum", "suppression", "Non", "-", "maximum", "suppression", "(", "NMS", ")", "is", "a", "basic", "method", "commonly", "used", "in", "the", "object", "detection", ",", "and", "its", "purpose", "is", "to", "remove", "duplicate", "boxes", ".", "The", "traditional", "NMS", "method", "is", "based", "on", "rectangular", "boxes", ",", "which", "is", "not", "the", "best", "choice", "for", "other", "shapes", ".", "In", "recent", "years", ",", "other", "NMS", "approaches", "were", "investigated", ",", "e.g.", ",", "locality", "-", "aware", "NMS", ",", "inclined", "NMS", ",", "Mask", "-", "NMS", "and", "polygonal", "NMS", "(", "PNMS", ")", ".", "As", "we", "consider", "the", "polygon", "in", "this", "study", ",", "both", "NMS", "and", "PNMS", "are", "compared", "in", "our", "experiment", ".", "section", ":", "Experiments", "subsection", ":", "Datasets", "subsubsection", ":", "ICDAR2015", "Incidental", "Scene", "Text", ".", "ICDAR2015", "Incidental", "Scene", "Text", "dataset", "is", "commonly", "used", "benchmark", "for", "detecting", "arbitrary", "-", "angle", "quadrangular", "text", "lines", ".", "It", "contains", "1000", "images", "for", "training", ",", "500", "images", "for", "testing", ".", "Some", "words", "which", "are", "too", "short", ",", "or", "unclear", "is", "annotated", "as", "do", "n’t", "cared", "samples", ".", "subsubsection", ":", "CTW1500", "Curve", "text", "dataset", "(", "CTW1500", ")", "is", "constructed", "by", "Yuliang", "et", "al", ".", ".", "Different", "from", "traditional", "text", "datasets", ",", "a", "text", "line", "is", "labelled", "by", "a", "polygon", "with", "14", "points", ".", "subsection", ":", "Implementation", "Details", "Since", "our", "proposed", "SLPR", "can", "be", "applied", "to", "any", "2", "-", "step", "object", "detection", "framework", ".", "We", "adopted", "Faster", "R", "-", "CNN", "in", "ICDAR2015", "Incidental", "Scene", "Text", ".", "And", "because", "also", "proposed", "a", "2", "-", "step", "framework", "based", "on", "R", "-", "FCN", "while", "presenting", "the", "CTW1500", "dataset", ",", "to", "perform", "a", "fair", "comparison", ",", "we", "directly", "used", "the", "network", "in", "from", ".", "All", "experiments", "were", "implemented", "in", "Caffe", "by", "using", "the", "NVIDIA", "GTX", "1080Ti", "GPU", ".", "subsubsection", ":", "ICDAR2015", "Incidental", "Scene", "Text", ".", "For", "Faster", "R", "-", "CNN", "structure", ",", "we", "used", "an", "additional", "anchor", "and", "replaced", "RoIPool", "with", "RoIAlign", "because", "the", "text", "line", "is", "smaller", "than", "other", "objects", ".", "We", "set", "anchor", "scales", "as", "[", "]", "and", "set", "ratios", "as", "[", "0.5", ",", "1", ",", "2", "]", ".", "The", "base", "network", "is", "VGG16", ",", "which", "is", "initialized", "by", "the", "pre", "-", "trained", "model", "on", "ImageNet", "database", ".", "We", "used", "stochastic", "gradient", "descent", "(", "SGD", ")", "with", "back", "-", "propagation", "and", "the", "maximum", "iteration", "was", ".", "Learning", "rates", "started", "from", ",", "decays", "to", "one", "-", "tenth", "every", "iterations", ".", "We", "set", "weight", "decay", "as", ",", "and", "momentum", "as", ".", "We", "used", "1000", "training", "incidental", "images", "in", "ICDAR2015", "Incidental", "Scene", "Text", "and", "the", "229", "training", "images", "from", "ICDAR", "2013", "to", "train", "our", "network", ".", "In", "order", "to", "prevent", "over", "-", "fitting", "we", "employed", "data", "augmentation", ".", "Specifically", ",", "we", "randomly", "resized", "the", "images", "to", "where", "the", "numbers", "represent", "the", "length", "of", "the", "short", "side", ",", "and", "randomly", "rotated", "the", "images", "among", ".", "subsubsection", ":", "CTW1500", "We", "used", "the", "curve", "text", "detector", "(", "CTD", ")", "network", "which", "is", "based", "R", "-", "FCN", "from", ".", "also", "added", "LSTM", "units", "named", "transverse", "and", "longitudinal", "offset", "connection", "(", "TLOC", ")", "to", "learn", "the", "correlation", "of", "points", ".", "But", "we", "removed", "it", ".", "As", "we", "only", "used", "PLS", "to", "restore", "polygon", ",", "Eq", ".", "(", "[", "reference", "]", ")", "was", "modified", "as", ":", "equals", "to", "when", "is", "true", ",", "otherwise", ".", "and", "are", "the", "height", "and", "width", "of", "the", "rectangle", ".", "Because", "most", "of", "the", "texts", "in", "this", "dataset", "are", "horizontal", "text", "line", ",", "to", "solve", "the", "imbalance", "between", "horizontal", "and", "vertical", "samples", ",", "we", "added", "to", "balance", "the", "losses", "between", "them", ".", "And", "when", "is", "close", "to", ",", "the", "text", "line", "may", "be", "judged", "as", "horizontal", "or", "vertical", ",", "so", "we", "set", "as", ".", "The", "base", "network", "is", "ResNet", "-", "50", ",", "which", "is", "initialized", "by", "the", "pre", "-", "trained", "model", "on", "ImageNet", "database", ".", "The", "max", "iteration", "was", ".", "The", "learning", "rate", "in", "this", "experiment", "was", "always", ".", "We", "set", "weight", "decay", "as", ",", "and", "momentum", "as", ".", "To", "conduct", "a", "fair", "comparison", ",", "we", "only", "used", "the", "training", "set", "in", "CTW1500", "to", "train", "our", "network", "and", "did", "not", "use", "data", "augmentation", ".", "subsection", ":", "Results", "subsubsection", ":", "ICDAR2015", "Incidental", "Scene", "Text", "Table", "[", "reference", "]", "shows", "the", "results", "of", "SLPR", "system", "with", "different", "settings", ".", "First", ",", "for", "the", "restoration", "of", "the", "quadrangle", "for", "the", "text", "region", ",", "BHVP", "using", "all", "the", "points", "can", "achieve", "better", "results", "than", "PLS", "using", "only", "the", "long", "-", "side", "points", ".", "Second", ",", "even", "we", "aim", "to", "detect", "the", "quadrangle", "in", "this", "dataset", ",", "PNMS", "still", "outperforms", "NMS", ".", "Finally", ",", "the", "use", "of", "multi", "-", "scale", "is", "one", "way", "to", "improve", "detection", "performance", "on", "different", "target", "sizes", ".", "We", "also", "test", "the", "multi", "-", "scale", "results", "of", "our", "system", "at", "(", "850", ",", "1000", ")", ",", "which", "yields", "about", "1", "%", "absolute", "improvement", "of", "Hmean", "measure", ".", "Fig", ".", "[", "reference", "]", "lists", "several", "challenging", "examples", "of", "detection", "results", "on", "ICDAR2015", "Incidental", "Scene", "Text", "dataset", ".", "Table", "[", "reference", "]", "gives", "the", "comparison", "of", "SLPR", "with", "state", "-", "of", "-", "the", "-", "art", "results", "on", "ICDAR2015", "Incidental", "Scene", "Text", ".", "We", "can", "observe", "that", "our", "method", "achieved", "the", "competitive", "results", "on", "this", "dataset", ".", "subsubsection", ":", "CTW1500", "Table", "[", "reference", "]", "shows", "the", "results", "of", "our", "method", "with", "different", "NMS", "settings", ".", "Different", "from", "the", "observation", "in", "ICDAR2015", "Incidental", "Scene", "Text", ",", "our", "method", "achieved", "the", "best", "result", "on", "NMS0.3", ",", "namely", "the", "traditional", "NMS", "method", "with", "the", "threshold", "0.3", "for", "calculating", "the", "IoU", "(", "Intersection", "-", "over", "-", "Union", ")", ".", "Table", "[", "reference", "]", "lists", "the", "results", "of", "our", "method", "compared", "with", "CTD", "and", "CTD", "+", "TLOC", ".", "We", "removed", "TLOC", "from", "as", "our", "base", "network", "which", "is", "the", "same", "as", "CTD", ".", "Clearly", ",", "the", "Hmean", "performance", "of", "our", "SLPR", "method", "could", "be", "increased", "by", "over", "the", "CTD", "method", ",", "demonstrating", "the", "effectiveness", "of", "our", "simple", "rules", "to", "set", "the", "regression", "points", ".", "Even", "compared", "with", "the", "CTD", "+", "TLOC", "method", "with", "an", "additional", "LSTM", "network", ",", "SLPR", "still", "achieved", "improvement", "of", "Hmean", "performance", ".", "Fig", ".", "[", "reference", "]", "gives", "several", "examples", "of", "the", "detection", "results", "of", "CTD", ",", "CTD", "+", "TLOC", "and", "our", "SLPR", ".", "We", "can", "observe", "that", "our", "method", "generated", "smoother", "regions", "and", "better", "detection", "results", "compared", "with", "CTD", ",", "which", "implied", "that", "the", "proposed", "SLPR", "can", "better", "handle", "the", "arbitrary", "-", "oriented", "case", "due", "to", "the", "novel", "design", "of", "the", "horizontally", "and", "vertically", "symmetrical", "scanning", "using", "sliding", "lines", ".", "section", ":", "Conclusion", "In", "this", "study", ",", "we", "propose", "a", "novel", "SLPR", "method", "for", "the", "text", "detection", "in", "arbitrary", "-", "shape", "case", ".", "Compared", "with", "the", "curve", "text", "detection", "method", "CTD", "+", "TLOC", ",", "SLPR", "is", "more", "concise", "without", "using", "LSTM", "and", "obtains", "better", "performance", ".", "In", "the", "traditional", "quadrangle", "dataset", "(", "ICDAR2015", "Incidental", "Scene", "Text", ")", ",", "SLPR", "also", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "performance", ".", "section", ":", "Acknowledgment", "This", "work", "was", "supported", "in", "part", "by", "the", "National", "Key", "R", "&", "D", "Program", "of", "China", "under", "contract", "No", ".", "2017YFB1002202", ",", "in", "part", "by", "the", "National", "Natural", "Science", "Foundation", "of", "China", "under", "Grants", "61671422", "and", "U1613211", ",", "in", "part", "by", "the", "MOE", "-", "Microsoft", "Key", "Laboratory", "of", "USTC", ".", "bibliography", ":", "References"], "entity": [[2, 6, ["Sliding", "Line", "Point", "Regression"], "Method"], [7, 12, ["Shape", "Robust", "Scene", "Text", "Detection"], "Task"], [13, 16, ["text", "detection", "methods"], "Method"], [31, 35, ["sliding", "line", "point", "regression"], "Method"], [42, 46, ["arbitrary", "-", "shape", "text"], "Task"], [50, 51, ["SLPR"], "Method"], [75, 76, ["SLPR"], "Method"], [86, 90, ["Faster", "R", "-", "CNN"], "Method"], [91, 94, ["R", "-", "FCN"], "Method"], [108, 111, ["region", "proposal", "network"], "Method"], [112, 113, ["RPN"], "Method"], [205, 210, ["ICDAR2015", "Incidental", "Scene", "Text", "benchmark"], "Material"], [211, 216, ["curve", "text", "detection", "dataset", "CTW1500"], "Material"], [220, 222, ["Text", "detection"], "Task"], [239, 242, ["digitization", "of", "text"], "Task"], [243, 245, ["text", "translation"], "Task"], [255, 258, ["scene", "text", "detection"], "Task"], [276, 280, ["Faster", "R", "-", "CNN"], "Method"], [281, 282, ["SSD"], "Method"], [367, 371, ["curve", "text", "line", "detection"], "Task"], [378, 381, ["arbitrary", "-", "oriented"], "Task"], [382, 385, ["curve", "text", "detection"], "Task"], [389, 393, ["sliding", "line", "point", "regression"], "Method"], [394, 395, ["SLPR"], "Method"], [406, 414, ["Faster", "R", "-", "CNN", "or", "R", "-", "FCN"], "Method"], [423, 426, ["region", "proposal", "network"], "Method"], [427, 428, ["RPN"], "Method"], [479, 480, ["RNN"], "Method"], [593, 596, ["R", "-", "CNN"], "Method"], [597, 600, ["R", "-", "FCN"], "Method"], [606, 609, ["sliding", "line", "method"], "Method"], [617, 618, ["regression"], "Task"], [644, 649, ["scene", "text", "detection", "and", "recognition"], "Task"], [657, 660, ["scene", "text", "detection"], "Task"], [682, 685, ["character", "based", "methods"], "Method"], [686, 689, ["word", "based", "methods"], "Method"], [690, 693, ["segmentation", "based", "methods"], "Method"], [694, 697, ["Character", "based", "methods"], "Method"], [749, 754, ["ICDAR2015", "Incidental", "Scene", "Text", "benchmark"], "Material"], [763, 767, ["semi", "-", "supervised", "method"], "Method"], [768, 770, ["finetune", "model"], "Method"], [779, 782, ["segmentation", "based", "methods"], "Method"], [787, 789, ["text", "detection"], "Task"], [793, 796, ["fully", "convolutional", "network"], "Method"], [797, 798, ["FCN"], "Method"], [819, 821, ["character", "components"], "Method"], [862, 864, ["segmentation", "methods"], "Method"], [871, 876, ["ICDAR2015", "Incidental", "Scene", "Text", "benchmark"], "Material"], [880, 882, ["object", "detection"], "Task"], [886, 888, ["text", "detection"], "Task"], [892, 895, ["R", "-", "CNN"], "Method"], [896, 897, ["SSD"], "Method"], [898, 901, ["R", "-", "FCN"], "Method"], [902, 903, ["YOLO"], "Method"], [905, 908, ["irregular", "convolutional", "filters"], "Method"], [912, 914, ["convolutional", "filters"], "Method"], [921, 924, ["long", "text", "detection"], "Task"], [927, 929, ["attention", "map"], "Method"], [940, 944, ["2", "-", "step", "methods"], "Method"], [947, 954, ["R", "-", "CNN", "or", "R", "-", "FCN"], "Method"], [971, 976, ["multi", "-", "scale", "pool", "operations"], "Method"], [978, 980, ["roipool", "layer"], "Method"], [1033, 1034, ["CTW1500"], "Material"], [1042, 1045, ["curve", "text", "detector"], "Method"], [1046, 1047, ["CTD"], "Method"], [1073, 1077, ["Faster", "R", "-", "CNN"], "Method"], [1078, 1081, ["R", "-", "FCN"], "Method"], [1111, 1114, ["R", "-", "CNN"], "Method"], [1126, 1127, ["RPN"], "Method"], [1198, 1200, ["neural", "net"], "Method"], [1277, 1279, ["neural", "net"], "Method"], [1440, 1442, ["computational", "complexity"], "Metric"], [1470, 1471, ["accuracy"], "Metric"], [1509, 1511, ["network", "complexity"], "Metric"], [1545, 1549, ["multi", "-", "task", "learning"], "Task"], [1568, 1572, ["multi", "-", "task", "learning"], "Task"], [1587, 1591, ["region", "proposal", "classification", "loss"], "Method"], [1593, 1596, ["box", "regression", "loss"], "Method"], [1605, 1606, ["RPN"], "Method"], [1616, 1618, ["classification", "loss"], "Metric"], [1619, 1622, ["box", "regression", "loss"], "Method"], [1648, 1649, ["SLPR"], "Method"], [1652, 1655, ["smooth", "L1", "loss"], "Method"], [1657, 1660, ["box", "regression", "task"], "Task"], [1891, 1894, ["Restoration", "of", "polygon"], "Task"], [1909, 1911, ["neural", "nets"], "Method"], [1929, 1930, ["subsubsection"], "Method"], [2080, 2081, ["subsubsection"], "Method"], [2090, 2091, ["BHVP"], "Method"], [2163, 2165, ["PLS", "method"], "Method"], [2170, 2171, ["BHVP"], "Method"], [2176, 2177, ["PLS"], "Method"], [2179, 2181, ["polygon", "case"], "Task"], [2201, 2206, ["Polygonal", "non", "-", "maximum", "suppression"], "Method"], [2206, 2210, ["Non", "-", "maximum", "suppression"], "Method"], [2211, 2212, ["NMS"], "Method"], [2221, 2223, ["object", "detection"], "Task"], [2280, 2281, ["PNMS"], "Method"], [2293, 2294, ["NMS"], "Method"], [2295, 2296, ["PNMS"], "Method"], [2315, 2316, ["ICDAR2015"], "Material"], [2316, 2320, ["Incidental", "Scene", "Text", "dataset"], "Material"], [2325, 2332, ["detecting", "arbitrary", "-", "angle", "quadrangular", "text", "lines"], "Task"], [2364, 2365, ["CTW1500"], "Material"], [2365, 2368, ["Curve", "text", "dataset"], "Material"], [2369, 2370, ["CTW1500"], "Material"], [2382, 2384, ["text", "datasets"], "Material"], [2386, 2388, ["text", "line"], "Material"], [2404, 2405, ["SLPR"], "Method"], [2420, 2423, ["R", "-", "CNN"], "Method"], [2434, 2438, ["2", "-", "step", "framework"], "Method"], [2440, 2443, ["R", "-", "FCN"], "Method"], [2468, 2469, ["Caffe"], "Method"], [2472, 2476, ["NVIDIA", "GTX", "1080Ti", "GPU"], "Method"], [2486, 2490, ["R", "-", "CNN", "structure"], "Task"], [2534, 2535, ["VGG16"], "Method"], [2541, 2545, ["pre", "-", "trained", "model"], "Method"], [2546, 2548, ["ImageNet", "database"], "Material"], [2551, 2554, ["stochastic", "gradient", "descent"], "Method"], [2555, 2556, ["SGD"], "Method"], [2558, 2561, ["back", "-", "propagation"], "Method"], [2567, 2569, ["Learning", "rates"], "Metric"], [2607, 2609, ["ICDAR", "2013"], "Material"], [2618, 2621, ["over", "-", "fitting"], "Task"], [2623, 2625, ["data", "augmentation"], "Method"], [2654, 2655, ["CTW1500"], "Material"], [2658, 2661, ["curve", "text", "detector"], "Method"], [2662, 2663, ["CTD"], "Method"], [2668, 2671, ["R", "-", "FCN"], "Method"], [2675, 2677, ["LSTM", "units"], "Method"], [2684, 2685, ["TLOC"], "Method"], [2794, 2797, ["ResNet", "-", "50"], "Method"], [2803, 2807, ["pre", "-", "trained", "model"], "Method"], [2808, 2810, ["ImageNet", "database"], "Material"], [2817, 2819, ["learning", "rate"], "Metric"], [2848, 2849, ["CTW1500"], "Material"], [2857, 2859, ["data", "augmentation"], "Method"], [2887, 2891, ["restoration", "of", "the", "quadrangle"], "Task"], [2896, 2897, ["BHVP"], "Method"], [2906, 2907, ["PLS"], "Method"], [2928, 2929, ["PNMS"], "Method"], [2931, 2932, ["NMS"], "Method"], [2938, 2941, ["multi", "-", "scale"], "Method"], [2946, 2947, ["detection"], "Task"], [2979, 2981, ["Hmean", "measure"], "Metric"], [2992, 2993, ["detection"], "Task"], [3009, 3010, ["SLPR"], "Method"], [3041, 3042, ["CTW1500"], "Material"], [3074, 3075, ["NMS0.3"], "Method"], [3088, 3089, ["IoU"], "Metric"], [3090, 3095, ["Intersection", "-", "over", "-", "Union"], "Metric"], [3109, 3110, ["CTD"], "Method"], [3128, 3129, ["CTD"], "Method"], [3133, 3134, ["Hmean"], "Metric"], [3172, 3174, ["LSTM", "network"], "Method"], [3175, 3176, ["SLPR"], "Method"], [3180, 3182, ["Hmean", "performance"], "Metric"], [3193, 3194, ["detection"], "Task"], [3196, 3197, ["CTD"], "Method"], [3203, 3204, ["SLPR"], "Method"], [3216, 3217, ["detection"], "Task"], [3220, 3221, ["CTD"], "Method"], [3227, 3228, ["SLPR"], "Method"], [3267, 3269, ["text", "detection"], "Task"], [3270, 3274, ["arbitrary", "-", "shape", "case"], "Task"], [3286, 3287, ["SLPR"], "Method"], [3292, 3293, ["LSTM"], "Method"], [3310, 3311, ["SLPR"], "Method"], [36, 37, ["SLPR"], "Method"], [81, 83, ["object", "detection"], "Task"], [289, 292, ["arbitrary", "-", "oriented"], "Task"], [402, 404, ["object", "detection"], "Task"], [583, 586, ["arbitrary", "-", "oriented"], "Task"], [587, 590, ["curve", "text", "detection"], "Task"], [1004, 1007, ["arbitrary", "-", "oriented"], "Task"], [1050, 1053, ["curve", "text", "detection"], "Task"], [1068, 1070, ["object", "detection"], "Task"], [1897, 1898, ["SLPR"], "Method"], [2193, 2194, ["ICDAR2015"], "Material"], [2235, 2236, ["NMS"], "Method"], [2258, 2259, ["NMS"], "Method"], [2268, 2269, ["NMS"], "Method"], [2271, 2272, ["NMS"], "Method"], [2275, 2276, ["NMS"], "Method"], [2278, 2279, ["NMS"], "Method"], [2310, 2311, ["ICDAR2015"], "Material"], [2413, 2415, ["object", "detection"], "Task"], [2424, 2425, ["ICDAR2015"], "Material"], [2446, 2447, ["CTW1500"], "Material"], [2479, 2480, ["ICDAR2015"], "Material"], [2597, 2598, ["ICDAR2015"], "Material"], [2865, 2866, ["ICDAR2015"], "Material"], [2877, 2878, ["SLPR"], "Method"], [2995, 2996, ["ICDAR2015"], "Material"], [2996, 3000, ["Incidental", "Scene", "Text", "dataset"], "Material"], [3020, 3021, ["ICDAR2015"], "Material"], [3054, 3055, ["NMS"], "Method"], [3062, 3063, ["ICDAR2015"], "Material"], [3079, 3080, ["NMS"], "Method"], [3111, 3112, ["CTD"], "Method"], [3137, 3138, ["SLPR"], "Method"], [3145, 3146, ["CTD"], "Method"], [3165, 3166, ["CTD"], "Method"], [3198, 3199, ["CTD"], "Method"], [3232, 3235, ["arbitrary", "-", "oriented"], "Task"], [3263, 3264, ["SLPR"], "Method"], [3278, 3281, ["curve", "text", "detection"], "Task"], [3282, 3283, ["CTD"], "Method"], [3304, 3305, ["ICDAR2015"], "Material"]]} {"doc_id": "35756f711a97166df11202ebe46820a36704ae77", "text": ["document", ":", "Unsupervised", "Representation", "Learning", "with", "Deep", "Convolutional", "Generative", "Adversarial", "Networks", "In", "recent", "years", ",", "supervised", "learning", "with", "convolutional", "networks", "(", "CNNs", ")", "has", "seen", "huge", "adoption", "in", "computer", "vision", "applications", ".", "Comparatively", ",", "unsupervised", "learning", "with", "CNNs", "has", "received", "less", "attention", ".", "In", "this", "work", "we", "hope", "to", "help", "bridge", "the", "gap", "between", "the", "success", "of", "CNNs", "for", "supervised", "learning", "and", "unsupervised", "learning", ".", "We", "introduce", "a", "class", "of", "CNNs", "called", "deep", "convolutional", "generative", "adversarial", "networks", "(", "DCGANs", ")", ",", "that", "have", "certain", "architectural", "constraints", ",", "and", "demonstrate", "that", "they", "are", "a", "strong", "candidate", "for", "unsupervised", "learning", ".", "Training", "on", "various", "image", "datasets", ",", "we", "show", "convincing", "evidence", "that", "our", "deep", "convolutional", "adversarial", "pair", "learns", "a", "hierarchy", "of", "representations", "from", "object", "parts", "to", "scenes", "in", "both", "the", "generator", "and", "discriminator", ".", "Additionally", ",", "we", "use", "the", "learned", "features", "for", "novel", "tasks", "-", "demonstrating", "their", "applicability", "as", "general", "image", "representations", ".", ".", "/", "section", ":", "Introduction", "Learning", "reusable", "feature", "representations", "from", "large", "unlabeled", "datasets", "has", "been", "an", "area", "of", "active", "research", ".", "In", "the", "context", "of", "computer", "vision", ",", "one", "can", "leverage", "the", "practically", "unlimited", "amount", "of", "unlabeled", "images", "and", "videos", "to", "learn", "good", "intermediate", "representations", ",", "which", "can", "then", "be", "used", "on", "a", "variety", "of", "supervised", "learning", "tasks", "such", "as", "image", "classification", ".", "We", "propose", "that", "one", "way", "to", "build", "good", "image", "representations", "is", "by", "training", "Generative", "Adversarial", "Networks", "(", "GANs", ")", "Goodfellow2014", ",", "and", "later", "reusing", "parts", "of", "the", "generator", "and", "discriminator", "networks", "as", "feature", "extractors", "for", "supervised", "tasks", ".", "GANs", "provide", "an", "attractive", "alternative", "to", "maximum", "likelihood", "techniques", ".", "One", "can", "additionally", "argue", "that", "their", "learning", "process", "and", "the", "lack", "of", "a", "heuristic", "cost", "function", "(", "such", "as", "pixel", "-", "wise", "independent", "mean", "-", "square", "error", ")", "are", "attractive", "to", "representation", "learning", ".", "GANs", "have", "been", "known", "to", "be", "unstable", "to", "train", ",", "often", "resulting", "in", "generators", "that", "produce", "nonsensical", "outputs", ".", "There", "has", "been", "very", "limited", "published", "research", "in", "trying", "to", "understand", "and", "visualize", "what", "GANs", "learn", ",", "and", "the", "intermediate", "representations", "of", "multi", "-", "layer", "GANs", ".", "In", "this", "paper", ",", "we", "make", "the", "following", "contributions", "We", "propose", "and", "evaluate", "a", "set", "of", "constraints", "on", "the", "architectural", "topology", "of", "Convolutional", "GANs", "that", "make", "them", "stable", "to", "train", "in", "most", "settings", ".", "We", "name", "this", "class", "of", "architectures", "Deep", "Convolutional", "GANs", "(", "DCGAN", ")", "We", "use", "the", "trained", "discriminators", "for", "image", "classification", "tasks", ",", "showing", "competitive", "performance", "with", "other", "unsupervised", "algorithms", ".", "We", "visualize", "the", "filters", "learnt", "by", "GANs", "and", "empirically", "show", "that", "specific", "filters", "have", "learned", "to", "draw", "specific", "objects", ".", "We", "show", "that", "the", "generators", "have", "interesting", "vector", "arithmetic", "properties", "allowing", "for", "easy", "manipulation", "of", "many", "semantic", "qualities", "of", "generated", "samples", ".", "section", ":", "Related", "Work", "subsection", ":", "Representation", "Learning", "from", "unlabeled", "data", "Unsupervised", "representation", "learning", "is", "a", "fairly", "well", "studied", "problem", "in", "general", "computer", "vision", "research", ",", "as", "well", "as", "in", "the", "context", "of", "images", ".", "A", "classic", "approach", "to", "unsupervised", "representation", "learning", "is", "to", "do", "clustering", "on", "the", "data", "(", "for", "example", "using", "K", "-", "means", ")", ",", "and", "leverage", "the", "clusters", "for", "improved", "classification", "scores", ".", "In", "the", "context", "of", "images", ",", "one", "can", "do", "hierarchical", "clustering", "of", "image", "patches", "coates2012learning", "to", "learn", "powerful", "image", "representations", ".", "Another", "popular", "method", "is", "to", "train", "auto", "-", "encoders", "(", "convolutionally", ",", "stacked", "vincent2010stacked", ",", "separating", "the", "what", "and", "where", "components", "of", "the", "code", "zhao2015stacked", ",", "ladder", "structures", "rasmus2015semi", ")", "that", "encode", "an", "image", "into", "a", "compact", "code", ",", "and", "decode", "the", "code", "to", "reconstruct", "the", "image", "as", "accurately", "as", "possible", ".", "These", "methods", "have", "also", "been", "shown", "to", "learn", "good", "feature", "representations", "from", "image", "pixels", ".", "Deep", "belief", "networks", "lee2009convolutional", "have", "also", "been", "shown", "to", "work", "well", "in", "learning", "hierarchical", "representations", ".", "subsection", ":", "Generating", "natural", "images", "Generative", "image", "models", "are", "well", "studied", "and", "fall", "into", "two", "categories", ":", "parametric", "and", "non", "-", "parametric", ".", "The", "non", "-", "parametric", "models", "often", "do", "matching", "from", "a", "database", "of", "existing", "images", ",", "often", "matching", "patches", "of", "images", ",", "and", "have", "been", "used", "in", "texture", "synthesis", "efros1999texture", ",", "super", "-", "resolution", "freeman2002example", "and", "in", "-", "painting", "hays2007scene", ".", "Parametric", "models", "for", "generating", "images", "has", "been", "explored", "extensively", "(", "for", "example", "on", "MNIST", "digits", "or", "for", "texture", "synthesis", "portilla2000parametric", ")", ".", "However", ",", "generating", "natural", "images", "of", "the", "real", "world", "have", "had", "not", "much", "success", "until", "recently", ".", "A", "variational", "sampling", "approach", "to", "generating", "images", "kingma2013auto", "has", "had", "some", "success", ",", "but", "the", "samples", "often", "suffer", "from", "being", "blurry", ".", "Another", "approach", "generates", "images", "using", "an", "iterative", "forward", "diffusion", "process", "sohl2015deep", ".", "Generative", "Adversarial", "Networks", "Goodfellow2014", "generated", "images", "suffering", "from", "being", "noisy", "and", "incomprehensible", ".", "A", "laplacian", "pyramid", "extension", "to", "this", "approach", "denton2015deep", "showed", "higher", "quality", "images", ",", "but", "they", "still", "suffered", "from", "the", "objects", "looking", "wobbly", "because", "of", "noise", "introduced", "in", "chaining", "multiple", "models", ".", "A", "recurrent", "network", "approach", "gregor2015draw", "and", "a", "deconvolution", "network", "approach", "dosovitskiy2014learning", "have", "also", "recently", "had", "some", "success", "with", "generating", "natural", "images", ".", "However", ",", "they", "have", "not", "leveraged", "the", "generators", "for", "supervised", "tasks", ".", "subsection", ":", "Visualizing", "the", "internals", "of", "CNNs", "One", "constant", "criticism", "of", "using", "neural", "networks", "has", "been", "that", "they", "are", "black", "-", "box", "methods", ",", "with", "little", "understanding", "of", "what", "the", "networks", "do", "in", "the", "form", "of", "a", "simple", "human", "-", "consumable", "algorithm", ".", "In", "the", "context", "of", "CNNs", ",", "Zeiler", "et", ".", "al", ".", "zeiler2014visualizing", "showed", "that", "by", "using", "deconvolutions", "and", "filtering", "the", "maximal", "activations", ",", "one", "can", "find", "the", "approximate", "purpose", "of", "each", "convolution", "filter", "in", "the", "network", ".", "Similarly", ",", "using", "a", "gradient", "descent", "on", "the", "inputs", "lets", "us", "inspect", "the", "ideal", "image", "that", "activates", "certain", "subsets", "of", "filters", "Inceptionism2015", ".", "section", ":", "Approach", "and", "Model", "Architecture", "Historical", "attempts", "to", "scale", "up", "GANs", "using", "CNNs", "to", "model", "images", "have", "been", "unsuccessful", ".", "This", "motivated", "the", "authors", "of", "LAPGAN", "denton2015deep", "to", "develop", "an", "alternative", "approach", "to", "iteratively", "upscale", "low", "resolution", "generated", "images", "which", "can", "be", "modeled", "more", "reliably", ".", "We", "also", "encountered", "difficulties", "attempting", "to", "scale", "GANs", "using", "CNN", "architectures", "commonly", "used", "in", "the", "supervised", "literature", ".", "However", ",", "after", "extensive", "model", "exploration", "we", "identified", "a", "family", "of", "architectures", "that", "resulted", "in", "stable", "training", "across", "a", "range", "of", "datasets", "and", "allowed", "for", "training", "higher", "resolution", "and", "deeper", "generative", "models", ".", "Core", "to", "our", "approach", "is", "adopting", "and", "modifying", "three", "recently", "demonstrated", "changes", "to", "CNN", "architectures", ".", "The", "first", "is", "the", "all", "convolutional", "net", "springenberg2014striving", "which", "replaces", "deterministic", "spatial", "pooling", "functions", "(", "such", "as", "maxpooling", ")", "with", "strided", "convolutions", ",", "allowing", "the", "network", "to", "learn", "its", "own", "spatial", "downsampling", ".", "We", "use", "this", "approach", "in", "our", "generator", ",", "allowing", "it", "to", "learn", "its", "own", "spatial", "upsampling", ",", "and", "discriminator", ".", "Second", "is", "the", "trend", "towards", "eliminating", "fully", "connected", "layers", "on", "top", "of", "convolutional", "features", ".", "The", "strongest", "example", "of", "this", "is", "global", "average", "pooling", "which", "has", "been", "utilized", "in", "state", "of", "the", "art", "image", "classification", "models", "Inceptionism2015", ".", "We", "found", "global", "average", "pooling", "increased", "model", "stability", "but", "hurt", "convergence", "speed", ".", "A", "middle", "ground", "of", "directly", "connecting", "the", "highest", "convolutional", "features", "to", "the", "input", "and", "output", "respectively", "of", "the", "generator", "and", "discriminator", "worked", "well", ".", "The", "first", "layer", "of", "the", "GAN", ",", "which", "takes", "a", "uniform", "noise", "distribution", "as", "input", ",", "could", "be", "called", "fully", "connected", "as", "it", "is", "just", "a", "matrix", "multiplication", ",", "but", "the", "result", "is", "reshaped", "into", "a", "4", "-", "dimensional", "tensor", "and", "used", "as", "the", "start", "of", "the", "convolution", "stack", ".", "For", "the", "discriminator", ",", "the", "last", "convolution", "layer", "is", "flattened", "and", "then", "fed", "into", "a", "single", "sigmoid", "output", ".", "See", "Fig", ".", "[", "reference", "]", "for", "a", "visualization", "of", "an", "example", "model", "architecture", ".", "Third", "is", "Batch", "Normalization", "ioffe2015batch", "which", "stabilizes", "learning", "by", "normalizing", "the", "input", "to", "each", "unit", "to", "have", "zero", "mean", "and", "unit", "variance", ".", "This", "helps", "deal", "with", "training", "problems", "that", "arise", "due", "to", "poor", "initialization", "and", "helps", "gradient", "flow", "in", "deeper", "models", ".", "This", "proved", "critical", "to", "get", "deep", "generators", "to", "begin", "learning", ",", "preventing", "the", "generator", "from", "collapsing", "all", "samples", "to", "a", "single", "point", "which", "is", "a", "common", "failure", "mode", "observed", "in", "GANs", ".", "Directly", "applying", "batchnorm", "to", "all", "layers", "however", ",", "resulted", "in", "sample", "oscillation", "and", "model", "instability", ".", "This", "was", "avoided", "by", "not", "applying", "batchnorm", "to", "the", "generator", "output", "layer", "and", "the", "discriminator", "input", "layer", ".", "The", "ReLU", "activation", "nair2010rectified", "is", "used", "in", "the", "generator", "with", "the", "exception", "of", "the", "output", "layer", "which", "uses", "the", "Tanh", "function", ".", "We", "observed", "that", "using", "a", "bounded", "activation", "allowed", "the", "model", "to", "learn", "more", "quickly", "to", "saturate", "and", "cover", "the", "color", "space", "of", "the", "training", "distribution", ".", "Within", "the", "discriminator", "we", "found", "the", "leaky", "rectified", "activation", "maas2013rectifier", "xu2015empirical", "to", "work", "well", ",", "especially", "for", "higher", "resolution", "modeling", ".", "This", "is", "in", "contrast", "to", "the", "original", "GAN", "paper", ",", "which", "used", "the", "maxout", "activation", "goodfellow2013maxout", ".", "Architecture", "guidelines", "for", "stable", "Deep", "Convolutional", "GANs", "Replace", "any", "pooling", "layers", "with", "strided", "convolutions", "(", "discriminator", ")", "and", "fractional", "-", "strided", "convolutions", "(", "generator", ")", ".", "Use", "batchnorm", "in", "both", "the", "generator", "and", "the", "discriminator", ".", "Remove", "fully", "connected", "hidden", "layers", "for", "deeper", "architectures", ".", "Use", "ReLU", "activation", "in", "generator", "for", "all", "layers", "except", "for", "the", "output", ",", "which", "uses", "Tanh", ".", "Use", "LeakyReLU", "activation", "in", "the", "discriminator", "for", "all", "layers", ".", "section", ":", "Details", "of", "adversarial", "training", "We", "trained", "DCGANs", "on", "three", "datasets", ",", "Large", "-", "scale", "Scene", "Understanding", "(", "LSUN", ")", "yu2015construction", ",", "Imagenet", "-", "1k", "and", "a", "newly", "assembled", "Faces", "dataset", ".", "Details", "on", "the", "usage", "of", "each", "of", "these", "datasets", "are", "given", "below", ".", "No", "pre", "-", "processing", "was", "applied", "to", "training", "images", "besides", "scaling", "to", "the", "range", "of", "the", "tanh", "activation", "function", "[", "-", "1", ",", "1", "]", ".", "All", "models", "were", "trained", "with", "mini", "-", "batch", "stochastic", "gradient", "descent", "(", "SGD", ")", "with", "a", "mini", "-", "batch", "size", "of", "128", ".", "All", "weights", "were", "initialized", "from", "a", "zero", "-", "centered", "Normal", "distribution", "with", "standard", "deviation", "0.02", ".", "In", "the", "LeakyReLU", ",", "the", "slope", "of", "the", "leak", "was", "set", "to", "0.2", "in", "all", "models", ".", "While", "previous", "GAN", "work", "has", "used", "momentum", "to", "accelerate", "training", ",", "we", "used", "the", "Adam", "optimizer", "kingma2014adam", "with", "tuned", "hyperparameters", ".", "We", "found", "the", "suggested", "learning", "rate", "of", "0.001", ",", "to", "be", "too", "high", ",", "using", "0.0002", "instead", ".", "Additionally", ",", "we", "found", "leaving", "the", "momentum", "term", "at", "the", "suggested", "value", "of", "0.9", "resulted", "in", "training", "oscillation", "and", "instability", "while", "reducing", "it", "to", "0.5", "helped", "stabilize", "training", ".", "subsection", ":", "LSUN", "As", "visual", "quality", "of", "samples", "from", "generative", "image", "models", "has", "improved", ",", "concerns", "of", "over", "-", "fitting", "and", "memorization", "of", "training", "samples", "have", "risen", ".", "To", "demonstrate", "how", "our", "model", "scales", "with", "more", "data", "and", "higher", "resolution", "generation", ",", "we", "train", "a", "model", "on", "the", "LSUN", "bedrooms", "dataset", "containing", "a", "little", "over", "3", "million", "training", "examples", ".", "Recent", "analysis", "has", "shown", "that", "there", "is", "a", "direct", "link", "between", "how", "fast", "models", "learn", "and", "their", "generalization", "performance", "hardt2015train", ".", "We", "show", "samples", "from", "one", "epoch", "of", "training", "(", "Fig", ".", "[", "reference", "]", ")", ",", "mimicking", "online", "learning", ",", "in", "addition", "to", "samples", "after", "convergence", "(", "Fig", ".", "[", "reference", "]", ")", ",", "as", "an", "opportunity", "to", "demonstrate", "that", "our", "model", "is", "not", "producing", "high", "quality", "samples", "via", "simply", "overfitting", "/", "memorizing", "training", "examples", ".", "No", "data", "augmentation", "was", "applied", "to", "the", "images", ".", "subsubsection", ":", "Deduplication", "To", "further", "decrease", "the", "likelihood", "of", "the", "generator", "memorizing", "input", "examples", "(", "Fig", ".", "[", "reference", "]", ")", "we", "perform", "a", "simple", "image", "de", "-", "duplication", "process", ".", "We", "fit", "a", "3072", "-", "128", "-", "3072", "de", "-", "noising", "dropout", "regularized", "RELU", "autoencoder", "on", "32x32", "downsampled", "center", "-", "crops", "of", "training", "examples", ".", "The", "resulting", "code", "layer", "activations", "are", "then", "binarized", "via", "thresholding", "the", "ReLU", "activation", "which", "has", "been", "shown", "to", "be", "an", "effective", "information", "preserving", "technique", "srivastava2014understanding", "and", "provides", "a", "convenient", "form", "of", "semantic", "-", "hashing", ",", "allowing", "for", "linear", "time", "de", "-", "duplication", ".", "Visual", "inspection", "of", "hash", "collisions", "showed", "high", "precision", "with", "an", "estimated", "false", "positive", "rate", "of", "less", "than", "1", "in", "100", ".", "Additionally", ",", "the", "technique", "detected", "and", "removed", "approximately", "275", ",", "000", "near", "duplicates", ",", "suggesting", "a", "high", "recall", ".", "subsection", ":", "Faces", "We", "scraped", "images", "containing", "human", "faces", "from", "random", "web", "image", "queries", "of", "peoples", "names", ".", "The", "people", "names", "were", "acquired", "from", "dbpedia", ",", "with", "a", "criterion", "that", "they", "were", "born", "in", "the", "modern", "era", ".", "This", "dataset", "has", "3", "M", "images", "from", "10", "K", "people", ".", "We", "run", "an", "OpenCV", "face", "detector", "on", "these", "images", ",", "keeping", "the", "detections", "that", "are", "sufficiently", "high", "resolution", ",", "which", "gives", "us", "approximately", "350", ",", "000", "face", "boxes", ".", "We", "use", "these", "face", "boxes", "for", "training", ".", "No", "data", "augmentation", "was", "applied", "to", "the", "images", ".", "subsection", ":", "Imagenet", "-", "1k", "We", "use", "Imagenet", "-", "1k", "deng2009imagenet", "as", "a", "source", "of", "natural", "images", "for", "unsupervised", "training", ".", "We", "train", "on", "min", "-", "resized", "center", "crops", ".", "No", "data", "augmentation", "was", "applied", "to", "the", "images", ".", "section", ":", "Empirical", "Validation", "of", "DCGANs", "capabilities", "subsection", ":", "Classifying", "CIFAR", "-", "10", "using", "GANs", "as", "a", "feature", "extractor", "One", "common", "technique", "for", "evaluating", "the", "quality", "of", "unsupervised", "representation", "learning", "algorithms", "is", "to", "apply", "them", "as", "a", "feature", "extractor", "on", "supervised", "datasets", "and", "evaluate", "the", "performance", "of", "linear", "models", "fitted", "on", "top", "of", "these", "features", ".", "On", "the", "CIFAR", "-", "10", "dataset", ",", "a", "very", "strong", "baseline", "performance", "has", "been", "demonstrated", "from", "a", "well", "tuned", "single", "layer", "feature", "extraction", "pipeline", "utilizing", "K", "-", "means", "as", "a", "feature", "learning", "algorithm", ".", "When", "using", "a", "very", "large", "amount", "of", "feature", "maps", "(", "4800", ")", "this", "technique", "achieves", "80.6", "%", "accuracy", ".", "An", "unsupervised", "multi", "-", "layered", "extension", "of", "the", "base", "algorithm", "reaches", "82.0", "%", "accuracy", "coates2011selecting", ".", "To", "evaluate", "the", "quality", "of", "the", "representations", "learned", "by", "DCGANs", "for", "supervised", "tasks", ",", "we", "train", "on", "Imagenet", "-", "1k", "and", "then", "use", "the", "discriminator", "’s", "convolutional", "features", "from", "all", "layers", ",", "maxpooling", "each", "layers", "representation", "to", "produce", "a", "spatial", "grid", ".", "These", "features", "are", "then", "flattened", "and", "concatenated", "to", "form", "a", "28672", "dimensional", "vector", "and", "a", "regularized", "linear", "L2", "-", "SVM", "classifier", "is", "trained", "on", "top", "of", "them", ".", "This", "achieves", "82.8", "%", "accuracy", ",", "out", "performing", "all", "K", "-", "means", "based", "approaches", ".", "Notably", ",", "the", "discriminator", "has", "many", "less", "feature", "maps", "(", "512", "in", "the", "highest", "layer", ")", "compared", "to", "K", "-", "means", "based", "techniques", ",", "but", "does", "result", "in", "a", "larger", "total", "feature", "vector", "size", "due", "to", "the", "many", "layers", "of", "spatial", "locations", ".", "The", "performance", "of", "DCGANs", "is", "still", "less", "than", "that", "of", "Exemplar", "CNNs", "dosovitskiy2014discriminative", ",", "a", "technique", "which", "trains", "normal", "discriminative", "CNNs", "in", "an", "unsupervised", "fashion", "to", "differentiate", "between", "specifically", "chosen", ",", "aggressively", "augmented", ",", "exemplar", "samples", "from", "the", "source", "dataset", ".", "Further", "improvements", "could", "be", "made", "by", "finetuning", "the", "discriminator", "’s", "representations", ",", "but", "we", "leave", "this", "for", "future", "work", ".", "Additionally", ",", "since", "our", "DCGAN", "was", "never", "trained", "on", "CIFAR", "-", "10", "this", "experiment", "also", "demonstrates", "the", "domain", "robustness", "of", "the", "learned", "features", ".", "subsection", ":", "Classifying", "SVHN", "digits", "using", "GANs", "as", "a", "feature", "extractor", "On", "the", "StreetView", "House", "Numbers", "dataset", "(", "SVHN", ")", "netzer2011reading", ",", "we", "use", "the", "features", "of", "the", "discriminator", "of", "a", "DCGAN", "for", "supervised", "purposes", "when", "labeled", "data", "is", "scarce", ".", "Following", "similar", "dataset", "preparation", "rules", "as", "in", "the", "CIFAR", "-", "10", "experiments", ",", "we", "split", "off", "a", "validation", "set", "of", "10", ",", "000", "examples", "from", "the", "non", "-", "extra", "set", "and", "use", "it", "for", "all", "hyperparameter", "and", "model", "selection", ".", "1000", "uniformly", "class", "distributed", "training", "examples", "are", "randomly", "selected", "and", "used", "to", "train", "a", "regularized", "linear", "L2", "-", "SVM", "classifier", "on", "top", "of", "the", "same", "feature", "extraction", "pipeline", "used", "for", "CIFAR", "-", "10", ".", "This", "achieves", "state", "of", "the", "art", "(", "for", "classification", "using", "1000", "labels", ")", "at", "22.48", "%", "test", "error", ",", "improving", "upon", "another", "modifcation", "of", "CNNs", "designed", "to", "leverage", "unlabled", "data", "zhao2015stacked", ".", "Additionally", ",", "we", "validate", "that", "the", "CNN", "architecture", "used", "in", "DCGAN", "is", "not", "the", "key", "contributing", "factor", "of", "the", "model", "’s", "performance", "by", "training", "a", "purely", "supervised", "CNN", "with", "the", "same", "architecture", "on", "the", "same", "data", "and", "optimizing", "this", "model", "via", "random", "search", "over", "64", "hyperparameter", "trials", "bergstra2012hpopt", ".", "It", "achieves", "a", "signficantly", "higher", "28.87", "%", "validation", "error", ".", "section", ":", "Investigating", "and", "visualizing", "the", "internals", "of", "the", "networks", "We", "investigate", "the", "trained", "generators", "and", "discriminators", "in", "a", "variety", "of", "ways", ".", "We", "do", "not", "do", "any", "kind", "of", "nearest", "neighbor", "search", "on", "the", "training", "set", ".", "Nearest", "neighbors", "in", "pixel", "or", "feature", "space", "are", "trivially", "fooled", "Theis2015d", "by", "small", "image", "transforms", ".", "We", "also", "do", "not", "use", "log", "-", "likelihood", "metrics", "to", "quantitatively", "assess", "the", "model", ",", "as", "it", "is", "a", "poor", "Theis2015d", "metric", ".", "subsection", ":", "Walking", "in", "the", "latent", "space", "The", "first", "experiment", "we", "did", "was", "to", "understand", "the", "landscape", "of", "the", "latent", "space", ".", "Walking", "on", "the", "manifold", "that", "is", "learnt", "can", "usually", "tell", "us", "about", "signs", "of", "memorization", "(", "if", "there", "are", "sharp", "transitions", ")", "and", "about", "the", "way", "in", "which", "the", "space", "is", "hierarchically", "collapsed", ".", "If", "walking", "in", "this", "latent", "space", "results", "in", "semantic", "changes", "to", "the", "image", "generations", "(", "such", "as", "objects", "being", "added", "and", "removed", ")", ",", "we", "can", "reason", "that", "the", "model", "has", "learned", "relevant", "and", "interesting", "representations", ".", "The", "results", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "subsection", ":", "Visualizing", "the", "Discriminator", "Features", "Previous", "work", "has", "demonstrated", "that", "supervised", "training", "of", "CNNs", "on", "large", "image", "datasets", "results", "in", "very", "powerful", "learned", "features", "zeiler2014visualizing", ".", "Additionally", ",", "supervised", "CNNs", "trained", "on", "scene", "classification", "learn", "object", "detectors", "Oquab14", ".", "We", "demonstrate", "that", "an", "unsupervised", "DCGAN", "trained", "on", "a", "large", "image", "dataset", "can", "also", "learn", "a", "hierarchy", "of", "features", "that", "are", "interesting", ".", "Using", "guided", "backpropagation", "as", "proposed", "by", "springenberg2014striving", ",", "we", "show", "in", "Fig", ".", "[", "reference", "]", "that", "the", "features", "learnt", "by", "the", "discriminator", "activate", "on", "typical", "parts", "of", "a", "bedroom", ",", "like", "beds", "and", "windows", ".", "For", "comparison", ",", "in", "the", "same", "figure", ",", "we", "give", "a", "baseline", "for", "randomly", "initialized", "features", "that", "are", "not", "activated", "on", "anything", "that", "is", "semantically", "relevant", "or", "interesting", ".", "subsection", ":", "Manipulating", "the", "Generator", "Representation", "subsubsection", ":", "Forgetting", "to", "draw", "certain", "objects", "In", "addition", "to", "the", "representations", "learnt", "by", "a", "discriminator", ",", "there", "is", "the", "question", "of", "what", "representations", "the", "generator", "learns", ".", "The", "quality", "of", "samples", "suggest", "that", "the", "generator", "learns", "specific", "object", "representations", "for", "major", "scene", "components", "such", "as", "beds", ",", "windows", ",", "lamps", ",", "doors", ",", "and", "miscellaneous", "furniture", ".", "In", "order", "to", "explore", "the", "form", "that", "these", "representations", "take", ",", "we", "conducted", "an", "experiment", "to", "attempt", "to", "remove", "windows", "from", "the", "generator", "completely", ".", "On", "150", "samples", ",", "52", "window", "bounding", "boxes", "were", "drawn", "manually", ".", "On", "the", "second", "highest", "convolution", "layer", "features", ",", "logistic", "regression", "was", "fit", "to", "predict", "whether", "a", "feature", "activation", "was", "on", "a", "window", "(", "or", "not", ")", ",", "by", "using", "the", "criterion", "that", "activations", "inside", "the", "drawn", "bounding", "boxes", "are", "positives", "and", "random", "samples", "from", "the", "same", "images", "are", "negatives", ".", "Using", "this", "simple", "model", ",", "all", "feature", "maps", "with", "weights", "greater", "than", "zero", "(", "200", "in", "total", ")", "were", "dropped", "from", "all", "spatial", "locations", ".", "Then", ",", "random", "new", "samples", "were", "generated", "with", "and", "without", "the", "feature", "map", "removal", ".", "The", "generated", "images", "with", "and", "without", "the", "window", "dropout", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "and", "interestingly", ",", "the", "network", "mostly", "forgets", "to", "draw", "windows", "in", "the", "bedrooms", ",", "replacing", "them", "with", "other", "objects", ".", "subsubsection", ":", "Vector", "arithmetic", "on", "face", "samples", "In", "the", "context", "of", "evaluating", "learned", "representations", "of", "words", "mikolov2013distributed", "demonstrated", "that", "simple", "arithmetic", "operations", "revealed", "rich", "linear", "structure", "in", "representation", "space", ".", "One", "canonical", "example", "demonstrated", "that", "the", "vector", "(", "”King", "”", ")", "-", "vector", "(", "”Man", "”", ")", "+", "vector", "(", "”Woman", "”", ")", "resulted", "in", "a", "vector", "whose", "nearest", "neighbor", "was", "the", "vector", "for", "Queen", ".", "We", "investigated", "whether", "similar", "structure", "emerges", "in", "the", "representation", "of", "our", "generators", ".", "We", "performed", "similar", "arithmetic", "on", "the", "vectors", "of", "sets", "of", "exemplar", "samples", "for", "visual", "concepts", ".", "Experiments", "working", "on", "only", "single", "samples", "per", "concept", "were", "unstable", ",", "but", "averaging", "the", "vector", "for", "three", "examplars", "showed", "consistent", "and", "stable", "generations", "that", "semantically", "obeyed", "the", "arithmetic", ".", "In", "addition", "to", "the", "object", "manipulation", "shown", "in", "(", "Fig", ".", "[", "reference", "]", ")", ",", "we", "demonstrate", "that", "face", "pose", "is", "also", "modeled", "linearly", "in", "space", "(", "Fig", ".", "[", "reference", "]", ")", ".", "These", "demonstrations", "suggest", "interesting", "applications", "can", "be", "developed", "using", "representations", "learned", "by", "our", "models", ".", "It", "has", "been", "previously", "demonstrated", "that", "conditional", "generative", "models", "can", "learn", "to", "convincingly", "model", "object", "attributes", "like", "scale", ",", "rotation", ",", "and", "position", "dosovitskiy2014learning", ".", "This", "is", "to", "our", "knowledge", "the", "first", "demonstration", "of", "this", "occurring", "in", "purely", "unsupervised", "models", ".", "Further", "exploring", "and", "developing", "the", "above", "mentioned", "vector", "arithmetic", "could", "dramatically", "reduce", "the", "amount", "of", "data", "needed", "for", "conditional", "generative", "modeling", "of", "complex", "image", "distributions", ".", "section", ":", "Conclusion", "and", "Future", "Work", "We", "propose", "a", "more", "stable", "set", "of", "architectures", "for", "training", "generative", "adversarial", "networks", "and", "we", "give", "evidence", "that", "adversarial", "networks", "learn", "good", "representations", "of", "images", "for", "supervised", "learning", "and", "generative", "modeling", ".", "There", "are", "still", "some", "forms", "of", "model", "instability", "remaining", "-", "we", "noticed", "as", "models", "are", "trained", "longer", "they", "sometimes", "collapse", "a", "subset", "of", "filters", "to", "a", "single", "oscillating", "mode", ".", "Further", "work", "is", "needed", "to", "tackle", "this", "from", "of", "instability", ".", "We", "think", "that", "extending", "this", "framework", "to", "other", "domains", "such", "as", "video", "(", "for", "frame", "prediction", ")", "and", "audio", "(", "pre", "-", "trained", "features", "for", "speech", "synthesis", ")", "should", "be", "very", "interesting", ".", "Further", "investigations", "into", "the", "properties", "of", "the", "learnt", "latent", "space", "would", "be", "interesting", "as", "well", ".", "subsubsection", ":", "Acknowledgments", "We", "are", "fortunate", "and", "thankful", "for", "all", "the", "advice", "and", "guidance", "we", "have", "received", "during", "this", "work", ",", "especially", "that", "of", "Ian", "Goodfellow", ",", "Tobias", "Springenberg", ",", "Arthur", "Szlam", "and", "Durk", "Kingma", ".", "Additionally", "we", "’d", "like", "to", "thank", "all", "of", "the", "folks", "at", "indico", "for", "providing", "support", ",", "resources", ",", "and", "conversations", ",", "especially", "the", "two", "other", "members", "of", "the", "indico", "research", "team", ",", "Dan", "Kuster", "and", "Nathan", "Lintz", ".", "Finally", ",", "we", "’d", "like", "to", "thank", "Nvidia", "for", "donating", "a", "Titan", "-", "X", "GPU", "used", "in", "this", "work", ".", "bibliography", ":", "References", "section", ":", "Supplementary", "Material", "subsection", ":", "Evaluating", "DCGANs", "capability", "to", "capture", "data", "distributions", "We", "propose", "to", "apply", "standard", "classification", "metrics", "to", "a", "conditional", "version", "of", "our", "model", ",", "evaluating", "the", "conditional", "distributions", "learned", ".", "We", "trained", "a", "DCGAN", "on", "MNIST", "(", "splitting", "off", "a", "10", "K", "validation", "set", ")", "as", "well", "as", "a", "permutation", "invariant", "GAN", "baseline", "and", "evaluated", "the", "models", "using", "a", "nearest", "neighbor", "classifier", "comparing", "real", "data", "to", "a", "set", "of", "generated", "conditional", "samples", ".", "We", "found", "that", "removing", "the", "scale", "and", "bias", "parameters", "from", "batchnorm", "produced", "better", "results", "for", "both", "models", ".", "We", "speculate", "that", "the", "noise", "introduced", "by", "batchnorm", "helps", "the", "generative", "models", "to", "better", "explore", "and", "generate", "from", "the", "underlying", "data", "distribution", ".", "The", "results", "are", "shown", "in", "Table", "[", "reference", "]", "which", "compares", "our", "models", "with", "other", "techniques", ".", "The", "DCGAN", "model", "achieves", "the", "same", "test", "error", "as", "a", "nearest", "neighbor", "classifier", "fitted", "on", "the", "training", "dataset", "-", "suggesting", "the", "DCGAN", "model", "has", "done", "a", "superb", "job", "at", "modeling", "the", "conditional", "distributions", "of", "this", "dataset", ".", "At", "one", "million", "samples", "per", "class", ",", "the", "DCGAN", "model", "outperforms", "InfiMNIST", "loosli", "-", "canu", "-", "bottou", "-", "2006", ",", "a", "hand", "developed", "data", "augmentation", "pipeline", "which", "uses", "translations", "and", "elastic", "deformations", "of", "training", "examples", ".", "The", "DCGAN", "is", "competitive", "with", "a", "probabilistic", "generative", "data", "augmentation", "technique", "utilizing", "learned", "per", "class", "transformations", "Hauberg2015", "while", "being", "more", "general", "as", "it", "directly", "models", "the", "data", "instead", "of", "transformations", "of", "the", "data", "."], "entity": [[2, 5, ["Unsupervised", "Representation", "Learning"], "Task"], [6, 11, ["Deep", "Convolutional", "Generative", "Adversarial", "Networks"], "Method"], [15, 17, ["supervised", "learning"], "Method"], [18, 20, ["convolutional", "networks"], "Method"], [21, 22, ["CNNs"], "Method"], [28, 31, ["computer", "vision", "applications"], "Task"], [34, 36, ["unsupervised", "learning"], "Method"], [37, 38, ["CNNs"], "Method"], [57, 58, ["CNNs"], "Method"], [59, 61, ["supervised", "learning"], "Task"], [62, 64, ["unsupervised", "learning"], "Task"], [70, 71, ["CNNs"], "Method"], [72, 77, ["deep", "convolutional", "generative", "adversarial", "networks"], "Method"], [78, 79, ["DCGANs"], "Method"], [96, 98, ["unsupervised", "learning"], "Task"], [111, 115, ["deep", "convolutional", "adversarial", "pair"], "Method"], [128, 131, ["generator", "and", "discriminator"], "Method"], [147, 150, ["general", "image", "representations"], "Method"], [157, 160, ["reusable", "feature", "representations"], "Method"], [176, 178, ["computer", "vision"], "Task"], [194, 196, ["intermediate", "representations"], "Method"], [206, 209, ["supervised", "learning", "tasks"], "Task"], [211, 213, ["image", "classification"], "Task"], [222, 224, ["image", "representations"], "Task"], [227, 230, ["Generative", "Adversarial", "Networks"], "Method"], [231, 232, ["GANs"], "Method"], [241, 245, ["generator", "and", "discriminator", "networks"], "Method"], [246, 248, ["feature", "extractors"], "Method"], [249, 251, ["supervised", "tasks"], "Task"], [252, 253, ["GANs"], "Method"], [258, 261, ["maximum", "likelihood", "techniques"], "Method"], [268, 270, ["learning", "process"], "Method"], [275, 278, ["heuristic", "cost", "function"], "Metric"], [281, 289, ["pixel", "-", "wise", "independent", "mean", "-", "square", "error"], "Method"], [293, 295, ["representation", "learning"], "Task"], [296, 297, ["GANs"], "Method"], [329, 331, ["GANs", "learn"], "Method"], [334, 341, ["intermediate", "representations", "of", "multi", "-", "layer", "GANs"], "Method"], [364, 366, ["Convolutional", "GANs"], "Method"], [382, 385, ["Deep", "Convolutional", "GANs"], "Method"], [386, 387, ["DCGAN"], "Method"], [394, 397, ["image", "classification", "tasks"], "Task"], [403, 405, ["unsupervised", "algorithms"], "Method"], [409, 410, ["filters"], "Method"], [412, 413, ["GANs"], "Method"], [454, 456, ["Representation", "Learning"], "Task"], [459, 462, ["Unsupervised", "representation", "learning"], "Method"], [469, 473, ["general", "computer", "vision", "research"], "Task"], [487, 490, ["unsupervised", "representation", "learning"], "Task"], [493, 494, ["clustering"], "Task"], [501, 504, ["K", "-", "means"], "Method"], [512, 514, ["classification", "scores"], "Metric"], [524, 529, ["hierarchical", "clustering", "of", "image", "patches"], "Method"], [533, 535, ["image", "representations"], "Method"], [542, 545, ["auto", "-", "encoders"], "Method"], [546, 547, ["convolutionally"], "Method"], [548, 549, ["stacked"], "Method"], [562, 564, ["ladder", "structures"], "Method"], [572, 574, ["compact", "code"], "Method"], [597, 599, ["feature", "representations"], "Method"], [603, 606, ["Deep", "belief", "networks"], "Method"], [615, 618, ["learning", "hierarchical", "representations"], "Task"], [621, 624, ["Generating", "natural", "images"], "Task"], [624, 627, ["Generative", "image", "models"], "Method"], [636, 641, ["parametric", "and", "non", "-", "parametric"], "Method"], [643, 647, ["non", "-", "parametric", "models"], "Method"], [649, 650, ["matching"], "Task"], [668, 670, ["texture", "synthesis"], "Task"], [672, 675, ["super", "-", "resolution"], "Task"], [677, 680, ["in", "-", "painting"], "Task"], [682, 684, ["Parametric", "models"], "Method"], [685, 687, ["generating", "images"], "Task"], [699, 701, ["texture", "synthesis"], "Task"], [722, 725, ["variational", "sampling", "approach"], "Method"], [726, 728, ["generating", "images"], "Task"], [749, 753, ["iterative", "forward", "diffusion", "process"], "Method"], [755, 758, ["Generative", "Adversarial", "Networks"], "Method"], [769, 772, ["laplacian", "pyramid", "extension"], "Method"], [800, 803, ["recurrent", "network", "approach"], "Method"], [806, 809, ["deconvolution", "network", "approach"], "Method"], [828, 829, ["generators"], "Method"], [830, 832, ["supervised", "tasks"], "Task"], [839, 840, ["CNNs"], "Method"], [845, 847, ["neural", "networks"], "Method"], [852, 856, ["black", "-", "box", "methods"], "Method"], [871, 875, ["human", "-", "consumable", "algorithm"], "Method"], [880, 881, ["CNNs"], "Method"], [892, 893, ["deconvolutions"], "Method"], [907, 909, ["convolution", "filter"], "Method"], [917, 919, ["gradient", "descent"], "Method"], [947, 948, ["GANs"], "Method"], [949, 950, ["CNNs"], "Method"], [962, 963, ["LAPGAN"], "Method"], [970, 976, ["iteratively", "upscale", "low", "resolution", "generated", "images"], "Task"], [990, 991, ["GANs"], "Method"], [992, 994, ["CNN", "architectures"], "Method"], [1005, 1007, ["model", "exploration"], "Task"], [1030, 1033, ["deeper", "generative", "models"], "Method"], [1047, 1049, ["CNN", "architectures"], "Method"], [1054, 1057, ["all", "convolutional", "net"], "Method"], [1060, 1064, ["deterministic", "spatial", "pooling", "functions"], "Method"], [1067, 1068, ["maxpooling"], "Method"], [1070, 1072, ["strided", "convolutions"], "Method"], [1097, 1099, ["spatial", "upsampling"], "Method"], [1101, 1102, ["discriminator"], "Method"], [1124, 1127, ["global", "average", "pooling"], "Method"], [1136, 1139, ["image", "classification", "models"], "Method"], [1143, 1146, ["global", "average", "pooling"], "Method"], [1147, 1149, ["model", "stability"], "Metric"], [1151, 1153, ["convergence", "speed"], "Metric"], [1172, 1173, ["generator"], "Method"], [1174, 1175, ["discriminator"], "Method"], [1183, 1184, ["GAN"], "Method"], [1204, 1206, ["matrix", "multiplication"], "Method"], [1225, 1227, ["convolution", "stack"], "Method"], [1230, 1231, ["discriminator"], "Method"], [1234, 1236, ["convolution", "layer"], "Method"], [1259, 1261, ["model", "architecture"], "Method"], [1264, 1266, ["Batch", "Normalization"], "Method"], [1269, 1270, ["learning"], "Task"], [1289, 1291, ["training", "problems"], "Task"], [1299, 1301, ["gradient", "flow"], "Method"], [1302, 1304, ["deeper", "models"], "Method"], [1310, 1312, ["deep", "generators"], "Method"], [1314, 1315, ["learning"], "Task"], [1335, 1336, ["GANs"], "Method"], [1339, 1340, ["batchnorm"], "Method"], [1359, 1360, ["batchnorm"], "Method"], [1362, 1365, ["generator", "output", "layer"], "Method"], [1367, 1370, ["discriminator", "input", "layer"], "Method"], [1372, 1374, ["ReLU", "activation"], "Method"], [1379, 1380, ["generator"], "Method"], [1425, 1428, ["leaky", "rectified", "activation"], "Method"], [1436, 1439, ["higher", "resolution", "modeling"], "Task"], [1461, 1464, ["Deep", "Convolutional", "GANs"], "Method"], [1466, 1468, ["pooling", "layers"], "Method"], [1469, 1471, ["strided", "convolutions"], "Method"], [1472, 1473, ["discriminator"], "Method"], [1475, 1479, ["fractional", "-", "strided", "convolutions"], "Method"], [1480, 1481, ["generator"], "Method"], [1484, 1485, ["batchnorm"], "Method"], [1488, 1489, ["generator"], "Method"], [1491, 1492, ["discriminator"], "Method"], [1499, 1501, ["deeper", "architectures"], "Method"], [1503, 1505, ["ReLU", "activation"], "Method"], [1506, 1507, ["generator"], "Method"], [1517, 1518, ["Tanh"], "Method"], [1524, 1525, ["discriminator"], "Method"], [1533, 1535, ["adversarial", "training"], "Method"], [1537, 1538, ["DCGANs"], "Method"], [1542, 1547, ["Large", "-", "scale", "Scene", "Understanding"], "Task"], [1548, 1549, ["LSUN"], "Method"], [1552, 1555, ["Imagenet", "-", "1k"], "Material"], [1558, 1561, ["assembled", "Faces", "dataset"], "Material"], [1606, 1612, ["mini", "-", "batch", "stochastic", "gradient", "descent"], "Method"], [1613, 1614, ["SGD"], "Method"], [1630, 1635, ["zero", "-", "centered", "Normal", "distribution"], "Method"], [1659, 1660, ["GAN"], "Method"], [1666, 1667, ["training"], "Task"], [1671, 1673, ["Adam", "optimizer"], "Method"], [1675, 1677, ["tuned", "hyperparameters"], "Method"], [1682, 1684, ["learning", "rate"], "Metric"], [1723, 1724, ["training"], "Task"], [1727, 1728, ["LSUN"], "Method"], [1729, 1731, ["visual", "quality"], "Metric"], [1734, 1737, ["generative", "image", "models"], "Method"], [1742, 1745, ["over", "-", "fitting"], "Method"], [1746, 1750, ["memorization", "of", "training", "samples"], "Task"], [1764, 1766, ["resolution", "generation"], "Task"], [1773, 1776, ["LSUN", "bedrooms", "dataset"], "Material"], [1802, 1804, ["generalization", "performance"], "Metric"], [1823, 1825, ["online", "learning"], "Method"], [1863, 1865, ["data", "augmentation"], "Method"], [1896, 1901, ["image", "de", "-", "duplication", "process"], "Method"], [1905, 1917, ["3072", "-", "128", "-", "3072", "de", "-", "noising", "dropout", "regularized", "RELU", "autoencoder"], "Method"], [1936, 1937, ["thresholding"], "Method"], [1938, 1940, ["ReLU", "activation"], "Method"], [1948, 1951, ["information", "preserving", "technique"], "Method"], [1958, 1961, ["semantic", "-", "hashing"], "Method"], [1964, 1969, ["linear", "time", "de", "-", "duplication"], "Task"], [1970, 1975, ["Visual", "inspection", "of", "hash", "collisions"], "Task"], [1977, 1978, ["precision"], "Metric"], [1981, 1984, ["false", "positive", "rate"], "Metric"], [2008, 2009, ["recall"], "Metric"], [2034, 2035, ["dbpedia"], "Material"], [2062, 2065, ["OpenCV", "face", "detector"], "Method"], [2097, 2099, ["data", "augmentation"], "Method"], [2107, 2110, ["Imagenet", "-", "1k"], "Material"], [2112, 2115, ["Imagenet", "-", "1k"], "Material"], [2123, 2125, ["unsupervised", "training"], "Task"], [2129, 2134, ["min", "-", "resized", "center", "crops"], "Method"], [2136, 2138, ["data", "augmentation"], "Method"], [2154, 2157, ["CIFAR", "-", "10"], "Material"], [2158, 2159, ["GANs"], "Method"], [2161, 2163, ["feature", "extractor"], "Method"], [2171, 2175, ["unsupervised", "representation", "learning", "algorithms"], "Method"], [2181, 2183, ["feature", "extractor"], "Method"], [2191, 2193, ["linear", "models"], "Method"], [2219, 2224, ["single", "layer", "feature", "extraction", "pipeline"], "Method"], [2225, 2228, ["K", "-", "means"], "Method"], [2230, 2233, ["feature", "learning", "algorithm"], "Method"], [2251, 2252, ["accuracy"], "Metric"], [2254, 2259, ["unsupervised", "multi", "-", "layered", "extension"], "Method"], [2266, 2267, ["accuracy"], "Metric"], [2278, 2279, ["DCGANs"], "Method"], [2280, 2282, ["supervised", "tasks"], "Task"], [2286, 2289, ["Imagenet", "-", "1k"], "Material"], [2303, 2305, ["layers", "representation"], "Method"], [2326, 2332, ["regularized", "linear", "L2", "-", "SVM", "classifier"], "Method"], [2343, 2344, ["accuracy"], "Metric"], [2348, 2353, ["K", "-", "means", "based", "approaches"], "Method"], [2357, 2358, ["discriminator"], "Method"], [2372, 2377, ["K", "-", "means", "based", "techniques"], "Method"], [2385, 2388, ["feature", "vector", "size"], "Metric"], [2400, 2401, ["DCGANs"], "Method"], [2407, 2409, ["Exemplar", "CNNs"], "Method"], [2415, 2418, ["normal", "discriminative", "CNNs"], "Method"], [2420, 2422, ["unsupervised", "fashion"], "Method"], [2446, 2449, ["discriminator", "’s", "representations"], "Method"], [2462, 2463, ["DCGAN"], "Method"], [2467, 2470, ["CIFAR", "-", "10"], "Material"], [2475, 2477, ["domain", "robustness"], "Metric"], [2484, 2487, ["Classifying", "SVHN", "digits"], "Task"], [2488, 2489, ["GANs"], "Method"], [2491, 2493, ["feature", "extractor"], "Method"], [2495, 2499, ["StreetView", "House", "Numbers", "dataset"], "Material"], [2500, 2501, ["SVHN"], "Material"], [2510, 2511, ["discriminator"], "Method"], [2513, 2514, ["DCGAN"], "Method"], [2515, 2517, ["supervised", "purposes"], "Task"], [2531, 2534, ["CIFAR", "-", "10"], "Material"], [2558, 2562, ["hyperparameter", "and", "model", "selection"], "Method"], [2577, 2583, ["regularized", "linear", "L2", "-", "SVM", "classifier"], "Method"], [2588, 2591, ["feature", "extraction", "pipeline"], "Method"], [2593, 2596, ["CIFAR", "-", "10"], "Material"], [2605, 2606, ["classification"], "Task"], [2613, 2615, ["test", "error"], "Metric"], [2621, 2622, ["CNNs"], "Method"], [2635, 2637, ["CNN", "architecture"], "Method"], [2639, 2640, ["DCGAN"], "Method"], [2655, 2657, ["supervised", "CNN"], "Method"], [2670, 2672, ["random", "search"], "Method"], [2685, 2687, ["validation", "error"], "Metric"], [2702, 2703, ["generators"], "Method"], [2704, 2705, ["discriminators"], "Method"], [2718, 2721, ["nearest", "neighbor", "search"], "Method"], [2738, 2741, ["small", "image", "transforms"], "Method"], [2747, 2751, ["log", "-", "likelihood", "metrics"], "Metric"], [2880, 2884, ["supervised", "training", "of", "CNNs"], "Method"], [2898, 2900, ["supervised", "CNNs"], "Method"], [2902, 2904, ["scene", "classification"], "Task"], [2905, 2907, ["object", "detectors"], "Method"], [2933, 2935, ["guided", "backpropagation"], "Method"], [2954, 2955, ["discriminator"], "Method"], [3001, 3003, ["Generator", "Representation"], "Method"], [3018, 3019, ["discriminator"], "Method"], [3028, 3029, ["generator"], "Method"], [3038, 3039, ["generator"], "Method"], [3041, 3043, ["object", "representations"], "Method"], [3083, 3084, ["generator"], "Method"], [3106, 3108, ["logistic", "regression"], "Method"], [3184, 3187, ["feature", "map", "removal"], "Method"], [3228, 3230, ["Vector", "arithmetic"], "Method"], [3238, 3242, ["learned", "representations", "of", "words"], "Task"], [3246, 3248, ["arithmetic", "operations"], "Method"], [3394, 3395, ["representations"], "Method"], [3406, 3409, ["conditional", "generative", "models"], "Method"], [3438, 3440, ["unsupervised", "models"], "Method"], [3448, 3450, ["vector", "arithmetic"], "Method"], [3459, 3462, ["conditional", "generative", "modeling"], "Method"], [3463, 3466, ["complex", "image", "distributions"], "Task"], [3483, 3486, ["generative", "adversarial", "networks"], "Method"], [3491, 3493, ["adversarial", "networks"], "Method"], [3499, 3501, ["supervised", "learning"], "Task"], [3502, 3504, ["generative", "modeling"], "Task"], [3560, 3562, ["frame", "prediction"], "Task"], [3566, 3570, ["pre", "-", "trained", "features"], "Task"], [3571, 3573, ["speech", "synthesis"], "Task"], [3703, 3705, ["data", "distributions"], "Task"], [3710, 3712, ["classification", "metrics"], "Method"], [3714, 3716, ["conditional", "version"], "Method"], [3729, 3730, ["DCGAN"], "Method"], [3731, 3732, ["MNIST"], "Material"], [3755, 3758, ["nearest", "neighbor", "classifier"], "Method"], [3779, 3780, ["batchnorm"], "Method"], [3794, 3795, ["batchnorm"], "Method"], [3797, 3799, ["generative", "models"], "Method"], [3833, 3835, ["test", "error"], "Metric"], [3837, 3840, ["nearest", "neighbor", "classifier"], "Method"], [3875, 3876, ["InfiMNIST"], "Method"], [3885, 3890, ["hand", "developed", "data", "augmentation", "pipeline"], "Method"], [3901, 3902, ["DCGAN"], "Method"], [3906, 3911, ["probabilistic", "generative", "data", "augmentation", "technique"], "Method"], [3913, 3916, ["per", "class", "transformations"], "Method"], [695, 696, ["MNIST"], "Material"], [1447, 1448, ["GAN"], "Method"], [2149, 2150, ["DCGANs"], "Method"], [2202, 2205, ["CIFAR", "-", "10"], "Material"], [2914, 2915, ["DCGAN"], "Method"], [3699, 3700, ["DCGANs"], "Method"], [3747, 3748, ["GAN"], "Method"], [3828, 3829, ["DCGAN"], "Method"], [3848, 3849, ["DCGAN"], "Method"], [3872, 3873, ["DCGAN"], "Method"]]} {"doc_id": "3858e5175799b97805b2b70ff54e8a7e0718870f", "text": ["document", ":", "Deep", "Learning", "For", "Smile", "Recognition", "Inspired", "by", "recent", "successes", "of", "deep", "learning", "in", "computer", "vision", ",", "we", "propose", "a", "novel", "application", "of", "deep", "convolutional", "neural", "networks", "to", "facial", "expression", "recognition", ",", "in", "particular", "smile", "recognition", ".", "A", "smile", "recognition", "test", "accuracy", "of", "99.45", "%", "is", "achieved", "for", "the", "Denver", "Intensity", "of", "Spontaneous", "Facial", "Action", "(", "DISFA", ")", "database", ",", "significantly", "outperforming", "existing", "approaches", "based", "on", "hand", "-", "crafted", "features", "with", "accuracies", "ranging", "from", "65.55", "%", "to", "79.67", "%", ".", "The", "novelty", "of", "this", "approach", "includes", "a", "comprehensive", "model", "selection", "of", "the", "architecture", "parameters", ",", "allowing", "to", "find", "an", "appropriate", "architecture", "for", "each", "expression", "such", "as", "smile", ".", "This", "is", "feasible", "because", "all", "experiments", "were", "run", "on", "a", "Tesla", "K40c", "GPU", ",", "allowing", "a", "speedup", "of", "factor", "10", "over", "traditional", "computations", "on", "a", "CPU", ".", "section", ":", "Introduction", "Neural", "networks", "are", "celebrating", "a", "comeback", "under", "the", "term", "”", "deep", "learning", "”", "for", "the", "last", "ten", "years", "by", "training", "many", "hidden", "layers", "allowing", "to", "self", "-", "learn", "complex", "feature", "hierarchies", ".", "This", "makes", "them", "of", "particular", "interest", "for", "computer", "vision", ",", "in", "which", "feature", "description", "is", "a", "long", "-", "standing", "issue", ".", "Many", "advances", "have", "been", "reported", "in", "this", "period", ",", "including", "new", "training", "methods", "and", "a", "paradigm", "shift", "of", "training", "from", "CPUs", "to", "GPUs", ".", "As", "a", "result", ",", "those", "advances", "allow", "to", "train", "more", "reliable", "models", "much", "faster", ".", "This", "has", "for", "example", "resulted", "in", "breakthroughs", "in", "signal", "processing", ".", "Nonetheless", ",", "deep", "neural", "networks", "are", "not", "a", "magic", "bullet", "and", "successful", "training", "is", "still", "heavily", "based", "on", "experimentation", ".", "The", "Facial", "Action", "Coding", "System", "(", "FACS", ")", "is", "a", "system", "to", "taxonomize", "any", "facial", "expression", "of", "a", "human", "being", "by", "their", "appearance", "on", "the", "face", ".", "Action", "units", "describe", "muscles", "or", "muscle", "groups", "in", "the", "face", ",", "are", "set", "or", "unset", "and", "the", "activation", "may", "be", "on", "different", "intensity", "levels", ".", "State", "-", "of", "-", "the", "art", "approaches", "in", "this", "field", "mostly", "rely", "on", "hand", "-", "crafted", "features", "leaving", "a", "lot", "of", "potential", "for", "higher", "accuracies", ".", "In", "contrast", "to", "other", "fields", "such", "as", "face", "or", "gesture", "recognition", ",", "only", "very", "few", "works", "on", "deep", "learning", "applied", "to", "facial", "expression", "recognition", "have", "been", "reported", "so", "far", "in", "which", "the", "architecture", "parameters", "are", "fixed", ".", "We", "are", "not", "aware", "of", "publications", "in", "which", "the", "architecture", "of", "a", "deep", "neural", "network", "for", "facial", "expression", "recognition", "is", "subject", "to", "extensive", "model", "selection", ".", "This", "allows", "to", "learn", "appropriate", "architectures", "per", "action", "unit", ".", "section", ":", "Deep", "neural", "networks", "Training", "neural", "networks", "is", "difficult", ",", "as", "their", "cost", "functions", "have", "many", "local", "minima", ".", "The", "more", "hidden", "layers", ",", "the", "more", "difficult", "the", "training", "of", "a", "neural", "network", ".", "Hence", ",", "training", "tends", "to", "converge", "to", "a", "local", "minimum", ",", "resulting", "in", "poor", "generalization", "of", "the", "network", ".", "In", "order", "to", "overcome", "these", "issues", ",", "a", "variety", "of", "new", "concepts", "have", "been", "proposed", "in", "the", "literature", ",", "of", "which", "only", "a", "few", "can", "be", "named", "in", "this", "chapter", ".", "Unsupervised", "pre", "-", "training", "methods", ",", "such", "as", "autoencoders", "allow", "to", "initialize", "the", "weights", "well", "in", "order", "for", "backpropagation", "to", "quickly", "optimize", "them", ".", "The", "Rectified", "Linear", "Unit", "(", "ReLU", ")", "and", "dropout", "are", "new", "regularization", "methods", ".", "The", "new", "training", "methods", "and", "other", "new", "concepts", "can", "also", "lead", "to", "significant", "improvements", "of", "shallow", "neural", "networks", "with", "just", "a", "few", "hidden", "layers", ".", "Convolutional", "neural", "networks", "(", "CNNs", ")", "were", "initially", "proposed", "by", "LeCun", "for", "the", "recognition", "of", "hand", "-", "written", "digits", ".", "A", "CNN", "consists", "of", "two", "layers", ":", "a", "convolutional", "layer", ",", "followed", "by", "a", "subsampling", "layer", ".", "Inspired", "by", "biological", "processes", "and", "exploiting", "the", "fact", "that", "nearby", "pixels", "are", "strongly", "correlated", ",", "CNNs", "are", "relatively", "insensitive", "to", "small", "translations", "or", "rotations", "of", "the", "image", "input", ".", "Training", "deep", "neural", "networks", "is", "slow", "due", "to", "the", "number", "of", "parameters", "in", "the", "model", ".", "As", "the", "training", "can", "be", "described", "in", "a", "vectorized", "form", ",", "it", "is", "possible", "to", "massively", "parallelize", "it", ".", "Modern", "GPUs", "have", "thousands", "of", "cores", "and", "are", "therefore", "an", "ideal", "candidate", "for", "the", "execution", "of", "the", "training", "of", "neural", "networks", ".", "Significant", "speedups", "of", "factor", "10", "or", "higher", "have", "been", "reported", ".", "A", "difficulty", "is", "to", "write", "GPU", "code", ".", "In", "the", "last", "few", "years", ",", "more", "abstract", "libraries", "have", "been", "released", ".", "section", ":", "DISFA", "database", "The", "Denver", "Intensity", "of", "Spontaneous", "Facial", "Action", "(", "DISFA", ")", "database", "consists", "of", "27", "videos", "of", "4844", "frames", "each", ",", "with", "130", ",", "788", "images", "in", "total", ".", "Action", "unit", "annotations", "are", "on", "different", "levels", "of", "intensity", ",", "which", "are", "ignored", "in", "the", "following", "experiments", "and", "action", "units", "are", "either", "set", "or", "unset", ".", "DISFA", "was", "selected", "from", "a", "wider", "range", "of", "databases", "popular", "in", "the", "field", "of", "facial", "expression", "recognition", "because", "of", "the", "high", "number", "of", "smiles", ",", "i.e.", "action", "unit", "12", ".", "In", "detail", ",", "30", ",", "792", "have", "this", "action", "unit", "set", ",", "82", ",", "176", "images", "have", "some", "action", "unit", "(", "s", ")", "set", "and", "48", ",", "612", "images", "have", "no", "action", "unit", "(", "s", ")", "set", "at", "all", ".", "Fig", ".", "[", "reference", "]", "contains", "a", "sample", "image", "of", "DISFA", ".", "[", "width=0.2", "]", "mouth", "[", "width=0.2", "]", "face", "In", "the", "original", "paper", "on", "DISFA", "multi", "-", "class", "SVMs", "were", "trained", "for", "the", "different", "levels", "0", "-", "5", "of", "action", "unit", "intensity", ".", "Test", "accuracies", "for", "the", "individual", "levels", "and", "for", "the", "binary", "action", "unit", "recognition", "problem", "are", "reported", "for", "three", "different", "hand", "-", "crafted", "feature", "description", "techniques", ".", "In", "those", "three", "cases", ",", "accuracies", "of", "65.55", "%", ",", "72.94", "%", "and", "79.67", "%", "for", "smile", "recognition", "are", "reported", ".", "section", ":", "Smile", "recognition", "In", "the", "following", "experiments", ",", "an", "aligned", "version", "of", "DISFA", "is", "used", ".", "In", "this", "aligned", "version", ",", "the", "faces", "have", "been", "cropped", "and", "annotated", "with", "facial", "landmark", "points", ".", "Facial", "landmark", "points", "allow", "to", "compute", "a", "bounding", "box", "to", "fit", "the", "mouth", "in", "all", "images", ".", "In", "the", "experiments", ",", "two", "inputs", "are", "used", ":", "the", "mouth", "and", "face", ",", "downscaled", "to", "and", "pixels", ",", "respectively", ".", "Both", "inputs", "are", "used", "to", "assess", "if", "the", "mouth", "alone", "is", "as", "expressive", "as", "or", "even", "more", "expressive", "than", "the", "entire", "face", "for", "smile", "recognition", ".", "subsection", ":", "Model", "The", "architecture", "of", "the", "network", "is", "as", "follows", ":", "The", "input", "images", "are", "fed", "into", "a", "convolution", "comprising", "a", "convolutional", "and", "a", "subsampling", "layer", ".", "That", "convolution", "may", "be", "followed", "by", "more", "convolutions", "to", "become", "gradually", "more", "invariant", "to", "distortions", "in", "the", "input", ".", "In", "the", "second", "stage", ",", "a", "regular", "neural", "network", "follows", "the", "convolutions", "in", "order", "to", "discriminate", "the", "features", "learned", "by", "the", "convolutions", ".", "The", "output", "layer", "consists", "of", "two", "units", "for", "smile", "or", "no", "smile", ".", "The", "novelty", "of", "this", "approach", "is", "that", "the", "exact", "number", "of", "convolutions", ",", "number", "of", "hidden", "layers", "and", "size", "of", "hidden", "layers", "are", "not", "fixed", "but", "subject", "to", "extensive", "model", "selection", "in", "Sec", ".", "[", "reference", "]", ".", "subsection", ":", "Experiment", "setting", "Due", "to", "training", "time", "constraints", ",", "some", "parameters", "have", "been", "fixed", "to", "reasonable", "and", "empirical", "values", ",", "such", "as", "the", "size", "of", "convolutions", "(", "pixels", ",", "32", "feature", "maps", ")", "and", "the", "size", "of", "subsamplings", "(", "pixels", "using", "max", "pooling", ")", ".", "All", "layers", "use", "ReLU", "units", ",", "except", "of", "softmax", "being", "used", "in", "the", "output", "layer", ".", "The", "learning", "rate", "is", "fixed", "to", "and", "not", "subject", "to", "model", "selection", "as", "it", "would", "significantly", "prolong", "the", "model", "selection", ".", "The", "same", "considerations", "apply", "to", "the", "momentum", ",", "which", "is", "fixed", "to", ".", "The", "entire", "database", "has", "been", "randomly", "split", "into", "a", "60%", "/", "20%", "/", "20", "%", "training", "/", "validation", "/", "test", "ratio", ".", "Training", "neural", "networks", "comes", "with", "uncertainties", ",", "mostly", "due", "to", "the", "random", "initialization", "of", "the", "weights", ",", "but", "also", "due", "to", "that", "random", "split", "of", "the", "data", ".", "Evaluations", "have", "shown", "that", "for", "10", "similar", "experiments", "carried", "out", ",", "the", "standard", "deviation", "of", "the", "test", "accuracy", "is", "0.041725", "%", ".", "Because", "of", "this", "low", "standard", "deviation", ",", "performing", "each", "experiment", "exactly", "once", "has", "only", "a", "very", "low", "bias", "and", "is", "therefore", "relatively", "safe", "to", "do", "for", "reasons", "of", "faster", "training", "time", ".", "Throughout", "the", "experiments", ",", "the", "classification", "rate", "is", "used", "as", "the", "accuracy", "measure", ".", "The", "model", "is", "implemented", "using", "Lasagne", "and", "the", "generated", "CUDA", "code", "is", "executed", "on", "a", "Tesla", "K40c", "as", "training", "on", "a", "GPU", "allows", "to", "perform", "a", "comprehensive", "model", "selection", "in", "a", "feasible", "amount", "of", "time", ".", "Stochastic", "gradient", "descent", "with", "a", "batch", "size", "of", "500", "is", "used", ".", "subsection", ":", "Parameter", "optimization", "Table", "[", "reference", "]", "contains", "the", "four", "parameters", "to", "be", "optimized", ":", "the", "number", "of", "convolutions", ",", "the", "number", "of", "hidden", "layers", ",", "the", "number", "of", "units", "per", "hidden", "layer", "and", "the", "dropout", "factor", ".", "Each", "parameter", "was", "optimized", "independently", "due", "to", "training", "time", "constraints", ".", "This", "may", "not", "lead", "to", "an", "optimal", "model", ",", "but", "has", "proven", "to", "work", "empirically", "well", ".", "Each", "model", "was", "trained", "for", "50", "epochs", "in", "the", "model", "selection", ".", "Parameters", "and", "possible", "values", "used", "in", "model", "selection", ".", "For", "both", "inputs", ",", "Table", "[", "reference", "]", "contains", "the", "final", "models", "selected", ".", "For", "the", "mouth", "input", ",", "there", "is", "a", "preference", "to", "more", "convolutions", "and", "more", "hidden", "layers", ".", "This", "is", "the", "case", "because", "slight", "translations", "or", "rotations", "in", "the", "mouth", "input", "have", "stronger", "consequences", "on", "the", "classification", "result", ".", "In", "the", "entire", "face", ",", "that", "sort", "of", "distortions", "may", "be", "less", "of", "a", "problem", "because", "other", "parts", "of", "the", "face", "such", "as", "the", "cheeks", "contribute", "to", "smile", "recognition", ",", "too", ".", "Selected", "parameter", "values", "for", "mouth", "and", "face", "input", ".", "subsection", ":", "Results", "and", "discussion", "Both", "final", "models", "were", "trained", "for", "1000", "epochs", ".", "The", "test", "accuracies", "of", "both", "models", "started", "to", "converge", "after", "about", "300", "epochs", ".", "For", "the", "mouth", "and", "face", "inputs", ",", "the", "best", "accuracies", "were", "achieved", "after", "700", "and", "1000", "epochs", "with", "99.45", "%", "and", "99.34", "%", ",", "respectively", ".", "Both", "models", "significantly", "outperform", "the", "state", "-", "of", "-", "the", "-", "art", "SVM", "baselines", "reported", "in", "Sec", ".", "[", "reference", "]", "ranging", "from", "65.55", "%", "to", "79.67", "%", ".", "Overall", ",", "there", "is", "no", "strong", "preference", "for", "either", "the", "mouth", "or", "face", "input", ".", "Further", "experiments", "with", "a", "reduced", "dataset", "containing", "only", "70", "%", "of", "the", "images", "that", "have", "no", "action", "unit", "(", "s", ")", "set", "at", "all", "support", "this", "hypothesis", ".", "Concretely", ",", "the", "test", "accuracies", "for", "the", "mouth", "and", "face", "input", "reduced", "to", "99.24", "%", "and", "99.26", "%", ",", "respectively", ".", "Thus", ",", "the", "difference", "between", "the", "two", "models", "has", "been", "further", "reduced", "and", "this", "time", "giving", "a", "very", "low", "preference", "for", "the", "face", "input", ".", "Nonetheless", ",", "this", "difference", "is", "not", "representative", "as", "it", "is", "within", "the", "experiment", "error", "standard", "deviation", "reported", "in", "Sec", ".", "[", "reference", "]", ".", "Training", "time", "per", "epoch", "are", "82", "seconds", "and", "41", "seconds", "for", "the", "mouth", "and", "face", "input", "models", ",", "respectively", ".", "Experiments", "have", "shown", "that", "the", "training", "time", "mostly", "depends", "on", "the", "number", "of", "convolutions", ".", "Using", "the", "Tesla", "K40c", "GPU", "has", "allowed", "to", "speed", "up", "the", "training", "time", "by", "factor", "ten", "over", "the", "use", "of", "a", "CPU", "to", "execute", "the", "CPU", "code", "generated", "by", "the", "library", ".", "This", "clearly", "demonstrates", "the", "importance", "of", "training", "on", "a", "GPU", "to", "do", "a", "comprehensive", "model", "selection", "in", "a", "feasible", "amount", "of", "time", ".", "section", ":", "Conclusions", "and", "future", "work", "Deep", "learning", "is", "an", "umbrella", "term", "for", "training", "neural", "networks", "with", "potentially", "many", "hidden", "layers", "using", "new", "training", "methods", "allowing", "to", "learn", "complex", "feature", "hierarchies", "from", "data", ".", "Applied", "to", "action", "unit", "recognition", "and", "smile", "recognition", "in", "particular", ",", "a", "deep", "convolutional", "neural", "network", "model", "with", "an", "overall", "accuracy", "of", "99.45", "%", "significantly", "outperforms", "existing", "approaches", ".", "The", "underlying", "extensive", "model", "selection", "allows", "to", "find", "for", "each", "action", "unit", "an", "appropriate", "architecture", "in", "order", "to", "maximize", "test", "accuracies", ".", "In", "the", "future", ",", "we", "will", "extend", "the", "model", "to", "images", "from", "multiple", "databases", "and", "to", "make", "predictions", "in", "image", "sequences", ".", "bibliography", ":", "References"], "entity": [[2, 4, ["Deep", "Learning"], "Method"], [5, 7, ["Smile", "Recognition"], "Task"], [12, 14, ["deep", "learning"], "Method"], [15, 17, ["computer", "vision"], "Task"], [24, 28, ["deep", "convolutional", "neural", "networks"], "Method"], [29, 32, ["facial", "expression", "recognition"], "Task"], [35, 37, ["smile", "recognition"], "Task"], [39, 43, ["smile", "recognition", "test", "accuracy"], "Metric"], [50, 56, ["Denver", "Intensity", "of", "Spontaneous", "Facial", "Action"], "Material"], [57, 58, ["DISFA"], "Material"], [72, 73, ["accuracies"], "Metric"], [89, 91, ["model", "selection"], "Method"], [139, 141, ["Neural", "networks"], "Method"], [149, 151, ["deep", "learning"], "Method"], [178, 180, ["computer", "vision"], "Task"], [183, 185, ["feature", "description"], "Task"], [203, 205, ["training", "methods"], "Method"], [212, 213, ["CPUs"], "Method"], [214, 215, ["GPUs"], "Method"], [239, 241, ["signal", "processing"], "Task"], [244, 247, ["deep", "neural", "networks"], "Method"], [263, 269, ["Facial", "Action", "Coding", "System", "(", "FACS"], "Method"], [338, 339, ["accuracies"], "Metric"], [347, 351, ["face", "or", "gesture", "recognition"], "Task"], [357, 359, ["deep", "learning"], "Method"], [361, 364, ["facial", "expression", "recognition"], "Task"], [389, 392, ["deep", "neural", "network"], "Method"], [393, 396, ["facial", "expression", "recognition"], "Task"], [400, 402, ["model", "selection"], "Method"], [415, 418, ["Deep", "neural", "networks"], "Method"], [418, 421, ["Training", "neural", "networks"], "Method"], [445, 447, ["neural", "network"], "Method"], [450, 451, ["training"], "Task"], [498, 503, ["Unsupervised", "pre", "-", "training", "methods"], "Method"], [506, 507, ["autoencoders"], "Method"], [516, 517, ["backpropagation"], "Method"], [523, 526, ["Rectified", "Linear", "Unit"], "Method"], [527, 528, ["ReLU"], "Method"], [530, 531, ["dropout"], "Method"], [533, 535, ["regularization", "methods"], "Method"], [538, 540, ["training", "methods"], "Method"], [551, 554, ["shallow", "neural", "networks"], "Method"], [561, 564, ["Convolutional", "neural", "networks"], "Method"], [565, 566, ["CNNs"], "Method"], [571, 572, ["LeCun"], "Method"], [574, 580, ["recognition", "of", "hand", "-", "written", "digits"], "Task"], [582, 583, ["CNN"], "Method"], [595, 597, ["subsampling", "layer"], "Method"], [600, 602, ["biological", "processes"], "Method"], [613, 614, ["CNNs"], "Method"], [628, 631, ["deep", "neural", "networks"], "Method"], [645, 646, ["training"], "Task"], [651, 653, ["vectorized", "form"], "Method"], [663, 664, ["GPUs"], "Method"], [679, 683, ["training", "of", "neural", "networks"], "Task"], [700, 702, ["GPU", "code"], "Method"], [721, 727, ["Denver", "Intensity", "of", "Spontaneous", "Facial", "Action"], "Material"], [728, 729, ["DISFA"], "Material"], [774, 775, ["DISFA"], "Material"], [788, 791, ["facial", "expression", "recognition"], "Task"], [854, 855, ["DISFA"], "Material"], [889, 890, ["accuracies"], "Metric"], [907, 913, ["hand", "-", "crafted", "feature", "description", "techniques"], "Method"], [919, 920, ["accuracies"], "Metric"], [930, 932, ["smile", "recognition"], "Task"], [937, 939, ["Smile", "recognition"], "Task"], [948, 949, ["DISFA"], "Material"], [1030, 1032, ["smile", "recognition"], "Task"], [1052, 1053, ["convolution"], "Method"], [1055, 1056, ["convolutional"], "Method"], [1058, 1060, ["subsampling", "layer"], "Method"], [1062, 1063, ["convolution"], "Method"], [1068, 1069, ["convolutions"], "Method"], [1086, 1089, ["regular", "neural", "network"], "Method"], [1091, 1092, ["convolutions"], "Method"], [1101, 1102, ["convolutions"], "Method"], [1104, 1106, ["output", "layer"], "Method"], [1145, 1147, ["model", "selection"], "Task"], [1196, 1198, ["max", "pooling"], "Method"], [1208, 1209, ["softmax"], "Method"], [1217, 1219, ["learning", "rate"], "Metric"], [1226, 1228, ["model", "selection"], "Task"], [1234, 1236, ["model", "selection"], "Task"], [1265, 1271, ["training", "/", "validation", "/", "test", "ratio"], "Metric"], [1272, 1275, ["Training", "neural", "networks"], "Method"], [1316, 1318, ["test", "accuracy"], "Metric"], [1351, 1353, ["training", "time"], "Task"], [1359, 1361, ["classification", "rate"], "Metric"], [1365, 1367, ["accuracy", "measure"], "Metric"], [1373, 1374, ["Lasagne"], "Method"], [1377, 1379, ["CUDA", "code"], "Method"], [1383, 1385, ["Tesla", "K40c"], "Method"], [1395, 1397, ["model", "selection"], "Task"], [1404, 1407, ["Stochastic", "gradient", "descent"], "Method"], [1418, 1420, ["Parameter", "optimization"], "Task"], [1492, 1494, ["model", "selection"], "Task"], [1501, 1503, ["model", "selection"], "Task"], [1553, 1554, ["classification"], "Task"], [1583, 1585, ["smile", "recognition"], "Task"], [1613, 1614, ["accuracies"], "Metric"], [1634, 1635, ["accuracies"], "Metric"], [1663, 1665, ["SVM", "baselines"], "Method"], [1727, 1728, ["accuracies"], "Metric"], [1793, 1795, ["Training", "time"], "Metric"], [1805, 1810, ["mouth", "and", "face", "input", "models"], "Method"], [1818, 1820, ["training", "time"], "Metric"], [1830, 1833, ["Tesla", "K40c", "GPU"], "Method"], [1839, 1841, ["training", "time"], "Metric"], [1874, 1876, ["model", "selection"], "Method"], [1889, 1891, ["Deep", "learning"], "Method"], [1893, 1895, ["umbrella", "term"], "Method"], [1897, 1899, ["neural", "networks"], "Method"], [1906, 1908, ["training", "methods"], "Method"], [1919, 1922, ["action", "unit", "recognition"], "Task"], [1923, 1925, ["smile", "recognition"], "Task"], [1937, 1938, ["accuracy"], "Metric"], [1949, 1951, ["model", "selection"], "Method"], [589, 590, ["convolutional"], "Method"], [718, 719, ["DISFA"], "Material"], [869, 870, ["DISFA"], "Material"], [898, 901, ["action", "unit", "recognition"], "Task"], [1930, 1931, ["convolutional"], "Method"]]} {"doc_id": "398c296d0cc7f9d180f84969f8937e6d3a413796", "text": ["document", ":", "Multi", "-", "column", "Deep", "Neural", "Networks", "for", "Image", "Classification", "Traditional", "methods", "of", "computer", "vision", "and", "machine", "learning", "can", "not", "match", "human", "performance", "on", "tasks", "such", "as", "the", "recognition", "of", "handwritten", "digits", "or", "traffic", "signs", ".", "Our", "biologically", "plausible", "deep", "artificial", "neural", "network", "architectures", "can", ".", "Small", "(", "often", "minimal", ")", "receptive", "fields", "of", "convolutional", "winner", "-", "take", "-", "all", "neurons", "yield", "large", "network", "depth", ",", "resulting", "in", "roughly", "as", "many", "sparsely", "connected", "neural", "layers", "as", "found", "in", "mammals", "between", "retina", "and", "visual", "cortex", ".", "Only", "winner", "neurons", "are", "trained", ".", "Several", "deep", "neural", "columns", "become", "experts", "on", "inputs", "preprocessed", "in", "different", "ways", ";", "their", "predictions", "are", "averaged", ".", "Graphics", "cards", "allow", "for", "fast", "training", ".", "On", "the", "very", "competitive", "MNIST", "handwriting", "benchmark", ",", "our", "method", "is", "the", "first", "to", "achieve", "near", "-", "human", "performance", ".", "On", "a", "traffic", "sign", "recognition", "benchmark", "it", "outperforms", "humans", "by", "a", "factor", "of", "two", ".", "We", "also", "improve", "the", "state", "-", "of", "-", "the", "-", "art", "on", "a", "plethora", "of", "common", "image", "classification", "benchmarks", ".", "04", "-", "12", "section", ":", "Introduction", "Recent", "publications", "suggest", "that", "unsupervised", "pre", "-", "training", "of", "deep", ",", "hierarchical", "neural", "networks", "improves", "supervised", "pattern", "classification", ".", "Here", "we", "train", "such", "nets", "by", "simple", "online", "back", "-", "propagation", ",", "setting", "new", ",", "greatly", "improved", "records", "on", "MNIST", ",", "Latin", "letters", ",", "Chinese", "characters", ",", "traffic", "signs", ",", "NORB", "(", "jittered", ",", "cluttered", ")", "and", "CIFAR10", "benchmarks", ".", "We", "focus", "on", "deep", "convolutional", "neural", "networks", "(", "DNN", ")", ",", "introduced", "by", ",", "improved", "by", ",", "refined", "and", "simplified", "by", ".", "Lately", ",", "DNN", "proved", "their", "mettle", "on", "data", "sets", "ranging", "from", "handwritten", "digits", "(", "MNIST", ")", ",", "handwritten", "characters", "to", "3D", "toys", "(", "NORB", ")", "and", "faces", ".", "DNNs", "fully", "unfold", "their", "potential", "when", "they", "are", "big", "and", "deep", ".", "But", "training", "them", "requires", "weeks", ",", "months", ",", "even", "years", "on", "CPUs", ".", "High", "data", "transfer", "latency", "prevents", "multi", "-", "threading", "and", "multi", "-", "CPU", "code", "from", "saving", "the", "situation", ".", "In", "recent", "years", ",", "however", ",", "fast", "parallel", "neural", "net", "code", "for", "graphics", "cards", "(", "GPUs", ")", "has", "overcome", "this", "problem", ".", "Carefully", "designed", "GPU", "code", "for", "image", "classification", "can", "be", "up", "to", "two", "orders", "of", "magnitude", "faster", "than", "its", "CPU", "counterpart", ".", "Hence", ",", "to", "train", "huge", "DNN", "in", "hours", "or", "days", ",", "we", "implement", "them", "on", "GPU", ",", "building", "upon", "the", "work", "of", ".", "The", "training", "algorithm", "is", "fully", "online", ",", "i.e.", "weight", "updates", "occur", "after", "each", "error", "back", "-", "propagation", "step", ".", "We", "will", "show", "that", "properly", "trained", "big", "and", "deep", "DNNs", "can", "outperform", "all", "previous", "methods", ",", "and", "demonstrate", "that", "unsupervised", "initialization", "/", "pretraining", "is", "not", "necessary", "(", "although", "we", "do", "n’t", "deny", "that", "it", "might", "help", "sometimes", ",", "especially", "for", "small", "datasets", ")", ".", "We", "also", "show", "how", "combining", "several", "DNN", "columns", "into", "a", "Multi", "-", "column", "DNN", "(", "MCDNN", ")", "further", "decreases", "the", "error", "rate", "by", "30", "-", "40", "%", ".", "section", ":", "Architecture", "The", "initially", "random", "weights", "of", "the", "DNN", "are", "iteratively", "trained", "to", "minimize", "the", "classification", "error", "on", "a", "set", "of", "labeled", "training", "images", ";", "generalization", "performance", "is", "then", "tested", "on", "a", "separate", "set", "of", "test", "images", ".", "Our", "architecture", "does", "this", "by", "combining", "several", "techniques", "in", "a", "novel", "way", ":", "(", "1", ")", "Unlike", "the", "shallow", "NN", "used", "in", "many", "1990s", "applications", ",", "ours", "are", "deep", ",", "inspired", "by", "the", "Neocognitron", ",", "with", "many", "(", "6", "-", "10", ")", "layers", "of", "non", "-", "linear", "neurons", "stacked", "on", "top", "of", "each", "other", ",", "comparable", "to", "the", "number", "of", "layers", "found", "between", "retina", "and", "visual", "cortex", "of", "macaque", "monkeys", ".", "(", "2", ")", "It", "was", "shown", "that", "such", "multi", "-", "layered", "DNN", "are", "hard", "to", "train", "by", "standard", "gradient", "descent", ",", "the", "method", "of", "choice", "from", "a", "mathematical", "/", "algorithmic", "point", "of", "view", ".", "Today", "’s", "computers", ",", "however", ",", "are", "fast", "enough", "for", "this", ",", "more", "than", "60000", "times", "faster", "than", "those", "of", "the", "early", "90s", ".", "Carefully", "designed", "code", "for", "massively", "parallel", "graphics", "processing", "units", "(", "GPUs", "normally", "used", "for", "video", "games", ")", "allows", "for", "gaining", "an", "additional", "speedup", "factor", "of", "50", "-", "100", "over", "serial", "code", "for", "standard", "computers", ".", "Given", "enough", "labeled", "data", ",", "our", "networks", "do", "not", "need", "additional", "heuristics", "such", "as", "unsupervised", "pre", "-", "training", "or", "carefully", "prewired", "synapses", ".", "(", "3", ")", "The", "DNN", "of", "this", "paper", "(", "Fig", ".", "[", "reference", "]", "a", ")", "have", "2", "-", "dimensional", "layers", "of", "winner", "-", "take", "-", "all", "neurons", "with", "overlapping", "receptive", "fields", "whose", "weights", "are", "shared", ".", "Given", "some", "input", "pattern", ",", "a", "simple", "max", "pooling", "technique", "determines", "winning", "neurons", "by", "partitioning", "layers", "into", "quadratic", "regions", "of", "local", "inhibition", ",", "selecting", "the", "most", "active", "neuron", "of", "each", "region", ".", "The", "winners", "of", "some", "layer", "represent", "a", "smaller", ",", "down", "-", "sampled", "layer", "with", "lower", "resolution", ",", "feeding", "the", "next", "layer", "in", "the", "hierarchy", ".", "The", "approach", "is", "inspired", "by", "Hubel", "and", "Wiesel", "’s", "seminal", "work", "on", "the", "cat", "’s", "primary", "visual", "cortex", ",", "which", "identified", "orientation", "-", "selective", "simple", "cells", "with", "overlapping", "local", "receptive", "fields", "and", "complex", "cells", "performing", "down", "-", "sampling", "-", "like", "operations", ".", "(", "4", ")", "Note", "that", "at", "some", "point", "down", "-", "sampling", "automatically", "leads", "to", "the", "first", "1", "-", "dimensional", "layer", ".", "From", "then", "on", ",", "only", "trivial", "1", "-", "dimensional", "winner", "-", "take", "-", "all", "regions", "are", "possible", ",", "that", "is", ",", "the", "top", "part", "of", "the", "hierarchy", "becomes", "a", "standard", "multi", "-", "layer", "perceptron", "(", "MLP", ")", ".", "Receptive", "fields", "and", "winner", "-", "take", "-", "all", "regions", "of", "our", "DNN", "often", "are", "(", "near", "-)", "minimal", ",", "e.g.", ",", "only", "2x2", "or", "3x3", "neurons", ".", "This", "results", "in", "(", "near", "-)", "maximal", "depth", "of", "layers", "with", "non", "-", "trivial", "(", "2", "-", "dimensional", ")", "winner", "-", "take", "-", "all", "regions", ".", "In", "fact", ",", "insisting", "on", "minimal", "2x2", "fields", "automatically", "defines", "the", "entire", "deep", "architecture", ",", "apart", "from", "the", "number", "of", "different", "convolutional", "kernels", "per", "layer", "and", "the", "depth", "of", "the", "plain", "MLP", "on", "top", ".", "(", "5", ")", "Only", "winner", "neurons", "are", "trained", ",", "that", "is", ",", "other", "neurons", "can", "not", "forget", "what", "they", "learnt", "so", "far", ",", "although", "they", "may", "be", "affected", "by", "weight", "changes", "in", "more", "peripheral", "layers", ".", "The", "resulting", "decrease", "of", "synaptic", "changes", "per", "time", "interval", "corresponds", "to", "biologically", "plausible", "reduction", "of", "energy", "consumption", ".", "Our", "training", "algorithm", "is", "fully", "online", ",", "i.e.", "weight", "updates", "occur", "after", "each", "gradient", "computation", "step", ".", "(", "6", ")", "Inspired", "by", "microcolumns", "of", "neurons", "in", "the", "cerebral", "cortex", ",", "we", "combine", "several", "DNN", "columns", "to", "form", "a", "Multi", "-", "column", "DNN", "(", "MCDNN", ")", ".", "Given", "some", "input", "pattern", ",", "the", "predictions", "of", "all", "columns", "are", "democratically", "averaged", ".", "Before", "training", ",", "the", "weights", "(", "synapses", ")", "of", "all", "columns", "are", "randomly", "initialized", ".", "Various", "columns", "can", "be", "trained", "on", "the", "same", "inputs", ",", "or", "on", "inputs", "preprocessed", "in", "different", "ways", ".", "The", "latter", "helps", "to", "reduce", "both", "error", "rate", "and", "number", "of", "columns", "required", "to", "reach", "a", "given", "accuracy", ".", "The", "MCDNN", "architecture", "and", "its", "training", "and", "testing", "procedures", "are", "illustrated", "in", "Figure", "[", "reference", "]", ".", "section", ":", "Experiments", "In", "the", "following", "we", "give", "a", "detailed", "description", "of", "all", "the", "experiments", "we", "performed", ".", "We", "evaluate", "our", "architecture", "on", "various", "commonly", "used", "object", "recognition", "benchmarks", "and", "improve", "the", "state", "-", "of", "-", "the", "-", "art", "on", "all", "of", "them", ".", "The", "description", "of", "the", "DNN", "architecture", "used", "for", "the", "various", "experiments", "is", "given", "in", "the", "following", "way", ":", "2x48x48", "-", "100C5", "-", "MP2", "-", "100C5", "-", "MP2", "-", "100C4", "-", "MP2", "-", "300N", "-", "100N", "-", "6N", "represents", "a", "net", "with", "2", "input", "images", "of", "size", "48x48", ",", "a", "convolutional", "layer", "with", "100", "maps", "and", "5x5", "filters", ",", "a", "max", "-", "pooling", "layer", "over", "non", "overlapping", "regions", "of", "size", "2x2", ",", "a", "convolutional", "layer", "with", "100", "maps", "and", "4x4", "filters", ",", "a", "max", "-", "pooling", "layer", "over", "non", "overlapping", "regions", "of", "size", "2x2", ",", "a", "fully", "connected", "layer", "with", "300", "hidden", "units", ",", "a", "fully", "connected", "layer", "with", "100", "hidden", "units", "and", "a", "fully", "connected", "output", "layer", "with", "6", "neurons", "(", "one", "per", "class", ")", ".", "We", "use", "a", "scaled", "hyperbolic", "tangent", "activation", "function", "for", "convolutional", "and", "fully", "connected", "layers", ",", "a", "linear", "activation", "function", "for", "max", "-", "pooling", "layers", "and", "a", "softmax", "activation", "function", "for", "the", "output", "layer", ".", "All", "DNN", "are", "trained", "using", "on", "-", "line", "gradient", "descent", "with", "an", "annealed", "learning", "rate", ".", "During", "training", ",", "images", "are", "continually", "translated", ",", "scaled", "and", "rotated", "(", "even", "elastically", "distorted", "in", "case", "of", "characters", ")", ",", "whereas", "only", "the", "original", "images", "are", "used", "for", "validation", ".", "Training", "ends", "once", "the", "validation", "error", "is", "zero", "or", "when", "the", "learning", "rate", "reaches", "its", "predetermined", "minimum", ".", "Initial", "weights", "are", "drawn", "from", "a", "uniform", "random", "distribution", "in", "the", "range", ".", "subsection", ":", "MNIST", "The", "original", "MNIST", "digits", "are", "normalized", "such", "that", "the", "width", "or", "height", "of", "the", "bounding", "box", "equals", "20", "pixels", ".", "Aspect", "ratios", "for", "various", "digits", "vary", "strongly", "and", "we", "therefore", "create", "six", "additional", "datasets", "by", "normalizing", "digit", "width", "to", "10", ",", "12", ",", "14", ",", "16", ",", "18", ",", "20", "pixels", ".", "This", "is", "like", "seeing", "the", "data", "from", "different", "angles", ".", "We", "train", "five", "DNN", "columns", "per", "normalization", ",", "resulting", "in", "a", "total", "of", "35", "columns", "for", "the", "entire", "MCDNN", ".", "All", "1x29x29", "-", "20C4", "-", "MP2", "-", "40C5", "-", "MP3", "-", "150N", "-", "10N", "DNN", "are", "trained", "for", "around", "800", "epochs", "with", "an", "annealed", "learning", "rate", "(", "i.e.", "initialized", "with", "0.001", "multiplied", "by", "a", "factor", "of", "0.993", "/", "epoch", "until", "it", "reaches", "0.00003", ")", ".", "Training", "a", "DNN", "takes", "almost", "14", "hours", "and", "after", "500", "training", "epochs", "little", "additional", "improvement", "is", "observed", ".", "During", "training", "the", "digits", "are", "randomly", "distorted", "before", "each", "epoch", "(", "see", "Fig", ".", "[", "reference", "]", "a", "for", "representative", "characters", "and", "their", "distorted", "versions", ")", ".", "The", "internal", "state", "of", "a", "single", "DNN", "is", "depicted", "in", "Figure", "[", "reference", "]", "c", ",", "where", "a", "particular", "digit", "is", "forward", "propagated", "through", "a", "trained", "network", "and", "all", "activations", "together", "with", "the", "network", "weights", "are", "plotted", ".", "Results", "of", "all", "individual", "nets", "and", "various", "MCDNN", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "MCDNN", "of", "5", "nets", "trained", "with", "the", "same", "preprocessor", "achieve", "better", "results", "than", "their", "constituent", "DNNs", ",", "except", "for", "original", "images", "(", "Tab", ".", "[", "reference", "]", ")", ".", "The", "MCDNN", "has", "a", "very", "low", "0.23", "%", "error", "rate", ",", "improving", "state", "of", "the", "art", "by", "at", "least", "34", "%", "(", "Tab", ".", "[", "reference", "]", ")", ".", "This", "is", "the", "first", "time", "an", "artificial", "method", "comes", "close", "to", "the", "0.2", "%", "error", "rate", "of", "humans", "on", "this", "task", ".", "Many", "of", "the", "wrongly", "classified", "digits", "either", "contain", "broken", "or", "strange", "strokes", ",", "or", "have", "wrong", "labels", ".", "The", "23", "errors", "(", "Fig", ".", "[", "reference", "]", "b", ")", "are", "associated", "with", "20", "correct", "second", "guesses", ".", "We", "also", "trained", "a", "single", "DNN", "on", "all", "7", "datasets", "simultaneously", "which", "yielded", "worse", "result", "(", "0.52", "%", ")", "than", "both", "MCDNN", "and", "their", "individual", "DNN", ".", "This", "shows", "that", "the", "improvements", "come", "from", "the", "MCDNN", "and", "not", "from", "using", "more", "preprocessed", "data", ".", "How", "are", "the", "MCDNN", "errors", "affected", "by", "the", "number", "of", "preprocessors", "?", "We", "train", "5", "DNNs", "on", "all", "7", "datasets", ".", "A", "MCDNN", "’", "out", "-", "of", "-", "7", "’", "(", "from", "1", "to", "7", ")", "averages", "nets", "trained", "on", "datasets", ".", "Table", "[", "reference", "]", "shows", "that", "more", "preprocessing", "results", "in", "lower", "MCDNN", "error", ".", "We", "also", "train", "5", "DNN", "for", "each", "odd", "normalization", ",", "i.e.", "W11", ",", "W13", ",", "W15", ",", "W17", "and", "W19", ".", "The", "60", "-", "net", "MCDNN", "performs", "(", "0.24", "%", ")", "similarly", "to", "the", "35", "-", "net", "MCDNN", ",", "indicating", "that", "additional", "preprocessing", "does", "not", "further", "improve", "recognition", ".", "We", "conclude", "that", "MCDNN", "outperform", "DNN", "trained", "on", "the", "same", "data", ",", "and", "that", "different", "preprocessors", "further", "decrease", "the", "error", "rate", ".", "subsection", ":", "NIST", "SD", "19", "The", "35", "-", "columns", "MCDNN", "architecture", "and", "preprocessing", "used", "for", "MNIST", "are", "also", "applied", "to", "Latin", "characters", "from", "NIST", "SD", "19", ".", "For", "all", "tasks", "our", "MCDNN", "achieves", "recognition", "rates", "1.5", "-", "5", "times", "better", "than", "any", "published", "result", "(", "Tab", ".", "[", "reference", "]", ")", ".", "In", "total", "there", "are", "82000", "characters", "in", "the", "test", "set", ",", "but", "there", "are", "many", "more", "easy", "to", "classify", "digits", "(", "58000", ")", "than", "hard", "to", "classify", "letters", "(", "24000", ")", ".", "This", "explains", "the", "lower", "overall", "error", "rate", "of", "the", "62", "-", "class", "problem", "compared", "to", "the", "52", "-", "class", "letters", "problem", ".", "From", "all", "errors", "of", "the", "62", "-", "class", "problem", "3", "%", "of", "the", "58000", "digits", "are", "misclassified", "and", "33", "%", "of", "the", "24000", "letters", "are", "misclassified", ".", "Letters", "are", "in", "general", "more", "difficult", "to", "classify", ",", "but", "there", "is", "also", "a", "higher", "amount", "of", "confusion", "between", "similar", "lower", "-", "and", "upper", "-", "case", "letters", "such", "as", "i", ",", "I", "and", "o", ",", "O", "for", "example", ".", "Indeed", ",", "error", "rates", "for", "the", "case", "insensitive", "task", "drop", "from", "21", "%", "to", "7.37", "%", ".", "If", "the", "confused", "upper", "-", "and", "lower", "-", "case", "classes", "are", "merged", ",", "resulting", "in", "37", "different", "classes", ",", "the", "error", "rate", "is", "only", "slightly", "bigger", "(", "7.99", "%", ")", ".", "Upper", "-", "case", "letters", "are", "far", "easier", "to", "classify", "(", "1.83", "%", "error", "rate", ")", "than", "lowercase", "letters", "(", "7.47", "%", ")", "due", "to", "the", "smaller", "writer", "dependent", "in", "-", "class", "variability", ".", "For", "a", "detailed", "analysis", "of", "all", "the", "errors", "and", "confusions", "between", "different", "classes", ",", "the", "confusion", "matrix", "is", "most", "informative", "(", "Supplementary", "material", "Fig", ".", "S1", ")", ".", "subsection", ":", "Chinese", "characters", "Compared", "to", "Latin", "character", "recognition", ",", "isolated", "Chinese", "character", "recognition", "is", "a", "much", "harder", "problem", ",", "mainly", "because", "of", "the", "much", "larger", "category", "set", ",", "but", "also", "because", "of", "wide", "variability", "of", "writing", "styles", ",", "and", "the", "confusion", "between", "similar", "characters", ".", "We", "use", "a", "dataset", "from", "the", "Institute", "of", "Automation", "of", "Chinese", "Academy", "of", "Sciences", "(", "CASIA", ")", ",", "which", "contains", "300", "samples", "for", "each", "of", "3755", "characters", "(", "in", "GB1", "set", ")", ".", "This", "resulted", "in", "a", "data", "set", "with", "more", "than", "1", "Million", "characters", "(", "3", "GB", "of", "data", ")", "which", "posed", "a", "major", "computational", "challenge", "even", "to", "our", "system", ".", "Without", "our", "fast", "GPU", "implementation", "the", "nets", "on", "this", "task", "would", "train", "for", "more", "than", "one", "year", ".", "Only", "the", "forward", "propagation", "of", "the", "training", "set", "takes", "27h", "on", "a", "normal", "CPU", ",", "and", "training", "a", "single", "epoch", "would", "consequently", "have", "lasted", "several", "days", ".", "On", "our", "fast", "GPU", "implementation", "on", "the", "other", "hand", ",", "training", "a", "single", "epoch", "takes", "3.4h", ",", "which", "makes", "it", "feasible", "to", "train", "a", "net", "within", "a", "few", "days", "instead", "of", "many", "months", ".", "We", "train", "following", "DNN", ",", "1x48x48", "-", "100C3", "-", "MP2", "-", "200C2", "-", "MP2", "-", "300C2", "-", "MP2", "-", "400C2", "-", "MP2", "-", "500N", "-", "3755N", ",", "on", "offline", "as", "well", "as", "on", "online", "characters", ".", "For", "the", "offline", "character", "recognition", "task", ",", "we", "resize", "all", "characters", "to", "40x40", "pixels", "and", "place", "them", "in", "the", "center", "of", "a", "48x48", "image", ".", "The", "contrast", "of", "each", "image", "is", "normalized", "independently", ".", "As", "suggested", "by", "the", "organizers", ",", "the", "first", "240", "writers", "from", "the", "database", "CASIA", "-", "HWDB1.1", "are", "used", "for", "training", "and", "the", "remaining", "60", "writers", "are", "used", "for", "testing", ".", "The", "total", "numbers", "of", "training", "and", "test", "characters", "are", "938679", "and", "234228", ",", "respectively", ".", "For", "the", "online", "dataset", ",", "we", "draw", "each", "character", "from", "its", "list", "of", "coordinates", ",", "resize", "the", "resulting", "images", "to", "40x40", "pixels", "and", "place", "them", "in", "the", "center", "of", "a", "48x48", "image", ".", "Additionally", ",", "we", "smooth", "-", "out", "the", "resulting", "images", "with", "a", "Gaussian", "blur", "filter", "over", "a", "3x3", "pixel", "neighborhood", "and", "uniform", "standard", "deviation", "of", "0.75", ".", "As", "suggested", "by", "the", "organizers", ",", "the", "characters", "of", "240", "writers", "from", "database", "CASIA", "-", "OLHWDB1.1", "are", "used", "for", "training", "the", "classifier", "and", "the", "characters", "of", "the", "remaining", "60", "writers", "are", "used", "for", "testing", ".", "The", "resulting", "numbers", "of", "training", "and", "test", "characters", "are", "939564", "and", "234800", ",", "respectively", ".", "All", "methods", "previously", "applied", "to", "this", "dataset", "perform", "some", "feature", "extraction", "followed", "by", "a", "dimensionality", "reduction", ",", "whereas", "our", "method", "directly", "works", "on", "raw", "pixel", "intensities", "and", "learns", "the", "feature", "extraction", "and", "dimensionality", "reduction", "in", "a", "supervised", "way", ".", "On", "the", "offline", "task", "we", "obtain", "an", "error", "rate", "of", "6.5", "%", "compared", "to", "10.01", "%", "of", "the", "best", "method", ".", "Even", "though", "much", "information", "is", "lost", "when", "drawing", "a", "character", "from", "it", "’s", "coordinate", "sequence", ",", "we", "obtain", "a", "recognition", "rate", "of", "5.61", "%", "on", "the", "online", "task", "compared", "to", "7.61", "%", "of", "the", "best", "method", ".", "We", "conclude", "that", "on", "this", "very", "hard", "classification", "problem", ",", "with", "many", "classes", "(", "3755", ")", "and", "relatively", "few", "samples", "per", "class", "(", "240", ")", ",", "our", "fully", "supervised", "DNN", "beats", "the", "current", "state", "-", "of", "-", "the", "-", "art", "methods", "by", "a", "large", "margin", ".", "subsection", ":", "Traffic", "signs", "Recognizing", "traffic", "signs", "is", "essential", "for", "the", "automotive", "industry", "’s", "efforts", "in", "the", "field", "of", "driver", "’s", "assistance", ",", "and", "for", "many", "other", "traffic", "-", "related", "applications", ".", "We", "use", "the", "GTSRB", "traffic", "sign", "dataset", ".", "The", "original", "color", "images", "contain", "one", "traffic", "sign", "each", ",", "with", "a", "border", "of", "10", "%", "around", "the", "sign", ".", "They", "vary", "in", "size", "from", "to", "pixels", "and", "are", "not", "necessarily", "square", ".", "The", "actual", "traffic", "sign", "is", "not", "always", "centered", "within", "the", "image", ";", "its", "bounding", "box", "is", "part", "of", "the", "annotations", ".", "The", "training", "set", "consists", "of", "26640", "images", ";", "the", "test", "set", "of", "12569", "images", ".", "We", "crop", "all", "images", "and", "process", "only", "within", "the", "bounding", "box", ".", "Our", "DNN", "implementation", "requires", "all", "training", "images", "to", "be", "of", "equal", "size", ".", "After", "visual", "inspection", "of", "the", "image", "size", "distribution", "we", "resize", "all", "images", "to", "pixels", ".", "As", "a", "consequence", ",", "scaling", "factors", "along", "both", "axes", "are", "different", "for", "traffic", "signs", "with", "rectangular", "bounding", "boxes", ".", "Resizing", "forces", "them", "to", "have", "square", "bounding", "boxes", ".", "Our", "MCDNN", "is", "the", "only", "artificial", "method", "to", "outperform", "humans", ",", "who", "produced", "twice", "as", "many", "errors", ".", "Since", "traffic", "signs", "greatly", "vary", "in", "illumination", "and", "contrast", ",", "standard", "image", "preprocessing", "methods", "are", "used", "to", "enhance", "/", "normalize", "them", "(", "Fig", ".", "[", "reference", "]", "a", "and", "supplementary", "material", ")", ".", "For", "each", "dataset", "five", "DNN", "are", "trained", "(", "architecture", ":", "3x48x48", "-", "100C7", "-", "MP2", "-", "150C4", "-", "150MP2", "-", "250C4", "-", "250MP2", "-", "300N", "-", "43N", ")", ",", "resulting", "in", "a", "MCDNN", "with", "25", "columns", ",", "achieving", "an", "error", "rate", "of", "0.54", "%", "on", "the", "test", "set", ".", "Figure", "[", "reference", "]", "b", "depicts", "all", "errors", ",", "plus", "ground", "truth", "and", "first", "and", "second", "predictions", ".", "Over", "80", "%", "of", "the", "68", "errors", "are", "associated", "with", "correct", "second", "predictions", ".", "Erroneously", "predicted", "class", "probabilities", "tend", "to", "be", "very", "low", "—", "here", "the", "MCDNN", "is", "quite", "unsure", "about", "its", "classifications", ".", "In", "general", ",", "however", ",", "it", "is", "very", "confident", "—", "most", "of", "its", "predicted", "class", "probabilities", "are", "close", "to", "one", "or", "zero", ".", "Rejecting", "only", "1", "%", "percent", "of", "all", "images", "(", "confidence", "below", "0.51", ")", "results", "in", "an", "even", "lower", "error", "rate", "of", "0.24", "%", ".", "To", "reach", "an", "error", "rate", "of", "0.01", "%", "(", "a", "single", "misclassification", ")", ",", "only", "6.67", "%", "of", "the", "images", "have", "to", "be", "rejected", "(", "confidence", "below", "0.94", ")", ".", "Our", "method", "outperforms", "the", "second", "best", "algorithm", "by", "a", "factor", "of", "3", ".", "It", "takes", "37", "hours", "to", "train", "the", "MCDNN", "with", "25", "columns", "on", "four", "GPUs", ".", "The", "trained", "MCDNN", "can", "check", "87", "images", "per", "second", "on", "one", "GPU", "(", "and", "2175", "images", "/", "s", "/", "DNN", ")", ".", "subsection", ":", "CIFAR", "10", "CIFAR10", "is", "a", "set", "of", "natural", "color", "images", "of", "32x32", "pixels", ".", "It", "contains", "10", "classes", ",", "each", "with", "5000", "training", "samples", "and", "1000", "test", "samples", ".", "Images", "vary", "greatly", "within", "each", "class", ".", "They", "are", "not", "necessarily", "centered", ",", "may", "contain", "only", "parts", "of", "the", "object", ",", "and", "show", "different", "backgrounds", ".", "Subjects", "may", "vary", "in", "size", "by", "an", "order", "of", "magnitude", "(", "i.e.", ",", "some", "images", "show", "only", "the", "head", "of", "a", "bird", ",", "others", "an", "entire", "bird", "from", "a", "distance", ")", ".", "Colors", "and", "textures", "of", "objects", "/", "animals", "also", "vary", "greatly", ".", "Our", "DNN", "input", "layers", "have", "three", "maps", ",", "one", "for", "each", "color", "channel", "(", "RGB", ")", ".", "We", "use", "a", "10", "-", "layer", "architecture", "with", "very", "small", "kernels", ":", "3x32x32", "-", "300C3", "-", "MP2", "-", "300C2", "-", "MP2", "-", "300C3", "-", "MP2", "-", "300C2", "-", "MP2", "-", "300N", "-", "100N", "-", "10N.", "Just", "like", "for", "MNIST", ",", "the", "initial", "learning", "rate", "0.001", "decays", "by", "a", "factor", "of", "0.993", "after", "every", "epoch", ".", "Transforming", "CIFAR", "color", "images", "to", "gray", "scale", "reduces", "input", "layer", "complexity", "but", "increases", "error", "rates", ".", "Hence", "we", "stick", "to", "the", "original", "color", "images", ".", "As", "for", "MNIST", ",", "augmenting", "the", "training", "set", "with", "randomly", "(", "by", "at", "most", "5", "%", ")", "translated", "images", "greatly", "decreases", "the", "error", "from", "28", "%", "to", "20", "%", "(", "the", "NN", "-", "inherent", "local", "translation", "invariance", "by", "itself", "is", "not", "sufficient", ")", ".", "By", "additional", "scaling", "(", "up", "to", "15", "%", ")", ",", "rotation", "(", "up", "to", ")", ",", "and", "up", "to", "15", "%", "translation", ",", "the", "individual", "net", "errors", "decrease", "by", "another", "3", "%", "(", "Tab", ".", "[", "reference", "]", ")", ".", "The", "above", "small", "maximal", "bounds", "prevent", "loss", "of", "too", "much", "information", "leaked", "beyond", "the", "pixels", "rectangle", ".", "We", "repeat", "the", "experiment", "with", "different", "random", "initializations", "and", "compute", "mean", "and", "standard", "deviation", "of", "the", "error", ",", "which", "is", "rather", "small", "for", "original", "images", ",", "showing", "that", "our", "DNN", "are", "robust", ".", "Our", "MCDNN", "obtains", "a", "very", "low", "error", "rate", "of", "11.21", "%", ",", "greatly", "rising", "the", "bar", "for", "this", "benchmark", ".", "The", "confusion", "matrix", "(", "Figure", "[", "reference", "]", ")", "shows", "that", "the", "MCDNN", "almost", "perfectly", "separates", "animals", "from", "artifacts", ",", "except", "for", "planes", "and", "birds", ",", "which", "seems", "natural", ",", "although", "humans", "easily", "distinguish", "almost", "all", "the", "incorrectly", "classified", "images", ",", "even", "if", "many", "are", "cluttered", "or", "contain", "only", "parts", "of", "the", "objects", "/", "animals", "(", "see", "false", "positive", "and", "false", "negative", "images", "in", "Figure", "[", "reference", "]", ")", ".", "There", "are", "many", "confusions", "between", "different", "animals", ";", "the", "frog", "class", "collects", "most", "false", "positives", "from", "other", "animal", "classes", ",", "with", "very", "few", "false", "negatives", ".", "As", "expected", ",", "cats", "are", "hard", "to", "tell", "from", "dogs", ",", "collectively", "causing", "15.25", "%", "of", "the", "errors", ".", "The", "MCDNN", "with", "8", "columns", "(", "four", "trained", "on", "original", "data", "and", "one", "trained", "for", "each", "preprocessing", "used", "also", "for", "traffic", "signs", ")", "reaches", "a", "low", "11.21", "%", "error", "rate", ",", "far", "better", "than", "any", "other", "algorithm", ".", "subsection", ":", "NORB", "We", "test", "a", "MCDNN", "with", "four", "columns", "on", "NORB", "(", "jittered", "-", "cluttered", ")", ",", "a", "collection", "of", "stereo", "images", "of", "3D", "models", "(", "Figure", "[", "reference", "]", ")", ".", "The", "objects", "are", "centrally", "placed", "on", "randomly", "chosen", "backgrounds", ",", "and", "there", "is", "also", "cluttering", "from", "a", "peripherally", "placed", "second", "object", ".", "This", "database", "is", "designed", "for", "experimenting", "with", "3D", "object", "recognition", "from", "shape", ".", "It", "contains", "images", "of", "50", "toys", "belonging", "to", "5", "generic", "categories", ":", "four", "-", "legged", "animals", ",", "human", "figures", ",", "airplanes", ",", "trucks", ",", "and", "cars", ".", "The", "objects", "were", "imaged", "by", "two", "cameras", "under", "6", "lighting", "conditions", ",", "9", "elevations", "(", "30", "to", "70", "degrees", "every", "5", "degrees", ")", ",", "and", "18", "azimuths", "(", "0", "to", "340", "every", "20", "degrees", ")", ".", "The", "training", "set", "has", "10", "folds", "of", "29160", "images", "each", "for", "a", "total", "of", "291600", "images", ";", "the", "testing", "set", "consists", "of", "two", "folds", "totalizing", "58320", "images", ".", "No", "preprocessing", "is", "used", "for", "this", "dataset", ".", "We", "scale", "down", "images", "from", "the", "original", "108x108", "to", "48x48", "pixels", ".", "This", "size", "is", "big", "enough", "to", "preserve", "the", "details", "present", "in", "images", "and", "small", "enough", "to", "allow", "fast", "training", ".", "We", "perform", "two", "rounds", "of", "experiments", ",", "using", "only", "the", "first", "two", "folds", "(", "to", "compare", "with", "previous", "results", "that", "do", "not", "use", "the", "entire", "training", "data", ")", "and", "using", "all", "training", "data", ".", "We", "tested", "several", "distortion", "parameters", "with", "small", "nets", "and", "found", "that", "maximum", "rotation", "of", ",", "maximum", "translation", "of", "15", "%", "and", "maximum", "scaling", "of", "15", "%", "are", "good", "choices", ",", "hence", "we", "use", "them", "for", "all", "NORB", "experiments", ".", "To", "compare", "to", "previous", "results", ",", "we", "first", "train", "only", "on", "the", "first", "2", "-", "folds", "of", "the", "data", ".", "The", "net", "architecture", "is", "deep", ",", "but", "has", "few", "maps", "per", "layer", ":", "2x48x48", "-", "50C5", "-", "MP2", "-", "50C5", "-", "MP2", "-", "50C4", "-", "MP2", "-", "300N", "-", "100N", "-", "6N.", "The", "learning", "rate", "setup", "is", ":", "eta", "start", "0.001", ";", "eta", "factor", "0.95", ";", "eta", "stop", "0.000003", ".", "Due", "to", "small", "net", "size", ",", "training", "is", "fast", "at", "156s", "/", "epoch", "for", "114", "epochs", ".", "Testing", "one", "sample", "requires", "0.5ms", ".", "Even", "when", "we", "use", "less", "data", "to", "train", ",", "the", "MCDNN", "greatly", "improves", "the", "state", "of", "the", "art", "from", "5", "%", "to", "3.57", "%", "(", "Table", "[", "reference", "]", ")", ".", "Our", "method", "is", "fast", "enough", "to", "process", "the", "entire", "training", "set", "though", ".", "We", "use", "the", "same", "architecture", "but", "double", "the", "number", "of", "maps", "when", "training", "with", "all", "10", "folds", ":", "2x48x48", "-", "100C5", "-", "MP2", "-", "100C5", "-", "MP2", "-", "100C4", "-", "MP2", "-", "300N", "-", "100N", "-", "6N.", "The", "learning", "rate", "setup", "remains", "the", "same", ".", "Training", "time", "increases", "to", "34min", "/", "epoch", "because", "the", "net", "is", "bigger", ",", "and", "we", "use", "five", "times", "more", "data", ".", "Testing", "one", "sample", "takes", "1.3ms", ".", "All", "of", "this", "pays", "off", ",", "resulting", "in", "a", "very", "low", "2.70", "%", "error", "rate", ",", "further", "improving", "the", "state", "of", "the", "art", ".", "Although", "NORB", "has", "only", "six", "classes", ",", "training", "and", "test", "instances", "sometimes", "differ", "greatly", ",", "making", "classification", "hard", ".", "More", "than", "50", "%", "of", "the", "errors", "are", "due", "to", "confusions", "between", "cars", "and", "trucks", ".", "Considering", "second", "predictions", ",", "too", ",", "the", "error", "rate", "drops", "from", "2.70", "%", "to", "0.42", "%", ",", "showing", "that", "84", "%", "of", "the", "errors", "are", "associated", "with", "a", "correct", "second", "prediction", ".", "section", ":", "Conclusion", "This", "is", "the", "first", "time", "human", "-", "competitive", "results", "are", "reported", "on", "widely", "used", "computer", "vision", "benchmarks", ".", "On", "many", "other", "image", "classification", "datasets", "our", "MCDNN", "improves", "the", "state", "-", "of", "-", "the", "-", "art", "by", "30", "-", "80", "%", "(", "Tab", ".", "[", "reference", "]", ")", ".", "We", "drastically", "improve", "recognition", "rates", "on", "MNIST", ",", "NIST", "SD", "19", ",", "Chinese", "characters", ",", "traffic", "signs", ",", "CIFAR10", "and", "NORB", ".", "Our", "method", "is", "fully", "supervised", "and", "does", "not", "use", "any", "additional", "unlabeled", "data", "source", ".", "Single", "DNN", "already", "are", "sufficient", "to", "obtain", "new", "state", "-", "of", "-", "the", "-", "art", "results", ";", "combining", "them", "into", "MCDNNs", "yields", "further", "dramatic", "performance", "boosts", ".", "section", ":", "Acknowledgment", "This", "work", "was", "partially", "supported", "by", "a", "FP7", "-", "ICT", "-", "2009", "-", "6", "EU", "Grant", "under", "Project", "Code", "270247", ":", "A", "Neuro", "-", "dynamic", "Framework", "for", "Cognitive", "Robotics", ":", "Scene", "Representations", ",", "Behavioral", "Sequences", ",", "and", "Learning", ".", "bibliography", ":", "References", "section", ":", "Supplementary", "Material", "subsection", ":", "Experiment", "details", "subsubsection", ":", "NIST", "SD", "19", "The", "confusion", "matrix", "of", "the", "62", "characters", "task", "(", "Fig", ".", "[", "reference", "]", ")", "shows", "that", "most", "of", "the", "errors", "are", "due", "to", "confusions", "between", "digits", "and", "letters", "and", "between", "lower", "-", "and", "upper", "-", "case", "letters", ".", "Not", "very", "surprisingly", ",", "the", "confusion", "matrix", "for", "the", "digit", "task", "(", "Fig", ".", "[", "reference", "]", ")", "shows", "that", "confusions", "between", "fours", "and", "nines", "are", "the", "most", "common", "error", "source", ".", "For", "the", "52", "letter", "task", "(", "case", "sensitive", ")", "the", "confusion", "matrix", "(", "Fig", ".", "[", "reference", "]", ")", "shows", "that", "the", "MCDNN", "has", "mainly", "problems", "with", "upper", "-", "and", "lower", "-", "case", "confusions", "of", "the", "same", "letter", ".", "Other", "hard", "-", "to", "-", "distinguish", "classes", "are", ":", "’", "q", "’", "and", "’", "g", "’", ",", "’", "l", "’", "and", "’", "i", "’", ".", "For", "the", "upper", "-", "case", "letter", "task", "the", "confusion", "matrix", "(", "Figure", "[", "reference", "]", ")", "shows", "that", "the", "MCDNN", "has", "problems", "with", "letters", "of", "similar", "shape", ",", "i.e.", "’", "D", "’", ",", "and", "’", "O", "’", ",", "’", "V", "’", "and", "’", "U", "’", "etc", ".", "The", "total", "error", "of", "1.82", "%", "is", "very", "low", "though", ".", "For", "the", "lower", "-", "case", "letter", "task", "the", "confusion", "matrix", "(", "Fig", ".", "[", "reference", "]", ")", "shows", "that", "like", "with", "upper", "-", "case", "letters", ",", "the", "MCDNN", "has", "problems", "with", "letters", "of", "similar", "shapes", ",", "i.e.", "’", "g", "’", ",", "and", "’", "q", "’", ",", "’", "v", "’", "and", "’", "u", "’", "etc", ".", "But", "the", "total", "error", "is", "much", "higher", "(", "7.47", "%", ")", "than", "for", "the", "upper", "-", "case", "letters", "task", ".", "For", "the", "merged", "-", "case", "letter", "task", "(", "37", "classes", ")", "the", "confusion", "matrix", "(", "Figure", "[", "reference", "]", ")", "shows", "that", "the", "MCDNN", "has", "mostly", "problems", "with", "letters", "of", "similar", "shapes", ",", "i.e.", "’", "l", "’", ",", "and", "’", "i", "’", ".", "All", "upper", "-", "lower", "-", "case", "confusions", "of", "identical", "letters", "from", "the", "52", "class", "task", "vanish", ",", "the", "error", "shrinks", "by", "a", "factor", "of", "almost", "three", "down", "to", "7.99", "%", ".", "The", "experiments", "on", "different", "subsets", "of", "the", "62", "character", "task", "clearly", "show", "that", "it", "is", "very", "hard", "to", "distinguish", "between", "small", "and", "capital", "letters", ".", "Also", ",", "digits", "0", "and", "1", "are", "hard", "to", "separate", "from", "letters", "O", "and", "I.", "Many", "of", "these", "problems", "could", "be", "alleviated", "by", "incorporating", "context", "where", "possible", ".", "subsubsection", ":", "Traffic", "signs", "High", "contrast", "variation", "among", "the", "images", "calls", "for", "normalization", ".", "We", "test", "the", "following", "standard", "contrast", "normalizations", ":", "Image", "Adjustment", "(", "Imadjust", ")", "increases", "image", "contrast", "by", "mapping", "pixel", "intensities", "to", "new", "values", "such", "that", "1", "%", "of", "the", "data", "is", "saturated", "at", "low", "and", "high", "intensities", ".", "Histogram", "Equalization", "(", "Histeq", ")", "enhances", "contrast", "by", "transforming", "pixel", "intensities", "such", "that", "the", "output", "image", "histogram", "is", "roughly", "uniform", ".", "Adaptive", "Histogram", "Equalization", "(", "Adapthisteq", ")", "operates", "(", "unlike", "Histeq", ")", "on", "tiles", "rather", "than", "the", "entire", "image", ",", "we", "tiled", "the", "image", "in", "8", "nonoverlapping", "regions", "of", "6x6", "pixels", ".", "Each", "tile", "’s", "contrast", "is", "enhanced", "such", "that", "its", "histogram", "becomes", "roughly", "uniform", ".", "Contrast", "Normalization", "(", "Conorm", ")", "enhances", "edges", ",", "filtering", "the", "input", "image", "by", "a", "difference", "of", "Gaussians", ",", "using", "a", "filter", "size", "of", "5x5", "pixels", ".", "Note", "that", "the", "above", "normalizations", ",", "except", "Conorm", ",", "are", "performed", "in", "a", "color", "space", "with", "image", "intensity", "as", "one", "of", "its", "components", ".", "For", "this", "purpose", "we", "transform", "the", "image", "from", "RGB", "-", "to", "Lab", "-", "space", ",", "then", "perform", "normalization", ",", "then", "transform", "the", "result", "back", "to", "RGB", "-", "space", ".", "The", "effect", "of", "the", "four", "different", "normalizations", "is", "summarized", "in", "Figure", "[", "reference", "]", ",", "where", "histograms", "of", "pixel", "intensities", "together", "with", "original", "and", "normalized", "images", "are", "shown", ".", "The", "DNN", "have", "three", "maps", "for", "the", "input", "layer", ",", "one", "for", "each", "color", "channel", "(", "RGB", ")", ".", "The", "rest", "of", "the", "net", "architecture", "is", "detailed", "in", "Table", "[", "reference", "]", ".", "We", "use", "a", "10", "-", "layer", "architecture", "with", "very", "small", "max", "-", "pooling", "kernels", ".", "subsubsection", ":", "CIFAR10", "subsubsection", ":", "NORB"], "entity": [[2, 8, ["Multi", "-", "column", "Deep", "Neural", "Networks"], "Method"], [9, 11, ["Image", "Classification"], "Task"], [14, 16, ["computer", "vision"], "Method"], [17, 19, ["machine", "learning"], "Method"], [29, 33, ["recognition", "of", "handwritten", "digits"], "Task"], [34, 36, ["traffic", "signs"], "Task"], [38, 45, ["biologically", "plausible", "deep", "artificial", "neural", "network", "architectures"], "Method"], [55, 62, ["convolutional", "winner", "-", "take", "-", "all", "neurons"], "Method"], [93, 96, ["deep", "neural", "columns"], "Method"], [114, 116, ["fast", "training"], "Task"], [139, 142, ["traffic", "sign", "recognition"], "Task"], [168, 171, ["image", "classification", "benchmarks"], "Task"], [182, 186, ["unsupervised", "pre", "-", "training"], "Method"], [187, 192, ["deep", ",", "hierarchical", "neural", "networks"], "Method"], [193, 196, ["supervised", "pattern", "classification"], "Task"], [204, 208, ["online", "back", "-", "propagation"], "Method"], [216, 217, ["MNIST"], "Material"], [218, 220, ["Latin", "letters"], "Material"], [221, 223, ["Chinese", "characters"], "Material"], [227, 228, ["NORB"], "Material"], [234, 236, ["CIFAR10", "benchmarks"], "Material"], [240, 244, ["deep", "convolutional", "neural", "networks"], "Method"], [245, 246, ["DNN"], "Method"], [261, 262, ["DNN"], "Method"], [273, 274, ["MNIST"], "Material"], [287, 288, ["DNNs"], "Method"], [313, 316, ["data", "transfer", "latency"], "Metric"], [337, 341, ["parallel", "neural", "net", "code"], "Method"], [342, 344, ["graphics", "cards"], "Method"], [345, 346, ["GPUs"], "Method"], [354, 356, ["GPU", "code"], "Method"], [357, 359, ["image", "classification"], "Task"], [370, 372, ["CPU", "counterpart"], "Method"], [378, 379, ["DNN"], "Method"], [388, 389, ["GPU"], "Method"], [397, 399, ["training", "algorithm"], "Method"], [409, 414, ["error", "back", "-", "propagation", "step"], "Method"], [421, 425, ["big", "and", "deep", "DNNs"], "Method"], [434, 436, ["unsupervised", "initialization"], "Method"], [437, 438, ["pretraining"], "Task"], [469, 473, ["Multi", "-", "column", "DNN"], "Method"], [474, 475, ["MCDNN"], "Method"], [479, 481, ["error", "rate"], "Metric"], [496, 497, ["DNN"], "Method"], [503, 505, ["classification", "error"], "Metric"], [513, 514, ["generalization"], "Task"], [544, 546, ["shallow", "NN"], "Method"], [559, 560, ["Neocognitron"], "Method"], [615, 617, ["gradient", "descent"], "Method"], [659, 664, ["massively", "parallel", "graphics", "processing", "units"], "Task"], [665, 666, ["GPUs"], "Method"], [669, 671, ["video", "games"], "Task"], [677, 679, ["speedup", "factor"], "Metric"], [684, 686, ["serial", "code"], "Method"], [704, 708, ["unsupervised", "pre", "-", "training"], "Method"], [717, 718, ["DNN"], "Method"], [730, 741, ["2", "-", "dimensional", "layers", "of", "winner", "-", "take", "-", "all", "neurons"], "Method"], [757, 760, ["max", "pooling", "technique"], "Method"], [842, 848, ["down", "-", "sampling", "-", "like", "operations"], "Method"], [900, 904, ["multi", "-", "layer", "perceptron"], "Method"], [905, 906, ["MLP"], "Method"], [919, 920, ["DNN"], "Method"], [973, 975, ["deep", "architecture"], "Method"], [982, 984, ["convolutional", "kernels"], "Method"], [992, 993, ["MLP"], "Method"], [1047, 1049, ["energy", "consumption"], "Metric"], [1051, 1053, ["training", "algorithm"], "Method"], [1088, 1092, ["Multi", "-", "column", "DNN"], "Method"], [1093, 1094, ["MCDNN"], "Method"], [1149, 1151, ["error", "rate"], "Metric"], [1160, 1161, ["accuracy"], "Metric"], [1163, 1165, ["MCDNN", "architecture"], "Method"], [1205, 1208, ["object", "recognition", "benchmarks"], "Task"], [1272, 1274, ["convolutional", "layer"], "Method"], [1282, 1286, ["max", "-", "pooling", "layer"], "Method"], [1295, 1297, ["convolutional", "layer"], "Method"], [1299, 1300, ["maps"], "Method"], [1301, 1303, ["4x4", "filters"], "Method"], [1305, 1309, ["max", "-", "pooling", "layer"], "Method"], [1318, 1321, ["fully", "connected", "layer"], "Method"], [1327, 1330, ["fully", "connected", "layer"], "Method"], [1336, 1340, ["fully", "connected", "output", "layer"], "Method"], [1352, 1357, ["scaled", "hyperbolic", "tangent", "activation", "function"], "Method"], [1358, 1363, ["convolutional", "and", "fully", "connected", "layers"], "Method"], [1365, 1368, ["linear", "activation", "function"], "Method"], [1369, 1373, ["max", "-", "pooling", "layers"], "Method"], [1375, 1378, ["softmax", "activation", "function"], "Method"], [1384, 1385, ["DNN"], "Method"], [1388, 1393, ["on", "-", "line", "gradient", "descent"], "Method"], [1395, 1398, ["annealed", "learning", "rate"], "Method"], [1428, 1429, ["validation"], "Task"], [1430, 1431, ["Training"], "Task"], [1434, 1436, ["validation", "error"], "Metric"], [1441, 1443, ["learning", "rate"], "Metric"], [1463, 1464, ["MNIST"], "Material"], [1532, 1533, ["normalization"], "Task"], [1544, 1545, ["MCDNN"], "Method"], [1549, 1561, ["20C4", "-", "MP2", "-", "40C5", "-", "MP3", "-", "150N", "-", "10N", "DNN"], "Method"], [1569, 1572, ["annealed", "learning", "rate"], "Metric"], [1593, 1594, ["DNN"], "Method"], [1642, 1643, ["DNN"], "Method"], [1681, 1682, ["MCDNN"], "Method"], [1690, 1691, ["MCDNN"], "Method"], [1705, 1706, ["DNNs"], "Method"], [1720, 1721, ["MCDNN"], "Method"], [1727, 1729, ["error", "rate"], "Metric"], [1762, 1764, ["error", "rate"], "Metric"], [1812, 1813, ["DNN"], "Method"], [1828, 1829, ["MCDNN"], "Method"], [1832, 1833, ["DNN"], "Method"], [1842, 1843, ["MCDNN"], "Method"], [1854, 1856, ["MCDNN", "errors"], "Metric"], [1866, 1867, ["DNNs"], "Method"], [1873, 1874, ["MCDNN"], "Method"], [1904, 1906, ["MCDNN", "error"], "Metric"], [1911, 1912, ["DNN"], "Method"], [1929, 1933, ["60", "-", "net", "MCDNN"], "Method"], [1944, 1945, ["MCDNN"], "Method"], [1949, 1950, ["preprocessing"], "Task"], [1954, 1955, ["recognition"], "Task"], [1959, 1960, ["MCDNN"], "Method"], [1961, 1962, ["DNN"], "Method"], [1975, 1977, ["error", "rate"], "Metric"], [1984, 1989, ["35", "-", "columns", "MCDNN", "architecture"], "Method"], [1990, 1991, ["preprocessing"], "Method"], [1993, 1994, ["MNIST"], "Material"], [1998, 2000, ["Latin", "characters"], "Material"], [2001, 2004, ["NIST", "SD", "19"], "Material"], [2009, 2010, ["MCDNN"], "Method"], [2011, 2013, ["recognition", "rates"], "Metric"], [2071, 2075, ["62", "-", "class", "problem"], "Task"], [2078, 2083, ["52", "-", "class", "letters", "problem"], "Task"], [2152, 2154, ["error", "rates"], "Metric"], [2156, 2159, ["case", "insensitive", "task"], "Task"], [2187, 2189, ["error", "rate"], "Metric"], [2210, 2212, ["error", "rate"], "Metric"], [2246, 2248, ["confusion", "matrix"], "Metric"], [2265, 2268, ["Latin", "character", "recognition"], "Task"], [2269, 2273, ["isolated", "Chinese", "character", "recognition"], "Task"], [2370, 2372, ["GPU", "implementation"], "Method"], [2387, 2389, ["forward", "propagation"], "Method"], [2398, 2399, ["CPU"], "Method"], [2415, 2417, ["GPU", "implementation"], "Method"], [2449, 2450, ["DNN"], "Method"], [2484, 2488, ["offline", "character", "recognition", "task"], "Task"], [2528, 2532, ["database", "CASIA", "-", "HWDB1.1"], "Material"], [2605, 2608, ["Gaussian", "blur", "filter"], "Method"], [2632, 2634, ["database", "CASIA"], "Material"], [2641, 2642, ["classifier"], "Method"], [2679, 2681, ["feature", "extraction"], "Method"], [2684, 2686, ["dimensionality", "reduction"], "Method"], [2699, 2701, ["feature", "extraction"], "Method"], [2702, 2704, ["dimensionality", "reduction"], "Task"], [2706, 2708, ["supervised", "way"], "Method"], [2711, 2713, ["offline", "task"], "Task"], [2716, 2718, ["error", "rate"], "Metric"], [2749, 2751, ["recognition", "rate"], "Metric"], [2756, 2758, ["online", "task"], "Task"], [2774, 2776, ["classification", "problem"], "Task"], [2815, 2817, ["Traffic", "signs"], "Task"], [2824, 2826, ["automotive", "industry"], "Task"], [2832, 2835, ["driver", "’s", "assistance"], "Task"], [2840, 2844, ["traffic", "-", "related", "applications"], "Task"], [2848, 2852, ["GTSRB", "traffic", "sign", "dataset"], "Material"], [2991, 2992, ["MCDNN"], "Method"], [3019, 3022, ["image", "preprocessing", "methods"], "Method"], [3045, 3046, ["DNN"], "Method"], [3073, 3074, ["MCDNN"], "Method"], [3080, 3082, ["error", "rate"], "Metric"], [3134, 3135, ["MCDNN"], "Method"], [3183, 3185, ["error", "rate"], "Metric"], [3192, 3194, ["error", "rate"], "Metric"], [3239, 3240, ["MCDNN"], "Method"], [3249, 3250, ["MCDNN"], "Method"], [3266, 3267, ["DNN"], "Method"], [3271, 3273, ["CIFAR", "10"], "Material"], [3273, 3274, ["CIFAR10"], "Material"], [3389, 3393, ["10", "-", "layer", "architecture"], "Method"], [3424, 3425, ["MNIST"], "Material"], [3428, 3430, ["learning", "rate"], "Metric"], [3442, 3445, ["CIFAR", "color", "images"], "Material"], [3454, 3456, ["error", "rates"], "Metric"], [3468, 3469, ["MNIST"], "Material"], [3488, 3489, ["error"], "Metric"], [3497, 3498, ["NN"], "Method"], [3535, 3537, ["net", "errors"], "Metric"], [3573, 3575, ["random", "initializations"], "Method"], [3577, 3584, ["mean", "and", "standard", "deviation", "of", "the", "error"], "Metric"], [3596, 3597, ["DNN"], "Method"], [3601, 3602, ["MCDNN"], "Method"], [3606, 3608, ["error", "rate"], "Metric"], [3621, 3623, ["confusion", "matrix"], "Metric"], [3632, 3633, ["MCDNN"], "Method"], [3736, 3737, ["MCDNN"], "Method"], [3763, 3765, ["error", "rate"], "Metric"], [3779, 3780, ["MCDNN"], "Method"], [3784, 3785, ["NORB"], "Material"], [3835, 3838, ["3D", "object", "recognition"], "Task"], [3933, 3934, ["preprocessing"], "Method"], [4042, 4043, ["NORB"], "Material"], [4066, 4068, ["net", "architecture"], "Method"], [4098, 4101, ["learning", "rate", "setup"], "Metric"], [4118, 4120, ["net", "size"], "Metric"], [4121, 4122, ["training"], "Task"], [4148, 4149, ["MCDNN"], "Method"], [4220, 4223, ["learning", "rate", "setup"], "Metric"], [4227, 4229, ["Training", "time"], "Metric"], [4267, 4269, ["error", "rate"], "Metric"], [4279, 4280, ["NORB"], "Material"], [4294, 4295, ["classification"], "Task"], [4320, 4322, ["error", "rate"], "Metric"], [4373, 4374, ["MCDNN"], "Method"], [4399, 4401, ["recognition", "rates"], "Metric"], [4402, 4403, ["MNIST"], "Material"], [4404, 4407, ["NIST", "SD", "19"], "Material"], [4408, 4410, ["Chinese", "characters"], "Material"], [4414, 4415, ["CIFAR10"], "Material"], [4416, 4417, ["NORB"], "Material"], [4422, 4423, ["supervised"], "Method"], [4434, 4435, ["DNN"], "Method"], [4453, 4454, ["MCDNNs"], "Method"], [4485, 4489, ["Neuro", "-", "dynamic", "Framework"], "Method"], [4490, 4492, ["Cognitive", "Robotics"], "Task"], [4493, 4495, ["Scene", "Representations"], "Task"], [4496, 4498, ["Behavioral", "Sequences"], "Task"], [4500, 4501, ["Learning"], "Task"], [4519, 4521, ["confusion", "matrix"], "Metric"], [4523, 4526, ["62", "characters", "task"], "Task"], [4562, 4564, ["confusion", "matrix"], "Metric"], [4566, 4568, ["digit", "task"], "Task"], [4591, 4594, ["52", "letter", "task"], "Task"], [4611, 4612, ["MCDNN"], "Method"], [4655, 4660, ["upper", "-", "case", "letter", "task"], "Task"], [4672, 4673, ["MCDNN"], "Method"], [4713, 4718, ["lower", "-", "case", "letter", "task"], "Task"], [4738, 4739, ["MCDNN"], "Method"], [4780, 4785, ["upper", "-", "case", "letters", "task"], "Task"], [4788, 4793, ["merged", "-", "case", "letter", "task"], "Task"], [4798, 4800, ["confusion", "matrix"], "Metric"], [4809, 4810, ["MCDNN"], "Method"], [4925, 4926, ["normalization"], "Task"], [4932, 4934, ["contrast", "normalizations"], "Method"], [4935, 4937, ["Image", "Adjustment"], "Task"], [4938, 4939, ["Imadjust"], "Method"], [4965, 4967, ["Histogram", "Equalization"], "Method"], [4968, 4969, ["Histeq"], "Method"], [4986, 4989, ["Adaptive", "Histogram", "Equalization"], "Method"], [4990, 4991, ["Adapthisteq"], "Method"], [4995, 4996, ["Histeq"], "Method"], [5031, 5033, ["Contrast", "Normalization"], "Method"], [5034, 5035, ["Conorm"], "Method"], [5047, 5048, ["Gaussians"], "Method"], [5061, 5062, ["normalizations"], "Method"], [5098, 5099, ["normalization"], "Method"], [5116, 5117, ["normalizations"], "Method"], [5140, 5141, ["DNN"], "Method"], [5162, 5164, ["net", "architecture"], "Method"], [5175, 5179, ["10", "-", "layer", "architecture"], "Method"], [5182, 5186, ["max", "-", "pooling", "kernels"], "Method"], [5189, 5190, ["CIFAR10"], "Material"], [5192, 5193, ["NORB"], "Material"], [121, 122, ["MNIST"], "Material"], [224, 226, ["traffic", "signs"], "Task"], [282, 283, ["NORB"], "Material"], [465, 466, ["DNN"], "Method"], [608, 609, ["DNN"], "Method"], [1083, 1084, ["DNN"], "Method"], [1227, 1228, ["DNN"], "Method"], [1466, 1467, ["MNIST"], "Material"], [1529, 1530, ["DNN"], "Method"], [2067, 2069, ["error", "rate"], "Metric"], [2796, 2797, ["DNN"], "Method"], [2818, 2820, ["traffic", "signs"], "Task"], [2935, 2936, ["DNN"], "Method"], [2974, 2976, ["traffic", "signs"], "Task"], [3009, 3011, ["traffic", "signs"], "Task"], [3370, 3371, ["DNN"], "Method"], [3755, 3757, ["traffic", "signs"], "Task"], [3775, 3776, ["NORB"], "Material"], [4411, 4413, ["traffic", "signs"], "Task"]]} {"doc_id": "3bf52841ade9f6adc55e102a4460af922e52bb7d", "text": ["We", "introduce", "a", "simple", "recurrent", "variational", "auto", "-", "encoder", "architecture", "that", "significantly", "improves", "image", "modeling", ".", "The", "system", "represents", "the", "state", "-", "of", "-", "the", "-", "art", "in", "latent", "variable", "models", "for", "both", "the", "ImageNet", "and", "Omniglot", "datasets", ".", "We", "show", "that", "it", "naturally", "separates", "global", "conceptual", "information", "from", "lower", "level", "details", ",", "thus", "addressing", "one", "of", "the", "fundamentally", "desired", "properties", "of", "unsupervised", "learning", ".", "Furthermore", ",", "the", "possibility", "of", "restricting", "ourselves", "to", "storing", "only", "global", "information", "about", "an", "image", "allows", "us", "to", "achieve", "high", "quality", "‘", "conceptual", "compression", "’", ".", "ConceptualCompression", "section", ":", "Introduction", "Images", "contain", "a", "large", "amount", "of", "information", "that", "is", "a", "priori", "stored", "independently", "in", "the", "pixels", ".", "In", "the", "semi", "-", "supervised", "learning", "regime", "where", "a", "large", "number", "of", "images", "is", "available", "but", "only", "a", "small", "number", "of", "labels", ",", "one", "would", "like", "to", "leverage", "this", "information", "to", "create", "representations", "that", "allow", "for", "better", "(", "and", "especially", "faster", ")", "generalization", ".", "Intuitively", "one", "expects", "such", "representations", "to", "explicitly", "extract", "global", "conceptual", "aspects", "of", "an", "image", ".", "In", "this", "paper", "we", "propose", "a", "method", "that", "is", "able", "to", "transform", "an", "image", "into", "a", "progression", "of", "increasingly", "detailed", "representations", ",", "ranging", "from", "global", "conceptual", "aspects", "to", "low", "level", "details", "(", "see", "Figures", "[", "reference", "]", "&", "[", "reference", "]", ")", ".", "At", "the", "same", "time", ",", "our", "model", "greatly", "improves", "latent", "variable", "image", "modeling", "compared", "to", "earlier", "implementations", "of", "deep", "variational", "auto", "-", "encoders", ".", "Furthermore", ",", "it", "has", "the", "advantage", "of", "being", "a", "simple", "homogeneous", "architecture", "not", "requiring", "complex", "design", "choices", ",", "which", "is", "similar", "to", "the", "recurrent", "structure", "of", "DRAW", "(", ")", ".", "Last", ",", "it", "provides", "an", "important", "insight", "into", "building", "good", "variational", "auto", "-", "encoder", "models", "of", "images", ":", "the", "use", "of", "multiple", "layers", "of", "stochastic", "variables", "that", "are", "all", "‘", "close", "’", "to", "the", "pixels", "significantly", "improves", "performance", ".", "The", "system", "’s", "ability", "to", "stratify", "information", "enables", "it", "to", "perform", "high", "quality", "lossy", "compression", ",", "by", "storing", "only", "a", "subset", "of", "latent", "variables", ",", "starting", "with", "the", "high", "level", "ones", ",", "and", "generating", "the", "remainder", "during", "decompression", "(", "see", "Figure", "[", "reference", "]", ")", ".", "Currently", "the", "ultimate", "arbiter", "of", "lossy", "compression", "remains", "human", "evaluation", ".", "Other", "simple", "measures", "such", "as", "the", "L2", "distance", "between", "compressed", "and", "original", "images", "are", "inappropriate", "–", "for", "example", "if", "a", "particular", "generated", "grass", "texture", "is", "sharp", ",", "but", "different", "from", "the", "one", "in", "the", "original", "image", ",", "it", "will", "yield", "a", "large", "L2", "distance", "yet", "should", ",", "at", "the", "same", "time", ",", "be", "considered", "conceptually", "close", "to", "the", "original", ".", "Achieving", "good", "lossy", "compression", "while", "storing", "only", "high", "level", "latent", "variables", "would", "imply", "that", "representations", "learned", "at", "a", "high", "level", "contain", "information", "similar", "to", "that", "used", "by", "humans", "to", "judge", "images", ".", "As", "humans", "outperform", "the", "best", "machines", "at", "learning", "abstract", "representations", ",", "human", "evaluation", "of", "lossy", "compression", "obtained", "by", "these", "generative", "models", "constitutes", "a", "reasonable", "test", "of", "the", "quality", "of", "representations", "learned", "by", "these", "models", ".", "In", "the", "following", "we", "discuss", "variational", "auto", "-", "encoders", "and", "compression", "in", "more", "detail", ",", "present", "the", "algorithm", "and", "demonstrate", "the", "results", "on", "generation", "quality", "and", "compression", ".", "subsection", ":", "Variational", "Auto", "-", "Encoders", "Numerous", "techniques", "exist", "for", "unsupervised", "learning", "in", "deep", "networks", ",", "e.g.", "sparse", "auto", "-", "encoders", "and", "sparse", "coding", ",", "denoising", "auto", "-", "encoders", ",", "deconvolutional", "networks", ",", "restricted", "Boltzmann", "machines", ",", "deep", "Boltzmann", "machines", ",", "generative", "adversarial", "networks", "and", "variational", "auto", "-", "encoders", ".", "In", "this", "paper", "we", "focus", "on", "the", "class", "of", "models", "in", "the", "variational", "auto", "-", "encoding", "framework", ".", "Since", "we", "are", "also", "interested", "in", "compression", ",", "we", "present", "them", "from", "an", "information", "-", "theoretic", "perspective", ".", "Variational", "auto", "-", "encoders", "typically", "consist", "of", "two", "neural", "networks", ":", "one", "that", "generates", "samples", "from", "latent", "variables", "(", "‘", "imagination", "’", ")", ",", "and", "one", "that", "infers", "latent", "variables", "from", "observations", "(", "‘", "recognition", "’", ")", ".", "The", "two", "networks", "share", "the", "latent", "variables", ".", "Intuitively", "speaking", "one", "might", "think", "of", "these", "variables", "as", "specifying", ",", "for", "a", "given", "image", ",", "at", "different", "levels", "of", "abstraction", ",", "whether", "a", "particular", "object", "such", "as", "a", "cat", "or", "a", "dog", "is", "present", "in", "the", "input", ",", "or", "perhaps", "what", "the", "exact", "position", "and", "intensity", "of", "an", "edge", "at", "a", "given", "location", "might", "be", ".", "During", "the", "recognition", "phase", "the", "network", "acquires", "information", "about", "the", "input", "and", "stores", "it", "in", "the", "latent", "variables", ",", "reducing", "their", "uncertainty", ".", "For", "example", ",", "at", "first", "not", "knowing", "whether", "a", "cat", "or", "a", "dog", "is", "present", "in", "the", "image", ",", "the", "network", "observes", "the", "input", "and", "becomes", "nearly", "certain", "that", "it", "is", "a", "cat", ".", "The", "reduction", "in", "uncertainty", "is", "quantitatively", "equal", "to", "the", "amount", "of", "information", "the", "network", "acquired", "about", "the", "input", ".", "During", "generation", "the", "network", "starts", "with", "uncertain", "latent", "variables", "and", "selects", "their", "values", "from", "a", "prior", "distribution", "that", "specifies", "this", "uncertainty", "(", "e.g.", "it", "chooses", "a", "dog", ")", ".", "Different", "choices", "will", "produce", "different", "samples", ".", "Variational", "auto", "-", "encoders", "provide", "a", "natural", "framework", "for", "unsupervised", "learning", "–", "we", "can", "build", "networks", "with", "layers", "of", "stochastic", "variables", "and", "expect", "that", ",", "after", "learning", ",", "the", "representations", "become", "increasingly", "more", "abstract", "for", "higher", "levels", "of", "the", "hierarchy", ".", "The", "questions", "then", "are", ":", "can", "such", "a", "framework", "indeed", "discover", "such", "representations", "both", "in", "principle", "and", "in", "practice", ",", "are", "such", "networks", "powerful", "enough", "for", "modeling", "real", "data", ",", "and", "what", "techniques", "one", "needs", "to", "make", "it", "work", "well", ".", "subsection", ":", "Conceptual", "Compression", "Variational", "auto", "-", "encoders", "can", "not", "only", "be", "used", "for", "representation", "learning", "but", "also", "for", "compression", ".", "The", "training", "objective", "of", "variational", "auto", "-", "encoders", "is", "to", "compress", "the", "total", "amount", "of", "information", "needed", "to", "encode", "the", "input", ".", "They", "achieve", "this", "by", "using", "information", "-", "carrying", "latent", "variables", "that", "express", "what", ",", "before", "compression", ",", "was", "encoded", "using", "a", "larger", "amount", "of", "information", "in", "the", "input", ".", "The", "information", "in", "the", "layers", "and", "the", "remaining", "information", "in", "the", "input", "can", "be", "encoded", "in", "practice", "as", "explained", "later", "in", "this", "paper", ".", "The", "amount", "of", "lossless", "compression", "one", "is", "able", "to", "achieve", "is", "bounded", "by", "the", "underlying", "entropy", "of", "the", "image", "distribution", ".", "Additionally", ",", "most", "image", "information", "as", "measured", "in", "bits", "is", "contained", "in", "the", "fine", "details", "of", "the", "image", ".", "Thus", "we", "might", "reasonably", "expect", "that", "lossless", "compression", "will", "never", "improve", "by", "more", "than", "a", "factor", "of", "two", "in", "comparison", "to", "current", "performance", ".", "Lossy", "compression", ",", "on", "the", "other", "hand", ",", "holds", "much", "more", "potential", "for", "improvement", ".", "In", "this", "case", "we", "want", "to", "compress", "an", "image", "by", "a", "certain", "amount", ",", "allowing", "some", "information", "loss", ",", "while", "maximizing", "both", "quality", "and", "similarity", "to", "the", "original", "image", ".", "As", "an", "example", ",", "at", "a", "low", "level", "of", "compression", "(", "close", "to", "lossless", "compression", ")", ",", "we", "could", "start", "by", "reducing", "pixel", "precision", ",", "e.g.", "from", "8", "bits", "to", "7", "bits", ".", "Then", ",", "as", "in", "JPEG", ",", "we", "could", "express", "a", "local", "8x8", "neighborhood", "in", "a", "discrete", "cosine", "transform", "basis", "and", "store", "only", "the", "most", "significant", "components", ".", "This", "way", ",", "instead", "of", "introducing", "quantization", "artifacts", "in", "the", "image", "that", "would", "appear", "if", "we", "kept", "decreasing", "pixel", "precision", ",", "we", "preserve", "higher", "level", "structures", "but", "to", "a", "lower", "level", "of", "precision", ".", "However", ",", "what", "can", "we", "do", "beyond", "that", "as", "we", "keep", "pushing", "the", "compression", "?", "We", "would", "like", "to", "preserve", "the", "most", "important", "aspects", "of", "the", "image", ".", "What", "determines", "what", "is", "important", "?", "Let", "us", "imagine", "that", "we", "are", "compressing", "images", "of", "cats", "and", "dogs", "and", "would", "like", "to", "compress", "an", "image", "down", "to", "one", "bit", ".", "What", "would", "that", "bit", "be", "?", "One", "would", "imagine", "that", "it", "should", "represent", "whether", "the", "image", "contains", "either", "a", "cat", "or", "a", "dog", ".", "How", "would", "we", "then", "get", "an", "image", "out", "of", "this", "single", "bit", "?", "If", "we", "have", "a", "good", "generative", "model", ",", "we", "can", "simply", "generate", "the", "entire", "image", "from", "this", "one", "latent", "variable", ",", "an", "image", "of", "a", "cat", "if", "the", "bit", "corresponds", "to", "‘", "cat", "’", ",", "and", "an", "image", "of", "a", "dog", "otherwise", ".", "Now", "let", "us", "imagine", "that", "instead", "of", "compressing", "to", "one", "bit", "we", "wanted", "to", "compress", "down", "to", "ten", "bits", ".", "Now", "we", "can", "store", "the", "most", "important", "properties", "of", "the", "animal", "as", "well", "–", "e.g.", "its", "type", ",", "color", ",", "and", "basic", "pose", ".", "The", "rest", "would", "be", "‘", "filled", "in", "’", "by", "the", "generative", "model", "that", "is", "conditioned", "on", "this", "information", ".", "If", "we", "increase", "the", "number", "of", "bits", "further", "we", "can", "preserve", "more", "and", "more", "about", "the", "image", ",", "while", "generating", "the", "fine", "details", "such", "as", "hair", ",", "or", "the", "exact", "pattern", "of", "the", "floor", ",", "etc", ".", "Most", "bits", "are", "in", "fact", "about", "such", "low", "level", "details", ".", "We", "call", "this", "kind", "of", "compression", "–", "compressing", "by", "giving", "priority", "to", "higher", "levels", "of", "representation", "and", "generating", "the", "remainder", "–", "‘", "conceptual", "compression", "’", ".", "We", "suggest", "that", "this", "should", "be", "the", "ultimate", "objective", "of", "lossy", "compression", ".", "Importantly", ",", "if", "we", "solve", "deep", "representation", "learning", "with", "latent", "variable", "generative", "models", "that", "generate", "high", "quality", "samples", ",", "we", "achieve", "the", "objective", "of", "lossy", "compression", "mentioned", "above", ".", "We", "can", "see", "this", "as", "follows", ".", "Assume", "that", "the", "network", "has", "learned", "a", "hierarchy", "of", "progressively", "more", "abstract", "representations", ".", "Then", ",", "to", "get", "different", "levels", "of", "compression", ",", "we", "can", "store", "only", "the", "corresponding", "number", "of", "topmost", "layers", "and", "generate", "the", "rest", ".", "By", "solving", "unsupervised", "deep", "learning", ",", "the", "network", "would", "order", "information", "according", "to", "its", "importance", "and", "store", "it", "with", "that", "priority", ".", "While", "the", "ultimate", "goal", "of", "unsupervised", "learning", "remains", "elusive", ",", "we", "make", "a", "step", "in", "this", "direction", ",", "and", "show", "that", "our", "network", "learns", "to", "order", "information", "from", "a", "rather", "global", "level", "to", "precise", "details", "in", "images", ",", "without", "being", "hand", "-", "engineered", "to", "do", "this", "explicitly", ",", "as", "illustrated", "in", "Figures", "[", "reference", "]", "and", "[", "reference", "]", ".", "This", "information", "separation", "already", "allows", "us", "to", "achieve", "better", "compression", "quality", "than", "JPEG", "and", "JPEG2000", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "While", "we", "are", "not", "bound", "by", "the", "same", "constraints", "as", "these", "algorithms", ",", "such", "as", "speed", "and", "memory", ",", "these", "results", "demonstrate", "the", "potential", "of", "this", "method", ",", "which", "will", "get", "better", "as", "latent", "variable", "generative", "models", "improve", ".", "subsection", ":", "The", "Importance", "of", "Recurrent", "Feedback", "What", "are", "the", "challenges", "involved", "in", "turning", "latent", "variable", "models", "into", "state", "-", "of", "-", "the", "-", "art", "generative", "models", "of", "images", "?", "Many", "successful", "vision", "architectures", "(", "e.g.", ")", "have", "highly", "over", "-", "complete", "representations", "that", "contain", "many", "more", "neurons", "in", "hidden", "layers", "than", "pixels", ".", "These", "representations", "need", "to", "be", "combined", "to", "get", "a", "very", "sharp", "distribution", "at", "the", "pixel", "level", "if", "the", "pixels", "are", "modeled", "independently", ".", "This", "distribution", "corresponds", "to", "salt", "and", "pepper", "noise", "which", "is", "not", "present", "in", "natural", "images", "to", "a", "perceptible", "level", ".", "This", "poses", "a", "major", "challenge", ".", "After", "experimenting", "with", "deep", "variational", "auto", "-", "encoders", "we", "concluded", "that", "it", "was", "exceedingly", "difficult", "to", "obtain", "satisfactory", "results", "with", "a", "single", "computational", "pass", "through", "the", "network", ".", "Instead", "we", "propose", "that", "the", "network", "needs", "the", "ability", "to", "correct", "itself", "over", "a", "number", "of", "time", "steps", ".", "Thus", ",", "sharp", "reconstructions", "should", "not", "be", "a", "property", "of", "high", "-", "precision", "values", "in", "the", "network", ",", "but", "should", "rather", "be", "the", "result", "of", "an", "iterative", "feedback", "mechanism", "that", "is", "robust", "to", "network", "parameter", "change", ".", "Such", "a", "mechanism", "is", "provided", "by", "the", "DRAW", "algorithm", ",", "which", "is", "a", "recurrent", "type", "of", "variational", "auto", "-", "encoder", ".", "At", "each", "time", "step", ",", "DRAW", "maintains", "a", "provisionary", "reconstruction", ",", "takes", "in", "information", "about", "a", "given", "image", ",", "stores", "it", "in", "latent", "variables", "and", "updates", "the", "reconstruction", ".", "Keeping", "track", "of", "the", "reconstruction", "aids", "the", "iterative", "feedback", "mechanism", "which", "is", "learned", "by", "back", "-", "propagation", ".", "Computation", "is", "both", "deep", "–", "in", "iterations", "–", "and", "close", "to", "the", "pixels", ".", "We", "introduce", "convolutional", "DRAW", ".", "It", "features", "convolutions", ",", "latent", "prior", "modeling", ",", "a", "Gaussian", "input", "distribution", "(", "for", "natural", "images", ")", "and", ",", "in", "some", "experiments", ",", "a", "multi", "-", "layer", "architecture", ".", "However", ",", "it", "does", "not", "use", "an", "explicit", "attentional", "mechanism", ".", "We", "note", "that", "even", "the", "single", "-", "layer", "version", "is", "already", "a", "deep", "generative", "model", "which", "can", "decide", "to", "process", "higher", "level", "information", "first", "before", "focusing", "on", "details", ",", "as", "we", "demonstrate", "to", "some", "degree", ".", "We", "also", "experiment", "with", "making", "convolutional", "DRAW", "hierarchical", "in", "a", "similar", "way", "that", "we", "would", "build", "conventional", "deep", "variational", "auto", "-", "encoders", "–", "stacking", "more", "layers", "of", "latent", "and", "deterministic", "variables", ".", "We", "believe", "that", "the", "recurrence", "is", "important", "not", "just", "for", "accurate", "pixel", "reconstructions", ",", "but", "also", "at", "higher", "levels", ".", "For", "example", ",", "when", "the", "network", "decides", "to", "generate", "edges", "at", "different", "locations", ",", "it", "needs", "to", "make", "sure", "that", "they", "are", "aligned", ".", "It", "is", "hard", "to", "imagine", "this", "happening", "in", "a", "single", "computational", "pass", "through", "the", "network", ".", "Similarly", "at", "higher", "levels", ",", "when", "it", "decides", "to", "generate", "objects", ",", "they", "need", "to", "be", "generated", "with", "the", "right", "relationship", "to", "one", "another", ".", "And", "finally", "at", "the", "scene", "level", ",", "one", "probably", "does", "not", "generate", "entire", "scenes", "at", "once", ",", "but", "rather", "one", "step", "at", "a", "time", ".", "subsection", ":", "Comparison", "to", "Non", "-", "variational", "Models", "Let", "us", "relate", "this", "discussion", "to", "two", "other", "families", "of", "generative", "models", ",", "specifically", "generative", "adversarial", "networks", "(", "GANs", ";", ")", "and", "auto", "-", "regressive", "pixel", "models", ".", "GANs", "have", "been", "demonstrated", "to", "be", "able", "to", "generate", "realistic", "looking", "images", ",", "with", "properly", "aligned", "edges", ",", "using", "a", "simple", "feedforward", "generative", "network", ".", "GANs", "also", "contain", "two", "networks", "–", "a", "generative", "network", "that", "is", "the", "same", "as", "in", "variational", "auto", "-", "encoders", ",", "and", "a", "classification", "network", ".", "The", "classification", "network", "is", "presented", "with", "both", "real", "and", "model", "-", "generated", "images", "and", "tries", "to", "classify", "them", "according", "to", "their", "true", "nature", "–", "real", "or", "model", "-", "generated", ".", "The", "generative", "network", "gets", "gradients", "from", "the", "classification", "network", ",", "changing", "its", "weights", "in", "an", "attempt", "to", "make", "the", "generated", "images", "be", "judged", "as", "real", "ones", "by", "the", "classification", "network", ".", "This", "makes", "the", "generation", "network", "produce", "realistic", "images", "that", "‘", "fool", "’", "the", "classification", "network", ".", "It", "needs", "to", "produce", "a", "wide", "diversity", "of", "images", ",", "not", "just", "one", "realistic", "image", ",", "because", "if", "it", "produced", "only", "one", "(", "or", "a", "small", "number", "of", "them", ")", ",", "the", "classification", "network", "would", "classify", "that", "image", "as", "generated", "and", "others", "as", "realistic", ",", "and", "be", "almost", "always", "correct", ".", "This", "actually", "happens", "in", "practice", ",", "and", "one", "has", "to", "apply", "a", "variety", "of", "techniques", ",", "e.g.", "as", "in", ",", "to", "obtain", "sufficient", "image", "diversity", ".", "However", "the", "extent", "of", "GANs", "’", "sampling", "diversity", "is", "unknown", "and", "currently", "there", "is", "no", "satisfactory", "measure", "for", "it", ".", "So", "while", "a", "given", "network", "does", "n’t", "produce", "just", "one", "image", ",", "it", "is", "possible", "that", "it", "produces", "only", "a", "tiny", "subset", "of", "possible", "realistic", "images", ",", "as", "it", "simply", "competes", "with", "the", "power", "of", "the", "classifier", ".", "For", "example", "if", "it", "generates", "a", "sharp", "image", ",", "it", "is", "unclear", "whether", "the", "system", "is", "also", "capable", "of", "generating", "its", "translated", "version", ",", "or", "simply", "a", "slightly", "distorted", "version", ".", "Finally", "there", "is", "another", "way", "to", "get", "low", "uncertainty", "at", "the", "pixel", "level", ":", "instead", "of", "predicting", "pixels", "independently", "given", "the", "latents", ",", "we", "can", "decide", "not", "to", "use", "latents", "and", "iterate", "sequentially", "over", "the", "pixels", ",", "predicting", "a", "given", "pixel", "from", "the", "previous", "ones", "(", "from", "top", "left", "to", "bottom", "right", ")", "in", "an", "autoregressive", "fashion", ".", "This", "is", "as", "‘", "close", "’", "to", "the", "pixels", "as", "one", "can", "possibly", "be", ",", "and", "furthermore", "the", "procedure", "is", "purely", "deterministic", ".", "The", "disadvantage", "is", "conceptual", "–", "the", "information", "and", "decisions", "are", "not", "done", "at", "a", "conceptual", "level", "but", "at", "the", "pixel", "level", ".", "For", "example", "when", "generating", "cats", "vs", "dogs", "the", "decisions", "at", "the", "first", "set", "of", "pixels", "(", "top", "left", "of", "the", "image", ")", "will", "contain", "no", "information", "about", "a", "hypothetical", "cat", "or", "dog", ".", "But", "as", "we", "get", "to", "the", "region", "where", "these", "objects", "are", ",", "we", "start", "choosing", "pixels", "that", "will", "slowly", "tip", "the", "probability", "of", "generating", "a", "cat", "vs", "a", "dog", "one", "way", "or", "the", "other", ".", "As", "we", "start", "generating", "an", "ear", ",", "it", "will", "more", "likely", "be", "a", "cat", "’s", "or", "a", "dog", "’s", "and", "so", "on", ".", "However", "this", "pixel", "level", "approach", "and", "our", "approach", "are", "orthogonal", "and", "can", "be", "easily", "combined", ",", "for", "example", "by", "feeding", "the", "output", "of", "convolutional", "DRAW", "into", "the", "conditional", "computation", "of", "a", "pixel", "level", "model", ".", "In", "this", "paper", "we", "study", "the", "latent", "variable", "approach", "and", "make", "the", "pixels", "independent", "given", "the", "latents", ".", "section", ":", "Convolutional", "DRAW", "In", "this", "section", "we", "describe", "the", "details", "of", "a", "single", "-", "layer", "version", "of", "the", "algorithm", ".", "Convolutional", "DRAW", "contains", "the", "following", "variables", ":", "input", ",", "reconstruction", ",", "reconstruction", "error", ",", "the", "state", "of", "the", "encoder", "recurrent", "net", ",", "the", "state", "of", "the", "decoder", "recurrent", "net", "and", "latent", "variable", ".", "The", "variables", ",", "and", "are", "recurrent", "(", "passed", "between", "different", "time", "steps", ")", "and", "are", "initialized", "with", "learned", "biases", ".", "Then", "at", "each", "time", "step", ",", "convolutional", "DRAW", "performs", "the", "following", "updates", ":", "where", "denotes", "a", "convolution", "and", "Rnn", "denotes", "a", "recurrent", "network", ".", "We", "use", "LSTM", "with", "convolutional", "operators", "instead", "of", "the", "usual", "linear", "ones", ".", "The", "final", "value", "of", "contains", "the", "parameters", "of", "the", "input", "distribution", ".", "For", "binary", "images", "we", "use", "the", "Bernoulli", "distribution", ".", "For", "natural", "images", "we", "use", "the", "Gaussian", "distribution", "with", "mean", "and", "log", "variance", "given", "by", "splitting", "the", "vector", "to", "obtain", "the", "input", "cost", "and", "the", "total", "cost", ":", "where", "the", "handling", "of", "real", "valued", "-", "ness", "of", "the", "inputs", "(", "[", "reference", "]", ",", "[", "reference", "]", ")", "is", "explained", "below", ",", "and", "being", "the", "standard", "setting", ".", "The", "algorithm", "is", "schematically", "illustrated", "in", "the", "first", "layer", "of", "Figure", "[", "reference", "]", ".", "The", "network", "is", "trained", "by", "calculating", "the", "gradient", "of", "this", "loss", "and", "using", "a", "stochastic", "gradient", "descent", "algorithm", ".", "Stochastic", "back", "-", "propagation", "through", "a", "sampling", "function", "is", "done", "as", "in", "variational", "auto", "-", "encoders", ".", "Both", "the", "approximate", "posterior", "and", "the", "prior", "are", "Gaussian", ",", "with", "mean", "and", "log", "variance", "being", "linear", "functions", "of", "or", ",", "respectively", ".", "Let", "us", "discuss", "how", "we", "handle", "the", "input", "distribution", "for", "natural", "images", ".", "Each", "pixel", "(", "per", "color", ")", "is", "one", "of", "256", "values", ".", "We", "could", "use", "a", "soft", "-", "max", "distribution", "to", "model", "it", ".", "This", "would", "result", "in", "a", "rather", "large", "output", "vector", "at", "every", "time", "step", "and", "also", "does", "not", "take", "advantage", "of", "the", "underlying", "real", "valued", "-", "ness", "of", "intensities", "and", "therefore", "we", "opted", "for", "a", "Gaussian", "distribution", ".", "However", "this", "still", "needs", "to", "be", "converted", "to", "a", "discrete", "distribution", "over", "256", "values", "to", "calculate", "the", "negative", "likelihood", "loss", ".", "Instead", "of", "this", ",", "we", "add", "uniform", "noise", "to", "the", "input", "with", "width", "corresponding", "to", "the", "spacing", "between", "discrete", "values", "and", "calculate", "where", "with", "if", "inputs", "are", "scaled", "to", "the", "interval", ".", "subsection", ":", "Multi", "-", "layer", "Architectures", "Next", "we", "explain", "how", "we", "can", "stack", "convolutional", "DRAW", "with", "a", "two", "layer", "example", ".", "The", "first", "layer", "is", "the", "same", "as", "the", "one", "just", "introduced", ".", "The", "second", "layer", "has", "the", "same", "structure", ":", "recurrent", "encoder", ",", "recurrent", "decoder", "and", "a", "stochastic", "layer", ".", "The", "input", "to", "the", "second", "layer", "is", "the", "mean", "of", "the", "approximate", "posterior", "of", "the", "first", "layer", ".", "The", "output", "of", "the", "second", "layer", "biases", "the", "prior", "of", "the", "latent", "variable", "of", "the", "first", "layer", "and", "is", "also", "passed", "as", "input", "into", "the", "first", "layer", "decoder", "recurrent", "net", ".", "This", "is", "illustrated", "in", "Figure", "[", "reference", "]", ".", "We", "do", "n’t", "use", "any", "reconstruction", "or", "error", "in", "the", "second", "layer", ".", "Here", "we", "describe", "a", "given", "computational", "step", "in", "detail", ".", "Indices", "and", "denote", "the", "variables", "of", "layers", "and", ",", "respectively", ".", "In", "addition", ",", "let", "be", "the", "mean", "of", ".", "Then", ",", "the", "update", "at", "a", "given", "time", "step", "is", "given", "by", "Systems", "with", "more", "layers", "can", "be", "built", "analogously", ".", "section", ":", "Compression", "Here", "we", "show", "how", "one", "can", "turn", "variational", "auto", "-", "encoders", "including", "convolutional", "DRAW", "into", "compression", "algorithms", ".", "We", "have", "not", "built", "the", "actual", "compressor", ",", "however", ",", "as", "we", "explain", ",", "we", "have", "strong", "reasons", "to", "believe", "it", "would", "perform", "as", "well", "as", "calculated", "here", ".", "Two", "basic", "approaches", "exist", ".", "The", "first", "one", "is", "less", "convenient", "because", "it", "needs", "to", "add", "extra", "data", "to", "the", "bitstream", "when", "compressing", "an", "image", "but", "has", "essentially", "a", "guaranteed", "compression", "rate", ".", "The", "other", "one", "may", "require", "some", "experimentation", "but", "is", "expected", "to", "yield", "a", "similar", "compression", "rate", "and", "can", "be", "used", "on", "a", "given", "image", "without", "needing", "extra", "data", ".", "The", "underlying", "compression", "mechanism", "for", "all", "cases", "is", "arithmetic", "coding", ".", "Arithmetic", "coding", "takes", "as", "input", "a", "sequence", "of", "discrete", "variables", "and", "a", "set", "of", "probabilities", "that", "predict", "the", "variable", "at", "time", "from", "previous", "variables", ".", "It", "then", "compresses", "this", "sequence", "to", "bits", "plus", "a", "constant", "of", "order", "one", ".", "Compressing", "inputs", "using", "variational", "auto", "-", "encoders", "proceeds", "as", "follows", ":", "discretize", "each", "latent", "variable", "in", "each", "layer", "using", "the", "width", "of", "(", "eq", ".", "[", "reference", "]", ")", ",", "treat", "the", "resulting", "variables", "as", "a", "sequence", "with", "predictions", "(", "eq", ".", "[", "reference", "]", ")", "and", "compress", "using", "arithmetic", "coding", ".", "For", "this", "to", "work", "as", "explained", ",", "several", "things", "are", "needed", ".", "First", ",", "the", "discretization", "should", "be", "independent", "of", "the", "input", ".", "This", "can", "be", "achieved", "by", "training", "the", "network", "with", "the", "variance", "of", "being", "a", "learned", "constant", "that", "does", "not", "depend", "on", "the", "input", ".", "We", "found", "that", "this", "does", "not", "have", "much", "effect", "on", "the", "likelihood", ".", "Second", ",", "one", "should", "evaluate", "the", "log", "likelihood", "using", "this", "discretization", ".", "One", "has", "to", "decide", "on", "the", "exact", "manner", "in", "which", "should", "be", "computed", "for", "each", "discrete", "value", ".", "Significant", "tuning", "might", "be", "required", "here", ",", "for", "the", "obtained", "likelihoods", "to", "be", "as", "good", "as", "the", "ones", "obtained", "with", "sampling", ".", "However", "this", "is", "likely", "to", "be", "fruitful", "since", "there", "exists", "a", "second", ",", "less", "convenient", "way", "to", "compress", "that", "is", "guaranteed", "to", "achieve", "this", "rate", ".", "This", "second", "approach", "uses", "bits", "-", "back", "coding", ".", "We", "explain", "only", "the", "basic", "idea", "here", ".", "We", "discretize", "the", "latents", "down", "to", "a", "very", "high", "level", "of", "precision", "and", "use", "to", "transmit", "the", "information", ".", "Because", "the", "discretization", "precision", "is", "high", ",", "the", "probabilities", "for", "discrete", "values", "are", "easily", "assigned", ".", "That", "will", "preserve", "the", "information", "but", "it", "will", "cost", "many", "bits", ",", "namely", "where", "is", "a", "probability", "under", "that", "discretization", ".", "Now", ",", "instead", "of", "sampling", "from", "the", "approximate", "posterior", "when", "encoding", "an", "input", ",", "we", "encode", "bits", "of", "other", "information", "into", "the", "choice", "of", ",", "that", "we", "also", "want", "to", "transmit", ".", "When", "is", "recovered", "at", "the", "receiving", "end", ",", "both", "the", "information", "about", "the", "current", "input", "and", "the", "other", "information", "is", "recovered", "and", "thus", "the", "information", "needed", "to", "encode", "the", "current", "input", "is", ".", "The", "expectation", "of", "this", "quantity", "is", "the", "KL", "-", "divergence", "in", "(", "[", "reference", "]", ")", ",", "which", "therefore", "measures", "the", "amount", "of", "information", "stored", "in", "a", "given", "latent", "layer", ".", "The", "disadvantage", "of", "this", "approach", "is", "that", "we", "can", "not", "encode", "a", "given", "input", "without", "also", "having", "some", "other", "information", "we", "want", "to", "transmit", ".", "However", ",", "this", "coding", "scheme", "works", "even", "if", "the", "variance", "of", "the", "approximate", "posterior", "is", "dependent", "on", "the", "input", ".", "section", ":", "Results", "For", "natural", "images", ",", "all", "models", "except", "otherwise", "specified", "were", "single", "-", "layer", ",", "with", ",", "a", "kernel", "size", "of", ",", "and", "stride", "2", "convolutions", "between", "input", "layers", "and", "hidden", "layers", "with", "latent", "feature", "maps", ".", "We", "trained", "the", "models", "on", "Cifar", "-", "10", "and", "ImageNet", "with", "and", "LSTM", "feature", "maps", "respectively", ".", "We", "use", "the", "version", "of", "ImageNet", "presented", "in", "that", "will", "soon", "be", "released", "as", "a", "standard", "dataset", ".", "We", "train", "the", "network", "with", "the", "Adam", "algorithm", "with", "learning", "rate", ".", "Occasionally", ",", "we", "find", "that", "the", "cost", "suddenly", "increases", "dramatically", ".", "This", "is", "probably", "due", "to", "the", "Gaussian", "nature", "of", "the", "distribution", ",", "when", "a", "given", "variable", "is", "produced", "too", "far", "from", "the", "mean", "relative", "to", "sigma", ".", "We", "observed", "this", "happening", "approximately", "once", "per", "run", ".", "To", "be", "able", "to", "keep", "training", "we", "store", "older", "parameters", ",", "detect", "such", "jumps", "and", "revert", "to", "the", "old", "parameters", "when", "they", "occur", ".", "The", "network", "then", "just", "keeps", "training", "as", "if", "nothing", "had", "happened", ".", "subsection", ":", "Modeling", "Quality", "subsubsection", ":", "Omniglot", "The", "recently", "introduced", "Omniglot", "dataset", "is", "comprised", "of", "character", "classes", "drawn", "from", "multiple", "alphabets", "with", "just", "samples", "per", "class", ".", "Referred", "to", "by", "many", "as", "the", "‘", "inverse", "of", "MNIST", "’", ",", "it", "was", "designed", "to", "study", "conceptual", "representations", "and", "generative", "models", "in", "a", "low", "-", "data", "regime", ".", "Table", "[", "reference", "]", "shows", "likelihoods", "of", "different", "models", "compared", "to", "ours", ".", "For", "our", "model", ",", "we", "only", "calculate", "the", "upper", "bound", "(", "variational", "bound", ")", "and", "therefore", "the", "true", "likelihood", "is", "actually", "better", ".", "Samples", "generated", "by", "the", "model", "are", "shown", "in", "Figure", "[", "reference", "]", ".", "subsubsection", ":", "Cifar", "-", "10", "Table", "[", "reference", "]", "shows", "likelihoods", "of", "different", "models", "on", "Cifar", "-", "10", ".", "We", "see", "that", "our", "method", "outperforms", "previous", "methods", "except", "for", "the", "just", "released", "Pixel", "RNN", "model", "of", ".", "As", "mentioned", ",", "the", "advantage", "of", "our", "model", "compared", "to", "such", "auto", "-", "regressive", "models", "is", "that", "it", "is", "a", "latent", "variable", "model", "that", "can", "be", "used", "for", "representation", "learning", "and", "lossy", "compression", ".", "At", "the", "same", "time", ",", "the", "two", "approaches", "are", "orthogonal", "and", "can", "be", "combined", ",", "for", "example", "by", "feeding", "the", "output", "of", "convolutional", "DRAW", "into", "the", "recurrent", "network", "of", "Pixel", "RNN", ".", "We", "also", "report", "the", "likelihood", "for", "a", "(", "non", "-", "recurrent", ")", "variational", "auto", "-", "encoder", "and", "standard", "DRAW", ".", "For", "the", "variational", "auto", "-", "encoder", "we", "tested", "architectures", "with", "multiple", "layers", ",", "both", "deterministic", "and", "stochastic", "but", "with", "standard", "functional", "forms", ",", "and", "this", "was", "the", "best", "result", "that", "we", "obtained", ".", "Convolutional", "DRAW", "performs", "significantly", "better", ".", "subsubsection", ":", "ImageNet", "Additionally", ",", "we", "trained", "on", "the", "version", "of", "ImageNet", "prepared", "in", "which", "was", "created", "with", "the", "aim", "of", "making", "a", "standardized", "dataset", "to", "test", "generative", "models", ".", "The", "results", "are", "in", "Table", "[", "reference", "]", ".", "Note", "that", "being", "a", "new", "dataset", ",", "no", "other", "methods", "have", "been", "reported", "on", "it", ".", "In", "Figure", "[", "reference", "]", "and", "Figure", "[", "reference", "]", "we", "show", "generations", "from", "the", "model", ".", "We", "trained", "networks", "with", "varying", "input", "cost", "scales", "as", "explained", "in", "the", "next", "section", ".", "The", "generations", "are", "sharp", "and", "contain", "many", "details", ",", "unlike", "previous", "versions", "of", "variational", "auto", "-", "encoder", "that", "tend", "to", "generate", "blurry", "images", ".", "subsection", ":", "Input", "Cost", "Scaling", "Each", "pixel", "(", "and", "color", "channel", ")", "of", "the", "data", "consists", "of", "256", "values", ",", "and", "as", "such", ",", "likelihood", "and", "lossless", "compression", "are", "well", "defined", ".", "When", "compressing", "the", "image", "there", "is", "much", "to", "be", "gained", "in", "capturing", "precise", "correlations", "between", "nearby", "pixels", ".", "There", "are", "a", "lot", "more", "bits", "in", "these", "low", "level", "details", "than", "in", "the", "higher", "level", "structure", "that", "we", "are", "actually", "interested", "in", "when", "learning", "higher", "level", "representations", ".", "The", "network", "might", "focus", "on", "these", "details", ",", "ignoring", "higher", "level", "structure", ".", "One", "way", "to", "make", "it", "focus", "less", "on", "the", "details", "is", "to", "scale", "down", "the", "cost", "of", "the", "input", "relative", "to", "the", "latents", ",", "that", "is", ",", "setting", "in", "(", "[", "reference", "]", ")", ".", "Generations", "for", "different", "cost", "scalings", "are", "shown", "in", "Figure", "[", "reference", "]", ",", "with", "the", "original", "objective", "being", "scale", ".", "We", "see", "that", "lower", "scales", "indeed", "have", "a", "‘", "cleaner", "’", "high", "level", "structure", ".", "Scale", "1", "contains", "a", "lot", "of", "information", "at", "the", "precise", "pixel", "values", "and", "the", "network", "tries", "to", "capture", "that", ",", "while", "not", "being", "good", "enough", "to", "properly", "align", "details", "and", "produce", "realistic", "patterns", ".", "This", "might", "be", "simply", "a", "matter", "of", "scaling", ",", "making", "layers", "larger", ",", "networks", "deeper", ",", "using", "more", "iterations", ",", "or", "using", "better", "functional", "forms", ".", "subsection", ":", "The", "Dependence", "on", "Computational", "Depth", "Convolutional", "DRAW", "uses", "many", "iterations", "and", "might", "be", "considered", "expensive", ".", "However", "we", "found", "that", "networks", "with", "a", "larger", "number", "of", "time", "steps", "train", "faster", "per", "data", "example", "as", "shown", "in", "Figure", "[", "reference", "]", "(", "left", ")", ".", "To", "study", "how", "they", "train", "with", "respect", "to", "real", "time", ",", "we", "multiply", "the", "time", "scale", "of", "each", "input", "by", "the", "number", "of", "iterations", "as", "seen", "in", "Figure", "[", "reference", "]", "(", "right", ")", ".", "We", "see", "that", "despite", "having", "to", "do", "several", "iterations", ",", "up", "to", "about", ",", "convolutional", "DRAW", "does", "not", "take", "more", "wall", "clock", "time", "to", "train", "than", "the", "same", "architecture", "with", "smaller", ".", "For", "larger", ",", "the", "training", "slows", "down", ",", "but", "it", "does", "eventually", "reach", "better", "performance", "than", "at", "lower", ".", "subsection", ":", "Information", "Distribution", "We", "look", "at", "how", "much", "information", "different", "levels", "and", "time", "steps", "contain", ".", "This", "information", "is", "simply", "the", "KL", "divergence", "in", "(", "[", "reference", "]", ")", "and", "(", "[", "reference", "]", ")", ".", "For", "a", "two", "layer", "system", "with", "one", "convolutional", "and", "one", "fully", "connected", "layer", ",", "this", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "We", "see", "that", "the", "higher", "level", "contains", "information", "mainly", "at", "the", "beginning", "of", "the", "computation", ",", "whereas", "the", "lower", "layer", "starts", "with", "low", "information", "which", "then", "gradually", "increases", ".", "This", "is", "desirable", "from", "a", "conceptual", "point", "of", "view", ".", "It", "suggests", "that", "the", "network", "first", "finds", "out", "about", "an", "overall", "structure", "of", "the", "image", "and", "then", "explains", "the", "details", "contained", "within", "that", "structure", ".", "Understanding", "the", "overall", "structure", "rapidly", "is", "also", "convenient", "if", "the", "algorithm", "needs", "to", "respond", "to", "observations", "in", "a", "timely", "manner", ".", "subsection", ":", "Lossy", "Compression", "We", "can", "compress", "an", "image", "with", "loss", "of", "information", "by", "storing", "only", "a", "subset", "of", "the", "latent", "variables", ",", "typically", "the", "high", "levels", "of", "the", "hierarchy", ".", "We", "can", "do", "this", "in", "multilayer", "convolutional", "DRAW", ",", "storing", "only", "higher", "levels", ".", "However", "we", "can", "also", "store", "only", "a", "subset", "of", "time", "steps", ",", "specifically", "a", "given", "number", "of", "time", "steps", "at", "the", "beginning", ",", "and", "let", "the", "network", "generate", "the", "rest", ".", "The", "units", "not", "stored", "should", "be", "generated", "from", "the", "prior", "distribution", "(", "Equation", "[", "reference", "]", ")", ".", "However", "we", "can", "also", "generate", "a", "more", "likely", "image", "by", "lowering", "the", "variance", "of", "the", "prior", "Gaussian", ".", "We", "show", "generations", "with", "full", "variance", "in", "row", "3", "of", "each", "block", "of", "Figure", "[", "reference", "]", "and", "with", "variance", "zero", "in", "row", "4", "of", "that", "figure", ".", "We", "see", "that", "using", "the", "original", "variance", ",", "the", "network", "generates", "sharp", "details", ".", "Because", "the", "generative", "model", "is", "not", "perfect", ",", "the", "resulting", "images", "are", "less", "realistic", "looking", "as", "we", "lower", "the", "number", "of", "stored", "time", "steps", ".", "For", "zero", "variance", "we", "see", "that", "the", "network", "starts", "with", "rough", "details", "making", "a", "smooth", "image", "and", "then", "refines", "it", "with", "more", "time", "steps", ".", "All", "these", "generations", "are", "produced", "with", "a", "single", "-", "layer", "convolutional", "DRAW", ",", "and", "thus", ",", "despite", "being", "single", "-", "layer", ",", "it", "achieves", "some", "level", "of", "‘", "conceptual", "compression", "’", "by", "first", "capturing", "the", "global", "structure", "of", "the", "image", "and", "then", "focusing", "on", "details", ".", "There", "is", "another", "dimension", "we", "can", "vary", "for", "lossy", "compression", "–", "the", "input", "scale", "introduced", "in", "subsection", "[", "reference", "]", ".", "Even", "if", "we", "store", "all", "the", "latent", "variables", "(", "but", "not", "the", "input", "bits", ")", ",", "the", "reconstructed", "images", "will", "get", "less", "detailed", "as", "we", "scale", "down", "the", "input", "cost", ".", "To", "build", "a", "really", "good", "compressor", ",", "at", "each", "compression", "rate", ",", "we", "need", "to", "find", "which", "of", "the", "networks", ",", "input", "scales", "and", "number", "of", "time", "steps", "would", "produce", "visually", "good", "images", ".", "For", "several", "compression", "levels", ",", "we", "have", "looked", "at", "images", "produced", "by", "different", "methods", "and", "selected", "qualitatively", "which", "network", "gave", "the", "best", "looking", "images", ".", "We", "have", "not", "done", "this", "per", "image", ",", "just", "per", "compression", "level", ".", "We", "then", "display", "compressed", "images", "that", "we", "have", "not", "seen", "with", "this", "selection", ".", "We", "compare", "our", "results", "to", "JPEG", "and", "JPEG2000", "compression", "which", "we", "obtained", "using", "ImageMagick", ".", "We", "found", "however", "that", "these", "compressors", "are", "unable", "to", "produce", "reasonable", "results", "for", "small", "images", "(", ")", "at", "high", "compression", "rates", ".", "Instead", ",", "we", "concatenated", "100", "images", "into", "one", "image", ",", "compressed", "that", "and", "extracted", "back", "the", "compressed", "small", "images", ".", "The", "number", "of", "bits", "per", "image", "reported", "is", "then", "the", "number", "of", "bits", "of", "this", "image", "divided", "by", "100", ".", "This", "is", "actually", "unfair", "to", "our", "algorithm", "since", "any", "correlations", "between", "nearby", "images", "can", "be", "exploited", ".", "Nevertheless", "we", "show", "the", "comparison", "in", "Figure", "[", "reference", "]", ".", "Our", "algorithm", "shows", "better", "quality", "than", "JPEG", "and", "JPEG", "2000", "at", "all", "levels", "where", "a", "corruption", "is", "easily", "detectable", ".", "Note", "that", "even", "when", "our", "algorithm", "is", "trained", "on", "one", "specific", "image", "size", ",", "it", "can", "be", "used", "on", "arbitrarily", "sized", "images", "for", "those", "networks", "that", "contain", "only", "convolutional", "operators", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "introduced", "convolutional", "DRAW", ",", "a", "state", "-", "of", "-", "the", "-", "art", "generative", "model", "which", "demonstrates", "the", "potential", "of", "sequential", "computation", "and", "recurrent", "neural", "networks", "in", "scaling", "up", "latent", "variable", "models", ".", "During", "inference", ",", "the", "algorithm", "sequentially", "arrives", "at", "a", "natural", "stratification", "of", "information", ",", "ranging", "from", "global", "aspects", "to", "low", "-", "level", "details", ".", "An", "interesting", "feature", "of", "the", "method", "is", "that", ",", "when", "we", "restrict", "ourselves", "to", "storing", "just", "the", "high", "level", "latent", "variables", ",", "we", "arrive", "at", "a", "‘", "conceptual", "compression", "’", "algorithm", "that", "rivals", "the", "quality", "of", "JPEG2000", ".", "As", "a", "generative", "model", ",", "it", "outperforms", "earlier", "latent", "variable", "models", "on", "both", "the", "Omniglot", "and", "ImageNet", "datasets", ".", "section", ":", "Acknowledgements", "We", "thank", "Aaron", "van", "den", "Oord", ",", "Diederik", "Kingma", "and", "Koray", "Kavukcuoglu", "for", "fruitful", "discussions", ".", "bibliography", ":", "References", "section", ":", "Appendix", "Figures", "[", "reference", "]", ",", "[", "reference", "]", "show", "image", "generations", "for", "input", "scaling", "and", "of", "(", "[", "reference", "]", ")", ",", "while", "Figures", "[", "reference", "]", ",", "[", "reference", "]", "show", "generations", ",", "also", "for", "and", "."], "entity": [[4, 10, ["recurrent", "variational", "auto", "-", "encoder", "architecture"], "Method"], [13, 15, ["image", "modeling"], "Task"], [28, 31, ["latent", "variable", "models"], "Method"], [34, 35, ["ImageNet"], "Method"], [36, 38, ["Omniglot", "datasets"], "Method"], [62, 64, ["unsupervised", "learning"], "Method"], [87, 89, ["conceptual", "compression"], "Metric"], [91, 92, ["ConceptualCompression"], "Method"], [114, 119, ["semi", "-", "supervised", "learning", "regime"], "Task"], [154, 155, ["generalization"], "Task"], [223, 227, ["latent", "variable", "image", "modeling"], "Task"], [232, 237, ["deep", "variational", "auto", "-", "encoders"], "Method"], [248, 250, ["homogeneous", "architecture"], "Method"], [261, 266, ["recurrent", "structure", "of", "DRAW", "("], "Method"], [278, 285, ["variational", "auto", "-", "encoder", "models", "of", "images"], "Method"], [320, 322, ["lossy", "compression"], "Task"], [344, 345, ["decompression"], "Task"], [358, 360, ["lossy", "compression"], "Task"], [361, 363, ["human", "evaluation"], "Task"], [406, 408, ["L2", "distance"], "Metric"], [426, 428, ["lossy", "compression"], "Metric"], [464, 466, ["abstract", "representations"], "Method"], [467, 469, ["human", "evaluation"], "Metric"], [470, 472, ["lossy", "compression"], "Metric"], [475, 477, ["generative", "models"], "Method"], [496, 500, ["variational", "auto", "-", "encoders"], "Method"], [501, 502, ["compression"], "Task"], [517, 518, ["compression"], "Task"], [521, 525, ["Variational", "Auto", "-", "Encoders"], "Method"], [529, 534, ["unsupervised", "learning", "in", "deep", "networks"], "Task"], [536, 540, ["sparse", "auto", "-", "encoders"], "Method"], [541, 543, ["sparse", "coding"], "Method"], [544, 548, ["denoising", "auto", "-", "encoders"], "Method"], [549, 551, ["deconvolutional", "networks"], "Method"], [552, 555, ["restricted", "Boltzmann", "machines"], "Method"], [556, 559, ["deep", "Boltzmann", "machines"], "Method"], [560, 563, ["generative", "adversarial", "networks"], "Method"], [564, 568, ["variational", "auto", "-", "encoders"], "Method"], [581, 586, ["variational", "auto", "-", "encoding", "framework"], "Method"], [593, 594, ["compression"], "Task"], [600, 604, ["information", "-", "theoretic", "perspective"], "Method"], [605, 609, ["Variational", "auto", "-", "encoders"], "Method"], [613, 615, ["neural", "networks"], "Method"], [639, 640, ["recognition"], "Task"], [710, 711, ["recognition"], "Task"], [766, 769, ["reduction", "in", "uncertainty"], "Metric"], [785, 786, ["generation"], "Task"], [820, 824, ["Variational", "auto", "-", "encoders"], "Method"], [829, 831, ["unsupervised", "learning"], "Task"], [835, 836, ["networks"], "Method"], [904, 906, ["Conceptual", "Compression"], "Method"], [906, 910, ["Variational", "auto", "-", "encoders"], "Method"], [916, 918, ["representation", "learning"], "Task"], [921, 922, ["compression"], "Task"], [927, 931, ["variational", "auto", "-", "encoders"], "Method"], [1001, 1003, ["lossless", "compression"], "Task"], [1013, 1014, ["entropy"], "Metric"], [1044, 1046, ["lossless", "compression"], "Method"], [1062, 1064, ["Lossy", "compression"], "Task"], [1099, 1100, ["quality"], "Metric"], [1101, 1102, ["similarity"], "Metric"], [1116, 1117, ["compression"], "Task"], [1120, 1122, ["lossless", "compression"], "Task"], [1144, 1145, ["JPEG"], "Method"], [1214, 1215, ["compression"], "Task"], [1301, 1303, ["generative", "model"], "Method"], [1393, 1395, ["generative", "model"], "Method"], [1455, 1456, ["compression"], "Method"], [1457, 1458, ["compressing"], "Task"], [1472, 1474, ["conceptual", "compression"], "Method"], [1486, 1488, ["lossy", "compression"], "Task"], [1494, 1497, ["deep", "representation", "learning"], "Method"], [1498, 1502, ["latent", "variable", "generative", "models"], "Method"], [1513, 1515, ["lossy", "compression"], "Task"], [1565, 1568, ["unsupervised", "deep", "learning"], "Method"], [1590, 1592, ["unsupervised", "learning"], "Task"], [1654, 1656, ["compression", "quality"], "Metric"], [1657, 1658, ["JPEG"], "Method"], [1659, 1660, ["JPEG2000"], "Method"], [1701, 1705, ["latent", "variable", "generative", "models"], "Method"], [1721, 1724, ["latent", "variable", "models"], "Method"], [1732, 1736, ["generative", "models", "of", "images"], "Method"], [1739, 1741, ["vision", "architectures"], "Method"], [1813, 1818, ["deep", "variational", "auto", "-", "encoders"], "Method"], [1883, 1886, ["iterative", "feedback", "mechanism"], "Method"], [1901, 1903, ["DRAW", "algorithm"], "Method"], [1907, 1914, ["recurrent", "type", "of", "variational", "auto", "-", "encoder"], "Method"], [1923, 1925, ["provisionary", "reconstruction"], "Method"], [1948, 1949, ["reconstruction"], "Task"], [1951, 1954, ["iterative", "feedback", "mechanism"], "Method"], [1958, 1961, ["back", "-", "propagation"], "Method"], [1962, 1963, ["Computation"], "Task"], [1978, 1980, ["convolutional", "DRAW"], "Method"], [1983, 1984, ["convolutions"], "Method"], [1985, 1988, ["latent", "prior", "modeling"], "Method"], [1990, 1993, ["Gaussian", "input", "distribution"], "Method"], [2005, 2009, ["multi", "-", "layer", "architecture"], "Method"], [2018, 2020, ["attentional", "mechanism"], "Method"], [2026, 2030, ["single", "-", "layer", "version"], "Method"], [2033, 2036, ["deep", "generative", "model"], "Method"], [2062, 2065, ["convolutional", "DRAW", "hierarchical"], "Method"], [2074, 2079, ["deep", "variational", "auto", "-", "encoders"], "Method"], [2093, 2094, ["recurrence"], "Method"], [2099, 2102, ["accurate", "pixel", "reconstructions"], "Task"], [2203, 2207, ["Non", "-", "variational", "Models"], "Method"], [2217, 2219, ["generative", "models"], "Method"], [2221, 2224, ["generative", "adversarial", "networks"], "Method"], [2225, 2226, ["GANs"], "Method"], [2229, 2234, ["auto", "-", "regressive", "pixel", "models"], "Method"], [2235, 2236, ["GANs"], "Method"], [2256, 2259, ["feedforward", "generative", "network"], "Method"], [2260, 2261, ["GANs"], "Method"], [2267, 2269, ["generative", "network"], "Method"], [2275, 2279, ["variational", "auto", "-", "encoders"], "Method"], [2282, 2284, ["classification", "network"], "Method"], [2286, 2288, ["classification", "network"], "Method"], [2316, 2318, ["generative", "network"], "Method"], [2322, 2324, ["classification", "network"], "Method"], [2343, 2345, ["classification", "network"], "Method"], [2359, 2361, ["classification", "network"], "Method"], [2394, 2396, ["classification", "network"], "Method"], [2443, 2447, ["GANs", "’", "sampling", "diversity"], "Metric"], [2495, 2496, ["classifier"], "Method"], [2583, 2585, ["autoregressive", "fashion"], "Method"], [2724, 2727, ["pixel", "level", "approach"], "Method"], [2745, 2747, ["convolutional", "DRAW"], "Method"], [2749, 2751, ["conditional", "computation"], "Method"], [2753, 2756, ["pixel", "level", "model"], "Method"], [2763, 2766, ["latent", "variable", "approach"], "Method"], [2777, 2779, ["Convolutional", "DRAW"], "Method"], [2788, 2792, ["single", "-", "layer", "version"], "Method"], [2796, 2798, ["Convolutional", "DRAW"], "Method"], [2814, 2817, ["encoder", "recurrent", "net"], "Method"], [2855, 2857, ["convolutional", "DRAW"], "Method"], [2865, 2866, ["convolution"], "Method"], [2867, 2868, ["Rnn"], "Method"], [2870, 2872, ["recurrent", "network"], "Method"], [2875, 2879, ["LSTM", "with", "convolutional", "operators"], "Method"], [2883, 2885, ["linear", "ones"], "Method"], [2904, 2906, ["Bernoulli", "distribution"], "Method"], [2913, 2915, ["Gaussian", "distribution"], "Method"], [2928, 2930, ["input", "cost"], "Metric"], [2932, 2934, ["total", "cost"], "Metric"], [2994, 2998, ["stochastic", "gradient", "descent", "algorithm"], "Method"], [2999, 3003, ["Stochastic", "back", "-", "propagation"], "Method"], [3005, 3007, ["sampling", "function"], "Method"], [3011, 3015, ["variational", "auto", "-", "encoders"], "Method"], [3068, 3072, ["soft", "-", "max", "distribution"], "Method"], [3110, 3112, ["Gaussian", "distribution"], "Method"], [3168, 3172, ["Multi", "-", "layer", "Architectures"], "Method"], [3179, 3181, ["convolutional", "DRAW"], "Method"], [3207, 3209, ["recurrent", "encoder"], "Method"], [3210, 3212, ["recurrent", "decoder"], "Method"], [3214, 3216, ["stochastic", "layer"], "Method"], [3260, 3265, ["first", "layer", "decoder", "recurrent", "net"], "Method"], [3341, 3342, ["Compression"], "Task"], [3349, 3353, ["variational", "auto", "-", "encoders"], "Method"], [3354, 3356, ["convolutional", "DRAW"], "Method"], [3357, 3359, ["compression", "algorithms"], "Method"], [3366, 3367, ["compressor"], "Method"], [3418, 3421, ["guaranteed", "compression", "rate"], "Metric"], [3436, 3438, ["compression", "rate"], "Metric"], [3453, 3455, ["compression", "mechanism"], "Method"], [3459, 3461, ["arithmetic", "coding"], "Method"], [3462, 3464, ["Arithmetic", "coding"], "Method"], [3504, 3508, ["variational", "auto", "-", "encoders"], "Method"], [3550, 3552, ["arithmetic", "coding"], "Method"], [3663, 3664, ["sampling"], "Method"], [3695, 3699, ["bits", "-", "back", "coding"], "Method"], [3719, 3720, ["precision"], "Metric"], [3888, 3890, ["coding", "scheme"], "Method"], [3930, 3933, ["stride", "2", "convolutions"], "Method"], [3949, 3952, ["Cifar", "-", "10"], "Material"], [3953, 3954, ["ImageNet"], "Method"], [3966, 3967, ["ImageNet"], "Method"], [3985, 3987, ["Adam", "algorithm"], "Method"], [3988, 3990, ["learning", "rate"], "Method"], [4076, 4078, ["Modeling", "Quality"], "Metric"], [4108, 4112, ["inverse", "of", "MNIST", "’"], "Method"], [4118, 4120, ["conceptual", "representations"], "Method"], [4121, 4123, ["generative", "models"], "Method"], [4181, 4184, ["Cifar", "-", "10"], "Material"], [4194, 4197, ["Cifar", "-", "10"], "Material"], [4210, 4214, ["released", "Pixel", "RNN", "model"], "Method"], [4227, 4231, ["auto", "-", "regressive", "models"], "Method"], [4236, 4239, ["latent", "variable", "model"], "Method"], [4244, 4246, ["representation", "learning"], "Task"], [4247, 4249, ["lossy", "compression"], "Task"], [4272, 4274, ["convolutional", "DRAW"], "Method"], [4276, 4281, ["recurrent", "network", "of", "Pixel", "RNN"], "Method"], [4290, 4298, ["non", "-", "recurrent", ")", "variational", "auto", "-", "encoder"], "Method"], [4299, 4301, ["standard", "DRAW"], "Method"], [4304, 4308, ["variational", "auto", "-", "encoder"], "Method"], [4316, 4319, ["deterministic", "and", "stochastic"], "Method"], [4335, 4337, ["Convolutional", "DRAW"], "Method"], [4343, 4344, ["ImageNet"], "Method"], [4352, 4353, ["ImageNet"], "Method"], [4368, 4370, ["generative", "models"], "Method"], [4441, 4445, ["variational", "auto", "-", "encoder"], "Method"], [4454, 4457, ["Input", "Cost", "Scaling"], "Method"], [4476, 4477, ["likelihood"], "Method"], [4478, 4480, ["lossless", "compression"], "Task"], [4527, 4530, ["higher", "level", "representations"], "Method"], [4681, 4683, ["Convolutional", "DRAW"], "Method"], [4769, 4771, ["convolutional", "DRAW"], "Method"], [4808, 4810, ["Information", "Distribution"], "Task"], [4845, 4848, ["two", "layer", "system"], "Method"], [4850, 4856, ["convolutional", "and", "one", "fully", "connected", "layer"], "Method"], [4953, 4955, ["Lossy", "Compression"], "Method"], [5107, 5109, ["generative", "model"], "Method"], [5183, 5185, ["conceptual", "compression"], "Task"], [5209, 5211, ["lossy", "compression"], "Task"], [5258, 5259, ["compressor"], "Method"], [5262, 5264, ["compression", "rate"], "Metric"], [5289, 5291, ["compression", "levels"], "Metric"], [5322, 5324, ["compression", "level"], "Metric"], [5352, 5353, ["ImageMagick"], "Material"], [5359, 5360, ["compressors"], "Method"], [5373, 5375, ["compression", "rates"], "Metric"], [5448, 5449, ["quality"], "Metric"], [5450, 5451, ["JPEG"], "Method"], [5492, 5494, ["convolutional", "operators"], "Method"], [5504, 5506, ["convolutional", "DRAW"], "Method"], [5515, 5517, ["generative", "model"], "Method"], [5522, 5524, ["sequential", "computation"], "Method"], [5525, 5528, ["recurrent", "neural", "networks"], "Method"], [5531, 5534, ["latent", "variable", "models"], "Method"], [5536, 5537, ["inference"], "Task"], [5586, 5590, ["conceptual", "compression", "’", "algorithm"], "Method"], [5593, 5594, ["quality"], "Metric"], [5595, 5596, ["JPEG2000"], "Method"], [5599, 5601, ["generative", "model"], "Method"], [5605, 5608, ["latent", "variable", "models"], "Method"], [5611, 5612, ["Omniglot"], "Material"], [5613, 5615, ["ImageNet", "datasets"], "Material"], [514, 515, ["generation"], "Task"], [2349, 2350, ["generation"], "Task"], [4080, 4081, ["Omniglot"], "Material"], [4084, 4085, ["Omniglot"], "Material"], [4988, 4990, ["convolutional", "DRAW"], "Method"], [5165, 5167, ["convolutional", "DRAW"], "Method"], [5344, 5345, ["JPEG"], "Method"], [5452, 5453, ["JPEG"], "Method"]]} {"doc_id": "3cdb1364c3e66443e1c2182474d44b2fb01cd584", "text": ["document", ":", "SegNet", ":", "A", "Deep", "Convolutional", "Encoder", "-", "Decoder", "Architecture", "for", "Image", "Segmentation", "We", "present", "a", "novel", "and", "practical", "deep", "fully", "convolutional", "neural", "network", "architecture", "for", "semantic", "pixel", "-", "wise", "segmentation", "termed", "SegNet", ".", "This", "core", "trainable", "segmentation", "engine", "consists", "of", "an", "encoder", "network", ",", "a", "corresponding", "decoder", "network", "followed", "by", "a", "pixel", "-", "wise", "classification", "layer", ".", "The", "architecture", "of", "the", "encoder", "network", "is", "topologically", "identical", "to", "the", "13", "convolutional", "layers", "in", "the", "VGG16", "network", ".", "The", "role", "of", "the", "decoder", "network", "is", "to", "map", "the", "low", "resolution", "encoder", "feature", "maps", "to", "full", "input", "resolution", "feature", "maps", "for", "pixel", "-", "wise", "classification", ".", "The", "novelty", "of", "SegNet", "lies", "is", "in", "the", "manner", "in", "which", "the", "decoder", "upsamples", "its", "lower", "resolution", "input", "feature", "map", "(", "s", ")", ".", "Specifically", ",", "the", "decoder", "uses", "pooling", "indices", "computed", "in", "the", "max", "-", "pooling", "step", "of", "the", "corresponding", "encoder", "to", "perform", "non", "-", "linear", "upsampling", ".", "This", "eliminates", "the", "need", "for", "learning", "to", "upsample", ".", "The", "upsampled", "maps", "are", "sparse", "and", "are", "then", "convolved", "with", "trainable", "filters", "to", "produce", "dense", "feature", "maps", ".", "We", "compare", "our", "proposed", "architecture", "with", "the", "widely", "adopted", "FCN", "and", "also", "with", "the", "well", "known", "DeepLab", "-", "LargeFOV", ",", "DeconvNet", "architectures", ".", "This", "comparison", "reveals", "the", "memory", "versus", "accuracy", "trade", "-", "off", "involved", "in", "achieving", "good", "segmentation", "performance", ".", "SegNet", "was", "primarily", "motivated", "by", "scene", "understanding", "applications", ".", "Hence", ",", "it", "is", "designed", "to", "be", "efficient", "both", "in", "terms", "of", "memory", "and", "computational", "time", "during", "inference", ".", "It", "is", "also", "significantly", "smaller", "in", "the", "number", "of", "trainable", "parameters", "than", "other", "competing", "architectures", "and", "can", "be", "trained", "end", "-", "to", "-", "end", "using", "stochastic", "gradient", "descent", ".", "We", "also", "performed", "a", "controlled", "benchmark", "of", "SegNet", "and", "other", "architectures", "on", "both", "road", "scenes", "and", "SUN", "RGB", "-", "D", "indoor", "scene", "segmentation", "tasks", ".", "These", "quantitative", "assessments", "show", "that", "SegNet", "provides", "good", "performance", "with", "competitive", "inference", "time", "and", "most", "efficient", "inference", "memory", "-", "wise", "as", "compared", "to", "other", "architectures", ".", "We", "also", "provide", "a", "Caffe", "implementation", "of", "SegNet", "and", "a", "web", "demo", "at", ".", "Deep", "Convolutional", "Neural", "Networks", ",", "Semantic", "Pixel", "-", "Wise", "Segmentation", ",", "Indoor", "Scenes", ",", "Road", "Scenes", ",", "Encoder", ",", "Decoder", ",", "Pooling", ",", "Upsampling", ".", "section", ":", "Introduction", "Semantic", "segmentation", "has", "a", "wide", "array", "of", "applications", "ranging", "from", "scene", "understanding", ",", "inferring", "support", "-", "relationships", "among", "objects", "to", "autonomous", "driving", ".", "Early", "methods", "that", "relied", "on", "low", "-", "level", "vision", "cues", "have", "fast", "been", "superseded", "by", "popular", "machine", "learning", "algorithms", ".", "In", "particular", ",", "deep", "learning", "has", "seen", "huge", "success", "lately", "in", "handwritten", "digit", "recognition", ",", "speech", ",", "categorising", "whole", "images", "and", "detecting", "objects", "in", "images", ".", "Now", "there", "is", "an", "active", "interest", "for", "semantic", "pixel", "-", "wise", "labelling", ",", ",", ",", ",", ",", ",", ",", ",", ",", ",", ",", ".", "However", ",", "some", "of", "these", "recent", "approaches", "have", "tried", "to", "directly", "adopt", "deep", "architectures", "designed", "for", "category", "prediction", "to", "pixel", "-", "wise", "labelling", ".", "The", "results", ",", "although", "very", "encouraging", ",", "appear", "coarse", ".", "This", "is", "primarily", "because", "max", "pooling", "and", "sub", "-", "sampling", "reduce", "feature", "map", "resolution", ".", "Our", "motivation", "to", "design", "SegNet", "arises", "from", "this", "need", "to", "map", "low", "resolution", "features", "to", "input", "resolution", "for", "pixel", "-", "wise", "classification", ".", "This", "mapping", "must", "produce", "features", "which", "are", "useful", "for", "accurate", "boundary", "localization", ".", "Our", "architecture", ",", "SegNet", ",", "is", "designed", "to", "be", "an", "efficient", "architecture", "for", "pixel", "-", "wise", "semantic", "segmentation", ".", "It", "is", "primarily", "motivated", "by", "road", "scene", "understanding", "applications", "which", "require", "the", "ability", "to", "model", "appearance", "(", "road", ",", "building", ")", ",", "shape", "(", "cars", ",", "pedestrians", ")", "and", "understand", "the", "spatial", "-", "relationship", "(", "context", ")", "between", "different", "classes", "such", "as", "road", "and", "side", "-", "walk", ".", "In", "typical", "road", "scenes", ",", "the", "majority", "of", "the", "pixels", "belong", "to", "large", "classes", "such", "as", "road", ",", "building", "and", "hence", "the", "network", "must", "produce", "smooth", "segmentations", ".", "The", "engine", "must", "also", "have", "the", "ability", "to", "delineate", "objects", "based", "on", "their", "shape", "despite", "their", "small", "size", ".", "Hence", "it", "is", "important", "to", "retain", "boundary", "information", "in", "the", "extracted", "image", "representation", ".", "From", "a", "computational", "perspective", ",", "it", "is", "necessary", "for", "the", "network", "to", "be", "efficient", "in", "terms", "of", "both", "memory", "and", "computation", "time", "during", "inference", ".", "The", "ability", "to", "train", "end", "-", "to", "-", "end", "in", "order", "to", "jointly", "optimise", "all", "the", "weights", "in", "the", "network", "using", "an", "efficient", "weight", "update", "technique", "such", "as", "stochastic", "gradient", "descent", "(", "SGD", ")", "is", "an", "additional", "benefit", "since", "it", "is", "more", "easily", "repeatable", ".", "The", "design", "of", "SegNet", "arose", "from", "a", "need", "to", "match", "these", "criteria", ".", "The", "encoder", "network", "in", "SegNet", "is", "topologically", "identical", "to", "the", "convolutional", "layers", "in", "VGG16", ".", "We", "remove", "the", "fully", "connected", "layers", "of", "VGG16", "which", "makes", "the", "SegNet", "encoder", "network", "significantly", "smaller", "and", "easier", "to", "train", "than", "many", "other", "recent", "architectures", ".", "The", "key", "component", "of", "SegNet", "is", "the", "decoder", "network", "which", "consists", "of", "a", "hierarchy", "of", "decoders", "one", "corresponding", "to", "each", "encoder", ".", "Of", "these", ",", "the", "appropriate", "decoders", "use", "the", "max", "-", "pooling", "indices", "received", "from", "the", "corresponding", "encoder", "to", "perform", "non", "-", "linear", "upsampling", "of", "their", "input", "feature", "maps", ".", "This", "idea", "was", "inspired", "from", "an", "architecture", "designed", "for", "unsupervised", "feature", "learning", ".", "Reusing", "max", "-", "pooling", "indices", "in", "the", "decoding", "process", "has", "several", "practical", "advantages", ";", "(", "i", ")", "it", "improves", "boundary", "delineation", ",", "(", "ii", ")", "it", "reduces", "the", "number", "of", "parameters", "enabling", "end", "-", "to", "-", "end", "training", ",", "and", "(", "iii", ")", "this", "form", "of", "upsampling", "can", "be", "incorporated", "into", "any", "encoder", "-", "decoder", "architecture", "such", "as", "with", "only", "a", "little", "modification", ".", "One", "of", "the", "main", "contributions", "of", "this", "paper", "is", "our", "analysis", "of", "the", "SegNet", "decoding", "technique", "and", "the", "widely", "used", "Fully", "Convolutional", "Network", "(", "FCN", ")", ".", "This", "is", "in", "order", "to", "convey", "the", "practical", "trade", "-", "offs", "involved", "in", "designing", "segmentation", "architectures", ".", "Most", "recent", "deep", "architectures", "for", "segmentation", "have", "identical", "encoder", "networks", ",", "i.e", "VGG16", ",", "but", "differ", "in", "the", "form", "of", "the", "decoder", "network", ",", "training", "and", "inference", ".", "Another", "common", "feature", "is", "they", "have", "trainable", "parameters", "in", "the", "order", "of", "hundreds", "of", "millions", "and", "thus", "encounter", "difficulties", "in", "performing", "end", "-", "to", "-", "end", "training", ".", "The", "difficulty", "of", "training", "these", "networks", "has", "led", "to", "multi", "-", "stage", "training", ",", "appending", "networks", "to", "a", "pre", "-", "trained", "architecture", "such", "as", "FCN", ",", "use", "of", "supporting", "aids", "such", "as", "region", "proposals", "for", "inference", ",", "disjoint", "training", "of", "classification", "and", "segmentation", "networks", "and", "use", "of", "additional", "training", "data", "for", "pre", "-", "training", "or", "for", "full", "training", ".", "In", "addition", ",", "performance", "boosting", "post", "-", "processing", "techniques", "have", "also", "been", "popular", ".", "Although", "all", "these", "factors", "improve", "performance", "on", "challenging", "benchmarks", ",", "it", "is", "unfortunately", "difficult", "from", "their", "quantitative", "results", "to", "disentangle", "the", "key", "design", "factors", "necessary", "to", "achieve", "good", "performance", ".", "We", "therefore", "analysed", "the", "decoding", "process", "used", "in", "some", "of", "these", "approaches", "and", "reveal", "their", "pros", "and", "cons", ".", "We", "evaluate", "the", "performance", "of", "SegNet", "on", "two", "scene", "segmentation", "tasks", ",", "CamVid", "road", "scene", "segmentation", "and", "SUN", "RGB", "-", "D", "indoor", "scene", "segmentation", ".", "Pascal", "VOC12", "has", "been", "the", "benchmark", "challenge", "for", "segmentation", "over", "the", "years", ".", "However", ",", "the", "majority", "of", "this", "task", "has", "one", "or", "two", "foreground", "classes", "surrounded", "by", "a", "highly", "varied", "background", ".", "This", "implicitly", "favours", "techniques", "used", "for", "detection", "as", "shown", "by", "the", "recent", "work", "on", "a", "decoupled", "classification", "-", "segmentation", "network", "where", "the", "classification", "network", "can", "be", "trained", "with", "a", "large", "set", "of", "weakly", "labelled", "data", "and", "the", "independent", "segmentation", "network", "performance", "is", "improved", ".", "The", "method", "of", "also", "use", "the", "feature", "maps", "of", "the", "classification", "network", "with", "an", "independent", "CRF", "post", "-", "processing", "technique", "to", "perform", "segmentation", ".", "The", "performance", "can", "also", "be", "boosted", "by", "the", "use", "additional", "inference", "aids", "such", "as", "region", "proposals", ",", ".", "Therefore", ",", "it", "is", "different", "from", "scene", "understanding", "where", "the", "idea", "is", "to", "exploit", "co", "-", "occurrences", "of", "objects", "and", "other", "spatial", "-", "context", "to", "perform", "robust", "segmentation", ".", "To", "demonstrate", "the", "efficacy", "of", "SegNet", ",", "we", "present", "a", "real", "-", "time", "online", "demo", "of", "road", "scene", "segmentation", "into", "11", "classes", "of", "interest", "for", "autonomous", "driving", "(", "see", "link", "in", "Fig", ".", "[", "reference", "]", ")", ".", "Some", "example", "test", "results", "produced", "on", "randomly", "sampled", "road", "scene", "images", "from", "Google", "and", "indoor", "test", "scenes", "from", "the", "SUN", "RGB", "-", "D", "dataset", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "The", "remainder", "of", "the", "paper", "is", "organized", "as", "follows", ".", "In", "Sec", ".", "[", "reference", "]", "we", "review", "related", "recent", "literature", ".", "We", "describe", "the", "SegNet", "architecture", "and", "its", "analysis", "in", "Sec", ".", "[", "reference", "]", ".", "In", "Sec", ".", "[", "reference", "]", "we", "evaluate", "the", "performance", "of", "SegNet", "on", "outdoor", "and", "indoor", "scene", "datasets", ".", "This", "is", "followed", "by", "a", "general", "discussion", "regarding", "our", "approach", "with", "pointers", "to", "future", "work", "in", "Sec", ".", "[", "reference", "]", ".", "We", "conclude", "in", "Sec", ".", "[", "reference", "]", ".", "section", ":", "Literature", "Review", "Semantic", "pixel", "-", "wise", "segmentation", "is", "an", "active", "topic", "of", "research", ",", "fuelled", "by", "challenging", "datasets", ".", "Before", "the", "arrival", "of", "deep", "networks", ",", "the", "best", "performing", "methods", "mostly", "relied", "on", "hand", "engineered", "features", "classifying", "pixels", "independently", ".", "Typically", ",", "a", "patch", "is", "fed", "into", "a", "classifier", "e.g.", "Random", "Forest", "or", "Boosting", "to", "predict", "the", "class", "probabilities", "of", "the", "center", "pixel", ".", "Features", "based", "on", "appearance", "or", "SfM", "and", "appearance", "have", "been", "explored", "for", "the", "CamVid", "road", "scene", "understanding", "test", ".", "These", "per", "-", "pixel", "noisy", "predictions", "(", "often", "called", "unary", "terms", ")", "from", "the", "classifiers", "are", "then", "smoothed", "by", "using", "a", "pair", "-", "wise", "or", "higher", "order", "CRF", "to", "improve", "the", "accuracy", ".", "More", "recent", "approaches", "have", "aimed", "to", "produce", "high", "quality", "unaries", "by", "trying", "to", "predict", "the", "labels", "for", "all", "the", "pixels", "in", "a", "patch", "as", "opposed", "to", "only", "the", "center", "pixel", ".", "This", "improves", "the", "results", "of", "Random", "Forest", "based", "unaries", "but", "thin", "structured", "classes", "are", "classified", "poorly", ".", "Dense", "depth", "maps", "computed", "from", "the", "CamVid", "video", "have", "also", "been", "used", "as", "input", "for", "classification", "using", "Random", "Forests", ".", "Another", "approach", "argues", "for", "the", "use", "of", "a", "combination", "of", "popular", "hand", "designed", "features", "and", "spatio", "-", "temporal", "super", "-", "pixelization", "to", "obtain", "higher", "accuracy", ".", "The", "best", "performing", "technique", "on", "the", "CamVid", "test", "addresses", "the", "imbalance", "among", "label", "frequencies", "by", "combining", "object", "detection", "outputs", "with", "classifier", "predictions", "in", "a", "CRF", "framework", ".", "The", "result", "of", "all", "these", "techniques", "indicate", "the", "need", "for", "improved", "features", "for", "classification", ".", "Indoor", "RGBD", "pixel", "-", "wise", "semantic", "segmentation", "has", "also", "gained", "popularity", "since", "the", "release", "of", "the", "NYU", "dataset", ".", "This", "dataset", "showed", "the", "usefulness", "of", "the", "depth", "channel", "to", "improve", "segmentation", ".", "Their", "approach", "used", "features", "such", "as", "RGB", "-", "SIFT", ",", "depth", "-", "SIFT", "and", "pixel", "location", "as", "input", "to", "a", "neural", "network", "classifier", "to", "predict", "pixel", "unaries", ".", "The", "noisy", "unaries", "are", "then", "smoothed", "using", "a", "CRF", ".", "Improvements", "were", "made", "using", "a", "richer", "feature", "set", "including", "LBP", "and", "region", "segmentation", "to", "obtain", "higher", "accuracy", "followed", "by", "a", "CRF", ".", "In", "more", "recent", "work", ",", "both", "class", "segmentation", "and", "support", "relationships", "are", "inferred", "together", "using", "a", "combination", "of", "RGB", "and", "depth", "based", "cues", ".", "Another", "approach", "focuses", "on", "real", "-", "time", "joint", "reconstruction", "and", "semantic", "segmentation", ",", "where", "Random", "Forests", "are", "used", "as", "the", "classifier", ".", "Gupta", "et", "al", ".", "use", "boundary", "detection", "and", "hierarchical", "grouping", "before", "performing", "category", "segmentation", ".", "The", "common", "attribute", "in", "all", "these", "approaches", "is", "the", "use", "of", "hand", "engineered", "features", "for", "classification", "of", "either", "RGB", "or", "RGBD", "images", ".", "The", "success", "of", "deep", "convolutional", "neural", "networks", "for", "object", "classification", "has", "more", "recently", "led", "researchers", "to", "exploit", "their", "feature", "learning", "capabilities", "for", "structured", "prediction", "problems", "such", "as", "segmentation", ".", "There", "have", "also", "been", "attempts", "to", "apply", "networks", "designed", "for", "object", "categorization", "to", "segmentation", ",", "particularly", "by", "replicating", "the", "deepest", "layer", "features", "in", "blocks", "to", "match", "image", "dimensions", ".", "However", ",", "the", "resulting", "classification", "is", "blocky", ".", "Another", "approach", "using", "recurrent", "neural", "networks", "merges", "several", "low", "resolution", "predictions", "to", "create", "input", "image", "resolution", "predictions", ".", "These", "techniques", "are", "already", "an", "improvement", "over", "hand", "engineered", "features", "but", "their", "ability", "to", "delineate", "boundaries", "is", "poor", ".", "Newer", "deep", "architectures", "particularly", "designed", "for", "segmentation", "have", "advanced", "the", "state", "-", "of", "-", "the", "-", "art", "by", "learning", "to", "decode", "or", "map", "low", "resolution", "image", "representations", "to", "pixel", "-", "wise", "predictions", ".", "The", "encoder", "network", "which", "produces", "these", "low", "resolution", "representations", "in", "all", "of", "these", "architectures", "is", "the", "VGG16", "classification", "network", "which", "has", "convolutional", "layers", "and", "fully", "connected", "layers", ".", "This", "encoder", "network", "weights", "are", "typically", "pre", "-", "trained", "on", "the", "large", "ImageNet", "object", "classification", "dataset", ".", "The", "decoder", "network", "varies", "between", "these", "architectures", "and", "is", "the", "part", "which", "is", "responsible", "for", "producing", "multi", "-", "dimensional", "features", "for", "each", "pixel", "for", "classification", ".", "Each", "decoder", "in", "the", "Fully", "Convolutional", "Network", "(", "FCN", ")", "architecture", "learns", "to", "upsample", "its", "input", "feature", "map", "(", "s", ")", "and", "combines", "them", "with", "the", "corresponding", "encoder", "feature", "map", "to", "produce", "the", "input", "to", "the", "next", "decoder", ".", "It", "is", "an", "architecture", "which", "has", "a", "large", "number", "of", "trainable", "parameters", "in", "the", "encoder", "network", "(", "134", "M", ")", "but", "a", "very", "small", "decoder", "network", "(", "0.5", "M", ")", ".", "The", "overall", "large", "size", "of", "this", "network", "makes", "it", "hard", "to", "train", "end", "-", "to", "-", "end", "on", "a", "relevant", "task", ".", "Therefore", ",", "the", "authors", "use", "a", "stage", "-", "wise", "training", "process", ".", "Here", "each", "decoder", "in", "the", "decoder", "network", "is", "progressively", "added", "to", "an", "existing", "trained", "network", ".", "The", "network", "is", "grown", "until", "no", "further", "increase", "in", "performance", "is", "observed", ".", "This", "growth", "is", "stopped", "after", "three", "decoders", "thus", "ignoring", "high", "resolution", "feature", "maps", "can", "certainly", "lead", "to", "loss", "of", "edge", "information", ".", "Apart", "from", "training", "related", "issues", ",", "the", "need", "to", "reuse", "the", "encoder", "feature", "maps", "in", "the", "decoder", "makes", "it", "memory", "intensive", "in", "test", "time", ".", "We", "study", "this", "network", "in", "more", "detail", "as", "it", "the", "core", "of", "other", "recent", "architectures", ".", "The", "predictive", "performance", "of", "FCN", "has", "been", "improved", "further", "by", "appending", "the", "FCN", "with", "a", "recurrent", "neural", "network", "(", "RNN", ")", "and", "fine", "-", "tuning", "them", "on", "large", "datasets", ",", ".", "The", "RNN", "layers", "mimic", "the", "sharp", "boundary", "delineation", "capabilities", "of", "CRFs", "while", "exploiting", "the", "feature", "representation", "power", "of", "FCN", "’s", ".", "They", "show", "a", "significant", "improvement", "over", "FCN", "-", "8", "but", "also", "show", "that", "this", "difference", "is", "reduced", "when", "more", "training", "data", "is", "used", "to", "train", "FCN", "-", "8", ".", "The", "main", "advantage", "of", "the", "CRF", "-", "RNN", "is", "revealed", "when", "it", "is", "jointly", "trained", "with", "an", "architecture", "such", "as", "the", "FCN", "-", "8", ".", "The", "fact", "that", "joint", "training", "helps", "is", "also", "shown", "in", "other", "recent", "results", ",", ".", "Interestingly", ",", "the", "deconvolutional", "network", "performs", "significantly", "better", "than", "FCN", "although", "at", "the", "cost", "of", "a", "more", "complex", "training", "and", "inference", ".", "This", "however", "raises", "the", "question", "as", "to", "whether", "the", "perceived", "advantage", "of", "the", "CRF", "-", "RNN", "would", "be", "reduced", "as", "the", "core", "feed", "-", "forward", "segmentation", "engine", "is", "made", "better", ".", "In", "any", "case", ",", "the", "CRF", "-", "RNN", "network", "can", "be", "appended", "to", "any", "deep", "segmentation", "architecture", "including", "SegNet", ".", "Multi", "-", "scale", "deep", "architectures", "are", "also", "being", "pursued", ".", "They", "come", "in", "two", "flavours", ",", "(", "i", ")", "those", "which", "use", "input", "images", "at", "a", "few", "scales", "and", "corresponding", "deep", "feature", "extraction", "networks", ",", "and", "(", "ii", ")", "those", "which", "combine", "feature", "maps", "from", "different", "layers", "of", "a", "single", "deep", "architecture", ".", "The", "common", "idea", "is", "to", "use", "features", "extracted", "at", "multiple", "scales", "to", "provide", "both", "local", "and", "global", "context", "and", "the", "using", "feature", "maps", "of", "the", "early", "encoding", "layers", "retain", "more", "high", "frequency", "detail", "leading", "to", "sharper", "class", "boundaries", ".", "Some", "of", "these", "architectures", "are", "difficult", "to", "train", "due", "to", "their", "parameter", "size", ".", "Thus", "a", "multi", "-", "stage", "training", "process", "is", "employed", "along", "with", "data", "augmentation", ".", "The", "inference", "is", "also", "expensive", "with", "multiple", "convolutional", "pathways", "for", "feature", "extraction", ".", "Others", "append", "a", "CRF", "to", "their", "multi", "-", "scale", "network", "and", "jointly", "train", "them", ".", "However", ",", "these", "are", "not", "feed", "-", "forward", "at", "test", "time", "and", "require", "optimization", "to", "determine", "the", "MAP", "labels", ".", "Several", "of", "the", "recently", "proposed", "deep", "architectures", "for", "segmentation", "are", "not", "feed", "-", "forward", "in", "inference", "time", ",", ",", ".", "They", "require", "either", "MAP", "inference", "over", "a", "CRF", ",", "or", "aids", "such", "as", "region", "proposals", "for", "inference", ".", "We", "believe", "the", "perceived", "performance", "increase", "obtained", "by", "using", "a", "CRF", "is", "due", "to", "the", "lack", "of", "good", "decoding", "techniques", "in", "their", "core", "feed", "-", "forward", "segmentation", "engine", ".", "SegNet", "on", "the", "other", "hand", "uses", "decoders", "to", "obtain", "features", "for", "accurate", "pixel", "-", "wise", "classification", ".", "The", "recently", "proposed", "Deconvolutional", "Network", "and", "its", "semi", "-", "supervised", "variant", "the", "Decoupled", "network", "use", "the", "max", "locations", "of", "the", "encoder", "feature", "maps", "(", "pooling", "indices", ")", "to", "perform", "non", "-", "linear", "upsampling", "in", "the", "decoder", "network", ".", "The", "authors", "of", "these", "architectures", ",", "independently", "of", "SegNet", "(", "first", "submitted", "to", "CVPR", "2015", ")", ",", "proposed", "this", "idea", "of", "decoding", "in", "the", "decoder", "network", ".", "However", ",", "their", "encoder", "network", "consists", "of", "the", "fully", "connected", "layers", "from", "the", "VGG", "-", "16", "network", "which", "consists", "of", "about", "of", "the", "parameters", "of", "their", "entire", "network", ".", "This", "makes", "training", "of", "their", "network", "very", "difficult", "and", "thus", "require", "additional", "aids", "such", "as", "the", "use", "of", "region", "proposals", "to", "enable", "training", ".", "Moreover", ",", "during", "inference", "these", "proposals", "are", "used", "and", "this", "increases", "inference", "time", "significantly", ".", "From", "a", "benchmarking", "point", "of", "view", ",", "this", "also", "makes", "it", "difficult", "to", "evaluate", "the", "performance", "of", "their", "architecture", "(", "encoder", "-", "decoder", "network", ")", "without", "other", "aids", ".", "In", "this", "work", "we", "discard", "the", "fully", "connected", "layers", "of", "the", "VGG16", "encoder", "network", "which", "enables", "us", "to", "train", "the", "network", "using", "the", "relevant", "training", "set", "using", "SGD", "optimization", ".", "Another", "recent", "method", "shows", "the", "benefit", "of", "reducing", "the", "number", "of", "parameters", "significantly", "without", "sacrificing", "performance", ",", "reducing", "memory", "consumption", "and", "improving", "inference", "time", ".", "Our", "work", "was", "inspired", "by", "the", "unsupervised", "feature", "learning", "architecture", "proposed", "by", "Ranzato", "et", "al", ".", ".", "The", "key", "learning", "module", "is", "an", "encoder", "-", "decoder", "network", ".", "An", "encoder", "consists", "of", "convolution", "with", "a", "filter", "bank", ",", "element", "-", "wise", "tanh", "non", "-", "linearity", ",", "max", "-", "pooling", "and", "sub", "-", "sampling", "to", "obtain", "the", "feature", "maps", ".", "For", "each", "sample", ",", "the", "indices", "of", "the", "max", "locations", "computed", "during", "pooling", "are", "stored", "and", "passed", "to", "the", "decoder", ".", "The", "decoder", "upsamples", "the", "feature", "maps", "by", "using", "the", "stored", "pooled", "indices", ".", "It", "convolves", "this", "upsampled", "map", "using", "a", "trainable", "decoder", "filter", "bank", "to", "reconstruct", "the", "input", "image", ".", "This", "architecture", "was", "used", "for", "unsupervised", "pre", "-", "training", "for", "classification", ".", "A", "somewhat", "similar", "decoding", "technique", "is", "used", "for", "visualizing", "trained", "convolutional", "networks", "for", "classification", ".", "The", "architecture", "of", "Ranzato", "et", "al", ".", "mainly", "focused", "on", "layer", "-", "wise", "feature", "learning", "using", "small", "input", "patches", ".", "This", "was", "extended", "by", "Kavukcuoglu", "et", ".", "al", ".", "to", "accept", "full", "image", "sizes", "as", "input", "to", "learn", "hierarchical", "encoders", ".", "Both", "these", "approaches", "however", "did", "not", "attempt", "to", "use", "deep", "encoder", "-", "decoder", "networks", "for", "unsupervised", "feature", "training", "as", "they", "discarded", "the", "decoders", "after", "each", "encoder", "training", ".", "Here", ",", "SegNet", "differs", "from", "these", "architectures", "as", "the", "deep", "encoder", "-", "decoder", "network", "is", "trained", "jointly", "for", "a", "supervised", "learning", "task", "and", "hence", "the", "decoders", "are", "an", "integral", "part", "of", "the", "network", "in", "test", "time", ".", "Other", "applications", "where", "pixel", "wise", "predictions", "are", "made", "using", "deep", "networks", "are", "image", "super", "-", "resolution", "and", "depth", "map", "prediction", "from", "a", "single", "image", ".", "The", "authors", "in", "discuss", "the", "need", "for", "learning", "to", "upsample", "from", "low", "resolution", "feature", "maps", "which", "is", "the", "central", "topic", "of", "this", "paper", ".", "section", ":", "Architecture", "SegNet", "has", "an", "encoder", "network", "and", "a", "corresponding", "decoder", "network", ",", "followed", "by", "a", "final", "pixelwise", "classification", "layer", ".", "This", "architecture", "is", "illustrated", "in", "Fig", ".", "[", "reference", "]", ".", "The", "encoder", "network", "consists", "of", "convolutional", "layers", "which", "correspond", "to", "the", "first", "convolutional", "layers", "in", "the", "VGG16", "network", "designed", "for", "object", "classification", ".", "We", "can", "therefore", "initialize", "the", "training", "process", "from", "weights", "trained", "for", "classification", "on", "large", "datasets", ".", "We", "can", "also", "discard", "the", "fully", "connected", "layers", "in", "favour", "of", "retaining", "higher", "resolution", "feature", "maps", "at", "the", "deepest", "encoder", "output", ".", "This", "also", "reduces", "the", "number", "of", "parameters", "in", "the", "SegNet", "encoder", "network", "significantly", "(", "from", "134", "M", "to", "14.7", "M", ")", "as", "compared", "to", "other", "recent", "architectures", ",", "(", "see", ".", "Table", "[", "reference", "]", ")", ".", "Each", "encoder", "layer", "has", "a", "corresponding", "decoder", "layer", "and", "hence", "the", "decoder", "network", "has", "layers", ".", "The", "final", "decoder", "output", "is", "fed", "to", "a", "multi", "-", "class", "soft", "-", "max", "classifier", "to", "produce", "class", "probabilities", "for", "each", "pixel", "independently", ".", "Each", "encoder", "in", "the", "encoder", "network", "performs", "convolution", "with", "a", "filter", "bank", "to", "produce", "a", "set", "of", "feature", "maps", ".", "These", "are", "then", "batch", "normalized", ",", ")", ".", "Then", "an", "element", "-", "wise", "rectified", "-", "linear", "non", "-", "linearity", "(", "ReLU", ")", "is", "applied", ".", "Following", "that", ",", "max", "-", "pooling", "with", "a", "window", "and", "stride", "(", "non", "-", "overlapping", "window", ")", "is", "performed", "and", "the", "resulting", "output", "is", "sub", "-", "sampled", "by", "a", "factor", "of", ".", "Max", "-", "pooling", "is", "used", "to", "achieve", "translation", "invariance", "over", "small", "spatial", "shifts", "in", "the", "input", "image", ".", "Sub", "-", "sampling", "results", "in", "a", "large", "input", "image", "context", "(", "spatial", "window", ")", "for", "each", "pixel", "in", "the", "feature", "map", ".", "While", "several", "layers", "of", "max", "-", "pooling", "and", "sub", "-", "sampling", "can", "achieve", "more", "translation", "invariance", "for", "robust", "classification", "correspondingly", "there", "is", "a", "loss", "of", "spatial", "resolution", "of", "the", "feature", "maps", ".", "The", "increasingly", "lossy", "(", "boundary", "detail", ")", "image", "representation", "is", "not", "beneficial", "for", "segmentation", "where", "boundary", "delineation", "is", "vital", ".", "Therefore", ",", "it", "is", "necessary", "to", "capture", "and", "store", "boundary", "information", "in", "the", "encoder", "feature", "maps", "before", "sub", "-", "sampling", "is", "performed", ".", "If", "memory", "during", "inference", "is", "not", "constrained", ",", "then", "all", "the", "encoder", "feature", "maps", "(", "after", "sub", "-", "sampling", ")", "can", "be", "stored", ".", "This", "is", "usually", "not", "the", "case", "in", "practical", "applications", "and", "hence", "we", "propose", "a", "more", "efficient", "way", "to", "store", "this", "information", ".", "It", "involves", "storing", "only", "the", "max", "-", "pooling", "indices", ",", "i.e", ",", "the", "locations", "of", "the", "maximum", "feature", "value", "in", "each", "pooling", "window", "is", "memorized", "for", "each", "encoder", "feature", "map", ".", "In", "principle", ",", "this", "can", "be", "done", "using", "2", "bits", "for", "each", "pooling", "window", "and", "is", "thus", "much", "more", "efficient", "to", "store", "as", "compared", "to", "memorizing", "feature", "map", "(", "s", ")", "in", "float", "precision", ".", "As", "we", "show", "later", "in", "this", "work", ",", "this", "lower", "memory", "storage", "results", "in", "a", "slight", "loss", "of", "accuracy", "but", "is", "still", "suitable", "for", "practical", "applications", ".", "The", "appropriate", "decoder", "in", "the", "decoder", "network", "upsamples", "its", "input", "feature", "map", "(", "s", ")", "using", "the", "memorized", "max", "-", "pooling", "indices", "from", "the", "corresponding", "encoder", "feature", "map", "(", "s", ")", ".", "This", "step", "produces", "sparse", "feature", "map", "(", "s", ")", ".", "This", "SegNet", "decoding", "technique", "is", "illustrated", "in", "Fig", ".", "[", "reference", "]", ".", "These", "feature", "maps", "are", "then", "convolved", "with", "a", "trainable", "decoder", "filter", "bank", "to", "produce", "dense", "feature", "maps", ".", "A", "batch", "normalization", "step", "is", "then", "applied", "to", "each", "of", "these", "maps", ".", "Note", "that", "the", "decoder", "corresponding", "to", "the", "first", "encoder", "(", "closest", "to", "the", "input", "image", ")", "produces", "a", "multi", "-", "channel", "feature", "map", ",", "although", "its", "encoder", "input", "has", "3", "channels", "(", "RGB", ")", ".", "This", "is", "unlike", "the", "other", "decoders", "in", "the", "network", "which", "produce", "feature", "maps", "with", "the", "same", "number", "of", "size", "and", "channels", "as", "their", "encoder", "inputs", ".", "The", "high", "dimensional", "feature", "representation", "at", "the", "output", "of", "the", "final", "decoder", "is", "fed", "to", "a", "trainable", "soft", "-", "max", "classifier", ".", "This", "soft", "-", "max", "classifies", "each", "pixel", "independently", ".", "The", "output", "of", "the", "soft", "-", "max", "classifier", "is", "a", "K", "channel", "image", "of", "probabilities", "where", "K", "is", "the", "number", "of", "classes", ".", "The", "predicted", "segmentation", "corresponds", "to", "the", "class", "with", "maximum", "probability", "at", "each", "pixel", ".", "We", "add", "here", "that", "two", "other", "architectures", ",", "DeconvNet", "and", "U", "-", "Net", "share", "a", "similar", "architecture", "to", "SegNet", "but", "with", "some", "differences", ".", "DeconvNet", "has", "a", "much", "larger", "parameterization", ",", "needs", "more", "computational", "resources", "and", "is", "harder", "to", "train", "end", "-", "to", "-", "end", "(", "Table", "[", "reference", "]", ")", ",", "primarily", "due", "to", "the", "use", "of", "fully", "connected", "layers", "(", "albeit", "in", "a", "convolutional", "manner", ")", "We", "report", "several", "comparisons", "with", "DeconvNet", "later", "in", "the", "paper", "Sec", ".", "[", "reference", "]", ".", "As", "compared", "to", "SegNet", ",", "U", "-", "Net", "(", "proposed", "for", "the", "medical", "imaging", "community", ")", "does", "not", "reuse", "pooling", "indices", "but", "instead", "transfers", "the", "entire", "feature", "map", "(", "at", "the", "cost", "of", "more", "memory", ")", "to", "the", "corresponding", "decoders", "and", "concatenates", "them", "to", "upsampled", "(", "via", "deconvolution", ")", "decoder", "feature", "maps", ".", "There", "is", "no", "conv5", "and", "max", "-", "pool", "5", "block", "in", "U", "-", "Net", "as", "in", "the", "VGG", "net", "architecture", ".", "SegNet", ",", "on", "the", "other", "hand", ",", "uses", "all", "of", "the", "pre", "-", "trained", "convolutional", "layer", "weights", "from", "VGG", "net", "as", "pre", "-", "trained", "weights", ".", "subsection", ":", "Decoder", "Variants", "Many", "segmentation", "architectures", "share", "the", "same", "encoder", "network", "and", "they", "only", "vary", "in", "the", "form", "of", "their", "decoder", "network", ".", "Of", "these", "we", "choose", "to", "compare", "the", "SegNet", "decoding", "technique", "with", "the", "widely", "used", "Fully", "Convolutional", "Network", "(", "FCN", ")", "decoding", "technique", ".", "In", "order", "to", "analyse", "SegNet", "and", "compare", "its", "performance", "with", "FCN", "(", "decoder", "variants", ")", "we", "use", "a", "smaller", "version", "of", "SegNet", ",", "termed", "SegNet", "-", "Basic", ",", "which", "has", "4", "encoders", "and", "4", "decoders", ".", "All", "the", "encoders", "in", "SegNet", "-", "Basic", "perform", "max", "-", "pooling", "and", "sub", "-", "sampling", "and", "the", "corresponding", "decoders", "upsample", "its", "input", "using", "the", "received", "max", "-", "pooling", "indices", ".", "Batch", "normalization", "is", "used", "after", "each", "convolutional", "layer", "in", "both", "the", "encoder", "and", "decoder", "network", ".", "No", "biases", "are", "used", "after", "convolutions", "and", "no", "ReLU", "non", "-", "linearity", "is", "present", "in", "the", "decoder", "network", ".", "Further", ",", "a", "constant", "kernel", "size", "of", "over", "all", "the", "encoder", "and", "decoder", "layers", "is", "chosen", "to", "provide", "a", "wide", "context", "for", "smooth", "labelling", "i.e.", "a", "pixel", "in", "the", "deepest", "layer", "feature", "map", "(", "layer", ")", "can", "be", "traced", "back", "to", "a", "context", "window", "in", "the", "input", "image", "of", "pixels", ".", "This", "small", "size", "of", "SegNet", "-", "Basic", "allows", "us", "to", "explore", "many", "different", "variants", "(", "decoders", ")", "and", "train", "them", "in", "reasonable", "time", ".", "Similarly", "we", "create", "FCN", "-", "Basic", ",", "a", "comparable", "version", "of", "FCN", "for", "our", "analysis", "which", "shares", "the", "same", "encoder", "network", "as", "SegNet", "-", "Basic", "but", "with", "the", "FCN", "decoding", "technique", "(", "see", "Fig", ".", "[", "reference", "]", ")", "used", "in", "all", "its", "decoders", ".", "On", "the", "left", "in", "Fig", ".", "[", "reference", "]", "is", "the", "decoding", "technique", "used", "by", "SegNet", "(", "also", "SegNet", "-", "Basic", ")", ",", "where", "there", "is", "no", "learning", "involved", "in", "the", "upsampling", "step", ".", "However", ",", "the", "upsampled", "maps", "are", "convolved", "with", "trainable", "multi", "-", "channel", "decoder", "filters", "to", "densify", "its", "sparse", "inputs", ".", "Each", "decoder", "filter", "has", "the", "same", "number", "of", "channels", "as", "the", "number", "of", "upsampled", "feature", "maps", ".", "A", "smaller", "variant", "is", "one", "where", "the", "decoder", "filters", "are", "single", "channel", ",", "i.e", "they", "only", "convolve", "their", "corresponding", "upsampled", "feature", "map", ".", "This", "variant", "(", "SegNet", "-", "Basic", "-", "SingleChannelDecoder", ")", "reduces", "the", "number", "of", "trainable", "parameters", "and", "inference", "time", "significantly", ".", "On", "the", "right", "in", "Fig", ".", "[", "reference", "]", "is", "the", "FCN", "(", "also", "FCN", "-", "Basic", ")", "decoding", "technique", ".", "The", "important", "design", "element", "of", "the", "FCN", "model", "is", "dimensionality", "reduction", "step", "of", "the", "encoder", "feature", "maps", ".", "This", "compresses", "the", "encoder", "feature", "maps", "which", "are", "then", "used", "in", "the", "corresponding", "decoders", ".", "Dimensionality", "reduction", "of", "the", "encoder", "feature", "maps", ",", "say", "of", "64", "channels", ",", "is", "performed", "by", "convolving", "them", "with", "trainable", "filters", ",", "where", "is", "the", "number", "of", "classes", ".", "The", "compressed", "channel", "final", "encoder", "layer", "feature", "maps", "are", "the", "input", "to", "the", "decoder", "network", ".", "In", "a", "decoder", "of", "this", "network", ",", "upsampling", "is", "performed", "by", "inverse", "convolution", "using", "a", "fixed", "or", "trainable", "multi", "-", "channel", "upsampling", "kernel", ".", "We", "set", "the", "kernel", "size", "to", ".", "This", "manner", "of", "upsampling", "is", "also", "termed", "as", "deconvolution", ".", "Note", "that", ",", "in", "comparison", ",", "SegNet", "the", "multi", "-", "channel", "convolution", "using", "trainable", "decoder", "filters", "is", "performed", "after", "upsampling", "to", "densifying", "feature", "maps", ".", "The", "upsampled", "feature", "map", "in", "FCN", "has", "channels", ".", "It", "is", "then", "added", "element", "-", "wise", "to", "the", "corresponding", "resolution", "encoder", "feature", "map", "to", "produce", "the", "output", "decoder", "feature", "map", ".", "The", "upsampling", "kernels", "are", "initialized", "using", "bilinear", "interpolation", "weights", ".", "The", "FCN", "decoder", "model", "requires", "storing", "encoder", "feature", "maps", "during", "inference", ".", "This", "can", "be", "memory", "intensive", "for", "embedded", "applications", ";", "for", "e.g.", "storing", "64", "feature", "maps", "of", "the", "first", "layer", "of", "FCN", "-", "Basic", "at", "resolution", "in", "32", "bit", "floating", "point", "precision", "takes", "11", "MB", ".", "This", "can", "be", "made", "smaller", "using", "dimensionality", "reduction", "to", "the", "11", "feature", "maps", "which", "requires", "1.9", "MB", "storage", ".", "SegNet", "on", "the", "other", "hand", "requires", "almost", "negligible", "storage", "cost", "for", "the", "pooling", "indices", "(", "MB", "if", "stored", "using", "2", "bits", "per", "pooling", "window", ")", ".", "We", "can", "also", "create", "a", "variant", "of", "the", "FCN", "-", "Basic", "model", "which", "discards", "the", "encoder", "feature", "map", "addition", "step", "and", "only", "learns", "the", "upsampling", "kernels", "(", "FCN", "-", "Basic", "-", "NoAddition", ")", ".", "In", "addition", "to", "the", "above", "variants", ",", "we", "study", "upsampling", "using", "fixed", "bilinear", "interpolation", "weights", "which", "therefore", "requires", "no", "learning", "for", "upsampling", "(", "Bilinear", "-", "Interpolation", ")", ".", "At", "the", "other", "extreme", ",", "we", "can", "add", "64", "encoder", "feature", "maps", "at", "each", "layer", "to", "the", "corresponding", "output", "feature", "maps", "from", "the", "SegNet", "decoder", "to", "create", "a", "more", "memory", "intensive", "variant", "of", "SegNet", "(", "SegNet", "-", "Basic", "-", "EncoderAddition", ")", ".", "Here", "both", "the", "pooling", "indices", "for", "upsampling", "are", "used", ",", "followed", "by", "a", "convolution", "step", "to", "densify", "its", "sparse", "input", ".", "This", "is", "then", "added", "element", "-", "wise", "to", "the", "corresponding", "encoder", "feature", "maps", "to", "produce", "a", "decoders", "output", ".", "Another", "and", "more", "memory", "intensive", "FCN", "-", "Basic", "variant", "(", "FCN", "-", "Basic", "-", "NoDimReduction", ")", "is", "where", "there", "is", "no", "dimensionality", "reduction", "performed", "for", "the", "encoder", "feature", "maps", ".", "This", "implies", "that", "unlike", "FCN", "-", "Basic", "the", "final", "encoder", "feature", "map", "is", "not", "compressed", "to", "channels", "before", "passing", "it", "to", "the", "decoder", "network", ".", "Therefore", ",", "the", "number", "of", "channels", "at", "the", "end", "of", "each", "decoder", "is", "the", "same", "as", "the", "corresponding", "encoder", "(", "i.e", ")", ".", "We", "also", "tried", "other", "generic", "variants", "where", "feature", "maps", "are", "simply", "upsampled", "by", "replication", ",", "or", "by", "using", "a", "fixed", "(", "and", "sparse", ")", "array", "of", "indices", "for", "upsampling", ".", "These", "performed", "quite", "poorly", "in", "comparison", "to", "the", "above", "variants", ".", "A", "variant", "without", "max", "-", "pooling", "and", "sub", "-", "sampling", "in", "the", "encoder", "network", "(", "decoders", "are", "redundant", ")", "consumes", "more", "memory", ",", "takes", "longer", "to", "converge", "and", "performs", "poorly", ".", "Finally", ",", "please", "note", "that", "to", "encourage", "reproduction", "of", "our", "results", "we", "release", "the", "Caffe", "implementation", "of", "all", "the", "variants", ".", "subsection", ":", "Training", "We", "use", "the", "CamVid", "road", "scenes", "dataset", "to", "benchmark", "the", "performance", "of", "the", "decoder", "variants", ".", "This", "dataset", "is", "small", ",", "consisting", "of", "367", "training", "and", "233", "testing", "RGB", "images", "(", "day", "and", "dusk", "scenes", ")", "at", "resolution", ".", "The", "challenge", "is", "to", "segment", "classes", "such", "as", "road", ",", "building", ",", "cars", ",", "pedestrians", ",", "signs", ",", "poles", ",", "side", "-", "walk", "etc", ".", "We", "perform", "local", "contrast", "normalization", "to", "the", "RGB", "input", ".", "The", "encoder", "and", "decoder", "weights", "were", "all", "initialized", "using", "the", "technique", "described", "in", "He", "et", "al", ".", ".", "To", "train", "all", "the", "variants", "we", "use", "stochastic", "gradient", "descent", "(", "SGD", ")", "with", "a", "fixed", "learning", "rate", "of", "0.1", "and", "momentum", "of", "0.9", "using", "our", "Caffe", "implementation", "of", "SegNet", "-", "Basic", ".", "We", "train", "the", "variants", "until", "the", "training", "loss", "converges", ".", "Before", "each", "epoch", ",", "the", "training", "set", "is", "shuffled", "and", "each", "mini", "-", "batch", "(", "12", "images", ")", "is", "then", "picked", "in", "order", "thus", "ensuring", "that", "each", "image", "is", "used", "only", "once", "in", "an", "epoch", ".", "We", "select", "the", "model", "which", "performs", "highest", "on", "a", "validation", "dataset", ".", "We", "use", "the", "cross", "-", "entropy", "loss", "as", "the", "objective", "function", "for", "training", "the", "network", ".", "The", "loss", "is", "summed", "up", "over", "all", "the", "pixels", "in", "a", "mini", "-", "batch", ".", "When", "there", "is", "large", "variation", "in", "the", "number", "of", "pixels", "in", "each", "class", "in", "the", "training", "set", "(", "e.g", "road", ",", "sky", "and", "building", "pixels", "dominate", "the", "CamVid", "dataset", ")", "then", "there", "is", "a", "need", "to", "weight", "the", "loss", "differently", "based", "on", "the", "true", "class", ".", "This", "is", "termed", "class", "balancing", ".", "We", "use", "median", "frequency", "balancing", "where", "the", "weight", "assigned", "to", "a", "class", "in", "the", "loss", "function", "is", "the", "ratio", "of", "the", "median", "of", "class", "frequencies", "computed", "on", "the", "entire", "training", "set", "divided", "by", "the", "class", "frequency", ".", "This", "implies", "that", "larger", "classes", "in", "the", "training", "set", "have", "a", "weight", "smaller", "than", "and", "the", "weights", "of", "the", "smallest", "classes", "are", "the", "highest", ".", "We", "also", "experimented", "with", "training", "the", "different", "variants", "without", "class", "balancing", "or", "equivalently", "using", "natural", "frequency", "balancing", ".", "subsection", ":", "Analysis", "To", "compare", "the", "quantitative", "performance", "of", "the", "different", "decoder", "variants", ",", "we", "use", "three", "commonly", "used", "performance", "measures", ":", "global", "accuracy", "(", "G", ")", "which", "measures", "the", "percentage", "of", "pixels", "correctly", "classified", "in", "the", "dataset", ",", "class", "average", "accuracy", "(", "C", ")", "is", "the", "mean", "of", "the", "predictive", "accuracy", "over", "all", "classes", "and", "mean", "intersection", "over", "union", "(", "mIoU", ")", "over", "all", "classes", "as", "used", "in", "the", "Pascal", "VOC12", "challenge", ".", "The", "mIoU", "metric", "is", "a", "more", "stringent", "metric", "than", "class", "average", "accuracy", "since", "it", "penalizes", "false", "positive", "predictions", ".", "However", ",", "mIoU", "metric", "is", "not", "optimized", "for", "directly", "through", "the", "class", "balanced", "cross", "-", "entropy", "loss", ".", "The", "mIoU", "metric", "otherwise", "known", "as", "the", "Jacard", "Index", "is", "most", "commonly", "used", "in", "benchmarking", ".", "However", ",", "Csurka", "et", "al", ".", "note", "that", "this", "metric", "does", "not", "always", "correspond", "to", "human", "qualitative", "judgements", "(", "ranks", ")", "of", "good", "quality", "segmentation", ".", "They", "show", "with", "examples", "that", "mIoU", "favours", "region", "smoothness", "and", "does", "not", "evaluate", "boundary", "accuracy", ",", "a", "point", "also", "alluded", "to", "recently", "by", "the", "authors", "of", "FCN", ".", "Hence", "they", "propose", "to", "complement", "the", "mIoU", "metric", "with", "a", "boundary", "measure", "based", "on", "the", "Berkeley", "contour", "matching", "score", "commonly", "used", "to", "evaluate", "unsupervised", "image", "segmentation", "quality", ".", "Csurka", "et", "al", ".", "simply", "extend", "this", "to", "semantic", "segmentation", "and", "show", "that", "the", "measure", "of", "semantic", "contour", "accuracy", "used", "in", "conjunction", "with", "the", "mIoU", "metric", "agrees", "more", "with", "human", "ranking", "of", "segmentation", "outputs", ".", "The", "key", "idea", "in", "computing", "a", "semantic", "contour", "score", "is", "to", "evaluate", "the", "F1", "-", "measure", "which", "involves", "computing", "the", "precision", "and", "recall", "values", "between", "the", "predicted", "and", "ground", "truth", "class", "boundary", "given", "a", "pixel", "tolerance", "distance", ".", "We", "used", "a", "value", "of", "of", "the", "image", "diagonal", "as", "the", "tolerance", "distance", ".", "The", "F1", "-", "measure", "for", "each", "class", "that", "is", "present", "in", "the", "ground", "truth", "test", "image", "is", "averaged", "to", "produce", "an", "image", "F1", "-", "measure", ".", "Then", "we", "compute", "the", "whole", "test", "set", "average", ",", "denoted", "the", "boundary", "F1", "-", "measure", "(", "BF", ")", "by", "average", "the", "image", "F1", "measures", ".", "We", "test", "each", "architectural", "variant", "after", "each", "iterations", "of", "optimization", "on", "the", "CamVid", "validation", "set", "until", "the", "training", "loss", "converges", ".", "With", "a", "training", "mini", "-", "batch", "size", "of", "12", "this", "corresponds", "to", "testing", "approximately", "every", "33", "epochs", "(", "passes", ")", "through", "the", "training", "set", ".", "We", "select", "the", "iteration", "wherein", "the", "global", "accuracy", "is", "highest", "amongst", "the", "evaluations", "on", "the", "validation", "set", ".", "We", "report", "all", "the", "three", "measures", "of", "performance", "at", "this", "point", "on", "the", "held", "-", "out", "CamVid", "test", "set", ".", "Although", "we", "use", "class", "balancing", "while", "training", "the", "variants", ",", "it", "is", "still", "important", "to", "achieve", "high", "global", "accuracy", "to", "result", "in", "an", "overall", "smooth", "segmentation", ".", "Another", "reason", "is", "that", "the", "contribution", "of", "segmentation", "towards", "autonomous", "driving", "is", "mainly", "for", "delineating", "classes", "such", "as", "roads", ",", "buildings", ",", "side", "-", "walk", ",", "sky", ".", "These", "classes", "dominate", "the", "majority", "of", "the", "pixels", "in", "an", "image", "and", "a", "high", "global", "accuracy", "corresponds", "to", "good", "segmentation", "of", "these", "important", "classes", ".", "We", "also", "observed", "that", "reporting", "the", "numerical", "performance", "when", "class", "average", "is", "highest", "can", "often", "correspond", "to", "low", "global", "accuracy", "indicating", "a", "perceptually", "noisy", "segmentation", "output", ".", "In", "Table", "[", "reference", "]", "we", "report", "the", "numerical", "results", "of", "our", "analysis", ".", "We", "also", "show", "the", "size", "of", "the", "trainable", "parameters", "and", "the", "highest", "resolution", "feature", "map", "or", "pooling", "indices", "storage", "memory", ",", "i.e", ",", "of", "the", "first", "layer", "feature", "maps", "after", "max", "-", "pooling", "and", "sub", "-", "sampling", ".", "We", "show", "the", "average", "time", "for", "one", "forward", "pass", "with", "our", "Caffe", "implementation", ",", "averaged", "over", "measurements", "using", "a", "input", "on", "an", "NVIDIA", "Titan", "GPU", "with", "cuDNN", "v3", "acceleration", ".", "We", "note", "that", "the", "upsampling", "layers", "in", "the", "SegNet", "variants", "are", "not", "optimised", "using", "cuDNN", "acceleration", ".", "We", "show", "the", "results", "for", "both", "testing", "and", "training", "for", "all", "the", "variants", "at", "the", "selected", "iteration", ".", "The", "results", "are", "also", "tabulated", "without", "class", "balancing", "(", "natural", "frequency", ")", "for", "training", "and", "testing", "accuracies", ".", "Below", "we", "analyse", "the", "results", "with", "class", "balancing", ".", "From", "the", "Table", "[", "reference", "]", ",", "we", "see", "that", "bilinear", "interpolation", "based", "upsampling", "without", "any", "learning", "performs", "the", "worst", "based", "on", "all", "the", "measures", "of", "accuracy", ".", "All", "the", "other", "methods", "which", "either", "use", "learning", "for", "upsampling", "(", "FCN", "-", "Basic", "and", "variants", ")", "or", "learning", "decoder", "filters", "after", "upsampling", "(", "SegNet", "-", "Basic", "and", "its", "variants", ")", "perform", "significantly", "better", ".", "This", "emphasizes", "the", "need", "to", "learn", "decoders", "for", "segmentation", ".", "This", "is", "also", "supported", "by", "experimental", "evidence", "gathered", "by", "other", "authors", "when", "comparing", "FCN", "with", "SegNet", "-", "type", "decoding", "techniques", ".", "When", "we", "compare", "SegNet", "-", "Basic", "and", "FCN", "-", "Basic", "we", "see", "that", "both", "perform", "equally", "well", "on", "this", "test", "over", "all", "the", "measures", "of", "accuracy", ".", "The", "difference", "is", "that", "SegNet", "uses", "less", "memory", "during", "inference", "since", "it", "only", "stores", "max", "-", "pooling", "indices", ".", "On", "the", "other", "hand", "FCN", "-", "Basic", "stores", "encoder", "feature", "maps", "in", "full", "which", "consumes", "much", "more", "memory", "(", "11", "times", "more", ")", ".", "SegNet", "-", "Basic", "has", "a", "decoder", "with", "64", "feature", "maps", "in", "each", "decoder", "layer", ".", "In", "comparison", "FCN", "-", "Basic", ",", "which", "uses", "dimensionality", "reduction", ",", "has", "fewer", "(", "11", ")", "feature", "maps", "in", "each", "decoder", "layer", ".", "This", "reduces", "the", "number", "of", "convolutions", "in", "the", "decoder", "network", "and", "hence", "FCN", "-", "Basic", "is", "faster", "during", "inference", "(", "forward", "pass", ")", ".", "From", "another", "perspective", ",", "the", "decoder", "network", "in", "SegNet", "-", "Basic", "makes", "it", "overall", "a", "larger", "network", "than", "FCN", "-", "Basic", ".", "This", "endows", "it", "with", "more", "flexibility", "and", "hence", "achieves", "higher", "training", "accuracy", "than", "FCN", "-", "Basic", "for", "the", "same", "number", "of", "iterations", ".", "Overall", "we", "see", "that", "SegNet", "-", "Basic", "has", "an", "advantage", "over", "FCN", "-", "Basic", "when", "inference", "time", "memory", "is", "constrained", "but", "where", "inference", "time", "can", "be", "compromised", "to", "some", "extent", ".", "SegNet", "-", "Basic", "is", "most", "similar", "to", "FCN", "-", "Basic", "-", "NoAddition", "in", "terms", "of", "their", "decoders", ",", "although", "the", "decoder", "of", "SegNet", "is", "larger", ".", "Both", "learn", "to", "produce", "dense", "feature", "maps", ",", "either", "directly", "by", "learning", "to", "perform", "deconvolution", "as", "in", "FCN", "-", "Basic", "-", "NoAddition", "or", "by", "first", "upsampling", "and", "then", "convolving", "with", "trained", "decoder", "filters", ".", "The", "performance", "of", "SegNet", "-", "Basic", "is", "superior", ",", "in", "part", "due", "to", "its", "larger", "decoder", "size", ".", "The", "accuracy", "of", "FCN", "-", "Basic", "-", "NoAddition", "is", "also", "lower", "as", "compared", "to", "FCN", "-", "Basic", ".", "This", "shows", "that", "it", "is", "vital", "to", "capture", "the", "information", "present", "in", "the", "encoder", "feature", "maps", "for", "better", "performance", ".", "In", "particular", ",", "note", "the", "large", "drop", "in", "the", "BF", "measure", "between", "these", "two", "variants", ".", "This", "can", "also", "explain", "the", "part", "of", "the", "reason", "why", "SegNet", "-", "Basic", "outperforms", "FCN", "-", "Basic", "-", "NoAddition", ".", "The", "size", "of", "the", "FCN", "-", "Basic", "-", "NoAddition", "-", "NoDimReduction", "model", "is", "slightly", "larger", "than", "SegNet", "-", "Basic", "since", "the", "final", "encoder", "feature", "maps", "are", "not", "compressed", "to", "match", "the", "number", "of", "classes", ".", "This", "makes", "it", "a", "fair", "comparison", "in", "terms", "of", "the", "size", "of", "the", "model", ".", "The", "performance", "of", "this", "FCN", "variant", "is", "poorer", "than", "SegNet", "-", "Basic", "in", "test", "but", "also", "its", "training", "accuracy", "is", "lower", "for", "the", "same", "number", "of", "training", "epochs", ".", "This", "shows", "that", "using", "a", "larger", "decoder", "is", "not", "enough", "but", "it", "is", "also", "important", "to", "capture", "encoder", "feature", "map", "information", "to", "learn", "better", ",", "particular", "the", "fine", "grained", "contour", "information", "(", "notice", "the", "drop", "in", "the", "BF", "measure", ")", ".", "Here", "it", "is", "also", "interesting", "to", "see", "that", "SegNet", "-", "Basic", "has", "a", "competitive", "training", "accuracy", "when", "compared", "to", "larger", "models", "such", "FCN", "-", "Basic", "-", "NoDimReduction", ".", "Another", "interesting", "comparison", "between", "FCN", "-", "Basic", "-", "NoAddition", "and", "SegNet", "-", "Basic", "-", "SingleChannelDecoder", "shows", "that", "using", "max", "-", "pooling", "indices", "for", "upsampling", "and", "an", "overall", "larger", "decoder", "leads", "to", "better", "performance", ".", "This", "also", "lends", "evidence", "to", "SegNet", "being", "a", "good", "architecture", "for", "segmentation", ",", "particularly", "when", "there", "is", "a", "need", "to", "find", "a", "compromise", "between", "storage", "cost", ",", "accuracy", "versus", "inference", "time", ".", "In", "the", "best", "case", ",", "when", "both", "memory", "and", "inference", "time", "is", "not", "constrained", ",", "larger", "models", "such", "as", "FCN", "-", "Basic", "-", "NoDimReduction", "and", "SegNet", "-", "EncoderAddition", "are", "both", "more", "accurate", "than", "the", "other", "variants", ".", "Particularly", ",", "discarding", "dimensionality", "reduction", "in", "the", "FCN", "-", "Basic", "model", "leads", "to", "the", "best", "performance", "amongst", "the", "FCN", "-", "Basic", "variants", "with", "a", "high", "BF", "score", ".", "This", "once", "again", "emphasizes", "the", "trade", "-", "off", "involved", "between", "memory", "and", "accuracy", "in", "segmentation", "architectures", ".", "The", "last", "two", "columns", "of", "Table", "[", "reference", "]", "show", "the", "result", "when", "no", "class", "balancing", "is", "used", "(", "natural", "frequency", ")", ".", "Here", ",", "we", "can", "observe", "that", "without", "weighting", "the", "results", "are", "poorer", "for", "all", "the", "variants", ",", "particularly", "for", "class", "average", "accuracy", "and", "mIoU", "metric", ".", "The", "global", "accuracy", "is", "the", "highest", "without", "weighting", "since", "the", "majority", "of", "the", "scene", "is", "dominated", "by", "sky", ",", "road", "and", "building", "pixels", ".", "Apart", "from", "this", "all", "the", "inference", "from", "the", "comparative", "analysis", "of", "variants", "holds", "true", "for", "natural", "frequency", "balancing", "too", ",", "including", "the", "trends", "for", "the", "BF", "measure", ".", "SegNet", "-", "Basic", "performs", "as", "well", "as", "FCN", "-", "Basic", "and", "is", "better", "than", "the", "larger", "FCN", "-", "Basic", "-", "NoAddition", "-", "NoDimReduction", ".", "The", "bigger", "but", "less", "efficient", "models", "FCN", "-", "Basic", "-", "NoDimReduction", "and", "SegNet", "-", "EncoderAddition", "perform", "better", "than", "the", "other", "variants", ".", "We", "can", "now", "summarize", "the", "above", "analysis", "with", "the", "following", "general", "points", ".", "The", "best", "performance", "is", "achieved", "when", "encoder", "feature", "maps", "are", "stored", "in", "full", ".", "This", "is", "reflected", "in", "the", "semantic", "contour", "delineation", "metric", "(", "BF", ")", "most", "clearly", ".", "When", "memory", "during", "inference", "is", "constrained", ",", "then", "compressed", "forms", "of", "encoder", "feature", "maps", "(", "dimensionality", "reduction", ",", "max", "-", "pooling", "indices", ")", "can", "be", "stored", "and", "used", "with", "an", "appropriate", "decoder", "(", "e.g.", "SegNet", "type", ")", "to", "improve", "performance", ".", "Larger", "decoders", "increase", "performance", "for", "a", "given", "encoder", "network", ".", "section", ":", "Benchmarking", "We", "quantify", "the", "performance", "of", "SegNet", "on", "two", "scene", "segmentation", "benchmarks", "using", "our", "Caffe", "implementation", ".", "The", "first", "task", "is", "road", "scene", "segmentation", "which", "is", "of", "current", "practical", "interest", "for", "various", "autonomous", "driving", "related", "problems", ".", "The", "second", "task", "is", "indoor", "scene", "segmentation", "which", "is", "of", "immediate", "interest", "to", "several", "augmented", "reality", "(", "AR", ")", "applications", ".", "The", "input", "RGB", "images", "for", "both", "tasks", "were", ".", "We", "benchmarked", "SegNet", "against", "several", "other", "well", "adopted", "deep", "architectures", "for", "segmentation", "such", "as", "FCN", ",", "DeepLab", "-", "LargFOV", "and", "DeconvNet", ".", "Our", "objective", "was", "to", "understand", "the", "performance", "of", "these", "architectures", "when", "trained", "end", "-", "to", "-", "end", "on", "the", "same", "datasets", ".", "To", "enable", "end", "-", "to", "-", "end", "training", "we", "added", "batch", "normalization", "layers", "after", "each", "convolutional", "layer", ".", "For", "DeepLab", "-", "LargeFOV", ",", "we", "changed", "the", "max", "pooling", "3", "stride", "to", "1", "to", "achieve", "a", "final", "predictive", "resolution", "of", ".", "We", "restricted", "the", "feature", "size", "in", "the", "fully", "connnected", "layers", "of", "DeconvNet", "to", "so", "as", "to", "enable", "training", "with", "the", "same", "batch", "size", "as", "other", "models", ".", "Here", "note", "that", "the", "authors", "of", "DeepLab", "-", "LargeFOV", "have", "also", "reported", "little", "loss", "in", "performance", "by", "reducing", "the", "size", "of", "the", "fully", "connected", "layers", ".", "In", "order", "to", "perform", "a", "controlled", "benchmark", "we", "used", "the", "same", "SGD", "solver", "with", "a", "fixed", "learning", "rate", "of", "and", "momentum", "of", ".", "The", "optimization", "was", "performed", "for", "more", "than", "100", "epochs", "through", "the", "dataset", "until", "no", "further", "performance", "increase", "was", "observed", ".", "Dropout", "of", "was", "added", "to", "the", "end", "of", "deeper", "convolutional", "layers", "in", "all", "models", "to", "prevent", "overfitting", "(", "see", "for", "example", "caffe", "prototxt", ")", ".", "For", "the", "road", "scenes", "which", "have", "classes", "we", "used", "a", "mini", "-", "batch", "size", "of", "and", "for", "indoor", "scenes", "with", "classes", "we", "used", "a", "mini", "-", "batch", "size", "of", ".", "subsection", ":", "Road", "Scene", "Segmentation", "A", "number", "of", "road", "scene", "datasets", "are", "available", "for", "semantic", "parsing", ".", "Of", "these", "we", "choose", "to", "benchmark", "SegNet", "using", "the", "CamVid", "dataset", "as", "it", "contains", "video", "sequences", ".", "This", "enables", "us", "to", "compare", "our", "proposed", "architecture", "with", "those", "which", "use", "motion", "and", "structure", "and", "video", "segments", ".", "We", "also", "combine", "to", "form", "an", "ensemble", "of", "3433", "images", "to", "train", "SegNet", "for", "an", "additional", "benchmark", ".", "For", "a", "web", "demo", "(", "see", "footnote", "[", "reference", "]", ")", "of", "road", "scene", "segmentation", ",", "we", "include", "the", "CamVid", "test", "set", "to", "this", "larger", "dataset", ".", "Here", ",", "we", "would", "like", "to", "note", "that", "another", "recent", "and", "independent", "segmentation", "benchmark", "on", "road", "scenes", "has", "been", "performed", "for", "SegNet", "and", "the", "other", "competing", "architectures", "used", "in", "this", "paper", ".", "However", ",", "the", "benchmark", "was", "not", "controlled", ",", "meaning", "that", "each", "architecture", "was", "trained", "with", "a", "separate", "recipe", "with", "varying", "input", "resolutions", "and", "sometimes", "with", "a", "validation", "set", "included", ".", "Therefore", ",", "we", "believe", "our", "more", "controlled", "benchmark", "can", "be", "used", "to", "complement", "their", "efforts", ".", "Building", "Tree", "Sky", "Car", "Sign", "-", "Symbol", "Road", "Pedestrian", "Fence", "Column", "-", "Pole", "Side", "-", "walk", "Bicyclist", "Class", "avg", ".", "Global", "avg", ".", "mIoU", "BF", "The", "qualitative", "comparisons", "of", "SegNet", "predictions", "with", "other", "deep", "architectures", "can", "be", "seen", "in", "Fig", ".", "[", "reference", "]", ".", "The", "qualitative", "results", "show", "the", "ability", "of", "the", "proposed", "architecture", "to", "segment", "smaller", "classes", "in", "road", "scenes", "while", "producing", "a", "smooth", "segmentation", "of", "the", "overall", "scene", ".", "Indeed", ",", "under", "the", "controlled", "benchmark", "setting", ",", "SegNet", "shows", "superior", "performance", "as", "compared", "to", "some", "of", "the", "larger", "models", ".", "DeepLab", "-", "LargeFOV", "is", "the", "most", "efficient", "model", "and", "with", "CRF", "post", "-", "processing", "can", "produce", "competitive", "results", "although", "smaller", "classes", "are", "lost", ".", "FCN", "with", "learnt", "deconvolution", "is", "clearly", "better", "than", "with", "fixed", "bilinear", "upsampling", ".", "DeconvNet", "is", "the", "largest", "model", "and", "the", "most", "inefficient", "to", "train", ".", "Its", "predictions", "do", "not", "retain", "small", "classes", ".", "We", "also", "use", "this", "benchmark", "to", "first", "compare", "SegNet", "with", "several", "non", "deep", "-", "learning", "methods", "including", "Random", "Forests", ",", "Boosting", "in", "combination", "with", "CRF", "based", "methods", ".", "This", "was", "done", "to", "give", "the", "user", "a", "perspective", "of", "the", "improvements", "in", "accuracy", "that", "has", "been", "achieved", "using", "deep", "networks", "compared", "to", "classical", "feature", "engineering", "based", "techniques", ".", "The", "results", "in", "Table", "[", "reference", "]", "show", "SegNet", "-", "Basic", ",", "SegNet", "obtain", "competitive", "results", "when", "compared", "with", "methods", "which", "use", "CRFs", ".", "This", "shows", "the", "ability", "of", "the", "deep", "architecture", "to", "extract", "meaningful", "features", "from", "the", "input", "image", "and", "map", "it", "to", "accurate", "and", "smooth", "class", "segment", "labels", ".", "The", "most", "interesting", "result", "here", "is", "the", "large", "performance", "improvement", "in", "class", "average", "and", "mIOU", "metrics", "that", "is", "obtained", "when", "a", "large", "training", "dataset", ",", "obtained", "by", "combining", ",", "is", "used", "to", "train", "SegNet", ".", "Correspondingly", ",", "the", "qualitative", "results", "of", "SegNet", "(", "see", "Fig", ".", "[", "reference", "]", ")", "are", "clearly", "superior", "to", "the", "rest", "of", "the", "methods", ".", "It", "is", "able", "to", "segment", "both", "small", "and", "large", "classes", "well", ".", "We", "remark", "here", "that", "we", "used", "median", "frequency", "class", "balancing", "in", "training", "SegNet", "-", "Basic", "and", "SegNet", ".", "In", "addition", ",", "there", "is", "an", "overall", "smooth", "quality", "of", "segmentation", "much", "like", "what", "is", "typically", "obtained", "with", "CRF", "post", "-", "processing", ".", "Although", "the", "fact", "that", "results", "improve", "with", "larger", "training", "sets", "is", "not", "surprising", ",", "the", "percentage", "improvement", "obtained", "using", "pre", "-", "trained", "encoder", "network", "and", "this", "training", "set", "indicates", "that", "this", "architecture", "can", "potentially", "be", "deployed", "for", "practical", "applications", ".", "Our", "random", "testing", "on", "urban", "and", "highway", "images", "from", "the", "internet", "(", "see", "Fig", ".", "[", "reference", "]", ")", "demonstrates", "that", "SegNet", "can", "absorb", "a", "large", "training", "set", "and", "generalize", "well", "to", "unseen", "images", ".", "It", "also", "indicates", "the", "contribution", "of", "the", "prior", "(", "CRF", ")", "can", "be", "lessened", "when", "sufficient", "amount", "of", "training", "data", "is", "made", "available", ".", "In", "Table", "[", "reference", "]", "we", "compare", "SegNet", "’s", "performance", "with", "now", "widely", "adopted", "fully", "convolutional", "architectures", "for", "segmentation", ".", "As", "compared", "to", "the", "experiment", "in", "Table", "[", "reference", "]", ",", "we", "did", "not", "use", "any", "class", "blancing", "for", "training", "any", "of", "the", "deep", "architectures", "including", "SegNet", ".", "This", "is", "because", "we", "found", "it", "difficult", "to", "train", "larger", "models", "such", "as", "DeconvNet", "with", "median", "frequency", "balancing", ".", "We", "benchmark", "performance", "at", "40", "K", ",", "80", "K", "and", "80", "K", "iterations", "which", "given", "the", "mini", "-", "batch", "size", "and", "training", "set", "size", "approximately", "corresponds", "to", "and", "100", "epochs", ".", "For", "the", "last", "test", "point", "we", "also", "report", "the", "maximum", "number", "of", "iterations", "(", "here", "atleast", "150", "epochs", ")", "beyond", "which", "we", "observed", "no", "accuracy", "improvements", "or", "when", "over", "-", "fitting", "set", "in", ".", "We", "report", "the", "metrics", "at", "three", "stages", "in", "the", "training", "phase", "to", "reveal", "how", "the", "metrics", "varied", "with", "training", "time", ",", "particularly", "for", "larger", "networks", ".", "This", "is", "important", "to", "understand", "if", "additional", "training", "time", "is", "justified", "when", "set", "against", "accuracy", "increases", ".", "Note", "also", "that", "for", "each", "evaluation", "we", "performed", "a", "complete", "run", "through", "the", "dataset", "to", "obtain", "batch", "norm", "statistics", "and", "then", "evaluated", "the", "test", "model", "with", "this", "statistic", "(", "see", "for", "code", ".", ")", ".", "These", "evaluations", "are", "expensive", "to", "perform", "on", "large", "training", "sets", "and", "hence", "we", "only", "report", "metrics", "at", "three", "time", "points", "in", "the", "training", "phase", ".", "From", "Table", "[", "reference", "]", "we", "immediately", "see", "that", "SegNet", ",", "DeconvNet", "achieve", "the", "highest", "scores", "in", "all", "the", "metrics", "as", "compared", "to", "other", "models", ".", "DeconvNet", "has", "a", "higher", "boundary", "delineation", "accuracy", "but", "SegNet", "is", "much", "more", "efficient", "as", "compared", "to", "DeconvNet", ".", "This", "can", "be", "seen", "from", "the", "compute", "statistics", "in", "Table", "[", "reference", "]", ".", "FCN", ",", "DeconvNet", "which", "have", "fully", "connected", "layers", "(", "turned", "into", "convolutional", "layers", ")", "train", "much", "more", "slowly", "and", "have", "comparable", "or", "higher", "forward", "-", "backward", "pass", "time", "with", "reference", "to", "SegNet", ".", "Here", "we", "note", "also", "that", "over", "-", "fitting", "was", "not", "an", "issue", "in", "training", "these", "larger", "models", ",", "since", "at", "comparable", "iterations", "to", "SegNet", "their", "metrics", "showed", "an", "increasing", "trend", ".", "For", "the", "FCN", "model", "learning", "the", "deconvolutional", "layers", "as", "opposed", "to", "fixing", "them", "with", "bi", "-", "linear", "interpolation", "weights", "improves", "performance", "particularly", "the", "BF", "score", ".", "It", "also", "achieves", "higher", "metrics", "in", "a", "far", "lesser", "time", ".", "This", "fact", "agrees", "with", "our", "earlier", "analysis", "in", "Sec", ".", "[", "reference", "]", ".", "Surprisingly", ",", "DeepLab", "-", "LargeFOV", "which", "is", "trained", "to", "predict", "labels", "at", "a", "resolution", "of", "produces", "competitive", "performance", "given", "that", "it", "is", "the", "smallest", "model", "in", "terms", "of", "parameterization", "and", "also", "has", "the", "fastest", "training", "time", "as", "per", "Table", "[", "reference", "]", ".", "However", ",", "the", "boundary", "accuracy", "is", "poorer", "and", "this", "is", "shared", "by", "the", "other", "architectures", ".", "DeconvNet", "’s", "BF", "score", "is", "higher", "than", "the", "other", "networks", "when", "trained", "for", "a", "very", "long", "time", ".", "Given", "our", "analysis", "in", "Sec", ".", "[", "reference", "]", "and", "the", "fact", "that", "it", "shares", "a", "SegNet", "type", "architecture", ".", "The", "impact", "of", "dense", "CRF", "post", "-", "processing", "can", "be", "seen", "in", "the", "last", "time", "point", "for", "DeepLab", "-", "LargeFOV", "-", "denseCRF", ".", "Both", "global", "and", "mIoU", "improve", "but", "class", "average", "diminshes", ".", "However", "a", "large", "improvement", "is", "obtained", "for", "the", "BF", "score", ".", "Note", "here", "that", "the", "dense", "CRF", "hyperparameters", "were", "obtained", "by", "an", "expensive", "grid", "-", "search", "process", "on", "a", "subset", "of", "the", "training", "set", "since", "no", "validation", "set", "was", "available", ".", "subsection", ":", "SUN", "RGB", "-", "D", "Indoor", "Scenes", "SUN", "RGB", "-", "D", "is", "a", "very", "challenging", "and", "large", "dataset", "of", "indoor", "scenes", "with", "training", "and", "testing", "images", ".", "The", "images", "are", "captured", "by", "different", "sensors", "and", "hence", "come", "in", "various", "resolutions", ".", "The", "task", "is", "to", "segment", "indoor", "scene", "classes", "including", "wall", ",", "floor", ",", "ceiling", ",", "table", ",", "chair", ",", "sofa", "etc", ".", "This", "task", "is", "made", "hard", "by", "the", "fact", "that", "object", "classes", "come", "in", "various", "shapes", ",", "sizes", "and", "in", "different", "poses", ".", "There", "are", "frequent", "partial", "occlusions", "since", "there", "are", "typically", "many", "different", "classes", "present", "in", "each", "of", "the", "test", "images", ".", "These", "factors", "make", "this", "one", "of", "the", "hardest", "segmentation", "challenges", ".", "We", "only", "use", "the", "RGB", "modality", "for", "our", "training", "and", "testing", ".", "Using", "the", "depth", "modality", "would", "necessitate", "architectural", "modifications", "/", "redesign", ".", "Also", "the", "quality", "of", "depth", "images", "from", "current", "cameras", "require", "careful", "post", "-", "processing", "to", "fill", "-", "in", "missing", "measurements", ".", "They", "may", "also", "require", "using", "fusion", "of", "many", "frames", "to", "robustly", "extract", "features", "for", "segmentation", ".", "Therefore", "we", "believe", "using", "depth", "for", "segmentation", "merits", "a", "separate", "body", "of", "work", "which", "is", "not", "in", "the", "scope", "of", "this", "paper", ".", "We", "also", "note", "that", "an", "earlier", "benchmark", "dataset", "NYUv2", "is", "included", "as", "part", "of", "this", "dataset", ".", "Road", "scene", "images", "have", "limited", "variation", ",", "both", "in", "terms", "of", "the", "classes", "of", "interest", "and", "their", "spatial", "arrangements", ".", "When", "captured", "from", "a", "moving", "vehicle", "where", "the", "camera", "position", "is", "nearly", "always", "parallel", "to", "the", "road", "surface", "limiting", "variability", "in", "view", "points", ".", "This", "makes", "it", "easier", "for", "deep", "networks", "to", "learn", "to", "segment", "them", "robustly", ".", "In", "comparison", ",", "images", "of", "indoor", "scenes", "are", "more", "complex", "since", "the", "view", "points", "can", "vary", "a", "lot", "and", "there", "is", "less", "regularity", "in", "both", "the", "number", "of", "classes", "present", "in", "a", "scene", "and", "their", "spatial", "arrangement", ".", "Another", "difficulty", "is", "caused", "by", "the", "widely", "varying", "sizes", "of", "the", "object", "classes", "in", "the", "scene", ".", "Some", "test", "samples", "from", "the", "recent", "SUN", "RGB", "-", "D", "dataset", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "We", "observe", "some", "scenes", "with", "few", "large", "classes", "and", "some", "others", "with", "dense", "clutter", "(", "bottom", "row", "and", "right", ")", ".", "The", "appearance", "(", "texture", "and", "shape", ")", "can", "also", "widely", "vary", "in", "indoor", "scenes", ".", "Therefore", ",", "we", "believe", "this", "is", "the", "hardest", "challenge", "for", "segmentation", "architectures", "and", "methods", "in", "computer", "vision", ".", "Other", "challenges", ",", "such", "as", "Pascal", "VOC12", "salient", "object", "segmentation", "have", "occupied", "researchers", "more", ",", "but", "we", "believe", "indoor", "scene", "segmentation", "is", "more", "challenging", "and", "has", "more", "current", "practical", "applications", "such", "as", "in", "AR", "and", "robotics", ".", "To", "encourage", "more", "research", "in", "this", "direction", "we", "compared", "well", "known", "deep", "architectures", "on", "the", "large", "SUN", "RGB", "-", "D", "dataset", ".", "The", "qualitative", "results", "of", "SegNet", "on", "samples", "of", "indoor", "scenes", "of", "different", "types", "such", "as", "bedroom", ",", "living", "room", ",", "laboratory", ",", "meeting", "room", ",", "bathroom", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "We", "see", "that", "SegNet", "obtains", "reasonable", "predictions", "when", "the", "size", "of", "the", "classes", "are", "large", "under", "different", "view", "points", ".", "This", "is", "particularly", "interesting", "since", "the", "input", "modality", "is", "only", "RGB", ".", "RGB", "images", "are", "also", "useful", "to", "segment", "thinner", "structures", "such", "as", "the", "legs", "of", "chairs", "and", "tables", ",", "lamps", "which", "is", "difficult", "to", "achieve", "using", "depth", "images", "from", "currently", "available", "sensors", ".", "This", "can", "be", "seen", "from", "the", "results", "of", "SegNet", ",", "DeconvNet", "in", "Fig", ".", "[", "reference", "]", ".", "It", "is", "also", "useful", "to", "segment", "decorative", "objects", "such", "as", "paintings", "on", "the", "wall", "for", "AR", "tasks", ".", "However", "as", "compared", "to", "outdoor", "scenes", "the", "segmentation", "quality", "is", "clearly", "more", "noisy", ".", "The", "quality", "drops", "significantly", "when", "clutter", "is", "increased", "(", "see", "the", "result", "sample", "in", "the", "middle", "column", ")", ".", "The", "quantitative", "results", "in", "Table", "[", "reference", "]", "show", "that", "all", "the", "deep", "architectures", "share", "low", "mIoU", "and", "boundary", "metrics", ".", "The", "global", "and", "class", "averages", "(", "correlates", "well", "with", "mIou", ")", "are", "also", "small", ".", "SegNet", "outperforms", "all", "other", "methods", "in", "terms", "of", "G", ",", "C", ",", "BF", "metrics", "and", "has", "a", "slightly", "lower", "mIoU", "than", "DeepLab", "-", "LargeFOV", ".", "As", "a", "stand", "alone", "experiment", "we", "trained", "SegNet", "with", "median", "frequency", "class", "balancing", "and", "the", "metrics", "were", "higher", "(", "see", "Table", "[", "reference", "]", ")", "and", "this", "agrees", "with", "our", "analysis", "in", "Sec", ".", "[", "reference", "]", ".", "Interestingly", ",", "using", "the", "grid", "search", "based", "optimal", "hyperparameters", "for", "the", "dense", "-", "CRF", "worsened", "all", "except", "the", "BF", "score", "metric", "for", "DeepLab", "-", "LargeFOV", "-", "denseCRF", ".", "More", "optimal", "settings", "could", "perhaps", "be", "found", "but", "the", "grid", "search", "process", "was", "too", "expensive", "given", "the", "large", "inference", "time", "for", "dense", "-", "CRFs", ".", "One", "reason", "for", "the", "overall", "poor", "performance", "is", "the", "large", "number", "of", "classes", "in", "this", "segmentation", "task", ",", "many", "of", "which", "occupy", "a", "small", "part", "of", "the", "image", "and", "appear", "infrequently", ".", "The", "accuracies", "reported", "in", "Table", "[", "reference", "]", "clearly", "show", "that", "larger", "classes", "have", "reasonable", "accuracy", "and", "smaller", "classes", "have", "lower", "accuracies", ".", "This", "can", "be", "improved", "with", "larger", "sized", "datasets", "and", "class", "distribution", "aware", "training", "techniques", ".", "Another", "reason", "for", "poor", "performance", "could", "lie", "in", "the", "inability", "of", "these", "deep", "architectures", "(", "all", "are", "based", "on", "the", "VGG", "architecture", ")", "to", "large", "variability", "in", "indoor", "scenes", ".", "This", "conjecture", "on", "our", "part", "is", "based", "on", "the", "fact", "that", "the", "smallest", "model", "DeepLab", "-", "LargeFOV", "produces", "the", "best", "accuracy", "in", "terms", "of", "mIoU", "and", "in", "comparison", ",", "larger", "parameterizations", "in", "DeconvNet", ",", "FCN", "did", "not", "improve", "perfomance", "even", "with", "much", "longer", "training", "(", "DeconvNet", ")", ".", "This", "suggests", "there", "could", "lie", "a", "common", "reason", "for", "poor", "performance", "across", "all", "architectures", ".", "More", "controlled", "datasets", "are", "needed", "to", "verify", "this", "hypothesis", ".", "section", ":", "Discussion", "and", "future", "work", "Deep", "learning", "models", "have", "often", "achieved", "increasing", "success", "due", "to", "the", "availability", "of", "massive", "datasets", "and", "expanding", "model", "depth", "and", "parameterisation", ".", "However", ",", "in", "practice", "factors", "like", "memory", "and", "computational", "time", "during", "training", "and", "testing", "are", "important", "factors", "to", "consider", "when", "choosing", "a", "model", "from", "a", "large", "bank", "of", "models", ".", "Training", "time", "becomes", "an", "important", "consideration", "particularly", "when", "the", "performance", "gain", "is", "not", "commensurate", "with", "increased", "training", "time", "as", "shown", "in", "our", "experiments", ".", "Test", "time", "memory", "and", "computational", "load", "are", "important", "to", "deploy", "models", "on", "specialised", "embedded", "devices", ",", "for", "example", ",", "in", "AR", "applications", ".", "From", "an", "overall", "efficiency", "viewpoint", ",", "we", "feel", "less", "attention", "has", "been", "paid", "to", "smaller", "and", "more", "memory", ",", "time", "efficient", "models", "for", "real", "-", "time", "applications", "such", "as", "road", "scene", "understanding", "and", "AR", ".", "This", "was", "the", "primary", "motivation", "behind", "the", "proposal", "of", "SegNet", ",", "which", "is", "significantly", "smaller", "and", "faster", "than", "other", "competing", "architectures", ",", "but", "which", "we", "have", "shown", "to", "be", "efficient", "for", "tasks", "such", "as", "road", "scene", "understanding", ".", "Segmentation", "challenges", "such", "as", "Pascal", "and", "MS", "-", "COCO", "are", "object", "segmentation", "challenges", "wherein", "a", "few", "classes", "are", "present", "in", "any", "test", "image", ".", "Scene", "segmentation", "is", "more", "challenging", "due", "to", "the", "high", "variability", "of", "indoor", "scenes", "and", "a", "need", "to", "segment", "a", "larger", "number", "of", "classes", "simultaneously", ".", "The", "task", "of", "outdoor", "and", "indoor", "scene", "segmentation", "are", "also", "more", "practically", "oriented", "with", "current", "applications", "such", "as", "autonomous", "driving", ",", "robotics", "and", "AR", ".", "The", "metrics", "we", "chose", "to", "benchmark", "various", "deep", "segmentation", "architectures", "like", "the", "boundary", "F1", "-", "measure", "(", "BF", ")", "was", "done", "to", "complement", "the", "existing", "metrics", "which", "are", "more", "biased", "towards", "region", "accuracies", ".", "It", "is", "clear", "from", "our", "experiments", "and", "other", "independent", "benchmarks", "that", "outdoor", "scene", "images", "captured", "from", "a", "moving", "car", "are", "easier", "to", "segment", "and", "deep", "architectures", "perform", "robustly", ".", "We", "hope", "our", "experiments", "will", "encourage", "researchers", "to", "engage", "their", "attention", "towards", "the", "more", "challenging", "indoor", "scene", "segmentation", "task", ".", "An", "important", "choice", "we", "had", "to", "make", "when", "benchmarking", "different", "deep", "architectures", "of", "varying", "parameterization", "was", "the", "manner", "in", "which", "to", "train", "them", ".", "Many", "of", "these", "architectures", "have", "used", "a", "host", "of", "supporting", "techniques", "and", "multi", "-", "stage", "training", "recipes", "to", "arrive", "at", "high", "accuracies", "on", "datasets", "but", "this", "makes", "it", "difficult", "to", "gather", "evidence", "about", "their", "true", "performance", "under", "time", "and", "memory", "constraints", ".", "Instead", "we", "chose", "to", "perform", "a", "controlled", "benchmarking", "where", "we", "used", "batch", "normalization", "to", "enable", "end", "-", "to", "-", "end", "training", "with", "the", "same", "solver", "(", "SGD", ")", ".", "However", ",", "we", "note", "that", "this", "approach", "can", "not", "entirely", "disentangle", "the", "effects", "of", "model", "versus", "solver", "(", "optimization", ")", "in", "achieving", "a", "particular", "result", ".", "This", "is", "mainly", "due", "to", "the", "fact", "that", "training", "these", "networks", "involves", "gradient", "back", "-", "propagation", "which", "is", "imperfect", "and", "the", "optimization", "is", "a", "non", "-", "convex", "problem", "in", "extremely", "large", "dimensions", ".", "Acknowledging", "these", "shortcomings", ",", "our", "hope", "is", "that", "this", "controlled", "analysis", "complements", "other", "benchmarks", "and", "reveals", "the", "practical", "trade", "-", "offs", "involved", "in", "different", "well", "known", "architectures", ".", "For", "the", "future", ",", "we", "would", "like", "to", "exploit", "our", "understanding", "of", "segmentation", "architectures", "gathered", "from", "our", "analysis", "to", "design", "more", "efficient", "architectures", "for", "real", "-", "time", "applications", ".", "We", "are", "also", "interested", "in", "estimating", "the", "model", "uncertainty", "for", "predictions", "from", "deep", "segmentation", "architectures", ",", ".", "section", ":", "Conclusion", "We", "presented", "SegNet", ",", "a", "deep", "convolutional", "network", "architecture", "for", "semantic", "segmentation", ".", "The", "main", "motivation", "behind", "SegNet", "was", "the", "need", "to", "design", "an", "efficient", "architecture", "for", "road", "and", "indoor", "scene", "understanding", "which", "is", "efficient", "both", "in", "terms", "of", "memory", "and", "computational", "time", ".", "We", "analysed", "SegNet", "and", "compared", "it", "with", "other", "important", "variants", "to", "reveal", "the", "practical", "trade", "-", "offs", "involved", "in", "designing", "architectures", "for", "segmentation", ",", "particularly", "training", "time", ",", "memory", "versus", "accuracy", ".", "Those", "architectures", "which", "store", "the", "encoder", "network", "feature", "maps", "in", "full", "perform", "best", "but", "consume", "more", "memory", "during", "inference", "time", ".", "SegNet", "on", "the", "other", "hand", "is", "more", "efficient", "since", "it", "only", "stores", "the", "max", "-", "pooling", "indices", "of", "the", "feature", "maps", "and", "uses", "them", "in", "its", "decoder", "network", "to", "achieve", "good", "performance", ".", "On", "large", "and", "well", "known", "datasets", "SegNet", "performs", "competitively", ",", "achieving", "high", "scores", "for", "road", "scene", "understanding", ".", "End", "-", "to", "-", "end", "learning", "of", "deep", "segmentation", "architectures", "is", "a", "harder", "challenge", "and", "we", "hope", "to", "see", "more", "attention", "paid", "to", "this", "important", "problem", ".", "bibliography", ":", "References", "[", "]", "Vijay", "Badrinarayanan", "obtained", "his", "Ph", ".", "D", "from", "INRIA", "Rennes", ",", "France", "in", "2009", ".", "He", "was", "a", "senior", "post", "-", "doctoral", "research", "associate", "at", "the", "Machine", "Intelligence", "Laboratory", ",", "Department", "of", "Engineering", ",", "University", "of", "Cambridge", ",", "U.K.", "He", "currently", "works", "as", "a", "Principal", "Engineer", ",", "Deep", "Learning", "at", "Magic", "Leap", ",", "Inc.", "in", "Mountain", "View", ",", "CA", ".", "His", "research", "interests", "are", "in", "probabilistic", "graphical", "models", ",", "deep", "learning", "applied", "to", "image", "and", "video", "based", "perception", "problems", ".", "[", "]", "Alex", "Kendall", "graduated", "with", "a", "Bachelor", "of", "Engineering", "with", "First", "Class", "Honours", "in", "2013", "from", "the", "University", "of", "Auckland", ",", "New", "Zealand", ".", "In", "2014", "he", "was", "awarded", "a", "Woolf", "Fisher", "Scholarship", "to", "study", "towards", "a", "Ph", ".", "D", "at", "the", "University", "of", "Cambridge", ",", "U.K.", "He", "is", "a", "member", "of", "the", "Machine", "Intelligence", "Laboratory", "and", "is", "interested", "in", "applications", "of", "deep", "learning", "for", "mobile", "robotics", ".", "[", "]", "Roberto", "Cipolla", "obtained", "a", "B.A.", "(", "Engineering", ")", "degree", "from", "the", "University", "of", "Cambridge", "in", "1984", ",", "an", "M.S.E.", "(", "Electrical", "Engineering", ")", "from", "the", "University", "of", "Pennsylvania", "in", "1985", "and", "a", "D.Phil", ".", "(", "Computer", "Vision", ")", "from", "the", "University", "of", "Oxford", "in", "1991", ".", "from", "1991", "-", "92", "was", "a", "Toshiba", "Fellow", "and", "engineer", "at", "the", "Toshiba", "Corporation", "Research", "and", "Development", "Centre", "in", "Kawasaki", ",", "Japan", ".", "He", "joined", "the", "Department", "of", "Engineering", ",", "University", "of", "Cambridge", "in", "1992", "as", "a", "Lecturer", "and", "a", "Fellow", "of", "Jesus", "College", ".", "He", "became", "a", "Reader", "in", "Information", "Engineering", "in", "1997", "and", "a", "Professor", "in", "2000", ".", "He", "became", "a", "Fellow", "of", "the", "Royal", "Academy", "of", "Engineering", "(", "FREng", ")", "in", "2010", ".", "His", "research", "interests", "are", "in", "computer", "vision", "and", "robotics", ".", "He", "has", "authored", "3", "books", ",", "edited", "9", "volumes", "and", "co", "-", "authored", "more", "than", "300", "papers", "."], "entity": [[2, 3, ["SegNet"], "Method"], [5, 11, ["Deep", "Convolutional", "Encoder", "-", "Decoder", "Architecture"], "Method"], [12, 14, ["Image", "Segmentation"], "Task"], [20, 26, ["deep", "fully", "convolutional", "neural", "network", "architecture"], "Method"], [27, 32, ["semantic", "pixel", "-", "wise", "segmentation"], "Task"], [33, 34, ["SegNet"], "Task"], [36, 40, ["core", "trainable", "segmentation", "engine"], "Method"], [43, 45, ["encoder", "network"], "Method"], [48, 50, ["decoder", "network"], "Method"], [53, 58, ["pixel", "-", "wise", "classification", "layer"], "Method"], [63, 65, ["encoder", "network"], "Method"], [71, 73, ["convolutional", "layers"], "Method"], [75, 77, ["VGG16", "network"], "Method"], [82, 84, ["decoder", "network"], "Method"], [100, 104, ["pixel", "-", "wise", "classification"], "Task"], [108, 109, ["SegNet"], "Method"], [132, 133, ["decoder"], "Method"], [139, 143, ["max", "-", "pooling", "step"], "Method"], [149, 153, ["non", "-", "linear", "upsampling"], "Task"], [173, 175, ["trainable", "filters"], "Method"], [190, 191, ["FCN"], "Method"], [197, 203, ["DeepLab", "-", "LargeFOV", ",", "DeconvNet", "architectures"], "Method"], [208, 214, ["memory", "versus", "accuracy", "trade", "-", "off"], "Metric"], [218, 219, ["segmentation"], "Task"], [221, 222, ["SegNet"], "Method"], [226, 229, ["scene", "understanding", "applications"], "Task"], [242, 243, ["memory"], "Metric"], [244, 246, ["computational", "time"], "Metric"], [247, 248, ["inference"], "Task"], [274, 277, ["stochastic", "gradient", "descent"], "Method"], [285, 286, ["SegNet"], "Method"], [291, 293, ["road", "scenes"], "Task"], [294, 302, ["SUN", "RGB", "-", "D", "indoor", "scene", "segmentation", "tasks"], "Task"], [308, 309, ["SegNet"], "Method"], [313, 316, ["competitive", "inference", "time"], "Metric"], [319, 323, ["inference", "memory", "-", "wise"], "Metric"], [333, 335, ["Caffe", "implementation"], "Method"], [336, 337, ["SegNet"], "Method"], [339, 341, ["web", "demo"], "Task"], [343, 347, ["Deep", "Convolutional", "Neural", "Networks"], "Method"], [348, 353, ["Semantic", "Pixel", "-", "Wise", "Segmentation"], "Task"], [354, 356, ["Indoor", "Scenes"], "Task"], [357, 359, ["Road", "Scenes"], "Task"], [360, 361, ["Encoder"], "Method"], [362, 363, ["Decoder"], "Method"], [364, 365, ["Pooling"], "Task"], [366, 367, ["Upsampling"], "Method"], [371, 373, ["Semantic", "segmentation"], "Task"], [381, 383, ["scene", "understanding"], "Task"], [384, 390, ["inferring", "support", "-", "relationships", "among", "objects"], "Task"], [391, 393, ["autonomous", "driving"], "Task"], [410, 413, ["machine", "learning", "algorithms"], "Method"], [417, 419, ["deep", "learning"], "Method"], [425, 428, ["handwritten", "digit", "recognition"], "Task"], [429, 430, ["speech"], "Task"], [431, 434, ["categorising", "whole", "images"], "Task"], [435, 439, ["detecting", "objects", "in", "images"], "Task"], [447, 452, ["semantic", "pixel", "-", "wise", "labelling"], "Task"], [476, 478, ["deep", "architectures"], "Method"], [480, 482, ["category", "prediction"], "Task"], [483, 487, ["pixel", "-", "wise", "labelling"], "Task"], [502, 504, ["max", "pooling"], "Method"], [505, 508, ["sub", "-", "sampling"], "Method"], [509, 512, ["feature", "map", "resolution"], "Method"], [517, 518, ["SegNet"], "Method"], [531, 535, ["pixel", "-", "wise", "classification"], "Task"], [546, 548, ["boundary", "localization"], "Task"], [552, 553, ["SegNet"], "Method"], [562, 567, ["pixel", "-", "wise", "semantic", "segmentation"], "Task"], [573, 577, ["road", "scene", "understanding", "applications"], "Task"], [618, 620, ["road", "scenes"], "Material"], [641, 643, ["smooth", "segmentations"], "Task"], [674, 676, ["image", "representation"], "Method"], [695, 699, ["memory", "and", "computation", "time"], "Metric"], [700, 701, ["inference"], "Task"], [725, 728, ["weight", "update", "technique"], "Method"], [730, 733, ["stochastic", "gradient", "descent"], "Method"], [734, 736, ["SGD", ")"], "Method"], [750, 751, ["SegNet"], "Task"], [761, 763, ["encoder", "network"], "Method"], [764, 765, ["SegNet"], "Method"], [770, 772, ["convolutional", "layers"], "Method"], [773, 774, ["VGG16"], "Method"], [778, 781, ["fully", "connected", "layers"], "Method"], [782, 783, ["VGG16"], "Method"], [786, 789, ["SegNet", "encoder", "network"], "Method"], [805, 806, ["SegNet"], "Method"], [808, 810, ["decoder", "network"], "Method"], [828, 829, ["decoders"], "Method"], [839, 840, ["encoder"], "Method"], [842, 846, ["non", "-", "linear", "upsampling"], "Method"], [861, 864, ["unsupervised", "feature", "learning"], "Task"], [872, 874, ["decoding", "process"], "Method"], [884, 886, ["boundary", "delineation"], "Task"], [901, 903, ["end", "training"], "Task"], [911, 912, ["upsampling"], "Method"], [917, 921, ["encoder", "-", "decoder", "architecture"], "Method"], [942, 945, ["SegNet", "decoding", "technique"], "Method"], [949, 952, ["Fully", "Convolutional", "Network"], "Method"], [953, 954, ["FCN"], "Method"], [970, 972, ["segmentation", "architectures"], "Method"], [975, 977, ["deep", "architectures"], "Method"], [978, 979, ["segmentation"], "Task"], [981, 983, ["encoder", "networks"], "Method"], [985, 986, ["VGG16"], "Method"], [994, 996, ["decoder", "network"], "Method"], [999, 1000, ["inference"], "Task"], [1022, 1028, ["end", "-", "to", "-", "end", "training"], "Task"], [1038, 1042, ["multi", "-", "stage", "training"], "Task"], [1047, 1051, ["pre", "-", "trained", "architecture"], "Method"], [1053, 1054, ["FCN"], "Method"], [1057, 1059, ["supporting", "aids"], "Method"], [1061, 1063, ["region", "proposals"], "Method"], [1064, 1065, ["inference"], "Task"], [1069, 1073, ["classification", "and", "segmentation", "networks"], "Method"], [1082, 1083, ["training"], "Task"], [1085, 1087, ["full", "training"], "Task"], [1092, 1097, ["boosting", "post", "-", "processing", "techniques"], "Method"], [1136, 1138, ["decoding", "process"], "Method"], [1156, 1157, ["SegNet"], "Method"], [1159, 1162, ["scene", "segmentation", "tasks"], "Task"], [1163, 1167, ["CamVid", "road", "scene", "segmentation"], "Task"], [1168, 1175, ["SUN", "RGB", "-", "D", "indoor", "scene", "segmentation"], "Task"], [1176, 1178, ["Pascal", "VOC12"], "Material"], [1184, 1185, ["segmentation"], "Task"], [1215, 1216, ["detection"], "Task"], [1224, 1229, ["decoupled", "classification", "-", "segmentation", "network"], "Method"], [1231, 1233, ["classification", "network"], "Method"], [1241, 1244, ["weakly", "labelled", "data"], "Material"], [1246, 1249, ["independent", "segmentation", "network"], "Method"], [1263, 1265, ["classification", "network"], "Method"], [1268, 1273, ["CRF", "post", "-", "processing", "technique"], "Method"], [1275, 1276, ["segmentation"], "Task"], [1291, 1293, ["region", "proposals"], "Method"], [1301, 1303, ["scene", "understanding"], "Task"], [1321, 1323, ["robust", "segmentation"], "Task"], [1329, 1330, ["SegNet"], "Method"], [1334, 1343, ["real", "-", "time", "online", "demo", "of", "road", "scene", "segmentation"], "Task"], [1349, 1351, ["autonomous", "driving"], "Task"], [1368, 1373, ["randomly", "sampled", "road", "scene", "images"], "Material"], [1374, 1379, ["Google", "and", "indoor", "test", "scenes"], "Material"], [1381, 1386, ["SUN", "RGB", "-", "D", "dataset"], "Material"], [1420, 1422, ["SegNet", "architecture"], "Method"], [1443, 1444, ["SegNet"], "Task"], [1445, 1450, ["outdoor", "and", "indoor", "scene", "datasets"], "Material"], [1486, 1491, ["Semantic", "pixel", "-", "wise", "segmentation"], "Task"], [1507, 1509, ["deep", "networks"], "Method"], [1532, 1533, ["classifier"], "Method"], [1534, 1536, ["Random", "Forest"], "Method"], [1537, 1538, ["Boosting"], "Method"], [1553, 1556, ["SfM", "and", "appearance"], "Method"], [1561, 1566, ["CamVid", "road", "scene", "understanding", "test"], "Task"], [1581, 1582, ["classifiers"], "Method"], [1592, 1595, ["higher", "order", "CRF"], "Method"], [1598, 1599, ["accuracy"], "Metric"], [1636, 1640, ["Random", "Forest", "based", "unaries"], "Method"], [1654, 1656, ["CamVid", "video"], "Material"], [1663, 1664, ["classification"], "Task"], [1665, 1667, ["Random", "Forests"], "Method"], [1683, 1689, ["spatio", "-", "temporal", "super", "-", "pixelization"], "Method"], [1692, 1693, ["accuracy"], "Metric"], [1700, 1702, ["CamVid", "test"], "Material"], [1714, 1716, ["classifier", "predictions"], "Method"], [1718, 1720, ["CRF", "framework"], "Method"], [1734, 1735, ["classification"], "Task"], [1736, 1743, ["Indoor", "RGBD", "pixel", "-", "wise", "semantic", "segmentation"], "Task"], [1752, 1754, ["NYU", "dataset"], "Material"], [1766, 1767, ["segmentation"], "Task"], [1788, 1791, ["neural", "network", "classifier"], "Method"], [1804, 1805, ["CRF"], "Method"], [1815, 1816, ["LBP"], "Method"], [1817, 1819, ["region", "segmentation"], "Method"], [1822, 1823, ["accuracy"], "Metric"], [1826, 1827, ["CRF"], "Method"], [1834, 1836, ["class", "segmentation"], "Task"], [1856, 1864, ["real", "-", "time", "joint", "reconstruction", "and", "semantic", "segmentation"], "Task"], [1866, 1868, ["Random", "Forests"], "Method"], [1872, 1873, ["classifier"], "Method"], [1879, 1881, ["boundary", "detection"], "Method"], [1882, 1884, ["hierarchical", "grouping"], "Method"], [1886, 1888, ["category", "segmentation"], "Task"], [1904, 1906, ["classification", "of"], "Task"], [1907, 1911, ["RGB", "or", "RGBD", "images"], "Material"], [1915, 1919, ["deep", "convolutional", "neural", "networks"], "Method"], [1920, 1922, ["object", "classification"], "Task"], [1930, 1933, ["feature", "learning", "capabilities"], "Method"], [1934, 1937, ["structured", "prediction", "problems"], "Task"], [1939, 1940, ["segmentation"], "Task"], [1948, 1949, ["networks"], "Method"], [1951, 1953, ["object", "categorization"], "Task"], [1954, 1955, ["segmentation"], "Task"], [1974, 1975, ["classification"], "Task"], [1981, 1984, ["recurrent", "neural", "networks"], "Method"], [1986, 1989, ["low", "resolution", "predictions"], "Method"], [2016, 2018, ["deep", "architectures"], "Method"], [2021, 2022, ["segmentation"], "Task"], [2038, 2042, ["low", "resolution", "image", "representations"], "Task"], [2043, 2047, ["pixel", "-", "wise", "predictions"], "Task"], [2049, 2051, ["encoder", "network"], "Method"], [2054, 2057, ["low", "resolution", "representations"], "Method"], [2064, 2067, ["VGG16", "classification", "network"], "Method"], [2069, 2071, ["convolutional", "layers"], "Method"], [2072, 2075, ["fully", "connected", "layers"], "Method"], [2077, 2080, ["encoder", "network", "weights"], "Method"], [2088, 2092, ["ImageNet", "object", "classification", "dataset"], "Material"], [2094, 2096, ["decoder", "network"], "Method"], [2117, 2118, ["classification"], "Task"], [2123, 2130, ["Fully", "Convolutional", "Network", "(", "FCN", ")", "architecture"], "Method"], [2172, 2174, ["encoder", "network"], "Method"], [2182, 2184, ["decoder", "network"], "Method"], [2217, 2222, ["stage", "-", "wise", "training", "process"], "Method"], [2228, 2230, ["decoder", "network"], "Method"], [2236, 2238, ["trained", "network"], "Method"], [2290, 2291, ["decoder"], "Method"], [2316, 2317, ["predictive"], "Task"], [2319, 2320, ["FCN"], "Method"], [2327, 2328, ["FCN"], "Method"], [2330, 2333, ["recurrent", "neural", "network"], "Method"], [2334, 2335, ["RNN"], "Method"], [2342, 2344, ["large", "datasets"], "Material"], [2347, 2349, ["RNN", "layers"], "Method"], [2356, 2357, ["CRFs"], "Method"], [2360, 2366, ["feature", "representation", "power", "of", "FCN", "’s"], "Method"], [2373, 2376, ["FCN", "-", "8"], "Method"], [2392, 2395, ["FCN", "-", "8"], "Method"], [2401, 2404, ["CRF", "-", "RNN"], "Method"], [2417, 2420, ["FCN", "-", "8"], "Method"], [2424, 2426, ["joint", "training"], "Method"], [2439, 2441, ["deconvolutional", "network"], "Method"], [2445, 2446, ["FCN"], "Method"], [2454, 2457, ["training", "and", "inference"], "Task"], [2471, 2474, ["CRF", "-", "RNN"], "Method"], [2479, 2485, ["core", "feed", "-", "forward", "segmentation", "engine"], "Method"], [2494, 2498, ["CRF", "-", "RNN", "network"], "Method"], [2503, 2506, ["deep", "segmentation", "architecture"], "Method"], [2507, 2508, ["SegNet"], "Method"], [2509, 2514, ["Multi", "-", "scale", "deep", "architectures"], "Method"], [2539, 2543, ["deep", "feature", "extraction", "networks"], "Method"], [2559, 2561, ["deep", "architecture"], "Method"], [2587, 2590, ["early", "encoding", "layers"], "Method"], [2617, 2622, ["multi", "-", "stage", "training", "process"], "Method"], [2626, 2628, ["data", "augmentation"], "Task"], [2630, 2631, ["inference"], "Task"], [2636, 2638, ["convolutional", "pathways"], "Method"], [2639, 2641, ["feature", "extraction"], "Task"], [2645, 2646, ["CRF"], "Method"], [2648, 2652, ["multi", "-", "scale", "network"], "Method"], [2670, 2671, ["optimization"], "Task"], [2682, 2684, ["deep", "architectures"], "Method"], [2685, 2686, ["segmentation"], "Task"], [2692, 2694, ["inference", "time"], "Task"], [2700, 2702, ["MAP", "inference"], "Method"], [2704, 2705, ["CRF"], "Method"], [2710, 2712, ["region", "proposals"], "Method"], [2713, 2714, ["inference"], "Task"], [2725, 2726, ["CRF"], "Method"], [2733, 2735, ["decoding", "techniques"], "Method"], [2738, 2743, ["feed", "-", "forward", "segmentation", "engine"], "Method"], [2744, 2745, ["SegNet"], "Method"], [2750, 2751, ["decoders"], "Method"], [2756, 2760, ["pixel", "-", "wise", "classification"], "Task"], [2764, 2766, ["Deconvolutional", "Network"], "Method"], [2768, 2772, ["semi", "-", "supervised", "variant"], "Method"], [2773, 2775, ["Decoupled", "network"], "Method"], [2790, 2794, ["non", "-", "linear", "upsampling"], "Method"], [2796, 2798, ["decoder", "network"], "Method"], [2807, 2808, ["SegNet"], "Method"], [2820, 2821, ["decoding"], "Task"], [2823, 2825, ["decoder", "network"], "Method"], [2829, 2831, ["encoder", "network"], "Method"], [2839, 2843, ["VGG", "-", "16", "network"], "Method"], [2877, 2878, ["training"], "Task"], [2882, 2883, ["inference"], "Task"], [2890, 2892, ["inference", "time"], "Task"], [2914, 2918, ["encoder", "-", "decoder", "network"], "Method"], [2929, 2932, ["fully", "connected", "layers"], "Method"], [2934, 2937, ["VGG16", "encoder", "network"], "Method"], [2950, 2952, ["SGD", "optimization"], "Method"], [2971, 2973, ["memory", "consumption"], "Metric"], [2975, 2977, ["inference", "time"], "Metric"], [2984, 2988, ["unsupervised", "feature", "learning", "architecture"], "Method"], [2997, 2999, ["learning", "module"], "Method"], [3001, 3005, ["encoder", "-", "decoder", "network"], "Method"], [3007, 3008, ["encoder"], "Method"], [3010, 3011, ["convolution"], "Method"], [3013, 3015, ["filter", "bank"], "Method"], [3016, 3023, ["element", "-", "wise", "tanh", "non", "-", "linearity"], "Method"], [3024, 3027, ["max", "-", "pooling"], "Method"], [3028, 3031, ["sub", "-", "sampling"], "Method"], [3056, 3057, ["decoder"], "Method"], [3074, 3076, ["upsampled", "map"], "Method"], [3078, 3082, ["trainable", "decoder", "filter", "bank"], "Method"], [3093, 3097, ["unsupervised", "pre", "-", "training"], "Task"], [3098, 3099, ["classification"], "Task"], [3103, 3105, ["decoding", "technique"], "Method"], [3109, 3112, ["trained", "convolutional", "networks"], "Method"], [3113, 3114, ["classification"], "Task"], [3125, 3130, ["layer", "-", "wise", "feature", "learning"], "Method"], [3153, 3155, ["hierarchical", "encoders"], "Method"], [3165, 3170, ["deep", "encoder", "-", "decoder", "networks"], "Method"], [3171, 3174, ["unsupervised", "feature", "training"], "Task"], [3181, 3183, ["encoder", "training"], "Method"], [3186, 3187, ["SegNet"], "Method"], [3193, 3198, ["deep", "encoder", "-", "decoder", "network"], "Method"], [3203, 3206, ["supervised", "learning", "task"], "Task"], [3224, 3227, ["pixel", "wise", "predictions"], "Task"], [3230, 3232, ["deep", "networks"], "Method"], [3233, 3237, ["image", "super", "-", "resolution"], "Task"], [3238, 3241, ["depth", "map", "prediction"], "Task"], [3253, 3254, ["learning"], "Task"], [3273, 3274, ["SegNet"], "Method"], [3276, 3278, ["encoder", "network"], "Method"], [3281, 3283, ["decoder", "network"], "Method"], [3288, 3291, ["pixelwise", "classification", "layer"], "Method"], [3304, 3306, ["encoder", "network"], "Method"], [3308, 3310, ["convolutional", "layers"], "Method"], [3315, 3317, ["convolutional", "layers"], "Method"], [3319, 3321, ["VGG16", "network"], "Method"], [3323, 3325, ["object", "classification"], "Task"], [3331, 3333, ["training", "process"], "Method"], [3334, 3335, ["weights"], "Method"], [3337, 3338, ["classification"], "Task"], [3339, 3341, ["large", "datasets"], "Material"], [3373, 3376, ["SegNet", "encoder", "network"], "Method"], [3402, 3404, ["encoder", "layer"], "Method"], [3412, 3414, ["decoder", "network"], "Method"], [3425, 3432, ["multi", "-", "class", "soft", "-", "max", "classifier"], "Method"], [3442, 3443, ["encoder"], "Method"], [3445, 3447, ["encoder", "network"], "Method"], [3448, 3449, ["convolution"], "Method"], [3451, 3453, ["filter", "bank"], "Method"], [3471, 3483, ["element", "-", "wise", "rectified", "-", "linear", "non", "-", "linearity", "(", "ReLU", ")"], "Method"], [3489, 3492, ["max", "-", "pooling"], "Method"], [3518, 3521, ["Max", "-", "pooling"], "Method"], [3536, 3539, ["Sub", "-", "sampling"], "Method"], [3562, 3565, ["max", "-", "pooling"], "Method"], [3566, 3569, ["sub", "-", "sampling"], "Method"], [3575, 3577, ["robust", "classification"], "Task"], [3592, 3599, ["lossy", "(", "boundary", "detail", ")", "image", "representation"], "Method"], [3603, 3604, ["segmentation"], "Task"], [3605, 3607, ["boundary", "delineation"], "Task"], [3627, 3630, ["sub", "-", "sampling"], "Method"], [3636, 3637, ["inference"], "Task"], [3649, 3652, ["sub", "-", "sampling"], "Method"], [3763, 3764, ["accuracy"], "Metric"], [3777, 3779, ["decoder", "network"], "Method"], [3815, 3818, ["SegNet", "decoding", "technique"], "Method"], [3828, 3830, ["feature", "maps"], "Method"], [3835, 3839, ["trainable", "decoder", "filter", "bank"], "Method"], [3846, 3849, ["batch", "normalization", "step"], "Method"], [3861, 3862, ["decoder"], "Method"], [3866, 3867, ["encoder"], "Method"], [3898, 3899, ["decoders"], "Method"], [3920, 3924, ["high", "dimensional", "feature", "representation"], "Method"], [3930, 3931, ["decoder"], "Method"], [3936, 3940, ["soft", "-", "max", "classifier"], "Method"], [3942, 3946, ["soft", "-", "max", "classifies"], "Method"], [3954, 3958, ["soft", "-", "max", "classifier"], "Method"], [3995, 3996, ["DeconvNet"], "Method"], [3997, 4000, ["U", "-", "Net"], "Method"], [4005, 4006, ["SegNet"], "Method"], [4011, 4012, ["DeconvNet"], "Method"], [4052, 4054, ["convolutional", "manner"], "Method"], [4060, 4061, ["DeconvNet"], "Method"], [4074, 4075, ["SegNet"], "Method"], [4076, 4079, ["U", "-", "Net"], "Method"], [4083, 4086, ["medical", "imaging", "community"], "Task"], [4110, 4111, ["decoders"], "Method"], [4118, 4123, ["deconvolution", ")", "decoder", "feature", "maps"], "Method"], [4135, 4138, ["U", "-", "Net"], "Method"], [4141, 4144, ["VGG", "net", "architecture"], "Method"], [4145, 4146, ["SegNet"], "Method"], [4163, 4165, ["VGG", "net"], "Method"], [4173, 4175, ["Decoder", "Variants"], "Method"], [4176, 4178, ["segmentation", "architectures"], "Method"], [4181, 4183, ["encoder", "network"], "Method"], [4192, 4194, ["decoder", "network"], "Method"], [4202, 4205, ["SegNet", "decoding", "technique"], "Method"], [4209, 4217, ["Fully", "Convolutional", "Network", "(", "FCN", ")", "decoding", "technique"], "Method"], [4222, 4223, ["SegNet"], "Method"], [4228, 4232, ["FCN", "(", "decoder", "variants"], "Method"], [4239, 4240, ["SegNet"], "Method"], [4242, 4245, ["SegNet", "-", "Basic"], "Method"], [4249, 4250, ["encoders"], "Method"], [4252, 4253, ["decoders"], "Method"], [4256, 4257, ["encoders"], "Method"], [4258, 4261, ["SegNet", "-", "Basic"], "Method"], [4262, 4265, ["max", "-", "pooling"], "Method"], [4266, 4269, ["sub", "-", "sampling"], "Method"], [4284, 4286, ["Batch", "normalization"], "Method"], [4290, 4292, ["convolutional", "layer"], "Method"], [4295, 4299, ["encoder", "and", "decoder", "network"], "Method"], [4305, 4306, ["convolutions"], "Method"], [4308, 4312, ["ReLU", "non", "-", "linearity"], "Method"], [4316, 4318, ["decoder", "network"], "Method"], [4329, 4333, ["encoder", "and", "decoder", "layers"], "Method"], [4341, 4343, ["smooth", "labelling"], "Task"], [4374, 4377, ["SegNet", "-", "Basic"], "Method"], [4385, 4386, ["decoders"], "Method"], [4397, 4400, ["FCN", "-", "Basic"], "Method"], [4405, 4406, ["FCN"], "Method"], [4413, 4415, ["encoder", "network"], "Method"], [4416, 4419, ["SegNet", "-", "Basic"], "Method"], [4422, 4425, ["FCN", "decoding", "technique"], "Method"], [4450, 4452, ["decoding", "technique"], "Method"], [4454, 4455, ["SegNet"], "Method"], [4457, 4460, ["SegNet", "-", "Basic"], "Method"], [4470, 4472, ["upsampling", "step"], "Method"], [4482, 4487, ["multi", "-", "channel", "decoder", "filters"], "Method"], [4494, 4496, ["decoder", "filter"], "Method"], [4517, 4519, ["decoder", "filters"], "Method"], [4536, 4541, ["SegNet", "-", "Basic", "-", "SingleChannelDecoder"], "Method"], [4549, 4551, ["inference", "time"], "Metric"], [4564, 4565, ["FCN"], "Method"], [4567, 4573, ["FCN", "-", "Basic", ")", "decoding", "technique"], "Method"], [4580, 4582, ["FCN", "model"], "Method"], [4583, 4586, ["dimensionality", "reduction", "step"], "Method"], [4588, 4591, ["encoder", "feature", "maps"], "Method"], [4605, 4606, ["decoders"], "Method"], [4607, 4609, ["Dimensionality", "reduction"], "Task"], [4611, 4614, ["encoder", "feature", "maps"], "Method"], [4626, 4628, ["trainable", "filters"], "Method"], [4649, 4651, ["decoder", "network"], "Method"], [4659, 4660, ["upsampling"], "Method"], [4663, 4665, ["inverse", "convolution"], "Method"], [4667, 4675, ["fixed", "or", "trainable", "multi", "-", "channel", "upsampling", "kernel"], "Method"], [4686, 4687, ["upsampling"], "Task"], [4691, 4692, ["deconvolution"], "Task"], [4699, 4700, ["SegNet"], "Method"], [4701, 4705, ["multi", "-", "channel", "convolution"], "Method"], [4706, 4709, ["trainable", "decoder", "filters"], "Method"], [4712, 4713, ["upsampling"], "Method"], [4719, 4722, ["upsampled", "feature", "map"], "Method"], [4723, 4724, ["FCN"], "Method"], [4750, 4752, ["upsampling", "kernels"], "Method"], [4755, 4758, ["bilinear", "interpolation", "weights"], "Method"], [4760, 4763, ["FCN", "decoder", "model"], "Method"], [4769, 4770, ["inference"], "Task"], [4777, 4779, ["embedded", "applications"], "Task"], [4791, 4794, ["FCN", "-", "Basic"], "Method"], [4812, 4814, ["dimensionality", "reduction"], "Method"], [4825, 4826, ["SegNet"], "Method"], [4833, 4835, ["storage", "cost"], "Metric"], [4859, 4863, ["FCN", "-", "Basic", "model"], "Method"], [4866, 4871, ["encoder", "feature", "map", "addition", "step"], "Method"], [4875, 4877, ["upsampling", "kernels"], "Method"], [4878, 4883, ["FCN", "-", "Basic", "-", "NoAddition"], "Method"], [4894, 4895, ["upsampling"], "Method"], [4897, 4900, ["bilinear", "interpolation", "weights"], "Method"], [4904, 4905, ["learning"], "Method"], [4906, 4907, ["upsampling"], "Task"], [4908, 4911, ["Bilinear", "-", "Interpolation"], "Method"], [4936, 4938, ["SegNet", "decoder"], "Method"], [4946, 4947, ["SegNet"], "Method"], [4948, 4953, ["SegNet", "-", "Basic", "-", "EncoderAddition"], "Method"], [4961, 4962, ["upsampling"], "Task"], [4968, 4970, ["convolution", "step"], "Method"], [4998, 5004, ["memory", "intensive", "FCN", "-", "Basic", "variant"], "Method"], [5005, 5010, ["FCN", "-", "Basic", "-", "NoDimReduction"], "Method"], [5016, 5018, ["dimensionality", "reduction"], "Method"], [5029, 5032, ["FCN", "-", "Basic"], "Method"], [5034, 5037, ["encoder", "feature", "map"], "Method"], [5047, 5049, ["decoder", "network"], "Method"], [5101, 5102, ["upsampling"], "Task"], [5117, 5120, ["max", "-", "pooling"], "Method"], [5121, 5124, ["sub", "-", "sampling"], "Method"], [5126, 5128, ["encoder", "network"], "Method"], [5129, 5130, ["decoders"], "Method"], [5172, 5176, ["CamVid", "road", "scenes", "dataset"], "Material"], [5182, 5184, ["decoder", "variants"], "Method"], [5197, 5199, ["RGB", "images"], "Material"], [5200, 5204, ["day", "and", "dusk", "scenes"], "Material"], [5235, 5238, ["local", "contrast", "normalization"], "Method"], [5240, 5242, ["RGB", "input"], "Material"], [5268, 5271, ["stochastic", "gradient", "descent"], "Method"], [5272, 5273, ["SGD"], "Method"], [5277, 5279, ["learning", "rate"], "Metric"], [5287, 5293, ["Caffe", "implementation", "of", "SegNet", "-", "Basic"], "Method"], [5300, 5302, ["training", "loss"], "Metric"], [5349, 5351, ["validation", "dataset"], "Material"], [5355, 5359, ["cross", "-", "entropy", "loss"], "Metric"], [5361, 5363, ["objective", "function"], "Metric"], [5410, 5412, ["CamVid", "dataset"], "Material"], [5432, 5434, ["class", "balancing"], "Task"], [5437, 5440, ["median", "frequency", "balancing"], "Method"], [5506, 5508, ["class", "balancing"], "Method"], [5511, 5514, ["natural", "frequency", "balancing"], "Method"], [5526, 5528, ["decoder", "variants"], "Method"], [5534, 5536, ["performance", "measures"], "Metric"], [5537, 5539, ["global", "accuracy"], "Metric"], [5540, 5541, ["G"], "Metric"], [5554, 5557, ["class", "average", "accuracy"], "Metric"], [5558, 5559, ["C"], "Metric"], [5565, 5567, ["predictive", "accuracy"], "Metric"], [5571, 5575, ["mean", "intersection", "over", "union"], "Metric"], [5576, 5577, ["mIoU"], "Method"], [5585, 5588, ["Pascal", "VOC12", "challenge"], "Material"], [5590, 5592, ["mIoU", "metric"], "Metric"], [5598, 5601, ["class", "average", "accuracy"], "Metric"], [5604, 5607, ["false", "positive", "predictions"], "Metric"], [5610, 5612, ["mIoU", "metric"], "Metric"], [5619, 5625, ["class", "balanced", "cross", "-", "entropy", "loss"], "Metric"], [5627, 5629, ["mIoU", "metric"], "Metric"], [5633, 5635, ["Jacard", "Index"], "Method"], [5640, 5641, ["benchmarking"], "Task"], [5666, 5667, ["segmentation"], "Task"], [5681, 5683, ["boundary", "accuracy"], "Metric"], [5694, 5695, ["FCN"], "Method"], [5702, 5704, ["mIoU", "metric"], "Method"], [5706, 5708, ["boundary", "measure"], "Method"], [5711, 5715, ["Berkeley", "contour", "matching", "score"], "Metric"], [5719, 5723, ["unsupervised", "image", "segmentation", "quality"], "Task"], [5732, 5734, ["semantic", "segmentation"], "Task"], [5740, 5743, ["semantic", "contour", "accuracy"], "Metric"], [5748, 5750, ["mIoU", "metric"], "Method"], [5753, 5758, ["human", "ranking", "of", "segmentation", "outputs"], "Metric"], [5765, 5768, ["semantic", "contour", "score"], "Metric"], [5772, 5775, ["F1", "-", "measure"], "Metric"], [5779, 5783, ["precision", "and", "recall", "values"], "Metric"], [5812, 5815, ["F1", "-", "measure"], "Method"], [5823, 5827, ["ground", "truth", "test", "image"], "Material"], [5842, 5845, ["test", "set", "average"], "Metric"], [5848, 5854, ["boundary", "F1", "-", "measure", "(", "BF"], "Metric"], [5858, 5861, ["image", "F1", "measures"], "Metric"], [5871, 5872, ["optimization"], "Task"], [5874, 5877, ["CamVid", "validation", "set"], "Material"], [5879, 5881, ["training", "loss"], "Metric"], [5914, 5916, ["global", "accuracy"], "Metric"], [5942, 5945, ["CamVid", "test", "set"], "Material"], [5949, 5951, ["class", "balancing"], "Method"], [5963, 5965, ["global", "accuracy"], "Metric"], [5970, 5972, ["smooth", "segmentation"], "Task"], [5980, 5981, ["segmentation"], "Task"], [5982, 5984, ["autonomous", "driving"], "Task"], [6015, 6017, ["global", "accuracy"], "Metric"], [6020, 6021, ["segmentation"], "Task"], [6032, 6034, ["numerical", "performance"], "Metric"], [6035, 6037, ["class", "average"], "Metric"], [6044, 6046, ["global", "accuracy"], "Metric"], [6092, 6096, ["first", "layer", "feature", "maps"], "Method"], [6097, 6100, ["max", "-", "pooling"], "Method"], [6101, 6104, ["sub", "-", "sampling"], "Method"], [6108, 6110, ["average", "time"], "Metric"], [6112, 6114, ["forward", "pass"], "Method"], [6116, 6118, ["Caffe", "implementation"], "Method"], [6127, 6130, ["NVIDIA", "Titan", "GPU"], "Method"], [6131, 6134, ["cuDNN", "v3", "acceleration"], "Method"], [6139, 6141, ["upsampling", "layers"], "Method"], [6143, 6145, ["SegNet", "variants"], "Method"], [6149, 6151, ["cuDNN", "acceleration"], "Method"], [6158, 6159, ["testing"], "Task"], [6160, 6161, ["training"], "Task"], [6183, 6187, ["training", "and", "testing", "accuracies"], "Metric"], [6194, 6196, ["class", "balancing"], "Task"], [6207, 6211, ["bilinear", "interpolation", "based", "upsampling"], "Method"], [6213, 6214, ["learning"], "Method"], [6223, 6224, ["accuracy"], "Metric"], [6232, 6233, ["learning"], "Method"], [6234, 6235, ["upsampling"], "Method"], [6236, 6241, ["FCN", "-", "Basic", "and", "variants"], "Method"], [6243, 6246, ["learning", "decoder", "filters"], "Method"], [6247, 6248, ["upsampling"], "Method"], [6249, 6252, ["SegNet", "-", "Basic"], "Method"], [6266, 6267, ["decoders"], "Method"], [6268, 6269, ["segmentation"], "Task"], [6283, 6284, ["FCN"], "Method"], [6285, 6290, ["SegNet", "-", "type", "decoding", "techniques"], "Method"], [6294, 6297, ["SegNet", "-", "Basic"], "Method"], [6298, 6299, ["FCN"], "Method"], [6316, 6317, ["accuracy"], "Metric"], [6322, 6323, ["SegNet"], "Method"], [6327, 6328, ["inference"], "Task"], [6341, 6344, ["FCN", "-", "Basic"], "Method"], [6361, 6364, ["SegNet", "-", "Basic"], "Method"], [6366, 6367, ["decoder"], "Method"], [6373, 6375, ["decoder", "layer"], "Method"], [6378, 6381, ["FCN", "-", "Basic"], "Method"], [6384, 6386, ["dimensionality", "reduction"], "Method"], [6407, 6409, ["decoder", "network"], "Method"], [6411, 6414, ["FCN", "-", "Basic"], "Method"], [6417, 6418, ["inference"], "Task"], [6428, 6430, ["decoder", "network"], "Method"], [6431, 6434, ["SegNet", "-", "Basic"], "Method"], [6441, 6444, ["FCN", "-", "Basic"], "Method"], [6455, 6457, ["training", "accuracy"], "Metric"], [6458, 6461, ["FCN", "-", "Basic"], "Method"], [6472, 6475, ["SegNet", "-", "Basic"], "Method"], [6479, 6482, ["FCN", "-", "Basic"], "Method"], [6499, 6502, ["SegNet", "-", "Basic"], "Method"], [6506, 6511, ["FCN", "-", "Basic", "-", "NoAddition"], "Method"], [6519, 6520, ["decoder"], "Method"], [6521, 6522, ["SegNet"], "Method"], [6539, 6540, ["deconvolution"], "Method"], [6542, 6547, ["FCN", "-", "Basic", "-", "NoAddition"], "Method"], [6550, 6551, ["upsampling"], "Method"], [6556, 6558, ["decoder", "filters"], "Method"], [6562, 6565, ["SegNet", "-", "Basic"], "Method"], [6574, 6576, ["decoder", "size"], "Metric"], [6578, 6579, ["accuracy"], "Metric"], [6580, 6585, ["FCN", "-", "Basic", "-", "NoAddition"], "Method"], [6591, 6594, ["FCN", "-", "Basic"], "Method"], [6624, 6626, ["BF", "measure"], "Metric"], [6641, 6644, ["SegNet", "-", "Basic"], "Method"], [6645, 6650, ["FCN", "-", "Basic", "-", "NoAddition"], "Method"], [6655, 6663, ["FCN", "-", "Basic", "-", "NoAddition", "-", "NoDimReduction", "model"], "Method"], [6667, 6670, ["SegNet", "-", "Basic"], "Method"], [6705, 6707, ["FCN", "variant"], "Method"], [6710, 6713, ["SegNet", "-", "Basic"], "Method"], [6718, 6720, ["training", "accuracy"], "Metric"], [6736, 6737, ["decoder"], "Method"], [6767, 6769, ["BF", "measure"], "Metric"], [6779, 6782, ["SegNet", "-", "Basic"], "Method"], [6785, 6787, ["training", "accuracy"], "Metric"], [6793, 6798, ["FCN", "-", "Basic", "-", "NoDimReduction"], "Method"], [6803, 6814, ["FCN", "-", "Basic", "-", "NoAddition", "and", "SegNet", "-", "Basic", "-", "SingleChannelDecoder"], "Method"], [6822, 6823, ["upsampling"], "Method"], [6838, 6839, ["SegNet"], "Method"], [6844, 6845, ["segmentation"], "Task"], [6857, 6859, ["storage", "cost"], "Metric"], [6860, 6861, ["accuracy"], "Metric"], [6862, 6864, ["inference", "time"], "Metric"], [6874, 6876, ["inference", "time"], "Metric"], [6884, 6893, ["FCN", "-", "Basic", "-", "NoDimReduction", "and", "SegNet", "-", "EncoderAddition"], "Method"], [6905, 6907, ["dimensionality", "reduction"], "Task"], [6909, 6913, ["FCN", "-", "Basic", "model"], "Method"], [6920, 6924, ["FCN", "-", "Basic", "variants"], "Method"], [6927, 6929, ["BF", "score"], "Metric"], [6940, 6941, ["memory"], "Metric"], [6942, 6943, ["accuracy"], "Metric"], [6944, 6946, ["segmentation", "architectures"], "Method"], [6989, 6992, ["class", "average", "accuracy"], "Metric"], [6993, 6995, ["mIoU", "metric"], "Metric"], [6997, 6999, ["global", "accuracy"], "Metric"], [7028, 7032, ["comparative", "analysis", "of", "variants"], "Method"], [7035, 7038, ["natural", "frequency", "balancing"], "Task"], [7045, 7047, ["BF", "measure"], "Metric"], [7048, 7051, ["SegNet", "-", "Basic"], "Method"], [7055, 7058, ["FCN", "-", "Basic"], "Method"], [7064, 7071, ["FCN", "-", "Basic", "-", "NoAddition", "-", "NoDimReduction"], "Method"], [7078, 7087, ["FCN", "-", "Basic", "-", "NoDimReduction", "and", "SegNet", "-", "EncoderAddition"], "Method"], [7126, 7130, ["semantic", "contour", "delineation", "metric"], "Metric"], [7131, 7132, ["BF"], "Method"], [7139, 7140, ["inference"], "Task"], [7151, 7153, ["dimensionality", "reduction"], "Method"], [7167, 7168, ["decoder"], "Method"], [7178, 7179, ["decoders"], "Method"], [7184, 7186, ["encoder", "network"], "Method"], [7195, 7196, ["SegNet"], "Method"], [7198, 7201, ["scene", "segmentation", "benchmarks"], "Task"], [7203, 7205, ["Caffe", "implementation"], "Method"], [7210, 7213, ["road", "scene", "segmentation"], "Task"], [7221, 7225, ["autonomous", "driving", "related", "problems"], "Task"], [7230, 7233, ["indoor", "scene", "segmentation"], "Task"], [7240, 7246, ["augmented", "reality", "(", "AR", ")", "applications"], "Task"], [7249, 7251, ["RGB", "images"], "Material"], [7258, 7259, ["SegNet"], "Method"], [7264, 7266, ["deep", "architectures"], "Method"], [7267, 7268, ["segmentation"], "Task"], [7270, 7271, ["FCN"], "Method"], [7272, 7275, ["DeepLab", "-", "LargFOV"], "Method"], [7276, 7277, ["DeconvNet"], "Method"], [7306, 7308, ["end", "training"], "Task"], [7315, 7317, ["convolutional", "layer"], "Method"], [7319, 7322, ["DeepLab", "-", "LargeFOV"], "Task"], [7336, 7338, ["predictive", "resolution"], "Metric"], [7347, 7350, ["fully", "connnected", "layers"], "Method"], [7351, 7352, ["DeconvNet"], "Method"], [7373, 7376, ["DeepLab", "-", "LargeFOV"], "Method"], [7404, 7406, ["SGD", "solver"], "Method"], [7409, 7411, ["learning", "rate"], "Metric"], [7417, 7418, ["optimization"], "Task"], [7457, 7459, ["caffe", "prototxt"], "Material"], [7463, 7465, ["road", "scenes"], "Material"], [7478, 7480, ["indoor", "scenes"], "Material"], [7493, 7496, ["Road", "Scene", "Segmentation"], "Task"], [7499, 7502, ["road", "scene", "datasets"], "Material"], [7505, 7507, ["semantic", "parsing"], "Task"], [7514, 7515, ["SegNet"], "Method"], [7517, 7519, ["CamVid", "dataset"], "Material"], [7522, 7524, ["video", "sequences"], "Material"], [7556, 7557, ["SegNet"], "Method"], [7574, 7577, ["road", "scene", "segmentation"], "Task"], [7581, 7584, ["CamVid", "test", "set"], "Material"], [7601, 7603, ["segmentation", "benchmark"], "Task"], [7604, 7606, ["road", "scenes"], "Task"], [7610, 7611, ["SegNet"], "Method"], [7687, 7689, ["Global", "avg"], "Metric"], [7690, 7691, ["mIoU"], "Metric"], [7696, 7698, ["SegNet", "predictions"], "Task"], [7700, 7702, ["deep", "architectures"], "Method"], [7727, 7729, ["road", "scenes"], "Task"], [7747, 7748, ["SegNet"], "Method"], [7760, 7763, ["DeepLab", "-", "LargeFOV"], "Method"], [7770, 7774, ["CRF", "post", "-", "processing"], "Method"], [7784, 7785, ["FCN"], "Method"], [7786, 7788, ["learnt", "deconvolution"], "Method"], [7793, 7796, ["fixed", "bilinear", "upsampling"], "Method"], [7797, 7798, ["DeconvNet"], "Method"], [7825, 7826, ["SegNet"], "Method"], [7828, 7833, ["non", "deep", "-", "learning", "methods"], "Method"], [7834, 7836, ["Random", "Forests"], "Method"], [7837, 7838, ["Boosting"], "Method"], [7841, 7844, ["CRF", "based", "methods"], "Method"], [7858, 7859, ["accuracy"], "Metric"], [7864, 7866, ["deep", "networks"], "Method"], [7869, 7873, ["feature", "engineering", "based", "techniques"], "Method"], [7882, 7883, ["SegNet"], "Method"], [7886, 7887, ["SegNet"], "Method"], [7896, 7897, ["CRFs"], "Method"], [7904, 7906, ["deep", "architecture"], "Method"], [7936, 7941, ["class", "average", "and", "mIOU", "metrics"], "Metric"], [7958, 7959, ["SegNet"], "Method"], [8003, 8007, ["median", "frequency", "class", "balancing"], "Method"], [8009, 8012, ["SegNet", "-", "Basic"], "Method"], [8013, 8014, ["SegNet"], "Method"], [8022, 8026, ["smooth", "quality", "of", "segmentation"], "Metric"], [8033, 8037, ["CRF", "post", "-", "processing"], "Method"], [8060, 8062, ["encoder", "network"], "Method"], [8079, 8081, ["random", "testing"], "Task"], [8082, 8086, ["urban", "and", "highway", "images"], "Material"], [8099, 8100, ["SegNet"], "Method"], [8110, 8112, ["unseen", "images"], "Material"], [8122, 8124, ["CRF", ")"], "Method"], [8131, 8133, ["training", "data"], "Material"], [8144, 8145, ["SegNet"], "Method"], [8151, 8154, ["fully", "convolutional", "architectures"], "Method"], [8155, 8156, ["segmentation"], "Task"], [8173, 8175, ["class", "blancing"], "Method"], [8180, 8182, ["deep", "architectures"], "Method"], [8183, 8184, ["SegNet"], "Method"], [8198, 8199, ["DeconvNet"], "Method"], [8200, 8203, ["median", "frequency", "balancing"], "Method"], [8259, 8260, ["accuracy"], "Metric"], [8302, 8304, ["training", "time"], "Metric"], [8309, 8310, ["accuracy"], "Metric"], [8381, 8382, ["SegNet"], "Method"], [8383, 8384, ["DeconvNet"], "Method"], [8398, 8399, ["DeconvNet"], "Method"], [8402, 8405, ["boundary", "delineation", "accuracy"], "Metric"], [8406, 8407, ["SegNet"], "Method"], [8414, 8415, ["DeconvNet"], "Method"], [8430, 8431, ["FCN"], "Method"], [8432, 8433, ["DeconvNet"], "Method"], [8441, 8443, ["convolutional", "layers"], "Method"], [8461, 8462, ["SegNet"], "Method"], [8496, 8498, ["FCN", "model"], "Method"], [8500, 8502, ["deconvolutional", "layers"], "Method"], [8517, 8519, ["BF", "score"], "Metric"], [8547, 8550, ["DeepLab", "-", "LargeFOV"], "Method"], [8579, 8581, ["training", "time"], "Metric"], [8591, 8593, ["boundary", "accuracy"], "Metric"], [8604, 8608, ["DeconvNet", "’s", "BF", "score"], "Metric"], [8638, 8641, ["SegNet", "type", "architecture"], "Method"], [8645, 8650, ["dense", "CRF", "post", "-", "processing"], "Method"], [8659, 8664, ["DeepLab", "-", "LargeFOV", "-", "denseCRF"], "Material"], [8666, 8667, ["global"], "Metric"], [8668, 8669, ["mIoU"], "Method"], [8671, 8674, ["class", "average", "diminshes"], "Metric"], [8683, 8685, ["BF", "score"], "Metric"], [8690, 8693, ["dense", "CRF", "hyperparameters"], "Method"], [8698, 8702, ["grid", "-", "search", "process"], "Method"], [8718, 8724, ["SUN", "RGB", "-", "D", "Indoor", "Scenes"], "Material"], [8724, 8728, ["SUN", "RGB", "-", "D"], "Material"], [8736, 8738, ["indoor", "scenes"], "Material"], [8739, 8743, ["training", "and", "testing", "images"], "Material"], [8830, 8832, ["segmentation", "challenges"], "Task"], [8860, 8862, ["depth", "images"], "Material"], [8867, 8870, ["post", "-", "processing"], "Method"], [8891, 8892, ["segmentation"], "Task"], [8899, 8900, ["segmentation"], "Task"], [8922, 8925, ["benchmark", "dataset", "NYUv2"], "Material"], [8933, 8936, ["Road", "scene", "images"], "Material"], [8982, 8984, ["deep", "networks"], "Method"], [8994, 8998, ["images", "of", "indoor", "scenes"], "Task"], [9051, 9057, ["recent", "SUN", "RGB", "-", "D", "dataset"], "Material"], [9099, 9101, ["indoor", "scenes"], "Material"], [9112, 9114, ["segmentation", "architectures"], "Method"], [9117, 9119, ["computer", "vision"], "Task"], [9125, 9130, ["Pascal", "VOC12", "salient", "object", "segmentation"], "Task"], [9138, 9141, ["indoor", "scene", "segmentation"], "Task"], [9153, 9156, ["AR", "and", "robotics"], "Task"], [9168, 9170, ["deep", "architectures"], "Method"], [9172, 9178, ["large", "SUN", "RGB", "-", "D", "dataset"], "Material"], [9183, 9184, ["SegNet"], "Method"], [9187, 9189, ["indoor", "scenes"], "Material"], [9217, 9218, ["SegNet"], "Method"], [9246, 9248, ["RGB", "images"], "Material"], [9271, 9273, ["depth", "images"], "Material"], [9311, 9313, ["AR", "tasks"], "Task"], [9321, 9323, ["segmentation", "quality"], "Metric"], [9329, 9330, ["quality"], "Metric"], [9359, 9361, ["deep", "architectures"], "Method"], [9362, 9367, ["low", "mIoU", "and", "boundary", "metrics"], "Metric"], [9369, 9373, ["global", "and", "class", "averages"], "Metric"], [9377, 9378, ["mIou"], "Method"], [9383, 9384, ["SegNet"], "Method"], [9391, 9392, ["G"], "Metric"], [9393, 9394, ["C"], "Metric"], [9395, 9397, ["BF", "metrics"], "Metric"], [9402, 9403, ["mIoU"], "Metric"], [9404, 9407, ["DeepLab", "-", "LargeFOV"], "Method"], [9415, 9416, ["SegNet"], "Method"], [9417, 9421, ["median", "frequency", "class", "balancing"], "Method"], [9450, 9455, ["grid", "search", "based", "optimal", "hyperparameters"], "Method"], [9457, 9460, ["dense", "-", "CRF"], "Method"], [9464, 9467, ["BF", "score", "metric"], "Metric"], [9468, 9473, ["DeepLab", "-", "LargeFOV", "-", "denseCRF"], "Material"], [9483, 9486, ["grid", "search", "process"], "Method"], [9492, 9494, ["inference", "time"], "Metric"], [9495, 9498, ["dense", "-", "CRFs"], "Method"], [9514, 9516, ["segmentation", "task"], "Task"], [9532, 9533, ["accuracies"], "Metric"], [9546, 9547, ["accuracy"], "Metric"], [9552, 9553, ["accuracies"], "Metric"], [9563, 9568, ["class", "distribution", "aware", "training", "techniques"], "Method"], [9581, 9583, ["deep", "architectures"], "Method"], [9589, 9591, ["VGG", "architecture"], "Method"], [9596, 9598, ["indoor", "scenes"], "Material"], [9612, 9616, ["model", "DeepLab", "-", "LargeFOV"], "Method"], [9619, 9620, ["accuracy"], "Metric"], [9623, 9624, ["mIoU"], "Metric"], [9631, 9632, ["DeconvNet"], "Method"], [9633, 9634, ["FCN"], "Method"], [9644, 9645, ["DeconvNet"], "Method"], [9678, 9681, ["Deep", "learning", "models"], "Method"], [9691, 9693, ["massive", "datasets"], "Material"], [9706, 9707, ["memory"], "Metric"], [9708, 9710, ["computational", "time"], "Metric"], [9711, 9712, ["training"], "Task"], [9713, 9714, ["testing"], "Task"], [9730, 9732, ["Training", "time"], "Metric"], [9739, 9741, ["performance", "gain"], "Metric"], [9746, 9748, ["training", "time"], "Metric"], [9754, 9757, ["Test", "time", "memory"], "Metric"], [9758, 9760, ["computational", "load"], "Metric"], [9767, 9769, ["embedded", "devices"], "Material"], [9774, 9776, ["AR", "applications"], "Task"], [9796, 9799, ["time", "efficient", "models"], "Method"], [9800, 9804, ["real", "-", "time", "applications"], "Task"], [9806, 9809, ["road", "scene", "understanding"], "Task"], [9810, 9811, ["AR"], "Task"], [9821, 9822, ["SegNet"], "Method"], [9846, 9849, ["road", "scene", "understanding"], "Task"], [9850, 9852, ["Segmentation", "challenges"], "Task"], [9854, 9855, ["Pascal"], "Task"], [9856, 9859, ["MS", "-", "COCO"], "Method"], [9860, 9863, ["object", "segmentation", "challenges"], "Task"], [9874, 9876, ["Scene", "segmentation"], "Task"], [9902, 9907, ["outdoor", "and", "indoor", "scene", "segmentation"], "Task"], [9917, 9919, ["autonomous", "driving"], "Task"], [9920, 9921, ["robotics"], "Task"], [9922, 9923, ["AR"], "Task"], [9931, 9934, ["deep", "segmentation", "architectures"], "Method"], [9941, 9942, ["BF"], "Method"], [9955, 9957, ["region", "accuracies"], "Metric"], [9969, 9972, ["outdoor", "scene", "images"], "Material"], [9982, 9984, ["deep", "architectures"], "Method"], [10002, 10006, ["indoor", "scene", "segmentation", "task"], "Task"], [10017, 10019, ["deep", "architectures"], "Method"], [10040, 10042, ["supporting", "techniques"], "Method"], [10043, 10048, ["multi", "-", "stage", "training", "recipes"], "Method"], [10052, 10053, ["accuracies"], "Metric"], [10084, 10086, ["batch", "normalization"], "Method"], [10088, 10094, ["end", "-", "to", "-", "end", "training"], "Task"], [10099, 10100, ["SGD"], "Method"], [10120, 10121, ["optimization"], "Task"], [10140, 10144, ["gradient", "back", "-", "propagation"], "Method"], [10149, 10150, ["optimization"], "Task"], [10152, 10156, ["non", "-", "convex", "problem"], "Task"], [10170, 10172, ["controlled", "analysis"], "Method"], [10201, 10203, ["segmentation", "architectures"], "Method"], [10213, 10217, ["real", "-", "time", "applications"], "Task"], [10228, 10229, ["predictions"], "Task"], [10230, 10233, ["deep", "segmentation", "architectures"], "Method"], [10240, 10241, ["SegNet"], "Method"], [10243, 10247, ["deep", "convolutional", "network", "architecture"], "Method"], [10248, 10250, ["semantic", "segmentation"], "Task"], [10255, 10256, ["SegNet"], "Method"], [10265, 10270, ["road", "and", "indoor", "scene", "understanding"], "Task"], [10277, 10281, ["memory", "and", "computational", "time"], "Metric"], [10284, 10285, ["SegNet"], "Method"], [10304, 10305, ["segmentation"], "Task"], [10307, 10309, ["training", "time"], "Metric"], [10310, 10313, ["memory", "versus", "accuracy"], "Metric"], [10332, 10334, ["inference", "time"], "Metric"], [10335, 10336, ["SegNet"], "Method"], [10361, 10363, ["decoder", "network"], "Method"], [10374, 10375, ["SegNet"], "Method"], [10382, 10385, ["road", "scene", "understanding"], "Task"], [10386, 10396, ["End", "-", "to", "-", "end", "learning", "of", "deep", "segmentation", "architectures"], "Task"], [10444, 10447, ["Machine", "Intelligence", "Laboratory"], "Task"], [10450, 10451, ["Engineering"], "Task"], [10465, 10467, ["Deep", "Learning"], "Task"], [10468, 10470, ["Magic", "Leap"], "Task"], [10483, 10486, ["probabilistic", "graphical", "models"], "Method"], [10487, 10489, ["deep", "learning"], "Method"], [10491, 10497, ["image", "and", "video", "based", "perception", "problems"], "Task"], [10552, 10555, ["Machine", "Intelligence", "Laboratory"], "Task"], [10561, 10563, ["deep", "learning"], "Task"], [10564, 10566, ["mobile", "robotics"], "Task"], [10589, 10591, ["Electrical", "Engineering"], "Task"], [10604, 10606, ["Computer", "Vision"], "Task"], [10665, 10667, ["Information", "Engineering"], "Task"], [10686, 10687, ["FREng"], "Method"], [10696, 10698, ["computer", "vision"], "Task"], [10699, 10700, ["robotics"], "Task"]]} {"doc_id": "3d597e16ef56119bb6e6d78e3f379623862f0752", "text": ["On", "gradient", "regularizers", "for", "MMD", "GANs", "section", ":", "Abstract", "We", "propose", "a", "principled", "method", "for", "gradient", "-", "based", "regularization", "of", "the", "critic", "of", "GAN", "-", "like", "models", "trained", "by", "adversarially", "optimizing", "the", "kernel", "of", "a", "Maximum", "Mean", "Discrepancy", "(", "MMD", ")", ".", "We", "show", "that", "controlling", "the", "gradient", "of", "the", "critic", "is", "vital", "to", "having", "a", "sensible", "loss", "function", ",", "and", "devise", "a", "method", "to", "enforce", "exact", ",", "analytical", "gradient", "constraints", "at", "no", "additional", "cost", "compared", "to", "existing", "approximate", "techniques", "based", "on", "additive", "regularizers", ".", "The", "new", "loss", "function", "is", "provably", "continuous", ",", "and", "experiments", "show", "that", "it", "stabilizes", "and", "accelerates", "training", ",", "giving", "image", "generation", "models", "that", "outperform", "state", "-", "of", "-", "the", "art", "methods", "on", "160", "×", "160", "CelebA", "and", "64", "×", "64", "unconditional", "ImageNet", ".", "section", ":", "however", ",", "when", "the", "MMD", "kernel", "is", "not", "based", "directly", "on", "image", "pixels", ",", "but", "on", "learned", "features", "of", "images", ".", "Wasserstein", "-", "inspired", "gradient", "regularization", "approaches", "can", "be", "used", "on", "the", "MMD", "critic", "when", "learning", "these", "features", ":", "[", "reference", "]", "uses", "weight", "clipping", "[", "reference", "]", ",", "and", "[", "reference", "][", "reference", "]", "use", "a", "gradient", "penalty", "[", "reference", "]", ".", "The", "recent", "Sobolev", "GAN", "[", "reference", "]", "uses", "a", "similar", "constraint", "on", "the", "expected", "gradient", "norm", ",", "but", "phrases", "it", "as", "estimating", "a", "Sobolev", "IPM", "rather", "than", "loosely", "approximating", "Wasserstein", ".", "This", "expectation", "can", "be", "taken", "over", "the", "same", "distribution", "as", "[", "reference", "]", ",", "but", "other", "measures", "are", "also", "proposed", ",", "such", "as", "(", "P", "+", "Q", "θ", ")", "/", "2", ".", "A", "second", "recent", "approach", ",", "the", "spectrally", "normalized", "GAN", "[", "reference", "]", ",", "controls", "the", "Lipschitz", "constant", "of", "the", "critic", "by", "enforcing", "the", "spectral", "norms", "of", "the", "weight", "matrices", "to", "be", "1", ".", "Gradient", "penalties", "also", "benefit", "GANs", "based", "on", "f", "-", "divergences", "[", "reference", "]", ":", "for", "instance", ",", "the", "spectral", "normalization", "technique", "of", "[", "reference", "]", "can", "be", "applied", "to", "the", "critic", "network", "of", "an", "f", "-", "GAN", ".", "Alternatively", ",", "a", "gradient", "penalty", "can", "be", "defined", "to", "approximate", "the", "effect", "of", "blurring", "P", "and", "Q", "θ", "with", "noise", "[", "reference", "]", ",", "which", "addresses", "the", "problem", "of", "non", "-", "overlapping", "support", "[", "reference", "]", ".", "This", "approach", "has", "recently", "been", "shown", "to", "yield", "locally", "convergent", "optimization", "in", "some", "cases", "with", "non", "-", "continuous", "distributions", ",", "where", "the", "original", "GAN", "does", "not", "[", "reference", "]", ".", "In", "this", "paper", ",", "we", "introduce", "a", "novel", "regularization", "for", "the", "MMD", "GAN", "critic", "of", "[", "reference", "][", "reference", "][", "reference", "]", ",", "which", "directly", "targets", "generator", "performance", ",", "rather", "than", "adopting", "regularization", "methods", "intended", "to", "approximate", "Wasserstein", "distances", "[", "reference", "][", "reference", "]", ".", "The", "new", "MMD", "regularizer", "derives", "from", "an", "approach", "widely", "used", "in", "semi", "-", "supervised", "learning", "[", "reference", "][", "reference", "]", ",", "where", "the", "aim", "is", "to", "define", "a", "classification", "function", "f", "which", "is", "positive", "on", "P", "(", "the", "positive", "class", ")", "and", "negative", "on", "Q", "θ", "(", "negative", "class", ")", ",", "in", "the", "absence", "of", "labels", "on", "many", "of", "the", "samples", ".", "The", "decision", "boundary", "between", "the", "classes", "is", "assumed", "to", "be", "in", "a", "region", "of", "low", "density", "for", "both", "P", "and", "Q", "θ", ":", "f", "should", "therefore", "be", "flat", "where", "P", "and", "Q", "θ", "have", "support", "(", "areas", "with", "constant", "label", ")", ",", "and", "have", "a", "larger", "slope", "in", "regions", "of", "low", "density", ".", "Bousquet", "et", "al", ".", "[", "reference", "]", "propose", "as", "their", "regularizer", "on", "f", "a", "sum", "of", "the", "variance", "and", "a", "density", "-", "weighted", "gradient", "norm", ".", "We", "adopt", "a", "related", "penalty", "on", "the", "MMD", "critic", ",", "with", "the", "difference", "that", "we", "only", "apply", "the", "penalty", "on", "P", ":", "thus", ",", "the", "critic", "is", "flatter", "where", "P", "has", "high", "mass", ",", "but", "does", "not", "vanish", "on", "the", "generator", "samples", "from", "Q", "θ", "(", "which", "we", "optimize", ")", ".", "In", "excluding", "Q", "θ", "from", "the", "critic", "function", "constraint", ",", "we", "also", "avoid", "the", "concern", "raised", "by", "[", "reference", "]", "that", "a", "critic", "depending", "on", "Q", "θ", "will", "change", "with", "the", "current", "minibatch", "-", "potentially", "leading", "to", "less", "stable", "learning", ".", "The", "resulting", "discrepancy", "is", "no", "longer", "an", "integral", "probability", "metric", ":", "it", "is", "asymmetric", ",", "and", "the", "critic", "function", "class", "depends", "on", "the", "target", "P", "being", "approximated", ".", "We", "first", "discuss", "in", "Section", "2", "how", "MMD", "-", "based", "losses", "can", "be", "used", "to", "learn", "implicit", "generative", "models", ",", "and", "how", "a", "naive", "approach", "could", "fail", ".", "This", "motivates", "our", "new", "discrepancies", ",", "introduced", "in", "Section", "3", ".", "Section", "4", "demonstrates", "that", "these", "losses", "outperform", "state", "-", "of", "-", "the", "-", "art", "models", "for", "image", "generation", ".", "section", ":", "Learning", "implicit", "generative", "models", "with", "MMD", "-", "based", "losses", "An", "IGM", "is", "a", "model", "Q", "θ", "which", "aims", "to", "approximate", "a", "target", "distribution", "P", "over", "a", "space", "X", "⊆", "R", "d", ".", "We", "will", "define", "Q", "θ", "by", "a", "generator", "function", "G", "θ", ":", "Z", "→", "X", ",", "implemented", "as", "a", "deep", "network", "with", "parameters", "θ", ",", "where", "Z", "is", "a", "space", "of", "latent", "codes", ",", "say", "R", "128", ".", "We", "assume", "a", "fixed", "distribution", "on", "Z", ",", "say", "Z", "∼", "Uniform", "[", "[", "reference", "]", "128", ",", "and", "call", "Q", "θ", "the", "distribution", "of", "G", "θ", "(", "Z", ")", ".", "We", "will", "consider", "learning", "by", "minimizing", "a", "discrepancy", "D", "between", "distributions", ",", "with", "D", "(", "P", ",", "Q", "θ", ")", "≥", "0", "and", "D", "(", "P", ",", "P", ")", "=", "0", ",", "which", "we", "call", "our", "loss", ".", "We", "aim", "to", "minimize", "D", "(", "P", ",", "Q", "θ", ")", "with", "stochastic", "gradient", "descent", "on", "an", "estimator", "of", "D.", "In", "the", "present", "work", ",", "we", "will", "build", "losses", "D", "based", "on", "the", "Maximum", "Mean", "Discrepancy", ",", "an", "integral", "probability", "metric", "where", "the", "critic", "class", "is", "the", "unit", "ball", "within", "H", "k", ",", "the", "reproducing", "kernel", "Hilbert", "space", "with", "a", "kernel", "k.", "The", "optimization", "in", "(", "1", ")", "admits", "a", "simple", "closed", "-", "form", "optimal", "critic", ",", "f", "There", "is", "also", "an", "unbiased", ",", "closed", "-", "form", "estimator", "of", "MMD", "2", "k", "with", "appealing", "statistical", "properties", "[", "reference", "]", "-", "in", "particular", ",", "its", "sample", "complexity", "is", "independent", "of", "the", "dimension", "of", "X", ",", "compared", "to", "the", "exponential", "dependence", "[", "reference", "]", "of", "the", "Wasserstein", "distance", "W", "(", "P", ",", "Q", ")", "=", "sup", "The", "MMD", "is", "continuous", "in", "the", "weak", "topology", "for", "any", "bounded", "kernel", "with", "Lipschitz", "embeddings", "[", "reference", "][", "reference", "]", "]", ",", "meaning", "that", "if", "P", "n", "converges", "in", "distribution", "to", "P", ",", "P", "n", "D", "−", "→", "P", ",", "then", "MMD", "(", "P", "n", ",", "P", ")", "→", "0", ".", "(", "W", "is", "continuous", "in", "the", "slightly", "stronger", "Wasserstein", "topology", "[", "reference", "][", "reference", "]", ";", "P", "n", "W", "−→", "P", "implies", "P", "n", "D", "−", "→", "P", ",", "and", "the", "two", "notions", "coincide", "if", "X", "is", "bounded", ".", ")", "Continuity", "means", "the", "loss", "can", "provide", "better", "signal", "to", "the", "generator", "as", "Q", "θ", "approaches", "P", ",", "as", "opposed", "to", "e.g.", "Jensen", "-", "Shannon", "where", "the", "loss", "could", "be", "constant", "until", "suddenly", "jumping", "to", "0", "[", "e.g.", "3", ",", "Example", "1", "]", ".", "The", "MMD", "is", "also", "strict", ",", "meaning", "it", "is", "zero", "iff", "P", "=", "Q", "θ", ",", "for", "characteristic", "kernels", "[", "reference", "]", ".", "The", "Gaussian", "kernel", "yields", "an", "MMD", "both", "continuous", "in", "the", "weak", "topology", "and", "strict", ".", "Thus", "in", "principle", ",", "one", "need", "not", "conduct", "any", "alternating", "optimization", "in", "an", "IGM", "at", "all", ",", "but", "merely", "choose", "generator", "parameters", "θ", "to", "minimize", "MMD", "k", ".", "Despite", "these", "appealing", "properties", ",", "using", "simple", "pixel", "-", "level", "kernels", "leads", "to", "poor", "generator", "samples", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "More", "recent", "MMD", "GANs", "[", "reference", "][", "reference", "][", "reference", "]", "achieve", "better", "results", "by", "using", "a", "parameterized", "family", "of", "kernels", ",", "{", "k", "ψ", "}", "ψ∈Ψ", ",", "in", "the", "Optimized", "MMD", "loss", "previously", "studied", "by", "[", "reference", "][", "reference", "]", ":", "We", "primarily", "consider", "kernels", "defined", "by", "some", "fixed", "kernel", "K", "on", "top", "of", "a", "learned", "low", "-", "dimensional", "representation", "In", "practice", ",", "K", "is", "a", "simple", "characteristic", "kernel", ",", "e.g.", "Gaussian", ",", "and", "φ", "ψ", "is", "usually", "a", "deep", "network", "with", "output", "dimension", "say", "s", "=", "16", "[", "reference", "]", "or", "even", "s", "=", "1", "(", "in", "our", "experiments", ")", ".", "If", "φ", "ψ", "is", "powerful", "enough", ",", "this", "choice", "is", "sufficient", ";", "we", "need", "not", "try", "to", "ensure", "each", "k", "ψ", "is", "characteristic", ",", "as", "did", "[", "reference", "]", ".", "Proposition", "1", ".", "Suppose", "k", "=", "K", "•", "φ", "ψ", ",", "with", "K", "characteristic", "and", "{", "φ", "ψ", "}", "rich", "enough", "that", "for", "any", "P", "=", "Q", ",", "there", "is", "a", "ψ", "∈", "Ψ", "for", "which", "φ", "ψ", "#", "P", "=", "φ", "ψ", "#", "Q.", "Proof", ".", "Letψ", "∈", "Ψ", "be", "such", "that", "φψ", "(", "P", ")", "=", "φψ", "(", "Q", ")", ".", "Then", ",", "since", "K", "is", "characteristic", ",", "To", "estimate", "D", "Ψ", "MMD", ",", "one", "can", "conduct", "alternating", "optimization", "to", "estimate", "aψ", "and", "then", "update", "the", "generator", "according", "to", "MMD", "kψ", ",", "similar", "to", "the", "scheme", "used", "in", "GANs", "and", "WGANs", ".", "(", "This", "form", "of", "estimator", "is", "justified", "by", "an", "envelope", "theorem", "[", "reference", "]", ",", "although", "it", "is", "invariably", "biased", "[", "reference", "]", ".", ")", "Unlike", "D", "GAN", "or", "W", ",", "fixing", "aψ", "and", "optimizing", "the", "generator", "still", "yields", "a", "sensible", "distance", "MMD", "kψ", ".", "Early", "attempts", "at", "minimizing", "D", "Ψ", "MMD", "in", "an", "IGM", ",", "though", ",", "were", "unsuccessful", "[", "reference", "]", ".", "This", "could", "be", "because", "for", "some", "kernel", "classes", ",", "D", "Ψ", "MMD", "is", "stronger", "than", "Wasserstein", "or", "MMD", ".", "Example", "1", "(", "DiracGAN", "[", "reference", "]", ")", ".", "We", "wish", "to", "model", "a", "point", "mass", "at", "the", "origin", "of", "R", ",", "P", "=", "δ", "0", ",", "with", "any", "possible", "point", "mass", ",", "Q", "θ", "=", "δ", "θ", "for", "θ", "∈", "R.", "We", "use", "a", "Gaussian", "kernel", "of", "any", "bandwidth", ",", "which", "can", "be", "written", "as", "shows", "that", "the", "Optimized", "MMD", "distance", "is", "not", "continuous", "in", "the", "weak", "or", "Wasserstein", "topologies", ".", "This", "also", "causes", "optimization", "issues", ".", "Figure", "1", "(", "a", ")", "shows", "gradient", "vector", "fields", "in", "parameter", "space", ",", ".", "Some", "sequences", "following", "v", "(", "e.g.", "A", ")", "converge", "to", "an", "optimal", "solution", "(", "0", ",", "ψ", ")", ",", "but", "some", "(", "B", ")", "move", "in", "the", "wrong", "direction", ",", "and", "others", "(", "C", ")", "are", "stuck", "because", "there", "is", "essentially", "no", "gradient", ".", "Figure", "1", "(", "c", ",", "red", ")", "shows", "that", "the", "optimal", "D", "Ψ", "MMD", "critic", "is", "very", "sharp", "near", "P", "and", "Q", ";", "this", "is", "less", "true", "for", "cases", "where", "the", "algorithm", "converged", ".", "We", "can", "avoid", "these", "issues", "if", "we", "ensure", "a", "bounded", "Lipschitz", "critic", ":", "are", "uniformly", "bounded", "and", "have", "a", "common", "Lipschitz", "constant", ":", "sup", "x∈X", ",", "ψ∈Ψ", "|f", "ψ", "(", "x", ")", "|", "<", "∞", "and", "sup", "ψ∈Ψ", "f", "ψ", "Lip", "<", "∞.", "In", "particular", ",", "this", "holds", "when", "k", "ψ", "=", "K", "•", "φ", "ψ", "and", "2", "f", "#", "P", "denotes", "the", "pushforward", "of", "a", "distribution", ":", "if", "X", "∼", "P", ",", "then", "f", "(", "X", ")", "∼", "f", "#", "P.", "[", "reference", "]", "[", "27", ",", "Theorem", "4", "]", "makes", "a", "similar", "claim", "to", "Proposition", "2", ",", "but", "its", "proof", "was", "incorrect", ":", "it", "tries", "to", "uniformly", "bound", "MMD", "k", "ψ", "≤", "W", "2", ",", "but", "the", "bound", "used", "is", "for", "a", "Wasserstein", "in", "terms", "of", "Proof", ".", "The", "main", "result", "is", "[", "reference", "][", "reference", "]", ".", "To", "show", "the", "claim", "for", "Indeed", ",", "if", "we", "put", "a", "box", "constraint", "on", "ψ", "[", "reference", "]", "or", "regularize", "the", "gradient", "of", "the", "critic", "function", "[", "reference", "]", ",", "the", "resulting", "MMD", "GAN", "generally", "matches", "or", "outperforms", "WGAN", "-", "based", "models", ".", "Unfortunately", ",", "though", ",", "an", "additive", "gradient", "penalty", "does", "n't", "substantially", "change", "the", "vector", "field", "of", "Figure", "1", "(", "a", ")", ",", "as", "shown", "in", "Figure", "5", "(", "Appendix", "B", ")", ".", "We", "will", "propose", "distances", "with", "much", "better", "convergence", "behavior", ".", "section", ":", "New", "discrepancies", "for", "learning", "implicit", "generative", "models", "Our", "aim", "here", "is", "to", "introduce", "a", "discrepancy", "that", "can", "provide", "useful", "gradient", "information", "when", "used", "as", "an", "IGM", "loss", ".", "Proofs", "of", "results", "in", "this", "section", "are", "deferred", "to", "Appendix", "A.", "section", ":", "Lipschitz", "Maximum", "Mean", "Discrepancy", "Proposition", "2", "shows", "that", "an", "MMD", "-", "like", "discrepancy", "can", "be", "continuous", "under", "the", "weak", "topology", "even", "when", "optimizing", "over", "kernels", ",", "if", "we", "directly", "restrict", "the", "critic", "functions", "to", "be", "Lipschitz", ".", "We", "can", "easily", "define", "such", "a", "distance", ",", "which", "we", "call", "the", "Lipschitz", "MMD", ":", "for", "some", "λ", ">", "0", ",", "LipMMD", "k", ",", "λ", "(", "P", ",", "Q", ")", ":", "=", "sup", "For", "a", "universal", "kernel", "k", ",", "we", "conjecture", "that", "lim", "λ→0", "LipMMD", "k", ",", "λ", "(", "P", ",", "Q", ")", "→", "W", "(", "P", ",", "Q", ")", ".", "But", "for", "any", "k", "and", "λ", ",", "LipMMD", "is", "upper", "-", "bounded", "by", "W", ",", "as", "(", "4", ")", "optimizes", "over", "a", "smaller", "set", "of", "functions", "than", "[", "reference", "]", ".", "LipMMD", "(", "P", ",", "Q", ")", ":", "=", "sup", "ψ∈Ψ", "LipMMD", "k", "ψ", ",", "λ", "(", "P", ",", "Q", ")", "is", "also", "upper", "-", "bounded", "by", "W", ",", "and", "hence", "is", "continuous", "in", "the", "Wasserstein", "topology", ".", "It", "also", "shows", "excellent", "empirical", "behavior", "on", "Example", "1", "(", "Figure", "1", "Constraining", "the", "mean", "gradient", "rather", "than", "the", "maximum", ",", "as", "we", "will", "do", "next", ",", "is", "far", "more", "tractable", ".", "section", ":", "Gradient", "-", "Constrained", "Maximum", "Mean", "Discrepancy", "We", "define", "the", "Gradient", "-", "Constrained", "MMD", "for", "λ", ">", "0", "and", "using", "some", "measure", "µ", "as", "where", "Rather", "than", "directly", "constraining", "the", "Lipschitz", "constant", ",", "the", "second", "term", "∇f", "2", "L", "2", "(", "µ", ")", "encourages", "the", "function", "f", "to", "be", "flat", "where", "µ", "has", "mass", ".", "In", "experiments", "we", "use", "µ", "=", "P", ",", "flattening", "the", "critic", "near", "the", "target", "sample", ".", "We", "add", "the", "first", "term", "following", "[", "reference", "]", ":", "in", "one", "dimension", "and", "with", "µ", "uniform", ",", "·", "S", "(", "µ", "),", "·", ",", "0", "is", "then", "an", "RKHS", "norm", "with", "the", "kernel", "κ", "(", "x", ",", "y", ")", "=", "exp", "(", "−", "x", "−", "y", ")", ",", "which", "is", "also", "a", "Sobolev", "space", ".", "The", "correspondence", "to", "a", "Sobolev", "norm", "is", "lost", "in", "higher", "dimensions", "[", "reference", "][", "reference", "]", "]", ",", "but", "we", "also", "found", "the", "first", "term", "to", "be", "beneficial", "in", "practice", ".", "We", "can", "exploit", "some", "properties", "of", "H", "k", "to", "compute", "(", "5", ")", "analytically", ".", "Call", "the", "difference", "in", "kernel", "mean", "M", "with", "mth", "entry", "η", "(", "X", "m", ")", ",", "and", "∇η", "(", "X", ")", "∈", "where", "K", "is", "the", "kernel", "matrix", ",", "and", "H", "that", "of", "derivatives", "of", "both", "arguments", "As", "long", "as", "P", "and", "Q", "have", "integrable", "first", "moments", ",", "and", "µ", "has", "second", "moments", ",", "Assumptions", "(", "A", ")", "to", "(", "D", ")", "are", "satisfied", "e.g.", "by", "a", "Gaussian", "or", "linear", "kernel", "on", "top", "of", "a", "differentiable", "φ", "ψ", ".", "We", "can", "thus", "estimate", "the", "GCMMD", "based", "on", "samples", "from", "P", ",", "Q", ",", "and", "µ", "by", "using", "the", "empirical", "meanη", "for", "η", ".", "This", "discrepancy", "indeed", "works", "well", "in", "practice", ":", "Appendix", "F.2", "shows", "that", "optimizing", "our", "estimate", "of", "D", "µ", ",", "Ψ", ",", "λ", "GCMMD", "=", "sup", "ψ∈Ψ", "GCMMD", "µ", ",", "k", "ψ", ",", "λ", "yields", "a", "good", "generative", "model", "on", "MNIST", ".", "But", "the", "linear", "system", "of", "size", "M", "+", "M", "d", "is", "impractical", ":", "even", "on", "28", "×", "28", "images", "and", "using", "a", "low", "-", "rank", "approximation", ",", "the", "model", "took", "days", "to", "converge", ".", "We", "therefore", "design", "a", "less", "expensive", "discrepancy", "in", "the", "next", "section", ".", "The", "GCMMD", "is", "related", "to", "some", "discrepancies", "previously", "used", "in", "IGM", "training", ".", "The", "Fisher", "GAN", "[", "reference", "]", "uses", "only", "the", "variance", "constraint", "f", "along", "with", "a", "vanishing", "boundary", "condition", "on", "f", "to", "ensure", "a", "well", "-", "defined", "solution", "(", "although", "this", "was", "not", "used", "in", "the", "implementation", ",", "and", "can", "cause", "very", "unintuitive", "critic", "behavior", ";", "see", "Appendix", "C", ")", ".", "The", "authors", "considered", "several", "choices", "of", "µ", ",", "including", "the", "WGAN", "-", "GP", "measure", "[", "reference", "]", "and", "mixtures", "(", "P", "+", "Q", "θ", ")", "/", "2", ".", "Rather", "than", "enforcing", "the", "constraints", "in", "closed", "form", "as", "we", "do", ",", "though", ",", "these", "models", "used", "additive", "regularization", ".", "We", "will", "compare", "to", "the", "Sobolev", "GAN", "in", "experiments", ".", "section", ":", "Scaled", "Maximum", "Mean", "Discrepancy", "We", "will", "now", "derive", "a", "lower", "bound", "on", "the", "Gradient", "-", "Constrained", "MMD", "which", "retains", "many", "of", "its", "attractive", "qualities", "but", "can", "be", "estimated", "in", "time", "linear", "in", "the", "dimension", "d.", "We", "then", "define", "the", "Scaled", "Maximum", "Mean", "Discrepancy", "based", "on", "this", "bound", "of", "Proposition", "4", ":", "4", "We", "use", "(", "m", ",", "i", ")", "to", "denote", "(", "m", "−", "1", ")", "d", "+", "i", ";", "thus", "∇η", "(", "X", ")", "stacks", "∇η", "(", "X1", ")", ",", ".", ".", ".", ",", "∇η", "(", "XM", ")", "into", "one", "vector", ".", "[", "reference", "]", "We", "use", "∂ik", "(", "x", ",", "y", ")", "to", "denote", "the", "partial", "derivative", "with", "respect", "to", "xi", ",", "and", "∂", "i", "+", "d", "k", "(", "x", ",", "y", ")", "that", "for", "yi", ".", "Because", "the", "constraint", "in", "the", "optimization", "of", "(", "7", ")", "is", "more", "restrictive", "than", "in", "that", "of", "(", "5", ")", ",", "we", "have", "that", "SMMD", "µ", ",", "k", ",", "λ", "(", "P", ",", "Q", ")", "≤", "GCMMD", "µ", ",", "k", ",", "λ", "(", "P", ",", "Q", ")", ".", "The", "Sobolev", "norm", "f", "S", "(", "µ", "),", "λ", ",", "and", "a", "fortiori", "the", "gradient", "norm", "under", "µ", ",", "is", "thus", "also", "controlled", "for", "the", "SMMD", "critic", ".", "We", "also", "show", "in", "Appendix", "F.1", "that", "SMMD", "µ", ",", "k", ",", "λ", "behaves", "similarly", "to", "GCMMD", "µ", ",", "k", ",", "λ", "on", "Gaussians", ".", "F", ".", "Estimating", "these", "terms", "based", "on", "samples", "from", "µ", "is", "straightforward", ",", "giving", "a", "natural", "estimator", "for", "the", "SMMD", ".", "Of", "course", ",", "if", "µ", "and", "k", "are", "fixed", ",", "the", "SMMD", "is", "simply", "a", "constant", "times", "the", "MMD", ",", "and", "so", "behaves", "in", "essentially", "the", "same", "way", "as", "the", "MMD", ".", "But", "optimizing", "the", "SMMD", "over", "a", "kernel", "family", "Ψ", ",", "D", "µ", ",", "Ψ", ",", "λ", "Figure", "1", "(", "b", ")", "shows", "the", "vector", "field", "for", "the", "Optimized", "SMMD", "loss", "in", "Example", "1", ",", "using", "the", "WGAN", "-", "GP", "measure", "µ", "=", "Uniform", "(", "0", ",", "θ", ")", ".", "The", "optimization", "surface", "is", "far", "more", "amenable", ":", "in", "particular", "the", "location", "C", ",", "which", "formerly", "had", "an", "extremely", "small", "gradient", "that", "made", "learning", "effectively", "impossible", ",", "now", "converges", "very", "quickly", "by", "first", "reducing", "the", "critic", "gradient", "until", "some", "signal", "is", "available", ".", "Figure", "1", "s", "a", "fully", "-", "connected", "L", "-", "layer", "network", "with", "Leaky", "-", "ReLU", "α", "activations", "whose", "layers", "do", "not", "increase", "in", "width", ",", "and", "K", "satisfying", "mild", "smoothness", "conditions", "Q", "K", "<", "∞", "(", "Assumptions", "(", "II", ")", "to", "(", "V", ")", "in", "Appendix", "A.2", ")", ".", "Let", "Ψ", "κ", "be", "the", "set", "of", "parameters", "where", "each", "layer", "'s", "weight", "matrices", "have", "condition", "number", "cond", "(", ".", "SMMD", "(", "P", "n", ",", "P", ")", "→", "0", ",", "even", "if", "µ", "is", "chosen", "to", "depend", "on", "P", "and", "Q.", "section", ":", "Uniform", "bounds", "vs", "bounds", "in", "expectation", "Controlling", "not", "necessarily", "imply", "a", "bound", "on", "f", "Lip", "≥", "sup", "x∈X", "∇f", "ψ", "(", "X", ")", ",", "and", "so", "does", "not", "in", "general", "give", "continuity", "via", "Proposition", "2", ".", "Theorem", "1", "implies", "that", "when", "the", "network", "'s", "weights", "are", "well", "-", "conditioned", ",", "it", "is", "sufficient", "to", "only", "control", "∇f", "ψ", "2", "L", "2", "(", "µ", ")", ",", "which", "is", "far", "easier", "in", "practice", "than", "controlling", "f", "Lip", ".", "If", "we", "instead", "tried", "to", "directly", "controlled", "f", "Lip", "with", "e.g.", "spectral", "normalization", "(", "SN", ")", "[", "reference", "]", ",", "we", "could", "significantly", "reduce", "the", "expressiveness", "of", "the", "parametric", "family", ".", "In", "Example", "1", ",", "constraining", "φ", "ψ", "Lip", "=", "1", "limits", "us", "to", "only", "Ψ", "=", "{", "1}.", "Thus", "D", "section", ":", "{", "1", "}", "MMD", "is", "simply", "the", "MMD", "with", "an", "RBF", "kernel", "of", "bandwidth", "1", ",", "which", "has", "poor", "gradients", "when", "θ", "is", "far", "from", "0", "(", "Figure", "1", "(", "c", ")", ",", "blue", ")", ".", "The", "CauchySchwartz", "bound", "of", "Proposition", "4", "allows", "jointly", "adjusting", "the", "smoothness", "of", "k", "ψ", "and", "the", "critic", "f", ",", "while", "SN", "must", "control", "the", "two", "independently", ".", "Relatedly", ",", "limiting", "φ", "Lip", "by", "limiting", "the", "Lipschitz", "norm", "of", "each", "layer", "could", "substantially", "reduce", "capacity", ",", "while", "∇f", "ψ", "L", "2", "(", "µ", ")", "need", "not", "be", "decomposed", "by", "layer", ".", "Another", "advantage", "is", "that", "µ", "provides", "a", "data", "-", "dependent", "measure", "of", "complexity", "as", "in", "[", "reference", "]", ":", "we", "do", "not", "needlessly", "prevent", "ourselves", "from", "using", "critics", "that", "behave", "poorly", "only", "far", "from", "the", "data", ".", "Spectral", "parametrization", "When", "the", "generator", "is", "near", "a", "local", "optimum", ",", "the", "critic", "might", "identify", "only", "one", "direction", "on", "which", "Q", "θ", "and", "P", "differ", ".", "If", "the", "generator", "parameterization", "is", "such", "that", "there", "is", "no", "local", "way", "for", "the", "generator", "to", "correct", "it", ",", "the", "critic", "may", "begin", "to", "single", "-", "mindedly", "focus", "on", "this", "difference", ",", "choosing", "redundant", "convolutional", "filters", "and", "causing", "the", "condition", "number", "of", "the", "weights", "to", "diverge", ".", "If", "this", "occurs", ",", "the", "generator", "will", "be", "motivated", "to", "fix", "this", "single", "direction", "while", "ignoring", "all", "other", "aspects", "of", "the", "distributions", ",", "after", "which", "it", "may", "become", "stuck", ".", "We", "can", "help", "avoid", "this", "collapse", "by", "using", "a", "critic", "parameterization", "that", "encourages", "diverse", "filters", "with", "higher", "-", "rank", "weight", "matrices", ".", "Miyato", "et", "al", ".", "[", "reference", "]", "propose", "to", "parameterize", "the", "weight", "matrices", "as", "W", "=", "γW", "/", "W", "op", ",", "where", "W", "op", "is", "the", "spectral", "norm", "ofW", ".", "This", "parametrization", "works", "particularly", "well", "with", "D", "section", ":", "Experiments", "We", "evaluated", "unsupervised", "image", "generation", "on", "three", "datasets", ":", "CIFAR", "-", "10", "[", "reference", "]", "(", "60", "000", "images", ",", "32", "×", "32", ")", ",", "CelebA", "[", "reference", "]", "(", "202", "599", "face", "images", ",", "resized", "and", "cropped", "to", "160", "×", "160", "as", "in", "[", "reference", "]", ")", ",", "and", "the", "more", "challenging", "ILSVRC2012", "(", "ImageNet", ")", "dataset", "[", "reference", "]", "(", "1", "281", "167", "images", ",", "resized", "to", "64", "×", "64", ")", ".", "Code", "for", "all", "of", "these", "experiments", "is", "available", "at", "github.com", "/", "MichaelArbel", "/", "Scaled", "-", "MMD", "-", "GAN", ".", "Losses", "All", "models", "are", "based", "on", "a", "scalar", "-", "output", "critic", "network", "φ", "ψ", ":", "X", "→", "R", ",", "except", "MMDGAN", "-", "GP", "where", "φ", "ψ", ":", "X", "→", "R", "16", "as", "in", "[", "reference", "]", ".", "The", "WGAN", "and", "Sobolev", "GAN", "use", "a", "critic", "f", "=", "φ", "ψ", ",", "while", "the", "GAN", "uses", "a", "discriminator", "D", "ψ", "(", "x", ")", "=", "1", "/(", "1", "+", "exp", "(", "−φ", "ψ", "(", "x", ")", ")", ")", ".", "The", "MMD", "-", "based", "methods", "use", "a", "kernel", "2", "/", "2", ")", ",", "except", "for", "MMDGAN", "-", "GP", "which", "uses", "a", "mixture", "of", "RQ", "kernels", "as", "in", "[", "reference", "]", ".", "Increasing", "the", "output", "dimension", "of", "the", "critic", "or", "using", "a", "different", "kernel", "did", "n't", "substantially", "change", "the", "performance", "of", "our", "proposed", "method", ".", "We", "also", "consider", "SMMD", "with", "a", "linear", "top", "-", "level", "kernel", ",", "k", "(", "x", ",", "y", ")", "=", "φ", "ψ", "(", "x", ")", "φ", "ψ", "(", "y", ")", ";", "because", "this", "becomes", "essentially", "identical", "to", "a", "WGAN", "(", "Appendix", "E", ")", ",", "we", "refer", "to", "this", "method", "as", "SWGAN", ".", "SMMD", "and", "SWGAN", "use", "µ", "=", "P", ";", "Sobolev", "GAN", "uses", "µ", "=", "(", "P", "+", "Q", ")/", "2", "as", "in", "[", "reference", "]", ".", "We", "choose", "λ", "and", "an", "overall", "scaling", "to", "obtain", "the", "losses", ":", ",", "SWGAN", ":", ".", "Architecture", "For", "CIFAR", "-", "10", ",", "we", "used", "the", "CNN", "architecture", "proposed", "by", "[", "reference", "]", "with", "a", "7", "-", "layer", "critic", "and", "a", "4", "-", "layer", "generator", ".", "For", "CelebA", ",", "we", "used", "a", "5", "-", "layer", "DCGAN", "discriminator", "and", "a", "10", "-", "layer", "ResNet", "generator", "as", "in", "[", "reference", "]", ".", "For", "ImageNet", ",", "we", "used", "a", "10", "-", "layer", "ResNet", "for", "both", "the", "generator", "and", "discriminator", ".", "In", "all", "experiments", "we", "used", "64", "filters", "for", "the", "smallest", "convolutional", "layer", ",", "and", "double", "it", "at", "each", "layer", "(", "CelebA", "/", "ImageNet", ")", "or", "every", "other", "layer", "(", "CIFAR", "-", "10", ")", ".", "The", "input", "codes", "for", "the", "generator", "are", "drawn", "from", "Uniform", "[", "−1", ",", "1", "]", "128", ".", "We", "consider", "two", "parameterizations", "for", "each", "critic", ":", "a", "standard", "one", "where", "the", "parameters", "can", "take", "any", "real", "value", ",", "and", "a", "spectral", "parametrization", "(", "denoted", "SN", "-", ")", "as", "above", "[", "reference", "]", ".", "Models", "without", "explicit", "gradient", "control", "(", "SN", "-", "GAN", ",", "SN", "-", "MMDGAN", ",", "SN", "-", "MMGAN", "-", "L2", ",", "SN", "-", "WGAN", ")", "fix", "γ", "=", "1", ",", "for", "spectral", "normalization", ";", "others", "learn", "γ", ",", "using", "a", "spectral", "parameterization", ".", "Training", "All", "models", "were", "trained", "for", "150", "000", "generator", "updates", "on", "a", "single", "GPU", ",", "except", "for", "ImageNet", "where", "the", "model", "was", "trained", "on", "3", "GPUs", "simultaneously", ".", "To", "limit", "communication", "overhead", "we", "averaged", "the", "MMD", "estimate", "on", "each", "GPU", ",", "giving", "the", "block", "MMD", "estimator", "[", "reference", "]", ".", "We", "always", "used", "64", "samples", "per", "GPU", "from", "each", "of", "P", "and", "Q", ",", "and", "5", "critic", "updates", "per", "generator", "step", ".", "We", "used", "initial", "learning", "rates", "of", "0.0001", "for", "CIFAR", "-", "10", "and", "CelebA", ",", "0.0002", "for", "ImageNet", ",", "and", "decayed", "these", "rates", "using", "the", "KID", "adaptive", "scheme", "of", "[", "reference", "]", ":", "every", "2", "000", "steps", ",", "generator", "samples", "are", "compared", "to", "those", "from", "20", "000", "steps", "ago", ",", "and", "if", "the", "relative", "KID", "test", "[", "reference", "]", "fails", "to", "show", "an", "improvement", "three", "consecutive", "times", ",", "the", "learning", "rate", "is", "decayed", "by", "0.8", ".", "We", "used", "the", "Adam", "optimizer", "[", "reference", "]", "with", "β", "1", "=", "0.5", ",", "β", "2", "=", "0.9", ".", "Evaluation", "To", "compare", "the", "sample", "quality", "of", "different", "models", ",", "we", "considered", "three", "different", "scores", "based", "on", "the", "Inception", "network", "[", "reference", "]", "trained", "for", "ImageNet", "classification", ",", "all", "using", "default", "parameters", "in", "the", "implementation", "of", "[", "reference", "]", ".", "The", "Inception", "Score", "(", "IS", ")", "[", "reference", "]", "is", "based", "on", "the", "entropy", "of", "predicted", "labels", ";", "higher", "values", "are", "better", ".", "Though", "standard", ",", "this", "metric", "has", "many", "issues", ",", "particularly", "on", "datasets", "other", "than", "ImageNet", "[", "reference", "][", "reference", "][", "reference", "]", ".", "The", "FID", "[", "reference", "]", "instead", "measures", "the", "similarity", "of", "samples", "from", "the", "generator", "and", "the", "target", "as", "the", "Wasserstein", "-", "2", "distance", "between", "Gaussians", "fit", "to", "their", "intermediate", "representations", ".", "It", "is", "more", "sensible", "than", "the", "IS", "and", "becoming", "standard", ",", "but", "its", "estimator", "is", "strongly", "biased", "[", "reference", "]", ".", "The", "KID", "[", "reference", "]", "is", "similar", "to", "FID", ",", "but", "by", "using", "a", "polynomial", "-", "kernel", "MMD", "its", "estimates", "enjoy", "better", "statistical", "properties", "and", "are", "easier", "to", "compare", ".", "(", "A", "similar", "score", "was", "recommended", "by", "[", "reference", "]", ".", ")", "Results", "Table", "1a", "presents", "the", "scores", "for", "models", "trained", "on", "both", "CIFAR", "-", "10", "and", "CelebA", "datasets", ".", "On", "CIFAR", "-", "10", ",", "SN", "-", "SWGAN", "and", "SN", "-", "SMMDGAN", "performed", "comparably", "to", "SN", "-", "GAN", ".", "But", "on", "CelebA", ",", "SN", "-", "SWGAN", "and", "SN", "-", "SMMDGAN", "dramatically", "outperformed", "the", "other", "methods", "with", "the", "same", "architecture", "in", "all", "three", "metrics", ".", "It", "also", "trained", "faster", ",", "and", "consistently", "outperformed", "other", "methods", "over", "multiple", "initializations", "(", "Figure", "2", "(", "a", ")", ")", ".", "It", "is", "worth", "noting", "that", "SN", "-", "SWGAN", "far", "outperformed", "WGAN", "-", "GP", "on", "both", "datasets", ".", "Table", "1b", "presents", "the", "scores", "for", "SMMDGAN", "and", "SN", "-", "SMMDGAN", "trained", "on", "ImageNet", ",", "and", "the", "scores", "of", "pre", "-", "trained", "models", "using", "BGAN", "[", "reference", "]", "and", "SN", "-", "GAN", "[", "reference", "]", ".", "[", "reference", "]", "The", "proposed", "methods", "substantially", "outperformed", "both", "methods", "in", "FID", "and", "KID", "scores", ".", "Figure", "3", "shows", "samples", "on", "ImageNet", "and", "CelebA", ";", "Appendix", "F.4", "has", "more", ".", "Spectrally", "normalized", "WGANs", "/", "MMDGANs", "To", "control", "for", "the", "contribution", "of", "the", "spectral", "parametrization", "to", "the", "performance", ",", "we", "evaluated", "variants", "of", "MMDGANs", ",", "WGANs", "and", "Sobolev", "-", "GAN", "using", "spectral", "normalization", "(", "in", "Table", "2", ",", "Appendix", "F.3", ")", ".", "WGAN", "and", "Sobolev", "-", "GAN", "led", "to", "unstable", "training", "and", "did", "n't", "converge", "at", "all", "(", "Figure", "11", ")", "despite", "many", "attempts", ".", "MMDGAN", "converged", "on", "CIFAR", "-", "10", "(", "Figure", "11", ")", "but", "was", "unstable", "on", "CelebA", "(", "Figure", "10", ")", ".", "The", "gradient", "control", "due", "to", "SN", "is", "thus", "probably", "too", "loose", "for", "these", "methods", ".", "This", "is", "reinforced", "by", "Figure", "2", "(", "c", ")", ",", "which", "shows", "that", "the", "expected", "gradient", "of", "the", "critic", "network", "is", "much", "better", "-", "controlled", "by", "SMMD", ",", "even", "when", "SN", "is", "used", ".", "We", "also", "considered", "variants", "of", "these", "models", "with", "a", "learned", "γ", "while", "also", "adding", "a", "gradient", "penalty", "and", "an", "L", "2", "penalty", "on", "critic", "activations", "[", "reference", "]", ".", "These", "generally", "behaved", "similarly", "to", "MMDGAN", ",", "and", "did", "n't", "lead", "to", "substantial", "improvements", ".", "We", "ran", "the", "same", "experiments", "on", "CelebA", ",", "but", "aborted", "the", "runs", "early", "when", "it", "became", "clear", "that", "training", "was", "not", "successful", ".", "Rank", "collapse", "We", "occasionally", "observed", "the", "failure", "mode", "for", "SMMD", "where", "the", "critic", "becomes", "low", "-", "rank", ",", "discussed", "in", "Section", "3.3", ",", "especially", "on", "CelebA", ";", "this", "failure", "was", "obvious", "even", "in", "the", "training", "objective", ".", "Figure", "2", "(", "b", ")", "is", "one", "of", "these", "examples", ".", "Spectral", "parametrization", "seemed", "to", "prevent", "this", "behavior", ".", "We", "also", "found", "one", "could", "avoid", "collapse", "by", "reverting", "to", "an", "earlier", "checkpoint", "and", "increasing", "the", "RKHS", "regularization", "parameter", "λ", ",", "but", "did", "not", "do", "this", "for", "any", "of", "the", "experiments", "here", ".", "section", ":", "Conclusion", "We", "studied", "gradient", "regularization", "for", "MMD", "-", "based", "critics", "in", "implicit", "generative", "models", ",", "clarifying", "how", "previous", "techniques", "relate", "to", "the", "D", "Ψ", "MMD", "loss", ".", "Based", "on", "these", "insights", ",", "we", "proposed", "the", "GradientConstrained", "MMD", "and", "its", "approximation", "the", "Scaled", "MMD", ",", "a", "new", "loss", "function", "for", "IGMs", "that", "controls", "gradient", "behavior", "in", "a", "principled", "way", "and", "obtains", "excellent", "performance", "in", "practice", ".", "One", "interesting", "area", "of", "future", "study", "for", "these", "distances", "is", "their", "behavior", "when", "used", "to", "diffuse", "particles", "distributed", "as", "Q", "towards", "particles", "distributed", "as", "P.", "Mroueh", "et", "al", ".", "[", "reference", "][", "reference", "]", "began", "such", "a", "study", "for", "the", "Sobolev", "GAN", "loss", ";", "[", "reference", "]", "proved", "convergence", "and", "studied", "discrete", "-", "time", "approximations", ".", "Another", "area", "to", "explore", "is", "the", "geometry", "of", "these", "losses", ",", "as", "studied", "by", "Bottou", "et", "al", ".", "[", "reference", "]", ",", "who", "showed", "potential", "advantages", "of", "the", "Wasserstein", "geometry", "over", "the", "MMD", ".", "Their", "results", ",", "though", ",", "do", "not", "address", "any", "distances", "based", "on", "optimized", "kernels", ";", "the", "new", "distances", "introduced", "here", "might", "have", "interesting", "geometry", "of", "their", "own", ".", "section", ":", "9", "section", ":", "A", "Proofs", "We", "first", "review", "some", "basic", "properties", "of", "Reproducing", "Kernel", "Hilbert", "Spaces", ".", "We", "consider", "here", "a", "separable", "RKHS", "H", "with", "basis", "(", "e", "i", ")", "i∈I", ",", "where", "I", "is", "either", "finite", "if", "H", "is", "finite", "-", "dimensional", ",", "or", "I", "=", "N", "otherwise", ".", "We", "also", "assume", "that", "the", "reproducing", "kernel", "k", "is", "continuously", "twice", "differentiable", ".", "We", "use", "a", "slightly", "nonstandard", "notation", "for", "derivatives", ":", "∂", "i", "f", "(", "x", ")", "denotes", "the", "ith", "partial", "derivative", "of", "f", "evaluated", "at", "x", ",", "and", "∂", "i", "∂", "j", "+", "d", "k", "(", "x", ",", "y", ")", "denotes", "We", "say", "that", "an", "operator", "A", ":", "H", "→", "H", "is", "Hilbert", "-", "Schmidt", "if", "A", "2", "HS", "=", "i∈I", "Ae", "i", "2", "H", "is", "finite", ".", "A", "HS", "is", "called", "the", "Hilbert", "-", "Schmidt", "norm", "of", "A.", "The", "space", "of", "Hilbert", "-", "Schmidt", "operators", "itself", "a", "Hilbert", "space", "with", "the", "inner", "product", "A", ",", "B", "HS", "=", "i∈I", "Ae", "i", ",", "Be", "i", "H", ".", "Moreover", ",", "we", "say", "that", "an", "operator", "A", "is", "trace", "-", "class", "if", "its", "trace", "norm", "is", "finite", ",", "i.e.", "A", "1", "=", "i∈I", "e", "i", ",", "(", "A", "*", "A", ")", "Given", "two", "vectors", "f", "and", "g", "in", "H", "and", "a", "Hilbert", "-", "Schmidt", "operator", "A", "we", "have", "the", "following", "properties", ":", "(", "i", ")", "The", "outer", "product", "f", "⊗", "g", "is", "a", "Hilbert", "-", "Schmidt", "operator", "with", "Hilbert", "-", "Schmidt", "norm", "given", "by", ":", "ii", ")", "The", "inner", "product", "between", "two", "rank", "-", "one", "operators", "f", "⊗", "g", "and", "u", "The", "following", "identity", "holds", ":", "f", ",", "Ag", "H", "=", "f", "⊗", "g", ",", "A", "HS", ".", "Define", "the", "following", "covariance", "-", "type", "operators", ":", "[", "reference", "]", "these", "are", "useful", "in", "that", ",", "using", "[", "reference", "]", "and", "[", "reference", "]", "section", ":", "A.1", "Definitions", "and", "estimators", "of", "the", "new", "distances", "We", "will", "need", "the", "following", "assumptions", "about", "the", "distributions", "P", "and", "Q", ",", "the", "measure", "µ", ",", "and", "the", "kernel", "k", ":", "(", "A", ")", "P", "and", "Q", "have", "integrable", "first", "moments", ".", "(", "B", ")", "k", "(", "x", ",", "x", ")", "grows", "at", "most", "linearly", "in", "x", ":", "for", "all", "x", "in", "X", ",", "k", "(", "x", ",", "x", ")", "≤", "C", "(", "x", "+", "1", ")", "for", "some", "constant", "C.", "is", "automatically", "satisfied", "by", "a", "K", "such", "as", "the", "Gaussian", ";", "when", "K", "is", "linear", ",", "it", "is", "true", "for", "a", "quite", "general", "class", "of", "networks", "φ", "ψ", "[", "7", ",", "Lemma", "1", "]", ".", "We", "will", "first", "give", "a", "form", "for", "the", "Gradient", "-", "Constrained", "MMD", "(", "5", ")", "in", "terms", "of", "the", "operator", "(", "10", ")", ":", "Proposition", "5", ".", "Under", "Assumptions", "(", "A", ")", "to", "(", "D", ")", ",", "the", "Gradient", "-", "Constrained", "MMD", "is", "given", "by", "Proof", "of", "Proposition", "5", ".", "Let", "f", "be", "a", "function", "in", "H.", "We", "will", "first", "express", "the", "squared", "λ", "-", "regularized", "Sobolev", "norm", "of", "f", "(", "6", ")", "as", "a", "quadratic", "form", "in", "H.", "Recalling", "the", "reproducing", "properties", "of", "(", "8", ")", "and", "(", "9", ")", ",", "we", "have", ":", "Using", "Property", "(", "ii", ")", "and", "the", "operator", "(", "10", ")", ",", "one", "further", "gets", "Under", "Assumption", "(", "D", ")", ",", "and", "using", "Lemma", "6", ",", "one", "can", "take", "the", "integral", "inside", "the", "inner", "product", ",", "which", "leads", "to", "f", "H", ".", "Finally", ",", "using", "Property", "(", "iii", ")", "it", "follows", "that", "Under", "Assumptions", "(", "A", ")", "and", "(", "B", ")", ",", "Lemma", "6", "applies", ",", "and", "it", "follows", "that", "k", "(", "x", ",", "·", ")", "is", "also", "Bochner", "integrable", "under", "P", "and", "Q.", "Thus", "where", "η", "is", "defined", "as", "this", "difference", "in", "mean", "embeddings", ".", "Since", "D", "µ", ",", "λ", "is", "symmetric", "positive", "definite", ",", "its", "square", "-", "root", "D", "µ", ",", "λ", "is", "well", "-", "defined", "and", "is", "also", "invertible", ".", "µ", ",", "λ", "g.", "Thus", "we", "can", "re", "-", "express", "the", "maximization", "problem", "in", "(", "5", ")", "in", "terms", "of", "g", ":", "Proposition", "5", ",", "though", ",", "involves", "inverting", "the", "infinite", "-", "dimensional", "operator", "D", "µ", ",", "λ", "and", "thus", "does", "n't", "directly", "give", "us", "a", "computable", "estimator", ".", "Proposition", "3", "solves", "this", "problem", "in", "the", "case", "where", "µ", "is", "a", "discrete", "measure", ":", "δ", "Xm", "be", "an", "empirical", "measure", "of", "M", "points", ".", "Let", "η", "(", "X", ")", "∈", "R", "M", "have", "mth", "entry", "η", "(", "X", "m", ")", ",", "and", "∇η", "(", "X", ")", "∈", "R", "M", "d", "have", "(", "m", ",", "i", ")", "th", "entry", "7", "∂", "i", "η", "(", "X", "m", ")", ".", "Then", "under", "Assumptions", "(", "A", ")", "to", "(", "D", ")", ",", "the", "Gradient", "-", "Constrained", "MMD", "is", "where", "K", "is", "the", "kernel", "matrix", ",", "and", "H", "that", "of", "derivatives", "of", "both", "arguments", "Before", "proving", "Proposition", "3", ",", "we", "note", "the", "following", "interesting", "alternate", "form", ".", "Letē", "i", "be", "the", "ith", "standard", "basis", "vector", "for", "R", "M", "+", "M", "d", ",", "and", "define", "T", ":", "H", "→", "R", "M", "+", "M", "d", "as", "the", "linear", "operator", "Thus", "we", "can", "write", "7", "We", "use", "(", "m", ",", "i", ")", "to", "denote", "(", "m", "−", "1", ")", "d", "+", "i", ";", "thus", "∇η", "(", "X", ")", "stacks", "∇η", "(", "X1", ")", ",", ".", ".", ".", ",", "∇η", "(", "XM", ")", "into", "one", "vector", ".", "Proof", "of", "Proposition", "3", ".", "Let", "g", "∈", "H", "be", "the", "solution", "to", "the", "regression", "problem", "D", "µ", ",", "λ", "g", "=", "η", ":", "Taking", "the", "inner", "product", "of", "both", "sides", "of", "(", "12", ")", "with", "k", "(", "X", "m", ",", "·", ")", "for", "each", "1", "≤", "m", "≤", "M", "yields", "the", "following", "M", "equations", ":", "Doing", "the", "same", "with", "∂", "j", "k", "(", "X", "m", ",", "·", ")", "gives", "M", "d", "equations", ":", ".", "(", "14", ")", "From", "(", "12", ")", ",", "it", "is", "clear", "that", "g", "is", "a", "linear", "combination", "of", "the", "form", ":", "where", "the", "coefficients", "α", ":", "satisfy", "the", "system", "of", "equations", "(", "13", ")", "and", "(", "14", ")", ".", "We", "can", "rewrite", "this", "system", "as", "where", "I", "M", ",", "I", "M", "d", "are", "the", "identity", "matrices", "of", "dimension", "M", ",", "M", "d.", "Since", "K", "and", "H", "must", "be", "positive", "semidefinite", ",", "an", "inverse", "exists", ".", "We", "conclude", "by", "noticing", "that", "The", "following", "result", "was", "key", "to", "our", "definition", "of", "the", "SMMD", "in", "Section", "3.3", ".", "Proposition", "4", ".", "Under", "Assumptions", "(", "A", ")", "to", "(", "D", ")", ",", "we", "have", "for", "all", "f", "∈", "H", "that", "Proof", "of", "Proposition", "4", ".", "The", "key", "idea", "here", "is", "to", "use", "the", "Cauchy", "-", "Schwarz", "inequality", "for", "the", "HilbertSchmidt", "inner", "product", ".", "(", "a", ")", "follows", "from", "the", "reproducing", "properties", "(", "8", ")", "and", "(", "9", ")", "and", "Property", "(", "ii", ")", ".", "(", "b", ")", "is", "obtained", "using", "Property", "(", "iii", ")", ",", "while", "(", "c", ")", "follows", "from", "the", "Cauchy", "-", "Schwarz", "inequality", "and", "Property", "(", "i", ")", ".", "Under", "Assumptions", "(", "A", ")", "and", "(", "B", ")", ",", "k", "(", "x", ",", "·", ")", "is", "Bochner", "integrable", "with", "respect", "to", "any", "probability", "distribution", "P", "with", "finite", "first", "moment", "and", "the", "following", "relation", "holds", ":", "Proof", ".", "The", "operator", "D", "x", "is", "positive", "self", "-", "adjoint", ".", "It", "is", "also", "trace", "-", "class", ",", "as", "by", "the", "triangle", "inequality", "By", "Assumption", "(", "D", ")", ",", "we", "have", "that", "The", "Bochner", "integrability", "of", "k", "(", "x", ",", "·", ")", "under", "a", "distribution", "P", "with", "finite", "moment", "follows", "directly", "from", "Assumptions", "(", "A", ")", "and", "(", "B", ")", ",", "since", "section", ":", "A.2", "Continuity", "of", "the", "Optimized", "Scaled", "MMD", "in", "the", "Wasserstein", "topology", "To", "prove", "Theorem", "1", ",", "we", "we", "will", "first", "need", "some", "new", "notation", ".", "We", "assume", "the", "kernel", "is", "k", "=", "K", "•", "φ", "ψ", ",", "i.e.", "k", "ψ", "(", "x", ",", "y", ")", "=", "K", "(", "φ", "ψ", "(", "x", ")", ",", "φ", "ψ", "(", "y", ")", ")", ",", "where", "the", "representation", "function", "φ", "ψ", "is", "a", "network", "φ", "ψ", "(", "X", ")", ":", "The", "intermediate", "representations", "h", "The", "elementwise", "activation", "function", "σ", "is", "given", "by", "σ", "0", "(", "x", ")", "=", "x", ",", "and", "for", "l", ">", "0", "the", "activation", "σ", "l", "is", "a", "leaky", "ReLU", "with", "leak", "coefficient", "0", "<", "α", "<", "1", ":", "The", "parameter", "ψ", "is", "the", "concatenation", "of", "all", "the", "layer", "parameters", ":", "We", "denote", "by", "Ψ", "the", "set", "of", "all", "such", "possible", "parameters", ",", "i.", ".", "Define", "the", "following", "restrictions", "of", "Ψ", ":", "Ψ", "κ", "is", "the", "set", "of", "those", "parameters", "such", "that", "W", "l", "have", "a", "small", "condition", "number", ",", "cond", "(", "W", ")", "=", "σ", "max", "(", "W", ")", "/", "σ", "min", "(", "W", ")", ".", "Ψ", "κ", "1", "is", "the", "set", "of", "per", "-", "layer", "normalized", "parameters", "with", "a", "condition", "number", "bounded", "by", "κ", ".", "Recall", "the", "definition", "of", "Scaled", "MMD", ",", "(", "7", ")", ",", "where", "λ", ">", "0", "and", "µ", "is", "a", "probability", "measure", ":", "The", "Optimized", "SMMD", "over", "the", "restricted", "set", "Ψ", "κ", "is", "given", "by", ":", "The", "constraint", "to", "ψ", "∈", "Ψ", "κ", "is", "critical", "to", "the", "proof", ".", "In", "practice", ",", "using", "a", "spectral", "parametrization", "helps", "enforce", "this", "assumption", ",", "as", "shown", "in", "Figures", "2", "and", "9", ".", "Other", "regularization", "methods", ",", "like", "orthogonal", "normalization", "[", "reference", "]", ",", "are", "also", "possible", ".", "We", "will", "use", "the", "following", "assumptions", ":", "(", "I", ")", "µ", "is", "a", "probability", "distribution", "absolutely", "continuous", "with", "respect", "to", "the", "Lebesgue", "measure", ".", "(", "II", ")", "The", "dimensions", "of", "the", "weights", "are", "decreasing", "per", "layer", ":", "(", "III", ")", "The", "non", "-", "linearity", "used", "is", "Leaky", "-", "ReLU", ",", "[", "reference", "]", ",", "with", "leak", "coefficient", "α", "∈", "(", "0", ",", "1", ")", ".", "(", "IV", ")", "The", "top", "-", "level", "kernel", "K", "is", "globally", "Lipschitz", "in", "the", "RKHS", "norm", ":", "there", "exists", "a", "positive", "constant", "(", "V", ")", "There", "is", "some", "γ", "K", ">", "0", "for", "which", "K", "satisfies", "Assumption", "(", "I", ")", "ensures", "that", "the", "points", "where", "φ", "ψ", "(", "X", ")", "is", "not", "differentiable", "are", "reached", "with", "probability", "0", "under", "µ.", "This", "assumption", "can", "be", "easily", "satisfied", "e.g.", "if", "we", "define", "µ", "by", "adding", "Gaussian", "noise", "to", "P.", "Assumption", "(", "II", ")", "helps", "ensure", "that", "the", "span", "of", "W", "l", "is", "never", "contained", "in", "the", "null", "space", "of", "W", "l", "+", "1", ".", "Using", "Leaky", "-", "ReLU", "as", "a", "non", "-", "linearity", ",", "Assumption", "(", "III", ")", ",", "further", "ensures", "that", "the", "network", "φ", "ψ", "is", "locally", "full", "-", "rank", "almost", "everywhere", ";", "this", "might", "not", "be", "true", "with", "ReLU", "activations", ",", "where", "it", "could", "be", "always", "0", ".", "Assumptions", "(", "II", ")", "and", "(", "III", ")", "can", "be", "easily", "satisfied", "by", "design", "of", "the", "network", ".", "Assumptions", "(", "IV", ")", "and", "(", "V", ")", "only", "depend", "on", "the", "top", "-", "level", "kernel", "K", "and", "are", "easy", "to", "satisfy", "in", "practice", ".", "In", "particular", ",", "they", "always", "hold", "for", "a", "smooth", "translation", "-", "invariant", "kernel", ",", "such", "as", "the", "Gaussian", ",", "as", "well", "as", "the", "linear", "kernel", ".", "We", "are", "now", "ready", "to", "prove", "Theorem", "1", ".", "Theorem", "1", ".", "Under", "Assumptions", "(", "I", ")", "to", "(", "V", ")", ",", "Proof", ".", "Define", "the", "pseudo", "-", "distance", "corresponding", "to", "the", "kernel", "k", "ψ", "Denote", "by", "W", "d", "ψ", "(", "P", ",", "Q", ")", "the", "optimal", "transport", "metric", "between", "P", "and", "Q", "using", "the", "cost", "d", "ψ", ",", "given", "by", "where", "Π", "is", "the", "set", "of", "couplings", "with", "marginals", "P", "and", "Q.", "By", "Lemma", "7", ",", "Recall", "that", "φ", "ψ", "is", "Lipschitz", ",", "φ", "ψ", "Lip", "<", "∞", ",", "so", "along", "with", "Assumption", "(", "IV", ")", "we", "have", "that", "where", "W", "is", "the", "standard", "Wasserstein", "distance", "(", "2", ")", ",", "and", "so", "We", "have", "that", "2", ",", "and", "hence", "Using", "Lemma", "8", ",", "we", "can", "write", "where", "we", "used", "φψ", "Lip", "≤", "L", "l=1", "W", "l", "=", "1", ".", "But", "by", "Lemma", "9", ",", "for", "Lebesgue", "-", "almost", "all", "X", ",", "Using", "Assumption", "(", "I", ")", ",", "this", "implies", "that", "Thus", "for", "any", "ψ", "∈", "Ψ", "κ", ",", "The", "desired", "bound", "on", "D", "µ", ",", "Ψ", "κ", ",", "λ", "SMMD", "follows", "immediately", ".", "Lemma", "7", ".", "Let", "(", "x", ",", "y", ")", "→", "k", "(", "x", ",", "y", ")", "be", "the", "continuous", "kernel", "of", "an", "RKHS", "H", "defined", "on", "a", "Polish", "space", "X", ",", "and", "define", "the", "corresponding", "pseudo", "-", "distance", "d", "k", "(", "x", ",", "y", ")", ":", "=", "k", "(", "x", ",", "·", ")", "−", "k", "(", "y", ",", "·", ")", "H", ".", "Then", "the", "following", "inequality", "holds", "for", "any", "distributions", "P", "and", "Q", "on", "X", ",", "including", "when", "the", "quantities", "are", "infinite", ":", "Proof", ".", "Let", "P", "and", "Q", "be", "two", "probability", "distributions", ",", "and", "let", "Π", "(", "P", ",", "Q", ")", "be", "the", "set", "of", "couplings", "between", "them", ".", "Let", "π", "*", "∈", "argmin", "(", "X", ",", "Y", ")", "∼π", "[", "c", "k", "(", "X", ",", "Y", ")", "]", "be", "an", "optimal", "coupling", ",", "which", "is", "guaranteed", "to", "exist", "Take", "a", "sample", "(", "X", ",", "Y", ")", "∼", "π", "and", "a", "function", "f", "∈", "H", "with", "f", "H", "≤", "1", ".", "By", "the", "Cauchy", "-", "Schwarz", "inequality", ",", "Taking", "the", "expectation", "with", "respect", "to", "π", ",", "we", "obtain", "The", "right", "-", "hand", "side", "is", "just", "the", "definition", "of", "W", "d", "k", "(", "P", ",", "Q", ")", ".", "By", "Jensen", "'s", "inequality", ",", "the", "left", "-", "hand", "side", "is", "lower", "-", "bounded", "by", "since", "π", "has", "marginals", "P", "and", "Q.", "We", "have", "shown", "so", "far", "that", "for", "any", "f", "∈", "H", "with", "f", "H", "≤", "1", ",", "the", "result", "follows", "by", "taking", "the", "supremum", "over", "f", ".", "Proof", ".", "Note", "that", "the", "condition", "number", "is", "unchanged", ",", "cond", "(", "W", "l", ")", "=", "cond", "(", "W", "l", ")", "≤", "κ", ",", "and", "W", "l", "=", "1", ",", "soψ", "∈", "Φ", "κ", "1", ".", "It", "is", "also", "easy", "to", "see", "from", "[", "reference", "]", "that", "Lemma", "9", ".", "Make", "Assumptions", "(", "II", ")", "and", "(", "III", ")", ",", "and", "let", "ψ", "∈", "Ψ", "κ", "1", ".", "Then", "the", "set", "of", "inputs", "for", "which", "any", "intermediate", "activation", "is", "exactly", "zero", ",", "has", "zero", "Lebesgue", "measure", ".", "Moreover", ",", "for", "any", "X", "/", "∈", "N", "ψ", ",", "∇", "X", "φ", "ψ", "(", "X", ")", "exists", "and", "Proof", ".", "First", ",", "note", "that", "the", "network", "representation", "at", "layer", "l", "is", "piecewise", "affine", ".", "Specifically", ",", "define", "M", "l", "X", "∈", "R", "d", "l", "by", ",", "using", "Assumption", "(", "III", ")", ",", "We", "will", "now", "show", "that", "the", "activation", "patterns", "are", "piecewise", "constant", ",", "so", "that", "Thus", ",", "take", "some", "X", "/", "∈", "N", "ψ", ",", "and", "find", "the", "smallest", "absolute", "value", "of", "its", "activations", ",", "=", ";", "clearly", ">", "0", ".", "For", "any", "X", "with", "X", "−", "X", "<", ",", "we", "know", "that", "for", "all", "l", "and", "k", ",", "sign", "h", "Proof", ".", "A", "more", "general", "version", "of", "this", "result", "can", "be", "found", "in", "[", "reference", "][", "reference", "]", ";", "we", "provide", "a", "proof", "here", "for", "completeness", ".", "If", "B", "has", "a", "nontrivial", "null", "space", ",", "σ", "min", "(", "B", ")", "=", "0", "and", "the", "inequality", "holds", ".", "Otherwise", ",", "let", "R", "n", "*", "denote", "R", "n", "\\", "{", "0}.", "Recall", "that", "for", "C", "∈", "R", "m×n", "with", "m", "≥", "n", ",", "Here", "A", "1", ",", "A", "2", ",", "A", "3", "and", "A", "4", "are", "defined", "by", "[", "reference", "]", "and", "are", "represented", "in", "Figure", "4", ":", "Figure", "4", ":", "Decomposition", "of", "R", "2", "into", "4", "regions", "A", "1", ",", "A", "2", ",", "A", "3", "and", "A", "4", "as", "defined", "in", "[", "reference", "]", ".", "As", "α", "approaches", "0", ",", "the", "area", "of", "sets", "A", "3", "and", "A", "4", "becomes", "negligible", ".", "It", "is", "easy", "to", "see", "that", "whenever", "µ", "has", "a", "density", ",", "the", "probability", "of", "the", "sets", "A", "3", "and", "A", "4", "goes", "to", "0", "are", "α", "→", "0", ".", "Hence", "one", "can", "deduce", "that", "On", "the", "other", "hand", ",", "the", "squared", "Lipschitz", "constant", "of", "φ", "is", "given", "by", "(", "1", "−", "γ", ")", "2", "+", "(", "1", "+", "α", "−", "γ", ")", "2", "which", "converges", "to", "2", "(", "1", "−", "γ", ")", "2", ".", "This", "shows", "that", "controlling", "the", "expectation", "of", "the", "gradient", "does", "n't", "allow", "to", "effectively", "control", "the", "Lipschitz", "constant", "of", "φ", ".", "section", ":", "Monotonicity", "of", "the", "dimensions", ":", "We", "would", "like", "to", "consider", "a", "second", "example", "where", "Assumption", "(", "II", ")", "does", "n't", "hold", ".", "Consider", "the", "following", "two", "layer", "network", "defined", "by", ":", "for", "β", ">", "0", ".", "Note", "that", "W", "β", "is", "a", "full", "rank", "matrix", ",", "but", "Assumption", "(", "II", ")", "does", "n't", "hold", ".", "Depending", "on", "the", "sign", "of", "the", "components", "of", "W", "β", "X", "one", "has", "the", "following", "expression", "for", "∇φ", "α", "(", "X", ")", "2", ":", "where", "(", "B", "i", ")", "1≤i≤6", "are", "defined", "by", "(", "26", ")", "section", ":", "20", "The", "squared", "Lipschitz", "constant", "is", "given", "by", "φ", "2", "L", "(", "1", "−", "γ", ")", "2", "+", "β", "2", "while", "the", "expected", "squared", "norm", "of", "the", "gradient", "of", "φ", "is", "given", "by", ":", "Again", "the", "set", "B", "4", "∪", "B", "5", "becomes", "negligible", "as", "β", "approaches", "0", "which", "implies", "that", "2", ".", "Note", "that", "unlike", "in", "the", "first", "example", "in", "(", "21", ")", ",", "the", "matrix", "W", "β", "has", "a", "bounded", "condition", "number", ".", "In", "this", "example", ",", "the", "columns", "of", "W", "0", "are", "all", "in", "the", "null", "space", "of", "[", "−1", "0", "1", "]", ",", "which", "implies", "∇φ", "0", "(", "X", ")", "=", "0", "for", "all", "X", "∈", "R", "2", ",", "even", "though", "all", "matrices", "have", "full", "rank", ".", "B", "DiracGAN", "vector", "fields", "for", "more", "losses", "Figure", "5", ":", "Vector", "fields", "for", "different", "losses", "with", "respect", "to", "the", "generator", "parameter", "θ", "and", "the", "feature", "representation", "parameter", "ψ", ";", "the", "losses", "use", "a", "Gaussian", "kernel", ",", "and", "are", "shown", "in", "[", "reference", "]", ".", "Following", "[", "reference", "]", ",", "P", "=", "δ", "0", ",", "Q", "=", "δ", "θ", "and", "φ", "ψ", "(", "x", ")", "=", "ψx", ".", "The", "curves", "show", "the", "result", "of", "taking", "simultaneous", "gradient", "steps", "in", "(", "θ", ",", "ψ", ")", "beginning", "from", "three", "initial", "parameter", "values", ".", "Figure", "5", "shows", "parameter", "vector", "fields", ",", "like", "those", "in", "Figure", "6", ",", "for", "Example", "1", "for", "a", "variety", "of", "different", "losses", ":", "section", ":", "21", "The", "squared", "MMD", "between", "δ", "0", "and", "δ", "θ", "under", "a", "Gaussian", "kernel", "of", "bandwidth", "1", "/", "ψ", "and", "is", "given", "by", ")", ".", "MMD", "-", "GP", "-", "unif", "uses", "a", "gradient", "penalty", "as", "in", "[", "reference", "]", "where", "each", "samples", "from", "µ", "*", "is", "obtained", "by", "first", "sampling", "X", "and", "Y", "from", "P", "and", "Q", "and", "then", "sampling", "uniformly", "between", "X", "and", "Y", ".", "MMD", "-", "GP", "uses", "the", "same", "gradient", "penalty", ",", "but", "the", "expectation", "is", "taken", "under", "P", "rather", "than", "µ", "*", ".", "SN", "-", "MMD", "refers", "to", "MMD", "with", "spectral", "normalization", ";", "here", "this", "means", "that", "ψ", "=", "1", ".", "Sobolev", "-", "MMD", "refers", "to", "the", "loss", "used", "in", "[", "reference", "]", "with", "the", "quadratic", "penalty", "only", ".", "GCMMD", "µ", ",", "k", ",", "λ", "is", "defined", "by", "(", "5", ")", ",", "with", "µ", "=", "N", "(", "0", ",", "10", "2", ")", ".", "C", "Vector", "fields", "of", "Gradient", "-", "Constrained", "MMD", "and", "Sobolev", "GAN", "critics", "Mroueh", "et", "al", ".", "[", "reference", "]", "argue", "that", "the", "gradient", "of", "the", "critic", "(", "...", ")", "defines", "a", "transportation", "plan", "for", "moving", "the", "distribution", "mass", "(", "from", "generated", "to", "reference", "distribution", ")", "and", "present", "the", "solution", "of", "Sobolev", "PDE", "for", "2", "-", "dimensional", "Gaussians", ".", "We", "observed", "that", "in", "this", "simple", "example", "the", "gradient", "of", "the", "Sobolev", "critic", "can", "be", "very", "high", "outside", "of", "the", "areas", "of", "high", "density", ",", "which", "is", "not", "the", "case", "with", "the", "Gradient", "-", "Constrained", "MMD", ".", "Figure", "6", "presents", "critic", "gradients", "in", "both", "cases", ",", "using", "µ", "=", "(", "P", "+", "Q", ")/", "2", "for", "both", ".", "(", "a", ")", "Gradient", "-", "Constrained", "MMD", "critic", "gradient", ".", "This", "unintuitive", "behavior", "is", "most", "likely", "related", "to", "the", "vanishing", "boundary", "condition", ",", "assummed", "by", "Sobolev", "GAN", ".", "Solving", "the", "actual", "Sobolev", "PDE", ",", "we", "found", "that", "the", "Sobolev", "critic", "has", "very", "high", "gradients", "close", "to", "the", "boundary", "in", "order", "to", "match", "the", "condition", ";", "moreover", ",", "these", "gradients", "point", "in", "opposite", "directions", "to", "the", "target", "distribution", ".", "and", "the", "corresponding", "critic", "function", "is", "Thus", "if", "we", "assume", "E", "X∼P", "φ", "(", "X", ")", ">", "E", "Y", "∼Q", "φ", "(", "Y", ")", ",", "as", "that", "is", "the", "goal", "of", "our", "critic", "training", ",", "we", "see", "that", "the", "MMD", "becomes", "identical", "to", "the", "WGAN", "loss", ",", "and", "the", "gradient", "penalty", "is", "applied", "to", "the", "same", "function", ".", "(", "MMD", "GANs", ",", "however", ",", "would", "typically", "train", "on", "the", "unbiased", "estimator", "of", "MMD", "2", ",", "giving", "a", "very", "slightly", "different", "loss", "function", ".", "[", "reference", "]", "also", "applied", "the", "gradient", "penalty", "to", "η", "rather", "than", "the", "true", "critic", "η", "/", "η", ".", ")", "The", "SMMD", "with", "a", "linear", "kernel", "is", "thus", "analogous", "to", "applying", "the", "scaling", "operator", "to", "a", "WGAN", ";", "hence", "the", "name", "SWGAN", ".", "F", "Additional", "experiments", "F.1", "Comparison", "of", "Gradient", "-", "Constrained", "MMD", "to", "Scaled", "MMD", "Figure", "7", "shows", "the", "behavior", "of", "the", "MMD", ",", "the", "Gradient", "-", "Constrained", "SMMD", ",", "and", "the", "Scaled", "MMD", "when", "comparing", "Gaussian", "distributions", ".", "We", "can", "see", "that", "MMD", "∝", "SMMD", "and", "the", "GradientConstrained", "MMD", "behave", "similarly", "in", "this", "case", ",", "and", "that", "optimizing", "the", "SMMD", "and", "the", "GradientConstrained", "MMD", "is", "also", "similar", ".", "Optimizing", "the", "MMD", "would", "yield", "an", "essentially", "constant", "distance", ".", "section", ":", "F.2", "IGMs", "with", "Optimized", "Gradient", "-", "Constrained", "MMD", "loss", "We", "implemented", "the", "estimator", "of", "Proposition", "3", "using", "the", "empirical", "mean", "estimator", "of", "η", ",", "and", "sharing", "samples", "for", "µ", "=", "P.", "To", "handle", "the", "large", "but", "approximately", "low", "-", "rank", "matrix", "system", ",", "we", "used", "an", "incomplete", "Cholesky", "decomposition", "[", "reference", "][", "reference", "]", "Then", "the", "Woodbury", "matrix", "identity", "allows", "an", "efficient", "evaluation", ":", "Even", "though", "only", "a", "small", "is", "required", "for", "a", "good", "approximation", ",", "and", "the", "full", "matrices", "K", ",", "G", ",", "and", "H", "need", "never", "be", "constructed", ",", "backpropagation", "through", "this", "procedure", "is", "slow", "and", "not", "especially", "GPU", "-", "friendly", ";", "training", "on", "CPU", "was", "faster", ".", "Thus", "we", "were", "only", "able", "to", "run", "the", "estimator", "on", "MNIST", ",", "and", "even", "that", "took", "days", "to", "conduct", "the", "optimization", "on", "powerful", "workstations", ".", "The", "learned", "models", ",", "however", ",", "were", "reasonable", ".", "Using", "a", "DCGAN", "architecture", ",", "batches", "of", "size", "64", ",", "and", "a", "procedure", "that", "otherwise", "agreed", "with", "the", "setup", "of", "Section", "4", ",", "samples", "with", "and", "without", "spectral", "normalization", "are", "shown", "in", "Figures", "8a", "and", "8b", ".", "After", "the", "points", "in", "training", "shown", ",", "however", ",", "the", "same", "rank", "collapse", "as", "discussed", "in", "Section", "4", "occurred", ".", "Here", "it", "seems", "that", "spectral", "normalization", "may", "have", "delayed", "the", "collapse", ",", "but", "not", "prevented", "it", ".", "Figure", "8c", "shows", "generator", "loss", "estimates", "through", "training", ",", "including", "the", "obvious", "peak", "at", "collapse", ";", "Figure", "8d", "shows", "KID", "scores", "based", "on", "the", "MNIST", "-", "trained", "convnet", "representation", "[", "reference", "]", ",", "including", "comparable", "SMMD", "models", "for", "context", ".", "The", "fact", "that", "SMMD", "models", "converged", "somewhat", "faster", "than", "Gradient", "-", "Constrained", "MMD", "models", "here", "may", "be", "more", "related", "to", "properties", "of", "the", "estimator", "of", "Proposition", "3", "rather", "than", "the", "distances", ";", "more", "work", "would", "be", "needed", "to", "fully", "compare", "the", "behavior", "of", "the", "two", "distances", ".", "Figure", "9", "shows", "the", "distribution", "of", "critic", "weight", "singular", "values", ",", "like", "Figure", "2", ",", "at", "more", "layers", ".", "Figure", "11", "and", "Table", "2", "show", "results", "for", "the", "spectral", "normalization", "variants", "considered", "in", "the", "experiments", ".", "MMDGAN", ",", "with", "neither", "spectral", "normalization", "nor", "a", "gradient", "penalty", ",", "did", "surprisingly", "well", "in", "this", "case", ",", "though", "it", "fails", "badly", "in", "other", "situations", ".", "Figure", "9", "compares", "the", "decay", "of", "singular", "values", "for", "layer", "of", "the", "critic", "'s", "network", "at", "both", "early", "and", "later", "stages", "of", "training", "in", "two", "cases", ":", "with", "or", "without", "the", "spectral", "parametrization", ".", "The", "model", "was", "trained", "on", "CelebA", "using", "SMMD", ".", "Figure", "11", "shows", "the", "evolution", "per", "iteration", "of", "Inception", "score", ",", "section", ":", "F.3", "Spectral", "normalization", "and", "Scaled", "MMD", "section", ":", "section", ":", "it", "is", "undefined", "when", "any", "h", "Because", "ψ", "∈", "Ψ", "κ", "1", ",", "we", "have", "≤", "1", ",", "and", "using", "Assumption", "(", "II", ")", "with", "Lemma", "10", "gives", "X", ")", "denote", "the", "full", "activation", "patterns", "up", "to", "level", "l", ",", "we", "can", "thus", "write", "There", "are", "only", "finitely", "many", "possible", "values", "for", "H", "l", "X", ";", "we", "denote", "the", "set", "of", "such", "values", "as", "H", "l", ".", "Then", "we", "have", "that", "Because", "each", "W", "H", "l", "k", "is", "of", "rank", "d", "l", ",", "each", "set", "in", "the", "union", "is", "either", "empty", "or", "an", "affine", "subspace", "of", "dimension", "d", "−", "d", "l", ".", "As", "each", "d", "l", ">", "0", ",", "each", "set", "in", "the", "finite", "union", "has", "zero", "Lebesgue", "measure", ",", "and", "N", "ψ", "also", "has", "zero", "Lebesgue", "measure", ".", "Thus", ",", "as", "Bx", "=", "0", "for", "x", "=", "0", ",", "section", ":", "A.2.1", "When", "some", "of", "the", "assumptions", "do", "n't", "hold", "Here", "we", "analyze", "through", "simple", "examples", "what", "happens", "when", "the", "condition", "number", "can", "be", "unbounded", ",", "and", "when", "Assumption", "(", "II", ")", ",", "about", "decreasing", "widths", "of", "the", "network", ",", "is", "violated", ".", "Condition", "Number", ":", "We", "start", "by", "a", "first", "example", "where", "the", "condition", "number", "can", "be", "arbitrarily", "high", ".", "We", "consider", "a", "two", "-", "layer", "network", "on", "R", "2", ",", "defined", "by", "where", "α", ">", "0", ".", "As", "α", "approaches", "0", "the", "matrix", "W", "α", "becomes", "singular", "which", "means", "that", "its", "condition", "number", "blows", "up", ".", "We", "are", "interested", "in", "analyzing", "the", "behavior", "of", "the", "Lipschitz", "constant", "of", "φ", "and", "the", "expected", "squared", "norm", "of", "its", "gradient", "under", "µ", "as", "α", "approaches", "0", ".", "One", "can", "easily", "compute", "the", "squared", "norm", "of", "the", "gradient", "of", "φ", "which", "is", "given", "by", "section", ":", "D", "An", "estimator", "for", "Lipschitz", "MMD", "We", "now", "describe", "briefly", "how", "to", "estimate", "the", "Lipschitz", "MMD", "in", "low", "dimensions", ".", "Recall", "that", "For", "f", "∈", "H", "k", ",", "it", "is", "the", "case", "that", "Thus", "we", "can", "approximate", "the", "constraint", "f", "2", "Lip", "+", "λ", "f", "2", "H", "k", "≤", "1", "by", "enforcing", "the", "constraint", "on", "a", "set", "of", "m", "points", "{", "Z", "i", "}", "reasonably", "densely", "covering", "the", "region", "around", "the", "supports", "of", "P", "and", "Q", ",", "rather", "22", "than", "enforcing", "it", "at", "every", "point", "in", "X", ".", "An", "estimator", "of", "the", "Lipschitz", "MMD", "based", "on", "X", "∼", "P", "n", "X", "and", "By", "the", "generalized", "representer", "theorem", ",", "the", "optimal", "f", "for", "(", "29", ")", "will", "be", "of", "the", "form", "Writing", "δ", "=", "(", "α", ",", "β", ",", "γ", ")", ",", "the", "objective", "function", "is", "linear", "in", "δ", ",", "The", "constraints", "are", "quadratic", ",", "built", "from", "the", "following", "matrices", ",", "where", "the", "X", "and", "Y", "samples", "are", "concatenated", "together", ",", "as", "are", "the", "derivatives", "with", "each", "dimension", "of", "the", "Z", "samples", ":", "Given", "these", "matrices", ",", "and", "letting", "where", "e", "(", "i", ",", "j", ")", "is", "the", "(", "i", ",", "j", ")", "th", "standard", "basis", "vector", "in", "R", "md", ",", "we", "have", "that", "Thus", "the", "optimization", "problem", "(", "29", ")", "is", "a", "linear", "problem", "with", "convex", "quadratic", "constraints", ",", "which", "can", "be", "solved", "by", "standard", "convex", "optimization", "software", ".", "The", "approximation", "is", "reasonable", "only", "if", "we", "can", "effectively", "cover", "the", "region", "of", "interest", "with", "densely", "spaced", "{", "Z", "i", "}", ";", "it", "requires", "a", "nontrivial", "amount", "of", "computation", "even", "for", "the", "very", "simple", "1", "-", "dimensional", "toy", "problem", "of", "Example", "1", ".", "One", "advantage", "of", "this", "estimator", ",", "though", ",", "is", "that", "finding", "its", "derivative", "with", "respect", "to", "the", "input", "points", "or", "the", "kernel", "parameterization", "is", "almost", "free", "once", "we", "have", "computed", "the", "estimate", ",", "as", "long", "as", "our", "solver", "has", "computed", "the", "dual", "variables", "µ", "corresponding", "to", "the", "constraints", "in", "[", "reference", "]", ".", "We", "just", "need", "to", "exploit", "the", "envelope", "theorem", "and", "then", "differentiate", "the", "KKT", "conditions", ",", "as", "done", "for", "instance", "in", "[", "reference", "]", ".", "The", "differential", "of", "(", "29", ")", "ends", "up", "being", ",", "assuming", "the", "optimum", "of", "(", "29", ")", "is", "atδ", "∈", "R", "n", "X", "+", "n", "Y", "+", "md", "and", "µ", "∈", "R", "m", ",", "d", "LipMMD", "k", ",", "λ", "(", "X", ",", "Y", ",", "Z", ")", "=", "δ", "T", "dK", "dB", "section", ":"], "entity": [[1, 3, ["gradient", "regularizers"], "Method"], [12, 14, ["principled", "method"], "Method"], [32, 33, ["kernel"], "Method"], [35, 38, ["Maximum", "Mean", "Discrepancy"], "Method"], [39, 40, ["MMD"], "Method"], [78, 80, ["approximate", "techniques"], "Method"], [82, 84, ["additive", "regularizers"], "Method"], [87, 89, ["loss", "function"], "Metric"], [101, 102, ["training"], "Task"], [120, 121, ["CelebA"], "Material"], [151, 157, ["Wasserstein", "-", "inspired", "gradient", "regularization", "approaches"], "Method"], [187, 189, ["gradient", "penalty"], "Method"], [216, 218, ["Sobolev", "IPM"], "Method"], [221, 223, ["approximating", "Wasserstein"], "Method"], [248, 252, ["P", "+", "Q", "θ"], "Method"], [289, 291, ["Gradient", "penalties"], "Method"], [293, 294, ["GANs"], "Method"], [296, 299, ["f", "-", "divergences"], "Method"], [307, 310, ["spectral", "normalization", "technique"], "Method"], [319, 321, ["critic", "network"], "Method"], [372, 375, ["locally", "convergent", "optimization"], "Task"], [387, 388, ["GAN"], "Method"], [402, 403, ["regularization"], "Method"], [420, 421, ["generator"], "Task"], [426, 428, ["regularization", "methods"], "Method"], [431, 433, ["Wasserstein", "distances"], "Method"], [450, 454, ["semi", "-", "supervised", "learning"], "Task"], [564, 565, ["regularizer"], "Method"], [574, 579, ["density", "-", "weighted", "gradient", "norm"], "Method"], [653, 654, ["critic"], "Method"], [669, 671, ["stable", "learning"], "Method"], [679, 682, ["integral", "probability", "metric"], "Metric"], [716, 719, ["implicit", "generative", "models"], "Method"], [755, 757, ["image", "generation"], "Task"], [761, 764, ["implicit", "generative", "models"], "Method"], [770, 771, ["IGM"], "Method"], [775, 776, ["θ"], "Method"], [781, 784, ["target", "distribution", "P"], "Method"], [799, 801, ["generator", "function"], "Method"], [811, 813, ["deep", "network"], "Method"], [850, 852, ["Q", "θ"], "Method"], [864, 865, ["learning"], "Task"], [903, 909, ["D", "(", "P", ",", "Q", "θ"], "Task"], [911, 914, ["stochastic", "gradient", "descent"], "Method"], [916, 919, ["estimator", "of", "D."], "Method"], [927, 929, ["losses", "D"], "Metric"], [932, 935, ["Maximum", "Mean", "Discrepancy"], "Method"], [962, 963, ["optimization"], "Task"], [970, 975, ["closed", "-", "form", "optimal", "critic"], "Method"], [1003, 1005, ["sample", "complexity"], "Metric"], [1023, 1028, ["Wasserstein", "distance", "W", "(", "P"], "Metric"], [1034, 1035, ["MMD"], "Method"], [1168, 1169, ["MMD"], "Method"], [1191, 1193, ["Gaussian", "kernel"], "Method"], [1195, 1196, ["MMD"], "Method"], [1214, 1216, ["alternating", "optimization"], "Task"], [1218, 1219, ["IGM"], "Task"], [1240, 1244, ["pixel", "-", "level", "kernels"], "Method"], [1309, 1311, ["kernel", "K"], "Method"], [1316, 1320, ["low", "-", "dimensional", "representation"], "Method"], [1327, 1329, ["characteristic", "kernel"], "Method"], [1331, 1332, ["Gaussian"], "Method"], [1334, 1336, ["φ", "ψ"], "Method"], [1339, 1341, ["deep", "network"], "Method"], [1363, 1365, ["φ", "ψ"], "Method"], [1471, 1473, ["alternating", "optimization"], "Method"], [1480, 1481, ["generator"], "Method"], [1492, 1493, ["GANs"], "Method"], [1494, 1495, ["WGANs"], "Method"], [1505, 1507, ["envelope", "theorem"], "Method"], [1532, 1533, ["generator"], "Method"], [1575, 1576, ["Wasserstein"], "Method"], [1577, 1578, ["MMD"], "Method"], [1624, 1626, ["Gaussian", "kernel"], "Method"], [1654, 1656, ["optimization", "issues"], "Task"], [1921, 1923, ["MMD", "GAN"], "Method"], [1927, 1931, ["WGAN", "-", "based", "models"], "Method"], [1937, 1940, ["additive", "gradient", "penalty"], "Method"], [1971, 1973, ["convergence", "behavior"], "Metric"], [1979, 1983, ["learning", "implicit", "generative", "models"], "Task"], [2001, 2003, ["IGM", "loss"], "Task"], [2122, 2123, ["LipMMD"], "Method"], [2146, 2147, ["LipMMD"], "Method"], [2316, 2318, ["RKHS", "norm"], "Method"], [2380, 2382, ["H", "k"], "Method"], [2393, 2395, ["kernel", "mean"], "Metric"], [2456, 2460, ["Gaussian", "or", "linear", "kernel"], "Method"], [2473, 2474, ["GCMMD"], "Method"], [2508, 2512, ["D", "µ", ",", "Ψ"], "Metric"], [2513, 2515, ["λ", "GCMMD"], "Method"], [2528, 2530, ["generative", "model"], "Method"], [2531, 2532, ["MNIST"], "Material"], [2535, 2543, ["linear", "system", "of", "size", "M", "+", "M", "d"], "Method"], [2555, 2559, ["low", "-", "rank", "approximation"], "Method"], [2580, 2581, ["GCMMD"], "Method"], [2589, 2591, ["IGM", "training"], "Task"], [2652, 2656, ["WGAN", "-", "GP", "measure"], "Method"], [2687, 2689, ["additive", "regularization"], "Method"], [2836, 2837, ["optimization"], "Task"], [2923, 2927, ["GCMMD", "µ", ",", "k"], "Method"], [2930, 2931, ["Gaussians"], "Method"], [2948, 2949, ["estimator"], "Method"], [2951, 2952, ["SMMD"], "Method"], [2964, 2965, ["SMMD"], "Method"], [2971, 2972, ["MMD"], "Method"], [2983, 2984, ["MMD"], "Method"], [2988, 2989, ["SMMD"], "Method"], [2991, 2994, ["kernel", "family", "Ψ"], "Method"], [3021, 3025, ["WGAN", "-", "GP", "measure"], "Method"], [3057, 3058, ["learning"], "Task"], [3081, 3088, ["fully", "-", "connected", "L", "-", "layer", "network"], "Method"], [3089, 3094, ["Leaky", "-", "ReLU", "α", "activations"], "Method"], [3146, 3147, ["SMMD"], "Method"], [3256, 3258, ["spectral", "normalization"], "Method"], [3259, 3260, ["SN"], "Method"], [3273, 3275, ["parametric", "family"], "Method"], [3301, 3302, ["MMD"], "Method"], [3305, 3306, ["MMD"], "Method"], [3308, 3313, ["RBF", "kernel", "of", "bandwidth", "1"], "Method"], [3354, 3355, ["SN"], "Method"], [3377, 3378, ["capacity"], "Metric"], [3403, 3407, ["dependent", "measure", "of", "complexity"], "Metric"], [3421, 3422, ["critics"], "Method"], [3431, 3433, ["Spectral", "parametrization"], "Method"], [3459, 3461, ["generator", "parameterization"], "Method"], [3490, 3493, ["redundant", "convolutional", "filters"], "Method"], [3543, 3545, ["critic", "parameterization"], "Method"], [3572, 3576, ["γW", "/", "W", "op"], "Method"], [3605, 3608, ["CIFAR", "-", "10"], "Material"], [3621, 3622, ["CelebA"], "Material"], [3649, 3654, ["ILSVRC2012", "(", "ImageNet", ")", "dataset"], "Material"], [3696, 3701, ["scalar", "-", "output", "critic", "network"], "Method"], [3709, 3712, ["MMDGAN", "-", "GP"], "Method"], [3733, 3735, ["critic", "f"], "Method"], [3736, 3738, ["φ", "ψ"], "Method"], [3741, 3742, ["GAN"], "Method"], [3744, 3747, ["discriminator", "D", "ψ"], "Method"], [3772, 3773, ["kernel"], "Method"], [3780, 3783, ["MMDGAN", "-", "GP"], "Method"], [3786, 3790, ["mixture", "of", "RQ", "kernels"], "Method"], [3822, 3823, ["SMMD"], "Method"], [3856, 3857, ["WGAN"], "Method"], [3868, 3869, ["SWGAN"], "Method"], [3870, 3871, ["SMMD"], "Method"], [3872, 3873, ["SWGAN"], "Method"], [3913, 3916, ["CIFAR", "-", "10"], "Material"], [3920, 3922, ["CNN", "architecture"], "Method"], [3929, 3933, ["7", "-", "layer", "critic"], "Method"], [3935, 3939, ["4", "-", "layer", "generator"], "Method"], [3941, 3942, ["CelebA"], "Material"], [3946, 3951, ["5", "-", "layer", "DCGAN", "discriminator"], "Method"], [3953, 3958, ["10", "-", "layer", "ResNet", "generator"], "Method"], [3965, 3966, ["ImageNet"], "Material"], [3970, 3974, ["10", "-", "layer", "ResNet"], "Method"], [3977, 3978, ["generator"], "Method"], [3979, 3980, ["discriminator"], "Method"], [3991, 3993, ["convolutional", "layer"], "Method"], [4001, 4002, ["CelebA"], "Material"], [4003, 4004, ["ImageNet"], "Material"], [4010, 4013, ["CIFAR", "-", "10"], "Material"], [4054, 4056, ["spectral", "parametrization"], "Method"], [4064, 4068, ["reference", "]", ".", "Models"], "Method"], [4069, 4072, ["explicit", "gradient", "control"], "Method"], [4073, 4076, ["SN", "-", "GAN"], "Method"], [4097, 4099, ["spectral", "normalization"], "Method"], [4106, 4108, ["spectral", "parameterization"], "Method"], [4126, 4127, ["ImageNet"], "Material"], [4148, 4149, ["GPU"], "Method"], [4175, 4177, ["critic", "updates"], "Method"], [4184, 4186, ["learning", "rates"], "Metric"], [4189, 4192, ["CIFAR", "-", "10"], "Material"], [4193, 4194, ["CelebA"], "Material"], [4197, 4198, ["ImageNet"], "Material"], [4205, 4208, ["KID", "adaptive", "scheme"], "Method"], [4249, 4251, ["learning", "rate"], "Metric"], [4259, 4261, ["Adam", "optimizer"], "Method"], [4279, 4281, ["sample", "quality"], "Metric"], [4293, 4295, ["Inception", "network"], "Method"], [4300, 4302, ["ImageNet", "classification"], "Task"], [4316, 4318, ["Inception", "Score"], "Metric"], [4319, 4320, ["IS"], "Metric"], [4352, 4353, ["ImageNet"], "Material"], [4362, 4363, ["FID"], "Metric"], [4385, 4386, ["Gaussians"], "Method"], [4389, 4391, ["intermediate", "representations"], "Method"], [4398, 4399, ["IS"], "Metric"], [4414, 4415, ["KID"], "Method"], [4421, 4422, ["FID"], "Metric"], [4435, 4437, ["statistical", "properties"], "Metric"], [4466, 4469, ["CIFAR", "-", "10"], "Material"], [4470, 4472, ["CelebA", "datasets"], "Material"], [4474, 4477, ["CIFAR", "-", "10"], "Material"], [4478, 4481, ["SN", "-", "SWGAN"], "Method"], [4482, 4485, ["SN", "-", "SMMDGAN"], "Method"], [4488, 4491, ["SN", "-", "GAN"], "Method"], [4494, 4495, ["CelebA"], "Material"], [4496, 4499, ["SN", "-", "SWGAN"], "Method"], [4500, 4503, ["SN", "-", "SMMDGAN"], "Method"], [4543, 4546, ["SN", "-", "SWGAN"], "Method"], [4548, 4551, ["WGAN", "-", "GP"], "Method"], [4561, 4562, ["SMMDGAN"], "Method"], [4563, 4566, ["SN", "-", "SMMDGAN"], "Method"], [4568, 4569, ["ImageNet"], "Material"], [4574, 4578, ["pre", "-", "trained", "models"], "Method"], [4579, 4580, ["BGAN"], "Method"], [4584, 4587, ["SN", "-", "GAN"], "Method"], [4602, 4603, ["FID"], "Metric"], [4604, 4606, ["KID", "scores"], "Metric"], [4612, 4613, ["ImageNet"], "Material"], [4614, 4615, ["CelebA"], "Material"], [4623, 4624, ["WGANs"], "Method"], [4625, 4626, ["MMDGANs"], "Method"], [4633, 4635, ["spectral", "parametrization"], "Method"], [4643, 4644, ["MMDGANs"], "Method"], [4645, 4646, ["WGANs"], "Method"], [4651, 4653, ["spectral", "normalization"], "Method"], [4662, 4663, ["WGAN"], "Method"], [4685, 4686, ["MMDGAN"], "Method"], [4688, 4691, ["CIFAR", "-", "10"], "Material"], [4699, 4700, ["CelebA"], "Material"], [4706, 4708, ["gradient", "control"], "Method"], [4710, 4711, ["SN"], "Method"], [4738, 4740, ["critic", "network"], "Method"], [4746, 4747, ["SMMD"], "Method"], [4750, 4751, ["SN"], "Method"], [4788, 4789, ["MMDGAN"], "Method"], [4804, 4805, ["CelebA"], "Material"], [4821, 4823, ["Rank", "collapse"], "Method"], [4830, 4831, ["SMMD"], "Method"], [4855, 4857, ["training", "objective"], "Metric"], [4869, 4871, ["Spectral", "parametrization"], "Method"], [4915, 4917, ["gradient", "regularization"], "Method"], [4923, 4926, ["implicit", "generative", "models"], "Method"], [4958, 4960, ["loss", "function"], "Metric"], [4961, 4962, ["IGMs"], "Method"], [5028, 5032, ["discrete", "-", "time", "approximations"], "Method"], [5061, 5063, ["Wasserstein", "geometry"], "Method"], [5065, 5066, ["MMD"], "Method"], [5079, 5081, ["optimized", "kernels"], "Method"], [5109, 5113, ["Reproducing", "Kernel", "Hilbert", "Spaces"], "Method"], [5152, 5155, ["reproducing", "kernel", "k"], "Method"], [5308, 5312, ["Hilbert", "-", "Schmidt", "operator"], "Method"], [5323, 5328, ["outer", "product", "f", "⊗", "g"], "Method"], [5330, 5334, ["Hilbert", "-", "Schmidt", "operator"], "Method"], [5335, 5339, ["Hilbert", "-", "Schmidt", "norm"], "Method"], [5378, 5382, ["covariance", "-", "type", "operators"], "Method"], [5429, 5431, ["kernel", "k"], "Method"], [5491, 5492, ["Gaussian"], "Method"], [5746, 5748, ["maximization", "problem"], "Task"], [5765, 5773, ["infinite", "-", "dimensional", "operator", "D", "µ", ",", "λ"], "Method"], [5781, 5783, ["computable", "estimator"], "Method"], [5925, 5927, ["linear", "operator"], "Method"], [5987, 5994, ["regression", "problem", "D", "µ", ",", "λ", "g"], "Task"], [6138, 6139, ["SMMD"], "Method"], [6177, 6181, ["Cauchy", "-", "Schwarz", "inequality"], "Method"], [6259, 6261, ["probability", "distribution"], "Method"], [6275, 6278, ["operator", "D", "x"], "Method"], [6306, 6312, ["Bochner", "integrability", "of", "k", "(", "x"], "Method"], [6346, 6348, ["Wasserstein", "topology"], "Method"], [6400, 6404, ["representation", "function", "φ", "ψ"], "Method"], [6414, 6416, ["intermediate", "representations"], "Method"], [6418, 6422, ["elementwise", "activation", "function", "σ"], "Method"], [6439, 6442, ["activation", "σ", "l"], "Method"], [6567, 6568, ["SMMD"], "Method"], [6596, 6598, ["spectral", "parametrization"], "Method"], [6612, 6614, ["regularization", "methods"], "Method"], [6616, 6618, ["orthogonal", "normalization"], "Method"], [6672, 6675, ["Leaky", "-", "ReLU"], "Method"], [6695, 6700, ["top", "-", "level", "kernel", "K"], "Method"], [6705, 6707, ["RKHS", "norm"], "Method"], [6794, 6797, ["Leaky", "-", "ReLU"], "Method"], [6812, 6815, ["network", "φ", "ψ"], "Method"], [6899, 6900, ["Gaussian"], "Method"], [6905, 6907, ["linear", "kernel"], "Method"], [6940, 6942, ["kernel", "k"], "Method"], [6963, 6966, ["cost", "d", "ψ"], "Metric"], [7114, 7116, ["RKHS", "H"], "Method"], [7119, 7121, ["Polish", "space"], "Material"], [7256, 7260, ["Cauchy", "-", "Schwarz", "inequality"], "Method"], [7452, 7454, ["network", "representation"], "Method"], [7803, 7806, ["two", "layer", "network"], "Method"], [7816, 7818, ["W", "β"], "Method"], [7873, 7876, ["squared", "Lipschitz", "constant"], "Metric"], [7893, 7901, ["expected", "squared", "norm", "of", "the", "gradient", "of", "φ"], "Metric"], [7937, 7940, ["matrix", "W", "β"], "Method"], [8025, 8027, ["Gaussian", "kernel"], "Method"], [8119, 8126, ["Gaussian", "kernel", "of", "bandwidth", "1", "/", "ψ"], "Method"], [8139, 8141, ["gradient", "penalty"], "Method"], [8194, 8197, ["SN", "-", "MMD"], "Method"], [8199, 8200, ["MMD"], "Method"], [8201, 8203, ["spectral", "normalization"], "Method"], [8230, 8234, ["GCMMD", "µ", ",", "k"], "Metric"], [8285, 8287, ["transportation", "plan"], "Method"], [8304, 8306, ["Sobolev", "PDE"], "Method"], [8307, 8311, ["2", "-", "dimensional", "Gaussians"], "Method"], [8401, 8403, ["Sobolev", "PDE"], "Method"], [8408, 8410, ["Sobolev", "critic"], "Method"], [8470, 8472, ["critic", "training"], "Task"], [8477, 8478, ["MMD"], "Method"], [8487, 8489, ["gradient", "penalty"], "Method"], [8518, 8520, ["loss", "function"], "Metric"], [8542, 8543, ["SMMD"], "Method"], [8545, 8547, ["linear", "kernel"], "Method"], [8553, 8555, ["scaling", "operator"], "Method"], [8557, 8558, ["WGAN"], "Method"], [8562, 8563, ["SWGAN"], "Method"], [8584, 8585, ["MMD"], "Method"], [8598, 8600, ["Gaussian", "distributions"], "Method"], [8605, 8608, ["MMD", "∝", "SMMD"], "Method"], [8622, 8623, ["SMMD"], "Method"], [8633, 8634, ["MMD"], "Method"], [8644, 8645, ["IGMs"], "Method"], [8661, 8666, ["empirical", "mean", "estimator", "of", "η"], "Method"], [8677, 8685, ["large", "but", "approximately", "low", "-", "rank", "matrix", "system"], "Task"], [8689, 8692, ["incomplete", "Cholesky", "decomposition"], "Method"], [8699, 8702, ["Woodbury", "matrix", "identity"], "Method"], [8734, 8735, ["backpropagation"], "Method"], [8749, 8750, ["CPU"], "Method"], [8763, 8764, ["MNIST"], "Material"], [8773, 8774, ["optimization"], "Task"], [8789, 8791, ["DCGAN", "architecture"], "Method"], [8814, 8816, ["spectral", "normalization"], "Method"], [8848, 8850, ["spectral", "normalization"], "Method"], [8864, 8867, ["generator", "loss", "estimates"], "Metric"], [8868, 8869, ["training"], "Task"], [8880, 8882, ["KID", "scores"], "Metric"], [8976, 8979, ["spectral", "normalization", "variants"], "Method"], [8984, 8985, ["MMDGAN"], "Method"], [8988, 8990, ["spectral", "normalization"], "Method"], [8992, 8994, ["gradient", "penalty"], "Method"], [9022, 9025, ["critic", "'s", "network"], "Method"], [9032, 9033, ["training"], "Task"], [9041, 9043, ["spectral", "parametrization"], "Method"], [9049, 9050, ["CelebA"], "Material"], [9051, 9052, ["SMMD"], "Method"], [9061, 9063, ["Inception", "score"], "Metric"], [9067, 9069, ["Spectral", "normalization"], "Method"], [9281, 9285, ["two", "-", "layer", "network"], "Method"], [9324, 9328, ["Lipschitz", "constant", "of", "φ"], "Metric"], [9330, 9333, ["expected", "squared", "norm"], "Metric"], [9363, 9364, ["estimator"], "Method"], [9465, 9468, ["generalized", "representer", "theorem"], "Method"], [9566, 9568, ["optimization", "problem"], "Task"], [9573, 9575, ["linear", "problem"], "Task"], [9586, 9589, ["convex", "optimization", "software"], "Method"], [9624, 9629, ["1", "-", "dimensional", "toy", "problem"], "Task"], [9654, 9656, ["kernel", "parameterization"], "Method"], [9692, 9694, ["envelope", "theorem"], "Method"], [4, 5, ["MMD"], "Method"], [23, 24, ["GAN"], "Method"], [57, 59, ["loss", "function"], "Metric"], [104, 106, ["image", "generation"], "Task"], [134, 135, ["MMD"], "Method"], [162, 163, ["MMD"], "Method"], [196, 197, ["GAN"], "Method"], [264, 265, ["GAN"], "Method"], [325, 326, ["GAN"], "Method"], [405, 406, ["MMD"], "Method"], [406, 407, ["GAN"], "Method"], [441, 442, ["MMD"], "Method"], [587, 588, ["MMD"], "Method"], [707, 708, ["MMD"], "Method"], [765, 766, ["MMD"], "Method"], [988, 989, ["MMD"], "Method"], [1074, 1075, ["MMD"], "Method"], [1230, 1231, ["MMD"], "Method"], [1261, 1262, ["MMD"], "Method"], [1290, 1291, ["MMD"], "Method"], [1466, 1467, ["MMD"], "Method"], [1483, 1484, ["MMD"], "Method"], [1523, 1524, ["GAN"], "Method"], [1538, 1539, ["MMD"], "Method"], [1547, 1548, ["MMD"], "Method"], [1571, 1572, ["MMD"], "Method"], [1639, 1640, ["MMD"], "Method"], [1728, 1729, ["MMD"], "Method"], [1859, 1860, ["MMD"], "Method"], [2018, 2021, ["Maximum", "Mean", "Discrepancy"], "Method"], [2026, 2027, ["MMD"], "Method"], [2067, 2068, ["MMD"], "Method"], [2220, 2223, ["Maximum", "Mean", "Discrepancy"], "Method"], [2229, 2230, ["MMD"], "Method"], [2594, 2595, ["GAN"], "Method"], [2696, 2697, ["GAN"], "Method"], [2703, 2706, ["Maximum", "Mean", "Discrepancy"], "Method"], [2718, 2719, ["MMD"], "Method"], [2742, 2745, ["Maximum", "Mean", "Discrepancy"], "Method"], [2855, 2856, ["SMMD"], "Method"], [2904, 2905, ["SMMD"], "Method"], [2914, 2915, ["SMMD"], "Method"], [2995, 2999, ["D", "µ", ",", "Ψ"], "Metric"], [3013, 3014, ["SMMD"], "Method"], [3599, 3601, ["image", "generation"], "Task"], [3685, 3686, ["MMD"], "Method"], [3687, 3688, ["GAN"], "Method"], [3730, 3731, ["GAN"], "Method"], [3766, 3767, ["MMD"], "Method"], [3879, 3880, ["GAN"], "Method"], [4058, 4059, ["SN"], "Method"], [4077, 4078, ["SN"], "Method"], [4079, 4080, ["MMDGAN"], "Method"], [4081, 4082, ["SN"], "Method"], [4087, 4088, ["SN"], "Method"], [4144, 4145, ["MMD"], "Method"], [4153, 4154, ["MMD"], "Method"], [4430, 4431, ["MMD"], "Method"], [4649, 4650, ["GAN"], "Method"], [4666, 4667, ["GAN"], "Method"], [4846, 4847, ["CelebA"], "Material"], [4918, 4919, ["MMD"], "Method"], [4936, 4937, ["MMD"], "Method"], [4948, 4949, ["MMD"], "Method"], [4954, 4955, ["MMD"], "Method"], [5018, 5019, ["GAN"], "Method"], [5528, 5529, ["MMD"], "Method"], [5558, 5559, ["MMD"], "Method"], [5867, 5868, ["MMD"], "Method"], [6343, 6344, ["MMD"], "Method"], [6548, 6549, ["MMD"], "Method"], [7081, 7085, ["D", "µ", ",", "Ψ"], "Metric"], [7088, 7089, ["SMMD"], "Method"], [8110, 8111, ["MMD"], "Method"], [8132, 8133, ["MMD"], "Method"], [8173, 8174, ["MMD"], "Method"], [8214, 8215, ["MMD"], "Method"], [8261, 8262, ["MMD"], "Method"], [8264, 8265, ["GAN"], "Method"], [8347, 8348, ["MMD"], "Method"], [8376, 8377, ["MMD"], "Method"], [8396, 8397, ["GAN"], "Method"], [8497, 8498, ["MMD"], "Method"], [8510, 8511, ["MMD"], "Method"], [8573, 8574, ["MMD"], "Method"], [8576, 8577, ["MMD"], "Method"], [8590, 8591, ["SMMD"], "Method"], [8595, 8596, ["MMD"], "Method"], [8611, 8612, ["MMD"], "Method"], [8626, 8627, ["MMD"], "Method"], [8650, 8651, ["MMD"], "Method"], [8885, 8886, ["MNIST"], "Material"], [8896, 8897, ["SMMD"], "Method"], [8904, 8905, ["SMMD"], "Method"], [8913, 8914, ["MMD"], "Method"], [9071, 9072, ["MMD"], "Method"], [9366, 9367, ["MMD"], "Method"], [9376, 9377, ["MMD"], "Method"], [9454, 9455, ["MMD"], "Method"]]} {"doc_id": "3f55d26dd638c849745b95e912c28d88445ba5e1", "text": ["document", ":", "Supervised", "Learning", "of", "Universal", "Sentence", "Representations", "from", "Natural", "Language", "Inference", "Data", "Many", "modern", "NLP", "systems", "rely", "on", "word", "embeddings", ",", "previously", "trained", "in", "an", "unsupervised", "manner", "on", "large", "corpora", ",", "as", "base", "features", ".", "Efforts", "to", "obtain", "embeddings", "for", "larger", "chunks", "of", "text", ",", "such", "as", "sentences", ",", "have", "however", "not", "been", "so", "successful", ".", "Several", "attempts", "at", "learning", "unsupervised", "representations", "of", "sentences", "have", "not", "reached", "satisfactory", "enough", "performance", "to", "be", "widely", "adopted", ".", "In", "this", "paper", ",", "we", "show", "how", "universal", "sentence", "representations", "trained", "using", "the", "supervised", "data", "of", "the", "Stanford", "Natural", "Language", "Inference", "datasets", "can", "consistently", "outperform", "unsupervised", "methods", "like", "SkipThought", "vectors", "on", "a", "wide", "range", "of", "transfer", "tasks", ".", "Much", "like", "how", "computer", "vision", "uses", "ImageNet", "to", "obtain", "features", ",", "which", "can", "then", "be", "transferred", "to", "other", "tasks", ",", "our", "work", "tends", "to", "indicate", "the", "suitability", "of", "natural", "language", "inference", "for", "transfer", "learning", "to", "other", "NLP", "tasks", ".", "Our", "encoder", "is", "publicly", "available", ".", "section", ":", "Introduction", "Distributed", "representations", "of", "words", "(", "or", "word", "embeddings", ")", "have", "shown", "to", "provide", "useful", "features", "for", "various", "tasks", "in", "natural", "language", "processing", "and", "computer", "vision", ".", "While", "there", "seems", "to", "be", "a", "consensus", "concerning", "the", "usefulness", "of", "word", "embeddings", "and", "how", "to", "learn", "them", ",", "this", "is", "not", "yet", "clear", "with", "regard", "to", "representations", "that", "carry", "the", "meaning", "of", "a", "full", "sentence", ".", "That", "is", ",", "how", "to", "capture", "the", "relationships", "among", "multiple", "words", "and", "phrases", "in", "a", "single", "vector", "remains", "an", "question", "to", "be", "solved", ".", "In", "this", "paper", ",", "we", "study", "the", "task", "of", "learning", "universal", "representations", "of", "sentences", ",", "i.e.", ",", "a", "sentence", "encoder", "model", "that", "is", "trained", "on", "a", "large", "corpus", "and", "subsequently", "transferred", "to", "other", "tasks", ".", "Two", "questions", "need", "to", "be", "solved", "in", "order", "to", "build", "such", "an", "encoder", ",", "namely", ":", "what", "is", "the", "preferable", "neural", "network", "architecture", ";", "and", "how", "and", "on", "what", "task", "should", "such", "a", "network", "be", "trained", ".", "Following", "existing", "work", "on", "learning", "word", "embeddings", ",", "most", "current", "approaches", "consider", "learning", "sentence", "encoders", "in", "an", "unsupervised", "manner", "like", "SkipThought", "or", "FastSent", ".", "Here", ",", "we", "investigate", "whether", "supervised", "learning", "can", "be", "leveraged", "instead", ",", "taking", "inspiration", "from", "previous", "results", "in", "computer", "vision", ",", "where", "many", "models", "are", "pretrained", "on", "the", "ImageNet", "before", "being", "transferred", ".", "We", "compare", "sentence", "embeddings", "trained", "on", "various", "supervised", "tasks", ",", "and", "show", "that", "sentence", "embeddings", "generated", "from", "models", "trained", "on", "a", "natural", "language", "inference", "(", "NLI", ")", "task", "reach", "the", "best", "results", "in", "terms", "of", "transfer", "accuracy", ".", "We", "hypothesize", "that", "the", "suitability", "of", "NLI", "as", "a", "training", "task", "is", "caused", "by", "the", "fact", "that", "it", "is", "a", "high", "-", "level", "understanding", "task", "that", "involves", "reasoning", "about", "the", "semantic", "relationships", "within", "sentences", ".", "Unlike", "in", "computer", "vision", ",", "where", "convolutional", "neural", "networks", "are", "predominant", ",", "there", "are", "multiple", "ways", "to", "encode", "a", "sentence", "using", "neural", "networks", ".", "Hence", ",", "we", "investigate", "the", "impact", "of", "the", "sentence", "encoding", "architecture", "on", "representational", "transferability", ",", "and", "compare", "convolutional", ",", "recurrent", "and", "even", "simpler", "word", "composition", "schemes", ".", "Our", "experiments", "show", "that", "an", "encoder", "based", "on", "a", "bi", "-", "directional", "LSTM", "architecture", "with", "max", "pooling", ",", "trained", "on", "the", "Stanford", "Natural", "Language", "Inference", "(", "SNLI", ")", "dataset", ",", "yields", "state", "-", "of", "-", "the", "-", "art", "sentence", "embeddings", "compared", "to", "all", "existing", "alternative", "unsupervised", "approaches", "like", "SkipThought", "or", "FastSent", ",", "while", "being", "much", "faster", "to", "train", ".", "We", "establish", "this", "finding", "on", "a", "broad", "and", "diverse", "set", "of", "transfer", "tasks", "that", "measures", "the", "ability", "of", "sentence", "representations", "to", "capture", "general", "and", "useful", "information", ".", "section", ":", "Related", "work", "Transfer", "learning", "using", "supervised", "features", "has", "been", "successful", "in", "several", "computer", "vision", "applications", ".", "Striking", "examples", "include", "face", "recognition", "and", "visual", "question", "answering", ",", "where", "image", "features", "trained", "on", "ImageNet", "and", "word", "embeddings", "trained", "on", "large", "unsupervised", "corpora", "are", "combined", ".", "In", "contrast", ",", "most", "approaches", "for", "sentence", "representation", "learning", "are", "unsupervised", ",", "arguably", "because", "the", "NLP", "community", "has", "not", "yet", "found", "the", "best", "supervised", "task", "for", "embedding", "the", "semantics", "of", "a", "whole", "sentence", ".", "Another", "reason", "is", "that", "neural", "networks", "are", "very", "good", "at", "capturing", "the", "biases", "of", "the", "task", "on", "which", "they", "are", "trained", ",", "but", "can", "easily", "forget", "the", "overall", "information", "or", "semantics", "of", "the", "input", "data", "by", "specializing", "too", "much", "on", "these", "biases", ".", "Learning", "models", "on", "large", "unsupervised", "task", "makes", "it", "harder", "for", "the", "model", "to", "specialize", ".", "littwin2016multiverse", "showed", "that", "co", "-", "adaptation", "of", "encoders", "and", "classifiers", ",", "when", "trained", "end", "-", "to", "-", "end", ",", "can", "negatively", "impact", "the", "generalization", "power", "of", "image", "features", "generated", "by", "an", "encoder", ".", "They", "propose", "a", "loss", "that", "incorporates", "multiple", "orthogonal", "classifiers", "to", "counteract", "this", "effect", ".", "Recent", "work", "on", "generating", "sentence", "embeddings", "range", "from", "models", "that", "compose", "word", "embeddings", "to", "more", "complex", "neural", "network", "architectures", ".", "SkipThought", "vectors", "propose", "an", "objective", "function", "that", "adapts", "the", "skip", "-", "gram", "model", "for", "words", "to", "the", "sentence", "level", ".", "By", "encoding", "a", "sentence", "to", "predict", "the", "sentences", "around", "it", ",", "and", "using", "the", "features", "in", "a", "linear", "model", ",", "they", "were", "able", "to", "demonstrate", "good", "performance", "on", "8", "transfer", "tasks", ".", "They", "further", "obtained", "better", "results", "using", "layer", "-", "norm", "regularization", "of", "their", "model", "in", ".", "hill2016learning", "showed", "that", "the", "task", "on", "which", "sentence", "embeddings", "are", "trained", "significantly", "impacts", "their", "quality", ".", "In", "addition", "to", "unsupervised", "methods", ",", "they", "included", "supervised", "training", "in", "their", "comparison", "—", "namely", ",", "on", "machine", "translation", "data", "(", "using", "the", "WMT’14", "English", "/", "French", "and", "English", "/", "German", "pairs", ")", ",", "dictionary", "definitions", "and", "image", "captioning", "data", "(", "see", "also", "kiela2017learning", ")", "from", "the", "COCO", "dataset", ".", "These", "models", "obtained", "significantly", "lower", "results", "compared", "to", "the", "unsupervised", "Skip", "-", "Thought", "approach", ".", "Recent", "work", "has", "explored", "training", "sentence", "encoders", "on", "the", "SNLI", "corpus", "and", "applying", "them", "on", "the", "SICK", "corpus", ",", "either", "using", "multi", "-", "task", "learning", "or", "pretraining", ".", "The", "results", "were", "inconclusive", "and", "did", "not", "reach", "the", "same", "level", "as", "simpler", "approaches", "that", "directly", "learn", "a", "classifier", "on", "top", "of", "unsupervised", "sentence", "embeddings", "instead", ".", "To", "our", "knowledge", ",", "this", "work", "is", "the", "first", "attempt", "to", "fully", "exploit", "the", "SNLI", "corpus", "for", "building", "generic", "sentence", "encoders", ".", "As", "we", "show", "in", "our", "experiments", ",", "we", "are", "able", "to", "consistently", "outperform", "unsupervised", "approaches", ",", "even", "if", "our", "models", "are", "trained", "on", "much", "less", "(", "but", "human", "-", "annotated", ")", "data", ".", "section", ":", "Approach", "This", "work", "combines", "two", "research", "directions", ",", "which", "we", "describe", "in", "what", "follows", ".", "First", ",", "we", "explain", "how", "the", "NLI", "task", "can", "be", "used", "to", "train", "universal", "sentence", "encoding", "models", "using", "the", "SNLI", "task", ".", "We", "subsequently", "describe", "the", "architectures", "that", "we", "investigated", "for", "the", "sentence", "encoder", ",", "which", ",", "in", "our", "opinion", ",", "covers", "a", "suitable", "range", "of", "sentence", "encoders", "currently", "in", "use", ".", "Specifically", ",", "we", "examine", "standard", "recurrent", "models", "such", "as", "LSTMs", "and", "GRUs", ",", "for", "which", "we", "investigate", "mean", "and", "max", "-", "pooling", "over", "the", "hidden", "representations", ";", "a", "self", "-", "attentive", "network", "that", "incorporates", "different", "views", "of", "the", "sentence", ";", "and", "a", "hierarchical", "convolutional", "network", "that", "can", "be", "seen", "as", "a", "tree", "-", "based", "method", "that", "blends", "different", "levels", "of", "abstraction", ".", "subsection", ":", "The", "Natural", "Language", "Inference", "task", "The", "SNLI", "dataset", "consists", "of", "570k", "human", "-", "generated", "English", "sentence", "pairs", ",", "manually", "labeled", "with", "one", "of", "three", "categories", ":", "entailment", ",", "contradiction", "and", "neutral", ".", "It", "captures", "natural", "language", "inference", ",", "also", "known", "in", "previous", "incarnations", "as", "Recognizing", "Textual", "Entailment", "(", "RTE", ")", ",", "and", "constitutes", "one", "of", "the", "largest", "high", "-", "quality", "labeled", "resources", "explicitly", "constructed", "in", "order", "to", "require", "understanding", "sentence", "semantics", ".", "We", "hypothesize", "that", "the", "semantic", "nature", "of", "NLI", "makes", "it", "a", "good", "candidate", "for", "learning", "universal", "sentence", "embeddings", "in", "a", "supervised", "way", ".", "That", "is", ",", "we", "aim", "to", "demonstrate", "that", "sentence", "encoders", "trained", "on", "natural", "language", "inference", "are", "able", "to", "learn", "sentence", "representations", "that", "capture", "universally", "useful", "features", ".", "Models", "can", "be", "trained", "on", "SNLI", "in", "two", "different", "ways", ":", "(", "i", ")", "sentence", "encoding", "-", "based", "models", "that", "explicitly", "separate", "the", "encoding", "of", "the", "individual", "sentences", "and", "(", "ii", ")", "joint", "methods", "that", "allow", "to", "use", "encoding", "of", "both", "sentences", "(", "to", "use", "cross", "-", "features", "or", "attention", "from", "one", "sentence", "to", "the", "other", ")", ".", "Since", "our", "goal", "is", "to", "train", "a", "generic", "sentence", "encoder", ",", "we", "adopt", "the", "first", "setting", ".", "As", "illustrated", "in", "Figure", "[", "reference", "]", ",", "a", "typical", "architecture", "of", "this", "kind", "uses", "a", "shared", "sentence", "encoder", "that", "outputs", "a", "representation", "for", "the", "premise", "and", "the", "hypothesis", ".", "Once", "the", "sentence", "vectors", "are", "generated", ",", "3", "matching", "methods", "are", "applied", "to", "extract", "relations", "between", "and", ":", "(", "i", ")", "concatenation", "of", "the", "two", "representations", ";", "(", "ii", ")", "element", "-", "wise", "product", ";", "and", "(", "iii", ")", "absolute", "element", "-", "wise", "difference", ".", "The", "resulting", "vector", ",", "which", "captures", "information", "from", "both", "the", "premise", "and", "the", "hypothesis", ",", "is", "fed", "into", "a", "3", "-", "class", "classifier", "consisting", "of", "multiple", "fully", "-", "connected", "layers", "culminating", "in", "a", "softmax", "layer", ".", "subsection", ":", "Sentence", "encoder", "architectures", "A", "wide", "variety", "of", "neural", "networks", "for", "encoding", "sentences", "into", "fixed", "-", "size", "representations", "exists", ",", "and", "it", "is", "not", "yet", "clear", "which", "one", "best", "captures", "generically", "useful", "information", ".", "We", "compare", "7", "different", "architectures", ":", "standard", "recurrent", "encoders", "with", "either", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "or", "Gated", "Recurrent", "Units", "(", "GRU", ")", ",", "concatenation", "of", "last", "hidden", "states", "of", "forward", "and", "backward", "GRU", ",", "Bi", "-", "directional", "LSTMs", "(", "BiLSTM", ")", "with", "either", "mean", "or", "max", "pooling", ",", "self", "-", "attentive", "network", "and", "hierarchical", "convolutional", "networks", ".", "subsubsection", ":", "LSTM", "and", "GRU", "Our", "first", ",", "and", "simplest", ",", "encoders", "apply", "recurrent", "neural", "networks", "using", "either", "LSTM", "or", "GRU", "modules", ",", "as", "in", "sequence", "to", "sequence", "encoders", ".", "For", "a", "sequence", "of", "words", ",", "the", "network", "computes", "a", "set", "of", "hidden", "representations", ",", "with", "(", "or", "using", "GRU", "units", "instead", ")", ".", "A", "sentence", "is", "represented", "by", "the", "last", "hidden", "vector", ",", ".", "We", "also", "consider", "a", "model", "BiGRU", "-", "last", "that", "concatenates", "the", "last", "hidden", "state", "of", "a", "forward", "GRU", ",", "and", "the", "last", "hidden", "state", "of", "a", "backward", "GRU", "to", "have", "the", "same", "architecture", "as", "for", "SkipThought", "vectors", ".", "subsubsection", ":", "BiLSTM", "with", "mean", "/", "max", "pooling", "For", "a", "sequence", "of", "T", "words", ",", "a", "bidirectional", "LSTM", "computes", "a", "set", "of", "T", "vectors", ".", "For", ",", ",", "is", "the", "concatenation", "of", "a", "forward", "LSTM", "and", "a", "backward", "LSTM", "that", "read", "the", "sentences", "in", "two", "opposite", "directions", ":", "We", "experiment", "with", "two", "ways", "of", "combining", "the", "varying", "number", "of", "to", "form", "a", "fixed", "-", "size", "vector", ",", "either", "by", "selecting", "the", "maximum", "value", "over", "each", "dimension", "of", "the", "hidden", "units", "(", "max", "pooling", ")", "or", "by", "considering", "the", "average", "of", "the", "representations", "(", "mean", "pooling", ")", ".", "subsubsection", ":", "Self", "-", "attentive", "network", "The", "self", "-", "attentive", "sentence", "encoder", "uses", "an", "attention", "mechanism", "over", "the", "hidden", "states", "of", "a", "BiLSTM", "to", "generate", "a", "representation", "of", "an", "input", "sentence", ".", "The", "attention", "mechanism", "is", "defined", "as", ":", "where", "are", "the", "output", "hidden", "vectors", "of", "a", "BiLSTM", ".", "These", "are", "fed", "to", "an", "affine", "transformation", "(", ",", ")", "which", "outputs", "a", "set", "of", "keys", ".", "The", "represent", "the", "score", "of", "similarity", "between", "the", "keys", "and", "a", "learned", "context", "query", "vector", ".", "These", "weights", "are", "used", "to", "produce", "the", "final", "representation", ",", "which", "is", "a", "weighted", "linear", "combination", "of", "the", "hidden", "vectors", ".", "Following", "lin2017structured", "we", "use", "a", "self", "-", "attentive", "network", "with", "multiple", "views", "of", "the", "input", "sentence", ",", "so", "that", "the", "model", "can", "learn", "which", "part", "of", "the", "sentence", "is", "important", "for", "the", "given", "task", ".", "Concretely", ",", "we", "have", "4", "context", "vectors", "which", "generate", "4", "representations", "that", "are", "then", "concatenated", "to", "obtain", "the", "sentence", "representation", ".", "Figure", "[", "reference", "]", "illustrates", "this", "architecture", ".", "subsubsection", ":", "Hierarchical", "ConvNet", "One", "of", "the", "currently", "best", "performing", "models", "on", "classification", "tasks", "is", "a", "convolutional", "architecture", "termed", "AdaSent", ",", "which", "concatenates", "different", "representations", "of", "the", "sentences", "at", "different", "level", "of", "abstractions", ".", "Inspired", "by", "this", "architecture", ",", "we", "introduce", "a", "faster", "version", "consisting", "of", "4", "convolutional", "layers", ".", "At", "every", "layer", ",", "a", "representation", "is", "computed", "by", "a", "max", "-", "pooling", "operation", "over", "the", "feature", "maps", "(", "see", "Figure", "[", "reference", "]", ")", ".", "The", "final", "representation", "concatenates", "representations", "at", "different", "levels", "of", "the", "input", "sentence", ".", "The", "model", "thus", "captures", "hierarchical", "abstractions", "of", "an", "input", "sentence", "in", "a", "fixed", "-", "size", "representation", ".", "subsection", ":", "Training", "details", "For", "all", "our", "models", "trained", "on", "SNLI", ",", "we", "use", "SGD", "with", "a", "learning", "rate", "of", "0.1", "and", "a", "weight", "decay", "of", "0.99", ".", "At", "each", "epoch", ",", "we", "divide", "the", "learning", "rate", "by", "5", "if", "the", "dev", "accuracy", "decreases", ".", "We", "use", "mini", "-", "batches", "of", "size", "64", "and", "training", "is", "stopped", "when", "the", "learning", "rate", "goes", "under", "the", "threshold", "of", ".", "For", "the", "classifier", ",", "we", "use", "a", "multi", "-", "layer", "perceptron", "with", "1", "hidden", "-", "layer", "of", "512", "hidden", "units", ".", "We", "use", "open", "-", "source", "GloVe", "vectors", "trained", "on", "Common", "Crawl", "840B", "with", "300", "dimensions", "as", "fixed", "word", "embeddings", ".", "section", ":", "Evaluation", "of", "sentence", "representations", "Our", "aim", "is", "to", "obtain", "general", "-", "purpose", "sentence", "embeddings", "that", "capture", "generic", "information", "that", "is", "useful", "for", "a", "broad", "set", "of", "tasks", ".", "To", "evaluate", "the", "quality", "of", "these", "representations", ",", "we", "use", "them", "as", "features", "in", "12", "transfer", "tasks", ".", "We", "present", "our", "sentence", "-", "embedding", "evaluation", "procedure", "in", "this", "section", ".", "We", "constructed", "a", "sentence", "evaluation", "tool", "called", "SentEval", "to", "automate", "evaluation", "on", "all", "the", "tasks", "mentioned", "in", "this", "paper", ".", "The", "tool", "uses", "Adam", "to", "fit", "a", "logistic", "regression", "classifier", ",", "with", "batch", "size", "64", ".", "paragraph", ":", "Binary", "and", "multi", "-", "class", "classification", "We", "use", "a", "set", "of", "binary", "classification", "tasks", "(", "see", "Table", "[", "reference", "]", ")", "that", "covers", "various", "types", "of", "sentence", "classification", ",", "including", "sentiment", "analysis", "(", "MR", ",", "SST", ")", ",", "question", "-", "type", "(", "TREC", ")", ",", "product", "reviews", "(", "CR", ")", ",", "subjectivity", "/", "objectivity", "(", "SUBJ", ")", "and", "opinion", "polarity", "(", "MPQA", ")", ".", "We", "generate", "sentence", "vectors", "and", "train", "a", "logistic", "regression", "on", "top", ".", "A", "linear", "classifier", "requires", "fewer", "parameters", "than", "an", "MLP", "and", "is", "thus", "suitable", "for", "small", "datasets", ",", "where", "transfer", "learning", "is", "especially", "well", "-", "suited", ".", "We", "tune", "the", "L2", "penalty", "of", "the", "logistic", "regression", "with", "grid", "-", "search", "on", "the", "validation", "set", ".", "paragraph", ":", "Entailment", "and", "semantic", "relatedness", "We", "also", "evaluate", "on", "the", "SICK", "dataset", "for", "both", "entailment", "(", "SICK", "-", "E", ")", "and", "semantic", "relatedness", "(", "SICK", "-", "R", ")", ".", "We", "use", "the", "same", "matching", "methods", "as", "in", "SNLI", "and", "learn", "a", "Logistic", "Regression", "on", "top", "of", "the", "joint", "representation", ".", "For", "semantic", "relatedness", "evaluation", ",", "we", "follow", "the", "approach", "of", "and", "learn", "to", "predict", "the", "probability", "distribution", "of", "relatedness", "scores", ".", "We", "report", "Pearson", "correlation", ".", "paragraph", ":", "STS14", "-", "Semantic", "Textual", "Similarity", "While", "semantic", "relatedness", "is", "supervised", "in", "the", "case", "of", "SICK", "-", "R", ",", "we", "also", "evaluate", "our", "embeddings", "on", "the", "6", "unsupervised", "SemEval", "tasks", "of", "STS14", ".", "This", "dataset", "includes", "subsets", "of", "news", "articles", ",", "forum", "discussions", ",", "image", "descriptions", "and", "headlines", "from", "news", "articles", "containing", "pairs", "of", "sentences", "(", "lower", "-", "cased", ")", ",", "labeled", "with", "a", "similarity", "score", "between", "0", "and", "5", ".", "These", "tasks", "evaluate", "how", "the", "cosine", "distance", "between", "two", "sentences", "correlate", "with", "a", "human", "-", "labeled", "similarity", "score", "through", "Pearson", "and", "Spearman", "correlations", ".", "paragraph", ":", "Paraphrase", "detection", "The", "Microsoft", "Research", "Paraphrase", "Corpus", "is", "composed", "of", "pairs", "of", "sentences", "which", "have", "been", "extracted", "from", "news", "sources", "on", "the", "Web", ".", "Sentence", "pairs", "have", "been", "human", "-", "annotated", "according", "to", "whether", "they", "capture", "a", "paraphrase", "/", "semantic", "equivalence", "relationship", ".", "We", "use", "the", "same", "approach", "as", "with", "SICK", "-", "E", ",", "except", "that", "our", "classifier", "has", "only", "2", "classes", ".", "paragraph", ":", "Caption", "-", "Image", "retrieval", "The", "caption", "-", "image", "retrieval", "task", "evaluates", "joint", "image", "and", "language", "feature", "models", ".", "The", "goal", "is", "either", "to", "rank", "a", "large", "collection", "of", "images", "by", "their", "relevance", "with", "respect", "to", "a", "given", "query", "caption", "(", "Image", "Retrieval", ")", ",", "or", "ranking", "captions", "by", "their", "relevance", "for", "a", "given", "query", "image", "(", "Caption", "Retrieval", ")", ".", "We", "use", "a", "pairwise", "ranking", "-", "loss", ":", "where", "consists", "of", "an", "image", "with", "one", "of", "its", "associated", "captions", ",", "and", "are", "negative", "examples", "of", "the", "ranking", "loss", ",", "is", "the", "margin", "and", "corresponds", "to", "the", "cosine", "similarity", ".", "and", "are", "learned", "linear", "transformations", "that", "project", "the", "caption", "and", "the", "image", "to", "the", "same", "embedding", "space", ".", "We", "use", "a", "margin", "and", "contrastive", "terms", ".", "We", "use", "the", "same", "splits", "as", "in", ",", "i.e.", ",", "we", "use", "113k", "images", "from", "the", "COCO", "dataset", "(", "each", "containing", "5", "captions", ")", "for", "training", ",", "5k", "images", "for", "validation", "and", "5k", "images", "for", "test", ".", "For", "evaluation", ",", "we", "split", "the", "5k", "images", "in", "5", "random", "sets", "of", "1k", "images", "on", "which", "we", "compute", "Recall@K", ",", "with", "K", "and", "median", "(", "Med", "r", ")", "over", "the", "5", "splits", ".", "For", "fair", "comparison", ",", "we", "also", "report", "SkipThought", "results", "in", "our", "setting", ",", "using", "2048", "-", "dimensional", "pretrained", "ResNet", "-", "101", "with", "113k", "training", "images", ".", "section", ":", "Empirical", "results", "In", "this", "section", ",", "we", "refer", "to", "”", "micro", "”", "and", "”", "macro", "”", "averages", "of", "development", "set", "(", "dev", ")", "results", "on", "transfer", "tasks", "whose", "metrics", "is", "accuracy", ":", "we", "compute", "a", "”", "macro", "”", "aggregated", "score", "that", "corresponds", "to", "the", "classical", "average", "of", "dev", "accuracies", ",", "and", "the", "”", "micro", "”", "score", "that", "is", "a", "sum", "of", "the", "dev", "accuracies", ",", "weighted", "by", "the", "number", "of", "dev", "samples", ".", "subsection", ":", "Architecture", "impact", "paragraph", ":", "Model", "We", "observe", "in", "Table", "[", "reference", "]", "that", "different", "models", "trained", "on", "the", "same", "NLI", "corpus", "lead", "to", "different", "transfer", "tasks", "results", ".", "The", "BiLSTM", "-", "4096", "with", "the", "max", "-", "pooling", "operation", "performs", "best", "on", "both", "SNLI", "and", "transfer", "tasks", ".", "Looking", "at", "the", "micro", "and", "macro", "averages", ",", "we", "see", "that", "it", "performs", "significantly", "better", "than", "the", "other", "models", "LSTM", ",", "GRU", ",", "BiGRU", "-", "last", ",", "BiLSTM", "-", "Mean", ",", "inner", "-", "attention", "and", "the", "hierarchical", "-", "ConvNet", ".", "Table", "[", "reference", "]", "also", "shows", "that", "better", "performance", "on", "the", "training", "task", "does", "not", "necessarily", "translate", "in", "better", "results", "on", "the", "transfer", "tasks", "like", "when", "comparing", "inner", "-", "attention", "and", "BiLSTM", "-", "Mean", "for", "instance", ".", "We", "hypothesize", "that", "some", "models", "are", "likely", "to", "over", "-", "specialize", "and", "adapt", "too", "well", "to", "the", "biases", "of", "a", "dataset", "without", "capturing", "general", "-", "purpose", "information", "of", "the", "input", "sentence", ".", "For", "example", ",", "the", "inner", "-", "attention", "model", "has", "the", "ability", "to", "focus", "only", "on", "certain", "parts", "of", "a", "sentence", "that", "are", "useful", "for", "the", "SNLI", "task", ",", "but", "not", "necessarily", "for", "the", "transfer", "tasks", ".", "On", "the", "other", "hand", ",", "BiLSTM", "-", "Mean", "does", "not", "make", "sharp", "choices", "on", "which", "part", "of", "the", "sentence", "is", "more", "important", "than", "others", ".", "The", "difference", "between", "the", "results", "seems", "to", "come", "from", "the", "different", "abilities", "of", "the", "models", "to", "incorporate", "general", "information", "while", "not", "focusing", "too", "much", "on", "specific", "features", "useful", "for", "the", "task", "at", "hand", ".", "For", "a", "given", "model", ",", "the", "transfer", "quality", "is", "also", "sensitive", "to", "the", "optimization", "algorithm", ":", "when", "training", "with", "Adam", "instead", "of", "SGD", ",", "we", "observed", "that", "the", "BiLSTM", "-", "max", "converged", "faster", "on", "SNLI", "(", "5", "epochs", "instead", "of", "10", ")", ",", "but", "obtained", "worse", "results", "on", "the", "transfer", "tasks", ",", "most", "likely", "because", "of", "the", "model", "and", "classifier", "’s", "increased", "capability", "to", "over", "-", "specialize", "on", "the", "training", "task", ".", "paragraph", ":", "Embedding", "size", "Figure", "[", "reference", "]", "compares", "the", "overall", "performance", "of", "different", "architectures", ",", "showing", "the", "evolution", "of", "micro", "averaged", "performance", "with", "regard", "to", "the", "embedding", "size", ".", "Since", "it", "is", "easier", "to", "linearly", "separate", "in", "high", "dimension", ",", "especially", "with", "logistic", "regression", ",", "it", "is", "not", "surprising", "that", "increased", "embedding", "sizes", "lead", "to", "increased", "performance", "for", "almost", "all", "models", ".", "However", ",", "this", "is", "particularly", "true", "for", "some", "models", "(", "BiLSTM", "-", "Max", ",", "HConvNet", ",", "inner", "-", "att", ")", ",", "which", "demonstrate", "unequal", "abilities", "to", "incorporate", "more", "information", "as", "the", "size", "grows", ".", "We", "hypothesize", "that", "such", "networks", "are", "able", "to", "incorporate", "information", "that", "is", "not", "directly", "relevant", "to", "the", "objective", "task", "(", "results", "on", "SNLI", "are", "relatively", "stable", "with", "regard", "to", "embedding", "size", ")", "but", "that", "can", "nevertheless", "be", "useful", "as", "features", "for", "transfer", "tasks", ".", "subsection", ":", "Task", "transfer", "We", "report", "in", "Table", "[", "reference", "]", "transfer", "tasks", "results", "for", "different", "architectures", "trained", "in", "different", "ways", ".", "We", "group", "models", "by", "the", "nature", "of", "the", "data", "on", "which", "they", "were", "trained", ".", "The", "first", "group", "corresponds", "to", "models", "trained", "with", "unsupervised", "unordered", "sentences", ".", "This", "includes", "bag", "-", "of", "-", "words", "models", "such", "as", "word2vec", "-", "SkipGram", ",", "the", "Unigram", "-", "TFIDF", "model", ",", "the", "Paragraph", "Vector", "model", ",", "the", "Sequential", "Denoising", "Auto", "-", "Encoder", "(", "SDAE", ")", "and", "the", "SIF", "model", ",", "all", "trained", "on", "the", "Toronto", "book", "corpus", ".", "The", "second", "group", "consists", "of", "models", "trained", "with", "unsupervised", "ordered", "sentences", "such", "as", "FastSent", "and", "SkipThought", "(", "also", "trained", "on", "the", "Toronto", "book", "corpus", ")", ".", "We", "also", "include", "the", "FastSent", "variant", "“", "FastSent", "+", "AE", "”", "and", "the", "SkipThought", "-", "LN", "version", "that", "uses", "layer", "normalization", ".", "We", "report", "results", "from", "models", "trained", "on", "supervised", "data", "in", "the", "third", "group", ",", "and", "also", "report", "some", "results", "of", "supervised", "methods", "trained", "directly", "on", "each", "task", "for", "comparison", "with", "transfer", "learning", "approaches", ".", "paragraph", ":", "Comparison", "with", "SkipThought", "The", "best", "performing", "sentence", "encoder", "to", "date", "is", "the", "SkipThought", "-", "LN", "model", ",", "which", "was", "trained", "on", "a", "very", "large", "corpora", "of", "ordered", "sentences", ".", "With", "much", "less", "data", "(", "570k", "compared", "to", "64", "M", "sentences", ")", "but", "with", "high", "-", "quality", "supervision", "from", "the", "SNLI", "dataset", ",", "we", "are", "able", "to", "consistently", "outperform", "the", "results", "obtained", "by", "SkipThought", "vectors", ".", "We", "train", "our", "model", "in", "less", "than", "a", "day", "on", "a", "single", "GPU", "compared", "to", "the", "best", "SkipThought", "-", "LN", "network", "trained", "for", "a", "month", ".", "Our", "BiLSTM", "-", "max", "trained", "on", "SNLI", "performs", "much", "better", "than", "released", "SkipThought", "vectors", "on", "MR", ",", "CR", ",", "MPQA", ",", "SST", ",", "MRPC", "-", "accuracy", ",", "SICK", "-", "R", ",", "SICK", "-", "E", "and", "STS14", "(", "see", "Table", "[", "reference", "]", ")", ".", "Except", "for", "the", "SUBJ", "dataset", ",", "it", "also", "performs", "better", "than", "SkipThought", "-", "LN", "on", "MR", ",", "CR", "and", "MPQA", ".", "We", "also", "observe", "by", "looking", "at", "the", "STS14", "results", "that", "the", "cosine", "metrics", "in", "our", "embedding", "space", "is", "much", "more", "semantically", "informative", "than", "in", "SkipThought", "embedding", "space", "(", "pearson", "score", "of", "0.68", "compared", "to", "0.29", "and", "0.44", "for", "ST", "and", "ST", "-", "LN", ")", ".", "We", "hypothesize", "that", "this", "is", "namely", "linked", "to", "the", "matching", "method", "of", "SNLI", "models", "which", "incorporates", "a", "notion", "of", "distance", "(", "element", "-", "wise", "product", "and", "absolute", "difference", ")", "during", "training", ".", "paragraph", ":", "NLI", "as", "a", "supervised", "training", "set", "Our", "findings", "indicate", "that", "our", "model", "trained", "on", "SNLI", "obtains", "much", "better", "overall", "results", "than", "models", "trained", "on", "other", "supervised", "tasks", "such", "as", "COCO", ",", "dictionary", "definitions", ",", "NMT", ",", "PPDB", "and", "SST", ".", "For", "SST", ",", "we", "tried", "exactly", "the", "same", "models", "as", "for", "SNLI", ";", "it", "is", "worth", "noting", "that", "SST", "is", "smaller", "than", "NLI", ".", "Our", "representations", "constitute", "higher", "-", "quality", "features", "for", "both", "classification", "and", "similarity", "tasks", ".", "One", "explanation", "is", "that", "the", "natural", "language", "inference", "task", "constrains", "the", "model", "to", "encode", "the", "semantic", "information", "of", "the", "input", "sentence", ",", "and", "that", "the", "information", "required", "to", "perform", "NLI", "is", "generally", "discriminative", "and", "informative", ".", "paragraph", ":", "Domain", "adaptation", "on", "SICK", "tasks", "Our", "transfer", "learning", "approach", "obtains", "better", "results", "than", "previous", "state", "-", "of", "-", "the", "-", "art", "on", "the", "SICK", "task", "-", "can", "be", "seen", "as", "an", "out", "-", "domain", "version", "of", "SNLI", "-", "for", "both", "entailment", "and", "relatedness", ".", "We", "obtain", "a", "pearson", "score", "of", "0.885", "on", "SICK", "-", "R", "while", "obtained", "0.868", ",", "and", "we", "obtain", "86.3", "%", "test", "accuracy", "on", "SICK", "-", "E", "while", "previous", "best", "hand", "-", "engineered", "models", "obtained", "84.5", "%", ".", "We", "also", "significantly", "outperformed", "previous", "transfer", "learning", "approaches", "on", "SICK", "-", "E", "that", "used", "the", "parameters", "of", "an", "LSTM", "model", "trained", "on", "SNLI", "to", "fine", "-", "tune", "on", "SICK", "(", "80.8", "%", "accuracy", ")", ".", "We", "hypothesize", "that", "our", "embeddings", "already", "contain", "the", "information", "learned", "from", "the", "in", "-", "domain", "task", ",", "and", "that", "learning", "only", "the", "classifier", "limits", "the", "number", "of", "parameters", "learned", "on", "the", "small", "out", "-", "domain", "task", ".", "paragraph", ":", "Image", "-", "caption", "retrieval", "results", "In", "Table", "[", "reference", "]", ",", "we", "report", "results", "for", "the", "COCO", "image", "-", "caption", "retrieval", "task", ".", "We", "report", "the", "mean", "recalls", "of", "5", "random", "splits", "of", "1", "K", "test", "images", ".", "When", "trained", "with", "ResNet", "features", "and", "30k", "more", "training", "data", ",", "the", "SkipThought", "vectors", "perform", "significantly", "better", "than", "the", "original", "setting", ",", "going", "from", "33.8", "to", "37.9", "for", "caption", "retrieval", "R@1", ",", "and", "from", "25.9", "to", "30.6", "on", "image", "retrieval", "R@1", ".", "Our", "approach", "pushes", "the", "results", "even", "further", ",", "from", "37.9", "to", "42.4", "on", "caption", "retrieval", ",", "and", "30.6", "to", "33.2", "on", "image", "retrieval", ".", "These", "results", "are", "comparable", "to", "previous", "approach", "of", "that", "did", "not", "do", "transfer", "but", "directly", "learned", "the", "sentence", "encoding", "on", "the", "image", "-", "caption", "retrieval", "task", ".", "This", "supports", "the", "claim", "that", "pre", "-", "trained", "representations", "such", "as", "ResNet", "image", "features", "and", "our", "sentence", "embeddings", "can", "achieve", "competitive", "results", "compared", "to", "features", "learned", "directly", "on", "the", "objective", "task", ".", "paragraph", ":", "MultiGenre", "NLI", "The", "MultiNLI", "corpus", "was", "recently", "released", "as", "a", "multi", "-", "genre", "version", "of", "SNLI", ".", "With", "433", "K", "sentence", "pairs", ",", "MultiNLI", "improves", "upon", "SNLI", "in", "its", "coverage", ":", "it", "contains", "ten", "distinct", "genres", "of", "written", "and", "spoken", "English", ",", "covering", "most", "of", "the", "complexity", "of", "the", "language", ".", "We", "augment", "Table", "4", "with", "our", "model", "trained", "on", "both", "SNLI", "and", "MultiNLI", "(", "AllNLI", ")", ".", "We", "observe", "a", "significant", "boost", "in", "performance", "overall", "compared", "to", "the", "model", "trained", "only", "on", "SLNI", ".", "Our", "model", "even", "reaches", "AdaSent", "performance", "on", "CR", ",", "suggesting", "that", "having", "a", "larger", "coverage", "for", "the", "training", "task", "helps", "learn", "even", "better", "general", "representations", ".", "On", "semantic", "textual", "similarity", "STS14", ",", "we", "are", "also", "competitive", "with", "PPDB", "based", "paragram", "-", "phrase", "embeddings", "with", "a", "pearson", "score", "of", "0.70", ".", "Interestingly", ",", "on", "caption", "-", "related", "transfer", "tasks", "such", "as", "the", "COCO", "image", "caption", "retrieval", "task", ",", "training", "our", "sentence", "encoder", "on", "other", "genres", "from", "MultiNLI", "does", "not", "degrade", "the", "performance", "compared", "to", "the", "model", "trained", "only", "SNLI", "(", "which", "contains", "mostly", "captions", ")", ",", "which", "confirms", "the", "generalization", "power", "of", "our", "embeddings", ".", "section", ":", "Conclusion", "This", "paper", "studies", "the", "effects", "of", "training", "sentence", "embeddings", "with", "supervised", "data", "by", "testing", "on", "12", "different", "transfer", "tasks", ".", "We", "showed", "that", "models", "learned", "on", "NLI", "can", "perform", "better", "than", "models", "trained", "in", "unsupervised", "conditions", "or", "on", "other", "supervised", "tasks", ".", "By", "exploring", "various", "architectures", ",", "we", "showed", "that", "a", "BiLSTM", "network", "with", "max", "pooling", "makes", "the", "best", "current", "universal", "sentence", "encoding", "methods", ",", "outperforming", "existing", "approaches", "like", "SkipThought", "vectors", ".", "We", "believe", "that", "this", "work", "only", "scratches", "the", "surface", "of", "possible", "combinations", "of", "models", "and", "tasks", "for", "learning", "generic", "sentence", "embeddings", ".", "Larger", "datasets", "that", "rely", "on", "natural", "language", "understanding", "for", "sentences", "could", "bring", "sentence", "embedding", "quality", "to", "the", "next", "level", ".", "bibliography", ":", "References", "section", ":", "Appendix", "paragraph", ":", "Max", "-", "pooling", "visualization", "for", "BiLSTM", "-", "max", "trained", "and", "untrained", "Our", "representations", "were", "trained", "to", "focus", "on", "parts", "of", "a", "sentence", "such", "that", "a", "classifier", "can", "easily", "tell", "the", "difference", "between", "contradictory", ",", "neutral", "or", "entailed", "sentences", ".", "In", "Table", "[", "reference", "]", "and", "Table", "[", "reference", "]", ",", "we", "investigate", "how", "the", "max", "-", "pooling", "operation", "selects", "the", "information", "from", "the", "hidden", "states", "of", "the", "BiLSTM", ",", "for", "our", "trained", "and", "untrained", "BiLSTM", "-", "max", "models", "(", "for", "both", "models", ",", "word", "embeddings", "are", "initialized", "with", "GloVe", "vectors", ")", ".", "For", "each", "time", "step", ",", "we", "report", "the", "number", "of", "times", "the", "max", "-", "pooling", "operation", "selected", "the", "hidden", "state", "(", "which", "can", "be", "seen", "as", "a", "sentence", "representation", "centered", "around", "word", ")", ".", "Without", "any", "training", ",", "the", "max", "-", "pooling", "is", "rather", "even", "across", "hidden", "states", ",", "although", "it", "seems", "to", "focus", "consistently", "more", "on", "the", "first", "and", "last", "hidden", "states", ".", "When", "trained", ",", "the", "model", "learns", "to", "focus", "on", "specific", "words", "that", "carry", "most", "of", "the", "meaning", "of", "the", "sentence", "without", "any", "explicit", "attention", "mechanism", ".", "Note", "that", "each", "hidden", "state", "also", "incorporates", "information", "from", "the", "sentence", "at", "different", "levels", ",", "explaining", "why", "the", "trained", "model", "also", "incorporates", "information", "from", "all", "hidden", "states", "."], "entity": [[2, 8, ["Supervised", "Learning", "of", "Universal", "Sentence", "Representations"], "Task"], [9, 13, ["Natural", "Language", "Inference", "Data"], "Material"], [15, 17, ["NLP", "systems"], "Method"], [19, 21, ["word", "embeddings"], "Method"], [26, 28, ["unsupervised", "manner"], "Method"], [61, 65, ["unsupervised", "representations", "of", "sentences"], "Method"], [83, 86, ["universal", "sentence", "representations"], "Method"], [93, 98, ["Stanford", "Natural", "Language", "Inference", "datasets"], "Material"], [101, 103, ["unsupervised", "methods"], "Method"], [104, 106, ["SkipThought", "vectors"], "Method"], [111, 113, ["transfer", "tasks"], "Task"], [117, 119, ["computer", "vision"], "Task"], [120, 121, ["ImageNet"], "Material"], [142, 145, ["natural", "language", "inference"], "Task"], [146, 148, ["transfer", "learning"], "Task"], [150, 152, ["NLP", "tasks"], "Task"], [162, 166, ["Distributed", "representations", "of", "words"], "Method"], [168, 170, ["word", "embeddings"], "Method"], [181, 184, ["natural", "language", "processing"], "Task"], [185, 187, ["computer", "vision"], "Task"], [259, 263, ["universal", "representations", "of", "sentences"], "Task"], [267, 270, ["sentence", "encoder", "model"], "Method"], [296, 297, ["encoder"], "Method"], [304, 307, ["neural", "network", "architecture"], "Method"], [325, 328, ["learning", "word", "embeddings"], "Task"], [334, 336, ["sentence", "encoders"], "Method"], [338, 340, ["unsupervised", "manner"], "Method"], [341, 342, ["SkipThought"], "Method"], [343, 344, ["FastSent"], "Method"], [350, 352, ["supervised", "learning"], "Method"], [363, 365, ["computer", "vision"], "Task"], [373, 374, ["ImageNet"], "Material"], [380, 382, ["sentence", "embeddings"], "Method"], [385, 387, ["supervised", "tasks"], "Task"], [391, 393, ["sentence", "embeddings"], "Method"], [399, 402, ["natural", "language", "inference"], "Task"], [403, 404, ["NLI"], "Task"], [413, 415, ["transfer", "accuracy"], "Metric"], [422, 423, ["NLI"], "Task"], [425, 427, ["training", "task"], "Task"], [436, 441, ["high", "-", "level", "understanding", "task"], "Task"], [453, 455, ["computer", "vision"], "Task"], [457, 460, ["convolutional", "neural", "networks"], "Method"], [472, 474, ["neural", "networks"], "Method"], [483, 486, ["sentence", "encoding", "architecture"], "Method"], [487, 489, ["representational", "transferability"], "Task"], [492, 495, ["convolutional", ",", "recurrent"], "Method"], [498, 501, ["word", "composition", "schemes"], "Method"], [507, 508, ["encoder"], "Method"], [517, 519, ["max", "pooling"], "Method"], [523, 527, ["Stanford", "Natural", "Language", "Inference"], "Material"], [528, 529, ["SNLI"], "Material"], [540, 542, ["sentence", "embeddings"], "Metric"], [547, 549, ["unsupervised", "approaches"], "Method"], [550, 551, ["SkipThought"], "Method"], [552, 553, ["FastSent"], "Method"], [572, 574, ["transfer", "tasks"], "Task"], [579, 581, ["sentence", "representations"], "Method"], [592, 594, ["Transfer", "learning"], "Method"], [602, 605, ["computer", "vision", "applications"], "Task"], [609, 611, ["face", "recognition"], "Task"], [612, 615, ["visual", "question", "answering"], "Task"], [621, 622, ["ImageNet"], "Method"], [623, 625, ["word", "embeddings"], "Method"], [639, 642, ["sentence", "representation", "learning"], "Task"], [643, 644, ["unsupervised"], "Task"], [648, 650, ["NLP", "community"], "Task"], [656, 658, ["supervised", "task"], "Task"], [671, 673, ["neural", "networks"], "Method"], [710, 712, ["Learning", "models"], "Method"], [714, 716, ["unsupervised", "task"], "Task"], [728, 735, ["co", "-", "adaptation", "of", "encoders", "and", "classifiers"], "Method"], [748, 750, ["generalization", "power"], "Metric"], [756, 757, ["encoder"], "Method"], [761, 762, ["loss"], "Method"], [765, 767, ["orthogonal", "classifiers"], "Method"], [776, 778, ["sentence", "embeddings"], "Task"], [783, 785, ["word", "embeddings"], "Method"], [788, 791, ["neural", "network", "architectures"], "Method"], [792, 794, ["SkipThought", "vectors"], "Method"], [796, 798, ["objective", "function"], "Method"], [801, 805, ["skip", "-", "gram", "model"], "Method"], [829, 831, ["linear", "model"], "Method"], [841, 843, ["transfer", "tasks"], "Task"], [850, 854, ["layer", "-", "norm", "regularization"], "Method"], [878, 880, ["unsupervised", "methods"], "Method"], [883, 885, ["supervised", "training"], "Method"], [901, 904, ["French", "and", "English"], "Material"], [905, 907, ["German", "pairs"], "Material"], [922, 924, ["COCO", "dataset"], "Material"], [934, 939, ["unsupervised", "Skip", "-", "Thought", "approach"], "Method"], [945, 947, ["sentence", "encoders"], "Method"], [956, 958, ["SICK", "corpus"], "Material"], [961, 965, ["multi", "-", "task", "learning"], "Method"], [966, 967, ["pretraining"], "Method"], [986, 987, ["classifier"], "Method"], [1013, 1016, ["generic", "sentence", "encoders"], "Task"], [1030, 1032, ["unsupervised", "approaches"], "Method"], [1073, 1075, ["NLI", "task"], "Task"], [1080, 1084, ["universal", "sentence", "encoding", "models"], "Method"], [1099, 1101, ["sentence", "encoder"], "Task"], [1113, 1115, ["sentence", "encoders"], "Method"], [1124, 1126, ["recurrent", "models"], "Method"], [1128, 1129, ["LSTMs"], "Method"], [1130, 1131, ["GRUs"], "Method"], [1136, 1137, ["mean"], "Method"], [1138, 1141, ["max", "-", "pooling"], "Method"], [1143, 1145, ["hidden", "representations"], "Method"], [1147, 1151, ["self", "-", "attentive", "network"], "Method"], [1161, 1164, ["hierarchical", "convolutional", "network"], "Method"], [1170, 1174, ["tree", "-", "based", "method"], "Method"], [1184, 1188, ["Natural", "Language", "Inference", "task"], "Task"], [1194, 1200, ["human", "-", "generated", "English", "sentence", "pairs"], "Material"], [1217, 1220, ["natural", "language", "inference"], "Task"], [1227, 1230, ["Recognizing", "Textual", "Entailment"], "Task"], [1231, 1232, ["RTE"], "Task"], [1262, 1263, ["NLI"], "Task"], [1270, 1273, ["universal", "sentence", "embeddings"], "Task"], [1275, 1277, ["supervised", "way"], "Task"], [1286, 1288, ["sentence", "encoders"], "Method"], [1290, 1293, ["natural", "language", "inference"], "Task"], [1297, 1299, ["sentence", "representations"], "Method"], [1310, 1311, ["SNLI"], "Material"], [1319, 1324, ["sentence", "encoding", "-", "based", "models"], "Method"], [1337, 1339, ["joint", "methods"], "Method"], [1370, 1373, ["generic", "sentence", "encoder"], "Method"], [1396, 1399, ["shared", "sentence", "encoder"], "Method"], [1418, 1420, ["matching", "methods"], "Method"], [1440, 1444, ["element", "-", "wise", "product"], "Method"], [1474, 1478, ["3", "-", "class", "classifier"], "Method"], [1481, 1485, ["fully", "-", "connected", "layers"], "Method"], [1488, 1490, ["softmax", "layer"], "Method"], [1493, 1496, ["Sentence", "encoder", "architectures"], "Method"], [1500, 1502, ["neural", "networks"], "Method"], [1503, 1505, ["encoding", "sentences"], "Task"], [1506, 1510, ["fixed", "-", "size", "representations"], "Task"], [1533, 1535, ["recurrent", "encoders"], "Method"], [1537, 1542, ["Long", "Short", "-", "Term", "Memory"], "Method"], [1543, 1544, ["LSTM"], "Method"], [1546, 1549, ["Gated", "Recurrent", "Units"], "Method"], [1550, 1551, ["GRU"], "Method"], [1562, 1563, ["GRU"], "Method"], [1564, 1568, ["Bi", "-", "directional", "LSTMs"], "Method"], [1569, 1570, ["BiLSTM"], "Method"], [1573, 1574, ["mean"], "Method"], [1575, 1577, ["max", "pooling"], "Method"], [1578, 1582, ["self", "-", "attentive", "network"], "Method"], [1583, 1586, ["hierarchical", "convolutional", "networks"], "Method"], [1589, 1590, ["LSTM"], "Method"], [1591, 1592, ["GRU"], "Method"], [1598, 1599, ["encoders"], "Method"], [1600, 1603, ["recurrent", "neural", "networks"], "Method"], [1612, 1616, ["sequence", "to", "sequence", "encoders"], "Method"], [1629, 1631, ["hidden", "representations"], "Method"], [1656, 1660, ["model", "BiGRU", "-", "last"], "Method"], [1692, 1693, ["BiLSTM"], "Method"], [1694, 1695, ["mean"], "Method"], [1696, 1698, ["max", "pooling"], "Method"], [1771, 1773, ["max", "pooling"], "Method"], [1789, 1793, ["Self", "-", "attentive", "network"], "Method"], [1794, 1799, ["self", "-", "attentive", "sentence", "encoder"], "Method"], [1801, 1803, ["attention", "mechanism"], "Method"], [1809, 1810, ["BiLSTM"], "Method"], [1820, 1822, ["attention", "mechanism"], "Method"], [1834, 1835, ["BiLSTM"], "Method"], [1841, 1843, ["affine", "transformation"], "Method"], [1882, 1885, ["weighted", "linear", "combination"], "Method"], [1895, 1899, ["self", "-", "attentive", "network"], "Method"], [1943, 1945, ["sentence", "representation"], "Method"], [1956, 1958, ["Hierarchical", "ConvNet"], "Method"], [1966, 1968, ["classification", "tasks"], "Task"], [1970, 1972, ["convolutional", "architecture"], "Method"], [1973, 1974, ["AdaSent"], "Method"], [2001, 2003, ["convolutional", "layers"], "Method"], [2014, 2018, ["max", "-", "pooling", "operation"], "Method"], [2055, 2059, ["fixed", "-", "size", "representation"], "Method"], [2070, 2071, ["SNLI"], "Material"], [2074, 2075, ["SGD"], "Method"], [2077, 2079, ["learning", "rate"], "Metric"], [2095, 2097, ["learning", "rate"], "Metric"], [2101, 2103, ["dev", "accuracy"], "Metric"], [2119, 2121, ["learning", "rate"], "Metric"], [2129, 2130, ["classifier"], "Method"], [2134, 2138, ["multi", "-", "layer", "perceptron"], "Method"], [2140, 2143, ["hidden", "-", "layer"], "Method"], [2153, 2155, ["GloVe", "vectors"], "Method"], [2172, 2174, ["sentence", "representations"], "Task"], [2182, 2184, ["sentence", "embeddings"], "Method"], [2213, 2215, ["transfer", "tasks"], "Task"], [2219, 2224, ["sentence", "-", "embedding", "evaluation", "procedure"], "Task"], [2231, 2234, ["sentence", "evaluation", "tool"], "Method"], [2235, 2236, ["SentEval"], "Material"], [2238, 2239, ["evaluation"], "Task"], [2251, 2252, ["Adam"], "Method"], [2255, 2258, ["logistic", "regression", "classifier"], "Method"], [2266, 2272, ["Binary", "and", "multi", "-", "class", "classification"], "Task"], [2277, 2280, ["binary", "classification", "tasks"], "Task"], [2292, 2294, ["sentence", "classification"], "Task"], [2296, 2298, ["sentiment", "analysis"], "Task"], [2299, 2300, ["MR"], "Task"], [2301, 2302, ["SST"], "Task"], [2304, 2307, ["question", "-", "type"], "Task"], [2308, 2309, ["TREC"], "Task"], [2311, 2313, ["product", "reviews"], "Task"], [2314, 2315, ["CR"], "Task"], [2327, 2328, ["MPQA"], "Method"], [2337, 2339, ["logistic", "regression"], "Method"], [2343, 2345, ["linear", "classifier"], "Method"], [2346, 2348, ["fewer", "parameters"], "Metric"], [2350, 2351, ["MLP"], "Method"], [2360, 2362, ["transfer", "learning"], "Method"], [2375, 2377, ["logistic", "regression"], "Method"], [2378, 2381, ["grid", "-", "search"], "Method"], [2388, 2389, ["Entailment"], "Task"], [2390, 2392, ["semantic", "relatedness"], "Metric"], [2397, 2399, ["SICK", "dataset"], "Material"], [2403, 2406, ["SICK", "-", "E"], "Metric"], [2408, 2410, ["semantic", "relatedness"], "Metric"], [2411, 2414, ["SICK", "-", "R"], "Metric"], [2420, 2422, ["matching", "methods"], "Method"], [2424, 2425, ["SNLI"], "Material"], [2428, 2430, ["Logistic", "Regression"], "Method"], [2434, 2436, ["joint", "representation"], "Method"], [2460, 2462, ["Pearson", "correlation"], "Metric"], [2467, 2470, ["Semantic", "Textual", "Similarity"], "Method"], [2479, 2482, ["SICK", "-", "R"], "Metric"], [2491, 2494, ["unsupervised", "SemEval", "tasks"], "Task"], [2495, 2496, ["STS14"], "Material"], [2528, 2530, ["similarity", "score"], "Metric"], [2548, 2553, ["human", "-", "labeled", "similarity", "score"], "Metric"], [2554, 2558, ["Pearson", "and", "Spearman", "correlations"], "Metric"], [2561, 2563, ["Paraphrase", "detection"], "Task"], [2564, 2568, ["Microsoft", "Research", "Paraphrase", "Corpus"], "Material"], [2583, 2584, ["Web"], "Material"], [2611, 2614, ["SICK", "-", "E"], "Metric"], [2618, 2619, ["classifier"], "Method"], [2626, 2627, ["Caption"], "Task"], [2628, 2630, ["Image", "retrieval"], "Task"], [2631, 2636, ["caption", "-", "image", "retrieval", "task"], "Task"], [2637, 2643, ["joint", "image", "and", "language", "feature", "models"], "Method"], [2666, 2668, ["Image", "Retrieval"], "Task"], [2671, 2673, ["ranking", "captions"], "Task"], [2682, 2684, ["Caption", "Retrieval"], "Task"], [2689, 2693, ["pairwise", "ranking", "-", "loss"], "Metric"], [2712, 2714, ["ranking", "loss"], "Metric"], [2728, 2730, ["linear", "transformations"], "Method"], [2767, 2769, ["COCO", "dataset"], "Material"], [2807, 2808, ["Recall@K"], "Metric"], [2840, 2843, ["ResNet", "-", "101"], "Method"], [2860, 2870, ["micro", "”", "and", "”", "macro", "”", "averages", "of", "development", "set"], "Metric"], [2875, 2877, ["transfer", "tasks"], "Task"], [2880, 2881, ["accuracy"], "Metric"], [2886, 2890, ["macro", "”", "aggregated", "score"], "Metric"], [2894, 2899, ["classical", "average", "of", "dev", "accuracies"], "Metric"], [2903, 2906, ["micro", "”", "score"], "Metric"], [2912, 2914, ["dev", "accuracies"], "Metric"], [2949, 2951, ["transfer", "tasks"], "Task"], [2954, 2957, ["BiLSTM", "-", "4096"], "Method"], [2959, 2963, ["max", "-", "pooling", "operation"], "Method"], [2991, 2992, ["LSTM"], "Method"], [2993, 2994, ["GRU"], "Method"], [2995, 2998, ["BiGRU", "-", "last"], "Method"], [2999, 3002, ["BiLSTM", "-", "Mean"], "Method"], [3003, 3006, ["inner", "-", "attention"], "Method"], [3008, 3011, ["hierarchical", "-", "ConvNet"], "Method"], [3023, 3025, ["training", "task"], "Task"], [3034, 3036, ["transfer", "tasks"], "Task"], [3043, 3046, ["BiLSTM", "-", "Mean"], "Method"], [3085, 3089, ["inner", "-", "attention", "model"], "Method"], [3114, 3116, ["transfer", "tasks"], "Task"], [3122, 3125, ["BiLSTM", "-", "Mean"], "Method"], [3182, 3184, ["transfer", "quality"], "Metric"], [3189, 3191, ["optimization", "algorithm"], "Method"], [3195, 3196, ["Adam"], "Method"], [3198, 3199, ["SGD"], "Method"], [3204, 3207, ["BiLSTM", "-", "max"], "Method"], [3210, 3211, ["SNLI"], "Material"], [3225, 3227, ["transfer", "tasks"], "Task"], [3235, 3236, ["classifier"], "Method"], [3245, 3247, ["training", "task"], "Task"], [3250, 3252, ["Embedding", "size"], "Metric"], [3268, 3271, ["micro", "averaged", "performance"], "Metric"], [3275, 3277, ["embedding", "size"], "Metric"], [3291, 3293, ["logistic", "regression"], "Method"], [3321, 3324, ["BiLSTM", "-", "Max"], "Method"], [3325, 3326, ["HConvNet"], "Method"], [3327, 3330, ["inner", "-", "att"], "Method"], [3362, 3364, ["objective", "task"], "Task"], [3367, 3368, ["SNLI"], "Material"], [3386, 3388, ["transfer", "tasks"], "Task"], [3391, 3393, ["Task", "transfer"], "Task"], [3400, 3402, ["transfer", "tasks"], "Task"], [3440, 3446, ["bag", "-", "of", "-", "words", "models"], "Method"], [3448, 3451, ["word2vec", "-", "SkipGram"], "Method"], [3453, 3457, ["Unigram", "-", "TFIDF", "model"], "Method"], [3459, 3462, ["Paragraph", "Vector", "model"], "Method"], [3464, 3469, ["Sequential", "Denoising", "Auto", "-", "Encoder"], "Method"], [3470, 3471, ["SDAE"], "Method"], [3474, 3476, ["SIF", "model"], "Method"], [3481, 3484, ["Toronto", "book", "corpus"], "Material"], [3498, 3499, ["FastSent"], "Method"], [3500, 3501, ["SkipThought"], "Method"], [3506, 3509, ["Toronto", "book", "corpus"], "Material"], [3515, 3522, ["FastSent", "variant", "“", "FastSent", "+", "AE", "”"], "Method"], [3524, 3528, ["SkipThought", "-", "LN", "version"], "Method"], [3530, 3532, ["layer", "normalization"], "Method"], [3553, 3555, ["supervised", "methods"], "Method"], [3563, 3566, ["transfer", "learning", "approaches"], "Method"], [3571, 3572, ["SkipThought"], "Method"], [3575, 3577, ["sentence", "encoder"], "Method"], [3581, 3585, ["SkipThought", "-", "LN", "model"], "Method"], [3631, 3633, ["SkipThought", "vectors"], "Method"], [3651, 3655, ["SkipThought", "-", "LN", "network"], "Method"], [3661, 3664, ["BiLSTM", "-", "max"], "Method"], [3666, 3667, ["SNLI"], "Material"], [3681, 3682, ["SST"], "Task"], [3683, 3686, ["MRPC", "-", "accuracy"], "Metric"], [3687, 3690, ["SICK", "-", "R"], "Metric"], [3691, 3694, ["SICK", "-", "E"], "Metric"], [3695, 3696, ["STS14"], "Method"], [3707, 3709, ["SUBJ", "dataset"], "Material"], [3715, 3718, ["SkipThought", "-", "LN"], "Method"], [3719, 3720, ["MR"], "Task"], [3721, 3722, ["CR"], "Task"], [3723, 3724, ["MPQA"], "Task"], [3736, 3738, ["cosine", "metrics"], "Metric"], [3753, 3755, ["pearson", "score"], "Metric"], [3765, 3768, ["ST", "-", "LN"], "Method"], [3779, 3781, ["matching", "method"], "Method"], [3804, 3805, ["NLI"], "Task"], [3818, 3819, ["SNLI"], "Material"], [3829, 3831, ["supervised", "tasks"], "Task"], [3833, 3834, ["COCO"], "Material"], [3835, 3837, ["dictionary", "definitions"], "Task"], [3838, 3839, ["NMT"], "Method"], [3840, 3841, ["PPDB"], "Method"], [3842, 3843, ["SST"], "Task"], [3845, 3846, ["SST"], "Task"], [3855, 3856, ["SNLI"], "Material"], [3862, 3863, ["SST"], "Task"], [3866, 3867, ["NLI"], "Task"], [3873, 3875, ["quality", "features"], "Metric"], [3877, 3881, ["classification", "and", "similarity", "tasks"], "Task"], [3911, 3912, ["NLI"], "Task"], [3920, 3922, ["Domain", "adaptation"], "Task"], [3923, 3925, ["SICK", "tasks"], "Task"], [3926, 3929, ["transfer", "learning", "approach"], "Method"], [3943, 3945, ["SICK", "task"], "Task"], [3951, 3955, ["out", "-", "domain", "version"], "Method"], [3956, 3957, ["SNLI"], "Material"], [3960, 3963, ["entailment", "and", "relatedness"], "Task"], [3967, 3969, ["pearson", "score"], "Metric"], [3972, 3975, ["SICK", "-", "R"], "Metric"], [3984, 3986, ["test", "accuracy"], "Metric"], [3987, 3990, ["SICK", "-", "E"], "Metric"], [3993, 3997, ["hand", "-", "engineered", "models"], "Method"], [4006, 4009, ["transfer", "learning", "approaches"], "Method"], [4010, 4013, ["SICK", "-", "E"], "Metric"], [4023, 4024, ["SNLI"], "Material"], [4033, 4034, ["accuracy"], "Metric"], [4048, 4052, ["in", "-", "domain", "task"], "Task"], [4058, 4059, ["classifier"], "Method"], [4067, 4072, ["small", "out", "-", "domain", "task"], "Task"], [4075, 4079, ["Image", "-", "caption", "retrieval"], "Task"], [4091, 4097, ["COCO", "image", "-", "caption", "retrieval", "task"], "Task"], [4102, 4103, ["recalls"], "Metric"], [4141, 4144, ["caption", "retrieval", "R@1"], "Task"], [4151, 4154, ["image", "retrieval", "R@1"], "Task"], [4168, 4170, ["caption", "retrieval"], "Task"], [4176, 4178, ["image", "retrieval"], "Task"], [4196, 4198, ["sentence", "encoding"], "Method"], [4200, 4205, ["image", "-", "caption", "retrieval", "task"], "Task"], [4213, 4215, ["trained", "representations"], "Method"], [4222, 4224, ["sentence", "embeddings"], "Method"], [4235, 4237, ["objective", "task"], "Task"], [4255, 4256, ["SNLI"], "Material"], [4263, 4264, ["MultiNLI"], "Material"], [4266, 4267, ["SNLI"], "Material"], [4301, 4302, ["SNLI"], "Material"], [4303, 4304, ["MultiNLI"], "Material"], [4305, 4306, ["AllNLI"], "Material"], [4323, 4324, ["SLNI"], "Material"], [4329, 4330, ["AdaSent"], "Method"], [4332, 4333, ["CR"], "Task"], [4342, 4344, ["training", "task"], "Task"], [4348, 4350, ["general", "representations"], "Method"], [4352, 4355, ["semantic", "textual", "similarity"], "Task"], [4362, 4368, ["PPDB", "based", "paragram", "-", "phrase", "embeddings"], "Method"], [4370, 4372, ["pearson", "score"], "Metric"], [4378, 4383, ["caption", "-", "related", "transfer", "tasks"], "Task"], [4386, 4391, ["COCO", "image", "caption", "retrieval", "task"], "Task"], [4394, 4396, ["sentence", "encoder"], "Method"], [4400, 4401, ["MultiNLI"], "Material"], [4412, 4413, ["SNLI"], "Material"], [4439, 4441, ["sentence", "embeddings"], "Task"], [4449, 4451, ["transfer", "tasks"], "Task"], [4458, 4459, ["NLI"], "Task"], [4471, 4473, ["supervised", "tasks"], "Task"], [4483, 4485, ["BiLSTM", "network"], "Method"], [4486, 4488, ["max", "pooling"], "Method"], [4492, 4496, ["universal", "sentence", "encoding", "methods"], "Method"], [4501, 4503, ["SkipThought", "vectors"], "Method"], [4521, 4525, ["learning", "generic", "sentence", "embeddings"], "Task"], [4538, 4541, ["sentence", "embedding", "quality"], "Metric"], [4554, 4558, ["Max", "-", "pooling", "visualization"], "Method"], [4559, 4562, ["BiLSTM", "-", "max"], "Method"], [4579, 4580, ["classifier"], "Method"], [4608, 4612, ["max", "-", "pooling", "operation"], "Method"], [4621, 4622, ["BiLSTM"], "Method"], [4658, 4662, ["max", "-", "pooling", "operation"], "Method"], [4673, 4675, ["sentence", "representation"], "Method"], [4685, 4688, ["max", "-", "pooling"], "Method"], [4733, 4735, ["attention", "mechanism"], "Method"], [514, 515, ["LSTM"], "Method"], [949, 950, ["SNLI"], "Material"], [1009, 1010, ["SNLI"], "Material"], [1086, 1087, ["SNLI"], "Material"], [1189, 1190, ["SNLI"], "Material"], [1605, 1606, ["LSTM"], "Method"], [1607, 1608, ["GRU"], "Method"], [1636, 1637, ["GRU"], "Method"], [1669, 1670, ["GRU"], "Method"], [1679, 1680, ["GRU"], "Method"], [1707, 1708, ["LSTM"], "Method"], [1724, 1725, ["LSTM"], "Method"], [1728, 1729, ["LSTM"], "Method"], [1783, 1784, ["mean"], "Method"], [2438, 2440, ["semantic", "relatedness"], "Metric"], [2471, 2473, ["semantic", "relatedness"], "Metric"], [2944, 2945, ["NLI"], "Task"], [2967, 2968, ["SNLI"], "Material"], [3106, 3107, ["SNLI"], "Material"], [3618, 3619, ["SNLI"], "Material"], [3677, 3678, ["CR"], "Task"], [3782, 3783, ["SNLI"], "Material"], [3887, 3890, ["natural", "language", "inference"], "Task"], [4019, 4020, ["LSTM"], "Method"], [4101, 4102, ["mean"], "Method"], [4241, 4242, ["NLI"], "Task"], [4243, 4244, ["MultiNLI"], "Material"], [4628, 4629, ["BiLSTM"], "Method"]]} {"doc_id": "3fc5ed18c2294596af072df929c8ee12c71f96a2", "text": ["document", ":", "Classical", "Structured", "Prediction", "Losses", "for", "Sequence", "to", "Sequence", "Learning", "There", "has", "been", "much", "recent", "work", "on", "training", "neural", "attention", "models", "at", "the", "sequence", "-", "level", "using", "either", "reinforcement", "learning", "-", "style", "methods", "or", "by", "optimizing", "the", "beam", ".", "In", "this", "paper", ",", "we", "survey", "a", "range", "of", "classical", "objective", "functions", "that", "have", "been", "widely", "used", "to", "train", "linear", "models", "for", "structured", "prediction", "and", "apply", "them", "to", "neural", "sequence", "to", "sequence", "models", ".", "Our", "experiments", "show", "that", "these", "losses", "can", "perform", "surprisingly", "well", "by", "slightly", "outperforming", "beam", "search", "optimization", "in", "a", "like", "for", "like", "setup", ".", "We", "also", "report", "new", "state", "of", "the", "art", "results", "on", "both", "IWSLT’14", "German", "-", "English", "translation", "as", "well", "as", "Gigaword", "abstractive", "summarization", ".", "On", "the", "large", "WMT’14", "English", "-", "French", "task", ",", "sequence", "-", "level", "training", "achieves", "41.5", "BLEU", "which", "is", "on", "par", "with", "the", "state", "of", "the", "art", ".", "section", ":", "Introduction", "Sequence", "to", "sequence", "models", "are", "usually", "trained", "with", "a", "simple", "token", "-", "level", "likelihood", "loss", "sutskever2014sequence", ",", "bahdanau2014neural", ".", "However", ",", "at", "test", "time", ",", "these", "models", "do", "not", "produce", "a", "single", "token", "but", "a", "whole", "sequence", ".", "In", "order", "to", "resolve", "this", "inconsistency", "and", "to", "potentially", "improve", "generation", ",", "recent", "work", "has", "focused", "on", "training", "these", "models", "at", "the", "sequence", "-", "level", ",", "for", "instance", "using", "REINFORCE", "ranzato2015sequence", ",", "actor", "-", "critic", "bahdanau2016ac", ",", "or", "with", "beam", "search", "optimization", "wiseman2016acl", ".", "Before", "the", "recent", "work", "on", "sequence", "level", "training", "for", "neural", "networks", ",", "there", "has", "been", "a", "large", "body", "of", "research", "on", "training", "linear", "models", "at", "the", "sequence", "level", ".", "For", "example", ",", "direct", "loss", "optimization", "has", "been", "popularized", "in", "machine", "translation", "with", "the", "Minimum", "Error", "Rate", "Training", "algorithm", "(", "MERT", ";", "Och", "2003", ")", "and", "expected", "risk", "minimization", "has", "an", "extensive", "history", "in", "NLP", ".", "This", "paper", "revisits", "several", "objective", "functions", "that", "have", "been", "commonly", "used", "for", "structured", "prediction", "tasks", "in", "NLP", "gimpel", "+", "smith2010acl", "and", "apply", "them", "to", "a", "neural", "sequence", "to", "sequence", "model", "gehring2017icml", "(", "§", "[", "reference", "]", ")", ".", "Specifically", ",", "we", "consider", "likelihood", "training", "at", "the", "sequence", "-", "level", ",", "a", "margin", "loss", "as", "well", "as", "expected", "risk", "training", ".", "We", "also", "investigate", "several", "combinations", "of", "global", "losses", "with", "token", "-", "level", "likelihood", ".", "This", "is", ",", "to", "our", "knowledge", ",", "the", "most", "comprehensive", "comparison", "of", "structured", "losses", "in", "the", "context", "of", "neural", "sequence", "to", "sequence", "models", "(", "§", "[", "reference", "]", ")", ".", "We", "experiment", "on", "the", "IWSLT’14", "German", "-", "English", "translation", "task", "cettolo2014report", "as", "well", "as", "the", "Gigaword", "abstractive", "summarization", "task", "rush2015abs", ".", "We", "achieve", "the", "best", "reported", "accuracy", "to", "date", "on", "both", "tasks", ".", "We", "find", "that", "the", "sequence", "level", "losses", "we", "survey", "perform", "similarly", "to", "one", "another", "and", "outperform", "beam", "search", "optimization", "wiseman2016acl", "on", "a", "comparable", "setup", ".", "On", "WMT’14", "English", "-", "French", ",", "we", "also", "illustrate", "the", "effectiveness", "of", "risk", "minimization", "on", "a", "larger", "translation", "task", ".", "Classical", "losses", "for", "structured", "prediction", "are", "still", "very", "competitive", "and", "effective", "for", "neural", "models", "(", "§", "[", "reference", "]", ",", "§", "[", "reference", "]", ")", ".", "section", ":", "Sequence", "to", "Sequence", "Learning", "The", "general", "architecture", "of", "our", "sequence", "to", "sequence", "models", "follows", "the", "encoder", "-", "decoder", "approach", "with", "soft", "attention", "first", "introduced", "in", ".", "As", "a", "main", "difference", ",", "in", "most", "of", "our", "experiments", "we", "parameterize", "the", "encoder", "and", "the", "decoder", "as", "convolutional", "neural", "networks", "instead", "of", "recurrent", "networks", "gehring2016convolutional", ",", "gehring2017icml", ".", "Our", "use", "of", "convolution", "is", "motivated", "by", "computational", "and", "accuracy", "considerations", ".", "However", ",", "the", "objective", "functions", "we", "present", "are", "model", "agnostic", "and", "equally", "applicable", "to", "recurrent", "and", "convolutional", "models", ".", "We", "demonstrate", "the", "applicability", "of", "our", "objective", "functions", "to", "recurrent", "models", "(", "LSTM", ")", "in", "our", "comparison", "to", "wiseman2016acl", "in", "§", "[", "reference", "]", ".", "Notation", ".", "We", "denote", "the", "source", "sentence", "as", ",", "an", "output", "sentence", "of", "our", "model", "as", ",", "and", "the", "reference", "or", "target", "sentence", "as", ".", "For", "some", "objectives", ",", "we", "choose", "a", "pseudo", "reference", "instead", ",", "such", "as", "a", "model", "output", "with", "the", "highest", "BLEU", "or", "ROUGE", "score", "among", "a", "set", "of", "candidate", "outputs", ",", ",", "generated", "by", "our", "model", ".", "Concretely", ",", "the", "encoder", "processes", "a", "source", "sentence", "containing", "words", "and", "outputs", "a", "sequence", "of", "states", ".", "The", "decoder", "takes", "and", "generates", "the", "output", "sequence", "left", "to", "right", ",", "one", "element", "at", "a", "time", ".", "For", "each", "output", ",", "the", "decoder", "computes", "hidden", "state", "based", "on", "the", "previous", "state", ",", "an", "embedding", "of", "the", "previous", "target", "language", "word", ",", "as", "well", "as", "a", "conditional", "input", "derived", "from", "the", "encoder", "output", ".", "The", "attention", "context", "is", "computed", "as", "a", "weighted", "sum", "of", "at", "each", "time", "step", ".", "The", "weights", "of", "this", "sum", "are", "referred", "to", "as", "attention", "scores", "and", "allow", "the", "network", "to", "focus", "on", "the", "most", "relevant", "parts", "of", "the", "input", "at", "each", "generation", "step", ".", "Attention", "scores", "are", "computed", "by", "comparing", "each", "encoder", "state", "to", "a", "combination", "of", "the", "previous", "decoder", "state", "and", "the", "last", "prediction", ";", "the", "result", "is", "normalized", "to", "be", "a", "distribution", "over", "input", "elements", ".", "At", "each", "generation", "step", ",", "the", "model", "scores", "for", "the", "possible", "next", "target", "words", "by", "transforming", "the", "decoder", "output", "via", "a", "linear", "layer", "with", "weights", "and", "bias", ":", ".", "This", "is", "turned", "into", "a", "distribution", "via", "a", "softmax", ":", ".", "Our", "encoder", "and", "decoder", "use", "gated", "convolutional", "neural", "networks", "which", "enable", "fast", "and", "accurate", "generation", ".", "Fast", "generation", "is", "essential", "to", "efficiently", "train", "on", "the", "model", "output", "as", "is", "done", "in", "this", "work", "as", "sequence", "-", "level", "losses", "require", "generating", "at", "training", "time", ".", "Both", "encoder", "and", "decoder", "networks", "share", "a", "simple", "block", "structure", "that", "computes", "intermediate", "states", "based", "on", "a", "fixed", "number", "of", "input", "tokens", "and", "we", "stack", "several", "blocks", "on", "top", "of", "each", "other", ".", "Each", "block", "contains", "a", "1", "-", "D", "convolution", "that", "takes", "as", "input", "feature", "vectors", "and", "outputs", "another", "vector", ";", "subsequent", "layers", "operate", "over", "the", "output", "elements", "of", "the", "previous", "layer", ".", "The", "output", "of", "the", "convolution", "is", "then", "fed", "into", "a", "gated", "linear", "unit", "dauphin2017icml", ".", "In", "the", "decoder", "network", ",", "we", "rely", "on", "causal", "convolution", "which", "rely", "only", "on", "states", "from", "the", "previous", "time", "steps", ".", "The", "parameters", "of", "our", "model", "are", "all", "the", "weight", "matrices", "in", "the", "encoder", "and", "decoder", "networks", ".", "Further", "details", "can", "be", "found", "in", "gehring2017icml", ".", "section", ":", "Objective", "Functions", "We", "compare", "several", "objective", "functions", "for", "training", "the", "model", "architecture", "described", "in", "§", "[", "reference", "]", ".", "The", "corresponding", "loss", "functions", "are", "either", "computed", "over", "individual", "tokens", "(", "§", "[", "reference", "]", ")", ",", "over", "entire", "sequences", "(", "§", "[", "reference", "]", ")", "or", "over", "a", "combination", "of", "tokens", "and", "sequences", "(", "§", "[", "reference", "]", ")", ".", "An", "overview", "of", "these", "loss", "functions", "is", "given", "in", "Figure", "[", "reference", "]", ".", "subsection", ":", "Token", "-", "Level", "Objectives", "Most", "prior", "work", "on", "sequence", "to", "sequence", "learning", "has", "focused", "on", "optimizing", "token", "-", "level", "loss", "functions", ",", "i.e.", ",", "functions", "for", "which", "the", "loss", "is", "computed", "additively", "over", "individual", "tokens", ".", "subsubsection", ":", "Token", "Negative", "Log", "Likelihood", "(", "TokNLL", ")", "Token", "-", "level", "likelihood", "(", "TokNLL", ",", "Equation", "[", "reference", "]", ")", "minimizes", "the", "negative", "log", "likelihood", "of", "individual", "reference", "tokens", ".", "It", "is", "the", "most", "common", "loss", "function", "optimized", "in", "related", "work", "and", "serves", "as", "a", "baseline", "for", "our", "comparison", ".", "subsubsection", ":", "Token", "NLL", "with", "Label", "Smoothing", "(", "TokLS", ")", "Likelihood", "training", "forces", "the", "model", "to", "make", "extreme", "zero", "or", "one", "predictions", "to", "distinguish", "between", "the", "ground", "truth", "and", "alternatives", ".", "This", "may", "result", "in", "a", "model", "that", "is", "too", "confident", "in", "its", "training", "predictions", ",", "which", "may", "hurt", "its", "generalization", "performance", ".", "Label", "smoothing", "addresses", "this", "by", "acting", "as", "a", "regularizer", "that", "makes", "the", "model", "less", "confident", "in", "its", "predictions", ".", "Specifically", ",", "we", "smooth", "the", "target", "distribution", "with", "a", "prior", "distribution", "that", "is", "independent", "of", "the", "current", "input", "szegedy2015inception", ",", "pereyra2017regularize", ",", "vaswani2017transformer", ".", "We", "use", "a", "uniform", "prior", "distribution", "over", "all", "words", "in", "the", "vocabulary", ",", ".", "One", "may", "also", "use", "a", "unigram", "distribution", "which", "has", "been", "shown", "to", "work", "better", "on", "some", "tasks", "pereyra2017regularize", ".", "Label", "smoothing", "is", "equivalent", "to", "adding", "the", "KL", "divergence", "between", "and", "the", "model", "prediction", "to", "the", "negative", "log", "likelihood", "(", "TokLS", ",", "Equation", "[", "reference", "]", ")", ".", "In", "practice", ",", "we", "implement", "label", "smoothing", "by", "modifying", "the", "ground", "truth", "distribution", "for", "word", "to", "be", "and", "for", "instead", "of", "and", "where", "is", "a", "smoothing", "parameter", ".", "subsection", ":", "Sequence", "-", "Level", "Objectives", "We", "also", "consider", "a", "class", "of", "objective", "functions", "that", "are", "computed", "over", "entire", "sequences", ",", "i.e.", ",", "sequence", "-", "level", "objectives", ".", "Training", "with", "these", "objectives", "requires", "generating", "and", "scoring", "multiple", "candidate", "output", "sequences", "for", "each", "input", "sequence", "during", "training", ",", "which", "is", "computationally", "expensive", "but", "allows", "us", "to", "directly", "optimize", "task", "-", "specific", "metrics", "such", "as", "BLEU", "or", "ROUGE", ".", "Unfortunately", ",", "these", "objectives", "are", "also", "typically", "defined", "over", "the", "entire", "space", "of", "possible", "output", "sequences", ",", "which", "is", "intractable", "to", "enumerate", "or", "score", "with", "our", "models", ".", "Instead", ",", "we", "compute", "our", "sequence", "losses", "over", "a", "subset", "of", "the", "output", "space", ",", ",", "generated", "by", "the", "model", ".", "We", "discuss", "approaches", "for", "generating", "this", "subset", "in", "§", "[", "reference", "]", ".", "subsubsection", ":", "Sequence", "Negative", "Log", "Likelihood", "(", "SeqNLL", ")", "Similar", "to", "TokNLL", ",", "we", "can", "minimize", "the", "negative", "log", "likelihood", "of", "an", "entire", "sequence", "rather", "than", "individual", "tokens", "(", "SeqNLL", ",", "Equation", "[", "reference", "]", ")", ".", "The", "log", "-", "likelihood", "of", "sequence", "is", "the", "sum", "of", "individual", "token", "log", "probabilities", ",", "normalized", "by", "the", "number", "of", "tokens", "to", "avoid", "bias", "towards", "shorter", "sequences", ":", "As", "target", "we", "choose", "a", "pseudo", "reference", "amongst", "the", "candidates", "which", "maximizes", "either", "BLEU", "or", "ROUGE", "with", "respect", "to", ",", "the", "gold", "reference", ":", "As", "is", "common", "practice", "when", "computing", "BLEU", "at", "the", "sentence", "-", "level", ",", "we", "smooth", "all", "initial", "counts", "to", "one", "(", "except", "for", "unigram", "counts", ")", "so", "that", "the", "geometric", "mean", "is", "not", "dominated", "by", "zero", "-", "valued", "-", "gram", "match", "counts", "lin2004orange", ".", "subsubsection", ":", "Expected", "Risk", "Minimization", "(", "Risk", ")", "This", "objective", "minimizes", "the", "expected", "value", "of", "a", "given", "cost", "function", "over", "the", "space", "of", "candidate", "sequences", "(", "Risk", ",", "Equation", "[", "reference", "]", ")", ".", "In", "this", "work", "we", "use", "task", "-", "specific", "cost", "functions", "designed", "to", "maximize", "BLEU", "or", "ROUGE", "lin2004rouge", ",", "e.g.", ",", ",", "for", "a", "given", "a", "candidate", "sequence", "and", "target", ".", "Different", "to", "SeqNLL", "(", "§", "[", "reference", "]", ")", ",", "this", "loss", "may", "increase", "the", "score", "of", "several", "candidates", "that", "have", "low", "cost", ",", "instead", "of", "focusing", "on", "a", "single", "sequence", "which", "may", "only", "be", "marginally", "better", "than", "any", "alternatives", ".", "Optimizing", "this", "loss", "is", "a", "particularly", "good", "strategy", "if", "the", "reference", "is", "not", "always", "reachable", ",", "although", "compared", "to", "classical", "phrase", "-", "based", "models", ",", "this", "is", "less", "of", "an", "issue", "with", "neural", "sequence", "to", "sequence", "models", "that", "predict", "individual", "words", "or", "even", "sub", "-", "word", "units", ".", "The", "Risk", "objective", "is", "similar", "to", "the", "REINFORCE", "objective", "used", "in", "Ranzato", "et", "al", ".", "ranzato2015sequence", ",", "since", "both", "objectives", "optimize", "an", "expected", "cost", "or", "reward", "williams1992reinforce", ".", "However", ",", "there", "are", "a", "few", "important", "differences", ":", "(", "1", ")", "whereas", "REINFORCE", "typically", "approximates", "the", "expectation", "with", "a", "single", "sampled", "sequence", ",", "the", "Risk", "objective", "considers", "multiple", "sequences", ";", "(", "2", ")", "whereas", "REINFORCE", "relies", "on", "a", "baseline", "rewardRanzato", "et", "al", ".", "estimate", "the", "baseline", "reward", "for", "REINFORCE", "with", "a", "separate", "linear", "regressor", "over", "the", "model", "’s", "current", "hidden", "state", ".", "to", "determine", "the", "sign", "of", "the", "gradients", "for", "the", "current", "sequence", ",", "for", "the", "Risk", "objective", "we", "instead", "estimate", "the", "expected", "cost", "over", "a", "set", "of", "candidate", "output", "sequences", "(", "see", "§", "[", "reference", "]", ")", ";", "and", "(", "3", ")", "while", "the", "baseline", "reward", "is", "different", "for", "every", "word", "in", "REINFORCE", ",", "the", "expected", "cost", "is", "the", "same", "for", "every", "word", "in", "risk", "minimization", "since", "it", "is", "computed", "on", "the", "sequence", "level", "based", "on", "the", "actual", "cost", ".", "subsubsection", ":", "Max", "-", "Margin", "MaxMargin", "(", "Equation", "[", "reference", "]", ")", "is", "a", "classical", "margin", "loss", "for", "structured", "prediction", "mmmn", ",", "structure_pred", "which", "enforces", "a", "margin", "between", "the", "model", "scores", "of", "the", "highest", "scoring", "candidate", "sequence", "and", "a", "reference", "sequence", ".", "We", "replace", "the", "human", "reference", "with", "a", "pseudo", "-", "reference", "since", "this", "setting", "performed", "slightly", "better", "in", "early", "experiments", ";", "is", "the", "candidate", "sequence", "with", "the", "highest", "BLEU", "score", ".", "The", "size", "of", "the", "margin", "varies", "between", "samples", "and", "is", "given", "by", "the", "difference", "between", "the", "cost", "of", "and", "the", "cost", "of", ".", "In", "practice", ",", "we", "scale", "the", "margin", "by", "a", "hyper", "-", "parameter", "determined", "on", "the", "validation", "set", ":", ".", "For", "this", "loss", "we", "use", "the", "unnormalized", "scores", "computed", "by", "the", "model", "before", "the", "final", "softmax", ":", "subsubsection", ":", "Multi", "-", "Margin", "MaxMargin", "only", "updates", "two", "elements", "in", "the", "candidate", "set", ".", "We", "therefore", "consider", "MultiMargin", "(", "Equation", "[", "reference", "]", ")", "which", "enforces", "a", "margin", "between", "every", "candidate", "sequence", "and", "a", "reference", "sequence", "herbrich199icann", ",", "hence", "the", "name", "Multi", "-", "Margin", ".", "Similar", "to", "MaxMargin", ",", "we", "replace", "the", "reference", "with", "the", "pseudo", "-", "reference", ".", "subsubsection", ":", "Softmax", "-", "Margin", "Finally", ",", "SoftmaxMargin", "(", "Equation", "[", "reference", "]", ")", "is", "another", "classic", "loss", "that", "has", "been", "proposed", "by", "gimpel", "+", "smith2010acl", "as", "another", "way", "to", "optimize", "task", "-", "specific", "costs", ".", "The", "loss", "augments", "the", "scores", "inside", "the", "of", "SeqNLL", "(", "Equation", "[", "reference", "]", ")", "by", "a", "cost", ".", "The", "intuition", "is", "that", "we", "want", "to", "penalize", "high", "cost", "outputs", "proportional", "to", "their", "cost", ".", "subsection", ":", "Combined", "Objectives", "We", "also", "experiment", "with", "two", "variants", "of", "combining", "sequence", "-", "level", "objectives", "(", "§", "[", "reference", "]", ")", "with", "token", "-", "level", "objectives", "(", "§", "[", "reference", "]", ")", ".", "First", ",", "we", "consider", "a", "weighted", "combination", "(", "Weighted", ")", "of", "both", "a", "sequence", "-", "level", "and", "token", "-", "level", "objective", "wu2016google", ",", "e.g.", ",", "for", "TokLS", "and", "Risk", "we", "have", ":", "where", "is", "a", "scaling", "constant", "that", "is", "tuned", "on", "a", "held", "-", "out", "validation", "set", ".", "Second", ",", "we", "consider", "a", "constrained", "combination", "(", "Constrained", ")", ",", "where", "for", "any", "given", "input", "we", "use", "either", "the", "token", "-", "level", "or", "sequence", "-", "level", "loss", ",", "but", "not", "both", ".", "The", "motivation", "is", "to", "maintain", "good", "token", "-", "level", "accuracy", "while", "optimizing", "on", "the", "sequence", "-", "level", ".", "In", "particular", ",", "a", "sample", "is", "processed", "with", "the", "sequence", "loss", "if", "the", "token", "loss", "under", "the", "current", "model", "is", "at", "least", "as", "good", "as", "the", "token", "loss", "of", "a", "baseline", "model", ".", "Otherwise", ",", "we", "update", "according", "to", "the", "token", "loss", ":", "In", "this", "work", "we", "use", "a", "fixed", "baseline", "model", "that", "was", "trained", "with", "a", "token", "-", "level", "loss", "to", "convergence", ".", "section", ":", "Candidate", "Generation", "Strategies", "The", "sequence", "-", "level", "objectives", "we", "consider", "(", "§", "[", "reference", "]", ")", "are", "defined", "over", "the", "entire", "space", "of", "possible", "output", "sequences", ",", "which", "is", "intractable", "to", "enumerate", "or", "score", "with", "our", "models", ".", "We", "therefore", "use", "a", "subset", "of", "candidate", "sequences", ",", "which", "we", "generate", "with", "our", "models", ".", "We", "consider", "two", "search", "strategies", "for", "generating", "the", "set", "of", "candidate", "sequences", ".", "The", "first", "is", "beam", "search", ",", "a", "greedy", "breadth", "-", "first", "search", "that", "maintains", "a", "“", "beam", "”", "of", "the", "top", "-", "scoring", "candidates", "at", "each", "generation", "step", ".", "Beam", "search", "is", "the", "de", "facto", "decoding", "strategy", "for", "achieving", "state", "-", "of", "-", "the", "-", "art", "results", "in", "machine", "translation", ".", "The", "second", "strategy", "is", "sampling", "chatterjee", ",", "which", "produces", "independent", "output", "sequences", "by", "sampling", "from", "the", "model", "’s", "conditional", "distribution", ".", "Whereas", "beam", "search", "focuses", "on", "high", "probability", "candidates", ",", "sampling", "introduces", "more", "diverse", "candidates", "(", "see", "comparison", "in", "§", "[", "reference", "]", ")", ".", "We", "also", "consider", "both", "online", "and", "offline", "candidate", "generation", "settings", "in", "§", "[", "reference", "]", ".", "In", "the", "online", "setting", ",", "we", "regenerate", "the", "candidate", "set", "every", "time", "we", "encounter", "an", "input", "sentence", "during", "training", ".", "In", "the", "offline", "setting", ",", "candidates", "are", "generated", "before", "training", "and", "are", "never", "regenerated", ".", "Offline", "generation", "is", "also", "embarrassingly", "parallel", "because", "all", "samples", "use", "the", "same", "model", ".", "The", "disadvantage", "is", "that", "the", "candidates", "become", "stale", ".", "Our", "model", "may", "perfectly", "be", "able", "to", "discriminate", "between", "them", "after", "only", "a", "single", "update", ",", "hindering", "the", "ability", "of", "the", "loss", "to", "correct", "eventual", "search", "errors", ".", "Finally", ",", "while", "some", "past", "work", "has", "added", "the", "reference", "target", "to", "the", "candidate", "set", ",", "i.e.", ",", ",", "we", "find", "this", "can", "destabilize", "training", "since", "the", "model", "learns", "to", "assign", "low", "probabilities", "nearly", "everywhere", ",", "ruining", "the", "candidates", "generated", "by", "the", "model", ",", "while", "still", "assigning", "a", "slightly", "higher", "score", "to", "the", "reference", "(", "cf", ".", "shen2016mrt", ")", ".", "Accordingly", "we", "do", "not", "add", "the", "reference", "translation", "to", "our", "candidate", "sets", ".", "section", ":", "Experimental", "Setup", "subsection", ":", "Translation", "We", "experiment", "on", "the", "IWSLT’14", "German", "to", "English", "cettolo2014report", "task", "using", "a", "similar", "setup", "as", "Ranzato", "et", "al", ".", "ranzato2015sequence", ",", "which", "allows", "us", "to", "compare", "to", "other", "recent", "studies", "that", "also", "adopted", "this", "setup", ",", "e.g.", ",", "wiseman2016acl", ".", "The", "training", "data", "consists", "of", "160", "K", "sentence", "pairs", "and", "the", "validation", "set", "comprises", "7", "K", "sentences", "randomly", "sampled", "and", "held", "-", "out", "from", "the", "train", "data", ".", "We", "test", "on", "the", "concatenation", "of", "all", "available", "test", "and", "dev", "sets", "of", "IWSLT", "2014", ",", "that", "is", "TED.tst2010", ",", "TED.tst2011", ",", "TED.tst2012", "and", "TED.dev2010", ",", "TEDX.dev2012", "which", "is", "of", "similar", "size", "to", "the", "validation", "set", ".", "All", "data", "is", "lowercased", "and", "tokenized", "with", "a", "byte", "-", "pair", "encoding", "(", "BPE", ")", "of", "14", ",", "000", "types", "sennrich2016bpe", "and", "we", "evaluate", "with", "case", "-", "insensitive", "BLEU", ".", "We", "also", "experiment", "on", "the", "much", "larger", "WMT’14", "English", "-", "French", "task", ".", "We", "remove", "sentences", "longer", "than", "175", "words", "as", "well", "as", "pairs", "with", "a", "source", "/", "target", "length", "ratio", "exceeding", "1.5", "resulting", "in", "35.5", "M", "sentence", "-", "pairs", "for", "training", ".", "The", "source", "and", "target", "vocabulary", "is", "based", "on", "40", "K", "BPE", "types", ".", "Results", "are", "reported", "on", "both", "newstest2014", "and", "a", "validation", "set", "held", "-", "out", "from", "the", "training", "data", "comprising", "26", ",", "658", "sentence", "pairs", ".", "We", "modify", "the", "fairseq", "-", "py", "toolkit", "to", "implement", "the", "objectives", "described", "in", "§", "[", "reference", "]", ".", "Our", "translation", "models", "have", "four", "convolutional", "encoder", "layers", "and", "three", "convolutional", "decoder", "layers", "with", "a", "kernel", "width", "of", "3", "and", "256", "dimensional", "hidden", "states", "and", "word", "embeddings", ".", "We", "optimize", "these", "models", "using", "Nesterov", "’s", "accelerated", "gradient", "method", "sutskever2013icml", "with", "a", "learning", "rate", "of", "0.25", "and", "momentum", "of", "0.99", ".", "Gradient", "vectors", "are", "renormalized", "to", "norm", "0.1", "pascanu2013difficulty", ".", "We", "train", "our", "baseline", "token", "-", "level", "models", "for", "200", "epochs", "and", "then", "anneal", "the", "learning", "by", "shrinking", "it", "by", "a", "factor", "of", "10", "after", "each", "subsequent", "epoch", "until", "the", "learning", "rate", "falls", "below", ".", "All", "sequence", "-", "level", "models", "are", "initialized", "with", "parameters", "of", "a", "token", "-", "level", "model", "before", "annealing", ".", "We", "then", "train", "sequence", "-", "level", "models", "for", "another", "10", "to", "20", "epochs", "depending", "on", "the", "objective", ".", "Our", "batches", "contain", "8", "K", "tokens", "and", "we", "normalize", "gradients", "by", "the", "number", "of", "non", "-", "padding", "tokens", "per", "mini", "-", "batch", ".", "We", "use", "weight", "normalization", "for", "all", "layers", "except", "for", "lookup", "tables", "salimans2016weight", ".", "Besides", "dropout", "on", "the", "embeddings", "and", "the", "decoder", "output", ",", "we", "also", "apply", "dropout", "to", "the", "input", "of", "the", "convolutional", "blocks", "at", "a", "rate", "of", "0.3", "srivastava2014dropout", ".", "We", "tuned", "the", "various", "parameters", "above", "and", "report", "accuracy", "on", "the", "test", "set", "by", "choosing", "the", "best", "configuration", "based", "on", "the", "validation", "set", ".", "We", "length", "normalize", "all", "scores", "and", "probabilities", "in", "the", "sequence", "-", "level", "losses", "by", "dividing", "by", "the", "number", "of", "tokens", "in", "the", "sequence", "so", "that", "scores", "are", "comparable", "between", "different", "lengths", ".", "Additionally", ",", "when", "generating", "candidate", "output", "sequences", "during", "training", "we", "limit", "the", "output", "sequence", "length", "to", "be", "less", "than", "200", "tokens", "for", "efficiency", ".", "We", "generally", "use", "16", "candidate", "sequences", "per", "training", "example", ",", "except", "for", "the", "ablations", "where", "we", "use", "5", "for", "faster", "experimental", "turnaround", ".", "subsection", ":", "Abstractive", "Summarization", "For", "summarization", "we", "use", "the", "Gigaword", "corpus", "as", "training", "data", "and", "pre", "-", "process", "it", "identically", "to", "rush2015abs", "resulting", "in", "3.8", "M", "training", "and", "190", "K", "validation", "examples", ".", "We", "evaluate", "on", "a", "Gigaword", "test", "set", "of", "2", ",", "000", "pairs", "identical", "to", "the", "one", "used", "by", "rush2015abs", "and", "report", "F1", "ROUGE", "similar", "to", "prior", "work", ".", "Our", "results", "are", "in", "terms", "of", "three", "variants", "of", "ROUGE", ",", "namely", ",", "ROUGE", "-", "1", "(", "RG", "-", "1", ",", "unigrams", ")", ",", "ROUGE", "-", "2", "(", "RG", "-", "2", ",", "bigrams", ")", ",", "and", "ROUGE", "-", "L", "(", "RG", "-", "L", ",", "longest", "-", "common", "substring", ")", ".", "Similar", "to", "ayana2016neural", "we", "use", "a", "source", "and", "target", "vocabulary", "of", "30k", "words", ".", "Our", "models", "for", "this", "task", "have", "12", "layers", "in", "the", "encoder", "and", "decoder", "each", "with", "256", "hidden", "units", "and", "kernel", "width", "3", ".", "We", "train", "on", "batches", "of", "8", ",", "000", "tokens", "with", "a", "learning", "rate", "of", "0.25", "for", "20", "epochs", "and", "then", "anneal", "as", "in", "§", "[", "reference", "]", ".", "section", ":", "Results", "subsection", ":", "Comparison", "of", "Sequence", "Level", "Losses", "First", ",", "we", "compare", "all", "objectives", "based", "on", "a", "weighted", "combination", "with", "token", "-", "level", "label", "smoothing", "(", "Equation", "[", "reference", "]", ")", ".", "We", "also", "show", "the", "likelihood", "baseline", "(", "MLE", ")", "of", "wiseman2016acl", ",", "their", "beam", "search", "optimization", "method", "(", "BSO", ")", ",", "the", "actor", "critic", "result", "of", "bahdanau2016ac", "as", "well", "as", "the", "best", "reported", "result", "on", "this", "dataset", "to", "date", "by", "huang2017npbmt", ".", "We", "show", "a", "like", "-", "for", "-", "like", "comparison", "to", "wiseman2016acl", "with", "a", "similar", "baseline", "model", "below", "(", "§", "[", "reference", "]", ")", ".", "Table", "[", "reference", "]", "shows", "that", "all", "sequence", "-", "level", "losses", "outperform", "token", "-", "level", "losses", ".", "Our", "baseline", "token", "-", "level", "results", "are", "several", "points", "above", "other", "figures", "in", "the", "literature", "and", "we", "further", "improve", "these", "results", "by", "up", "to", "0.61", "BLEU", "with", "Risk", "training", ".", "subsection", ":", "Combination", "with", "Token", "-", "Level", "Loss", "Next", ",", "we", "compare", "various", "strategies", "to", "combine", "sequence", "-", "level", "and", "token", "-", "level", "objectives", "(", "cf", ".", "§", "[", "reference", "]", ")", ".", "For", "these", "experiments", "we", "use", "5", "candidate", "sequences", "per", "training", "example", "for", "faster", "experimental", "turnaround", ".", "We", "consider", "Risk", "as", "sequence", "-", "level", "loss", "and", "label", "smoothing", "as", "token", "-", "level", "loss", ".", "Table", "[", "reference", "]", "shows", "that", "combined", "objectives", "perform", "better", "than", "pure", "Risk", ".", "The", "weighted", "combination", "(", "Equation", "[", "reference", "]", ")", "with", "performs", "best", ",", "outperforming", "constrained", "combination", "(", "Equation", "[", "reference", "]", ")", ".", "We", "also", "compare", "to", "randomly", "choosing", "between", "token", "-", "level", "and", "sequence", "-", "level", "updates", "and", "find", "it", "underperforms", "the", "more", "principled", "constrained", "strategy", ".", "In", "the", "remaining", "experiments", "we", "use", "the", "weighted", "strategy", ".", "subsection", ":", "Effect", "of", "initialization", "So", "far", "we", "initialized", "sequence", "-", "level", "models", "with", "parameters", "from", "a", "token", "-", "level", "model", "trained", "with", "label", "smoothing", ".", "Table", "[", "reference", "]", "shows", "that", "initializing", "weighted", "Risk", "with", "token", "-", "level", "label", "smoothing", "achieves", "0.7", "-", "0.8", "better", "BLEU", "compared", "to", "initializing", "with", "parameters", "from", "token", "-", "level", "likelihood", ".", "The", "improvement", "of", "initializing", "with", "TokNLL", "is", "only", "0.3", "BLEU", "with", "respect", "to", "the", "TokNLL", "baseline", ",", "whereas", ",", "the", "improvement", "from", "initializing", "with", "TokLS", "is", "0.6", "-", "0.8", "BLEU", ".", "We", "believe", "that", "the", "regularization", "provided", "by", "label", "smoothing", "leads", "to", "models", "with", "less", "sharp", "distributions", "that", "are", "a", "better", "starting", "point", "for", "sequence", "-", "level", "training", ".", "subsection", ":", "Online", "vs.", "Offline", "Candidate", "Generation", "Next", ",", "we", "consider", "the", "question", "if", "refreshing", "the", "candidate", "subset", "at", "every", "training", "step", "(", "online", ")", "results", "in", "better", "accuracy", "compared", "to", "generating", "candidates", "before", "training", "and", "keeping", "the", "set", "static", "throughout", "training", "(", "offline", ")", ".", "Table", "[", "reference", "]", "shows", "that", "offline", "generation", "gives", "lower", "accuracy", ".", "However", "the", "online", "setting", "is", "much", "slower", ",", "since", "regenerating", "the", "candidate", "set", "requires", "incremental", "(", "left", "to", "right", ")", "inference", "with", "our", "model", "which", "is", "very", "slow", "compared", "to", "efficient", "forward", "/", "backward", "over", "large", "batches", "of", "pre", "-", "generated", "hypothesis", ".", "In", "our", "setting", ",", "offline", "generation", "has", "26", "times", "higher", "throughput", "than", "the", "online", "generation", "setting", ",", "despite", "the", "high", "inference", "speed", "of", "fairseq", "gehring2017icml", ".", "subsection", ":", "Beam", "Search", "vs.", "Sampling", "and", "Candidate", "Set", "Size", "So", "far", "we", "generated", "candidates", "with", "beam", "search", ",", "however", ",", "we", "may", "also", "sample", "to", "obtain", "a", "more", "diverse", "set", "of", "candidates", "shen2016mrt", ".", "Figure", "[", "reference", "]", "compares", "beam", "search", "and", "sampling", "for", "various", "candidate", "set", "sizes", "on", "the", "validation", "set", ".", "Beam", "search", "performs", "better", "for", "all", "candidate", "set", "sizes", "considered", ".", "In", "other", "experiments", ",", "we", "rely", "on", "a", "candidate", "set", "size", "of", "16", "which", "strikes", "a", "good", "balance", "between", "efficiency", "and", "accuracy", ".", "subsection", ":", "Comparison", "to", "Beam", "-", "Search", "Optimization", "Next", ",", "we", "compare", "classical", "sequence", "-", "level", "training", "to", "the", "recently", "proposed", "Beam", "Search", "Optimization", "wiseman2016acl", ".", "To", "enable", "a", "fair", "comparison", ",", "we", "re", "-", "implement", "their", "baseline", ",", "a", "single", "layer", "LSTM", "encoder", "/", "decoder", "model", "with", "256", "-", "dimensional", "hidden", "layers", "and", "word", "embeddings", "as", "well", "as", "attention", "and", "input", "feeding", "luong2015effective", ".", "This", "baseline", "is", "trained", "with", "Adagrad", "duchi2011adaptive", "using", "a", "learning", "rate", "of", "for", "five", "epochs", ",", "with", "batches", "of", "64", "sequences", ".", "For", "sequence", "-", "level", "training", "we", "initialize", "weights", "with", "the", "baseline", "parameters", "and", "train", "with", "Adam", "kingma2014adam", "for", "another", "10", "epochs", "with", "learning", "rate", "and", "16", "candidate", "sequences", "per", "training", "example", ".", "We", "conduct", "experiments", "with", "Risk", "since", "it", "performed", "best", "in", "trial", "experiments", ".", "Different", "from", "other", "sequence", "-", "level", "experiments", "(", "§", "[", "reference", "]", ")", ",", "we", "rescale", "the", "BLEU", "scores", "in", "each", "candidate", "set", "by", "the", "difference", "between", "the", "maximum", "and", "minimum", "scores", "of", "each", "sentence", ".", "This", "avoids", "short", "sentences", "dominating", "the", "sequence", "updates", ",", "since", "candidate", "sets", "for", "short", "sentences", "have", "a", "wider", "range", "of", "BLEU", "scores", "compared", "to", "longer", "sentences", ";", "a", "similar", "rescaling", "was", "used", "by", "bahdanau2016ac", ".", "Table", "[", "reference", "]", "shows", "the", "results", "from", "wiseman2016acl", "for", "their", "token", "-", "level", "likelihood", "baseline", "(", "MLE", ")", ",", "best", "beam", "search", "optimization", "results", "(", "BSO", ")", ",", "as", "well", "as", "our", "reimplemented", "baseline", ".", "Risk", "significantly", "improves", "BLEU", "compared", "to", "our", "baseline", "at", "+", "2.75", "BLEU", ",", "which", "is", "slightly", "better", "than", "the", "+", "2.33", "BLEU", "improvement", "reported", "for", "Beam", "Search", "Optimization", "(", "cf", ".", "wiseman2016acl", ")", ".", "This", "shows", "that", "classical", "objectives", "for", "structured", "prediction", "are", "still", "very", "competitive", ".", "subsection", ":", "WMT’14", "English", "-", "French", "results", "Next", ",", "we", "experiment", "on", "the", "much", "larger", "WMT’14", "English", "-", "French", "task", "using", "the", "same", "model", "setup", "as", "gehring2017icml", ".", "We", "TokLS", "for", "15", "epochs", "and", "then", "switch", "to", "sequence", "-", "level", "training", "for", "another", "epoch", ".", "Table", "[", "reference", "]", "shows", "that", "sequence", "-", "level", "training", "can", "improve", "an", "already", "very", "strong", "model", "by", "another", "+", "0.37", "BLEU", ".", "Next", ",", "we", "improve", "the", "baseline", "by", "adding", "self", "-", "attention", "paulus2017summary", ",", "vaswani2017transformer", "to", "the", "decoder", "network", "(", "TokLS", "+", "selfatt", ")", "which", "results", "in", "a", "smaller", "gain", "of", "+", "0.2", "BLEU", "by", "Risk", ".", "If", "we", "train", "Risk", "only", "on", "the", "news", "-", "commentary", "portion", "of", "the", "training", "data", ",", "then", "we", "achieve", "state", "of", "the", "art", "accuracy", "on", "this", "dataset", "of", "41.5", "BLEU", "yingce2017deliberation", ".", "subsection", ":", "Abstractive", "Summarization", "Our", "final", "experiment", "evaluates", "sequence", "-", "level", "training", "on", "Gigaword", "headline", "summarization", ".", "There", "has", "been", "much", "prior", "art", "on", "this", "dataset", "originally", "introduced", "by", "rush2015abs", "who", "experiment", "with", "a", "feed", "-", "forward", "network", "(", "ABS", "+", ")", ".", "ayana2016neural", "report", "a", "likelihood", "baseline", "(", "RNN", "MLE", ")", "and", "also", "experiment", "with", "risk", "training", "(", "RNN", "MRT", ")", ".", "Different", "to", "their", "setup", "we", "did", "not", "find", "a", "softmax", "temperature", "to", "be", "beneficial", ",", "and", "we", "use", "beam", "search", "instead", "of", "sampling", "to", "obtain", "the", "candidate", "set", "(", "cf", ".", "§", "[", "reference", "]", ")", ".", "suzuki2017cutting", "improve", "over", "an", "MLE", "RNN", "baseline", "by", "limiting", "generation", "of", "repeated", "phrases", ".", "zhou2017seass", "also", "consider", "an", "MLE", "RNN", "baseline", "and", "add", "an", "additional", "gating", "mechanism", "for", "the", "encoder", ".", "li2017drgd", "equip", "the", "decoder", "of", "a", "similar", "network", "with", "additional", "latent", "variables", "to", "accommodate", "the", "uncertainty", "of", "this", "task", ".", "Table", "[", "reference", "]", "shows", "that", "our", "baseline", "(", "TokLS", ")", "outperforms", "all", "prior", "approaches", "in", "terms", "of", "ROUGE", "-", "2", "and", "ROUGE", "-", "L", "and", "it", "is", "on", "par", "to", "the", "best", "previous", "result", "for", "ROUGE", "-", "1", ".", "We", "optimize", "all", "three", "ROUGE", "metrics", "separately", "and", "find", "that", "Risk", "can", "further", "improve", "our", "strong", "baseline", ".", "We", "also", "compared", "Risk", "only", "training", "to", "Weighted", "on", "this", "dataset", "(", "cf", ".", "§", "[", "reference", "]", ")", "but", "accuracy", "was", "generally", "lower", "on", "the", "validation", "set", ":", "RG", "-", "1", "(", "36.59", "Risk", "only", "vs.", "36.67", "Weighted", ")", ",", "RG", "-", "2", "(", "17.34", "vs.", "18.05", ")", ",", "and", "RG", "-", "L", "(", "33.66", "vs.", "33.98", ")", ".", "section", ":", "Conclusion", "We", "present", "a", "comprehensive", "comparison", "of", "classical", "losses", "for", "structured", "prediction", "and", "apply", "them", "to", "a", "strong", "neural", "sequence", "to", "sequence", "model", ".", "We", "found", "that", "combining", "sequence", "-", "level", "and", "token", "-", "level", "losses", "is", "necessary", "to", "perform", "best", ",", "and", "so", "is", "training", "on", "candidates", "decoded", "with", "the", "current", "model", ".", "We", "show", "that", "sequence", "-", "level", "training", "improves", "state", "-", "of", "-", "the", "-", "art", "baselines", "both", "for", "IWSLT’14", "German", "-", "English", "translation", "and", "Gigaword", "abstractive", "sentence", "summarization", ".", "Structured", "prediction", "losses", "are", "very", "competitive", "to", "recent", "work", "on", "reinforcement", "or", "beam", "optimization", ".", "Classical", "expected", "risk", "can", "slightly", "outperform", "beam", "search", "optimization", "wiseman2016acl", "in", "a", "like", "-", "for", "-", "like", "setup", ".", "Future", "work", "may", "investigate", "better", "use", "of", "already", "generated", "candidates", "since", "invoking", "generation", "for", "each", "batch", "slows", "down", "training", "by", "a", "large", "factor", ",", "e.g.", ",", "mixing", "with", "fresh", "and", "older", "candidates", "inspired", "by", "MERT", ".", "bibliography", ":", "References"], "entity": [[3, 6, ["Structured", "Prediction", "Losses"], "Method"], [7, 11, ["Sequence", "to", "Sequence", "Learning"], "Task"], [19, 22, ["neural", "attention", "models"], "Method"], [24, 27, ["sequence", "-", "level"], "Task"], [29, 34, ["reinforcement", "learning", "-", "style", "methods"], "Method"], [50, 52, ["objective", "functions"], "Metric"], [59, 61, ["linear", "models"], "Method"], [62, 64, ["structured", "prediction"], "Task"], [68, 73, ["neural", "sequence", "to", "sequence", "models"], "Method"], [87, 90, ["beam", "search", "optimization"], "Method"], [108, 113, ["IWSLT’14", "German", "-", "English", "translation"], "Material"], [116, 119, ["Gigaword", "abstractive", "summarization"], "Task"], [123, 128, ["WMT’14", "English", "-", "French", "task"], "Material"], [129, 133, ["sequence", "-", "level", "training"], "Method"], [135, 136, ["BLEU"], "Metric"], [150, 154, ["Sequence", "to", "sequence", "models"], "Method"], [160, 164, ["token", "-", "level", "likelihood"], "Method"], [198, 199, ["generation"], "Task"], [217, 218, ["REINFORCE"], "Method"], [220, 223, ["actor", "-", "critic"], "Method"], [227, 230, ["beam", "search", "optimization"], "Method"], [237, 240, ["sequence", "level", "training"], "Task"], [241, 243, ["neural", "networks"], "Method"], [254, 256, ["linear", "models"], "Method"], [264, 267, ["direct", "loss", "optimization"], "Method"], [271, 273, ["machine", "translation"], "Task"], [275, 280, ["Minimum", "Error", "Rate", "Training", "algorithm"], "Method"], [281, 282, ["MERT"], "Method"], [287, 290, ["expected", "risk", "minimization"], "Method"], [295, 296, ["NLP"], "Task"], [301, 303, ["objective", "functions"], "Method"], [309, 312, ["structured", "prediction", "tasks"], "Task"], [313, 315, ["NLP", "gimpel"], "Task"], [322, 327, ["neural", "sequence", "to", "sequence", "model"], "Method"], [339, 341, ["likelihood", "training"], "Task"], [348, 350, ["margin", "loss"], "Method"], [389, 394, ["neural", "sequence", "to", "sequence", "models"], "Method"], [405, 411, ["IWSLT’14", "German", "-", "English", "translation", "task"], "Material"], [416, 420, ["Gigaword", "abstractive", "summarization", "task"], "Task"], [427, 428, ["accuracy"], "Metric"], [450, 453, ["beam", "search", "optimization"], "Method"], [460, 464, ["WMT’14", "English", "-", "French"], "Material"], [471, 473, ["risk", "minimization"], "Method"], [476, 478, ["translation", "task"], "Task"], [480, 481, ["losses"], "Method"], [482, 484, ["structured", "prediction"], "Task"], [491, 493, ["neural", "models"], "Method"], [507, 511, ["Sequence", "to", "Sequence", "Learning"], "Task"], [516, 520, ["sequence", "to", "sequence", "models"], "Method"], [522, 526, ["encoder", "-", "decoder", "approach"], "Method"], [546, 547, ["encoder"], "Method"], [549, 550, ["decoder"], "Method"], [551, 554, ["convolutional", "neural", "networks"], "Method"], [556, 559, ["recurrent", "networks", "gehring2016convolutional"], "Method"], [565, 566, ["convolution"], "Method"], [569, 573, ["computational", "and", "accuracy", "considerations"], "Metric"], [582, 584, ["model", "agnostic"], "Method"], [588, 592, ["recurrent", "and", "convolutional", "models"], "Method"], [599, 601, ["objective", "functions"], "Method"], [602, 604, ["recurrent", "models"], "Method"], [605, 606, ["LSTM"], "Method"], [662, 663, ["BLEU"], "Metric"], [682, 683, ["encoder"], "Method"], [697, 698, ["decoder"], "Method"], [850, 852, ["linear", "layer"], "Method"], [866, 867, ["softmax"], "Method"], [870, 873, ["encoder", "and", "decoder"], "Method"], [874, 878, ["gated", "convolutional", "neural", "networks"], "Method"], [883, 884, ["generation"], "Task"], [885, 887, ["Fast", "generation"], "Task"], [914, 918, ["encoder", "and", "decoder", "networks"], "Method"], [921, 923, ["block", "structure"], "Method"], [950, 954, ["1", "-", "D", "convolution"], "Method"], [981, 982, ["convolution"], "Method"], [987, 990, ["gated", "linear", "unit"], "Method"], [994, 996, ["decoder", "network"], "Method"], [1000, 1002, ["causal", "convolution"], "Method"], [1025, 1029, ["encoder", "and", "decoder", "networks"], "Method"], [1040, 1042, ["Objective", "Functions"], "Metric"], [1045, 1047, ["objective", "functions"], "Metric"], [1050, 1052, ["model", "architecture"], "Method"], [1104, 1106, ["loss", "functions"], "Method"], [1116, 1120, ["Token", "-", "Level", "Objectives"], "Metric"], [1124, 1128, ["sequence", "to", "sequence", "learning"], "Task"], [1154, 1158, ["Token", "Negative", "Log", "Likelihood"], "Method"], [1159, 1160, ["TokNLL"], "Method"], [1188, 1190, ["loss", "function"], "Method"], [1205, 1207, ["Token", "NLL"], "Method"], [1208, 1210, ["Label", "Smoothing"], "Method"], [1211, 1212, ["TokLS"], "Method"], [1213, 1215, ["Likelihood", "training"], "Method"], [1253, 1254, ["generalization"], "Task"], [1256, 1258, ["Label", "smoothing"], "Task"], [1264, 1265, ["regularizer"], "Method"], [1318, 1320, ["unigram", "distribution"], "Method"], [1332, 1334, ["Label", "smoothing"], "Task"], [1344, 1346, ["model", "prediction"], "Method"], [1365, 1367, ["label", "smoothing"], "Method"], [1445, 1449, ["task", "-", "specific", "metrics"], "Metric"], [1451, 1452, ["BLEU"], "Metric"], [1453, 1454, ["ROUGE"], "Metric"], [1519, 1523, ["Sequence", "Negative", "Log", "Likelihood"], "Method"], [1524, 1525, ["SeqNLL"], "Method"], [1528, 1529, ["TokNLL"], "Method"], [1595, 1596, ["BLEU"], "Metric"], [1597, 1598, ["ROUGE"], "Metric"], [1612, 1613, ["BLEU"], "Metric"], [1652, 1655, ["Expected", "Risk", "Minimization"], "Method"], [1656, 1657, ["Risk"], "Method"], [1667, 1669, ["cost", "function"], "Metric"], [1689, 1694, ["task", "-", "specific", "cost", "functions"], "Method"], [1697, 1698, ["BLEU"], "Metric"], [1699, 1700, ["ROUGE"], "Metric"], [1716, 1717, ["SeqNLL"], "Method"], [1775, 1779, ["phrase", "-", "based", "models"], "Method"], [1787, 1792, ["neural", "sequence", "to", "sequence", "models"], "Method"], [1825, 1827, ["expected", "cost"], "Metric"], [1844, 1845, ["REINFORCE"], "Method"], [1866, 1867, ["REINFORCE"], "Method"], [1880, 1881, ["REINFORCE"], "Method"], [1884, 1886, ["linear", "regressor"], "Method"], [1914, 1916, ["expected", "cost"], "Metric"], [1948, 1950, ["expected", "cost"], "Metric"], [1957, 1959, ["risk", "minimization"], "Task"], [1975, 1978, ["Max", "-", "Margin"], "Method"], [1978, 1979, ["MaxMargin"], "Method"], [1988, 1990, ["margin", "loss"], "Method"], [1991, 1994, ["structured", "prediction", "mmmn"], "Task"], [1995, 1996, ["structure_pred"], "Method"], [2102, 2103, ["softmax"], "Method"], [2109, 2110, ["MaxMargin"], "Method"], [2122, 2123, ["MultiMargin"], "Method"], [2152, 2153, ["MaxMargin"], "Method"], [2166, 2169, ["Softmax", "-", "Margin"], "Method"], [2171, 2172, ["SoftmaxMargin"], "Method"], [2195, 2199, ["task", "-", "specific", "costs"], "Task"], [2208, 2209, ["SeqNLL"], "Method"], [2274, 2276, ["weighted", "combination"], "Method"], [2286, 2290, ["token", "-", "level", "objective"], "Metric"], [2295, 2296, ["TokLS"], "Task"], [2322, 2324, ["constrained", "combination"], "Method"], [2356, 2360, ["token", "-", "level", "accuracy"], "Metric"], [2364, 2367, ["sequence", "-", "level"], "Metric"], [2381, 2383, ["token", "loss"], "Metric"], [2398, 2400, ["baseline", "model"], "Method"], [2417, 2420, ["fixed", "baseline", "model"], "Method"], [2434, 2437, ["Candidate", "Generation", "Strategies"], "Method"], [2491, 2493, ["search", "strategies"], "Method"], [2504, 2506, ["beam", "search"], "Method"], [2508, 2513, ["greedy", "breadth", "-", "first", "search"], "Method"], [2530, 2532, ["Beam", "search"], "Method"], [2535, 2538, ["facto", "decoding", "strategy"], "Method"], [2549, 2551, ["machine", "translation"], "Task"], [2557, 2558, ["chatterjee"], "Method"], [2568, 2572, ["model", "’s", "conditional", "distribution"], "Method"], [2574, 2576, ["beam", "search"], "Method"], [2601, 2607, ["online", "and", "offline", "candidate", "generation", "settings"], "Task"], [2615, 2617, ["online", "setting"], "Task"], [2635, 2637, ["offline", "setting"], "Task"], [2648, 2650, ["Offline", "generation"], "Task"], [2778, 2779, ["Translation"], "Task"], [2783, 2787, ["IWSLT’14", "German", "to", "English"], "Material"], [2860, 2862, ["IWSLT", "2014"], "Material"], [2865, 2866, ["TED.tst2010"], "Material"], [2867, 2868, ["TED.tst2011"], "Material"], [2869, 2870, ["TED.tst2012"], "Material"], [2871, 2872, ["TED.dev2010"], "Material"], [2873, 2874, ["TEDX.dev2012"], "Material"], [2892, 2896, ["byte", "-", "pair", "encoding"], "Method"], [2897, 2898, ["BPE"], "Method"], [2921, 2926, ["WMT’14", "English", "-", "French", "task"], "Material"], [2997, 3001, ["fairseq", "-", "py", "toolkit"], "Method"], [3013, 3015, ["translation", "models"], "Method"], [3017, 3020, ["convolutional", "encoder", "layers"], "Method"], [3022, 3025, ["convolutional", "decoder", "layers"], "Method"], [3053, 3055, ["learning", "rate"], "Metric"], [3075, 3079, ["token", "-", "level", "models"], "Method"], [3101, 3103, ["learning", "rate"], "Metric"], [3107, 3111, ["sequence", "-", "level", "models"], "Method"], [3117, 3121, ["token", "-", "level", "model"], "Method"], [3122, 3123, ["annealing"], "Method"], [3127, 3131, ["sequence", "-", "level", "models"], "Method"], [3167, 3169, ["weight", "normalization"], "Method"], [3191, 3192, ["dropout"], "Method"], [3214, 3215, ["accuracy"], "Metric"], [3299, 3300, ["ablations"], "Task"], [3311, 3313, ["Abstractive", "Summarization"], "Task"], [3314, 3315, ["summarization"], "Task"], [3318, 3320, ["Gigaword", "corpus"], "Material"], [3346, 3349, ["Gigaword", "test", "set"], "Material"], [3379, 3380, ["ROUGE"], "Metric"], [3383, 3386, ["ROUGE", "-", "1"], "Metric"], [3387, 3390, ["RG", "-", "1"], "Metric"], [3391, 3392, ["unigrams"], "Metric"], [3394, 3397, ["ROUGE", "-", "2"], "Metric"], [3398, 3401, ["RG", "-", "2"], "Metric"], [3402, 3403, ["bigrams"], "Metric"], [3406, 3409, ["ROUGE", "-", "L"], "Metric"], [3410, 3413, ["RG", "-", "L"], "Metric"], [3414, 3418, ["longest", "-", "common", "substring"], "Metric"], [3444, 3447, ["encoder", "and", "decoder"], "Method"], [3468, 3470, ["learning", "rate"], "Metric"], [3492, 3495, ["Sequence", "Level", "Losses"], "Metric"], [3504, 3506, ["weighted", "combination"], "Method"], [3507, 3512, ["token", "-", "level", "label", "smoothing"], "Method"], [3523, 3525, ["likelihood", "baseline"], "Method"], [3526, 3527, ["MLE"], "Method"], [3537, 3538, ["BSO"], "Method"], [3541, 3543, ["actor", "critic"], "Method"], [3627, 3628, ["BLEU"], "Metric"], [3636, 3640, ["Token", "-", "Level", "Loss"], "Method"], [3685, 3689, ["sequence", "-", "level", "loss"], "Task"], [3690, 3692, ["label", "smoothing"], "Method"], [3693, 3697, ["token", "-", "level", "loss"], "Task"], [3713, 3715, ["weighted", "combination"], "Method"], [3726, 3728, ["constrained", "combination"], "Method"], [3757, 3759, ["constrained", "strategy"], "Method"], [3767, 3769, ["weighted", "strategy"], "Method"], [3779, 3783, ["sequence", "-", "level", "models"], "Method"], [3787, 3791, ["token", "-", "level", "model"], "Method"], [3793, 3795, ["label", "smoothing"], "Method"], [3806, 3811, ["token", "-", "level", "label", "smoothing"], "Method"], [3816, 3817, ["BLEU"], "Metric"], [3833, 3834, ["TokNLL"], "Method"], [3837, 3838, ["BLEU"], "Metric"], [3852, 3853, ["TokLS"], "Method"], [3857, 3858, ["BLEU"], "Metric"], [3863, 3864, ["regularization"], "Method"], [3866, 3868, ["label", "smoothing"], "Method"], [3882, 3886, ["sequence", "-", "level", "training"], "Method"], [3891, 3894, ["Offline", "Candidate", "Generation"], "Task"], [3915, 3916, ["accuracy"], "Metric"], [3939, 3941, ["offline", "generation"], "Task"], [3943, 3944, ["accuracy"], "Metric"], [3947, 3949, ["online", "setting"], "Task"], [3976, 3979, ["forward", "/", "backward"], "Method"], [3992, 3994, ["offline", "generation"], "Task"], [3998, 3999, ["throughput"], "Metric"], [4001, 4004, ["online", "generation", "setting"], "Task"], [4008, 4010, ["inference", "speed"], "Metric"], [4011, 4013, ["fairseq", "gehring2017icml"], "Method"], [4016, 4018, ["Beam", "Search"], "Method"], [4019, 4020, ["Sampling"], "Method"], [4030, 4032, ["beam", "search"], "Method"], [4054, 4056, ["beam", "search"], "Method"], [4057, 4058, ["sampling"], "Method"], [4068, 4070, ["Beam", "search"], "Method"], [4098, 4099, ["efficiency"], "Metric"], [4100, 4101, ["accuracy"], "Metric"], [4106, 4110, ["Beam", "-", "Search", "Optimization"], "Method"], [4115, 4119, ["sequence", "-", "level", "training"], "Method"], [4123, 4126, ["Beam", "Search", "Optimization"], "Method"], [4142, 4149, ["single", "layer", "LSTM", "encoder", "/", "decoder", "model"], "Method"], [4172, 4173, ["Adagrad"], "Method"], [4176, 4178, ["learning", "rate"], "Metric"], [4190, 4194, ["sequence", "-", "level", "training"], "Method"], [4204, 4206, ["Adam", "kingma2014adam"], "Method"], [4211, 4213, ["learning", "rate"], "Metric"], [4316, 4321, ["token", "-", "level", "likelihood", "baseline"], "Method"], [4322, 4323, ["MLE"], "Method"], [4326, 4329, ["beam", "search", "optimization"], "Method"], [4331, 4332, ["BSO"], "Method"], [4338, 4340, ["reimplemented", "baseline"], "Method"], [4341, 4342, ["Risk"], "Method"], [4344, 4345, ["BLEU"], "Metric"], [4352, 4353, ["BLEU"], "Metric"], [4366, 4369, ["Beam", "Search", "Optimization"], "Method"], [4381, 4383, ["structured", "prediction"], "Task"], [4390, 4394, ["WMT’14", "English", "-", "French"], "Material"], [4403, 4408, ["WMT’14", "English", "-", "French", "task"], "Material"], [4425, 4429, ["sequence", "-", "level", "training"], "Method"], [4439, 4443, ["sequence", "-", "level", "training"], "Method"], [4454, 4455, ["BLEU"], "Metric"], [4472, 4474, ["decoder", "network"], "Method"], [4475, 4476, ["TokLS"], "Method"], [4477, 4478, ["selfatt"], "Method"], [4488, 4489, ["BLEU"], "Metric"], [4515, 4516, ["accuracy"], "Metric"], [4521, 4522, ["BLEU"], "Metric"], [4526, 4528, ["Abstractive", "Summarization"], "Task"], [4532, 4536, ["sequence", "-", "level", "training"], "Method"], [4537, 4540, ["Gigaword", "headline", "summarization"], "Task"], [4558, 4562, ["feed", "-", "forward", "network"], "Method"], [4563, 4565, ["ABS", "+"], "Method"], [4570, 4572, ["likelihood", "baseline"], "Method"], [4580, 4582, ["risk", "training"], "Method"], [4583, 4585, ["RNN", "MRT"], "Method"], [4605, 4607, ["beam", "search"], "Method"], [4649, 4651, ["gating", "mechanism"], "Method"], [4658, 4659, ["decoder"], "Method"], [4661, 4663, ["similar", "network"], "Method"], [4684, 4685, ["TokLS"], "Method"], [4693, 4696, ["ROUGE", "-", "2"], "Metric"], [4697, 4700, ["ROUGE", "-", "L"], "Metric"], [4711, 4714, ["ROUGE", "-", "1"], "Metric"], [4753, 4754, ["accuracy"], "Metric"], [4774, 4777, ["RG", "-", "2"], "Metric"], [4784, 4787, ["RG", "-", "L"], "Metric"], [4805, 4807, ["structured", "prediction"], "Task"], [4813, 4818, ["neural", "sequence", "to", "sequence", "model"], "Method"], [4852, 4856, ["sequence", "-", "level", "training"], "Method"], [4867, 4872, ["IWSLT’14", "German", "-", "English", "translation"], "Material"], [4873, 4877, ["Gigaword", "abstractive", "sentence", "summarization"], "Material"], [4878, 4881, ["Structured", "prediction", "losses"], "Method"], [4888, 4892, ["reinforcement", "or", "beam", "optimization"], "Task"], [4894, 4896, ["expected", "risk"], "Method"], [4899, 4902, ["beam", "search", "optimization"], "Method"], [4905, 4911, ["like", "-", "for", "-", "like", "setup"], "Task"], [4930, 4931, ["training"], "Task"], [4946, 4947, ["MERT"], "Method"], [354, 356, ["risk", "training"], "Method"], [366, 370, ["token", "-", "level", "likelihood"], "Method"], [664, 665, ["ROUGE"], "Metric"], [1676, 1677, ["Risk"], "Method"], [1804, 1805, ["Risk"], "Method"], [1810, 1811, ["REINFORCE"], "Method"], [1856, 1857, ["Risk"], "Method"], [1908, 1909, ["Risk"], "Method"], [1945, 1946, ["REINFORCE"], "Method"], [2042, 2043, ["BLEU"], "Metric"], [2297, 2298, ["Risk"], "Method"], [2912, 2913, ["BLEU"], "Metric"], [3364, 3365, ["ROUGE"], "Metric"], [3532, 3535, ["beam", "search", "optimization"], "Method"], [3629, 3630, ["Risk"], "Method"], [3683, 3684, ["Risk"], "Method"], [3710, 3711, ["Risk"], "Method"], [3804, 3805, ["Risk"], "Method"], [3823, 3827, ["token", "-", "level", "likelihood"], "Method"], [4225, 4226, ["Risk"], "Method"], [4251, 4252, ["BLEU"], "Metric"], [4290, 4291, ["BLEU"], "Metric"], [4362, 4363, ["BLEU"], "Metric"], [4490, 4491, ["Risk"], "Method"], [4495, 4496, ["Risk"], "Method"], [4574, 4575, ["MLE"], "Method"], [4628, 4629, ["MLE"], "Method"], [4642, 4643, ["MLE"], "Method"], [4719, 4720, ["ROUGE"], "Metric"], [4725, 4726, ["Risk"], "Method"], [4736, 4737, ["Risk"], "Method"], [4762, 4765, ["RG", "-", "1"], "Metric"], [4767, 4768, ["Risk"], "Method"]]} {"doc_id": "40a63746a710baf4a694fd5a4dd8b5a3d9fc2846", "text": ["document", ":", "Invertible", "Conditional", "GANs", "for", "image", "editing", "Generative", "Adversarial", "Networks", "(", "GANs", ")", "have", "recently", "demonstrated", "to", "successfully", "approximate", "complex", "data", "distributions", ".", "A", "relevant", "extension", "of", "this", "model", "is", "conditional", "GANs", "(", "cGANs", ")", ",", "where", "the", "introduction", "of", "external", "information", "allows", "to", "determine", "specific", "representations", "of", "the", "generated", "images", ".", "In", "this", "work", ",", "we", "evaluate", "encoders", "to", "inverse", "the", "mapping", "of", "a", "cGAN", ",", "i.e.", ",", "mapping", "a", "real", "image", "into", "a", "latent", "space", "and", "a", "conditional", "representation", ".", "This", "allows", ",", "for", "example", ",", "to", "reconstruct", "and", "modify", "real", "images", "of", "faces", "conditioning", "on", "arbitrary", "attributes", ".", "Additionally", ",", "we", "evaluate", "the", "design", "of", "cGANs", ".", "The", "combination", "of", "an", "encoder", "with", "a", "cGAN", ",", "which", "we", "call", "Invertible", "cGAN", "(", "IcGAN", ")", ",", "enables", "to", "re", "-", "generate", "real", "images", "with", "deterministic", "complex", "modifications", ".", "section", ":", "Introduction", "Image", "editing", "can", "be", "performed", "at", "different", "levels", "of", "complexity", "and", "abstraction", ".", "Common", "operations", "consist", "in", "simply", "applying", "a", "filter", "to", "an", "image", "to", ",", "for", "example", ",", "augment", "the", "contrast", "or", "convert", "to", "grayscale", ".", "These", ",", "however", ",", "are", "low", "-", "complex", "operations", "that", "do", "not", "necessarily", "require", "to", "comprehend", "the", "scene", "or", "object", "that", "the", "image", "is", "representing", ".", "On", "the", "other", "hand", ",", "if", "one", "would", "want", "to", "modify", "the", "attributes", "of", "a", "face", "(", "e.g.", "add", "a", "smile", ",", "change", "the", "hair", "color", "or", "even", "the", "gender", ")", ",", "this", "is", "a", "more", "complex", "and", "challenging", "modification", "to", "perform", ".", "In", "this", "case", ",", "in", "order", "to", "obtain", "realistic", "results", ",", "a", "skilled", "human", "with", "an", "image", "edition", "software", "would", "often", "be", "required", ".", "A", "solution", "to", "automatically", "perform", "these", "non", "-", "trivial", "operations", "relies", "on", "generative", "models", ".", "Natural", "image", "generation", "has", "been", "a", "strong", "research", "topic", "for", "many", "years", ",", "but", "it", "has", "not", "been", "until", "2015", "that", "promising", "results", "have", "been", "achieved", "with", "deep", "learning", "techniques", "combined", "with", "generative", "modeling", "Gregor2015", ",", "Radford2015", ".", "Generative", "Adversarial", "Networks", "(", "GANs", ")", "is", "one", "of", "the", "state", "-", "of", "-", "the", "-", "art", "approaches", "for", "image", "generation", ".", "GANs", "are", "especially", "interesting", "as", "they", "are", "directly", "optimized", "towards", "generating", "the", "most", "plausible", "and", "realistic", "data", ",", "as", "opposed", "to", "other", "models", "(", "e.g.", "Variational", "Autoencoders", ")", ",", "which", "focus", "on", "an", "image", "reconstruction", "loss", ".", "Additionally", ",", "GANs", "are", "able", "to", "explicitly", "control", "generated", "images", "features", "with", "a", "conditional", "extension", ",", "conditional", "GANs", "(", "cGANs", ")", ".", "However", ",", "the", "GAN", "framework", "lacks", "an", "inference", "mechanism", ",", "i.e.", ",", "finding", "the", "latent", "representation", "of", "an", "input", "image", ",", "which", "is", "a", "necessary", "step", "for", "being", "able", "to", "reconstruct", "and", "modify", "real", "images", ".", "In", "order", "to", "overcome", "this", "limitation", ",", "in", "this", "paper", "we", "introduce", "Invertible", "Conditional", "GANs", "(", "IcGANs", ")", "for", "complex", "image", "editing", "as", "the", "union", "of", "an", "encoder", "used", "jointly", "with", "a", "cGAN", ".", "This", "model", "allows", "to", "map", "real", "images", "into", "a", "high", "-", "feature", "space", "(", "encoder", ")", "and", "perform", "meaningful", "modifications", "on", "them", "(", "cGAN", ")", ".", "As", "a", "result", ",", "we", "can", "explicitly", "control", "the", "attributes", "of", "a", "real", "image", "(", "Figure", "[", "reference", "]", ")", ",", "which", "could", "be", "potentially", "useful", "in", "several", "applications", ",", "be", "it", "creative", "processes", ",", "data", "augmentation", "or", "face", "profiling", ".", "[", "width=0.85", "]", "images", "/", "1.intro", "/", "pull_figure", "The", "summary", "of", "contributions", "of", "our", "work", "is", "the", "following", ":", "Proposing", "IcGANs", ",", "composed", "of", "two", "crucial", "parts", ":", "an", "encoder", "and", "a", "cGAN", ".", "We", "apply", "this", "model", "to", "MNIST", "LeCun1998", "and", "CelebA", "celeba", "datasets", ",", "which", "allows", "performing", "meaningful", "and", "realistic", "editing", "operations", "on", "them", "by", "arbitrarily", "changing", "the", "conditional", "information", ".", "Introducing", "an", "encoder", "in", "the", "conditional", "GAN", "framework", "to", "compress", "a", "real", "image", "into", "a", "latent", "representation", "and", "conditional", "vector", ".", "We", "consider", "several", "designs", "and", "training", "procedures", "to", "leverage", "the", "performance", "obtained", "from", "available", "conditional", "information", ".", "Evaluating", "and", "refining", "cGANs", "through", "conditional", "position", "and", "conditional", "sampling", "to", "enhance", "the", "quality", "of", "generated", "images", ".", "section", ":", "Related", "work", "There", "are", "different", "approaches", "for", "generative", "models", ".", "Among", "them", ",", "there", "are", "two", "promising", "ones", "that", "are", "recently", "pushing", "the", "state", "-", "of", "-", "the", "-", "art", "with", "highly", "plausible", "generated", "images", ".", "The", "first", "one", "is", "Variational", "Autoencoders", "(", "VAE", ")", "Gregor2015", ",", "Kingma2013", ",", "Rezende2014", ",", "Kingma2014", ",", "which", "impose", "a", "prior", "representation", "space", "(", "e.g.", "normal", "distribution", ")", "in", "order", "to", "regularize", "and", "constrain", "the", "model", "to", "sample", "from", "it", ".", "However", ",", "VAEs", "main", "limitation", "is", "the", "pixel", "-", "wise", "reconstruction", "error", "used", "as", "a", "loss", "function", ",", "which", "causes", "the", "output", "images", "to", "look", "blurry", ".", "The", "second", "approach", "is", "Generative", "Adversarial", "Nets", "(", "GANs", ")", ".", "Originally", "proposed", "by", "Goodfellow", "et", "al", ".", "Goodfellow2014", ",", "GANs", "have", "been", "improved", "with", "a", "deeper", "architecture", "(", "DCGAN", ")", "by", "Radford", "et", "al", ".", "Radford2015", ".", "The", "latest", "advances", "introduced", "several", "techniques", "that", "improve", "the", "overall", "performance", "for", "training", "GANs", "Salimans2016", "and", "an", "unsupervised", "approach", "to", "disentangle", "feature", "representations", "Chen2016", ".", "Additionally", ",", "the", "most", "advanced", "and", "recent", "work", "on", "cGANs", "trains", "a", "model", "to", "generate", "realistic", "images", "from", "text", "descriptions", "and", "landmarks", ".", "Our", "work", "is", "considered", "in", "the", "content", "of", "the", "GAN", "framework", ".", "The", "baseline", "will", "be", "the", "work", "of", "Radford", "’s", "et", "al", ".", "(", "DCGANs", ")", "Radford2015", ",", "which", "we", "will", "add", "a", "conditional", "extension", ".", "The", "difference", "of", "our", "approach", "to", "prior", "work", "is", "that", "we", "also", "propose", "an", "encoder", "(", "Invertible", "cGAN", ")", "with", "which", "we", "can", ",", "given", "an", "input", "image", ",", "to", "obtain", "its", "representation", "as", "a", "latent", "variable", "and", "a", "conditional", "vector", ".", "Then", ",", "we", "can", "modify", "and", "to", "re", "-", "generate", "the", "original", "image", "with", "complex", "variations", ".", "Dumoulin", "et", "al", ".", "Dumoulin2016", "and", "Donahue", "et", "al", ".", "Donahue2016", "also", "proposed", "an", "encoder", "in", "GANs", ",", "but", "in", "a", "non", "-", "conditional", "and", "jointly", "trained", "setting", ".", "Additionally", ",", "Makhzani", "et", "al", ".", "Makhzani2015", "and", "Larsen", "et", "al", ".", "Larsen2015", "proposed", "a", "similar", "idea", "to", "this", "paper", "by", "combining", "a", "VAE", "and", "a", "GAN", "with", "promising", "results", ".", "Reed", "et", "al", ".", "Reed2016", "implemented", "an", "encoder", "in", "a", "similar", "fashion", "to", "our", "approach", ".", "This", "paper", "builds", "alongside", "their", "work", "in", "a", "complementary", "manner", ".", "In", "our", "case", ",", "we", "analyze", "more", "deeply", "the", "encoder", "by", "including", "conditional", "information", "encoding", "and", "testing", "different", "architectures", "and", "training", "approaches", ".", "Also", ",", "we", "evaluate", "unexplored", "design", "decisions", "for", "building", "a", "cGAN", ".", "section", ":", "Background", ":", "Generative", "Adversarial", "Networks", "A", "GAN", "is", "composed", "of", "two", "neural", "networks", ",", "a", "generator", "and", "a", "discriminator", ".", "Both", "networks", "are", "iteratively", "trained", "competing", "against", "each", "other", "in", "a", "minimax", "game", ".", "The", "generator", "aims", "to", "approximate", "the", "underlying", "unknown", "data", "distribution", "to", "fool", "the", "discriminator", ",", "whilst", "the", "discriminator", "is", "focused", "on", "being", "able", "to", "tell", "which", "samples", "are", "real", "or", "generated", ".", "On", "convergence", ",", "we", "want", ",", "where", "is", "the", "generator", "distribution", ".", "More", "formally", ",", "considering", "the", "function", ",", "where", "and", "are", "the", "parameters", "of", "the", "generator", "and", "discriminator", "respectively", ",", "we", "can", "formulate", "GAN", "training", "as", "optimizing", "where", "is", "a", "vector", "noise", "sampled", "from", "a", "known", "simple", "distribution", "(", "e.g.", "normal", ")", ".", "GAN", "framework", "can", "be", "extended", "with", "conditional", "GANs", "(", "cGANs", ")", "Mirza2014", ".", "They", "are", "quite", "similar", "to", "vanilla", "(", "non", "-", "conditional", ")", "GANs", ",", "the", "only", "difference", "is", "that", ",", "in", "this", "case", ",", "we", "have", "extra", "information", "(", "e.g.", "class", "labels", ",", "attribute", "information", ")", "for", "a", "given", "real", "sample", ".", "Conditional", "information", "strictly", "depends", "on", "real", "samples", ",", "but", "we", "can", "model", "a", "density", "model", "in", "order", "to", "sample", "generated", "labels", "for", "generated", "data", ".", "Then", ",", "Equation", "[", "reference", "]", "can", "be", "reformulated", "for", "the", "cGAN", "extension", "as", "Once", "a", "cGAN", "is", "trained", ",", "it", "allows", "us", "to", "generate", "samples", "using", "two", "level", "of", "variations", ":", "constrained", "and", "unconstrained", ".", "Constrained", "variations", "are", "modeled", "with", "as", "it", "directly", "correlates", "with", "features", "of", "the", "data", "that", "are", "explicitly", "correlated", "with", "and", "the", "data", "itself", ".", "Then", ",", "all", "the", "other", "variations", "of", "the", "data", "not", "modeled", "by", "(", "unconstrained", "variations", ")", "are", "encoded", "in", ".", "section", ":", "Invertible", "Conditional", "GANs", "We", "introduce", "Invertible", "Conditional", "GANs", "(", "IcGANs", ")", ",", "which", "are", "composed", "of", "a", "cGAN", "and", "an", "encoder", ".", "Even", "though", "encoders", "have", "recently", "been", "introduced", "into", "the", "GAN", "framework", ",", "we", "are", "the", "first", "ones", "to", "include", "and", "leverage", "the", "conditional", "information", "into", "the", "design", "of", "the", "encoding", "process", ".", "In", "section", "[", "reference", "]", "we", "explain", "how", "and", "why", "an", "encoder", "is", "included", "in", "the", "GAN", "framework", "for", "a", "conditional", "setting", ".", "In", "section", "[", "reference", "]", ",", "we", "introduce", "our", "approach", "to", "refine", "cGANs", "on", "two", "aspects", ":", "conditional", "position", "and", "conditional", "sampling", ".", "The", "model", "architecture", "is", "described", "in", "section", "[", "reference", "]", ".", "subsection", ":", "Encoder", "A", "generator", "from", "a", "GAN", "framework", "does", "not", "have", "the", "capability", "to", "map", "a", "real", "image", "to", "its", "latent", "representation", ".", "To", "overcome", "this", "problem", ",", "we", "can", "train", "an", "encoder", "/", "inference", "network", "that", "approximately", "inverses", "this", "mapping", ".", "This", "inversion", "would", "allow", "us", "to", "have", "a", "latent", "representation", "from", "a", "real", "image", "and", ",", "then", ",", "we", "would", "be", "able", "to", "explore", "the", "latent", "space", "by", "interpolating", "or", "adding", "variations", "on", "it", ",", "which", "would", "result", "in", "variations", "on", "the", "generated", "image", ".", "If", "combined", "with", "a", "cGAN", ",", "once", "the", "latent", "representation", "has", "been", "obtained", ",", "explicitly", "controlled", "variations", "can", "be", "added", "to", "an", "input", "image", "via", "conditional", "information", "(", "e.g.", "generate", "a", "certain", "digit", "in", "MNIST", "or", "specify", "face", "attributes", "on", "a", "face", "dataset", ")", ".", "We", "call", "this", "combination", "Invertible", "cGAN", ",", "as", "now", "the", "mapping", "can", "be", "inverted", ":", "and", ",", "where", "is", "an", "input", "image", "and", "its", "reconstruction", ".", "See", "Figure", "[", "reference", "]", "for", "an", "example", "on", "how", "a", "trained", "IcGAN", "is", "used", ".", "[", "width=0.92", "]", "images", "/", "4.icgans", "/", "icgan_overview", "Our", "approach", "consists", "of", "training", "an", "encoder", "once", "the", "cGAN", "has", "been", "trained", ",", "as", "similarly", "considered", "by", "Reed", "et", "al", ".", "In", "our", "case", ",", "however", ",", "the", "encoder", "is", "composed", "of", "two", "sub", "-", "encoders", ":", ",", "which", "encodes", "an", "image", "to", ",", "and", ",", "which", "encodes", "an", "image", "to", ".", "To", "train", "we", "use", "the", "generator", "to", "create", "a", "dataset", "of", "generated", "images", "and", "their", "latent", "vectors", ",", "and", "then", "minimize", "a", "squared", "reconstruction", "loss", "(", "Eq", ".", "[", "reference", "]", ")", ".", "For", ",", "we", "initially", "used", "generated", "images", "and", "their", "conditional", "information", "for", "training", ".", "However", ",", "we", "found", "that", "generated", "images", "tend", "to", "be", "noisier", "than", "real", "ones", "and", ",", "in", "this", "specific", "case", ",", "we", "could", "improve", "by", "directly", "training", "with", "real", "images", "and", "labels", "from", "the", "dataset", "(", "Eq", ".", "[", "reference", "]", ")", ".", "Although", "and", "might", "seem", "completely", "independent", ",", "we", "can", "adopt", "different", "strategies", "to", "make", "them", "interact", "and", "leverage", "the", "conditional", "information", "(", "for", "an", "evaluation", "of", "them", ",", "see", "section", "[", "reference", "]", ")", ":", "SNG", ":", "One", "single", "encoder", "with", "shared", "layers", "and", "two", "outputs", ".", "That", "is", ",", "and", "are", "embedded", "in", "a", "single", "encoder", ".", "IND", ":", "Two", "independent", "encoders", ".", "and", "are", "trained", "separately", ".", "IND", "-", "COND", ":", "Two", "encoders", ",", "where", "is", "conditioned", "on", "the", "output", "of", "encoder", ".", "Recently", ",", "Dumoulin", "et", "al", ".", "Dumoulin2016", "and", "Donahue", "et", "al", ".", "Donahue2016", "proposed", "different", "approaches", "on", "how", "to", "train", "an", "encoder", "in", "the", "GAN", "framework", ".", "One", "of", "the", "most", "interesting", "approaches", "consists", "in", "jointly", "training", "the", "encoder", "with", "both", "the", "discriminator", "and", "the", "generator", ".", "Although", "this", "approach", "is", "promising", ",", "our", "work", "has", "been", "completely", "independent", "of", "these", "articles", "and", "focuses", "on", "another", "direction", ",", "since", "we", "consider", "the", "encoder", "in", "a", "conditional", "setting", ".", "Consequently", ",", "we", "implemented", "our", "aforementioned", "approach", "which", "performs", "nearly", "equally", "Donahue2016", "to", "their", "strategy", ".", "subsection", ":", "Conditional", "GAN", "We", "consider", "two", "main", "design", "decisions", "concerning", "cGANs", ".", "The", "first", "one", "is", "to", "find", "the", "optimal", "conditional", "position", "on", "the", "generator", "and", "discriminator", ",", "which", ",", "to", "our", "knowledge", ",", "has", "not", "been", "previously", "addressed", ".", "Secondly", ",", "we", "discuss", "the", "best", "approach", "to", "sample", "conditional", "information", "for", "the", "generator", ".", "Conditional", "position", "In", "the", "cGAN", ",", "the", "conditional", "information", "vector", "needs", "to", "be", "introduced", "in", "both", "the", "generator", "and", "the", "discriminator", ".", "In", "the", "generator", ",", "and", "(", "where", "are", "always", "concatenated", "in", "the", "filter", "dimension", "at", "the", "input", "level", "Reed2016", ",", "Mirza2014", ",", "Gauthier2014", ".", "As", "for", "the", "discriminator", ",", "different", "authors", "insert", "in", "different", "parts", "of", "the", "model", "Reed2016", ",", "Mirza2014", ",", "Gauthier2014", ".", "We", "expect", "that", "the", "earlier", "is", "positioned", "in", "the", "model", "the", "better", "since", "the", "model", "is", "allowed", "to", "have", "more", "learning", "interactions", "with", ".", "Experiments", "regarding", "the", "optimal", "position", "will", "be", "detailed", "in", "section", "[", "reference", "]", ".", "Conditional", "sampling", "There", "are", "two", "types", "of", "conditional", "information", ",", "and", ".", "The", "first", "one", "is", "trivially", "sampled", "from", "and", "is", "used", "for", "training", "the", "discriminator", "with", "a", "real", "image", "and", "its", "associated", "label", ".", "The", "second", "one", "is", "sampled", "from", "and", "serves", "as", "input", "to", "the", "generator", "along", "with", "a", "latent", "vector", "to", "generate", "an", "image", ",", "and", "it", "can", "be", "sampled", "using", "different", "approaches", ":", "Kernel", "density", "estimation", ":", "also", "known", "as", "Parzen", "window", "estimation", ",", "it", "consists", "in", "randomly", "sampling", "from", "a", "kernel", "(", "e.g.", "Gaussian", "kernel", "with", "a", "cross", "-", "validated", ")", ".", "Direct", "interpolation", ":", "interpolate", "between", "label", "vectors", "from", "the", "training", "set", "Reed2016", ".", "The", "reasoning", "behind", "this", "approach", "is", "that", "interpolations", "can", "belong", "to", "the", "label", "distribution", ".", "Sampling", "from", "the", "training", "set", ",", ":", "Use", "directly", "the", "real", "labels", "from", "the", "training", "set", ".", "As", "Gauthier", "Gauthier2014", "pointed", "out", ",", "unlike", "the", "previous", "two", "approaches", ",", "this", "method", "could", "overfit", "the", "model", "by", "using", "the", "conditional", "information", "to", "reproduce", "the", "images", "of", "the", "training", "set", ".", "However", ",", "this", "is", "only", "likely", "to", "occur", "if", "the", "conditional", "information", "is", ",", "to", "some", "extent", ",", "unique", "for", "each", "image", ".", "In", "the", "case", "where", "the", "attributes", "of", "an", "image", "are", "binary", ",", "one", "attribute", "vector", "could", "describe", "a", "varied", "and", "large", "enough", "subset", "of", "images", ",", "preventing", "the", "model", "from", "overfitting", "given", ".", "Kernel", "density", "estimation", "and", "direct", "interpolation", "are", ",", "at", "the", "end", ",", "two", "different", "ways", "to", "interpolate", "on", ".", "Nevertheless", ",", "interpolation", "is", "mostly", "suitable", "when", "the", "attribute", "information", "is", "composed", "of", "real", "vectors", ",", "not", "binary", "ones", ".", "It", "is", "not", "the", "case", "of", "the", "binary", "conditional", "information", "of", "the", "datasets", "used", "in", "this", "paper", "(", "see", "section", "[", "reference", "]", "for", "dataset", "information", ")", ".", "Directly", "interpolating", "binary", "vectors", "would", "not", "create", "plausible", "conditional", "information", ",", "as", "an", "interpolated", "vector", "would", "not", "belong", "to", "nor", ".", "Using", "a", "kernel", "density", "estimation", "would", "not", "make", "sense", "either", ",", "as", "all", "the", "binary", "labels", "would", "fall", "in", "the", "corners", "of", "a", "hypercube", ".", "Therefore", ",", "we", "will", "directly", "sample", "from", ".", "subsection", ":", "Model", "architecture", "Conditional", "GAN", "The", "work", "of", "this", "paper", "is", "based", "on", "the", "Torch", "implementation", "of", "the", "DCGAN", ".", "We", "use", "the", "recommended", "configuration", "for", "the", "DCGAN", ",", "which", "trains", "with", "the", "Adam", "optimizer", "adam2014", "(", ")", "with", "a", "learning", "rate", "of", "and", "a", "mini", "-", "batch", "size", "of", "(", "samples", "drawn", "independently", "at", "each", "update", "step", ")", "during", "epochs", ".", "The", "output", "image", "size", "used", "as", "a", "baseline", "is", ".", "Also", ",", "we", "train", "the", "cGAN", "with", "the", "matching", "-", "aware", "discriminator", "method", "from", "Reed", "et", "al", ".", ".", "In", "Figure", "[", "reference", "]", "we", "show", "an", "overview", "architecture", "of", "both", "generator", "and", "discriminator", "for", "the", "cGAN", ".", "For", "a", "more", "detailed", "description", "of", "the", "model", "see", "Table", "[", "reference", "]", ".", "[", "]", "[", "width=0.50", "]", "images", "/", "5.impl", "/", "cGAN_G_arch", "[", "]", "[", "width=0.475", "]", "images", "/", "5.impl", "/", "cGAN_D_arch", "width=1.0", "Encoder", "For", "simplicity", ",", "we", "show", "the", "architecture", "of", "the", "IND", "encoders", "(", "Table", "[", "reference", "]", ")", ",", "as", "they", "are", "the", "ones", "that", "give", "the", "best", "performance", ".", "Batch", "Normalization", "and", "non", "-", "linear", "activation", "functions", "are", "removed", "from", "the", "last", "layer", "to", "guarantee", "that", "the", "output", "distribution", "is", "similar", "to", ".", "Additionally", ",", "after", "trying", "different", "configurations", ",", "we", "have", "replaced", "the", "last", "two", "convolutional", "layers", "with", "two", "fully", "connected", "layers", "at", "the", "end", "of", "the", "encoder", ",", "which", "yields", "a", "lower", "error", ".", "The", "training", "configuration", "(", "Adam", "optimizer", ",", "batch", "size", ",", "etc", ")", "is", "the", "same", "as", "the", "one", "used", "for", "the", "cGAN", "model", ".", "width=0.6", "section", ":", "Experiments", "subsection", ":", "Datasets", "We", "use", "two", "image", "datasets", "of", "different", "complexity", "and", "variation", ",", "MNIST", "LeCun1998", "and", "CelebFaces", "Attributes", "(", "CelebA", ")", "celeba", ".", "MNIST", "is", "a", "digit", "dataset", "of", "grayscale", "images", "composed", "of", "60", ",", "000", "training", "images", "and", "10", ",", "000", "test", "images", ".", "Each", "sample", "is", "a", "centered", "image", "labeled", "with", "the", "class", "of", "the", "digit", "(", "0", "to", "9", ")", ".", "CelebA", "is", "a", "dataset", "composed", "of", "202", ",", "599", "face", "colored", "images", "and", "40", "attribute", "binary", "vectors", ".", "We", "use", "the", "aligned", "and", "cropped", "version", "and", "scale", "the", "images", "down", "to", ".", "We", "also", "use", "the", "official", "train", "and", "test", "partitions", ",", "182", "K", "for", "training", "and", "20", "K", "for", "testing", ".", "Of", "the", "original", "40", "attributes", ",", "we", "filter", "those", "that", "do", "not", "have", "a", "clear", "visual", "impact", "on", "the", "generated", "images", ",", "which", "leaves", "a", "total", "of", "18", "attributes", ".", "We", "will", "evaluate", "the", "quality", "of", "generated", "samples", "of", "both", "datasets", ".", "However", ",", "a", "quantitative", "evaluation", "will", "be", "performed", "on", "CelebA", "only", ",", "as", "it", "is", "considerably", "more", "complex", "than", "MNIST", ".", "subsection", ":", "Evaluating", "the", "conditional", "GAN", "The", "goals", "of", "this", "experiment", "are", "two", ".", "First", ",", "we", "evaluate", "the", "general", "performance", "of", "the", "cGAN", "with", "an", "attribute", "predictor", "network", "(", "Anet", ")", "on", "CelebA", "dataset", ".", "Second", ",", "we", "test", "the", "impact", "of", "adding", "in", "different", "layers", "of", "the", "cGAN", "(", "section", "[", "reference", "]", ",", "conditional", "position", ")", ".", "We", "use", "an", "Anet", "as", "a", "way", "to", "make", "a", "quantitative", "evaluation", "in", "a", "similar", "manner", "as", "Salimans", "et", "al", ".", "Inception", "model", ",", "as", "the", "output", "given", "by", "this", "Anet", "(", "i.e.", ",", "which", "attributes", "are", "detected", "on", "a", "generated", "sample", ")", "is", "a", "good", "indicator", "of", "the", "generator", "ability", "to", "model", "them", ".", "In", "other", "words", ",", "if", "the", "predicted", "Anet", "attributes", "are", "closer", "to", "the", "original", "attributes", "used", "to", "generate", "an", "image", ",", "we", "expect", "that", "the", "generator", "has", "successfully", "learned", "the", "capability", "to", "generate", "new", "images", "considering", "the", "semantic", "meaning", "of", "the", "attributes", ".", "Therefore", ",", "we", "use", "the", "generator", "to", "create", "images", "conditioned", "on", "attribute", "vectors", "(", "i.e.", ")", ",", "and", "make", "the", "Anet", "predict", "them", ".", "Using", "the", "Anet", "output", ",", "we", "build", "a", "confusion", "matrix", "for", "each", "attribute", "and", "compute", "the", "mean", "accuracy", "and", "F1", "-", "Score", "to", "test", "the", "model", "and", "its", "inserted", "optimal", "position", "of", "in", "both", "generator", "and", "discriminator", ".", "width=0.80", "In", "Table", "[", "reference", "]", "we", "can", "see", "how", "cGANs", "have", "successfully", "learned", "to", "generate", "the", "visual", "representations", "of", "the", "conditional", "attributes", "with", "an", "overall", "accuracy", "of", "%", ".", "The", "best", "accuracy", "is", "achieved", "by", "inserting", "in", "the", "first", "convolutional", "layer", "of", "the", "discriminator", "and", "at", "the", "input", "level", "for", "the", "generator", ".", "Thus", ",", "we", "are", "going", "to", "use", "this", "configuration", "for", "the", "IcGAN", ".", "Both", "accuracy", "and", "F1", "-", "Score", "are", "similar", "as", "long", "as", "is", "not", "inserted", "in", "the", "last", "convolutional", "layers", ",", "in", "which", "case", "the", "performance", "considerably", "drops", ",", "especially", "in", "the", "generator", ".", "Then", ",", "these", "results", "reinforce", "our", "initial", "intuition", "of", "being", "added", "at", "an", "early", "stage", "of", "the", "model", "to", "allow", "learning", "interactions", "with", "it", ".", "subsection", ":", "Evaluating", "the", "encoder", "In", "this", "experiment", ",", "we", "prioritize", "the", "visual", "quality", "of", "reconstructed", "samples", "as", "an", "evaluation", "criterion", ".", "Among", "the", "different", "encoder", "configurations", "of", "section", "[", "reference", "]", ",", "IND", "and", "IND", "-", "COND", "yield", "a", "similar", "qualitative", "performance", ",", "being", "IND", "slightly", "superior", ".", "A", "comparison", "of", "these", "different", "configurations", "is", "shown", "in", "Figure", "[", "reference", "]", "a", "and", "in", "Figure", "[", "reference", "]", "b", "we", "focus", "on", "IND", "reconstructed", "samples", ".", "On", "another", "level", ",", "the", "fact", "that", "the", "generator", "is", "able", ",", "via", "an", "encoder", ",", "to", "reconstruct", "unseen", "images", "from", "the", "test", "set", "shows", "that", "the", "cGAN", "is", "generalizing", "and", "suggests", "that", "it", "does", "not", "suffer", "from", "overfitting", ",", "i.e.", ",", "it", "is", "not", "just", "memorizing", "and", "reproducing", "training", "samples", ".", "Additionally", ",", "we", "compare", "the", "different", "encoder", "configurations", "in", "a", "quantitative", "manner", "by", "using", "the", "minimal", "squared", "reconstruction", "loss", "as", "a", "criterion", ".", "Each", "encoder", "is", "trained", "minimizing", "with", "respect", "to", "latent", "representations", "(", ")", "or", "conditional", "information", "(", ")", ".", "Then", ",", "we", "quantitatively", "evaluate", "different", "model", "architectures", "using", "as", "a", "metric", "on", "a", "test", "set", "of", "150", "K", "CelebA", "generated", "images", ".", "We", "find", "that", "the", "encoder", "that", "yields", "the", "lowest", "is", "also", "IND", "(", "0.429", ")", ",", "followed", "closely", "by", "IND", "-", "CND", "(", "0.432", ")", ",", "and", "being", "SNG", "the", "worst", "case", "(", "0.500", ")", ".", "Furthermore", ",", "we", "can", "see", "an", "interesting", "property", "of", "minimizing", "a", "loss", "based", "on", "the", "latent", "space", "instead", "of", "a", "pixel", "-", "wise", "image", "reconstruction", ":", "reconstructed", "images", "tend", "to", "accurately", "keep", "high", "-", "level", "features", "of", "an", "input", "image", "(", "e.g.", "how", "a", "face", "generally", "looks", ")", "in", "detriment", "to", "more", "local", "details", "such", "as", "the", "exact", "position", "of", "the", "hair", ",", "eyes", "or", "face", ".", "Consequently", ",", "a", "latent", "space", "based", "encoder", "is", "invariant", "to", "these", "local", "details", ",", "making", "it", "an", "interesting", "approach", "for", "encoding", "purposes", ".", "For", "example", ",", "notice", "how", "the", "reconstructions", "in", "the", "last", "row", "of", "CelebA", "samples", "in", "Figure", "[", "reference", "]", "b", "fill", "the", "occluded", "part", "of", "the", "face", "by", "a", "hand", ".", "Another", "advantage", "with", "respect", "to", "element", "-", "wise", "encoders", "such", "as", "VAE", "is", "that", "GAN", "based", "reconstructions", "do", "not", "look", "blurry", ".", "[", "]", "[", "width=0.3255", "]", "images", "/", "6.exp", "/", "encoders_comparison", "[", "]", "[", "width=0.525", "]", "images", "/", "6.exp", "/", "encoder_reconstructions_all", "subsection", ":", "Evaluating", "the", "IcGAN", "In", "order", "to", "test", "that", "the", "model", "is", "able", "to", "correctly", "encode", "and", "re", "-", "generate", "a", "real", "image", "by", "preserving", "its", "main", "attributes", ",", "we", "take", "real", "samples", "from", "MNIST", "and", "CelebA", "test", "sets", "and", "reconstruct", "them", "with", "modifications", "on", "the", "conditional", "information", ".", "The", "result", "of", "this", "procedure", "is", "shown", "in", "Figure", "[", "reference", "]", ",", "where", "we", "show", "a", "subset", "of", "9", "of", "the", "18", "for", "CelebA", "attributes", "for", "image", "clarity", ".", "We", "can", "see", "that", ",", "in", "MNIST", ",", "we", "are", "able", "to", "get", "the", "hand", "-", "written", "style", "of", "real", "unseen", "digits", "and", "replicate", "these", "style", "on", "all", "the", "other", "digits", ".", "On", "the", "other", "hand", ",", "in", "CelebA", "we", "can", "see", "how", "reconstructed", "faces", "generally", "match", "the", "specified", "attribute", ".", "Additionally", ",", "we", "noticed", "that", "faces", "with", "uncommon", "conditions", "(", "e.g.", ",", "looking", "away", "from", "the", "camera", ",", "face", "not", "centered", ")", "were", "the", "most", "likely", "to", "be", "noisy", ".", "Furthermore", ",", "attributes", "such", "as", "mustache", "often", "fail", "to", "be", "generated", "especially", "on", "women", "samples", ",", "which", "might", "indicate", "that", "the", "generator", "is", "limited", "to", "some", "unusual", "attribute", "combinations", ".", "[", "]", "[", "width=0.95", "]", "images", "/", "6.exp", "/", "icgan_mnist_final_small", "[", "]", "[", "width=0.95", "]", "images", "/", "6.exp", "/", "icgan_celeba_final_smaller", "Manipulating", "the", "latent", "space", "The", "latent", "feature", "representation", "and", "conditional", "information", "learned", "by", "the", "generator", "can", "be", "further", "explored", "beyond", "encoding", "real", "images", "or", "randomly", "sampling", ".", "In", "order", "to", "do", "so", ",", "we", "linearly", "interpolate", "both", "and", "with", "pairs", "of", "reconstructed", "images", "from", "the", "CelebA", "test", "set", "(", "Figure", "[", "reference", "]", "a", ")", ".", "All", "the", "interpolated", "faces", "are", "plausible", "and", "the", "transition", "between", "faces", "is", "smooth", ",", "demonstrating", "that", "the", "IcGAN", "learned", "manifold", "is", "also", "consistent", "between", "interpolations", ".", "Then", ",", "this", "is", "also", "a", "good", "indicator", "that", "the", "model", "is", "generalizing", "the", "face", "representation", "properly", ",", "as", "it", "is", "not", "directly", "memorizing", "training", "samples", ".", "In", "addition", ",", "we", "perform", "in", "Figure", "[", "reference", "]", "b", "an", "attribute", "transfer", "between", "pairs", "of", "faces", ".", "We", "infer", "the", "latent", "representation", "and", "attribute", "information", "of", "two", "real", "faces", "from", "the", "test", "set", ",", "swap", "between", "those", "faces", "and", "re", "-", "generate", "them", ".", "As", "we", "previously", "noticed", ",", "the", "results", "suggest", "that", "encodes", "pose", ",", "illumination", "and", "background", "information", ",", "while", "tends", "to", "represent", "unique", "features", "of", "the", "face", ".", "[", "]", "[", "width=0.555", "]", "images", "/", "6.exp", "/", "icgan_interpolations", "[", "]", "[", "width=0.44", "]", "images", "/", "6.exp", "/", "icgan_attributeTransfer", "section", ":", "Conclusions", "We", "introduce", "an", "encoder", "in", "a", "conditional", "setting", "within", "the", "GAN", "framework", ",", "a", "model", "which", "we", "call", "Invertible", "Conditional", "GANs", "(", "IcGANs", ")", ".", "It", "solves", "the", "problem", "of", "GANs", "lacking", "the", "ability", "to", "infer", "real", "samples", "to", "a", "latent", "representation", ",", "while", "also", "allowing", "to", "explicitly", "control", "complex", "attributes", "of", "generated", "samples", "with", "conditional", "information", ".", "We", "also", "refine", "the", "performance", "of", "cGANS", "by", "testing", "the", "optimal", "position", "in", "which", "the", "conditional", "information", "is", "inserted", "in", "the", "model", ".", "We", "have", "found", "that", "for", "the", "generator", ",", "should", "be", "added", "at", "the", "input", "level", ",", "whereas", "the", "discriminator", "works", "best", "when", "is", "at", "the", "first", "layer", ".", "Additionally", ",", "we", "evaluate", "several", "ways", "to", "training", "an", "encoder", ".", "Training", "two", "independent", "encoders", "–", "one", "for", "encoding", "and", "another", "for", "encoding", "–", "has", "proven", "to", "be", "the", "best", "option", "in", "our", "experiments", ".", "The", "results", "obtained", "with", "a", "complex", "face", "dataset", ",", "CelebA", ",", "are", "satisfactory", "and", "promising", ".", "Acknowledgments", "This", "work", "is", "funded", "by", "the", "Projects", "TIN2013", "-", "41751", "-", "P", "of", "the", "Spanish", "Ministry", "of", "Science", "and", "the", "CHIST", "ERA", "project", "PCIN", "-", "2015", "-", "226", ".", "bibliography", ":", "References"], "entity": [[2, 5, ["Invertible", "Conditional", "GANs"], "Method"], [6, 8, ["image", "editing"], "Task"], [8, 11, ["Generative", "Adversarial", "Networks"], "Method"], [12, 13, ["GANs"], "Method"], [31, 33, ["conditional", "GANs"], "Method"], [34, 35, ["cGANs"], "Method"], [59, 60, ["encoders"], "Method"], [80, 82, ["conditional", "representation"], "Method"], [109, 110, ["cGANs"], "Method"], [115, 116, ["encoder"], "Method"], [118, 119, ["cGAN"], "Method"], [123, 125, ["Invertible", "cGAN"], "Method"], [126, 127, ["IcGAN"], "Method"], [144, 146, ["Image", "editing"], "Task"], [153, 154, ["complexity"], "Metric"], [164, 165, ["filter"], "Method"], [266, 269, ["image", "edition", "software"], "Method"], [286, 288, ["generative", "models"], "Method"], [289, 292, ["Natural", "image", "generation"], "Task"], [316, 319, ["deep", "learning", "techniques"], "Method"], [321, 323, ["generative", "modeling"], "Method"], [327, 330, ["Generative", "Adversarial", "Networks"], "Method"], [331, 332, ["GANs"], "Method"], [346, 348, ["image", "generation"], "Task"], [349, 350, ["GANs"], "Method"], [374, 376, ["Variational", "Autoencoders"], "Method"], [382, 385, ["image", "reconstruction", "loss"], "Task"], [388, 389, ["GANs"], "Method"], [399, 401, ["conditional", "extension"], "Method"], [402, 404, ["conditional", "GANs"], "Method"], [405, 406, ["cGANs"], "Method"], [415, 417, ["inference", "mechanism"], "Method"], [422, 424, ["latent", "representation"], "Method"], [456, 459, ["Invertible", "Conditional", "GANs"], "Method"], [460, 461, ["IcGANs"], "Method"], [463, 466, ["complex", "image", "editing"], "Task"], [471, 472, ["encoder"], "Method"], [476, 477, ["cGAN"], "Method"], [492, 493, ["encoder"], "Method"], [501, 502, ["cGAN"], "Method"], [536, 538, ["creative", "processes"], "Task"], [539, 541, ["data", "augmentation"], "Task"], [542, 544, ["face", "profiling"], "Task"], [565, 566, ["IcGANs"], "Method"], [574, 575, ["encoder"], "Method"], [577, 578, ["cGAN"], "Method"], [584, 585, ["MNIST"], "Material"], [587, 588, ["CelebA"], "Material"], [610, 611, ["encoder"], "Method"], [623, 625, ["latent", "representation"], "Method"], [634, 636, ["training", "procedures"], "Method"], [649, 650, ["cGANs"], "Method"], [651, 653, ["conditional", "position"], "Method"], [654, 656, ["conditional", "sampling"], "Method"], [673, 675, ["generative", "models"], "Method"], [706, 708, ["Variational", "Autoencoders"], "Method"], [709, 710, ["VAE"], "Method"], [745, 746, ["VAEs"], "Method"], [750, 755, ["pixel", "-", "wise", "reconstruction", "error"], "Metric"], [774, 777, ["Generative", "Adversarial", "Nets"], "Method"], [778, 779, ["GANs"], "Method"], [790, 791, ["GANs"], "Method"], [796, 798, ["deeper", "architecture"], "Method"], [799, 800, ["DCGAN"], "Method"], [821, 822, ["GANs"], "Method"], [825, 827, ["unsupervised", "approach"], "Method"], [829, 831, ["feature", "representations"], "Method"], [842, 843, ["cGANs"], "Method"], [881, 882, ["DCGANs"], "Method"], [890, 892, ["conditional", "extension"], "Method"], [907, 908, ["encoder"], "Method"], [909, 911, ["Invertible", "cGAN"], "Method"], [966, 967, ["encoder"], "Method"], [968, 969, ["GANs"], "Method"], [973, 980, ["non", "-", "conditional", "and", "jointly", "trained", "setting"], "Task"], [1004, 1005, ["VAE"], "Method"], [1007, 1008, ["GAN"], "Method"], [1048, 1049, ["encoder"], "Method"], [1051, 1054, ["conditional", "information", "encoding"], "Method"], [1059, 1061, ["training", "approaches"], "Method"], [1072, 1073, ["cGAN"], "Method"], [1078, 1081, ["Generative", "Adversarial", "Networks"], "Method"], [1082, 1083, ["GAN"], "Method"], [1087, 1089, ["neural", "networks"], "Method"], [1091, 1092, ["generator"], "Method"], [1094, 1095, ["discriminator"], "Method"], [1107, 1109, ["minimax", "game"], "Task"], [1123, 1124, ["discriminator"], "Method"], [1143, 1144, ["convergence"], "Task"], [1168, 1169, ["generator"], "Method"], [1170, 1171, ["discriminator"], "Method"], [1202, 1204, ["conditional", "GANs"], "Method"], [1205, 1206, ["cGANs"], "Method"], [1263, 1265, ["density", "model"], "Method"], [1291, 1292, ["cGAN"], "Method"], [1357, 1360, ["Invertible", "Conditional", "GANs"], "Method"], [1362, 1365, ["Invertible", "Conditional", "GANs"], "Method"], [1366, 1367, ["IcGANs"], "Method"], [1374, 1375, ["cGAN"], "Method"], [1377, 1378, ["encoder"], "Method"], [1381, 1382, ["encoders"], "Method"], [1408, 1410, ["encoding", "process"], "Method"], [1422, 1423, ["encoder"], "Method"], [1431, 1433, ["conditional", "setting"], "Task"], [1446, 1447, ["cGANs"], "Method"], [1454, 1456, ["conditional", "sampling"], "Method"], [1472, 1473, ["generator"], "Method"], [1489, 1491, ["latent", "representation"], "Method"], [1501, 1505, ["encoder", "/", "inference", "network"], "Method"], [1519, 1521, ["latent", "representation"], "Method"], [1560, 1561, ["cGAN"], "Method"], [1564, 1566, ["latent", "representation"], "Method"], [1606, 1607, ["cGAN"], "Method"], [1639, 1640, ["IcGAN"], "Method"], [1650, 1651, ["icgan_overview"], "Method"], [1657, 1658, ["encoder"], "Method"], [1660, 1661, ["cGAN"], "Method"], [1680, 1681, ["encoder"], "Method"], [1685, 1688, ["sub", "-", "encoders"], "Method"], [1709, 1710, ["generator"], "Method"], [1726, 1729, ["squared", "reconstruction", "loss"], "Metric"], [1829, 1830, ["SNG"], "Method"], [1833, 1834, ["encoder"], "Method"], [1850, 1851, ["encoder"], "Method"], [1855, 1857, ["independent", "encoders"], "Method"], [1863, 1866, ["IND", "-", "COND"], "Method"], [1868, 1869, ["encoders"], "Method"], [1900, 1901, ["encoder"], "Method"], [1917, 1918, ["encoder"], "Method"], [1921, 1922, ["discriminator"], "Method"], [1924, 1925, ["generator"], "Method"], [1951, 1952, ["encoder"], "Method"], [1954, 1956, ["conditional", "setting"], "Task"], [1984, 1985, ["cGANs"], "Method"], [1998, 2001, ["generator", "and", "discriminator"], "Method"], [2027, 2028, ["generator"], "Method"], [2033, 2034, ["cGAN"], "Method"], [2046, 2047, ["generator"], "Method"], [2049, 2050, ["discriminator"], "Method"], [2053, 2054, ["generator"], "Method"], [2133, 2135, ["Conditional", "sampling"], "Method"], [2158, 2159, ["discriminator"], "Method"], [2200, 2203, ["Kernel", "density", "estimation"], "Method"], [2207, 2210, ["Parzen", "window", "estimation"], "Method"], [2214, 2216, ["randomly", "sampling"], "Method"], [2218, 2219, ["kernel"], "Method"], [2221, 2223, ["Gaussian", "kernel"], "Method"], [2225, 2228, ["cross", "-", "validated"], "Method"], [2230, 2232, ["Direct", "interpolation"], "Method"], [2363, 2366, ["Kernel", "density", "estimation"], "Method"], [2367, 2369, ["direct", "interpolation"], "Method"], [2384, 2385, ["interpolation"], "Method"], [2453, 2456, ["kernel", "density", "estimation"], "Method"], [2499, 2501, ["Torch", "implementation"], "Method"], [2503, 2504, ["DCGAN"], "Method"], [2512, 2513, ["DCGAN"], "Method"], [2518, 2521, ["Adam", "optimizer", "adam2014"], "Method"], [2525, 2527, ["learning", "rate"], "Metric"], [2562, 2563, ["cGAN"], "Method"], [2565, 2570, ["matching", "-", "aware", "discriminator", "method"], "Method"], [2588, 2591, ["generator", "and", "discriminator"], "Method"], [2593, 2594, ["cGAN"], "Method"], [2640, 2642, ["IND", "encoders"], "Method"], [2660, 2662, ["Batch", "Normalization"], "Method"], [2721, 2723, ["Adam", "optimizer"], "Method"], [2738, 2740, ["cGAN", "model"], "Method"], [2759, 2760, ["MNIST"], "Material"], [2762, 2764, ["CelebFaces", "Attributes"], "Material"], [2765, 2766, ["CelebA"], "Material"], [2767, 2768, ["celeba"], "Material"], [2769, 2770, ["MNIST"], "Material"], [2810, 2811, ["CelebA"], "Material"], [2913, 2914, ["CelebA"], "Material"], [2923, 2924, ["MNIST"], "Material"], [2948, 2949, ["cGAN"], "Method"], [2951, 2954, ["attribute", "predictor", "network"], "Method"], [2955, 2956, ["Anet"], "Method"], [2974, 2975, ["cGAN"], "Method"], [3006, 3008, ["Inception", "model"], "Method"], [3123, 3125, ["mean", "accuracy"], "Metric"], [3126, 3129, ["F1", "-", "Score"], "Metric"], [3143, 3144, ["discriminator"], "Method"], [3155, 3156, ["cGANs"], "Method"], [3162, 3164, ["visual", "representations"], "Method"], [3171, 3172, ["accuracy"], "Metric"], [3177, 3178, ["accuracy"], "Metric"], [3185, 3187, ["convolutional", "layer"], "Method"], [3189, 3190, ["discriminator"], "Method"], [3197, 3198, ["generator"], "Method"], [3210, 3211, ["IcGAN"], "Method"], [3213, 3214, ["accuracy"], "Metric"], [3215, 3218, ["F1", "-", "Score"], "Metric"], [3243, 3244, ["generator"], "Task"], [3282, 3287, ["visual", "quality", "of", "reconstructed", "samples"], "Metric"], [3289, 3291, ["evaluation", "criterion"], "Metric"], [3303, 3304, ["IND"], "Method"], [3305, 3308, ["IND", "-", "COND"], "Method"], [3315, 3316, ["IND"], "Method"], [3343, 3346, ["IND", "reconstructed", "samples"], "Material"], [3355, 3356, ["generator"], "Method"], [3361, 3362, ["encoder"], "Method"], [3374, 3375, ["cGAN"], "Method"], [3414, 3418, ["minimal", "squared", "reconstruction", "loss"], "Metric"], [3423, 3424, ["encoder"], "Method"], [3446, 3448, ["model", "architectures"], "Method"], [3459, 3462, ["CelebA", "generated", "images"], "Material"], [3467, 3468, ["encoder"], "Method"], [3482, 3485, ["IND", "-", "CND"], "Method"], [3491, 3492, ["SNG"], "Method"], [3510, 3511, ["loss"], "Task"], [3519, 3524, ["pixel", "-", "wise", "image", "reconstruction"], "Method"], [3569, 3573, ["latent", "space", "based", "encoder"], "Method"], [3586, 3588, ["encoding", "purposes"], "Task"], [3625, 3629, ["element", "-", "wise", "encoders"], "Method"], [3631, 3632, ["VAE"], "Method"], [3666, 3667, ["IcGAN"], "Method"], [3697, 3698, ["MNIST"], "Material"], [3699, 3700, ["CelebA"], "Material"], [3748, 3749, ["MNIST"], "Material"], [3878, 3881, ["latent", "feature", "representation"], "Method"], [3897, 3899, ["randomly", "sampling"], "Task"], [3969, 3971, ["face", "representation"], "Method"], [3994, 3996, ["attribute", "transfer"], "Task"], [4081, 4082, ["encoder"], "Method"], [4084, 4086, ["conditional", "setting"], "Task"], [4096, 4099, ["Invertible", "Conditional", "GANs"], "Method"], [4100, 4101, ["IcGANs"], "Method"], [4108, 4109, ["GANs"], "Method"], [4118, 4120, ["latent", "representation"], "Method"], [4142, 4143, ["cGANS"], "Method"], [4165, 4166, ["generator"], "Method"], [4177, 4178, ["discriminator"], "Method"], [4196, 4197, ["encoder"], "Method"], [4201, 4202, ["encoders"], "Method"], [4205, 4206, ["encoding"], "Task"], [4209, 4210, ["encoding"], "Task"], [4231, 4232, ["CelebA"], "Material"], [66, 67, ["cGAN"], "Method"], [411, 412, ["GAN"], "Method"], [614, 615, ["GAN"], "Method"], [865, 866, ["GAN"], "Method"], [1176, 1177, ["GAN"], "Method"], [1196, 1197, ["GAN"], "Method"], [1220, 1221, ["GANs"], "Method"], [1286, 1287, ["cGAN"], "Method"], [1388, 1389, ["GAN"], "Method"], [1427, 1428, ["GAN"], "Method"], [1475, 1476, ["GAN"], "Method"], [1590, 1591, ["MNIST"], "Material"], [1903, 1904, ["GAN"], "Method"], [1976, 1977, ["GAN"], "Method"], [2489, 2490, ["GAN"], "Method"], [2930, 2931, ["GAN"], "Method"], [2958, 2959, ["CelebA"], "Material"], [3601, 3602, ["CelebA"], "Material"], [3634, 3635, ["GAN"], "Method"], [3736, 3737, ["CelebA"], "Material"], [3780, 3781, ["CelebA"], "Material"], [3918, 3919, ["CelebA"], "Material"], [3946, 3947, ["IcGAN"], "Method"], [4088, 4089, ["GAN"], "Method"]]} {"doc_id": "4265269bc894caa97efbfcfe5b83da7413f86a30", "text": ["Asymmetric", "Tri", "-", "training", "for", "Unsupervised", "Domain", "Adaptation", "section", ":", "Abstract", "Deep", "-", "layered", "models", "trained", "on", "a", "large", "number", "of", "labeled", "samples", "boost", "the", "accuracy", "of", "many", "tasks", ".", "It", "is", "important", "to", "apply", "such", "models", "to", "different", "domains", "because", "collecting", "many", "labeled", "samples", "in", "various", "domains", "is", "expensive", ".", "In", "unsupervised", "domain", "adaptation", ",", "one", "needs", "to", "train", "a", "classifier", "that", "works", "well", "on", "a", "target", "domain", "when", "provided", "with", "labeled", "source", "samples", "and", "unlabeled", "target", "samples", ".", "Although", "many", "methods", "aim", "to", "match", "the", "distributions", "of", "source", "and", "target", "samples", ",", "simply", "matching", "the", "distribution", "can", "not", "ensure", "accuracy", "on", "the", "target", "domain", ".", "To", "learn", "discriminative", "representations", "for", "the", "target", "domain", ",", "we", "assume", "that", "artificially", "labeling", "target", "samples", "can", "result", "in", "a", "good", "representation", ".", "Tri", "-", "training", "leverages", "three", "classifiers", "equally", "to", "give", "pseudo", "-", "labels", "to", "unlabeled", "samples", ",", "but", "the", "method", "does", "not", "assume", "labeling", "samples", "generated", "from", "a", "different", "domain", ".", "In", "this", "paper", ",", "we", "propose", "an", "asymmetric", "tri", "-", "training", "method", "for", "unsupervised", "domain", "adaptation", ",", "where", "we", "assign", "pseudo", "-", "labels", "to", "unlabeled", "samples", "and", "train", "neural", "networks", "as", "if", "they", "are", "true", "labels", ".", "In", "our", "work", ",", "we", "use", "three", "networks", "asymmetrically", ".", "By", "asymmetric", ",", "we", "mean", "that", "two", "networks", "are", "used", "to", "label", "unlabeled", "target", "samples", "and", "one", "network", "is", "trained", "by", "the", "samples", "to", "obtain", "targetdiscriminative", "representations", ".", "We", "evaluate", "our", "method", "on", "digit", "recognition", "and", "sentiment", "analysis", "datasets", ".", "Our", "proposed", "method", "achieves", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "the", "benchmark", "digit", "recognition", "datasets", "of", "domain", "adaptation", ".", "section", ":", "Introduction", "With", "the", "development", "of", "deep", "neural", "networks", "including", "deep", "convolutional", "neural", "networks", "(", "CNN", ")", "[", "reference", "]", "The", "University", "of", "Tokyo", ",", "Tokyo", ",", "Japan", ".", "Correspondence", "to", ":", "Kuniaki", "Saito", "<", "k", "-", "saito@mi.t.u", "-", "tokyo.ac.jp", ">", ",", "Yoshitaka", "Ushiku", "<", "ushiku@mi.t.u", "-", "tokyo.ac.jp", ">", ",", "Tatsuya", "Harada", "<", "harada@mi.t.u", "-", "tokyo.ac.jp>.", "[", "reference", "]", ",", "the", "recognition", "abilities", "of", "images", "and", "languages", "have", "improved", "dramatically", ".", "Training", "deep", "-", "layered", "networks", "with", "a", "large", "number", "of", "labeled", "samples", "enables", "us", "to", "correctly", "categorize", "samples", "in", "diverse", "domains", ".", "In", "addition", ",", "the", "transfer", "learning", "of", "CNN", "is", "utilized", "in", "many", "studies", ".", "For", "object", "detection", "or", "segmentation", ",", "we", "can", "transfer", "the", "knowledge", "of", "a", "CNN", "trained", "with", "a", "large", "-", "scale", "dataset", "by", "fine", "-", "tuning", "it", "on", "a", "relatively", "small", "dataset", "[", "reference", "][", "reference", "]", ".", "Moreover", ",", "features", "from", "a", "CNN", "trained", "on", "ImageNet", "[", "reference", "]", ")", "are", "useful", "for", "multimodal", "learning", "tasks", "including", "image", "captioning", "[", "reference", "]", "and", "visual", "question", "answering", "[", "reference", "]", ".", "One", "of", "the", "problems", "of", "neural", "networks", "is", "that", "although", "they", "perform", "well", "on", "the", "samples", "generated", "from", "the", "same", "distribution", "as", "the", "training", "samples", ",", "they", "may", "find", "it", "difficult", "to", "correctly", "recognize", "samples", "from", "different", "distributions", "at", "the", "test", "time", ".", "One", "example", "is", "images", "collected", "from", "the", "Internet", ",", "which", "may", "come", "in", "abundance", "and", "be", "fully", "labeled", ".", "They", "have", "a", "distribution", "different", "from", "the", "images", "taken", "from", "a", "camera", ".", "Thus", ",", "a", "classifier", "that", "performs", "well", "on", "various", "domains", "is", "important", "for", "practical", "use", ".", "To", "realize", "this", ",", "it", "is", "necessary", "to", "learn", "domain", "-", "invariantly", "discriminative", "representations", ".", "However", ",", "acquiring", "such", "representations", "is", "not", "easy", "because", "it", "is", "often", "difficult", "to", "collect", "a", "large", "number", "of", "labeled", "samples", "and", "because", "samples", "from", "different", "domains", "have", "domain", "-", "specific", "characteristics", ".", "In", "unsupervised", "domain", "adaptation", ",", "we", "try", "to", "train", "a", "classifier", "that", "works", "well", "on", "a", "target", "domain", "on", "the", "condition", "that", "we", "are", "provided", "labeled", "source", "samples", "and", "unlabeled", "target", "samples", "during", "training", ".", "Most", "of", "the", "previous", "deep", "domain", "adaptation", "methods", "have", "been", "proposed", "mainly", "under", "the", "assumption", "that", "the", "adaptation", "can", "be", "realized", "by", "matching", "the", "distribution", "of", "features", "from", "different", "domains", ".", "These", "methods", "aimed", "to", "obtain", "domain", "-", "invariant", "features", "by", "minimizing", "the", "divergence", "between", "domains", "as", "well", "as", "a", "category", "loss", "on", "the", "source", "domain", "[", "reference", "][", "reference", "]", ".", "However", ",", "as", "shown", "in", "[", "reference", "]", ",", "theoretically", ",", "if", "a", "classifier", "that", "works", "well", "on", "both", "the", "source", "and", "the", "target", "domains", "does", "not", "exist", ",", "we", "can", "not", "expect", "a", "discriminative", "classifier", "for", "the", "target", "domain", ".", "That", "is", ",", "even", "if", "the", "distributions", "are", "matched", "on", "the", "nondiscriminative", "representations", ",", "the", "classifier", "may", "not", "work", ":", "$", "'", "1.0;\"#", "-", "'\"'.", "/", "7#", "(", "8'7", "/", "$#34\"'$", "*", "Figure", "1", ".", "Outline", "of", "our", "model", ".", "We", "assign", "pseudo", "-", "labels", "to", "unlabeled", "target", "samples", "based", "on", "the", "predictions", "from", "two", "classifiers", "trained", "on", "source", "samples", ".", "well", "on", "the", "target", "domain", ".", "Since", "directly", "learning", "discriminative", "representations", "for", "the", "target", "domain", ",", "in", "the", "absence", "of", "target", "labels", ",", "is", "considered", "very", "difficult", ",", "we", "propose", "to", "assign", "pseudo", "-", "labels", "to", "target", "samples", "and", "train", "targetspecific", "networks", "as", "if", "they", "were", "true", "labels", ".", "Co", "-", "training", "and", "tri", "-", "training", "[", "reference", "]", "leverage", "multiple", "classifiers", "to", "artificially", "label", "unlabeled", "samples", "and", "retrain", "the", "classifiers", ".", "However", ",", "the", "methods", "do", "not", "assume", "labeling", "samples", "from", "different", "domains", ".", "Since", "our", "goal", "is", "to", "classify", "unlabeled", "target", "samples", "that", "have", "different", "characteristics", "from", "labeled", "source", "samples", ",", "we", "propose", "asymmetric", "tri", "-", "training", "for", "unsupervised", "domain", "adaptation", ".", "By", "asymmetric", ",", "we", "mean", "that", "we", "assign", "different", "roles", "to", "three", "classifiers", ".", "In", "this", "paper", ",", "we", "propose", "a", "novel", "tri", "-", "training", "method", "for", "unsupervised", "domain", "adaptation", ",", "where", "we", "assign", "pseudolabels", "to", "unlabeled", "samples", "and", "train", "neural", "networks", "utilizing", "the", "samples", ".", "As", "described", "in", "Fig", ".", "1", ",", "two", "networks", "are", "used", "to", "label", "unlabeled", "target", "samples", "and", "the", "remaining", "network", "is", "trained", "by", "the", "pseudo", "-", "labeled", "target", "samples", ".", "Our", "method", "does", "not", "need", "any", "special", "implementations", ".", "We", "evaluate", "our", "method", "on", "the", "digit", "classification", "task", ",", "traffic", "sign", "classification", "task", "and", "sentiment", "analysis", "task", "using", "the", "Amazon", "Review", "dataset", ",", "and", "demonstrate", "state", "-", "of", "-", "the", "-", "art", "performance", "in", "nearly", "all", "experiments", ".", "In", "particular", ",", "in", "the", "adaptation", "scenario", ",", "MNIST→SVHN", ",", "our", "method", "outperformed", "other", "methods", "by", "more", "than", "10", "%", ".", "section", ":", "Related", "Work", "As", "many", "methods", "have", "been", "proposed", "to", "tackle", "various", "tasks", "in", "domain", "adaptation", ",", "we", "present", "details", "of", "the", "research", "most", "closely", "related", "to", "our", "paper", ".", "A", "number", "of", "previous", "methods", "attempted", "to", "realize", "adaptation", "by", "utilizing", "the", "measurement", "of", "divergence", "between", "different", "domains", "[", "reference", "][", "reference", "][", "reference", "]", ".", "The", "methods", "are", "based", "on", "the", "theory", "proposed", "in", "[", "reference", "]", ",", "which", "states", "that", "the", "expected", "loss", "for", "a", "target", "domain", "is", "bounded", "by", "three", "terms", ":", "(", "i", ")", "expected", "loss", "for", "the", "source", "domain", ";", "(", "ii", ")", "domain", "divergence", "between", "source", "and", "target", ";", "and", "(", "iii", ")", "the", "minimum", "value", "of", "a", "shared", "expected", "loss", ".", "The", "shared", "expected", "loss", "means", "the", "sum", "of", "the", "loss", "on", "the", "source", "and", "target", "domain", ".", "As", "the", "third", "term", ",", "which", "is", "usually", "considered", "to", "be", "very", "low", ",", "can", "not", "be", "evaluated", "when", "labeled", "target", "samples", "are", "absent", ",", "most", "methods", "try", "to", "minimize", "the", "first", "term", "and", "the", "second", "term", ".", "With", "regards", "to", "training", "deep", "architectures", ",", "the", "maximum", "mean", "discrepancy", "(", "MMD", ")", "or", "a", "loss", "of", "domain", "classifier", "network", "is", "utilized", "to", "measure", "the", "divergence", "corresponding", "to", "the", "second", "term", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "However", ",", "the", "third", "term", "is", "very", "important", "in", "training", "CNN", ",", "which", "simultaneously", "extract", "representations", "and", "recognize", "them", ".", "The", "third", "term", "can", "easily", "be", "large", "when", "the", "representations", "are", "not", "discriminative", "for", "the", "target", "domain", ".", "Therefore", ",", "we", "focus", "on", "how", "to", "learn", "target", "-", "discriminative", "representations", "considering", "the", "third", "term", ".", "In", "[", "reference", "]", ")", "the", "focus", "was", "on", "the", "point", "we", "have", "stated", "and", "a", "target", "-", "specific", "classifier", "was", "constructed", "using", "a", "residual", "network", "structure", ".", "Different", "from", "their", "method", ",", "we", "constructed", "a", "target", "-", "specific", "network", "by", "providing", "artificially", "labeled", "target", "samples", ".", "Several", "transductive", "methods", "use", "similarity", "of", "features", "to", "provide", "labels", "for", "unlabeled", "samples", "[", "reference", "][", "reference", "]", ".", "For", "unsupervised", "domain", "adaptation", ",", "in", "[", "reference", "]", ",", "a", "method", "was", "proposed", "to", "learn", "labeling", "metrics", "by", "using", "the", "k", "-", "nearest", "neighbors", "between", "unlabeled", "target", "samples", "and", "labeled", "source", "samples", ".", "In", "contrast", "to", "this", "method", ",", "our", "method", "explicitly", "and", "simply", "backpropagates", "the", "category", "loss", "for", "target", "samples", "based", "on", "pseudo", "-", "labeled", "samples", ".", "Our", "approach", "does", "not", "require", "any", "special", "modules", ".", "Many", "methods", "proposed", "to", "give", "pseudo", "-", "labels", "to", "unlabeled", "samples", "by", "utilizing", "the", "predictions", "of", "a", "classifier", "and", "retraining", "it", "including", "the", "pseudo", "-", "labeled", "samples", ",", "which", "is", "called", "self", "-", "training", ".", "The", "underlying", "assumption", "of", "selftraining", "is", "that", "one", "'s", "own", "high", "-", "confidence", "predictions", "are", "correct", "[", "reference", "]", ".", "As", "the", "predictions", "are", "mostly", "correct", ",", "utilizing", "samples", "with", "high", "confidence", "will", "further", "improve", "the", "performance", "of", "the", "classifier", ".", "Co", "-", "training", "utilizes", "two", "classifiers", ",", "which", "have", "different", "views", "on", "one", "sample", ",", "to", "provide", "pseudo", "-", "labels", "[", "reference", "][", "reference", "]", ".", "Then", ",", "the", "unlabeled", "samples", "are", "added", "to", "training", "set", "if", "at", "least", "one", "classifier", "is", "confident", "about", "the", "predictions", ".", "The", "generalization", "ability", "of", "co", "-", "training", "is", "theoretically", "ensured", "[", "reference", "][", "reference", "]", ")", "under", "some", "assumptions", "and", "applied", "to", "various", "tasks", "[", "reference", "][", "reference", "]", ".", "In", "[", "reference", "]", ",", "the", "idea", "of", "co", "-", "training", "was", "incorporated", "into", "domain", "adaptation", ".", "Tri", "-", "training", "can", "be", "regarded", "as", "the", "extension", "of", "co", "-", "training", "[", "reference", "]", ".", "Similar", "to", "co", "-", "training", ",", "tritraining", "uses", "the", "output", "of", "three", "different", "classifiers", "to", "give", "pseudo", "-", "labels", "to", "unlabeled", "samples", ".", "Tri", "-", "training", "does", "not", "require", "partitioning", "features", "into", "different", "views", ";", "instead", ",", "tri", "-", "training", "initializes", "each", "classifier", "differently", ".", "However", ",", "Figure", "2", ".", "The", "proposed", "method", "includes", "a", "shared", "feature", "extractor", "(", "F", ")", ",", "classifiers", "for", "labeled", "samples", "(", "F1", ",", "F2", ")", ",", "which", "learn", "from", "labeled", "source", "samples", ",", "and", "newly", "labeled", "target", "samples", ".", "In", "addition", ",", "a", "target", "-", "specific", "classifier", "(", "Ft", ")", "learns", "from", "pseudo", "-", "labeled", "target", "samples", ".", "Our", "method", "first", "trains", "networks", "from", "only", "labeled", "source", "samples", ",", "then", "labels", "the", "target", "samples", "based", "on", "the", "output", "of", "F1", ",", "F2", ".", "We", "train", "all", "architectures", "using", "them", "as", "if", "they", "are", "correctly", "labeled", "samples", ".", "tri", "-", "training", "does", "not", "assume", "that", "the", "unlabeled", "samples", "follow", "the", "different", "distributions", "from", "the", "ones", "which", "labeled", "samples", "are", "generated", "from", ".", "Therefore", ",", "we", "develop", "a", "tritraining", "method", "suitable", "for", "domain", "adaptation", "by", "using", "three", "classifiers", "asymmetrically", ".", "In", "[", "reference", "]", ",", "the", "effect", "of", "pseudo", "-", "labels", "in", "a", "neural", "network", "was", "investigated", ".", "They", "argued", "that", "the", "effect", "of", "training", "a", "classifier", "with", "pseudo", "-", "labels", "is", "equivalent", "to", "entropy", "regularization", ",", "thus", "leading", "to", "a", "low", "-", "density", "separation", "between", "classes", ".", "In", "addition", ",", "in", "our", "experiment", ",", "we", "observe", "that", "target", "samples", "are", "separated", "in", "hidden", "features", ".", "section", ":", "Method", "In", "this", "section", ",", "we", "provide", "details", "of", "the", "proposed", "model", "for", "domain", "adaptation", ".", "We", "aim", "to", "construct", "a", "targetspecific", "network", "by", "utilizing", "pseudo", "-", "labeled", "target", "samples", ".", "Simultaneously", ",", "we", "expect", "two", "labeling", "networks", "to", "acquire", "target", "-", "discriminative", "representations", "and", "gradually", "increase", "accuracy", "on", "the", "target", "domain", ".", "We", "show", "our", "proposed", "network", "structure", "in", "Fig", ".", "2", ".", "Here", "F", "denotes", "the", "network", "which", "outputs", "shared", "features", "among", "three", "networks", ",", "F", "1", "and", "F", "2", "classify", "features", "generated", "from", "F", ".", "Their", "predictions", "are", "utilized", "to", "give", "pseudo", "-", "labels", ".", "The", "classifier", "F", "t", "classifies", "features", "generated", "from", "F", ",", "which", "is", "a", "target", "-", "specific", "network", ".", "Here", "F", "1", ",", "F", "2", "learn", "from", "source", "and", "pseudo", "-", "labeled", "target", "samples", "and", "F", "t", "learns", "only", "from", "pseudo", "-", "labeled", "target", "samples", ".", "The", "shared", "network", "F", "learns", "from", "all", "gradients", "from", "F", "1", ",", "F", "2", ",", "F", "t", ".", "Without", "such", "a", "shared", "network", ",", "another", "option", "for", "the", "network", "architecture", "we", "can", "think", "of", "is", "training", "three", "networks", "separately", ",", "but", "this", "is", "inefficient", "in", "terms", "of", "training", "and", "implementation", ".", "Furthermore", ",", "by", "building", "a", "shared", "network", "F", ",", "F", "1", "and", "F", "2", "can", "also", "harness", "the", "target", "-", "discriminative", "representations", "learned", "by", "the", "feedback", "from", "F", "t", ".", "The", "set", "of", "source", "samples", "is", "defined", "as", "(", "x", "i", ",", "y", "i", ")", "∼", "T", ",", "and", "the", "pseudolabeled", "target", "set", "is", "section", ":", "Loss", "for", "Multiview", "Features", "Network", "In", "the", "existing", "works", "[", "reference", "]", "on", "co", "-", "training", "for", "domain", "adaptation", ",", "given", "features", "are", "divided", "into", "separate", "parts", "and", "considered", "to", "be", "different", "views", ".", "As", "we", "aim", "to", "label", "target", "samples", "with", "high", "accuracy", ",", "we", "expect", "F", "1", ",", "F", "2", "to", "classify", "samples", "based", "on", "different", "viewpoints", ".", "Therefore", ",", "we", "make", "a", "constraint", "for", "the", "weight", "of", "F", "1", ",", "F", "2", "to", "make", "their", "inputs", "different", "to", "each", "other", ".", "We", "add", "the", "term", "|W", "1", "T", "W", "2", "|", "to", "the", "cost", "function", ",", "where", "W", "1", ",", "W", "2", "denote", "fully", "connected", "layers", "'", "weights", "of", "F", "1", "and", "F", "2", "which", "are", "first", "applied", "to", "the", "feature", "F", "(", "x", "i", ")", ".", "Each", "network", "will", "learn", "from", "different", "features", "with", "this", "constraint", ".", "The", "objective", "for", "learning", "F", "1", ",", "F", "2", "is", "defined", "as", "where", "L", "y", "denotes", "the", "standard", "softmax", "cross", "-", "entropy", "loss", "function", ".", "We", "decided", "the", "trade", "-", "off", "parameter", "λ", "based", "on", "validation", "split", ".", "section", ":", "Learning", "Procedure", "and", "Labeling", "Method", "Pseudo", "-", "labeled", "target", "samples", "will", "provide", "targetdiscriminative", "information", "to", "the", "network", ".", "However", ",", "since", "they", "certainly", "contain", "false", "labels", ",", "we", "have", "to", "pick", "up", "reliable", "pseudo", "-", "labels", ".", "Our", "labeling", "and", "learning", "method", "is", "aimed", "at", "realizing", "this", ".", "The", "entire", "procedure", "of", "training", "the", "network", "is", "shown", "in", "Algorithm", "1", ".", "First", ",", "we", "train", "the", "entire", "network", "with", "source", "training", "set", "S.", "Here", "F", "1", ",", "F", "2", "are", "optimized", "by", "Eq", ".", "(", "1", ")", "and", "F", "t", "is", "trained", "on", "standard", "category", "loss", ".", "After", "training", "on", "S", ",", "to", "provide", "pseudo", "-", "labels", ",", "we", "use", "predictions", "of", "F", "1", "and", "F", "2", ",", "namely", "y", "1", ",", "y", "2", "obtained", "from", "x", "k", ".", "When", "C", "1", ",", "C", "2", "denote", "the", "class", "which", "has", "the", "maximum", "predicted", "probability", "for", "y", "1", ",", "y", "2", ",", "we", "assign", "a", "pseudo", "-", "label", "to", "x", "k", "if", "the", "following", "two", "conditions", "are", "satisfied", ".", "First", ",", "we", "require", "C", "1", "=", "C", "2", "to", "give", "pseudo", "-", "labels", ",", "which", "means", "two", "different", "classifiers", "agree", "with", "the", "prediction", ".", "The", "second", "requirement", "is", "that", "the", "maximizing", "probability", "of", "y", "1", "or", "y", "2", "exceeds", "the", "threshold", "parameter", ",", "which", "we", "set", "as", "0.9", "or", "0.95", "in", "the", "experiment", ".", "We", "suppose", "that", "unless", "one", "of", "two", "classifiers", "is", "confident", "of", "the", "prediction", ",", "the", "prediction", "is", "not", "reliable", ".", "If", "the", "two", "requirements", "are", "satisfied", ",", "x", "k", ",", "ŷ", "k", "=", "C", "1", "=", "C", "2", "is", "added", "to", "T", "l", ".", "To", "prevent", "the", "overfitting", "to", "pseudo", "-", "labels", ",", "we", "resample", "the", "candidate", "for", "labeling", "samples", "in", "each", "step", ".", "We", "set", "the", "Algorithm", "1", "iter", "denotes", "the", "iteration", "of", "training", ".", "The", "function", "Labeling", "means", "the", "method", "of", "labeling", ".", "We", "assign", "pseudo", "-", "labels", "to", "samples", "when", "the", "predictions", "of", "F", "1", "and", "F", "2", "agree", "and", "at", "least", "one", "of", "them", "is", "confident", "of", "their", "predictions", ".", "number", "of", "the", "initial", "candidates", "N", "init", "as", "5", ",", "000", ".", "We", "gradually", "increase", "the", "number", "of", "the", "candidates", "N", "t", "=", "k", "/", "20", "*", "n", ",", "where", "n", "denotes", "the", "number", "of", "all", "target", "samples", "and", "k", "denotes", "the", "number", "of", "steps", ",", "and", "we", "set", "the", "maximum", "number", "of", "pseudo", "-", "labeled", "candidates", "as", "40", ",", "000", ".", "After", "the", "pseudo", "-", "labeled", "training", "set", "T", "l", "is", "composed", ",", "F", ",", "F", "1", ",", "F", "2", "are", "updated", "by", "the", "objective", "Eq", ".", "(", "1", ")", "on", "the", "labeled", "training", "set", "L", "=", "S", "∪", "T", "l", ".", "Then", ",", "F", ",", "F", "t", "are", "simply", "optimized", "by", "the", "category", "loss", "for", "T", "l", ".", "Discriminative", "representations", "will", "be", "learned", "by", "constructing", "a", "target", "-", "specific", "network", "trained", "only", "on", "target", "samples", ".", "However", ",", "if", "only", "noisy", "pseudo", "-", "labeled", "samples", "are", "used", "for", "training", ",", "the", "network", "may", "not", "learn", "useful", "representations", ".", "Then", ",", "we", "use", "both", "source", "samples", "and", "pseudo", "-", "labeled", "samples", "for", "training", "F", ",", "F", "1", ",", "F", "2", "to", "ensure", "the", "accuracy", ".", "Also", ",", "as", "the", "learning", "proceeds", ",", "F", "will", "learn", "target", "-", "discriminative", "representations", ",", "resulting", "in", "an", "improvement", "in", "accuracy", "in", "F", "1", ",", "F", "2", ".", "This", "cycle", "will", "gradually", "enhance", "the", "accuracy", "in", "the", "target", "domain", ".", "section", ":", "Batch", "Normalization", "for", "Domain", "Adaptation", "Batch", "normalization", "(", "BN", ")", "[", "reference", "]", ",", "which", "whitens", "the", "output", "of", "the", "hidden", "layer", "in", "a", "CNN", ",", "is", "an", "effective", "technique", "to", "accelerate", "training", "speed", "and", "enhance", "the", "accuracy", "of", "the", "model", ".", "In", "addition", ",", "in", "domain", "adaptation", ",", "whitening", "the", "hidden", "layer", "'s", "output", "is", "effective", "for", "improving", "the", "performance", ",", "which", "make", "the", "distribution", "in", "different", "domains", "similar", "[", "reference", "][", "reference", "]", ".", "Input", "samples", "of", "F", "1", ",", "F", "2", "include", "both", "pseudo", "-", "labeled", "target", "samples", "and", "source", "samples", ".", "Introducing", "BN", "will", "be", "useful", "for", "matching", "the", "distribution", "and", "improves", "the", "performance", ".", "We", "add", "the", "BN", "layer", "in", "the", "last", "layer", "in", "F", ".", "section", ":", "Analysis", "In", "this", "section", ",", "we", "provide", "a", "theoretical", "analysis", "to", "our", "approach", ".", "First", ",", "we", "provide", "an", "insight", "into", "existing", "theory", ",", "then", "we", "introduce", "a", "simple", "expansion", "of", "the", "theory", "related", "to", "our", "method", ".", "In", "[", "reference", "]", ",", "an", "equation", "was", "introduced", "showing", "that", "the", "upper", "bound", "of", "the", "expected", "error", "in", "the", "target", "domain", "depends", "on", "three", "terms", ",", "which", "include", "the", "divergence", "between", "different", "domains", "and", "the", "error", "of", "an", "ideal", "joint", "hypothesis", ".", "The", "divergence", "between", "source", "and", "target", "domain", ",", "H∆H", "-", "distance", ",", "is", "defined", "as", "follows", ":", "This", "distance", "is", "frequently", "used", "to", "measure", "the", "adaptability", "between", "different", "domains", ".", "section", ":", "The", "ideal", "joint", "hypothesis", "is", "defined", "as", ",", "and", "its", "corresponding", "error", ",", "where", "R", "denotes", "the", "expected", "error", "on", "each", "hypothesis", ".", "The", "theorem", "is", "as", "follows", ".", "This", "theorem", "means", "that", "the", "expected", "error", "on", "the", "target", "domain", "is", "upper", "bounded", "by", "three", "terms", ",", "the", "expected", "error", "on", "the", "source", "domain", ",", "the", "domain", "divergence", "measured", "by", "the", "disagreement", "of", "the", "hypothesis", ",", "and", "the", "error", "of", "the", "ideal", "joint", "hypothesis", ".", "In", "the", "existing", "work", "[", "reference", "][", "reference", "]", ",", "C", "was", "disregarded", "because", "it", "was", "considered", "to", "be", "negligibly", "small", ".", "If", "we", "are", "provided", "with", "fixed", "features", ",", "we", "do", "not", "need", "to", "consider", "the", "term", "because", "the", "term", "is", "also", "fixed", ".", "However", ",", "if", "we", "assume", "that", "x", "s", "∼", "S", ",", "x", "t", "∼", "T", "are", "obtained", "from", "the", "last", "fully", "connected", "layer", "of", "deep", "models", ",", "we", "note", "that", "C", "is", "determined", "by", "the", "output", "of", "the", "layer", ",", "and", "further", "note", "the", "necessity", "of", "considering", "this", "term", ".", "We", "consider", "the", "pseudo", "-", "labeled", "target", "samples", "set", "given", "false", "labels", "at", "the", "ratio", "of", "ρ", ".", "The", "shared", "error", "of", "h", "*", "on", "S", ",", "T", "l", "is", "denoted", "as", "C", "′", ".", "Then", ",", "the", "following", "inequality", "holds", ":", "We", "show", "a", "simple", "derivation", "of", "the", "inequality", "in", "the", "Supplementary", "material", ".", "In", "Theorem", "1", ",", "we", "can", "not", "measure", "C", "in", "the", "absence", "of", "labeled", "target", "samples", ".", "We", "can", "approximately", "evaluate", "and", "minimize", "it", "by", "using", "pseudo", "-", "labels", ".", "Furthermore", ",", "when", "we", "consider", "the", "second", "term", "on", "the", "right", "-", "hand", "side", ",", "our", "method", "is", "expected", "to", "reduce", "this", "term", ".", "This", "term", "intuitively", "denotes", "the", "discrepancy", "between", "different", "domains", "in", "the", "disagreement", "of", "two", "classifiers", ".", "If", "we", "regard", "certain", "h", "and", "h", "′", "as", "F", "1", "and", "F", "2", ",", "respectively", ",", "]", "is", "expected", "to", "be", "low", ",", "although", "we", "use", "the", "training", "set", "T", "l", "instead", "of", "genuine", "labeled", "target", "samples", ".", "Thus", ",", "our", "method", "will", "consider", "both", "the", "second", "and", "the", "third", "term", "in", "Theorem", "1", ".", "section", ":", "Experiment", "and", "Evaluation", "We", "perform", "extensive", "evaluations", "of", "our", "method", "on", "image", "datasets", "and", "a", "sentiment", "analysis", "dataset", ".", "We", "evaluate", "the", "accuracy", "of", "target", "-", "specific", "networks", "in", "all", "experiments", ".", "Visual", "Domain", "Adaptation", "For", "visual", "domain", "adaptation", ",", "we", "perform", "our", "evaluation", "on", "the", "digits", "datasets", "and", "traffic", "signs", "datasets", ".", "Digits", "datasets", "include", "MNIST", "[", "reference", "]", ",", "[", "reference", "]", ",", "Street", "View", "House", "Numbers", "(", "SVHN", ")", "[", "reference", "]", ",", "and", "Synthetic", "Digits", "(", "SYN", "DIGITS", ")", "[", "reference", "]", ".", "We", "further", "evaluate", "our", "method", "on", "traffic", "sign", "datasets", "including", "Synthetic", "Traffic", "Signs", "(", "SYN", "SIGNS", ")", "[", "reference", "]", "and", "German", "Traffic", "Signs", "Recognition", "Benchmark", "[", "reference", "]", ".", "In", "total", ",", "five", "adaptation", "scenarios", "are", "evaluated", "in", "this", "experiment", ".", "As", "the", "datasets", "used", "for", "evaluation", "are", "varied", "in", "previous", "works", ",", "we", "extensively", "evaluate", "our", "method", "on", "the", "five", "scenarios", ".", "We", "do", "not", "evaluate", "our", "method", "on", "Office", "[", "reference", "]", ",", "which", "is", "the", "most", "commonly", "used", "dataset", "for", "visual", "domain", "adaptation", ".", "As", "pointed", "out", "by", "[", "reference", "]", ",", "some", "labels", "in", "that", "dataset", "are", "noisy", "and", "some", "images", "contain", "other", "classes", "'", "objects", ".", "Furthermore", ",", "many", "previous", "studies", "have", "evaluated", "the", "fine", "-", "tuning", "of", "pretrained", "networks", "using", "ImageNet", ".", "This", "protocol", "assumes", "the", "existence", "of", "another", "source", "domain", ".", "In", "our", "work", ",", "we", "want", "to", "evaluate", "the", "situation", "where", "we", "have", "access", "to", "only", "one", "source", "domain", "and", "one", "target", "domain", ".", "section", ":", "Adaptation", "in", "Amazon", "Reviews", "To", "investigate", "the", "behavior", "on", "language", "datasets", ",", "we", "also", "evaluated", "our", "method", "on", "the", "Amazon", "Reviews", "dataset", "[", "reference", "]", "with", "the", "same", "preprocessing", "as", "used", "by", "[", "reference", "][", "reference", "]", ".", "The", "dataset", "contains", "reviews", "on", "four", "types", "of", "products", ":", "books", ",", "DVDs", ",", "electronics", ",", "and", "kitchen", "appliances", ".", "We", "evaluate", "our", "method", "on", "12", "domain", "adaptation", "scenarios", ".", "The", "results", "are", "shown", "in", "Table", "1", ".", "Baseline", "Methods", "We", "compare", "our", "method", "with", "five", "methods", "for", "unsupervised", "domain", "adaptation", "including", "state", "-", "of", "-", "the", "art", "methods", "in", "visual", "domain", "adaptation", ";", "Maximum", "Mean", "Discrepancy", "(", "MMD", ")", "[", "reference", "]", ",", "Domain", "Adversarial", "Neural", "Network", "(", "DANN", ")", "[", "reference", "]", ",", "Deep", "Reconstruction", "Classification", "Network", "(", "DRCN", ")", "[", "reference", "]", ",", "Domain", "Separation", "Networks", "(", "DSN", ")", "[", "reference", "]", ",", "and", "k", "-", "Nearest", "Neighbor", "based", "adaptation", "(", "kNN", "-", "Ad", ")", "[", "reference", "]", ".", "We", "cite", "the", "results", "of", "MMD", "from", "[", "reference", "]", ".", "In", "addition", ",", "we", "compare", "our", "method", "with", "CNN", "trained", "only", "on", "source", "samples", ".", "We", "compare", "our", "method", "with", "Variational", "Fair", "AutoEncoder", "(", "VFAE", ")", "[", "reference", "]", "and", "DANN", "[", "reference", "]", "in", "the", "Amazon", "Reviews", "experiment", ".", "section", ":", "Implementation", "Detail", "In", "experiments", "on", "image", "datasets", ",", "we", "employ", "the", "architecture", "of", "CNN", "used", "in", "[", "reference", "]", ".", "For", "a", "fair", "comparison", ",", "we", "separate", "the", "network", "at", "the", "hidden", "layer", "from", "which", "[", "reference", "]", "constructed", "discriminator", "networks", ".", "Therefore", ",", "when", "considering", "one", "classifier", ",", "for", "example", ",", "F", "1", "•", "F", ",", "the", "architecture", "is", "identical", "to", "previous", "work", ".", "We", "also", "follow", "[", "reference", "]", "in", "the", "other", "protocols", ".", "We", "set", "the", "threshold", "value", "for", "the", "labeling", "method", "as", "0.95", "in", "MNIST→SVHN", ".", "In", "other", "scenarios", ",", "we", "set", "it", "as", "0.9", ".", "We", "use", "MomentumSGD", "for", "optimization", "and", "set", "the", "momentum", "as", "0.9", ",", "while", "the", "learning", "rate", "is", "determined", "on", "validation", "splits", "and", "uses", "either", "[", "0.01", ",", "0.05", "]", ".", "λ", "is", "set", "0.01", "in", "all", "scenarios", ".", "In", "our", "Supplementary", "material", ",", "we", "provide", "details", "of", "the", "network", "architecture", "and", "hyper", "-", "parameters", ".", "For", "experiments", "on", "the", "Amazon", "Reviews", "dataset", ",", "we", "use", "a", "similar", "architecture", "to", "that", "used", "in", "[", "reference", "]", ":", "with", "sigmoid", "activated", ",", "one", "dense", "hidden", "layer", "with", "50", "hidden", "units", ",", "and", "softmax", "output", ".", "We", "extend", "the", "architecture", "to", "our", "method", "similarly", "in", "the", "architecture", "of", "CNN", ".", "λ", "is", "set", "as", "0.001", "based", "on", "the", "validation", ".", "Since", "the", "input", "is", "sparse", ",", "we", "use", "Adagrad", "[", "reference", "]", "for", "optimization", ".", "We", "repeat", "this", "evaluation", "10", "times", "and", "report", "mean", "accuracy", ".", "section", ":", "Experimental", "Result", "In", "Tables", "1", "and", "3", ",", "we", "show", "the", "main", "results", "of", "the", "experiments", ".", "When", "training", "only", "on", "source", "samples", ",", "the", "effect", "of", "the", "BN", "is", "not", "clear", "as", "in", "Tables", "1", ".", "However", ",", "in", "all", "image", "recognition", "experiments", ",", "the", "effect", "of", "BN", "in", "our", "method", "is", "clear", ";", "at", "the", "same", "time", ",", "the", "effect", "of", "our", "method", "is", "also", "clear", "when", "we", "do", "not", "use", "BN", "in", "the", "network", "architecture", ".", "The", "effect", "of", "the", "weight", "constraint", "is", "obvious", "in", "MNIST→SVHN", ".", "[", "reference", "]", "and", "[", "reference", "]", "in", "parentheses", ".", "MNIST→MNIST", "-", "M", ":", "last", "pooling", "layer", "MNIST→MNIST", "-", "M", "First", ",", "we", "evaluate", "the", "adaptation", "scenario", "between", "the", "hand", "-", "written", "digits", "dataset", "MNIST", "and", "its", "transformed", "dataset", "MNIST", "-", "M.", "MNIST", "-", "M", "is", "composed", "by", "merging", "the", "clip", "of", "the", "background", "from", "BSDS500", "datasets", "[", "reference", "]", ".", "A", "patch", "is", "randomly", "taken", "from", "the", "images", "in", "BSDS500", ",", "merged", "to", "MNIST", "digits", ".", "Even", "with", "this", "simple", "domain", "shift", ",", "the", "adaptation", "performance", "of", "CNN", "is", "much", "worse", "than", "the", "case", "where", "it", "was", "trained", "on", "target", "samples", ".", "From", "59", ",", "001", "target", "training", "samples", ",", "we", "randomly", "select", "1", ",", "000", "labeled", "target", "samples", "as", "a", "validation", "split", "and", "tuned", "hyper", "-", "parameters", ".", "Our", "method", "outperforms", "the", "other", "existing", "method", "by", "about", "7", "%", ".", "Visualization", "of", "features", "in", "the", "last", "pooling", "layer", "is", "shown", "in", "Fig", ".", "3", "(", "a", ")(", "b", ")", ".", "We", "can", "observe", "that", "the", "red", "target", "samples", "are", "more", "dispersed", "when", "adaptation", "is", "achieved", ".", "We", "show", "the", "comparison", "of", "the", "accuracy", "between", "the", "actual", "labeling", "accuracy", "on", "target", "samples", "during", "training", "and", "the", "test", "accuracy", "in", "Fig", ".", "4", ".", "The", "test", "accuracy", "is", "very", "low", "at", "first", ",", "but", "as", "the", "steps", "increase", ",", "the", "accuracy", "becomes", "closer", "to", "that", "of", "the", "labeling", "accuracy", ".", "In", "this", "adaptation", ",", "we", "can", "clearly", "see", "that", "the", "actual", "labeling", "accuracy", "gradually", "improves", "with", "the", "accuracy", "of", "the", "network", ".", "SVHN↔MNIST", "We", "increase", "the", "gap", "between", "distributions", "in", "this", "experiment", ".", "We", "evaluate", "adaptation", "between", "SVHN", "[", "reference", "]", "and", "MNIST", "in", "a", "ten", "-", "class", "classification", "problem", ".", "SVHN", "and", "MNIST", "have", "distinct", "appearance", ",", "thus", "this", "adaptation", "is", "a", "challenging", "scenario", "especially", "in", "MNIST→SVHN", ".", "SVHN", "is", "colored", "and", "some", "images", "contain", "multiple", "digits", ".", "Therefore", ",", "a", "classifier", "trained", "on", "SVHN", "is", "expected", "to", "perform", "well", "on", "MNIST", ",", "but", "the", "reverse", "is", "not", "true", ".", "MNIST", "does", "not", "include", "any", "samples", "containing", "multiple", "digits", "and", "most", "samples", "are", "centered", "in", "images", ",", "thus", "adaptation", "from", "MNIST", "to", "SVHN", "is", "rather", "difficult", ".", "In", "both", "settings", ",", "we", "use", "1", ",", "000", "labeled", "target", "samples", "to", "find", "the", "optimal", "hyperparameters", ".", "We", "evaluate", "our", "method", "on", "both", "adaptation", "scenarios", "and", "achieved", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "both", "datasets", ".", "In", "particular", ",", "for", "the", "adaptation", "MNIST→SVHN", ",", "we", "outperformed", "other", "methods", "by", "more", "than", "10", "%", ".", "In", "Fig", ".", "3", "(", "c", ")(", "d", ")", ",", "we", "visualize", "the", "representations", "in", "MNIST→SVHN", ".", "Although", "the", "distributions", "seem", "to", "be", "separated", "between", "domains", ",", "the", "red", "SVHN", "samples", "become", "more", "discriminative", "using", "our", "method", "compared", "with", "non", "-", "adapted", "embedding", ".", "We", "also", "show", "the", "comparison", "between", "actual", "labeling", "method", "accuracy", "and", "testing", "accuracy", "in", "Fig", ".", "4", "(", "b", ")(", "c", ")", ".", "In", "this", "figure", ",", "we", "can", "see", "that", "the", "labeling", "accuracy", "rapidly", "drops", "in", "the", "initial", "adaptation", "stage", ".", "On", "the", "other", "hand", ",", "testing", "accuracy", "continues", "to", "improve", ",", "and", "finally", "exceeds", "the", "labeling", "accuracy", ".", "There", "are", "two", "questions", "about", "this", "interesting", "phenomenon", ".", "The", "first", "question", "is", "why", "does", "the", "labeling", "method", "continue", "to", "decrease", "despite", "the", "increase", "in", "the", "test", "accuracy", "?", "Target", "samples", "given", "pseudo", "-", "labels", "always", "include", "mistakenly", "labeled", "samples", "whereas", "those", "given", "no", "labels", "are", "ignored", "in", "our", "method", ".", "Therefore", ",", "the", "error", "will", "be", "reinforced", "in", "the", "target", "samples", "that", "are", "included", "in", "training", "set", ".", "The", "second", "question", "is", "why", "does", "the", "test", "accuracy", "continue", "to", "increase", "despite", "the", "lower", "labeling", "accuracy", "?", "The", "assumed", "reasons", "are", "that", "the", "network", "already", "acquires", "target", "discriminative", "representations", "in", "this", "phase", "and", "they", "can", "improve", "the", "accuracy", "using", "source", "samples", "and", "correctly", "labeled", "target", "samples", ".", "In", "Fig", ".", "4", "(", "f", ")", ",", "we", "also", "show", "the", "comparison", "of", "accuracy", "of", "the", "three", "networks", "F", "1", ",", "F", "2", ",", "F", "t", "in", "SVHN→MNIST", ".", "The", "accuracy", "of", "three", "networks", "is", "nearly", "the", "same", "in", "every", "step", ".", "The", "same", "thing", "is", "observed", "in", "other", "scenarios", ".", "From", "this", "result", ",", "we", "can", "state", "that", "the", "target", "-", "discriminative", "representations", "are", "shared", "in", "all", "three", "networks", ".", "SYN", "DIGITS→SVHN", "In", "this", "experiment", ",", "we", "aimed", "to", "address", "a", "common", "adaptation", "scenario", "from", "synthetic", "images", "to", "real", "images", ".", "The", "datasets", "of", "synthetic", "numbers", "[", "reference", "]", "consist", "of", "500", ",", "000", "images", "generated", "from", "Windows", "fonts", "by", "varying", "the", "text", ",", "positioning", ",", "orientation", ",", "background", "and", "stroke", "colors", ",", "and", "the", "amount", "of", "blur", ".", "We", "use", "479", ",", "400", "source", "samples", "and", "73", ",", "257", "target", "samples", "for", "training", ",", "and", "26", ",", "032", "target", "samples", "for", "testing", ".", "We", "use", "1", ",", "000", "SVHN", "samples", "as", "a", "validation", "set", ".", "Our", "method", "also", "outperforms", "other", "methods", "in", "this", "experiment", ".", "In", "this", "experiment", ",", "the", "effect", "of", "BN", "is", "not", "clear", "compared", "with", "other", "scenarios", ".", "The", "domain", "gap", "is", "considered", "small", "in", "this", "scenario", "as", "the", "performance", "of", "the", "source", "-", "only", "classifier", "shows", ".", "In", "Fig", ".", "4", "(", "d", ")", ",", "although", "the", "labeling", "accuracy", "is", "dropping", ",", "the", "accuracy", "of", "the", "learned", "network", "'s", "prediction", "is", "improving", "as", "in", "MNIST↔SVHN", ".", "SYN", "SIGNS→GTSRB", "This", "setting", "is", "similar", "to", "the", "pre", "-", "Table", "2", ".", "Results", "of", "Gradient", "stop", "experiment", ".", "When", "stopping", "gradients", "from", "Ft", ",", "we", "do", "not", "use", "backward", "gradients", "from", "Ft", "to", "F", ",", "and", "F", "learns", "only", "from", "F1", ",", "F2", ".", "When", "stopping", "gradients", "from", "F1", ",", "F2", ",", "we", "do", "not", "use", "backward", "gradients", "from", "F1", ",", "F2", "to", "F", ",", "and", "F", "learns", "from", "Ft", ".", "None", "denotes", "our", "proposed", "method", ",", "we", "backward", "all", "gradients", "from", "all", "branches", "to", "F", ".", "In", "these", "three", "adaptation", "scenarios", ",", "our", "method", "shows", "stable", "performance", ".", "vious", "setting", ",", "adaptation", "from", "synthetic", "images", "to", "real", "images", ",", "but", "we", "have", "a", "larger", "number", "of", "classes", ",", "namely", "43", "classes", "instead", "of", "10", ".", "We", "use", "the", "SYN", "SIGNS", "dataset", "[", "reference", "]", "for", "the", "source", "dataset", "and", "the", "GTSRB", "dataset", "[", "reference", "]", "for", "the", "target", "dataset", ",", "which", "consist", "of", "real", "traffic", "sign", "images", ".", "We", "select", "randomly", "31", ",", "367", "samples", "for", "target", "training", "samples", "and", "evaluate", "accuracy", "on", "the", "rest", "of", "the", "samples", ".", "A", "total", "of", "3", ",", "000", "labeled", "target", "samples", "are", "used", "for", "validation", ".", "In", "this", "scenario", ",", "our", "method", "outperforms", "other", "methods", ".", "This", "result", "shows", "that", "our", "method", "is", "effective", "for", "the", "adaptation", "from", "synthesized", "images", "to", "real", "images", ",", "which", "have", "diverse", "classes", ".", "In", "Fig", ".", "4", "(", "e", ")", ",", "the", "same", "tendency", "as", "in", "MNIST↔SVHN", "is", "observed", "in", "this", "adaptation", "scenario", ".", "section", ":", "Gradient", "Stop", "Experiment", "We", "evaluate", "the", "effect", "of", "the", "target", "-", "specific", "network", "in", "our", "method", ".", "We", "stop", "the", "gradient", "from", "upper", "layer", "networks", "F", "1", ",", "F", "2", ",", "and", "F", "t", "to", "examine", "the", "effect", "of", "F", "t", ".", "Table", "2", "shows", "three", "scenarios", "including", "the", "case", "where", "we", "stop", "the", "gradient", "from", "F", "1", ",", "F", "2", ",", "and", "F", "t", ".", "In", "all", "scenarios", ",", "when", "we", "backward", "all", "gradients", "from", "F", "1", ",", "F", "2", ",", "F", "t", ",", "we", "obtain", "clear", "performance", "improvements", ".", "In", "the", "experiment", "MNIST→MNIST", "-", "M", ",", "we", "can", "assume", "that", "only", "the", "backpropagation", "from", "F", "1", ",", "F", "2", "can", "not", "construct", "discriminative", "representations", "for", "target", "samples", "and", "confirm", "the", "effect", "of", "F", "t", ".", "For", "the", "adaptation", "MNIST→SVHN", ",", "the", "best", "performance", "is", "realized", "when", "F", "receives", "all", "gradients", "from", "upper", "networks", ".", "Backwarding", "all", "gradients", "will", "ensure", "both", "target", "-", "specific", "discriminative", "representations", "in", "difficult", "adaptations", ".", "In", "SYN", "SIGNS→GTSRB", ",", "backwarding", "only", "from", "F", "t", "produces", "the", "worst", "performance", "because", "these", "domains", "are", "similar", "and", "noisy", "pseudo", "-", "labeled", "target", "samples", "worsen", "the", "performance", ".", "section", ":", "A", "-", "distance", "From", "the", "theoretical", "results", "in", "[", "reference", "]", ",", "A", "-", "distance", "is", "usually", "used", "as", "a", "measure", "of", "domain", "discrepancy", ".", "The", "way", "of", "estimating", "empirical", "A", "-", "distance", "is", "simple", ",", "in", "which", "we", "train", "a", "classifier", "to", "classify", "a", "domain", "from", "each", "domains", "'", "feature", ".", "Then", ",", "the", "approximate", "distance", "is", "calculated", "Table", "3", ".", "Amazon", "Reviews", "experimental", "results", ".", "The", "accuracy", "(", "%", ")", "of", "the", "proposed", "method", "is", "shown", "with", "the", "result", "of", "VFAE", "[", "reference", "]", "and", "DANN", "[", "reference", "]", ".", "asd", "A", "=", "2", "(", "1", "−", "2ǫ", ")", ",", "where", "ǫ", "is", "the", "generalization", "error", "of", "the", "classifier", ".", "In", "Fig", ".", "4", "(", "g", ")", ",", "we", "show", "the", "A", "-", "distance", "calculated", "from", "each", "CNN", "features", ".", "We", "used", "linear", "SVM", "to", "calculate", "the", "distance", ".", "From", "this", "graph", ",", "we", "can", "see", "that", "our", "method", "certainly", "reduces", "the", "A", "-", "distance", "compared", "with", "the", "CNN", "trained", "on", "only", "source", "samples", ".", "In", "addition", ",", "when", "comparing", "DANN", "and", "our", "method", ",", "although", "DANN", "reduces", "A", "-", "distance", "much", "more", "than", "our", "method", ",", "our", "method", "shows", "superior", "performance", ".", "This", "indicates", "that", "minimizing", "the", "domain", "discrepancy", "is", "not", "necessarily", "an", "appropriate", "way", "to", "achieve", "better", "performance", ".", "Amazon", "Reviews", "Reviews", "are", "encoded", "in", "5", ",", "000", "dimensional", "vectors", "of", "bag", "-", "of", "-", "words", "unigrams", "and", "bigrams", "with", "binary", "labels", ".", "Negative", "labels", "are", "attached", "to", "the", "samples", "if", "they", "are", "ranked", "with", "1", "-", "3", "stars", ".", "Positive", "labels", "are", "attached", "if", "they", "are", "ranked", "with", "4", "or", "5", "stars", ".", "We", "have", "2", ",", "000", "labeled", "source", "samples", "and", "2", ",", "000", "unlabeled", "target", "samples", "for", "training", ",", "and", "between", "3", ",", "000", "and", "6", ",", "000", "samples", "for", "testing", ".", "We", "use", "200", "of", "labeled", "target", "samples", "for", "validation", ".", "From", "the", "results", "in", "Table", "3", ",", "our", "method", "performs", "better", "than", "VFAE", "[", "reference", "]", "and", "DANN", "[", "reference", "]", "in", "nine", "settings", "out", "of", "twelve", ".", "Our", "method", "is", "effective", "in", "learning", "a", "shallow", "network", "on", "different", "domains", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "have", "proposed", "a", "novel", "asymmetric", "tri", "-", "training", "method", "for", "unsupervised", "domain", "adaptation", ",", "which", "is", "simply", "implemented", ".", "We", "aimed", "to", "learn", "discriminative", "representations", "by", "utilizing", "pseudo", "-", "labels", "assigned", "to", "unlabeled", "target", "samples", ".", "We", "utilized", "three", "classifiers", ",", "two", "networks", "assign", "pseudo", "-", "labels", "to", "unlabeled", "target", "samples", "and", "the", "remaining", "network", "learns", "from", "them", ".", "We", "evaluated", "our", "method", "both", "on", "domain", "adaptation", "on", "a", "visual", "recognition", "task", "and", "a", "sentiment", "analysis", "task", ",", "outperforming", "other", "methods", ".", "In", "particular", ",", "our", "method", "outperformed", "the", "other", "methods", "by", "more", "than", "10", "%", "in", "the", "MNIST→SVHN", "adaptation", "task", ".", "section", ":", "Acknowledgement", "Proof", "of", "Theorem", "We", "introduce", "the", "derivation", "of", "theorem", "of", "the", "main", "paper", ".", "The", "ideal", "joint", "hypothesis", "is", "defined", "as", "h", "*", "=", "arg", "min", ",", "and", "its", "corresponding", "error", ",", "where", "R", "denotes", "the", "expected", "error", "on", "each", "hypothesis", ".", "We", "consider", "the", "pseudo", "-", "labeled", "target", "samples", "set", "given", "false", "labels", "at", "the", "ratio", "of", "ρ", ".", "The", "minimum", "shared", "error", "on", "S", ",", "T", "l", "is", "denoted", "as", "C", "′", ".", "Then", ",", "the", "following", "inequality", "holds", ":", "Proof", ".", "The", "probabiliy", "of", "false", "labels", "in", "the", "pseudo", "-", "labeled", "set", "T", "l", "is", "ρ", ".", "When", "we", "consider", "0", "-", "1", "loss", "function", "for", "l", ",", "the", "difference", "between", "the", "error", "based", "on", "the", "true", "labeled", "set", "and", "pseudo", "-", "labeled", "set", "is", "Then", ",", "the", "difference", "in", "the", "expected", "error", "is", ",", "From", "the", "characteritic", "of", "the", "loss", "function", ",", "the", "triangle", "inequality", "will", "hold", ",", "then", "From", "this", "result", ",", "the", "main", "inequality", "holds", ".", "section", ":", "CNN", "Architectures", "and", "training", "detail", "Four", "types", "of", "architectures", "are", "used", "for", "our", "method", ",", "which", "is", "based", "on", "[", "reference", "]", ".", "The", "network", "topology", "is", "shown", "in", "Figs", "6", ",", "7", "and", "8", ".", "The", "other", "hyperparameters", "are", "decided", "on", "the", "validation", "splits", ".", "The", "learning", "rate", "is", "set", "to", "0.05", "in", "SVHN↔MNIST", ".", "In", "the", "other", "scenarios", ",", "it", "is", "set", "to", "0.01", ".", "The", "batchsize", "for", "training", "F", "t", ",", "F", "is", "set", "as", "128", ",", "the", "batchsize", "for", "training", "F", "1", ",", "F", "2", ",", "F", "is", "set", "as", "64", "in", "all", "scenarios", ".", "In", "MNIST→MNIST", "-", "M", ",", "the", "dropout", "rate", "used", "in", "the", "experiment", "is", "0.2", "for", "training", "F", "t", ",", "0.5", "for", "training", "F", "1", ",", "F", "2", ".", "In", "MNIST→SVHN", ",", "we", "did", "not", "use", "dropout", ".", "We", "decreased", "learning", "rate", "to", "0.001", "after", "step", "10", ".", "In", "SVHN→MNIST", ",", "the", "dropout", "rate", "used", "in", "the", "experiment", "is", "0.5", ".", "In", "SYNDIGITS→SVHN", ",", "the", "dropout", "rate", "used", "in", "the", "experiment", "is", "0.5", ".", "In", "SYNSIGNS→GTSRB", ",", "the", "dropout", "rate", "used", "in", "the", "experiment", "is", "0.5", ".", "section", ":", "Supplementary", "experiments", "on", "MNIST→MNIST", "-", "M", "We", "observe", "the", "behavior", "of", "our", "model", "when", "increasing", "the", "number", "of", "steps", "up", "to", "one", "hundred", ".", "We", "show", "the", "result", "in", "Fig", ".", "5", ".", "Our", "model", "'s", "accuracy", "gets", "about", "97", "%", ".", "In", "our", "main", "experiments", ",", "we", "set", "the", "number", "of", "steps", "thirty", ",", "but", "from", "this", "experiment", ",", "further", "improvements", "can", "be", "expected", ".", "Figure", "6", ".", "The", "architecture", "used", "for", "MNIST→MNIST", "-", "M.", "We", "added", "BN", "layer", "in", "the", "last", "convolution", "layer", "and", "FC", "layers", "in", "F1", ",", "F2", ".", "We", "also", "used", "dropout", "in", "our", "experiment", ".", "Figure", "8", ".", "The", "architecture", "used", "in", "the", "adaptation", "Synthetic", "Signs→GTSRB", ".", "We", "added", "a", "BN", "layer", "after", "the", "last", "convolution", "layer", "in", "F", "and", "also", "used", "dropout", ".", "section", ":", "section", ":", "This", "work", "was", "funded", "by", "ImPACT", "Program", "of", "Council", "for", "Science", ",", "Technology", "and", "Innovation", "(", "Cabinet", "Office", ",", "Government", "of", "Japan", ")", "and", "supported", "by", "CREST", ",", "JST", ".", "section", ":"], "entity": [[0, 4, ["Asymmetric", "Tri", "-", "training"], "Method"], [5, 8, ["Unsupervised", "Domain", "Adaptation"], "Task"], [11, 15, ["Deep", "-", "layered", "models"], "Method"], [25, 26, ["accuracy"], "Metric"], [52, 55, ["unsupervised", "domain", "adaptation"], "Task"], [61, 62, ["classifier"], "Method"], [101, 102, ["accuracy"], "Metric"], [109, 111, ["discriminative", "representations"], "Method"], [130, 133, ["Tri", "-", "training"], "Method"], [135, 136, ["classifiers"], "Method"], [173, 176, ["unsupervised", "domain", "adaptation"], "Task"], [188, 190, ["neural", "networks"], "Method"], [232, 234, ["targetdiscriminative", "representations"], "Method"], [240, 242, ["digit", "recognition"], "Task"], [266, 268, ["domain", "adaptation"], "Task"], [276, 279, ["deep", "neural", "networks"], "Method"], [331, 332, ["recognition"], "Task"], [342, 346, ["deep", "-", "layered", "networks"], "Method"], [378, 382, ["object", "detection", "or", "segmentation"], "Task"], [390, 391, ["CNN"], "Method"], [419, 420, ["CNN"], "Method"], [422, 423, ["ImageNet"], "Material"], [430, 433, ["multimodal", "learning", "tasks"], "Task"], [434, 436, ["image", "captioning"], "Task"], [440, 443, ["visual", "question", "answering"], "Task"], [452, 454, ["neural", "networks"], "Method"], [497, 498, ["Internet"], "Material"], [525, 526, ["classifier"], "Method"], [547, 552, ["domain", "-", "invariantly", "discriminative", "representations"], "Method"], [587, 590, ["unsupervised", "domain", "adaptation"], "Task"], [596, 597, ["classifier"], "Method"], [625, 629, ["deep", "domain", "adaptation", "methods"], "Method"], [638, 639, ["adaptation"], "Task"], [696, 697, ["classifier"], "Method"], [717, 719, ["discriminative", "classifier"], "Method"], [735, 737, ["nondiscriminative", "representations"], "Method"], [739, 740, ["classifier"], "Method"], [779, 780, ["classifiers"], "Method"], [794, 796, ["discriminative", "representations"], "Method"], [825, 827, ["targetspecific", "networks"], "Method"], [834, 837, ["Co", "-", "training"], "Method"], [838, 841, ["tri", "-", "training"], "Method"], [846, 847, ["classifiers"], "Method"], [895, 898, ["unsupervised", "domain", "adaptation"], "Task"], [926, 929, ["unsupervised", "domain", "adaptation"], "Task"], [939, 941, ["neural", "networks"], "Method"], [990, 993, ["digit", "classification", "task"], "Task"], [994, 998, ["traffic", "sign", "classification", "task"], "Task"], [999, 1002, ["sentiment", "analysis", "task"], "Task"], [1004, 1007, ["Amazon", "Review", "dataset"], "Material"], [1028, 1030, ["adaptation", "scenario"], "Task"], [1031, 1032, ["MNIST→SVHN"], "Method"], [1059, 1061, ["domain", "adaptation"], "Task"], [1083, 1084, ["adaptation"], "Task"], [1133, 1135, ["expected", "loss"], "Metric"], [1159, 1162, ["shared", "expected", "loss"], "Metric"], [1164, 1167, ["shared", "expected", "loss"], "Metric"], [1222, 1224, ["deep", "architectures"], "Method"], [1226, 1229, ["maximum", "mean", "discrepancy"], "Method"], [1230, 1231, ["MMD"], "Method"], [1234, 1239, ["loss", "of", "domain", "classifier", "network"], "Method"], [1270, 1271, ["CNN"], "Method"], [1306, 1310, ["target", "-", "discriminative", "representations"], "Method"], [1333, 1335, ["specific", "classifier"], "Method"], [1339, 1342, ["residual", "network", "structure"], "Method"], [1351, 1355, ["target", "-", "specific", "network"], "Method"], [1363, 1365, ["transductive", "methods"], "Method"], [1382, 1385, ["unsupervised", "domain", "adaptation"], "Task"], [1466, 1467, ["classifier"], "Method"], [1480, 1483, ["self", "-", "training"], "Task"], [1523, 1524, ["classifier"], "Method"], [1525, 1528, ["Co", "-", "training"], "Method"], [1530, 1531, ["classifiers"], "Method"], [1565, 1566, ["classifier"], "Method"], [1573, 1575, ["generalization", "ability"], "Metric"], [1576, 1579, ["co", "-", "training"], "Method"], [1610, 1613, ["co", "-", "training"], "Method"], [1616, 1618, ["domain", "adaptation"], "Method"], [1619, 1622, ["Tri", "-", "training"], "Method"], [1629, 1632, ["co", "-", "training"], "Method"], [1638, 1641, ["co", "-", "training"], "Method"], [1642, 1643, ["tritraining"], "Method"], [1649, 1650, ["classifiers"], "Method"], [1659, 1662, ["Tri", "-", "training"], "Method"], [1673, 1676, ["tri", "-", "training"], "Method"], [1678, 1679, ["classifier"], "Method"], [1691, 1694, ["shared", "feature", "extractor"], "Method"], [1695, 1696, ["F"], "Method"], [1698, 1699, ["classifiers"], "Method"], [1725, 1729, ["target", "-", "specific", "classifier"], "Method"], [1730, 1731, ["Ft"], "Method"], [1779, 1782, ["tri", "-", "training"], "Method"], [1808, 1810, ["tritraining", "method"], "Method"], [1812, 1814, ["domain", "adaptation"], "Task"], [1817, 1818, ["classifiers"], "Method"], [1833, 1835, ["neural", "network"], "Method"], [1846, 1847, ["classifier"], "Method"], [1854, 1856, ["entropy", "regularization"], "Method"], [1901, 1903, ["domain", "adaptation"], "Task"], [1909, 1911, ["targetspecific", "network"], "Method"], [1924, 1926, ["labeling", "networks"], "Method"], [1935, 1936, ["accuracy"], "Metric"], [1945, 1947, ["network", "structure"], "Method"], [1987, 1989, ["classifier", "F"], "Method"], [1999, 2003, ["target", "-", "specific", "network"], "Method"], [2032, 2035, ["shared", "network", "F"], "Method"], [2052, 2054, ["shared", "network"], "Method"], [2059, 2061, ["network", "architecture"], "Method"], [2078, 2079, ["training"], "Task"], [2080, 2081, ["implementation"], "Task"], [2087, 2090, ["shared", "network", "F"], "Method"], [2138, 2139, ["Loss"], "Task"], [2140, 2143, ["Multiview", "Features", "Network"], "Task"], [2151, 2154, ["co", "-", "training"], "Method"], [2155, 2157, ["domain", "adaptation"], "Task"], [2181, 2182, ["accuracy"], "Metric"], [2282, 2285, ["learning", "F", "1"], "Task"], [2297, 2303, ["softmax", "cross", "-", "entropy", "loss", "function"], "Method"], [2319, 2321, ["Learning", "Procedure"], "Task"], [2322, 2324, ["Labeling", "Method"], "Method"], [2357, 2361, ["labeling", "and", "learning", "method"], "Method"], [2407, 2409, ["F", "t"], "Method"], [2413, 2415, ["category", "loss"], "Method"], [2549, 2550, ["classifiers"], "Method"], [2557, 2558, ["prediction"], "Task"], [2616, 2617, ["training"], "Task"], [2619, 2621, ["function", "Labeling"], "Method"], [2625, 2626, ["labeling"], "Task"], [2776, 2778, ["Discriminative", "representations"], "Method"], [2784, 2788, ["target", "-", "specific", "network"], "Method"], [2840, 2841, ["accuracy"], "Metric"], [2852, 2856, ["target", "-", "discriminative", "representations"], "Method"], [2862, 2863, ["accuracy"], "Metric"], [2876, 2877, ["accuracy"], "Metric"], [2884, 2886, ["Batch", "Normalization"], "Method"], [2887, 2889, ["Domain", "Adaptation"], "Task"], [2889, 2894, ["Batch", "normalization", "(", "BN", ")"], "Method"], [2908, 2909, ["CNN"], "Method"], [2916, 2918, ["training", "speed"], "Metric"], [2921, 2922, ["accuracy"], "Metric"], [2930, 2932, ["domain", "adaptation"], "Task"], [2980, 2981, ["BN"], "Method"], [3061, 3063, ["expected", "error"], "Metric"], [3084, 3087, ["ideal", "joint", "hypothesis"], "Method"], [3154, 3156, ["expected", "error"], "Metric"], [3168, 3170, ["expected", "error"], "Metric"], [3191, 3194, ["ideal", "joint", "hypothesis"], "Method"], [3260, 3263, ["fully", "connected", "layer"], "Method"], [3264, 3266, ["deep", "models"], "Method"], [3309, 3311, ["shared", "error"], "Metric"], [3495, 3496, ["accuracy"], "Metric"], [3497, 3501, ["target", "-", "specific", "networks"], "Method"], [3505, 3508, ["Visual", "Domain", "Adaptation"], "Task"], [3509, 3512, ["visual", "domain", "adaptation"], "Task"], [3529, 3530, ["MNIST"], "Material"], [3538, 3542, ["Street", "View", "House", "Numbers"], "Material"], [3543, 3544, ["SVHN"], "Material"], [3550, 3552, ["Synthetic", "Digits"], "Material"], [3553, 3555, ["SYN", "DIGITS"], "Material"], [3570, 3573, ["Synthetic", "Traffic", "Signs"], "Material"], [3574, 3576, ["SYN", "SIGNS"], "Material"], [3581, 3586, ["German", "Traffic", "Signs", "Recognition", "Benchmark"], "Material"], [3631, 3632, ["Office"], "Material"], [3644, 3647, ["visual", "domain", "adaptation"], "Task"], [3680, 3686, ["fine", "-", "tuning", "of", "pretrained", "networks"], "Task"], [3687, 3688, ["ImageNet"], "Material"], [3725, 3726, ["Adaptation"], "Task"], [3727, 3729, ["Amazon", "Reviews"], "Task"], [3744, 3747, ["Amazon", "Reviews", "dataset"], "Material"], [3789, 3792, ["domain", "adaptation", "scenarios"], "Task"], [3811, 3814, ["unsupervised", "domain", "adaptation"], "Task"], [3823, 3826, ["visual", "domain", "adaptation"], "Task"], [3827, 3830, ["Maximum", "Mean", "Discrepancy"], "Method"], [3831, 3832, ["MMD"], "Method"], [3837, 3841, ["Domain", "Adversarial", "Neural", "Network"], "Method"], [3842, 3843, ["DANN"], "Method"], [3848, 3852, ["Deep", "Reconstruction", "Classification", "Network"], "Method"], [3853, 3855, ["DRCN", ")"], "Method"], [3859, 3862, ["Domain", "Separation", "Networks"], "Method"], [3863, 3864, ["DSN"], "Method"], [3870, 3874, ["k", "-", "Nearest", "Neighbor"], "Method"], [3874, 3876, ["based", "adaptation"], "Method"], [3877, 3880, ["kNN", "-", "Ad"], "Method"], [3890, 3891, ["MMD"], "Method"], [3904, 3905, ["CNN"], "Method"], [3916, 3919, ["Variational", "Fair", "AutoEncoder"], "Method"], [3920, 3921, ["VFAE"], "Method"], [3932, 3935, ["Amazon", "Reviews", "experiment"], "Material"], [3951, 3952, ["CNN"], "Method"], [3977, 3979, ["discriminator", "networks"], "Method"], [3985, 3986, ["classifier"], "Method"], [4021, 4023, ["labeling", "method"], "Method"], [4026, 4027, ["MNIST→SVHN"], "Method"], [4040, 4041, ["MomentumSGD"], "Method"], [4042, 4043, ["optimization"], "Task"], [4052, 4054, ["learning", "rate"], "Metric"], [4086, 4088, ["network", "architecture"], "Method"], [4097, 4100, ["Amazon", "Reviews", "dataset"], "Material"], [4119, 4122, ["dense", "hidden", "layer"], "Method"], [4143, 4144, ["CNN"], "Method"], [4163, 4164, ["Adagrad"], "Method"], [4168, 4169, ["optimization"], "Task"], [4178, 4180, ["mean", "accuracy"], "Metric"], [4211, 4212, ["BN"], "Method"], [4224, 4226, ["image", "recognition"], "Task"], [4231, 4232, ["BN"], "Method"], [4256, 4257, ["BN"], "Method"], [4259, 4261, ["network", "architecture"], "Method"], [4271, 4272, ["MNIST→SVHN"], "Method"], [4283, 4286, ["MNIST→MNIST", "-", "M"], "Method"], [4287, 4290, ["last", "pooling", "layer"], "Method"], [4298, 4300, ["adaptation", "scenario"], "Task"], [4307, 4308, ["MNIST"], "Material"], [4328, 4330, ["BSDS500", "datasets"], "Material"], [4343, 4344, ["BSDS500"], "Material"], [4354, 4356, ["domain", "shift"], "Method"], [4358, 4359, ["adaptation"], "Task"], [4361, 4362, ["CNN"], "Method"], [4415, 4418, ["Visualization", "of", "features"], "Task"], [4447, 4448, ["adaptation"], "Task"], [4457, 4458, ["accuracy"], "Metric"], [4461, 4463, ["labeling", "accuracy"], "Metric"], [4470, 4472, ["test", "accuracy"], "Metric"], [4478, 4480, ["test", "accuracy"], "Metric"], [4493, 4494, ["accuracy"], "Metric"], [4500, 4502, ["labeling", "accuracy"], "Metric"], [4514, 4516, ["labeling", "accuracy"], "Metric"], [4520, 4521, ["accuracy"], "Metric"], [4525, 4526, ["SVHN↔MNIST"], "Method"], [4538, 4539, ["adaptation"], "Method"], [4540, 4541, ["SVHN"], "Material"], [4545, 4546, ["MNIST"], "Material"], [4548, 4553, ["ten", "-", "class", "classification", "problem"], "Task"], [4554, 4555, ["SVHN"], "Material"], [4556, 4557, ["MNIST"], "Material"], [4570, 4571, ["MNIST→SVHN"], "Method"], [4572, 4573, ["SVHN"], "Material"], [4585, 4586, ["classifier"], "Method"], [4588, 4589, ["SVHN"], "Material"], [4595, 4596, ["MNIST"], "Material"], [4604, 4605, ["MNIST"], "Material"], [4624, 4625, ["MNIST"], "Material"], [4626, 4627, ["SVHN"], "Material"], [4676, 4678, ["adaptation", "MNIST→SVHN"], "Method"], [4704, 4705, ["MNIST→SVHN"], "Method"], [4728, 4732, ["non", "-", "adapted", "embedding"], "Method"], [4740, 4743, ["labeling", "method", "accuracy"], "Metric"], [4744, 4746, ["testing", "accuracy"], "Metric"], [4765, 4767, ["labeling", "accuracy"], "Metric"], [4780, 4782, ["testing", "accuracy"], "Metric"], [4790, 4792, ["labeling", "accuracy"], "Metric"], [4809, 4811, ["labeling", "method"], "Method"], [4819, 4821, ["test", "accuracy"], "Metric"], [4869, 4871, ["test", "accuracy"], "Metric"], [4877, 4879, ["labeling", "accuracy"], "Metric"], [4889, 4892, ["target", "discriminative", "representations"], "Method"], [4900, 4901, ["accuracy"], "Metric"], [4924, 4925, ["accuracy"], "Metric"], [4938, 4939, ["SVHN→MNIST"], "Method"], [4941, 4942, ["accuracy"], "Metric"], [4971, 4975, ["target", "-", "discriminative", "representations"], "Method"], [4983, 4984, ["DIGITS→SVHN"], "Method"], [4994, 4996, ["adaptation", "scenario"], "Task"], [5095, 5096, ["BN"], "Method"], [5118, 5122, ["source", "-", "only", "classifier"], "Method"], [5134, 5136, ["labeling", "accuracy"], "Metric"], [5140, 5141, ["accuracy"], "Metric"], [5146, 5147, ["prediction"], "Task"], [5151, 5152, ["MNIST↔SVHN"], "Method"], [5256, 5257, ["adaptation"], "Task"], [5283, 5286, ["SYN", "SIGNS", "dataset"], "Material"], [5295, 5297, ["GTSRB", "dataset"], "Material"], [5326, 5327, ["accuracy"], "Metric"], [5346, 5347, ["validation"], "Task"], [5368, 5369, ["adaptation"], "Task"], [5399, 5401, ["adaptation", "scenario"], "Task"], [5404, 5406, ["Gradient", "Stop"], "Task"], [5497, 5501, ["experiment", "MNIST→MNIST", "-", "M"], "Method"], [5518, 5520, ["discriminative", "representations"], "Method"], [5533, 5535, ["adaptation", "MNIST→SVHN"], "Method"], [5559, 5561, ["discriminative", "representations"], "Method"], [5566, 5568, ["SYN", "SIGNS→GTSRB"], "Method"], [5596, 5599, ["A", "-", "distance"], "Metric"], [5608, 5611, ["A", "-", "distance"], "Metric"], [5616, 5620, ["measure", "of", "domain", "discrepancy"], "Metric"], [5637, 5638, ["classifier"], "Method"], [5658, 5660, ["Amazon", "Reviews"], "Material"], [5664, 5665, ["accuracy"], "Metric"], [5683, 5684, ["DANN"], "Method"], [5702, 5704, ["generalization", "error"], "Metric"], [5706, 5707, ["classifier"], "Method"], [5719, 5722, ["A", "-", "distance"], "Metric"], [5730, 5732, ["linear", "SVM"], "Method"], [5750, 5753, ["A", "-", "distance"], "Metric"], [5756, 5757, ["CNN"], "Method"], [5768, 5769, ["DANN"], "Method"], [5774, 5775, ["DANN"], "Method"], [5776, 5779, ["A", "-", "distance"], "Metric"], [5809, 5812, ["Amazon", "Reviews", "Reviews"], "Material"], [5821, 5827, ["bag", "-", "of", "-", "words", "unigrams"], "Method"], [5828, 5829, ["bigrams"], "Method"], [5917, 5918, ["VFAE"], "Method"], [5922, 5923, ["DANN"], "Method"], [5940, 5942, ["shallow", "network"], "Task"], [5964, 5967, ["unsupervised", "domain", "adaptation"], "Task"], [5977, 5979, ["discriminative", "representations"], "Method"], [5993, 5994, ["classifiers"], "Method"], [6019, 6021, ["domain", "adaptation"], "Task"], [6023, 6026, ["visual", "recognition", "task"], "Task"], [6028, 6031, ["sentiment", "analysis", "task"], "Task"], [6052, 6055, ["MNIST→SVHN", "adaptation", "task"], "Method"], [6120, 6123, ["minimum", "shared", "error"], "Metric"], [6174, 6175, ["error"], "Metric"], [6193, 6195, ["expected", "error"], "Metric"], [6270, 6272, ["learning", "rate"], "Metric"], [6277, 6278, ["SVHN↔MNIST"], "Method"], [6323, 6326, ["MNIST→MNIST", "-", "M"], "Method"], [6328, 6330, ["dropout", "rate"], "Metric"], [6351, 6352, ["MNIST→SVHN"], "Method"], [6361, 6363, ["learning", "rate"], "Metric"], [6370, 6371, ["SVHN→MNIST"], "Method"], [6373, 6375, ["dropout", "rate"], "Metric"], [6383, 6384, ["SYNDIGITS→SVHN"], "Method"], [6386, 6388, ["dropout", "rate"], "Metric"], [6396, 6397, ["SYNSIGNS→GTSRB"], "Method"], [6399, 6401, ["dropout", "rate"], "Metric"], [6413, 6416, ["MNIST→MNIST", "-", "M"], "Method"], [6446, 6447, ["accuracy"], "Metric"], [6483, 6486, ["MNIST→MNIST", "-", "M."], "Method"], [6496, 6498, ["FC", "layers"], "Method"], [6519, 6522, ["adaptation", "Synthetic", "Signs→GTSRB"], "Method"], [6526, 6528, ["BN", "layer"], "Method"], [6538, 6539, ["dropout"], "Method"], [168, 171, ["tri", "-", "training"], "Method"], [285, 286, ["CNN"], "Method"], [370, 371, ["CNN"], "Method"], [891, 894, ["tri", "-", "training"], "Method"], [921, 924, ["tri", "-", "training"], "Method"], [3926, 3927, ["DANN"], "Method"], [4290, 4293, ["MNIST→MNIST", "-", "M"], "Method"], [4312, 4313, ["MNIST"], "Material"], [4315, 4316, ["MNIST"], "Material"], [4347, 4348, ["MNIST"], "Material"], [4718, 4719, ["SVHN"], "Material"], [5071, 5072, ["SVHN"], "Material"], [5153, 5155, ["SYN", "SIGNS→GTSRB"], "Method"], [5394, 5395, ["MNIST↔SVHN"], "Method"], [5626, 5629, ["A", "-", "distance"], "Metric"], [5678, 5679, ["VFAE"], "Method"], [5725, 5726, ["CNN"], "Method"], [5959, 5962, ["tri", "-", "training"], "Method"], [6223, 6224, ["CNN"], "Method"], [6561, 6562, ["Office"], "Material"]]} {"doc_id": "44fcb91c6c436902edb35b4c453c34fb353b14ad", "text": ["A", "Convolutional", "Encoder", "Model", "for", "Neural", "Machine", "Translation", "section", ":", "Abstract", "The", "prevalent", "approach", "to", "neural", "machine", "translation", "relies", "on", "bi", "-", "directional", "LSTMs", "to", "encode", "the", "source", "sentence", ".", "We", "present", "a", "faster", "and", "simpler", "architecture", "based", "on", "a", "succession", "of", "convolutional", "layers", ".", "This", "allows", "to", "encode", "the", "source", "sentence", "simultaneously", "compared", "to", "recurrent", "networks", "for", "which", "computation", "is", "constrained", "by", "temporal", "dependencies", ".", "On", "WMT'16", "EnglishRomanian", "translation", "we", "achieve", "competitive", "accuracy", "to", "the", "state", "-", "of", "-", "the", "-", "art", "and", "on", "WMT'15", "English", "-", "German", "we", "outperform", "several", "recently", "published", "results", ".", "Our", "models", "obtain", "almost", "the", "same", "accuracy", "as", "a", "very", "deep", "LSTM", "setup", "on", "WMT'14", "English", "-", "French", "translation", ".", "We", "speed", "up", "CPU", "decoding", "by", "more", "than", "two", "times", "at", "the", "same", "or", "higher", "accuracy", "as", "a", "strong", "bidirectional", "LSTM", ".", "1", "section", ":", "Introduction", "Neural", "machine", "translation", "(", "NMT", ")", "is", "an", "end", "-", "to", "-", "end", "approach", "to", "machine", "translation", ".", "The", "most", "successful", "approach", "to", "date", "encodes", "the", "source", "sentence", "with", "a", "bi", "-", "directional", "recurrent", "neural", "network", "(", "RNN", ")", "into", "a", "variable", "length", "representation", "and", "then", "generates", "the", "translation", "left", "-", "to", "-", "right", "with", "another", "RNN", "where", "both", "components", "interface", "via", "a", "soft", "-", "attention", "mechanism", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "Recurrent", "networks", "are", "typically", "parameterized", "as", "long", "short", "term", "memory", "networks", "(", "LSTM", ";", "[", "reference", "]", "or", "gated", "recurrent", "units", "[", "reference", "][", "reference", "]", ",", "often", "with", "residual", "or", "skip", "connections", "[", "reference", "][", "reference", "]", "to", "enable", "stacking", "of", "several", "layers", "(", "§", "2", ")", ".", "There", "have", "been", "several", "attempts", "to", "use", "convolutional", "encoder", "models", "for", "neural", "machine", "trans", "-", "lation", "in", "the", "past", "but", "they", "were", "either", "only", "applied", "to", "rescoring", "n", "-", "best", "lists", "of", "classical", "systems", "[", "reference", "]", "or", "were", "not", "competitive", "to", "recurrent", "alternatives", "[", "reference", "]", ".", "This", "is", "despite", "several", "attractive", "properties", "of", "convolutional", "networks", ".", "For", "example", ",", "convolutional", "networks", "operate", "over", "a", "fixed", "-", "size", "window", "of", "the", "input", "sequence", "which", "enables", "the", "simultaneous", "computation", "of", "all", "features", "for", "a", "source", "sentence", ".", "This", "contrasts", "to", "RNNs", "which", "maintain", "a", "hidden", "state", "of", "the", "entire", "past", "that", "prevents", "parallel", "computation", "within", "a", "sequence", ".", "A", "succession", "of", "convolutional", "layers", "provides", "a", "shorter", "path", "to", "capture", "relationships", "between", "elements", "of", "a", "sequence", "compared", "to", "RNNs", ".", "[", "reference", "]", "This", "also", "eases", "learning", "because", "the", "resulting", "tree", "-", "structure", "applies", "a", "fixed", "number", "of", "non", "-", "linearities", "compared", "to", "a", "recurrent", "neural", "network", "for", "which", "the", "number", "of", "non", "-", "linearities", "vary", "depending", "on", "the", "time", "-", "step", ".", "Because", "processing", "is", "bottom", "-", "up", ",", "all", "words", "undergo", "the", "same", "number", "of", "transformations", ",", "whereas", "for", "RNNs", "the", "first", "word", "is", "over", "-", "processed", "and", "the", "last", "word", "is", "transformed", "only", "once", ".", "In", "this", "paper", "we", "show", "that", "an", "architecture", "based", "on", "convolutional", "layers", "is", "very", "competitive", "to", "recurrent", "encoders", ".", "We", "investigate", "simple", "average", "pooling", "as", "well", "as", "parameterized", "convolutions", "as", "an", "alternative", "to", "recurrent", "encoders", "and", "enable", "very", "deep", "convolutional", "encoders", "by", "using", "residual", "connections", "[", "reference", "][", "reference", "]", ".", "We", "experiment", "on", "several", "standard", "datasets", "and", "compare", "our", "approach", "to", "variants", "of", "recurrent", "encoders", "such", "as", "uni", "-", "directional", "and", "bi", "-", "directional", "LSTMs", ".", "On", "WMT'16", "English", "-", "Romanian", "translation", "we", "achieve", "accuracy", "that", "is", "very", "competitive", "to", "the", "current", "state", "-", "of", "-", "the", "-", "art", "result", ".", "We", "perform", "competitively", "on", "WMT'15", "English", "-", "German", ",", "and", "nearly", "match", "the", "performance", "of", "the", "best", "WMT'14", "English", "-", "French", "system", "based", "on", "a", "deep", "LSTM", "setup", "when", "comparing", "on", "a", "commonly", "used", "subset", "2", "For", "kernel", "width", "k", "and", "sequence", "length", "n", "we", "require", "max", "1", ",", "forwards", "on", "a", "succession", "of", "stacked", "convolutional", "layers", "compared", "to", "n", "forwards", "with", "an", "RNN", ".", "section", ":", "arXiv:1611.02344v3", "[", "cs", ".", "CL", "]", "25", "Jul", "2017", "of", "the", "training", "data", "[", "reference", "][", "reference", "]", ".", "section", ":", "Recurrent", "Neural", "Machine", "Translation", "The", "general", "architecture", "of", "the", "models", "in", "this", "work", "follows", "the", "encoder", "-", "decoder", "approach", "with", "soft", "attention", "first", "introduced", "in", "[", "reference", "]", ".", "A", "source", "sentence", "x", "=", "(", "x", "1", ",", ".", ".", ".", ",", "x", "m", ")", "of", "m", "words", "is", "processed", "by", "an", "encoder", "which", "outputs", "a", "sequence", "of", "states", "z", "=", "(", "z", "1", ".", ".", ".", ".", ",", "z", "m", ")", ".", "The", "decoder", "is", "an", "RNN", "network", "that", "computes", "a", "new", "hidden", "state", "s", "i", "+", "1", "based", "on", "the", "previous", "state", "s", "i", ",", "an", "embedding", "g", "i", "of", "the", "previous", "target", "language", "word", "y", "i", ",", "as", "well", "as", "a", "conditional", "input", "c", "i", "derived", "from", "the", "encoder", "output", "z.", "We", "use", "LSTMs", "(", "Hochreiter", "and", "Schmidhuber", ",", "1997", ")", "for", "all", "decoder", "networks", "whose", "state", "s", "i", "comprises", "of", "a", "cell", "vector", "and", "a", "hidden", "vector", "h", "i", "which", "is", "output", "by", "the", "LSTM", "at", "each", "time", "step", ".", "We", "input", "c", "i", "into", "the", "LSTM", "by", "concatenating", "it", "to", "g", "i", ".", "The", "translation", "model", "computes", "a", "distribution", "over", "the", "V", "possible", "target", "words", "y", "i", "+", "1", "by", "transforming", "the", "LSTM", "output", "h", "i", "via", "a", "linear", "layer", "with", "weights", "W", "o", "and", "bias", "b", "o", ":", "The", "conditional", "input", "c", "i", "at", "time", "i", "is", "computed", "via", "a", "simple", "dot", "-", "product", "style", "attention", "mechanism", "[", "reference", "]", ".", "Specifically", ",", "we", "transform", "the", "decoder", "hidden", "state", "h", "i", "by", "a", "linear", "layer", "with", "weights", "W", "d", "and", "b", "d", "to", "match", "the", "size", "of", "the", "embedding", "of", "the", "previous", "target", "word", "g", "i", "and", "then", "sum", "the", "two", "representations", "to", "yield", "d", "i", ".", "Conditional", "input", "c", "i", "is", "a", "weighted", "sum", "of", "attention", "scores", "a", "i", "∈", "R", "m", "and", "encoder", "outputs", "z.", "The", "attention", "scores", "a", "i", "are", "determined", "by", "a", "dot", "product", "between", "h", "i", "with", "each", "z", "j", ",", "followed", "by", "a", "softmax", "over", "the", "source", "sequence", ":", "In", "preliminary", "experiments", ",", "we", "did", "not", "find", "the", "MLP", "attention", "of", "[", "reference", "]", "to", "perform", "significantly", "better", "in", "terms", "of", "BLEU", "nor", "perplexity", ".", "However", ",", "we", "found", "the", "dot", "-", "product", "attention", "to", "be", "more", "favorable", "in", "terms", "of", "training", "and", "evaluation", "speed", ".", "We", "use", "bi", "-", "directional", "LSTMs", "to", "implement", "recurrent", "encoders", "similar", "to", "[", "reference", "]", "which", "achieved", "some", "of", "the", "best", "WMT14", "EnglishFrench", "results", "reported", "to", "date", ".", "First", ",", "each", "word", "of", "the", "input", "sequence", "x", "is", "embedded", "in", "distributional", "space", "resulting", "in", "e", "=", "(", "e", "1", ",", ".", ".", ".", ",", "e", "m", ")", ".", "The", "embeddings", "are", "input", "to", "two", "stacks", "of", "uni", "-", "directional", "RNNs", "where", "the", "output", "of", "each", "layer", "is", "reversed", "before", "being", "fed", "into", "the", "next", "layer", ".", "The", "first", "stack", "takes", "the", "original", "sequence", "while", "the", "second", "takes", "the", "reversed", "input", "sequence", ";", "the", "output", "of", "the", "second", "stack", "is", "reversed", "so", "that", "the", "final", "outputs", "of", "the", "stacks", "align", ".", "Finally", ",", "the", "top", "-", "level", "hidden", "states", "of", "the", "two", "stacks", "are", "concatenated", "and", "fed", "into", "a", "linear", "layer", "to", "yield", "z.", "We", "denote", "this", "encoder", "architecture", "as", "BiLSTM", ".", "3", "Non", "-", "recurrent", "Encoders", "section", ":", "Pooling", "Encoder", "A", "simple", "baseline", "for", "non", "-", "recurrent", "encoders", "is", "the", "pooling", "model", "described", "in", "[", "reference", "]", "which", "simply", "averages", "the", "embeddings", "of", "k", "consecutive", "words", ".", "Averaging", "word", "embeddings", "does", "not", "convey", "positional", "information", "besides", "that", "the", "words", "in", "the", "input", "are", "somewhat", "close", "to", "each", "other", ".", "As", "a", "remedy", ",", "we", "add", "position", "embeddings", "to", "encode", "the", "absolute", "position", "of", "each", "source", "word", "within", "a", "sentence", ".", "Each", "source", "embedding", "e", "j", "therefore", "contains", "a", "position", "embedding", "l", "j", "as", "well", "as", "the", "word", "embedding", "w", "j", ".", "Position", "embeddings", "have", "also", "been", "found", "helpful", "in", "memory", "networks", "for", "question", "-", "answering", "and", "language", "modeling", "[", "reference", "]", ".", "Similar", "to", "the", "recurrent", "encoder", "(", "§", "2", ")", ",", "the", "attention", "scores", "a", "ij", "are", "computed", "from", "the", "pooled", "representations", "z", "j", ",", "however", ",", "the", "conditional", "input", "c", "i", "is", "a", "weighted", "sum", "of", "the", "embeddings", "e", "j", ",", "not", "z", "j", ",", "i.e.", ",", "The", "input", "sequence", "is", "padded", "prior", "to", "pooling", "such", "that", "the", "encoder", "output", "matches", "the", "input", "length", "|z|", "=", "|x|.", "We", "set", "k", "to", "5", "in", "all", "experiments", "as", "[", "reference", "]", ".", "section", ":", "Convolutional", "Encoder", "A", "straightforward", "extension", "of", "pooling", "is", "to", "learn", "the", "kernel", "in", "a", "convolutional", "neural", "network", "(", "CNN", ")", ".", "The", "encoder", "output", "z", "j", "contains", "information", "about", "a", "fixed", "-", "sized", "context", "depending", "on", "the", "kernel", "width", "k", "but", "the", "desired", "context", "width", "may", "vary", ".", "This", "can", "be", "addressed", "by", "stacking", "several", "layers", "of", "convolutions", "followed", "by", "non", "-", "linearities", ":", "additional", "layers", "increase", "the", "total", "context", "size", "while", "non", "-", "linearities", "can", "modulate", "the", "effective", "size", "of", "the", "context", "as", "needed", ".", "For", "instance", ",", "stacking", "5", "convolutions", "with", "kernel", "width", "k", "=", "3", "results", "in", "an", "input", "field", "of", "11", "words", ",", "i.e.", ",", "each", "output", "depends", "on", "11", "input", "words", ",", "and", "the", "non", "-", "linearities", "allow", "the", "encoder", "to", "exploit", "the", "full", "input", "field", ",", "or", "to", "concentrate", "on", "fewer", "words", "as", "needed", ".", "To", "ease", "learning", "for", "deep", "encoders", ",", "we", "add", "residual", "connections", "from", "the", "input", "of", "each", "convolution", "to", "the", "output", "and", "then", "apply", "the", "non", "-", "linear", "activation", "function", "to", "the", "output", "(", "tanh", ";", "[", "reference", "]", ";", "the", "non", "-", "linearities", "are", "therefore", "not", "'", "bypassed", "'", ".", "Multi", "-", "layer", "CNNs", "are", "constructed", "by", "stacking", "several", "blocks", "on", "top", "of", "each", "other", ".", "The", "CNNs", "do", "not", "contain", "pooling", "layers", "which", "are", "commonly", "used", "for", "down", "-", "sampling", ",", "i.e.", ",", "the", "full", "source", "sequence", "length", "will", "be", "retained", "after", "the", "network", "has", "been", "applied", ".", "Similar", "to", "the", "pooling", "model", ",", "the", "convolutional", "encoder", "uses", "position", "embeddings", ".", "The", "final", "encoder", "consists", "of", "two", "stacked", "convolutional", "networks", "(", "Figure", "1", ")", ":", "CNN", "-", "a", "produces", "the", "encoder", "output", "z", "j", "to", "compute", "the", "attention", "scores", "a", "i", ",", "while", "the", "conditional", "input", "c", "i", "to", "the", "decoder", "is", "computed", "by", "summing", "the", "outputs", "of", "CNN", "-", "c", ",", "In", "practice", ",", "we", "found", "that", "two", "different", "CNNs", "resulted", "in", "better", "perplexity", "as", "well", "as", "BLEU", "compared", "to", "using", "a", "single", "one", "(", "§", "5.3", ")", ".", "We", "also", "found", "this", "to", "perform", "better", "than", "directly", "summing", "the", "e", "i", "without", "transformation", "as", "for", "the", "pooling", "model", ".", "section", ":", "Related", "Work", "There", "are", "several", "past", "attempts", "to", "use", "convolutional", "encoders", "for", "neural", "machine", "translation", ",", "however", ",", "to", "our", "knowledge", "none", "of", "them", "were", "able", "to", "match", "the", "performance", "of", "recurrent", "encoders", ".", "(", "Kalchbrenner", "and", "Blunsom", ",", "2013", ")", "introduce", "a", "convolutional", "sentence", "encoder", "in", "which", "a", "multi", "-", "layer", "CNN", "generates", "a", "fixed", "sized", "embedding", "for", "a", "source", "sentence", ",", "or", "an", "n", "-", "gram", "representation", "followed", "by", "transposed", "convolutions", "for", "directly", "generating", "a", "per", "-", "token", "decoder", "input", ".", "The", "latter", "requires", "the", "length", "of", "the", "translation", "prior", "to", "generation", "and", "both", "models", "were", "evaluated", "by", "rescoring", "the", "output", "of", "an", "existing", "translation", "system", ".", "[", "reference", "]", "propose", "a", "gated", "recursive", "CNN", "which", "is", "repeatedly", "applied", "until", "a", "fixed", "-", "size", "representation", "is", "obtained", "but", "the", "recurrent", "encoder", "achieves", "higher", "accuracy", ".", "In", "follow", "-", "up", "work", ",", "the", "authors", "improved", "the", "model", "via", "a", "soft", "-", "attention", "mechanism", "but", "did", "not", "reconsider", "convolutional", "encoder", "models", "[", "reference", "]", ".", "Concurrently", "to", "our", "work", ",", "[", "reference", "]", "have", "introduced", "convolutional", "translation", "models", "without", "an", "explicit", "attention", "mechanism", "but", "their", "approach", "does", "not", "yet", "result", "in", "state", "-", "ofthe", "-", "art", "accuracy", ".", "[", "reference", "]", ")", "also", "proposed", "a", "multi", "-", "layer", "CNN", "to", "generate", "a", "fixed", "-", "size", "encoder", "representation", "but", "their", "work", "lacks", "quantitative", "evaluation", "in", "terms", "of", "BLEU", ".", "[", "reference", "]", "and", "[", "reference", "]", "applied", "convolutional", "models", "to", "score", "phrase", "-", "pairs", "of", "traditional", "phrasebased", "and", "dependency", "-", "based", "translation", "models", ".", "Convolutional", "architectures", "have", "also", "been", "successful", "in", "language", "modeling", "but", "so", "far", "failed", "to", "outperform", "LSTMs", "[", "reference", "]", ".", "section", ":", "Experimental", "Setup", "section", ":", "Datasets", "We", "evaluate", "different", "encoders", "and", "ablate", "architectural", "choices", "on", "a", "small", "dataset", "from", "the", "GermanEnglish", "machine", "translation", "track", "of", "IWSLT", "2014", "(", "Cettolo", "et", "al", ".", ",", "2014", ")", "with", "a", "similar", "setting", "to", "[", "reference", "]", ".", "Unless", "otherwise", "stated", ",", "we", "restrict", "training", "sentences", "to", "have", "no", "more", "than", "175", "words", ";", "test", "sentences", "are", "not", "filtered", ".", "This", "is", "a", "higher", "threshold", "compared", "to", "other", "publications", "but", "ensures", "proper", "training", "of", "the", "position", "embeddings", "for", "non", "-", "recurrent", "encoders", ";", "the", "length", "threshold", "did", "not", "significantly", "effect", "recurrent", "encoders", ".", "Length", "filtering", "results", "in", "167", "K", "sentence", "pairs", "and", "we", "test", "on", "the", "concatenation", "of", "tst2010", ",", "tst2011", ",", "tst2012", ",", "tst2013", "and", "dev2010", "comprising", "6948", "sentence", "pairs", ".", "[", "reference", "]", "Our", "final", "results", "are", "on", "three", "major", "WMT", "tasks", ":", "WMT'16", "English", "-", "Romanian", ".", "We", "use", "the", "same", "data", "and", "pre", "-", "processing", "as", "[", "reference", "]", "and", "train", "on", "2.8", "M", "sentence", "pairs", ".", "[", "reference", "]", "Our", "model", "is", "word", "-", "based", "instead", "of", "relying", "on", "byte", "-", "pair", "encoding", "[", "reference", "]", ".", "We", "evaluate", "on", "newstest2016", ".", "WMT'15", "English", "-", "German", ".", "We", "use", "all", "available", "parallel", "training", "data", ",", "namely", "Europarl", "v7", ",", "Com", "-", "mon", "Crawl", "and", "News", "Commentary", "v10", "and", "apply", "the", "standard", "Moses", "tokenization", "to", "obtain", "3.9", "M", "sentence", "pairs", "[", "reference", "]", ".", "We", "report", "results", "on", "newstest2015", ".", "WMT'14", "English", "-", "French", ".", "We", "use", "a", "commonly", "used", "subset", "of", "12", "M", "sentence", "pairs", "[", "reference", "]", ",", "and", "remove", "sentences", "longer", "than", "150", "words", ".", "This", "results", "in", "10.7", "M", "sentence", "-", "pairs", "for", "training", ".", "Results", "are", "reported", "on", "ntst14", ".", "A", "small", "subset", "of", "the", "training", "data", "serves", "as", "validation", "set", "(", "5", "%", "for", "IWSLT'14", "and", "1", "%", "for", "WMT", ")", "for", "early", "stopping", "and", "learning", "rate", "annealing", "(", "§", "4.3", ")", ".", "For", "IWSLT'14", ",", "we", "replace", "words", "that", "occur", "fewer", "than", "3", "times", "with", "a", "<", "unk", ">", "symbol", ",", "which", "results", "in", "a", "vocabulary", "of", "24158", "English", "and", "35882", "German", "word", "types", ".", "For", "WMT", "datasets", ",", "we", "retain", "200", "K", "source", "and", "80", "K", "target", "words", ".", "For", "English", "-", "French", "only", ",", "we", "set", "the", "target", "vocabulary", "to", "30", "K", "types", "to", "be", "comparable", "with", "previous", "work", ".", "section", ":", "Model", "parameters", "We", "use", "512", "hidden", "units", "for", "both", "recurrent", "encoders", "and", "decoders", ".", "We", "reset", "the", "decoder", "hidden", "states", "to", "zero", "between", "sentences", ".", "For", "the", "convolutional", "encoder", ",", "512", "hidden", "units", "are", "used", "for", "each", "layer", "in", "CNN", "-", "a", ",", "while", "layers", "in", "CNN", "-", "c", "contain", "256", "units", "each", ".", "All", "embeddings", ",", "including", "the", "output", "produced", "by", "the", "decoder", "before", "the", "final", "linear", "layer", ",", "are", "of", "256", "dimensions", ".", "On", "the", "WMT", "corpora", ",", "we", "find", "that", "we", "can", "improve", "the", "performance", "of", "the", "bidirectional", "LSTM", "models", "(", "BiLSTM", ")", "by", "using", "512", "-", "dimensional", "word", "embeddings", ".", "Model", "weights", "are", "initialized", "from", "a", "uniform", "distribution", "within", "[", "−0.05", ",", "0.05", "]", ".", "For", "convolutional", "layers", ",", "we", "use", "a", "uniform", "distribution", "of", "−kd", "−0.5", ",", "kd", "−0.5", ",", "where", "k", "is", "the", "kernel", "width", "(", "we", "use", "3", "throughout", "this", "work", ")", "and", "d", "is", "the", "input", "size", "for", "the", "first", "layer", "and", "the", "number", "of", "hidden", "units", "for", "subsequent", "layers", "[", "reference", "]", ".", "For", "CNN", "-", "c", ",", "we", "transform", "the", "input", "and", "output", "with", "a", "linear", "layer", "each", "to", "match", "the", "smaller", "embedding", "size", ".", "The", "model", "parameters", "were", "tuned", "on", "IWSLT'14", "and", "cross", "-", "validated", "on", "the", "larger", "WMT", "corpora", ".", "section", ":", "Optimization", "Recurrent", "models", "are", "trained", "with", "Adam", "as", "we", "found", "them", "to", "benefit", "from", "aggressive", "optimization", ".", "We", "use", "a", "step", "width", "of", "3.125", "·", "10", "−4", "and", "early", "stopping", "based", "on", "validation", "perplexity", "(", "Kingma", "and", "Ba", ",", "2014", ")", ".", "For", "non", "-", "recurrent", "encoders", ",", "we", "obtain", "best", "results", "with", "stochastic", "gradient", "descent", "(", "SGD", ")", "and", "annealing", ":", "we", "use", "a", "learning", "rate", "of", "0.1", "and", "once", "the", "validation", "perplexity", "stops", "improving", ",", "we", "reduce", "the", "learning", "rate", "by", "an", "order", "of", "magnitude", "each", "epoch", "until", "it", "falls", "below", "10", "−4", ".", "For", "all", "models", ",", "we", "use", "mini", "-", "batches", "of", "32", "sentences", "for", "IWSLT'14", "and", "64", "for", "WMT", ".", "We", "use", "truncated", "back", "-", "propagation", "through", "time", "to", "limit", "the", "length", "of", "target", "sequences", "per", "mini", "-", "batch", "to", "25", "words", ".", "Gradients", "are", "normalized", "by", "the", "mini", "-", "batch", "size", ".", "We", "re", "-", "normalize", "the", "gradients", "if", "their", "norm", "exceeds", "25", "[", "reference", "]", ".", "Gradients", "of", "convolutional", "layers", "are", "scaled", "by", "sqrt", "(", "dim", "(", "input", ")", ")", "−1", "similar", "to", "[", "reference", "]", ".", "We", "use", "dropout", "on", "the", "embeddings", "and", "decoder", "outputs", "h", "i", "with", "a", "rate", "of", "0.2", "for", "IWSLT'14", "and", "0.1", "for", "WMT", "[", "reference", "]", ".", "All", "models", "are", "implemented", "in", "Torch", "[", "reference", "]", "and", "trained", "on", "a", "single", "GPU", ".", "section", ":", "Evaluation", "We", "report", "accuracy", "of", "single", "systems", "by", "training", "several", "identical", "models", "with", "different", "random", "seeds", "(", "5", "for", "IWSLT'14", ",", "3", "for", "WMT", ")", "and", "pick", "the", "one", "with", "the", "best", "validation", "perplexity", "for", "final", "BLEU", "evaluation", ".", "Translations", "are", "generated", "by", "a", "beam", "search", "and", "we", "normalize", "log", "-", "likelihood", "scores", "by", "sentence", "length", ".", "On", "IWSLT'14", "we", "use", "a", "beam", "width", "of", "10", "and", "for", "WMT", "models", "we", "tune", "beam", "width", "and", "word", "penalty", "on", "a", "separate", "test", "set", ",", "that", "is", "newsdev2016", "for", "WMT'16", "English", "-", "Romanian", ",", "newstest2014", "for", "WMT'15", "English", "-", "German", "and", "ntst1213", "for", "WMT'14", "English", "-", "French", ".", "[", "reference", "]", "The", "word", "penalty", "adds", "a", "constant", "factor", "to", "log", "-", "likelihoods", ",", "except", "for", "the", "end", "-", "of", "-", "sentence", "token", ".", "Prior", "to", "scoring", "the", "generated", "translations", "against", "the", "respective", "references", ",", "we", "perform", "unknown", "word", "replacement", "based", "on", "attention", "scores", "[", "reference", "]", ".", "Unknown", "words", "are", "replaced", "by", "looking", "up", "the", "source", "word", "with", "the", "maximum", "attention", "score", "in", "a", "pre", "-", "computed", "dictionary", ".", "If", "the", "dictionary", "contains", "no", "translation", ",", "then", "we", "simply", "copy", "the", "source", "word", ".", "Dictionaries", "were", "extracted", "from", "the", "aligned", "training", "data", "that", "was", "aligned", "with", "fast", "align", "[", "reference", "]", ".", "Each", "source", "word", "is", "mapped", "to", "the", "target", "word", "it", "is", "most", "frequently", "aligned", "to", ".", "For", "convolutional", "encoders", "with", "stacked", "CNN", "-", "c", "layers", "we", "noticed", "for", "some", "models", "that", "the", "attention", "maxima", "were", "consistently", "shifted", "by", "one", "word", ".", "We", "determine", "this", "per", "-", "model", "offset", "on", "the", "abovementioned", "development", "sets", "and", "correct", "for", "it", ".", "Finally", ",", "we", "compute", "case", "-", "sensitive", "tokenized", "BLEU", ",", "except", "for", "WMT'16", "English", "-", "Romanian", "where", "we", "use", "detokenized", "BLEU", "to", "be", "comparable", "with", "[", "reference", "]", "5", "Results", "section", ":", "Recurrent", "vs.", "Non", "-", "recurrent", "Encoders", "We", "first", "compare", "recurrent", "and", "non", "-", "recurrent", "encoders", "in", "terms", "of", "perplexity", "and", "BLEU", "on", "IWSLT'14", "with", "and", "without", "position", "embeddings", "(", "§", "3.1", ")", "and", "include", "a", "phrase", "-", "based", "system", "[", "reference", "]", ".", "Table", "1", "shows", "that", "a", "single", "-", "layer", "convolutional", "model", "with", "position", "embeddings", "(", "Convolutional", ")", "can", "outperform", "both", "a", "uni", "-", "directional", "LSTM", "encoder", "(", "LSTM", ")", "as", "well", "as", "a", "bi", "-", "directional", "LSTM", "encoder", "(", "BiLSTM", ")", ".", "Next", ",", "we", "increase", "the", "depth", "of", "the", "convolutional", "encoder", ".", "We", "choose", "a", "5", "Specifically", ",", "we", "select", "a", "beam", "from", "{", "5", ",", "10", "}", "and", "a", "word", "penalty", "from", "{", "0", ",", "−0.5", ",", "−1", ",", "−1.5", "}", "6", "https:", "//", "github.com", "/", "moses", "-", "smt", "/", "mosesdecoder", "/", "blob", "/", "617e8c8ed1630fb1d1", "/", "scripts", "/", "generic", "/", "{multi", "-", "bleu.perl", ",", "mteval", "-", "v13a.pl", "}", "good", "setting", "by", "independently", "varying", "the", "number", "of", "layers", "in", "CNN", "-", "a", "and", "CNN", "-", "c", "between", "1", "and", "10", "and", "obtained", "best", "validation", "set", "perplexity", "with", "six", "layers", "for", "CNN", "-", "a", "and", "three", "layers", "for", "CNN", "-", "c", ".", "This", "configuration", "outperforms", "BiLSTM", "by", "0.7", "BLEU", "(", "Deep", "Convolutional", "6", "/", "3", ")", ".", "We", "investigate", "depth", "in", "the", "convolutional", "encoder", "more", "in", "§", "5.3", ".", "Among", "recurrent", "encoders", ",", "the", "BiLSTM", "is", "2.3", "BLEU", "better", "than", "the", "uni", "-", "directional", "version", ".", "The", "simple", "pooling", "encoder", "which", "does", "not", "contain", "any", "parameters", "is", "only", "1.3", "BLEU", "lower", "than", "a", "unidirectional", "LSTM", "encoder", "and", "3.6", "BLEU", "lower", "than", "BiLSTM", ".", "The", "results", "without", "position", "embeddings", "(", "words", ")", "show", "that", "position", "information", "is", "crucial", "for", "convolutional", "encoders", ".", "In", "particular", "for", "shallow", "models", "(", "Pooling", "and", "Convolutional", ")", ",", "whereas", "deeper", "models", "are", "less", "effected", ".", "Recurrent", "encoders", "do", "not", "benefit", "from", "explicit", "position", "information", "because", "this", "information", "can", "be", "naturally", "extracted", "through", "the", "sequential", "computation", ".", "When", "tuning", "model", "settings", ",", "we", "generally", "observe", "good", "correlation", "between", "perplexity", "and", "BLEU", ".", "However", ",", "for", "convolutional", "encoders", "perplexity", "gains", "translate", "to", "smaller", "BLEU", "improvements", "compared", "to", "recurrent", "counterparts", "(", "Table", "1", ")", ".", "We", "observe", "a", "similar", "trend", "on", "larger", "datasets", ".", "section", ":", "Evaluation", "on", "WMT", "Corpora", "Next", ",", "we", "evaluate", "the", "BiLSTM", "encoder", "and", "the", "convolutional", "encoder", "architecture", "on", "three", "larger", "tasks", "and", "compare", "against", "previously", "published", "results", ".", "On", "WMT'16", "English", "-", "Romanian", "translation", "we", "compare", "to", "[", "reference", "]", ",", "the", "winning", "single", "system", "entry", "for", "this", "language", "pair", ".", "Their", "model", "consists", "of", "a", "bi", "-", "directional", "GRU", "encoder", ",", "a", "GRU", "decoder", "and", "MLP", "-", "based", "attention", ".", "Table", "2", ":", "Accuracy", "on", "three", "WMT", "tasks", ",", "including", "results", "published", "in", "previous", "work", ".", "For", "deep", "convolutional", "encoders", ",", "we", "include", "the", "number", "of", "layers", "in", "CNN", "-", "a", "and", "CNN", "-", "c", ",", "respectively", ".", "They", "use", "byte", "pair", "encoding", "(", "BPE", ")", "to", "achieve", "openvocabulary", "translation", "and", "dropout", "in", "all", "components", "of", "the", "neural", "network", "to", "achieve", "28.1", "BLEU", ";", "we", "use", "the", "same", "pre", "-", "processing", "but", "no", "BPE", "(", "§", "4", ")", ".", "The", "results", "(", "Table", "2", ")", "show", "that", "a", "deep", "convolutional", "encoder", "can", "perform", "competitively", "to", "the", "state", "of", "the", "art", "on", "this", "dataset", "[", "reference", "]", ".", "Our", "bi", "-", "directional", "LSTM", "encoder", "baseline", "is", "0.6", "BLEU", "lower", "than", "the", "state", "of", "the", "art", "but", "uses", "only", "512", "hidden", "units", "compared", "to", "1024", ".", "A", "singlelayer", "convolutional", "encoder", "with", "embedding", "size", "256", "performs", "at", "27.1", "BLEU", ".", "Increasing", "the", "number", "of", "convolutional", "layers", "to", "8", "in", "CNN", "-", "a", "and", "4", "in", "CNN", "-", "c", "achieves", "27.8", "BLEU", "which", "outperforms", "our", "baseline", "and", "is", "competitive", "to", "the", "state", "of", "the", "art", ".", "On", "WMT'15", "English", "to", "German", ",", "we", "compare", "to", "a", "BiLSTM", "baseline", "and", "prior", "work", ":", "[", "reference", "]", "introduce", "a", "large", "output", "vocabulary", ";", "the", "decoder", "of", "(", "Chung", "et", "al", ".", ",", "2016", ")", "operates", "on", "the", "character", "-", "level", ";", "[", "reference", "]", "uses", "LSTMs", "instead", "of", "GRUs", "and", "feeds", "the", "conditional", "input", "to", "the", "output", "layer", "as", "well", "as", "to", "the", "decoder", ".", "Our", "single", "-", "layer", "BiLSTM", "baseline", "is", "competitive", "to", "prior", "work", "and", "a", "two", "-", "layer", "BiLSTM", "encoder", "performs", "0.6", "BLEU", "better", "at", "24.1", "BLEU", ".", "Previous", "work", "also", "used", "multi", "-", "layer", "setups", ",", "e.g.", ",", "[", "reference", "]", "has", "two", "layers", "both", "in", "the", "encoder", "and", "the", "decoder", "with", "1024", "hidden", "units", ",", "and", "[", "reference", "]", "use", "1000", "hidden", "units", "per", "LSTM", ".", "We", "use", "512", "hidden", "units", "for", "both", "LSTM", "and", "convolutional", "encoders", ".", "Our", "convolutional", "model", "with", "15", "layers", "in", "CNN", "-", "a", "and", "5", "layers", "in", "CNN", "-", "c", "outperforms", "the", "BiLSTM", "encoder", "with", "both", "a", "single", "decoder", "layer", "or", "two", "decoder", "layers", ".", "Finally", ",", "we", "evaluate", "on", "the", "larger", "WMT'14", "English", "-", "French", "corpus", ".", "On", "this", "dataset", "the", "recurrent", "architectures", "benefit", "from", "an", "additional", "layer", "both", "in", "the", "encoder", "and", "the", "decoder", ".", "For", "a", "singlelayer", "decoder", ",", "a", "deep", "convolutional", "encoder", "outperforms", "the", "BiLSTM", "accuracy", "by", "0.3", "BLEU", "and", "for", "a", "two", "-", "layer", "decoder", ",", "our", "very", "deep", "convolutional", "encoder", "with", "up", "to", "20", "layers", "outperforms", "the", "BiLSTM", "by", "0.4", "BLEU", ".", "It", "has", "40", "%", "fewer", "parameters", "than", "the", "BiLSTM", "due", "to", "the", "smaller", "embedding", "sizes", ".", "We", "also", "outperform", "several", "previous", "systems", ",", "including", "the", "very", "deep", "encoder", "-", "decoder", "model", "proposed", "by", "[", "reference", "]", ".", "Our", "best", "result", "is", "just", "0.2", "BLEU", "below", "[", "reference", "]", "who", "use", "a", "very", "deep", "LSTM", "setup", "with", "a", "9", "-", "layer", "encoder", ",", "a", "7", "-", "layer", "decoder", ",", "shortcut", "connections", "and", "extensive", "regularization", "with", "dropout", "and", "L2", "regularization", ".", "section", ":", "Convolutional", "Encoder", "Architecture", "Details", "We", "next", "motivate", "our", "design", "of", "the", "convolutional", "encoder", "(", "§", "3.2", ")", ".", "We", "use", "the", "smaller", "IWSLT'14", "German", "-", "English", "setup", "without", "unknown", "word", "replacement", "to", "enable", "fast", "experimental", "turn", "-", "around", ".", "BLEU", "results", "are", "averaged", "over", "three", "training", "runs", "initialized", "with", "different", "seeds", ".", "Figure", "2", "shows", "accuracy", "for", "a", "different", "number", "of", "layers", "of", "both", "CNNs", "with", "and", "without", "residual", "connections", ".", "Our", "first", "observation", "is", "that", "computing", "the", "conditional", "input", "c", "i", "directly", "over", "embeddings", "e", "(", "line", "\"", "without", "CNN", "-", "c", "\"", ")", "is", "already", "working", "well", "at", "28.3", "BLEU", "with", "a", "single", "CNN", "-", "a", "layer", "and", "at", "29.1", "BLEU", "for", "CNN", "-", "a", "with", "7", "layers", "(", "Figure", "2a", ")", ".", "Increasing", "the", "number", "of", "CNN", "-", "c", "layers", "is", "beneficial", "up", "to", "three", "layers", "and", "beyond", "this", "we", "did", "not", "observe", "further", "improvements", ".", "Similarly", ",", "increasing", "the", "number", "of", "layers", "in", "CNN", "-", "a", "beyond", "six", "does", "not", "increase", "accuracy", "on", "this", "relatively", "small", "dataset", ".", "In", "general", ",", "choosing", "two", "to", "three", "times", "as", "many", "layers", "in", "CNN", "-", "a", "as", "in", "CNN", "-", "c", "is", "a", "good", "rule", "of", "thumb", ".", "Without", "residual", "connections", ",", "the", "model", "fails", "to", "utilize", "the", "increase", "in", "modeling", "power", "from", "additional", "layers", ",", "and", "performance", "drops", "significantly", "for", "deeper", "encoders", "(", "Figure", "2b", ")", ".", "Our", "convolutional", "architecture", "relies", "on", "two", "sets", "of", "networks", ",", "CNN", "-", "a", "for", "attention", "score", "computation", "a", "i", "and", "CNN", "-", "c", "for", "the", "conditional", "input", "c", "i", "to", "be", "fed", "to", "the", "decoder", ".", "We", "found", "that", "using", "the", "same", "network", "for", "both", "tasks", ",", "similar", "to", "recurrent", "encoders", ",", "resulted", "in", "poor", "accuracy", "of", "22.9", "BLEU", ".", "This", "compares", "to", "28.5", "BLEU", "for", "separate", "singlelayer", "networks", ",", "or", "28.3", "BLEU", "when", "aggregating", "embeddings", "for", "c", "i", ".", "Increasing", "the", "number", "of", "layers", "in", "the", "single", "network", "setup", "did", "not", "help", ".", "Figure", "2", "(", "a", ")", "suggests", "that", "the", "attention", "weights", "(", "CNN", "-", "a", ")", "need", "to", "integrate", "information", "from", "a", "wide", "context", "which", "can", "be", "done", "with", "a", "deep", "stack", ".", "At", "the", "same", "time", ",", "the", "vectors", "which", "are", "averaged", "(", "CNN", "-", "c", ")", "seem", "to", "benefit", "from", "a", "shallower", ",", "more", "local", "representation", "closer", "to", "the", "input", "words", ".", "Two", "stacks", "are", "an", "easy", "way", "to", "achieve", "these", "contradicting", "requirements", ".", "In", "Appendix", "A", "we", "visualize", "attention", "scores", "and", "find", "that", "alignments", "for", "CNN", "encoders", "are", "less", "sharp", "compared", "to", "BiLSTMs", ",", "however", ",", "this", "does", "not", "affect", "the", "effectiveness", "of", "unknown", "word", "replacement", "once", "we", "adjust", "for", "shifted", "maxima", ".", "In", "Appendix", "B", "we", "investigate", "whether", "deep", "convolutional", "encoders", "are", "required", "for", "translating", "long", "sentences", "and", "observe", "that", "even", "relatively", "shallow", "encoders", "perform", "well", "on", "long", "sentences", ".", "section", ":", "Training", "and", "Generation", "Speed", "For", "training", ",", "we", "use", "the", "fast", "CuDNN", "LSTM", "implementation", "for", "layers", "without", "attention", "and", "experiment", "on", "IWSLT'14", "with", "batch", "size", "32", ".", "The", "single", "-", "layer", "BiLSTM", "model", "trains", "at", "4300", "target", "words", "/", "second", ",", "while", "the", "6", "/", "3", "deep", "convolutional", "encoder", "compares", "at", "6400", "words", "/", "second", "on", "an", "NVidia", "Tesla", "M40", "GPU", ".", "We", "do", "not", "observe", "shorter", "overall", "training", "time", "since", "SGD", "converges", "slower", "than", "Adam", "which", "we", "use", "for", "BiLSTM", "models", ".", "We", "measure", "generation", "speed", "on", "an", "Intel", "Haswell", "CPU", "clocked", "at", "2.50GHz", "with", "a", "single", "thread", "for", "BLAS", "operations", ".", "We", "use", "vocabulary", "selection", "which", "can", "speed", "up", "generation", "by", "up", "to", "a", "factor", "of", "ten", "at", "no", "cost", "in", "accuracy", "via", "making", "the", "time", "to", "compute", "the", "final", "output", "layer", "negligible", "[", "reference", "][", "reference", "]", ".", "This", "shifts", "the", "focus", "from", "the", "efficiency", "of", "the", "encoder", "to", "the", "efficiency", "of", "the", "decoder", ".", "On", "IWSLT'14", "(", "Table", "3a", ")", "the", "convolutional", "encoder", "increases", "the", "speed", "of", "the", "overall", "model", "by", "a", "factor", "of", "1.35", "compared", "to", "the", "BiLSTM", "encoder", "while", "improving", "accuracy", "by", "0.7", "BLEU", ".", "In", "this", "setup", "both", "encoders", "models", "have", "the", "same", "hidden", "layer", "and", "embedding", "sizes", ".", "On", "the", "larger", "WMT'15", "English", "-", "German", "task", "(", "Table", "3b", ")", "the", "convolutional", "encoder", "speeds", "up", "generation", "by", "2.1", "times", "compared", "to", "a", "two", "-", "layer", "BiL", "-", "STM", ".", "This", "corresponds", "to", "231", "source", "words", "/", "second", "with", "beam", "size", "5", ".", "Our", "best", "model", "on", "this", "dataset", "generates", "203", "words", "/", "second", "but", "at", "slightly", "lower", "accuracy", "compared", "to", "the", "full", "vocabulary", "setting", "in", "Table", "2", ".", "The", "recurrent", "encoder", "uses", "larger", "embeddings", "than", "the", "convolutional", "encoder", "which", "were", "required", "for", "the", "models", "to", "match", "in", "accuracy", ".", "The", "smaller", "embedding", "size", "is", "not", "the", "only", "reason", "for", "the", "speed", "-", "up", ".", "In", "Table", "3a", "(", "a", ")", ",", "we", "compare", "a", "Conv", "6", "/", "3", "encoder", "and", "a", "BiLSTM", "with", "equal", "embedding", "sizes", ".", "The", "convolutional", "encoder", "is", "still", "1.34x", "faster", "(", "at", "0.7", "higher", "BLEU", ")", "although", "it", "requires", "roughly", "1.6x", "as", "many", "FLOPs", ".", "We", "believe", "that", "this", "is", "likely", "due", "to", "better", "cache", "locality", "for", "convolutional", "layers", "on", "CPUs", ":", "an", "LSTM", "with", "fused", "gates", "7", "requires", "two", "big", "matrix", "multiplications", "with", "different", "weights", "as", "well", "as", "additions", ",", "multiplications", "and", "non", "-", "linearities", "for", "each", "source", "word", ",", "while", "the", "output", "of", "each", "convolutional", "layer", "can", "be", "computed", "as", "whole", "with", "a", "single", "matrix", "multiply", ".", "For", "comparison", ",", "the", "quantized", "deep", "LSTM", "-", "7", "Our", "bi", "-", "directional", "LSTM", "implementation", "is", "based", "on", "torch", "rnnlib", "which", "uses", "fused", "LSTM", "gates", "(", "https:", "//", "github.com", "/", "facebookresearch", "/", "torch", "-", "rnnlib", "/", ")", "and", "which", "we", "consider", "an", "efficient", "implementation", ".", "section", ":", "Conclusion", "We", "introduced", "a", "simple", "encoder", "model", "for", "neural", "machine", "translation", "based", "on", "convolutional", "networks", ".", "This", "approach", "is", "more", "parallelizable", "than", "recurrent", "networks", "and", "provides", "a", "shorter", "path", "to", "capture", "long", "-", "range", "dependencies", "in", "the", "source", ".", "We", "find", "it", "essential", "to", "use", "source", "position", "embeddings", "as", "well", "as", "different", "CNNs", "for", "attention", "score", "computation", "and", "conditional", "input", "aggregation", ".", "Our", "experiments", "show", "that", "convolutional", "encoders", "perform", "on", "par", "or", "better", "than", "baselines", "based", "on", "bi", "-", "directional", "LSTM", "encoders", ".", "In", "comparison", "to", "other", "recent", "work", ",", "our", "deep", "convolutional", "encoder", "is", "competitive", "to", "the", "best", "published", "results", "to", "date", "(", "WMT'16", "English", "-", "Romanian", ")", "which", "are", "obtained", "with", "significantly", "more", "complex", "models", "(", "WMT'14", "English", "-", "French", ")", "or", "stem", "from", "improvements", "that", "are", "orthogonal", "to", "our", "work", "(", "WMT'15", "English", "-", "German", ")", ".", "Our", "architecture", "also", "leads", "to", "large", "generation", "speed", "improvements", ":", "translation", "models", "with", "our", "convolutional", "encoder", "can", "translate", "twice", "as", "fast", "as", "strong", "baselines", "with", "bi", "-", "directional", "recurrent", "encoders", ".", "Future", "work", "includes", "better", "training", "to", "enable", "faster", "convergence", "with", "the", "convolutional", "encoder", "to", "better", "leverage", "the", "higher", "processing", "speed", ".", "Our", "fast", "architecture", "is", "interesting", "for", "character", "level", "encoders", "where", "the", "input", "is", "significantly", "longer", "than", "for", "words", ".", "Also", ",", "we", "plan", "to", "investigate", "the", "effectiveness", "of", "our", "architecture", "on", "other", "sequence", "-", "tosequence", "tasks", ",", "e.g.", "summarization", ",", "constituency", "parsing", ",", "dialog", "modeling", ".", "section", ":", "A", "Alignment", "Visualization", "In", "Figure", "4", "and", "Figure", "5", ",", "we", "plot", "attention", "scores", "for", "a", "sample", "WMT'15", "English", "-", "German", "and", "WMT'14", "English", "-", "French", "translation", "with", "BiLSTM", "and", "deep", "convolutional", "encoders", ".", "The", "translation", "is", "on", "the", "x", "-", "axis", "and", "the", "source", "sentence", "on", "the", "y", "-", "axis", ".", "The", "attention", "scores", "of", "the", "BiLSTM", "output", "are", "sharp", "but", "do", "not", "necessarily", "represent", "a", "correct", "alignment", ".", "For", "CNN", "encoders", "the", "scores", "are", "less", "focused", "but", "still", "indicate", "an", "approximate", "source", "location", ",", "e.g.", ",", "in", "Figure", "4b", ",", "when", "moving", "the", "clause", "\"", "over", "1", ",", "000", "people", "were", "taken", "hostage", "\"", "to", "the", "back", "of", "the", "translation", ".", "For", "some", "models", ",", "attention", "maxima", "are", "consistently", "shifted", "by", "one", "token", "as", "both", "in", "Figure", "4b", "and", "Figure", "5b", ".", "Interestingly", ",", "convolutional", "encoders", "tend", "to", "focus", "on", "the", "last", "token", "(", "Figure", "4b", ")", "or", "both", "the", "first", "and", "last", "tokens", "(", "Figure", "5b", ")", ".", "Motivated", "by", "the", "hypothesis", "that", "the", "this", "may", "be", "due", "to", "the", "decoder", "depending", "on", "the", "length", "of", "the", "source", "sentence", "(", "which", "it", "can", "not", "determine", "without", "position", "embeddings", ")", ",", "we", "explicitly", "provided", "a", "distributed", "representation", "of", "the", "input", "length", "to", "the", "decoder", "and", "attention", "module", ".", "However", ",", "this", "did", "not", "cause", "a", "change", "in", "attention", "patterns", "nor", "did", "it", "improve", "translation", "accuracy", ".", "One", "characteristic", "of", "our", "convolutional", "encoder", "architecture", "is", "that", "the", "context", "over", "which", "outputs", "are", "computed", "depends", "on", "the", "number", "of", "layers", ".", "With", "bi", "-", "directional", "RNNs", ",", "every", "encoder", "output", "depends", "on", "the", "entire", "source", "sentence", ".", "In", "Figure", "3", ",", "we", "evaluate", "whether", "limited", "context", "affects", "the", "translation", "quality", "on", "longer", "sentences", "of", "WMT'15", "English", "-", "German", "which", "often", "requires", "moving", "verbs", "over", "long", "distances", ".", "We", "sort", "the", "newstest2015", "test", "set", "by", "source", "length", ",", "partition", "it", "into", "15", "equallysized", "buckets", ",", "and", "compare", "the", "BLEU", "scores", "of", "models", "listed", "in", "Table", "2", "on", "a", "per", "-", "bucket", "basis", ".", "section", ":", "B", "Performance", "by", "Sentence", "Length", "There", "is", "no", "clear", "evidence", "for", "sub", "-", "par", "translations", "on", "sentences", "that", "are", "longer", "than", "the", "observable", "context", "per", "encoder", "output", ".", "We", "include", "a", "small", "encoder", "with", "a", "6", "-", "layer", "CNN", "-", "c", "and", "a", "3", "-", "layer", "CNN", "-", "a", "in", "the", "comparison", "which", "performs", "worse", "than", "a", "2", "-", "layer", "BiLSTM", "(", "23.3", "BLEU", "vs23.6", ")", ".", "With", "6", "convolutional", "layers", "at", "kernel", "width", "3", ",", "each", "encoder", "output", "contains", "information", "of", "13", "adjacent", "source", "words", ".", "Looking", "at", "the", "accuracy", "for", "sentences", "with", "15", "words", "or", "more", ",", "this", "relatively", "shallow", "CNN", "is", "either", "on", "par", "or", "better", "than", "the", "BiLSTM", "for", "5", "out", "of", "10", "buckets", ";", "the", "BiLSTM", "has", "access", "to", "the", "entire", "source", "context", ".", "Similar", "observations", "can", "be", "made", "for", "the", "deeper", "convolutional", "encoders", ".", "section", ":"], "entity": [[1, 4, ["Convolutional", "Encoder", "Model"], "Method"], [5, 8, ["Neural", "Machine", "Translation"], "Task"], [15, 18, ["neural", "machine", "translation"], "Task"], [20, 24, ["bi", "-", "directional", "LSTMs"], "Method"], [40, 44, ["succession", "of", "convolutional", "layers"], "Method"], [55, 57, ["recurrent", "networks"], "Method"], [59, 60, ["computation"], "Task"], [67, 70, ["WMT'16", "EnglishRomanian", "translation"], "Task"], [73, 74, ["accuracy"], "Metric"], [85, 89, ["WMT'15", "English", "-", "German"], "Material"], [102, 103, ["accuracy"], "Metric"], [106, 109, ["deep", "LSTM", "setup"], "Method"], [119, 121, ["CPU", "decoding"], "Task"], [131, 132, ["accuracy"], "Metric"], [135, 137, ["bidirectional", "LSTM"], "Method"], [142, 145, ["Neural", "machine", "translation"], "Task"], [146, 147, ["NMT"], "Task"], [154, 156, ["end", "approach"], "Method"], [157, 159, ["machine", "translation"], "Task"], [172, 178, ["bi", "-", "directional", "recurrent", "neural", "network"], "Method"], [179, 180, ["RNN"], "Method"], [183, 186, ["variable", "length", "representation"], "Method"], [198, 199, ["RNN"], "Method"], [205, 209, ["soft", "-", "attention", "mechanism"], "Method"], [219, 221, ["Recurrent", "networks"], "Method"], [225, 230, ["long", "short", "term", "memory", "networks"], "Method"], [231, 232, ["LSTM"], "Method"], [237, 240, ["gated", "recurrent", "units"], "Method"], [275, 278, ["convolutional", "encoder", "models"], "Method"], [279, 284, ["neural", "machine", "trans", "-", "lation"], "Task"], [310, 312, ["recurrent", "alternatives"], "Method"], [323, 325, ["convolutional", "networks"], "Method"], [329, 331, ["convolutional", "networks"], "Method"], [358, 359, ["RNNs"], "Method"], [377, 381, ["succession", "of", "convolutional", "layers"], "Method"], [395, 396, ["RNNs"], "Method"], [403, 404, ["learning"], "Task"], [407, 410, ["tree", "-", "structure"], "Method"], [421, 424, ["recurrent", "neural", "network"], "Method"], [441, 442, ["processing"], "Task"], [458, 459, ["RNNs"], "Method"], [485, 487, ["convolutional", "layers"], "Method"], [491, 493, ["recurrent", "encoders"], "Method"], [497, 499, ["average", "pooling"], "Method"], [502, 504, ["parameterized", "convolutions"], "Method"], [508, 510, ["recurrent", "encoders"], "Method"], [513, 516, ["deep", "convolutional", "encoders"], "Method"], [539, 541, ["recurrent", "encoders"], "Method"], [543, 546, ["uni", "-", "directional"], "Method"], [547, 551, ["bi", "-", "directional", "LSTMs"], "Method"], [560, 561, ["accuracy"], "Metric"], [581, 585, ["WMT'15", "English", "-", "German"], "Material"], [602, 605, ["deep", "LSTM", "setup"], "Method"], [626, 627, ["forwards"], "Method"], [631, 634, ["stacked", "convolutional", "layers"], "Method"], [636, 638, ["n", "forwards"], "Method"], [640, 641, ["RNN"], "Method"], [680, 684, ["encoder", "-", "decoder", "approach"], "Method"], [717, 718, ["encoder"], "Method"], [739, 740, ["decoder"], "Method"], [742, 744, ["RNN", "network"], "Method"], [791, 792, ["LSTMs"], "Method"], [801, 803, ["decoder", "networks"], "Method"], [823, 824, ["LSTM"], "Method"], [835, 836, ["LSTM"], "Method"], [844, 846, ["translation", "model"], "Method"], [868, 870, ["linear", "layer"], "Method"], [892, 898, ["dot", "-", "product", "style", "attention", "mechanism"], "Method"], [914, 916, ["linear", "layer"], "Method"], [990, 991, ["softmax"], "Method"], [1018, 1019, ["BLEU"], "Metric"], [1020, 1021, ["perplexity"], "Metric"], [1027, 1031, ["dot", "-", "product", "attention"], "Method"], [1038, 1042, ["training", "and", "evaluation", "speed"], "Metric"], [1045, 1049, ["bi", "-", "directional", "LSTMs"], "Method"], [1051, 1053, ["recurrent", "encoders"], "Method"], [1064, 1066, ["WMT14", "EnglishFrench"], "Material"], [1181, 1183, ["linear", "layer"], "Method"], [1189, 1191, ["encoder", "architecture"], "Method"], [1192, 1193, ["BiLSTM"], "Method"], [1195, 1199, ["Non", "-", "recurrent", "Encoders"], "Method"], [1201, 1203, ["Pooling", "Encoder"], "Method"], [1207, 1211, ["non", "-", "recurrent", "encoders"], "Method"], [1213, 1215, ["pooling", "model"], "Method"], [1230, 1233, ["Averaging", "word", "embeddings"], "Method"], [1294, 1296, ["Position", "embeddings"], "Method"], [1302, 1304, ["memory", "networks"], "Task"], [1305, 1308, ["question", "-", "answering"], "Task"], [1309, 1311, ["language", "modeling"], "Task"], [1318, 1320, ["recurrent", "encoder"], "Method"], [1334, 1336, ["pooled", "representations"], "Method"], [1369, 1370, ["pooling"], "Task"], [1397, 1399, ["Convolutional", "Encoder"], "Method"], [1403, 1404, ["pooling"], "Method"], [1408, 1409, ["kernel"], "Method"], [1411, 1414, ["convolutional", "neural", "network"], "Method"], [1415, 1416, ["CNN"], "Method"], [1452, 1455, ["layers", "of", "convolutions"], "Method"], [1521, 1522, ["encoder"], "Method"], [1540, 1541, ["learning"], "Task"], [1542, 1544, ["deep", "encoders"], "Task"], [1605, 1606, ["CNNs"], "Method"], [1609, 1611, ["pooling", "layers"], "Method"], [1616, 1619, ["down", "-", "sampling"], "Task"], [1640, 1642, ["pooling", "model"], "Method"], [1644, 1646, ["convolutional", "encoder"], "Method"], [1647, 1649, ["position", "embeddings"], "Method"], [1652, 1653, ["encoder"], "Method"], [1664, 1665, ["CNN"], "Method"], [1697, 1700, ["CNN", "-", "c"], "Method"], [1709, 1710, ["CNNs"], "Method"], [1713, 1714, ["perplexity"], "Metric"], [1717, 1718, ["BLEU"], "Metric"], [1747, 1749, ["pooling", "model"], "Method"], [1761, 1763, ["convolutional", "encoders"], "Method"], [1764, 1767, ["neural", "machine", "translation"], "Task"], [1783, 1785, ["recurrent", "encoders"], "Method"], [1795, 1798, ["convolutional", "sentence", "encoder"], "Method"], [1807, 1810, ["fixed", "sized", "embedding"], "Method"], [1817, 1821, ["n", "-", "gram", "representation"], "Method"], [1823, 1825, ["transposed", "convolutions"], "Method"], [1845, 1846, ["generation"], "Task"], [1858, 1860, ["translation", "system"], "Method"], [1875, 1879, ["fixed", "-", "size", "representation"], "Method"], [1883, 1885, ["recurrent", "encoder"], "Method"], [1887, 1888, ["accuracy"], "Metric"], [1902, 1906, ["soft", "-", "attention", "mechanism"], "Method"], [1910, 1913, ["convolutional", "encoder", "models"], "Method"], [1927, 1930, ["convolutional", "translation", "models"], "Method"], [1933, 1935, ["attention", "mechanism"], "Method"], [1948, 1949, ["accuracy"], "Metric"], [1964, 1969, ["fixed", "-", "size", "encoder", "representation"], "Method"], [1973, 1975, ["quantitative", "evaluation"], "Metric"], [1978, 1979, ["BLEU"], "Metric"], [1988, 1990, ["convolutional", "models"], "Method"], [1997, 2004, ["phrasebased", "and", "dependency", "-", "based", "translation", "models"], "Method"], [2005, 2007, ["Convolutional", "architectures"], "Method"], [2012, 2014, ["language", "modeling"], "Task"], [2020, 2021, ["LSTMs"], "Method"], [2035, 2036, ["encoders"], "Method"], [2046, 2050, ["GermanEnglish", "machine", "translation", "track"], "Material"], [2051, 2053, ["IWSLT", "2014"], "Material"], [2110, 2114, ["non", "-", "recurrent", "encoders"], "Method"], [2122, 2124, ["recurrent", "encoders"], "Method"], [2125, 2127, ["Length", "filtering"], "Task"], [2140, 2141, ["tst2010"], "Material"], [2142, 2143, ["tst2011"], "Material"], [2144, 2145, ["tst2012"], "Material"], [2146, 2147, ["tst2013"], "Material"], [2148, 2149, ["dev2010"], "Material"], [2167, 2171, ["WMT'16", "English", "-", "Romanian"], "Material"], [2199, 2202, ["word", "-", "based"], "Method"], [2206, 2210, ["byte", "-", "pair", "encoding"], "Method"], [2233, 2235, ["Europarl", "v7"], "Material"], [2236, 2240, ["Com", "-", "mon", "Crawl"], "Method"], [2241, 2243, ["News", "Commentary"], "Method"], [2248, 2250, ["Moses", "tokenization"], "Method"], [2266, 2270, ["WMT'14", "English", "-", "French"], "Material"], [2326, 2327, ["IWSLT'14"], "Material"], [2331, 2332, ["WMT"], "Material"], [2334, 2336, ["early", "stopping"], "Task"], [2337, 2340, ["learning", "rate", "annealing"], "Task"], [2346, 2347, ["IWSLT'14"], "Material"], [2394, 2397, ["English", "-", "French"], "Material"], [2426, 2428, ["recurrent", "encoders"], "Method"], [2429, 2430, ["decoders"], "Method"], [2444, 2446, ["convolutional", "encoder"], "Method"], [2456, 2459, ["CNN", "-", "a"], "Method"], [2463, 2466, ["CNN", "-", "c"], "Method"], [2480, 2481, ["decoder"], "Method"], [2484, 2486, ["linear", "layer"], "Method"], [2515, 2520, ["512", "-", "dimensional", "word", "embeddings"], "Method"], [2590, 2593, ["CNN", "-", "c"], "Method"], [2602, 2604, ["linear", "layer"], "Method"], [2618, 2619, ["IWSLT'14"], "Material"], [2631, 2632, ["Optimization"], "Task"], [2632, 2634, ["Recurrent", "models"], "Method"], [2637, 2638, ["Adam"], "Method"], [2645, 2647, ["aggressive", "optimization"], "Method"], [2663, 2665, ["validation", "perplexity"], "Method"], [2674, 2678, ["non", "-", "recurrent", "encoders"], "Method"], [2684, 2687, ["stochastic", "gradient", "descent"], "Method"], [2688, 2689, ["SGD"], "Method"], [2691, 2692, ["annealing"], "Method"], [2696, 2698, ["learning", "rate"], "Metric"], [2703, 2705, ["validation", "perplexity"], "Metric"], [2711, 2713, ["learning", "rate"], "Metric"], [2740, 2741, ["IWSLT'14"], "Material"], [2744, 2745, ["WMT"], "Material"], [2748, 2752, ["truncated", "back", "-", "propagation"], "Method"], [2817, 2818, ["dropout"], "Method"], [2832, 2833, ["IWSLT'14"], "Material"], [2836, 2837, ["WMT"], "Material"], [2846, 2847, ["Torch"], "Method"], [2855, 2856, ["GPU"], "Method"], [2862, 2863, ["accuracy"], "Metric"], [2878, 2879, ["IWSLT'14"], "Material"], [2882, 2883, ["WMT"], "Material"], [2891, 2893, ["validation", "perplexity"], "Metric"], [2895, 2897, ["BLEU", "evaluation"], "Metric"], [2903, 2905, ["beam", "search"], "Method"], [2917, 2918, ["IWSLT'14"], "Material"], [2946, 2950, ["WMT'16", "English", "-", "Romanian"], "Material"], [2951, 2952, ["newstest2014"], "Material"], [2953, 2957, ["WMT'15", "English", "-", "German"], "Material"], [2960, 2964, ["WMT'14", "English", "-", "French"], "Material"], [2969, 2971, ["word", "penalty"], "Method"], [3086, 3088, ["convolutional", "encoders"], "Method"], [3139, 3143, ["WMT'16", "English", "-", "Romanian"], "Material"], [3159, 3165, ["Recurrent", "vs.", "Non", "-", "recurrent", "Encoders"], "Method"], [3168, 3174, ["recurrent", "and", "non", "-", "recurrent", "encoders"], "Method"], [3177, 3178, ["perplexity"], "Metric"], [3179, 3180, ["BLEU"], "Metric"], [3181, 3182, ["IWSLT'14"], "Material"], [3194, 3198, ["phrase", "-", "based", "system"], "Method"], [3207, 3212, ["single", "-", "layer", "convolutional", "model"], "Method"], [3213, 3215, ["position", "embeddings"], "Method"], [3216, 3217, ["Convolutional"], "Method"], [3222, 3227, ["uni", "-", "directional", "LSTM", "encoder"], "Method"], [3228, 3229, ["LSTM"], "Method"], [3240, 3241, ["BiLSTM"], "Method"], [3251, 3253, ["convolutional", "encoder"], "Method"], [3321, 3324, ["CNN", "-", "a"], "Method"], [3325, 3328, ["CNN", "-", "c"], "Method"], [3335, 3338, ["validation", "set", "perplexity"], "Metric"], [3342, 3345, ["CNN", "-", "a"], "Method"], [3349, 3352, ["CNN", "-", "c"], "Method"], [3356, 3357, ["BiLSTM"], "Method"], [3359, 3360, ["BLEU"], "Metric"], [3361, 3366, ["Deep", "Convolutional", "6", "/", "3"], "Method"], [3373, 3375, ["convolutional", "encoder"], "Method"], [3381, 3383, ["recurrent", "encoders"], "Method"], [3385, 3386, ["BiLSTM"], "Method"], [3388, 3389, ["BLEU"], "Metric"], [3399, 3401, ["pooling", "encoder"], "Method"], [3410, 3411, ["BLEU"], "Metric"], [3414, 3417, ["unidirectional", "LSTM", "encoder"], "Method"], [3419, 3420, ["BLEU"], "Metric"], [3422, 3423, ["BiLSTM"], "Method"], [3439, 3441, ["convolutional", "encoders"], "Method"], [3445, 3447, ["shallow", "models"], "Method"], [3448, 3449, ["Pooling"], "Method"], [3450, 3451, ["Convolutional"], "Method"], [3454, 3456, ["deeper", "models"], "Method"], [3460, 3462, ["Recurrent", "encoders"], "Method"], [3478, 3480, ["sequential", "computation"], "Method"], [3490, 3491, ["correlation"], "Metric"], [3492, 3493, ["perplexity"], "Metric"], [3494, 3495, ["BLEU"], "Metric"], [3499, 3502, ["convolutional", "encoders", "perplexity"], "Method"], [3510, 3512, ["recurrent", "counterparts"], "Method"], [3541, 3544, ["convolutional", "encoder", "architecture"], "Method"], [3583, 3588, ["bi", "-", "directional", "GRU", "encoder"], "Method"], [3590, 3592, ["GRU", "decoder"], "Method"], [3593, 3597, ["MLP", "-", "based", "attention"], "Method"], [3601, 3602, ["Accuracy"], "Metric"], [3615, 3618, ["deep", "convolutional", "encoders"], "Method"], [3626, 3629, ["CNN", "-", "a"], "Method"], [3630, 3633, ["CNN", "-", "c"], "Method"], [3638, 3641, ["byte", "pair", "encoding"], "Method"], [3642, 3643, ["BPE"], "Method"], [3646, 3648, ["openvocabulary", "translation"], "Task"], [3655, 3657, ["neural", "network"], "Method"], [3660, 3661, ["BLEU"], "Metric"], [3666, 3669, ["pre", "-", "processing"], "Method"], [3671, 3672, ["BPE"], "Method"], [3686, 3689, ["deep", "convolutional", "encoder"], "Method"], [3714, 3715, ["BLEU"], "Metric"], [3733, 3736, ["singlelayer", "convolutional", "encoder"], "Method"], [3743, 3744, ["BLEU"], "Metric"], [3754, 3757, ["CNN", "-", "a"], "Method"], [3760, 3763, ["CNN", "-", "c"], "Method"], [3765, 3766, ["BLEU"], "Metric"], [3781, 3785, ["WMT'15", "English", "to", "German"], "Material"], [3827, 3828, ["LSTMs"], "Method"], [3830, 3831, ["GRUs"], "Method"], [3867, 3868, ["BLEU"], "Metric"], [3871, 3872, ["BLEU"], "Metric"], [3893, 3894, ["encoder"], "Method"], [3896, 3897, ["decoder"], "Method"], [3911, 3912, ["LSTM"], "Method"], [3926, 3928, ["convolutional", "model"], "Method"], [3932, 3935, ["CNN", "-", "a"], "Method"], [3939, 3942, ["CNN", "-", "c"], "Method"], [3950, 3952, ["decoder", "layer"], "Method"], [3954, 3956, ["decoder", "layers"], "Method"], [3974, 3976, ["recurrent", "architectures"], "Method"], [3984, 3985, ["encoder"], "Method"], [3987, 3988, ["decoder"], "Method"], [3991, 3993, ["singlelayer", "decoder"], "Method"], [3995, 3998, ["deep", "convolutional", "encoder"], "Method"], [4004, 4005, ["BLEU"], "Metric"], [4008, 4012, ["two", "-", "layer", "decoder"], "Method"], [4015, 4018, ["deep", "convolutional", "encoder"], "Method"], [4025, 4026, ["BiLSTM"], "Method"], [4028, 4029, ["BLEU"], "Metric"], [4038, 4039, ["BiLSTM"], "Method"], [4055, 4061, ["very", "deep", "encoder", "-", "decoder", "model"], "Method"], [4073, 4074, ["BLEU"], "Metric"], [4082, 4085, ["deep", "LSTM", "setup"], "Method"], [4087, 4091, ["9", "-", "layer", "encoder"], "Method"], [4093, 4097, ["7", "-", "layer", "decoder"], "Method"], [4098, 4100, ["shortcut", "connections"], "Method"], [4102, 4103, ["regularization"], "Method"], [4104, 4108, ["dropout", "and", "L2", "regularization"], "Method"], [4111, 4114, ["Convolutional", "Encoder", "Architecture"], "Method"], [4122, 4124, ["convolutional", "encoder"], "Method"], [4150, 4151, ["BLEU"], "Metric"], [4166, 4167, ["accuracy"], "Metric"], [4175, 4176, ["CNNs"], "Method"], [4201, 4204, ["CNN", "-", "c"], "Method"], [4212, 4213, ["BLEU"], "Metric"], [4223, 4224, ["BLEU"], "Metric"], [4225, 4228, ["CNN", "-", "a"], "Method"], [4268, 4271, ["CNN", "-", "a"], "Method"], [4276, 4277, ["accuracy"], "Metric"], [4295, 4298, ["CNN", "-", "a"], "Method"], [4300, 4303, ["CNN", "-", "c"], "Method"], [4341, 4343, ["convolutional", "architecture"], "Method"], [4350, 4353, ["CNN", "-", "a"], "Method"], [4354, 4357, ["attention", "score", "computation"], "Task"], [4360, 4363, ["CNN", "-", "c"], "Method"], [4374, 4375, ["decoder"], "Method"], [4389, 4391, ["recurrent", "encoders"], "Method"], [4395, 4396, ["accuracy"], "Metric"], [4398, 4399, ["BLEU"], "Metric"], [4404, 4405, ["BLEU"], "Metric"], [4407, 4409, ["singlelayer", "networks"], "Method"], [4412, 4413, ["BLEU"], "Metric"], [4445, 4448, ["CNN", "-", "a"], "Method"], [4463, 4465, ["deep", "stack"], "Method"], [4477, 4480, ["CNN", "-", "c"], "Method"], [4521, 4523, ["CNN", "encoders"], "Method"], [4528, 4529, ["BiLSTMs"], "Method"], [4539, 4542, ["unknown", "word", "replacement"], "Task"], [4555, 4558, ["deep", "convolutional", "encoders"], "Method"], [4569, 4571, ["shallow", "encoders"], "Method"], [4581, 4583, ["Generation", "Speed"], "Metric"], [4584, 4585, ["training"], "Task"], [4600, 4601, ["IWSLT'14"], "Material"], [4622, 4628, ["6", "/", "3", "deep", "convolutional", "encoder"], "Method"], [4647, 4649, ["training", "time"], "Metric"], [4650, 4651, ["SGD"], "Method"], [4654, 4655, ["Adam"], "Method"], [4664, 4666, ["generation", "speed"], "Metric"], [4684, 4686, ["vocabulary", "selection"], "Method"], [4690, 4691, ["generation"], "Task"], [4702, 4703, ["accuracy"], "Metric"], [4729, 4730, ["encoder"], "Method"], [4735, 4736, ["decoder"], "Method"], [4738, 4739, ["IWSLT'14"], "Material"], [4744, 4746, ["convolutional", "encoder"], "Method"], [4765, 4766, ["accuracy"], "Metric"], [4768, 4769, ["BLEU"], "Metric"], [4774, 4776, ["encoders", "models"], "Method"], [4798, 4800, ["convolutional", "encoder"], "Method"], [4802, 4803, ["generation"], "Task"], [4811, 4815, ["layer", "BiL", "-", "STM"], "Method"], [4844, 4845, ["accuracy"], "Metric"], [4856, 4858, ["recurrent", "encoder"], "Method"], [4863, 4865, ["convolutional", "encoder"], "Method"], [4874, 4875, ["accuracy"], "Metric"], [4878, 4880, ["embedding", "size"], "Metric"], [4901, 4906, ["Conv", "6", "/", "3", "encoder"], "Method"], [4908, 4909, ["BiLSTM"], "Method"], [4915, 4917, ["convolutional", "encoder"], "Method"], [4925, 4926, ["BLEU"], "Metric"], [4948, 4950, ["convolutional", "layers"], "Method"], [4951, 4952, ["CPUs"], "Method"], [4962, 4964, ["matrix", "multiplications"], "Method"], [4987, 4989, ["convolutional", "layer"], "Method"], [5018, 5020, ["torch", "rnnlib"], "Method"], [5052, 5054, ["encoder", "model"], "Method"], [5055, 5058, ["neural", "machine", "translation"], "Task"], [5060, 5062, ["convolutional", "networks"], "Method"], [5069, 5071, ["recurrent", "networks"], "Method"], [5099, 5100, ["CNNs"], "Method"], [5101, 5104, ["attention", "score", "computation"], "Task"], [5105, 5108, ["conditional", "input", "aggregation"], "Task"], [5113, 5115, ["convolutional", "encoders"], "Method"], [5138, 5141, ["deep", "convolutional", "encoder"], "Method"], [5151, 5155, ["WMT'16", "English", "-", "Romanian"], "Material"], [5165, 5169, ["WMT'14", "English", "-", "French"], "Material"], [5181, 5185, ["WMT'15", "English", "-", "German"], "Material"], [5193, 5196, ["generation", "speed", "improvements"], "Metric"], [5197, 5199, ["translation", "models"], "Method"], [5201, 5203, ["convolutional", "encoder"], "Method"], [5212, 5217, ["bi", "-", "directional", "recurrent", "encoders"], "Method"], [5229, 5231, ["convolutional", "encoder"], "Method"], [5236, 5238, ["processing", "speed"], "Metric"], [5240, 5242, ["fast", "architecture"], "Method"], [5245, 5248, ["character", "level", "encoders"], "Task"], [5271, 5275, ["sequence", "-", "tosequence", "tasks"], "Task"], [5276, 5278, ["e.g.", "summarization"], "Task"], [5279, 5281, ["constituency", "parsing"], "Task"], [5282, 5284, ["dialog", "modeling"], "Task"], [5288, 5290, ["Alignment", "Visualization"], "Task"], [5299, 5301, ["attention", "scores"], "Metric"], [5304, 5308, ["WMT'15", "English", "-", "German"], "Material"], [5309, 5313, ["WMT'14", "English", "-", "French"], "Material"], [5315, 5316, ["BiLSTM"], "Method"], [5317, 5320, ["deep", "convolutional", "encoders"], "Method"], [5344, 5345, ["BiLSTM"], "Method"], [5358, 5360, ["CNN", "encoders"], "Method"], [5423, 5425, ["convolutional", "encoders"], "Method"], [5460, 5461, ["decoder"], "Method"], [5484, 5486, ["distributed", "representation"], "Method"], [5492, 5496, ["decoder", "and", "attention", "module"], "Method"], [5512, 5514, ["translation", "accuracy"], "Metric"], [5519, 5522, ["convolutional", "encoder", "architecture"], "Method"], [5539, 5543, ["bi", "-", "directional", "RNNs"], "Method"], [5565, 5567, ["translation", "quality"], "Metric"], [5571, 5575, ["WMT'15", "English", "-", "German"], "Material"], [5624, 5626, ["Sentence", "Length"], "Metric"], [5653, 5654, ["encoder"], "Method"], [5684, 5685, ["BLEU"], "Metric"], [5690, 5692, ["convolutional", "layers"], "Method"], [5711, 5712, ["accuracy"], "Metric"], [5732, 5733, ["BiLSTM"], "Method"], [5741, 5742, ["BiLSTM"], "Method"], [5757, 5760, ["deeper", "convolutional", "encoders"], "Method"], [110, 114, ["WMT'14", "English", "-", "French"], "Material"], [553, 557, ["WMT'16", "English", "-", "Romanian"], "Material"], [594, 598, ["WMT'14", "English", "-", "French"], "Material"], [666, 669, ["Neural", "Machine", "Translation"], "Task"], [862, 863, ["LSTM"], "Method"], [1109, 1112, ["uni", "-", "directional"], "Method"], [1591, 1592, ["CNNs"], "Method"], [1657, 1659, ["convolutional", "networks"], "Method"], [1804, 1805, ["CNN"], "Method"], [1868, 1869, ["CNN"], "Method"], [1960, 1961, ["CNN"], "Method"], [2164, 2165, ["WMT"], "Material"], [2178, 2181, ["pre", "-", "processing"], "Method"], [2219, 2223, ["WMT'15", "English", "-", "German"], "Material"], [2379, 2380, ["WMT"], "Material"], [2494, 2495, ["WMT"], "Material"], [2507, 2509, ["bidirectional", "LSTM"], "Method"], [2511, 2512, ["BiLSTM"], "Method"], [2537, 2539, ["convolutional", "layers"], "Method"], [2626, 2627, ["WMT"], "Material"], [2796, 2798, ["convolutional", "layers"], "Method"], [2927, 2928, ["WMT"], "Material"], [3090, 3091, ["CNN"], "Method"], [3135, 3136, ["BLEU"], "Metric"], [3147, 3148, ["BLEU"], "Metric"], [3237, 3238, ["LSTM"], "Method"], [3392, 3395, ["uni", "-", "directional"], "Method"], [3506, 3507, ["BLEU"], "Metric"], [3530, 3531, ["WMT"], "Material"], [3537, 3538, ["BiLSTM"], "Method"], [3556, 3560, ["WMT'16", "English", "-", "Romanian"], "Material"], [3604, 3605, ["WMT"], "Material"], [3709, 3710, ["LSTM"], "Method"], [3749, 3751, ["convolutional", "layers"], "Method"], [3790, 3791, ["BiLSTM"], "Method"], [3851, 3852, ["BiLSTM"], "Method"], [3863, 3864, ["BiLSTM"], "Method"], [3920, 3921, ["LSTM"], "Method"], [3944, 3945, ["BiLSTM"], "Method"], [3964, 3968, ["WMT'14", "English", "-", "French"], "Material"], [4000, 4001, ["BiLSTM"], "Method"], [4133, 4134, ["IWSLT'14"], "Material"], [4216, 4217, ["CNN"], "Method"], [4240, 4241, ["CNN"], "Method"], [4591, 4592, ["LSTM"], "Method"], [4610, 4611, ["BiLSTM"], "Method"], [4659, 4660, ["BiLSTM"], "Method"], [4761, 4762, ["BiLSTM"], "Method"], [4788, 4792, ["WMT'15", "English", "-", "German"], "Material"], [4954, 4955, ["LSTM"], "Method"], [5006, 5007, ["LSTM"], "Method"], [5013, 5014, ["LSTM"], "Method"], [5023, 5024, ["LSTM"], "Method"], [5127, 5128, ["LSTM"], "Method"], [5604, 5605, ["BLEU"], "Metric"], [5659, 5660, ["CNN"], "Method"], [5667, 5668, ["CNN"], "Method"], [5681, 5682, ["BiLSTM"], "Method"], [5723, 5724, ["CNN"], "Method"]]} {"doc_id": "4592364a6c1eb227ffb3cbb6bb9f4e6431e89f2f", "text": ["document", ":", "Learning", "to", "Compare", ":", "Relation", "Network", "for", "Few", "-", "Shot", "Learning", "We", "present", "a", "conceptually", "simple", ",", "flexible", ",", "and", "general", "framework", "for", "few", "-", "shot", "learning", ",", "where", "a", "classifier", "must", "learn", "to", "recognise", "new", "classes", "given", "only", "few", "examples", "from", "each", ".", "Our", "method", ",", "called", "the", "Relation", "Network", "(", "RN", ")", ",", "is", "trained", "end", "-", "to", "-", "end", "from", "scratch", ".", "During", "meta", "-", "learning", ",", "it", "learns", "to", "learn", "a", "deep", "distance", "metric", "to", "compare", "a", "small", "number", "of", "images", "within", "episodes", ",", "each", "of", "which", "is", "designed", "to", "simulate", "the", "few", "-", "shot", "setting", ".", "Once", "trained", ",", "a", "RN", "is", "able", "to", "classify", "images", "of", "new", "classes", "by", "computing", "relation", "scores", "between", "query", "images", "and", "the", "few", "examples", "of", "each", "new", "class", "without", "further", "updating", "the", "network", ".", "Besides", "providing", "improved", "performance", "on", "few", "-", "shot", "learning", ",", "our", "framework", "is", "easily", "extended", "to", "zero", "-", "shot", "learning", ".", "Extensive", "experiments", "on", "five", "benchmarks", "demonstrate", "that", "our", "simple", "approach", "provides", "a", "unified", "and", "effective", "approach", "for", "both", "of", "these", "two", "tasks", ".", "section", ":", "Introduction", "Deep", "learning", "models", "have", "achieved", "great", "success", "in", "visual", "recognition", "tasks", ".", "However", ",", "these", "supervised", "learning", "models", "need", "large", "amounts", "of", "labelled", "data", "and", "many", "iterations", "to", "train", "their", "large", "number", "of", "parameters", ".", "This", "severely", "limits", "their", "scalability", "to", "new", "classes", "due", "to", "annotation", "cost", ",", "but", "more", "fundamentally", "limits", "their", "applicability", "to", "newly", "emerging", "(", "eg", ".", "new", "consumer", "devices", ")", "or", "rare", "(", "eg", ".", "rare", "animals", ")", "categories", "where", "numerous", "annotated", "images", "may", "simply", "never", "exist", ".", "In", "contrast", ",", "humans", "are", "very", "good", "at", "recognising", "objects", "with", "very", "little", "direct", "supervision", ",", "or", "none", "at", "all", ",", "few", "-", "shot", "or", "zero", "-", "shot", "learning", ".", "For", "example", ",", "children", "have", "no", "problem", "generalising", "the", "concept", "of", "“", "zebra", "”", "from", "a", "single", "picture", "in", "a", "book", ",", "or", "hearing", "its", "description", "as", "looking", "like", "a", "stripy", "horse", ".", "Motivated", "by", "the", "failure", "of", "conventional", "deep", "learning", "methods", "to", "work", "well", "on", "one", "or", "few", "examples", "per", "class", ",", "and", "inspired", "by", "the", "few", "-", "and", "zero", "-", "shot", "learning", "ability", "of", "humans", ",", "there", "has", "been", "a", "recent", "resurgence", "of", "interest", "in", "machine", "one", "/", "few", "-", "shot", "and", "zero", "-", "shot", "learning", ".", "Few", "-", "shot", "learning", "aims", "to", "recognise", "novel", "visual", "categories", "from", "very", "few", "labelled", "examples", ".", "The", "availability", "of", "only", "one", "or", "very", "few", "examples", "challenges", "the", "standard", "‘", "fine", "-", "tuning", "’", "practice", "in", "deep", "learning", ".", "Data", "augmentation", "and", "regularisation", "techniques", "can", "alleviate", "overfitting", "in", "such", "a", "limited", "-", "data", "regime", ",", "but", "they", "do", "not", "solve", "it", ".", "Therefore", "contemporary", "approaches", "to", "few", "-", "shot", "learning", "often", "decompose", "training", "into", "an", "auxiliary", "meta", "learning", "phase", "where", "transferrable", "knowledge", "is", "learned", "in", "the", "form", "of", "good", "initial", "conditions", ",", "embeddings", "or", "optimisation", "strategies", ".", "The", "target", "few", "-", "shot", "learning", "problem", "is", "then", "learned", "by", "fine", "-", "tuning", "with", "the", "learned", "optimisation", "strategy", "or", "computed", "in", "a", "feed", "-", "forward", "pass", "without", "updating", "network", "weights", ".", "Zero", "-", "shot", "learning", "also", "suffers", "from", "a", "related", "challenge", ".", "Recognisers", "are", "trained", "by", "a", "single", "example", "in", "the", "form", "of", "a", "class", "description", "(", "c.f", ".", ",", "single", "exemplar", "image", "in", "one", "-", "shot", ")", ",", "making", "data", "insufficiency", "for", "gradient", "-", "based", "learning", "a", "challenge", ".", "While", "promising", ",", "most", "existing", "few", "-", "shot", "learning", "approaches", "either", "require", "complex", "inference", "mechanisms", ",", "complex", "recurrent", "neural", "network", "(", "RNN", ")", "architectures", ",", "or", "fine", "-", "tuning", "the", "target", "problem", ".", "Our", "approach", "is", "most", "related", "to", "others", "that", "aim", "to", "train", "an", "effective", "metric", "for", "one", "-", "shot", "learning", ".", "Where", "they", "focus", "on", "the", "learning", "of", "the", "transferrable", "embedding", "and", "pre", "-", "define", "a", "fixed", "metric", "(", "e.g.", ",", "as", "Euclidean", ")", ",", "we", "further", "aim", "to", "learn", "a", "transferrable", "deep", "metric", "for", "comparing", "the", "relation", "between", "images", "(", "few", "-", "shot", "learning", ")", ",", "or", "between", "images", "and", "class", "descriptions", "(", "zero", "-", "shot", "learning", ")", ".", "By", "expressing", "the", "inductive", "bias", "of", "a", "deeper", "solution", "(", "multiple", "non", "-", "linear", "learned", "stages", "at", "both", "embedding", "and", "relation", "modules", ")", ",", "we", "make", "it", "easier", "to", "learn", "a", "generalisable", "solution", "to", "the", "problem", ".", "Specifically", ",", "we", "propose", "a", "two", "-", "branch", "Relation", "Network", "(", "RN", ")", "that", "performs", "few", "-", "shot", "recognition", "by", "learning", "to", "compare", "query", "images", "against", "few", "-", "shot", "labeled", "sample", "images", ".", "First", "an", "embedding", "module", "generates", "representations", "of", "the", "query", "and", "training", "images", ".", "Then", "these", "embeddings", "are", "compared", "by", "a", "relation", "module", "that", "determines", "if", "they", "are", "from", "matching", "categories", "or", "not", ".", "Defining", "an", "episode", "-", "based", "strategy", "inspired", "by", ",", "the", "embedding", "and", "relation", "modules", "are", "meta", "-", "learned", "end", "-", "to", "-", "end", "to", "support", "few", "-", "shot", "learning", ".", "This", "can", "be", "seen", "as", "extending", "the", "strategy", "of", "to", "include", "a", "learnable", "non", "-", "linear", "comparator", ",", "instead", "of", "a", "fixed", "linear", "comparator", ".", "Our", "approach", "outperforms", "prior", "approaches", ",", "while", "being", "simpler", "(", "no", "RNNs", ")", "and", "faster", "(", "no", "fine", "-", "tuning", ")", ".", "Our", "proposed", "strategy", "also", "directly", "generalises", "to", "zero", "-", "shot", "learning", ".", "In", "this", "case", "the", "sample", "branch", "embeds", "a", "single", "-", "shot", "category", "description", "rather", "than", "a", "single", "exemplar", "training", "image", ",", "and", "the", "relation", "module", "learns", "to", "compare", "query", "image", "and", "category", "description", "embeddings", ".", "Overall", "our", "contribution", "is", "to", "provide", "a", "clean", "framework", "that", "elegantly", "encompasses", "both", "few", "and", "zero", "-", "shot", "learning", ".", "Our", "evaluation", "on", "four", "benchmarks", "show", "that", "it", "provides", "compelling", "performance", "across", "the", "board", "while", "being", "simpler", "and", "faster", "than", "the", "alternatives", ".", "section", ":", "Related", "Work", "The", "study", "of", "one", "or", "few", "-", "shot", "object", "recognition", "has", "been", "of", "interest", "for", "some", "time", ".", "Earlier", "work", "on", "few", "-", "shot", "learning", "tended", "to", "involve", "generative", "models", "with", "complex", "iterative", "inference", "strategies", ".", "With", "the", "success", "of", "discriminative", "deep", "learning", "-", "based", "approaches", "in", "the", "data", "-", "rich", "many", "-", "shot", "setting", ",", "there", "has", "been", "a", "surge", "of", "interest", "in", "generalising", "such", "deep", "learning", "approaches", "to", "the", "few", "-", "shot", "learning", "setting", ".", "Many", "of", "these", "approaches", "use", "a", "meta", "-", "learning", "or", "learning", "-", "to", "-", "learn", "strategy", "in", "the", "sense", "that", "they", "extract", "some", "transferrable", "knowledge", "from", "a", "set", "of", "auxiliary", "tasks", "(", "meta", "-", "learning", ",", "learning", "-", "to", "-", "learn", ")", ",", "which", "then", "helps", "them", "to", "learn", "the", "target", "few", "-", "shot", "problem", "well", "without", "suffering", "from", "the", "overfitting", "that", "might", "be", "expected", "when", "applying", "deep", "models", "to", "sparse", "data", "problems", ".", "Learning", "to", "Fine", "-", "Tune", "The", "successful", "MAML", "approach", "aimed", "to", "meta", "-", "learn", "an", "initial", "condition", "(", "set", "of", "neural", "network", "weights", ")", "that", "is", "good", "for", "fine", "-", "tuning", "on", "few", "-", "shot", "problems", ".", "The", "strategy", "here", "is", "to", "search", "for", "the", "weight", "configuration", "of", "a", "given", "neural", "network", "such", "that", "it", "can", "be", "effectively", "fine", "-", "tuned", "on", "a", "sparse", "data", "problem", "within", "a", "few", "gradient", "-", "descent", "update", "steps", ".", "Many", "distinct", "target", "problems", "are", "sampled", "from", "a", "multiple", "task", "training", "set", ";", "the", "base", "neural", "network", "model", "is", "then", "fine", "-", "tuned", "to", "solve", "each", "of", "them", ",", "and", "the", "success", "at", "each", "target", "problem", "after", "fine", "-", "tuning", "drives", "updates", "in", "the", "base", "model", "–", "thus", "driving", "the", "production", "of", "an", "easy", "to", "fine", "-", "tune", "initial", "condition", ".", "The", "few", "-", "shot", "optimisation", "approach", "goes", "further", "in", "meta", "-", "learning", "not", "only", "a", "good", "initial", "condition", "but", "an", "LSTM", "-", "based", "optimizer", "that", "is", "trained", "to", "be", "specifically", "effective", "for", "fine", "-", "tuning", ".", "However", "both", "of", "these", "approaches", "suffer", "from", "the", "need", "to", "fine", "-", "tune", "on", "the", "target", "problem", ".", "In", "contrast", ",", "our", "approach", "solves", "target", "problems", "in", "an", "entirely", "feed", "-", "forward", "manner", "with", "no", "model", "updates", "required", ",", "making", "it", "more", "convenient", "for", "low", "-", "latency", "or", "low", "-", "power", "applications", ".", "RNN", "Memory", "Based", "Another", "category", "of", "approaches", "leverage", "recurrent", "neural", "networks", "with", "memories", ".", "Here", "the", "idea", "is", "typically", "that", "an", "RNN", "iterates", "over", "an", "examples", "of", "given", "problem", "and", "accumulates", "the", "knowledge", "required", "to", "solve", "that", "problem", "in", "its", "hidden", "activations", ",", "or", "external", "memory", ".", "New", "examples", "can", "be", "classified", ",", "for", "example", "by", "comparing", "them", "to", "historic", "information", "stored", "in", "the", "memory", ".", "So", "‘", "learning", "’", "a", "single", "target", "problem", "can", "occur", "in", "unrolling", "the", "RNN", ",", "while", "learning", "-", "to", "-", "learn", "means", "training", "the", "weights", "of", "the", "RNN", "by", "learning", "many", "distinct", "problems", ".", "While", "appealing", ",", "these", "architectures", "face", "issues", "in", "ensuring", "that", "they", "reliably", "store", "all", "the", ",", "potentially", "long", "term", ",", "historical", "information", "of", "relevance", "without", "forgetting", ".", "In", "our", "approach", "we", "avoid", "the", "complexity", "of", "recurrent", "networks", ",", "and", "the", "issues", "involved", "in", "ensuring", "the", "adequacy", "of", "their", "memory", ".", "Instead", "our", "learning", "-", "to", "-", "learn", "approach", "is", "defined", "entirely", "with", "simple", "and", "fast", "feed", "forward", "CNNs", ".", "Embedding", "and", "Metric", "Learning", "Approaches", "The", "prior", "approaches", "entail", "some", "complexity", "when", "learning", "the", "target", "few", "-", "shot", "problem", ".", "Another", "category", "of", "approach", "aims", "to", "learn", "a", "set", "of", "projection", "functions", "that", "take", "query", "and", "sample", "images", "from", "the", "target", "problem", "and", "classify", "them", "in", "a", "feed", "forward", "manner", ".", "One", "approach", "is", "to", "parameterise", "the", "weights", "of", "a", "feed", "-", "forward", "classifier", "in", "terms", "of", "the", "sample", "set", ".", "The", "meta", "-", "learning", "here", "is", "to", "train", "the", "auxiliary", "parameterisation", "net", "that", "learns", "how", "to", "paramaterise", "a", "given", "feed", "-", "forward", "classification", "problem", "in", "terms", "of", "a", "few", "-", "shot", "sample", "set", ".", "Metric", "-", "learning", "based", "approaches", "aim", "to", "learn", "a", "set", "of", "projection", "functions", "such", "that", "when", "represented", "in", "this", "embedding", ",", "images", "are", "easy", "to", "recognise", "using", "simple", "nearest", "neighbour", "or", "linear", "classifiers", ".", "In", "this", "case", "the", "meta", "-", "learned", "transferrable", "knowledge", "are", "the", "projection", "functions", "and", "the", "target", "problem", "is", "a", "simple", "feed", "-", "forward", "computation", ".", "The", "most", "related", "methodologies", "to", "ours", "are", "the", "prototypical", "networks", "of", "and", "the", "siamese", "networks", "of", ".", "These", "approaches", "focus", "on", "learning", "embeddings", "that", "transform", "the", "data", "such", "that", "it", "can", "be", "recognised", "with", "a", "fixed", "nearest", "-", "neighbour", "or", "linear", "classifier", ".", "In", "contrast", ",", "our", "framework", "further", "defines", "a", "relation", "classifier", "CNN", ",", "in", "the", "style", "of", "(", "While", "focuses", "on", "reasoning", "about", "relation", "between", "two", "objects", "in", "a", "same", "image", "which", "is", "to", "address", "a", "different", "problem", ".", ")", ".", "Compared", "to", ",", "this", "can", "be", "seen", "as", "providing", "a", "learnable", "rather", "than", "fixed", "metric", ",", "or", "non", "-", "linear", "rather", "than", "linear", "classifier", ".", "Compared", "to", "we", "benefit", "from", "an", "episodic", "training", "strategy", "with", "an", "end", "-", "to", "-", "end", "manner", "from", "scratch", ",", "and", "compared", "to", "we", "avoid", "the", "complexity", "of", "set", "-", "to", "-", "set", "RNN", "embedding", "of", "the", "sample", "-", "set", ",", "and", "simply", "rely", "on", "pooling", ".", "Zero", "-", "Shot", "Learning", "Our", "approach", "is", "designed", "for", "few", "-", "shot", "learning", ",", "but", "elegantly", "spans", "the", "space", "into", "zero", "-", "shot", "learning", "(", "ZSL", ")", "by", "modifying", "the", "sample", "branch", "to", "input", "a", "single", "category", "description", "rather", "than", "single", "training", "image", ".", "When", "applied", "to", "ZSL", "our", "architecture", "is", "related", "to", "methods", "that", "learn", "to", "align", "images", "and", "category", "embeddings", "and", "perform", "recognition", "by", "predicting", "if", "an", "image", "and", "category", "embedding", "pair", "match", ".", "Similarly", "to", "the", "case", "with", "the", "prior", "metric", "-", "based", "few", "-", "shot", "approaches", ",", "most", "of", "these", "apply", "a", "fixed", "manually", "defined", "similarity", "metric", "or", "linear", "classifier", "after", "combining", "the", "image", "and", "category", "embedding", ".", "In", "contrast", ",", "we", "again", "benefit", "from", "a", "deeper", "end", "-", "to", "-", "end", "architecture", "including", "a", "learned", "non", "-", "linear", "metric", "in", "the", "form", "of", "our", "learned", "convolutional", "relation", "network", ";", "as", "well", "as", "from", "an", "episode", "-", "based", "training", "strategy", ".", "section", ":", "Methodology", "subsection", ":", "Problem", "Definition", "We", "consider", "the", "task", "of", "few", "-", "shot", "classifier", "learning", ".", "Formally", ",", "we", "have", "three", "datasets", ":", "a", "training", "set", ",", "a", "support", "set", ",", "and", "a", "testing", "set", ".", "The", "support", "set", "and", "testing", "set", "share", "the", "same", "label", "space", ",", "but", "the", "training", "set", "has", "its", "own", "label", "space", "that", "is", "disjoint", "with", "support", "/", "testing", "set", ".", "If", "the", "support", "set", "contains", "labelled", "examples", "for", "each", "of", "unique", "classes", ",", "the", "target", "few", "-", "shot", "problem", "is", "called", "-", "way", "-", "shot", ".", "With", "the", "support", "set", "only", ",", "we", "can", "in", "principle", "train", "a", "classifier", "to", "assign", "a", "class", "label", "to", "each", "sample", "in", "the", "test", "set", ".", "However", ",", "due", "to", "the", "lack", "of", "labelled", "samples", "in", "the", "support", "set", ",", "the", "performance", "of", "such", "a", "classifier", "is", "usually", "not", "satisfactory", ".", "Therefore", "we", "aim", "to", "perform", "meta", "-", "learning", "on", "the", "training", "set", ",", "in", "order", "to", "extract", "transferrable", "knowledge", "that", "will", "allow", "us", "to", "perform", "better", "few", "-", "shot", "learning", "on", "the", "support", "set", "and", "thus", "classify", "the", "test", "set", "more", "successfully", ".", "An", "effective", "way", "to", "exploit", "the", "training", "set", "is", "to", "mimic", "the", "few", "-", "shot", "learning", "setting", "via", "episode", "based", "training", ",", "as", "proposed", "in", ".", "In", "each", "training", "iteration", ",", "an", "episode", "is", "formed", "by", "randomly", "selecting", "classes", "from", "the", "training", "set", "with", "labelled", "samples", "from", "each", "of", "the", "classes", "to", "act", "as", "the", "sample", "set", "(", ")", ",", "as", "well", "as", "a", "fraction", "of", "the", "remainder", "of", "those", "classes", "’", "samples", "to", "serve", "as", "the", "query", "set", ".", "This", "sample", "/", "query", "set", "split", "is", "designed", "to", "simulate", "the", "support", "/", "test", "set", "that", "will", "be", "encountered", "at", "test", "time", ".", "A", "model", "trained", "from", "sample", "/", "query", "set", "can", "be", "further", "fine", "-", "tuned", "using", "the", "support", "set", ",", "if", "desired", ".", "In", "this", "work", "we", "adopt", "such", "an", "episode", "-", "based", "training", "strategy", ".", "In", "our", "few", "-", "shot", "experiments", "(", "see", "Section", "[", "reference", "]", ")", "we", "consider", "one", "-", "shot", "(", ",", "Figure", "[", "reference", "]", ")", "and", "five", "-", "shot", "(", ")", "settings", ".", "We", "also", "address", "the", "zero", "-", "shot", "learning", "case", "as", "explained", "in", "Section", "[", "reference", "]", ".", "subsection", ":", "Model", "One", "-", "Shot", "Our", "Relation", "Network", "(", "RN", ")", "consists", "of", "two", "modules", ":", "an", "embedding", "module", "and", "a", "relation", "module", ",", "as", "illustrated", "in", "Figure", "[", "reference", "]", ".", "Samples", "in", "the", "query", "set", ",", "and", "samples", "in", "the", "sample", "set", "are", "fed", "through", "the", "embedding", "module", ",", "which", "produces", "feature", "maps", "and", ".", "The", "feature", "maps", "and", "are", "combined", "with", "operator", ".", "In", "this", "work", "we", "assume", "to", "be", "concatenation", "of", "feature", "maps", "in", "depth", ",", "although", "other", "choices", "are", "possible", ".", "The", "combined", "feature", "map", "of", "the", "sample", "and", "query", "are", "fed", "into", "the", "relation", "module", ",", "which", "eventually", "produces", "a", "scalar", "in", "range", "of", "to", "representing", "the", "similarity", "between", "and", ",", "which", "is", "called", "relation", "score", ".", "Thus", ",", "in", "the", "-", "way", "one", "-", "shot", "setting", ",", "we", "generate", "relation", "scores", "for", "the", "relation", "between", "one", "query", "input", "and", "training", "sample", "set", "examples", ",", "K", "-", "shot", "For", "-", "shot", "where", ",", "we", "element", "-", "wise", "sum", "over", "the", "embedding", "module", "outputs", "of", "all", "samples", "from", "each", "training", "class", "to", "form", "this", "class", "’", "feature", "map", ".", "This", "pooled", "class", "-", "level", "feature", "map", "is", "combined", "with", "the", "query", "image", "feature", "map", "as", "above", ".", "Thus", ",", "the", "number", "of", "relation", "scores", "for", "one", "query", "is", "always", "in", "both", "one", "-", "shot", "or", "few", "-", "shot", "setting", ".", "Objective", "function", "We", "use", "mean", "square", "error", "(", "MSE", ")", "loss", "(", "Eq", ".", "(", "[", "reference", "]", ")", ")", "to", "train", "our", "model", ",", "regressing", "the", "relation", "score", "to", "the", "ground", "truth", ":", "matched", "pairs", "have", "similarity", "and", "the", "mismatched", "pair", "have", "similarity", ".", "The", "choice", "of", "MSE", "is", "somewhat", "non", "-", "standard", ".", "Our", "problem", "may", "seem", "to", "be", "a", "classification", "problem", "with", "a", "label", "space", ".", "However", "conceptually", "we", "are", "predicting", "relation", "scores", ",", "which", "can", "be", "considered", "a", "regression", "problem", "despite", "that", "for", "ground", "-", "truth", "we", "can", "only", "automatically", "generate", "targets", ".", "subsection", ":", "Zero", "-", "shot", "Learning", "Zero", "-", "shot", "learning", "is", "analogous", "to", "one", "-", "shot", "learning", "in", "that", "one", "datum", "is", "given", "to", "define", "each", "class", "to", "recognise", ".", "However", "instead", "of", "being", "given", "a", "support", "set", "with", "one", "-", "shot", "image", "for", "each", "of", "training", "classes", ",", "it", "contains", "a", "semantic", "class", "embedding", "vector", "for", "each", ".", "Modifying", "our", "framework", "to", "deal", "with", "the", "zero", "-", "shot", "case", "is", "straightforward", ":", "as", "a", "different", "modality", "of", "semantic", "vectors", "is", "used", "for", "the", "support", "set", "(", "attribute", "vectors", "instead", "of", "images", ")", ",", "we", "use", "a", "second", "heterogeneous", "embedding", "module", "besides", "the", "embedding", "module", "used", "for", "the", "image", "query", "set", ".", "Then", "the", "relation", "net", "is", "applied", "as", "before", ".", "Therefore", ",", "the", "relation", "score", "for", "each", "query", "input", "will", "be", ":", "The", "objective", "function", "for", "zero", "-", "shot", "learning", "is", "the", "same", "as", "that", "for", "few", "-", "shot", "learning", ".", "subsection", ":", "Network", "Architecture", "As", "most", "few", "-", "shot", "learning", "models", "utilise", "four", "convolutional", "blocks", "for", "embedding", "module", ",", "we", "follow", "the", "same", "architecture", "setting", "for", "fair", "comparison", ",", "see", "Figure", "[", "reference", "]", ".", "More", "concretely", ",", "each", "convolutional", "block", "contains", "a", "64", "-", "filter", "convolution", ",", "a", "batch", "normalisation", "and", "a", "ReLU", "nonlinearity", "layer", "respectively", ".", "The", "first", "two", "blocks", "also", "contain", "a", "max", "-", "pooling", "layer", "while", "the", "latter", "two", "do", "not", ".", "We", "do", "so", "because", "we", "need", "the", "output", "feature", "maps", "for", "further", "convolutional", "layers", "in", "the", "relation", "module", ".", "The", "relation", "module", "consists", "of", "two", "convolutional", "blocks", "and", "two", "fully", "-", "connected", "layers", ".", "Each", "of", "convolutional", "block", "is", "a", "convolution", "with", "64", "filters", "followed", "by", "batch", "normalisation", ",", "ReLU", "non", "-", "linearity", "and", "max", "-", "pooling", ".", "The", "output", "size", "of", "last", "max", "pooling", "layer", "is", "and", "for", "Omniglot", "and", "mini", "ImageNet", "respectively", ".", "The", "two", "fully", "-", "connected", "layers", "are", "8", "and", "1", "dimensional", ",", "respectively", ".", "All", "fully", "-", "connected", "layers", "are", "ReLU", "except", "the", "output", "layer", "is", "Sigmoid", "in", "order", "to", "generate", "relation", "scores", "in", "a", "reasonable", "range", "for", "all", "versions", "of", "our", "network", "architecture", ".", "The", "zero", "-", "shot", "learning", "architecture", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "In", "this", "architecture", ",", "the", "DNN", "subnet", "is", "an", "existing", "network", "(", "e.g.", ",", "Inception", "or", "ResNet", ")", "pretrained", "on", "ImageNet", ".", "section", ":", "Experiments", "We", "evaluate", "our", "approach", "on", "two", "related", "tasks", ":", "few", "-", "shot", "classification", "on", "Omniglot", "and", "mini", "Imagenet", ",", "and", "zero", "-", "shot", "classification", "on", "Animals", "with", "Attributes", "(", "AwA", ")", "and", "Caltech", "-", "UCSD", "Birds", "-", "200", "-", "2011", "(", "CUB", ")", ".", "All", "the", "experiments", "are", "implemented", "based", "on", "PyTorch", ".", "subsection", ":", "Few", "-", "shot", "Recognition", "Settings", "Few", "-", "shot", "learning", "in", "all", "experiments", "uses", "Adam", "with", "initial", "learning", "rate", ",", "annealed", "by", "half", "for", "every", "100", ",", "000", "episodes", ".", "All", "our", "models", "are", "end", "-", "to", "-", "end", "trained", "from", "scratch", "with", "no", "additional", "dataset", ".", "Baselines", "We", "compare", "against", "various", "state", "of", "the", "art", "baselines", "for", "few", "-", "shot", "recognition", ",", "including", "neural", "statistician", ",", "Matching", "Nets", "with", "and", "without", "fine", "-", "tuning", ",", "MANN", ",", "Siamese", "Nets", "with", "Memory", ",", "Convolutional", "Siamese", "Nets", ",", "MAML", ",", "Meta", "Nets", ",", "Prototypical", "Nets", "and", "Meta", "-", "Learner", "LSTM", ".", "subsubsection", ":", "Omniglot", "Dataset", "Omniglot", "contains", "1623", "characters", "(", "classes", ")", "from", "50", "different", "alphabets", ".", "Each", "class", "contains", "20", "samples", "drawn", "by", "different", "people", ".", "Following", ",", "we", "augment", "new", "classes", "through", ",", "and", "rotations", "of", "existing", "data", "and", "use", "1200", "original", "classes", "plus", "rotations", "for", "training", "and", "remaining", "423", "classes", "plus", "rotations", "for", "testing", ".", "All", "input", "images", "are", "resized", "to", ".", "Training", "Besides", "the", "sample", "images", ",", "the", "5", "-", "way", "1", "-", "shot", "contains", "19", "query", "images", ",", "the", "5", "-", "way", "5", "-", "shot", "has", "15", "query", "images", ",", "the", "20", "-", "way", "1", "-", "shot", "has", "10", "query", "images", "and", "the", "20", "-", "way", "5", "-", "shot", "has", "5", "query", "images", "for", "each", "of", "the", "sampled", "classes", "in", "each", "training", "episode", ".", "This", "means", "for", "example", "that", "there", "are", "images", "in", "one", "training", "episode", "/", "mini", "-", "batch", "for", "the", "5", "-", "way", "1", "-", "shot", "experiments", ".", "Results", "Following", ",", "we", "computed", "few", "-", "shot", "classification", "accuracies", "on", "Omniglot", "by", "averaging", "over", "1000", "randomly", "generated", "episodes", "from", "the", "testing", "set", ".", "For", "the", "1", "-", "shot", "and", "5", "-", "shot", "experiments", ",", "we", "batch", "one", "and", "five", "query", "images", "per", "class", "respectively", "for", "evaluation", "during", "testing", ".", "The", "results", "are", "shown", "in", "Table", "[", "reference", "]", ".", "We", "achieved", "state", "-", "of", "-", "the", "-", "art", "performance", "under", "all", "experiments", "setting", "with", "higher", "averaged", "accuracies", "and", "lower", "standard", "deviations", ",", "except", "5", "-", "way", "5", "-", "shot", "where", "our", "model", "is", "0.1", "%", "lower", "in", "accuracy", "than", ".", "This", "is", "despite", "that", "many", "alternatives", "have", "significantly", "more", "complicated", "machinery", ",", "or", "fine", "-", "tune", "on", "the", "target", "problem", ",", "while", "we", "do", "not", ".", "subsubsection", ":", "mini", "ImageNet", "paragraph", ":", "Dataset", "The", "mini", "Imagenet", "dataset", ",", "originally", "proposed", "by", ",", "consists", "of", "60", ",", "000", "colour", "images", "with", "100", "classes", ",", "each", "having", "600", "examples", ".", "We", "followed", "the", "split", "introduced", "by", ",", "with", "64", ",", "16", ",", "and", "20", "classes", "for", "training", ",", "validation", "and", "testing", ",", "respectively", ".", "The", "16", "validation", "classes", "is", "used", "for", "monitoring", "generalisation", "performance", "only", ".", "Training", "Following", "the", "standard", "setting", "adopted", "by", "most", "existing", "few", "-", "shot", "learning", "work", ",", "we", "conducted", "5", "way", "1", "-", "shot", "and", "5", "-", "shot", "classification", ".", "Beside", "the", "sample", "images", ",", "the", "5", "-", "way", "1", "-", "shot", "contains", "15", "query", "images", ",", "and", "the", "5", "-", "way", "5", "-", "shot", "has", "10", "query", "images", "for", "each", "of", "the", "sampled", "classes", "in", "each", "training", "episode", ".", "This", "means", "for", "example", "that", "there", "are", "images", "in", "one", "training", "episode", "/", "mini", "-", "batch", "for", "5", "-", "way", "1", "-", "shot", "experiments", ".", "We", "resize", "input", "images", "to", ".", "Our", "model", "is", "trained", "end", "-", "to", "-", "end", "from", "scratch", ",", "with", "random", "initialisation", ",", "and", "no", "additional", "training", "set", ".", "Results", "Following", ",", "we", "batch", "15", "query", "images", "per", "class", "in", "each", "episode", "for", "evaluation", "in", "both", "1", "-", "shot", "and", "5", "-", "shot", "scenarios", "and", "the", "few", "-", "shot", "classification", "accuracies", "are", "computed", "by", "averaging", "over", "600", "randomly", "generated", "episodes", "from", "the", "test", "set", ".", "From", "Table", "[", "reference", "]", ",", "we", "can", "see", "that", "our", "model", "achieved", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "5", "-", "way", "1", "-", "shot", "settings", "and", "competitive", "results", "on", "5", "-", "way", "5", "-", "shot", ".", "However", ",", "the", "1", "-", "shot", "result", "reported", "by", "prototypical", "networks", "reqired", "to", "be", "trained", "on", "30", "-", "way", "15", "queries", "per", "training", "episode", ",", "and", "5", "-", "shot", "result", "was", "trained", "on", "20", "-", "way", "15", "queries", "per", "training", "episode", ".", "When", "trained", "with", "5", "-", "way", "15", "query", "per", "training", "episode", ",", "only", "got", "for", "1", "-", "shot", "evaluation", ",", "clearly", "weaker", "than", "ours", ".", "In", "contrast", ",", "all", "our", "models", "are", "trained", "on", "5", "-", "way", ",", "1", "query", "for", "1", "-", "shot", "and", "5", "queries", "for", "5", "-", "shot", "per", "training", "episode", ",", "with", "much", "less", "training", "queries", "than", ".", "subsection", ":", "Zero", "-", "shot", "Recognition", "Datasets", "and", "settings", "We", "follow", "two", "ZSL", "settings", ":", "the", "old", "setting", "and", "the", "new", "GBU", "setting", "provided", "by", "for", "training", "/", "test", "splits", ".", "Under", "the", "old", "setting", ",", "adopted", "by", "most", "existing", "ZSL", "works", "before", ",", "some", "of", "the", "test", "classes", "also", "appear", "in", "the", "ImageNet", "1", "K", "classes", ",", "which", "have", "been", "used", "to", "pretrain", "the", "image", "embedding", "network", ",", "thus", "violating", "the", "zero", "-", "shot", "assumption", ".", "In", "contrast", ",", "the", "new", "GBU", "setting", "ensures", "that", "none", "of", "the", "test", "classes", "of", "the", "datasets", "appear", "in", "the", "ImageNet", "1", "K", "classes", ".", "Under", "both", "settings", ",", "the", "test", "set", "can", "comprise", "only", "the", "unseen", "class", "samples", "(", "conventional", "test", "set", "setting", ")", "or", "a", "mixture", "of", "seen", "and", "unseen", "class", "samples", ".", "The", "latter", ",", "termed", "generalised", "zero", "-", "shot", "learning", "(", "GZSL", ")", ",", "is", "more", "realistic", "in", "practice", ".", "Two", "widely", "used", "ZSL", "benchmarks", "are", "selected", "for", "the", "old", "setting", ":", "AwA", "(", "Animals", "with", "Attributes", ")", "consists", "of", "30", ",", "745", "images", "of", "50", "classes", "of", "animals", ".", "It", "has", "a", "fixed", "split", "for", "evaluation", "with", "40", "training", "classes", "and", "10", "test", "classes", ".", "CUB", "(", "Caltech", "-", "UCSD", "Birds", "-", "200", "-", "2011", ")", "contains", "11", ",", "788", "images", "of", "200", "bird", "species", "with", "150", "seen", "classes", "and", "50", "disjoint", "unseen", "classes", ".", "Three", "datasets", "are", "selected", "for", "GBU", "setting", ":", "AwA1", ",", "AwA2", "and", "CUB", ".", "The", "newly", "released", "AwA2", "consists", "of", "37", ",", "322", "images", "of", "50", "classes", "which", "is", "an", "extension", "of", "AwA", "while", "AwA1", "is", "same", "as", "AwA", "but", "under", "the", "GBU", "setting", ".", "Semantic", "representation", "For", "AwA", ",", "we", "use", "the", "continuous", "85", "-", "dimension", "class", "-", "level", "attribute", "vector", "from", ",", "which", "has", "been", "used", "by", "all", "recent", "works", ".", "For", "CUB", ",", "a", "continuous", "312", "-", "dimension", "class", "-", "level", "attribute", "vector", "is", "used", ".", "Implementation", "details", "Two", "different", "embedding", "modules", "are", "used", "for", "the", "two", "input", "modalities", "in", "zero", "-", "shot", "learning", ".", "Unless", "otherwise", "specified", ",", "we", "use", "Inception", "-", "V2", "as", "the", "query", "image", "embedding", "DNN", "in", "the", "old", "and", "conventional", "setting", "and", "ResNet101", "for", "the", "GBU", "and", "generalised", "setting", ",", "taking", "the", "top", "pooling", "units", "as", "image", "embedding", "with", "dimension", "and", "respectively", ".", "This", "DNN", "is", "pre", "-", "trained", "on", "ILSVRC", "2012", "1", "K", "classification", "without", "fine", "-", "tuning", ",", "as", "in", "recent", "deep", "ZSL", "works", ".", "A", "MLP", "network", "is", "used", "for", "embedding", "semantic", "attribute", "vectors", ".", "The", "size", "of", "hidden", "layer", "FC1", "(", "Figure", "[", "reference", "]", ")", "is", "set", "to", "1024", "and", "1200", "for", "AwA", "and", "CUB", "respectively", ",", "and", "the", "output", "size", "FC2", "is", "set", "to", "the", "same", "dimension", "as", "the", "image", "embedding", "for", "both", "datasets", ".", "For", "the", "relation", "module", ",", "the", "image", "and", "semantic", "embeddings", "are", "concatenated", "before", "being", "fed", "into", "MLPs", "with", "hidden", "layer", "FC3", "size", "400", "and", "1200", "for", "AwA", "and", "CUB", ",", "respectively", ".", "We", "add", "weight", "decay", "(", "L2", "regularisation", ")", "in", "FC1", "&", "2", "as", "there", "is", "a", "hubness", "problem", "in", "cross", "-", "modal", "mapping", "for", "ZSL", "which", "can", "be", "best", "solved", "by", "mapping", "the", "semantic", "feature", "vector", "to", "the", "visual", "feature", "space", "with", "regularisation", ".", "After", "that", ",", "FC3", "&", "4", "(", "relation", "module", ")", "are", "used", "to", "compute", "the", "relation", "between", "the", "semantic", "representation", "(", "in", "the", "visual", "feature", "space", ")", "and", "the", "visual", "representation", ".", "Since", "the", "hubness", "problem", "does", "not", "existing", "in", "this", "step", ",", "no", "L2", "regularisation", "/", "weight", "decay", "is", "needed", ".", "All", "the", "ZSL", "models", "are", "trained", "with", "weight", "decay", "in", "the", "embedding", "network", ".", "The", "learning", "rate", "is", "initialised", "to", "with", "Adam", "and", "then", "annealed", "by", "half", "every", "200", ",", "000", "iterations", ".", "Results", "under", "the", "old", "setting", "The", "conventional", "evaluation", "for", "ZSL", "followed", "by", "the", "majority", "of", "prior", "work", "is", "to", "assume", "that", "the", "test", "data", "all", "comes", "from", "unseen", "classes", ".", "We", "evaluate", "this", "setting", "first", ".", "We", "compare", "15", "alternative", "approaches", "in", "Table", "[", "reference", "]", ".", "With", "only", "the", "attribute", "vector", "used", "as", "the", "sample", "class", "embedding", ",", "our", "model", "achieves", "competitive", "result", "on", "AwA", "and", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "the", "more", "challenging", "CUB", "dataset", ",", "outperforming", "the", "most", "related", "alternative", "prototypical", "networks", "by", "a", "big", "margin", ".", "Note", "that", "only", "inductive", "methods", "are", "considered", ".", "Some", "recent", "methods", "are", "tranductive", "in", "that", "they", "use", "all", "test", "data", "at", "once", "for", "model", "training", ",", "which", "gives", "them", "a", "big", "advantage", "at", "the", "cost", "of", "making", "a", "very", "strong", "assumption", "that", "may", "not", "be", "met", "in", "practical", "applications", ",", "so", "we", "do", "not", "compare", "with", "them", "here", ".", "Results", "under", "the", "GBU", "setting", "We", "follow", "the", "evaluation", "setting", "of", ".", "We", "compare", "our", "model", "with", "11", "alternative", "ZSL", "models", "in", "Table", "[", "reference", "]", ".", "The", "10", "shallow", "models’", "results", "are", "from", "and", "the", "result", "of", "the", "state", "-", "of", "-", "the", "-", "art", "method", "DEM", "is", "from", "the", "authors", "’", "GitHub", "page", ".", "We", "can", "see", "that", "on", "AwA2", "and", "CUB", ",", "Our", "model", "is", "particularly", "strong", "under", "the", "more", "realistic", "GZSL", "setting", "measured", "using", "the", "harmonic", "mean", "(", "H", ")", "metric", ".", "While", "on", "AwA1", ",", "our", "method", "is", "only", "outperformed", "by", "DEM", ".", "section", ":", "Why", "does", "Relation", "Network", "Work", "?", "subsection", ":", "Relationship", "to", "existing", "models", "Related", "prior", "few", "-", "shot", "work", "uses", "fixed", "pre", "-", "specified", "distance", "metrics", "such", "as", "Euclidean", "or", "cosine", "distance", "to", "perform", "classification", ".", "These", "studies", "can", "be", "seen", "as", "distance", "metric", "learning", ",", "but", "where", "all", "the", "learning", "occurs", "in", "the", "feature", "embedding", ",", "and", "a", "fixed", "metric", "is", "used", "given", "the", "learned", "embedding", ".", "Also", "related", "are", "conventional", "metric", "learning", "approaches", "that", "focus", "on", "learning", "a", "shallow", "(", "linear", ")", "Mahalanobis", "metric", "for", "a", "fixed", "feature", "representation", ".", "In", "contrast", "to", "prior", "work", "’s", "fixed", "metric", "or", "fixed", "features", "and", "shallow", "learned", "metric", ",", "Relation", "Network", "can", "be", "seen", "as", "both", "learning", "a", "deep", "embedding", "and", "learning", "a", "deep", "non", "-", "linear", "metric", "(", "similarity", "function", ")", ".", "These", "are", "mutually", "tuned", "end", "-", "to", "-", "end", "to", "support", "each", "other", "in", "few", "short", "learning", ".", "Why", "might", "this", "be", "particularly", "useful", "?", "By", "using", "a", "flexible", "function", "approximator", "to", "learn", "similarity", ",", "we", "learn", "a", "good", "metric", "in", "a", "data", "driven", "way", "and", "do", "not", "have", "to", "manually", "choose", "the", "right", "metric", "(", "Euclidean", ",", "cosine", ",", "Mahalanobis", ")", ".", "Fixed", "metrics", "like", "assume", "that", "features", "are", "solely", "compared", "element", "-", "wise", ",", "and", "the", "most", "related", "assumes", "linear", "separability", "after", "the", "embedding", ".", "These", "are", "thus", "critically", "dependent", "on", "the", "efficacy", "of", "the", "learned", "embedding", "network", ",", "and", "hence", "limited", "by", "the", "extent", "to", "which", "the", "embedding", "networks", "generate", "inadequately", "discriminative", "representations", ".", "In", "contrast", ",", "by", "deep", "learning", "a", "non", "-", "linear", "similarity", "metric", "jointly", "with", "the", "embedding", ",", "Relation", "Network", "can", "better", "identify", "matching", "/", "mismatching", "pairs", ".", "[", "b", "]", "0.5", "[", "b", "]", "0.5", "[", "b", "]", "0.5", "[", "b", "]", "0.5", "subsection", ":", "Visualisation", "To", "illustrate", "the", "previous", "point", "about", "adequacy", "of", "learned", "input", "embeddings", ",", "we", "show", "a", "synthetic", "example", "where", "existing", "approaches", "definitely", "fail", "and", "our", "Relation", "Network", "can", "succeed", "due", "to", "using", "a", "deep", "relation", "module", ".", "Assuming", "2D", "query", "and", "sample", "input", "embeddings", "to", "a", "relation", "module", ",", "Fig", ".", "[", "reference", "]", "(", "a", ")", "shows", "the", "space", "of", "2D", "sample", "inputs", "for", "a", "fixed", "2D", "query", "input", ".", "Each", "sample", "input", "(", "pixel", ")", "is", "colored", "according", "to", "whether", "it", "matches", "the", "fixed", "query", "or", "not", ".", "This", "represents", "a", "case", "where", "the", "output", "of", "the", "embedding", "modules", "is", "not", "discriminative", "enough", "for", "trivial", "(", "Euclidean", "NN", ")", "comparison", "between", "query", "and", "sample", "set", ".", "In", "Fig", ".", "[", "reference", "]", "(", "c", ")", "we", "attempt", "to", "learn", "matching", "via", "a", "Mahalanobis", "metric", "learning", "relation", "module", ",", "and", "we", "can", "see", "the", "result", "is", "inadequate", ".", "In", "Fig", ".", "[", "reference", "]", "(", "d", ")", "we", "learn", "a", "further", "2", "-", "hidden", "layer", "MLP", "embedding", "of", "query", "and", "sample", "inputs", "as", "well", "as", "the", "subsequent", "Mahalanobis", "metric", ",", "which", "is", "also", "not", "adequate", ".", "Only", "by", "learning", "the", "full", "deep", "relation", "module", "for", "similarity", "can", "we", "solve", "this", "problem", "in", "Fig", ".", "[", "reference", "]", "(", "b", ")", ".", "In", "a", "real", "problem", "the", "difficulty", "of", "comparing", "embeddings", "may", "not", "be", "this", "extreme", ",", "but", "it", "can", "still", "be", "challenging", ".", "We", "qualitatively", "illustrate", "the", "challenge", "of", "matching", "two", "example", "Omniglot", "query", "images", "(", "embeddings", "projected", "to", "2D", ",", "Figure", "[", "reference", "]", "(", "left", ")", ")", "by", "showing", "an", "analogous", "plot", "of", "real", "sample", "images", "colored", "by", "match", "(", "cyan", ")", "or", "mismatch", "(", "magenta", ")", "to", "two", "example", "queries", "(", "yellow", ")", ".", "Under", "standard", "assumptions", "the", "cyan", "matching", "samples", "should", "be", "nearest", "neighbours", "to", "the", "yellow", "query", "image", "with", "some", "metric", "(", "Euclidean", ",", "Cosine", ",", "Mahalanobis", ")", ".", "But", "we", "can", "see", "that", "the", "match", "relation", "is", "more", "complex", "than", "this", ".", "In", "Figure", "[", "reference", "]", "(", "right", ")", ",", "we", "instead", "plot", "the", "same", "two", "example", "queries", "in", "terms", "of", "a", "2D", "PCA", "representation", "of", "each", "query", "-", "sample", "pair", ",", "as", "represented", "by", "the", "relation", "module", "’s", "penultimate", "layer", ".", "We", "can", "see", "that", "the", "relation", "network", "has", "mapped", "the", "data", "into", "a", "space", "where", "the", "(", "mis", ")", "matched", "pairs", "are", "linearly", "separable", ".", "section", ":", "Conclusion", "We", "proposed", "a", "simple", "method", "called", "the", "Relation", "Network", "for", "few", "-", "shot", "and", "zero", "-", "shot", "learning", ".", "Relation", "network", "learns", "an", "embedding", "and", "a", "deep", "non", "-", "linear", "distance", "metric", "for", "comparing", "query", "and", "sample", "items", ".", "Training", "the", "network", "end", "-", "to", "-", "end", "with", "episodic", "training", "tunes", "the", "embedding", "and", "distance", "metric", "for", "effective", "few", "-", "shot", "learning", ".", "This", "approach", "is", "far", "simpler", "and", "more", "efficient", "than", "recent", "few", "-", "shot", "meta", "-", "learning", "approaches", ",", "and", "produces", "state", "-", "of", "-", "the", "-", "art", "results", ".", "It", "further", "proves", "effective", "at", "both", "conventional", "and", "generalised", "zero", "-", "shot", "settings", ".", "Acknowledgements", "This", "work", "was", "supported", "by", "the", "ERC", "grant", "ERC", "-", "2012", "-", "AdG", "321162", "-", "HELIOS", ",", "EPSRC", "grant", "Seebibyte", "EP", "/", "M013774", "/", "1", ",", "EPSRC", "/", "MURI", "grant", "EP", "/", "N019474", "/", "1", ",", "EPSRC", "grant", "EP", "/", "R026173", "/", "1", ",", "and", "the", "European", "Union", "’s", "Horizon", "2020", "research", "and", "innovation", "program", "(", "grant", "agreement", "no", ".", "640891", ")", ".", "We", "gratefully", "acknowledge", "the", "support", "of", "NVIDIA", "Corporation", "with", "the", "donation", "of", "the", "Titan", "Xp", "GPU", "and", "the", "ESPRC", "funded", "Tier", "2", "facility", ",", "JADE", "used", "for", "this", "research", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["Learning"], "Task"], [6, 8, ["Relation", "Network"], "Method"], [9, 13, ["Few", "-", "Shot", "Learning"], "Task"], [25, 29, ["few", "-", "shot", "learning"], "Task"], [32, 33, ["classifier"], "Method"], [51, 53, ["Relation", "Network"], "Method"], [54, 55, ["RN"], "Method"], [68, 71, ["meta", "-", "learning"], "Method"], [98, 102, ["few", "-", "shot", "setting"], "Task"], [107, 108, ["RN"], "Method"], [142, 146, ["few", "-", "shot", "learning"], "Task"], [153, 157, ["zero", "-", "shot", "learning"], "Task"], [184, 187, ["Deep", "learning", "models"], "Method"], [192, 195, ["visual", "recognition", "tasks"], "Task"], [199, 202, ["supervised", "learning", "models"], "Method"], [229, 231, ["annotation", "cost"], "Metric"], [274, 276, ["recognising", "objects"], "Task"], [287, 290, ["few", "-", "shot"], "Task"], [291, 295, ["zero", "-", "shot", "learning"], "Task"], [353, 355, ["few", "-"], "Task"], [356, 360, ["zero", "-", "shot", "learning"], "Task"], [376, 379, ["few", "-", "shot"], "Task"], [380, 384, ["zero", "-", "shot", "learning"], "Task"], [385, 389, ["Few", "-", "shot", "learning"], "Task"], [393, 395, ["visual", "categories"], "Task"], [414, 419, ["fine", "-", "tuning", "’", "practice"], "Method"], [420, 422, ["deep", "learning"], "Task"], [423, 428, ["Data", "augmentation", "and", "regularisation", "techniques"], "Method"], [430, 431, ["overfitting"], "Task"], [450, 454, ["few", "-", "shot", "learning"], "Task"], [456, 457, ["training"], "Task"], [460, 463, ["meta", "learning", "phase"], "Method"], [478, 480, ["optimisation", "strategies"], "Method"], [492, 495, ["fine", "-", "tuning"], "Method"], [498, 500, ["optimisation", "strategy"], "Method"], [504, 508, ["feed", "-", "forward", "pass"], "Method"], [509, 512, ["updating", "network", "weights"], "Method"], [513, 517, ["Zero", "-", "shot", "learning"], "Method"], [524, 525, ["Recognisers"], "Method"], [536, 538, ["class", "description"], "Method"], [552, 554, ["data", "insufficiency"], "Task"], [555, 559, ["gradient", "-", "based", "learning"], "Method"], [575, 577, ["inference", "mechanisms"], "Method"], [608, 609, ["metric"], "Metric"], [623, 625, ["transferrable", "embedding"], "Method"], [645, 648, ["transferrable", "deep", "metric"], "Method"], [655, 659, ["few", "-", "shot", "learning"], "Task"], [668, 672, ["zero", "-", "shot", "learning"], "Task"], [692, 696, ["embedding", "and", "relation", "modules"], "Method"], [719, 721, ["Relation", "Network"], "Method"], [722, 723, ["RN"], "Method"], [746, 748, ["embedding", "module"], "Method"], [764, 766, ["relation", "module"], "Method"], [779, 783, ["episode", "-", "based", "strategy"], "Method"], [787, 791, ["embedding", "and", "relation", "modules"], "Method"], [802, 806, ["few", "-", "shot", "learning"], "Task"], [828, 831, ["fixed", "linear", "comparator"], "Method"], [843, 844, ["RNNs"], "Method"], [849, 852, ["fine", "-", "tuning"], "Method"], [861, 865, ["zero", "-", "shot", "learning"], "Task"], [889, 891, ["relation", "module"], "Method"], [897, 900, ["category", "description", "embeddings"], "Method"], [969, 973, ["few", "-", "shot", "learning"], "Task"], [976, 978, ["generative", "models"], "Method"], [980, 983, ["iterative", "inference", "strategies"], "Method"], [996, 1003, ["data", "-", "rich", "many", "-", "shot", "setting"], "Task"], [1031, 1041, ["meta", "-", "learning", "or", "learning", "-", "to", "-", "learn", "strategy"], "Method"], [1057, 1060, ["meta", "-", "learning"], "Method"], [1092, 1094, ["deep", "models"], "Method"], [1095, 1098, ["sparse", "data", "problems"], "Task"], [1106, 1108, ["MAML", "approach"], "Method"], [1127, 1130, ["fine", "-", "tuning"], "Task"], [1149, 1151, ["neural", "network"], "Method"], [1162, 1165, ["sparse", "data", "problem"], "Task"], [1168, 1173, ["gradient", "-", "descent", "update", "steps"], "Method"], [1188, 1192, ["base", "neural", "network", "model"], "Method"], [1244, 1247, ["meta", "-", "learning"], "Method"], [1255, 1259, ["LSTM", "-", "based", "optimizer"], "Method"], [1267, 1270, ["fine", "-", "tuning"], "Task"], [1295, 1297, ["target", "problems"], "Task"], [1300, 1304, ["feed", "-", "forward", "manner"], "Method"], [1315, 1323, ["low", "-", "latency", "or", "low", "-", "power", "applications"], "Task"], [1332, 1335, ["recurrent", "neural", "networks"], "Method"], [1336, 1337, ["memories"], "Method"], [1345, 1346, ["RNN"], "Method"], [1403, 1404, ["RNN"], "Method"], [1417, 1418, ["RNN"], "Method"], [1459, 1461, ["recurrent", "networks"], "Method"], [1476, 1482, ["learning", "-", "to", "-", "learn", "approach"], "Method"], [1489, 1492, ["feed", "forward", "CNNs"], "Method"], [1493, 1498, ["Embedding", "and", "Metric", "Learning", "Approaches"], "Method"], [1523, 1525, ["projection", "functions"], "Method"], [1540, 1543, ["feed", "forward", "manner"], "Method"], [1553, 1557, ["feed", "-", "forward", "classifier"], "Method"], [1565, 1568, ["meta", "-", "learning"], "Method"], [1573, 1576, ["auxiliary", "parameterisation", "net"], "Method"], [1583, 1588, ["feed", "-", "forward", "classification", "problem"], "Task"], [1598, 1603, ["Metric", "-", "learning", "based", "approaches"], "Method"], [1609, 1611, ["projection", "functions"], "Method"], [1626, 1628, ["nearest", "neighbour"], "Method"], [1629, 1631, ["linear", "classifiers"], "Method"], [1643, 1645, ["projection", "functions"], "Method"], [1652, 1656, ["feed", "-", "forward", "computation"], "Method"], [1665, 1667, ["prototypical", "networks"], "Method"], [1670, 1672, ["siamese", "networks"], "Method"], [1679, 1680, ["embeddings"], "Method"], [1693, 1699, ["nearest", "-", "neighbour", "or", "linear", "classifier"], "Method"], [1708, 1711, ["relation", "classifier", "CNN"], "Method"], [1757, 1760, ["non", "-", "linear"], "Method"], [1762, 1764, ["linear", "classifier"], "Method"], [1771, 1774, ["episodic", "training", "strategy"], "Method"], [1810, 1811, ["pooling"], "Method"], [1812, 1816, ["Zero", "-", "Shot", "Learning"], "Method"], [1821, 1825, ["few", "-", "shot", "learning"], "Task"], [1832, 1836, ["zero", "-", "shot", "learning"], "Task"], [1837, 1838, ["ZSL"], "Method"], [1859, 1860, ["ZSL"], "Method"], [1876, 1877, ["recognition"], "Task"], [1909, 1913, ["manually", "defined", "similarity", "metric"], "Method"], [1914, 1916, ["linear", "classifier"], "Method"], [1919, 1923, ["image", "and", "category", "embedding"], "Method"], [1933, 1939, ["end", "-", "to", "-", "end", "architecture"], "Method"], [1952, 1955, ["convolutional", "relation", "network"], "Method"], [1961, 1966, ["episode", "-", "based", "training", "strategy"], "Method"], [2073, 2074, ["classifier"], "Method"], [2106, 2107, ["classifier"], "Method"], [2117, 2120, ["meta", "-", "learning"], "Method"], [2138, 2142, ["few", "-", "shot", "learning"], "Task"], [2173, 2176, ["episode", "based", "training"], "Method"], [2287, 2292, ["episode", "-", "based", "training", "strategy"], "Method"], [2330, 2335, ["zero", "-", "shot", "learning", "case"], "Task"], [2350, 2352, ["Relation", "Network"], "Method"], [2353, 2354, ["RN"], "Method"], [2361, 2363, ["embedding", "module"], "Method"], [2365, 2367, ["relation", "module"], "Method"], [2392, 2394, ["embedding", "module"], "Method"], [2408, 2409, ["operator"], "Method"], [2443, 2445, ["relation", "module"], "Method"], [2464, 2466, ["relation", "score"], "Metric"], [2510, 2512, ["embedding", "module"], "Method"], [2530, 2535, ["class", "-", "level", "feature", "map"], "Method"], [2560, 2563, ["one", "-", "shot"], "Task"], [2564, 2568, ["few", "-", "shot", "setting"], "Task"], [2569, 2571, ["Objective", "function"], "Metric"], [2573, 2576, ["mean", "square", "error"], "Metric"], [2577, 2578, ["MSE"], "Metric"], [2596, 2598, ["relation", "score"], "Metric"], [2617, 2618, ["MSE"], "Metric"], [2631, 2633, ["classification", "problem"], "Task"], [2651, 2653, ["regression", "problem"], "Task"], [2668, 2672, ["Zero", "-", "shot", "Learning"], "Task"], [2672, 2676, ["Zero", "-", "shot", "learning"], "Method"], [2732, 2736, ["zero", "-", "shot", "case"], "Task"], [2764, 2767, ["heterogeneous", "embedding", "module"], "Method"], [2769, 2771, ["embedding", "module"], "Method"], [2780, 2782, ["relation", "net"], "Method"], [2790, 2792, ["relation", "score"], "Metric"], [2800, 2802, ["objective", "function"], "Metric"], [2803, 2807, ["zero", "-", "shot", "learning"], "Task"], [2813, 2817, ["few", "-", "shot", "learning"], "Task"], [2820, 2822, ["Network", "Architecture"], "Method"], [2831, 2833, ["convolutional", "blocks"], "Method"], [2834, 2836, ["embedding", "module"], "Method"], [2857, 2859, ["convolutional", "block"], "Method"], [2861, 2865, ["64", "-", "filter", "convolution"], "Method"], [2867, 2869, ["batch", "normalisation"], "Method"], [2871, 2874, ["ReLU", "nonlinearity", "layer"], "Method"], [2883, 2887, ["max", "-", "pooling", "layer"], "Method"], [2906, 2908, ["convolutional", "layers"], "Method"], [2910, 2912, ["relation", "module"], "Method"], [2914, 2916, ["relation", "module"], "Method"], [2919, 2921, ["convolutional", "blocks"], "Method"], [2923, 2927, ["fully", "-", "connected", "layers"], "Method"], [2930, 2932, ["convolutional", "block"], "Method"], [2934, 2935, ["convolution"], "Method"], [2936, 2938, ["64", "filters"], "Method"], [2940, 2942, ["batch", "normalisation"], "Method"], [2943, 2947, ["ReLU", "non", "-", "linearity"], "Method"], [2948, 2951, ["max", "-", "pooling"], "Method"], [2957, 2960, ["max", "pooling", "layer"], "Method"], [2963, 2964, ["Omniglot"], "Material"], [2965, 2967, ["mini", "ImageNet"], "Material"], [2971, 2975, ["fully", "-", "connected", "layers"], "Method"], [2984, 2988, ["fully", "-", "connected", "layers"], "Method"], [2989, 2990, ["ReLU"], "Method"], [3011, 3013, ["network", "architecture"], "Method"], [3033, 3035, ["DNN", "subnet"], "Method"], [3042, 3043, ["Inception"], "Method"], [3044, 3045, ["ResNet"], "Method"], [3048, 3049, ["ImageNet"], "Material"], [3067, 3068, ["Omniglot"], "Material"], [3069, 3071, ["mini", "Imagenet"], "Material"], [3073, 3077, ["zero", "-", "shot", "classification"], "Task"], [3078, 3081, ["Animals", "with", "Attributes"], "Material"], [3082, 3083, ["AwA"], "Material"], [3085, 3093, ["Caltech", "-", "UCSD", "Birds", "-", "200", "-", "2011"], "Material"], [3094, 3095, ["CUB"], "Material"], [3104, 3105, ["PyTorch"], "Method"], [3108, 3112, ["Few", "-", "shot", "Recognition"], "Task"], [3113, 3117, ["Few", "-", "shot", "learning"], "Task"], [3121, 3122, ["Adam"], "Method"], [3124, 3126, ["learning", "rate"], "Metric"], [3171, 3173, ["neural", "statistician"], "Method"], [3174, 3176, ["Matching", "Nets"], "Method"], [3179, 3182, ["fine", "-", "tuning"], "Method"], [3183, 3184, ["MANN"], "Method"], [3185, 3187, ["Siamese", "Nets"], "Method"], [3188, 3189, ["Memory"], "Method"], [3190, 3193, ["Convolutional", "Siamese", "Nets"], "Method"], [3194, 3195, ["MAML"], "Method"], [3196, 3198, ["Meta", "Nets"], "Method"], [3199, 3201, ["Prototypical", "Nets"], "Method"], [3202, 3206, ["Meta", "-", "Learner", "LSTM"], "Method"], [3211, 3212, ["Omniglot"], "Material"], [3278, 3284, ["5", "-", "way", "1", "-", "shot"], "Material"], [3290, 3296, ["5", "-", "way", "5", "-", "shot"], "Material"], [3302, 3308, ["20", "-", "way", "1", "-", "shot"], "Material"], [3387, 3395, ["1", "-", "shot", "and", "5", "-", "shot", "experiments"], "Task"], [3437, 3439, ["averaged", "accuracies"], "Metric"], [3459, 3460, ["accuracy"], "Metric"], [3490, 3492, ["mini", "ImageNet"], "Material"], [3538, 3539, ["validation"], "Task"], [3540, 3541, ["testing"], "Task"], [3551, 3554, ["monitoring", "generalisation", "performance"], "Task"], [3565, 3569, ["few", "-", "shot", "learning"], "Task"], [3574, 3583, ["way", "1", "-", "shot", "and", "5", "-", "shot", "classification"], "Task"], [3590, 3596, ["5", "-", "way", "1", "-", "shot"], "Material"], [3603, 3609, ["5", "-", "way", "5", "-", "shot"], "Material"], [3641, 3647, ["5", "-", "way", "1", "-", "shot"], "Material"], [3668, 3670, ["random", "initialisation"], "Method"], [3694, 3702, ["1", "-", "shot", "and", "5", "-", "shot", "scenarios"], "Task"], [3756, 3762, ["5", "-", "way", "5", "-", "shot"], "Material"], [3772, 3774, ["prototypical", "networks"], "Method"], [3820, 3824, ["1", "-", "shot", "evaluation"], "Task"], [3846, 3849, ["1", "-", "shot"], "Material"], [3869, 3873, ["Zero", "-", "shot", "Recognition"], "Task"], [3879, 3881, ["ZSL", "settings"], "Method"], [3888, 3890, ["GBU", "setting"], "Method"], [3907, 3908, ["ZSL"], "Method"], [3920, 3924, ["ImageNet", "1", "K", "classes"], "Material"], [3932, 3935, ["image", "embedding", "network"], "Method"], [3949, 3951, ["GBU", "setting"], "Method"], [3964, 3968, ["ImageNet", "1", "K", "classes"], "Material"], [4003, 4008, ["generalised", "zero", "-", "shot", "learning"], "Method"], [4009, 4010, ["GZSL"], "Method"], [4021, 4023, ["ZSL", "benchmarks"], "Method"], [4030, 4031, ["AwA"], "Material"], [4032, 4035, ["Animals", "with", "Attributes"], "Material"], [4099, 4101, ["GBU", "setting"], "Task"], [4102, 4103, ["AwA1"], "Material"], [4104, 4105, ["AwA2"], "Material"], [4106, 4107, ["CUB"], "Material"], [4111, 4112, ["AwA2"], "Material"], [4126, 4127, ["AwA"], "Material"], [4128, 4129, ["AwA1"], "Material"], [4132, 4133, ["AwA"], "Material"], [4139, 4141, ["Semantic", "representation"], "Method"], [4142, 4143, ["AwA"], "Material"], [4168, 4169, ["CUB"], "Material"], [4187, 4189, ["embedding", "modules"], "Method"], [4197, 4201, ["zero", "-", "shot", "learning"], "Task"], [4208, 4211, ["Inception", "-", "V2"], "Method"], [4213, 4217, ["query", "image", "embedding", "DNN"], "Method"], [4224, 4225, ["ResNet101"], "Method"], [4227, 4231, ["GBU", "and", "generalised", "setting"], "Task"], [4235, 4237, ["pooling", "units"], "Method"], [4238, 4240, ["image", "embedding"], "Method"], [4246, 4247, ["DNN"], "Method"], [4252, 4257, ["ILSVRC", "2012", "1", "K", "classification"], "Task"], [4258, 4261, ["fine", "-", "tuning"], "Task"], [4265, 4268, ["deep", "ZSL", "works"], "Method"], [4270, 4272, ["MLP", "network"], "Method"], [4283, 4286, ["hidden", "layer", "FC1"], "Method"], [4299, 4300, ["AwA"], "Material"], [4301, 4302, ["CUB"], "Material"], [4308, 4309, ["FC2"], "Method"], [4317, 4319, ["image", "embedding"], "Task"], [4325, 4327, ["relation", "module"], "Method"], [4339, 4340, ["MLPs"], "Method"], [4341, 4344, ["hidden", "layer", "FC3"], "Method"], [4349, 4350, ["AwA"], "Material"], [4351, 4352, ["CUB"], "Material"], [4357, 4359, ["weight", "decay"], "Method"], [4360, 4362, ["L2", "regularisation"], "Method"], [4364, 4365, ["FC1"], "Method"], [4371, 4373, ["hubness", "problem"], "Task"], [4374, 4378, ["cross", "-", "modal", "mapping"], "Task"], [4379, 4380, ["ZSL"], "Task"], [4397, 4398, ["regularisation"], "Method"], [4402, 4403, ["FC3"], "Method"], [4406, 4408, ["relation", "module"], "Method"], [4417, 4419, ["semantic", "representation"], "Method"], [4428, 4430, ["visual", "representation"], "Method"], [4433, 4435, ["hubness", "problem"], "Task"], [4443, 4448, ["L2", "regularisation", "/", "weight", "decay"], "Method"], [4453, 4455, ["ZSL", "models"], "Method"], [4462, 4464, ["embedding", "network"], "Method"], [4466, 4468, ["learning", "rate"], "Metric"], [4472, 4473, ["Adam"], "Method"], [4493, 4494, ["ZSL"], "Task"], [4539, 4542, ["sample", "class", "embedding"], "Method"], [4549, 4550, ["AwA"], "Material"], [4571, 4573, ["prototypical", "networks"], "Method"], [4581, 4583, ["inductive", "methods"], "Method"], [4601, 4603, ["model", "training"], "Task"], [4656, 4658, ["ZSL", "models"], "Method"], [4666, 4668, ["shallow", "models’"], "Method"], [4698, 4699, ["AwA2"], "Material"], [4700, 4701, ["CUB"], "Material"], [4716, 4722, ["harmonic", "mean", "(", "H", ")", "metric"], "Metric"], [4725, 4726, ["AwA1"], "Material"], [4733, 4734, ["DEM"], "Method"], [4739, 4741, ["Relation", "Network"], "Method"], [4760, 4762, ["distance", "metrics"], "Metric"], [4764, 4768, ["Euclidean", "or", "cosine", "distance"], "Method"], [4770, 4771, ["classification"], "Task"], [4778, 4781, ["distance", "metric", "learning"], "Method"], [4786, 4787, ["learning"], "Task"], [4790, 4792, ["feature", "embedding"], "Method"], [4808, 4811, ["metric", "learning", "approaches"], "Method"], [4816, 4822, ["shallow", "(", "linear", ")", "Mahalanobis", "metric"], "Method"], [4825, 4827, ["feature", "representation"], "Method"], [4844, 4846, ["Relation", "Network"], "Method"], [4853, 4855, ["deep", "embedding"], "Method"], [4882, 4885, ["few", "short", "learning"], "Task"], [4897, 4899, ["function", "approximator"], "Method"], [4931, 4933, ["Fixed", "metrics"], "Metric"], [4949, 4951, ["linear", "separability"], "Method"], [4953, 4954, ["embedding"], "Method"], [4966, 4968, ["embedding", "network"], "Method"], [4978, 4980, ["embedding", "networks"], "Method"], [4982, 4984, ["discriminative", "representations"], "Method"], [4989, 4991, ["deep", "learning"], "Task"], [4992, 4997, ["non", "-", "linear", "similarity", "metric"], "Method"], [5000, 5001, ["embedding"], "Method"], [5002, 5004, ["Relation", "Network"], "Method"], [5030, 5031, ["Visualisation"], "Task"], [5055, 5057, ["Relation", "Network"], "Method"], [5076, 5078, ["relation", "module"], "Method"], [5129, 5131, ["embedding", "modules"], "Method"], [5138, 5142, ["Euclidean", "NN", ")", "comparison"], "Method"], [5161, 5162, ["matching"], "Task"], [5192, 5203, ["2", "-", "hidden", "layer", "MLP", "embedding", "of", "query", "and", "sample", "inputs"], "Method"], [5208, 5210, ["Mahalanobis", "metric"], "Metric"], [5226, 5227, ["similarity"], "Task"], [5249, 5251, ["comparing", "embeddings"], "Task"], [5342, 5343, ["Mahalanobis"], "Method"], [5380, 5383, ["2D", "PCA", "representation"], "Method"], [5405, 5407, ["relation", "network"], "Method"], [5435, 5437, ["Relation", "Network"], "Method"], [5438, 5441, ["few", "-", "shot"], "Task"], [5442, 5446, ["zero", "-", "shot", "learning"], "Task"], [5447, 5449, ["Relation", "network"], "Method"], [5451, 5452, ["embedding"], "Method"], [5454, 5460, ["deep", "non", "-", "linear", "distance", "metric"], "Method"], [5476, 5478, ["episodic", "training"], "Method"], [5480, 5484, ["embedding", "and", "distance", "metric"], "Metric"], [5486, 5490, ["few", "-", "shot", "learning"], "Task"], [5528, 5533, ["generalised", "zero", "-", "shot", "settings"], "Task"], [335, 337, ["deep", "learning"], "Task"], [483, 485, ["few", "-"], "Task"], [546, 549, ["one", "-", "shot"], "Task"], [567, 569, ["few", "-"], "Task"], [583, 584, ["RNN"], "Method"], [610, 613, ["one", "-", "shot"], "Task"], [726, 728, ["few", "-"], "Task"], [737, 739, ["few", "-"], "Task"], [916, 920, ["zero", "-", "shot", "learning"], "Task"], [953, 955, ["few", "-"], "Task"], [989, 991, ["deep", "learning"], "Task"], [1014, 1016, ["deep", "learning"], "Task"], [1019, 1021, ["few", "-"], "Task"], [1076, 1078, ["few", "-"], "Task"], [1131, 1133, ["few", "-"], "Task"], [1236, 1238, ["few", "-"], "Task"], [1324, 1325, ["RNN"], "Method"], [1508, 1510, ["few", "-"], "Task"], [1592, 1594, ["few", "-"], "Task"], [1798, 1799, ["RNN"], "Method"], [1898, 1900, ["few", "-"], "Task"], [1979, 1981, ["few", "-"], "Task"], [2050, 2052, ["few", "-"], "Task"], [2167, 2169, ["few", "-"], "Task"], [2295, 2297, ["few", "-"], "Task"], [2308, 2311, ["one", "-", "shot"], "Task"], [2473, 2476, ["one", "-", "shot"], "Task"], [2679, 2682, ["one", "-", "shot"], "Task"], [2705, 2708, ["one", "-", "shot"], "Task"], [2824, 2826, ["few", "-"], "Task"], [3015, 3019, ["zero", "-", "shot", "learning"], "Task"], [3062, 3064, ["few", "-"], "Task"], [3165, 3167, ["few", "-"], "Task"], [3209, 3210, ["Omniglot"], "Material"], [3353, 3359, ["5", "-", "way", "1", "-", "shot"], "Material"], [3366, 3368, ["few", "-"], "Task"], [3372, 3373, ["Omniglot"], "Material"], [3445, 3451, ["5", "-", "way", "5", "-", "shot"], "Material"], [3496, 3498, ["mini", "Imagenet"], "Material"], [3704, 3706, ["few", "-"], "Task"], [3745, 3751, ["5", "-", "way", "1", "-", "shot"], "Material"], [4064, 4065, ["CUB"], "Material"], [4066, 4074, ["Caltech", "-", "UCSD", "Birds", "-", "200", "-", "2011"], "Material"], [4563, 4564, ["CUB"], "Material"], [4711, 4712, ["GZSL"], "Method"], [4751, 4753, ["few", "-"], "Task"], [5064, 5066, ["relation", "module"], "Method"], [5167, 5169, ["relation", "module"], "Method"], [5223, 5225, ["relation", "module"], "Method"], [5273, 5274, ["Omniglot"], "Material"], [5394, 5396, ["relation", "module"], "Method"], [5501, 5503, ["few", "-"], "Task"]]} {"doc_id": "4682e5d4654b6040f213475c8b75ca039661b5f6", "text": ["document", ":", "EAST", ":", "An", "Efficient", "and", "Accurate", "Scene", "Text", "Detector", "Previous", "approaches", "for", "scene", "text", "detection", "have", "already", "achieved", "promising", "performances", "across", "various", "benchmarks", ".", "However", ",", "they", "usually", "fall", "short", "when", "dealing", "with", "challenging", "scenarios", ",", "even", "when", "equipped", "with", "deep", "neural", "network", "models", ",", "because", "the", "overall", "performance", "is", "determined", "by", "the", "interplay", "of", "multiple", "stages", "and", "components", "in", "the", "pipelines", ".", "In", "this", "work", ",", "we", "propose", "a", "simple", "yet", "powerful", "pipeline", "that", "yields", "fast", "and", "accurate", "text", "detection", "in", "natural", "scenes", ".", "The", "pipeline", "directly", "predicts", "words", "or", "text", "lines", "of", "arbitrary", "orientations", "and", "quadrilateral", "shapes", "in", "full", "images", ",", "eliminating", "unnecessary", "intermediate", "steps", "(", "e.g.", ",", "candidate", "aggregation", "and", "word", "partitioning", ")", ",", "with", "a", "single", "neural", "network", ".", "The", "simplicity", "of", "our", "pipeline", "allows", "concentrating", "efforts", "on", "designing", "loss", "functions", "and", "neural", "network", "architecture", ".", "Experiments", "on", "standard", "datasets", "including", "ICDAR", "2015", ",", "COCO", "-", "Text", "and", "MSRA", "-", "TD500", "demonstrate", "that", "the", "proposed", "algorithm", "significantly", "outperforms", "state", "-", "of", "-", "the", "-", "art", "methods", "in", "terms", "of", "both", "accuracy", "and", "efficiency", ".", "On", "the", "ICDAR", "2015", "dataset", ",", "the", "proposed", "algorithm", "achieves", "an", "F", "-", "score", "of", "0.7820", "at", "13.2fps", "at", "720p", "resolution", ".", "Megvii", "Technology", "Inc.", ",", "Beijing", ",", "China", "{", "zxy", ",", "yaocong", ",", "wenhe", ",", "wangyuzhi", ",", "zsc", ",", "hwr", ",", "liangjiajun}@megvii.com", "section", ":", "Introduction", "Recently", ",", "extracting", "and", "understanding", "textual", "information", "embodied", "in", "natural", "scenes", "have", "become", "increasingly", "important", "and", "popular", ",", "which", "is", "evidenced", "by", "the", "unprecedented", "large", "numbers", "of", "participants", "of", "the", "ICDAR", "series", "contests", "and", "the", "launch", "of", "the", "TRAIT", "2016", "evaluation", "by", "NIST", ".", "Text", "detection", ",", "as", "a", "prerequisite", "of", "the", "subsequent", "processes", ",", "plays", "a", "critical", "role", "in", "the", "whole", "procedure", "of", "textual", "information", "extraction", "and", "understanding", ".", "Previous", "text", "detection", "approaches", "have", "already", "obtained", "promising", "performances", "on", "various", "benchmarks", "in", "this", "field", ".", "The", "core", "of", "text", "detection", "is", "the", "design", "of", "features", "to", "distinguish", "text", "from", "backgrounds", ".", "Traditionally", ",", "features", "are", "manually", "designed", "to", "capture", "the", "properties", "of", "scene", "text", ",", "while", "in", "deep", "learning", "based", "methods", "effective", "features", "are", "directly", "learned", "from", "training", "data", ".", "compat=1.8", "every", "tick", "label", "/", ".append", "style", "=", "font=", "grid", "style", "=", "dotted", ",", "gray", "[", "xmin=0", ",", "xmax=17", ",", "ymin=0.49", ",", "ymax=0.85", ",", "grid", "=", "both", ",", "axis", "lines", "=", "middle", ",", "line", "width=.8pt", ",", "width=1.1height=.7xlabel=", "Speed", "(", "FPS", ")", ",", "ylabel=", "F", "-", "score", "]", "[", "only", "marks", ",", "draw", "=", "black", ",", "fill", "=", "blue", ",", "mark", "size=3pt", "]", "table", "x", "y", "+", "7.14", "+", "0.6085", "+", "1.61", "+", "0.6477", "+", "0.476", "+", "0.5358", ";", "[", "only", "marks", ",", "draw", "=", "black", ",", "fill", "=", "red", ",", "mark", "size=3pt", "]", "table", "x", "y", "+", "13.245e", "+", "00", "+", "0.7820", "+", "16.779e", "+", "00", "+", "0.7571", ";", "at", "(", "axis", "cs:7.3", ",", "0.58", ")", "[", "anchor", "=", "base", "west", ",", "text", "=", "black", ",", "rotate=0.0", ",", "align", "=", "left", "]", "Tian", "(", "0.609@7.14fps", ")", ";", "at", "(", "axis", "cs:1.3", ",", "0.6677", ")", "[", "anchor", "=", "base", "west", ",", "text", "=", "black", ",", "rotate=0.0", ",", "align", "=", "left", "]", "Yao", "(", "0.648@1.61fps", ")", ";", "at", "(", "axis", "cs:0.6", ",", "0.505", ")", "[", "anchor", "=", "base", "west", ",", "text", "=", "black", ",", "rotate=0.0", ",", "align", "=", "left", "]", "Zhang", "(", "0.532@0.476fps", ")", ";", "at", "(", "axis", "cs:13.0", ",", "0.78", ")", "[", "anchor", "=", "base", "east", ",", "text", "=", "black", ",", "rotate=0.0", ",", "align", "=", "right", "]", "Ours", "+", "PVANet2x", "(", "0.782@13.2fps", ")", ";", "at", "(", "axis", "cs:16.8", ",", "0.69", ")", "[", "anchor", "=", "base", "east", ",", "text", "=", "black", ",", "rotate=0.0", ",", "align", "=", "right", "]", "Ours", "+", "PVANet", "(", "0.757@16.8fps", ")", ";", "However", ",", "existing", "methods", ",", "either", "conventional", "or", "deep", "neural", "network", "based", ",", "mostly", "consist", "of", "several", "stages", "and", "components", ",", "which", "are", "probably", "sub", "-", "optimal", "and", "time", "-", "consuming", ".", "Therefore", ",", "the", "accuracy", "and", "efficiency", "of", "such", "methods", "are", "still", "far", "from", "satisfactory", ".", "In", "this", "paper", ",", "we", "propose", "a", "fast", "and", "accurate", "scene", "text", "detection", "pipeline", "that", "has", "only", "two", "stages", ".", "The", "pipeline", "utilizes", "a", "fully", "convolutional", "network", "(", "FCN", ")", "model", "that", "directly", "produces", "word", "or", "text", "-", "line", "level", "predictions", ",", "excluding", "redundant", "and", "slow", "intermediate", "steps", ".", "The", "produced", "text", "predictions", ",", "which", "can", "be", "either", "rotated", "rectangles", "or", "quadrangles", ",", "are", "sent", "to", "Non", "-", "Maximum", "Suppression", "to", "yield", "final", "results", ".", "Compared", "with", "existing", "methods", ",", "the", "proposed", "algorithm", "achieves", "significantly", "enhanced", "performance", ",", "while", "running", "much", "faster", ",", "according", "to", "the", "qualitative", "and", "quantitative", "experiments", "on", "standard", "benchmarks", ".", "Specifically", ",", "the", "proposed", "algorithm", "achieves", "an", "F", "-", "score", "of", "0.7820", "on", "ICDAR", "2015", "(", "0.8072", "when", "tested", "in", "multi", "-", "scale", ")", ",", "0.7608", "on", "MSRA", "-", "TD500", "and", "0.3945", "on", "COCO", "-", "Text", ",", "outperforming", "previous", "state", "-", "of", "-", "the", "-", "art", "algorithms", "in", "performance", "while", "taking", "much", "less", "time", "on", "average", "(", "13.2fps", "at", "720p", "resolution", "on", "a", "Titan", "-", "X", "GPU", "for", "our", "best", "performing", "model", ",", "16.8fps", "for", "our", "fastest", "model", ")", ".", "The", "contributions", "of", "this", "work", "are", "three", "-", "fold", ":", "We", "propose", "a", "scene", "text", "detection", "method", "that", "consists", "of", "two", "stages", ":", "a", "Fully", "Convolutional", "Network", "and", "an", "NMS", "merging", "stage", ".", "The", "FCN", "directly", "produces", "text", "regions", ",", "excluding", "redundant", "and", "time", "-", "consuming", "intermediate", "steps", ".", "The", "pipeline", "is", "flexible", "to", "produce", "either", "word", "level", "or", "line", "level", "predictions", ",", "whose", "geometric", "shapes", "can", "be", "rotated", "boxes", "or", "quadrangles", ",", "depending", "on", "specific", "applications", ".", "The", "proposed", "algorithm", "significantly", "outperforms", "state", "-", "of", "-", "the", "-", "art", "methods", "in", "both", "accuracy", "and", "speed", ".", "section", ":", "Related", "Work", "Scene", "text", "detection", "and", "recognition", "have", "been", "active", "research", "topics", "in", "computer", "vision", "for", "a", "long", "period", "of", "time", ".", "Numerous", "inspiring", "ideas", "and", "effective", "approaches", "have", "been", "investigated", ".", "Comprehensive", "reviews", "and", "detailed", "analyses", "can", "be", "found", "in", "survey", "papers", ".", "This", "section", "will", "focus", "on", "works", "that", "are", "mostly", "relevant", "to", "the", "proposed", "algorithm", ".", "Conventional", "approaches", "rely", "on", "manually", "designed", "features", ".", "Stroke", "Width", "Transform", "(", "SWT", ")", "and", "Maximally", "Stable", "Extremal", "Regions", "(", "MSER", ")", "based", "methods", "generally", "seek", "character", "candidates", "via", "edge", "detection", "or", "extremal", "region", "extraction", ".", "Zhang", "made", "use", "of", "the", "local", "symmetry", "property", "of", "text", "and", "designed", "various", "features", "for", "text", "region", "detection", ".", "FASText", "is", "a", "fast", "text", "detection", "system", "that", "adapted", "and", "modified", "the", "well", "-", "known", "FAST", "key", "point", "detector", "for", "stroke", "extraction", ".", "However", ",", "these", "methods", "fall", "behind", "of", "those", "based", "on", "deep", "neural", "networks", ",", "in", "terms", "of", "both", "accuracy", "and", "adaptability", ",", "especially", "when", "dealing", "with", "challenging", "scenarios", ",", "such", "as", "low", "resolution", "and", "geometric", "distortion", ".", "Recently", ",", "the", "area", "of", "scene", "text", "detection", "has", "entered", "a", "new", "era", "that", "deep", "neural", "network", "based", "algorithms", "have", "gradually", "become", "the", "mainstream", ".", "Huang", "first", "found", "candidates", "using", "MSER", "and", "then", "employed", "a", "deep", "convolutional", "network", "as", "a", "strong", "classifier", "to", "prune", "false", "positives", ".", "The", "method", "of", "Jaderberg", "scanned", "the", "image", "in", "a", "sliding", "-", "window", "fashion", "and", "produced", "a", "dense", "heatmap", "for", "each", "scale", "with", "a", "convolutional", "neural", "network", "model", ".", "Later", ",", "Jaderberg", "employed", "both", "a", "CNN", "and", "an", "ACF", "to", "hunt", "word", "candidates", "and", "further", "refined", "them", "using", "regression", ".", "Tian", "developed", "vertical", "anchors", "and", "constructed", "a", "CNN", "-", "RNN", "joint", "model", "to", "detect", "horizontal", "text", "lines", ".", "Different", "from", "these", "methods", ",", "Zhang", "proposed", "to", "utilize", "FCN", "for", "heatmap", "generation", "and", "to", "use", "component", "projection", "for", "orientation", "estimation", ".", "These", "methods", "obtained", "excellent", "performance", "on", "standard", "benchmarks", ".", "However", ",", "as", "illustrated", "in", "Fig", ".", "[", "reference", "]", "(", "a", "-", "d", ")", ",", "they", "mostly", "consist", "of", "multiple", "stages", "and", "components", ",", "such", "as", "false", "positive", "removal", "by", "post", "filtering", ",", "candidate", "aggregation", ",", "line", "formation", "and", "word", "partition", ".", "The", "multitude", "of", "stages", "and", "components", "may", "require", "exhaustive", "tuning", ",", "leading", "to", "sub", "-", "optimal", "performance", ",", "and", "add", "to", "processing", "time", "of", "the", "whole", "pipeline", ".", "In", "this", "paper", ",", "we", "devise", "a", "deep", "FCN", "-", "based", "pipeline", "that", "directly", "targets", "the", "final", "goal", "of", "text", "detection", ":", "word", "or", "text", "-", "line", "level", "detection", ".", "As", "depicted", "in", "Fig", ".", "[", "reference", "]", "(", "e", ")", ",", "the", "model", "abandons", "unnecessary", "intermediate", "components", "and", "steps", ",", "and", "allows", "for", "end", "-", "to", "-", "end", "training", "and", "optimization", ".", "The", "resultant", "system", ",", "equipped", "with", "a", "single", ",", "light", "-", "weighted", "neural", "network", ",", "surpasses", "all", "previous", "methods", "by", "an", "obvious", "margin", "in", "both", "performance", "and", "speed", ".", "section", ":", "Methodology", "The", "key", "component", "of", "the", "proposed", "algorithm", "is", "a", "neural", "network", "model", ",", "which", "is", "trained", "to", "directly", "predict", "the", "existence", "of", "text", "instances", "and", "their", "geometries", "from", "full", "images", ".", "The", "model", "is", "a", "fully", "-", "convolutional", "neural", "network", "adapted", "for", "text", "detection", "that", "outputs", "dense", "per", "-", "pixel", "predictions", "of", "words", "or", "text", "lines", ".", "This", "eliminates", "intermediate", "steps", "such", "as", "candidate", "proposal", ",", "text", "region", "formation", "and", "word", "partition", ".", "The", "post", "-", "processing", "steps", "only", "include", "thresholding", "and", "NMS", "on", "predicted", "geometric", "shapes", ".", "The", "detector", "is", "named", "as", "EAST", ",", "since", "it", "is", "an", "E", "fficient", "and", "A", "ccuracy", "S", "cene", "T", "ext", "detection", "pipeline", ".", "subsection", ":", "Pipeline", "A", "high", "-", "level", "overview", "of", "our", "pipeline", "is", "illustrated", "in", "Fig", ".", "[", "reference", "]", "(", "e", ")", ".", "The", "algorithm", "follows", "the", "general", "design", "of", "DenseBox", ",", "in", "which", "an", "image", "is", "fed", "into", "the", "FCN", "and", "multiple", "channels", "of", "pixel", "-", "level", "text", "score", "map", "and", "geometry", "are", "generated", ".", "One", "of", "the", "predicted", "channels", "is", "a", "score", "map", "whose", "pixel", "values", "are", "in", "the", "range", "of", ".", "The", "remaining", "channels", "represent", "geometries", "that", "encloses", "the", "word", "from", "the", "view", "of", "each", "pixel", ".", "The", "score", "stands", "for", "the", "confidence", "of", "the", "geometry", "shape", "predicted", "at", "the", "same", "location", ".", "We", "have", "experimented", "with", "two", "geometry", "shapes", "for", "text", "regions", ",", "rotated", "box", "(", "RBOX", ")", "and", "quadrangle", "(", "QUAD", ")", ",", "and", "designed", "different", "loss", "functions", "for", "each", "geometry", ".", "Thresholding", "is", "then", "applied", "to", "each", "predicted", "region", ",", "where", "the", "geometries", "whose", "scores", "are", "over", "the", "predefined", "threshold", "is", "considered", "valid", "and", "saved", "for", "later", "non", "-", "maximum", "-", "suppression", ".", "Results", "after", "NMS", "are", "considered", "the", "final", "output", "of", "the", "pipeline", ".", "subsection", ":", "Network", "Design", "Several", "factors", "must", "be", "taken", "into", "account", "when", "designing", "neural", "networks", "for", "text", "detection", ".", "Since", "the", "sizes", "of", "word", "regions", ",", "as", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "vary", "tremendously", ",", "determining", "the", "existence", "of", "large", "words", "would", "require", "features", "from", "late", "-", "stage", "of", "a", "neural", "network", ",", "while", "predicting", "accurate", "geometry", "enclosing", "a", "small", "word", "regions", "need", "low", "-", "level", "information", "in", "early", "stages", ".", "Therefore", "the", "network", "must", "use", "features", "from", "different", "levels", "to", "fulfill", "these", "requirements", ".", "HyperNet", "meets", "these", "conditions", "on", "features", "maps", ",", "but", "merging", "a", "large", "number", "of", "channels", "on", "large", "feature", "maps", "would", "significantly", "increase", "the", "computation", "overhead", "for", "later", "stages", ".", "In", "remedy", "of", "this", ",", "we", "adopt", "the", "idea", "from", "U", "-", "shape", "to", "merge", "feature", "maps", "gradually", ",", "while", "keeping", "the", "up", "-", "sampling", "branches", "small", ".", "Together", "we", "end", "up", "with", "a", "network", "that", "can", "both", "utilize", "different", "levels", "of", "features", "and", "keep", "a", "small", "computation", "cost", ".", "A", "schematic", "view", "of", "our", "model", "is", "depicted", "in", "Fig", ".", "[", "reference", "]", ".", "The", "model", "can", "be", "decomposed", "in", "to", "three", "parts", ":", "feature", "extractor", "stem", ",", "feature", "-", "merging", "branch", "and", "output", "layer", ".", "The", "stem", "can", "be", "a", "convolutional", "network", "pre", "-", "trained", "on", "ImageNet", "dataset", ",", "with", "interleaving", "convolution", "and", "pooling", "layers", ".", "Four", "levels", "of", "feature", "maps", ",", "denoted", "as", ",", "are", "extracted", "from", "the", "stem", ",", "whose", "sizes", "are", ",", ",", "and", "of", "the", "input", "image", ",", "respectively", ".", "In", "Fig", ".", "[", "reference", "]", ",", "PVANet", "is", "depicted", ".", "In", "our", "experiments", ",", "we", "also", "adopted", "the", "well", "-", "known", "VGG16", "model", ",", "where", "feature", "maps", "after", "pooling", "-", "2", "to", "pooling", "-", "5", "are", "extracted", ".", "In", "the", "feature", "-", "merging", "branch", ",", "we", "gradually", "merge", "them", ":", "where", "is", "the", "merge", "base", ",", "and", "is", "the", "merged", "feature", "map", ",", "and", "the", "operator", "represents", "concatenation", "along", "the", "channel", "axis", ".", "In", "each", "merging", "stage", ",", "the", "feature", "map", "from", "the", "last", "stage", "is", "first", "fed", "to", "an", "unpooling", "layer", "to", "double", "its", "size", ",", "and", "then", "concatenated", "with", "the", "current", "feature", "map", ".", "Next", ",", "a", "bottleneck", "cuts", "down", "the", "number", "of", "channels", "and", "reduces", "computation", ",", "followed", "by", "a", "that", "fuses", "the", "information", "to", "finally", "produce", "the", "output", "of", "this", "merging", "stage", ".", "Following", "the", "last", "merging", "stage", ",", "a", "layer", "produces", "the", "final", "feature", "map", "of", "the", "merging", "branch", "and", "feed", "it", "to", "the", "output", "layer", ".", "The", "number", "of", "output", "channels", "for", "each", "convolution", "is", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "We", "keep", "the", "number", "of", "channels", "for", "convolutions", "in", "branch", "small", ",", "which", "adds", "only", "a", "fraction", "of", "computation", "overhead", "over", "the", "stem", ",", "making", "the", "network", "computation", "-", "efficient", ".", "The", "final", "output", "layer", "contains", "several", "operations", "to", "project", "32", "channels", "of", "feature", "maps", "into", "1", "channel", "of", "score", "map", "and", "a", "multi", "-", "channel", "geometry", "map", ".", "The", "geometry", "output", "can", "be", "either", "one", "of", "RBOX", "or", "QUAD", ",", "summarized", "in", "Tab", ".", "[", "reference", "]", "For", "RBOX", ",", "the", "geometry", "is", "represented", "by", "4", "channels", "of", "axis", "-", "aligned", "bounding", "box", "(", "AABB", ")", "and", "1", "channel", "rotation", "angle", ".", "The", "formulation", "of", "is", "the", "same", "as", "that", "in", ",", "where", "the", "4", "channels", "represents", "4", "distances", "from", "the", "pixel", "location", "to", "the", "top", ",", "right", ",", "bottom", ",", "left", "boundaries", "of", "the", "rectangle", "respectively", ".", "For", "QUAD", ",", "we", "use", "8", "numbers", "to", "denote", "the", "coordinate", "shift", "from", "four", "corner", "vertices", "of", "the", "quadrangle", "to", "the", "pixel", "location", ".", "As", "each", "distance", "offset", "contains", "two", "numbers", ",", "the", "geometry", "output", "contains", "8", "channels", ".", "subsection", ":", "Label", "Generation", "subsubsection", ":", "Score", "Map", "Generation", "for", "Quadrangle", "Without", "loss", "of", "generality", ",", "we", "only", "consider", "the", "case", "where", "the", "geometry", "is", "a", "quadrangle", ".", "The", "positive", "area", "of", "the", "quadrangle", "on", "the", "score", "map", "is", "designed", "to", "be", "roughly", "a", "shrunk", "version", "of", "the", "original", "one", ",", "illustrated", "in", "Fig", ".", "[", "reference", "]", "(", "a", ")", ".", "For", "a", "quadrangle", ",", "where", "are", "vertices", "on", "the", "quadrangle", "in", "clockwise", "order", ".", "To", "shrink", ",", "we", "first", "compute", "a", "reference", "length", "for", "each", "vertex", "as", "where", "is", "the", "distance", "between", "and", ".", "We", "first", "shrink", "the", "two", "longer", "edges", "of", "a", "quadrangle", ",", "and", "then", "the", "two", "shorter", "ones", ".", "For", "each", "pair", "of", "two", "opposing", "edges", ",", "we", "determine", "the", "“", "longer", "”", "pair", "by", "comparing", "the", "mean", "of", "their", "lengths", ".", "For", "each", "edge", ",", "we", "shrink", "it", "by", "moving", "its", "two", "endpoints", "inward", "along", "the", "edge", "by", "and", "respectively", ".", "subsubsection", ":", "Geometry", "Map", "Generation", "As", "discussed", "in", "Sec", ".", "[", "reference", "]", ",", "the", "geometry", "map", "is", "either", "one", "of", "RBOX", "or", "QUAD", ".", "The", "generation", "process", "for", "RBOX", "is", "illustrated", "in", "Fig", ".", "[", "reference", "]", "(", "c", "-", "e", ")", ".", "For", "those", "datasets", "whose", "text", "regions", "are", "annotated", "in", "QUAD", "style", "(", ",", "ICDAR", "2015", ")", ",", "we", "first", "generate", "a", "rotated", "rectangle", "that", "covers", "the", "region", "with", "minimal", "area", ".", "Then", "for", "each", "pixel", "which", "has", "positive", "score", ",", "we", "calculate", "its", "distances", "to", "the", "4", "boundaries", "of", "the", "text", "box", ",", "and", "put", "them", "to", "the", "4", "channels", "of", "RBOX", "ground", "truth", ".", "For", "the", "QUAD", "ground", "truth", ",", "the", "value", "of", "each", "pixel", "with", "positive", "score", "in", "the", "8", "-", "channel", "geometry", "map", "is", "its", "coordinate", "shift", "from", "the", "4", "vertices", "of", "the", "quadrangle", ".", "subsection", ":", "Loss", "Functions", "The", "loss", "can", "be", "formulated", "as", "where", "and", "represents", "the", "losses", "for", "the", "score", "map", "and", "the", "geometry", ",", "respectively", ",", "and", "weighs", "the", "importance", "between", "two", "losses", ".", "In", "our", "experiment", ",", "we", "set", "to", "1", ".", "subsubsection", ":", "Loss", "for", "Score", "Map", "In", "most", "state", "-", "of", "-", "the", "-", "art", "detection", "pipelines", ",", "training", "images", "are", "carefully", "processed", "by", "balanced", "sampling", "and", "hard", "negative", "mining", "to", "tackle", "with", "the", "imbalanced", "distribution", "of", "target", "objects", ".", "Doing", "so", "would", "potentially", "improve", "the", "network", "performance", ".", "However", ",", "using", "such", "techniques", "inevitably", "introduces", "a", "non", "-", "differentiable", "stage", "and", "more", "parameters", "to", "tune", "and", "a", "more", "complicated", "pipeline", ",", "which", "contradicts", "our", "design", "principle", ".", "To", "facilitate", "a", "simpler", "training", "procedure", ",", "we", "use", "class", "-", "balanced", "cross", "-", "entropy", "introduced", "in", ",", "given", "by", "where", "is", "the", "prediction", "of", "the", "score", "map", ",", "and", "is", "the", "ground", "truth", ".", "The", "parameter", "is", "the", "balancing", "factor", "between", "positive", "and", "negative", "samples", ",", "given", "by", "This", "balanced", "cross", "-", "entropy", "loss", "is", "first", "adopted", "in", "text", "detection", "by", "Yao", "as", "the", "objective", "function", "for", "score", "map", "prediction", ".", "We", "find", "it", "works", "well", "in", "practice", ".", "subsubsection", ":", "Loss", "for", "Geometries", "One", "challenge", "for", "text", "detection", "is", "that", "the", "sizes", "of", "text", "in", "natural", "scene", "images", "vary", "tremendously", ".", "Directly", "using", "L1", "or", "L2", "loss", "for", "regression", "would", "guide", "the", "loss", "bias", "towards", "larger", "and", "longer", "text", "regions", ".", "As", "we", "need", "to", "generate", "accurate", "text", "geometry", "prediction", "for", "both", "large", "and", "small", "text", "regions", ",", "the", "regression", "loss", "should", "be", "scale", "-", "invariant", ".", "Therefore", ",", "we", "adopt", "the", "IoU", "loss", "in", "the", "AABB", "part", "of", "RBOX", "regression", ",", "and", "a", "scale", "-", "normalized", "smoothed", "-", "L1", "loss", "for", "QUAD", "regression", ".", "paragraph", ":", "RBOX", "For", "the", "AABB", "part", ",", "we", "adopt", "IoU", "loss", "in", ",", "since", "it", "is", "invariant", "against", "objects", "of", "different", "scales", ".", "where", "represents", "the", "predicted", "AABB", "geometry", "and", "is", "its", "corresponding", "ground", "truth", ".", "It", "is", "easy", "to", "see", "that", "the", "width", "and", "height", "of", "the", "intersected", "rectangle", "are", "where", ",", ",", "and", "represents", "the", "distance", "from", "a", "pixel", "to", "the", "top", ",", "right", ",", "bottom", "and", "left", "boundary", "of", "its", "corresponding", "rectangle", ",", "respectively", ".", "The", "union", "area", "is", "given", "by", "Therefore", ",", "both", "the", "intersection", "/", "union", "area", "can", "be", "computed", "easily", ".", "Next", ",", "the", "loss", "of", "rotation", "angle", "is", "computed", "as", "where", "is", "the", "prediction", "to", "the", "rotation", "angle", "and", "represents", "the", "ground", "truth", ".", "Finally", ",", "the", "overall", "geometry", "loss", "is", "the", "weighted", "sum", "of", "AABB", "loss", "and", "angle", "loss", ",", "given", "by", "Where", "is", "set", "to", "in", "our", "experiments", ".", "Note", "that", "we", "compute", "regardless", "of", "rotation", "angle", ".", "This", "can", "be", "seen", "as", "an", "approximation", "of", "quadrangle", "IoU", "when", "the", "angle", "is", "perfectly", "predicted", ".", "Although", "it", "is", "not", "the", "case", "during", "training", ",", "it", "could", "still", "impose", "the", "correct", "gradient", "for", "the", "network", "to", "learn", "to", "predict", ".", "paragraph", ":", "QUAD", "We", "extend", "the", "smoothed", "-", "L1", "loss", "proposed", "in", "by", "adding", "an", "extra", "normalization", "term", "designed", "for", "word", "quadrangles", ",", "which", "is", "typically", "longer", "in", "one", "direction", ".", "Let", "all", "coordinate", "values", "of", "be", "an", "ordered", "set", "then", "the", "loss", "can", "be", "written", "as", "where", "the", "normalization", "term", "is", "the", "shorted", "edge", "length", "of", "the", "quadrangle", ",", "given", "by", "and", "is", "the", "set", "of", "all", "equivalent", "quadrangles", "of", "with", "different", "vertices", "ordering", ".", "This", "ordering", "permutation", "is", "required", "since", "the", "annotations", "of", "quadrangles", "in", "the", "public", "training", "datasets", "are", "inconsistent", ".", "subsection", ":", "Training", "The", "network", "is", "trained", "end", "-", "to", "-", "end", "using", "ADAM", "optimizer", ".", "To", "speed", "up", "learning", ",", "we", "uniformly", "sample", "512x512", "crops", "from", "images", "to", "form", "a", "minibatch", "of", "size", "24", ".", "Learning", "rate", "of", "ADAM", "starts", "from", "1e", "-", "3", ",", "decays", "to", "one", "-", "tenth", "every", "27300", "minibatches", ",", "and", "stops", "at", "1e", "-", "5", ".", "The", "network", "is", "trained", "until", "performance", "stops", "improving", ".", "subsection", ":", "Locality", "-", "Aware", "NMS", "To", "form", "the", "final", "results", ",", "the", "geometries", "survived", "after", "thresholding", "should", "be", "merged", "by", "NMS", ".", "A", "naïve", "NMS", "algorithm", "runs", "in", ",", "where", "is", "the", "number", "of", "candidate", "geometries", ",", "which", "is", "unacceptable", "as", "we", "are", "facing", "tens", "of", "thousands", "of", "geometries", "from", "dense", "predictions", ".", "Under", "the", "assumption", "that", "the", "geometries", "from", "nearby", "pixels", "tend", "to", "be", "highly", "correlated", ",", "we", "proposed", "to", "merge", "the", "geometries", "row", "by", "row", ",", "and", "while", "merging", "geometries", "in", "the", "same", "row", ",", "we", "will", "iteratively", "merge", "the", "geometry", "currently", "encountered", "with", "the", "last", "merged", "one", ".", "This", "improved", "technique", "runs", "in", "in", "best", "scenarios", ".", "Even", "though", "its", "worst", "case", "is", "the", "same", "as", "the", "naïve", "one", ",", "as", "long", "as", "the", "locality", "assumption", "holds", ",", "the", "algorithm", "runs", "sufficiently", "fast", "in", "practice", ".", "The", "procedure", "is", "summarized", "in", "Algorithm", "[", "reference", "]", "It", "is", "worth", "mentioning", "that", ",", "in", ",", "the", "coordinates", "of", "merged", "quadrangle", "are", "weight", "-", "averaged", "by", "the", "scores", "of", "two", "given", "quadrangles", ".", "To", "be", "specific", ",", "if", ",", "then", "and", ",", "where", "is", "one", "of", "the", "coordinates", "of", "subscripted", "by", ",", "and", "is", "the", "score", "of", "geometry", ".", "In", "fact", ",", "there", "is", "a", "subtle", "difference", "that", "we", "are", "”", "averaging", "”", "rather", "than", "”", "selecting", "”", "geometries", ",", "as", "in", "a", "standard", "NMS", "procedure", "will", "do", ",", "acting", "as", "a", "voting", "mechanism", ",", "which", "in", "turn", "introduces", "a", "stabilization", "effect", "when", "feeding", "videos", ".", "Nonetheless", ",", "we", "still", "adopt", "the", "word", "”", "NMS", "”", "for", "functional", "description", ".", "[", "t", "]", "Locality", "-", "Aware", "NMS", "[", "1", "]", "NMSLocality", ",", "in", "row", "first", "order", "section", ":", "Experiments", "To", "compare", "the", "proposed", "algorithm", "with", "existing", "methods", ",", "we", "conducted", "qualitative", "and", "quantitative", "experiments", "on", "three", "public", "benchmarks", ":", "ICDAR2015", ",", "COCO", "-", "Text", "and", "MSRA", "-", "TD500", ".", "subsection", ":", "Benchmark", "Datasets", "ICDAR", "2015", "is", "used", "in", "Challenge", "4", "of", "ICDAR", "2015", "Robust", "Reading", "Competition", ".", "It", "includes", "a", "total", "of", "1500", "pictures", ",", "1000", "of", "which", "are", "used", "for", "training", "and", "the", "remaining", "are", "for", "testing", ".", "The", "text", "regions", "are", "annotated", "by", "4", "vertices", "of", "the", "quadrangle", ",", "corresponding", "to", "the", "QUAD", "geometry", "in", "this", "paper", ".", "We", "also", "generate", "RBOX", "output", "by", "fitting", "a", "rotated", "rectangle", "which", "has", "the", "minimum", "area", ".", "These", "images", "are", "taken", "by", "Google", "Glass", "in", "an", "incidental", "way", ".", "Therefore", "text", "in", "the", "scene", "can", "be", "in", "arbitrary", "orientations", ",", "or", "suffer", "from", "motion", "blur", "and", "low", "resolution", ".", "We", "also", "used", "the", "229", "training", "images", "from", "ICDAR", "2013", ".", "COCO", "-", "Text", "is", "the", "largest", "text", "detection", "dataset", "to", "date", ".", "It", "reuses", "the", "images", "from", "MS", "-", "COCO", "dataset", ".", "A", "total", "of", "63", ",", "686", "images", "are", "annotated", ",", "in", "which", "43", ",", "686", "are", "chosen", "to", "be", "the", "training", "set", "and", "the", "rest", "20", ",", "000", "for", "testing", ".", "Word", "regions", "are", "annotated", "in", "the", "form", "of", "axis", "-", "aligned", "bounding", "box", "(", "AABB", ")", ",", "which", "is", "a", "special", "case", "of", "RBOX", ".", "For", "this", "dataset", ",", "we", "set", "angle", "to", "zero", ".", "We", "use", "the", "same", "data", "processing", "and", "test", "method", "as", "in", "ICDAR", "2015", ".", "MSRA", "-", "TD500", "is", "a", "dataset", "comprises", "of", "300", "training", "images", "and", "200", "test", "images", ".", "Text", "regions", "are", "of", "arbitrary", "orientations", "and", "annotated", "at", "sentence", "level", ".", "Different", "from", "the", "other", "datasets", ",", "it", "contains", "text", "in", "both", "English", "and", "Chinese", ".", "The", "text", "regions", "are", "annotated", "in", "RBOX", "format", ".", "Since", "the", "number", "of", "training", "images", "is", "too", "few", "to", "learn", "a", "deep", "model", ",", "we", "also", "harness", "400", "images", "from", "HUST", "-", "TR400", "dataset", "as", "training", "data", ".", "subsection", ":", "Base", "Networks", "As", "except", "for", "COCO", "-", "Text", ",", "all", "text", "detection", "datasets", "are", "relatively", "small", "compared", "to", "the", "datasets", "for", "general", "object", "detection", ",", "therefore", "if", "a", "single", "network", "is", "adopted", "for", "all", "the", "benchmarks", ",", "it", "may", "suffer", "from", "either", "over", "-", "fitting", "or", "under", "-", "fitting", ".", "We", "experimented", "with", "three", "different", "base", "networks", ",", "with", "different", "output", "geometries", ",", "on", "all", "the", "datasets", "to", "evaluate", "the", "proposed", "framework", ".", "These", "networks", "are", "summarized", "in", "Tab", ".", "[", "reference", "]", ".", "VGG16", "is", "widely", "used", "as", "base", "network", "in", "many", "tasks", "to", "support", "subsequent", "task", "-", "specific", "fine", "-", "tuning", ",", "including", "text", "detection", ".", "There", "are", "two", "drawbacks", "of", "this", "network", ":", "(", "1", ")", ".", "The", "receptive", "field", "for", "this", "network", "is", "small", ".", "Each", "pixel", "in", "output", "of", "conv5_3", "only", "has", "a", "receptive", "field", "of", "196", ".", "(", "2", ")", ".", "It", "is", "a", "rather", "large", "network", ".", "PVANET", "is", "a", "light", "weight", "network", "introduced", "in", ",", "aiming", "as", "a", "substitution", "of", "the", "feature", "extractor", "in", "Faster", "-", "RCNN", "framework", ".", "Since", "it", "is", "too", "small", "for", "GPU", "to", "fully", "utilizes", "computation", "parallelism", ",", "we", "also", "adopt", "PVANET2x", "that", "doubles", "the", "channels", "of", "the", "original", "PVANET", ",", "exploiting", "more", "computation", "parallelism", "while", "running", "slightly", "slower", "than", "PVANET", ".", "This", "is", "detailed", "in", "Sec", ".", "[", "reference", "]", ".", "The", "receptive", "field", "of", "the", "output", "of", "the", "last", "convolution", "layer", "is", "809", ",", "which", "is", "much", "larger", "than", "VGG16", ".", "The", "models", "are", "pre", "-", "trained", "on", "the", "ImageNet", "dataset", ".", "subsection", ":", "Qualitative", "Results", "(", "a", ")", "(", "b", ")", "(", "c", ")", "(", "a", ")", "(", "c", ")", "(", "b", ")", "(", "d", ")", "Fig", ".", "[", "reference", "]", "depicts", "several", "detection", "examples", "by", "the", "proposed", "algorithm", ".", "It", "is", "able", "to", "handle", "various", "challenging", "scenarios", ",", "such", "as", "non", "-", "uniform", "illumination", ",", "low", "resolution", ",", "varying", "orientation", "and", "perspective", "distortion", ".", "Moreover", ",", "due", "to", "the", "voting", "mechanism", "in", "the", "NMS", "procedure", ",", "the", "proposed", "method", "shows", "a", "high", "level", "of", "stability", "on", "videos", "with", "various", "forms", "of", "text", "instances", ".", "The", "intermediate", "results", "of", "the", "proposed", "method", "are", "illustrated", "in", "Fig", ".", "[", "reference", "]", ".", "As", "can", "be", "seen", ",", "the", "trained", "model", "produces", "highly", "accurate", "geometry", "maps", "and", "score", "map", ",", "in", "which", "detections", "of", "text", "instances", "in", "varying", "orientations", "are", "easily", "formed", ".", "subsection", ":", "Quantitative", "Results", "As", "shown", "in", "Tab", ".", "[", "reference", "]", "and", "Tab", ".", "[", "reference", "]", ",", "our", "approach", "outperforms", "previous", "state", "-", "of", "-", "the", "-", "art", "methods", "by", "a", "large", "margin", "on", "ICDAR", "2015", "and", "COCO", "-", "Text", ".", "In", "ICDAR", "2015", "Challenge", "4", ",", "when", "images", "are", "fed", "at", "their", "original", "scale", ",", "the", "proposed", "method", "achieves", "an", "F", "-", "score", "of", "0.7820", ".", "When", "tested", "at", "multiple", "scales", "using", "the", "same", "network", ",", "our", "method", "reaches", "0.8072", "in", "F", "-", "score", ",", "which", "is", "nearly", "0.16", "higher", "than", "the", "best", "method", "in", "terms", "of", "absolute", "value", "(", "0.8072", "vs.", "0.6477", ")", ".", "Comparing", "the", "results", "using", "VGG16", "network", ",", "the", "proposed", "method", "also", "outperforms", "best", "previous", "work", "by", "0.0924", "when", "using", "QUAD", "output", ",", "0.116", "when", "using", "RBOX", "output", ".", "Meanwhile", "these", "networks", "are", "quite", "efficient", ",", "as", "will", "be", "shown", "in", "Sec", ".", "[", "reference", "]", ".", "In", "COCO", "-", "Text", ",", "all", "of", "the", "three", "settings", "of", "the", "proposed", "algorithm", "result", "in", "higher", "accuracy", "than", "previous", "top", "performer", ".", "Specifically", ",", "the", "improvement", "over", "in", "F", "-", "score", "is", "0.0614", "while", "that", "in", "recall", "is", "0.053", ",", "which", "confirm", "the", "advantage", "of", "the", "proposed", "algorithm", ",", "considering", "that", "COCO", "-", "Text", "is", "the", "largest", "and", "most", "challenging", "benchmark", "to", "date", ".", "Note", "that", "we", "also", "included", "the", "results", "from", "as", "reference", ",", "but", "these", "results", "are", "actually", "not", "valid", "baselines", ",", "since", "the", "methods", "(", "A", ",", "B", "and", "C", ")", "are", "used", "in", "data", "annotation", ".", "The", "improvements", "of", "the", "proposed", "algorithm", "over", "previous", "methods", "prove", "that", "a", "simple", "text", "detection", "pipeline", ",", "which", "directly", "targets", "the", "final", "goal", "and", "eliminating", "redundant", "processes", ",", "can", "beat", "elaborated", "pipelines", ",", "even", "those", "integrated", "with", "large", "neural", "network", "models", ".", "As", "shown", "in", "Tab", ".", "[", "reference", "]", ",", "on", "MSRA", "-", "TD500", "all", "of", "the", "three", "settings", "of", "our", "method", "achieve", "excellent", "results", ".", "The", "F", "-", "score", "of", "the", "best", "performer", "(", "Ours", "+", "PVANET2x", ")", "is", "slightly", "higher", "than", "that", "of", ".", "Compared", "with", "the", "method", "of", "Zhang", ",", "the", "previous", "published", "state", "-", "of", "-", "the", "-", "art", "system", ",", "the", "best", "performer", "(", "Ours", "+", "PVANET2x", ")", "obtains", "an", "improvement", "of", "0.0208", "in", "F", "-", "score", "and", "0.0428", "in", "precision", ".", "Note", "that", "on", "MSRA", "-", "TD500", "our", "algorithm", "equipped", "with", "VGG16", "performs", "much", "poorer", "than", "that", "with", "PVANET", "and", "PVANET2x", "(", "0.7023", "vs.", "0.7445", "and", "0.7608", ")", ",", "the", "main", "reason", "is", "that", "the", "effective", "receptive", "field", "of", "VGG16", "is", "smaller", "than", "that", "of", "PVANET", "and", "PVANET2x", ",", "while", "the", "evaluation", "protocol", "of", "MSRA", "-", "TD500", "requires", "text", "detection", "algorithms", "output", "line", "level", "instead", "of", "word", "level", "predictions", ".", "In", "addition", ",", "we", "also", "evaluated", "Ours", "+", "PVANET2x", "on", "the", "ICDAR", "2013", "benchmark", ".", "It", "achieves", "0.8267", ",", "0.9264", "and", "0.8737", "in", "recall", ",", "precision", "and", "F", "-", "score", ",", "which", "are", "comparable", "with", "the", "previous", "state", "-", "of", "-", "the", "-", "art", "method", ",", "which", "obtains", "0.8298", ",", "0.9298", "and", "0.8769", "in", "recall", ",", "precision", "and", "F", "-", "score", ",", "respectively", ".", "subsection", ":", "Speed", "Comparison", "The", "overall", "speed", "comparison", "is", "demonstrated", "in", "Tab", ".", "[", "reference", "]", ".", "The", "numbers", "we", "reported", "are", "averages", "from", "running", "through", "500", "test", "images", "from", "the", "ICDAR", "2015", "dataset", "at", "their", "original", "resolution", "(", "1280x720", ")", "using", "our", "best", "performing", "networks", ".", "These", "experiments", "were", "conducted", "on", "a", "server", "using", "a", "single", "NVIDIA", "Titan", "X", "graphic", "card", "with", "Maxwell", "architecture", "and", "an", "Intel", "E5", "-", "2670", "v3", "@", "2.30GHz", "CPU", ".", "For", "the", "proposed", "method", ",", "the", "post", "-", "processing", "includes", "thresholding", "and", "NMS", ",", "while", "others", "should", "refer", "to", "their", "original", "paper", ".", "While", "the", "proposed", "method", "significantly", "outperforms", "state", "-", "of", "-", "the", "-", "art", "methods", ",", "the", "computation", "cost", "is", "kept", "very", "low", ",", "attributing", "to", "the", "simple", "and", "efficient", "pipeline", ".", "As", "can", "be", "observed", "from", "Tab", ".", "[", "reference", "]", ",", "the", "fastest", "setting", "of", "our", "method", "runs", "at", "a", "speed", "of", "16.8", "FPS", ",", "while", "slowest", "setting", "runs", "at", "6.52", "FPS", ".", "Even", "the", "best", "performing", "model", "Ours", "+", "PVANET2x", "runs", "at", "a", "speed", "of", "13.2", "FPS", ".", "This", "confirm", "that", "our", "method", "is", "among", "the", "most", "efficient", "text", "detectors", "that", "achieve", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "benchmarks", ".", "subsection", ":", "Limitations", "The", "maximal", "size", "of", "text", "instances", "the", "detector", "can", "handle", "is", "proportional", "to", "the", "receptive", "field", "of", "the", "network", ".", "This", "limits", "the", "capability", "of", "the", "network", "to", "predict", "even", "longer", "text", "regions", "like", "text", "lines", "running", "across", "the", "images", ".", "Also", ",", "the", "algorithm", "might", "miss", "or", "give", "imprecise", "predictions", "for", "vertical", "text", "instances", "as", "they", "take", "only", "a", "small", "portion", "of", "text", "regions", "in", "the", "ICDAR", "2015", "training", "set", ".", "section", ":", "Conclusion", "and", "Future", "Work", "We", "have", "presented", "a", "scene", "text", "detector", "that", "directly", "produces", "word", "or", "line", "level", "predictions", "from", "full", "images", "with", "a", "single", "neural", "network", ".", "By", "incorporating", "proper", "loss", "functions", ",", "the", "detector", "can", "predict", "either", "rotated", "rectangles", "or", "quadrangles", "for", "text", "regions", ",", "depending", "on", "specific", "applications", ".", "The", "experiments", "on", "standard", "benchmarks", "confirm", "that", "the", "proposed", "algorithm", "substantially", "outperforms", "previous", "methods", "in", "terms", "of", "both", "accuracy", "and", "efficiency", ".", "Possible", "directions", "for", "future", "research", "include", ":", "(", "1", ")", "adapting", "the", "geometry", "formulation", "to", "allow", "direct", "detection", "of", "curved", "text", ";", "(", "2", ")", "integrating", "the", "detector", "with", "a", "text", "recognizer", ";", "(", "3", ")", "extending", "the", "idea", "to", "general", "object", "detection", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["EAST"], "Method"], [5, 11, ["Efficient", "and", "Accurate", "Scene", "Text", "Detector"], "Method"], [14, 17, ["scene", "text", "detection"], "Task"], [42, 46, ["deep", "neural", "network", "models"], "Method"], [81, 86, ["text", "detection", "in", "natural", "scenes"], "Task"], [112, 114, ["candidate", "aggregation"], "Task"], [115, 117, ["word", "partitioning"], "Task"], [122, 124, ["neural", "network"], "Method"], [135, 137, ["loss", "functions"], "Method"], [138, 141, ["neural", "network", "architecture"], "Method"], [147, 149, ["ICDAR", "2015"], "Material"], [150, 153, ["COCO", "-", "Text"], "Material"], [154, 157, ["MSRA", "-", "TD500"], "Material"], [176, 177, ["accuracy"], "Metric"], [178, 179, ["efficiency"], "Metric"], [191, 194, ["F", "-", "score"], "Metric"], [202, 204, ["Megvii", "Technology"], "Method"], [228, 233, ["extracting", "and", "understanding", "textual", "information"], "Task"], [256, 259, ["ICDAR", "series", "contests"], "Task"], [268, 269, ["NIST"], "Metric"], [270, 272, ["Text", "detection"], "Task"], [290, 295, ["textual", "information", "extraction", "and", "understanding"], "Task"], [297, 300, ["text", "detection", "approaches"], "Method"], [315, 317, ["text", "detection"], "Task"], [344, 348, ["deep", "learning", "based", "methods"], "Method"], [396, 397, ["FPS"], "Metric"], [400, 403, ["F", "-", "score"], "Metric"], [603, 608, ["Ours", "+", "PVANet", "(", "0.757@16.8fps"], "Method"], [618, 621, ["deep", "neural", "network"], "Method"], [645, 646, ["accuracy"], "Metric"], [681, 688, ["fully", "convolutional", "network", "(", "FCN", ")", "model"], "Method"], [725, 727, ["Maximum", "Suppression"], "Method"], [768, 771, ["F", "-", "score"], "Metric"], [774, 776, ["ICDAR", "2015"], "Material"], [781, 784, ["multi", "-", "scale"], "Task"], [788, 791, ["MSRA", "-", "TD500"], "Material"], [794, 797, ["COCO", "-", "Text"], "Material"], [865, 868, ["Fully", "Convolutional", "Network"], "Method"], [870, 873, ["NMS", "merging", "stage"], "Method"], [875, 876, ["FCN"], "Method"], [934, 935, ["accuracy"], "Metric"], [936, 937, ["speed"], "Metric"], [942, 945, ["Scene", "text", "detection"], "Task"], [946, 947, ["recognition"], "Task"], [953, 955, ["computer", "vision"], "Task"], [1007, 1010, ["Stroke", "Width", "Transform"], "Method"], [1011, 1012, ["SWT"], "Method"], [1014, 1018, ["Maximally", "Stable", "Extremal", "Regions"], "Method"], [1019, 1020, ["MSER"], "Method"], [1028, 1030, ["edge", "detection"], "Method"], [1031, 1034, ["extremal", "region", "extraction"], "Method"], [1050, 1053, ["text", "region", "detection"], "Task"], [1054, 1055, ["FASText"], "Method"], [1057, 1061, ["fast", "text", "detection", "system"], "Method"], [1069, 1073, ["FAST", "key", "point", "detector"], "Method"], [1074, 1076, ["stroke", "extraction"], "Task"], [1087, 1090, ["deep", "neural", "networks"], "Method"], [1095, 1096, ["accuracy"], "Metric"], [1097, 1098, ["adaptability"], "Metric"], [1119, 1122, ["scene", "text", "detection"], "Task"], [1128, 1133, ["deep", "neural", "network", "based", "algorithms"], "Method"], [1144, 1145, ["MSER"], "Method"], [1149, 1152, ["deep", "convolutional", "network"], "Method"], [1155, 1156, ["classifier"], "Method"], [1178, 1179, ["heatmap"], "Method"], [1184, 1188, ["convolutional", "neural", "network", "model"], "Method"], [1195, 1196, ["CNN"], "Method"], [1198, 1199, ["ACF"], "Method"], [1208, 1209, ["regression"], "Method"], [1212, 1214, ["vertical", "anchors"], "Method"], [1217, 1222, ["CNN", "-", "RNN", "joint", "model"], "Method"], [1237, 1238, ["FCN"], "Method"], [1239, 1241, ["heatmap", "generation"], "Task"], [1244, 1246, ["component", "projection"], "Method"], [1247, 1249, ["orientation", "estimation"], "Task"], [1286, 1289, ["false", "positive", "removal"], "Task"], [1290, 1292, ["post", "filtering"], "Method"], [1293, 1295, ["candidate", "aggregation"], "Task"], [1296, 1298, ["line", "formation"], "Task"], [1299, 1301, ["word", "partition"], "Task"], [1323, 1325, ["processing", "time"], "Metric"], [1337, 1342, ["deep", "FCN", "-", "based", "pipeline"], "Method"], [1349, 1351, ["text", "detection"], "Task"], [1352, 1359, ["word", "or", "text", "-", "line", "level", "detection"], "Task"], [1388, 1390, ["end", "training"], "Task"], [1391, 1392, ["optimization"], "Task"], [1402, 1407, ["light", "-", "weighted", "neural", "network"], "Method"], [1420, 1421, ["speed"], "Metric"], [1434, 1437, ["neural", "network", "model"], "Method"], [1460, 1465, ["fully", "-", "convolutional", "neural", "network"], "Method"], [1467, 1469, ["text", "detection"], "Task"], [1471, 1478, ["dense", "per", "-", "pixel", "predictions", "of", "words"], "Task"], [1488, 1490, ["candidate", "proposal"], "Task"], [1491, 1494, ["text", "region", "formation"], "Task"], [1495, 1497, ["word", "partition"], "Task"], [1505, 1506, ["thresholding"], "Method"], [1507, 1508, ["NMS"], "Method"], [1518, 1519, ["EAST"], "Method"], [1533, 1535, ["detection", "pipeline"], "Method"], [1566, 1567, ["DenseBox"], "Method"], [1576, 1577, ["FCN"], "Method"], [1599, 1601, ["score", "map"], "Method"], [1667, 1669, ["loss", "functions"], "Method"], [1673, 1674, ["Thresholding"], "Method"], [1699, 1704, ["non", "-", "maximum", "-", "suppression"], "Task"], [1707, 1708, ["NMS"], "Method"], [1719, 1721, ["Network", "Design"], "Method"], [1730, 1732, ["neural", "networks"], "Method"], [1733, 1735, ["text", "detection"], "Task"], [1770, 1772, ["neural", "network"], "Method"], [1805, 1806, ["HyperNet"], "Method"], [1810, 1812, ["features", "maps"], "Task"], [1828, 1830, ["computation", "overhead"], "Metric"], [1881, 1883, ["computation", "cost"], "Metric"], [1909, 1912, ["feature", "extractor", "stem"], "Method"], [1918, 1920, ["output", "layer"], "Method"], [1926, 1928, ["convolutional", "network"], "Method"], [1932, 1934, ["ImageNet", "dataset"], "Material"], [1936, 1941, ["interleaving", "convolution", "and", "pooling", "layers"], "Method"], [1977, 1978, ["PVANet"], "Method"], [1992, 1994, ["VGG16", "model"], "Method"], [2011, 2015, ["feature", "-", "merging", "branch"], "Task"], [2061, 2063, ["unpooling", "layer"], "Method"], [2105, 2107, ["merging", "stage"], "Method"], [2140, 2141, ["convolution"], "Method"], [2168, 2170, ["computation", "overhead"], "Metric"], [2176, 2178, ["network", "computation"], "Method"], [2199, 2201, ["score", "map"], "Method"], [2203, 2208, ["multi", "-", "channel", "geometry", "map"], "Method"], [2229, 2230, ["RBOX"], "Method"], [2245, 2246, ["AABB"], "Method"], [2330, 2332, ["Label", "Generation"], "Task"], [2332, 2333, ["subsubsection"], "Method"], [2334, 2337, ["Score", "Map", "Generation"], "Method"], [2338, 2339, ["Quadrangle"], "Task"], [2364, 2366, ["score", "map"], "Metric"], [2487, 2490, ["Geometry", "Map", "Generation"], "Task"], [2506, 2507, ["RBOX"], "Method"], [2508, 2509, ["QUAD"], "Method"], [2511, 2513, ["generation", "process"], "Method"], [2514, 2515, ["RBOX"], "Method"], [2596, 2599, ["QUAD", "ground", "truth"], "Task"], [2629, 2631, ["Loss", "Functions"], "Method"], [2673, 2675, ["Score", "Map"], "Task"], [2684, 2686, ["detection", "pipelines"], "Method"], [2693, 2695, ["balanced", "sampling"], "Method"], [2696, 2699, ["hard", "negative", "mining"], "Method"], [2773, 2775, ["score", "map"], "Metric"], [2797, 2802, ["balanced", "cross", "-", "entropy", "loss"], "Method"], [2806, 2808, ["text", "detection"], "Task"], [2812, 2814, ["objective", "function"], "Metric"], [2815, 2818, ["score", "map", "prediction"], "Task"], [2831, 2832, ["Geometries"], "Task"], [2835, 2837, ["text", "detection"], "Task"], [2857, 2858, ["regression"], "Task"], [2876, 2879, ["text", "geometry", "prediction"], "Task"], [2901, 2903, ["IoU", "loss"], "Metric"], [2905, 2910, ["AABB", "part", "of", "RBOX", "regression"], "Method"], [2913, 2920, ["scale", "-", "normalized", "smoothed", "-", "L1", "loss"], "Metric"], [2921, 2923, ["QUAD", "regression"], "Method"], [2929, 2931, ["AABB", "part"], "Task"], [3049, 3052, ["overall", "geometry", "loss"], "Metric"], [3054, 3059, ["weighted", "sum", "of", "AABB", "loss"], "Metric"], [3060, 3062, ["angle", "loss"], "Metric"], [3090, 3092, ["quadrangle", "IoU"], "Task"], [3129, 3133, ["smoothed", "-", "L1", "loss"], "Method"], [3139, 3141, ["normalization", "term"], "Method"], [3230, 3232, ["ADAM", "optimizer"], "Method"], [3236, 3237, ["learning"], "Task"], [3253, 3255, ["Learning", "rate"], "Metric"], [3256, 3257, ["ADAM"], "Method"], [3290, 3294, ["Locality", "-", "Aware", "NMS"], "Method"], [3309, 3310, ["NMS"], "Method"], [3313, 3315, ["NMS", "algorithm"], "Method"], [3513, 3515, ["NMS", "procedure"], "Method"], [3521, 3523, ["voting", "mechanism"], "Method"], [3543, 3545, ["NMS", "”"], "Method"], [3546, 3548, ["functional", "description"], "Task"], [3552, 3556, ["Locality", "-", "Aware", "NMS"], "Method"], [3559, 3560, ["NMSLocality"], "Method"], [3588, 3589, ["ICDAR2015"], "Material"], [3590, 3593, ["COCO", "-", "Text"], "Material"], [3594, 3597, ["MSRA", "-", "TD500"], "Material"], [3602, 3604, ["ICDAR", "2015"], "Material"], [3715, 3717, ["ICDAR", "2013"], "Material"], [3718, 3721, ["COCO", "-", "Text"], "Material"], [3735, 3739, ["MS", "-", "COCO", "dataset"], "Material"], [3785, 3786, ["AABB"], "Method"], [3794, 3795, ["RBOX"], "Method"], [3817, 3819, ["ICDAR", "2015"], "Material"], [3820, 3823, ["MSRA", "-", "TD500"], "Material"], [3859, 3860, ["English"], "Material"], [3861, 3862, ["Chinese"], "Material"], [3869, 3871, ["RBOX", "format"], "Material"], [3884, 3886, ["deep", "model"], "Method"], [3893, 3897, ["HUST", "-", "TR400", "dataset"], "Material"], [3903, 3905, ["Base", "Networks"], "Method"], [3908, 3911, ["COCO", "-", "Text"], "Material"], [3924, 3927, ["general", "object", "detection"], "Task"], [3958, 3960, ["base", "networks"], "Method"], [3987, 3988, ["VGG16"], "Method"], [3992, 3994, ["base", "network"], "Method"], [4000, 4006, ["task", "-", "specific", "fine", "-", "tuning"], "Task"], [4008, 4010, ["text", "detection"], "Task"], [4037, 4038, ["conv5_3"], "Method"], [4057, 4058, ["PVANET"], "Method"], [4060, 4063, ["light", "weight", "network"], "Method"], [4072, 4074, ["feature", "extractor"], "Method"], [4075, 4079, ["Faster", "-", "RCNN", "framework"], "Method"], [4096, 4097, ["PVANET2x"], "Method"], [4104, 4105, ["PVANET"], "Method"], [4115, 4116, ["PVANET"], "Method"], [4136, 4138, ["convolution", "layer"], "Method"], [4146, 4147, ["VGG16"], "Method"], [4156, 4158, ["ImageNet", "dataset"], "Material"], [4228, 4230, ["voting", "mechanism"], "Method"], [4232, 4234, ["NMS", "procedure"], "Method"], [4243, 4244, ["stability"], "Metric"], [4280, 4282, ["geometry", "maps"], "Metric"], [4283, 4285, ["score", "map"], "Metric"], [4335, 4337, ["ICDAR", "2015"], "Material"], [4338, 4341, ["COCO", "-", "Text"], "Material"], [4362, 4365, ["F", "-", "score"], "Metric"], [4383, 4386, ["F", "-", "score"], "Metric"], [4399, 4401, ["absolute", "value"], "Metric"], [4411, 4413, ["VGG16", "network"], "Method"], [4454, 4457, ["COCO", "-", "Text"], "Material"], [4470, 4471, ["accuracy"], "Metric"], [4482, 4485, ["F", "-", "score"], "Metric"], [4490, 4491, ["recall"], "Metric"], [4505, 4508, ["COCO", "-", "Text"], "Material"], [4551, 4553, ["data", "annotation"], "Task"], [4567, 4570, ["text", "detection", "pipeline"], "Method"], [4591, 4595, ["large", "neural", "network", "models"], "Method"], [4622, 4625, ["F", "-", "score"], "Metric"], [4630, 4633, ["Ours", "+", "PVANET2x"], "Method"], [4646, 4647, ["Zhang"], "Method"], [4664, 4667, ["Ours", "+", "PVANET2x"], "Method"], [4674, 4677, ["F", "-", "score"], "Metric"], [4680, 4681, ["precision"], "Metric"], [4685, 4688, ["MSRA", "-", "TD500"], "Material"], [4692, 4693, ["VGG16"], "Method"], [4699, 4700, ["PVANET"], "Method"], [4701, 4702, ["PVANET2x"], "Method"], [4720, 4721, ["VGG16"], "Method"], [4726, 4727, ["PVANET"], "Method"], [4728, 4729, ["PVANET2x"], "Method"], [4732, 4734, ["evaluation", "protocol"], "Metric"], [4735, 4738, ["MSRA", "-", "TD500"], "Material"], [4739, 4742, ["text", "detection", "algorithms"], "Method"], [4762, 4765, ["ICDAR", "2013", "benchmark"], "Material"], [4774, 4775, ["recall"], "Metric"], [4776, 4777, ["precision"], "Metric"], [4778, 4781, ["F", "-", "score"], "Metric"], [4805, 4806, ["recall"], "Metric"], [4807, 4808, ["precision"], "Metric"], [4809, 4812, ["F", "-", "score"], "Metric"], [4817, 4819, ["Speed", "Comparison"], "Metric"], [4821, 4823, ["speed", "comparison"], "Metric"], [4872, 4877, ["NVIDIA", "Titan", "X", "graphic", "card"], "Method"], [4878, 4880, ["Maxwell", "architecture"], "Method"], [4897, 4900, ["post", "-", "processing"], "Task"], [4901, 4902, ["thresholding"], "Method"], [4903, 4904, ["NMS"], "Method"], [4930, 4932, ["computation", "cost"], "Metric"], [4943, 4944, ["pipeline"], "Method"], [4983, 4986, ["Ours", "+", "PVANET2x"], "Method"], [5004, 5006, ["text", "detectors"], "Method"], [5029, 5030, ["detector"], "Method"], [5104, 5107, ["scene", "text", "detector"], "Method"], [5110, 5115, ["word", "or", "line", "level", "predictions"], "Task"], [5121, 5123, ["neural", "network"], "Method"], [5166, 5167, ["accuracy"], "Metric"], [5168, 5169, ["efficiency"], "Metric"], [5182, 5184, ["geometry", "formulation"], "Method"], [5186, 5191, ["direct", "detection", "of", "curved", "text"], "Task"], [5200, 5202, ["text", "recognizer"], "Method"], [5211, 5213, ["object", "detection"], "Task"], [182, 184, ["ICDAR", "2015"], "Material"], [667, 670, ["scene", "text", "detection"], "Task"], [854, 857, ["scene", "text", "detection"], "Task"], [2542, 2544, ["ICDAR", "2015"], "Material"], [2934, 2936, ["IoU", "loss"], "Metric"], [3610, 3612, ["ICDAR", "2015"], "Material"], [4343, 4345, ["ICDAR", "2015"], "Material"], [4606, 4609, ["MSRA", "-", "TD500"], "Material"], [4846, 4848, ["ICDAR", "2015"], "Material"], [4968, 4969, ["FPS"], "Metric"], [4976, 4977, ["FPS"], "Metric"], [4992, 4993, ["FPS"], "Metric"], [5089, 5091, ["ICDAR", "2015"], "Material"]]}