diff --git "a/parse/train/HJr4QJ26W/HJr4QJ26W_middle.json" "b/parse/train/HJr4QJ26W/HJr4QJ26W_middle.json" new file mode 100644--- /dev/null +++ "b/parse/train/HJr4QJ26W/HJr4QJ26W_middle.json" @@ -0,0 +1,35862 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 80, + 506, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 508, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 508, + 98 + ], + "score": 1.0, + "content": "IMPROVING IMAGE GENERATIVE MODELS WITH HU-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 102, + 258, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 258, + 118 + ], + "score": 1.0, + "content": "MAN INTERACTIONS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 185, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 185, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 209, + 469, + 341 + ], + "lines": [ + { + "bbox": [ + 142, + 209, + 470, + 222 + ], + "spans": [ + { + "bbox": [ + 142, + 209, + 470, + 222 + ], + "score": 1.0, + "content": "GANs provide a framework for training generative models which mimic a data", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 220, + 469, + 233 + ], + "spans": [ + { + "bbox": [ + 141, + 220, + 469, + 233 + ], + "score": 1.0, + "content": "distribution. However, in many cases we wish to train a generative model to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 232, + 469, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 232, + 469, + 244 + ], + "score": 1.0, + "content": "optimize some auxiliary objective function within the data it generates, such", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 243, + 469, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 243, + 469, + 255 + ], + "score": 1.0, + "content": "as making more aesthetically pleasing images. In some cases, these objective", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 253, + 470, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 470, + 267 + ], + "score": 1.0, + "content": "functions are difficult to evaluate, e.g. they may require human interaction. Here,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 265, + 470, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 265, + 470, + 276 + ], + "score": 1.0, + "content": "we develop a system for efficiently training a GAN to increase a generic rate of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 276, + 470, + 287 + ], + "spans": [ + { + "bbox": [ + 141, + 276, + 470, + 287 + ], + "score": 1.0, + "content": "positive user interactions, for example aesthetic ratings. To do this, we build a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 287, + 470, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 287, + 470, + 298 + ], + "score": 1.0, + "content": "model of human behavior in the targeted domain from a relatively small set of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 297, + 469, + 309 + ], + "spans": [ + { + "bbox": [ + 142, + 297, + 469, + 309 + ], + "score": 1.0, + "content": "interactions, and then use this behavioral model as an auxiliary loss function to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "improve the generative model. As a proof of concept, we demonstrate that this", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 320, + 469, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 320, + 469, + 333 + ], + "score": 1.0, + "content": "system is successful at improving positive interaction rates simulated from a variety", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 330, + 428, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 428, + 343 + ], + "score": 1.0, + "content": "of objectives, and characterize some factors that affect its performance.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 108, + 360, + 206, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 359, + 208, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 208, + 376 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 504, + 451 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "Generative image models have improved rapidly in the past few years, in part because of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "success of Generative Adversarial Networks, or GANs (Goodfellow et al., 2014). GANs attempt to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "train a “generator” to create images which mimic real images, by training it to fool an adversarial", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "“discriminator,” which attempts to discern whether images are real or fake. This is one solution to the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "difficult problem of learning when we don’t know how to write down an objective function for image", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 440, + 412, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 412, + 452 + ], + "score": 1.0, + "content": "quality: take an empirical distribution of “good” images, and try to match it.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 470 + ], + "score": 1.0, + "content": "Often, we want to impose additional constraints on our goal distribution besides simply matching", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "empirical data. If we can write down an objective which reflects our goals (even approximately), we", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 477, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 493 + ], + "score": 1.0, + "content": "can often simply incorporate this into the loss function to achieve our goals. For example, when trying", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 488, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 503 + ], + "score": 1.0, + "content": "to generate art, we would like our network to be creative and innovative rather than just imitating", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "previous styles, and including a penalty in the loss for producing recognized styles appears to make", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "GANs more creative (Elgammal et al., 2017). Conditioning on image content class, training the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 521, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 536 + ], + "score": 1.0, + "content": "discriminator to classify image content as well as making real/fake judgements, and including a loss", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 532, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 506, + 547 + ], + "score": 1.0, + "content": "term for fooling the discriminator on class both allows for targeted image generation and improves", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 544, + 273, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 273, + 557 + ], + "score": 1.0, + "content": "overall performance (Odena et al., 2016).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 504, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "However, sometimes it is not easy to write an explicit objective that reflects our goals. Often the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "only effective way to evaluate machine learning systems on complex tasks is by asking humans to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "determine the quality of their results (Christiano et al., 2017, e.g.) or by actually trying them out", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "in the real world. Can we incorporate this kind of feedback to efficiently guide a generative model", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "toward producing better results? Can we do so without a prohibitively expensive and slow amount", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "of data collection? In this paper, we tackle a specific problem of this kind: generating images that", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "cause more positive user interactions. We imagine interactions are measured by a generic Positive", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 637, + 402, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 402, + 651 + ], + "score": 1.0, + "content": "Interaction Rate (PIR), which could come from a wide variety of sources.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "For example, users might be asked to rate how aesthetically pleasing an image is from 1 to 5 stars.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "The PIR could be computed as a weighted sum of how frequently different ratings were chosen.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "Alternatively, these images could be used in the background of web pages. We can assess user", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "interactions with a webpage in a variety of ways (time on page, clicks, shares, etc.), and summarize", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "these interactions as the PIR. In both of these tasks, we don’t know exactly what features will affect", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "the PIR, and we certainly don’t know how to explicitly compute the PIR for an image. However,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "we can empirically determine the quality of an image by actually showing it to users, and in this", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 80, + 506, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 508, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 508, + 98 + ], + "score": 1.0, + "content": "IMPROVING IMAGE GENERATIVE MODELS WITH HU-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 102, + 258, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 258, + 118 + ], + "score": 1.0, + "content": "MAN INTERACTIONS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 112, + 136, + 245, + 158 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 185, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 185, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 209, + 469, + 341 + ], + "lines": [ + { + "bbox": [ + 142, + 209, + 470, + 222 + ], + "spans": [ + { + "bbox": [ + 142, + 209, + 470, + 222 + ], + "score": 1.0, + "content": "GANs provide a framework for training generative models which mimic a data", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 220, + 469, + 233 + ], + "spans": [ + { + "bbox": [ + 141, + 220, + 469, + 233 + ], + "score": 1.0, + "content": "distribution. However, in many cases we wish to train a generative model to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 232, + 469, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 232, + 469, + 244 + ], + "score": 1.0, + "content": "optimize some auxiliary objective function within the data it generates, such", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 243, + 469, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 243, + 469, + 255 + ], + "score": 1.0, + "content": "as making more aesthetically pleasing images. In some cases, these objective", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 253, + 470, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 470, + 267 + ], + "score": 1.0, + "content": "functions are difficult to evaluate, e.g. they may require human interaction. Here,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 265, + 470, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 265, + 470, + 276 + ], + "score": 1.0, + "content": "we develop a system for efficiently training a GAN to increase a generic rate of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 276, + 470, + 287 + ], + "spans": [ + { + "bbox": [ + 141, + 276, + 470, + 287 + ], + "score": 1.0, + "content": "positive user interactions, for example aesthetic ratings. To do this, we build a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 287, + 470, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 287, + 470, + 298 + ], + "score": 1.0, + "content": "model of human behavior in the targeted domain from a relatively small set of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 297, + 469, + 309 + ], + "spans": [ + { + "bbox": [ + 142, + 297, + 469, + 309 + ], + "score": 1.0, + "content": "interactions, and then use this behavioral model as an auxiliary loss function to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "improve the generative model. As a proof of concept, we demonstrate that this", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 320, + 469, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 320, + 469, + 333 + ], + "score": 1.0, + "content": "system is successful at improving positive interaction rates simulated from a variety", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 330, + 428, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 428, + 343 + ], + "score": 1.0, + "content": "of objectives, and characterize some factors that affect its performance.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5, + "bbox_fs": [ + 141, + 209, + 470, + 343 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 360, + 206, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 359, + 208, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 208, + 376 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 504, + 451 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "Generative image models have improved rapidly in the past few years, in part because of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "success of Generative Adversarial Networks, or GANs (Goodfellow et al., 2014). GANs attempt to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "train a “generator” to create images which mimic real images, by training it to fool an adversarial", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "“discriminator,” which attempts to discern whether images are real or fake. This is one solution to the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "difficult problem of learning when we don’t know how to write down an objective function for image", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 440, + 412, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 412, + 452 + ], + "score": 1.0, + "content": "quality: take an empirical distribution of “good” images, and try to match it.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 385, + 506, + 452 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 470 + ], + "score": 1.0, + "content": "Often, we want to impose additional constraints on our goal distribution besides simply matching", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "empirical data. If we can write down an objective which reflects our goals (even approximately), we", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 477, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 493 + ], + "score": 1.0, + "content": "can often simply incorporate this into the loss function to achieve our goals. For example, when trying", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 488, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 503 + ], + "score": 1.0, + "content": "to generate art, we would like our network to be creative and innovative rather than just imitating", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "previous styles, and including a penalty in the loss for producing recognized styles appears to make", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "GANs more creative (Elgammal et al., 2017). Conditioning on image content class, training the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 521, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 536 + ], + "score": 1.0, + "content": "discriminator to classify image content as well as making real/fake judgements, and including a loss", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 532, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 506, + 547 + ], + "score": 1.0, + "content": "term for fooling the discriminator on class both allows for targeted image generation and improves", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 544, + 273, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 273, + 557 + ], + "score": 1.0, + "content": "overall performance (Odena et al., 2016).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 455, + 506, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 504, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "However, sometimes it is not easy to write an explicit objective that reflects our goals. Often the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "only effective way to evaluate machine learning systems on complex tasks is by asking humans to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "determine the quality of their results (Christiano et al., 2017, e.g.) or by actually trying them out", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "in the real world. Can we incorporate this kind of feedback to efficiently guide a generative model", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "toward producing better results? Can we do so without a prohibitively expensive and slow amount", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "of data collection? In this paper, we tackle a specific problem of this kind: generating images that", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "cause more positive user interactions. We imagine interactions are measured by a generic Positive", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 637, + 402, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 402, + 651 + ], + "score": 1.0, + "content": "Interaction Rate (PIR), which could come from a wide variety of sources.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5, + "bbox_fs": [ + 104, + 561, + 506, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "For example, users might be asked to rate how aesthetically pleasing an image is from 1 to 5 stars.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "The PIR could be computed as a weighted sum of how frequently different ratings were chosen.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "Alternatively, these images could be used in the background of web pages. We can assess user", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "interactions with a webpage in a variety of ways (time on page, clicks, shares, etc.), and summarize", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "these interactions as the PIR. In both of these tasks, we don’t know exactly what features will affect", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "the PIR, and we certainly don’t know how to explicitly compute the PIR for an image. However,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "we can empirically determine the quality of an image by actually showing it to users, and in this", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 327, + 507, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 507, + 342 + ], + "score": 1.0, + "content": "paper we show how to use a small amount of this data (results on 1000 images) to efficiently tune a", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "generative model to produce images which increase PIR. In this work we focus on simulated PIR", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 351, + 507, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 507, + 363 + ], + "score": 1.0, + "content": "values as a proof of concept, but in future work we will investigate PIR values from real interactions.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 655, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 79, + 477, + 271 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 133, + 79, + 477, + 271 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 133, + 79, + 477, + 271 + ], + "spans": [ + { + "bbox": [ + 133, + 79, + 477, + 271 + ], + "score": 0.973, + "type": "image", + "image_path": "ec679fbe76f983e591cddbdeceb8798f656c1b196c970471f04538046cd10173.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 133, + 79, + 477, + 143.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 133, + 143.0, + 477, + 207.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 133, + 207.0, + 477, + 271.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 240, + 278, + 372, + 290 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 239, + 277, + 372, + 291 + ], + "spans": [ + { + "bbox": [ + 239, + 277, + 372, + 291 + ], + "score": 1.0, + "content": "Figure 1: Diagram of our system", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 328, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 105, + 327, + 507, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 507, + 342 + ], + "score": 1.0, + "content": "paper we show how to use a small amount of this data (results on 1000 images) to efficiently tune a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "generative model to produce images which increase PIR. In this work we focus on simulated PIR", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 351, + 507, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 507, + 363 + ], + "score": 1.0, + "content": "values as a proof of concept, but in future work we will investigate PIR values from real interactions.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 107, + 398, + 182, + 410 + ], + "lines": [ + { + "bbox": [ + 104, + 396, + 185, + 413 + ], + "spans": [ + { + "bbox": [ + 104, + 396, + 185, + 413 + ], + "score": 1.0, + "content": "2 APPROACH", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "The most straight-forward way to improve an image GAN might be to evaluate the images the model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "produces with real users at each training step. However, this process is far too slow. Instead, we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "want to be able to collect a batch of PIR data on a batch of images, and then use this batch of data", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "score": 1.0, + "content": "to improve the generative model for many gradient steps; we want to do this despite the fact that", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "the images the generator is producing may evolve to be very different from the original images we", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "collected PIR data on. In order to do this, we use the batch of image and PIR data to train a “PIR", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "Estimator Model” which predicts PIRs on images. We then use these estimated PIRs at each step as a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 510, + 128, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 128, + 524 + ], + "score": 1.0, + "content": "loss.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "Our approach is inspired by the work of Christiano and colleagues (Christiano et al., 2017), who", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "score": 1.0, + "content": "integrated human preference ratings between action sequences into training of a reinforcement", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "learning model by using the preference data to estimate a reward function. However, our problem and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "approach differ in several key ways. First, we are optimizing a generative image model rather than a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "RL model. This is more difficult in some ways, since the output space is much higher-dimensional", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "than typical RL problems, which means that scalar feedback (like a PIR) may be harder for the system", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "to learn from. This difficulty is partially offset by the fact that we assume we get “reward” (PIR)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "score": 1.0, + "content": "information for an image when we evaluate, instead of just getting preferences which we have to map", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "to rewards. Perhaps most importantly, we use our PIR estimation model as a fully-differentiable loss", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "function for training, instead of just using its estimated rewards. This allows us to more effectively", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 638, + 383, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 383, + 651 + ], + "score": 1.0, + "content": "exploit its knowledge of the objective function (but risks overfitting).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "Our system consists of three components: A generative image model, users who interact with the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 666, + 507, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 507, + 679 + ], + "score": 1.0, + "content": "generated images in some way, and a PIR estimator that models user interactions given an image.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "See Fig. 1 for a diagram of the system’s general operation. The generative model produces images,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "which are served to users. Using interaction data from these users, we train the PIR estimator model,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "which predicts PIRs given a background image, and then incorporate this estimated PIR into the loss", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "of the generative model to tune it to produce higher quality images. Below, we discuss each of these", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 721, + 217, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 217, + 732 + ], + "score": 1.0, + "content": "components in more detail.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 79, + 477, + 271 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 133, + 79, + 477, + 271 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 133, + 79, + 477, + 271 + ], + "spans": [ + { + "bbox": [ + 133, + 79, + 477, + 271 + ], + "score": 0.973, + "type": "image", + "image_path": "ec679fbe76f983e591cddbdeceb8798f656c1b196c970471f04538046cd10173.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 133, + 79, + 477, + 143.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 133, + 143.0, + 477, + 207.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 133, + 207.0, + 477, + 271.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 240, + 278, + 372, + 290 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 239, + 277, + 372, + 291 + ], + "spans": [ + { + "bbox": [ + 239, + 277, + 372, + 291 + ], + "score": 1.0, + "content": "Figure 1: Diagram of our system", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 328, + 505, + 362 + ], + "lines": [], + "index": 5, + "bbox_fs": [ + 105, + 327, + 507, + 363 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 398, + 182, + 410 + ], + "lines": [ + { + "bbox": [ + 104, + 396, + 185, + 413 + ], + "spans": [ + { + "bbox": [ + 104, + 396, + 185, + 413 + ], + "score": 1.0, + "content": "2 APPROACH", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "The most straight-forward way to improve an image GAN might be to evaluate the images the model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "produces with real users at each training step. However, this process is far too slow. Instead, we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "want to be able to collect a batch of PIR data on a batch of images, and then use this batch of data", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "score": 1.0, + "content": "to improve the generative model for many gradient steps; we want to do this despite the fact that", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "the images the generator is producing may evolve to be very different from the original images we", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "collected PIR data on. In order to do this, we use the batch of image and PIR data to train a “PIR", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "Estimator Model” which predicts PIRs on images. We then use these estimated PIRs at each step as a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 510, + 128, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 128, + 524 + ], + "score": 1.0, + "content": "loss.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 434, + 506, + 524 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "Our approach is inspired by the work of Christiano and colleagues (Christiano et al., 2017), who", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "score": 1.0, + "content": "integrated human preference ratings between action sequences into training of a reinforcement", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "learning model by using the preference data to estimate a reward function. However, our problem and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "approach differ in several key ways. First, we are optimizing a generative image model rather than a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "RL model. This is more difficult in some ways, since the output space is much higher-dimensional", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "than typical RL problems, which means that scalar feedback (like a PIR) may be harder for the system", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "to learn from. This difficulty is partially offset by the fact that we assume we get “reward” (PIR)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "score": 1.0, + "content": "information for an image when we evaluate, instead of just getting preferences which we have to map", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "to rewards. Perhaps most importantly, we use our PIR estimation model as a fully-differentiable loss", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "function for training, instead of just using its estimated rewards. This allows us to more effectively", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 638, + 383, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 383, + 651 + ], + "score": 1.0, + "content": "exploit its knowledge of the objective function (but risks overfitting).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 528, + 506, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "Our system consists of three components: A generative image model, users who interact with the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 666, + 507, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 507, + 679 + ], + "score": 1.0, + "content": "generated images in some way, and a PIR estimator that models user interactions given an image.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "See Fig. 1 for a diagram of the system’s general operation. The generative model produces images,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "which are served to users. Using interaction data from these users, we train the PIR estimator model,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "which predicts PIRs given a background image, and then incorporate this estimated PIR into the loss", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "of the generative model to tune it to produce higher quality images. Below, we discuss each of these", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 721, + 217, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 217, + 732 + ], + "score": 1.0, + "content": "components in more detail.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 654, + 507, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 221, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 222, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 222, + 95 + ], + "score": 1.0, + "content": "2.1 GENERATIVE MODEL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 102, + 505, + 158 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 176, + 116 + ], + "score": 1.0, + "content": "We begin with a", + "type": "text" + }, + { + "bbox": [ + 177, + 102, + 204, + 114 + ], + "score": 0.76, + "content": "\\mathrm { G A N ^ { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 101, + 505, + 116 + ], + "score": 1.0, + "content": "(Goodfellow et al., 2014) which we pre-trained to produce images from", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 418, + 127 + ], + "score": 1.0, + "content": "a target distribution (specifically, landscapes of mountains and coasts). Let", + "type": "text" + }, + { + "bbox": [ + 418, + 114, + 446, + 126 + ], + "score": 0.9, + "content": "D _ { \\mathrm { s o u r c e } }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "be the source", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 234, + 138 + ], + "score": 1.0, + "content": "estimated by the discriminator,", + "type": "text" + }, + { + "bbox": [ + 234, + 126, + 243, + 135 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 124, + 318, + 138 + ], + "score": 1.0, + "content": "the generator, and", + "type": "text" + }, + { + "bbox": [ + 319, + 127, + 326, + 135 + ], + "score": 0.77, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 124, + 505, + 138 + ], + "score": 1.0, + "content": "a noise input to the generator sampled from", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 135, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 284, + 149 + ], + "score": 1.0, + "content": "a multivariate standard normal distribution", + "type": "text" + }, + { + "bbox": [ + 284, + 136, + 317, + 148 + ], + "score": 0.93, + "content": "\\mathcal { N } ( 0 , I )", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 135, + 338, + 149 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 338, + 136, + 346, + 146 + ], + "score": 0.8, + "content": "\\mathcal { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 135, + 505, + 149 + ], + "score": 1.0, + "content": "be the set of real images shown to the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 146, + 197, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 197, + 159 + ], + "score": 1.0, + "content": "discriminator. Define:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 161, + 160, + 450, + 207 + ], + "lines": [ + { + "bbox": [ + 161, + 160, + 450, + 207 + ], + "spans": [ + { + "bbox": [ + 161, + 160, + 450, + 207 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { L _ { \\mathrm { f a k e i m a g e } } = } & { } & { E _ { z \\sim \\mathcal { N } ( 0 , I ) } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( G ( z ) ) = \\mathrm { f a k e } ) \\right] } \\\\ { L _ { \\mathrm { f a k e i m a g e f o o l s } } = } & { } & { E _ { z \\sim \\mathcal { N } ( 0 , I ) } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( G ( z ) ) = \\mathrm { r e a l } ) \\right] } \\\\ { L _ { \\mathrm { r e a l i m a g e } } = } & { } & { E _ { i \\sim \\mathcal { Z } } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( i ) = \\mathrm { r e a l } ) \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "0088b43259ac008782f8d88cec6de49629fc3381a9b69629b272e6a48dd3be36.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 161, + 160, + 450, + 175.66666666666666 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 161, + 175.66666666666666, + 450, + 191.33333333333331 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 161, + 191.33333333333331, + 450, + 206.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 506, + 236 + ], + "score": 1.0, + "content": "Then the discriminator and generator are trained to maximize the following losses (respectively),", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 234, + 306, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 234, + 147, + 245 + ], + "score": 1.0, + "content": "where the", + "type": "text" + }, + { + "bbox": [ + 148, + 235, + 160, + 244 + ], + "score": 0.87, + "content": "w _ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 234, + 306, + 245 + ], + "score": 1.0, + "content": "are weights set as hyperparameters:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "interline_equation", + "bbox": [ + 190, + 247, + 421, + 276 + ], + "lines": [ + { + "bbox": [ + 190, + 247, + 421, + 276 + ], + "spans": [ + { + "bbox": [ + 190, + 247, + 421, + 276 + ], + "score": 0.84, + "content": "\\begin{array} { r l r } { L _ { \\mathrm { d i s c r i m i n a t o r } } = } & { { } } & { L _ { \\mathrm { f a k e i m a g e } } + L _ { \\mathrm { r e a l i m a g e } } } \\\\ { L _ { \\mathrm { G e n e r a t o r } } = } & { { } } & { L _ { \\mathrm { f a k e i m a g e f o o l s } } } \\end{array}", + "type": "interline_equation", + "image_path": "6cd424a9880748934e609233a2c271a2dfa0d7a3a7cf644a9761608799819961.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 190, + 247, + 421, + 261.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 190, + 261.5, + 421, + 276.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 505, + 323 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 505, + 291 + ], + "score": 1.0, + "content": "Note that there is a difference between these losses and the standard GAN formulation given", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 288, + 507, + 304 + ], + "spans": [ + { + "bbox": [ + 104, + 288, + 297, + 304 + ], + "score": 1.0, + "content": "in (Goodfellow et al., 2014) – we maximize", + "type": "text" + }, + { + "bbox": [ + 297, + 290, + 317, + 300 + ], + "score": 0.31, + "content": "L _ { \\mathrm { f a k e } }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 288, + 354, + 304 + ], + "score": 1.0, + "content": "image fools", + "type": "text" + }, + { + "bbox": [ + 354, + 290, + 393, + 301 + ], + "score": 0.6, + "content": "= \\ \\log \\ P", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 288, + 507, + 304 + ], + "score": 1.0, + "content": "(classified real) rather than", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 300, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 158, + 312 + ], + "score": 1.0, + "content": "minimizing", + "type": "text" + }, + { + "bbox": [ + 158, + 300, + 201, + 312 + ], + "score": 0.72, + "content": "\\log { ( 1 - P }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 301, + 506, + 312 + ], + "score": 1.0, + "content": "(classified real)). This seems to result in the generation of slightly better", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 312, + 185, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 185, + 324 + ], + "score": 1.0, + "content": "images in practice.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 328, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "This GAN was trained on a dataset consisting of landscape images of mountains and coastlines (see", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 505, + 352 + ], + "score": 1.0, + "content": "Appendix C.2 for details of the architecture and training). It is worth noting that this generative", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 351, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 505, + 362 + ], + "score": 1.0, + "content": "model is not photorealistic (see Fig. 3a for some samples). Its expressive capacity is limited, and it", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 360, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 374 + ], + "score": 1.0, + "content": "has clear output modes with limited intra-mode variability. However, for our purposes this may not", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "matter. Indeed, it is in some ways more interesting if we can tweak this model to optimize for many", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 383, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 505, + 395 + ], + "score": 1.0, + "content": "objective functions, since its limited expressive capacity will make it more difficult for us to estimate", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "and pursue the real objective – a limited set of images will effectively give us fewer points to estimate", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 405, + 504, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 504, + 417 + ], + "score": 1.0, + "content": "the PIR function from, and will reduce the space in which the model can easily produce images, thus", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "reducing the possibility of getting very optimal images from the model. For example, a model which", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 427, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 439 + ], + "score": 1.0, + "content": "produces images of birds may not produce data points which provide good estimates of a PIR based", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "on how much the image looks like a car, and even if it could, it may not be able to produce images", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "which are “more car-like.” If we are able to succeed in improving PIRs with this generative model, it", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 460, + 388, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 388, + 472 + ], + "score": 1.0, + "content": "is likely that a better generative model would yield even better results.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 108, + 484, + 221, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 484, + 222, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 222, + 496 + ], + "score": 1.0, + "content": "2.2 USER INTERACTIONS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "We will show these images to users in a variety of ways, depending on our target domain. For the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "score": 1.0, + "content": "purposes of this paper, however, we will use simulated interaction data (see Section 3 for details).", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 526, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 506, + 539 + ], + "score": 1.0, + "content": "Of course, since showing images to users is an expensive prospect, we wanted to limit the size of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "the datasets we used to train the model. Typical datasets used to train vision models are on the order", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "of millions of images, (e.g. ImageNet (Russakovsky et al., 2015)), but it is completely infeasible", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 559, + 504, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 504, + 571 + ], + "score": 1.0, + "content": "to collect user data on this number of images. We estimated that we could show 1000 images each", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "1000 times to generate our datasets. We used these dataset sizes and number of impressions for all", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 581, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 506, + 595 + ], + "score": 1.0, + "content": "experiments discussed here, and added noise to the PIRs that was binomially distributed according to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 592, + 473, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 473, + 605 + ], + "score": 1.0, + "content": "the number of times each image was shown and the “true” PIR simulate from the objective.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 108, + 617, + 234, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 235, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 235, + 630 + ], + "score": 1.0, + "content": "2.3 PIR ESTIMATOR MODEL", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "The final component of our system is the PIR estimator model, which learns to predict PIR from a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 293, + 661 + ], + "score": 1.0, + "content": "background image. We denote this model by", + "type": "text" + }, + { + "bbox": [ + 293, + 649, + 372, + 660 + ], + "score": 0.85, + "content": "R : { \\mathrm { i m a g e } } [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 649, + 505, + 661 + ], + "score": 1.0, + "content": ". We parameterize this model as", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "a deep neural network. Specifically, we take the Inception v2 architecture (Szegedy et al., 2016),", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "score": 1.0, + "content": "remove the output layer, and replace it with a fully-connected layer to PIR estimates. We initialize the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 681, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 506, + 694 + ], + "score": 1.0, + "content": "Inception v2 parameters from a version of the model trained on [dataset redacted for blind review].", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 693, + 249, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 249, + 704 + ], + "score": 1.0, + "content": "See Appendix C.3 for more details.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 711, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 119, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "1In fact, we used an ACGAN (Odena et al., 2016) because it generated better initial images, but we present", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 720, + 438, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 438, + 732 + ], + "score": 1.0, + "content": "the results here in terms of a GAN for clarity. See Appendix C.1 for details of the ACGAN.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 221, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 222, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 222, + 95 + ], + "score": 1.0, + "content": "2.1 GENERATIVE MODEL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 102, + 505, + 158 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 176, + 116 + ], + "score": 1.0, + "content": "We begin with a", + "type": "text" + }, + { + "bbox": [ + 177, + 102, + 204, + 114 + ], + "score": 0.76, + "content": "\\mathrm { G A N ^ { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 101, + 505, + 116 + ], + "score": 1.0, + "content": "(Goodfellow et al., 2014) which we pre-trained to produce images from", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 418, + 127 + ], + "score": 1.0, + "content": "a target distribution (specifically, landscapes of mountains and coasts). Let", + "type": "text" + }, + { + "bbox": [ + 418, + 114, + 446, + 126 + ], + "score": 0.9, + "content": "D _ { \\mathrm { s o u r c e } }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "be the source", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 234, + 138 + ], + "score": 1.0, + "content": "estimated by the discriminator,", + "type": "text" + }, + { + "bbox": [ + 234, + 126, + 243, + 135 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 124, + 318, + 138 + ], + "score": 1.0, + "content": "the generator, and", + "type": "text" + }, + { + "bbox": [ + 319, + 127, + 326, + 135 + ], + "score": 0.77, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 124, + 505, + 138 + ], + "score": 1.0, + "content": "a noise input to the generator sampled from", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 135, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 284, + 149 + ], + "score": 1.0, + "content": "a multivariate standard normal distribution", + "type": "text" + }, + { + "bbox": [ + 284, + 136, + 317, + 148 + ], + "score": 0.93, + "content": "\\mathcal { N } ( 0 , I )", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 135, + 338, + 149 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 338, + 136, + 346, + 146 + ], + "score": 0.8, + "content": "\\mathcal { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 135, + 505, + 149 + ], + "score": 1.0, + "content": "be the set of real images shown to the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 146, + 197, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 197, + 159 + ], + "score": 1.0, + "content": "discriminator. Define:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 101, + 505, + 159 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 161, + 160, + 450, + 207 + ], + "lines": [ + { + "bbox": [ + 161, + 160, + 450, + 207 + ], + "spans": [ + { + "bbox": [ + 161, + 160, + 450, + 207 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { L _ { \\mathrm { f a k e i m a g e } } = } & { } & { E _ { z \\sim \\mathcal { N } ( 0 , I ) } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( G ( z ) ) = \\mathrm { f a k e } ) \\right] } \\\\ { L _ { \\mathrm { f a k e i m a g e f o o l s } } = } & { } & { E _ { z \\sim \\mathcal { N } ( 0 , I ) } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( G ( z ) ) = \\mathrm { r e a l } ) \\right] } \\\\ { L _ { \\mathrm { r e a l i m a g e } } = } & { } & { E _ { i \\sim \\mathcal { Z } } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( i ) = \\mathrm { r e a l } ) \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "0088b43259ac008782f8d88cec6de49629fc3381a9b69629b272e6a48dd3be36.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 161, + 160, + 450, + 175.66666666666666 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 161, + 175.66666666666666, + 450, + 191.33333333333331 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 161, + 191.33333333333331, + 450, + 206.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 506, + 236 + ], + "score": 1.0, + "content": "Then the discriminator and generator are trained to maximize the following losses (respectively),", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 234, + 306, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 234, + 147, + 245 + ], + "score": 1.0, + "content": "where the", + "type": "text" + }, + { + "bbox": [ + 148, + 235, + 160, + 244 + ], + "score": 0.87, + "content": "w _ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 234, + 306, + 245 + ], + "score": 1.0, + "content": "are weights set as hyperparameters:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 221, + 506, + 245 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 190, + 247, + 421, + 276 + ], + "lines": [ + { + "bbox": [ + 190, + 247, + 421, + 276 + ], + "spans": [ + { + "bbox": [ + 190, + 247, + 421, + 276 + ], + "score": 0.84, + "content": "\\begin{array} { r l r } { L _ { \\mathrm { d i s c r i m i n a t o r } } = } & { { } } & { L _ { \\mathrm { f a k e i m a g e } } + L _ { \\mathrm { r e a l i m a g e } } } \\\\ { L _ { \\mathrm { G e n e r a t o r } } = } & { { } } & { L _ { \\mathrm { f a k e i m a g e f o o l s } } } \\end{array}", + "type": "interline_equation", + "image_path": "6cd424a9880748934e609233a2c271a2dfa0d7a3a7cf644a9761608799819961.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 190, + 247, + 421, + 261.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 190, + 261.5, + 421, + 276.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 505, + 323 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 505, + 291 + ], + "score": 1.0, + "content": "Note that there is a difference between these losses and the standard GAN formulation given", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 288, + 507, + 304 + ], + "spans": [ + { + "bbox": [ + 104, + 288, + 297, + 304 + ], + "score": 1.0, + "content": "in (Goodfellow et al., 2014) – we maximize", + "type": "text" + }, + { + "bbox": [ + 297, + 290, + 317, + 300 + ], + "score": 0.31, + "content": "L _ { \\mathrm { f a k e } }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 288, + 354, + 304 + ], + "score": 1.0, + "content": "image fools", + "type": "text" + }, + { + "bbox": [ + 354, + 290, + 393, + 301 + ], + "score": 0.6, + "content": "= \\ \\log \\ P", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 288, + 507, + 304 + ], + "score": 1.0, + "content": "(classified real) rather than", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 300, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 158, + 312 + ], + "score": 1.0, + "content": "minimizing", + "type": "text" + }, + { + "bbox": [ + 158, + 300, + 201, + 312 + ], + "score": 0.72, + "content": "\\log { ( 1 - P }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 301, + 506, + 312 + ], + "score": 1.0, + "content": "(classified real)). This seems to result in the generation of slightly better", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 312, + 185, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 185, + 324 + ], + "score": 1.0, + "content": "images in practice.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 277, + 507, + 324 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 328, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "This GAN was trained on a dataset consisting of landscape images of mountains and coastlines (see", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 505, + 352 + ], + "score": 1.0, + "content": "Appendix C.2 for details of the architecture and training). It is worth noting that this generative", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 351, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 505, + 362 + ], + "score": 1.0, + "content": "model is not photorealistic (see Fig. 3a for some samples). Its expressive capacity is limited, and it", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 360, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 374 + ], + "score": 1.0, + "content": "has clear output modes with limited intra-mode variability. However, for our purposes this may not", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "matter. Indeed, it is in some ways more interesting if we can tweak this model to optimize for many", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 383, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 505, + 395 + ], + "score": 1.0, + "content": "objective functions, since its limited expressive capacity will make it more difficult for us to estimate", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "and pursue the real objective – a limited set of images will effectively give us fewer points to estimate", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 405, + 504, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 504, + 417 + ], + "score": 1.0, + "content": "the PIR function from, and will reduce the space in which the model can easily produce images, thus", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "reducing the possibility of getting very optimal images from the model. For example, a model which", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 427, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 439 + ], + "score": 1.0, + "content": "produces images of birds may not produce data points which provide good estimates of a PIR based", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "on how much the image looks like a car, and even if it could, it may not be able to produce images", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "which are “more car-like.” If we are able to succeed in improving PIRs with this generative model, it", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 460, + 388, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 388, + 472 + ], + "score": 1.0, + "content": "is likely that a better generative model would yield even better results.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 328, + 506, + 472 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 484, + 221, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 484, + 222, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 222, + 496 + ], + "score": 1.0, + "content": "2.2 USER INTERACTIONS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "We will show these images to users in a variety of ways, depending on our target domain. For the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "score": 1.0, + "content": "purposes of this paper, however, we will use simulated interaction data (see Section 3 for details).", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 526, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 506, + 539 + ], + "score": 1.0, + "content": "Of course, since showing images to users is an expensive prospect, we wanted to limit the size of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "the datasets we used to train the model. Typical datasets used to train vision models are on the order", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "of millions of images, (e.g. ImageNet (Russakovsky et al., 2015)), but it is completely infeasible", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 559, + 504, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 504, + 571 + ], + "score": 1.0, + "content": "to collect user data on this number of images. We estimated that we could show 1000 images each", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "1000 times to generate our datasets. We used these dataset sizes and number of impressions for all", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 581, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 506, + 595 + ], + "score": 1.0, + "content": "experiments discussed here, and added noise to the PIRs that was binomially distributed according to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 592, + 473, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 473, + 605 + ], + "score": 1.0, + "content": "the number of times each image was shown and the “true” PIR simulate from the objective.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 505, + 506, + 605 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 617, + 234, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 235, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 235, + 630 + ], + "score": 1.0, + "content": "2.3 PIR ESTIMATOR MODEL", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "The final component of our system is the PIR estimator model, which learns to predict PIR from a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 293, + 661 + ], + "score": 1.0, + "content": "background image. We denote this model by", + "type": "text" + }, + { + "bbox": [ + 293, + 649, + 372, + 660 + ], + "score": 0.85, + "content": "R : { \\mathrm { i m a g e } } [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 649, + 505, + 661 + ], + "score": 1.0, + "content": ". We parameterize this model as", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "a deep neural network. Specifically, we take the Inception v2 architecture (Szegedy et al., 2016),", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "score": 1.0, + "content": "remove the output layer, and replace it with a fully-connected layer to PIR estimates. We initialize the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 681, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 506, + 694 + ], + "score": 1.0, + "content": "Inception v2 parameters from a version of the model trained on [dataset redacted for blind review].", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 693, + 249, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 249, + 704 + ], + "score": 1.0, + "content": "See Appendix C.3 for more details.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 637, + 506, + 704 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "Why did we not make estimated PIR simply another auxiliary output from the discriminator, like", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "class in the ACGAN (Appendix C.1)? Because the PIR estimator needs to be held constant in order", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "score": 1.0, + "content": "to provide an accurate training objective. If the PIR estimates were produced by the discriminator,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "then as the discriminator changed to accurately discriminate the evolving generator images, the PIR", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 138 + ], + "score": 1.0, + "content": "estimates would tend to drift without a ground-truth to train them on. Separating the discriminator", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 150 + ], + "score": 1.0, + "content": "and the PIR estimator allows us to freeze the PIR estimator while still letting the discriminator adapt.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 162, + 191, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 192, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 192, + 175 + ], + "score": 1.0, + "content": "2.4 INTEGRATION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Once we have trained a PIR estimator model, we have to use it to improve the GAN. We do this as", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 158, + 206 + ], + "score": 1.0, + "content": "follows. Let", + "type": "text" + }, + { + "bbox": [ + 159, + 194, + 167, + 204 + ], + "score": 0.78, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 193, + 372, + 206 + ], + "score": 1.0, + "content": "denote the PIR estimator model, as above. Define", + "type": "text" + }, + { + "bbox": [ + 372, + 194, + 391, + 205 + ], + "score": 0.89, + "content": "L _ { \\mathrm { P I R } }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 193, + 505, + 206 + ], + "score": 1.0, + "content": "to be the expectation of the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 205, + 370, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 370, + 217 + ], + "score": 1.0, + "content": "estimated PIR produced over images sampled from the generator:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 222, + 374, + 236 + ], + "lines": [ + { + "bbox": [ + 237, + 222, + 374, + 236 + ], + "spans": [ + { + "bbox": [ + 237, + 222, + 374, + 236 + ], + "score": 0.91, + "content": "{ \\cal L } _ { \\mathrm { P I R } } = { \\cal E } _ { z \\sim { \\cal N } ( 0 , I ) , c \\sim { \\cal C } } \\left[ R ( G ( z ) ) \\right]", + "type": "interline_equation", + "image_path": "31e98b68a936132de4225252446cd8cbd2a29d8084557596bd57aef4e9dd4d44.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 237, + 222, + 374, + 236 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 241, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 443, + 254 + ], + "score": 1.0, + "content": "Then we simply supplement the generator loss by adding this term times a weight", + "type": "text" + }, + { + "bbox": [ + 444, + 243, + 468, + 252 + ], + "score": 0.88, + "content": "w _ { P I R }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 240, + 506, + 254 + ], + "score": 1.0, + "content": ", set as a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 252, + 174, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 174, + 264 + ], + "score": 1.0, + "content": "hyperparameter:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 269, + 388, + 282 + ], + "lines": [ + { + "bbox": [ + 223, + 269, + 388, + 282 + ], + "spans": [ + { + "bbox": [ + 223, + 269, + 388, + 282 + ], + "score": 0.91, + "content": "L _ { \\mathrm { G e n e r a t o r } } = L _ { \\mathrm { f a k e i m a g e f o o l s } } + w _ { P I R } L _ { P I R }", + "type": "interline_equation", + "image_path": "f70a2347392d7ceed6a2bbc190de7dcbf0d19af1844a2d8217ec78d2b91149cf.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 223, + 269, + 388, + 282 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 287, + 505, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 136, + 300 + ], + "score": 1.0, + "content": "We set", + "type": "text" + }, + { + "bbox": [ + 136, + 288, + 195, + 299 + ], + "score": 0.9, + "content": "w _ { P I R } = 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "as this made the magnitude of the PIR loss and the other loss terms roughly", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "comparable. Otherwise we used the same parameters as in the GAN training above, except that", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 236, + 322 + ], + "score": 1.0, + "content": "we reduced the learning rate to", + "type": "text" + }, + { + "bbox": [ + 237, + 309, + 258, + 320 + ], + "score": 0.9, + "content": "1 0 ^ { - 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "to allow the system to adapt more smoothly to the multiple", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 320, + 281, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 281, + 334 + ], + "score": 1.0, + "content": "objectives, and we trained for 50,000 steps.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 108, + 348, + 154, + 360 + ], + "lines": [ + { + "bbox": [ + 104, + 346, + 156, + 363 + ], + "spans": [ + { + "bbox": [ + 104, + 346, + 156, + 363 + ], + "score": 1.0, + "content": "3 DATA", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 372, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "In this paper, we use simulated interaction data as proof of concept. This raises an issue: what", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 383, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 397 + ], + "score": 1.0, + "content": "functions should we use to simulate user interactions? Human behavior is complex, and if we already", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "knew precisely what guided user interactions, there would be no need to actually collect human", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "behavioral data at all. Since we don’t know what features will guide human behavior, the next best", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 417, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 429 + ], + "score": 1.0, + "content": "thing we can do is to ensure that our system is able to alter the image generation model in a broad", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 426, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 442 + ], + "score": 1.0, + "content": "variety of ways, ranging from low level features (like making the images more colorful) to altering", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "complex semantic features (such as including more plants in outdoor scenery). We also want to avoid", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 449, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 463 + ], + "score": 1.0, + "content": "hand-engineering tasks to the greatest extent possible. We present an overview of our approaches to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 461, + 358, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 358, + 473 + ], + "score": 1.0, + "content": "simulating PIR data below, see Appendix C.4 for more details.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 108, + 485, + 202, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 203, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 203, + 499 + ], + "score": 1.0, + "content": "3.1 VGG FEATURES", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 506, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "The first approach we took to evaluating our system’s ability to train for different features was to use", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 517, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 506, + 530 + ], + "score": 1.0, + "content": "activity from hidden layers of a computer vision model, specifically VGG 16 (Simonyan & Zisserman,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "2014) trained on ImageNet (Russakovsky et al., 2015). In particular, we took the activity of a single", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "filter in a layer of VGG relative to the overall activity of that layer. This approach to simulating PIRs", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "has several benefits. First, it gives a wide variety of complex objectives that can nevertheless be easily", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "computed to simulate data. Second, models like VGG exhibit hierarchical organization, where lower", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "levels generally respond to lower-level features such as edges and colors, while higher levels respond", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "to higher-level semantic features such as faces (Zeiler & Fergus, 2014), and the represented features", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "relate to those in human and macaque visual cortex (Yamins et al., 2014). Thus VGG features give a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 604, + 463, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 463, + 618 + ], + "score": 1.0, + "content": "wide range of objectives which we may relate to the human perception we wish to target.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 506, + 634 + ], + "score": 1.0, + "content": "There are some caveats to this approach, however. First, although the higher layers of CNNs are", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "somewhat selective for “abstract” object categories, they are also fooled by adversarial images", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "that humans would not be, and directly optimizing inputs for these high level features does not", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "actually produce semantically meaningful images (Nguyen et al., 2014). Thus, even if our system", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "succeeds in increasing activity in a targeted layer which is semantically selective, it will likely do so", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "by adversarially exploiting particulars of VGG 16’s parameterization of the classification problem", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "(although the fact that we are not backpropagating through the true objective will make this harder).", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "It is not necessarily a failure of the system if it exploits simple features of the objective it is given to", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 507, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 507, + 723 + ], + "score": 1.0, + "content": "increase PIRs – indeed, it should be seen as a success, as long as it is generalizes to novel images.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "However, success on this task does not necessarily guarantee success on modifying semantic content", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 43.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "Why did we not make estimated PIR simply another auxiliary output from the discriminator, like", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "class in the ACGAN (Appendix C.1)? Because the PIR estimator needs to be held constant in order", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "score": 1.0, + "content": "to provide an accurate training objective. If the PIR estimates were produced by the discriminator,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "then as the discriminator changed to accurately discriminate the evolving generator images, the PIR", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 138 + ], + "score": 1.0, + "content": "estimates would tend to drift without a ground-truth to train them on. Separating the discriminator", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 150 + ], + "score": 1.0, + "content": "and the PIR estimator allows us to freeze the PIR estimator while still letting the discriminator adapt.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 83, + 506, + 150 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 162, + 191, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 192, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 192, + 175 + ], + "score": 1.0, + "content": "2.4 INTEGRATION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Once we have trained a PIR estimator model, we have to use it to improve the GAN. We do this as", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 158, + 206 + ], + "score": 1.0, + "content": "follows. Let", + "type": "text" + }, + { + "bbox": [ + 159, + 194, + 167, + 204 + ], + "score": 0.78, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 193, + 372, + 206 + ], + "score": 1.0, + "content": "denote the PIR estimator model, as above. Define", + "type": "text" + }, + { + "bbox": [ + 372, + 194, + 391, + 205 + ], + "score": 0.89, + "content": "L _ { \\mathrm { P I R } }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 193, + 505, + 206 + ], + "score": 1.0, + "content": "to be the expectation of the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 205, + 370, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 370, + 217 + ], + "score": 1.0, + "content": "estimated PIR produced over images sampled from the generator:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 182, + 505, + 217 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 222, + 374, + 236 + ], + "lines": [ + { + "bbox": [ + 237, + 222, + 374, + 236 + ], + "spans": [ + { + "bbox": [ + 237, + 222, + 374, + 236 + ], + "score": 0.91, + "content": "{ \\cal L } _ { \\mathrm { P I R } } = { \\cal E } _ { z \\sim { \\cal N } ( 0 , I ) , c \\sim { \\cal C } } \\left[ R ( G ( z ) ) \\right]", + "type": "interline_equation", + "image_path": "31e98b68a936132de4225252446cd8cbd2a29d8084557596bd57aef4e9dd4d44.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 237, + 222, + 374, + 236 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 241, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 443, + 254 + ], + "score": 1.0, + "content": "Then we simply supplement the generator loss by adding this term times a weight", + "type": "text" + }, + { + "bbox": [ + 444, + 243, + 468, + 252 + ], + "score": 0.88, + "content": "w _ { P I R }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 240, + 506, + 254 + ], + "score": 1.0, + "content": ", set as a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 252, + 174, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 174, + 264 + ], + "score": 1.0, + "content": "hyperparameter:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 240, + 506, + 264 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 269, + 388, + 282 + ], + "lines": [ + { + "bbox": [ + 223, + 269, + 388, + 282 + ], + "spans": [ + { + "bbox": [ + 223, + 269, + 388, + 282 + ], + "score": 0.91, + "content": "L _ { \\mathrm { G e n e r a t o r } } = L _ { \\mathrm { f a k e i m a g e f o o l s } } + w _ { P I R } L _ { P I R }", + "type": "interline_equation", + "image_path": "f70a2347392d7ceed6a2bbc190de7dcbf0d19af1844a2d8217ec78d2b91149cf.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 223, + 269, + 388, + 282 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 287, + 505, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 136, + 300 + ], + "score": 1.0, + "content": "We set", + "type": "text" + }, + { + "bbox": [ + 136, + 288, + 195, + 299 + ], + "score": 0.9, + "content": "w _ { P I R } = 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "as this made the magnitude of the PIR loss and the other loss terms roughly", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "comparable. Otherwise we used the same parameters as in the GAN training above, except that", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 236, + 322 + ], + "score": 1.0, + "content": "we reduced the learning rate to", + "type": "text" + }, + { + "bbox": [ + 237, + 309, + 258, + 320 + ], + "score": 0.9, + "content": "1 0 ^ { - 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "to allow the system to adapt more smoothly to the multiple", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 320, + 281, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 281, + 334 + ], + "score": 1.0, + "content": "objectives, and we trained for 50,000 steps.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 287, + 506, + 334 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 348, + 154, + 360 + ], + "lines": [ + { + "bbox": [ + 104, + 346, + 156, + 363 + ], + "spans": [ + { + "bbox": [ + 104, + 346, + 156, + 363 + ], + "score": 1.0, + "content": "3 DATA", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 372, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "In this paper, we use simulated interaction data as proof of concept. This raises an issue: what", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 383, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 397 + ], + "score": 1.0, + "content": "functions should we use to simulate user interactions? Human behavior is complex, and if we already", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "knew precisely what guided user interactions, there would be no need to actually collect human", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "behavioral data at all. Since we don’t know what features will guide human behavior, the next best", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 417, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 429 + ], + "score": 1.0, + "content": "thing we can do is to ensure that our system is able to alter the image generation model in a broad", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 426, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 442 + ], + "score": 1.0, + "content": "variety of ways, ranging from low level features (like making the images more colorful) to altering", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "complex semantic features (such as including more plants in outdoor scenery). We also want to avoid", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 449, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 463 + ], + "score": 1.0, + "content": "hand-engineering tasks to the greatest extent possible. We present an overview of our approaches to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 461, + 358, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 358, + 473 + ], + "score": 1.0, + "content": "simulating PIR data below, see Appendix C.4 for more details.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 372, + 506, + 473 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 485, + 202, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 203, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 203, + 499 + ], + "score": 1.0, + "content": "3.1 VGG FEATURES", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 506, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "The first approach we took to evaluating our system’s ability to train for different features was to use", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 517, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 506, + 530 + ], + "score": 1.0, + "content": "activity from hidden layers of a computer vision model, specifically VGG 16 (Simonyan & Zisserman,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "2014) trained on ImageNet (Russakovsky et al., 2015). In particular, we took the activity of a single", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "filter in a layer of VGG relative to the overall activity of that layer. This approach to simulating PIRs", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "has several benefits. First, it gives a wide variety of complex objectives that can nevertheless be easily", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "computed to simulate data. Second, models like VGG exhibit hierarchical organization, where lower", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "levels generally respond to lower-level features such as edges and colors, while higher levels respond", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "to higher-level semantic features such as faces (Zeiler & Fergus, 2014), and the represented features", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "relate to those in human and macaque visual cortex (Yamins et al., 2014). Thus VGG features give a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 604, + 463, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 463, + 618 + ], + "score": 1.0, + "content": "wide range of objectives which we may relate to the human perception we wish to target.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 506, + 506, + 618 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 506, + 634 + ], + "score": 1.0, + "content": "There are some caveats to this approach, however. First, although the higher layers of CNNs are", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "somewhat selective for “abstract” object categories, they are also fooled by adversarial images", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "that humans would not be, and directly optimizing inputs for these high level features does not", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "actually produce semantically meaningful images (Nguyen et al., 2014). Thus, even if our system", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "succeeds in increasing activity in a targeted layer which is semantically selective, it will likely do so", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "by adversarially exploiting particulars of VGG 16’s parameterization of the classification problem", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "(although the fact that we are not backpropagating through the true objective will make this harder).", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "It is not necessarily a failure of the system if it exploits simple features of the objective it is given to", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 507, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 507, + 723 + ], + "score": 1.0, + "content": "increase PIRs – indeed, it should be seen as a success, as long as it is generalizes to novel images.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "However, success on this task does not necessarily guarantee success on modifying semantic content", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "when interacting with actual humans. It may be easier for the PIR estimator model (which is based on", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 506, + 106 + ], + "score": 1.0, + "content": "a CNN) to learn objectives which come from another CNN than more general possible objectives. The", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "fact that adversarial examples can sometimes transfer between networks with different architectures", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "(Liu et al., 2016) suggests that the computations being performed by these networks are somewhat", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 484, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 484, + 139 + ], + "score": 1.0, + "content": "architecture invariant. Thus CNN objectives may be easier for our estimator than human ones.", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 622, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "when interacting with actual humans. It may be easier for the PIR estimator model (which is based on", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 506, + 106 + ], + "score": 1.0, + "content": "a CNN) to learn objectives which come from another CNN than more general possible objectives. The", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "fact that adversarial examples can sometimes transfer between networks with different architectures", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "(Liu et al., 2016) suggests that the computations being performed by these networks are somewhat", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 484, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 484, + 139 + ], + "score": 1.0, + "content": "architecture invariant. Thus CNN objectives may be easier for our estimator than human ones.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "We have tried to minimize these problems to the greatest extent possible by using different network", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "architectures (Inception V2 and VGG 16, respectively) trained on different datasets ([hidden] and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "ImageNet (Russakovsky et al., 2015), respectively) for the estimator and the objective. However,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "we cannot be certain that the network is not “cheating” in some way on the VGG 16 tasks, so our", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "results must be considered with this qualification in mind. Despite this, we think that evaluating", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "our system’s ability to optimize for objectives generated from various layers of VGG will show its", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "ability to optimize for a variety of complex objectives, and thus will serve as a useful indicator of its", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 276, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 276, + 232 + ], + "score": 1.0, + "content": "potential to improve PIRs from real users.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 249, + 212, + 259 + ], + "lines": [ + { + "bbox": [ + 106, + 248, + 213, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 213, + 261 + ], + "score": 1.0, + "content": "3.2 MULTIPLE FILTERS", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 505, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "After using our system on the tasks above, we noted that its performance was quite poor at layers 5,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "6, and 7 of VGG compared to other tasks (see Fig. 2). This could suggest that our system was unable", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "score": 1.0, + "content": "to capture the complex features represented at the higher levels of VGG. However, we also noticed", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "score": 1.0, + "content": "that the feature representations at these layers tended to be quite sparse, so many of the simulated", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "score": 1.0, + "content": "PIRs we generated were actually zero to within the bin width of our PIR estimator (see Appendix B", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "Fig. 8 for a plot of how this affected learning). In order to evaluate whether the poor performance", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "score": 1.0, + "content": "of our system at the higher layers of VGG was due to the number of zeros or to the complexity of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 471, + 360 + ], + "score": 1.0, + "content": "the features, we created less sparse features from these layers by simply targeting a set of", + "type": "text" + }, + { + "bbox": [ + 471, + 348, + 478, + 357 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "filters", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 359, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 505, + 370 + ], + "score": 1.0, + "content": "sampled without replacement from the layer, rather than a single filter. The single filter cases above", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 301, + 381 + ], + "score": 1.0, + "content": "can be thought of as a special case of this, where", + "type": "text" + }, + { + "bbox": [ + 302, + 370, + 326, + 380 + ], + "score": 0.9, + "content": "k = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 370, + 473, + 381 + ], + "score": 1.0, + "content": ". To complement these, we also tried", + "type": "text" + }, + { + "bbox": [ + 473, + 370, + 503, + 380 + ], + "score": 0.9, + "content": "k = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 370, + 506, + 381 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 386, + 505, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 398 + ], + "score": 1.0, + "content": "This can also be thought of as perhaps a more realistic simulation of human behavior, in the sense", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "that it is highly unlikely that there is a single feature which influences human PIRs. Rather, there are", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 409, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 505, + 421 + ], + "score": 1.0, + "content": "probably many related features which influence PIR in various ways. Thus it is important to evaluate", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 420, + 346, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 346, + 431 + ], + "score": 1.0, + "content": "our system’s ability to target these types of features as well.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 107, + 448, + 167, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 446, + 169, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 169, + 461 + ], + "score": 1.0, + "content": "3.3 COLORS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 469, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "Finally, we also considered some simpler objectives based on targeting specific colors in the output", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "images, or targeting vertical bands of two different colors, one in each half of the image, or three", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 492, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 504 + ], + "score": 1.0, + "content": "colors, one in each third of the image. These objectives provide a useful complement to the VGG", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 503, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 506, + 515 + ], + "score": 1.0, + "content": "objectives above. Although the single color objectives may be relevant to the classification task", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "VGG 16 performs, the split color tasks are less likely to be relevant to classification. Note that it is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 525, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 506, + 537 + ], + "score": 1.0, + "content": "important that we split the images along the width instead of the height dimension, as there may well", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 534, + 507, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 507, + 550 + ], + "score": 1.0, + "content": "be semantically relevant features corresponding to color divisions along the height dimension, e.g.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "a blue upper half and green lower half likely correlates with outdoor images, which would provide", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "useful class information. By contrast, it is harder to imagine circumstances where different colors", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "on the left and right halves of the image are semantically predictive, especially since flipping left", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 580, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 505, + 591 + ], + "score": 1.0, + "content": "to right is usually included in the data augmentation for computer vision systems. Thus success on", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 590, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 482, + 604 + ], + "score": 1.0, + "content": "optimizing for these objectives would increase our confidence in the generality of our system.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 108, + 621, + 172, + 634 + ], + "lines": [ + { + "bbox": [ + 104, + 619, + 174, + 637 + ], + "spans": [ + { + "bbox": [ + 104, + 619, + 174, + 637 + ], + "score": 1.0, + "content": "4 RESULTS", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "We present our results in terms of the change in mean PIR from 1000 images produced by the GAN", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "before tuning to 1000 images produced after tuning, or in terms of the effect size of this change", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 145, + 684 + ], + "score": 1.0, + "content": "(Cohen’s", + "type": "text" + }, + { + "bbox": [ + 146, + 671, + 152, + 681 + ], + "score": 0.62, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 671, + 505, + 684 + ], + "score": 1.0, + "content": ", i.e. the change in mean PIR standardized by the standard deviation of the PIRs in the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 681, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 696 + ], + "score": 1.0, + "content": "pre- and post-tuning image sets). We assess whether these changes are significant by performing a", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 693, + 500, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 294, + 706 + ], + "score": 1.0, + "content": "Welch’s t-test (with a significance threshold of", + "type": "text" + }, + { + "bbox": [ + 294, + 693, + 337, + 703 + ], + "score": 0.85, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 693, + 500, + 706 + ], + "score": 1.0, + "content": ") between the pre- and post-tuning PIRs.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Overall, our system was quite successful at improving PIRs across a range of simulated objective", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 385, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 385, + 732 + ], + "score": 1.0, + "content": "functions (see Fig. 2). Below, we discuss these results in more detail.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 137 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 105, + 83, + 506, + 139 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "We have tried to minimize these problems to the greatest extent possible by using different network", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "architectures (Inception V2 and VGG 16, respectively) trained on different datasets ([hidden] and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "ImageNet (Russakovsky et al., 2015), respectively) for the estimator and the objective. However,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "we cannot be certain that the network is not “cheating” in some way on the VGG 16 tasks, so our", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "results must be considered with this qualification in mind. Despite this, we think that evaluating", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "our system’s ability to optimize for objectives generated from various layers of VGG will show its", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "ability to optimize for a variety of complex objectives, and thus will serve as a useful indicator of its", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 276, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 276, + 232 + ], + "score": 1.0, + "content": "potential to improve PIRs from real users.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 143, + 506, + 232 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 249, + 212, + 259 + ], + "lines": [ + { + "bbox": [ + 106, + 248, + 213, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 213, + 261 + ], + "score": 1.0, + "content": "3.2 MULTIPLE FILTERS", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 505, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "After using our system on the tasks above, we noted that its performance was quite poor at layers 5,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "6, and 7 of VGG compared to other tasks (see Fig. 2). This could suggest that our system was unable", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "score": 1.0, + "content": "to capture the complex features represented at the higher levels of VGG. However, we also noticed", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "score": 1.0, + "content": "that the feature representations at these layers tended to be quite sparse, so many of the simulated", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "score": 1.0, + "content": "PIRs we generated were actually zero to within the bin width of our PIR estimator (see Appendix B", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "Fig. 8 for a plot of how this affected learning). In order to evaluate whether the poor performance", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "score": 1.0, + "content": "of our system at the higher layers of VGG was due to the number of zeros or to the complexity of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 471, + 360 + ], + "score": 1.0, + "content": "the features, we created less sparse features from these layers by simply targeting a set of", + "type": "text" + }, + { + "bbox": [ + 471, + 348, + 478, + 357 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "filters", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 359, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 505, + 370 + ], + "score": 1.0, + "content": "sampled without replacement from the layer, rather than a single filter. The single filter cases above", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 301, + 381 + ], + "score": 1.0, + "content": "can be thought of as a special case of this, where", + "type": "text" + }, + { + "bbox": [ + 302, + 370, + 326, + 380 + ], + "score": 0.9, + "content": "k = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 370, + 473, + 381 + ], + "score": 1.0, + "content": ". To complement these, we also tried", + "type": "text" + }, + { + "bbox": [ + 473, + 370, + 503, + 380 + ], + "score": 0.9, + "content": "k = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 370, + 506, + 381 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 270, + 506, + 381 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 386, + 505, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 398 + ], + "score": 1.0, + "content": "This can also be thought of as perhaps a more realistic simulation of human behavior, in the sense", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "that it is highly unlikely that there is a single feature which influences human PIRs. Rather, there are", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 409, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 505, + 421 + ], + "score": 1.0, + "content": "probably many related features which influence PIR in various ways. Thus it is important to evaluate", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 420, + 346, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 346, + 431 + ], + "score": 1.0, + "content": "our system’s ability to target these types of features as well.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 386, + 505, + 431 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 448, + 167, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 446, + 169, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 169, + 461 + ], + "score": 1.0, + "content": "3.3 COLORS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 469, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "Finally, we also considered some simpler objectives based on targeting specific colors in the output", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "images, or targeting vertical bands of two different colors, one in each half of the image, or three", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 492, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 504 + ], + "score": 1.0, + "content": "colors, one in each third of the image. These objectives provide a useful complement to the VGG", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 503, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 506, + 515 + ], + "score": 1.0, + "content": "objectives above. Although the single color objectives may be relevant to the classification task", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "VGG 16 performs, the split color tasks are less likely to be relevant to classification. Note that it is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 525, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 506, + 537 + ], + "score": 1.0, + "content": "important that we split the images along the width instead of the height dimension, as there may well", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 534, + 507, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 507, + 550 + ], + "score": 1.0, + "content": "be semantically relevant features corresponding to color divisions along the height dimension, e.g.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "a blue upper half and green lower half likely correlates with outdoor images, which would provide", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "useful class information. By contrast, it is harder to imagine circumstances where different colors", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "on the left and right halves of the image are semantically predictive, especially since flipping left", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 580, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 505, + 591 + ], + "score": 1.0, + "content": "to right is usually included in the data augmentation for computer vision systems. Thus success on", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 590, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 482, + 604 + ], + "score": 1.0, + "content": "optimizing for these objectives would increase our confidence in the generality of our system.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 469, + 507, + 604 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 621, + 172, + 634 + ], + "lines": [ + { + "bbox": [ + 104, + 619, + 174, + 637 + ], + "spans": [ + { + "bbox": [ + 104, + 619, + 174, + 637 + ], + "score": 1.0, + "content": "4 RESULTS", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "We present our results in terms of the change in mean PIR from 1000 images produced by the GAN", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "before tuning to 1000 images produced after tuning, or in terms of the effect size of this change", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 145, + 684 + ], + "score": 1.0, + "content": "(Cohen’s", + "type": "text" + }, + { + "bbox": [ + 146, + 671, + 152, + 681 + ], + "score": 0.62, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 671, + 505, + 684 + ], + "score": 1.0, + "content": ", i.e. the change in mean PIR standardized by the standard deviation of the PIRs in the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 681, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 696 + ], + "score": 1.0, + "content": "pre- and post-tuning image sets). We assess whether these changes are significant by performing a", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 693, + 500, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 294, + 706 + ], + "score": 1.0, + "content": "Welch’s t-test (with a significance threshold of", + "type": "text" + }, + { + "bbox": [ + 294, + 693, + 337, + 703 + ], + "score": 0.85, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 693, + 500, + 706 + ], + "score": 1.0, + "content": ") between the pre- and post-tuning PIRs.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 649, + 506, + 706 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Overall, our system was quite successful at improving PIRs across a range of simulated objective", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 385, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 385, + 732 + ], + "score": 1.0, + "content": "functions (see Fig. 2). Below, we discuss these results in more detail.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5, + "bbox_fs": [ + 106, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 82, + 482, + 325 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 82, + 482, + 325 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 82, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 127, + 82, + 482, + 325 + ], + "score": 0.972, + "type": "image", + "image_path": "452a4c3d84260f277ccf32fa578a0b0396ce5df970f4fe40eb77375a745ada9e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 127, + 82, + 482, + 163.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 127, + 163.0, + 482, + 244.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 244.0, + 482, + 325.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 334, + 506, + 401 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 335, + 507, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 507, + 348 + ], + "score": 1.0, + "content": "Figure 2: Effect size (number of standard deviations change in mean PIR) on a variety of tasks.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "Values greater than zero indicate improvement. Each panel represents a set of tasks (such as single", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "filters from VGG), each color/column represents a subset (such as a single layer within VGG), and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "each point represents one objective (such as a single filter from that layer) for which we optimized", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 380, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 482, + 391 + ], + "score": 1.0, + "content": "the generative model. Points for which the change in mean PIR is not significant by a Welch’s", + "type": "text" + }, + { + "bbox": [ + 482, + 380, + 487, + 389 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 380, + 505, + 391 + ], + "score": 1.0, + "content": "-test", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 389, + 328, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 183, + 403 + ], + "score": 1.0, + "content": "with a threshold of", + "type": "text" + }, + { + "bbox": [ + 184, + 390, + 227, + 401 + ], + "score": 0.89, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 389, + 328, + 403 + ], + "score": 1.0, + "content": "are partially transparent.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "title", + "bbox": [ + 108, + 423, + 209, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 210, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 210, + 436 + ], + "score": 1.0, + "content": "4.1 VGG OBJECTIVES", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 506, + 457 + ], + "score": 1.0, + "content": "Our system largely succeed at increasing PIRs on a variety of VGG objectives (see Fig. 2). However,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "score": 1.0, + "content": "there are several interesting patterns to note. First, the system is not particularly successful at targeting", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "single filters from the pool5, fc6, and fc7 layers. However, we believe this is due to the fact that filters", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 477, + 504, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 504, + 490 + ], + "score": 1.0, + "content": "in these layers produce relatively sparse activation, see section 3.2. Indeed, the performance of the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "system seems much more consistent when it is optimizing for sets of 20 filters than for single filters.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "score": 1.0, + "content": "Even when using 20 filters, however, there is a noticeable decline in the effect size of the improvement", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 514, + 504, + 528 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 323, + 528 + ], + "score": 1.0, + "content": "the system is able to make at higher layers of VGG", + "type": "text" + }, + { + "bbox": [ + 324, + 516, + 372, + 527 + ], + "score": 0.88, + "content": "\\beta = - 0 . 1 3", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 514, + 414, + 528 + ], + "score": 1.0, + "content": "per layer,", + "type": "text" + }, + { + "bbox": [ + 414, + 516, + 454, + 527 + ], + "score": 0.8, + "content": "t = - 7 . 8", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 514, + 458, + 528 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 458, + 515, + 504, + 527 + ], + "score": 0.85, + "content": "\\dot { p } < 1 0 ^ { - 1 0 }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "score": 1.0, + "content": "in a linear model controlling for initial standard deviation and percent zeros). This suggests that as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "the objectives grow more complex, the system may be finding less accurate approximations to them.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "However, the system’s continuing (if diminished) success at the higher layers of VGG suggests that", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 559, + 426, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 426, + 572 + ], + "score": 1.0, + "content": "our model is capable of at least partially capturing complex objective functions.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 108, + 584, + 217, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 218, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 218, + 597 + ], + "score": 1.0, + "content": "4.2 COLOR OBJECTIVES", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "Overall, the system performed quite well at optimizing for the color objectives, particularly the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "single and two-color results (see Fig. 2). It had more difficulty optimizing for the three-color results,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "and indeed had produced only very small improvements after the usual 50,000 tuning steps for the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "generative model, but after 500,000 steps it was able to produce significant improvements for two out", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 649, + 425, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 425, + 662 + ], + "score": 1.0, + "content": "of the three objectives (these longer training results are the ones included here).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "Because the color objectives are easiest to assess visually, we have included results for a variety of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "these objectives in Fig. 3. For the single color objectives, the improvement is quite clear, for example", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "the images in Fig. 3d appear much more blue than the pre-training ones. For the two color objectives,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "it appears that the system found the “trick” of reducing the third color, for example the red-green", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 709, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 721 + ], + "score": 1.0, + "content": "split images in Fig. 3e appear much less blue than the pre-training images. Even on the three-color", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "images where the system struggled, there are some visible signs of improvement, for example on the", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 82, + 482, + 325 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 82, + 482, + 325 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 82, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 127, + 82, + 482, + 325 + ], + "score": 0.972, + "type": "image", + "image_path": "452a4c3d84260f277ccf32fa578a0b0396ce5df970f4fe40eb77375a745ada9e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 127, + 82, + 482, + 163.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 127, + 163.0, + 482, + 244.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 244.0, + 482, + 325.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 334, + 506, + 401 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 335, + 507, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 507, + 348 + ], + "score": 1.0, + "content": "Figure 2: Effect size (number of standard deviations change in mean PIR) on a variety of tasks.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "Values greater than zero indicate improvement. Each panel represents a set of tasks (such as single", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "filters from VGG), each color/column represents a subset (such as a single layer within VGG), and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "each point represents one objective (such as a single filter from that layer) for which we optimized", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 380, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 482, + 391 + ], + "score": 1.0, + "content": "the generative model. Points for which the change in mean PIR is not significant by a Welch’s", + "type": "text" + }, + { + "bbox": [ + 482, + 380, + 487, + 389 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 380, + 505, + 391 + ], + "score": 1.0, + "content": "-test", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 389, + 328, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 183, + 403 + ], + "score": 1.0, + "content": "with a threshold of", + "type": "text" + }, + { + "bbox": [ + 184, + 390, + 227, + 401 + ], + "score": 0.89, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 389, + 328, + 403 + ], + "score": 1.0, + "content": "are partially transparent.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "title", + "bbox": [ + 108, + 423, + 209, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 210, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 210, + 436 + ], + "score": 1.0, + "content": "4.1 VGG OBJECTIVES", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 506, + 457 + ], + "score": 1.0, + "content": "Our system largely succeed at increasing PIRs on a variety of VGG objectives (see Fig. 2). However,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "score": 1.0, + "content": "there are several interesting patterns to note. First, the system is not particularly successful at targeting", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "single filters from the pool5, fc6, and fc7 layers. However, we believe this is due to the fact that filters", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 477, + 504, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 504, + 490 + ], + "score": 1.0, + "content": "in these layers produce relatively sparse activation, see section 3.2. Indeed, the performance of the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "system seems much more consistent when it is optimizing for sets of 20 filters than for single filters.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 443, + 506, + 501 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "score": 1.0, + "content": "Even when using 20 filters, however, there is a noticeable decline in the effect size of the improvement", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 514, + 504, + 528 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 323, + 528 + ], + "score": 1.0, + "content": "the system is able to make at higher layers of VGG", + "type": "text" + }, + { + "bbox": [ + 324, + 516, + 372, + 527 + ], + "score": 0.88, + "content": "\\beta = - 0 . 1 3", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 514, + 414, + 528 + ], + "score": 1.0, + "content": "per layer,", + "type": "text" + }, + { + "bbox": [ + 414, + 516, + 454, + 527 + ], + "score": 0.8, + "content": "t = - 7 . 8", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 514, + 458, + 528 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 458, + 515, + 504, + 527 + ], + "score": 0.85, + "content": "\\dot { p } < 1 0 ^ { - 1 0 }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "score": 1.0, + "content": "in a linear model controlling for initial standard deviation and percent zeros). This suggests that as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "the objectives grow more complex, the system may be finding less accurate approximations to them.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "However, the system’s continuing (if diminished) success at the higher layers of VGG suggests that", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 559, + 426, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 426, + 572 + ], + "score": 1.0, + "content": "our model is capable of at least partially capturing complex objective functions.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 505, + 506, + 572 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 584, + 217, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 218, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 218, + 597 + ], + "score": 1.0, + "content": "4.2 COLOR OBJECTIVES", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "Overall, the system performed quite well at optimizing for the color objectives, particularly the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "single and two-color results (see Fig. 2). It had more difficulty optimizing for the three-color results,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "and indeed had produced only very small improvements after the usual 50,000 tuning steps for the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "generative model, but after 500,000 steps it was able to produce significant improvements for two out", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 649, + 425, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 425, + 662 + ], + "score": 1.0, + "content": "of the three objectives (these longer training results are the ones included here).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 604, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "Because the color objectives are easiest to assess visually, we have included results for a variety of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "these objectives in Fig. 3. For the single color objectives, the improvement is quite clear, for example", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "the images in Fig. 3d appear much more blue than the pre-training ones. For the two color objectives,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "it appears that the system found the “trick” of reducing the third color, for example the red-green", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 709, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 721 + ], + "score": 1.0, + "content": "split images in Fig. 3e appear much less blue than the pre-training images. Even on the three-color", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "images where the system struggled, there are some visible signs of improvement, for example on the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "green-blue-red task the system has started producing a number of images with a blue streak in the", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 366, + 140, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 140, + 380 + ], + "score": 1.0, + "content": "middle.", + "type": "text", + "cross_page": true + } + ], + "index": 5 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 665, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 80, + 500, + 313 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 80, + 500, + 313 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 80, + 500, + 313 + ], + "spans": [ + { + "bbox": [ + 111, + 80, + 500, + 313 + ], + "score": 0.975, + "type": "image", + "image_path": "a677d9bfd60d90b339e6056f4eedc86e60ca0e0077525f722cfbef7db0088cd1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 80, + 500, + 157.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 157.66666666666669, + 500, + 235.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 235.33333333333337, + 500, + 313.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 120, + 323, + 487, + 335 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 322, + 488, + 336 + ], + "spans": [ + { + "bbox": [ + 123, + 322, + 488, + 336 + ], + "score": 1.0, + "content": "Figure 3: Color objective sample images. Samples are randomly drawn, not cherry-picked.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 108, + 357, + 504, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "green-blue-red task the system has started producing a number of images with a blue streak in the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 366, + 140, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 140, + 380 + ], + "score": 1.0, + "content": "middle.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 247, + 404 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 248, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 248, + 405 + ], + "score": 1.0, + "content": "4.3 SUPPLEMENTAL ANALYSES", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "We also conducted several supplemental analyses which can be found in detail Appendix A. In", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "summary, the initial variability in the PIR of the images used to train the system is strongly correlated", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 437, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 506, + 448 + ], + "score": 1.0, + "content": "with the amount of improvement the system makes in the PIR, the system fairly consistently underes-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "timates the performance it achieves (because of a detail of training procedure, see the Appendix), and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "iterating the process of improving PIRs yields better results for objectives from a lower layer of VGG", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 468, + 173, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 173, + 482 + ], + "score": 1.0, + "content": "but not a higher.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 497, + 190, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 192, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 192, + 512 + ], + "score": 1.0, + "content": "5 DISCUSSION", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "Overall, our system appears to be relatively successful. It can optimize a generative model to produce", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "images which target a wide variety of objectives, ranging from low-level visual features such as", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 544, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 557 + ], + "score": 1.0, + "content": "colors and early features of VGG to features computed at the top layers of VGG. This success across", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "a wide variety of objective functions allows us to be somewhat confident that our system will be able", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 566, + 351, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 351, + 579 + ], + "score": 1.0, + "content": "to achieve success in optimizing for real human interactions.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 583, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "Furthermore, the system did not require an inordinate amount of training data. In fact, we were able", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "to successfully estimate many different objective functions from only 1000 images, several orders of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "magnitude fewer than is typically used to train CNNs for vision tasks. Furthermore, these images", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "came from a very biased and narrow distribution (samples from our generative model) which is", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "score": 1.0, + "content": "reflective of neither the images that were used to pre-train the Inception model in the PIR estimator,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 639, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 649 + ], + "score": 1.0, + "content": "nor the images the VGG model (which produced the simulated objectives) was trained on. Our", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "success from this small amount of data suggests that not only will our system be able to optimize for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 659, + 464, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 464, + 672 + ], + "score": 1.0, + "content": "real human interactions, it will be able to do so from a feasible number of training points.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "These results are exciting – the model is able to approximate apparently complex objective functions", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "from a small amount of data, even though this data comes from a very biased distribution that is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "unrelated to most the objectives in question. But what is really being learned? In the case of the color", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "images, it’s clear that the model is doing something close to correct. However, for the objectives", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "derived from VGG we have no way to really assess whether the model is making the images better or", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 80, + 500, + 313 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 80, + 500, + 313 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 80, + 500, + 313 + ], + "spans": [ + { + "bbox": [ + 111, + 80, + 500, + 313 + ], + "score": 0.975, + "type": "image", + "image_path": "a677d9bfd60d90b339e6056f4eedc86e60ca0e0077525f722cfbef7db0088cd1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 80, + 500, + 157.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 157.66666666666669, + 500, + 235.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 235.33333333333337, + 500, + 313.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 120, + 323, + 487, + 335 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 322, + 488, + 336 + ], + "spans": [ + { + "bbox": [ + 123, + 322, + 488, + 336 + ], + "score": 1.0, + "content": "Figure 3: Color objective sample images. Samples are randomly drawn, not cherry-picked.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 108, + 357, + 504, + 379 + ], + "lines": [], + "index": 4.5, + "bbox_fs": [ + 105, + 356, + 505, + 380 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 247, + 404 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 248, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 248, + 405 + ], + "score": 1.0, + "content": "4.3 SUPPLEMENTAL ANALYSES", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "We also conducted several supplemental analyses which can be found in detail Appendix A. In", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "summary, the initial variability in the PIR of the images used to train the system is strongly correlated", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 437, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 506, + 448 + ], + "score": 1.0, + "content": "with the amount of improvement the system makes in the PIR, the system fairly consistently underes-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "timates the performance it achieves (because of a detail of training procedure, see the Appendix), and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "iterating the process of improving PIRs yields better results for objectives from a lower layer of VGG", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 468, + 173, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 173, + 482 + ], + "score": 1.0, + "content": "but not a higher.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 414, + 506, + 482 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 497, + 190, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 192, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 192, + 512 + ], + "score": 1.0, + "content": "5 DISCUSSION", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "Overall, our system appears to be relatively successful. It can optimize a generative model to produce", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "images which target a wide variety of objectives, ranging from low-level visual features such as", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 544, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 557 + ], + "score": 1.0, + "content": "colors and early features of VGG to features computed at the top layers of VGG. This success across", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "a wide variety of objective functions allows us to be somewhat confident that our system will be able", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 566, + 351, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 351, + 579 + ], + "score": 1.0, + "content": "to achieve success in optimizing for real human interactions.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 523, + 506, + 579 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 583, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "Furthermore, the system did not require an inordinate amount of training data. In fact, we were able", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "to successfully estimate many different objective functions from only 1000 images, several orders of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "magnitude fewer than is typically used to train CNNs for vision tasks. Furthermore, these images", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "came from a very biased and narrow distribution (samples from our generative model) which is", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "score": 1.0, + "content": "reflective of neither the images that were used to pre-train the Inception model in the PIR estimator,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 639, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 649 + ], + "score": 1.0, + "content": "nor the images the VGG model (which produced the simulated objectives) was trained on. Our", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "success from this small amount of data suggests that not only will our system be able to optimize for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 659, + 464, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 464, + 672 + ], + "score": 1.0, + "content": "real human interactions, it will be able to do so from a feasible number of training points.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 582, + 506, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "These results are exciting – the model is able to approximate apparently complex objective functions", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "from a small amount of data, even though this data comes from a very biased distribution that is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "unrelated to most the objectives in question. But what is really being learned? In the case of the color", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "images, it’s clear that the model is doing something close to correct. However, for the objectives", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "derived from VGG we have no way to really assess whether the model is making the images better or", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "just more adversarial. For instance, when we are optimizing for the logit for “magpie,” it’s almost", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 507, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 507, + 107 + ], + "score": 1.0, + "content": "certainly the case that the result of this optimization will not look more like a magpie to a human,", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "even if VGG does rate the images as more “magpie-like.” On the other hand, this is not necessarily a", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "failure of the system – it is accurately capturing the objective function it is given. What remains to be", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "seen is whether it can capture how background images influence human behavior as well as it can", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 303, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 303, + 150 + ], + "score": 1.0, + "content": "capture the vagaries of deep vision architectures.", + "type": "text", + "cross_page": true + } + ], + "index": 5 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 676, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 148 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "just more adversarial. For instance, when we are optimizing for the logit for “magpie,” it’s almost", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 507, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 507, + 107 + ], + "score": 1.0, + "content": "certainly the case that the result of this optimization will not look more like a magpie to a human,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "even if VGG does rate the images as more “magpie-like.” On the other hand, this is not necessarily a", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "failure of the system – it is accurately capturing the objective function it is given. What remains to be", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "seen is whether it can capture how background images influence human behavior as well as it can", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 303, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 303, + 150 + ], + "score": 1.0, + "content": "capture the vagaries of deep vision architectures.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 167 + ], + "score": 1.0, + "content": "We believe there are many domains where a system similar to ours could be useful. We mentioned", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "producing better webpage backgrounds and making more aesthetic images above, but there are many", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 507, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 507, + 190 + ], + "score": 1.0, + "content": "potential applications for improving GANs with a limited amount of human feedback. For example,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "a model could be trained to produce better music (e.g. song skip rates on streaming generated music", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 197, + 242, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 242, + 211 + ], + "score": 1.0, + "content": "could be treated as inverse PIRs).", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 108, + 224, + 290, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 224, + 292, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 292, + 237 + ], + "score": 1.0, + "content": "5.1 TRADING IMAGE DIVERSITY FOR PIR", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "When tuning the GAN, the decrease in the PIR loss is usually accompanied by an increase in the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "generator loss, and often by a partial collapse of the generator output (for example, the optimized", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 268, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 280 + ], + "score": 1.0, + "content": "images generally seem to have fewer output modes than the pre-training images in Fig. 3). This is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 279, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 506, + 290 + ], + "score": 1.0, + "content": "not especially surprising – because we weighted the PIR loss very highly, the model is rewarded for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "trading some image diversity for image optimality. Depending on the desired application, the weight", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "on the PIR loss could be adjusted as necessary to trade off between producing images close to the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "data distribution and optimizing PIR. At its most extreme, one could down-weight the generator loss", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "entirely, and train until the model just produces a single optimal image. However, the generator likely", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "score": 1.0, + "content": "provides some regularization by constraining the images to be somewhat close to the real images,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "score": 1.0, + "content": "which will reduce overfitting to an imperfect estimate of the PIR function. Furthermore, in many", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "score": 1.0, + "content": "settings we will want to generate a variety of images (e.g. backgrounds for different websites). For", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 366, + 401, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 401, + 379 + ], + "score": 1.0, + "content": "these reasons, we chose to keep the generator loss when tuning the GAN.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 108, + 392, + 222, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 391, + 225, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 225, + 405 + ], + "score": 1.0, + "content": "5.2 FUTURE DIRECTIONS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "There are a number of future directions suggested by this work. A number of possible improvements", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "are discussed in Appendix C.5. However, we also think this work has potential applications from the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 435, + 507, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 507, + 448 + ], + "score": 1.0, + "content": "perspective of distillation or imitation approaches, which attempt to train one network to emulate an-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "other (Hinton et al., 2015; Parisotto et al., 2015, e.g), as well as from the perspective of understanding", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "the computations that these vision architectures perform. As far as we are aware, these results are the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "first to show that a deep vision model can be tuned rapidly from relatively little data to produce outputs", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "which accurately emulate the behavior of hidden layers of another deep vision architecture trained on", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 491, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 502 + ], + "score": 1.0, + "content": "a different dataset. This suggests both that the inductive biases shared among these architectures are", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "causing them to find similar solutions (which is also supported by work on transferable adversarial", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "score": 1.0, + "content": "examples (Liu et al., 2016, e.g.)), and that these networks final layers represent the computations of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "earlier hidden layers in a way that is somewhat accessible. It’s possible that using our system with", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "objectives from CNN layers as we did here might help to understand the features those layers are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "attending to, by analyzing the distribution of images that are produced. In this sense, our system can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "be thought of as offering a new approach to multifaceted feature visualization (Nguyen et al., 2016),", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "because our system attempts to optimize a distribution of images for an objective and encourages", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 577, + 429, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 429, + 592 + ], + "score": 1.0, + "content": "diversity in the distribution produced, rather than just optimizing a single image.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 108, + 607, + 201, + 619 + ], + "lines": [ + { + "bbox": [ + 104, + 605, + 203, + 623 + ], + "spans": [ + { + "bbox": [ + 104, + 605, + 203, + 623 + ], + "score": 1.0, + "content": "6 CONCLUSIONS", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "We have described a system for efficiently tuning a generative image model according to a slow-to-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "evaluate objective function. We have demonstrated the success of this system at targeting a variety of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "objective functions simulated from different layers of a deep vision model, as well as from low-level", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "visual features of the images, and have shown that it can do so from a small amount of data. We", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "have quantified some of the features that affect its performance, including the variability of the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "training PIR data and the number of zeros it contains. Our system’s success on a wide variety of", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "objectives suggests that it will be able to improve real user interactions, or other objectives which", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 709, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 724 + ], + "score": 1.0, + "content": "are slow and expensive to evaluate. This may have many exciting applications, such as improving", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 721, + 273, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 273, + 734 + ], + "score": 1.0, + "content": "machine-generated images, music, or art.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 46 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 148 + ], + "lines": [], + "index": 2.5, + "bbox_fs": [ + 105, + 82, + 507, + 150 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 167 + ], + "score": 1.0, + "content": "We believe there are many domains where a system similar to ours could be useful. We mentioned", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "producing better webpage backgrounds and making more aesthetic images above, but there are many", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 507, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 507, + 190 + ], + "score": 1.0, + "content": "potential applications for improving GANs with a limited amount of human feedback. For example,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "a model could be trained to produce better music (e.g. song skip rates on streaming generated music", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 197, + 242, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 242, + 211 + ], + "score": 1.0, + "content": "could be treated as inverse PIRs).", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 153, + 507, + 211 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 224, + 290, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 224, + 292, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 292, + 237 + ], + "score": 1.0, + "content": "5.1 TRADING IMAGE DIVERSITY FOR PIR", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "When tuning the GAN, the decrease in the PIR loss is usually accompanied by an increase in the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "generator loss, and often by a partial collapse of the generator output (for example, the optimized", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 268, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 280 + ], + "score": 1.0, + "content": "images generally seem to have fewer output modes than the pre-training images in Fig. 3). This is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 279, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 506, + 290 + ], + "score": 1.0, + "content": "not especially surprising – because we weighted the PIR loss very highly, the model is rewarded for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "trading some image diversity for image optimality. Depending on the desired application, the weight", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "on the PIR loss could be adjusted as necessary to trade off between producing images close to the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "data distribution and optimizing PIR. At its most extreme, one could down-weight the generator loss", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "entirely, and train until the model just produces a single optimal image. However, the generator likely", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "score": 1.0, + "content": "provides some regularization by constraining the images to be somewhat close to the real images,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "score": 1.0, + "content": "which will reduce overfitting to an imperfect estimate of the PIR function. Furthermore, in many", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "score": 1.0, + "content": "settings we will want to generate a variety of images (e.g. backgrounds for different websites). For", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 366, + 401, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 401, + 379 + ], + "score": 1.0, + "content": "these reasons, we chose to keep the generator loss when tuning the GAN.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 245, + 507, + 379 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 392, + 222, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 391, + 225, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 225, + 405 + ], + "score": 1.0, + "content": "5.2 FUTURE DIRECTIONS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "There are a number of future directions suggested by this work. A number of possible improvements", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "are discussed in Appendix C.5. However, we also think this work has potential applications from the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 435, + 507, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 507, + 448 + ], + "score": 1.0, + "content": "perspective of distillation or imitation approaches, which attempt to train one network to emulate an-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "other (Hinton et al., 2015; Parisotto et al., 2015, e.g), as well as from the perspective of understanding", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "the computations that these vision architectures perform. As far as we are aware, these results are the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "first to show that a deep vision model can be tuned rapidly from relatively little data to produce outputs", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "which accurately emulate the behavior of hidden layers of another deep vision architecture trained on", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 491, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 502 + ], + "score": 1.0, + "content": "a different dataset. This suggests both that the inductive biases shared among these architectures are", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "causing them to find similar solutions (which is also supported by work on transferable adversarial", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "score": 1.0, + "content": "examples (Liu et al., 2016, e.g.)), and that these networks final layers represent the computations of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "earlier hidden layers in a way that is somewhat accessible. It’s possible that using our system with", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "objectives from CNN layers as we did here might help to understand the features those layers are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "attending to, by analyzing the distribution of images that are produced. In this sense, our system can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "be thought of as offering a new approach to multifaceted feature visualization (Nguyen et al., 2016),", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "because our system attempts to optimize a distribution of images for an objective and encourages", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 577, + 429, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 429, + 592 + ], + "score": 1.0, + "content": "diversity in the distribution produced, rather than just optimizing a single image.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 414, + 507, + 592 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 607, + 201, + 619 + ], + "lines": [ + { + "bbox": [ + 104, + 605, + 203, + 623 + ], + "spans": [ + { + "bbox": [ + 104, + 605, + 203, + 623 + ], + "score": 1.0, + "content": "6 CONCLUSIONS", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "We have described a system for efficiently tuning a generative image model according to a slow-to-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "evaluate objective function. We have demonstrated the success of this system at targeting a variety of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "objective functions simulated from different layers of a deep vision model, as well as from low-level", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "visual features of the images, and have shown that it can do so from a small amount of data. We", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "have quantified some of the features that affect its performance, including the variability of the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "training PIR data and the number of zeros it contains. Our system’s success on a wide variety of", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "objectives suggests that it will be able to improve real user interactions, or other objectives which", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 709, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 724 + ], + "score": 1.0, + "content": "are slow and expensive to evaluate. This may have many exciting applications, such as improving", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 721, + 273, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 273, + 734 + ], + "score": 1.0, + "content": "machine-generated images, music, or art.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 633, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 101, + 504, + 124 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 506, + 115 + ], + "score": 1.0, + "content": "Paul Christiano, Jan Leike, Tom B Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 112, + 361, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 112, + 361, + 124 + ], + "score": 1.0, + "content": "reinforcement learning from human preferences. arXi, 2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 108, + 133, + 504, + 167 + ], + "lines": [ + { + "bbox": [ + 107, + 134, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 107, + 134, + 505, + 146 + ], + "score": 1.0, + "content": "Ahmed Elgammal, Bingchen Liu, Mohamed Elhoseiny, and Marian Mazzone. CAN: Creative", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "Adversarial Networks Generating ”Art” by Learning About Styles and Deviating from Style", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 156, + 253, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 156, + 253, + 168 + ], + "score": 1.0, + "content": "Norms. arXiv, (Iccc):1–22, 2017.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 177, + 505, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 506, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 506, + 191 + ], + "score": 1.0, + "content": "Ij Goodfellow, J Pouget-Abadie, and Mehdi Mirza. Generative Adversarial Networks. arXiv, pp. 1–9,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 188, + 375, + 200 + ], + "spans": [ + { + "bbox": [ + 116, + 188, + 375, + 200 + ], + "score": 1.0, + "content": "2014. ISSN 10495258. doi: 10.1001/jamainternmed.2016.8245.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 104, + 210, + 506, + 234 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 506, + 223 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. arXiv,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 221, + 352, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 221, + 352, + 234 + ], + "score": 1.0, + "content": "pp. 1–9, 2015. ISSN 0022-2488. doi: 10.1063/1.4931082.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 506, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 506, + 256 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. Iclr, pp. 1–15,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 254, + 496, + 267 + ], + "spans": [ + { + "bbox": [ + 116, + 254, + 496, + 267 + ], + "score": 1.0, + "content": "2015. ISSN 09252312. doi: http://doi.acm.org.ezproxy.lib.ucf.edu/10.1145/1830483.1830503.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 107, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 107, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into Transferable Adversarial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 286, + 347, + 299 + ], + "spans": [ + { + "bbox": [ + 115, + 286, + 347, + 299 + ], + "score": 1.0, + "content": "Examples and Black-box Attacks. arXiv, (2):1–24, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 505, + 342 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep Neural Networks are Easily Fooled: High", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 319, + 507, + 333 + ], + "spans": [ + { + "bbox": [ + 115, + 319, + 507, + 333 + ], + "score": 1.0, + "content": "Confidence Predictions for Unrecognizable Images. arXiv, 2014. doi: 10.1109/CVPR.2015.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 330, + 156, + 343 + ], + "spans": [ + { + "bbox": [ + 115, + 330, + 156, + 343 + ], + "score": 1.0, + "content": "7298640.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 105, + 353, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "Anh Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted Feature Visualization: Uncovering the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 117, + 364, + 492, + 377 + ], + "spans": [ + { + "bbox": [ + 117, + 364, + 492, + 377 + ], + "score": 1.0, + "content": "Different Types of Features Learned By Each Neuron in Deep Neural Networks. arXiv, 2016.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 506, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 465, + 399 + ], + "score": 1.0, + "content": "Anh Nguyen, Jason Yosinski, Yoshua Bengio, Alexey Dosovitskiy, and Jeff Clune. Plug", + "type": "text" + }, + { + "bbox": [ + 465, + 386, + 483, + 398 + ], + "score": 0.57, + "content": "\\{ \\& \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 384, + 505, + 399 + ], + "score": 1.0, + "content": "Play", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 397, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 116, + 397, + 506, + 409 + ], + "score": 1.0, + "content": "Generative Networks: Conditional Iterative Generation of Images in Latent Space. ICCV, (3):33,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 407, + 267, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 267, + 419 + ], + "score": 1.0, + "content": "2017. doi: 10.1109/CVPR.2017.374.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 105, + 429, + 505, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 443 + ], + "score": 1.0, + "content": "Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional Image Synthesis With Auxiliary", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 440, + 353, + 453 + ], + "spans": [ + { + "bbox": [ + 116, + 440, + 353, + 453 + ], + "score": 1.0, + "content": "Classifier GANs. arXiv, pp. 1–14, 2016. ISSN 1938-7228.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 485 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Actor-Mimic: Deep Multitask and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 473, + 348, + 486 + ], + "spans": [ + { + "bbox": [ + 117, + 473, + 348, + 486 + ], + "score": 1.0, + "content": "Transfer Reinforcement Learning. arXiv, pp. 1–16, 2015.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 506, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 507, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 507, + 509 + ], + "score": 1.0, + "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 115, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C Berg, and Li Fei-Fei. ImageNet", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 517, + 507, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 517, + 507, + 531 + ], + "score": 1.0, + "content": "Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 527, + 391, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 391, + 541 + ], + "score": 1.0, + "content": "211–252, 2015. ISSN 15731405. doi: 10.1007/s11263-015-0816-y.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 548, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 564 + ], + "score": 1.0, + "content": "Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 561, + 473, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 561, + 473, + 574 + ], + "score": 1.0, + "content": "Recognition. arXiv, pp. 1–10, 2014. ISSN 09505849. doi: 10.1016/j.infsof.2008.09.005.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 506, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethink-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 115, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "ing the Inception Architecture for Computer Vision. Proceedings of the IEEE Computer Society", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 115, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2818–2826, 2016. ISSN", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 617, + 285, + 627 + ], + "spans": [ + { + "bbox": [ + 116, + 617, + 285, + 627 + ], + "score": 1.0, + "content": "08866236. doi: 10.1002/2014GB005021.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 506, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 506, + 649 + ], + "score": 1.0, + "content": "Daniel L K Yamins, Ha Hong, Charles F Cadieu, Ethan A Solomon, Darren Seibert, and James J", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 116, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "DiCarlo. Performance-optimized hierarchical models predict neural responses in higher visual", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 660, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 116, + 660, + 506, + 672 + ], + "score": 1.0, + "content": "cortex. Proceedings of the National Academy of Sciences of the United States of America, 111(23):", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 671, + 396, + 683 + ], + "spans": [ + { + "bbox": [ + 116, + 671, + 396, + 683 + ], + "score": 1.0, + "content": "8619–8624, 2014. ISSN 1091-6490. doi: 10.1073/pnas.1403112111.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 693, + 506, + 725 + ], + "lines": [ + { + "bbox": [ + 105, + 692, + 506, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 506, + 705 + ], + "score": 1.0, + "content": "Matthew D Zeiler and Rob Fergus. Visualizing and Understanding Convolutional Networks", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 704, + 507, + 716 + ], + "spans": [ + { + "bbox": [ + 116, + 704, + 507, + 716 + ], + "score": 1.0, + "content": "arXiv:1311.2901v3 [cs.CV] 28 Nov 2013. Computer Vision–ECCV 2014, 8689:818–833, 2014.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 714, + 377, + 726 + ], + "spans": [ + { + "bbox": [ + 116, + 714, + 377, + 726 + ], + "score": 1.0, + "content": "ISSN 978-3-319-10589-5. doi: 10.1007/978-3-319-10590-1 53.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 101, + 504, + 124 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 506, + 115 + ], + "score": 1.0, + "content": "Paul Christiano, Jan Leike, Tom B Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 112, + 361, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 112, + 361, + 124 + ], + "score": 1.0, + "content": "reinforcement learning from human preferences. arXi, 2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 99, + 506, + 124 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 133, + 504, + 167 + ], + "lines": [ + { + "bbox": [ + 107, + 134, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 107, + 134, + 505, + 146 + ], + "score": 1.0, + "content": "Ahmed Elgammal, Bingchen Liu, Mohamed Elhoseiny, and Marian Mazzone. CAN: Creative", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "Adversarial Networks Generating ”Art” by Learning About Styles and Deviating from Style", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 156, + 253, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 156, + 253, + 168 + ], + "score": 1.0, + "content": "Norms. arXiv, (Iccc):1–22, 2017.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 107, + 134, + 505, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 177, + 505, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 506, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 506, + 191 + ], + "score": 1.0, + "content": "Ij Goodfellow, J Pouget-Abadie, and Mehdi Mirza. Generative Adversarial Networks. arXiv, pp. 1–9,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 188, + 375, + 200 + ], + "spans": [ + { + "bbox": [ + 116, + 188, + 375, + 200 + ], + "score": 1.0, + "content": "2014. ISSN 10495258. doi: 10.1001/jamainternmed.2016.8245.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 177, + 506, + 200 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 210, + 506, + 234 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 506, + 223 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. arXiv,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 221, + 352, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 221, + 352, + 234 + ], + "score": 1.0, + "content": "pp. 1–9, 2015. ISSN 0022-2488. doi: 10.1063/1.4931082.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 210, + 506, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 506, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 506, + 256 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. Iclr, pp. 1–15,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 254, + 496, + 267 + ], + "spans": [ + { + "bbox": [ + 116, + 254, + 496, + 267 + ], + "score": 1.0, + "content": "2015. ISSN 09252312. doi: http://doi.acm.org.ezproxy.lib.ucf.edu/10.1145/1830483.1830503.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 243, + 506, + 267 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 107, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 107, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into Transferable Adversarial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 286, + 347, + 299 + ], + "spans": [ + { + "bbox": [ + 115, + 286, + 347, + 299 + ], + "score": 1.0, + "content": "Examples and Black-box Attacks. arXiv, (2):1–24, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 107, + 276, + 505, + 299 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 505, + 342 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep Neural Networks are Easily Fooled: High", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 319, + 507, + 333 + ], + "spans": [ + { + "bbox": [ + 115, + 319, + 507, + 333 + ], + "score": 1.0, + "content": "Confidence Predictions for Unrecognizable Images. arXiv, 2014. doi: 10.1109/CVPR.2015.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 330, + 156, + 343 + ], + "spans": [ + { + "bbox": [ + 115, + 330, + 156, + 343 + ], + "score": 1.0, + "content": "7298640.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 308, + 507, + 343 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 353, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "Anh Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted Feature Visualization: Uncovering the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 117, + 364, + 492, + 377 + ], + "spans": [ + { + "bbox": [ + 117, + 364, + 492, + 377 + ], + "score": 1.0, + "content": "Different Types of Features Learned By Each Neuron in Deep Neural Networks. arXiv, 2016.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 352, + 505, + 377 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 506, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 465, + 399 + ], + "score": 1.0, + "content": "Anh Nguyen, Jason Yosinski, Yoshua Bengio, Alexey Dosovitskiy, and Jeff Clune. Plug", + "type": "text" + }, + { + "bbox": [ + 465, + 386, + 483, + 398 + ], + "score": 0.57, + "content": "\\{ \\& \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 384, + 505, + 399 + ], + "score": 1.0, + "content": "Play", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 397, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 116, + 397, + 506, + 409 + ], + "score": 1.0, + "content": "Generative Networks: Conditional Iterative Generation of Images in Latent Space. ICCV, (3):33,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 407, + 267, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 267, + 419 + ], + "score": 1.0, + "content": "2017. doi: 10.1109/CVPR.2017.374.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 384, + 506, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 429, + 505, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 443 + ], + "score": 1.0, + "content": "Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional Image Synthesis With Auxiliary", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 440, + 353, + 453 + ], + "spans": [ + { + "bbox": [ + 116, + 440, + 353, + 453 + ], + "score": 1.0, + "content": "Classifier GANs. arXiv, pp. 1–14, 2016. ISSN 1938-7228.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 428, + 506, + 453 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 485 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Actor-Mimic: Deep Multitask and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 473, + 348, + 486 + ], + "spans": [ + { + "bbox": [ + 117, + 473, + 348, + 486 + ], + "score": 1.0, + "content": "Transfer Reinforcement Learning. arXiv, pp. 1–16, 2015.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 461, + 505, + 486 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 506, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 507, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 507, + 509 + ], + "score": 1.0, + "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 115, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C Berg, and Li Fei-Fei. ImageNet", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 517, + 507, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 517, + 507, + 531 + ], + "score": 1.0, + "content": "Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 527, + 391, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 391, + 541 + ], + "score": 1.0, + "content": "211–252, 2015. ISSN 15731405. doi: 10.1007/s11263-015-0816-y.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 495, + 507, + 541 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 548, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 564 + ], + "score": 1.0, + "content": "Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 561, + 473, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 561, + 473, + 574 + ], + "score": 1.0, + "content": "Recognition. arXiv, pp. 1–10, 2014. ISSN 09505849. doi: 10.1016/j.infsof.2008.09.005.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 548, + 505, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 506, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethink-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 115, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "ing the Inception Architecture for Computer Vision. Proceedings of the IEEE Computer Society", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 115, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2818–2826, 2016. ISSN", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 617, + 285, + 627 + ], + "spans": [ + { + "bbox": [ + 116, + 617, + 285, + 627 + ], + "score": 1.0, + "content": "08866236. doi: 10.1002/2014GB005021.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 583, + 506, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 506, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 506, + 649 + ], + "score": 1.0, + "content": "Daniel L K Yamins, Ha Hong, Charles F Cadieu, Ethan A Solomon, Darren Seibert, and James J", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 116, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "DiCarlo. Performance-optimized hierarchical models predict neural responses in higher visual", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 660, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 116, + 660, + 506, + 672 + ], + "score": 1.0, + "content": "cortex. Proceedings of the National Academy of Sciences of the United States of America, 111(23):", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 671, + 396, + 683 + ], + "spans": [ + { + "bbox": [ + 116, + 671, + 396, + 683 + ], + "score": 1.0, + "content": "8619–8624, 2014. ISSN 1091-6490. doi: 10.1073/pnas.1403112111.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 638, + 506, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 693, + 506, + 725 + ], + "lines": [ + { + "bbox": [ + 105, + 692, + 506, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 506, + 705 + ], + "score": 1.0, + "content": "Matthew D Zeiler and Rob Fergus. Visualizing and Understanding Convolutional Networks", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 704, + 507, + 716 + ], + "spans": [ + { + "bbox": [ + 116, + 704, + 507, + 716 + ], + "score": 1.0, + "content": "arXiv:1311.2901v3 [cs.CV] 28 Nov 2013. Computer Vision–ECCV 2014, 8689:818–833, 2014.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 714, + 377, + 726 + ], + "spans": [ + { + "bbox": [ + 116, + 714, + 377, + 726 + ], + "score": 1.0, + "content": "ISSN 978-3-319-10589-5. doi: 10.1007/978-3-319-10590-1 53.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 692, + 507, + 726 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 240, + 417, + 539 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 187, + 240, + 417, + 539 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 240, + 417, + 539 + ], + "spans": [ + { + "bbox": [ + 187, + 240, + 417, + 539 + ], + "score": 0.956, + "type": "image", + "image_path": "98240ce4438c7e3b31bfe707d1ab0bc9c597be9cf48d72834b5da56731a9b559.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 240, + 417, + 253.5909090909091 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 187, + 253.5909090909091, + 417, + 267.1818181818182 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 187, + 267.1818181818182, + 417, + 280.77272727272725 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 187, + 280.77272727272725, + 417, + 294.3636363636363 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 187, + 294.3636363636363, + 417, + 307.9545454545454 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 187, + 307.9545454545454, + 417, + 321.54545454545445 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 187, + 321.54545454545445, + 417, + 335.1363636363635 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 187, + 335.1363636363635, + 417, + 348.7272727272726 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 187, + 348.7272727272726, + 417, + 362.31818181818164 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 187, + 362.31818181818164, + 417, + 375.9090909090907 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 187, + 375.9090909090907, + 417, + 389.4999999999998 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 187, + 389.4999999999998, + 417, + 403.09090909090884 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 187, + 403.09090909090884, + 417, + 416.6818181818179 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 187, + 416.6818181818179, + 417, + 430.27272727272697 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 187, + 430.27272727272697, + 417, + 443.86363636363603 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 187, + 443.86363636363603, + 417, + 457.4545454545451 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 187, + 457.4545454545451, + 417, + 471.04545454545416 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 187, + 471.04545454545416, + 417, + 484.6363636363632 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 187, + 484.6363636363632, + 417, + 498.2272727272723 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 187, + 498.2272727272723, + 417, + 511.81818181818136 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 187, + 511.81818181818136, + 417, + 525.4090909090904 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 187, + 525.4090909090904, + 417, + 538.9999999999995 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 549, + 504, + 573 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "Figure 4: Change in mean PIR vs. estimated change in mean PIR. Points for which the change in", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 265, + 574 + ], + "score": 1.0, + "content": "mean PIR is not significant by a Welch’s", + "type": "text" + }, + { + "bbox": [ + 266, + 561, + 271, + 571 + ], + "score": 0.64, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 560, + 363, + 574 + ], + "score": 1.0, + "content": "-test with a threshold of", + "type": "text" + }, + { + "bbox": [ + 363, + 561, + 407, + 571 + ], + "score": 0.89, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "are partially transparent.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + } + ], + "index": 16.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 240, + 417, + 539 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 187, + 240, + 417, + 539 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 240, + 417, + 539 + ], + "spans": [ + { + "bbox": [ + 187, + 240, + 417, + 539 + ], + "score": 0.956, + "type": "image", + "image_path": "98240ce4438c7e3b31bfe707d1ab0bc9c597be9cf48d72834b5da56731a9b559.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 240, + 417, + 253.5909090909091 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 187, + 253.5909090909091, + 417, + 267.1818181818182 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 187, + 267.1818181818182, + 417, + 280.77272727272725 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 187, + 280.77272727272725, + 417, + 294.3636363636363 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 187, + 294.3636363636363, + 417, + 307.9545454545454 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 187, + 307.9545454545454, + 417, + 321.54545454545445 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 187, + 321.54545454545445, + 417, + 335.1363636363635 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 187, + 335.1363636363635, + 417, + 348.7272727272726 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 187, + 348.7272727272726, + 417, + 362.31818181818164 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 187, + 362.31818181818164, + 417, + 375.9090909090907 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 187, + 375.9090909090907, + 417, + 389.4999999999998 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 187, + 389.4999999999998, + 417, + 403.09090909090884 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 187, + 403.09090909090884, + 417, + 416.6818181818179 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 187, + 416.6818181818179, + 417, + 430.27272727272697 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 187, + 430.27272727272697, + 417, + 443.86363636363603 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 187, + 443.86363636363603, + 417, + 457.4545454545451 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 187, + 457.4545454545451, + 417, + 471.04545454545416 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 187, + 471.04545454545416, + 417, + 484.6363636363632 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 187, + 484.6363636363632, + 417, + 498.2272727272723 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 187, + 498.2272727272723, + 417, + 511.81818181818136 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 187, + 511.81818181818136, + 417, + 525.4090909090904 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 187, + 525.4090909090904, + 417, + 538.9999999999995 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 549, + 504, + 573 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "Figure 4: Change in mean PIR vs. estimated change in mean PIR. Points for which the change in", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 265, + 574 + ], + "score": 1.0, + "content": "mean PIR is not significant by a Welch’s", + "type": "text" + }, + { + "bbox": [ + 266, + 561, + 271, + 571 + ], + "score": 0.64, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 560, + 363, + 574 + ], + "score": 1.0, + "content": "-test with a threshold of", + "type": "text" + }, + { + "bbox": [ + 363, + 561, + 407, + 571 + ], + "score": 0.89, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "are partially transparent.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + } + ], + "index": 16.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 158, + 82, + 439, + 312 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 158, + 82, + 439, + 312 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 158, + 82, + 439, + 312 + ], + "spans": [ + { + "bbox": [ + 158, + 82, + 439, + 312 + ], + "score": 0.965, + "type": "image", + "image_path": "4b39472a7f5980f4b2ddf37b5e22ac2583f56e4f0262421e27809b90bbe04416.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 158, + 82, + 439, + 95.52941176470588 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 158, + 95.52941176470588, + 439, + 109.05882352941177 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 158, + 109.05882352941177, + 439, + 122.58823529411765 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 158, + 122.58823529411765, + 439, + 136.11764705882354 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 158, + 136.11764705882354, + 439, + 149.64705882352942 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 158, + 149.64705882352942, + 439, + 163.1764705882353 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 158, + 163.1764705882353, + 439, + 176.7058823529412 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 158, + 176.7058823529412, + 439, + 190.23529411764707 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 158, + 190.23529411764707, + 439, + 203.76470588235296 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 158, + 203.76470588235296, + 439, + 217.29411764705884 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 158, + 217.29411764705884, + 439, + 230.82352941176472 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 158, + 230.82352941176472, + 439, + 244.3529411764706 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 158, + 244.3529411764706, + 439, + 257.88235294117646 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 158, + 257.88235294117646, + 439, + 271.4117647058823 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 158, + 271.4117647058823, + 439, + 284.9411764705882 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 158, + 284.9411764705882, + 439, + 298.47058823529403 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 158, + 298.47058823529403, + 439, + 311.9999999999999 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 324, + 505, + 346 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "Figure 5: Change in mean PIR vs. initial variability. Points for which the change in mean PIR is not", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 334, + 450, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 204, + 348 + ], + "score": 1.0, + "content": "significant by a Welch’s", + "type": "text" + }, + { + "bbox": [ + 204, + 335, + 209, + 344 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 334, + 305, + 348 + ], + "score": 1.0, + "content": "-test with a threshold of", + "type": "text" + }, + { + "bbox": [ + 305, + 335, + 349, + 345 + ], + "score": 0.9, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 334, + 450, + 348 + ], + "score": 1.0, + "content": "are partially transparent.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + } + ], + "index": 12.75 + }, + { + "type": "title", + "bbox": [ + 109, + 368, + 221, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 223, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 223, + 382 + ], + "score": 1.0, + "content": "A OTHER ANALYSES", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 203, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 204, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 204, + 406 + ], + "score": 1.0, + "content": "A.1 INTROSPECTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 143, + 429 + ], + "score": 1.0, + "content": "Because", + "type": "text" + }, + { + "bbox": [ + 143, + 415, + 168, + 426 + ], + "score": 0.88, + "content": "L _ { P I R }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 413, + 381, + 429 + ], + "score": 1.0, + "content": "is just the expected value of the PIR, by looking at", + "type": "text" + }, + { + "bbox": [ + 382, + 415, + 407, + 426 + ], + "score": 0.9, + "content": "L _ { P I R }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 413, + 506, + 429 + ], + "score": 1.0, + "content": "before and after tuning", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 425, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 439 + ], + "score": 1.0, + "content": "the generative model, we can tell how well the system thinks it is doing, i.e. how much it estimates", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "score": 1.0, + "content": "that it improved PIR. This comparison reveals the interesting pattern that the system is overly", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "score": 1.0, + "content": "pessimistic about its performance. In fact, it tends to underestimate its performance by a factor of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 457, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 169, + 472 + ], + "score": 1.0, + "content": "more than 1.5", + "type": "text" + }, + { + "bbox": [ + 169, + 459, + 209, + 470 + ], + "score": 0.86, + "content": "\\beta = 1 . 6 7", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 457, + 506, + 472 + ], + "score": 1.0, + "content": "when regressing change in mean PIR on predicted change in mean PIR,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 468, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 483 + ], + "score": 1.0, + "content": "see Fig. 4). However, it does so fairly consistently. This effect appears to be driven by the system", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 480, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 494 + ], + "score": 1.0, + "content": "consistently underestimating the (absolute) PIRs, which is probably caused by our change in the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "softmax temperature between training the PIR estimator and tuning the generative model (which we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 503, + 342, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 342, + 515 + ], + "score": 1.0, + "content": "empirically found improves performance, as noted above).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 506, + 532 + ], + "score": 1.0, + "content": "This is in contrast to the possible a priori expectation that the model would systematically over-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 531, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 506, + 542 + ], + "score": 1.0, + "content": "estimate its performance, because it is overfitting to an imperfectly estimated objective function.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 541, + 507, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 507, + 554 + ], + "score": 1.0, + "content": "Although decreasing the softmax temperature between training and using the PIR obscures this effect,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "score": 1.0, + "content": "we do see some evidence of this; the more complex objectives (which the system produced lower", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "score": 1.0, + "content": "effect sizes on) seem to both have lower estimated changes in mean PIR and true changes in PIR", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 574, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 506, + 586 + ], + "score": 1.0, + "content": "which are even lower than the estimated ones (see Fig. 4). Thus although the system is somewhat", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 585, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 598 + ], + "score": 1.0, + "content": "aware of its reduced effectiveness with these objectives (as evidenced by the lower estimates of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "change in mean PIR), it is not reducing its estimates sufficiently to account for the true difficulty of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "the objectives (as evidenced by the fact that the true change in PIR is even lower than the estimates).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 618, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 506, + 631 + ], + "score": 1.0, + "content": "However, the system was generally still able to obtain positive results on these objectives (see Fig. 2).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 108, + 644, + 225, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 226, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 226, + 656 + ], + "score": 1.0, + "content": "A.2 INITIAL VARIABILITY", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "There is a general trend (see Fig. 5) that the variability in PIR in the initial dataset is strongly positively", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 675, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 356, + 690 + ], + "score": 1.0, + "content": "related with the change in PIR the system is able to produce", + "type": "text" + }, + { + "bbox": [ + 357, + 677, + 395, + 688 + ], + "score": 0.82, + "content": "\\beta = 0 . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 675, + 399, + 690 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 399, + 677, + 435, + 688 + ], + "score": 0.74, + "content": "t = 1 0 . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 675, + 439, + 690 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 439, + 676, + 483, + 688 + ], + "score": 0.79, + "content": "p \\bar { < } \\bar { 1 0 } ^ { - 1 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 675, + 507, + 690 + ], + "score": 1.0, + "content": ", in a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "linear model controlling for initial mean PIR and initial percent of values that are zero). In fact, initial", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 239, + 711 + ], + "score": 1.0, + "content": "standard deviation explains about", + "type": "text" + }, + { + "bbox": [ + 239, + 699, + 259, + 709 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "of the variance in the change in mean PIR. This is perhaps not", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "too surprising – more variability means that the generative model has capacity to produce higher PIR", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "images without too much tweaking, and that the PIR estimator model gets a wider range of values to", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 158, + 82, + 439, + 312 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 158, + 82, + 439, + 312 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 158, + 82, + 439, + 312 + ], + "spans": [ + { + "bbox": [ + 158, + 82, + 439, + 312 + ], + "score": 0.965, + "type": "image", + "image_path": "4b39472a7f5980f4b2ddf37b5e22ac2583f56e4f0262421e27809b90bbe04416.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 158, + 82, + 439, + 95.52941176470588 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 158, + 95.52941176470588, + 439, + 109.05882352941177 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 158, + 109.05882352941177, + 439, + 122.58823529411765 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 158, + 122.58823529411765, + 439, + 136.11764705882354 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 158, + 136.11764705882354, + 439, + 149.64705882352942 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 158, + 149.64705882352942, + 439, + 163.1764705882353 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 158, + 163.1764705882353, + 439, + 176.7058823529412 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 158, + 176.7058823529412, + 439, + 190.23529411764707 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 158, + 190.23529411764707, + 439, + 203.76470588235296 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 158, + 203.76470588235296, + 439, + 217.29411764705884 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 158, + 217.29411764705884, + 439, + 230.82352941176472 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 158, + 230.82352941176472, + 439, + 244.3529411764706 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 158, + 244.3529411764706, + 439, + 257.88235294117646 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 158, + 257.88235294117646, + 439, + 271.4117647058823 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 158, + 271.4117647058823, + 439, + 284.9411764705882 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 158, + 284.9411764705882, + 439, + 298.47058823529403 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 158, + 298.47058823529403, + 439, + 311.9999999999999 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 324, + 505, + 346 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "Figure 5: Change in mean PIR vs. initial variability. Points for which the change in mean PIR is not", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 334, + 450, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 204, + 348 + ], + "score": 1.0, + "content": "significant by a Welch’s", + "type": "text" + }, + { + "bbox": [ + 204, + 335, + 209, + 344 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 334, + 305, + 348 + ], + "score": 1.0, + "content": "-test with a threshold of", + "type": "text" + }, + { + "bbox": [ + 305, + 335, + 349, + 345 + ], + "score": 0.9, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 334, + 450, + 348 + ], + "score": 1.0, + "content": "are partially transparent.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + } + ], + "index": 12.75 + }, + { + "type": "title", + "bbox": [ + 109, + 368, + 221, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 223, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 223, + 382 + ], + "score": 1.0, + "content": "A OTHER ANALYSES", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 203, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 204, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 204, + 406 + ], + "score": 1.0, + "content": "A.1 INTROSPECTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 143, + 429 + ], + "score": 1.0, + "content": "Because", + "type": "text" + }, + { + "bbox": [ + 143, + 415, + 168, + 426 + ], + "score": 0.88, + "content": "L _ { P I R }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 413, + 381, + 429 + ], + "score": 1.0, + "content": "is just the expected value of the PIR, by looking at", + "type": "text" + }, + { + "bbox": [ + 382, + 415, + 407, + 426 + ], + "score": 0.9, + "content": "L _ { P I R }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 413, + 506, + 429 + ], + "score": 1.0, + "content": "before and after tuning", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 425, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 439 + ], + "score": 1.0, + "content": "the generative model, we can tell how well the system thinks it is doing, i.e. how much it estimates", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "score": 1.0, + "content": "that it improved PIR. This comparison reveals the interesting pattern that the system is overly", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "score": 1.0, + "content": "pessimistic about its performance. In fact, it tends to underestimate its performance by a factor of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 457, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 169, + 472 + ], + "score": 1.0, + "content": "more than 1.5", + "type": "text" + }, + { + "bbox": [ + 169, + 459, + 209, + 470 + ], + "score": 0.86, + "content": "\\beta = 1 . 6 7", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 457, + 506, + 472 + ], + "score": 1.0, + "content": "when regressing change in mean PIR on predicted change in mean PIR,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 468, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 483 + ], + "score": 1.0, + "content": "see Fig. 4). However, it does so fairly consistently. This effect appears to be driven by the system", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 480, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 494 + ], + "score": 1.0, + "content": "consistently underestimating the (absolute) PIRs, which is probably caused by our change in the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "softmax temperature between training the PIR estimator and tuning the generative model (which we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 503, + 342, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 342, + 515 + ], + "score": 1.0, + "content": "empirically found improves performance, as noted above).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 413, + 506, + 515 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 506, + 532 + ], + "score": 1.0, + "content": "This is in contrast to the possible a priori expectation that the model would systematically over-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 531, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 506, + 542 + ], + "score": 1.0, + "content": "estimate its performance, because it is overfitting to an imperfectly estimated objective function.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 541, + 507, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 507, + 554 + ], + "score": 1.0, + "content": "Although decreasing the softmax temperature between training and using the PIR obscures this effect,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "score": 1.0, + "content": "we do see some evidence of this; the more complex objectives (which the system produced lower", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "score": 1.0, + "content": "effect sizes on) seem to both have lower estimated changes in mean PIR and true changes in PIR", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 574, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 506, + 586 + ], + "score": 1.0, + "content": "which are even lower than the estimated ones (see Fig. 4). Thus although the system is somewhat", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 585, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 598 + ], + "score": 1.0, + "content": "aware of its reduced effectiveness with these objectives (as evidenced by the lower estimates of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "change in mean PIR), it is not reducing its estimates sufficiently to account for the true difficulty of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "the objectives (as evidenced by the fact that the true change in PIR is even lower than the estimates).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 618, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 506, + 631 + ], + "score": 1.0, + "content": "However, the system was generally still able to obtain positive results on these objectives (see Fig. 2).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 519, + 507, + 631 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 644, + 225, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 226, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 226, + 656 + ], + "score": 1.0, + "content": "A.2 INITIAL VARIABILITY", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "There is a general trend (see Fig. 5) that the variability in PIR in the initial dataset is strongly positively", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 675, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 356, + 690 + ], + "score": 1.0, + "content": "related with the change in PIR the system is able to produce", + "type": "text" + }, + { + "bbox": [ + 357, + 677, + 395, + 688 + ], + "score": 0.82, + "content": "\\beta = 0 . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 675, + 399, + 690 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 399, + 677, + 435, + 688 + ], + "score": 0.74, + "content": "t = 1 0 . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 675, + 439, + 690 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 439, + 676, + 483, + 688 + ], + "score": 0.79, + "content": "p \\bar { < } \\bar { 1 0 } ^ { - 1 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 675, + 507, + 690 + ], + "score": 1.0, + "content": ", in a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "linear model controlling for initial mean PIR and initial percent of values that are zero). In fact, initial", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 239, + 711 + ], + "score": 1.0, + "content": "standard deviation explains about", + "type": "text" + }, + { + "bbox": [ + 239, + 699, + 259, + 709 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "of the variance in the change in mean PIR. This is perhaps not", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "too surprising – more variability means that the generative model has capacity to produce higher PIR", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "images without too much tweaking, and that the PIR estimator model gets a wider range of values to", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "score": 1.0, + "content": "learn from. Still, when attempting to use this system in practice, it is important to keep in mind that", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 420, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 431 + ], + "score": 1.0, + "content": "starting with a sufficiently expressive generative model will more likely produce better results than", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 430, + 250, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 250, + 442 + ], + "score": 1.0, + "content": "starting with a more limited model.", + "type": "text", + "cross_page": true + } + ], + "index": 21 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 666, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 82, + 420, + 316 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 187, + 82, + 420, + 316 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 82, + 420, + 316 + ], + "spans": [ + { + "bbox": [ + 187, + 82, + 420, + 316 + ], + "score": 0.971, + "type": "image", + "image_path": "881a640f7021a2d81e1a10fbb7a3eae014fc91ad5806b8582eceaadd61ac6fd7.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 82, + 420, + 95.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 187, + 95.0, + 420, + 108.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 187, + 108.0, + 420, + 121.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 187, + 121.0, + 420, + 134.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 187, + 134.0, + 420, + 147.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 187, + 147.0, + 420, + 160.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 187, + 160.0, + 420, + 173.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 187, + 173.0, + 420, + 186.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 187, + 186.0, + 420, + 199.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 187, + 199.0, + 420, + 212.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 187, + 212.0, + 420, + 225.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 187, + 225.0, + 420, + 238.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 187, + 238.0, + 420, + 251.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 187, + 251.0, + 420, + 264.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 187, + 264.0, + 420, + 277.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 187, + 277.0, + 420, + 290.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 187, + 290.0, + 420, + 303.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 187, + 303.0, + 420, + 316.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 205, + 327, + 406, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 205, + 327, + 405, + 340 + ], + "spans": [ + { + "bbox": [ + 205, + 327, + 405, + 340 + ], + "score": 1.0, + "content": "Figure 6: Effect size evolution over two iterations", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + } + ], + "index": 13.25 + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "score": 1.0, + "content": "learn from. Still, when attempting to use this system in practice, it is important to keep in mind that", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 420, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 431 + ], + "score": 1.0, + "content": "starting with a sufficiently expressive generative model will more likely produce better results than", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 430, + 250, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 250, + 442 + ], + "score": 1.0, + "content": "starting with a more limited model.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 107, + 508, + 180, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 507, + 182, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 182, + 520 + ], + "score": 1.0, + "content": "A.3 ITERATION", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 549, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 506, + 563 + ], + "score": 1.0, + "content": "Given that our model improves PIRs, an obvious question is whether we can iterate the process.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Once we have increased PIRs, can we train a new PIR estimator on samples from our new generative", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "model, and use that to increase PIRs again? If we could iterate this many times, we might be able", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "to create much larger improvements in PIR than we can on a single step. On the other hand, it is", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "possible that after a single step of optimization we will effectively have saturated the easily achievable", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 605, + 435, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 435, + 617 + ], + "score": 1.0, + "content": "improvement in the model, and further steps will not result in much improvement.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "To evaluate this, we took the subset of models trained on single filters from VGG layers pool2 and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "fc8, and used the set of images generated from the post-tuning model along with the original set of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "score": 1.0, + "content": "images to tune their PIR estimators for another 25,000 steps, and then tuned the generative model", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "using this updated PIR estimator for another 50,000 steps. We then evaluated them as before, see Fig.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "6 for the results. The second iteration results were mixed, while the pool2 models all improved from", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "the first step to the second, none of them improved as much as they had on the first step, and many of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "the fc8 models actually performed worse after the second step. However, it is possible that further", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "hyperparameter tuning could improve these results, and it is certainly the case that running multiple", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "steps of iteration and selecting the best model by experimentation could yield better results, so this is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 721, + 217, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 217, + 733 + ], + "score": 1.0, + "content": "worth investigating further.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 82, + 420, + 316 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 187, + 82, + 420, + 316 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 82, + 420, + 316 + ], + "spans": [ + { + "bbox": [ + 187, + 82, + 420, + 316 + ], + "score": 0.971, + "type": "image", + "image_path": "881a640f7021a2d81e1a10fbb7a3eae014fc91ad5806b8582eceaadd61ac6fd7.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 82, + 420, + 95.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 187, + 95.0, + 420, + 108.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 187, + 108.0, + 420, + 121.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 187, + 121.0, + 420, + 134.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 187, + 134.0, + 420, + 147.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 187, + 147.0, + 420, + 160.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 187, + 160.0, + 420, + 173.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 187, + 173.0, + 420, + 186.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 187, + 186.0, + 420, + 199.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 187, + 199.0, + 420, + 212.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 187, + 212.0, + 420, + 225.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 187, + 225.0, + 420, + 238.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 187, + 238.0, + 420, + 251.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 187, + 251.0, + 420, + 264.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 187, + 264.0, + 420, + 277.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 187, + 277.0, + 420, + 290.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 187, + 290.0, + 420, + 303.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 187, + 303.0, + 420, + 316.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 205, + 327, + 406, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 205, + 327, + 405, + 340 + ], + "spans": [ + { + "bbox": [ + 205, + 327, + 405, + 340 + ], + "score": 1.0, + "content": "Figure 6: Effect size evolution over two iterations", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + } + ], + "index": 13.25 + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 442 + ], + "lines": [], + "index": 20, + "bbox_fs": [ + 105, + 408, + 506, + 442 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 508, + 180, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 507, + 182, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 182, + 520 + ], + "score": 1.0, + "content": "A.3 ITERATION", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 549, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 506, + 563 + ], + "score": 1.0, + "content": "Given that our model improves PIRs, an obvious question is whether we can iterate the process.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Once we have increased PIRs, can we train a new PIR estimator on samples from our new generative", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "model, and use that to increase PIRs again? If we could iterate this many times, we might be able", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "to create much larger improvements in PIR than we can on a single step. On the other hand, it is", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "possible that after a single step of optimization we will effectively have saturated the easily achievable", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 605, + 435, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 435, + 617 + ], + "score": 1.0, + "content": "improvement in the model, and further steps will not result in much improvement.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 549, + 506, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "To evaluate this, we took the subset of models trained on single filters from VGG layers pool2 and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "fc8, and used the set of images generated from the post-tuning model along with the original set of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "score": 1.0, + "content": "images to tune their PIR estimators for another 25,000 steps, and then tuned the generative model", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "using this updated PIR estimator for another 50,000 steps. We then evaluated them as before, see Fig.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "6 for the results. The second iteration results were mixed, while the pool2 models all improved from", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "the first step to the second, none of them improved as much as they had on the first step, and many of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "the fc8 models actually performed worse after the second step. However, it is possible that further", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "hyperparameter tuning could improve these results, and it is certainly the case that running multiple", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "steps of iteration and selecting the best model by experimentation could yield better results, so this is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 721, + 217, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 217, + 733 + ], + "score": 1.0, + "content": "worth investigating further.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 621, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 113, + 502, + 381 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 113, + 502, + 381 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 113, + 502, + 381 + ], + "spans": [ + { + "bbox": [ + 108, + 113, + 502, + 381 + ], + "score": 0.972, + "type": "image", + "image_path": "bb5568a77c88b87dcfddadac91fed93fcab073ec0eaf412bb76545bbf1d6be46.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 113, + 502, + 202.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 202.33333333333331, + 502, + 291.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 291.66666666666663, + 502, + 380.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 391, + 506, + 447 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 392, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 506, + 404 + ], + "score": 1.0, + "content": "Figure 7: Change in mean PIR on a variety of tasks. Each panel represents a set of tasks (such as", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "single filters from VGG), each color/column represents a subset (such as a single layer within VGG),", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "and each point represents one objective (such as a single filter from that layer) for which we optimized", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 482, + 438 + ], + "score": 1.0, + "content": "the generative model. Points for which the change in mean PIR is not significant by a Welch’s", + "type": "text" + }, + { + "bbox": [ + 482, + 426, + 487, + 434 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 423, + 506, + 438 + ], + "score": 1.0, + "content": "-test", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 435, + 327, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 183, + 448 + ], + "score": 1.0, + "content": "with a threshold of", + "type": "text" + }, + { + "bbox": [ + 184, + 435, + 227, + 446 + ], + "score": 0.89, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 435, + 327, + 448 + ], + "score": 1.0, + "content": "are partially transparent.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 107, + 469, + 235, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 468, + 236, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 236, + 483 + ], + "score": 1.0, + "content": "C DETAILED METHODS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 107, + 493, + 171, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 172, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 172, + 507 + ], + "score": 1.0, + "content": "C.1 ACGAN", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "score": 1.0, + "content": "We describe our results using the standard GAN framework for clarity, but we actually used an", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "ACGAN (Odena et al., 2016), which allows for conditioning for various user-specific features. This", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 536, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 284, + 548 + ], + "score": 1.0, + "content": "requires the following adjustments. Defining", + "type": "text" + }, + { + "bbox": [ + 284, + 537, + 291, + 546 + ], + "score": 0.79, + "content": "\\mathcal { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 536, + 432, + 548 + ], + "score": 1.0, + "content": "to be the set of possible classes and", + "type": "text" + }, + { + "bbox": [ + 433, + 537, + 444, + 547 + ], + "score": 0.88, + "content": "\\mathcal { T } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 536, + 506, + 548 + ], + "score": 1.0, + "content": "to be the set of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 547, + 281, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 253, + 559 + ], + "score": 1.0, + "content": "real images corresponding to a class", + "type": "text" + }, + { + "bbox": [ + 253, + 547, + 276, + 557 + ], + "score": 0.88, + "content": "c \\in { \\mathcal { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 547, + 281, + 559 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 153, + 562, + 458, + 637 + ], + "lines": [ + { + "bbox": [ + 153, + 562, + 458, + 637 + ], + "spans": [ + { + "bbox": [ + 153, + 562, + 458, + 637 + ], + "score": 0.94, + "content": "\\begin{array} { r l r } { L _ { \\mathrm { f a k e i m a g e } } = } & { } & { E _ { z \\sim \\mathcal { N } ( 0 , I ) , c \\sim \\mathcal { C } } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( G ( z , c ) ) = \\mathrm { f a k e } ) \\right] } \\\\ { L _ { \\mathrm { f a k e i m a g e ~ f o o l s } } = } & { } & { E _ { z \\sim \\mathcal { N } ( 0 , I ) , c \\sim \\mathcal { C } } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( G ( z , c ) ) = \\mathrm { r e a l } ) \\right] } \\\\ { L _ { \\mathrm { r e a l ~ i m a g e } } = } & { } & { E _ { c \\sim \\mathcal { C } , i \\sim \\mathcal { L } _ { c } } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( i ) = \\mathrm { r e a l } ) \\right] } \\\\ { L _ { \\mathrm { f a k e i m a g e c l a s s } } = } & { } & { E _ { z \\sim \\mathcal { N } ( 0 , I ) , c \\sim \\mathcal { C } } \\left[ \\log P ( D _ { \\mathrm { c l a s s } } ( G ( z , c ) ) = c ) \\right] } \\\\ { L _ { \\mathrm { r e a l ~ i m a g e c l a s s } } = } & { } & { E _ { c \\sim \\mathcal { C } , i \\sim \\mathcal { Z } _ { c } } \\left[ \\log P ( D _ { \\mathrm { c l a s s } } ( i ) = c ) \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "46d61d28718b47f1447cdcde2244b97da3d2efab9461ec27d33671ddc7176ccd.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 153, + 562, + 458, + 587.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 153, + 587.0, + 458, + 612.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 153, + 612.0, + 458, + 637.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 638, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 416, + 653 + ], + "score": 1.0, + "content": "Then the discriminator and generator losses are modified as follows (letting", + "type": "text" + }, + { + "bbox": [ + 416, + 641, + 486, + 652 + ], + "score": 0.89, + "content": "w _ { f a k e c l a s s } = 1 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 638, + 506, + 653 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 651, + 181, + 663 + ], + "spans": [ + { + "bbox": [ + 107, + 651, + 176, + 663 + ], + "score": 0.87, + "content": "w _ { r e a l c l a s s } = 1 . 0 \\AA", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 651, + 181, + 663 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 666, + 449, + 695 + ], + "lines": [ + { + "bbox": [ + 162, + 666, + 449, + 695 + ], + "spans": [ + { + "bbox": [ + 162, + 666, + 449, + 695 + ], + "score": 0.88, + "content": "\\begin{array} { r l r } { L _ { \\mathrm { d i s c r i m i n a t o r } } = } & { { } } & { L _ { \\mathrm { f a k e ~ i m a g e } } + L _ { \\mathrm { r e a l ~ i m a g e } } + w _ { r e a l c l a s s } L _ { \\mathrm { r e a l ~ c l a s s } } } \\\\ { L _ { \\mathrm { G e n e r a t o r } } = } & { { } } & { L _ { \\mathrm { f a k e ~ i m a g e ~ f o o l s } } + w _ { f a k e c l a s s } L _ { \\mathrm { f a k e ~ c l a s s } } } \\end{array}", + "type": "interline_equation", + "image_path": "acc8f92bc4ac62832f05e8e89b7f6e8675ca13042430cbc8dea8acace320e3b8.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 162, + 666, + 449, + 675.6666666666666 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 162, + 675.6666666666666, + 449, + 685.3333333333333 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 162, + 685.3333333333333, + 449, + 694.9999999999999 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Note that unlike the standard ACGAN formulation given in (Odena et al., 2016), we do not include", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 107, + 710, + 142, + 721 + ], + "score": 0.89, + "content": "L _ { \\mathrm { f a k e \\ c l a s s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "in the discriminator’s loss, to keep the discriminator from cooperating with the generator on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 721, + 196, + 731 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 196, + 731 + ], + "score": 1.0, + "content": "the classification task.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 82, + 259, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 260, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 260, + 96 + ], + "score": 1.0, + "content": "B SUPPLEMENTAL FIGURES", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 113, + 502, + 381 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 113, + 502, + 381 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 113, + 502, + 381 + ], + "spans": [ + { + "bbox": [ + 108, + 113, + 502, + 381 + ], + "score": 0.972, + "type": "image", + "image_path": "bb5568a77c88b87dcfddadac91fed93fcab073ec0eaf412bb76545bbf1d6be46.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 113, + 502, + 202.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 202.33333333333331, + 502, + 291.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 291.66666666666663, + 502, + 380.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 391, + 506, + 447 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 392, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 506, + 404 + ], + "score": 1.0, + "content": "Figure 7: Change in mean PIR on a variety of tasks. Each panel represents a set of tasks (such as", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "single filters from VGG), each color/column represents a subset (such as a single layer within VGG),", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "and each point represents one objective (such as a single filter from that layer) for which we optimized", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 482, + 438 + ], + "score": 1.0, + "content": "the generative model. Points for which the change in mean PIR is not significant by a Welch’s", + "type": "text" + }, + { + "bbox": [ + 482, + 426, + 487, + 434 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 423, + 506, + 438 + ], + "score": 1.0, + "content": "-test", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 435, + 327, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 183, + 448 + ], + "score": 1.0, + "content": "with a threshold of", + "type": "text" + }, + { + "bbox": [ + 184, + 435, + 227, + 446 + ], + "score": 0.89, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 435, + 327, + 448 + ], + "score": 1.0, + "content": "are partially transparent.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 107, + 469, + 235, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 468, + 236, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 236, + 483 + ], + "score": 1.0, + "content": "C DETAILED METHODS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 107, + 493, + 171, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 172, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 172, + 507 + ], + "score": 1.0, + "content": "C.1 ACGAN", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "score": 1.0, + "content": "We describe our results using the standard GAN framework for clarity, but we actually used an", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "ACGAN (Odena et al., 2016), which allows for conditioning for various user-specific features. This", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 536, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 284, + 548 + ], + "score": 1.0, + "content": "requires the following adjustments. Defining", + "type": "text" + }, + { + "bbox": [ + 284, + 537, + 291, + 546 + ], + "score": 0.79, + "content": "\\mathcal { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 536, + 432, + 548 + ], + "score": 1.0, + "content": "to be the set of possible classes and", + "type": "text" + }, + { + "bbox": [ + 433, + 537, + 444, + 547 + ], + "score": 0.88, + "content": "\\mathcal { T } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 536, + 506, + 548 + ], + "score": 1.0, + "content": "to be the set of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 547, + 281, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 253, + 559 + ], + "score": 1.0, + "content": "real images corresponding to a class", + "type": "text" + }, + { + "bbox": [ + 253, + 547, + 276, + 557 + ], + "score": 0.88, + "content": "c \\in { \\mathcal { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 547, + 281, + 559 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 514, + 506, + 559 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 153, + 562, + 458, + 637 + ], + "lines": [ + { + "bbox": [ + 153, + 562, + 458, + 637 + ], + "spans": [ + { + "bbox": [ + 153, + 562, + 458, + 637 + ], + "score": 0.94, + "content": "\\begin{array} { r l r } { L _ { \\mathrm { f a k e i m a g e } } = } & { } & { E _ { z \\sim \\mathcal { N } ( 0 , I ) , c \\sim \\mathcal { C } } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( G ( z , c ) ) = \\mathrm { f a k e } ) \\right] } \\\\ { L _ { \\mathrm { f a k e i m a g e ~ f o o l s } } = } & { } & { E _ { z \\sim \\mathcal { N } ( 0 , I ) , c \\sim \\mathcal { C } } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( G ( z , c ) ) = \\mathrm { r e a l } ) \\right] } \\\\ { L _ { \\mathrm { r e a l ~ i m a g e } } = } & { } & { E _ { c \\sim \\mathcal { C } , i \\sim \\mathcal { L } _ { c } } \\left[ \\log P ( D _ { \\mathrm { s o u r c e } } ( i ) = \\mathrm { r e a l } ) \\right] } \\\\ { L _ { \\mathrm { f a k e i m a g e c l a s s } } = } & { } & { E _ { z \\sim \\mathcal { N } ( 0 , I ) , c \\sim \\mathcal { C } } \\left[ \\log P ( D _ { \\mathrm { c l a s s } } ( G ( z , c ) ) = c ) \\right] } \\\\ { L _ { \\mathrm { r e a l ~ i m a g e c l a s s } } = } & { } & { E _ { c \\sim \\mathcal { C } , i \\sim \\mathcal { Z } _ { c } } \\left[ \\log P ( D _ { \\mathrm { c l a s s } } ( i ) = c ) \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "46d61d28718b47f1447cdcde2244b97da3d2efab9461ec27d33671ddc7176ccd.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 153, + 562, + 458, + 587.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 153, + 587.0, + 458, + 612.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 153, + 612.0, + 458, + 637.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 638, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 416, + 653 + ], + "score": 1.0, + "content": "Then the discriminator and generator losses are modified as follows (letting", + "type": "text" + }, + { + "bbox": [ + 416, + 641, + 486, + 652 + ], + "score": 0.89, + "content": "w _ { f a k e c l a s s } = 1 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 638, + 506, + 653 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 651, + 181, + 663 + ], + "spans": [ + { + "bbox": [ + 107, + 651, + 176, + 663 + ], + "score": 0.87, + "content": "w _ { r e a l c l a s s } = 1 . 0 \\AA", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 651, + 181, + 663 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 638, + 506, + 663 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 666, + 449, + 695 + ], + "lines": [ + { + "bbox": [ + 162, + 666, + 449, + 695 + ], + "spans": [ + { + "bbox": [ + 162, + 666, + 449, + 695 + ], + "score": 0.88, + "content": "\\begin{array} { r l r } { L _ { \\mathrm { d i s c r i m i n a t o r } } = } & { { } } & { L _ { \\mathrm { f a k e ~ i m a g e } } + L _ { \\mathrm { r e a l ~ i m a g e } } + w _ { r e a l c l a s s } L _ { \\mathrm { r e a l ~ c l a s s } } } \\\\ { L _ { \\mathrm { G e n e r a t o r } } = } & { { } } & { L _ { \\mathrm { f a k e ~ i m a g e ~ f o o l s } } + w _ { f a k e c l a s s } L _ { \\mathrm { f a k e ~ c l a s s } } } \\end{array}", + "type": "interline_equation", + "image_path": "acc8f92bc4ac62832f05e8e89b7f6e8675ca13042430cbc8dea8acace320e3b8.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 162, + 666, + 449, + 675.6666666666666 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 162, + 675.6666666666666, + 449, + 685.3333333333333 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 162, + 685.3333333333333, + 449, + 694.9999999999999 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Note that unlike the standard ACGAN formulation given in (Odena et al., 2016), we do not include", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 107, + 710, + 142, + 721 + ], + "score": 0.89, + "content": "L _ { \\mathrm { f a k e \\ c l a s s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "in the discriminator’s loss, to keep the discriminator from cooperating with the generator on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 721, + 196, + 731 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 196, + 731 + ], + "score": 1.0, + "content": "the classification task.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 698, + 505, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "We modified the generator network by adding one-hot class inputs, and the discriminator by adding", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 379, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 379, + 106 + ], + "score": 1.0, + "content": "class outputs alongside the source output, as in (Odena et al., 2016).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 108, + 118, + 268, + 129 + ], + "lines": [ + { + "bbox": [ + 106, + 117, + 270, + 131 + ], + "spans": [ + { + "bbox": [ + 106, + 117, + 270, + 131 + ], + "score": 1.0, + "content": "C.2 GENERATOR & DISCRIMINATOR", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 138, + 505, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 139, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 505, + 151 + ], + "score": 1.0, + "content": "We parameterized the generator as a deep neural network, which begins with a fully-connected", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 281, + 163 + ], + "score": 1.0, + "content": "mapping from the latent (noise) space to a", + "type": "text" + }, + { + "bbox": [ + 281, + 150, + 332, + 160 + ], + "score": 0.91, + "content": "4 \\times 4 \\times 5 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "dimensional image, and then successively", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 161, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 427, + 173 + ], + "score": 1.0, + "content": "upsampled (a factor of 2 by nearest neighbor), padded and applied a convolution (", + "type": "text" + }, + { + "bbox": [ + 427, + 161, + 451, + 172 + ], + "score": 0.88, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 161, + 506, + 173 + ], + "score": 1.0, + "content": "kernel, stride", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 205, + 185 + ], + "score": 1.0, + "content": "of 1) and a leaky ReLU", + "type": "text" + }, + { + "bbox": [ + 205, + 172, + 240, + 183 + ], + "score": 0.85, + "content": "\\alpha = 0 . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 171, + 505, + 185 + ], + "score": 1.0, + "content": ") nonlinearity repeatedly. We repeated this process 5 times (except", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 181, + 506, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 197 + ], + "score": 1.0, + "content": "with no upsampling on the first step, and a tanh nonlinearity on the last), while stepping the image", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 193, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 506, + 207 + ], + "score": 1.0, + "content": "depth down as follows: 512, 512, 256, 128, 64, and finally 3 (RGB) for the output image. This means", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 242, + 217 + ], + "score": 1.0, + "content": "that the final output images were", + "type": "text" + }, + { + "bbox": [ + 243, + 205, + 276, + 216 + ], + "score": 0.9, + "content": "6 4 \\times 6 4", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 205, + 505, + 217 + ], + "score": 1.0, + "content": ". We parameterized the discriminator as a convolutional", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 216, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 305, + 228 + ], + "score": 1.0, + "content": "network with 7 layers, 6 convolutions (kernels all", + "type": "text" + }, + { + "bbox": [ + 306, + 216, + 329, + 226 + ], + "score": 0.89, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 216, + 506, + 228 + ], + "score": 1.0, + "content": "; strides 2, 1, 2, 1, 2, 1; dropout after the 1st,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 226, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 506, + 239 + ], + "score": 1.0, + "content": "3rd, and 5th layers; filter depth 16, 32, 64, 128, 256, 512; batch normalization after each layer) and a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 416, + 250 + ], + "score": 1.0, + "content": "fully connected layer to a single output for real/fake. We used a leaky ReLU", + "type": "text" + }, + { + "bbox": [ + 416, + 238, + 451, + 249 + ], + "score": 0.85, + "content": "\\alpha = 0 . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 237, + 505, + 250 + ], + "score": 1.0, + "content": ") nonlinearity", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 248, + 351, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 351, + 261 + ], + "score": 1.0, + "content": "after each layer, except the final layer, where we used a tanh.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "This GAN was trained on a dataset consisting of landscape images of mountains and coastlines", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "obtained from the web. The generator was trained with the Adam optimizer (Kingma & Ba, 2015),", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 286, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 408, + 300 + ], + "score": 1.0, + "content": "and the discriminator with RMSProp. The learning rates for both were set to", + "type": "text" + }, + { + "bbox": [ + 409, + 287, + 430, + 298 + ], + "score": 0.9, + "content": "1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 286, + 506, + 300 + ], + "score": 1.0, + "content": ", and for Adam we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 296, + 507, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 120, + 313 + ], + "score": 1.0, + "content": "set", + "type": "text" + }, + { + "bbox": [ + 120, + 298, + 158, + 310 + ], + "score": 0.91, + "content": "\\beta _ { 1 } = 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 296, + 404, + 313 + ], + "score": 1.0, + "content": ". We used a latent size of 64 units. The model was trained for", + "type": "text" + }, + { + "bbox": [ + 405, + 298, + 444, + 309 + ], + "score": 0.91, + "content": "1 . 1 \\times 1 0 ^ { 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 296, + 507, + 313 + ], + "score": 1.0, + "content": "gradient steps,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 309, + 330, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 330, + 322 + ], + "score": 1.0, + "content": "when the generated images appeared to stop improving.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 108, + 334, + 202, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 203, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 203, + 347 + ], + "score": 1.0, + "content": "C.3 PIR ESTIMATOR", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 465 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 368 + ], + "score": 1.0, + "content": "Instead of predicting PIR as a scalar directly, we predict it by classifying into 100 bins via a softmax,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "which performs better empirically. This choice was motivated by noting that the scalar version was", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "having trouble fitting some highly multi-modal distributions that appear in the data. We trained the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 386, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 326, + 401 + ], + "score": 1.0, + "content": "PIR estimator with the Adam optimizer (learning rate", + "type": "text" + }, + { + "bbox": [ + 326, + 387, + 361, + 398 + ], + "score": 0.9, + "content": "5 \\cdot 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 386, + 506, + 401 + ], + "score": 1.0, + "content": "). When evaluating and when using", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 397, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 412 + ], + "score": 1.0, + "content": "this model for improving the GAN we froze the weights of the PIR estimator. We also reduced", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 409, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 505, + 423 + ], + "score": 1.0, + "content": "the output softmax’s temperature to 0.01, so it was behaving almost like a max, which empirically", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "improved results. Intuitively, a low softmax temperature in training allows the system to rapidly get", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "gradients from many different output bins and adjust the distribution appropriately, whereas when", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "actually using the system we want to be conservative with our estimates and not be too biased by low", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 453, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 288, + 466 + ], + "score": 1.0, + "content": "probability bins far from the modal estimate.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 108, + 478, + 208, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 477, + 210, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 210, + 491 + ], + "score": 1.0, + "content": "C.4 SIMULATED DATA", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 108, + 498, + 212, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 212, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 212, + 511 + ], + "score": 1.0, + "content": "C.4.1 VGG FEATURES", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 518, + 506, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 518, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 506, + 531 + ], + "score": 1.0, + "content": "The first approach we took to evaluating our system’s ability to train for different features was to use", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "activity from hidden layers of a computer vision model, specifically VGG 16 (Simonyan & Zisserman,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 442, + 552 + ], + "score": 1.0, + "content": "2014) trained on ImageNet (Russakovsky et al., 2015). In particular, we took the", + "type": "text" + }, + { + "bbox": [ + 442, + 541, + 453, + 551 + ], + "score": 0.86, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "norm of the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 551, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 347, + 564 + ], + "score": 1.0, + "content": "activity of one filter within a layer, and normalized it by the", + "type": "text" + }, + { + "bbox": [ + 348, + 552, + 357, + 562 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 551, + 506, + 564 + ], + "score": 1.0, + "content": "norm of the total layer’s activity, i.e.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 561, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 135, + 576 + ], + "score": 1.0, + "content": "letting", + "type": "text" + }, + { + "bbox": [ + 135, + 562, + 179, + 574 + ], + "score": 0.91, + "content": "\\mathrm { v G G } _ { l , f } ( i )", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 561, + 344, + 576 + ], + "score": 1.0, + "content": "be the vector of unit activations in filter", + "type": "text" + }, + { + "bbox": [ + 344, + 563, + 352, + 573 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 561, + 387, + 576 + ], + "score": 1.0, + "content": "of layer", + "type": "text" + }, + { + "bbox": [ + 387, + 563, + 392, + 572 + ], + "score": 0.64, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 561, + 482, + 576 + ], + "score": 1.0, + "content": "of VGG 16 on image", + "type": "text" + }, + { + "bbox": [ + 482, + 563, + 487, + 572 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 561, + 506, + 576 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 573, + 372, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 333, + 586 + ], + "score": 1.0, + "content": "computed PIR for that image and a given layer and filter", + "type": "text" + }, + { + "bbox": [ + 333, + 573, + 357, + 584 + ], + "score": 0.67, + "content": "l ^ { * } , f ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 573, + 372, + 586 + ], + "score": 1.0, + "content": "as:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 589, + 388, + 630 + ], + "lines": [ + { + "bbox": [ + 225, + 589, + 388, + 630 + ], + "spans": [ + { + "bbox": [ + 225, + 589, + 388, + 630 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\mathrm { P I R } _ { l ^ { * } , f ^ { * } } ( i ) = \\sqrt { \\frac { \\left| \\mathrm { V G G } _ { l ^ { * } , f ^ { * } } ( i ) \\right| _ { 2 } ^ { 2 } } { \\sum _ { f \\in l ^ { * } } \\left| \\mathrm { V G G } _ { l ^ { * } , f } ( i ) \\right| _ { 2 } ^ { 2 } } } } \\end{array}", + "type": "interline_equation", + "image_path": "a81b33135838f00b492ca1e7e4e55f5c17ef46cc5a03d415774ab2fef327b534.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 225, + 589, + 388, + 602.6666666666666 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 225, + 602.6666666666666, + 388, + 616.3333333333333 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 225, + 616.3333333333333, + 388, + 629.9999999999999 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 647 + ], + "score": 1.0, + "content": "(Note that if we did not normalize by the activity in the whole layer, the system might be able to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "“cheat” to improve the PIR by just increasing the contrast of the images, which will likely increase", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 656, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 506, + 668 + ], + "score": 1.0, + "content": "overall network activity.) As noted above, we also added binomially distributed noise to these PIRs.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "title", + "bbox": [ + 107, + 679, + 222, + 690 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 224, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 224, + 692 + ], + "score": 1.0, + "content": "C.4.2 MULTIPLE FILTERS", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "After using our system on the tasks above, we noted that its performance was quite poor at layers 5,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "6, and 7 of VGG compared to other tasks (see Fig. 2). This could suggest that our system was unable", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "to capture the complex features represented at the higher levels of VGG. However, we also noticed", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "We modified the generator network by adding one-hot class inputs, and the discriminator by adding", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 379, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 379, + 106 + ], + "score": 1.0, + "content": "class outputs alongside the source output, as in (Odena et al., 2016).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 80, + 506, + 106 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 118, + 268, + 129 + ], + "lines": [ + { + "bbox": [ + 106, + 117, + 270, + 131 + ], + "spans": [ + { + "bbox": [ + 106, + 117, + 270, + 131 + ], + "score": 1.0, + "content": "C.2 GENERATOR & DISCRIMINATOR", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 138, + 505, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 139, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 505, + 151 + ], + "score": 1.0, + "content": "We parameterized the generator as a deep neural network, which begins with a fully-connected", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 281, + 163 + ], + "score": 1.0, + "content": "mapping from the latent (noise) space to a", + "type": "text" + }, + { + "bbox": [ + 281, + 150, + 332, + 160 + ], + "score": 0.91, + "content": "4 \\times 4 \\times 5 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "dimensional image, and then successively", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 161, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 427, + 173 + ], + "score": 1.0, + "content": "upsampled (a factor of 2 by nearest neighbor), padded and applied a convolution (", + "type": "text" + }, + { + "bbox": [ + 427, + 161, + 451, + 172 + ], + "score": 0.88, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 161, + 506, + 173 + ], + "score": 1.0, + "content": "kernel, stride", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 205, + 185 + ], + "score": 1.0, + "content": "of 1) and a leaky ReLU", + "type": "text" + }, + { + "bbox": [ + 205, + 172, + 240, + 183 + ], + "score": 0.85, + "content": "\\alpha = 0 . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 171, + 505, + 185 + ], + "score": 1.0, + "content": ") nonlinearity repeatedly. We repeated this process 5 times (except", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 181, + 506, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 197 + ], + "score": 1.0, + "content": "with no upsampling on the first step, and a tanh nonlinearity on the last), while stepping the image", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 193, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 506, + 207 + ], + "score": 1.0, + "content": "depth down as follows: 512, 512, 256, 128, 64, and finally 3 (RGB) for the output image. This means", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 242, + 217 + ], + "score": 1.0, + "content": "that the final output images were", + "type": "text" + }, + { + "bbox": [ + 243, + 205, + 276, + 216 + ], + "score": 0.9, + "content": "6 4 \\times 6 4", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 205, + 505, + 217 + ], + "score": 1.0, + "content": ". We parameterized the discriminator as a convolutional", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 216, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 305, + 228 + ], + "score": 1.0, + "content": "network with 7 layers, 6 convolutions (kernels all", + "type": "text" + }, + { + "bbox": [ + 306, + 216, + 329, + 226 + ], + "score": 0.89, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 216, + 506, + 228 + ], + "score": 1.0, + "content": "; strides 2, 1, 2, 1, 2, 1; dropout after the 1st,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 226, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 506, + 239 + ], + "score": 1.0, + "content": "3rd, and 5th layers; filter depth 16, 32, 64, 128, 256, 512; batch normalization after each layer) and a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 416, + 250 + ], + "score": 1.0, + "content": "fully connected layer to a single output for real/fake. We used a leaky ReLU", + "type": "text" + }, + { + "bbox": [ + 416, + 238, + 451, + 249 + ], + "score": 0.85, + "content": "\\alpha = 0 . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 237, + 505, + 250 + ], + "score": 1.0, + "content": ") nonlinearity", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 248, + 351, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 351, + 261 + ], + "score": 1.0, + "content": "after each layer, except the final layer, where we used a tanh.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 139, + 506, + 261 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "This GAN was trained on a dataset consisting of landscape images of mountains and coastlines", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "obtained from the web. The generator was trained with the Adam optimizer (Kingma & Ba, 2015),", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 286, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 408, + 300 + ], + "score": 1.0, + "content": "and the discriminator with RMSProp. The learning rates for both were set to", + "type": "text" + }, + { + "bbox": [ + 409, + 287, + 430, + 298 + ], + "score": 0.9, + "content": "1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 286, + 506, + 300 + ], + "score": 1.0, + "content": ", and for Adam we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 296, + 507, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 120, + 313 + ], + "score": 1.0, + "content": "set", + "type": "text" + }, + { + "bbox": [ + 120, + 298, + 158, + 310 + ], + "score": 0.91, + "content": "\\beta _ { 1 } = 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 296, + 404, + 313 + ], + "score": 1.0, + "content": ". We used a latent size of 64 units. The model was trained for", + "type": "text" + }, + { + "bbox": [ + 405, + 298, + 444, + 309 + ], + "score": 0.91, + "content": "1 . 1 \\times 1 0 ^ { 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 296, + 507, + 313 + ], + "score": 1.0, + "content": "gradient steps,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 309, + 330, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 330, + 322 + ], + "score": 1.0, + "content": "when the generated images appeared to stop improving.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 264, + 507, + 322 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 334, + 202, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 203, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 203, + 347 + ], + "score": 1.0, + "content": "C.3 PIR ESTIMATOR", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 465 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 368 + ], + "score": 1.0, + "content": "Instead of predicting PIR as a scalar directly, we predict it by classifying into 100 bins via a softmax,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "which performs better empirically. This choice was motivated by noting that the scalar version was", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "having trouble fitting some highly multi-modal distributions that appear in the data. We trained the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 386, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 326, + 401 + ], + "score": 1.0, + "content": "PIR estimator with the Adam optimizer (learning rate", + "type": "text" + }, + { + "bbox": [ + 326, + 387, + 361, + 398 + ], + "score": 0.9, + "content": "5 \\cdot 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 386, + 506, + 401 + ], + "score": 1.0, + "content": "). When evaluating and when using", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 397, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 412 + ], + "score": 1.0, + "content": "this model for improving the GAN we froze the weights of the PIR estimator. We also reduced", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 409, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 505, + 423 + ], + "score": 1.0, + "content": "the output softmax’s temperature to 0.01, so it was behaving almost like a max, which empirically", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "improved results. Intuitively, a low softmax temperature in training allows the system to rapidly get", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "gradients from many different output bins and adjust the distribution appropriately, whereas when", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "actually using the system we want to be conservative with our estimates and not be too biased by low", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 453, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 288, + 466 + ], + "score": 1.0, + "content": "probability bins far from the modal estimate.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 354, + 506, + 466 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 478, + 208, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 477, + 210, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 210, + 491 + ], + "score": 1.0, + "content": "C.4 SIMULATED DATA", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 108, + 498, + 212, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 212, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 212, + 511 + ], + "score": 1.0, + "content": "C.4.1 VGG FEATURES", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 518, + 506, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 518, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 506, + 531 + ], + "score": 1.0, + "content": "The first approach we took to evaluating our system’s ability to train for different features was to use", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "activity from hidden layers of a computer vision model, specifically VGG 16 (Simonyan & Zisserman,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 442, + 552 + ], + "score": 1.0, + "content": "2014) trained on ImageNet (Russakovsky et al., 2015). In particular, we took the", + "type": "text" + }, + { + "bbox": [ + 442, + 541, + 453, + 551 + ], + "score": 0.86, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "norm of the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 551, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 347, + 564 + ], + "score": 1.0, + "content": "activity of one filter within a layer, and normalized it by the", + "type": "text" + }, + { + "bbox": [ + 348, + 552, + 357, + 562 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 551, + 506, + 564 + ], + "score": 1.0, + "content": "norm of the total layer’s activity, i.e.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 561, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 135, + 576 + ], + "score": 1.0, + "content": "letting", + "type": "text" + }, + { + "bbox": [ + 135, + 562, + 179, + 574 + ], + "score": 0.91, + "content": "\\mathrm { v G G } _ { l , f } ( i )", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 561, + 344, + 576 + ], + "score": 1.0, + "content": "be the vector of unit activations in filter", + "type": "text" + }, + { + "bbox": [ + 344, + 563, + 352, + 573 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 561, + 387, + 576 + ], + "score": 1.0, + "content": "of layer", + "type": "text" + }, + { + "bbox": [ + 387, + 563, + 392, + 572 + ], + "score": 0.64, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 561, + 482, + 576 + ], + "score": 1.0, + "content": "of VGG 16 on image", + "type": "text" + }, + { + "bbox": [ + 482, + 563, + 487, + 572 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 561, + 506, + 576 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 573, + 372, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 333, + 586 + ], + "score": 1.0, + "content": "computed PIR for that image and a given layer and filter", + "type": "text" + }, + { + "bbox": [ + 333, + 573, + 357, + 584 + ], + "score": 0.67, + "content": "l ^ { * } , f ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 573, + 372, + 586 + ], + "score": 1.0, + "content": "as:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 518, + 506, + 586 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 589, + 388, + 630 + ], + "lines": [ + { + "bbox": [ + 225, + 589, + 388, + 630 + ], + "spans": [ + { + "bbox": [ + 225, + 589, + 388, + 630 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\mathrm { P I R } _ { l ^ { * } , f ^ { * } } ( i ) = \\sqrt { \\frac { \\left| \\mathrm { V G G } _ { l ^ { * } , f ^ { * } } ( i ) \\right| _ { 2 } ^ { 2 } } { \\sum _ { f \\in l ^ { * } } \\left| \\mathrm { V G G } _ { l ^ { * } , f } ( i ) \\right| _ { 2 } ^ { 2 } } } } \\end{array}", + "type": "interline_equation", + "image_path": "a81b33135838f00b492ca1e7e4e55f5c17ef46cc5a03d415774ab2fef327b534.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 225, + 589, + 388, + 602.6666666666666 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 225, + 602.6666666666666, + 388, + 616.3333333333333 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 225, + 616.3333333333333, + 388, + 629.9999999999999 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 647 + ], + "score": 1.0, + "content": "(Note that if we did not normalize by the activity in the whole layer, the system might be able to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "“cheat” to improve the PIR by just increasing the contrast of the images, which will likely increase", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 656, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 506, + 668 + ], + "score": 1.0, + "content": "overall network activity.) As noted above, we also added binomially distributed noise to these PIRs.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 633, + 506, + 668 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 679, + 222, + 690 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 224, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 224, + 692 + ], + "score": 1.0, + "content": "C.4.2 MULTIPLE FILTERS", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "After using our system on the tasks above, we noted that its performance was quite poor at layers 5,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "6, and 7 of VGG compared to other tasks (see Fig. 2). This could suggest that our system was unable", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "to capture the complex features represented at the higher levels of VGG. However, we also noticed", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "that the feature representations at these layers tended to be quite sparse, so many of the simulated", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 342, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 355 + ], + "score": 1.0, + "content": "PIRs we generated were actually zero to within the bin width of our PIR estimator. Respectively,", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 353, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 192, + 366 + ], + "score": 1.0, + "content": "these layers had only", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 192, + 353, + 212, + 364 + ], + "score": 0.84, + "content": "20 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 212, + 353, + 215, + 366 + ], + "score": 1.0, + "content": ",", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 215, + 353, + 235, + 364 + ], + "score": 0.85, + "content": "2 5 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 235, + 353, + 255, + 366 + ], + "score": 1.0, + "content": ", and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 255, + 353, + 275, + 364 + ], + "score": 0.88, + "content": "24 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 275, + 353, + 506, + 366 + ], + "score": 1.0, + "content": "non-zero PIRs (collapsing across filters), and around half", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 316, + 377 + ], + "score": 1.0, + "content": "the filters in each (resp. 6, 4, and 5) were producing", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 316, + 364, + 347, + 375 + ], + "score": 0.88, + "content": "> 9 0 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 347, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "zero PIRs. (By contrast, the layer with", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 313, + 388 + ], + "score": 1.0, + "content": "the next greatest number of zero PIRs, fc8, still had", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 314, + 376, + 333, + 386 + ], + "score": 0.86, + "content": "69 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 334, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "nonzero PIRs overall, and had no filters in", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 132, + 399 + ], + "score": 1.0, + "content": "which", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 133, + 387, + 153, + 397 + ], + "score": 0.85, + "content": "90 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 153, + 385, + 506, + 399 + ], + "score": 1.0, + "content": "or more of the PIRs were zero.) In a few cases on layers 5, 6, and 7, all of the generated", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 396, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 411 + ], + "score": 1.0, + "content": "PIRs were zero. This clearly makes learning infeasible, and indeed we noted that there was a strong", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "score": 1.0, + "content": "relationship between number of non-zero simulated PIRs in the training dataset and the ability of our", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "system to improve PIR (see Fig. 8). This is somewhat troubling, since probably most images in the", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 429, + 313, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 313, + 443 + ], + "score": 1.0, + "content": "real world will not produce a PIR that is truly zero.", + "type": "text", + "cross_page": true + } + ], + "index": 25 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 83, + 429, + 285 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 176, + 83, + 429, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 176, + 83, + 429, + 285 + ], + "spans": [ + { + "bbox": [ + 176, + 83, + 429, + 285 + ], + "score": 0.967, + "type": "image", + "image_path": "1cc28adcf4d57ca6b2d09c22ef5c94c72c0d24e85a98baf26a61262dbd9e4b51.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 176, + 83, + 429, + 96.46666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 176, + 96.46666666666667, + 429, + 109.93333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 176, + 109.93333333333334, + 429, + 123.4 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 176, + 123.4, + 429, + 136.86666666666667 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 176, + 136.86666666666667, + 429, + 150.33333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 176, + 150.33333333333334, + 429, + 163.8 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 176, + 163.8, + 429, + 177.26666666666668 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 176, + 177.26666666666668, + 429, + 190.73333333333335 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 176, + 190.73333333333335, + 429, + 204.20000000000002 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 176, + 204.20000000000002, + 429, + 217.66666666666669 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 176, + 217.66666666666669, + 429, + 231.13333333333335 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 176, + 231.13333333333335, + 429, + 244.60000000000002 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 176, + 244.60000000000002, + 429, + 258.06666666666666 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 176, + 258.06666666666666, + 429, + 271.5333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 176, + 271.5333333333333, + 429, + 284.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 127, + 295, + 483, + 307 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 294, + 483, + 308 + ], + "spans": [ + { + "bbox": [ + 127, + 294, + 483, + 308 + ], + "score": 1.0, + "content": "Figure 8: Percent non-zero PIRs vs. effect size (Single-filter VGG tasks and color tasks)", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 11.0 + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "that the feature representations at these layers tended to be quite sparse, so many of the simulated", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 342, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 355 + ], + "score": 1.0, + "content": "PIRs we generated were actually zero to within the bin width of our PIR estimator. Respectively,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 353, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 192, + 366 + ], + "score": 1.0, + "content": "these layers had only", + "type": "text" + }, + { + "bbox": [ + 192, + 353, + 212, + 364 + ], + "score": 0.84, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 353, + 215, + 366 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 215, + 353, + 235, + 364 + ], + "score": 0.85, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 353, + 255, + 366 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 255, + 353, + 275, + 364 + ], + "score": 0.88, + "content": "24 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 353, + 506, + 366 + ], + "score": 1.0, + "content": "non-zero PIRs (collapsing across filters), and around half", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 316, + 377 + ], + "score": 1.0, + "content": "the filters in each (resp. 6, 4, and 5) were producing", + "type": "text" + }, + { + "bbox": [ + 316, + 364, + 347, + 375 + ], + "score": 0.88, + "content": "> 9 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "zero PIRs. (By contrast, the layer with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 313, + 388 + ], + "score": 1.0, + "content": "the next greatest number of zero PIRs, fc8, still had", + "type": "text" + }, + { + "bbox": [ + 314, + 376, + 333, + 386 + ], + "score": 0.86, + "content": "69 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "nonzero PIRs overall, and had no filters in", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 132, + 399 + ], + "score": 1.0, + "content": "which", + "type": "text" + }, + { + "bbox": [ + 133, + 387, + 153, + 397 + ], + "score": 0.85, + "content": "90 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 385, + 506, + 399 + ], + "score": 1.0, + "content": "or more of the PIRs were zero.) In a few cases on layers 5, 6, and 7, all of the generated", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 396, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 411 + ], + "score": 1.0, + "content": "PIRs were zero. This clearly makes learning infeasible, and indeed we noted that there was a strong", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "score": 1.0, + "content": "relationship between number of non-zero simulated PIRs in the training dataset and the ability of our", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "system to improve PIR (see Fig. 8). This is somewhat troubling, since probably most images in the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 429, + 313, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 313, + 443 + ], + "score": 1.0, + "content": "real world will not produce a PIR that is truly zero.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 447, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "in order to evaluate whether the poor performance of our system at the higher layers of VGG was", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "due to the number of zeros or to the complexity of the features, we created less sparse features from", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 266, + 482 + ], + "score": 1.0, + "content": "these layers by simply targeting a set of", + "type": "text" + }, + { + "bbox": [ + 266, + 469, + 273, + 479 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 468, + 506, + 482 + ], + "score": 1.0, + "content": "filters sampled without replacement from the layer, rather", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 479, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 364, + 494 + ], + "score": 1.0, + "content": "than a single filter. We did this by taking the norm across the", + "type": "text" + }, + { + "bbox": [ + 365, + 480, + 371, + 490 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 479, + 505, + 494 + ], + "score": 1.0, + "content": "target filters, or equivalently by", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 490, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 244, + 504 + ], + "score": 1.0, + "content": "summing the squared norms of the", + "type": "text" + }, + { + "bbox": [ + 244, + 491, + 251, + 501 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 490, + 506, + 504 + ], + "score": 1.0, + "content": "filters before taking the square root, and then normalizing by the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 295, + 515 + ], + "score": 1.0, + "content": "activity in the layer as before. Formally, letting", + "type": "text" + }, + { + "bbox": [ + 295, + 503, + 333, + 513 + ], + "score": 0.89, + "content": "a _ { 1 } , . . . , a _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 502, + 377, + 515 + ], + "score": 1.0, + "content": "be a set of", + "type": "text" + }, + { + "bbox": [ + 378, + 502, + 384, + 512 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "filter indices sampled without", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 512, + 470, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 180, + 526 + ], + "score": 1.0, + "content": "replacement from", + "type": "text" + }, + { + "bbox": [ + 180, + 513, + 206, + 525 + ], + "score": 0.77, + "content": "\\{ 0 , . . . ,", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 512, + 306, + 526 + ], + "score": 1.0, + "content": "number of filters in layer", + "type": "text" + }, + { + "bbox": [ + 307, + 514, + 312, + 525 + ], + "score": 0.61, + "content": "\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 512, + 470, + 526 + ], + "score": 1.0, + "content": ", we computed the PIR for an image as:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 532, + 390, + 573 + ], + "lines": [ + { + "bbox": [ + 220, + 532, + 390, + 573 + ], + "spans": [ + { + "bbox": [ + 220, + 532, + 390, + 573 + ], + "score": 0.95, + "content": "\\mathrm { P I R } _ { l ^ { * } , f ^ { * } , k } ( i ) = \\sqrt { \\frac { \\sum _ { j = 1 } ^ { j = k } \\left| \\mathrm { V G G } _ { l ^ { * } , a _ { j } } ( i ) \\right| _ { 2 } ^ { 2 } } { \\sum _ { f \\in l ^ { * } } \\left| \\mathrm { V G G } _ { l ^ { * } , f } ( i ) \\right| _ { 2 } ^ { 2 } } }", + "type": "interline_equation", + "image_path": "0d856d11f2fd4ca5070e7b117c391bd33d6cd193ae6a897db80141e6bad859d5.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 220, + 532, + 390, + 545.6666666666666 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 220, + 545.6666666666666, + 390, + 559.3333333333333 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 220, + 559.3333333333333, + 390, + 572.9999999999999 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 579, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 412, + 592 + ], + "score": 1.0, + "content": "The single filter cases above can be thought of as a special case of this, where", + "type": "text" + }, + { + "bbox": [ + 412, + 580, + 437, + 590 + ], + "score": 0.89, + "content": "k = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 579, + 505, + 592 + ], + "score": 1.0, + "content": ". To complement", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 590, + 492, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 185, + 603 + ], + "score": 1.0, + "content": "these, we also tried", + "type": "text" + }, + { + "bbox": [ + 185, + 591, + 215, + 601 + ], + "score": 0.9, + "content": "k = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 590, + 492, + 603 + ], + "score": 1.0, + "content": ". As above, we also added binomially distributed noise to these PIRs.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "This can also be thought of as perhaps a more realistic simulation of human behavior, in the sense", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "that it is highly unlikely that there is a single feature which influences human PIRs. Rather, there are", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "probably many related features which influence PIR in various ways. Thus it is important to evaluate", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 641, + 345, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 345, + 653 + ], + "score": 1.0, + "content": "our system’s ability to target these types of features as well.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + }, + { + "type": "title", + "bbox": [ + 107, + 667, + 177, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 180, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 180, + 680 + ], + "score": 1.0, + "content": "C.4.3 COLORS", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Finally, we also considered some simpler objectives based on targeting specific colors in the output", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "images. Analogously to the VGG features, we computed the PIRs from the vector norm of a given", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "image in the targeted color, normalized by the total image value. We considered several objectives of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 719, + 147, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 147, + 735 + ], + "score": 1.0, + "content": "this type:", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 83, + 429, + 285 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 176, + 83, + 429, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 176, + 83, + 429, + 285 + ], + "spans": [ + { + "bbox": [ + 176, + 83, + 429, + 285 + ], + "score": 0.967, + "type": "image", + "image_path": "1cc28adcf4d57ca6b2d09c22ef5c94c72c0d24e85a98baf26a61262dbd9e4b51.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 176, + 83, + 429, + 96.46666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 176, + 96.46666666666667, + 429, + 109.93333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 176, + 109.93333333333334, + 429, + 123.4 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 176, + 123.4, + 429, + 136.86666666666667 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 176, + 136.86666666666667, + 429, + 150.33333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 176, + 150.33333333333334, + 429, + 163.8 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 176, + 163.8, + 429, + 177.26666666666668 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 176, + 177.26666666666668, + 429, + 190.73333333333335 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 176, + 190.73333333333335, + 429, + 204.20000000000002 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 176, + 204.20000000000002, + 429, + 217.66666666666669 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 176, + 217.66666666666669, + 429, + 231.13333333333335 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 176, + 231.13333333333335, + 429, + 244.60000000000002 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 176, + 244.60000000000002, + 429, + 258.06666666666666 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 176, + 258.06666666666666, + 429, + 271.5333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 176, + 271.5333333333333, + 429, + 284.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 127, + 295, + 483, + 307 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 294, + 483, + 308 + ], + "spans": [ + { + "bbox": [ + 127, + 294, + 483, + 308 + ], + "score": 1.0, + "content": "Figure 8: Percent non-zero PIRs vs. effect size (Single-filter VGG tasks and color tasks)", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 11.0 + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 505, + 442 + ], + "lines": [], + "index": 20.5, + "bbox_fs": [ + 105, + 331, + 506, + 443 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 447, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "in order to evaluate whether the poor performance of our system at the higher layers of VGG was", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "due to the number of zeros or to the complexity of the features, we created less sparse features from", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 266, + 482 + ], + "score": 1.0, + "content": "these layers by simply targeting a set of", + "type": "text" + }, + { + "bbox": [ + 266, + 469, + 273, + 479 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 468, + 506, + 482 + ], + "score": 1.0, + "content": "filters sampled without replacement from the layer, rather", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 479, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 364, + 494 + ], + "score": 1.0, + "content": "than a single filter. We did this by taking the norm across the", + "type": "text" + }, + { + "bbox": [ + 365, + 480, + 371, + 490 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 479, + 505, + 494 + ], + "score": 1.0, + "content": "target filters, or equivalently by", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 490, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 244, + 504 + ], + "score": 1.0, + "content": "summing the squared norms of the", + "type": "text" + }, + { + "bbox": [ + 244, + 491, + 251, + 501 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 490, + 506, + 504 + ], + "score": 1.0, + "content": "filters before taking the square root, and then normalizing by the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 295, + 515 + ], + "score": 1.0, + "content": "activity in the layer as before. Formally, letting", + "type": "text" + }, + { + "bbox": [ + 295, + 503, + 333, + 513 + ], + "score": 0.89, + "content": "a _ { 1 } , . . . , a _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 502, + 377, + 515 + ], + "score": 1.0, + "content": "be a set of", + "type": "text" + }, + { + "bbox": [ + 378, + 502, + 384, + 512 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "filter indices sampled without", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 512, + 470, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 180, + 526 + ], + "score": 1.0, + "content": "replacement from", + "type": "text" + }, + { + "bbox": [ + 180, + 513, + 206, + 525 + ], + "score": 0.77, + "content": "\\{ 0 , . . . ,", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 512, + 306, + 526 + ], + "score": 1.0, + "content": "number of filters in layer", + "type": "text" + }, + { + "bbox": [ + 307, + 514, + 312, + 525 + ], + "score": 0.61, + "content": "\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 512, + 470, + 526 + ], + "score": 1.0, + "content": ", we computed the PIR for an image as:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 447, + 506, + 526 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 532, + 390, + 573 + ], + "lines": [ + { + "bbox": [ + 220, + 532, + 390, + 573 + ], + "spans": [ + { + "bbox": [ + 220, + 532, + 390, + 573 + ], + "score": 0.95, + "content": "\\mathrm { P I R } _ { l ^ { * } , f ^ { * } , k } ( i ) = \\sqrt { \\frac { \\sum _ { j = 1 } ^ { j = k } \\left| \\mathrm { V G G } _ { l ^ { * } , a _ { j } } ( i ) \\right| _ { 2 } ^ { 2 } } { \\sum _ { f \\in l ^ { * } } \\left| \\mathrm { V G G } _ { l ^ { * } , f } ( i ) \\right| _ { 2 } ^ { 2 } } }", + "type": "interline_equation", + "image_path": "0d856d11f2fd4ca5070e7b117c391bd33d6cd193ae6a897db80141e6bad859d5.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 220, + 532, + 390, + 545.6666666666666 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 220, + 545.6666666666666, + 390, + 559.3333333333333 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 220, + 559.3333333333333, + 390, + 572.9999999999999 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 579, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 412, + 592 + ], + "score": 1.0, + "content": "The single filter cases above can be thought of as a special case of this, where", + "type": "text" + }, + { + "bbox": [ + 412, + 580, + 437, + 590 + ], + "score": 0.89, + "content": "k = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 579, + 505, + 592 + ], + "score": 1.0, + "content": ". To complement", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 590, + 492, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 185, + 603 + ], + "score": 1.0, + "content": "these, we also tried", + "type": "text" + }, + { + "bbox": [ + 185, + 591, + 215, + 601 + ], + "score": 0.9, + "content": "k = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 590, + 492, + 603 + ], + "score": 1.0, + "content": ". As above, we also added binomially distributed noise to these PIRs.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 579, + 505, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "This can also be thought of as perhaps a more realistic simulation of human behavior, in the sense", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "that it is highly unlikely that there is a single feature which influences human PIRs. Rather, there are", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "probably many related features which influence PIR in various ways. Thus it is important to evaluate", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 641, + 345, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 345, + 653 + ], + "score": 1.0, + "content": "our system’s ability to target these types of features as well.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 607, + 505, + 653 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 667, + 177, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 180, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 180, + 680 + ], + "score": 1.0, + "content": "C.4.3 COLORS", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Finally, we also considered some simpler objectives based on targeting specific colors in the output", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "images. Analogously to the VGG features, we computed the PIRs from the vector norm of a given", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "image in the targeted color, normalized by the total image value. We considered several objectives of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 719, + 147, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 147, + 735 + ], + "score": 1.0, + "content": "this type:", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 687, + 506, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 494, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 495, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 495, + 96 + ], + "score": 1.0, + "content": "Single color: Optimizing for a single color of output image, e.g., for red the objective would be.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 266, + 99, + 380, + 133 + ], + "lines": [ + { + "bbox": [ + 266, + 99, + 380, + 133 + ], + "spans": [ + { + "bbox": [ + 266, + 99, + 380, + 133 + ], + "score": 0.94, + "content": "\\mathrm { P I R } _ { \\mathrm { r e d } } ( i ) = \\sqrt { \\frac { | i ( : , : , \\mathrm { r e d } ) | _ { 2 } ^ { 2 } } { | \\mathbf { i } ( : , : , : ) | _ { 2 } ^ { 2 } } }", + "type": "interline_equation", + "image_path": "49a2729840f850b1a0f27b66a0f0bf20944f3d6ea451216655f6921b402a27a7.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 266, + 99, + 380, + 116.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 266, + 116.0, + 380, + 133.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 139, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 153 + ], + "score": 1.0, + "content": "Two color: We split the image horizontally into a left and right half, and then computed PIR from", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 150, + 390, + 163 + ], + "spans": [ + { + "bbox": [ + 141, + 150, + 390, + 163 + ], + "score": 1.0, + "content": "one color in the left half and a different color in the right half.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 167, + 441, + 207 + ], + "lines": [ + { + "bbox": [ + 206, + 167, + 441, + 207 + ], + "spans": [ + { + "bbox": [ + 206, + 167, + 441, + 207 + ], + "score": 0.95, + "content": "\\mathsf { P I R } _ { \\mathrm { r e d \\ b l u e } } ( i ) = \\sqrt { \\frac { \\left| i ( : , : \\frac { \\mathrm { w i d t h } } { 2 } , \\mathrm { r e d } ) \\right| _ { 2 } ^ { 2 } + \\left| i ( : , \\frac { \\mathrm { w i d t h } } { 2 } : , \\mathsf { b l u e } ) \\right| _ { 2 } ^ { 2 } } { | \\dot { \\mathbf { \\mathbf { \\mathbf { \\mathbf { \\Phi } } } } } ( : , : , : ) | _ { 2 } ^ { 2 } } }", + "type": "interline_equation", + "image_path": "440e59433ffa75d02fe3d75cd6e9f9e9263257489294a25621957b7c08e324d4.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 206, + 167, + 441, + 180.33333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 206, + 180.33333333333334, + 441, + 193.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 206, + 193.66666666666669, + 441, + 207.00000000000003 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 504, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 213, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 505, + 226 + ], + "score": 1.0, + "content": "Three color: Similar to two color, but split the image into thirds, and computed PIR from a different", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 225, + 220, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 220, + 236 + ], + "score": 1.0, + "content": "color in each third.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 505, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "(As above, we also added binomially distributed noise to these PIRs.) These objectives provide", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "a useful complement to the VGG objectives discussed in section 3.1. Although the single color", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "objectives may be relevant to the classification task VGG 16 performs, the split color tasks are less", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "likely to be relevant to classification. Note that it is important that we split the images along the width", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 288, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 303 + ], + "score": 1.0, + "content": "instead of the height dimension, as there may well be semantically relevant features corresponding", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "to color divisions along the height dimension, e.g. a blue upper half and green lower half likely", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 312, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 323 + ], + "score": 1.0, + "content": "correlates with outdoor images, which would provide useful class information. By contrast, it is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 321, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 336 + ], + "score": 1.0, + "content": "harder to imagine circumstances where different colors on the left and right halves of the image", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "are semantically predictive, especially since flipping left to right is usually included in the data", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "augmentation for computer vision systems. Thus success on optimizing for these objectives would", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 356, + 328, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 328, + 368 + ], + "score": 1.0, + "content": "increase our confidence in the generality of our system.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 108, + 380, + 245, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 380, + 247, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 247, + 393 + ], + "score": 1.0, + "content": "C.5 POSSIBLE IMPROVEMENTS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 400, + 505, + 544 + ], + "lines": [ + { + "bbox": [ + 106, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "There are a number of techniques that could be explored to improve our system. As we mentioned", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 411, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 425 + ], + "score": 1.0, + "content": "above, iterating for multiple steps of PIR collection and generative model tuning is worth exploring", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "further. Also, some form of data scaling might allow the system to perform better on tasks with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "score": 1.0, + "content": "low variance. We briefly tried normalizing all data for an objective to have mean 0.5 and standard", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "deviation 0.25, but did not achieve particularly good results from this, possibly because there were", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 456, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 468 + ], + "score": 1.0, + "content": "many outliers getting clipped to 0 or 1. Still, there are many other possibilities for scaling data that", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "could potentially result in some improvement in performance. Also, one alternative approach to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 477, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 490 + ], + "score": 1.0, + "content": "training a GAN to produce high-PIR images would be to use the PIR estimator objective in the Plug &", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "score": 1.0, + "content": "Play Generative Networks framework (Nguyen et al., 2017) instead of using it to tune the GAN. This", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 499, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 513 + ], + "score": 1.0, + "content": "could be an interesting direction to explore, but its success would probably depend on expressiveness", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "of the initial generative model. With the mediocre model we started with, it’s probably better to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "actually tune the model itself, which may allow it to explore parts of image space which it had not", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 532, + 153, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 153, + 545 + ], + "score": 1.0, + "content": "previously.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 28 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 494, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 495, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 495, + 96 + ], + "score": 1.0, + "content": "Single color: Optimizing for a single color of output image, e.g., for red the objective would be.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 81, + 495, + 96 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 266, + 99, + 380, + 133 + ], + "lines": [ + { + "bbox": [ + 266, + 99, + 380, + 133 + ], + "spans": [ + { + "bbox": [ + 266, + 99, + 380, + 133 + ], + "score": 0.94, + "content": "\\mathrm { P I R } _ { \\mathrm { r e d } } ( i ) = \\sqrt { \\frac { | i ( : , : , \\mathrm { r e d } ) | _ { 2 } ^ { 2 } } { | \\mathbf { i } ( : , : , : ) | _ { 2 } ^ { 2 } } }", + "type": "interline_equation", + "image_path": "49a2729840f850b1a0f27b66a0f0bf20944f3d6ea451216655f6921b402a27a7.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 266, + 99, + 380, + 116.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 266, + 116.0, + 380, + 133.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 139, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 153 + ], + "score": 1.0, + "content": "Two color: We split the image horizontally into a left and right half, and then computed PIR from", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 150, + 390, + 163 + ], + "spans": [ + { + "bbox": [ + 141, + 150, + 390, + 163 + ], + "score": 1.0, + "content": "one color in the left half and a different color in the right half.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 138, + 505, + 163 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 167, + 441, + 207 + ], + "lines": [ + { + "bbox": [ + 206, + 167, + 441, + 207 + ], + "spans": [ + { + "bbox": [ + 206, + 167, + 441, + 207 + ], + "score": 0.95, + "content": "\\mathsf { P I R } _ { \\mathrm { r e d \\ b l u e } } ( i ) = \\sqrt { \\frac { \\left| i ( : , : \\frac { \\mathrm { w i d t h } } { 2 } , \\mathrm { r e d } ) \\right| _ { 2 } ^ { 2 } + \\left| i ( : , \\frac { \\mathrm { w i d t h } } { 2 } : , \\mathsf { b l u e } ) \\right| _ { 2 } ^ { 2 } } { | \\dot { \\mathbf { \\mathbf { \\mathbf { \\mathbf { \\Phi } } } } } ( : , : , : ) | _ { 2 } ^ { 2 } } }", + "type": "interline_equation", + "image_path": "440e59433ffa75d02fe3d75cd6e9f9e9263257489294a25621957b7c08e324d4.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 206, + 167, + 441, + 180.33333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 206, + 180.33333333333334, + 441, + 193.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 206, + 193.66666666666669, + 441, + 207.00000000000003 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 504, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 213, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 505, + 226 + ], + "score": 1.0, + "content": "Three color: Similar to two color, but split the image into thirds, and computed PIR from a different", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 225, + 220, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 220, + 236 + ], + "score": 1.0, + "content": "color in each third.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 213, + 505, + 236 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 505, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "(As above, we also added binomially distributed noise to these PIRs.) These objectives provide", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "a useful complement to the VGG objectives discussed in section 3.1. Although the single color", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "objectives may be relevant to the classification task VGG 16 performs, the split color tasks are less", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "likely to be relevant to classification. Note that it is important that we split the images along the width", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 288, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 303 + ], + "score": 1.0, + "content": "instead of the height dimension, as there may well be semantically relevant features corresponding", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "to color divisions along the height dimension, e.g. a blue upper half and green lower half likely", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 312, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 323 + ], + "score": 1.0, + "content": "correlates with outdoor images, which would provide useful class information. By contrast, it is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 321, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 336 + ], + "score": 1.0, + "content": "harder to imagine circumstances where different colors on the left and right halves of the image", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "are semantically predictive, especially since flipping left to right is usually included in the data", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "augmentation for computer vision systems. Thus success on optimizing for these objectives would", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 356, + 328, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 328, + 368 + ], + "score": 1.0, + "content": "increase our confidence in the generality of our system.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 245, + 506, + 368 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 380, + 245, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 380, + 247, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 247, + 393 + ], + "score": 1.0, + "content": "C.5 POSSIBLE IMPROVEMENTS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 400, + 505, + 544 + ], + "lines": [ + { + "bbox": [ + 106, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "There are a number of techniques that could be explored to improve our system. As we mentioned", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 411, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 425 + ], + "score": 1.0, + "content": "above, iterating for multiple steps of PIR collection and generative model tuning is worth exploring", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "further. Also, some form of data scaling might allow the system to perform better on tasks with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "score": 1.0, + "content": "low variance. We briefly tried normalizing all data for an objective to have mean 0.5 and standard", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "deviation 0.25, but did not achieve particularly good results from this, possibly because there were", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 456, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 468 + ], + "score": 1.0, + "content": "many outliers getting clipped to 0 or 1. Still, there are many other possibilities for scaling data that", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "could potentially result in some improvement in performance. Also, one alternative approach to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 477, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 490 + ], + "score": 1.0, + "content": "training a GAN to produce high-PIR images would be to use the PIR estimator objective in the Plug &", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "score": 1.0, + "content": "Play Generative Networks framework (Nguyen et al., 2017) instead of using it to tune the GAN. This", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 499, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 513 + ], + "score": 1.0, + "content": "could be an interesting direction to explore, but its success would probably depend on expressiveness", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "of the initial generative model. With the mediocre model we started with, it’s probably better to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "actually tune the model itself, which may allow it to explore parts of image space which it had not", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 532, + 153, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 153, + 545 + ], + "score": 1.0, + "content": "previously.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 401, + 506, + 545 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file