diff --git "a/parse/train/HJsjkMb0Z/HJsjkMb0Z_middle.json" "b/parse/train/HJsjkMb0Z/HJsjkMb0Z_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/train/HJsjkMb0Z/HJsjkMb0Z_middle.json"
@@ -0,0 +1,34478 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 78,
+ 419,
+ 96
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 76,
+ 421,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 114,
+ 95
+ ],
+ "score": 0.45,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 114,
+ 76,
+ 421,
+ 98
+ ],
+ "score": 1.0,
+ "content": "-REVNET: DEEP INVERTIBLE NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 114,
+ 395,
+ 149
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 113,
+ 392,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 113,
+ 392,
+ 128
+ ],
+ "score": 1.0,
+ "content": "Jorn-Henrik Jacobsen ¨ †‡, Arnold Smeulders †, Edouard Oyallon §",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 112,
+ 125,
+ 221,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 125,
+ 221,
+ 140
+ ],
+ "score": 1.0,
+ "content": "†University of Amsterdam",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 138,
+ 282,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 138,
+ 282,
+ 151
+ ],
+ "score": 1.0,
+ "content": "joern.jacobsen@bethgelab.org",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 178,
+ 333,
+ 190
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 177,
+ 335,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 177,
+ 335,
+ 191
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 204,
+ 468,
+ 368
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 203,
+ 469,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 203,
+ 469,
+ 216
+ ],
+ "score": 1.0,
+ "content": "It is widely believed that the success of deep convolutional networks is based on",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 214,
+ 470,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 214,
+ 470,
+ 228
+ ],
+ "score": 1.0,
+ "content": "progressively discarding uninformative variability about the input with respect to",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 224,
+ 470,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 224,
+ 470,
+ 239
+ ],
+ "score": 1.0,
+ "content": "the problem at hand. This is supported empirically by the difficulty of recovering",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 237,
+ 469,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 237,
+ 469,
+ 249
+ ],
+ "score": 1.0,
+ "content": "images from their hidden representations, in most commonly used network archi-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 248,
+ 470,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 248,
+ 470,
+ 260
+ ],
+ "score": 1.0,
+ "content": "tectures. In this paper we show via a one-to-one mapping that this loss of infor-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 259,
+ 470,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 259,
+ 470,
+ 271
+ ],
+ "score": 1.0,
+ "content": "mation is not a necessary condition to learn representations that generalize well",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 142,
+ 270,
+ 469,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 270,
+ 469,
+ 281
+ ],
+ "score": 1.0,
+ "content": "on complicated problems, such as ImageNet. Via a cascade of homeomorphic",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 281,
+ 470,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 281,
+ 223,
+ 292
+ ],
+ "score": 1.0,
+ "content": "layers, we build the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 223,
+ 281,
+ 228,
+ 290
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 229,
+ 281,
+ 470,
+ 292
+ ],
+ "score": 1.0,
+ "content": "-RevNet, a network that can be fully inverted up to the final",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 291,
+ 470,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 291,
+ 470,
+ 304
+ ],
+ "score": 1.0,
+ "content": "projection onto the classes, i.e. no information is discarded. Building an invertible",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 142,
+ 303,
+ 470,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 303,
+ 470,
+ 314
+ ],
+ "score": 1.0,
+ "content": "architecture is difficult, for one, because the local inversion is ill-conditioned, we",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 314,
+ 469,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 314,
+ 469,
+ 325
+ ],
+ "score": 1.0,
+ "content": "overcome this by providing an explicit inverse. An analysis of i-RevNets learned",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 325,
+ 469,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 325,
+ 469,
+ 336
+ ],
+ "score": 1.0,
+ "content": "representations suggests an alternative explanation for the success of deep net-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 336,
+ 469,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 336,
+ 469,
+ 347
+ ],
+ "score": 1.0,
+ "content": "works by a progressive contraction and linear separation with depth. To shed light",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 345,
+ 469,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 345,
+ 309,
+ 358
+ ],
+ "score": 1.0,
+ "content": "on the nature of the model learned by the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 347,
+ 314,
+ 356
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 315,
+ 345,
+ 469,
+ 358
+ ],
+ "score": 1.0,
+ "content": "-RevNet we reconstruct linear interpo-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 357,
+ 330,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 357,
+ 330,
+ 370
+ ],
+ "score": 1.0,
+ "content": "lations between natural image representations.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 390,
+ 205,
+ 403
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 208,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 208,
+ 406
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 415,
+ 505,
+ 537
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 415,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 415,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "A CNN may be very effective in classifying images of all sorts (He et al., 2016; Krizhevsky et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 427,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 427,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "2012), but the cascade of linear and nonlinear operators reveals little about the contribution of the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 504,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 504,
+ 451
+ ],
+ "score": 1.0,
+ "content": "internal representation to the classification. The learning process is characterized by a steady re-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 447,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 447,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "duction of large amounts of uninformative variability in the images while simultaneously revealing",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "score": 1.0,
+ "content": "the essence of the visual class. It is widely believed that this process is based on progressively dis-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 470,
+ 505,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 470,
+ 505,
+ 483
+ ],
+ "score": 1.0,
+ "content": "carding uninformative variability about the input with respect to the problem at hand (Dosovitskiy",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 481,
+ 505,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 505,
+ 494
+ ],
+ "score": 1.0,
+ "content": "& Brox, 2016; Mahendran & Vedaldi, 2016; Shwartz-Ziv & Tishby, 2017; Achille & Soatto, 2017).",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "However, the extent to which information is discarded is lost somewhere in the intermediate non-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 517
+ ],
+ "score": 1.0,
+ "content": "linear processing steps. In this paper, we aim to provide insight into the variability reduction process",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 514,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "by proposing an invertible convolutional network, that does not discard any information about the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 104,
+ 525,
+ 133,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 525,
+ 133,
+ 540
+ ],
+ "score": 1.0,
+ "content": "input.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 542,
+ 504,
+ 597
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": "The difficulty to recover images from their hidden representations is found in many commonly used",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 553,
+ 506,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 553,
+ 506,
+ 566
+ ],
+ "score": 1.0,
+ "content": "network architectures (Dosovitskiy & Brox, 2016; Mahendran & Vedaldi, 2016). This poses the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": "question if a substantial loss of information is necessary for successful classification. We show",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 575,
+ 506,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 575,
+ 506,
+ 588
+ ],
+ "score": 1.0,
+ "content": "information does not have to be discarded. By using homeomorphic layers, the invariance can be",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 586,
+ 297,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 297,
+ 599
+ ],
+ "score": 1.0,
+ "content": "built only at the very last layer via a projection.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 603,
+ 505,
+ 691
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 615
+ ],
+ "score": 1.0,
+ "content": "In Shwartz-Ziv & Tishby (2017), minimal sufficient statistics are proposed as a candidate to explain",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 626
+ ],
+ "score": 1.0,
+ "content": "the reduction of variability. Tishby & Zaslavsky (2015) introduces the information bottleneck princi-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 637
+ ],
+ "score": 1.0,
+ "content": "ple which states that an optimal representation must reduce the mutual information between an input",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 648
+ ],
+ "score": 1.0,
+ "content": "and its representation to reduce as much uninformative variability as possible. At the same time, the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 647,
+ 505,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 505,
+ 659
+ ],
+ "score": 1.0,
+ "content": "network should maximize the mutual information between the desired output and its representation",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 504,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 504,
+ 670
+ ],
+ "score": 1.0,
+ "content": "to effectively preserve each class from collapsing onto other classes. The effect of the information",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 668,
+ 506,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 668,
+ 506,
+ 681
+ ],
+ "score": 1.0,
+ "content": "bottleneck was demonstrated on small datasets in Shwartz-Ziv & Tishby (2017); Achille & Soatto",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 140,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 140,
+ 693
+ ],
+ "score": 1.0,
+ "content": "(2017).",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 40.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 700,
+ 400,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 699,
+ 279,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 699,
+ 279,
+ 712
+ ],
+ "score": 1.0,
+ "content": "‡Now at Bethgelab, University of Tubingen ¨",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 118,
+ 709,
+ 400,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 709,
+ 400,
+ 724
+ ],
+ "score": 1.0,
+ "content": "§CVN, CentraleSupelec, Universit ´ e Paris-Saclay ; Galen team, INRIA Saclay ´",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 303,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 303,
+ 731
+ ],
+ "score": 1.0,
+ "content": "SequeL team, INRIA Lille ; DI, ENS, Universite PSL´",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "1",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 78,
+ 419,
+ 96
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 76,
+ 421,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 114,
+ 95
+ ],
+ "score": 0.45,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 114,
+ 76,
+ 421,
+ 98
+ ],
+ "score": 1.0,
+ "content": "-REVNET: DEEP INVERTIBLE NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 113,
+ 114,
+ 395,
+ 149
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 113,
+ 392,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 113,
+ 392,
+ 128
+ ],
+ "score": 1.0,
+ "content": "Jorn-Henrik Jacobsen ¨ †‡, Arnold Smeulders †, Edouard Oyallon §",
+ "type": "text"
+ }
+ ],
+ "index": 1,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 112,
+ 125,
+ 221,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 125,
+ 221,
+ 140
+ ],
+ "score": 1.0,
+ "content": "†University of Amsterdam",
+ "type": "text"
+ }
+ ],
+ "index": 2,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 112,
+ 138,
+ 282,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 138,
+ 282,
+ 151
+ ],
+ "score": 1.0,
+ "content": "joern.jacobsen@bethgelab.org",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ }
+ ],
+ "index": 2,
+ "bbox_fs": [
+ 111,
+ 113,
+ 392,
+ 151
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 178,
+ 333,
+ 190
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 177,
+ 335,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 177,
+ 335,
+ 191
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 204,
+ 468,
+ 368
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 203,
+ 469,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 203,
+ 469,
+ 216
+ ],
+ "score": 1.0,
+ "content": "It is widely believed that the success of deep convolutional networks is based on",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 214,
+ 470,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 214,
+ 470,
+ 228
+ ],
+ "score": 1.0,
+ "content": "progressively discarding uninformative variability about the input with respect to",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 224,
+ 470,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 224,
+ 470,
+ 239
+ ],
+ "score": 1.0,
+ "content": "the problem at hand. This is supported empirically by the difficulty of recovering",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 237,
+ 469,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 237,
+ 469,
+ 249
+ ],
+ "score": 1.0,
+ "content": "images from their hidden representations, in most commonly used network archi-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 248,
+ 470,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 248,
+ 470,
+ 260
+ ],
+ "score": 1.0,
+ "content": "tectures. In this paper we show via a one-to-one mapping that this loss of infor-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 259,
+ 470,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 259,
+ 470,
+ 271
+ ],
+ "score": 1.0,
+ "content": "mation is not a necessary condition to learn representations that generalize well",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 142,
+ 270,
+ 469,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 270,
+ 469,
+ 281
+ ],
+ "score": 1.0,
+ "content": "on complicated problems, such as ImageNet. Via a cascade of homeomorphic",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 281,
+ 470,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 281,
+ 223,
+ 292
+ ],
+ "score": 1.0,
+ "content": "layers, we build the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 223,
+ 281,
+ 228,
+ 290
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 229,
+ 281,
+ 470,
+ 292
+ ],
+ "score": 1.0,
+ "content": "-RevNet, a network that can be fully inverted up to the final",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 291,
+ 470,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 291,
+ 470,
+ 304
+ ],
+ "score": 1.0,
+ "content": "projection onto the classes, i.e. no information is discarded. Building an invertible",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 142,
+ 303,
+ 470,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 303,
+ 470,
+ 314
+ ],
+ "score": 1.0,
+ "content": "architecture is difficult, for one, because the local inversion is ill-conditioned, we",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 314,
+ 469,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 314,
+ 469,
+ 325
+ ],
+ "score": 1.0,
+ "content": "overcome this by providing an explicit inverse. An analysis of i-RevNets learned",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 325,
+ 469,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 325,
+ 469,
+ 336
+ ],
+ "score": 1.0,
+ "content": "representations suggests an alternative explanation for the success of deep net-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 336,
+ 469,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 336,
+ 469,
+ 347
+ ],
+ "score": 1.0,
+ "content": "works by a progressive contraction and linear separation with depth. To shed light",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 345,
+ 469,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 345,
+ 309,
+ 358
+ ],
+ "score": 1.0,
+ "content": "on the nature of the model learned by the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 347,
+ 314,
+ 356
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 315,
+ 345,
+ 469,
+ 358
+ ],
+ "score": 1.0,
+ "content": "-RevNet we reconstruct linear interpo-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 357,
+ 330,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 357,
+ 330,
+ 370
+ ],
+ "score": 1.0,
+ "content": "lations between natural image representations.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 141,
+ 203,
+ 470,
+ 370
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 390,
+ 205,
+ 403
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 208,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 208,
+ 406
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 415,
+ 505,
+ 537
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 415,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 415,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "A CNN may be very effective in classifying images of all sorts (He et al., 2016; Krizhevsky et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 427,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 427,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "2012), but the cascade of linear and nonlinear operators reveals little about the contribution of the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 504,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 504,
+ 451
+ ],
+ "score": 1.0,
+ "content": "internal representation to the classification. The learning process is characterized by a steady re-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 447,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 447,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "duction of large amounts of uninformative variability in the images while simultaneously revealing",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "score": 1.0,
+ "content": "the essence of the visual class. It is widely believed that this process is based on progressively dis-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 470,
+ 505,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 470,
+ 505,
+ 483
+ ],
+ "score": 1.0,
+ "content": "carding uninformative variability about the input with respect to the problem at hand (Dosovitskiy",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 481,
+ 505,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 505,
+ 494
+ ],
+ "score": 1.0,
+ "content": "& Brox, 2016; Mahendran & Vedaldi, 2016; Shwartz-Ziv & Tishby, 2017; Achille & Soatto, 2017).",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "However, the extent to which information is discarded is lost somewhere in the intermediate non-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 517
+ ],
+ "score": 1.0,
+ "content": "linear processing steps. In this paper, we aim to provide insight into the variability reduction process",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 514,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "by proposing an invertible convolutional network, that does not discard any information about the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 104,
+ 525,
+ 133,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 525,
+ 133,
+ 540
+ ],
+ "score": 1.0,
+ "content": "input.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 104,
+ 415,
+ 506,
+ 540
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 542,
+ 504,
+ 597
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": "The difficulty to recover images from their hidden representations is found in many commonly used",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 553,
+ 506,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 553,
+ 506,
+ 566
+ ],
+ "score": 1.0,
+ "content": "network architectures (Dosovitskiy & Brox, 2016; Mahendran & Vedaldi, 2016). This poses the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": "question if a substantial loss of information is necessary for successful classification. We show",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 575,
+ 506,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 575,
+ 506,
+ 588
+ ],
+ "score": 1.0,
+ "content": "information does not have to be discarded. By using homeomorphic layers, the invariance can be",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 586,
+ 297,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 297,
+ 599
+ ],
+ "score": 1.0,
+ "content": "built only at the very last layer via a projection.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 105,
+ 542,
+ 506,
+ 599
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 603,
+ 505,
+ 691
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 615
+ ],
+ "score": 1.0,
+ "content": "In Shwartz-Ziv & Tishby (2017), minimal sufficient statistics are proposed as a candidate to explain",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 626
+ ],
+ "score": 1.0,
+ "content": "the reduction of variability. Tishby & Zaslavsky (2015) introduces the information bottleneck princi-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 637
+ ],
+ "score": 1.0,
+ "content": "ple which states that an optimal representation must reduce the mutual information between an input",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 648
+ ],
+ "score": 1.0,
+ "content": "and its representation to reduce as much uninformative variability as possible. At the same time, the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 647,
+ 505,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 505,
+ 659
+ ],
+ "score": 1.0,
+ "content": "network should maximize the mutual information between the desired output and its representation",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 504,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 504,
+ 670
+ ],
+ "score": 1.0,
+ "content": "to effectively preserve each class from collapsing onto other classes. The effect of the information",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 668,
+ 506,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 668,
+ 506,
+ 681
+ ],
+ "score": 1.0,
+ "content": "bottleneck was demonstrated on small datasets in Shwartz-Ziv & Tishby (2017); Achille & Soatto",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 140,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 140,
+ 693
+ ],
+ "score": 1.0,
+ "content": "(2017).",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 603,
+ 506,
+ 693
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 504,
+ 148
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "However, in this work, we show it is not a necessary condition and we build a cascade of home-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "omorphic layers, which preserves the mutual information between input and hidden representation",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 116
+ ],
+ "score": 1.0,
+ "content": "and shows that the loss of information can only occur at the final layer. This way we demonstrate",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 505,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 505,
+ 127
+ ],
+ "score": 1.0,
+ "content": "that a loss of information can be avoided while maintaining discriminability, even for large-scale",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 139
+ ],
+ "score": 1.0,
+ "content": "problems like ImageNet. One way to reduce variability is progressive contraction with respect to a",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 336,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 154,
+ 150
+ ],
+ "score": 1.0,
+ "content": "meaningful",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 137,
+ 164,
+ 147
+ ],
+ "score": 0.83,
+ "content": "\\ell ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 164,
+ 136,
+ 336,
+ 150
+ ],
+ "score": 1.0,
+ "content": "metric in the intermediate representations.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 154,
+ 505,
+ 242
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 168
+ ],
+ "score": 1.0,
+ "content": "Several works (Oyallon, 2017; Zeiler & Fergus, 2014) observed a phenomenon of progressive sep-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "aration and contraction in non-invertible networks on limited datasets. Those progressive improve-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": "ments can be interpreted as the creation of progressively stronger invariants for classification. Ide-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 188,
+ 505,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 188,
+ 505,
+ 199
+ ],
+ "score": 1.0,
+ "content": "ally, the contraction should not be too brutal to avoid removing important information from the",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "intermediate signal. This shows that a good trade-off between discriminability and invariance has",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 223
+ ],
+ "score": 1.0,
+ "content": "to be progressively built. In this paper, we extend some findings of Zeiler & Fergus (2014); Oy-",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 232
+ ],
+ "score": 1.0,
+ "content": "allon (2017) to ImageNet (Russakovsky et al., 2015) and, most importantly, show that a loss of",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 230,
+ 382,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 230,
+ 382,
+ 243
+ ],
+ "score": 1.0,
+ "content": "information is not necessary for observing a progressive contraction.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 248,
+ 505,
+ 369
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 247,
+ 506,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 247,
+ 506,
+ 260
+ ],
+ "score": 1.0,
+ "content": "The duality between invariance and separation of the classes is discussed in Mallat (2016). Here,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 271
+ ],
+ "score": 1.0,
+ "content": "intra-class variabilities are modeled as Lie groups that are processed by performing a parallel trans-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 506,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 506,
+ 282
+ ],
+ "score": 1.0,
+ "content": "port along those symmetries. Filters are adapted through learning to the specific bias of the dataset",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "and avoid to contract along discriminative directions. However, using groups beyond the Euclidean",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "case for image classification is hard. Mainly because groups associated with abstract variabilities",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 304,
+ 505,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 304,
+ 505,
+ 315
+ ],
+ "score": 1.0,
+ "content": "are difficult to estimate due to their high-dimensional nature, as well as the appropriate degree of",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "invariance required. An illustration of this framework on the Euclidean group is given by the scatter-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 505,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 505,
+ 337
+ ],
+ "score": 1.0,
+ "content": "ing transform (Mallat, 2012), which builds invariance to small translations while being recoverable",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "to a certain extent. In this work, we introduce a network that cannot discard any information ex-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 359
+ ],
+ "score": 1.0,
+ "content": "cept at the final classification stage, while we demonstrate numerically progressive contraction and",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 358,
+ 234,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 234,
+ 370
+ ],
+ "score": 1.0,
+ "content": "separation of the signal classes.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 374,
+ 505,
+ 440
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 177,
+ 387
+ ],
+ "score": 1.0,
+ "content": "We introduce the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 177,
+ 375,
+ 182,
+ 385
+ ],
+ "score": 0.7,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 374,
+ 334,
+ 387
+ ],
+ "score": 1.0,
+ "content": "-RevNet, an invertible deep network.1",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 335,
+ 375,
+ 340,
+ 384
+ ],
+ "score": 0.59,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 374,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": "-RevNets retain all information about the",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "input signal in any of their intermediate representations up until the last layer. Our architecture builds",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "score": 1.0,
+ "content": "upon the recently introduced RevNet (Gomez et al., 2017), where we replace the non-invertible",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 506,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 327,
+ 420
+ ],
+ "score": 1.0,
+ "content": "components of the original RevNets by invertible ones.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 408,
+ 332,
+ 417
+ ],
+ "score": 0.65,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 407,
+ 506,
+ 420
+ ],
+ "score": 1.0,
+ "content": "-RevNets achieve the same performance on",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 431
+ ],
+ "score": 1.0,
+ "content": "Imagenet compared to similar non-invertible RevNet and ResNet architectures (Gomez et al., 2017;",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 172,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 172,
+ 441
+ ],
+ "score": 1.0,
+ "content": "He et al., 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 441,
+ 504,
+ 484
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "To shed light on the mechanism underlying the generalization-ability of the learned representation,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 165,
+ 464
+ ],
+ "score": 1.0,
+ "content": "we show that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 452,
+ 170,
+ 461
+ ],
+ "score": 0.63,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 170,
+ 451,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "-RevNets progressively separate and contract signals with depth. Our results are",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 474
+ ],
+ "score": 1.0,
+ "content": "evidence for an effective reduction of variability through a contraction with a recoverable input",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 473,
+ 297,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 297,
+ 486
+ ],
+ "score": 1.0,
+ "content": "obtained from a series of one-to-one mappings.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 502,
+ 210,
+ 515
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 501,
+ 213,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 501,
+ 213,
+ 518
+ ],
+ "score": 1.0,
+ "content": "2 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 528,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "Several recent works show that significant information about the input images is lost with depth in",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "successful Imagenet classification CNNs (Dosovitskiy & Brox, 2016; Mahendran & Vedaldi, 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "To understand the loss of information, the references propose to invert the representations by means",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "of learned or hand-engineered priors. The approximate inversions indicate increased geometric",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 586
+ ],
+ "score": 1.0,
+ "content": "and photometric invariance with depth. Multiple other works report progressive properties of deep",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "networks that may be linked to discarded information in the representations as well, such as lin-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 593,
+ 505,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 505,
+ 607
+ ],
+ "score": 1.0,
+ "content": "earization (Radford et al., 2015), linear separability (Zeiler & Fergus, 2014), contraction (Oyallon,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "2017) and low-dimensional embeddings (Aubry & Russell, 2015). However, it is not clear from",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "above observations if the loss of information is a necessity for the observed progressive phenomena.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "In this work, we show that progressive separation and contraction can be obtained while at the same",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 312,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 312,
+ 651
+ ],
+ "score": 1.0,
+ "content": "time allowing an exact reconstruction of the signal.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 655,
+ 504,
+ 709
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Multiple frameworks have been introduced that permit to learn invertible representations under cer-",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 667,
+ 504,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 504,
+ 677
+ ],
+ "score": 1.0,
+ "content": "tain conditions. Parseval networks (Cisse et al., 2017) have been introduced to increase the robust-",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "ness of learned representations with respect to adversarial attacks. In this framework, the spectrum",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "of convolutional operators is constrained to norm 1 during learning. The linear operator is thus",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 145,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 145,
+ 711
+ ],
+ "score": 1.0,
+ "content": "injective.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 119,
+ 721,
+ 366,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 719,
+ 367,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 719,
+ 367,
+ 734
+ ],
+ "score": 1.0,
+ "content": "1Code is available at: https://github.com/jhjacobsen/pytorch-i-revnet",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 310,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 310,
+ 763
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 13,
+ "width": 8
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 504,
+ 148
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "However, in this work, we show it is not a necessary condition and we build a cascade of home-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "omorphic layers, which preserves the mutual information between input and hidden representation",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 116
+ ],
+ "score": 1.0,
+ "content": "and shows that the loss of information can only occur at the final layer. This way we demonstrate",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 505,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 505,
+ 127
+ ],
+ "score": 1.0,
+ "content": "that a loss of information can be avoided while maintaining discriminability, even for large-scale",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 139
+ ],
+ "score": 1.0,
+ "content": "problems like ImageNet. One way to reduce variability is progressive contraction with respect to a",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 336,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 154,
+ 150
+ ],
+ "score": 1.0,
+ "content": "meaningful",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 137,
+ 164,
+ 147
+ ],
+ "score": 0.83,
+ "content": "\\ell ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 164,
+ 136,
+ 336,
+ 150
+ ],
+ "score": 1.0,
+ "content": "metric in the intermediate representations.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 150
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 154,
+ 505,
+ 242
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 168
+ ],
+ "score": 1.0,
+ "content": "Several works (Oyallon, 2017; Zeiler & Fergus, 2014) observed a phenomenon of progressive sep-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "aration and contraction in non-invertible networks on limited datasets. Those progressive improve-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": "ments can be interpreted as the creation of progressively stronger invariants for classification. Ide-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 188,
+ 505,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 188,
+ 505,
+ 199
+ ],
+ "score": 1.0,
+ "content": "ally, the contraction should not be too brutal to avoid removing important information from the",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "intermediate signal. This shows that a good trade-off between discriminability and invariance has",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 223
+ ],
+ "score": 1.0,
+ "content": "to be progressively built. In this paper, we extend some findings of Zeiler & Fergus (2014); Oy-",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 232
+ ],
+ "score": 1.0,
+ "content": "allon (2017) to ImageNet (Russakovsky et al., 2015) and, most importantly, show that a loss of",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 230,
+ 382,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 230,
+ 382,
+ 243
+ ],
+ "score": 1.0,
+ "content": "information is not necessary for observing a progressive contraction.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 105,
+ 153,
+ 506,
+ 243
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 248,
+ 505,
+ 369
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 247,
+ 506,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 247,
+ 506,
+ 260
+ ],
+ "score": 1.0,
+ "content": "The duality between invariance and separation of the classes is discussed in Mallat (2016). Here,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 271
+ ],
+ "score": 1.0,
+ "content": "intra-class variabilities are modeled as Lie groups that are processed by performing a parallel trans-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 506,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 506,
+ 282
+ ],
+ "score": 1.0,
+ "content": "port along those symmetries. Filters are adapted through learning to the specific bias of the dataset",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "and avoid to contract along discriminative directions. However, using groups beyond the Euclidean",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "case for image classification is hard. Mainly because groups associated with abstract variabilities",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 304,
+ 505,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 304,
+ 505,
+ 315
+ ],
+ "score": 1.0,
+ "content": "are difficult to estimate due to their high-dimensional nature, as well as the appropriate degree of",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "invariance required. An illustration of this framework on the Euclidean group is given by the scatter-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 505,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 505,
+ 337
+ ],
+ "score": 1.0,
+ "content": "ing transform (Mallat, 2012), which builds invariance to small translations while being recoverable",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "to a certain extent. In this work, we introduce a network that cannot discard any information ex-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 359
+ ],
+ "score": 1.0,
+ "content": "cept at the final classification stage, while we demonstrate numerically progressive contraction and",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 358,
+ 234,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 234,
+ 370
+ ],
+ "score": 1.0,
+ "content": "separation of the signal classes.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 105,
+ 247,
+ 506,
+ 370
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 374,
+ 505,
+ 440
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 177,
+ 387
+ ],
+ "score": 1.0,
+ "content": "We introduce the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 177,
+ 375,
+ 182,
+ 385
+ ],
+ "score": 0.7,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 374,
+ 334,
+ 387
+ ],
+ "score": 1.0,
+ "content": "-RevNet, an invertible deep network.1",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 335,
+ 375,
+ 340,
+ 384
+ ],
+ "score": 0.59,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 374,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": "-RevNets retain all information about the",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "input signal in any of their intermediate representations up until the last layer. Our architecture builds",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "score": 1.0,
+ "content": "upon the recently introduced RevNet (Gomez et al., 2017), where we replace the non-invertible",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 506,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 327,
+ 420
+ ],
+ "score": 1.0,
+ "content": "components of the original RevNets by invertible ones.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 408,
+ 332,
+ 417
+ ],
+ "score": 0.65,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 407,
+ 506,
+ 420
+ ],
+ "score": 1.0,
+ "content": "-RevNets achieve the same performance on",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 431
+ ],
+ "score": 1.0,
+ "content": "Imagenet compared to similar non-invertible RevNet and ResNet architectures (Gomez et al., 2017;",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 172,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 172,
+ 441
+ ],
+ "score": 1.0,
+ "content": "He et al., 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 374,
+ 506,
+ 441
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 441,
+ 504,
+ 484
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "To shed light on the mechanism underlying the generalization-ability of the learned representation,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 165,
+ 464
+ ],
+ "score": 1.0,
+ "content": "we show that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 452,
+ 170,
+ 461
+ ],
+ "score": 0.63,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 170,
+ 451,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "-RevNets progressively separate and contract signals with depth. Our results are",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 474
+ ],
+ "score": 1.0,
+ "content": "evidence for an effective reduction of variability through a contraction with a recoverable input",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 473,
+ 297,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 297,
+ 486
+ ],
+ "score": 1.0,
+ "content": "obtained from a series of one-to-one mappings.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 105,
+ 439,
+ 506,
+ 486
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 502,
+ 210,
+ 515
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 501,
+ 213,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 501,
+ 213,
+ 518
+ ],
+ "score": 1.0,
+ "content": "2 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 528,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "Several recent works show that significant information about the input images is lost with depth in",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "successful Imagenet classification CNNs (Dosovitskiy & Brox, 2016; Mahendran & Vedaldi, 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "To understand the loss of information, the references propose to invert the representations by means",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "of learned or hand-engineered priors. The approximate inversions indicate increased geometric",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 586
+ ],
+ "score": 1.0,
+ "content": "and photometric invariance with depth. Multiple other works report progressive properties of deep",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "networks that may be linked to discarded information in the representations as well, such as lin-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 593,
+ 505,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 505,
+ 607
+ ],
+ "score": 1.0,
+ "content": "earization (Radford et al., 2015), linear separability (Zeiler & Fergus, 2014), contraction (Oyallon,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "2017) and low-dimensional embeddings (Aubry & Russell, 2015). However, it is not clear from",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "above observations if the loss of information is a necessity for the observed progressive phenomena.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "In this work, we show that progressive separation and contraction can be obtained while at the same",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 312,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 312,
+ 651
+ ],
+ "score": 1.0,
+ "content": "time allowing an exact reconstruction of the signal.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 41,
+ "bbox_fs": [
+ 105,
+ 529,
+ 506,
+ 651
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 655,
+ 504,
+ 709
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Multiple frameworks have been introduced that permit to learn invertible representations under cer-",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 667,
+ 504,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 504,
+ 677
+ ],
+ "score": 1.0,
+ "content": "tain conditions. Parseval networks (Cisse et al., 2017) have been introduced to increase the robust-",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "ness of learned representations with respect to adversarial attacks. In this framework, the spectrum",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "of convolutional operators is constrained to norm 1 during learning. The linear operator is thus",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 145,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 145,
+ 711
+ ],
+ "score": 1.0,
+ "content": "injective.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 49,
+ "bbox_fs": [
+ 105,
+ 655,
+ 506,
+ 711
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 182
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 94
+ ],
+ "score": 1.0,
+ "content": "As a consequence, the input of Parseval networks can be recovered if but only if the built-in non-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 504,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 504,
+ 106
+ ],
+ "score": 1.0,
+ "content": "linearities are invertible as well, which is typically not the case. Bruna et al. (2013) derive condi-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "tions under which pooling representations are, but our method directly overcomes this issue. The",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 114,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 114,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "Scattering transform (Mallat, 2012) is an example of predefined deep representation, approximately",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "invariant to translations, that can be reconstructed when the degree of invariance specified is small.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 137,
+ 505,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 137,
+ 505,
+ 150
+ ],
+ "score": 1.0,
+ "content": "Yet, it requires a gradient descent optimization and no guarantee of convergences are known. In",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 160
+ ],
+ "score": 1.0,
+ "content": "summary, the references make clear that invertibility requires special care in designing the architec-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "ture or special care in designing the optimization procedure. In this paper, we introduce a network,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 171,
+ 383,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 383,
+ 183
+ ],
+ "score": 1.0,
+ "content": "that overcomes these issues and has an exact inverse by construction.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 187,
+ 505,
+ 297
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 505,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 505,
+ 199
+ ],
+ "score": 1.0,
+ "content": "Our main inspiration for this work is the recent reversible residual network (RevNet), introduced",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 198,
+ 504,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 198,
+ 504,
+ 210
+ ],
+ "score": 1.0,
+ "content": "in Gomez et al. (2017). RevNets are in turn closely related to NICE and Real-NVP architec-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "tures (Dinh et al., 2016; 2014), which make use of constrained Jacobian determinants for generative",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 220,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 220,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "modeling. All these architectures are similar to the lifting scheme (Sweldens, 1998) and Feistel",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 244
+ ],
+ "score": 1.0,
+ "content": "cipher diagrams (Menezes et al., 1996), as we will show. RevNets illustrate how to build invert-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 241,
+ 505,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 505,
+ 255
+ ],
+ "score": 1.0,
+ "content": "ible ResNet-type blocks that avoid storing intermediate activations necessary for the backward pass.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 265
+ ],
+ "score": 1.0,
+ "content": "However, RevNets still employ multiple non-invertible operators like max-pooling and downsam-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 505,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 505,
+ 276
+ ],
+ "score": 1.0,
+ "content": "pling operators as part of the network. As such, RevNets are not invertible by construction. In this",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 274,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": "paper, we show how to build an invertible type of RevNet architecture that performs competitively",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 405,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 275,
+ 298
+ ],
+ "score": 1.0,
+ "content": "with RevNets on Imagenet, which we call",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 276,
+ 286,
+ 280,
+ 296
+ ],
+ "score": 0.62,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 286,
+ 405,
+ 298
+ ],
+ "score": 1.0,
+ "content": "-RevNet for invertible RevNet.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 314,
+ 204,
+ 326
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 312,
+ 206,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 312,
+ 150,
+ 329
+ ],
+ "score": 1.0,
+ "content": "3 THE",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 315,
+ 155,
+ 325
+ ],
+ "score": 0.69,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 312,
+ 206,
+ 329
+ ],
+ "score": 1.0,
+ "content": "-REVNET",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 339,
+ 504,
+ 373
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 338,
+ 505,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 338,
+ 328,
+ 352
+ ],
+ "score": 1.0,
+ "content": "This section introduces the general framework of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 340,
+ 333,
+ 350
+ ],
+ "score": 0.67,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 338,
+ 505,
+ 352
+ ],
+ "score": 1.0,
+ "content": "-RevNet architecture and explains how to",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 297,
+ 363
+ ],
+ "score": 1.0,
+ "content": "explicitly build an inverse or a left-inverse to an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 352,
+ 302,
+ 361
+ ],
+ "score": 0.58,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 302,
+ 351,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "-RevNet. Its practical implementation is discussed,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 362,
+ 310,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 310,
+ 373
+ ],
+ "score": 1.0,
+ "content": "and we demonstrate competitive numerical results.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 388,
+ 268,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 271,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 271,
+ 400
+ ],
+ "score": 1.0,
+ "content": "3.1 AN INVERTIBLE ARCHITECTURE",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 124,
+ 420,
+ 491,
+ 539
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 124,
+ 420,
+ 491,
+ 539
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 124,
+ 420,
+ 491,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 420,
+ 491,
+ 539
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "5cd194f32b37730890221ba1d7413ff2ff91a6729965b3a9be0fa94b8d97be5b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 25,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 124,
+ 420,
+ 491,
+ 459.6666666666667
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 124,
+ 459.6666666666667,
+ 491,
+ 499.33333333333337
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 124,
+ 499.33333333333337,
+ 491,
+ 539.0
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 551,
+ 505,
+ 600
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 551,
+ 505,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 551,
+ 259,
+ 564
+ ],
+ "score": 1.0,
+ "content": "Figure 1: The main component of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 553,
+ 264,
+ 562
+ ],
+ "score": 0.63,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 551,
+ 505,
+ 564
+ ],
+ "score": 1.0,
+ "content": "-RevNet and its inverse. RevNet blocks are interleaved with",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 210,
+ 576
+ ],
+ "score": 1.0,
+ "content": "convolutional bottlenecks",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 563,
+ 223,
+ 576
+ ],
+ "score": 0.89,
+ "content": "{ \\mathcal { F } } _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 223,
+ 563,
+ 330,
+ 576
+ ],
+ "score": 1.0,
+ "content": "and reshuffling operations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 563,
+ 342,
+ 576
+ ],
+ "score": 0.89,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 563,
+ 505,
+ 576
+ ],
+ "score": 1.0,
+ "content": "to ensure invertibility of the architecture",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 575,
+ 437,
+ 588
+ ],
+ "score": 1.0,
+ "content": "and computational efficiency. The input is processed through a splitting operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 437,
+ 574,
+ 446,
+ 586
+ ],
+ "score": 0.81,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 446,
+ 575,
+ 505,
+ 588
+ ],
+ "score": 1.0,
+ "content": ", and output is",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 475,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 172,
+ 601
+ ],
+ "score": 1.0,
+ "content": "merged through",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 587,
+ 185,
+ 599
+ ],
+ "score": 0.88,
+ "content": "\\tilde { \\mathcal { M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 588,
+ 475,
+ 601
+ ],
+ "score": 1.0,
+ "content": ". Observe that the inverse network is obtained with minimal adaptations.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28.5
+ }
+ ],
+ "index": 26.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 620,
+ 505,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 620,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 620,
+ 157,
+ 633
+ ],
+ "score": 1.0,
+ "content": "We describe",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 622,
+ 162,
+ 631
+ ],
+ "score": 0.72,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 163,
+ 620,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "-RevNets in their general setting. Their foundations are largely grounded in the recent",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 631,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 631,
+ 298,
+ 645
+ ],
+ "score": 1.0,
+ "content": "RevNet architecture (Gomez et al., 2017). In an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 299,
+ 633,
+ 303,
+ 642
+ ],
+ "score": 0.66,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 304,
+ 631,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "-RevNet, an initial input is split into two sublayers",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 107,
+ 643,
+ 505,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 644,
+ 140,
+ 657
+ ],
+ "score": 0.92,
+ "content": "( x _ { 0 } , \\tilde { x } _ { 0 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 643,
+ 315,
+ 658
+ ],
+ "score": 1.0,
+ "content": "of equal size, thanks to a splitting operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 643,
+ 374,
+ 656
+ ],
+ "score": 0.92,
+ "content": "\\tilde { S } x \\triangleq ( x _ { 0 } , \\tilde { x } _ { 0 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 643,
+ 505,
+ 658
+ ],
+ "score": 1.0,
+ "content": ", in this paper we choose to split",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 656,
+ 505,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 657,
+ 344,
+ 670
+ ],
+ "score": 1.0,
+ "content": "the channel dimension as is done in RevNets. The operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 656,
+ 352,
+ 667
+ ],
+ "score": 0.84,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 353,
+ 657,
+ 505,
+ 670
+ ],
+ "score": 1.0,
+ "content": "is linear, injective, reduces the spatial",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "resolution of the coefficients and can potentially increase the layer size, as wider layers usually",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 323,
+ 691
+ ],
+ "score": 1.0,
+ "content": "improve the classification performance (Zagoruyko",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 680,
+ 333,
+ 689
+ ],
+ "score": 0.27,
+ "content": "\\&",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 678,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "Komodakis, 2016). We can thus build a",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 689,
+ 505,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 689,
+ 167,
+ 703
+ ],
+ "score": 1.0,
+ "content": "pseudo inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 690,
+ 182,
+ 701
+ ],
+ "score": 0.89,
+ "content": "{ \\tilde { S } } ^ { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 689,
+ 375,
+ 703
+ ],
+ "score": 1.0,
+ "content": "that will be used for the inversion. Recall that if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 690,
+ 383,
+ 701
+ ],
+ "score": 0.87,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 384,
+ 689,
+ 456,
+ 703
+ ],
+ "score": 1.0,
+ "content": "is invertible, then",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 456,
+ 689,
+ 501,
+ 702
+ ],
+ "score": 0.92,
+ "content": "\\tilde { S } ^ { + } = \\tilde { S } ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 689,
+ 505,
+ 703
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 708,
+ 503,
+ 733
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 707,
+ 505,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 707,
+ 421,
+ 721
+ ],
+ "score": 1.0,
+ "content": "The number of coefficients of the next block is maintained, and at each depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 709,
+ 427,
+ 720
+ ],
+ "score": 0.8,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 427,
+ 707,
+ 505,
+ 721
+ ],
+ "score": 1.0,
+ "content": ", the representation",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 107,
+ 719,
+ 444,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 721,
+ 125,
+ 733
+ ],
+ "score": 0.9,
+ "content": "\\Phi _ { j } x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 125,
+ 719,
+ 276,
+ 734
+ ],
+ "score": 1.0,
+ "content": "is again decoupled into two variables",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 720,
+ 340,
+ 733
+ ],
+ "score": 0.93,
+ "content": "\\Phi _ { j } x \\triangleq ( x _ { j } , \\tilde { x } _ { j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 719,
+ 444,
+ 734
+ ],
+ "score": 1.0,
+ "content": "that play interlaced roles.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38.5
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published 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": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 182
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 94
+ ],
+ "score": 1.0,
+ "content": "As a consequence, the input of Parseval networks can be recovered if but only if the built-in non-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 504,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 504,
+ 106
+ ],
+ "score": 1.0,
+ "content": "linearities are invertible as well, which is typically not the case. Bruna et al. (2013) derive condi-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "tions under which pooling representations are, but our method directly overcomes this issue. The",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 114,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 114,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "Scattering transform (Mallat, 2012) is an example of predefined deep representation, approximately",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "invariant to translations, that can be reconstructed when the degree of invariance specified is small.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 137,
+ 505,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 137,
+ 505,
+ 150
+ ],
+ "score": 1.0,
+ "content": "Yet, it requires a gradient descent optimization and no guarantee of convergences are known. In",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 160
+ ],
+ "score": 1.0,
+ "content": "summary, the references make clear that invertibility requires special care in designing the architec-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "ture or special care in designing the optimization procedure. In this paper, we introduce a network,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 171,
+ 383,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 383,
+ 183
+ ],
+ "score": 1.0,
+ "content": "that overcomes these issues and has an exact inverse by construction.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 105,
+ 82,
+ 505,
+ 183
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 187,
+ 505,
+ 297
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 505,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 505,
+ 199
+ ],
+ "score": 1.0,
+ "content": "Our main inspiration for this work is the recent reversible residual network (RevNet), introduced",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 198,
+ 504,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 198,
+ 504,
+ 210
+ ],
+ "score": 1.0,
+ "content": "in Gomez et al. (2017). RevNets are in turn closely related to NICE and Real-NVP architec-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "tures (Dinh et al., 2016; 2014), which make use of constrained Jacobian determinants for generative",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 220,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 220,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "modeling. All these architectures are similar to the lifting scheme (Sweldens, 1998) and Feistel",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 244
+ ],
+ "score": 1.0,
+ "content": "cipher diagrams (Menezes et al., 1996), as we will show. RevNets illustrate how to build invert-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 241,
+ 505,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 505,
+ 255
+ ],
+ "score": 1.0,
+ "content": "ible ResNet-type blocks that avoid storing intermediate activations necessary for the backward pass.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 265
+ ],
+ "score": 1.0,
+ "content": "However, RevNets still employ multiple non-invertible operators like max-pooling and downsam-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 505,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 505,
+ 276
+ ],
+ "score": 1.0,
+ "content": "pling operators as part of the network. As such, RevNets are not invertible by construction. In this",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 274,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": "paper, we show how to build an invertible type of RevNet architecture that performs competitively",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 405,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 275,
+ 298
+ ],
+ "score": 1.0,
+ "content": "with RevNets on Imagenet, which we call",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 276,
+ 286,
+ 280,
+ 296
+ ],
+ "score": 0.62,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 286,
+ 405,
+ 298
+ ],
+ "score": 1.0,
+ "content": "-RevNet for invertible RevNet.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 105,
+ 187,
+ 506,
+ 298
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 314,
+ 204,
+ 326
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 312,
+ 206,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 312,
+ 150,
+ 329
+ ],
+ "score": 1.0,
+ "content": "3 THE",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 315,
+ 155,
+ 325
+ ],
+ "score": 0.69,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 312,
+ 206,
+ 329
+ ],
+ "score": 1.0,
+ "content": "-REVNET",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 339,
+ 504,
+ 373
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 338,
+ 505,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 338,
+ 328,
+ 352
+ ],
+ "score": 1.0,
+ "content": "This section introduces the general framework of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 340,
+ 333,
+ 350
+ ],
+ "score": 0.67,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 338,
+ 505,
+ 352
+ ],
+ "score": 1.0,
+ "content": "-RevNet architecture and explains how to",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 297,
+ 363
+ ],
+ "score": 1.0,
+ "content": "explicitly build an inverse or a left-inverse to an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 352,
+ 302,
+ 361
+ ],
+ "score": 0.58,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 302,
+ 351,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "-RevNet. Its practical implementation is discussed,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 362,
+ 310,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 310,
+ 373
+ ],
+ "score": 1.0,
+ "content": "and we demonstrate competitive numerical results.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 106,
+ 338,
+ 505,
+ 373
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 388,
+ 268,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 271,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 271,
+ 400
+ ],
+ "score": 1.0,
+ "content": "3.1 AN INVERTIBLE ARCHITECTURE",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 124,
+ 420,
+ 491,
+ 539
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 124,
+ 420,
+ 491,
+ 539
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 124,
+ 420,
+ 491,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 420,
+ 491,
+ 539
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "5cd194f32b37730890221ba1d7413ff2ff91a6729965b3a9be0fa94b8d97be5b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 25,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 124,
+ 420,
+ 491,
+ 459.6666666666667
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 124,
+ 459.6666666666667,
+ 491,
+ 499.33333333333337
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 124,
+ 499.33333333333337,
+ 491,
+ 539.0
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 551,
+ 505,
+ 600
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 551,
+ 505,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 551,
+ 259,
+ 564
+ ],
+ "score": 1.0,
+ "content": "Figure 1: The main component of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 553,
+ 264,
+ 562
+ ],
+ "score": 0.63,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 551,
+ 505,
+ 564
+ ],
+ "score": 1.0,
+ "content": "-RevNet and its inverse. RevNet blocks are interleaved with",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 210,
+ 576
+ ],
+ "score": 1.0,
+ "content": "convolutional bottlenecks",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 563,
+ 223,
+ 576
+ ],
+ "score": 0.89,
+ "content": "{ \\mathcal { F } } _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 223,
+ 563,
+ 330,
+ 576
+ ],
+ "score": 1.0,
+ "content": "and reshuffling operations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 563,
+ 342,
+ 576
+ ],
+ "score": 0.89,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 563,
+ 505,
+ 576
+ ],
+ "score": 1.0,
+ "content": "to ensure invertibility of the architecture",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 575,
+ 437,
+ 588
+ ],
+ "score": 1.0,
+ "content": "and computational efficiency. The input is processed through a splitting operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 437,
+ 574,
+ 446,
+ 586
+ ],
+ "score": 0.81,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 446,
+ 575,
+ 505,
+ 588
+ ],
+ "score": 1.0,
+ "content": ", and output is",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 475,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 172,
+ 601
+ ],
+ "score": 1.0,
+ "content": "merged through",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 587,
+ 185,
+ 599
+ ],
+ "score": 0.88,
+ "content": "\\tilde { \\mathcal { M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 588,
+ 475,
+ 601
+ ],
+ "score": 1.0,
+ "content": ". Observe that the inverse network is obtained with minimal adaptations.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28.5
+ }
+ ],
+ "index": 26.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 620,
+ 505,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 620,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 620,
+ 157,
+ 633
+ ],
+ "score": 1.0,
+ "content": "We describe",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 622,
+ 162,
+ 631
+ ],
+ "score": 0.72,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 163,
+ 620,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "-RevNets in their general setting. Their foundations are largely grounded in the recent",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 631,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 631,
+ 298,
+ 645
+ ],
+ "score": 1.0,
+ "content": "RevNet architecture (Gomez et al., 2017). In an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 299,
+ 633,
+ 303,
+ 642
+ ],
+ "score": 0.66,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 304,
+ 631,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "-RevNet, an initial input is split into two sublayers",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 107,
+ 643,
+ 505,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 644,
+ 140,
+ 657
+ ],
+ "score": 0.92,
+ "content": "( x _ { 0 } , \\tilde { x } _ { 0 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 643,
+ 315,
+ 658
+ ],
+ "score": 1.0,
+ "content": "of equal size, thanks to a splitting operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 643,
+ 374,
+ 656
+ ],
+ "score": 0.92,
+ "content": "\\tilde { S } x \\triangleq ( x _ { 0 } , \\tilde { x } _ { 0 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 643,
+ 505,
+ 658
+ ],
+ "score": 1.0,
+ "content": ", in this paper we choose to split",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 656,
+ 505,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 657,
+ 344,
+ 670
+ ],
+ "score": 1.0,
+ "content": "the channel dimension as is done in RevNets. The operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 656,
+ 352,
+ 667
+ ],
+ "score": 0.84,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 353,
+ 657,
+ 505,
+ 670
+ ],
+ "score": 1.0,
+ "content": "is linear, injective, reduces the spatial",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "resolution of the coefficients and can potentially increase the layer size, as wider layers usually",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 323,
+ 691
+ ],
+ "score": 1.0,
+ "content": "improve the classification performance (Zagoruyko",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 680,
+ 333,
+ 689
+ ],
+ "score": 0.27,
+ "content": "\\&",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 678,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "Komodakis, 2016). We can thus build a",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 689,
+ 505,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 689,
+ 167,
+ 703
+ ],
+ "score": 1.0,
+ "content": "pseudo inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 690,
+ 182,
+ 701
+ ],
+ "score": 0.89,
+ "content": "{ \\tilde { S } } ^ { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 689,
+ 375,
+ 703
+ ],
+ "score": 1.0,
+ "content": "that will be used for the inversion. Recall that if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 690,
+ 383,
+ 701
+ ],
+ "score": 0.87,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 384,
+ 689,
+ 456,
+ 703
+ ],
+ "score": 1.0,
+ "content": "is invertible, then",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 456,
+ 689,
+ 501,
+ 702
+ ],
+ "score": 0.92,
+ "content": "\\tilde { S } ^ { + } = \\tilde { S } ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 689,
+ 505,
+ 703
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 104,
+ 620,
+ 506,
+ 703
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 708,
+ 503,
+ 733
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 707,
+ 505,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 707,
+ 421,
+ 721
+ ],
+ "score": 1.0,
+ "content": "The number of coefficients of the next block is maintained, and at each depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 709,
+ 427,
+ 720
+ ],
+ "score": 0.8,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 427,
+ 707,
+ 505,
+ 721
+ ],
+ "score": 1.0,
+ "content": ", the representation",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 107,
+ 719,
+ 444,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 721,
+ 125,
+ 733
+ ],
+ "score": 0.9,
+ "content": "\\Phi _ { j } x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 125,
+ 719,
+ 276,
+ 734
+ ],
+ "score": 1.0,
+ "content": "is again decoupled into two variables",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 720,
+ 340,
+ 733
+ ],
+ "score": 0.93,
+ "content": "\\Phi _ { j } x \\triangleq ( x _ { j } , \\tilde { x } _ { j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 719,
+ 444,
+ 734
+ ],
+ "score": 1.0,
+ "content": "that play interlaced roles.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38.5,
+ "bbox_fs": [
+ 105,
+ 707,
+ 505,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 162
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 241,
+ 95
+ ],
+ "score": 1.0,
+ "content": "The strategy implemented by an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 83,
+ 246,
+ 92
+ ],
+ "score": 0.7,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "-RevNet consists in an alternation between additions, and non-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 504,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 171,
+ 106
+ ],
+ "score": 1.0,
+ "content": "linear operators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 94,
+ 183,
+ 106
+ ],
+ "score": 0.89,
+ "content": "{ \\mathcal { F } } _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 93,
+ 465,
+ 106
+ ],
+ "score": 1.0,
+ "content": ", while progressively down-sampling the signal thanks to the operators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 94,
+ 476,
+ 106
+ ],
+ "score": 0.88,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 93,
+ 504,
+ 106
+ ],
+ "score": 1.0,
+ "content": ". Here,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 119,
+ 117
+ ],
+ "score": 0.87,
+ "content": "{ \\mathcal { F } } _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 105,
+ 309,
+ 117
+ ],
+ "score": 1.0,
+ "content": "consists of convolutions and non-linearity on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 310,
+ 105,
+ 321,
+ 116
+ ],
+ "score": 0.84,
+ "content": "\\tilde { x } _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": ". The pair of the final layer is concatenated",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 114,
+ 506,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 114,
+ 218,
+ 131
+ ],
+ "score": 1.0,
+ "content": "through a merging operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 115,
+ 232,
+ 128
+ ],
+ "score": 0.83,
+ "content": "\\tilde { \\mathcal { M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 114,
+ 291,
+ 131
+ ],
+ "score": 1.0,
+ "content": ". We will omit",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 116,
+ 349,
+ 129
+ ],
+ "score": 0.89,
+ "content": "\\tilde { \\mathcal { M } } , \\tilde { \\mathcal { M } } ^ { - 1 } , \\tilde { \\mathcal { S } } ^ { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 349,
+ 114,
+ 367,
+ 131
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 367,
+ 116,
+ 375,
+ 128
+ ],
+ "score": 0.84,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 114,
+ 506,
+ 131
+ ],
+ "score": 1.0,
+ "content": "for the sake of simplicity, when",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 505,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 315,
+ 141
+ ],
+ "score": 1.0,
+ "content": "not necessary. Figure 1 describes the blocks of an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 130,
+ 320,
+ 138
+ ],
+ "score": 0.67,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 128,
+ 505,
+ 141
+ ],
+ "score": 1.0,
+ "content": "-RevNet. The design is similar to the Feistel",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 151
+ ],
+ "score": 1.0,
+ "content": "cipher diagrams (Menezes et al., 1996) or a lifting scheme (Sweldens, 1998), which are invertible",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 150,
+ 451,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 150,
+ 451,
+ 163
+ ],
+ "score": 1.0,
+ "content": "and efficient implementations of complex transforms like second generation wavelets.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 167,
+ 505,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 505,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 505,
+ 179
+ ],
+ "score": 1.0,
+ "content": "In this way, we avoid the non-invertible modules of a RevNet (e.g. max-pooling or strides) which",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 505,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 505,
+ 191
+ ],
+ "score": 1.0,
+ "content": "are necessary to train them in a reasonable time and are designed to build invariance w.r.t. translation",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 189,
+ 505,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 189,
+ 454,
+ 201
+ ],
+ "score": 1.0,
+ "content": "variability. Our method shows we can replace them by linear and invertible modules",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 454,
+ 190,
+ 466,
+ 202
+ ],
+ "score": 0.88,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 189,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": ", that can",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 200,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 200,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "reduce the spatial resolution (we refer to it as a spatial down-sampling for the sake of simplicity)",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 212,
+ 397,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 212,
+ 397,
+ 223
+ ],
+ "score": 1.0,
+ "content": "while maintaining the layer’s size by increasing the number of channels.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 228,
+ 505,
+ 261
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 505,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 505,
+ 242
+ ],
+ "score": 1.0,
+ "content": "We keep the computational cost manageable by tightly coupling downsampling and increase in",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 239,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 239,
+ 417,
+ 252
+ ],
+ "score": 1.0,
+ "content": "width of the network. Reducing the spatial resolution can be undesirable, so",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 417,
+ 240,
+ 429,
+ 251
+ ],
+ "score": 0.88,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 239,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "can potentially be",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 250,
+ 473,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 273,
+ 262
+ ],
+ "score": 1.0,
+ "content": "the identity. We refer to such networks as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 274,
+ 251,
+ 279,
+ 260
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 279,
+ 250,
+ 473,
+ 262
+ ],
+ "score": 1.0,
+ "content": "-RevNets. This leads to the following equations:",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 181,
+ 275,
+ 420,
+ 305
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 181,
+ 275,
+ 420,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 181,
+ 275,
+ 420,
+ 305
+ ],
+ "score": 0.92,
+ "content": "\\left\\{ \\begin{array} { l l } { x _ { j + 1 } = S _ { j + 1 } \\tilde { x } _ { j } } \\\\ { \\tilde { x } _ { j + 1 } = x _ { j } + \\mathcal { F } _ { j + 1 } \\tilde { x } _ { j } } \\end{array} \\right. \\iff \\quad \\left\\{ \\begin{array} { l l } { \\tilde { x } _ { j } = S _ { j + 1 } ^ { - 1 } x _ { j + 1 } } \\\\ { x _ { j } = \\tilde { x } _ { j + 1 } - \\mathcal { F } _ { j + 1 } \\tilde { x } _ { j } } \\end{array} \\right.",
+ "type": "interline_equation",
+ "image_path": "da0995a53b3995df67a5ec1405289067fd48910cb6ad5f7bd3f7226e7dcd5d2a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 181,
+ 275,
+ 420,
+ 290.0
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 181,
+ 290.0,
+ 420,
+ 305.0
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 235,
+ 314,
+ 503,
+ 336
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 235,
+ 313,
+ 504,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 313,
+ 416,
+ 326
+ ],
+ "score": 1.0,
+ "content": "Our downsampling layer can be written for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 316,
+ 424,
+ 324
+ ],
+ "score": 0.78,
+ "content": "u",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 424,
+ 313,
+ 504,
+ 326
+ ],
+ "score": 1.0,
+ "content": "the spatial variable",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 234,
+ 325,
+ 336,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 234,
+ 325,
+ 252,
+ 336
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 326,
+ 259,
+ 335
+ ],
+ "score": 0.8,
+ "content": "\\lambda",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 260,
+ 325,
+ 336,
+ 336
+ ],
+ "score": 1.0,
+ "content": "the channel index:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 320,
+ 342,
+ 420,
+ 356
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 342,
+ 420,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 342,
+ 420,
+ 356
+ ],
+ "score": 0.9,
+ "content": "S _ { j } x ( u , \\lambda ) = x ( \\Psi ( u , \\lambda ) )",
+ "type": "interline_equation",
+ "image_path": "82df32856045c35be0eba3c2ee24262b514350416991477d0f1a642428812415.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 20,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 320,
+ 342,
+ 420,
+ 356
+ ],
+ "spans": [],
+ "index": 20
+ }
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 111,
+ 343,
+ 222,
+ 389
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 111,
+ 343,
+ 222,
+ 389
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 343,
+ 222,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 343,
+ 222,
+ 389
+ ],
+ "score": 0.949,
+ "type": "image",
+ "image_path": "e760aaaf5fd2f19f2d2458df6118500adce3d86267e7acf371876f61f9b09078.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 20.0,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 111,
+ 343,
+ 222,
+ 366.0
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 111,
+ 366.0,
+ 222,
+ 389.0
+ ],
+ "spans": [],
+ "index": 21
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 415,
+ 227,
+ 438
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 227,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 227,
+ 427
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Illustration of the",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 425,
+ 211,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 211,
+ 440
+ ],
+ "score": 1.0,
+ "content": "invertible down-sampling",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ }
+ ],
+ "index": 23.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 236,
+ 362,
+ 504,
+ 452
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 235,
+ 362,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 362,
+ 264,
+ 375
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 362,
+ 273,
+ 372
+ ],
+ "score": 0.81,
+ "content": "\\Psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 274,
+ 362,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "is some invertible mapping. In principle, any invert-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 235,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "ible downsampling operation like e.g. dilated convolutions (Yu &",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 235,
+ 384,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 384,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "Koltun, 2015) can be considered here. We use the inverse of the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 235,
+ 395,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 395,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "operation described in Shi et al. (2016) as illustrated in Figure 2,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 235,
+ 406,
+ 505,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 406,
+ 505,
+ 418
+ ],
+ "score": 1.0,
+ "content": "since it preserves roughly the spatial ordering, and thus permits to",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 234,
+ 417,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 234,
+ 417,
+ 495,
+ 430
+ ],
+ "score": 1.0,
+ "content": "avoid mixing different neighborhoods via the next convolution.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 417,
+ 505,
+ 429
+ ],
+ "score": 0.76,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 234,
+ 429,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 234,
+ 429,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "is similar, but also linearly increases the channel dimensionality, for",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 235,
+ 440,
+ 351,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 440,
+ 351,
+ 452
+ ],
+ "score": 1.0,
+ "content": "example by concatenating 0.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 457,
+ 505,
+ 559
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 171,
+ 471
+ ],
+ "score": 1.0,
+ "content": "The final layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 457,
+ 274,
+ 470
+ ],
+ "score": 0.91,
+ "content": "\\Phi x \\triangleq \\Phi _ { J } x = \\left( x _ { J } , { \\tilde { x } } _ { J } \\right)",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 456,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "is then averaged along the spatial dimension, followed",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 482
+ ],
+ "score": 1.0,
+ "content": "by a ReLU non-linearity and finally a linear projection on the class probes, which are fed to a",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 285,
+ 493
+ ],
+ "score": 1.0,
+ "content": "supervised training algorithm. From a given",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 481,
+ 290,
+ 490
+ ],
+ "score": 0.74,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 479,
+ 471,
+ 493
+ ],
+ "score": 1.0,
+ "content": "-RevNet, it is possible to define a left-inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 471,
+ 480,
+ 486,
+ 491
+ ],
+ "score": 0.88,
+ "content": "\\Phi ^ { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 479,
+ 506,
+ 493
+ ],
+ "score": 1.0,
+ "content": ", i.e.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 107,
+ 489,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 492,
+ 155,
+ 503
+ ],
+ "score": 0.92,
+ "content": "\\Phi ^ { + } \\Phi x = x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 489,
+ 232,
+ 506
+ ],
+ "score": 1.0,
+ "content": "or even an inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 492,
+ 251,
+ 503
+ ],
+ "score": 0.91,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 489,
+ 272,
+ 506
+ ],
+ "score": 1.0,
+ "content": ", i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 492,
+ 370,
+ 503
+ ],
+ "score": 0.91,
+ "content": "\\Phi ^ { - 1 } \\Phi x = \\Phi ^ { - 1 } \\Phi x = x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 371,
+ 489,
+ 380,
+ 506
+ ],
+ "score": 1.0,
+ "content": "if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 381,
+ 491,
+ 389,
+ 502
+ ],
+ "score": 0.85,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 389,
+ 489,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "is invertible. In these cases,",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 503,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 503,
+ 280,
+ 516
+ ],
+ "score": 1.0,
+ "content": "the convolutional sections are as well some",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 505,
+ 285,
+ 514
+ ],
+ "score": 0.72,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 286,
+ 503,
+ 342,
+ 516
+ ],
+ "score": 1.0,
+ "content": "-RevNets. An",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 342,
+ 505,
+ 346,
+ 514
+ ],
+ "score": 0.65,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 347,
+ 503,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "-RevNet is the dual of its inverse, in the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 513,
+ 506,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 513,
+ 237,
+ 528
+ ],
+ "score": 1.0,
+ "content": "sense that it requires to replace",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 515,
+ 271,
+ 527
+ ],
+ "score": 0.93,
+ "content": "( S _ { j } , \\mathcal { F } _ { j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 272,
+ 513,
+ 287,
+ 528
+ ],
+ "score": 1.0,
+ "content": "by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 287,
+ 514,
+ 336,
+ 529
+ ],
+ "score": 0.94,
+ "content": "( S _ { j } ^ { - 1 } , - \\mathcal { F } _ { j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 337,
+ 513,
+ 396,
+ 528
+ ],
+ "score": 1.0,
+ "content": "at each depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 396,
+ 515,
+ 402,
+ 526
+ ],
+ "score": 0.8,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 402,
+ 513,
+ 460,
+ 528
+ ],
+ "score": 1.0,
+ "content": ", and to apply",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 461,
+ 513,
+ 475,
+ 525
+ ],
+ "score": 0.88,
+ "content": "{ \\tilde { S } } ^ { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 475,
+ 513,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "on the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 104,
+ 525,
+ 506,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 525,
+ 506,
+ 538
+ ],
+ "score": 1.0,
+ "content": "output. In consequence, its implementation is simple and specified by Equation (1). In Subsection",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 241,
+ 549
+ ],
+ "score": 1.0,
+ "content": "4.2, we discuss that the inverse of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 537,
+ 250,
+ 546
+ ],
+ "score": 0.83,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 536,
+ 505,
+ 549
+ ],
+ "score": 1.0,
+ "content": "does not suffer from significant round-off errors, while however",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 548,
+ 505,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 548,
+ 505,
+ 560
+ ],
+ "score": 1.0,
+ "content": "being very sensitive to small variations of an input on a large subspace, as shown in Subsection 4.1.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 573,
+ 341,
+ 584
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 342,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 342,
+ 586
+ ],
+ "score": 1.0,
+ "content": "3.2 ARCHITECTURE, TRAINING AND PERFORMANCES",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 594,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 392,
+ 606
+ ],
+ "score": 1.0,
+ "content": "In this subsection, we describe two models that we trained: an injective",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 595,
+ 397,
+ 604
+ ],
+ "score": 0.68,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 397,
+ 594,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "-RevNet (a) and a bijective",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 107,
+ 604,
+ 506,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 606,
+ 111,
+ 615
+ ],
+ "score": 0.69,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 111,
+ 604,
+ 506,
+ 618
+ ],
+ "score": 1.0,
+ "content": "-RevNet (b), with fewer parameters. The hyper-parameters were selected to be either close to the",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "ResNet and RevNet baselines in terms of the number of layers (a) or parameters (b) while keeping",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "performance competitive. For the same reasons as in Gomez et al. (2017), our scheme also allows",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "avoiding storing any intermediate activations at training time, making memory consumption for very",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 128,
+ 662
+ ],
+ "score": 1.0,
+ "content": "deep",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 650,
+ 133,
+ 659
+ ],
+ "score": 0.66,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 648,
+ 506,
+ 662
+ ],
+ "score": 1.0,
+ "content": "-RevNets not an issue in practice. We compare our implementation with a RevNet with 56",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 204,
+ 672
+ ],
+ "score": 1.0,
+ "content": "layers corresponding to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 660,
+ 227,
+ 670
+ ],
+ "score": 0.81,
+ "content": "2 8 M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "parameters, as provided in the open source release of Gomez et al.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 671,
+ 461,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 322,
+ 683
+ ],
+ "score": 1.0,
+ "content": "(2017), and with a standard ResNet of 50 layers, with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 671,
+ 344,
+ 681
+ ],
+ "score": 0.6,
+ "content": "2 6 M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 671,
+ 461,
+ 683
+ ],
+ "score": 1.0,
+ "content": "parameters (He et al., 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 45.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 154,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Each block",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 688,
+ 167,
+ 700
+ ],
+ "score": 0.89,
+ "content": "{ \\mathcal { F } } _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "is a bottleneck block, which consists of a succession of 3 convolutional operators,",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "each preceded by Batchnormalization (Ioffe & Szegedy, 2015) and ReLU non-linearity. The second",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "layer has four times fewer channels than the other two, while their corresponding kernel sizes are",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 237,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 156,
+ 732
+ ],
+ "score": 1.0,
+ "content": "respectively",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 157,
+ 721,
+ 233,
+ 732
+ ],
+ "score": 0.92,
+ "content": "1 \\times 1 , 3 \\times 3 , 1 \\times 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 234,
+ 721,
+ 237,
+ 732
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 51.5
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 759
+ ],
+ "lines": []
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 162
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 241,
+ 95
+ ],
+ "score": 1.0,
+ "content": "The strategy implemented by an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 83,
+ 246,
+ 92
+ ],
+ "score": 0.7,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "-RevNet consists in an alternation between additions, and non-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 504,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 171,
+ 106
+ ],
+ "score": 1.0,
+ "content": "linear operators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 94,
+ 183,
+ 106
+ ],
+ "score": 0.89,
+ "content": "{ \\mathcal { F } } _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 93,
+ 465,
+ 106
+ ],
+ "score": 1.0,
+ "content": ", while progressively down-sampling the signal thanks to the operators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 94,
+ 476,
+ 106
+ ],
+ "score": 0.88,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 93,
+ 504,
+ 106
+ ],
+ "score": 1.0,
+ "content": ". Here,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 119,
+ 117
+ ],
+ "score": 0.87,
+ "content": "{ \\mathcal { F } } _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 105,
+ 309,
+ 117
+ ],
+ "score": 1.0,
+ "content": "consists of convolutions and non-linearity on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 310,
+ 105,
+ 321,
+ 116
+ ],
+ "score": 0.84,
+ "content": "\\tilde { x } _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": ". The pair of the final layer is concatenated",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 114,
+ 506,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 114,
+ 218,
+ 131
+ ],
+ "score": 1.0,
+ "content": "through a merging operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 115,
+ 232,
+ 128
+ ],
+ "score": 0.83,
+ "content": "\\tilde { \\mathcal { M } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 114,
+ 291,
+ 131
+ ],
+ "score": 1.0,
+ "content": ". We will omit",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 116,
+ 349,
+ 129
+ ],
+ "score": 0.89,
+ "content": "\\tilde { \\mathcal { M } } , \\tilde { \\mathcal { M } } ^ { - 1 } , \\tilde { \\mathcal { S } } ^ { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 349,
+ 114,
+ 367,
+ 131
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 367,
+ 116,
+ 375,
+ 128
+ ],
+ "score": 0.84,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 114,
+ 506,
+ 131
+ ],
+ "score": 1.0,
+ "content": "for the sake of simplicity, when",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 505,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 315,
+ 141
+ ],
+ "score": 1.0,
+ "content": "not necessary. Figure 1 describes the blocks of an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 130,
+ 320,
+ 138
+ ],
+ "score": 0.67,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 128,
+ 505,
+ 141
+ ],
+ "score": 1.0,
+ "content": "-RevNet. The design is similar to the Feistel",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 151
+ ],
+ "score": 1.0,
+ "content": "cipher diagrams (Menezes et al., 1996) or a lifting scheme (Sweldens, 1998), which are invertible",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 150,
+ 451,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 150,
+ 451,
+ 163
+ ],
+ "score": 1.0,
+ "content": "and efficient implementations of complex transforms like second generation wavelets.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 104,
+ 82,
+ 506,
+ 163
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 167,
+ 505,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 505,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 505,
+ 179
+ ],
+ "score": 1.0,
+ "content": "In this way, we avoid the non-invertible modules of a RevNet (e.g. max-pooling or strides) which",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 505,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 505,
+ 191
+ ],
+ "score": 1.0,
+ "content": "are necessary to train them in a reasonable time and are designed to build invariance w.r.t. translation",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 189,
+ 505,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 189,
+ 454,
+ 201
+ ],
+ "score": 1.0,
+ "content": "variability. Our method shows we can replace them by linear and invertible modules",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 454,
+ 190,
+ 466,
+ 202
+ ],
+ "score": 0.88,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 189,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": ", that can",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 200,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 200,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "reduce the spatial resolution (we refer to it as a spatial down-sampling for the sake of simplicity)",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 212,
+ 397,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 212,
+ 397,
+ 223
+ ],
+ "score": 1.0,
+ "content": "while maintaining the layer’s size by increasing the number of channels.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9,
+ "bbox_fs": [
+ 105,
+ 167,
+ 505,
+ 223
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 228,
+ 505,
+ 261
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 505,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 505,
+ 242
+ ],
+ "score": 1.0,
+ "content": "We keep the computational cost manageable by tightly coupling downsampling and increase in",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 239,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 239,
+ 417,
+ 252
+ ],
+ "score": 1.0,
+ "content": "width of the network. Reducing the spatial resolution can be undesirable, so",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 417,
+ 240,
+ 429,
+ 251
+ ],
+ "score": 0.88,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 239,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "can potentially be",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 250,
+ 473,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 273,
+ 262
+ ],
+ "score": 1.0,
+ "content": "the identity. We refer to such networks as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 274,
+ 251,
+ 279,
+ 260
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 279,
+ 250,
+ 473,
+ 262
+ ],
+ "score": 1.0,
+ "content": "-RevNets. This leads to the following equations:",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 105,
+ 227,
+ 505,
+ 262
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 181,
+ 275,
+ 420,
+ 305
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 181,
+ 275,
+ 420,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 181,
+ 275,
+ 420,
+ 305
+ ],
+ "score": 0.92,
+ "content": "\\left\\{ \\begin{array} { l l } { x _ { j + 1 } = S _ { j + 1 } \\tilde { x } _ { j } } \\\\ { \\tilde { x } _ { j + 1 } = x _ { j } + \\mathcal { F } _ { j + 1 } \\tilde { x } _ { j } } \\end{array} \\right. \\iff \\quad \\left\\{ \\begin{array} { l l } { \\tilde { x } _ { j } = S _ { j + 1 } ^ { - 1 } x _ { j + 1 } } \\\\ { x _ { j } = \\tilde { x } _ { j + 1 } - \\mathcal { F } _ { j + 1 } \\tilde { x } _ { j } } \\end{array} \\right.",
+ "type": "interline_equation",
+ "image_path": "da0995a53b3995df67a5ec1405289067fd48910cb6ad5f7bd3f7226e7dcd5d2a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 181,
+ 275,
+ 420,
+ 290.0
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 181,
+ 290.0,
+ 420,
+ 305.0
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 235,
+ 314,
+ 503,
+ 336
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 235,
+ 313,
+ 504,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 313,
+ 416,
+ 326
+ ],
+ "score": 1.0,
+ "content": "Our downsampling layer can be written for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 316,
+ 424,
+ 324
+ ],
+ "score": 0.78,
+ "content": "u",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 424,
+ 313,
+ 504,
+ 326
+ ],
+ "score": 1.0,
+ "content": "the spatial variable",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 234,
+ 325,
+ 336,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 234,
+ 325,
+ 252,
+ 336
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 326,
+ 259,
+ 335
+ ],
+ "score": 0.8,
+ "content": "\\lambda",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 260,
+ 325,
+ 336,
+ 336
+ ],
+ "score": 1.0,
+ "content": "the channel index:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 234,
+ 313,
+ 504,
+ 336
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 320,
+ 342,
+ 420,
+ 356
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 342,
+ 420,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 342,
+ 420,
+ 356
+ ],
+ "score": 0.9,
+ "content": "S _ { j } x ( u , \\lambda ) = x ( \\Psi ( u , \\lambda ) )",
+ "type": "interline_equation",
+ "image_path": "82df32856045c35be0eba3c2ee24262b514350416991477d0f1a642428812415.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 20,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 320,
+ 342,
+ 420,
+ 356
+ ],
+ "spans": [],
+ "index": 20
+ }
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 111,
+ 343,
+ 222,
+ 389
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 111,
+ 343,
+ 222,
+ 389
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 343,
+ 222,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 343,
+ 222,
+ 389
+ ],
+ "score": 0.949,
+ "type": "image",
+ "image_path": "e760aaaf5fd2f19f2d2458df6118500adce3d86267e7acf371876f61f9b09078.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 20.0,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 111,
+ 343,
+ 222,
+ 366.0
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 111,
+ 366.0,
+ 222,
+ 389.0
+ ],
+ "spans": [],
+ "index": 21
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 415,
+ 227,
+ 438
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 227,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 227,
+ 427
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Illustration of the",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 425,
+ 211,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 211,
+ 440
+ ],
+ "score": 1.0,
+ "content": "invertible down-sampling",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ }
+ ],
+ "index": 23.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 236,
+ 362,
+ 504,
+ 452
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 235,
+ 362,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 362,
+ 264,
+ 375
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 362,
+ 273,
+ 372
+ ],
+ "score": 0.81,
+ "content": "\\Psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 274,
+ 362,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "is some invertible mapping. In principle, any invert-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 235,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "ible downsampling operation like e.g. dilated convolutions (Yu &",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 235,
+ 384,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 384,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "Koltun, 2015) can be considered here. We use the inverse of the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 235,
+ 395,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 395,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "operation described in Shi et al. (2016) as illustrated in Figure 2,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 235,
+ 406,
+ 505,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 406,
+ 505,
+ 418
+ ],
+ "score": 1.0,
+ "content": "since it preserves roughly the spatial ordering, and thus permits to",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 234,
+ 417,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 234,
+ 417,
+ 495,
+ 430
+ ],
+ "score": 1.0,
+ "content": "avoid mixing different neighborhoods via the next convolution.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 417,
+ 505,
+ 429
+ ],
+ "score": 0.76,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 234,
+ 429,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 234,
+ 429,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "is similar, but also linearly increases the channel dimensionality, for",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 235,
+ 440,
+ 351,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 235,
+ 440,
+ 351,
+ 452
+ ],
+ "score": 1.0,
+ "content": "example by concatenating 0.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 234,
+ 362,
+ 505,
+ 452
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 457,
+ 505,
+ 559
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 171,
+ 471
+ ],
+ "score": 1.0,
+ "content": "The final layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 457,
+ 274,
+ 470
+ ],
+ "score": 0.91,
+ "content": "\\Phi x \\triangleq \\Phi _ { J } x = \\left( x _ { J } , { \\tilde { x } } _ { J } \\right)",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 456,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "is then averaged along the spatial dimension, followed",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 482
+ ],
+ "score": 1.0,
+ "content": "by a ReLU non-linearity and finally a linear projection on the class probes, which are fed to a",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 285,
+ 493
+ ],
+ "score": 1.0,
+ "content": "supervised training algorithm. From a given",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 481,
+ 290,
+ 490
+ ],
+ "score": 0.74,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 479,
+ 471,
+ 493
+ ],
+ "score": 1.0,
+ "content": "-RevNet, it is possible to define a left-inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 471,
+ 480,
+ 486,
+ 491
+ ],
+ "score": 0.88,
+ "content": "\\Phi ^ { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 479,
+ 506,
+ 493
+ ],
+ "score": 1.0,
+ "content": ", i.e.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 107,
+ 489,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 492,
+ 155,
+ 503
+ ],
+ "score": 0.92,
+ "content": "\\Phi ^ { + } \\Phi x = x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 489,
+ 232,
+ 506
+ ],
+ "score": 1.0,
+ "content": "or even an inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 492,
+ 251,
+ 503
+ ],
+ "score": 0.91,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 489,
+ 272,
+ 506
+ ],
+ "score": 1.0,
+ "content": ", i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 492,
+ 370,
+ 503
+ ],
+ "score": 0.91,
+ "content": "\\Phi ^ { - 1 } \\Phi x = \\Phi ^ { - 1 } \\Phi x = x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 371,
+ 489,
+ 380,
+ 506
+ ],
+ "score": 1.0,
+ "content": "if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 381,
+ 491,
+ 389,
+ 502
+ ],
+ "score": 0.85,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 389,
+ 489,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "is invertible. In these cases,",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 503,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 503,
+ 280,
+ 516
+ ],
+ "score": 1.0,
+ "content": "the convolutional sections are as well some",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 505,
+ 285,
+ 514
+ ],
+ "score": 0.72,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 286,
+ 503,
+ 342,
+ 516
+ ],
+ "score": 1.0,
+ "content": "-RevNets. An",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 342,
+ 505,
+ 346,
+ 514
+ ],
+ "score": 0.65,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 347,
+ 503,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "-RevNet is the dual of its inverse, in the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 513,
+ 506,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 513,
+ 237,
+ 528
+ ],
+ "score": 1.0,
+ "content": "sense that it requires to replace",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 515,
+ 271,
+ 527
+ ],
+ "score": 0.93,
+ "content": "( S _ { j } , \\mathcal { F } _ { j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 272,
+ 513,
+ 287,
+ 528
+ ],
+ "score": 1.0,
+ "content": "by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 287,
+ 514,
+ 336,
+ 529
+ ],
+ "score": 0.94,
+ "content": "( S _ { j } ^ { - 1 } , - \\mathcal { F } _ { j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 337,
+ 513,
+ 396,
+ 528
+ ],
+ "score": 1.0,
+ "content": "at each depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 396,
+ 515,
+ 402,
+ 526
+ ],
+ "score": 0.8,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 402,
+ 513,
+ 460,
+ 528
+ ],
+ "score": 1.0,
+ "content": ", and to apply",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 461,
+ 513,
+ 475,
+ 525
+ ],
+ "score": 0.88,
+ "content": "{ \\tilde { S } } ^ { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 475,
+ 513,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "on the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 104,
+ 525,
+ 506,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 525,
+ 506,
+ 538
+ ],
+ "score": 1.0,
+ "content": "output. In consequence, its implementation is simple and specified by Equation (1). In Subsection",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 241,
+ 549
+ ],
+ "score": 1.0,
+ "content": "4.2, we discuss that the inverse of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 537,
+ 250,
+ 546
+ ],
+ "score": 0.83,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 536,
+ 505,
+ 549
+ ],
+ "score": 1.0,
+ "content": "does not suffer from significant round-off errors, while however",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 548,
+ 505,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 548,
+ 505,
+ 560
+ ],
+ "score": 1.0,
+ "content": "being very sensitive to small variations of an input on a large subspace, as shown in Subsection 4.1.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 104,
+ 456,
+ 506,
+ 560
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 573,
+ 341,
+ 584
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 342,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 342,
+ 586
+ ],
+ "score": 1.0,
+ "content": "3.2 ARCHITECTURE, TRAINING AND PERFORMANCES",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 594,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 392,
+ 606
+ ],
+ "score": 1.0,
+ "content": "In this subsection, we describe two models that we trained: an injective",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 595,
+ 397,
+ 604
+ ],
+ "score": 0.68,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 397,
+ 594,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "-RevNet (a) and a bijective",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 107,
+ 604,
+ 506,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 606,
+ 111,
+ 615
+ ],
+ "score": 0.69,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 111,
+ 604,
+ 506,
+ 618
+ ],
+ "score": 1.0,
+ "content": "-RevNet (b), with fewer parameters. The hyper-parameters were selected to be either close to the",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "ResNet and RevNet baselines in terms of the number of layers (a) or parameters (b) while keeping",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "performance competitive. For the same reasons as in Gomez et al. (2017), our scheme also allows",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "avoiding storing any intermediate activations at training time, making memory consumption for very",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 128,
+ 662
+ ],
+ "score": 1.0,
+ "content": "deep",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 650,
+ 133,
+ 659
+ ],
+ "score": 0.66,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 648,
+ 506,
+ 662
+ ],
+ "score": 1.0,
+ "content": "-RevNets not an issue in practice. We compare our implementation with a RevNet with 56",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 204,
+ 672
+ ],
+ "score": 1.0,
+ "content": "layers corresponding to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 660,
+ 227,
+ 670
+ ],
+ "score": 0.81,
+ "content": "2 8 M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "parameters, as provided in the open source release of Gomez et al.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 671,
+ 461,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 322,
+ 683
+ ],
+ "score": 1.0,
+ "content": "(2017), and with a standard ResNet of 50 layers, with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 671,
+ 344,
+ 681
+ ],
+ "score": 0.6,
+ "content": "2 6 M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 671,
+ 461,
+ 683
+ ],
+ "score": 1.0,
+ "content": "parameters (He et al., 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 45.5,
+ "bbox_fs": [
+ 105,
+ 594,
+ 506,
+ 683
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 154,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Each block",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 688,
+ 167,
+ 700
+ ],
+ "score": 0.89,
+ "content": "{ \\mathcal { F } } _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "is a bottleneck block, which consists of a succession of 3 convolutional operators,",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "each preceded by Batchnormalization (Ioffe & Szegedy, 2015) and ReLU non-linearity. The second",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "layer has four times fewer channels than the other two, while their corresponding kernel sizes are",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 237,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 156,
+ 732
+ ],
+ "score": 1.0,
+ "content": "respectively",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 157,
+ 721,
+ 233,
+ 732
+ ],
+ "score": 0.92,
+ "content": "1 \\times 1 , 3 \\times 3 , 1 \\times 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 234,
+ 721,
+ 237,
+ 732
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 51.5,
+ "bbox_fs": [
+ 105,
+ 687,
+ 506,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 162,
+ 79,
+ 449,
+ 144
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 162,
+ 79,
+ 449,
+ 144
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 162,
+ 79,
+ 449,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 79,
+ 449,
+ 144
+ ],
+ "score": 0.963,
+ "html": "| Architecture | Injective | Bijective | Top-1error | Parameters |
| ResNet | | | 24.7 | 26M |
| RevNet | = | | 25.2 | 28M |
| i-RevNet (a) | yes | = | 24.7 | 181M |
| i-RevNet (b) | yes | yes | 26.7 | 29M |
",
+ "type": "table",
+ "image_path": "251ee7588dce774e1181820f0000e4bc7c393edb175e76ba0ff406a4c8e79c70.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 162,
+ 79,
+ 449,
+ 100.66666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 162,
+ 100.66666666666667,
+ 449,
+ 122.33333333333334
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 162,
+ 122.33333333333334,
+ 449,
+ 144.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 156,
+ 503,
+ 179
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 505,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 505,
+ 169
+ ],
+ "score": 1.0,
+ "content": "Table 1: Comparison of different architectures trained on ILSVRC-2012, in terms of classification",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 168,
+ 251,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 168,
+ 251,
+ 180
+ ],
+ "score": 1.0,
+ "content": "accuracy and number of parameters",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 221,
+ 503,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 504,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 504,
+ 233
+ ],
+ "score": 1.0,
+ "content": "The final representation is spatially averaged and projected onto the 1000 classes after a ReLU non-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "linearity. We now discuss how we progressively decrease the spatial resolution, while increasing the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 243,
+ 334,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 318,
+ 257
+ ],
+ "score": 1.0,
+ "content": "number of channels per layer by use of the operators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 318,
+ 243,
+ 330,
+ 255
+ ],
+ "score": 0.89,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 330,
+ 243,
+ 334,
+ 257
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 259,
+ 505,
+ 304
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "We first describe the model (a), that consists of 56 layers which have been optimized to match",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 283
+ ],
+ "score": 1.0,
+ "content": "the performances of a RevNet or a ResNet with approximatively the same number of layers. In",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 282,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 282,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": "particular, we explain how we progressively decrease the spatial resolution, while increasing the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 336,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 320,
+ 306
+ ],
+ "score": 1.0,
+ "content": "number of channels per block by use of the operators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 321,
+ 293,
+ 332,
+ 305
+ ],
+ "score": 0.89,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 332,
+ 293,
+ 336,
+ 306
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 443
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 196,
+ 325
+ ],
+ "score": 1.0,
+ "content": "The splitting operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 310,
+ 205,
+ 322
+ ],
+ "score": 0.82,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 311,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "consists in a linear and injective embedding that downsamples by a factor",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 506,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 322,
+ 117,
+ 333
+ ],
+ "score": 0.76,
+ "content": "4 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 117,
+ 321,
+ 506,
+ 337
+ ],
+ "score": 1.0,
+ "content": "the spatial resolution by increasing the number of output channels from 48 to 96 by simply adding",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 333,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 333,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": "0. The latter permits to increase the initial layer size, and consequently, the size of the next layers as",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 104,
+ 344,
+ 504,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 344,
+ 403,
+ 358
+ ],
+ "score": 1.0,
+ "content": "performed in Gomez et al. (2017); it is thus not a bijective yet an injective",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 403,
+ 345,
+ 408,
+ 355
+ ],
+ "score": 0.61,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 408,
+ 344,
+ 482,
+ 358
+ ],
+ "score": 1.0,
+ "content": "-RevNet. At depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 345,
+ 504,
+ 357
+ ],
+ "score": 0.53,
+ "content": "j , { \\mathcal { S } } _ { j }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 356,
+ 505,
+ 368
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 505,
+ 368
+ ],
+ "score": 1.0,
+ "content": "allows us to reduce the number of computations while maintaining good classification performance.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 365,
+ 506,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 365,
+ 397,
+ 380
+ ],
+ "score": 1.0,
+ "content": "It will correspond to a downsampling operator respectively at the depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 367,
+ 457,
+ 378
+ ],
+ "score": 0.55,
+ "content": "3 j = 1 5 , 2 7 , 4 ",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 458,
+ 365,
+ 506,
+ 380
+ ],
+ "score": 1.0,
+ "content": "5 (3j as one",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 377,
+ 505,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 377,
+ 505,
+ 390
+ ],
+ "score": 1.0,
+ "content": "block corresponds to three layers), similar to a normal RevNet. The spatial resolution of these layers",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 104,
+ 387,
+ 506,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 387,
+ 193,
+ 401
+ ],
+ "score": 1.0,
+ "content": "is reduced by a factor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 388,
+ 204,
+ 399
+ ],
+ "score": 0.83,
+ "content": "2 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 387,
+ 506,
+ 401
+ ],
+ "score": 1.0,
+ "content": "while increasing the number of channels by a factor of 4 respectively to 48,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 399,
+ 506,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 399,
+ 506,
+ 412
+ ],
+ "score": 1.0,
+ "content": "192, 768 and 3072. Furthermore, it means that the corresponding spatial resolutions for an input of",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 409,
+ 506,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 409,
+ 124,
+ 423
+ ],
+ "score": 1.0,
+ "content": "size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 410,
+ 145,
+ 421
+ ],
+ "score": 0.87,
+ "content": "2 2 4 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 409,
+ 213,
+ 423
+ ],
+ "score": 1.0,
+ "content": "are respectively",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 214,
+ 410,
+ 304,
+ 421
+ ],
+ "score": 0.47,
+ "content": "1 1 2 ^ { 2 } , 5 6 ^ { 2 } , 2 8 ^ { 2 } , 1 4 ^ { 2 } , 7 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 305,
+ 409,
+ 506,
+ 423
+ ],
+ "score": 1.0,
+ "content": ". The total number of coefficients at each layer is",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 150,
+ 435
+ ],
+ "score": 1.0,
+ "content": "then about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 151,
+ 421,
+ 175,
+ 432
+ ],
+ "score": 0.81,
+ "content": "0 . 3 M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 420,
+ 280,
+ 435
+ ],
+ "score": 1.0,
+ "content": ". All the remaining blocks",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 422,
+ 292,
+ 433
+ ],
+ "score": 0.88,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 292,
+ 420,
+ 506,
+ 435
+ ],
+ "score": 1.0,
+ "content": "are kept fix to the identity as explained in the section",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 431,
+ 137,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 137,
+ 445
+ ],
+ "score": 1.0,
+ "content": "above.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 529
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "Architecture (b) is bijective, it consists of 300 layers (100 blocks), whose total numbers of parame-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 493,
+ 474
+ ],
+ "score": 1.0,
+ "content": "ters have been optimized to match those of a RevNet with 56 layers. Initially, the input is split via",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 460,
+ 501,
+ 472
+ ],
+ "score": 0.82,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 461,
+ 505,
+ 474
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 350,
+ 485
+ ],
+ "score": 1.0,
+ "content": "which corresponds to an invertible spatial downsampling of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 351,
+ 471,
+ 361,
+ 483
+ ],
+ "score": 0.85,
+ "content": "2 ^ { \\frac { 5 } { 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 362,
+ 471,
+ 505,
+ 485
+ ],
+ "score": 1.0,
+ "content": "that increases the number of chan-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 505,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 463,
+ 496
+ ],
+ "score": 1.0,
+ "content": "nels from 3 to 12. It thus keeps the dimension constant and permits building a bijective",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 463,
+ 484,
+ 468,
+ 494
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 468,
+ 483,
+ 505,
+ 496
+ ],
+ "score": 1.0,
+ "content": "-RevNet.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 168,
+ 507
+ ],
+ "score": 1.0,
+ "content": "Then, at depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 495,
+ 213,
+ 506
+ ],
+ "score": 0.5,
+ "content": "3 j = 3 , 2 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 493,
+ 396,
+ 507
+ ],
+ "score": 1.0,
+ "content": ", 69, 285, the spatial resolution is reduced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 396,
+ 494,
+ 407,
+ 505
+ ],
+ "score": 0.81,
+ "content": "2 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 407,
+ 493,
+ 423,
+ 507
+ ],
+ "score": 1.0,
+ "content": "via",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 495,
+ 435,
+ 507
+ ],
+ "score": 0.87,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 493,
+ 506,
+ 507
+ ],
+ "score": 1.0,
+ "content": ". Contrary to the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 386,
+ 520
+ ],
+ "score": 1.0,
+ "content": "architecture (a), the dimensionality of each layer is constantly equal to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 387,
+ 506,
+ 422,
+ 517
+ ],
+ "score": 0.92,
+ "content": "3 \\times 2 2 4 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 505,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": ", until the final layer,",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 518,
+ 267,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 518,
+ 267,
+ 528
+ ],
+ "score": 1.0,
+ "content": "with channel sizes of 24, 96, 384, 1536.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 534,
+ 297,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 534,
+ 297,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 534,
+ 297,
+ 546
+ ],
+ "score": 1.0,
+ "content": "For both networks, the training on Imagenet",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 545,
+ 297,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 297,
+ 558
+ ],
+ "score": 1.0,
+ "content": "follows the same setup as Gomez et al. (2017).",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 556,
+ 297,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 556,
+ 297,
+ 568
+ ],
+ "score": 1.0,
+ "content": "We train with SGD and momentum of 0.9. We",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 567,
+ 298,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 567,
+ 221,
+ 579
+ ],
+ "score": 1.0,
+ "content": "regularized the model with a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 221,
+ 567,
+ 231,
+ 577
+ ],
+ "score": 0.83,
+ "content": "\\ell ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 567,
+ 298,
+ 579
+ ],
+ "score": 1.0,
+ "content": "weight decay of",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 298,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 578,
+ 128,
+ 589
+ ],
+ "score": 0.88,
+ "content": "1 0 ^ { - 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 129,
+ 577,
+ 298,
+ 590
+ ],
+ "score": 1.0,
+ "content": "and batch normalization. The dataset is",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 589,
+ 298,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 162,
+ 600
+ ],
+ "score": 1.0,
+ "content": "processed for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 162,
+ 589,
+ 184,
+ 600
+ ],
+ "score": 0.65,
+ "content": "6 0 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 590,
+ 298,
+ 600
+ ],
+ "score": 1.0,
+ "content": "iterations on a batch size of",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 298,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 599,
+ 298,
+ 613
+ ],
+ "score": 1.0,
+ "content": "256, distributed on 4GPUs. The initial learning",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 611,
+ 297,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 611,
+ 275,
+ 622
+ ],
+ "score": 1.0,
+ "content": "rate is 0.1, dropped by a factor of ten every",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 275,
+ 611,
+ 297,
+ 622
+ ],
+ "score": 0.54,
+ "content": "1 6 0 \\mathrm { k }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 622,
+ 297,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 622,
+ 297,
+ 633
+ ],
+ "score": 1.0,
+ "content": "iterations. The dataset was augmented accord-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 298,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 298,
+ 645
+ ],
+ "score": 1.0,
+ "content": "ing to Gomez et al. (2017). The images values",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 297,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 297,
+ 656
+ ],
+ "score": 1.0,
+ "content": "are mapped to [0, 1] while following geometric",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 297,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 297,
+ 668
+ ],
+ "score": 1.0,
+ "content": "transformations were applied: random scaling,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 297,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 297,
+ 678
+ ],
+ "score": 1.0,
+ "content": "random horizontal flipping, random cropping",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 676,
+ 298,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 676,
+ 136,
+ 689
+ ],
+ "score": 1.0,
+ "content": "of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 137,
+ 676,
+ 157,
+ 687
+ ],
+ "score": 0.87,
+ "content": "2 2 4 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 158,
+ 676,
+ 298,
+ 689
+ ],
+ "score": 1.0,
+ "content": ", and finally color distortions. No",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 297,
+ 699
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 297,
+ 699
+ ],
+ "score": 1.0,
+ "content": "other regularizations were incorporated into the",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 698,
+ 297,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 297,
+ 711
+ ],
+ "score": 1.0,
+ "content": "classification pipeline. At test time, we rescale",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 708,
+ 298,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 183,
+ 722
+ ],
+ "score": 1.0,
+ "content": "the image size to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 709,
+ 204,
+ 720
+ ],
+ "score": 0.85,
+ "content": "2 5 6 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 708,
+ 298,
+ 722
+ ],
+ "score": 1.0,
+ "content": "and perform a center",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 180,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 155,
+ 732
+ ],
+ "score": 1.0,
+ "content": "crop of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 156,
+ 720,
+ 176,
+ 731
+ ],
+ "score": 0.88,
+ "content": "2 2 4 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 721,
+ 180,
+ 732
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 39.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 308,
+ 539,
+ 500,
+ 691
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 308,
+ 539,
+ 500,
+ 691
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 539,
+ 500,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 539,
+ 500,
+ 691
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "9b84f3e11c0bdeb5dd1a42d4cbb16eb51d1e10b7988637944ff79b527e7a6073.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 54,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 308,
+ 539,
+ 500,
+ 552.8181818181819
+ ],
+ "spans": [],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 308,
+ 552.8181818181819,
+ 500,
+ 566.6363636363637
+ ],
+ "spans": [],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 308,
+ 566.6363636363637,
+ 500,
+ 580.4545454545456
+ ],
+ "spans": [],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 308,
+ 580.4545454545456,
+ 500,
+ 594.2727272727275
+ ],
+ "spans": [],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 308,
+ 594.2727272727275,
+ 500,
+ 608.0909090909093
+ ],
+ "spans": [],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 308,
+ 608.0909090909093,
+ 500,
+ 621.9090909090912
+ ],
+ "spans": [],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 308,
+ 621.9090909090912,
+ 500,
+ 635.7272727272731
+ ],
+ "spans": [],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 308,
+ 635.7272727272731,
+ 500,
+ 649.545454545455
+ ],
+ "spans": [],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 308,
+ 649.545454545455,
+ 500,
+ 663.3636363636368
+ ],
+ "spans": [],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 308,
+ 663.3636363636368,
+ 500,
+ 677.1818181818187
+ ],
+ "spans": [],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 308,
+ 677.1818181818187,
+ 500,
+ 691.0000000000006
+ ],
+ "spans": [],
+ "index": 59
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 303,
+ 697,
+ 504,
+ 720
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 696,
+ 504,
+ 709
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 696,
+ 425,
+ 709
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Training loss of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 698,
+ 430,
+ 707
+ ],
+ "score": 0.73,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 430,
+ 696,
+ 504,
+ 709
+ ],
+ "score": 1.0,
+ "content": "-RevNet (b), com-",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ },
+ {
+ "bbox": [
+ 303,
+ 708,
+ 445,
+ 720
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 708,
+ 445,
+ 720
+ ],
+ "score": 1.0,
+ "content": "pared to the ResNet, on ImageNet.",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ }
+ ],
+ "index": 60.5
+ }
+ ],
+ "index": 57.25
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published 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": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 162,
+ 79,
+ 449,
+ 144
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 162,
+ 79,
+ 449,
+ 144
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 162,
+ 79,
+ 449,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 79,
+ 449,
+ 144
+ ],
+ "score": 0.963,
+ "html": "| Architecture | Injective | Bijective | Top-1error | Parameters |
| ResNet | | | 24.7 | 26M |
| RevNet | = | | 25.2 | 28M |
| i-RevNet (a) | yes | = | 24.7 | 181M |
| i-RevNet (b) | yes | yes | 26.7 | 29M |
",
+ "type": "table",
+ "image_path": "251ee7588dce774e1181820f0000e4bc7c393edb175e76ba0ff406a4c8e79c70.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 162,
+ 79,
+ 449,
+ 100.66666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 162,
+ 100.66666666666667,
+ 449,
+ 122.33333333333334
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 162,
+ 122.33333333333334,
+ 449,
+ 144.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 156,
+ 503,
+ 179
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 505,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 505,
+ 169
+ ],
+ "score": 1.0,
+ "content": "Table 1: Comparison of different architectures trained on ILSVRC-2012, in terms of classification",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 168,
+ 251,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 168,
+ 251,
+ 180
+ ],
+ "score": 1.0,
+ "content": "accuracy and number of parameters",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 221,
+ 503,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 504,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 504,
+ 233
+ ],
+ "score": 1.0,
+ "content": "The final representation is spatially averaged and projected onto the 1000 classes after a ReLU non-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "linearity. We now discuss how we progressively decrease the spatial resolution, while increasing the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 243,
+ 334,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 318,
+ 257
+ ],
+ "score": 1.0,
+ "content": "number of channels per layer by use of the operators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 318,
+ 243,
+ 330,
+ 255
+ ],
+ "score": 0.89,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 330,
+ 243,
+ 334,
+ 257
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 221,
+ 505,
+ 257
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 259,
+ 505,
+ 304
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "We first describe the model (a), that consists of 56 layers which have been optimized to match",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 283
+ ],
+ "score": 1.0,
+ "content": "the performances of a RevNet or a ResNet with approximatively the same number of layers. In",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 282,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 282,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": "particular, we explain how we progressively decrease the spatial resolution, while increasing the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 336,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 320,
+ 306
+ ],
+ "score": 1.0,
+ "content": "number of channels per block by use of the operators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 321,
+ 293,
+ 332,
+ 305
+ ],
+ "score": 0.89,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 332,
+ 293,
+ 336,
+ 306
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 105,
+ 259,
+ 505,
+ 306
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 443
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 196,
+ 325
+ ],
+ "score": 1.0,
+ "content": "The splitting operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 310,
+ 205,
+ 322
+ ],
+ "score": 0.82,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 311,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "consists in a linear and injective embedding that downsamples by a factor",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 506,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 322,
+ 117,
+ 333
+ ],
+ "score": 0.76,
+ "content": "4 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 117,
+ 321,
+ 506,
+ 337
+ ],
+ "score": 1.0,
+ "content": "the spatial resolution by increasing the number of output channels from 48 to 96 by simply adding",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 333,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 333,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": "0. The latter permits to increase the initial layer size, and consequently, the size of the next layers as",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 104,
+ 344,
+ 504,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 344,
+ 403,
+ 358
+ ],
+ "score": 1.0,
+ "content": "performed in Gomez et al. (2017); it is thus not a bijective yet an injective",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 403,
+ 345,
+ 408,
+ 355
+ ],
+ "score": 0.61,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 408,
+ 344,
+ 482,
+ 358
+ ],
+ "score": 1.0,
+ "content": "-RevNet. At depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 345,
+ 504,
+ 357
+ ],
+ "score": 0.53,
+ "content": "j , { \\mathcal { S } } _ { j }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 356,
+ 505,
+ 368
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 505,
+ 368
+ ],
+ "score": 1.0,
+ "content": "allows us to reduce the number of computations while maintaining good classification performance.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 365,
+ 506,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 365,
+ 397,
+ 380
+ ],
+ "score": 1.0,
+ "content": "It will correspond to a downsampling operator respectively at the depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 367,
+ 457,
+ 378
+ ],
+ "score": 0.55,
+ "content": "3 j = 1 5 , 2 7 , 4 ",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 458,
+ 365,
+ 506,
+ 380
+ ],
+ "score": 1.0,
+ "content": "5 (3j as one",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 377,
+ 505,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 377,
+ 505,
+ 390
+ ],
+ "score": 1.0,
+ "content": "block corresponds to three layers), similar to a normal RevNet. The spatial resolution of these layers",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 104,
+ 387,
+ 506,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 387,
+ 193,
+ 401
+ ],
+ "score": 1.0,
+ "content": "is reduced by a factor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 388,
+ 204,
+ 399
+ ],
+ "score": 0.83,
+ "content": "2 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 387,
+ 506,
+ 401
+ ],
+ "score": 1.0,
+ "content": "while increasing the number of channels by a factor of 4 respectively to 48,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 399,
+ 506,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 399,
+ 506,
+ 412
+ ],
+ "score": 1.0,
+ "content": "192, 768 and 3072. Furthermore, it means that the corresponding spatial resolutions for an input of",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 409,
+ 506,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 409,
+ 124,
+ 423
+ ],
+ "score": 1.0,
+ "content": "size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 410,
+ 145,
+ 421
+ ],
+ "score": 0.87,
+ "content": "2 2 4 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 409,
+ 213,
+ 423
+ ],
+ "score": 1.0,
+ "content": "are respectively",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 214,
+ 410,
+ 304,
+ 421
+ ],
+ "score": 0.47,
+ "content": "1 1 2 ^ { 2 } , 5 6 ^ { 2 } , 2 8 ^ { 2 } , 1 4 ^ { 2 } , 7 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 305,
+ 409,
+ 506,
+ 423
+ ],
+ "score": 1.0,
+ "content": ". The total number of coefficients at each layer is",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 150,
+ 435
+ ],
+ "score": 1.0,
+ "content": "then about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 151,
+ 421,
+ 175,
+ 432
+ ],
+ "score": 0.81,
+ "content": "0 . 3 M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 420,
+ 280,
+ 435
+ ],
+ "score": 1.0,
+ "content": ". All the remaining blocks",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 422,
+ 292,
+ 433
+ ],
+ "score": 0.88,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 292,
+ 420,
+ 506,
+ 435
+ ],
+ "score": 1.0,
+ "content": "are kept fix to the identity as explained in the section",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 431,
+ 137,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 137,
+ 445
+ ],
+ "score": 1.0,
+ "content": "above.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 104,
+ 310,
+ 506,
+ 445
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 529
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "Architecture (b) is bijective, it consists of 300 layers (100 blocks), whose total numbers of parame-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 493,
+ 474
+ ],
+ "score": 1.0,
+ "content": "ters have been optimized to match those of a RevNet with 56 layers. Initially, the input is split via",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 460,
+ 501,
+ 472
+ ],
+ "score": 0.82,
+ "content": "\\tilde { \\cal S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 461,
+ 505,
+ 474
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 350,
+ 485
+ ],
+ "score": 1.0,
+ "content": "which corresponds to an invertible spatial downsampling of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 351,
+ 471,
+ 361,
+ 483
+ ],
+ "score": 0.85,
+ "content": "2 ^ { \\frac { 5 } { 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 362,
+ 471,
+ 505,
+ 485
+ ],
+ "score": 1.0,
+ "content": "that increases the number of chan-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 505,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 463,
+ 496
+ ],
+ "score": 1.0,
+ "content": "nels from 3 to 12. It thus keeps the dimension constant and permits building a bijective",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 463,
+ 484,
+ 468,
+ 494
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 468,
+ 483,
+ 505,
+ 496
+ ],
+ "score": 1.0,
+ "content": "-RevNet.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 168,
+ 507
+ ],
+ "score": 1.0,
+ "content": "Then, at depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 495,
+ 213,
+ 506
+ ],
+ "score": 0.5,
+ "content": "3 j = 3 , 2 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 493,
+ 396,
+ 507
+ ],
+ "score": 1.0,
+ "content": ", 69, 285, the spatial resolution is reduced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 396,
+ 494,
+ 407,
+ 505
+ ],
+ "score": 0.81,
+ "content": "2 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 407,
+ 493,
+ 423,
+ 507
+ ],
+ "score": 1.0,
+ "content": "via",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 495,
+ 435,
+ 507
+ ],
+ "score": 0.87,
+ "content": "S _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 493,
+ 506,
+ 507
+ ],
+ "score": 1.0,
+ "content": ". Contrary to the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 386,
+ 520
+ ],
+ "score": 1.0,
+ "content": "architecture (a), the dimensionality of each layer is constantly equal to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 387,
+ 506,
+ 422,
+ 517
+ ],
+ "score": 0.92,
+ "content": "3 \\times 2 2 4 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 505,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": ", until the final layer,",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 518,
+ 267,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 518,
+ 267,
+ 528
+ ],
+ "score": 1.0,
+ "content": "with channel sizes of 24, 96, 384, 1536.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 105,
+ 449,
+ 506,
+ 528
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 534,
+ 297,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 534,
+ 297,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 534,
+ 297,
+ 546
+ ],
+ "score": 1.0,
+ "content": "For both networks, the training on Imagenet",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 545,
+ 297,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 297,
+ 558
+ ],
+ "score": 1.0,
+ "content": "follows the same setup as Gomez et al. (2017).",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 556,
+ 297,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 556,
+ 297,
+ 568
+ ],
+ "score": 1.0,
+ "content": "We train with SGD and momentum of 0.9. We",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 567,
+ 298,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 567,
+ 221,
+ 579
+ ],
+ "score": 1.0,
+ "content": "regularized the model with a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 221,
+ 567,
+ 231,
+ 577
+ ],
+ "score": 0.83,
+ "content": "\\ell ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 567,
+ 298,
+ 579
+ ],
+ "score": 1.0,
+ "content": "weight decay of",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 298,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 578,
+ 128,
+ 589
+ ],
+ "score": 0.88,
+ "content": "1 0 ^ { - 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 129,
+ 577,
+ 298,
+ 590
+ ],
+ "score": 1.0,
+ "content": "and batch normalization. The dataset is",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 589,
+ 298,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 162,
+ 600
+ ],
+ "score": 1.0,
+ "content": "processed for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 162,
+ 589,
+ 184,
+ 600
+ ],
+ "score": 0.65,
+ "content": "6 0 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 590,
+ 298,
+ 600
+ ],
+ "score": 1.0,
+ "content": "iterations on a batch size of",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 298,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 599,
+ 298,
+ 613
+ ],
+ "score": 1.0,
+ "content": "256, distributed on 4GPUs. The initial learning",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 611,
+ 297,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 611,
+ 275,
+ 622
+ ],
+ "score": 1.0,
+ "content": "rate is 0.1, dropped by a factor of ten every",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 275,
+ 611,
+ 297,
+ 622
+ ],
+ "score": 0.54,
+ "content": "1 6 0 \\mathrm { k }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 622,
+ 297,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 622,
+ 297,
+ 633
+ ],
+ "score": 1.0,
+ "content": "iterations. The dataset was augmented accord-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 298,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 298,
+ 645
+ ],
+ "score": 1.0,
+ "content": "ing to Gomez et al. (2017). The images values",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 297,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 297,
+ 656
+ ],
+ "score": 1.0,
+ "content": "are mapped to [0, 1] while following geometric",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 297,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 297,
+ 668
+ ],
+ "score": 1.0,
+ "content": "transformations were applied: random scaling,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 297,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 297,
+ 678
+ ],
+ "score": 1.0,
+ "content": "random horizontal flipping, random cropping",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 676,
+ 298,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 676,
+ 136,
+ 689
+ ],
+ "score": 1.0,
+ "content": "of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 137,
+ 676,
+ 157,
+ 687
+ ],
+ "score": 0.87,
+ "content": "2 2 4 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 158,
+ 676,
+ 298,
+ 689
+ ],
+ "score": 1.0,
+ "content": ", and finally color distortions. No",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 297,
+ 699
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 297,
+ 699
+ ],
+ "score": 1.0,
+ "content": "other regularizations were incorporated into the",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 698,
+ 297,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 297,
+ 711
+ ],
+ "score": 1.0,
+ "content": "classification pipeline. At test time, we rescale",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 708,
+ 298,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 183,
+ 722
+ ],
+ "score": 1.0,
+ "content": "the image size to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 709,
+ 204,
+ 720
+ ],
+ "score": 0.85,
+ "content": "2 5 6 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 708,
+ 298,
+ 722
+ ],
+ "score": 1.0,
+ "content": "and perform a center",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 180,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 155,
+ 732
+ ],
+ "score": 1.0,
+ "content": "crop of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 156,
+ 720,
+ 176,
+ 731
+ ],
+ "score": 0.88,
+ "content": "2 2 4 ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 721,
+ 180,
+ 732
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 39.5,
+ "bbox_fs": [
+ 105,
+ 534,
+ 298,
+ 732
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 308,
+ 539,
+ 500,
+ 691
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 308,
+ 539,
+ 500,
+ 691
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 539,
+ 500,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 539,
+ 500,
+ 691
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "9b84f3e11c0bdeb5dd1a42d4cbb16eb51d1e10b7988637944ff79b527e7a6073.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 54,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 308,
+ 539,
+ 500,
+ 552.8181818181819
+ ],
+ "spans": [],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 308,
+ 552.8181818181819,
+ 500,
+ 566.6363636363637
+ ],
+ "spans": [],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 308,
+ 566.6363636363637,
+ 500,
+ 580.4545454545456
+ ],
+ "spans": [],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 308,
+ 580.4545454545456,
+ 500,
+ 594.2727272727275
+ ],
+ "spans": [],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 308,
+ 594.2727272727275,
+ 500,
+ 608.0909090909093
+ ],
+ "spans": [],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 308,
+ 608.0909090909093,
+ 500,
+ 621.9090909090912
+ ],
+ "spans": [],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 308,
+ 621.9090909090912,
+ 500,
+ 635.7272727272731
+ ],
+ "spans": [],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 308,
+ 635.7272727272731,
+ 500,
+ 649.545454545455
+ ],
+ "spans": [],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 308,
+ 649.545454545455,
+ 500,
+ 663.3636363636368
+ ],
+ "spans": [],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 308,
+ 663.3636363636368,
+ 500,
+ 677.1818181818187
+ ],
+ "spans": [],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 308,
+ 677.1818181818187,
+ 500,
+ 691.0000000000006
+ ],
+ "spans": [],
+ "index": 59
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 303,
+ 697,
+ 504,
+ 720
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 696,
+ 504,
+ 709
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 696,
+ 425,
+ 709
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Training loss of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 698,
+ 430,
+ 707
+ ],
+ "score": 0.73,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 430,
+ 696,
+ 504,
+ 709
+ ],
+ "score": 1.0,
+ "content": "-RevNet (b), com-",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ },
+ {
+ "bbox": [
+ 303,
+ 708,
+ 445,
+ 720
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 708,
+ 445,
+ 720
+ ],
+ "score": 1.0,
+ "content": "pared to the ResNet, on ImageNet.",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ }
+ ],
+ "index": 60.5
+ }
+ ],
+ "index": 57.25
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 182
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 389,
+ 95
+ ],
+ "score": 1.0,
+ "content": "We report the training loss (i.e. Cross entropy) curves in Figure 3 of our",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 390,
+ 83,
+ 394,
+ 93
+ ],
+ "score": 0.51,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 395,
+ 81,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "-RevNet (b) and the ResNet",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "baseline, displayed is a moving average over 100 iterations. Observe that the decrease of both",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "training-losses are very similar which indicates that the constraint of invertibility does not interfere",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 114,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 114,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "negatively with the learning process. However, we observed one third longer wall-clock times for",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 126,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 111,
+ 136
+ ],
+ "score": 0.66,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 111,
+ 126,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "-RevNets compared to plain RevNets because the channel size becomes larger. The Table 1 reports",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 137,
+ 505,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 137,
+ 207,
+ 149
+ ],
+ "score": 1.0,
+ "content": "the performances of our",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 138,
+ 212,
+ 147
+ ],
+ "score": 0.73,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 137,
+ 505,
+ 149
+ ],
+ "score": 1.0,
+ "content": "-RevNets, with comparable RevNet and ResNet. First, we compare the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 147,
+ 506,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 149,
+ 111,
+ 158
+ ],
+ "score": 0.69,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 111,
+ 147,
+ 506,
+ 162
+ ],
+ "score": 1.0,
+ "content": "-RevNet (a) with the RevNet and ResNet. Indeed, those CNNs have the same number of layers,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 138,
+ 172
+ ],
+ "score": 1.0,
+ "content": "and the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 138,
+ 160,
+ 143,
+ 169
+ ],
+ "score": 0.53,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 143,
+ 158,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "-RevNet (a) increases the channel width of the initial layer as done in Gomez et al. (2017).",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 169,
+ 482,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 482,
+ 184
+ ],
+ "score": 1.0,
+ "content": "The drawback of this technique is that the kernel sizes will be larger for all subsequent layers.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 187,
+ 505,
+ 276
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 506,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 125,
+ 199
+ ],
+ "score": 1.0,
+ "content": "The",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 188,
+ 130,
+ 197
+ ],
+ "score": 0.66,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 131,
+ 187,
+ 506,
+ 199
+ ],
+ "score": 1.0,
+ "content": "-RevNet (a) has about 6 times more parameters than a RevNet and a ResNet but leads to a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 506,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 394,
+ 210
+ ],
+ "score": 1.0,
+ "content": "similar accuracy on the validation set of ImageNet. On the contrary, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 394,
+ 199,
+ 399,
+ 208
+ ],
+ "score": 0.59,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 399,
+ 199,
+ 506,
+ 210
+ ],
+ "score": 1.0,
+ "content": "-RevNet (b) is designed to",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "have roughly the same number of parameters as the RevNet and ResNet, while being bijective. Its",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 197,
+ 232
+ ],
+ "score": 1.0,
+ "content": "accuracy decreases by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 220,
+ 219,
+ 231
+ ],
+ "score": 0.86,
+ "content": "1 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 220,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "absolute percent on ImageNet compared to the RevNet baseline, which",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 243
+ ],
+ "score": 1.0,
+ "content": "is not surprising because the number of channels was not drastically increased in the earlier layers",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 242,
+ 506,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 242,
+ 506,
+ 254
+ ],
+ "score": 1.0,
+ "content": "as done in the baselines (Gomez et al., 2017; Krizhevsky et al., 2012; He et al., 2016); we did not",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "score": 1.0,
+ "content": "explore wide ranges of hyper-parameters, thus the gap between (a) and (b) can likely be reduced",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 264,
+ 221,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 264,
+ 221,
+ 277
+ ],
+ "score": 1.0,
+ "content": "with additional engineering.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 12.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 292,
+ 265,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 267,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 267,
+ 305
+ ],
+ "score": 1.0,
+ "content": "4 ANALYSIS OF THE INVERSE",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 316,
+ 505,
+ 360
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 316,
+ 505,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 316,
+ 253,
+ 329
+ ],
+ "score": 1.0,
+ "content": "We now analyze the representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 317,
+ 262,
+ 326
+ ],
+ "score": 0.83,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 316,
+ 419,
+ 329
+ ],
+ "score": 1.0,
+ "content": "built by our bijective neural network",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 317,
+ 424,
+ 326
+ ],
+ "score": 0.76,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 424,
+ 316,
+ 505,
+ 329
+ ],
+ "score": 1.0,
+ "content": "-RevNet (b) and its",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 505,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 137,
+ 340
+ ],
+ "score": 1.0,
+ "content": "inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 137,
+ 327,
+ 156,
+ 337
+ ],
+ "score": 0.92,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 326,
+ 403,
+ 340
+ ],
+ "score": 1.0,
+ "content": ", as trained on ILSVRC-2012. We first explain why obtaining",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 403,
+ 327,
+ 422,
+ 338
+ ],
+ "score": 0.9,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 326,
+ 505,
+ 340
+ ],
+ "score": 1.0,
+ "content": "is challenging, even",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 337,
+ 505,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 337,
+ 505,
+ 351
+ ],
+ "score": 1.0,
+ "content": "locally. We then discuss the reconstruction, while displaying in the image space linear interpolations",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 349,
+ 207,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 207,
+ 362
+ ],
+ "score": 1.0,
+ "content": "between representations.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 374,
+ 275,
+ 385
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 277,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 277,
+ 387
+ ],
+ "score": 1.0,
+ "content": "4.1 AN ILL-CONDITIONED INVERSION",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 394,
+ 296,
+ 548
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 297,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 297,
+ 406
+ ],
+ "score": 1.0,
+ "content": "In the previous section, we have described the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 297,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 406,
+ 111,
+ 415
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 111,
+ 404,
+ 297,
+ 418
+ ],
+ "score": 1.0,
+ "content": "-RevNet architecture, that permits defining a",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 416,
+ 297,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 416,
+ 297,
+ 429
+ ],
+ "score": 1.0,
+ "content": "deep network with an explicit inverse. We ex-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 297,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 297,
+ 441
+ ],
+ "score": 1.0,
+ "content": "plain now why this is normally difficult, by",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 297,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 297,
+ 450
+ ],
+ "score": 1.0,
+ "content": "studying its local inversion. We study the lo-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 449,
+ 295,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 449,
+ 208,
+ 461
+ ],
+ "score": 1.0,
+ "content": "cal stability of a network",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 450,
+ 216,
+ 460
+ ],
+ "score": 0.82,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 217,
+ 449,
+ 276,
+ 461
+ ],
+ "score": 1.0,
+ "content": "and its inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 449,
+ 295,
+ 460
+ ],
+ "score": 0.89,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 461,
+ 297,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 297,
+ 471
+ ],
+ "score": 1.0,
+ "content": "w.r.t. to its input, which means that we will",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 472,
+ 297,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 472,
+ 297,
+ 483
+ ],
+ "score": 1.0,
+ "content": "quantify locally the variations of the network",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 482,
+ 297,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 297,
+ 493
+ ],
+ "score": 1.0,
+ "content": "and its inverse w.r.t. to small variations of an",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 298,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 147,
+ 505
+ ],
+ "score": 1.0,
+ "content": "input. As",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 147,
+ 493,
+ 156,
+ 504
+ ],
+ "score": 0.82,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 493,
+ 298,
+ 505
+ ],
+ "score": 1.0,
+ "content": "is differentiable (and its inverse as",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 504,
+ 297,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 297,
+ 516
+ ],
+ "score": 1.0,
+ "content": "well), an equivalent way to perform this study",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 515,
+ 297,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 297,
+ 527
+ ],
+ "score": 1.0,
+ "content": "is to analyze the singular values of the differ-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 526,
+ 297,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 526,
+ 131,
+ 538
+ ],
+ "score": 1.0,
+ "content": "ential",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 526,
+ 146,
+ 536
+ ],
+ "score": 0.89,
+ "content": "\\partial \\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 526,
+ 234,
+ 538
+ ],
+ "score": 1.0,
+ "content": "at some point, as for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 526,
+ 257,
+ 538
+ ],
+ "score": 0.92,
+ "content": "( a , b )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 257,
+ 526,
+ 297,
+ 538
+ ],
+ "score": 1.0,
+ "content": "close the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 537,
+ 174,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 537,
+ 174,
+ 550
+ ],
+ "score": 1.0,
+ "content": "following holds:",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 151,
+ 553,
+ 252,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 151,
+ 553,
+ 252,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 151,
+ 553,
+ 252,
+ 567
+ ],
+ "score": 0.9,
+ "content": "\\Phi \\boldsymbol { a } \\approx \\Phi \\boldsymbol { b } + \\partial \\Phi _ { b } ( \\boldsymbol { a } - \\boldsymbol { b } ) .",
+ "type": "interline_equation",
+ "image_path": "15d1e4ab319c8f85cfdb3eb8af46b5fd022df59920c092b7db8b9ba844949389.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 37,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 151,
+ 553,
+ 252,
+ 567
+ ],
+ "spans": [],
+ "index": 37
+ }
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 307,
+ 397,
+ 499,
+ 539
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 307,
+ 397,
+ 499,
+ 539
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 397,
+ 499,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 397,
+ 499,
+ 539
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "ce28b51d37c721b4a781c7d9d9b2bcd0ff508b71714b9075ab8d17c86d3d3d9a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 42.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 307,
+ 397,
+ 499,
+ 411.2
+ ],
+ "spans": [],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 307,
+ 411.2,
+ 499,
+ 425.4
+ ],
+ "spans": [],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 307,
+ 425.4,
+ 499,
+ 439.59999999999997
+ ],
+ "spans": [],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 307,
+ 439.59999999999997,
+ 499,
+ 453.79999999999995
+ ],
+ "spans": [],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 307,
+ 453.79999999999995,
+ 499,
+ 467.99999999999994
+ ],
+ "spans": [],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 307,
+ 467.99999999999994,
+ 499,
+ 482.19999999999993
+ ],
+ "spans": [],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 307,
+ 482.19999999999993,
+ 499,
+ 496.3999999999999
+ ],
+ "spans": [],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 307,
+ 496.3999999999999,
+ 499,
+ 510.5999999999999
+ ],
+ "spans": [],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 307,
+ 510.5999999999999,
+ 499,
+ 524.8
+ ],
+ "spans": [],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 307,
+ 524.8,
+ 499,
+ 539.0
+ ],
+ "spans": [],
+ "index": 47
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 304,
+ 543,
+ 504,
+ 565
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 542,
+ 506,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 542,
+ 506,
+ 555
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Normalized sorted singular values of",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 304,
+ 552,
+ 329,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 554,
+ 324,
+ 565
+ ],
+ "score": 0.82,
+ "content": "\\partial \\Phi _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 324,
+ 552,
+ 329,
+ 568
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 48.5
+ }
+ ],
+ "index": 45.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 572,
+ 504,
+ 594
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 570,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 570,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "Ideally, a well-conditioned operator has all its singular values constant equal to 1, for instance as",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 583,
+ 340,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 583,
+ 340,
+ 595
+ ],
+ "score": 1.0,
+ "content": "achieved by the isometric operators of Cisse et al. (2017).",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 50.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 506,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 273,
+ 613
+ ],
+ "score": 1.0,
+ "content": "In our numerical application to an image",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 274,
+ 602,
+ 281,
+ 610
+ ],
+ "score": 0.45,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 600,
+ 285,
+ 613
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 600,
+ 304,
+ 611
+ ],
+ "score": 0.75,
+ "content": "\\partial \\Phi _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 304,
+ 600,
+ 506,
+ 613
+ ],
+ "score": 1.0,
+ "content": "corresponds to a very large matrix (square of the",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 611,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 611,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "number of coefficients of the image at least) whose computations are expensive. Figure 4 corre-",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 466,
+ 635
+ ],
+ "score": 1.0,
+ "content": "sponds to the singular values of the differential (i.e. the square roots of the eigen values of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 622,
+ 498,
+ 632
+ ],
+ "score": 0.9,
+ "content": "\\partial \\Phi ^ { * } \\partial \\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 499,
+ 622,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "),",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "in decreasing order, for a given natural image from ImageNet. The example we plot is typical of",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 106,
+ 643,
+ 505,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 169,
+ 656
+ ],
+ "score": 1.0,
+ "content": "the behavior of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 644,
+ 183,
+ 654
+ ],
+ "score": 0.85,
+ "content": "\\partial \\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 644,
+ 392,
+ 656
+ ],
+ "score": 1.0,
+ "content": ". Observe there is a fast decay: numerically, the first",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 643,
+ 408,
+ 654
+ ],
+ "score": 0.9,
+ "content": "\\mathrm { 1 0 ^ { 3 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 408,
+ 644,
+ 425,
+ 656
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 426,
+ 644,
+ 442,
+ 654
+ ],
+ "score": 0.88,
+ "content": "1 0 ^ { \\bar { 4 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 442,
+ 644,
+ 505,
+ 656
+ ],
+ "score": 1.0,
+ "content": "singular values",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 232,
+ 668
+ ],
+ "score": 1.0,
+ "content": "are responsible respectively for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 233,
+ 655,
+ 253,
+ 666
+ ],
+ "score": 0.89,
+ "content": "8 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 654,
+ 270,
+ 668
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 270,
+ 655,
+ 290,
+ 666
+ ],
+ "score": 0.89,
+ "content": "9 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 654,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "of the cumulated energy (i.e. sum of squared singular",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 199,
+ 678
+ ],
+ "score": 1.0,
+ "content": "values). This indicates",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 666,
+ 208,
+ 676
+ ],
+ "score": 0.83,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "linearizes the space locally in a considerably smaller space in comparison",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 442,
+ 689
+ ],
+ "score": 1.0,
+ "content": "to the original input dimension. However, the dimensionality is still quite large (i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 442,
+ 677,
+ 465,
+ 688
+ ],
+ "score": 0.83,
+ "content": "> 1 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 465,
+ 676,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": ") and thus",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 504,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 192,
+ 700
+ ],
+ "score": 1.0,
+ "content": "we can not infer that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 688,
+ 201,
+ 698
+ ],
+ "score": 0.82,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 202,
+ 688,
+ 495,
+ 700
+ ],
+ "score": 1.0,
+ "content": "lays locally in a low-dimensional manifold. It also proves that inversing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 688,
+ 504,
+ 698
+ ],
+ "score": 0.8,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 60
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 710
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 710
+ ],
+ "score": 1.0,
+ "content": "is difficult and is an ill-conditioned problem. Thus obtaining implicitly this inverse would be a chal-",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ },
+ {
+ "bbox": [
+ 104,
+ 709,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 709,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "lenging task that we avoided, thanks to the formal reconstruction algorithm provided by Subsection",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 126,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 126,
+ 732
+ ],
+ "score": 1.0,
+ "content": "3.1.",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ }
+ ],
+ "index": 57.5
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published 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": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 182
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 389,
+ 95
+ ],
+ "score": 1.0,
+ "content": "We report the training loss (i.e. Cross entropy) curves in Figure 3 of our",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 390,
+ 83,
+ 394,
+ 93
+ ],
+ "score": 0.51,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 395,
+ 81,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "-RevNet (b) and the ResNet",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "baseline, displayed is a moving average over 100 iterations. Observe that the decrease of both",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "training-losses are very similar which indicates that the constraint of invertibility does not interfere",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 114,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 114,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "negatively with the learning process. However, we observed one third longer wall-clock times for",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 126,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 111,
+ 136
+ ],
+ "score": 0.66,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 111,
+ 126,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "-RevNets compared to plain RevNets because the channel size becomes larger. The Table 1 reports",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 137,
+ 505,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 137,
+ 207,
+ 149
+ ],
+ "score": 1.0,
+ "content": "the performances of our",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 138,
+ 212,
+ 147
+ ],
+ "score": 0.73,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 137,
+ 505,
+ 149
+ ],
+ "score": 1.0,
+ "content": "-RevNets, with comparable RevNet and ResNet. First, we compare the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 147,
+ 506,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 149,
+ 111,
+ 158
+ ],
+ "score": 0.69,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 111,
+ 147,
+ 506,
+ 162
+ ],
+ "score": 1.0,
+ "content": "-RevNet (a) with the RevNet and ResNet. Indeed, those CNNs have the same number of layers,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 138,
+ 172
+ ],
+ "score": 1.0,
+ "content": "and the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 138,
+ 160,
+ 143,
+ 169
+ ],
+ "score": 0.53,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 143,
+ 158,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "-RevNet (a) increases the channel width of the initial layer as done in Gomez et al. (2017).",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 169,
+ 482,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 482,
+ 184
+ ],
+ "score": 1.0,
+ "content": "The drawback of this technique is that the kernel sizes will be larger for all subsequent layers.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 105,
+ 81,
+ 506,
+ 184
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 187,
+ 505,
+ 276
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 506,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 125,
+ 199
+ ],
+ "score": 1.0,
+ "content": "The",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 188,
+ 130,
+ 197
+ ],
+ "score": 0.66,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 131,
+ 187,
+ 506,
+ 199
+ ],
+ "score": 1.0,
+ "content": "-RevNet (a) has about 6 times more parameters than a RevNet and a ResNet but leads to a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 506,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 394,
+ 210
+ ],
+ "score": 1.0,
+ "content": "similar accuracy on the validation set of ImageNet. On the contrary, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 394,
+ 199,
+ 399,
+ 208
+ ],
+ "score": 0.59,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 399,
+ 199,
+ 506,
+ 210
+ ],
+ "score": 1.0,
+ "content": "-RevNet (b) is designed to",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "have roughly the same number of parameters as the RevNet and ResNet, while being bijective. Its",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 197,
+ 232
+ ],
+ "score": 1.0,
+ "content": "accuracy decreases by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 220,
+ 219,
+ 231
+ ],
+ "score": 0.86,
+ "content": "1 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 220,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "absolute percent on ImageNet compared to the RevNet baseline, which",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 243
+ ],
+ "score": 1.0,
+ "content": "is not surprising because the number of channels was not drastically increased in the earlier layers",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 242,
+ 506,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 242,
+ 506,
+ 254
+ ],
+ "score": 1.0,
+ "content": "as done in the baselines (Gomez et al., 2017; Krizhevsky et al., 2012; He et al., 2016); we did not",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "score": 1.0,
+ "content": "explore wide ranges of hyper-parameters, thus the gap between (a) and (b) can likely be reduced",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 264,
+ 221,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 264,
+ 221,
+ 277
+ ],
+ "score": 1.0,
+ "content": "with additional engineering.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 105,
+ 187,
+ 506,
+ 277
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 292,
+ 265,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 267,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 267,
+ 305
+ ],
+ "score": 1.0,
+ "content": "4 ANALYSIS OF THE INVERSE",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 316,
+ 505,
+ 360
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 316,
+ 505,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 316,
+ 253,
+ 329
+ ],
+ "score": 1.0,
+ "content": "We now analyze the representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 317,
+ 262,
+ 326
+ ],
+ "score": 0.83,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 316,
+ 419,
+ 329
+ ],
+ "score": 1.0,
+ "content": "built by our bijective neural network",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 317,
+ 424,
+ 326
+ ],
+ "score": 0.76,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 424,
+ 316,
+ 505,
+ 329
+ ],
+ "score": 1.0,
+ "content": "-RevNet (b) and its",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 505,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 137,
+ 340
+ ],
+ "score": 1.0,
+ "content": "inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 137,
+ 327,
+ 156,
+ 337
+ ],
+ "score": 0.92,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 326,
+ 403,
+ 340
+ ],
+ "score": 1.0,
+ "content": ", as trained on ILSVRC-2012. We first explain why obtaining",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 403,
+ 327,
+ 422,
+ 338
+ ],
+ "score": 0.9,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 326,
+ 505,
+ 340
+ ],
+ "score": 1.0,
+ "content": "is challenging, even",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 337,
+ 505,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 337,
+ 505,
+ 351
+ ],
+ "score": 1.0,
+ "content": "locally. We then discuss the reconstruction, while displaying in the image space linear interpolations",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 349,
+ 207,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 207,
+ 362
+ ],
+ "score": 1.0,
+ "content": "between representations.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19.5,
+ "bbox_fs": [
+ 105,
+ 316,
+ 505,
+ 362
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 374,
+ 275,
+ 385
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 277,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 277,
+ 387
+ ],
+ "score": 1.0,
+ "content": "4.1 AN ILL-CONDITIONED INVERSION",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 394,
+ 296,
+ 548
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 297,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 297,
+ 406
+ ],
+ "score": 1.0,
+ "content": "In the previous section, we have described the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 297,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 406,
+ 111,
+ 415
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 111,
+ 404,
+ 297,
+ 418
+ ],
+ "score": 1.0,
+ "content": "-RevNet architecture, that permits defining a",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 416,
+ 297,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 416,
+ 297,
+ 429
+ ],
+ "score": 1.0,
+ "content": "deep network with an explicit inverse. We ex-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 297,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 297,
+ 441
+ ],
+ "score": 1.0,
+ "content": "plain now why this is normally difficult, by",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 297,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 297,
+ 450
+ ],
+ "score": 1.0,
+ "content": "studying its local inversion. We study the lo-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 449,
+ 295,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 449,
+ 208,
+ 461
+ ],
+ "score": 1.0,
+ "content": "cal stability of a network",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 450,
+ 216,
+ 460
+ ],
+ "score": 0.82,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 217,
+ 449,
+ 276,
+ 461
+ ],
+ "score": 1.0,
+ "content": "and its inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 449,
+ 295,
+ 460
+ ],
+ "score": 0.89,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 461,
+ 297,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 297,
+ 471
+ ],
+ "score": 1.0,
+ "content": "w.r.t. to its input, which means that we will",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 472,
+ 297,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 472,
+ 297,
+ 483
+ ],
+ "score": 1.0,
+ "content": "quantify locally the variations of the network",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 482,
+ 297,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 297,
+ 493
+ ],
+ "score": 1.0,
+ "content": "and its inverse w.r.t. to small variations of an",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 298,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 147,
+ 505
+ ],
+ "score": 1.0,
+ "content": "input. As",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 147,
+ 493,
+ 156,
+ 504
+ ],
+ "score": 0.82,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 493,
+ 298,
+ 505
+ ],
+ "score": 1.0,
+ "content": "is differentiable (and its inverse as",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 504,
+ 297,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 297,
+ 516
+ ],
+ "score": 1.0,
+ "content": "well), an equivalent way to perform this study",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 515,
+ 297,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 297,
+ 527
+ ],
+ "score": 1.0,
+ "content": "is to analyze the singular values of the differ-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 526,
+ 297,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 526,
+ 131,
+ 538
+ ],
+ "score": 1.0,
+ "content": "ential",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 526,
+ 146,
+ 536
+ ],
+ "score": 0.89,
+ "content": "\\partial \\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 526,
+ 234,
+ 538
+ ],
+ "score": 1.0,
+ "content": "at some point, as for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 526,
+ 257,
+ 538
+ ],
+ "score": 0.92,
+ "content": "( a , b )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 257,
+ 526,
+ 297,
+ 538
+ ],
+ "score": 1.0,
+ "content": "close the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 537,
+ 174,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 537,
+ 174,
+ 550
+ ],
+ "score": 1.0,
+ "content": "following holds:",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 105,
+ 394,
+ 298,
+ 550
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 151,
+ 553,
+ 252,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 151,
+ 553,
+ 252,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 151,
+ 553,
+ 252,
+ 567
+ ],
+ "score": 0.9,
+ "content": "\\Phi \\boldsymbol { a } \\approx \\Phi \\boldsymbol { b } + \\partial \\Phi _ { b } ( \\boldsymbol { a } - \\boldsymbol { b } ) .",
+ "type": "interline_equation",
+ "image_path": "15d1e4ab319c8f85cfdb3eb8af46b5fd022df59920c092b7db8b9ba844949389.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 37,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 151,
+ 553,
+ 252,
+ 567
+ ],
+ "spans": [],
+ "index": 37
+ }
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 307,
+ 397,
+ 499,
+ 539
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 307,
+ 397,
+ 499,
+ 539
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 397,
+ 499,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 397,
+ 499,
+ 539
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "ce28b51d37c721b4a781c7d9d9b2bcd0ff508b71714b9075ab8d17c86d3d3d9a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 42.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 307,
+ 397,
+ 499,
+ 411.2
+ ],
+ "spans": [],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 307,
+ 411.2,
+ 499,
+ 425.4
+ ],
+ "spans": [],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 307,
+ 425.4,
+ 499,
+ 439.59999999999997
+ ],
+ "spans": [],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 307,
+ 439.59999999999997,
+ 499,
+ 453.79999999999995
+ ],
+ "spans": [],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 307,
+ 453.79999999999995,
+ 499,
+ 467.99999999999994
+ ],
+ "spans": [],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 307,
+ 467.99999999999994,
+ 499,
+ 482.19999999999993
+ ],
+ "spans": [],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 307,
+ 482.19999999999993,
+ 499,
+ 496.3999999999999
+ ],
+ "spans": [],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 307,
+ 496.3999999999999,
+ 499,
+ 510.5999999999999
+ ],
+ "spans": [],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 307,
+ 510.5999999999999,
+ 499,
+ 524.8
+ ],
+ "spans": [],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 307,
+ 524.8,
+ 499,
+ 539.0
+ ],
+ "spans": [],
+ "index": 47
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 304,
+ 543,
+ 504,
+ 565
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 542,
+ 506,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 542,
+ 506,
+ 555
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Normalized sorted singular values of",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 304,
+ 552,
+ 329,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 554,
+ 324,
+ 565
+ ],
+ "score": 0.82,
+ "content": "\\partial \\Phi _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 324,
+ 552,
+ 329,
+ 568
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 48.5
+ }
+ ],
+ "index": 45.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 572,
+ 504,
+ 594
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 570,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 570,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "Ideally, a well-conditioned operator has all its singular values constant equal to 1, for instance as",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 583,
+ 340,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 583,
+ 340,
+ 595
+ ],
+ "score": 1.0,
+ "content": "achieved by the isometric operators of Cisse et al. (2017).",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 50.5,
+ "bbox_fs": [
+ 105,
+ 570,
+ 506,
+ 595
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 506,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 273,
+ 613
+ ],
+ "score": 1.0,
+ "content": "In our numerical application to an image",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 274,
+ 602,
+ 281,
+ 610
+ ],
+ "score": 0.45,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 600,
+ 285,
+ 613
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 600,
+ 304,
+ 611
+ ],
+ "score": 0.75,
+ "content": "\\partial \\Phi _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 304,
+ 600,
+ 506,
+ 613
+ ],
+ "score": 1.0,
+ "content": "corresponds to a very large matrix (square of the",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 611,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 611,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "number of coefficients of the image at least) whose computations are expensive. Figure 4 corre-",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 466,
+ 635
+ ],
+ "score": 1.0,
+ "content": "sponds to the singular values of the differential (i.e. the square roots of the eigen values of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 622,
+ 498,
+ 632
+ ],
+ "score": 0.9,
+ "content": "\\partial \\Phi ^ { * } \\partial \\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 499,
+ 622,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "),",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "in decreasing order, for a given natural image from ImageNet. The example we plot is typical of",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 106,
+ 643,
+ 505,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 169,
+ 656
+ ],
+ "score": 1.0,
+ "content": "the behavior of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 644,
+ 183,
+ 654
+ ],
+ "score": 0.85,
+ "content": "\\partial \\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 644,
+ 392,
+ 656
+ ],
+ "score": 1.0,
+ "content": ". Observe there is a fast decay: numerically, the first",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 643,
+ 408,
+ 654
+ ],
+ "score": 0.9,
+ "content": "\\mathrm { 1 0 ^ { 3 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 408,
+ 644,
+ 425,
+ 656
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 426,
+ 644,
+ 442,
+ 654
+ ],
+ "score": 0.88,
+ "content": "1 0 ^ { \\bar { 4 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 442,
+ 644,
+ 505,
+ 656
+ ],
+ "score": 1.0,
+ "content": "singular values",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 232,
+ 668
+ ],
+ "score": 1.0,
+ "content": "are responsible respectively for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 233,
+ 655,
+ 253,
+ 666
+ ],
+ "score": 0.89,
+ "content": "8 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 654,
+ 270,
+ 668
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 270,
+ 655,
+ 290,
+ 666
+ ],
+ "score": 0.89,
+ "content": "9 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 654,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "of the cumulated energy (i.e. sum of squared singular",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 199,
+ 678
+ ],
+ "score": 1.0,
+ "content": "values). This indicates",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 666,
+ 208,
+ 676
+ ],
+ "score": 0.83,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "linearizes the space locally in a considerably smaller space in comparison",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 442,
+ 689
+ ],
+ "score": 1.0,
+ "content": "to the original input dimension. However, the dimensionality is still quite large (i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 442,
+ 677,
+ 465,
+ 688
+ ],
+ "score": 0.83,
+ "content": "> 1 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 465,
+ 676,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": ") and thus",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 504,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 192,
+ 700
+ ],
+ "score": 1.0,
+ "content": "we can not infer that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 688,
+ 201,
+ 698
+ ],
+ "score": 0.82,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 202,
+ 688,
+ 495,
+ 700
+ ],
+ "score": 1.0,
+ "content": "lays locally in a low-dimensional manifold. It also proves that inversing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 688,
+ 504,
+ 698
+ ],
+ "score": 0.8,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 60
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 710
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 710
+ ],
+ "score": 1.0,
+ "content": "is difficult and is an ill-conditioned problem. Thus obtaining implicitly this inverse would be a chal-",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ },
+ {
+ "bbox": [
+ 104,
+ 709,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 709,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "lenging task that we avoided, thanks to the formal reconstruction algorithm provided by Subsection",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 126,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 126,
+ 732
+ ],
+ "score": 1.0,
+ "content": "3.1.",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ }
+ ],
+ "index": 57.5,
+ "bbox_fs": [
+ 104,
+ 600,
+ 506,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 64,
+ 505,
+ 288
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 64,
+ 505,
+ 288
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 64,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 64,
+ 505,
+ 288
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "a6fe6b5bd4ff7c42fa5cbd145f44da0d3e48b046022b4d75069d5e7eefe3144c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 64,
+ 505,
+ 138.66666666666669
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 138.66666666666669,
+ 505,
+ 213.33333333333337
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 213.33333333333337,
+ 505,
+ 288.00000000000006
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 297,
+ 503,
+ 320
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 506,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 365,
+ 311
+ ],
+ "score": 1.0,
+ "content": "Figure 5: This graphic displays several reconstructed sequences",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 365,
+ 297,
+ 388,
+ 309
+ ],
+ "score": 0.92,
+ "content": "\\{ x ^ { t } \\} _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 388,
+ 295,
+ 506,
+ 311
+ ],
+ "score": 1.0,
+ "content": ". The left image corresponds",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 307,
+ 235,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 117,
+ 321
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 308,
+ 128,
+ 318
+ ],
+ "score": 0.87,
+ "content": "{ \\bar { \\mathbf { \\Gamma } } } _ { x } 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 128,
+ 307,
+ 219,
+ 321
+ ],
+ "score": 1.0,
+ "content": "and the right image to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 308,
+ 230,
+ 318
+ ],
+ "score": 0.87,
+ "content": "x ^ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 307,
+ 235,
+ 321
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 343,
+ 339,
+ 353
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 340,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 340,
+ 354
+ ],
+ "score": 1.0,
+ "content": "4.2 LINEAR INTERPOLATION AND RECONSTRUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 363,
+ 505,
+ 419
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 364,
+ 505,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 364,
+ 505,
+ 376
+ ],
+ "score": 1.0,
+ "content": "Visualizing or understanding the important directions in the representation of inner layers of a CNN,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 375,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 375,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": "and in particular, the final layer is complex because typically the cascade is either not invertible or",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 385,
+ 506,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 506,
+ 398
+ ],
+ "score": 1.0,
+ "content": "unstable. One approach to reconstruct from an output layer consists in finding the input image that",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "matches the activation through via gradient descent. However, this technique leads only to a partial",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 407,
+ 338,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 407,
+ 338,
+ 419
+ ],
+ "score": 1.0,
+ "content": "or informal reconstruction (Mahendran & Vedaldi, 2015).",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 479
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "Another method consists in embedding the representation in a lower dimensional space and com-",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "paring the common attributes of nearest neighbors (Szegedy et al., 2013). It is also possible to train",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 104,
+ 445,
+ 506,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 445,
+ 506,
+ 459
+ ],
+ "score": 1.0,
+ "content": "a CNN to reconstruct the representation (Dosovitskiy & Brox, 2016). Yet these methods require a",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 471
+ ],
+ "score": 1.0,
+ "content": "priori knowledge in order to find the appropriate embeddings or training sets. We now discuss the",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 468,
+ 272,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 468,
+ 230,
+ 480
+ ],
+ "score": 1.0,
+ "content": "improvements achieved by the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 230,
+ 469,
+ 235,
+ 478
+ ],
+ "score": 0.7,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 468,
+ 272,
+ 480
+ ],
+ "score": 1.0,
+ "content": "-RevNet.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 484,
+ 505,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 427,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Our main claim is that while the local inversion is ill-conditioned, the inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 428,
+ 484,
+ 447,
+ 495
+ ],
+ "score": 0.9,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 447,
+ 484,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "computations",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 506,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 506,
+ 509
+ ],
+ "score": 1.0,
+ "content": "do not involve significant round-off errors. The forward pass of the network does not seem to",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 506,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 473,
+ 519
+ ],
+ "score": 1.0,
+ "content": "suffer from significant instabilities, thus it seems coherent to assume that this will hold for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 506,
+ 492,
+ 517
+ ],
+ "score": 0.9,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 506,
+ 506,
+ 519
+ ],
+ "score": 1.0,
+ "content": "as",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 518,
+ 505,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 518,
+ 505,
+ 530
+ ],
+ "score": 1.0,
+ "content": "well. For example, adding constraints beyond vanishing moments in the case of a Lifting scheme",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "is difficult (Sweldens, 1998; Mallat, 1999), and this is a weakness of this method. We validate our",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 540,
+ 415,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 371,
+ 551
+ ],
+ "score": 1.0,
+ "content": "claim by computing the empirical relative error on several subsets",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 540,
+ 381,
+ 550
+ ],
+ "score": 0.84,
+ "content": "\\mathcal { X }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 541,
+ 415,
+ 551
+ ],
+ "score": 1.0,
+ "content": "of data:",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 238,
+ 565,
+ 372,
+ 597
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 238,
+ 565,
+ 372,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 238,
+ 565,
+ 372,
+ 597
+ ],
+ "score": 0.94,
+ "content": "\\epsilon ( \\mathcal { X } ) = \\frac { 1 } { | \\mathcal { X } | } \\sum _ { \\boldsymbol { x } \\in \\mathcal { X } } \\frac { \\| \\boldsymbol { x } - \\Phi ^ { - 1 } \\Phi \\boldsymbol { x } \\| } { \\| \\boldsymbol { x } \\| }",
+ "type": "interline_equation",
+ "image_path": "faf617bbc82767bf83b06069a10c7f59321cf2682e02e0004f4de34f1200046d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 22.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 238,
+ 565,
+ 372,
+ 581.0
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 238,
+ 581.0,
+ 372,
+ 597.0
+ ],
+ "spans": [],
+ "index": 23
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 608,
+ 505,
+ 653
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 607,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 607,
+ 263,
+ 621
+ ],
+ "score": 1.0,
+ "content": "We evaluate this measure on a subset",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 609,
+ 276,
+ 619
+ ],
+ "score": 0.89,
+ "content": "\\mathcal { X } _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 607,
+ 291,
+ 621
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 608,
+ 340,
+ 620
+ ],
+ "score": 0.91,
+ "content": "| \\mathcal { X } _ { 1 } | = 1 0 ^ { 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 341,
+ 607,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "independent uniform noises and on the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 163,
+ 632
+ ],
+ "score": 1.0,
+ "content": "validation set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 163,
+ 620,
+ 176,
+ 631
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { X } _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 618,
+ 277,
+ 632
+ ],
+ "score": 1.0,
+ "content": "of ImageNet. We report",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 620,
+ 354,
+ 632
+ ],
+ "score": 0.9,
+ "content": "\\epsilon ( \\mathcal { X } _ { 1 } ) = 5 \\times 1 0 ^ { - 6 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 355,
+ 618,
+ 374,
+ 632
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 374,
+ 619,
+ 451,
+ 632
+ ],
+ "score": 0.92,
+ "content": "\\epsilon ( \\mathcal { X } _ { 2 } ) = 3 \\times 1 0 ^ { - 6 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 618,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "respectively,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 630,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 630,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "which are close to the machine error and indicates that the inversion does not suffer from significant",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 641,
+ 175,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 641,
+ 175,
+ 653
+ ],
+ "score": 1.0,
+ "content": "round-off errors.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 658,
+ 505,
+ 702
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 656,
+ 507,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 656,
+ 206,
+ 672
+ ],
+ "score": 1.0,
+ "content": "Given a pair of images",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 207,
+ 658,
+ 242,
+ 670
+ ],
+ "score": 0.94,
+ "content": "\\{ x ^ { 0 } , x ^ { 1 } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 656,
+ 507,
+ 672
+ ],
+ "score": 1.0,
+ "content": ", we propose to study linear interpolations between the pair of",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 668,
+ 506,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 668,
+ 171,
+ 683
+ ],
+ "score": 1.0,
+ "content": "representations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 669,
+ 221,
+ 681
+ ],
+ "score": 0.91,
+ "content": "\\{ \\Phi x ^ { \\bar { 0 } } , \\Phi x ^ { \\mathrm { { 1 } } } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 221,
+ 668,
+ 506,
+ 683
+ ],
+ "score": 1.0,
+ "content": ", in the feature domain. Those interpolations correspond to existing",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 680,
+ 506,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 680,
+ 147,
+ 693
+ ],
+ "score": 1.0,
+ "content": "images as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 148,
+ 680,
+ 167,
+ 690
+ ],
+ "score": 0.91,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 680,
+ 506,
+ 693
+ ],
+ "score": 1.0,
+ "content": "is an exact inverse. We reconstruct a convex path between two input points; it means",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 690,
+ 136,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 690,
+ 136,
+ 703
+ ],
+ "score": 1.0,
+ "content": "that if:",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 252,
+ 702,
+ 358,
+ 716
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 252,
+ 702,
+ 358,
+ 716
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 252,
+ 702,
+ 358,
+ 716
+ ],
+ "score": 0.91,
+ "content": "\\phi ^ { t } = t \\Phi x ^ { 0 } + ( 1 - t ) \\Phi x ^ { 1 } ,",
+ "type": "interline_equation",
+ "image_path": "b0fc9ca852c1841fc195fdab2bbbe01c79b13bcab5247de6e4056b08676f6395.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 32,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 252,
+ 702,
+ 358,
+ 716
+ ],
+ "spans": [],
+ "index": 32
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 720,
+ 343,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 718,
+ 344,
+ 735
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 718,
+ 129,
+ 735
+ ],
+ "score": 1.0,
+ "content": "then:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 129,
+ 720,
+ 181,
+ 732
+ ],
+ "score": 0.93,
+ "content": "x ^ { t } = \\Phi ^ { - 1 } \\phi ^ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 718,
+ 344,
+ 735
+ ],
+ "score": 1.0,
+ "content": "is a signal that corresponds to an image.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 303,
+ 751,
+ 309,
+ 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": [
+ 106,
+ 64,
+ 505,
+ 288
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 64,
+ 505,
+ 288
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 64,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 64,
+ 505,
+ 288
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "a6fe6b5bd4ff7c42fa5cbd145f44da0d3e48b046022b4d75069d5e7eefe3144c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 64,
+ 505,
+ 138.66666666666669
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 138.66666666666669,
+ 505,
+ 213.33333333333337
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 213.33333333333337,
+ 505,
+ 288.00000000000006
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 297,
+ 503,
+ 320
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 506,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 365,
+ 311
+ ],
+ "score": 1.0,
+ "content": "Figure 5: This graphic displays several reconstructed sequences",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 365,
+ 297,
+ 388,
+ 309
+ ],
+ "score": 0.92,
+ "content": "\\{ x ^ { t } \\} _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 388,
+ 295,
+ 506,
+ 311
+ ],
+ "score": 1.0,
+ "content": ". The left image corresponds",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 307,
+ 235,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 117,
+ 321
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 308,
+ 128,
+ 318
+ ],
+ "score": 0.87,
+ "content": "{ \\bar { \\mathbf { \\Gamma } } } _ { x } 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 128,
+ 307,
+ 219,
+ 321
+ ],
+ "score": 1.0,
+ "content": "and the right image to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 308,
+ 230,
+ 318
+ ],
+ "score": 0.87,
+ "content": "x ^ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 307,
+ 235,
+ 321
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 343,
+ 339,
+ 353
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 340,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 340,
+ 354
+ ],
+ "score": 1.0,
+ "content": "4.2 LINEAR INTERPOLATION AND RECONSTRUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 363,
+ 505,
+ 419
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 364,
+ 505,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 364,
+ 505,
+ 376
+ ],
+ "score": 1.0,
+ "content": "Visualizing or understanding the important directions in the representation of inner layers of a CNN,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 375,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 375,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": "and in particular, the final layer is complex because typically the cascade is either not invertible or",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 385,
+ 506,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 506,
+ 398
+ ],
+ "score": 1.0,
+ "content": "unstable. One approach to reconstruct from an output layer consists in finding the input image that",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "matches the activation through via gradient descent. However, this technique leads only to a partial",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 407,
+ 338,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 407,
+ 338,
+ 419
+ ],
+ "score": 1.0,
+ "content": "or informal reconstruction (Mahendran & Vedaldi, 2015).",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 8,
+ "bbox_fs": [
+ 105,
+ 364,
+ 506,
+ 419
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 479
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "Another method consists in embedding the representation in a lower dimensional space and com-",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "paring the common attributes of nearest neighbors (Szegedy et al., 2013). It is also possible to train",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 104,
+ 445,
+ 506,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 445,
+ 506,
+ 459
+ ],
+ "score": 1.0,
+ "content": "a CNN to reconstruct the representation (Dosovitskiy & Brox, 2016). Yet these methods require a",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 471
+ ],
+ "score": 1.0,
+ "content": "priori knowledge in order to find the appropriate embeddings or training sets. We now discuss the",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 468,
+ 272,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 468,
+ 230,
+ 480
+ ],
+ "score": 1.0,
+ "content": "improvements achieved by the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 230,
+ 469,
+ 235,
+ 478
+ ],
+ "score": 0.7,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 468,
+ 272,
+ 480
+ ],
+ "score": 1.0,
+ "content": "-RevNet.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 104,
+ 424,
+ 506,
+ 480
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 484,
+ 505,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 427,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Our main claim is that while the local inversion is ill-conditioned, the inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 428,
+ 484,
+ 447,
+ 495
+ ],
+ "score": 0.9,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 447,
+ 484,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "computations",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 506,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 506,
+ 509
+ ],
+ "score": 1.0,
+ "content": "do not involve significant round-off errors. The forward pass of the network does not seem to",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 506,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 473,
+ 519
+ ],
+ "score": 1.0,
+ "content": "suffer from significant instabilities, thus it seems coherent to assume that this will hold for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 506,
+ 492,
+ 517
+ ],
+ "score": 0.9,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 506,
+ 506,
+ 519
+ ],
+ "score": 1.0,
+ "content": "as",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 518,
+ 505,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 518,
+ 505,
+ 530
+ ],
+ "score": 1.0,
+ "content": "well. For example, adding constraints beyond vanishing moments in the case of a Lifting scheme",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "is difficult (Sweldens, 1998; Mallat, 1999), and this is a weakness of this method. We validate our",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 540,
+ 415,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 371,
+ 551
+ ],
+ "score": 1.0,
+ "content": "claim by computing the empirical relative error on several subsets",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 540,
+ 381,
+ 550
+ ],
+ "score": 0.84,
+ "content": "\\mathcal { X }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 541,
+ 415,
+ 551
+ ],
+ "score": 1.0,
+ "content": "of data:",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 105,
+ 484,
+ 506,
+ 551
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 238,
+ 565,
+ 372,
+ 597
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 238,
+ 565,
+ 372,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 238,
+ 565,
+ 372,
+ 597
+ ],
+ "score": 0.94,
+ "content": "\\epsilon ( \\mathcal { X } ) = \\frac { 1 } { | \\mathcal { X } | } \\sum _ { \\boldsymbol { x } \\in \\mathcal { X } } \\frac { \\| \\boldsymbol { x } - \\Phi ^ { - 1 } \\Phi \\boldsymbol { x } \\| } { \\| \\boldsymbol { x } \\| }",
+ "type": "interline_equation",
+ "image_path": "faf617bbc82767bf83b06069a10c7f59321cf2682e02e0004f4de34f1200046d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 22.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 238,
+ 565,
+ 372,
+ 581.0
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 238,
+ 581.0,
+ 372,
+ 597.0
+ ],
+ "spans": [],
+ "index": 23
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 608,
+ 505,
+ 653
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 607,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 607,
+ 263,
+ 621
+ ],
+ "score": 1.0,
+ "content": "We evaluate this measure on a subset",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 609,
+ 276,
+ 619
+ ],
+ "score": 0.89,
+ "content": "\\mathcal { X } _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 607,
+ 291,
+ 621
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 608,
+ 340,
+ 620
+ ],
+ "score": 0.91,
+ "content": "| \\mathcal { X } _ { 1 } | = 1 0 ^ { 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 341,
+ 607,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "independent uniform noises and on the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 163,
+ 632
+ ],
+ "score": 1.0,
+ "content": "validation set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 163,
+ 620,
+ 176,
+ 631
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { X } _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 618,
+ 277,
+ 632
+ ],
+ "score": 1.0,
+ "content": "of ImageNet. We report",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 620,
+ 354,
+ 632
+ ],
+ "score": 0.9,
+ "content": "\\epsilon ( \\mathcal { X } _ { 1 } ) = 5 \\times 1 0 ^ { - 6 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 355,
+ 618,
+ 374,
+ 632
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 374,
+ 619,
+ 451,
+ 632
+ ],
+ "score": 0.92,
+ "content": "\\epsilon ( \\mathcal { X } _ { 2 } ) = 3 \\times 1 0 ^ { - 6 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 618,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "respectively,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 630,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 630,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "which are close to the machine error and indicates that the inversion does not suffer from significant",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 641,
+ 175,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 641,
+ 175,
+ 653
+ ],
+ "score": 1.0,
+ "content": "round-off errors.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 105,
+ 607,
+ 506,
+ 653
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 658,
+ 505,
+ 702
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 656,
+ 507,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 656,
+ 206,
+ 672
+ ],
+ "score": 1.0,
+ "content": "Given a pair of images",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 207,
+ 658,
+ 242,
+ 670
+ ],
+ "score": 0.94,
+ "content": "\\{ x ^ { 0 } , x ^ { 1 } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 656,
+ 507,
+ 672
+ ],
+ "score": 1.0,
+ "content": ", we propose to study linear interpolations between the pair of",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 668,
+ 506,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 668,
+ 171,
+ 683
+ ],
+ "score": 1.0,
+ "content": "representations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 669,
+ 221,
+ 681
+ ],
+ "score": 0.91,
+ "content": "\\{ \\Phi x ^ { \\bar { 0 } } , \\Phi x ^ { \\mathrm { { 1 } } } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 221,
+ 668,
+ 506,
+ 683
+ ],
+ "score": 1.0,
+ "content": ", in the feature domain. Those interpolations correspond to existing",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 680,
+ 506,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 680,
+ 147,
+ 693
+ ],
+ "score": 1.0,
+ "content": "images as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 148,
+ 680,
+ 167,
+ 690
+ ],
+ "score": 0.91,
+ "content": "\\Phi ^ { - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 680,
+ 506,
+ 693
+ ],
+ "score": 1.0,
+ "content": "is an exact inverse. We reconstruct a convex path between two input points; it means",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 690,
+ 136,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 690,
+ 136,
+ 703
+ ],
+ "score": 1.0,
+ "content": "that if:",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 104,
+ 656,
+ 507,
+ 703
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 252,
+ 702,
+ 358,
+ 716
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 252,
+ 702,
+ 358,
+ 716
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 252,
+ 702,
+ 358,
+ 716
+ ],
+ "score": 0.91,
+ "content": "\\phi ^ { t } = t \\Phi x ^ { 0 } + ( 1 - t ) \\Phi x ^ { 1 } ,",
+ "type": "interline_equation",
+ "image_path": "b0fc9ca852c1841fc195fdab2bbbe01c79b13bcab5247de6e4056b08676f6395.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 32,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 252,
+ 702,
+ 358,
+ 716
+ ],
+ "spans": [],
+ "index": 32
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 720,
+ 343,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 718,
+ 344,
+ 735
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 718,
+ 129,
+ 735
+ ],
+ "score": 1.0,
+ "content": "then:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 129,
+ 720,
+ 181,
+ 732
+ ],
+ "score": 0.93,
+ "content": "x ^ { t } = \\Phi ^ { - 1 } \\phi ^ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 718,
+ 344,
+ 735
+ ],
+ "score": 1.0,
+ "content": "is a signal that corresponds to an image.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 105,
+ 718,
+ 344,
+ 735
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 119,
+ 62,
+ 492,
+ 218
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 119,
+ 62,
+ 492,
+ 218
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 62,
+ 492,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 62,
+ 492,
+ 218
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "293bac6aae019b36fb6785db214dbbee593d7a2ee2f0f04de109fb47c7cb0a9a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 119,
+ 62,
+ 492,
+ 114.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 119,
+ 114.0,
+ 492,
+ 166.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 119,
+ 166.0,
+ 492,
+ 218.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 227,
+ 503,
+ 250
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 224,
+ 240
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Accuracy at depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 228,
+ 231,
+ 239
+ ],
+ "score": 0.83,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 226,
+ 505,
+ 240
+ ],
+ "score": 1.0,
+ "content": "for a linear SVM and a 1-nearest neighbor classifier applied to the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 198,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 181,
+ 252
+ ],
+ "score": 1.0,
+ "content": "spatially averaged",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 181,
+ 239,
+ 193,
+ 251
+ ],
+ "score": 0.89,
+ "content": "\\Phi _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 194,
+ 237,
+ 198,
+ 252
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 275,
+ 505,
+ 308
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 170,
+ 288
+ ],
+ "score": 1.0,
+ "content": "We discretized",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 275,
+ 191,
+ 287
+ ],
+ "score": 0.48,
+ "content": "[ 0 , 1 ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 192,
+ 273,
+ 213,
+ 288
+ ],
+ "score": 1.0,
+ "content": "into",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 213,
+ 275,
+ 258,
+ 287
+ ],
+ "score": 0.94,
+ "content": "\\{ t _ { 1 } , . . . , t _ { k } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 273,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": ", adapt the step size manually and reconstruct the sequence",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 284,
+ 506,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 118,
+ 299
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 119,
+ 286,
+ 173,
+ 298
+ ],
+ "score": 0.93,
+ "content": "\\{ x ^ { t _ { 1 } } , . . . , x ^ { t _ { k } } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 284,
+ 506,
+ 299
+ ],
+ "score": 1.0,
+ "content": ". Results are displayed in the Figure 5. We selected images from the basel face",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 296,
+ 475,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 475,
+ 310
+ ],
+ "score": 1.0,
+ "content": "dataset (Paysan et al., 2009), describable texture dataset (Cimpoi et al., 2014) and imagenet.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 314,
+ 505,
+ 369
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "We now interpret the results. First, observe that a linear interpolation in the feature space is not",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 505,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 505,
+ 338
+ ],
+ "score": 1.0,
+ "content": "a linear interpolation in the image space and that intermediary images are noisy, even for small",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 336,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "deformations, yet they mostly remain recognizable. However, some geometric transformations such",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 505,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 505,
+ 360
+ ],
+ "score": 1.0,
+ "content": "as a 3D-rotation seem to have been linearized, as suggested in Aubry & Russell (2015). In the next",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 357,
+ 411,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 411,
+ 371
+ ],
+ "score": 1.0,
+ "content": "section, we thus investigate how the linear separation progresses with depth.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 389,
+ 213,
+ 402
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 214,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 214,
+ 403
+ ],
+ "score": 1.0,
+ "content": "5 A CONTRACTION",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 416,
+ 505,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 292,
+ 429
+ ],
+ "score": 1.0,
+ "content": "In this section, we study again the bijective",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 292,
+ 417,
+ 297,
+ 426
+ ],
+ "score": 0.69,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 298,
+ 416,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "-RevNet. We first show that a localized or linear",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "classifier progressively improves with depth. Then, we describe the linear subspace spanned by",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 506,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 439,
+ 115,
+ 449
+ ],
+ "score": 0.76,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 115,
+ 438,
+ 506,
+ 451
+ ],
+ "score": 1.0,
+ "content": ", namely the feature space, showing that the classification can be performed on a much smaller",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 450,
+ 269,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 269,
+ 461
+ ],
+ "score": 1.0,
+ "content": "subspace, which can be built via a PCA.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 112,
+ 479,
+ 365,
+ 489
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 478,
+ 367,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 478,
+ 367,
+ 491
+ ],
+ "score": 1.0,
+ "content": "5.1 PROGRESSIVE LINEAR SEPARATION AND CONTRACTION",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 500,
+ 505,
+ 533
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 505,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 259,
+ 512
+ ],
+ "score": 1.0,
+ "content": "We show that both a ResNet and an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 501,
+ 264,
+ 510
+ ],
+ "score": 0.64,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 500,
+ 505,
+ 512
+ ],
+ "score": 1.0,
+ "content": "-RevNet build a progressively more linearly separable and",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 496,
+ 523
+ ],
+ "score": 1.0,
+ "content": "contracted representation as measured in Oyallon (2017). Observe this property holds for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 512,
+ 500,
+ 521
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 511,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 522,
+ 359,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 359,
+ 535
+ ],
+ "score": 1.0,
+ "content": "RevNet despite the fact that it can not discard any information.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 539,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 538,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 538,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "We investigate these properties in each block, with the following experimental protocol. To reduce",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 562
+ ],
+ "score": 1.0,
+ "content": "the computational burden we used a subset of 100 randomly selected imagenet classes, that consist",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 117,
+ 574
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 561,
+ 161,
+ 572
+ ],
+ "score": 0.91,
+ "content": "N = 1 2 0 k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 560,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "images, and keep the same subset during all our following experiments. At each depth",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 572,
+ 506,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 112,
+ 584
+ ],
+ "score": 0.72,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 113,
+ 572,
+ 207,
+ 586
+ ],
+ "score": 1.0,
+ "content": ", we extract the features",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 572,
+ 259,
+ 585
+ ],
+ "score": 0.93,
+ "content": "\\{ \\Phi _ { j } x ^ { n } \\} _ { n \\leq N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 260,
+ 572,
+ 506,
+ 586
+ ],
+ "score": 1.0,
+ "content": "of the training set, we average them along the spatial variable",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 583,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 583,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "and standardize them in order to avoid any ill-conditioning effects. We used both a nearest neighbor",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 593,
+ 506,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 455,
+ 606
+ ],
+ "score": 1.0,
+ "content": "classifier and a linear SVM. The former is a localized classifier that indicates that the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 456,
+ 594,
+ 465,
+ 604
+ ],
+ "score": 0.82,
+ "content": "\\ell ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 593,
+ 506,
+ 606
+ ],
+ "score": 1.0,
+ "content": "metric is",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": "progressively more important for classification, while a linear SVM measures the linear separation",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 628
+ ],
+ "score": 1.0,
+ "content": "of the different classes. The parameters of the linear SVM are cross-validated on a small subset of",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "the training set, prior to training on the 100 classes. We evaluate both classifiers for each model on",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 637,
+ 399,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 399,
+ 650
+ ],
+ "score": 1.0,
+ "content": "the validation set of ImageNet and report the Top-1 accuracy in Figure 6.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 654,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "We observe that both classifiers progressively improve similarly with depth for each model, the lin-",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "ear SVM performing slightly better than the nearest neighbor classifier because it is the more robust",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 337,
+ 690
+ ],
+ "score": 1.0,
+ "content": "and discriminative classifier of the two. In the case of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 678,
+ 342,
+ 687
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 676,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "-RevNet, the classification performed by",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 180,
+ 700
+ ],
+ "score": 1.0,
+ "content": "the CNN leads to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 181,
+ 687,
+ 200,
+ 699
+ ],
+ "score": 0.88,
+ "content": "7 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 200,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": ", and the linear SVM performs slightly better because we did not fine-tune",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "the model to 100 classes. Observe that there is a more intense jump of performance on the 3 last",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "layers, which seems to indicate that the former layers have prepared the representation to be more",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 318,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 318,
+ 733
+ ],
+ "score": 1.0,
+ "content": "contracted and linearly separated for the final layers.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published 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": "image",
+ "bbox": [
+ 119,
+ 62,
+ 492,
+ 218
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 119,
+ 62,
+ 492,
+ 218
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 62,
+ 492,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 62,
+ 492,
+ 218
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "293bac6aae019b36fb6785db214dbbee593d7a2ee2f0f04de109fb47c7cb0a9a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 119,
+ 62,
+ 492,
+ 114.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 119,
+ 114.0,
+ 492,
+ 166.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 119,
+ 166.0,
+ 492,
+ 218.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 227,
+ 503,
+ 250
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 224,
+ 240
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Accuracy at depth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 228,
+ 231,
+ 239
+ ],
+ "score": 0.83,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 226,
+ 505,
+ 240
+ ],
+ "score": 1.0,
+ "content": "for a linear SVM and a 1-nearest neighbor classifier applied to the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 198,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 181,
+ 252
+ ],
+ "score": 1.0,
+ "content": "spatially averaged",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 181,
+ 239,
+ 193,
+ 251
+ ],
+ "score": 0.89,
+ "content": "\\Phi _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 194,
+ 237,
+ 198,
+ 252
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 275,
+ 505,
+ 308
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 170,
+ 288
+ ],
+ "score": 1.0,
+ "content": "We discretized",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 275,
+ 191,
+ 287
+ ],
+ "score": 0.48,
+ "content": "[ 0 , 1 ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 192,
+ 273,
+ 213,
+ 288
+ ],
+ "score": 1.0,
+ "content": "into",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 213,
+ 275,
+ 258,
+ 287
+ ],
+ "score": 0.94,
+ "content": "\\{ t _ { 1 } , . . . , t _ { k } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 273,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": ", adapt the step size manually and reconstruct the sequence",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 284,
+ 506,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 118,
+ 299
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 119,
+ 286,
+ 173,
+ 298
+ ],
+ "score": 0.93,
+ "content": "\\{ x ^ { t _ { 1 } } , . . . , x ^ { t _ { k } } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 284,
+ 506,
+ 299
+ ],
+ "score": 1.0,
+ "content": ". Results are displayed in the Figure 5. We selected images from the basel face",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 296,
+ 475,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 475,
+ 310
+ ],
+ "score": 1.0,
+ "content": "dataset (Paysan et al., 2009), describable texture dataset (Cimpoi et al., 2014) and imagenet.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 273,
+ 506,
+ 310
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 314,
+ 505,
+ 369
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "We now interpret the results. First, observe that a linear interpolation in the feature space is not",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 505,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 505,
+ 338
+ ],
+ "score": 1.0,
+ "content": "a linear interpolation in the image space and that intermediary images are noisy, even for small",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 336,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "deformations, yet they mostly remain recognizable. However, some geometric transformations such",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 505,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 505,
+ 360
+ ],
+ "score": 1.0,
+ "content": "as a 3D-rotation seem to have been linearized, as suggested in Aubry & Russell (2015). In the next",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 357,
+ 411,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 411,
+ 371
+ ],
+ "score": 1.0,
+ "content": "section, we thus investigate how the linear separation progresses with depth.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 105,
+ 314,
+ 505,
+ 371
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 389,
+ 213,
+ 402
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 214,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 214,
+ 403
+ ],
+ "score": 1.0,
+ "content": "5 A CONTRACTION",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 416,
+ 505,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 292,
+ 429
+ ],
+ "score": 1.0,
+ "content": "In this section, we study again the bijective",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 292,
+ 417,
+ 297,
+ 426
+ ],
+ "score": 0.69,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 298,
+ 416,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "-RevNet. We first show that a localized or linear",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "classifier progressively improves with depth. Then, we describe the linear subspace spanned by",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 506,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 439,
+ 115,
+ 449
+ ],
+ "score": 0.76,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 115,
+ 438,
+ 506,
+ 451
+ ],
+ "score": 1.0,
+ "content": ", namely the feature space, showing that the classification can be performed on a much smaller",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 450,
+ 269,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 269,
+ 461
+ ],
+ "score": 1.0,
+ "content": "subspace, which can be built via a PCA.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 105,
+ 416,
+ 506,
+ 461
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 112,
+ 479,
+ 365,
+ 489
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 478,
+ 367,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 478,
+ 367,
+ 491
+ ],
+ "score": 1.0,
+ "content": "5.1 PROGRESSIVE LINEAR SEPARATION AND CONTRACTION",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 500,
+ 505,
+ 533
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 505,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 259,
+ 512
+ ],
+ "score": 1.0,
+ "content": "We show that both a ResNet and an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 501,
+ 264,
+ 510
+ ],
+ "score": 0.64,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 500,
+ 505,
+ 512
+ ],
+ "score": 1.0,
+ "content": "-RevNet build a progressively more linearly separable and",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 496,
+ 523
+ ],
+ "score": 1.0,
+ "content": "contracted representation as measured in Oyallon (2017). Observe this property holds for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 512,
+ 500,
+ 521
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 511,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 522,
+ 359,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 359,
+ 535
+ ],
+ "score": 1.0,
+ "content": "RevNet despite the fact that it can not discard any information.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20,
+ "bbox_fs": [
+ 105,
+ 500,
+ 505,
+ 535
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 539,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 538,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 538,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "We investigate these properties in each block, with the following experimental protocol. To reduce",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 562
+ ],
+ "score": 1.0,
+ "content": "the computational burden we used a subset of 100 randomly selected imagenet classes, that consist",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 117,
+ 574
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 561,
+ 161,
+ 572
+ ],
+ "score": 0.91,
+ "content": "N = 1 2 0 k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 560,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "images, and keep the same subset during all our following experiments. At each depth",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 572,
+ 506,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 112,
+ 584
+ ],
+ "score": 0.72,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 113,
+ 572,
+ 207,
+ 586
+ ],
+ "score": 1.0,
+ "content": ", we extract the features",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 572,
+ 259,
+ 585
+ ],
+ "score": 0.93,
+ "content": "\\{ \\Phi _ { j } x ^ { n } \\} _ { n \\leq N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 260,
+ 572,
+ 506,
+ 586
+ ],
+ "score": 1.0,
+ "content": "of the training set, we average them along the spatial variable",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 583,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 583,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "and standardize them in order to avoid any ill-conditioning effects. We used both a nearest neighbor",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 593,
+ 506,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 455,
+ 606
+ ],
+ "score": 1.0,
+ "content": "classifier and a linear SVM. The former is a localized classifier that indicates that the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 456,
+ 594,
+ 465,
+ 604
+ ],
+ "score": 0.82,
+ "content": "\\ell ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 593,
+ 506,
+ 606
+ ],
+ "score": 1.0,
+ "content": "metric is",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": "progressively more important for classification, while a linear SVM measures the linear separation",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 628
+ ],
+ "score": 1.0,
+ "content": "of the different classes. The parameters of the linear SVM are cross-validated on a small subset of",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "the training set, prior to training on the 100 classes. We evaluate both classifiers for each model on",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 637,
+ 399,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 399,
+ 650
+ ],
+ "score": 1.0,
+ "content": "the validation set of ImageNet and report the Top-1 accuracy in Figure 6.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 538,
+ 506,
+ 650
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 654,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "We observe that both classifiers progressively improve similarly with depth for each model, the lin-",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "ear SVM performing slightly better than the nearest neighbor classifier because it is the more robust",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 337,
+ 690
+ ],
+ "score": 1.0,
+ "content": "and discriminative classifier of the two. In the case of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 678,
+ 342,
+ 687
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 676,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "-RevNet, the classification performed by",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 180,
+ 700
+ ],
+ "score": 1.0,
+ "content": "the CNN leads to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 181,
+ 687,
+ 200,
+ 699
+ ],
+ "score": 0.88,
+ "content": "7 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 200,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": ", and the linear SVM performs slightly better because we did not fine-tune",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "the model to 100 classes. Observe that there is a more intense jump of performance on the 3 last",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "layers, which seems to indicate that the former layers have prepared the representation to be more",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 318,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 318,
+ 733
+ ],
+ "score": 1.0,
+ "content": "contracted and linearly separated for the final layers.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 654,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 82,
+ 505,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "The results suggest a low-dimensional embedding of the data, but this is difficult to validate as",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "estimating local dimensionality in high dimensions is an open problem. However, in the next section,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 104,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 104,
+ 463,
+ 117
+ ],
+ "score": 1.0,
+ "content": "we try to compute the dimension of the discriminative part of the representation built by an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 463,
+ 105,
+ 468,
+ 114
+ ],
+ "score": 0.68,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 468,
+ 104,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "-RevNet.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 130,
+ 356,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 357,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 357,
+ 142
+ ],
+ "score": 1.0,
+ "content": "5.2 DIMENSIONALITY ANALYSIS OF THE FEATURE SPACE",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 150,
+ 505,
+ 194
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "In this section, we investigate if we can refine the dimensionality of informative variabilities in the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 162,
+ 505,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 162,
+ 173,
+ 173
+ ],
+ "score": 1.0,
+ "content": "final layer of an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 162,
+ 178,
+ 172
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 179,
+ 162,
+ 505,
+ 173
+ ],
+ "score": 1.0,
+ "content": "-RevNet. Indeed, the cascade of convolutional operators has been trained on the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "score": 1.0,
+ "content": "training set to separate the 1000 different classes while being a homeomorphism on its feature space.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 182,
+ 367,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 182,
+ 367,
+ 197
+ ],
+ "score": 1.0,
+ "content": "Thus, the dimensionality of the feature space is potentially large.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 200,
+ 505,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "As shown in the previous subsection, the final layer is progressively prepared to be projected on",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 211,
+ 506,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 506,
+ 223
+ ],
+ "score": 1.0,
+ "content": "the final probes corresponding to the classes. This indicates that the non-informative variabilities",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 221,
+ 506,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 221,
+ 408,
+ 235
+ ],
+ "score": 1.0,
+ "content": "for classification can be removed via a linear projection on the final layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 223,
+ 417,
+ 232
+ ],
+ "score": 0.77,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 417,
+ 221,
+ 506,
+ 235
+ ],
+ "score": 1.0,
+ "content": ", which lie in a space",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "of dimension 1000, at most. However, this projection has been built via supervision, which can",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "still retain directions that have been contracted and thus will not be selected by an algorithm such as",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 505,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 505,
+ 268
+ ],
+ "score": 1.0,
+ "content": "PCA. We show in fact a PCA retains the necessary information for classification in a small subspace.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 271,
+ 296,
+ 469
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 297,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 271,
+ 285
+ ],
+ "score": 1.0,
+ "content": "To do so, we build the linear projectors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 271,
+ 274,
+ 283,
+ 283
+ ],
+ "score": 0.85,
+ "content": "\\pi _ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 283,
+ 271,
+ 297,
+ 285
+ ],
+ "score": 1.0,
+ "content": "on",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 283,
+ 296,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 283,
+ 194,
+ 295
+ ],
+ "score": 1.0,
+ "content": "the subspace of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 284,
+ 201,
+ 293
+ ],
+ "score": 0.7,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 283,
+ 296,
+ 295
+ ],
+ "score": 1.0,
+ "content": "first principal compo-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 294,
+ 297,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 297,
+ 305
+ ],
+ "score": 1.0,
+ "content": "nents, and we propose to measure the classifica-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 305,
+ 297,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 305,
+ 297,
+ 317
+ ],
+ "score": 1.0,
+ "content": "tion power of the projected representation with",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 316,
+ 297,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 316,
+ 297,
+ 328
+ ],
+ "score": 1.0,
+ "content": "a supervised classifier, e.g. nearest neighbor or",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 297,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 297,
+ 339
+ ],
+ "score": 1.0,
+ "content": "a linear SVM, on the previous 100 class task.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 335,
+ 295,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 248,
+ 352
+ ],
+ "score": 1.0,
+ "content": "Again, the feature representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 338,
+ 295,
+ 350
+ ],
+ "score": 0.92,
+ "content": "\\{ \\Phi x ^ { n } \\} _ { n \\leq N }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 348,
+ 298,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 348,
+ 298,
+ 361
+ ],
+ "score": 1.0,
+ "content": "are spatially averaged to remove the translation",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 360,
+ 297,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 297,
+ 371
+ ],
+ "score": 1.0,
+ "content": "variability, and standardized on the training set.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 370,
+ 298,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 370,
+ 298,
+ 383
+ ],
+ "score": 1.0,
+ "content": "We apply both classifiers, and we report the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 379,
+ 298,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 379,
+ 214,
+ 395
+ ],
+ "score": 1.0,
+ "content": "classification accuracy of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 214,
+ 381,
+ 272,
+ 394
+ ],
+ "score": 0.93,
+ "content": "\\{ \\pi _ { d } \\Phi x ^ { n } \\} _ { n \\leq N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 379,
+ 298,
+ 395
+ ],
+ "score": 1.0,
+ "content": "w.r.t.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 392,
+ 297,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 392,
+ 118,
+ 405
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 118,
+ 393,
+ 125,
+ 403
+ ],
+ "score": 0.77,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 126,
+ 392,
+ 297,
+ 405
+ ],
+ "score": 1.0,
+ "content": "on the Figure 7. A linear projection re-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 297,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 404,
+ 297,
+ 415
+ ],
+ "score": 1.0,
+ "content": "moves some information that can not be recov-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 414,
+ 297,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 414,
+ 297,
+ 426
+ ],
+ "score": 1.0,
+ "content": "ered by a linear classifier, therefore we observe",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 425,
+ 297,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 297,
+ 437
+ ],
+ "score": 1.0,
+ "content": "that the classification accuracy only decreases",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 437,
+ 297,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 437,
+ 173,
+ 447
+ ],
+ "score": 1.0,
+ "content": "significantly for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 437,
+ 210,
+ 447
+ ],
+ "score": 0.91,
+ "content": "d \\leq 2 0 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 210,
+ 437,
+ 297,
+ 447
+ ],
+ "score": 1.0,
+ "content": ". This shows that the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 447,
+ 297,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 447,
+ 297,
+ 459
+ ],
+ "score": 1.0,
+ "content": "signal indeed lies in a subspace much lower di-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 459,
+ 297,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 459,
+ 297,
+ 470
+ ],
+ "score": 1.0,
+ "content": "mensional than the original feature dimensions",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 308,
+ 275,
+ 500,
+ 416
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 308,
+ 275,
+ 500,
+ 416
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 275,
+ 500,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 275,
+ 500,
+ 416
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "93bdd4b0bf41956f8e5f34e143db6c00cc121e2956561dec013eee6385bf9daf.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 36.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 308,
+ 275,
+ 500,
+ 289.1
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 308,
+ 289.1,
+ 500,
+ 303.20000000000005
+ ],
+ "spans": [],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 308,
+ 303.20000000000005,
+ 500,
+ 317.30000000000007
+ ],
+ "spans": [],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 308,
+ 317.30000000000007,
+ 500,
+ 331.4000000000001
+ ],
+ "spans": [],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 308,
+ 331.4000000000001,
+ 500,
+ 345.5000000000001
+ ],
+ "spans": [],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 308,
+ 345.5000000000001,
+ 500,
+ 359.60000000000014
+ ],
+ "spans": [],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 308,
+ 359.60000000000014,
+ 500,
+ 373.70000000000016
+ ],
+ "spans": [],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 308,
+ 373.70000000000016,
+ 500,
+ 387.8000000000002
+ ],
+ "spans": [],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 308,
+ 387.8000000000002,
+ 500,
+ 401.9000000000002
+ ],
+ "spans": [],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 308,
+ 401.9000000000002,
+ 500,
+ 416.0000000000002
+ ],
+ "spans": [],
+ "index": 41
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 304,
+ 423,
+ 504,
+ 456
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 423,
+ 505,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 423,
+ 505,
+ 435
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Accuracy of a linear SVM and nearest",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 303,
+ 434,
+ 504,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 434,
+ 504,
+ 447
+ ],
+ "score": 1.0,
+ "content": "neighbor against the number of principal compo-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 304,
+ 446,
+ 365,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 446,
+ 365,
+ 456
+ ],
+ "score": 1.0,
+ "content": "nents retained.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 39.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 469,
+ 504,
+ 492
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 504,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 504,
+ 482
+ ],
+ "score": 1.0,
+ "content": "that can be extracted simply with a PCA that only considers directions of largest variances, illustrat-",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 306,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 306,
+ 493
+ ],
+ "score": 1.0,
+ "content": "ing a successful contraction of the representation.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 508,
+ 195,
+ 520
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 198,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 198,
+ 524
+ ],
+ "score": 1.0,
+ "content": "6 CONCLUSION",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 47
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 533,
+ 505,
+ 610
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "Invertible representations and their relationship to loss of information are on the agenda of deep",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "score": 1.0,
+ "content": "learning for some time. Understanding how transformations in feature space are related to the cor-",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 556,
+ 505,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 556,
+ 505,
+ 568
+ ],
+ "score": 1.0,
+ "content": "responding input is an important step towards interpretable deep networks, invertible deep networks",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 567,
+ 506,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 567,
+ 506,
+ 578
+ ],
+ "score": 1.0,
+ "content": "may play an important role in such analysis since, for example, one could potentially back-track a",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "score": 1.0,
+ "content": "property from the feature space to the input space. To the best of our knowledge, this work pro-",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "vides the first empirical evidence that learning invertible representations that do not discard any",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 414,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 599,
+ 414,
+ 613
+ ],
+ "score": 1.0,
+ "content": "information about their input on large-scale supervised problems is possible.",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ }
+ ],
+ "index": 51
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 615,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 243,
+ 629
+ ],
+ "score": 1.0,
+ "content": "To achieve this we introduce the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 243,
+ 617,
+ 248,
+ 626
+ ],
+ "score": 0.7,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 248,
+ 615,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "-RevNet class of CNN which is fully invertible and permits to",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 337,
+ 639
+ ],
+ "score": 1.0,
+ "content": "exactly recover the input from its last convolutional layer.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 628,
+ 342,
+ 637
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 627,
+ 506,
+ 639
+ ],
+ "score": 1.0,
+ "content": "-RevNets achieve the same classification",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "accuracy in the classification of complex datasets as illustrated on ILSVRC-2012, when compared",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "score": 1.0,
+ "content": "to the RevNet (Gomez et al., 2017) and ResNet (He et al., 2016) architectures with a similar number",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 506,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 424,
+ 674
+ ],
+ "score": 1.0,
+ "content": "of layers. Furthermore, the inverse network is obtained for free when training an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 661,
+ 429,
+ 670
+ ],
+ "score": 0.67,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 658,
+ 506,
+ 674
+ ],
+ "score": 1.0,
+ "content": "-RevNet, requiring",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 400,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 400,
+ 683
+ ],
+ "score": 1.0,
+ "content": "only minimal adaption to recover inputs from the hidden representations.",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ }
+ ],
+ "index": 57.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "The absence of loss of information is surprising, given the wide believe, that discarding information",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "is essential for learning representations that generalize well to unseen data. We show that this is not",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "the case and propose to explain the generalization property with empirical evidence of progressive",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 318,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 318,
+ 732
+ ],
+ "score": 1.0,
+ "content": "separation and contraction with depth, on ImageNet.",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ }
+ ],
+ "index": 62.5
+ }
+ ],
+ "page_idx": 8,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "9",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 82,
+ 505,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "The results suggest a low-dimensional embedding of the data, but this is difficult to validate as",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "estimating local dimensionality in high dimensions is an open problem. However, in the next section,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 104,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 104,
+ 463,
+ 117
+ ],
+ "score": 1.0,
+ "content": "we try to compute the dimension of the discriminative part of the representation built by an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 463,
+ 105,
+ 468,
+ 114
+ ],
+ "score": 0.68,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 468,
+ 104,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "-RevNet.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 106,
+ 82,
+ 505,
+ 117
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 130,
+ 356,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 357,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 357,
+ 142
+ ],
+ "score": 1.0,
+ "content": "5.2 DIMENSIONALITY ANALYSIS OF THE FEATURE SPACE",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 150,
+ 505,
+ 194
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "In this section, we investigate if we can refine the dimensionality of informative variabilities in the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 162,
+ 505,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 162,
+ 173,
+ 173
+ ],
+ "score": 1.0,
+ "content": "final layer of an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 162,
+ 178,
+ 172
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 179,
+ 162,
+ 505,
+ 173
+ ],
+ "score": 1.0,
+ "content": "-RevNet. Indeed, the cascade of convolutional operators has been trained on the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "score": 1.0,
+ "content": "training set to separate the 1000 different classes while being a homeomorphism on its feature space.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 182,
+ 367,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 182,
+ 367,
+ 197
+ ],
+ "score": 1.0,
+ "content": "Thus, the dimensionality of the feature space is potentially large.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 151,
+ 506,
+ 197
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 200,
+ 505,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "As shown in the previous subsection, the final layer is progressively prepared to be projected on",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 211,
+ 506,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 506,
+ 223
+ ],
+ "score": 1.0,
+ "content": "the final probes corresponding to the classes. This indicates that the non-informative variabilities",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 221,
+ 506,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 221,
+ 408,
+ 235
+ ],
+ "score": 1.0,
+ "content": "for classification can be removed via a linear projection on the final layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 223,
+ 417,
+ 232
+ ],
+ "score": 0.77,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 417,
+ 221,
+ 506,
+ 235
+ ],
+ "score": 1.0,
+ "content": ", which lie in a space",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "of dimension 1000, at most. However, this projection has been built via supervision, which can",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "still retain directions that have been contracted and thus will not be selected by an algorithm such as",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 505,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 505,
+ 268
+ ],
+ "score": 1.0,
+ "content": "PCA. We show in fact a PCA retains the necessary information for classification in a small subspace.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 105,
+ 199,
+ 506,
+ 268
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 271,
+ 296,
+ 469
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 297,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 271,
+ 285
+ ],
+ "score": 1.0,
+ "content": "To do so, we build the linear projectors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 271,
+ 274,
+ 283,
+ 283
+ ],
+ "score": 0.85,
+ "content": "\\pi _ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 283,
+ 271,
+ 297,
+ 285
+ ],
+ "score": 1.0,
+ "content": "on",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 283,
+ 296,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 283,
+ 194,
+ 295
+ ],
+ "score": 1.0,
+ "content": "the subspace of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 284,
+ 201,
+ 293
+ ],
+ "score": 0.7,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 283,
+ 296,
+ 295
+ ],
+ "score": 1.0,
+ "content": "first principal compo-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 294,
+ 297,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 297,
+ 305
+ ],
+ "score": 1.0,
+ "content": "nents, and we propose to measure the classifica-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 305,
+ 297,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 305,
+ 297,
+ 317
+ ],
+ "score": 1.0,
+ "content": "tion power of the projected representation with",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 316,
+ 297,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 316,
+ 297,
+ 328
+ ],
+ "score": 1.0,
+ "content": "a supervised classifier, e.g. nearest neighbor or",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 297,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 297,
+ 339
+ ],
+ "score": 1.0,
+ "content": "a linear SVM, on the previous 100 class task.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 335,
+ 295,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 248,
+ 352
+ ],
+ "score": 1.0,
+ "content": "Again, the feature representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 338,
+ 295,
+ 350
+ ],
+ "score": 0.92,
+ "content": "\\{ \\Phi x ^ { n } \\} _ { n \\leq N }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 348,
+ 298,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 348,
+ 298,
+ 361
+ ],
+ "score": 1.0,
+ "content": "are spatially averaged to remove the translation",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 360,
+ 297,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 297,
+ 371
+ ],
+ "score": 1.0,
+ "content": "variability, and standardized on the training set.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 370,
+ 298,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 370,
+ 298,
+ 383
+ ],
+ "score": 1.0,
+ "content": "We apply both classifiers, and we report the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 379,
+ 298,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 379,
+ 214,
+ 395
+ ],
+ "score": 1.0,
+ "content": "classification accuracy of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 214,
+ 381,
+ 272,
+ 394
+ ],
+ "score": 0.93,
+ "content": "\\{ \\pi _ { d } \\Phi x ^ { n } \\} _ { n \\leq N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 379,
+ 298,
+ 395
+ ],
+ "score": 1.0,
+ "content": "w.r.t.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 392,
+ 297,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 392,
+ 118,
+ 405
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 118,
+ 393,
+ 125,
+ 403
+ ],
+ "score": 0.77,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 126,
+ 392,
+ 297,
+ 405
+ ],
+ "score": 1.0,
+ "content": "on the Figure 7. A linear projection re-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 297,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 404,
+ 297,
+ 415
+ ],
+ "score": 1.0,
+ "content": "moves some information that can not be recov-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 414,
+ 297,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 414,
+ 297,
+ 426
+ ],
+ "score": 1.0,
+ "content": "ered by a linear classifier, therefore we observe",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 425,
+ 297,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 297,
+ 437
+ ],
+ "score": 1.0,
+ "content": "that the classification accuracy only decreases",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 437,
+ 297,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 437,
+ 173,
+ 447
+ ],
+ "score": 1.0,
+ "content": "significantly for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 437,
+ 210,
+ 447
+ ],
+ "score": 0.91,
+ "content": "d \\leq 2 0 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 210,
+ 437,
+ 297,
+ 447
+ ],
+ "score": 1.0,
+ "content": ". This shows that the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 447,
+ 297,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 447,
+ 297,
+ 459
+ ],
+ "score": 1.0,
+ "content": "signal indeed lies in a subspace much lower di-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 459,
+ 297,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 459,
+ 297,
+ 470
+ ],
+ "score": 1.0,
+ "content": "mensional than the original feature dimensions",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 105,
+ 271,
+ 298,
+ 470
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 308,
+ 275,
+ 500,
+ 416
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 308,
+ 275,
+ 500,
+ 416
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 275,
+ 500,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 275,
+ 500,
+ 416
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "93bdd4b0bf41956f8e5f34e143db6c00cc121e2956561dec013eee6385bf9daf.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 36.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 308,
+ 275,
+ 500,
+ 289.1
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 308,
+ 289.1,
+ 500,
+ 303.20000000000005
+ ],
+ "spans": [],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 308,
+ 303.20000000000005,
+ 500,
+ 317.30000000000007
+ ],
+ "spans": [],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 308,
+ 317.30000000000007,
+ 500,
+ 331.4000000000001
+ ],
+ "spans": [],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 308,
+ 331.4000000000001,
+ 500,
+ 345.5000000000001
+ ],
+ "spans": [],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 308,
+ 345.5000000000001,
+ 500,
+ 359.60000000000014
+ ],
+ "spans": [],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 308,
+ 359.60000000000014,
+ 500,
+ 373.70000000000016
+ ],
+ "spans": [],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 308,
+ 373.70000000000016,
+ 500,
+ 387.8000000000002
+ ],
+ "spans": [],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 308,
+ 387.8000000000002,
+ 500,
+ 401.9000000000002
+ ],
+ "spans": [],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 308,
+ 401.9000000000002,
+ 500,
+ 416.0000000000002
+ ],
+ "spans": [],
+ "index": 41
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 304,
+ 423,
+ 504,
+ 456
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 423,
+ 505,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 423,
+ 505,
+ 435
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Accuracy of a linear SVM and nearest",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 303,
+ 434,
+ 504,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 434,
+ 504,
+ 447
+ ],
+ "score": 1.0,
+ "content": "neighbor against the number of principal compo-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 304,
+ 446,
+ 365,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 446,
+ 365,
+ 456
+ ],
+ "score": 1.0,
+ "content": "nents retained.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 39.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 469,
+ 504,
+ 492
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 504,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 504,
+ 482
+ ],
+ "score": 1.0,
+ "content": "that can be extracted simply with a PCA that only considers directions of largest variances, illustrat-",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 306,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 306,
+ 493
+ ],
+ "score": 1.0,
+ "content": "ing a successful contraction of the representation.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45.5,
+ "bbox_fs": [
+ 105,
+ 468,
+ 504,
+ 493
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 508,
+ 195,
+ 520
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 198,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 198,
+ 524
+ ],
+ "score": 1.0,
+ "content": "6 CONCLUSION",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 47
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 533,
+ 505,
+ 610
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "Invertible representations and their relationship to loss of information are on the agenda of deep",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "score": 1.0,
+ "content": "learning for some time. Understanding how transformations in feature space are related to the cor-",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 556,
+ 505,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 556,
+ 505,
+ 568
+ ],
+ "score": 1.0,
+ "content": "responding input is an important step towards interpretable deep networks, invertible deep networks",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 567,
+ 506,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 567,
+ 506,
+ 578
+ ],
+ "score": 1.0,
+ "content": "may play an important role in such analysis since, for example, one could potentially back-track a",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "score": 1.0,
+ "content": "property from the feature space to the input space. To the best of our knowledge, this work pro-",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "vides the first empirical evidence that learning invertible representations that do not discard any",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 414,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 599,
+ 414,
+ 613
+ ],
+ "score": 1.0,
+ "content": "information about their input on large-scale supervised problems is possible.",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ }
+ ],
+ "index": 51,
+ "bbox_fs": [
+ 105,
+ 532,
+ 506,
+ 613
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 615,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 243,
+ 629
+ ],
+ "score": 1.0,
+ "content": "To achieve this we introduce the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 243,
+ 617,
+ 248,
+ 626
+ ],
+ "score": 0.7,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 248,
+ 615,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "-RevNet class of CNN which is fully invertible and permits to",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 337,
+ 639
+ ],
+ "score": 1.0,
+ "content": "exactly recover the input from its last convolutional layer.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 628,
+ 342,
+ 637
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 627,
+ 506,
+ 639
+ ],
+ "score": 1.0,
+ "content": "-RevNets achieve the same classification",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "accuracy in the classification of complex datasets as illustrated on ILSVRC-2012, when compared",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "score": 1.0,
+ "content": "to the RevNet (Gomez et al., 2017) and ResNet (He et al., 2016) architectures with a similar number",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 506,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 424,
+ 674
+ ],
+ "score": 1.0,
+ "content": "of layers. Furthermore, the inverse network is obtained for free when training an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 661,
+ 429,
+ 670
+ ],
+ "score": 0.67,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 658,
+ 506,
+ 674
+ ],
+ "score": 1.0,
+ "content": "-RevNet, requiring",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 400,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 400,
+ 683
+ ],
+ "score": 1.0,
+ "content": "only minimal adaption to recover inputs from the hidden representations.",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ }
+ ],
+ "index": 57.5,
+ "bbox_fs": [
+ 105,
+ 615,
+ 506,
+ 683
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "The absence of loss of information is surprising, given the wide believe, that discarding information",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "is essential for learning representations that generalize well to unseen data. We show that this is not",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "the case and propose to explain the generalization property with empirical evidence of progressive",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 318,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 318,
+ 732
+ ],
+ "score": 1.0,
+ "content": "separation and contraction with depth, on ImageNet.",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ }
+ ],
+ "index": 62.5,
+ "bbox_fs": [
+ 105,
+ 687,
+ 505,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 225,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 80,
+ 226,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 80,
+ 226,
+ 96
+ ],
+ "score": 1.0,
+ "content": "ACKNOWLEDGEMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 106,
+ 505,
+ 172
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 119
+ ],
+ "score": 1.0,
+ "content": "Jorn-Henrik Jacobsen was partially funded by the STW perspective program ImaGene. Edouard ¨",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 130
+ ],
+ "score": 1.0,
+ "content": "Oyallon was partially funded by the ERC grant InvariantClass 320959, via a grant for PhD Students",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 141
+ ],
+ "score": 1.0,
+ "content": "of the Conseil regional dIle-de-France (RDM-IdF), and a postdoctoral grant from the from DPEI ´",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 152
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 152
+ ],
+ "score": 1.0,
+ "content": "of Inria (AAR 2017POD057) for the collaboration with CWI. We thank Berkay Kicanaoglu for the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 163
+ ],
+ "score": 1.0,
+ "content": "Basel Face data, Mathieu Andreux, Eugene Belilovsky, Amal Rannen, Patrick Putzky and Kyriacos",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 281,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 281,
+ 174
+ ],
+ "score": 1.0,
+ "content": "Shiarlis for feedback on drafts of the paper.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 189,
+ 175,
+ 201
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 189,
+ 176,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 189,
+ 176,
+ 202
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 208,
+ 504,
+ 230
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 207,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 207,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "Alessandro Achille and Stefano Soatto. On the emergence of invariance and disentangling in deep",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 219,
+ 345,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 219,
+ 345,
+ 231
+ ],
+ "score": 1.0,
+ "content": "representations. arXiv preprint arXiv:1706.01350, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 237,
+ 504,
+ 271
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 504,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 504,
+ 252
+ ],
+ "score": 1.0,
+ "content": "Mathieu Aubry and Bryan C Russell. Understanding deep features with computer-generated im-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 114,
+ 249,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 249,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "agery. In Proceedings of the IEEE International Conference on Computer Vision, pp. 2875–2883,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 260,
+ 142,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 260,
+ 142,
+ 272
+ ],
+ "score": 1.0,
+ "content": "2015.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 279,
+ 505,
+ 302
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 278,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 278,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "Joan Bruna, Arthur Szlam, and Yann LeCun. Signal recovery from pooling representations. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 291,
+ 249,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 291,
+ 249,
+ 302
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1311.4025, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 344
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. De-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 321,
+ 505,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 321,
+ 505,
+ 334
+ ],
+ "score": 1.0,
+ "content": "scribing textures in the wild. In Proceedings of the IEEE Conference on Computer Vision and",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 116,
+ 332,
+ 291,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 332,
+ 291,
+ 344
+ ],
+ "score": 1.0,
+ "content": "Pattern Recognition, pp. 3606–3613, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 351,
+ 504,
+ 385
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "Moustapha Cisse, Piotr Bojanowski, Edouard Grave, Yann Dauphin, and Nicolas Usunier. Parseval",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 362,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 362,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "networks: Improving robustness to adversarial examples. In International Conference on Machine",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 115,
+ 373,
+ 238,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 373,
+ 238,
+ 385
+ ],
+ "score": 1.0,
+ "content": "Learning, pp. 854–863, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 392,
+ 502,
+ 415
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 503,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 503,
+ 406
+ ],
+ "score": 1.0,
+ "content": "Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components esti-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 404,
+ 308,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 404,
+ 308,
+ 415
+ ],
+ "score": 1.0,
+ "content": "mation. arXiv preprint arXiv:1410.8516, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 422,
+ 503,
+ 445
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 421,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 421,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 434,
+ 254,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 434,
+ 254,
+ 445
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1605.08803, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 452,
+ 506,
+ 487
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 504,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 504,
+ 466
+ ],
+ "score": 1.0,
+ "content": "Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 463,
+ 505,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 463,
+ 505,
+ 477
+ ],
+ "score": 1.0,
+ "content": "In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4829–",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 474,
+ 167,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 474,
+ 167,
+ 488
+ ],
+ "score": 1.0,
+ "content": "4837, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 494,
+ 504,
+ 517
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 495,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 495,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "Aidan N Gomez, Mengye Ren, Raquel Urtasun, and Roger B Grosse. The reversible residual net-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 116,
+ 506,
+ 487,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 506,
+ 487,
+ 518
+ ],
+ "score": 1.0,
+ "content": "work: Backpropagation without storing activations. arXiv preprint arXiv:1707.04585, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 524,
+ 505,
+ 559
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 504,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 504,
+ 539
+ ],
+ "score": 1.0,
+ "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 534,
+ 505,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 534,
+ 505,
+ 550
+ ],
+ "score": 1.0,
+ "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 545,
+ 182,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 545,
+ 182,
+ 559
+ ],
+ "score": 1.0,
+ "content": "770–778, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 565,
+ 504,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 580
+ ],
+ "score": 1.0,
+ "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 115,
+ 577,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 577,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 588,
+ 142,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 588,
+ 142,
+ 600
+ ],
+ "score": 1.0,
+ "content": "2015.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 607,
+ 504,
+ 641
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 504,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 504,
+ 621
+ ],
+ "score": 1.0,
+ "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 618,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 618,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 114,
+ 627,
+ 143,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 627,
+ 143,
+ 642
+ ],
+ "score": 1.0,
+ "content": "2012.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 648,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 659,
+ 506,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 659,
+ 506,
+ 674
+ ],
+ "score": 1.0,
+ "content": "them. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 116,
+ 671,
+ 191,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 671,
+ 191,
+ 682
+ ],
+ "score": 1.0,
+ "content": "5188–5196, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 690,
+ 505,
+ 713
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 704
+ ],
+ "score": 1.0,
+ "content": "Aravindh Mahendran and Andrea Vedaldi. Visualizing deep convolutional neural networks using",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 702,
+ 462,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 702,
+ 462,
+ 714
+ ],
+ "score": 1.0,
+ "content": "natural pre-images. International Journal of Computer Vision, 120(3):233–255, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 720,
+ 417,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 719,
+ 417,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 719,
+ 417,
+ 734
+ ],
+ "score": 1.0,
+ "content": "Stephane Mallat. ´ A wavelet tour of signal processing. Academic press, 1999.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "page_idx": 9,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 301,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 765
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 765
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 15,
+ "width": 14
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 225,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 80,
+ 226,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 80,
+ 226,
+ 96
+ ],
+ "score": 1.0,
+ "content": "ACKNOWLEDGEMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 106,
+ 505,
+ 172
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 119
+ ],
+ "score": 1.0,
+ "content": "Jorn-Henrik Jacobsen was partially funded by the STW perspective program ImaGene. Edouard ¨",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 130
+ ],
+ "score": 1.0,
+ "content": "Oyallon was partially funded by the ERC grant InvariantClass 320959, via a grant for PhD Students",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 141
+ ],
+ "score": 1.0,
+ "content": "of the Conseil regional dIle-de-France (RDM-IdF), and a postdoctoral grant from the from DPEI ´",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 152
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 152
+ ],
+ "score": 1.0,
+ "content": "of Inria (AAR 2017POD057) for the collaboration with CWI. We thank Berkay Kicanaoglu for the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 163
+ ],
+ "score": 1.0,
+ "content": "Basel Face data, Mathieu Andreux, Eugene Belilovsky, Amal Rannen, Patrick Putzky and Kyriacos",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 281,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 281,
+ 174
+ ],
+ "score": 1.0,
+ "content": "Shiarlis for feedback on drafts of the paper.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 105,
+ 105,
+ 506,
+ 174
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 189,
+ 175,
+ 201
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 189,
+ 176,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 189,
+ 176,
+ 202
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 208,
+ 504,
+ 230
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 207,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 207,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "Alessandro Achille and Stefano Soatto. On the emergence of invariance and disentangling in deep",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 219,
+ 345,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 219,
+ 345,
+ 231
+ ],
+ "score": 1.0,
+ "content": "representations. arXiv preprint arXiv:1706.01350, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 106,
+ 207,
+ 505,
+ 231
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 237,
+ 504,
+ 271
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 504,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 504,
+ 252
+ ],
+ "score": 1.0,
+ "content": "Mathieu Aubry and Bryan C Russell. Understanding deep features with computer-generated im-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 114,
+ 249,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 249,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "agery. In Proceedings of the IEEE International Conference on Computer Vision, pp. 2875–2883,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 260,
+ 142,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 260,
+ 142,
+ 272
+ ],
+ "score": 1.0,
+ "content": "2015.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 237,
+ 505,
+ 272
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 279,
+ 505,
+ 302
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 278,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 278,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "Joan Bruna, Arthur Szlam, and Yann LeCun. Signal recovery from pooling representations. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 291,
+ 249,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 291,
+ 249,
+ 302
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1311.4025, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 105,
+ 278,
+ 505,
+ 302
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 344
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. De-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 321,
+ 505,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 321,
+ 505,
+ 334
+ ],
+ "score": 1.0,
+ "content": "scribing textures in the wild. In Proceedings of the IEEE Conference on Computer Vision and",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 116,
+ 332,
+ 291,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 332,
+ 291,
+ 344
+ ],
+ "score": 1.0,
+ "content": "Pattern Recognition, pp. 3606–3613, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 106,
+ 309,
+ 505,
+ 344
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 351,
+ 504,
+ 385
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "Moustapha Cisse, Piotr Bojanowski, Edouard Grave, Yann Dauphin, and Nicolas Usunier. Parseval",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 362,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 362,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "networks: Improving robustness to adversarial examples. In International Conference on Machine",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 115,
+ 373,
+ 238,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 373,
+ 238,
+ 385
+ ],
+ "score": 1.0,
+ "content": "Learning, pp. 854–863, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 106,
+ 351,
+ 506,
+ 385
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 392,
+ 502,
+ 415
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 503,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 503,
+ 406
+ ],
+ "score": 1.0,
+ "content": "Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components esti-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 404,
+ 308,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 404,
+ 308,
+ 415
+ ],
+ "score": 1.0,
+ "content": "mation. arXiv preprint arXiv:1410.8516, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21.5,
+ "bbox_fs": [
+ 105,
+ 391,
+ 503,
+ 415
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 422,
+ 503,
+ 445
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 421,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 421,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 434,
+ 254,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 434,
+ 254,
+ 445
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1605.08803, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 104,
+ 421,
+ 505,
+ 445
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 452,
+ 506,
+ 487
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 504,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 504,
+ 466
+ ],
+ "score": 1.0,
+ "content": "Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 463,
+ 505,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 463,
+ 505,
+ 477
+ ],
+ "score": 1.0,
+ "content": "In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4829–",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 474,
+ 167,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 474,
+ 167,
+ 488
+ ],
+ "score": 1.0,
+ "content": "4837, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 106,
+ 452,
+ 505,
+ 488
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 494,
+ 504,
+ 517
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 495,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 495,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "Aidan N Gomez, Mengye Ren, Raquel Urtasun, and Roger B Grosse. The reversible residual net-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 116,
+ 506,
+ 487,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 506,
+ 487,
+ 518
+ ],
+ "score": 1.0,
+ "content": "work: Backpropagation without storing activations. arXiv preprint arXiv:1707.04585, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 106,
+ 495,
+ 505,
+ 518
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 524,
+ 505,
+ 559
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 504,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 504,
+ 539
+ ],
+ "score": 1.0,
+ "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 534,
+ 505,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 534,
+ 505,
+ 550
+ ],
+ "score": 1.0,
+ "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 545,
+ 182,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 545,
+ 182,
+ 559
+ ],
+ "score": 1.0,
+ "content": "770–778, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 105,
+ 523,
+ 505,
+ 559
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 565,
+ 504,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 580
+ ],
+ "score": 1.0,
+ "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 115,
+ 577,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 577,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 588,
+ 142,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 588,
+ 142,
+ 600
+ ],
+ "score": 1.0,
+ "content": "2015.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 105,
+ 565,
+ 505,
+ 600
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 607,
+ 504,
+ 641
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 504,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 504,
+ 621
+ ],
+ "score": 1.0,
+ "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 618,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 618,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 114,
+ 627,
+ 143,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 627,
+ 143,
+ 642
+ ],
+ "score": 1.0,
+ "content": "2012.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 105,
+ 606,
+ 506,
+ 642
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 648,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 659,
+ 506,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 659,
+ 506,
+ 674
+ ],
+ "score": 1.0,
+ "content": "them. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 116,
+ 671,
+ 191,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 671,
+ 191,
+ 682
+ ],
+ "score": 1.0,
+ "content": "5188–5196, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 105,
+ 647,
+ 506,
+ 682
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 690,
+ 505,
+ 713
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 704
+ ],
+ "score": 1.0,
+ "content": "Aravindh Mahendran and Andrea Vedaldi. Visualizing deep convolutional neural networks using",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 702,
+ 462,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 702,
+ 462,
+ 714
+ ],
+ "score": 1.0,
+ "content": "natural pre-images. International Journal of Computer Vision, 120(3):233–255, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 106,
+ 688,
+ 505,
+ 714
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 720,
+ 417,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 719,
+ 417,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 719,
+ 417,
+ 734
+ ],
+ "score": 1.0,
+ "content": "Stephane Mallat. ´ A wavelet tour of signal processing. Academic press, 1999.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 44,
+ "bbox_fs": [
+ 106,
+ 719,
+ 417,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Stephane Mallat. Group invariant scattering. ´ Communications on Pure and Applied Mathematics,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 93,
+ 220,
+ 104
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 93,
+ 220,
+ 104
+ ],
+ "score": 1.0,
+ "content": "65(10):1331–1398, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 111,
+ 504,
+ 135
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "score": 1.0,
+ "content": "Stephane Mallat. Understanding deep convolutional networks. ´ Phil. Trans. R. Soc. A, 374(2065):",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 116,
+ 123,
+ 186,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 123,
+ 186,
+ 135
+ ],
+ "score": 1.0,
+ "content": "20150203, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 141,
+ 504,
+ 165
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 140,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 140,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "Alfred J Menezes, Paul C Van Oorschot, and Scott A Vanstone. Handbook of applied cryptography.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 116,
+ 153,
+ 189,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 153,
+ 189,
+ 165
+ ],
+ "score": 1.0,
+ "content": "CRC press, 1996.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 171,
+ 505,
+ 195
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 170,
+ 505,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 170,
+ 505,
+ 186
+ ],
+ "score": 1.0,
+ "content": "Edouard Oyallon. Building a regular decision boundary with deep networks. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 116,
+ 183,
+ 220,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 183,
+ 220,
+ 195
+ ],
+ "score": 1.0,
+ "content": "arXiv:1703.01775, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 201,
+ 504,
+ 236
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 506,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 506,
+ 215
+ ],
+ "score": 1.0,
+ "content": "Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 213,
+ 506,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 213,
+ 506,
+ 226
+ ],
+ "score": 1.0,
+ "content": "model for pose and illumination invariant face recognition. In Advanced Video and Signal Based",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 223,
+ 502,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 223,
+ 502,
+ 236
+ ],
+ "score": 1.0,
+ "content": "Surveillance, 2009. AVSS’09. Sixth IEEE International Conference on, pp. 296–301. Ieee, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 242,
+ 504,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 241,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 241,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 254,
+ 468,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 254,
+ 468,
+ 266
+ ],
+ "score": 1.0,
+ "content": "convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 272,
+ 504,
+ 307
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 286
+ ],
+ "score": 1.0,
+ "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 116,
+ 284,
+ 505,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 284,
+ 505,
+ 296
+ ],
+ "score": 1.0,
+ "content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 115,
+ 295,
+ 474,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 295,
+ 474,
+ 308
+ ],
+ "score": 1.0,
+ "content": "recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 358
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel ´",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 324,
+ 505,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 324,
+ 505,
+ 337
+ ],
+ "score": 1.0,
+ "content": "Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 116,
+ 335,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 335,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "sub-pixel convolutional neural network. In Proceedings of the IEEE Conference on Computer",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 116,
+ 346,
+ 336,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 346,
+ 336,
+ 359
+ ],
+ "score": 1.0,
+ "content": "Vision and Pattern Recognition, pp. 1874–1883, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 365,
+ 503,
+ 388
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 365,
+ 505,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 365,
+ 505,
+ 378
+ ],
+ "score": 1.0,
+ "content": "Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via informa-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 116,
+ 377,
+ 300,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 377,
+ 300,
+ 387
+ ],
+ "score": 1.0,
+ "content": "tion. arXiv preprint arXiv:1703.00810, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 395,
+ 504,
+ 418
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "Wim Sweldens. The lifting scheme: A construction of second generation wavelets. SIAM journal",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 115,
+ 406,
+ 312,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 406,
+ 312,
+ 418
+ ],
+ "score": 1.0,
+ "content": "on mathematical analysis, 29(2):511–546, 1998.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 425,
+ 505,
+ 448
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 505,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 505,
+ 438
+ ],
+ "score": 1.0,
+ "content": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 115,
+ 437,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 437,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 455,
+ 505,
+ 478
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 505,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 505,
+ 468
+ ],
+ "score": 1.0,
+ "content": "Naftali Tishby and Noga Zaslavsky. Deep learning and the information bottleneck principle. In",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 116,
+ 466,
+ 403,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 466,
+ 403,
+ 478
+ ],
+ "score": 1.0,
+ "content": "Information Theory Workshop (ITW), 2015 IEEE, pp. 1–5. IEEE, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 484,
+ 504,
+ 508
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 114,
+ 496,
+ 254,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 496,
+ 254,
+ 508
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1511.07122, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 514,
+ 505,
+ 538
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 513,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 514,
+ 297,
+ 527
+ ],
+ "score": 1.0,
+ "content": "Sergey Zagoruyko and Nikos Komodakis.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 312,
+ 513,
+ 426,
+ 527
+ ],
+ "score": 1.0,
+ "content": "Wide residual networks.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 438,
+ 514,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 525,
+ 219,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 525,
+ 219,
+ 537
+ ],
+ "score": 1.0,
+ "content": "arXiv:1605.07146, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 568
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 116,
+ 555,
+ 403,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 555,
+ 403,
+ 568
+ ],
+ "score": 1.0,
+ "content": "European conference on computer vision, pp. 818–833. Springer, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5
+ }
+ ],
+ "page_idx": 10,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 26,
+ 293,
+ 38
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 39
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 39
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 310,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 765
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 765
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 15,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Stephane Mallat. Group invariant scattering. ´ Communications on Pure and Applied Mathematics,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 93,
+ 220,
+ 104
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 93,
+ 220,
+ 104
+ ],
+ "score": 1.0,
+ "content": "65(10):1331–1398, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 105,
+ 81,
+ 506,
+ 104
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 111,
+ 504,
+ 135
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "score": 1.0,
+ "content": "Stephane Mallat. Understanding deep convolutional networks. ´ Phil. Trans. R. Soc. A, 374(2065):",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 116,
+ 123,
+ 186,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 123,
+ 186,
+ 135
+ ],
+ "score": 1.0,
+ "content": "20150203, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 106,
+ 112,
+ 505,
+ 135
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 141,
+ 504,
+ 165
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 140,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 140,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "Alfred J Menezes, Paul C Van Oorschot, and Scott A Vanstone. Handbook of applied cryptography.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 116,
+ 153,
+ 189,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 153,
+ 189,
+ 165
+ ],
+ "score": 1.0,
+ "content": "CRC press, 1996.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 105,
+ 140,
+ 505,
+ 165
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 171,
+ 505,
+ 195
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 170,
+ 505,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 170,
+ 505,
+ 186
+ ],
+ "score": 1.0,
+ "content": "Edouard Oyallon. Building a regular decision boundary with deep networks. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 116,
+ 183,
+ 220,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 183,
+ 220,
+ 195
+ ],
+ "score": 1.0,
+ "content": "arXiv:1703.01775, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 105,
+ 170,
+ 505,
+ 195
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 201,
+ 504,
+ 236
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 506,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 506,
+ 215
+ ],
+ "score": 1.0,
+ "content": "Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 213,
+ 506,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 213,
+ 506,
+ 226
+ ],
+ "score": 1.0,
+ "content": "model for pose and illumination invariant face recognition. In Advanced Video and Signal Based",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 223,
+ 502,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 223,
+ 502,
+ 236
+ ],
+ "score": 1.0,
+ "content": "Surveillance, 2009. AVSS’09. Sixth IEEE International Conference on, pp. 296–301. Ieee, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9,
+ "bbox_fs": [
+ 105,
+ 201,
+ 506,
+ 236
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 242,
+ 504,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 241,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 241,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 254,
+ 468,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 254,
+ 468,
+ 266
+ ],
+ "score": 1.0,
+ "content": "convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 106,
+ 241,
+ 505,
+ 266
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 272,
+ 504,
+ 307
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 286
+ ],
+ "score": 1.0,
+ "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 116,
+ 284,
+ 505,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 284,
+ 505,
+ 296
+ ],
+ "score": 1.0,
+ "content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 115,
+ 295,
+ 474,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 295,
+ 474,
+ 308
+ ],
+ "score": 1.0,
+ "content": "recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14,
+ "bbox_fs": [
+ 106,
+ 271,
+ 505,
+ 308
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 358
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel ´",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 324,
+ 505,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 324,
+ 505,
+ 337
+ ],
+ "score": 1.0,
+ "content": "Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 116,
+ 335,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 335,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "sub-pixel convolutional neural network. In Proceedings of the IEEE Conference on Computer",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 116,
+ 346,
+ 336,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 346,
+ 336,
+ 359
+ ],
+ "score": 1.0,
+ "content": "Vision and Pattern Recognition, pp. 1874–1883, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 106,
+ 313,
+ 505,
+ 359
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 365,
+ 503,
+ 388
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 365,
+ 505,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 365,
+ 505,
+ 378
+ ],
+ "score": 1.0,
+ "content": "Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via informa-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 116,
+ 377,
+ 300,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 377,
+ 300,
+ 387
+ ],
+ "score": 1.0,
+ "content": "tion. arXiv preprint arXiv:1703.00810, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 106,
+ 365,
+ 505,
+ 387
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 395,
+ 504,
+ 418
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "Wim Sweldens. The lifting scheme: A construction of second generation wavelets. SIAM journal",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 115,
+ 406,
+ 312,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 406,
+ 312,
+ 418
+ ],
+ "score": 1.0,
+ "content": "on mathematical analysis, 29(2):511–546, 1998.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 106,
+ 394,
+ 505,
+ 418
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 425,
+ 505,
+ 448
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 505,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 505,
+ 438
+ ],
+ "score": 1.0,
+ "content": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 115,
+ 437,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 437,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 106,
+ 425,
+ 505,
+ 448
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 455,
+ 505,
+ 478
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 505,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 505,
+ 468
+ ],
+ "score": 1.0,
+ "content": "Naftali Tishby and Noga Zaslavsky. Deep learning and the information bottleneck principle. In",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 116,
+ 466,
+ 403,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 466,
+ 403,
+ 478
+ ],
+ "score": 1.0,
+ "content": "Information Theory Workshop (ITW), 2015 IEEE, pp. 1–5. IEEE, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 106,
+ 455,
+ 505,
+ 478
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 484,
+ 504,
+ 508
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 114,
+ 496,
+ 254,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 496,
+ 254,
+ 508
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1511.07122, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 105,
+ 484,
+ 505,
+ 508
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 514,
+ 505,
+ 538
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 513,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 514,
+ 297,
+ 527
+ ],
+ "score": 1.0,
+ "content": "Sergey Zagoruyko and Nikos Komodakis.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 312,
+ 513,
+ 426,
+ 527
+ ],
+ "score": 1.0,
+ "content": "Wide residual networks.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 438,
+ 514,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 525,
+ 219,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 525,
+ 219,
+ 537
+ ],
+ "score": 1.0,
+ "content": "arXiv:1605.07146, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 106,
+ 513,
+ 505,
+ 537
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 568
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 116,
+ 555,
+ 403,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 555,
+ 403,
+ 568
+ ],
+ "score": 1.0,
+ "content": "European conference on computer vision, pp. 818–833. Springer, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 105,
+ 544,
+ 505,
+ 568
+ ]
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file