diff --git "a/parse/dev/JavFPcsscd5/JavFPcsscd5_middle.json" "b/parse/dev/JavFPcsscd5/JavFPcsscd5_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/dev/JavFPcsscd5/JavFPcsscd5_middle.json"
@@ -0,0 +1,34824 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 97,
+ 504,
+ 137
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 96,
+ 504,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 96,
+ 504,
+ 119
+ ],
+ "score": 1.0,
+ "content": "The Effects of Regularization and Data Augmentation",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 227,
+ 117,
+ 384,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 227,
+ 117,
+ 384,
+ 139
+ ],
+ "score": 1.0,
+ "content": "are Class Dependent",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 119,
+ 179,
+ 230,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 178,
+ 218,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 178,
+ 218,
+ 191
+ ],
+ "score": 1.0,
+ "content": "Randall Balestriero",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 123,
+ 190,
+ 226,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 190,
+ 226,
+ 201
+ ],
+ "score": 1.0,
+ "content": "Meta AI Research, FAIR",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 150,
+ 201,
+ 199,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 150,
+ 201,
+ 199,
+ 212
+ ],
+ "score": 1.0,
+ "content": "NYC, USA",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 118,
+ 212,
+ 231,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 212,
+ 231,
+ 223
+ ],
+ "score": 1.0,
+ "content": "rbalestriero@meta.com",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 246,
+ 180,
+ 348,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 269,
+ 178,
+ 326,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 269,
+ 178,
+ 326,
+ 191
+ ],
+ "score": 1.0,
+ "content": "Leon Bottou",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 246,
+ 190,
+ 348,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 246,
+ 190,
+ 348,
+ 201
+ ],
+ "score": 1.0,
+ "content": "Meta AI Research, FAIR",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 271,
+ 201,
+ 322,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 271,
+ 201,
+ 322,
+ 213
+ ],
+ "score": 1.0,
+ "content": "NYC, USA",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 259,
+ 212,
+ 336,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 259,
+ 212,
+ 336,
+ 224
+ ],
+ "score": 1.0,
+ "content": "leonb@meta.com",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 364,
+ 180,
+ 492,
+ 224
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 400,
+ 179,
+ 457,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 400,
+ 179,
+ 457,
+ 191
+ ],
+ "score": 1.0,
+ "content": "Yann LeCun",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 363,
+ 190,
+ 493,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 363,
+ 190,
+ 493,
+ 202
+ ],
+ "score": 1.0,
+ "content": "Meta AI Research, FAIR, NYU",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 403,
+ 201,
+ 453,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 403,
+ 201,
+ 453,
+ 213
+ ],
+ "score": 1.0,
+ "content": "NYC, USA",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 387,
+ 213,
+ 469,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 213,
+ 469,
+ 224
+ ],
+ "score": 1.0,
+ "content": "ylecun@meta.com",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 283,
+ 252,
+ 328,
+ 265
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 281,
+ 251,
+ 331,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 281,
+ 251,
+ 331,
+ 267
+ ],
+ "score": 1.0,
+ "content": "Abstract",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 142,
+ 278,
+ 469,
+ 485
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 278,
+ 471,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 278,
+ 471,
+ 290
+ ],
+ "score": 1.0,
+ "content": "Regularization is a fundamental technique to improve a model’s generalization per-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 301
+ ],
+ "score": 1.0,
+ "content": "formances by limiting its complexity. Deep Neural Networks (DNNs), which tend",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 299,
+ 470,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 299,
+ 470,
+ 312
+ ],
+ "score": 1.0,
+ "content": "to overfit their training data, heavily rely on regularizers such as Data-Augmentation",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 310,
+ 470,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 310,
+ 470,
+ 323
+ ],
+ "score": 1.0,
+ "content": "(DA) or weight-decay with hyper-parameters found from structural risk minimiza-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 142,
+ 322,
+ 470,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 322,
+ 470,
+ 333
+ ],
+ "score": 1.0,
+ "content": "tion, i.e., cross-validation. In this study, we demonstrate that the optimal regular-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 332,
+ 470,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 332,
+ 470,
+ 345
+ ],
+ "score": 1.0,
+ "content": "ization’s hyper-parameters found from cross-validation over all classes leads to",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 142,
+ 343,
+ 470,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 343,
+ 470,
+ 356
+ ],
+ "score": 1.0,
+ "content": "disastrous model performances on a minority of classes. For example, a resnet50",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 354,
+ 469,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 354,
+ 398,
+ 366
+ ],
+ "score": 1.0,
+ "content": "trained on Imagenet sees its “barn spider” test accuracy falls from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 399,
+ 354,
+ 419,
+ 365
+ ],
+ "score": 0.88,
+ "content": "6 8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 354,
+ 429,
+ 366
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 354,
+ 449,
+ 365
+ ],
+ "score": 0.88,
+ "content": "4 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 449,
+ 354,
+ 469,
+ 366
+ ],
+ "score": 1.0,
+ "content": "only",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 142,
+ 365,
+ 469,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 365,
+ 469,
+ 377
+ ],
+ "score": 1.0,
+ "content": "by introducing random crop DA during training. Even more surprising, such unfair",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 389
+ ],
+ "score": 1.0,
+ "content": "impact of regularization also appears when introducing uninformative regularizers",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 387,
+ 470,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 387,
+ 470,
+ 399
+ ],
+ "score": 1.0,
+ "content": "such as weight decay or dropout. Those results demonstrate that our search for ever",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 398,
+ 470,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 398,
+ 470,
+ 411
+ ],
+ "score": 1.0,
+ "content": "increasing generalization performance —averaged over all classes and samples—",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 141,
+ 408,
+ 469,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 408,
+ 469,
+ 421
+ ],
+ "score": 1.0,
+ "content": "has left us with models and regularizers that silently sacrifice performances on",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 141,
+ 420,
+ 469,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 420,
+ 469,
+ 432
+ ],
+ "score": 1.0,
+ "content": "some classes. This scenario can become dangerous when deploying a model on",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 141,
+ 431,
+ 470,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 431,
+ 470,
+ 442
+ ],
+ "score": 1.0,
+ "content": "downstream tasks e.g. an Imagenet pre-trained resnet50 deployed on INaturalist",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 141,
+ 441,
+ 470,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 441,
+ 267,
+ 453
+ ],
+ "score": 1.0,
+ "content": "sees its performances fall from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 267,
+ 441,
+ 287,
+ 452
+ ],
+ "score": 0.89,
+ "content": "7 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 287,
+ 441,
+ 298,
+ 453
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 298,
+ 441,
+ 318,
+ 452
+ ],
+ "score": 0.89,
+ "content": "3 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 318,
+ 441,
+ 352,
+ 453
+ ],
+ "score": 1.0,
+ "content": "on class",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 441,
+ 379,
+ 452
+ ],
+ "score": 0.51,
+ "content": "\\# 8 8 8 9",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 379,
+ 441,
+ 470,
+ 453
+ ],
+ "score": 1.0,
+ "content": "when introducing ran-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 141,
+ 452,
+ 470,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 452,
+ 470,
+ 465
+ ],
+ "score": 1.0,
+ "content": "dom crop DA during the Imagenet pre-training phase. Those results demonstrate",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 141,
+ 464,
+ 469,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 464,
+ 469,
+ 475
+ ],
+ "score": 1.0,
+ "content": "that finding a correct measure of a model’s complexity without class-dependent",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 141,
+ 475,
+ 329,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 475,
+ 329,
+ 487
+ ],
+ "score": 1.0,
+ "content": "preference remains an open research question.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 507,
+ 190,
+ 520
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 192,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 192,
+ 522
+ ],
+ "score": 1.0,
+ "content": "1 Introduction",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 532,
+ 505,
+ 663
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "Machine learning and deep learning aim at learning systems to solve as accurately as possible a",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 543,
+ 507,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 507,
+ 556
+ ],
+ "score": 1.0,
+ "content": "given task at hand [LeCun et al., 1998, Bishop and Nasrabadi, 2006, Jordan and Mitchell, 2015].",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 555,
+ 506,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 555,
+ 506,
+ 567
+ ],
+ "score": 1.0,
+ "content": "This process often takes the form of (i) being given a finite dataset, a (differentiable) loss function,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 579
+ ],
+ "score": 1.0,
+ "content": "and a performance measure, (ii) splitting the dataset into train/valid/test sets to optimizing the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 506,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 506,
+ 589
+ ],
+ "score": 1.0,
+ "content": "system’s parameters e.g. from gradient updates of the loss on the train set while cross-validating",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 599
+ ],
+ "score": 1.0,
+ "content": "hyper-parameters using the valid set, and (iii) assessing the system’s performance on the test set. As",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "the training set is finite, and the optimal design of the system is unknown, it is common to employ",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "regularization during the optimization phase to reduce over-fitting [Tikhonov, 1943, Tihonov, 1963]",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "i.e. to decrease the system’s performance gap between train set and test set samples [Simard et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 630,
+ 506,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 630,
+ 506,
+ 643
+ ],
+ "score": 1.0,
+ "content": "1991, Chapelle et al., 2000, Bottou, 2012, Neyshabur et al., 2014]. Central to our study is the fact that",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 642,
+ 505,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 505,
+ 654
+ ],
+ "score": 1.0,
+ "content": "hyper-parameter selection is done via cross-validation by maximizing the valid set performance with",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 477,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 653,
+ 477,
+ 665
+ ],
+ "score": 1.0,
+ "content": "ad-hoc statistics e.g. the average accuracy over all samples for classes in classification tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 669,
+ 506,
+ 713
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 668,
+ 507,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 668,
+ 507,
+ 681
+ ],
+ "score": 1.0,
+ "content": "Cross-validation commonly involves many different types of regularization along with their “strengths”",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 680,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "[Goodfellow et al., 2016, He et al., 2021]. Most variants of regularization take one of two forms:",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 104,
+ 689,
+ 506,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 689,
+ 506,
+ 705
+ ],
+ "score": 1.0,
+ "content": "Data-Augmentation (DA) and weight-decay. DA is a data-driven and informed regularization strategy",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 702,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 702,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "that artificially increase the number of training samples [Shorten and Khoshgoftaar, 2019]. As",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 48.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 731,
+ 385,
+ 742
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 730,
+ 386,
+ 743
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 730,
+ 386,
+ 743
+ ],
+ "score": 1.0,
+ "content": "36th Conference on Neural Information Processing Systems (NeurIPS 2022).",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 97,
+ 504,
+ 137
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 96,
+ 504,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 96,
+ 504,
+ 119
+ ],
+ "score": 1.0,
+ "content": "The Effects of Regularization and Data Augmentation",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 227,
+ 117,
+ 384,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 227,
+ 117,
+ 384,
+ 139
+ ],
+ "score": 1.0,
+ "content": "are Class Dependent",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 119,
+ 179,
+ 230,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 178,
+ 218,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 178,
+ 218,
+ 191
+ ],
+ "score": 1.0,
+ "content": "Randall Balestriero",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 123,
+ 190,
+ 226,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 190,
+ 226,
+ 201
+ ],
+ "score": 1.0,
+ "content": "Meta AI Research, FAIR",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 150,
+ 201,
+ 199,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 150,
+ 201,
+ 199,
+ 212
+ ],
+ "score": 1.0,
+ "content": "NYC, USA",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 118,
+ 212,
+ 231,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 212,
+ 231,
+ 223
+ ],
+ "score": 1.0,
+ "content": "rbalestriero@meta.com",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 118,
+ 178,
+ 231,
+ 223
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 246,
+ 180,
+ 348,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 269,
+ 178,
+ 326,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 269,
+ 178,
+ 326,
+ 191
+ ],
+ "score": 1.0,
+ "content": "Leon Bottou",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 246,
+ 190,
+ 348,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 246,
+ 190,
+ 348,
+ 201
+ ],
+ "score": 1.0,
+ "content": "Meta AI Research, FAIR",
+ "type": "text"
+ }
+ ],
+ "index": 6,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 271,
+ 201,
+ 322,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 271,
+ 201,
+ 322,
+ 213
+ ],
+ "score": 1.0,
+ "content": "NYC, USA",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 259,
+ 212,
+ 336,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 259,
+ 212,
+ 336,
+ 224
+ ],
+ "score": 1.0,
+ "content": "leonb@meta.com",
+ "type": "text"
+ }
+ ],
+ "index": 12,
+ "is_list_start_line": true
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 246,
+ 178,
+ 348,
+ 224
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 364,
+ 180,
+ 492,
+ 224
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 400,
+ 179,
+ 457,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 400,
+ 179,
+ 457,
+ 191
+ ],
+ "score": 1.0,
+ "content": "Yann LeCun",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 363,
+ 190,
+ 493,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 363,
+ 190,
+ 493,
+ 202
+ ],
+ "score": 1.0,
+ "content": "Meta AI Research, FAIR, NYU",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 403,
+ 201,
+ 453,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 403,
+ 201,
+ 453,
+ 213
+ ],
+ "score": 1.0,
+ "content": "NYC, USA",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 387,
+ 213,
+ 469,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 213,
+ 469,
+ 224
+ ],
+ "score": 1.0,
+ "content": "ylecun@meta.com",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 363,
+ 179,
+ 493,
+ 224
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 283,
+ 252,
+ 328,
+ 265
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 281,
+ 251,
+ 331,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 281,
+ 251,
+ 331,
+ 267
+ ],
+ "score": 1.0,
+ "content": "Abstract",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 142,
+ 278,
+ 469,
+ 485
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 278,
+ 471,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 278,
+ 471,
+ 290
+ ],
+ "score": 1.0,
+ "content": "Regularization is a fundamental technique to improve a model’s generalization per-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 301
+ ],
+ "score": 1.0,
+ "content": "formances by limiting its complexity. Deep Neural Networks (DNNs), which tend",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 299,
+ 470,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 299,
+ 470,
+ 312
+ ],
+ "score": 1.0,
+ "content": "to overfit their training data, heavily rely on regularizers such as Data-Augmentation",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 310,
+ 470,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 310,
+ 470,
+ 323
+ ],
+ "score": 1.0,
+ "content": "(DA) or weight-decay with hyper-parameters found from structural risk minimiza-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 142,
+ 322,
+ 470,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 322,
+ 470,
+ 333
+ ],
+ "score": 1.0,
+ "content": "tion, i.e., cross-validation. In this study, we demonstrate that the optimal regular-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 332,
+ 470,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 332,
+ 470,
+ 345
+ ],
+ "score": 1.0,
+ "content": "ization’s hyper-parameters found from cross-validation over all classes leads to",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 142,
+ 343,
+ 470,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 343,
+ 470,
+ 356
+ ],
+ "score": 1.0,
+ "content": "disastrous model performances on a minority of classes. For example, a resnet50",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 354,
+ 469,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 354,
+ 398,
+ 366
+ ],
+ "score": 1.0,
+ "content": "trained on Imagenet sees its “barn spider” test accuracy falls from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 399,
+ 354,
+ 419,
+ 365
+ ],
+ "score": 0.88,
+ "content": "6 8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 354,
+ 429,
+ 366
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 354,
+ 449,
+ 365
+ ],
+ "score": 0.88,
+ "content": "4 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 449,
+ 354,
+ 469,
+ 366
+ ],
+ "score": 1.0,
+ "content": "only",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 142,
+ 365,
+ 469,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 365,
+ 469,
+ 377
+ ],
+ "score": 1.0,
+ "content": "by introducing random crop DA during training. Even more surprising, such unfair",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 389
+ ],
+ "score": 1.0,
+ "content": "impact of regularization also appears when introducing uninformative regularizers",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 387,
+ 470,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 387,
+ 470,
+ 399
+ ],
+ "score": 1.0,
+ "content": "such as weight decay or dropout. Those results demonstrate that our search for ever",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 398,
+ 470,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 398,
+ 470,
+ 411
+ ],
+ "score": 1.0,
+ "content": "increasing generalization performance —averaged over all classes and samples—",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 141,
+ 408,
+ 469,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 408,
+ 469,
+ 421
+ ],
+ "score": 1.0,
+ "content": "has left us with models and regularizers that silently sacrifice performances on",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 141,
+ 420,
+ 469,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 420,
+ 469,
+ 432
+ ],
+ "score": 1.0,
+ "content": "some classes. This scenario can become dangerous when deploying a model on",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 141,
+ 431,
+ 470,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 431,
+ 470,
+ 442
+ ],
+ "score": 1.0,
+ "content": "downstream tasks e.g. an Imagenet pre-trained resnet50 deployed on INaturalist",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 141,
+ 441,
+ 470,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 441,
+ 267,
+ 453
+ ],
+ "score": 1.0,
+ "content": "sees its performances fall from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 267,
+ 441,
+ 287,
+ 452
+ ],
+ "score": 0.89,
+ "content": "7 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 287,
+ 441,
+ 298,
+ 453
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 298,
+ 441,
+ 318,
+ 452
+ ],
+ "score": 0.89,
+ "content": "3 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 318,
+ 441,
+ 352,
+ 453
+ ],
+ "score": 1.0,
+ "content": "on class",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 441,
+ 379,
+ 452
+ ],
+ "score": 0.51,
+ "content": "\\# 8 8 8 9",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 379,
+ 441,
+ 470,
+ 453
+ ],
+ "score": 1.0,
+ "content": "when introducing ran-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 141,
+ 452,
+ 470,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 452,
+ 470,
+ 465
+ ],
+ "score": 1.0,
+ "content": "dom crop DA during the Imagenet pre-training phase. Those results demonstrate",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 141,
+ 464,
+ 469,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 464,
+ 469,
+ 475
+ ],
+ "score": 1.0,
+ "content": "that finding a correct measure of a model’s complexity without class-dependent",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 141,
+ 475,
+ 329,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 475,
+ 329,
+ 487
+ ],
+ "score": 1.0,
+ "content": "preference remains an open research question.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 141,
+ 278,
+ 471,
+ 487
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 507,
+ 190,
+ 520
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 192,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 192,
+ 522
+ ],
+ "score": 1.0,
+ "content": "1 Introduction",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 532,
+ 505,
+ 663
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "Machine learning and deep learning aim at learning systems to solve as accurately as possible a",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 543,
+ 507,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 507,
+ 556
+ ],
+ "score": 1.0,
+ "content": "given task at hand [LeCun et al., 1998, Bishop and Nasrabadi, 2006, Jordan and Mitchell, 2015].",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 555,
+ 506,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 555,
+ 506,
+ 567
+ ],
+ "score": 1.0,
+ "content": "This process often takes the form of (i) being given a finite dataset, a (differentiable) loss function,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 579
+ ],
+ "score": 1.0,
+ "content": "and a performance measure, (ii) splitting the dataset into train/valid/test sets to optimizing the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 506,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 506,
+ 589
+ ],
+ "score": 1.0,
+ "content": "system’s parameters e.g. from gradient updates of the loss on the train set while cross-validating",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 599
+ ],
+ "score": 1.0,
+ "content": "hyper-parameters using the valid set, and (iii) assessing the system’s performance on the test set. As",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "the training set is finite, and the optimal design of the system is unknown, it is common to employ",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "regularization during the optimization phase to reduce over-fitting [Tikhonov, 1943, Tihonov, 1963]",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "i.e. to decrease the system’s performance gap between train set and test set samples [Simard et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 630,
+ 506,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 630,
+ 506,
+ 643
+ ],
+ "score": 1.0,
+ "content": "1991, Chapelle et al., 2000, Bottou, 2012, Neyshabur et al., 2014]. Central to our study is the fact that",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 642,
+ 505,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 505,
+ 654
+ ],
+ "score": 1.0,
+ "content": "hyper-parameter selection is done via cross-validation by maximizing the valid set performance with",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 477,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 653,
+ 477,
+ 665
+ ],
+ "score": 1.0,
+ "content": "ad-hoc statistics e.g. the average accuracy over all samples for classes in classification tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 532,
+ 507,
+ 665
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 669,
+ 506,
+ 713
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 668,
+ 507,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 668,
+ 507,
+ 681
+ ],
+ "score": 1.0,
+ "content": "Cross-validation commonly involves many different types of regularization along with their “strengths”",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 680,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "[Goodfellow et al., 2016, He et al., 2021]. Most variants of regularization take one of two forms:",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 104,
+ 689,
+ 506,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 689,
+ 506,
+ 705
+ ],
+ "score": 1.0,
+ "content": "Data-Augmentation (DA) and weight-decay. DA is a data-driven and informed regularization strategy",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 702,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 702,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "that artificially increase the number of training samples [Shorten and Khoshgoftaar, 2019]. As",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": "opposed to most explicit regularizers e.g. Tikhonov regularization [Krogh and Hertz, 1991], also",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "denoted as weight decay, DA’s regularization is implicit as it is not a function of a model’s parameter,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 356,
+ 506,
+ 368
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 506,
+ 368
+ ],
+ "score": 1.0,
+ "content": "but a function of the training samples [Neyshabur et al., 2014, Hernández-García and König, 2018,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 379
+ ],
+ "score": 1.0,
+ "content": "LeJeune et al., 2019]; although some DA strategies can be turned into explicit regularizers Balestriero",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 505,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 505,
+ 390
+ ],
+ "score": 1.0,
+ "content": "et al. [2022]. Nevertheless, a key distinction between DA and weight decay is that DA tends to",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 388,
+ 506,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 388,
+ 506,
+ 401
+ ],
+ "score": 1.0,
+ "content": "require more domain knowledge to be successful than weight decay. Most —if not all— of current",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 412
+ ],
+ "score": 1.0,
+ "content": "state-of-the-art employ such regularizers [Huang et al., 2018, Chen et al., 2020b, Liu et al., 2021, Tan",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 411,
+ 232,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 411,
+ 232,
+ 423
+ ],
+ "score": 1.0,
+ "content": "and Le, 2021, Liu et al., 2022].",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 48.5,
+ "bbox_fs": [
+ 104,
+ 668,
+ 507,
+ 714
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 118,
+ 77,
+ 498,
+ 215
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 118,
+ 77,
+ 498,
+ 215
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 77,
+ 498,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 77,
+ 498,
+ 215
+ ],
+ "score": 0.92,
+ "type": "image",
+ "image_path": "0a582fb6f48f5b79c70845ea818ec095486cbfc5c65229ebf025a2f650a7a805.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 118,
+ 77,
+ 498,
+ 123.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 118,
+ 123.0,
+ 498,
+ 169.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 118,
+ 169.0,
+ 498,
+ 215.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 223,
+ 505,
+ 314
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 506,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 506,
+ 236
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Structural risk minimization minimizes the empirical risk of several models of varying complexity, and",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 235,
+ 506,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 235,
+ 506,
+ 245
+ ],
+ "score": 1.0,
+ "content": "selects the one offering the best compromise between under-fitting and over-fitting [Vapnik and Chervonenkis,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 506,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 506,
+ 256
+ ],
+ "score": 1.0,
+ "content": "1974]. In deep learning, one commonly control the model’s complexity by picking different DN architectures",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 265
+ ],
+ "score": 1.0,
+ "content": "and/or by applying different levels and flavors of regularization. The key observation of our study is that when the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 264,
+ 505,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 264,
+ 505,
+ 275
+ ],
+ "score": 1.0,
+ "content": "model complexity is calibrated by DA (see Figs. 2, 5 and 6), or weight-decay (see Fig. 3), the class-conditional",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 273,
+ 506,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 506,
+ 285
+ ],
+ "score": 1.0,
+ "content": "empirical risks do not align between classes i.e. cross-validation produces models that perform well on the",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 283,
+ 506,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 283,
+ 506,
+ 296
+ ],
+ "score": 1.0,
+ "content": "majority of classes but arbitrarily poorly on a few of them as depicted on the left-hand-side. In an ideal setting",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 305
+ ],
+ "score": 1.0,
+ "content": "where the control of the model’s complexity is well aligned with the task and model, one would observe the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 304,
+ 432,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 304,
+ 432,
+ 315
+ ],
+ "score": 1.0,
+ "content": "right-hand-side ideal scenario where the same model complexity is optimal for all classes.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 4.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 506,
+ 422
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": "opposed to most explicit regularizers e.g. Tikhonov regularization [Krogh and Hertz, 1991], also",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "denoted as weight decay, DA’s regularization is implicit as it is not a function of a model’s parameter,",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 356,
+ 506,
+ 368
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 506,
+ 368
+ ],
+ "score": 1.0,
+ "content": "but a function of the training samples [Neyshabur et al., 2014, Hernández-García and König, 2018,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 379
+ ],
+ "score": 1.0,
+ "content": "LeJeune et al., 2019]; although some DA strategies can be turned into explicit regularizers Balestriero",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 505,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 505,
+ 390
+ ],
+ "score": 1.0,
+ "content": "et al. [2022]. Nevertheless, a key distinction between DA and weight decay is that DA tends to",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 388,
+ 506,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 388,
+ 506,
+ 401
+ ],
+ "score": 1.0,
+ "content": "require more domain knowledge to be successful than weight decay. Most —if not all— of current",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 412
+ ],
+ "score": 1.0,
+ "content": "state-of-the-art employ such regularizers [Huang et al., 2018, Chen et al., 2020b, Liu et al., 2021, Tan",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 411,
+ 232,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 411,
+ 232,
+ 423
+ ],
+ "score": 1.0,
+ "content": "and Le, 2021, Liu et al., 2022].",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 426,
+ 505,
+ 558
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 427,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 427,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "In this paper, we will demonstrate that when cross-validation is employed to select the regularization",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 450
+ ],
+ "score": 1.0,
+ "content": "settings maximizing the validation performance, a significant bias is introduced into the trained model:",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 448,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "the regularized model exhibits strong per-class favoritism i.e. while the average test performance",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 506,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 506,
+ 473
+ ],
+ "score": 1.0,
+ "content": "is improved, it is at the cost of producing a model with significant performance drop on some of",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 471,
+ 506,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 471,
+ 506,
+ 483
+ ],
+ "score": 1.0,
+ "content": "the classes as illustrated in the schematic of Fig. 1. For readers familiar with statistical estimation",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 482,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": "results e.g. the bias-variance trade-off [Kohavi et al., 1996, Von Luxburg and Schölkopf, 2011] or",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "bayesian estimation e.g. Tikhonov regularization [Box and Tiao, 2011, Gruber, 2017], it should not be",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 516
+ ],
+ "score": 1.0,
+ "content": "surprising that regularization produces bias (more details and background provided in Appendix A).",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 513,
+ 506,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 513,
+ 506,
+ 527
+ ],
+ "score": 1.0,
+ "content": "In fact, it is beneficial to introduce bias through regularization if it results in a significant reduction of",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "the estimator variance —when one minimizes the average empirical risk. However, the potentially",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 535,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 535,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "dangerous effect of regularization that this study brings forward is that the bias introduced by",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 547,
+ 391,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 547,
+ 391,
+ 559
+ ],
+ "score": 1.0,
+ "content": "regularization is class-dependent, including on transfer learning tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 563,
+ 505,
+ 662
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "To thoroughly validate this observation, we propose a variety of controlled experiments in Section 2.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 587
+ ],
+ "score": 1.0,
+ "content": "First, we carefully quantify the impact of DA, weight decay and dropout on the per-class performance",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 584,
+ 506,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 506,
+ 598
+ ],
+ "score": 1.0,
+ "content": "of a model in Section 2.1, demonstrating that current deep learning finds itself in the scenario depicted",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 609
+ ],
+ "score": 1.0,
+ "content": "on the left of Fig. 1. Then, we consider the task of transfer learning in Section 2.2 where it is again",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 607,
+ 506,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 607,
+ 506,
+ 621
+ ],
+ "score": 1.0,
+ "content": "possible to identify again a per-class bias on the target dataset even though the regularization was",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "applied on a different (source) training set. This latter scenario is particularly relevant in current times",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 629,
+ 506,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 629,
+ 506,
+ 642
+ ],
+ "score": 1.0,
+ "content": "where it is common to deploy a large pre-trained model on a variety of tasks and raise an important",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "issue: selecting the —on average— best performing pre-trained model can lead to catastrophic",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 651,
+ 383,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 651,
+ 383,
+ 663
+ ],
+ "score": 1.0,
+ "content": "individual class performance even on for different downstream tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 667,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "Our next Section 3 will aim at exploring possible explanations and solutions. First, we will provide",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "a brief theoretical justification on why and when DA can be the cause of model bias (Section 3.1)",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "regardless of the task and data at hand. This will shed light to a first possible issue: the DA parameters",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "that make the transformed input preserve its label information vary depending on the class underlying",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 505,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 505,
+ 724
+ ],
+ "score": 1.0,
+ "content": "statistics. In short, DA silently introduces class-imbalance in the training set. We propose a dedicated",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 350,
+ 71,
+ 457,
+ 81
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 349,
+ 69,
+ 459,
+ 83
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 349,
+ 69,
+ 459,
+ 83
+ ],
+ "score": 1.0,
+ "content": "Ideal Complexity Measure",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 160,
+ 70,
+ 254,
+ 81
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 160,
+ 67,
+ 255,
+ 84
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 67,
+ 255,
+ 84
+ ],
+ "score": 1.0,
+ "content": "Current Deep Learning",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 741,
+ 310,
+ 753
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 741,
+ 310,
+ 753
+ ],
+ "score": 1.0,
+ "content": "2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 118,
+ 77,
+ 498,
+ 215
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 118,
+ 77,
+ 498,
+ 215
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 77,
+ 498,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 77,
+ 498,
+ 215
+ ],
+ "score": 0.92,
+ "type": "image",
+ "image_path": "0a582fb6f48f5b79c70845ea818ec095486cbfc5c65229ebf025a2f650a7a805.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 118,
+ 77,
+ 498,
+ 123.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 118,
+ 123.0,
+ 498,
+ 169.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 118,
+ 169.0,
+ 498,
+ 215.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 223,
+ 505,
+ 314
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 506,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 506,
+ 236
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Structural risk minimization minimizes the empirical risk of several models of varying complexity, and",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 235,
+ 506,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 235,
+ 506,
+ 245
+ ],
+ "score": 1.0,
+ "content": "selects the one offering the best compromise between under-fitting and over-fitting [Vapnik and Chervonenkis,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 506,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 506,
+ 256
+ ],
+ "score": 1.0,
+ "content": "1974]. In deep learning, one commonly control the model’s complexity by picking different DN architectures",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 265
+ ],
+ "score": 1.0,
+ "content": "and/or by applying different levels and flavors of regularization. The key observation of our study is that when the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 264,
+ 505,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 264,
+ 505,
+ 275
+ ],
+ "score": 1.0,
+ "content": "model complexity is calibrated by DA (see Figs. 2, 5 and 6), or weight-decay (see Fig. 3), the class-conditional",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 273,
+ 506,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 506,
+ 285
+ ],
+ "score": 1.0,
+ "content": "empirical risks do not align between classes i.e. cross-validation produces models that perform well on the",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 283,
+ 506,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 283,
+ 506,
+ 296
+ ],
+ "score": 1.0,
+ "content": "majority of classes but arbitrarily poorly on a few of them as depicted on the left-hand-side. In an ideal setting",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 305
+ ],
+ "score": 1.0,
+ "content": "where the control of the model’s complexity is well aligned with the task and model, one would observe the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 304,
+ 432,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 304,
+ 432,
+ 315
+ ],
+ "score": 1.0,
+ "content": "right-hand-side ideal scenario where the same model complexity is optimal for all classes.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 4.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 506,
+ 422
+ ],
+ "lines": [],
+ "index": 15.5,
+ "bbox_fs": [
+ 104,
+ 334,
+ 506,
+ 423
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 426,
+ 505,
+ 558
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 427,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 427,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "In this paper, we will demonstrate that when cross-validation is employed to select the regularization",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 450
+ ],
+ "score": 1.0,
+ "content": "settings maximizing the validation performance, a significant bias is introduced into the trained model:",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 448,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "the regularized model exhibits strong per-class favoritism i.e. while the average test performance",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 506,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 506,
+ 473
+ ],
+ "score": 1.0,
+ "content": "is improved, it is at the cost of producing a model with significant performance drop on some of",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 471,
+ 506,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 471,
+ 506,
+ 483
+ ],
+ "score": 1.0,
+ "content": "the classes as illustrated in the schematic of Fig. 1. For readers familiar with statistical estimation",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 482,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": "results e.g. the bias-variance trade-off [Kohavi et al., 1996, Von Luxburg and Schölkopf, 2011] or",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "bayesian estimation e.g. Tikhonov regularization [Box and Tiao, 2011, Gruber, 2017], it should not be",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 516
+ ],
+ "score": 1.0,
+ "content": "surprising that regularization produces bias (more details and background provided in Appendix A).",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 513,
+ 506,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 513,
+ 506,
+ 527
+ ],
+ "score": 1.0,
+ "content": "In fact, it is beneficial to introduce bias through regularization if it results in a significant reduction of",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 525,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "the estimator variance —when one minimizes the average empirical risk. However, the potentially",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 535,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 535,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "dangerous effect of regularization that this study brings forward is that the bias introduced by",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 547,
+ 391,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 547,
+ 391,
+ 559
+ ],
+ "score": 1.0,
+ "content": "regularization is class-dependent, including on transfer learning tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 105,
+ 427,
+ 506,
+ 559
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 563,
+ 505,
+ 662
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "To thoroughly validate this observation, we propose a variety of controlled experiments in Section 2.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 587
+ ],
+ "score": 1.0,
+ "content": "First, we carefully quantify the impact of DA, weight decay and dropout on the per-class performance",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 584,
+ 506,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 506,
+ 598
+ ],
+ "score": 1.0,
+ "content": "of a model in Section 2.1, demonstrating that current deep learning finds itself in the scenario depicted",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 609
+ ],
+ "score": 1.0,
+ "content": "on the left of Fig. 1. Then, we consider the task of transfer learning in Section 2.2 where it is again",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 607,
+ 506,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 607,
+ 506,
+ 621
+ ],
+ "score": 1.0,
+ "content": "possible to identify again a per-class bias on the target dataset even though the regularization was",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "applied on a different (source) training set. This latter scenario is particularly relevant in current times",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 629,
+ 506,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 629,
+ 506,
+ 642
+ ],
+ "score": 1.0,
+ "content": "where it is common to deploy a large pre-trained model on a variety of tasks and raise an important",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "issue: selecting the —on average— best performing pre-trained model can lead to catastrophic",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 651,
+ 383,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 651,
+ 383,
+ 663
+ ],
+ "score": 1.0,
+ "content": "individual class performance even on for different downstream tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 563,
+ 506,
+ 663
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 667,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "Our next Section 3 will aim at exploring possible explanations and solutions. First, we will provide",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "a brief theoretical justification on why and when DA can be the cause of model bias (Section 3.1)",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "regardless of the task and data at hand. This will shed light to a first possible issue: the DA parameters",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "that make the transformed input preserve its label information vary depending on the class underlying",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 505,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 505,
+ 724
+ ],
+ "score": 1.0,
+ "content": "statistics. In short, DA silently introduces class-imbalance in the training set. We propose a dedicated",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 271
+ ],
+ "score": 1.0,
+ "content": "analysis of the label-preserving property of DA on different classes and models in Section 3.2. We",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 282
+ ],
+ "score": 1.0,
+ "content": "then take on the task of searching for a possible solution by first reviewing known theoretical studies",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 294
+ ],
+ "score": 1.0,
+ "content": "quantifying the interplay between regularization and bias in Section 3.3. Lastly, we propose some",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "solutions of our own in Section 3.4 built from the gained insights of Section 3.2 using label-distillation",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 303,
+ 505,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 505,
+ 315
+ ],
+ "score": 1.0,
+ "content": "and adaptive DA. All the codebase used to train the various models and to generate the figures is in the",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 314,
+ 180,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 314,
+ 180,
+ 326
+ ],
+ "score": 1.0,
+ "content": "supplementary files.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 43,
+ "bbox_fs": [
+ 105,
+ 667,
+ 505,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 70,
+ 502,
+ 190
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 70,
+ 502,
+ 190
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 502,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 502,
+ 190
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "cf86150e77df9c74677df4284187ee43dd9c86df8e08cd627401fed3e0886a04.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 502,
+ 110.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 110.0,
+ 502,
+ 150.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 150.0,
+ 502,
+ 190.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 191,
+ 505,
+ 241
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 189,
+ 506,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 189,
+ 304,
+ 203
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Varying the random crop DA lower bound",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 305,
+ 192,
+ 310,
+ 199
+ ],
+ "score": 0.35,
+ "content": "\\mathbf { \\hat { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 311,
+ 189,
+ 352,
+ 203
+ ],
+ "score": 1.0,
+ "content": "-axis) from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 191,
+ 374,
+ 200
+ ],
+ "score": 0.87,
+ "content": "100 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 189,
+ 384,
+ 203
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 385,
+ 191,
+ 398,
+ 200
+ ],
+ "score": 0.85,
+ "content": "8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 189,
+ 506,
+ 203
+ ],
+ "score": 1.0,
+ "content": "provides greater average test",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 212
+ ],
+ "score": 1.0,
+ "content": "accuracy (blue) but makes the per-class performance fall for some of the classes. Images of each class are",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 210,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 210,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "provided in Fig. 12, in the appendix. See Fig. 8 for the convnext and ViT experiments. Results obtained by",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 232
+ ],
+ "score": 1.0,
+ "content": "averaging over 20 runs, official PyTorch resnet50 implementation trained on Imagenet with horizontal flip and",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 230,
+ 249,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 230,
+ 249,
+ 241
+ ],
+ "score": 1.0,
+ "content": "varying random crop lower bound DA.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 259,
+ 505,
+ 325
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 271
+ ],
+ "score": 1.0,
+ "content": "analysis of the label-preserving property of DA on different classes and models in Section 3.2. We",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 282
+ ],
+ "score": 1.0,
+ "content": "then take on the task of searching for a possible solution by first reviewing known theoretical studies",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 294
+ ],
+ "score": 1.0,
+ "content": "quantifying the interplay between regularization and bias in Section 3.3. Lastly, we propose some",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "solutions of our own in Section 3.4 built from the gained insights of Section 3.2 using label-distillation",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 303,
+ 505,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 505,
+ 315
+ ],
+ "score": 1.0,
+ "content": "and adaptive DA. All the codebase used to train the various models and to generate the figures is in the",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 314,
+ 180,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 314,
+ 180,
+ 326
+ ],
+ "score": 1.0,
+ "content": "supplementary files.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 336,
+ 501,
+ 364
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 501,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 501,
+ 353
+ ],
+ "score": 1.0,
+ "content": "2 Maximizing the Average Model Performance by Cross-Validation Silently",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 123,
+ 351,
+ 446,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 351,
+ 446,
+ 366
+ ],
+ "score": 1.0,
+ "content": "Produce Poor Final Performances on a Minority of the Classes",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 370,
+ 505,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 370,
+ 506,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 370,
+ 506,
+ 384
+ ],
+ "score": 1.0,
+ "content": "We now turn to the empirical validation of Fig. 1 i.e. quantifying the amount of class-dependent",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 381,
+ 506,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 506,
+ 394
+ ],
+ "score": 1.0,
+ "content": "bias caused by DA, weight decay and dropout in various realistic scenarios (Section 2.1). We then",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "demonstrate how the bias introduced by regularization transfers to downstream tasks e.g. when",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 403,
+ 505,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 403,
+ 505,
+ 415
+ ],
+ "score": 1.0,
+ "content": "deploying an Imagenet (source) trained model on the INaturalist (target) dataset in Section 2.2; that",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 414,
+ 504,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 414,
+ 504,
+ 426
+ ],
+ "score": 1.0,
+ "content": "scenario is key as it demonstrates the potential harm of selecting the best performing model on the",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 425,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 425,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "source dataset which could turn out to also be the most biased model against the target dataset class",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 435,
+ 152,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 152,
+ 448
+ ],
+ "score": 1.0,
+ "content": "of interest.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 459,
+ 474,
+ 482
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 459,
+ 474,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 124,
+ 471
+ ],
+ "score": 1.0,
+ "content": "2.1",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 127,
+ 459,
+ 474,
+ 471
+ ],
+ "score": 1.0,
+ "content": "Precisely Measuring the Per-Class Effect of Data-Augmentation and Uninformed",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 128,
+ 471,
+ 321,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 471,
+ 321,
+ 483
+ ],
+ "score": 1.0,
+ "content": "Regularization with Controlled Experiments",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 575
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 506,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 506,
+ 500
+ ],
+ "score": 1.0,
+ "content": "This section aims at quantifying precisely the amount of downward or upward per-class performance",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 498,
+ 506,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 498,
+ 506,
+ 511
+ ],
+ "score": 1.0,
+ "content": "shift that came as a result from using DA or uninformed regularization e.g. weight-decay or dropout.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 521
+ ],
+ "score": 1.0,
+ "content": "In fact, it is crucial to remember that regularization, or any other form of structural risk minimization,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 520,
+ 506,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 506,
+ 532
+ ],
+ "score": 1.0,
+ "content": "improves generalization performances by increasing the bias of the estimator so that the estimator’s",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 531,
+ 506,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 531,
+ 506,
+ 543
+ ],
+ "score": 1.0,
+ "content": "variance is decreased by a greater amount. However, nothing guarantees the fairness of this bias i.e.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "for it to be equally distributed amongst the dataset classes. We thus propose a sensitivity analysis",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "by training a large collection of models with varying regularization policies to precisely assess their",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 564,
+ 278,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 564,
+ 278,
+ 576
+ ],
+ "score": 1.0,
+ "content": "impact on the class-dependent model bias.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 579,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 579,
+ 504,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 579,
+ 504,
+ 591
+ ],
+ "score": 1.0,
+ "content": "Data-Augmentation. DA samples have been known to sometimes disregard the semantic information",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 589,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 589,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "of the original samples [Krizhevsky et al., 2012]. Nevertheless, DA remains applied universally",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 615
+ ],
+ "score": 1.0,
+ "content": "across tasks and datasets [Shorten and Khoshgoftaar, 2019] as it provides significant performance",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 612,
+ 506,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 612,
+ 506,
+ 625
+ ],
+ "score": 1.0,
+ "content": "improvements, even in semi-supervised and unsupervised settings Guo et al. [2018], Xie et al. [2020],",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 623,
+ 506,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 623,
+ 506,
+ 637
+ ],
+ "score": 1.0,
+ "content": "Misra and Maaten [2020]. To measure the impact of DA onto per-class performances, we propose in",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 648
+ ],
+ "score": 1.0,
+ "content": "Fig. 2 a sensitivity analysis by training the same architecture on Imagenet with varying DA policies.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 645,
+ 505,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 645,
+ 505,
+ 659
+ ],
+ "score": 1.0,
+ "content": "In particular, we consider a given DA (random crop in this case) and we vary the support of the",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 655,
+ 506,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 147,
+ 670
+ ],
+ "score": 1.0,
+ "content": "parameter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 148,
+ 658,
+ 156,
+ 666
+ ],
+ "score": 0.76,
+ "content": "\\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 655,
+ 506,
+ 670
+ ],
+ "score": 1.0,
+ "content": "which represents how much of the original image is kept in the crop (examples at the top",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 667,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 372,
+ 680
+ ],
+ "score": 1.0,
+ "content": "of Fig. 5). We train multiple Deep Neural Networks (DNN)s using",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 372,
+ 667,
+ 422,
+ 679
+ ],
+ "score": 0.93,
+ "content": "\\alpha \\in [ 1 0 0 , \\tau ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 667,
+ 443,
+ 680
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 669,
+ 450,
+ 677
+ ],
+ "score": 0.74,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 667,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "varying from",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 104,
+ 676,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 676,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "100 to 8 and for each case, we report our metrics averaged over 20 trained models. We observe a",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "clear relation between increase in the strength of the DA, increase in the average test accuracy overall",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 507,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 507,
+ 714
+ ],
+ "score": 1.0,
+ "content": "classes, and decrease in some per-class test accuracies. For example, on a resnet50 Imagenet setting,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 471,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 324,
+ 724
+ ],
+ "score": 1.0,
+ "content": "the accuracy on the “academic gown” class goes from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 711,
+ 343,
+ 721
+ ],
+ "score": 0.86,
+ "content": "62 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 344,
+ 711,
+ 354,
+ 724
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 355,
+ 711,
+ 374,
+ 721
+ ],
+ "score": 0.87,
+ "content": "40 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 711,
+ 419,
+ 724
+ ],
+ "score": 1.0,
+ "content": "steadily as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 713,
+ 427,
+ 721
+ ],
+ "score": 0.76,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 427,
+ 711,
+ 471,
+ 724
+ ],
+ "score": 1.0,
+ "content": "decreases.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 70,
+ 502,
+ 190
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 70,
+ 502,
+ 190
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 502,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 502,
+ 190
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "cf86150e77df9c74677df4284187ee43dd9c86df8e08cd627401fed3e0886a04.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 502,
+ 110.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 110.0,
+ 502,
+ 150.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 150.0,
+ 502,
+ 190.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 191,
+ 505,
+ 241
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 189,
+ 506,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 189,
+ 304,
+ 203
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Varying the random crop DA lower bound",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 305,
+ 192,
+ 310,
+ 199
+ ],
+ "score": 0.35,
+ "content": "\\mathbf { \\hat { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 311,
+ 189,
+ 352,
+ 203
+ ],
+ "score": 1.0,
+ "content": "-axis) from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 191,
+ 374,
+ 200
+ ],
+ "score": 0.87,
+ "content": "100 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 189,
+ 384,
+ 203
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 385,
+ 191,
+ 398,
+ 200
+ ],
+ "score": 0.85,
+ "content": "8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 189,
+ 506,
+ 203
+ ],
+ "score": 1.0,
+ "content": "provides greater average test",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 212
+ ],
+ "score": 1.0,
+ "content": "accuracy (blue) but makes the per-class performance fall for some of the classes. Images of each class are",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 210,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 210,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "provided in Fig. 12, in the appendix. See Fig. 8 for the convnext and ViT experiments. Results obtained by",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 232
+ ],
+ "score": 1.0,
+ "content": "averaging over 20 runs, official PyTorch resnet50 implementation trained on Imagenet with horizontal flip and",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 230,
+ 249,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 230,
+ 249,
+ 241
+ ],
+ "score": 1.0,
+ "content": "varying random crop lower bound DA.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 259,
+ 505,
+ 325
+ ],
+ "lines": [],
+ "index": 10.5,
+ "bbox_fs": [
+ 105,
+ 259,
+ 506,
+ 326
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 336,
+ 501,
+ 364
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 501,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 501,
+ 353
+ ],
+ "score": 1.0,
+ "content": "2 Maximizing the Average Model Performance by Cross-Validation Silently",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 123,
+ 351,
+ 446,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 351,
+ 446,
+ 366
+ ],
+ "score": 1.0,
+ "content": "Produce Poor Final Performances on a Minority of the Classes",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 370,
+ 505,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 370,
+ 506,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 370,
+ 506,
+ 384
+ ],
+ "score": 1.0,
+ "content": "We now turn to the empirical validation of Fig. 1 i.e. quantifying the amount of class-dependent",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 381,
+ 506,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 506,
+ 394
+ ],
+ "score": 1.0,
+ "content": "bias caused by DA, weight decay and dropout in various realistic scenarios (Section 2.1). We then",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "demonstrate how the bias introduced by regularization transfers to downstream tasks e.g. when",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 403,
+ 505,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 403,
+ 505,
+ 415
+ ],
+ "score": 1.0,
+ "content": "deploying an Imagenet (source) trained model on the INaturalist (target) dataset in Section 2.2; that",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 414,
+ 504,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 414,
+ 504,
+ 426
+ ],
+ "score": 1.0,
+ "content": "scenario is key as it demonstrates the potential harm of selecting the best performing model on the",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 425,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 425,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "source dataset which could turn out to also be the most biased model against the target dataset class",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 435,
+ 152,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 152,
+ 448
+ ],
+ "score": 1.0,
+ "content": "of interest.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 105,
+ 370,
+ 506,
+ 448
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 459,
+ 474,
+ 482
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 459,
+ 474,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 124,
+ 471
+ ],
+ "score": 1.0,
+ "content": "2.1",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 127,
+ 459,
+ 474,
+ 471
+ ],
+ "score": 1.0,
+ "content": "Precisely Measuring the Per-Class Effect of Data-Augmentation and Uninformed",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 128,
+ 471,
+ 321,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 471,
+ 321,
+ 483
+ ],
+ "score": 1.0,
+ "content": "Regularization with Controlled Experiments",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 575
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 506,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 506,
+ 500
+ ],
+ "score": 1.0,
+ "content": "This section aims at quantifying precisely the amount of downward or upward per-class performance",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 498,
+ 506,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 498,
+ 506,
+ 511
+ ],
+ "score": 1.0,
+ "content": "shift that came as a result from using DA or uninformed regularization e.g. weight-decay or dropout.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 521
+ ],
+ "score": 1.0,
+ "content": "In fact, it is crucial to remember that regularization, or any other form of structural risk minimization,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 520,
+ 506,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 506,
+ 532
+ ],
+ "score": 1.0,
+ "content": "improves generalization performances by increasing the bias of the estimator so that the estimator’s",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 531,
+ 506,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 531,
+ 506,
+ 543
+ ],
+ "score": 1.0,
+ "content": "variance is decreased by a greater amount. However, nothing guarantees the fairness of this bias i.e.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 542,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "for it to be equally distributed amongst the dataset classes. We thus propose a sensitivity analysis",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "by training a large collection of models with varying regularization policies to precisely assess their",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 564,
+ 278,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 564,
+ 278,
+ 576
+ ],
+ "score": 1.0,
+ "content": "impact on the class-dependent model bias.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 105,
+ 486,
+ 506,
+ 576
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 579,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 579,
+ 504,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 579,
+ 504,
+ 591
+ ],
+ "score": 1.0,
+ "content": "Data-Augmentation. DA samples have been known to sometimes disregard the semantic information",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 589,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 589,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "of the original samples [Krizhevsky et al., 2012]. Nevertheless, DA remains applied universally",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 615
+ ],
+ "score": 1.0,
+ "content": "across tasks and datasets [Shorten and Khoshgoftaar, 2019] as it provides significant performance",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 612,
+ 506,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 612,
+ 506,
+ 625
+ ],
+ "score": 1.0,
+ "content": "improvements, even in semi-supervised and unsupervised settings Guo et al. [2018], Xie et al. [2020],",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 623,
+ 506,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 623,
+ 506,
+ 637
+ ],
+ "score": 1.0,
+ "content": "Misra and Maaten [2020]. To measure the impact of DA onto per-class performances, we propose in",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 648
+ ],
+ "score": 1.0,
+ "content": "Fig. 2 a sensitivity analysis by training the same architecture on Imagenet with varying DA policies.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 645,
+ 505,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 645,
+ 505,
+ 659
+ ],
+ "score": 1.0,
+ "content": "In particular, we consider a given DA (random crop in this case) and we vary the support of the",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 655,
+ 506,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 147,
+ 670
+ ],
+ "score": 1.0,
+ "content": "parameter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 148,
+ 658,
+ 156,
+ 666
+ ],
+ "score": 0.76,
+ "content": "\\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 655,
+ 506,
+ 670
+ ],
+ "score": 1.0,
+ "content": "which represents how much of the original image is kept in the crop (examples at the top",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 667,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 372,
+ 680
+ ],
+ "score": 1.0,
+ "content": "of Fig. 5). We train multiple Deep Neural Networks (DNN)s using",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 372,
+ 667,
+ 422,
+ 679
+ ],
+ "score": 0.93,
+ "content": "\\alpha \\in [ 1 0 0 , \\tau ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 667,
+ 443,
+ 680
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 669,
+ 450,
+ 677
+ ],
+ "score": 0.74,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 667,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "varying from",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 104,
+ 676,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 676,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "100 to 8 and for each case, we report our metrics averaged over 20 trained models. We observe a",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "clear relation between increase in the strength of the DA, increase in the average test accuracy overall",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 507,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 507,
+ 714
+ ],
+ "score": 1.0,
+ "content": "classes, and decrease in some per-class test accuracies. For example, on a resnet50 Imagenet setting,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 471,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 324,
+ 724
+ ],
+ "score": 1.0,
+ "content": "the accuracy on the “academic gown” class goes from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 711,
+ 343,
+ 721
+ ],
+ "score": 0.86,
+ "content": "62 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 344,
+ 711,
+ 354,
+ 724
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 355,
+ 711,
+ 374,
+ 721
+ ],
+ "score": 0.87,
+ "content": "40 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 711,
+ 419,
+ 724
+ ],
+ "score": 1.0,
+ "content": "steadily as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 713,
+ 427,
+ 721
+ ],
+ "score": 0.76,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 427,
+ 711,
+ 471,
+ 724
+ ],
+ "score": 1.0,
+ "content": "decreases.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 39,
+ "bbox_fs": [
+ 104,
+ 579,
+ 507,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 70,
+ 501,
+ 190
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 70,
+ 501,
+ 190
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 501,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 501,
+ 190
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "a9f2b68ae848c586ddff72ac17bb6397ed8cefb20460c7f6deba497956f957c6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 501,
+ 110.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 110.0,
+ 501,
+ 150.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 150.0,
+ 501,
+ 190.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 191,
+ 505,
+ 240
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 189,
+ 505,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 189,
+ 505,
+ 202
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Varying the amount of weight decay, an uninformed regularizer employed throughout, surprisingly",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 200,
+ 506,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 200,
+ 506,
+ 212
+ ],
+ "score": 1.0,
+ "content": "exhibits similar class-dependent bias as the DA scenario of Fig. 2. Images for each class are provided in Fig. 13,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "in the appendix. Results obtained by averaging over 20 runs, official PyTorch resnet50 implementation trained",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 220,
+ 505,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 220,
+ 505,
+ 231
+ ],
+ "score": 1.0,
+ "content": "on Imagenet with varying weight decay, see Fig. 16 for DenseNet121 results with the same trend and Fig. 8 for",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 230,
+ 235,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 230,
+ 235,
+ 243
+ ],
+ "score": 1.0,
+ "content": "the convnext and ViT experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 254,
+ 505,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 506,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 506,
+ 268
+ ],
+ "score": 1.0,
+ "content": "Uninformed Regularization. As per the arguments given in Sections 3.1 and 3.2, it would be",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 266,
+ 506,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 266,
+ 506,
+ 278
+ ],
+ "score": 1.0,
+ "content": "natural to assume that what makes DA responsible for creating class-dependent bias in DNNs is",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": "our misfortune in defining correct augmentation policies. Hence, uninformed weight decay or",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 505,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 505,
+ 300
+ ],
+ "score": 1.0,
+ "content": "dropout should behave differently and more fairly. We demonstrate here that such regularizers are",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 312
+ ],
+ "score": 1.0,
+ "content": "also unfair between classes. We thus propose to train multiple models with varying weight-decay",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 308,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "and dropout parameters. In our setting, weight-decay is applied to all the DNN parameters except",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 333
+ ],
+ "score": 1.0,
+ "content": "for the ones of batch-normalization layers, as commonly done [Hastie et al., 2009, Leclerc et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 329,
+ 506,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 329,
+ 506,
+ 345
+ ],
+ "score": 1.0,
+ "content": "2022]. We report in Fig. 3 the per-class performance of a resnet50 trained on Imagenet with varying",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 342,
+ 505,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 205,
+ 354
+ ],
+ "score": 1.0,
+ "content": "weight decay coefficient",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 343,
+ 213,
+ 353
+ ],
+ "score": 0.82,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 342,
+ 505,
+ 354
+ ],
+ "score": 1.0,
+ "content": "(as was done for DA in Fig. 2) and we observe that different classes have",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 352,
+ 505,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 352,
+ 305,
+ 366
+ ],
+ "score": 1.0,
+ "content": "different test accuracy sensitivities to variations in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 306,
+ 354,
+ 313,
+ 364
+ ],
+ "score": 0.79,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 352,
+ 505,
+ 366
+ ],
+ "score": 1.0,
+ "content": ". Some will see their generalization performance",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 376
+ ],
+ "score": 1.0,
+ "content": "increase, while others will have decreasing generalization performances. We further confirm such",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 375,
+ 506,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 375,
+ 506,
+ 387
+ ],
+ "score": 1.0,
+ "content": "findings in Figs. 9 and 10 for dropout where the same per-class trend is observed. In short, even for",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "uninformative regularizers such as weight decay or dropout, a per-class bias is introduced, reducing",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 410
+ ],
+ "score": 1.0,
+ "content": "performances for some of the classes. Although weight-decay is one of the most popular regularizer",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "that is uninformed on the data and task at hand, recent studies have demonstrated that techniques",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 430
+ ],
+ "score": 1.0,
+ "content": "such as model pruning —which can be seen as a post-training model complexity reduction i.e.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 507,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 507,
+ 442
+ ],
+ "score": 1.0,
+ "content": "regularization— also produce increased bias towards under-represented features [Hooker et al., 2019,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "2020]. More recently, Balestriero et al. [2022] obtained the close-form explicit regularizer of DA",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 505,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 505,
+ 464
+ ],
+ "score": 1.0,
+ "content": "from which it is possible to quantify the sample-dependent aspect of DA’s regularization. From our",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "findings, it seems that classes sharing the same type of features are thus impacted different by DAs.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 478,
+ 505,
+ 621
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 506,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 506,
+ 491
+ ],
+ "score": 1.0,
+ "content": "Formal Statistical Test. To further convey our claim, we now propose a formal statistical test",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 502
+ ],
+ "score": 1.0,
+ "content": "[Neyman and Pearson, 1933, Fisher, 1955] on the hypothesis that the per-class accuracy is significantly",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 513
+ ],
+ "score": 1.0,
+ "content": "higher when DA is applied for each class (details provided in Appendix A.2). We obtain that there is",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 510,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 510,
+ 291,
+ 523
+ ],
+ "score": 1.0,
+ "content": "enough evidence to reject the hypothesis with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 292,
+ 511,
+ 312,
+ 522
+ ],
+ "score": 0.89,
+ "content": "9 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 510,
+ 373,
+ 523
+ ],
+ "score": 1.0,
+ "content": "confidence for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 373,
+ 511,
+ 396,
+ 522
+ ],
+ "score": 0.89,
+ "content": "4 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 396,
+ 510,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "of all the classes, and with",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 521,
+ 506,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 522,
+ 126,
+ 533
+ ],
+ "score": 0.87,
+ "content": "9 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 127,
+ 521,
+ 185,
+ 535
+ ],
+ "score": 1.0,
+ "content": "confidence for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 522,
+ 208,
+ 532
+ ],
+ "score": 0.89,
+ "content": "2 . { \\dot { 6 } } \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 521,
+ 506,
+ 535
+ ],
+ "score": 1.0,
+ "content": "of all the classes. Hence there is sufficient evidence to say that the per-class",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 341,
+ 546
+ ],
+ "score": 1.0,
+ "content": "test accuracies is not increased when introducing DA for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 341,
+ 533,
+ 363,
+ 543
+ ],
+ "score": 0.88,
+ "content": "4 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 533,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "of the 1000 Imagenet classes. We",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 543,
+ 506,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 506,
+ 557
+ ],
+ "score": 1.0,
+ "content": "provide in Table 1 the same statistical test but applied on a variety of settings including different",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 554,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "architectures (resnet50, densenet121, ViT-small and ConvNext-Tiny) and across the random crop DA",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "and the weight decay controlled experiments. We should highlight however that this is not necessarily",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "a meaningful measure since for example one regularization might not have any negative impact on the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "classes, but can provide a beneficial gain that is drastically different between classes. Hence, although",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 613
+ ],
+ "score": 1.0,
+ "content": "the per-class performance does not drop by introduce the regularizer, the inter-class performance gap",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 344,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 344,
+ 622
+ ],
+ "score": 1.0,
+ "content": "can be increased by it, which is an equally harmful impact.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 625,
+ 502,
+ 658
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 625,
+ 504,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 625,
+ 504,
+ 637
+ ],
+ "score": 1.0,
+ "content": "The next Section 2.2 proposes to study the scenario of introducing a pre-trained model, on a different",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 636,
+ 504,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 636,
+ 504,
+ 648
+ ],
+ "score": 1.0,
+ "content": "downstream task to show that the class-dependent effect of regularization remains present and unfair",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 647,
+ 257,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 647,
+ 257,
+ 660
+ ],
+ "score": 1.0,
+ "content": "towards specific downstream classes.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 671,
+ 407,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 408,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 408,
+ 684
+ ],
+ "score": 1.0,
+ "content": "2.2 The Class-Dependent Bias Transfers to Other Downstream Tasks",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 44
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 689,
+ 504,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "The last experiment we propose is to quantify the amount of class-dependent bias that transfers to",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "other downstream tasks, a common situation in transfer learning and in system deployment to the",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": "real world [Pan and Yang, 2009]. We thus want to measure how regularization applied during the",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 741,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 741,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 11,
+ "width": 9
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 70,
+ 501,
+ 190
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 70,
+ 501,
+ 190
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 501,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 501,
+ 190
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "a9f2b68ae848c586ddff72ac17bb6397ed8cefb20460c7f6deba497956f957c6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 70,
+ 501,
+ 110.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 110.0,
+ 501,
+ 150.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 150.0,
+ 501,
+ 190.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 191,
+ 505,
+ 240
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 189,
+ 505,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 189,
+ 505,
+ 202
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Varying the amount of weight decay, an uninformed regularizer employed throughout, surprisingly",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 200,
+ 506,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 200,
+ 506,
+ 212
+ ],
+ "score": 1.0,
+ "content": "exhibits similar class-dependent bias as the DA scenario of Fig. 2. Images for each class are provided in Fig. 13,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "in the appendix. Results obtained by averaging over 20 runs, official PyTorch resnet50 implementation trained",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 220,
+ 505,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 220,
+ 505,
+ 231
+ ],
+ "score": 1.0,
+ "content": "on Imagenet with varying weight decay, see Fig. 16 for DenseNet121 results with the same trend and Fig. 8 for",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 230,
+ 235,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 230,
+ 235,
+ 243
+ ],
+ "score": 1.0,
+ "content": "the convnext and ViT experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 254,
+ 505,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 506,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 506,
+ 268
+ ],
+ "score": 1.0,
+ "content": "Uninformed Regularization. As per the arguments given in Sections 3.1 and 3.2, it would be",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 266,
+ 506,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 266,
+ 506,
+ 278
+ ],
+ "score": 1.0,
+ "content": "natural to assume that what makes DA responsible for creating class-dependent bias in DNNs is",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": "our misfortune in defining correct augmentation policies. Hence, uninformed weight decay or",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 505,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 505,
+ 300
+ ],
+ "score": 1.0,
+ "content": "dropout should behave differently and more fairly. We demonstrate here that such regularizers are",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 312
+ ],
+ "score": 1.0,
+ "content": "also unfair between classes. We thus propose to train multiple models with varying weight-decay",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 308,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "and dropout parameters. In our setting, weight-decay is applied to all the DNN parameters except",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 333
+ ],
+ "score": 1.0,
+ "content": "for the ones of batch-normalization layers, as commonly done [Hastie et al., 2009, Leclerc et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 329,
+ 506,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 329,
+ 506,
+ 345
+ ],
+ "score": 1.0,
+ "content": "2022]. We report in Fig. 3 the per-class performance of a resnet50 trained on Imagenet with varying",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 342,
+ 505,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 205,
+ 354
+ ],
+ "score": 1.0,
+ "content": "weight decay coefficient",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 343,
+ 213,
+ 353
+ ],
+ "score": 0.82,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 342,
+ 505,
+ 354
+ ],
+ "score": 1.0,
+ "content": "(as was done for DA in Fig. 2) and we observe that different classes have",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 352,
+ 505,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 352,
+ 305,
+ 366
+ ],
+ "score": 1.0,
+ "content": "different test accuracy sensitivities to variations in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 306,
+ 354,
+ 313,
+ 364
+ ],
+ "score": 0.79,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 352,
+ 505,
+ 366
+ ],
+ "score": 1.0,
+ "content": ". Some will see their generalization performance",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 376
+ ],
+ "score": 1.0,
+ "content": "increase, while others will have decreasing generalization performances. We further confirm such",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 375,
+ 506,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 375,
+ 506,
+ 387
+ ],
+ "score": 1.0,
+ "content": "findings in Figs. 9 and 10 for dropout where the same per-class trend is observed. In short, even for",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "uninformative regularizers such as weight decay or dropout, a per-class bias is introduced, reducing",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 410
+ ],
+ "score": 1.0,
+ "content": "performances for some of the classes. Although weight-decay is one of the most popular regularizer",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "that is uninformed on the data and task at hand, recent studies have demonstrated that techniques",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 430
+ ],
+ "score": 1.0,
+ "content": "such as model pruning —which can be seen as a post-training model complexity reduction i.e.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 507,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 507,
+ 442
+ ],
+ "score": 1.0,
+ "content": "regularization— also produce increased bias towards under-represented features [Hooker et al., 2019,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "2020]. More recently, Balestriero et al. [2022] obtained the close-form explicit regularizer of DA",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 505,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 505,
+ 464
+ ],
+ "score": 1.0,
+ "content": "from which it is possible to quantify the sample-dependent aspect of DA’s regularization. From our",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "findings, it seems that classes sharing the same type of features are thus impacted different by DAs.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 105,
+ 254,
+ 507,
+ 475
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 478,
+ 505,
+ 621
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 506,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 506,
+ 491
+ ],
+ "score": 1.0,
+ "content": "Formal Statistical Test. To further convey our claim, we now propose a formal statistical test",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 502
+ ],
+ "score": 1.0,
+ "content": "[Neyman and Pearson, 1933, Fisher, 1955] on the hypothesis that the per-class accuracy is significantly",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 513
+ ],
+ "score": 1.0,
+ "content": "higher when DA is applied for each class (details provided in Appendix A.2). We obtain that there is",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 510,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 510,
+ 291,
+ 523
+ ],
+ "score": 1.0,
+ "content": "enough evidence to reject the hypothesis with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 292,
+ 511,
+ 312,
+ 522
+ ],
+ "score": 0.89,
+ "content": "9 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 510,
+ 373,
+ 523
+ ],
+ "score": 1.0,
+ "content": "confidence for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 373,
+ 511,
+ 396,
+ 522
+ ],
+ "score": 0.89,
+ "content": "4 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 396,
+ 510,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "of all the classes, and with",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 521,
+ 506,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 522,
+ 126,
+ 533
+ ],
+ "score": 0.87,
+ "content": "9 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 127,
+ 521,
+ 185,
+ 535
+ ],
+ "score": 1.0,
+ "content": "confidence for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 522,
+ 208,
+ 532
+ ],
+ "score": 0.89,
+ "content": "2 . { \\dot { 6 } } \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 521,
+ 506,
+ 535
+ ],
+ "score": 1.0,
+ "content": "of all the classes. Hence there is sufficient evidence to say that the per-class",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 341,
+ 546
+ ],
+ "score": 1.0,
+ "content": "test accuracies is not increased when introducing DA for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 341,
+ 533,
+ 363,
+ 543
+ ],
+ "score": 0.88,
+ "content": "4 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 533,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "of the 1000 Imagenet classes. We",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 543,
+ 506,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 506,
+ 557
+ ],
+ "score": 1.0,
+ "content": "provide in Table 1 the same statistical test but applied on a variety of settings including different",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 554,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "architectures (resnet50, densenet121, ViT-small and ConvNext-Tiny) and across the random crop DA",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "and the weight decay controlled experiments. We should highlight however that this is not necessarily",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "a meaningful measure since for example one regularization might not have any negative impact on the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "classes, but can provide a beneficial gain that is drastically different between classes. Hence, although",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 613
+ ],
+ "score": 1.0,
+ "content": "the per-class performance does not drop by introduce the regularizer, the inter-class performance gap",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 344,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 344,
+ 622
+ ],
+ "score": 1.0,
+ "content": "can be increased by it, which is an equally harmful impact.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 105,
+ 477,
+ 506,
+ 622
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 625,
+ 502,
+ 658
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 625,
+ 504,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 625,
+ 504,
+ 637
+ ],
+ "score": 1.0,
+ "content": "The next Section 2.2 proposes to study the scenario of introducing a pre-trained model, on a different",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 636,
+ 504,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 636,
+ 504,
+ 648
+ ],
+ "score": 1.0,
+ "content": "downstream task to show that the class-dependent effect of regularization remains present and unfair",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 647,
+ 257,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 647,
+ 257,
+ 660
+ ],
+ "score": 1.0,
+ "content": "towards specific downstream classes.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 106,
+ 625,
+ 504,
+ 660
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 671,
+ 407,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 408,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 408,
+ 684
+ ],
+ "score": 1.0,
+ "content": "2.2 The Class-Dependent Bias Transfers to Other Downstream Tasks",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 44
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 689,
+ 504,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "The last experiment we propose is to quantify the amount of class-dependent bias that transfers to",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "other downstream tasks, a common situation in transfer learning and in system deployment to the",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": "real world [Pan and Yang, 2009]. We thus want to measure how regularization applied during the",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 415,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 415,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "pre-training phase on a source dataset impacts the per-class accuracy of that model on the target",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 140,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 140,
+ 441
+ ],
+ "score": 1.0,
+ "content": "dataset.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 46,
+ "bbox_fs": [
+ 105,
+ 688,
+ 506,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 107,
+ 126,
+ 504,
+ 189
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 70,
+ 505,
+ 121
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 69,
+ 506,
+ 82
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 69,
+ 506,
+ 82
+ ],
+ "score": 1.0,
+ "content": "Table 1: Percentage of Imagenet classes for which the test set performance (per-class) is statistically not greater",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 80,
+ 505,
+ 91
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 505,
+ 91
+ ],
+ "score": 1.0,
+ "content": "when applying random crop DA (or weight decay) as measured by the statistical test from Section 2. We observe",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 89,
+ 506,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 89,
+ 506,
+ 102
+ ],
+ "score": 1.0,
+ "content": "that although this measure is highly conservative since a regularizer (DA or weight decay) might now have a",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 100,
+ 506,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 506,
+ 113
+ ],
+ "score": 1.0,
+ "content": "negative impact on a per-class performance but still increase the performance gap between difference classes,",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 109,
+ 502,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 109,
+ 502,
+ 123
+ ],
+ "score": 1.0,
+ "content": "already, a nonzero proportion of classes are negatively impact by introduce random crop DA or weight decay.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 107,
+ 126,
+ 504,
+ 189
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 126,
+ 504,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 126,
+ 504,
+ 189
+ ],
+ "score": 0.961,
+ "html": "
| confidencel | resnet50 | resnet50 random crop|weight decay | convnext-tinyl random crop | weight decay | /|convnext-tiny|ViT random crop|densenet121 random crop | weight decay |
| 90% | 5.9 | 39.4 | 0.1 | 85.5 | 1.2 | 33.8 |
| 95% | 4.5 | 30.3 | 0.1 | 78.8 | 1.0 | 17.1 |
| 99% | 2.6 | 15.3 | 0.0 | 63.9 | 0.7 | 3.3 |
",
+ "type": "table",
+ "image_path": "d14ff8e494a07e4226d3489db971d8c8bdf600feec86e6ea8b3c94cbb8d04650.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 126,
+ 504,
+ 147.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 147.0,
+ 504,
+ 168.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 107,
+ 168.0,
+ 504,
+ 189.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ }
+ ],
+ "index": 4.0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 204,
+ 502,
+ 325
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 204,
+ 502,
+ 325
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 204,
+ 502,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 204,
+ 502,
+ 325
+ ],
+ "score": 0.966,
+ "type": "image",
+ "image_path": "a8a51a7e77d4c0f0b5f318e9248d7d51d5b4a27be610a277dd1e5d171670e4be.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 204,
+ 502,
+ 244.33333333333334
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 107,
+ 244.33333333333334,
+ 502,
+ 284.6666666666667
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 107,
+ 284.6666666666667,
+ 502,
+ 325.0
+ ],
+ "spans": [],
+ "index": 10
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 393
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 344
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Varying the random crop lower-bound during pre-training (Imagenet) also produces unfair per-class bi-",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 355
+ ],
+ "score": 1.0,
+ "content": "ases on the downstream task (INaturalist). Hence, selecting the pre-trained model with best average test accuracy",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 353,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 353,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "on the source dataset might result in deploying a model with the worst performance on the classes of interest in",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 363,
+ 506,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 363,
+ 506,
+ 374
+ ],
+ "score": 1.0,
+ "content": "the target task. Images for each class are provided in Fig. 17, in the appendix. Results obtained by averaging over",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 372,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 372,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "20 runs, official PyTorch resnet50 implementation trained on Imagenet with varying random crop lower bound",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 382,
+ 336,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 382,
+ 336,
+ 393
+ ],
+ "score": 1.0,
+ "content": "and transferred to INaturalist with frozen backbone parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13.5
+ }
+ ],
+ "index": 11.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 416,
+ 504,
+ 439
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 415,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 415,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "pre-training phase on a source dataset impacts the per-class accuracy of that model on the target",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 140,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 140,
+ 441
+ ],
+ "score": 1.0,
+ "content": "dataset.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 443,
+ 505,
+ 554
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 444,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 444,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "In order to keep the setting similar to Section 2.1, we adopt a resnet50 model with random crop DA.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 506,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 415,
+ 468
+ ],
+ "score": 1.0,
+ "content": "That model is pre-trained on Imagenet dataset (source) with varying value of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 457,
+ 423,
+ 465
+ ],
+ "score": 0.71,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 423,
+ 455,
+ 506,
+ 468
+ ],
+ "score": 1.0,
+ "content": "(random crop lower",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 505,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 505,
+ 478
+ ],
+ "score": 1.0,
+ "content": "bound) and then, the trained model is transferred to the INaturalist dataset [Van Horn et al., 2018]",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 477,
+ 505,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 477,
+ 505,
+ 489
+ ],
+ "score": 1.0,
+ "content": "(target) that consists of 10,000 classes. When transferring the model to INaturalist, the parameters",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 500
+ ],
+ "score": 1.0,
+ "content": "are kept frozen, and only a linear classifier is trained on top of it. We report in Fig. 4 the performance",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 506,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 250,
+ 511
+ ],
+ "score": 1.0,
+ "content": "of the trained models with varying",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 500,
+ 258,
+ 509
+ ],
+ "score": 0.72,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 497,
+ 506,
+ 511
+ ],
+ "score": 1.0,
+ "content": "on different INaturalist classes. We observe once again that",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 523
+ ],
+ "score": 1.0,
+ "content": "the best resnet50 —on average— is not necessarily the one that should be deployed as there exists a",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 520,
+ 506,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 520,
+ 252,
+ 533
+ ],
+ "score": 1.0,
+ "content": "strong per-class bias that varies with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 522,
+ 259,
+ 530
+ ],
+ "score": 0.74,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 520,
+ 506,
+ 533
+ ],
+ "score": 1.0,
+ "content": ". As a result, picking the best performing model from a source",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 531,
+ 505,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 531,
+ 505,
+ 544
+ ],
+ "score": 1.0,
+ "content": "dataset to a target dataset, might leave the pipeline to perform poorly since that model might also be",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 421,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 421,
+ 555
+ ],
+ "score": 1.0,
+ "content": "the one that is the most biased against the class of interest in the target dataset.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 558,
+ 505,
+ 603
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "This result should motivate the design of novel regularizers that do not reduce performances between",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 569,
+ 505,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 505,
+ 583
+ ],
+ "score": 1.0,
+ "content": "classes at different regimes. Additionally, due to the cost of training multiple models with varying",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 581,
+ 506,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 506,
+ 592
+ ],
+ "score": 1.0,
+ "content": "regularization settings, one might wonder on the possible alternative solutions to detect trends such",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 359,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 359,
+ 604
+ ],
+ "score": 1.0,
+ "content": "as shown in Fig. 4 only when given a single pre-trained model.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 613,
+ 483,
+ 640
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 612,
+ 484,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 612,
+ 484,
+ 628
+ ],
+ "score": 1.0,
+ "content": "3 Understanding Why and When Regularization Produces Models With",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 123,
+ 627,
+ 277,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 627,
+ 277,
+ 642
+ ],
+ "score": 1.0,
+ "content": "Class-Dependent Preferences",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 645,
+ 506,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": "The first part of our study (Section 2) empirically validated that DNN regularization produces unfair",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 656,
+ 506,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 506,
+ 669
+ ],
+ "score": 1.0,
+ "content": "model complexity control over different classes, resulting in a model performing poorly on a few of",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 667,
+ 506,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 506,
+ 679
+ ],
+ "score": 1.0,
+ "content": "the classes although being highly performing on average. We now provide in Sections 3.1 and 3.2",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "some intuition on why DA can be a source of bias regardless of the task, dataset and model at hand.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "Then, Section 3.3 reviews existing works trying to confront regularization and model bias, and as we",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "will see, an out-of-the-box solution does not seem to exist when there are only a few classes suffering",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 243,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 243,
+ 723
+ ],
+ "score": 1.0,
+ "content": "from regularization (Section 3.4).",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 740,
+ 309,
+ 753
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 740,
+ 309,
+ 753
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 107,
+ 126,
+ 504,
+ 189
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 70,
+ 505,
+ 121
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 69,
+ 506,
+ 82
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 69,
+ 506,
+ 82
+ ],
+ "score": 1.0,
+ "content": "Table 1: Percentage of Imagenet classes for which the test set performance (per-class) is statistically not greater",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 80,
+ 505,
+ 91
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 505,
+ 91
+ ],
+ "score": 1.0,
+ "content": "when applying random crop DA (or weight decay) as measured by the statistical test from Section 2. We observe",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 89,
+ 506,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 89,
+ 506,
+ 102
+ ],
+ "score": 1.0,
+ "content": "that although this measure is highly conservative since a regularizer (DA or weight decay) might now have a",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 100,
+ 506,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 506,
+ 113
+ ],
+ "score": 1.0,
+ "content": "negative impact on a per-class performance but still increase the performance gap between difference classes,",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 109,
+ 502,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 109,
+ 502,
+ 123
+ ],
+ "score": 1.0,
+ "content": "already, a nonzero proportion of classes are negatively impact by introduce random crop DA or weight decay.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 107,
+ 126,
+ 504,
+ 189
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 126,
+ 504,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 126,
+ 504,
+ 189
+ ],
+ "score": 0.961,
+ "html": "| confidencel | resnet50 | resnet50 random crop|weight decay | convnext-tinyl random crop | weight decay | /|convnext-tiny|ViT random crop|densenet121 random crop | weight decay |
| 90% | 5.9 | 39.4 | 0.1 | 85.5 | 1.2 | 33.8 |
| 95% | 4.5 | 30.3 | 0.1 | 78.8 | 1.0 | 17.1 |
| 99% | 2.6 | 15.3 | 0.0 | 63.9 | 0.7 | 3.3 |
",
+ "type": "table",
+ "image_path": "d14ff8e494a07e4226d3489db971d8c8bdf600feec86e6ea8b3c94cbb8d04650.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 126,
+ 504,
+ 147.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 147.0,
+ 504,
+ 168.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 107,
+ 168.0,
+ 504,
+ 189.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ }
+ ],
+ "index": 4.0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 204,
+ 502,
+ 325
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 204,
+ 502,
+ 325
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 204,
+ 502,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 204,
+ 502,
+ 325
+ ],
+ "score": 0.966,
+ "type": "image",
+ "image_path": "a8a51a7e77d4c0f0b5f318e9248d7d51d5b4a27be610a277dd1e5d171670e4be.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 204,
+ 502,
+ 244.33333333333334
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 107,
+ 244.33333333333334,
+ 502,
+ 284.6666666666667
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 107,
+ 284.6666666666667,
+ 502,
+ 325.0
+ ],
+ "spans": [],
+ "index": 10
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 393
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 344
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Varying the random crop lower-bound during pre-training (Imagenet) also produces unfair per-class bi-",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 355
+ ],
+ "score": 1.0,
+ "content": "ases on the downstream task (INaturalist). Hence, selecting the pre-trained model with best average test accuracy",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 353,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 353,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "on the source dataset might result in deploying a model with the worst performance on the classes of interest in",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 363,
+ 506,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 363,
+ 506,
+ 374
+ ],
+ "score": 1.0,
+ "content": "the target task. Images for each class are provided in Fig. 17, in the appendix. Results obtained by averaging over",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 372,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 372,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "20 runs, official PyTorch resnet50 implementation trained on Imagenet with varying random crop lower bound",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 382,
+ 336,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 382,
+ 336,
+ 393
+ ],
+ "score": 1.0,
+ "content": "and transferred to INaturalist with frozen backbone parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13.5
+ }
+ ],
+ "index": 11.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 416,
+ 504,
+ 439
+ ],
+ "lines": [],
+ "index": 17.5,
+ "bbox_fs": [
+ 105,
+ 415,
+ 505,
+ 441
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 443,
+ 505,
+ 554
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 444,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 444,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "In order to keep the setting similar to Section 2.1, we adopt a resnet50 model with random crop DA.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 506,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 415,
+ 468
+ ],
+ "score": 1.0,
+ "content": "That model is pre-trained on Imagenet dataset (source) with varying value of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 457,
+ 423,
+ 465
+ ],
+ "score": 0.71,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 423,
+ 455,
+ 506,
+ 468
+ ],
+ "score": 1.0,
+ "content": "(random crop lower",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 505,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 505,
+ 478
+ ],
+ "score": 1.0,
+ "content": "bound) and then, the trained model is transferred to the INaturalist dataset [Van Horn et al., 2018]",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 477,
+ 505,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 477,
+ 505,
+ 489
+ ],
+ "score": 1.0,
+ "content": "(target) that consists of 10,000 classes. When transferring the model to INaturalist, the parameters",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 500
+ ],
+ "score": 1.0,
+ "content": "are kept frozen, and only a linear classifier is trained on top of it. We report in Fig. 4 the performance",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 506,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 250,
+ 511
+ ],
+ "score": 1.0,
+ "content": "of the trained models with varying",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 500,
+ 258,
+ 509
+ ],
+ "score": 0.72,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 497,
+ 506,
+ 511
+ ],
+ "score": 1.0,
+ "content": "on different INaturalist classes. We observe once again that",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 506,
+ 523
+ ],
+ "score": 1.0,
+ "content": "the best resnet50 —on average— is not necessarily the one that should be deployed as there exists a",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 520,
+ 506,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 520,
+ 252,
+ 533
+ ],
+ "score": 1.0,
+ "content": "strong per-class bias that varies with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 522,
+ 259,
+ 530
+ ],
+ "score": 0.74,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 520,
+ 506,
+ 533
+ ],
+ "score": 1.0,
+ "content": ". As a result, picking the best performing model from a source",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 531,
+ 505,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 531,
+ 505,
+ 544
+ ],
+ "score": 1.0,
+ "content": "dataset to a target dataset, might leave the pipeline to perform poorly since that model might also be",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 421,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 421,
+ 555
+ ],
+ "score": 1.0,
+ "content": "the one that is the most biased against the class of interest in the target dataset.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 105,
+ 444,
+ 506,
+ 555
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 558,
+ 505,
+ 603
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "This result should motivate the design of novel regularizers that do not reduce performances between",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 569,
+ 505,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 505,
+ 583
+ ],
+ "score": 1.0,
+ "content": "classes at different regimes. Additionally, due to the cost of training multiple models with varying",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 581,
+ 506,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 506,
+ 592
+ ],
+ "score": 1.0,
+ "content": "regularization settings, one might wonder on the possible alternative solutions to detect trends such",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 359,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 359,
+ 604
+ ],
+ "score": 1.0,
+ "content": "as shown in Fig. 4 only when given a single pre-trained model.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 105,
+ 558,
+ 506,
+ 604
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 613,
+ 483,
+ 640
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 612,
+ 484,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 612,
+ 484,
+ 628
+ ],
+ "score": 1.0,
+ "content": "3 Understanding Why and When Regularization Produces Models With",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 123,
+ 627,
+ 277,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 627,
+ 277,
+ 642
+ ],
+ "score": 1.0,
+ "content": "Class-Dependent Preferences",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 645,
+ 506,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": "The first part of our study (Section 2) empirically validated that DNN regularization produces unfair",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 656,
+ 506,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 506,
+ 669
+ ],
+ "score": 1.0,
+ "content": "model complexity control over different classes, resulting in a model performing poorly on a few of",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 667,
+ 506,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 506,
+ 679
+ ],
+ "score": 1.0,
+ "content": "the classes although being highly performing on average. We now provide in Sections 3.1 and 3.2",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "some intuition on why DA can be a source of bias regardless of the task, dataset and model at hand.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "Then, Section 3.3 reviews existing works trying to confront regularization and model bias, and as we",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "will see, an out-of-the-box solution does not seem to exist when there are only a few classes suffering",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 243,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 243,
+ 723
+ ],
+ "score": 1.0,
+ "content": "from regularization (Section 3.4).",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 38,
+ "bbox_fs": [
+ 105,
+ 645,
+ 506,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 105,
+ 72,
+ 497,
+ 95
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 71,
+ 498,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 71,
+ 498,
+ 86
+ ],
+ "score": 1.0,
+ "content": "3.1 A Data-Augmentation Policy That is Not Label-Preserving For All Classes Will Create",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 129,
+ 83,
+ 264,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 83,
+ 264,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Class-Imbalance Performances",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 103,
+ 505,
+ 147
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "To provide a simple explanation on how DA causes bias in a trained model, we propose the following",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 505,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 505,
+ 127
+ ],
+ "score": 1.0,
+ "content": "derivation that holds for any signal e.g. timeseries, images, videos. Without loss of generality, we will",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 125,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 125,
+ 176,
+ 138
+ ],
+ "score": 1.0,
+ "content": "consider here the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 126,
+ 186,
+ 136
+ ],
+ "score": 0.87,
+ "content": "\\ell _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 125,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "loss which was shown to perform as well as the cross-entropy even on Imagenet",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 136,
+ 204,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 204,
+ 148
+ ],
+ "score": 1.0,
+ "content": "[Hui and Belkin, 2020].",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 152,
+ 506,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 505,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 257,
+ 165
+ ],
+ "score": 1.0,
+ "content": "Dataset notations. Given a sample",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 257,
+ 153,
+ 290,
+ 163
+ ],
+ "score": 0.9,
+ "content": "\\textbf { \\em x } \\in { \\mathcal { X } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 151,
+ 313,
+ 165
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 314,
+ 152,
+ 355,
+ 163
+ ],
+ "score": 0.92,
+ "content": "\\boldsymbol { \\mathcal { X } } \\subset \\mathbb { R } ^ { D }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 355,
+ 151,
+ 414,
+ 165
+ ],
+ "score": 1.0,
+ "content": ", we consider",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 414,
+ 152,
+ 463,
+ 165
+ ],
+ "score": 0.92,
+ "content": "{ \\pmb y } \\triangleq f ^ { * } ( { \\pmb x } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 463,
+ 151,
+ 505,
+ 165
+ ],
+ "score": 1.0,
+ "content": "to be the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 163,
+ 506,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 163,
+ 380,
+ 177
+ ],
+ "score": 1.0,
+ "content": "ground-truth target value. Hence our hope is to learn an approximator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 164,
+ 390,
+ 175
+ ],
+ "score": 0.87,
+ "content": "f _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 390,
+ 163,
+ 506,
+ 177
+ ],
+ "score": 1.0,
+ "content": "that is as close as possible to",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 107,
+ 174,
+ 506,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 176,
+ 118,
+ 187
+ ],
+ "score": 0.88,
+ "content": "f ^ { * }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 118,
+ 174,
+ 178,
+ 189
+ ],
+ "score": 1.0,
+ "content": "everywhere in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 179,
+ 176,
+ 189,
+ 186
+ ],
+ "score": 0.78,
+ "content": "\\mathcal { X }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 189,
+ 174,
+ 375,
+ 189
+ ],
+ "score": 1.0,
+ "content": ", although we only observe a finite training set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 175,
+ 503,
+ 187
+ ],
+ "score": 0.89,
+ "content": "\\mathbb { X } \\triangleq \\{ ( \\pmb { x } _ { 1 } , \\pmb { y } _ { 1 } ) , \\dots , ( \\pmb { x } _ { N } , \\pmb { y } _ { N } ) \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 504,
+ 174,
+ 506,
+ 189
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 186,
+ 496,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 186,
+ 200,
+ 200
+ ],
+ "score": 1.0,
+ "content": "Given an output vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 200,
+ 189,
+ 208,
+ 197
+ ],
+ "score": 0.67,
+ "content": "\\textbf { \\em u }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 186,
+ 374,
+ 200
+ ],
+ "score": 1.0,
+ "content": "we also define the level-set of a mapping",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 187,
+ 382,
+ 199
+ ],
+ "score": 0.84,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 382,
+ 186,
+ 406,
+ 200
+ ],
+ "score": 1.0,
+ "content": "to be",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 187,
+ 492,
+ 198
+ ],
+ "score": 0.88,
+ "content": "\\{ { \\pmb x } \\in { \\mathcal { X } } : f ( { \\pmb x } ) = { \\pmb u } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 492,
+ 186,
+ 496,
+ 200
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 198,
+ 505,
+ 241
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 196,
+ 506,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 196,
+ 401,
+ 210
+ ],
+ "score": 1.0,
+ "content": "Data-Augmentation notations. Additionally, one employs a DA policy",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 401,
+ 198,
+ 482,
+ 208
+ ],
+ "score": 0.83,
+ "content": "\\mathcal { T } : \\mathbb { R } ^ { D } \\times \\mathcal { K } \\mapsto \\mathbb { R } ^ { D }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 196,
+ 506,
+ 210
+ ],
+ "score": 1.0,
+ "content": "such",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 208,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 208,
+ 263,
+ 220
+ ],
+ "score": 1.0,
+ "content": "that given a transformation parameter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 263,
+ 209,
+ 292,
+ 219
+ ],
+ "score": 0.89,
+ "content": "\\alpha \\in { \\mathcal { K } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 292,
+ 208,
+ 296,
+ 220
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 208,
+ 322,
+ 221
+ ],
+ "score": 0.85,
+ "content": "{ \\mathcal T } _ { \\alpha } ( { \\pmb x } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 323,
+ 208,
+ 465,
+ 220
+ ],
+ "score": 1.0,
+ "content": "produces the transformed view of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 210,
+ 473,
+ 218
+ ],
+ "score": 0.72,
+ "content": "_ { \\textbf { \\em x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 473,
+ 208,
+ 505,
+ 220
+ ],
+ "score": 1.0,
+ "content": ". Often,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 219,
+ 211,
+ 232
+ ],
+ "score": 1.0,
+ "content": "one also defines a density",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 221,
+ 218,
+ 231
+ ],
+ "score": 0.8,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 219,
+ 231,
+ 232
+ ],
+ "score": 1.0,
+ "content": "on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 220,
+ 241,
+ 230
+ ],
+ "score": 0.82,
+ "content": "\\kappa",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 241,
+ 219,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "that helps in sampling transformation parameters that are a priori",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 231,
+ 223,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 231,
+ 223,
+ 241
+ ],
+ "score": 1.0,
+ "content": "known to be the most useful.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 243,
+ 505,
+ 276
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 242,
+ 506,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 242,
+ 328,
+ 257
+ ],
+ "score": 1.0,
+ "content": "Theorem 1. Whenever the transformations produced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 244,
+ 357,
+ 255
+ ],
+ "score": 0.85,
+ "content": "\\mathcal { T } _ { \\alpha } , \\forall \\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 242,
+ 473,
+ 257
+ ],
+ "score": 1.0,
+ "content": "do not respect the level-set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 244,
+ 484,
+ 255
+ ],
+ "score": 0.89,
+ "content": "f ^ { * }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 485,
+ 242,
+ 506,
+ 257
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": "whenever the model has enough capacity to minimize the training loss, the DA will create irreducible",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 169,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 135,
+ 277
+ ],
+ "score": 1.0,
+ "content": "bias in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 136,
+ 265,
+ 146,
+ 276
+ ],
+ "score": 0.87,
+ "content": "f _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 264,
+ 169,
+ 277
+ ],
+ "score": 1.0,
+ "content": "as in",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 111,
+ 280,
+ 486,
+ 324
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 111,
+ 280,
+ 486,
+ 324
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 280,
+ 486,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 280,
+ 486,
+ 324
+ ],
+ "score": 0.798,
+ "type": "image",
+ "image_path": "2ef5e8d9a430737307b40dbef9e6c37e868afc1ea40042ca3717c1fd319a031f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 18,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 111,
+ 280,
+ 486,
+ 294.6666666666667
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 111,
+ 294.6666666666667,
+ 486,
+ 309.33333333333337
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 111,
+ 309.33333333333337,
+ 486,
+ 324.00000000000006
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 333,
+ 505,
+ 367
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 314,
+ 346
+ ],
+ "score": 1.0,
+ "content": "The main idea of the proof, provided in Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 314,
+ 334,
+ 322,
+ 344
+ ],
+ "score": 0.27,
+ "content": "\\mathbf { B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 334,
+ 506,
+ 346
+ ],
+ "score": 1.0,
+ "content": ", is to show that if a transformation does not",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 345,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 444,
+ 357
+ ],
+ "score": 1.0,
+ "content": "move samples on the level-set of the true function (left-hand-side of Eq. (1)), then",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 345,
+ 455,
+ 357
+ ],
+ "score": 0.88,
+ "content": "f _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 345,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "will learn a",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 355,
+ 463,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 355,
+ 334,
+ 369
+ ],
+ "score": 1.0,
+ "content": "different level-set (since it has 0 training error), and thus",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 334,
+ 356,
+ 395,
+ 368
+ ],
+ "score": 0.93,
+ "content": "\\| f ^ { * } - f _ { \\theta } \\| > 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 396,
+ 355,
+ 411,
+ 369
+ ],
+ "score": 1.0,
+ "content": "i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 412,
+ 356,
+ 422,
+ 367
+ ],
+ "score": 0.85,
+ "content": "f _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 423,
+ 355,
+ 463,
+ 369
+ ],
+ "score": 1.0,
+ "content": "is biased.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 371,
+ 505,
+ 449
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 384
+ ],
+ "score": 1.0,
+ "content": "Whenever the left-hand-side of Eq. (1) is 0, the DA is denoted as label-preserving [Cui et al., 2015,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 468,
+ 395
+ ],
+ "score": 1.0,
+ "content": "Taylor and Nitschke, 2018]. From the above, we see that unless the target y associated to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 468,
+ 383,
+ 494,
+ 395
+ ],
+ "score": 0.91,
+ "content": "{ \\mathcal T } _ { \\alpha } ( { \\pmb x } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 383,
+ 505,
+ 395
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 439,
+ 407
+ ],
+ "score": 1.0,
+ "content": "modified accordingly to encode the shift in the target function level-set produced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 439,
+ 394,
+ 451,
+ 405
+ ],
+ "score": 0.86,
+ "content": "\\mathcal { T } _ { \\alpha }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 394,
+ 471,
+ 407
+ ],
+ "score": 1.0,
+ "content": ", any",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 471,
+ 395,
+ 486,
+ 404
+ ],
+ "score": 0.48,
+ "content": "D A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 394,
+ 506,
+ 407
+ ],
+ "score": 1.0,
+ "content": "that",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 417
+ ],
+ "score": 1.0,
+ "content": "is not label-preserving will introduce a bias. Some DAs propose to incorporate label transformation",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 157,
+ 428
+ ],
+ "score": 1.0,
+ "content": "i.e. not only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 417,
+ 165,
+ 426
+ ],
+ "score": 0.73,
+ "content": "_ { \\textbf { \\em x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 416,
+ 200,
+ 428
+ ],
+ "score": 1.0,
+ "content": "but also",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 200,
+ 418,
+ 208,
+ 428
+ ],
+ "score": 0.77,
+ "content": "\\textbf { { y } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 416,
+ 505,
+ 428
+ ],
+ "score": 1.0,
+ "content": "is augmented to better inform on the uncertainty that has been added into",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 107,
+ 426,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 426,
+ 132,
+ 438
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { T } _ { \\boldsymbol { \\theta } } ( \\mathbf { \\mathscr { x } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 426,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": ". This is for example the case for MixUp [Zhang et al., 2017], ManifoldMixUp [Verma et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 437,
+ 325,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 325,
+ 450
+ ],
+ "score": 1.0,
+ "content": "2019], CutMix [Yun et al., 2019] and their extensions.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 454,
+ 505,
+ 498
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 507,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 507,
+ 468
+ ],
+ "score": 1.0,
+ "content": "Our goal in the next Section 3.2 is to demonstrate how DAs such as random crop, color jittering,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 506,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 338,
+ 478
+ ],
+ "score": 1.0,
+ "content": "or CutOut are only label preserving for some values of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 466,
+ 346,
+ 475
+ ],
+ "score": 0.76,
+ "content": "\\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 346,
+ 465,
+ 506,
+ 478
+ ],
+ "score": 1.0,
+ "content": "that vary with the sample class. As a",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 488
+ ],
+ "score": 1.0,
+ "content": "consequence, while the use of the DA improves the average test performance, it is at the cost of a",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 487,
+ 349,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 487,
+ 349,
+ 499
+ ],
+ "score": 1.0,
+ "content": "significant reduction in performance for some of the classes.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 31.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 507,
+ 504,
+ 520
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 505,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 505,
+ 521
+ ],
+ "score": 1.0,
+ "content": "3.2 The Same Data-Augmentation can be Label-Preserving or Not Between Different Classes",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 525,
+ 505,
+ 591
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "In the previous Section 3.1 we provided a general argument on the sufficient conditions for DA to",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "produce a biased model. We hope in this section to provide a more concrete example that applies",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "score": 1.0,
+ "content": "to current DNN training. To that end, we will demonstrate that a DA can be label-preserving or not",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "depending on the sample’s class, hence, since the same DA policy is employed for all classes, the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 104,
+ 568,
+ 507,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 568,
+ 507,
+ 582
+ ],
+ "score": 1.0,
+ "content": "augmented dataset will exhibit a class-imbalance in favor of the classes for which the DA is most",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 177,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 177,
+ 594
+ ],
+ "score": 1.0,
+ "content": "label-preserving.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 596,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 507,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 259,
+ 610
+ ],
+ "score": 1.0,
+ "content": "To measure by how much a given DA,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 596,
+ 271,
+ 607
+ ],
+ "score": 0.86,
+ "content": "\\mathcal { T } _ { \\alpha }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 272,
+ 595,
+ 507,
+ 610
+ ],
+ "score": 1.0,
+ "content": ", is label-preserving, we propose to take 6 popular architec-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "tures that are pre-trained on Imagenet [Deng et al., 2009] from the official PyTorch [Paszke et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "2019] repository, and to evaluate their performances for varying DA settings (top of Fig. 5). We",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 506,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 506,
+ 642
+ ],
+ "score": 1.0,
+ "content": "observe that considering the dataset as a whole is not a good indicator of the optimal DA value to",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 653
+ ],
+ "score": 1.0,
+ "content": "employ since per-class accuracy performance (bottom of Fig. 5) vary drastically. For example, for",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 650,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 650,
+ 268,
+ 663
+ ],
+ "score": 1.0,
+ "content": "some classes, any level of transformation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 653,
+ 276,
+ 660
+ ],
+ "score": 0.76,
+ "content": "\\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 650,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": "can produce augmented samples with enough information",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 662,
+ 505,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 662,
+ 505,
+ 674
+ ],
+ "score": 1.0,
+ "content": "to be correctly classified, while for other classes, even a small amount of DA makes the samples",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 673,
+ 165,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 673,
+ 165,
+ 685
+ ],
+ "score": 1.0,
+ "content": "unpredictable.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 44.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 689,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 506,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 506,
+ 701
+ ],
+ "score": 1.0,
+ "content": "To further ensure that the observed relation between label-preservation, sample class, and amount of",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 168,
+ 712
+ ],
+ "score": 1.0,
+ "content": "transformation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 702,
+ 177,
+ 710
+ ],
+ "score": 0.76,
+ "content": "\\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "is sound, we provide in Fig. 6 the per-class test accuracy on different models, all",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "exhibit the same trends. In short, we identify that when creating an augmented dataset by applying the",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 105,
+ 72,
+ 497,
+ 95
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 71,
+ 498,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 71,
+ 498,
+ 86
+ ],
+ "score": 1.0,
+ "content": "3.1 A Data-Augmentation Policy That is Not Label-Preserving For All Classes Will Create",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 129,
+ 83,
+ 264,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 83,
+ 264,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Class-Imbalance Performances",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 103,
+ 505,
+ 147
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "To provide a simple explanation on how DA causes bias in a trained model, we propose the following",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 505,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 505,
+ 127
+ ],
+ "score": 1.0,
+ "content": "derivation that holds for any signal e.g. timeseries, images, videos. Without loss of generality, we will",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 125,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 125,
+ 176,
+ 138
+ ],
+ "score": 1.0,
+ "content": "consider here the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 126,
+ 186,
+ 136
+ ],
+ "score": 0.87,
+ "content": "\\ell _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 125,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "loss which was shown to perform as well as the cross-entropy even on Imagenet",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 136,
+ 204,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 204,
+ 148
+ ],
+ "score": 1.0,
+ "content": "[Hui and Belkin, 2020].",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 105,
+ 102,
+ 505,
+ 148
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 152,
+ 506,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 505,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 257,
+ 165
+ ],
+ "score": 1.0,
+ "content": "Dataset notations. Given a sample",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 257,
+ 153,
+ 290,
+ 163
+ ],
+ "score": 0.9,
+ "content": "\\textbf { \\em x } \\in { \\mathcal { X } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 151,
+ 313,
+ 165
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 314,
+ 152,
+ 355,
+ 163
+ ],
+ "score": 0.92,
+ "content": "\\boldsymbol { \\mathcal { X } } \\subset \\mathbb { R } ^ { D }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 355,
+ 151,
+ 414,
+ 165
+ ],
+ "score": 1.0,
+ "content": ", we consider",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 414,
+ 152,
+ 463,
+ 165
+ ],
+ "score": 0.92,
+ "content": "{ \\pmb y } \\triangleq f ^ { * } ( { \\pmb x } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 463,
+ 151,
+ 505,
+ 165
+ ],
+ "score": 1.0,
+ "content": "to be the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 163,
+ 506,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 163,
+ 380,
+ 177
+ ],
+ "score": 1.0,
+ "content": "ground-truth target value. Hence our hope is to learn an approximator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 164,
+ 390,
+ 175
+ ],
+ "score": 0.87,
+ "content": "f _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 390,
+ 163,
+ 506,
+ 177
+ ],
+ "score": 1.0,
+ "content": "that is as close as possible to",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 107,
+ 174,
+ 506,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 176,
+ 118,
+ 187
+ ],
+ "score": 0.88,
+ "content": "f ^ { * }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 118,
+ 174,
+ 178,
+ 189
+ ],
+ "score": 1.0,
+ "content": "everywhere in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 179,
+ 176,
+ 189,
+ 186
+ ],
+ "score": 0.78,
+ "content": "\\mathcal { X }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 189,
+ 174,
+ 375,
+ 189
+ ],
+ "score": 1.0,
+ "content": ", although we only observe a finite training set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 175,
+ 503,
+ 187
+ ],
+ "score": 0.89,
+ "content": "\\mathbb { X } \\triangleq \\{ ( \\pmb { x } _ { 1 } , \\pmb { y } _ { 1 } ) , \\dots , ( \\pmb { x } _ { N } , \\pmb { y } _ { N } ) \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 504,
+ 174,
+ 506,
+ 189
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 186,
+ 496,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 186,
+ 200,
+ 200
+ ],
+ "score": 1.0,
+ "content": "Given an output vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 200,
+ 189,
+ 208,
+ 197
+ ],
+ "score": 0.67,
+ "content": "\\textbf { \\em u }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 186,
+ 374,
+ 200
+ ],
+ "score": 1.0,
+ "content": "we also define the level-set of a mapping",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 187,
+ 382,
+ 199
+ ],
+ "score": 0.84,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 382,
+ 186,
+ 406,
+ 200
+ ],
+ "score": 1.0,
+ "content": "to be",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 187,
+ 492,
+ 198
+ ],
+ "score": 0.88,
+ "content": "\\{ { \\pmb x } \\in { \\mathcal { X } } : f ( { \\pmb x } ) = { \\pmb u } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 492,
+ 186,
+ 496,
+ 200
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 105,
+ 151,
+ 506,
+ 200
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 198,
+ 505,
+ 241
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 196,
+ 506,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 196,
+ 401,
+ 210
+ ],
+ "score": 1.0,
+ "content": "Data-Augmentation notations. Additionally, one employs a DA policy",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 401,
+ 198,
+ 482,
+ 208
+ ],
+ "score": 0.83,
+ "content": "\\mathcal { T } : \\mathbb { R } ^ { D } \\times \\mathcal { K } \\mapsto \\mathbb { R } ^ { D }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 196,
+ 506,
+ 210
+ ],
+ "score": 1.0,
+ "content": "such",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 208,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 208,
+ 263,
+ 220
+ ],
+ "score": 1.0,
+ "content": "that given a transformation parameter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 263,
+ 209,
+ 292,
+ 219
+ ],
+ "score": 0.89,
+ "content": "\\alpha \\in { \\mathcal { K } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 292,
+ 208,
+ 296,
+ 220
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 208,
+ 322,
+ 221
+ ],
+ "score": 0.85,
+ "content": "{ \\mathcal T } _ { \\alpha } ( { \\pmb x } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 323,
+ 208,
+ 465,
+ 220
+ ],
+ "score": 1.0,
+ "content": "produces the transformed view of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 210,
+ 473,
+ 218
+ ],
+ "score": 0.72,
+ "content": "_ { \\textbf { \\em x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 473,
+ 208,
+ 505,
+ 220
+ ],
+ "score": 1.0,
+ "content": ". Often,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 219,
+ 211,
+ 232
+ ],
+ "score": 1.0,
+ "content": "one also defines a density",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 221,
+ 218,
+ 231
+ ],
+ "score": 0.8,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 219,
+ 231,
+ 232
+ ],
+ "score": 1.0,
+ "content": "on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 220,
+ 241,
+ 230
+ ],
+ "score": 0.82,
+ "content": "\\kappa",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 241,
+ 219,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "that helps in sampling transformation parameters that are a priori",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 231,
+ 223,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 231,
+ 223,
+ 241
+ ],
+ "score": 1.0,
+ "content": "known to be the most useful.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 196,
+ 506,
+ 241
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 243,
+ 505,
+ 276
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 242,
+ 506,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 242,
+ 328,
+ 257
+ ],
+ "score": 1.0,
+ "content": "Theorem 1. Whenever the transformations produced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 244,
+ 357,
+ 255
+ ],
+ "score": 0.85,
+ "content": "\\mathcal { T } _ { \\alpha } , \\forall \\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 242,
+ 473,
+ 257
+ ],
+ "score": 1.0,
+ "content": "do not respect the level-set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 244,
+ 484,
+ 255
+ ],
+ "score": 0.89,
+ "content": "f ^ { * }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 485,
+ 242,
+ 506,
+ 257
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": "whenever the model has enough capacity to minimize the training loss, the DA will create irreducible",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 169,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 135,
+ 277
+ ],
+ "score": 1.0,
+ "content": "bias in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 136,
+ 265,
+ 146,
+ 276
+ ],
+ "score": 0.87,
+ "content": "f _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 264,
+ 169,
+ 277
+ ],
+ "score": 1.0,
+ "content": "as in",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 105,
+ 242,
+ 506,
+ 277
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 111,
+ 280,
+ 486,
+ 324
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 111,
+ 280,
+ 486,
+ 324
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 280,
+ 486,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 280,
+ 486,
+ 324
+ ],
+ "score": 0.798,
+ "type": "image",
+ "image_path": "2ef5e8d9a430737307b40dbef9e6c37e868afc1ea40042ca3717c1fd319a031f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 18,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 111,
+ 280,
+ 486,
+ 294.6666666666667
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 111,
+ 294.6666666666667,
+ 486,
+ 309.33333333333337
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 111,
+ 309.33333333333337,
+ 486,
+ 324.00000000000006
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 333,
+ 505,
+ 367
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 314,
+ 346
+ ],
+ "score": 1.0,
+ "content": "The main idea of the proof, provided in Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 314,
+ 334,
+ 322,
+ 344
+ ],
+ "score": 0.27,
+ "content": "\\mathbf { B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 334,
+ 506,
+ 346
+ ],
+ "score": 1.0,
+ "content": ", is to show that if a transformation does not",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 345,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 444,
+ 357
+ ],
+ "score": 1.0,
+ "content": "move samples on the level-set of the true function (left-hand-side of Eq. (1)), then",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 345,
+ 455,
+ 357
+ ],
+ "score": 0.88,
+ "content": "f _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 345,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "will learn a",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 355,
+ 463,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 355,
+ 334,
+ 369
+ ],
+ "score": 1.0,
+ "content": "different level-set (since it has 0 training error), and thus",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 334,
+ 356,
+ 395,
+ 368
+ ],
+ "score": 0.93,
+ "content": "\\| f ^ { * } - f _ { \\theta } \\| > 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 396,
+ 355,
+ 411,
+ 369
+ ],
+ "score": 1.0,
+ "content": "i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 412,
+ 356,
+ 422,
+ 367
+ ],
+ "score": 0.85,
+ "content": "f _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 423,
+ 355,
+ 463,
+ 369
+ ],
+ "score": 1.0,
+ "content": "is biased.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 105,
+ 334,
+ 506,
+ 369
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 371,
+ 505,
+ 449
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 384
+ ],
+ "score": 1.0,
+ "content": "Whenever the left-hand-side of Eq. (1) is 0, the DA is denoted as label-preserving [Cui et al., 2015,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 468,
+ 395
+ ],
+ "score": 1.0,
+ "content": "Taylor and Nitschke, 2018]. From the above, we see that unless the target y associated to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 468,
+ 383,
+ 494,
+ 395
+ ],
+ "score": 0.91,
+ "content": "{ \\mathcal T } _ { \\alpha } ( { \\pmb x } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 383,
+ 505,
+ 395
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 439,
+ 407
+ ],
+ "score": 1.0,
+ "content": "modified accordingly to encode the shift in the target function level-set produced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 439,
+ 394,
+ 451,
+ 405
+ ],
+ "score": 0.86,
+ "content": "\\mathcal { T } _ { \\alpha }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 394,
+ 471,
+ 407
+ ],
+ "score": 1.0,
+ "content": ", any",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 471,
+ 395,
+ 486,
+ 404
+ ],
+ "score": 0.48,
+ "content": "D A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 394,
+ 506,
+ 407
+ ],
+ "score": 1.0,
+ "content": "that",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 417
+ ],
+ "score": 1.0,
+ "content": "is not label-preserving will introduce a bias. Some DAs propose to incorporate label transformation",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 157,
+ 428
+ ],
+ "score": 1.0,
+ "content": "i.e. not only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 417,
+ 165,
+ 426
+ ],
+ "score": 0.73,
+ "content": "_ { \\textbf { \\em x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 416,
+ 200,
+ 428
+ ],
+ "score": 1.0,
+ "content": "but also",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 200,
+ 418,
+ 208,
+ 428
+ ],
+ "score": 0.77,
+ "content": "\\textbf { { y } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 416,
+ 505,
+ 428
+ ],
+ "score": 1.0,
+ "content": "is augmented to better inform on the uncertainty that has been added into",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 107,
+ 426,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 426,
+ 132,
+ 438
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { T } _ { \\boldsymbol { \\theta } } ( \\mathbf { \\mathscr { x } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 426,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": ". This is for example the case for MixUp [Zhang et al., 2017], ManifoldMixUp [Verma et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 437,
+ 325,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 325,
+ 450
+ ],
+ "score": 1.0,
+ "content": "2019], CutMix [Yun et al., 2019] and their extensions.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 105,
+ 371,
+ 506,
+ 450
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 454,
+ 505,
+ 498
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 507,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 507,
+ 468
+ ],
+ "score": 1.0,
+ "content": "Our goal in the next Section 3.2 is to demonstrate how DAs such as random crop, color jittering,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 506,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 338,
+ 478
+ ],
+ "score": 1.0,
+ "content": "or CutOut are only label preserving for some values of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 466,
+ 346,
+ 475
+ ],
+ "score": 0.76,
+ "content": "\\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 346,
+ 465,
+ 506,
+ 478
+ ],
+ "score": 1.0,
+ "content": "that vary with the sample class. As a",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 488
+ ],
+ "score": 1.0,
+ "content": "consequence, while the use of the DA improves the average test performance, it is at the cost of a",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 487,
+ 349,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 487,
+ 349,
+ 499
+ ],
+ "score": 1.0,
+ "content": "significant reduction in performance for some of the classes.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 31.5,
+ "bbox_fs": [
+ 105,
+ 452,
+ 507,
+ 499
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 507,
+ 504,
+ 520
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 505,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 505,
+ 521
+ ],
+ "score": 1.0,
+ "content": "3.2 The Same Data-Augmentation can be Label-Preserving or Not Between Different Classes",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 525,
+ 505,
+ 591
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "In the previous Section 3.1 we provided a general argument on the sufficient conditions for DA to",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "produce a biased model. We hope in this section to provide a more concrete example that applies",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "score": 1.0,
+ "content": "to current DNN training. To that end, we will demonstrate that a DA can be label-preserving or not",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "depending on the sample’s class, hence, since the same DA policy is employed for all classes, the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 104,
+ 568,
+ 507,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 568,
+ 507,
+ 582
+ ],
+ "score": 1.0,
+ "content": "augmented dataset will exhibit a class-imbalance in favor of the classes for which the DA is most",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 177,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 177,
+ 594
+ ],
+ "score": 1.0,
+ "content": "label-preserving.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 104,
+ 525,
+ 507,
+ 594
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 596,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 507,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 259,
+ 610
+ ],
+ "score": 1.0,
+ "content": "To measure by how much a given DA,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 596,
+ 271,
+ 607
+ ],
+ "score": 0.86,
+ "content": "\\mathcal { T } _ { \\alpha }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 272,
+ 595,
+ 507,
+ 610
+ ],
+ "score": 1.0,
+ "content": ", is label-preserving, we propose to take 6 popular architec-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "tures that are pre-trained on Imagenet [Deng et al., 2009] from the official PyTorch [Paszke et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "2019] repository, and to evaluate their performances for varying DA settings (top of Fig. 5). We",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 506,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 506,
+ 642
+ ],
+ "score": 1.0,
+ "content": "observe that considering the dataset as a whole is not a good indicator of the optimal DA value to",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 653
+ ],
+ "score": 1.0,
+ "content": "employ since per-class accuracy performance (bottom of Fig. 5) vary drastically. For example, for",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 650,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 650,
+ 268,
+ 663
+ ],
+ "score": 1.0,
+ "content": "some classes, any level of transformation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 653,
+ 276,
+ 660
+ ],
+ "score": 0.76,
+ "content": "\\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 650,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": "can produce augmented samples with enough information",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 662,
+ 505,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 662,
+ 505,
+ 674
+ ],
+ "score": 1.0,
+ "content": "to be correctly classified, while for other classes, even a small amount of DA makes the samples",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 673,
+ 165,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 673,
+ 165,
+ 685
+ ],
+ "score": 1.0,
+ "content": "unpredictable.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 44.5,
+ "bbox_fs": [
+ 105,
+ 595,
+ 507,
+ 685
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 689,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 506,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 506,
+ 701
+ ],
+ "score": 1.0,
+ "content": "To further ensure that the observed relation between label-preservation, sample class, and amount of",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 168,
+ 712
+ ],
+ "score": 1.0,
+ "content": "transformation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 702,
+ 177,
+ 710
+ ],
+ "score": 0.76,
+ "content": "\\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "is sound, we provide in Fig. 6 the per-class test accuracy on different models, all",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "exhibit the same trends. In short, we identify that when creating an augmented dataset by applying the",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "same DA across classes, the number of per-class samples that actually contain enough information",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "about their true labels will become largely imbalance between classes, even if the original dataset",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 620
+ ],
+ "score": 1.0,
+ "content": "was balanced. Any model trained on the augmented dataset will thus focus on the classes for which",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 253,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 253,
+ 633
+ ],
+ "score": 1.0,
+ "content": "the DA is the most label-preserving.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 50,
+ "bbox_fs": [
+ 105,
+ 689,
+ 506,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 69,
+ 502,
+ 331
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 69,
+ 502,
+ 331
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 69,
+ 502,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 69,
+ 502,
+ 331
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "9be1a27a21ea55563825551ac6014351c52449c918b05b84800756d2bc783a05.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 69,
+ 502,
+ 156.33333333333331
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 156.33333333333331,
+ 502,
+ 243.66666666666663
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 243.66666666666663,
+ 502,
+ 330.99999999999994
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 413
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 345
+ ],
+ "score": 1.0,
+ "content": "Figure 5: Top: examples of an augmented image of class “bird”. Middle: average accuracy (train set in dashed",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 354
+ ],
+ "score": 1.0,
+ "content": "line and test set in plain) on Imagenet, using 6 popular architectures. Bottom: per-class performances from the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 365
+ ],
+ "score": 1.0,
+ "content": "middle scenario along with 9 images of the corresponding classes. We observe that the random crop DA seems",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 362,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 362,
+ 332,
+ 375
+ ],
+ "score": 1.0,
+ "content": "to loose its label-preserving property on average when less than",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 333,
+ 363,
+ 350,
+ 372
+ ],
+ "score": 0.87,
+ "content": "30 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 350,
+ 362,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "of the image is kept in the crop but looking",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 373,
+ 435,
+ 384
+ ],
+ "score": 1.0,
+ "content": "at the per-class performance, we observe that such DA can be label-preserving with only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 435,
+ 373,
+ 449,
+ 382
+ ],
+ "score": 0.84,
+ "content": "8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 373,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "of the original",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 453,
+ 394
+ ],
+ "score": 1.0,
+ "content": "image for some classes, while for other classes the label information starts to reduce at around",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 453,
+ 383,
+ 471,
+ 392
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 472,
+ 383,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": ". Results",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 404
+ ],
+ "score": 1.0,
+ "content": "obtained from the official Imagenet pre-trained PyTorch models. CutOut and ColorJitter cases are provided in",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 403,
+ 263,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 263,
+ 414
+ ],
+ "score": 1.0,
+ "content": "Figs. 14 and 15 and exhibit the same trend.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 6.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 110,
+ 418,
+ 497,
+ 527
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 110,
+ 418,
+ 497,
+ 527
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 418,
+ 497,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 418,
+ 497,
+ 527
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "55ce512e0d33f69dd7ef53b1d5f49f2f68c03d0a122ec2f4392926e06cf02245.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 418,
+ 497,
+ 454.3333333333333
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 110,
+ 454.3333333333333,
+ 497,
+ 490.66666666666663
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 110,
+ 490.66666666666663,
+ 497,
+ 527.0
+ ],
+ "spans": [],
+ "index": 13
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 528,
+ 505,
+ 569
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 527,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 527,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Reprise of the bottom left of Fig. 5 for three different DAs (each column) and using the same",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 550
+ ],
+ "score": 1.0,
+ "content": "6 popular architectures (different lines). We observe that across DAs, different architectures agree on the",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 548,
+ 505,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 548,
+ 208,
+ 560
+ ],
+ "score": 1.0,
+ "content": "label-preserving regimes for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 548,
+ 219,
+ 558
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { T } _ { \\alpha }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 220,
+ 548,
+ 505,
+ 560
+ ],
+ "score": 1.0,
+ "content": "i.e. even an ensemble of model would not reduce the class-dependent bias of the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 558,
+ 428,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 558,
+ 428,
+ 570
+ ],
+ "score": 1.0,
+ "content": "final prediction. Results obtained from the official Imagenet pre-trained PyTorch models.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 15.5
+ }
+ ],
+ "index": 13.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 586,
+ 505,
+ 631
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "same DA across classes, the number of per-class samples that actually contain enough information",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "about their true labels will become largely imbalance between classes, even if the original dataset",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 620
+ ],
+ "score": 1.0,
+ "content": "was balanced. Any model trained on the augmented dataset will thus focus on the classes for which",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 253,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 253,
+ 633
+ ],
+ "score": 1.0,
+ "content": "the DA is the most label-preserving.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 635,
+ 504,
+ 658
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 505,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 505,
+ 648
+ ],
+ "score": 1.0,
+ "content": "Beyond the above intuitive and natural understand we obtained in term of data-augmentation, there",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 647,
+ 354,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 647,
+ 354,
+ 658
+ ],
+ "score": 1.0,
+ "content": "exists theoretical studies that we propose to summarize below",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 105,
+ 667,
+ 493,
+ 689
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 493,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 493,
+ 680
+ ],
+ "score": 1.0,
+ "content": "3.3 Do Existing Studies Provide Answers Into the Inter-Play Between Regularization and",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 128,
+ 677,
+ 229,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 677,
+ 229,
+ 690
+ ],
+ "score": 1.0,
+ "content": "Class-Dependent Bias?",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 699,
+ 503,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "We explored in Section 3.2 a possible explanation of the class-dependent bias we observed in Section 2",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "hinting at the need to use class-dependent DA. We propose here to summarize existing studies that",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "score": 1.0,
+ "content": "7",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 69,
+ 502,
+ 331
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 69,
+ 502,
+ 331
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 69,
+ 502,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 69,
+ 502,
+ 331
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "9be1a27a21ea55563825551ac6014351c52449c918b05b84800756d2bc783a05.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 69,
+ 502,
+ 156.33333333333331
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 156.33333333333331,
+ 502,
+ 243.66666666666663
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 243.66666666666663,
+ 502,
+ 330.99999999999994
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 413
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 345
+ ],
+ "score": 1.0,
+ "content": "Figure 5: Top: examples of an augmented image of class “bird”. Middle: average accuracy (train set in dashed",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 354
+ ],
+ "score": 1.0,
+ "content": "line and test set in plain) on Imagenet, using 6 popular architectures. Bottom: per-class performances from the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 365
+ ],
+ "score": 1.0,
+ "content": "middle scenario along with 9 images of the corresponding classes. We observe that the random crop DA seems",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 362,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 362,
+ 332,
+ 375
+ ],
+ "score": 1.0,
+ "content": "to loose its label-preserving property on average when less than",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 333,
+ 363,
+ 350,
+ 372
+ ],
+ "score": 0.87,
+ "content": "30 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 350,
+ 362,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "of the image is kept in the crop but looking",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 373,
+ 435,
+ 384
+ ],
+ "score": 1.0,
+ "content": "at the per-class performance, we observe that such DA can be label-preserving with only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 435,
+ 373,
+ 449,
+ 382
+ ],
+ "score": 0.84,
+ "content": "8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 373,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "of the original",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 453,
+ 394
+ ],
+ "score": 1.0,
+ "content": "image for some classes, while for other classes the label information starts to reduce at around",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 453,
+ 383,
+ 471,
+ 392
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 472,
+ 383,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": ". Results",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 404
+ ],
+ "score": 1.0,
+ "content": "obtained from the official Imagenet pre-trained PyTorch models. CutOut and ColorJitter cases are provided in",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 403,
+ 263,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 263,
+ 414
+ ],
+ "score": 1.0,
+ "content": "Figs. 14 and 15 and exhibit the same trend.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 6.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 110,
+ 418,
+ 497,
+ 527
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 110,
+ 418,
+ 497,
+ 527
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 418,
+ 497,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 418,
+ 497,
+ 527
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "55ce512e0d33f69dd7ef53b1d5f49f2f68c03d0a122ec2f4392926e06cf02245.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 418,
+ 497,
+ 454.3333333333333
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 110,
+ 454.3333333333333,
+ 497,
+ 490.66666666666663
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 110,
+ 490.66666666666663,
+ 497,
+ 527.0
+ ],
+ "spans": [],
+ "index": 13
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 528,
+ 505,
+ 569
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 527,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 527,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Reprise of the bottom left of Fig. 5 for three different DAs (each column) and using the same",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 550
+ ],
+ "score": 1.0,
+ "content": "6 popular architectures (different lines). We observe that across DAs, different architectures agree on the",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 548,
+ 505,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 548,
+ 208,
+ 560
+ ],
+ "score": 1.0,
+ "content": "label-preserving regimes for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 548,
+ 219,
+ 558
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { T } _ { \\alpha }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 220,
+ 548,
+ 505,
+ 560
+ ],
+ "score": 1.0,
+ "content": "i.e. even an ensemble of model would not reduce the class-dependent bias of the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 558,
+ 428,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 558,
+ 428,
+ 570
+ ],
+ "score": 1.0,
+ "content": "final prediction. Results obtained from the official Imagenet pre-trained PyTorch models.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 15.5
+ }
+ ],
+ "index": 13.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 586,
+ 505,
+ 631
+ ],
+ "lines": [],
+ "index": 19.5,
+ "bbox_fs": [
+ 105,
+ 586,
+ 506,
+ 633
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 635,
+ 504,
+ 658
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 505,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 505,
+ 648
+ ],
+ "score": 1.0,
+ "content": "Beyond the above intuitive and natural understand we obtained in term of data-augmentation, there",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 647,
+ 354,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 647,
+ 354,
+ 658
+ ],
+ "score": 1.0,
+ "content": "exists theoretical studies that we propose to summarize below",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 105,
+ 635,
+ 505,
+ 658
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 105,
+ 667,
+ 493,
+ 689
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 493,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 493,
+ 680
+ ],
+ "score": 1.0,
+ "content": "3.3 Do Existing Studies Provide Answers Into the Inter-Play Between Regularization and",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 128,
+ 677,
+ 229,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 677,
+ 229,
+ 690
+ ],
+ "score": 1.0,
+ "content": "Class-Dependent Bias?",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 699,
+ 503,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "We explored in Section 3.2 a possible explanation of the class-dependent bias we observed in Section 2",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "hinting at the need to use class-dependent DA. We propose here to summarize existing studies that",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 281
+ ],
+ "score": 1.0,
+ "content": "are attempting to better understand the impact of DA and regularization in general onto a model’s",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 391,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 391,
+ 292
+ ],
+ "score": 1.0,
+ "content": "performance and bias. Then, Section 3.4 will explore their application.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 106,
+ 699,
+ 505,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 72,
+ 325,
+ 216
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 105,
+ 72,
+ 325,
+ 216
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 325,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 325,
+ 216
+ ],
+ "score": 0.963,
+ "type": "image",
+ "image_path": "382a91d715d38ca9b29e90abc548d2f9ac19acd3a0987ee91d1f7d922caba770.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 325,
+ 84.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 84.0,
+ 325,
+ 96.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 96.0,
+ 325,
+ 108.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 108.0,
+ 325,
+ 120.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 120.0,
+ 325,
+ 132.0
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 132.0,
+ 325,
+ 144.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 144.0,
+ 325,
+ 156.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 156.0,
+ 325,
+ 168.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 168.0,
+ 325,
+ 180.0
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 180.0,
+ 325,
+ 192.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 192.0,
+ 325,
+ 204.0
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 204.0,
+ 325,
+ 216.0
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 328,
+ 76,
+ 504,
+ 217
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 326,
+ 76,
+ 506,
+ 89
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 76,
+ 506,
+ 89
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Reprise of Fig. 2 but now implement-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 326,
+ 87,
+ 506,
+ 99
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 87,
+ 506,
+ 99
+ ],
+ "score": 1.0,
+ "content": "ing the class-dependent DA as prescribed in Sec-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 326,
+ 96,
+ 505,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 96,
+ 505,
+ 109
+ ],
+ "score": 1.0,
+ "content": "tion 3.3 i.e. we only apply DA (random crop) to",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 326,
+ 107,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 107,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "the classes that see their per-class accuracy im-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 326,
+ 117,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 117,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "prove when this DA is employed during training..",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 326,
+ 127,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 127,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "We observe here that the bias of the model intro-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 327,
+ 137,
+ 504,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 327,
+ 137,
+ 504,
+ 148
+ ],
+ "score": 1.0,
+ "content": "duced from random crop on the majority of the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 327,
+ 147,
+ 505,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 327,
+ 147,
+ 505,
+ 158
+ ],
+ "score": 1.0,
+ "content": "classes spills-over to the minority of the classes",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 326,
+ 156,
+ 505,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 156,
+ 505,
+ 168
+ ],
+ "score": 1.0,
+ "content": "that do not benefit from that DA, even though",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 326,
+ 167,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 167,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "such classes never received that DA during train-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 326,
+ 177,
+ 504,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 177,
+ 504,
+ 188
+ ],
+ "score": 1.0,
+ "content": "ing. Results are averaged over 5 runs and employ",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 326,
+ 187,
+ 504,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 187,
+ 504,
+ 198
+ ],
+ "score": 1.0,
+ "content": "the official resnet50 implementation trained on",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 326,
+ 196,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 196,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Imagenet with horizontal flip but a per-class ran-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 326,
+ 207,
+ 381,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 207,
+ 381,
+ 218
+ ],
+ "score": 1.0,
+ "content": "dom crop DA.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 18.5
+ }
+ ],
+ "index": 12.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 269,
+ 504,
+ 291
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 281
+ ],
+ "score": 1.0,
+ "content": "are attempting to better understand the impact of DA and regularization in general onto a model’s",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 391,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 391,
+ 292
+ ],
+ "score": 1.0,
+ "content": "performance and bias. Then, Section 3.4 will explore their application.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 296,
+ 505,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 506,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 506,
+ 308
+ ],
+ "score": 1.0,
+ "content": "Input independent DA exacerbates the bias already present in a dataset. Especially relevant to",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 304,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 304,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "our results is a recent result of Xu et al. [2020]. In this work, it was theorized that when the underlying",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 318,
+ 506,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 318,
+ 506,
+ 330
+ ],
+ "score": 1.0,
+ "content": "dataset contains inherent biases, training on the original data is more effective than employing an i.i.d.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 328,
+ 506,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 328,
+ 506,
+ 342
+ ],
+ "score": 1.0,
+ "content": "DA policy, i.e. applying the same random augmentation to all samples/classes, to produce an unbiased",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 339,
+ 506,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 339,
+ 506,
+ 352
+ ],
+ "score": 1.0,
+ "content": "model. In short, the DA exacerbates the already present biases and makes the trained model further",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 363
+ ],
+ "score": 1.0,
+ "content": "away from the unbiased optimum. The difficulty of this result lies in defining bias for real images.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "As per our experiments from Section 3, we observe that bias can take many form e.g. one class might",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "naturally represent its object always under the same angle. This is particularly true say for boats",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 384,
+ 506,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 384,
+ 506,
+ 396
+ ],
+ "score": 1.0,
+ "content": "which are rarely captured upside-down. Hence, simply having classes with different natural statistics",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 393,
+ 507,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 393,
+ 507,
+ 408
+ ],
+ "score": 1.0,
+ "content": "could be enough for Xu et al. [2020] to prohibit the use of DA in current datasets. Furthermore,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 405,
+ 506,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 506,
+ 418
+ ],
+ "score": 1.0,
+ "content": "Raghunathan et al. [2020] obtained a surprising result combining both DA and regularization. In",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 397,
+ 429
+ ],
+ "score": 1.0,
+ "content": "that case, it was found that the minimum norm interpolant on the original",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 398,
+ 417,
+ 422,
+ 427
+ ],
+ "score": 0.28,
+ "content": "+ \\textrm Ḋ \\textmu Ḋ Ḍ Ḍ _ { \\mathrm { Ḋ } } \\textrm Ḋ \\textmu Ḍ Ḍ",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 423,
+ 416,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "dataset could have a",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "larger standard error than the minimum norm interpolant on the original dataset alone, even when",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 506,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 506,
+ 450
+ ],
+ "score": 1.0,
+ "content": "using label-preserving DA. However, this phenomenon only occurs as long as the model remains",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 449,
+ 378,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 324,
+ 461
+ ],
+ "score": 1.0,
+ "content": "over-parametrized, even when considering the original",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 449,
+ 345,
+ 460
+ ],
+ "score": 0.41,
+ "content": "+ \\mathrm { D A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 450,
+ 378,
+ 461
+ ],
+ "score": 1.0,
+ "content": "dataset.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 547
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 506,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 506,
+ 473
+ ],
+ "score": 1.0,
+ "content": "Input dependent DA can reduce sample/class bias. Recall that Section 3.2 brought forward one",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 470,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 470,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "possible explanation on how DA can be the source of training-set by introduce class-imbalance due to",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "score": 1.0,
+ "content": "the same DA being label-preserving for some classes and not for others. From this, a direct solution",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 491,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 491,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "would be to adapt the “strength”. This solution, formalized in Xu et al. [2020], consists in measuring",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 516
+ ],
+ "score": 1.0,
+ "content": "the bias of a model and adapt the DA policy accordingly to correct it. This has been done in different",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 514,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 514,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "flavors e.g. in McLaughlin et al. [2015], Iosifidis and Ntoutsi [2018], Jaipuria et al. [2020]. One",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 104,
+ 523,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 523,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "limitation of this direction is that it requires to estimate a model bias and adapt the DA accordingly",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 303,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 303,
+ 549
+ ],
+ "score": 1.0,
+ "content": "which can be challenging for large scale models.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 46.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 635
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "score": 1.0,
+ "content": "Learned DA e.g. from a GAN can produce even more class-dependent bias. One natural extension",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 106,
+ 558,
+ 506,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 558,
+ 506,
+ 570
+ ],
+ "score": 1.0,
+ "content": "of the hands-on adaptivity of a DA to a measured model bias would be to learn a DA to maximize a",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 567,
+ 505,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 567,
+ 505,
+ 583
+ ],
+ "score": 1.0,
+ "content": "model’s performance, for example. This line of work has led to many learn DA policies e.g. using",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "Generative Adversarial Networks [Hu and Li, 2019]. Yet, it has been shown that learning a DA to",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 505,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 505,
+ 603
+ ],
+ "score": 1.0,
+ "content": "maximize some aggregated measure of performance will produce even more bias in a model [Hu",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 106,
+ 601,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 601,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "and Li, 2019]. In fact, and as per the controlled experiments from Section 2, the learned DA will",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 106,
+ 612,
+ 506,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 612,
+ 506,
+ 626
+ ],
+ "score": 1.0,
+ "content": "entirely disregard a minority of the classes if it means that the produced DA can drastically improve",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 106,
+ 624,
+ 419,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 624,
+ 419,
+ 636
+ ],
+ "score": 1.0,
+ "content": "performances on all others, effectively maximizing the average performances.",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ }
+ ],
+ "index": 54.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 635,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 634,
+ 505,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 634,
+ 505,
+ 647
+ ],
+ "score": 1.0,
+ "content": "Inherent tradeoff between DA and model robustness. In addition to the implication of DA",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 106,
+ 645,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "into bias, there exists an intertwined relationship between DA and model robustness. In fact, even",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ },
+ {
+ "bbox": [
+ 106,
+ 657,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 657,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "assuming the use of perfectly adapted DAs, there exists an inherent tradeoff between accuracy and",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ },
+ {
+ "bbox": [
+ 106,
+ 667,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": "robustness that holds even in the infinite data limit [Tsipras et al., 2018, Fawzi et al., 2018, Zhang et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 692
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 692
+ ],
+ "score": 1.0,
+ "content": "2019]. For example, Min et al. [2021] proved in the robust linear classification regime that (i) more",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "data improves generalization in a weak adversary regime, (ii) more data can improve generalization",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 714
+ ],
+ "score": 1.0,
+ "content": "up to a point where additional data starts to hurt generalization in a medium adversary regime, and",
+ "type": "text"
+ }
+ ],
+ "index": 65
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 474,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 474,
+ 724
+ ],
+ "score": 1.0,
+ "content": "that (iii) more data immediately decreases generalization error in a strong adversary regime.",
+ "type": "text"
+ }
+ ],
+ "index": 66
+ }
+ ],
+ "index": 62.5
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 12,
+ "width": 9
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 72,
+ 325,
+ 216
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 105,
+ 72,
+ 325,
+ 216
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 325,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 325,
+ 216
+ ],
+ "score": 0.963,
+ "type": "image",
+ "image_path": "382a91d715d38ca9b29e90abc548d2f9ac19acd3a0987ee91d1f7d922caba770.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 325,
+ 84.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 84.0,
+ 325,
+ 96.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 96.0,
+ 325,
+ 108.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 108.0,
+ 325,
+ 120.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 120.0,
+ 325,
+ 132.0
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 132.0,
+ 325,
+ 144.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 144.0,
+ 325,
+ 156.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 156.0,
+ 325,
+ 168.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 168.0,
+ 325,
+ 180.0
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 180.0,
+ 325,
+ 192.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 192.0,
+ 325,
+ 204.0
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 204.0,
+ 325,
+ 216.0
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 328,
+ 76,
+ 504,
+ 217
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 326,
+ 76,
+ 506,
+ 89
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 76,
+ 506,
+ 89
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Reprise of Fig. 2 but now implement-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 326,
+ 87,
+ 506,
+ 99
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 87,
+ 506,
+ 99
+ ],
+ "score": 1.0,
+ "content": "ing the class-dependent DA as prescribed in Sec-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 326,
+ 96,
+ 505,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 96,
+ 505,
+ 109
+ ],
+ "score": 1.0,
+ "content": "tion 3.3 i.e. we only apply DA (random crop) to",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 326,
+ 107,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 107,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "the classes that see their per-class accuracy im-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 326,
+ 117,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 117,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "prove when this DA is employed during training..",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 326,
+ 127,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 127,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "We observe here that the bias of the model intro-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 327,
+ 137,
+ 504,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 327,
+ 137,
+ 504,
+ 148
+ ],
+ "score": 1.0,
+ "content": "duced from random crop on the majority of the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 327,
+ 147,
+ 505,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 327,
+ 147,
+ 505,
+ 158
+ ],
+ "score": 1.0,
+ "content": "classes spills-over to the minority of the classes",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 326,
+ 156,
+ 505,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 156,
+ 505,
+ 168
+ ],
+ "score": 1.0,
+ "content": "that do not benefit from that DA, even though",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 326,
+ 167,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 167,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "such classes never received that DA during train-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 326,
+ 177,
+ 504,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 177,
+ 504,
+ 188
+ ],
+ "score": 1.0,
+ "content": "ing. Results are averaged over 5 runs and employ",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 326,
+ 187,
+ 504,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 187,
+ 504,
+ 198
+ ],
+ "score": 1.0,
+ "content": "the official resnet50 implementation trained on",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 326,
+ 196,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 196,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Imagenet with horizontal flip but a per-class ran-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 326,
+ 207,
+ 381,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 207,
+ 381,
+ 218
+ ],
+ "score": 1.0,
+ "content": "dom crop DA.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 18.5
+ }
+ ],
+ "index": 12.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 269,
+ 504,
+ 291
+ ],
+ "lines": [],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 268,
+ 505,
+ 292
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 296,
+ 505,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 506,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 506,
+ 308
+ ],
+ "score": 1.0,
+ "content": "Input independent DA exacerbates the bias already present in a dataset. Especially relevant to",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 304,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 304,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "our results is a recent result of Xu et al. [2020]. In this work, it was theorized that when the underlying",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 318,
+ 506,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 318,
+ 506,
+ 330
+ ],
+ "score": 1.0,
+ "content": "dataset contains inherent biases, training on the original data is more effective than employing an i.i.d.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 328,
+ 506,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 328,
+ 506,
+ 342
+ ],
+ "score": 1.0,
+ "content": "DA policy, i.e. applying the same random augmentation to all samples/classes, to produce an unbiased",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 339,
+ 506,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 339,
+ 506,
+ 352
+ ],
+ "score": 1.0,
+ "content": "model. In short, the DA exacerbates the already present biases and makes the trained model further",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 363
+ ],
+ "score": 1.0,
+ "content": "away from the unbiased optimum. The difficulty of this result lies in defining bias for real images.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "As per our experiments from Section 3, we observe that bias can take many form e.g. one class might",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "naturally represent its object always under the same angle. This is particularly true say for boats",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 384,
+ 506,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 384,
+ 506,
+ 396
+ ],
+ "score": 1.0,
+ "content": "which are rarely captured upside-down. Hence, simply having classes with different natural statistics",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 393,
+ 507,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 393,
+ 507,
+ 408
+ ],
+ "score": 1.0,
+ "content": "could be enough for Xu et al. [2020] to prohibit the use of DA in current datasets. Furthermore,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 405,
+ 506,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 506,
+ 418
+ ],
+ "score": 1.0,
+ "content": "Raghunathan et al. [2020] obtained a surprising result combining both DA and regularization. In",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 397,
+ 429
+ ],
+ "score": 1.0,
+ "content": "that case, it was found that the minimum norm interpolant on the original",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 398,
+ 417,
+ 422,
+ 427
+ ],
+ "score": 0.28,
+ "content": "+ \\textrm Ḋ \\textmu Ḋ Ḍ Ḍ _ { \\mathrm { Ḋ } } \\textrm Ḋ \\textmu Ḍ Ḍ",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 423,
+ 416,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "dataset could have a",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "larger standard error than the minimum norm interpolant on the original dataset alone, even when",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 506,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 506,
+ 450
+ ],
+ "score": 1.0,
+ "content": "using label-preserving DA. However, this phenomenon only occurs as long as the model remains",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 449,
+ 378,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 324,
+ 461
+ ],
+ "score": 1.0,
+ "content": "over-parametrized, even when considering the original",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 449,
+ 345,
+ 460
+ ],
+ "score": 0.41,
+ "content": "+ \\mathrm { D A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 450,
+ 378,
+ 461
+ ],
+ "score": 1.0,
+ "content": "dataset.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 104,
+ 296,
+ 507,
+ 461
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 547
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 506,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 506,
+ 473
+ ],
+ "score": 1.0,
+ "content": "Input dependent DA can reduce sample/class bias. Recall that Section 3.2 brought forward one",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 470,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 470,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "possible explanation on how DA can be the source of training-set by introduce class-imbalance due to",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "score": 1.0,
+ "content": "the same DA being label-preserving for some classes and not for others. From this, a direct solution",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 491,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 491,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "would be to adapt the “strength”. This solution, formalized in Xu et al. [2020], consists in measuring",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 506,
+ 516
+ ],
+ "score": 1.0,
+ "content": "the bias of a model and adapt the DA policy accordingly to correct it. This has been done in different",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 514,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 514,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "flavors e.g. in McLaughlin et al. [2015], Iosifidis and Ntoutsi [2018], Jaipuria et al. [2020]. One",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 104,
+ 523,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 523,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "limitation of this direction is that it requires to estimate a model bias and adapt the DA accordingly",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 303,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 303,
+ 549
+ ],
+ "score": 1.0,
+ "content": "which can be challenging for large scale models.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 46.5,
+ "bbox_fs": [
+ 104,
+ 460,
+ 506,
+ 549
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 635
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 506,
+ 560
+ ],
+ "score": 1.0,
+ "content": "Learned DA e.g. from a GAN can produce even more class-dependent bias. One natural extension",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 106,
+ 558,
+ 506,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 558,
+ 506,
+ 570
+ ],
+ "score": 1.0,
+ "content": "of the hands-on adaptivity of a DA to a measured model bias would be to learn a DA to maximize a",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 567,
+ 505,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 567,
+ 505,
+ 583
+ ],
+ "score": 1.0,
+ "content": "model’s performance, for example. This line of work has led to many learn DA policies e.g. using",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "Generative Adversarial Networks [Hu and Li, 2019]. Yet, it has been shown that learning a DA to",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 505,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 505,
+ 603
+ ],
+ "score": 1.0,
+ "content": "maximize some aggregated measure of performance will produce even more bias in a model [Hu",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 106,
+ 601,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 601,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "and Li, 2019]. In fact, and as per the controlled experiments from Section 2, the learned DA will",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 106,
+ 612,
+ 506,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 612,
+ 506,
+ 626
+ ],
+ "score": 1.0,
+ "content": "entirely disregard a minority of the classes if it means that the produced DA can drastically improve",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 106,
+ 624,
+ 419,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 624,
+ 419,
+ 636
+ ],
+ "score": 1.0,
+ "content": "performances on all others, effectively maximizing the average performances.",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ }
+ ],
+ "index": 54.5,
+ "bbox_fs": [
+ 105,
+ 546,
+ 506,
+ 636
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 635,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 634,
+ 505,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 634,
+ 505,
+ 647
+ ],
+ "score": 1.0,
+ "content": "Inherent tradeoff between DA and model robustness. In addition to the implication of DA",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 106,
+ 645,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "into bias, there exists an intertwined relationship between DA and model robustness. In fact, even",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ },
+ {
+ "bbox": [
+ 106,
+ 657,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 657,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "assuming the use of perfectly adapted DAs, there exists an inherent tradeoff between accuracy and",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ },
+ {
+ "bbox": [
+ 106,
+ 667,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": "robustness that holds even in the infinite data limit [Tsipras et al., 2018, Fawzi et al., 2018, Zhang et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 692
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 692
+ ],
+ "score": 1.0,
+ "content": "2019]. For example, Min et al. [2021] proved in the robust linear classification regime that (i) more",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": "data improves generalization in a weak adversary regime, (ii) more data can improve generalization",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 714
+ ],
+ "score": 1.0,
+ "content": "up to a point where additional data starts to hurt generalization in a medium adversary regime, and",
+ "type": "text"
+ }
+ ],
+ "index": 65
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 474,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 474,
+ 724
+ ],
+ "score": 1.0,
+ "content": "that (iii) more data immediately decreases generalization error in a strong adversary regime.",
+ "type": "text"
+ }
+ ],
+ "index": 66
+ }
+ ],
+ "index": 62.5,
+ "bbox_fs": [
+ 105,
+ 634,
+ 506,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 105,
+ 72,
+ 483,
+ 84
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 69,
+ 484,
+ 88
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 69,
+ 484,
+ 88
+ ],
+ "score": 1.0,
+ "content": "3.4 Class-Dependent Data-Augmentation Seems Insufficient for Performance Recovery",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 97,
+ 505,
+ 153
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 97,
+ 506,
+ 111
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 97,
+ 506,
+ 111
+ ],
+ "score": 1.0,
+ "content": "We observed in Section 2 that DA could lead to disastrous per-class performances on a minority of",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 108,
+ 505,
+ 122
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 108,
+ 505,
+ 122
+ ],
+ "score": 1.0,
+ "content": "classes. We now propose to implement one solution from Section 3.3 that consists in simply not",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 119,
+ 506,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 119,
+ 506,
+ 133
+ ],
+ "score": 1.0,
+ "content": "applying the harmful DA to the classes suffering from it. As will become clear, applying the DA on",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 129,
+ 505,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 129,
+ 505,
+ 145
+ ],
+ "score": 1.0,
+ "content": "all other classes will be enough to skew the training of the model preventing any performing recovery",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 141,
+ 176,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 141,
+ 176,
+ 154
+ ],
+ "score": 1.0,
+ "content": "on those classes.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 158,
+ 505,
+ 420
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 171
+ ],
+ "score": 1.0,
+ "content": "To motivate the need for a better control of the per-class performance of a model, we first present an",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 167,
+ 505,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 365,
+ 183
+ ],
+ "score": 1.0,
+ "content": "illustrative argument. A standard resnet50 on Imagenet reaches",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 366,
+ 169,
+ 398,
+ 180
+ ],
+ "score": 0.89,
+ "content": "7 7 . 1 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 167,
+ 505,
+ 183
+ ],
+ "score": 1.0,
+ "content": "top-1 with a random crop",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 170,
+ 194
+ ],
+ "score": 1.0,
+ "content": "lower bound of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 170,
+ 180,
+ 185,
+ 190
+ ],
+ "score": 0.87,
+ "content": "8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 178,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": ". Using precise cross-validation, one could reach 77.29 by using a random crop",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 189,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 189,
+ 169,
+ 205
+ ],
+ "score": 1.0,
+ "content": "lower bound of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 191,
+ 189,
+ 201
+ ],
+ "score": 0.89,
+ "content": "1 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 189,
+ 189,
+ 231,
+ 205
+ ],
+ "score": 1.0,
+ "content": "instead of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 191,
+ 246,
+ 201
+ ],
+ "score": 0.88,
+ "content": "8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 189,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "on all classes. Yet, and most interestingly, if one were able to get",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 201,
+ 505,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 505,
+ 214
+ ],
+ "score": 1.0,
+ "content": "the best per-class performance —as per varying the lower-bound as in Fig. 2 and picking for each",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 226
+ ],
+ "score": 1.0,
+ "content": "class the best per-class performance of any of the models— one could reach 79.37. Beyond pure",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 224,
+ 505,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 224,
+ 505,
+ 236
+ ],
+ "score": 1.0,
+ "content": "average test performance, controlling the worst-case per-class performance is of crucial importance",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 234,
+ 506,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 506,
+ 247
+ ],
+ "score": 1.0,
+ "content": "for fairness [Du et al., 2020, Veitch et al., 2021]. We thus explore one of the solution that we",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 245,
+ 506,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 245,
+ 506,
+ 258
+ ],
+ "score": 1.0,
+ "content": "reviewed in Section 3.3 that consists in only applying the random crop DA to the classes whose test",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 270
+ ],
+ "score": 1.0,
+ "content": "performances increased with the DA’s level. We obtain in Fig. 7 that such class-specific strategy",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 268,
+ 506,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 506,
+ 280
+ ],
+ "score": 1.0,
+ "content": "is not sufficient to guarantee that the classes negatively impacted by the random crop DA see their",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 278,
+ 506,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 278,
+ 506,
+ 291
+ ],
+ "score": 1.0,
+ "content": "performance to be constant across the DA level applied to all the other classes. This finding is also",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 301
+ ],
+ "score": 1.0,
+ "content": "supported by our weight decay experiment in Figs. 3 and 16 in which the regularization of the DN",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 299,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 299,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "impacted classes differently. In fact, first notice that applying weight decay only when seeing some",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 324
+ ],
+ "score": 1.0,
+ "content": "specific classes would simply amount (on average) to reducing the weight decay hyper-parameter",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "score": 1.0,
+ "content": "proportionally to how many classes are considered to be without regularization. In a similar fashion,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 333,
+ 506,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 333,
+ 506,
+ 344
+ ],
+ "score": 1.0,
+ "content": "DA produces an implicit regularizer [Balestriero et al., 2022] and applying a class-specific DA level",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 356
+ ],
+ "score": 1.0,
+ "content": "reduces the impact of the implicit regularizer. But since the amount of classes for which we do not",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 354,
+ 506,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 355,
+ 450,
+ 366
+ ],
+ "score": 1.0,
+ "content": "apply random crop is quite small compared to the total number of classes (between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 450,
+ 354,
+ 465,
+ 365
+ ],
+ "score": 0.85,
+ "content": "1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 355,
+ 484,
+ 366
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 484,
+ 354,
+ 499,
+ 365
+ ],
+ "score": 0.84,
+ "content": "5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 500,
+ 355,
+ 506,
+ 366
+ ],
+ "score": 1.0,
+ "content": "),",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 365,
+ 505,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 365,
+ 505,
+ 377
+ ],
+ "score": 1.0,
+ "content": "it means that the DA’s implicit regularizer remains nearly the same and thus the model’s bias is",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 376,
+ 507,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 507,
+ 389
+ ],
+ "score": 1.0,
+ "content": "nearly the same regardless if that DA is applied or not onto those classes. This is what we observe,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 387,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "applying the random crop DA to the classes that benefit from it is enough to bias the model and",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 506,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 506,
+ 411
+ ],
+ "score": 1.0,
+ "content": "degrade the performances on some classes at the same pace than when applying the DA to all classes",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 410,
+ 271,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 271,
+ 421
+ ],
+ "score": 1.0,
+ "content": "unconditionally (compare Figs. 2 and 7).",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 425,
+ 505,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "As a result, we observe that no readily and easily implemented solution provides us with a strategy to",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 436,
+ 506,
+ 449
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 436,
+ 506,
+ 449
+ ],
+ "score": 1.0,
+ "content": "prevent deep learning to fall into the scenario depicted on the left of Fig. 1. Those findings however",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 447,
+ 505,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 447,
+ 505,
+ 460
+ ],
+ "score": 1.0,
+ "content": "motivate the search of novel model complexity controls that is fair among classes. From a more",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 459,
+ 505,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 459,
+ 505,
+ 470
+ ],
+ "score": 1.0,
+ "content": "theoretical viewpoint, it might also be possible to better understand if even such a fair per-class model",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 481
+ ],
+ "score": 1.0,
+ "content": "complexity could exist, which is not clear as natural image classes tend to have inherently different",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 146,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 146,
+ 492
+ ],
+ "score": 1.0,
+ "content": "statistics.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 519,
+ 272,
+ 533
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 519,
+ 273,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 519,
+ 273,
+ 535
+ ],
+ "score": 1.0,
+ "content": "4 Conclusions and Limitations",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 552,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "We proposed in this study to understand the impact of regularization, in particular data-augmentation",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 564,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 564,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "and weight decay, into the final performances of a deep network. We obtained that the use of",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 574,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "regularization increases the average test performances at the cost of significant performance drops on",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 598
+ ],
+ "score": 1.0,
+ "content": "some specific classes. By focusing on maximizing aggregate performance statistics we have produced",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 609
+ ],
+ "score": 1.0,
+ "content": "learning mechanisms that can be potentially harmful, especially in transfer learning tasks. In fact,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 608,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 608,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "we have also observed that varying the amount of regularization employed during pre-training of a",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "specific dataset impacts the per-class performances of that pre-trained model on different downstream",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 628,
+ 506,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 506,
+ 642
+ ],
+ "score": 1.0,
+ "content": "tasks e.g. going from Imagenet to INaturalist. Lastly, commonly prescribed solutions e.g. class-",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "dependent data-augmentation do not seem to help indicating that the sole use of an augmentation on",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 104,
+ 650,
+ 506,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 650,
+ 506,
+ 664
+ ],
+ "score": 1.0,
+ "content": "some classes is enough to bias the model on all classes. Hence, there remains a vast research area to",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 662,
+ 505,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 662,
+ 505,
+ 675
+ ],
+ "score": 1.0,
+ "content": "explore in order to turn deep learning model selection from the current regime to a more ideal one",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 673,
+ 199,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 673,
+ 199,
+ 685
+ ],
+ "score": 1.0,
+ "content": "(left to right of Fig. 1).",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 42.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 689,
+ 504,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "The main limitation of this work is its focus on computer vision datasets and models. It is possible",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "that our observation will be further conveyed in other regimes or not, and we leave such analysis for",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 158,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 158,
+ 722
+ ],
+ "score": 1.0,
+ "content": "future work.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "page_idx": 8,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "score": 1.0,
+ "content": "9",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 105,
+ 72,
+ 483,
+ 84
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 69,
+ 484,
+ 88
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 69,
+ 484,
+ 88
+ ],
+ "score": 1.0,
+ "content": "3.4 Class-Dependent Data-Augmentation Seems Insufficient for Performance Recovery",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 97,
+ 505,
+ 153
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 97,
+ 506,
+ 111
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 97,
+ 506,
+ 111
+ ],
+ "score": 1.0,
+ "content": "We observed in Section 2 that DA could lead to disastrous per-class performances on a minority of",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 108,
+ 505,
+ 122
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 108,
+ 505,
+ 122
+ ],
+ "score": 1.0,
+ "content": "classes. We now propose to implement one solution from Section 3.3 that consists in simply not",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 119,
+ 506,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 119,
+ 506,
+ 133
+ ],
+ "score": 1.0,
+ "content": "applying the harmful DA to the classes suffering from it. As will become clear, applying the DA on",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 129,
+ 505,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 129,
+ 505,
+ 145
+ ],
+ "score": 1.0,
+ "content": "all other classes will be enough to skew the training of the model preventing any performing recovery",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 141,
+ 176,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 141,
+ 176,
+ 154
+ ],
+ "score": 1.0,
+ "content": "on those classes.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 105,
+ 97,
+ 506,
+ 154
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 158,
+ 505,
+ 420
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 171
+ ],
+ "score": 1.0,
+ "content": "To motivate the need for a better control of the per-class performance of a model, we first present an",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 167,
+ 505,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 365,
+ 183
+ ],
+ "score": 1.0,
+ "content": "illustrative argument. A standard resnet50 on Imagenet reaches",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 366,
+ 169,
+ 398,
+ 180
+ ],
+ "score": 0.89,
+ "content": "7 7 . 1 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 167,
+ 505,
+ 183
+ ],
+ "score": 1.0,
+ "content": "top-1 with a random crop",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 170,
+ 194
+ ],
+ "score": 1.0,
+ "content": "lower bound of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 170,
+ 180,
+ 185,
+ 190
+ ],
+ "score": 0.87,
+ "content": "8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 178,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": ". Using precise cross-validation, one could reach 77.29 by using a random crop",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 189,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 189,
+ 169,
+ 205
+ ],
+ "score": 1.0,
+ "content": "lower bound of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 191,
+ 189,
+ 201
+ ],
+ "score": 0.89,
+ "content": "1 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 189,
+ 189,
+ 231,
+ 205
+ ],
+ "score": 1.0,
+ "content": "instead of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 191,
+ 246,
+ 201
+ ],
+ "score": 0.88,
+ "content": "8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 189,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "on all classes. Yet, and most interestingly, if one were able to get",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 201,
+ 505,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 505,
+ 214
+ ],
+ "score": 1.0,
+ "content": "the best per-class performance —as per varying the lower-bound as in Fig. 2 and picking for each",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 226
+ ],
+ "score": 1.0,
+ "content": "class the best per-class performance of any of the models— one could reach 79.37. Beyond pure",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 224,
+ 505,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 224,
+ 505,
+ 236
+ ],
+ "score": 1.0,
+ "content": "average test performance, controlling the worst-case per-class performance is of crucial importance",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 234,
+ 506,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 506,
+ 247
+ ],
+ "score": 1.0,
+ "content": "for fairness [Du et al., 2020, Veitch et al., 2021]. We thus explore one of the solution that we",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 245,
+ 506,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 245,
+ 506,
+ 258
+ ],
+ "score": 1.0,
+ "content": "reviewed in Section 3.3 that consists in only applying the random crop DA to the classes whose test",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 270
+ ],
+ "score": 1.0,
+ "content": "performances increased with the DA’s level. We obtain in Fig. 7 that such class-specific strategy",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 268,
+ 506,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 506,
+ 280
+ ],
+ "score": 1.0,
+ "content": "is not sufficient to guarantee that the classes negatively impacted by the random crop DA see their",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 278,
+ 506,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 278,
+ 506,
+ 291
+ ],
+ "score": 1.0,
+ "content": "performance to be constant across the DA level applied to all the other classes. This finding is also",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 301
+ ],
+ "score": 1.0,
+ "content": "supported by our weight decay experiment in Figs. 3 and 16 in which the regularization of the DN",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 299,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 299,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "impacted classes differently. In fact, first notice that applying weight decay only when seeing some",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 324
+ ],
+ "score": 1.0,
+ "content": "specific classes would simply amount (on average) to reducing the weight decay hyper-parameter",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "score": 1.0,
+ "content": "proportionally to how many classes are considered to be without regularization. In a similar fashion,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 333,
+ 506,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 333,
+ 506,
+ 344
+ ],
+ "score": 1.0,
+ "content": "DA produces an implicit regularizer [Balestriero et al., 2022] and applying a class-specific DA level",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 356
+ ],
+ "score": 1.0,
+ "content": "reduces the impact of the implicit regularizer. But since the amount of classes for which we do not",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 354,
+ 506,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 355,
+ 450,
+ 366
+ ],
+ "score": 1.0,
+ "content": "apply random crop is quite small compared to the total number of classes (between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 450,
+ 354,
+ 465,
+ 365
+ ],
+ "score": 0.85,
+ "content": "1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 355,
+ 484,
+ 366
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 484,
+ 354,
+ 499,
+ 365
+ ],
+ "score": 0.84,
+ "content": "5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 500,
+ 355,
+ 506,
+ 366
+ ],
+ "score": 1.0,
+ "content": "),",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 365,
+ 505,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 365,
+ 505,
+ 377
+ ],
+ "score": 1.0,
+ "content": "it means that the DA’s implicit regularizer remains nearly the same and thus the model’s bias is",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 376,
+ 507,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 507,
+ 389
+ ],
+ "score": 1.0,
+ "content": "nearly the same regardless if that DA is applied or not onto those classes. This is what we observe,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 387,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "applying the random crop DA to the classes that benefit from it is enough to bias the model and",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 506,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 506,
+ 411
+ ],
+ "score": 1.0,
+ "content": "degrade the performances on some classes at the same pace than when applying the DA to all classes",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 410,
+ 271,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 271,
+ 421
+ ],
+ "score": 1.0,
+ "content": "unconditionally (compare Figs. 2 and 7).",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 105,
+ 158,
+ 507,
+ 421
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 425,
+ 505,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "As a result, we observe that no readily and easily implemented solution provides us with a strategy to",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 436,
+ 506,
+ 449
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 436,
+ 506,
+ 449
+ ],
+ "score": 1.0,
+ "content": "prevent deep learning to fall into the scenario depicted on the left of Fig. 1. Those findings however",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 447,
+ 505,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 447,
+ 505,
+ 460
+ ],
+ "score": 1.0,
+ "content": "motivate the search of novel model complexity controls that is fair among classes. From a more",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 459,
+ 505,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 459,
+ 505,
+ 470
+ ],
+ "score": 1.0,
+ "content": "theoretical viewpoint, it might also be possible to better understand if even such a fair per-class model",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 481
+ ],
+ "score": 1.0,
+ "content": "complexity could exist, which is not clear as natural image classes tend to have inherently different",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 146,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 146,
+ 492
+ ],
+ "score": 1.0,
+ "content": "statistics.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 105,
+ 423,
+ 506,
+ 492
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 519,
+ 272,
+ 533
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 519,
+ 273,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 519,
+ 273,
+ 535
+ ],
+ "score": 1.0,
+ "content": "4 Conclusions and Limitations",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 552,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "We proposed in this study to understand the impact of regularization, in particular data-augmentation",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 564,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 564,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "and weight decay, into the final performances of a deep network. We obtained that the use of",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 574,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "regularization increases the average test performances at the cost of significant performance drops on",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 598
+ ],
+ "score": 1.0,
+ "content": "some specific classes. By focusing on maximizing aggregate performance statistics we have produced",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 609
+ ],
+ "score": 1.0,
+ "content": "learning mechanisms that can be potentially harmful, especially in transfer learning tasks. In fact,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 608,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 608,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "we have also observed that varying the amount of regularization employed during pre-training of a",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "specific dataset impacts the per-class performances of that pre-trained model on different downstream",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 628,
+ 506,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 506,
+ 642
+ ],
+ "score": 1.0,
+ "content": "tasks e.g. going from Imagenet to INaturalist. Lastly, commonly prescribed solutions e.g. class-",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "dependent data-augmentation do not seem to help indicating that the sole use of an augmentation on",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 104,
+ 650,
+ 506,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 650,
+ 506,
+ 664
+ ],
+ "score": 1.0,
+ "content": "some classes is enough to bias the model on all classes. Hence, there remains a vast research area to",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 662,
+ 505,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 662,
+ 505,
+ 675
+ ],
+ "score": 1.0,
+ "content": "explore in order to turn deep learning model selection from the current regime to a more ideal one",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 673,
+ 199,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 673,
+ 199,
+ 685
+ ],
+ "score": 1.0,
+ "content": "(left to right of Fig. 1).",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 104,
+ 553,
+ 506,
+ 685
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 689,
+ 504,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "The main limitation of this work is its focus on computer vision datasets and models. It is possible",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "that our observation will be further conveyed in other regimes or not, and we leave such analysis for",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 158,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 158,
+ 722
+ ],
+ "score": 1.0,
+ "content": "future work.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 50,
+ "bbox_fs": [
+ 106,
+ 688,
+ 505,
+ 722
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 49,
+ 507,
+ 727
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "score": 1.0,
+ "content": "References",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 87,
+ 506,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 87,
+ 506,
+ 102
+ ],
+ "score": 1.0,
+ "content": "Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 113,
+ 99,
+ 241,
+ 111
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 99,
+ 241,
+ 111
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1907.02893, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 118,
+ 506,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 506,
+ 130
+ ],
+ "score": 1.0,
+ "content": "Randall Balestriero, Ishan Misra, and Yann LeCun. A data-augmentation is worth a thousand samples: Exact",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 114,
+ 127,
+ 479,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 127,
+ 479,
+ 142
+ ],
+ "score": 1.0,
+ "content": "quantification from analytical augmented sample moments. arXiv preprint arXiv:2202.08325, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 145,
+ 487,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 145,
+ 487,
+ 159
+ ],
+ "score": 1.0,
+ "content": "Dimitri P Bertsekas. Constrained optimization and Lagrange multiplier methods. Academic press, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 163,
+ 507,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 163,
+ 507,
+ 177
+ ],
+ "score": 1.0,
+ "content": "Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning, volume 4. Springer,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 114,
+ 174,
+ 140,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 174,
+ 140,
+ 186
+ ],
+ "score": 1.0,
+ "content": "2006.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 193,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 193,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "Léon Bottou. Stochastic gradient descent tricks. In Neural networks: Tricks of the trade, pages 421–436.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 203,
+ 176,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 203,
+ 176,
+ 215
+ ],
+ "score": 1.0,
+ "content": "Springer, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 233
+ ],
+ "score": 1.0,
+ "content": "George EP Box and George C Tiao. Bayesian inference in statistical analysis, volume 40. John Wiley & Sons,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 114,
+ 228,
+ 141,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 228,
+ 141,
+ 244
+ ],
+ "score": 1.0,
+ "content": "2011.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 249,
+ 506,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 249,
+ 506,
+ 261
+ ],
+ "score": 1.0,
+ "content": "Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. Advances in",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 259,
+ 295,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 259,
+ 295,
+ 271
+ ],
+ "score": 1.0,
+ "content": "neural information processing systems, 13, 2000.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 276,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 276,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": "Shuxiao Chen, Edgar Dobriban, and Jane Lee. A group-theoretic framework for data augmentation. Advances in",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 114,
+ 288,
+ 354,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 288,
+ 354,
+ 299
+ ],
+ "score": 1.0,
+ "content": "Neural Information Processing Systems, 33:21321–21333, 2020a.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 305,
+ 505,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 305,
+ 505,
+ 317
+ ],
+ "score": 1.0,
+ "content": "Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 316,
+ 302,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 316,
+ 302,
+ 328
+ ],
+ "score": 1.0,
+ "content": "learning. arXiv preprint arXiv:2003.04297, 2020b.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 346
+ ],
+ "score": 1.0,
+ "content": "Xiaodong Cui, Vaibhava Goel, and Brian Kingsbury. Data augmentation for deep neural network acoustic",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 114,
+ 343,
+ 502,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 343,
+ 502,
+ 357
+ ],
+ "score": 1.0,
+ "content": "modeling. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 23(9):1469–1477, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 374
+ ],
+ "score": 1.0,
+ "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 373,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 373,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 116,
+ 383,
+ 140,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 383,
+ 140,
+ 393
+ ],
+ "score": 1.0,
+ "content": "2009.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 401,
+ 505,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 505,
+ 412
+ ],
+ "score": 1.0,
+ "content": "Mengnan Du, Fan Yang, Na Zou, and Xia Hu. Fairness in deep learning: A computational perspective. IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 411,
+ 260,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 411,
+ 260,
+ 423
+ ],
+ "score": 1.0,
+ "content": "Intelligent Systems, 36(4):25–34, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Analysis of classifiers’ robustness to adversarial perturba-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 439,
+ 293,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 439,
+ 293,
+ 451
+ ],
+ "score": 1.0,
+ "content": "tions. Machine learning, 107(3):481–508, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 469
+ ],
+ "score": 1.0,
+ "content": "Ronald Fisher. Statistical methods and scientific induction. Journal of the Royal Statistical Society: Series B",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 114,
+ 468,
+ 254,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 468,
+ 254,
+ 479
+ ],
+ "score": 1.0,
+ "content": "(Methodological), 17(1):69–78, 1955.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 485,
+ 426,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 426,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 502,
+ 507,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 507,
+ 517
+ ],
+ "score": 1.0,
+ "content": "Marvin HJ Gruber. Improving efficiency by shrinkage: The James-Stein and ridge regression estimators.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 514,
+ 181,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 514,
+ 181,
+ 525
+ ],
+ "score": 1.0,
+ "content": "Routledge, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 532,
+ 506,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 532,
+ 506,
+ 544
+ ],
+ "score": 1.0,
+ "content": "Xifeng Guo, En Zhu, Xinwang Liu, and Jianping Yin. Deep embedded clustering with data augmentation. In",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 113,
+ 542,
+ 370,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 542,
+ 370,
+ 556
+ ],
+ "score": 1.0,
+ "content": "Asian conference on machine learning, pages 550–565. PMLR, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 560,
+ 506,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 506,
+ 572
+ ],
+ "score": 1.0,
+ "content": "Trevor Hastie, Robert Tibshirani, Jerome H Friedman, and Jerome H Friedman. The elements of statistical",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 571,
+ 392,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 571,
+ 392,
+ 583
+ ],
+ "score": 1.0,
+ "content": "learning: data mining, inference, and prediction, volume 2. Springer, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 600
+ ],
+ "score": 1.0,
+ "content": "Xin He, Kaiyong Zhao, and Xiaowen Chu. Automl: A survey of the state-of-the-art. Knowledge-Based Systems,",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 597,
+ 187,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 597,
+ 187,
+ 611
+ ],
+ "score": 1.0,
+ "content": "212:106622, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 630
+ ],
+ "score": 1.0,
+ "content": "Alex Hernández-García and Peter König. Data augmentation instead of explicit regularization. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 627,
+ 209,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 627,
+ 209,
+ 639
+ ],
+ "score": 1.0,
+ "content": "arXiv:1806.03852, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": "Sara Hooker, Aaron Courville, Gregory Clark, Yann Dauphin, and Andrea Frome. What do compressed deep",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 655,
+ 352,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 655,
+ 352,
+ 667
+ ],
+ "score": 1.0,
+ "content": "neural networks forget? arXiv preprint arXiv:1911.05248, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 673,
+ 506,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 673,
+ 506,
+ 685
+ ],
+ "score": 1.0,
+ "content": "Sara Hooker, Nyalleng Moorosi, Gregory Clark, Samy Bengio, and Emily Denton. Characterising bias in",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 114,
+ 682,
+ 339,
+ 696
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 682,
+ 339,
+ 696
+ ],
+ "score": 1.0,
+ "content": "compressed models. arXiv preprint arXiv:2010.03058, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 702,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 702,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Mengxiao Hu and Jinlong Li. Exploring bias in gan-based data augmentation for small samples. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 114,
+ 711,
+ 209,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 711,
+ 209,
+ 724
+ ],
+ "score": 1.0,
+ "content": "arXiv:1905.08495, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 22.5
+ }
+ ],
+ "page_idx": 9,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "10",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 105,
+ 49,
+ 507,
+ 727
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "score": 1.0,
+ "content": "References",
+ "type": "text"
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 87,
+ 506,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 87,
+ 506,
+ 102
+ ],
+ "score": 1.0,
+ "content": "Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 1,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 113,
+ 99,
+ 241,
+ 111
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 99,
+ 241,
+ 111
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1907.02893, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 2,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 118,
+ 506,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 506,
+ 130
+ ],
+ "score": 1.0,
+ "content": "Randall Balestriero, Ishan Misra, and Yann LeCun. A data-augmentation is worth a thousand samples: Exact",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 127,
+ 479,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 127,
+ 479,
+ 142
+ ],
+ "score": 1.0,
+ "content": "quantification from analytical augmented sample moments. arXiv preprint arXiv:2202.08325, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 4,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 145,
+ 487,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 145,
+ 487,
+ 159
+ ],
+ "score": 1.0,
+ "content": "Dimitri P Bertsekas. Constrained optimization and Lagrange multiplier methods. Academic press, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 5,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 163,
+ 507,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 163,
+ 507,
+ 177
+ ],
+ "score": 1.0,
+ "content": "Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning, volume 4. Springer,",
+ "type": "text"
+ }
+ ],
+ "index": 6,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 174,
+ 140,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 174,
+ 140,
+ 186
+ ],
+ "score": 1.0,
+ "content": "2006.",
+ "type": "text"
+ }
+ ],
+ "index": 7,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 193,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 193,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "Léon Bottou. Stochastic gradient descent tricks. In Neural networks: Tricks of the trade, pages 421–436.",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 203,
+ 176,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 203,
+ 176,
+ 215
+ ],
+ "score": 1.0,
+ "content": "Springer, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 233
+ ],
+ "score": 1.0,
+ "content": "George EP Box and George C Tiao. Bayesian inference in statistical analysis, volume 40. John Wiley & Sons,",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 228,
+ 141,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 228,
+ 141,
+ 244
+ ],
+ "score": 1.0,
+ "content": "2011.",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 249,
+ 506,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 249,
+ 506,
+ 261
+ ],
+ "score": 1.0,
+ "content": "Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. Advances in",
+ "type": "text"
+ }
+ ],
+ "index": 12,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 259,
+ 295,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 259,
+ 295,
+ 271
+ ],
+ "score": 1.0,
+ "content": "neural information processing systems, 13, 2000.",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 276,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 276,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": "Shuxiao Chen, Edgar Dobriban, and Jane Lee. A group-theoretic framework for data augmentation. Advances in",
+ "type": "text"
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 288,
+ 354,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 288,
+ 354,
+ 299
+ ],
+ "score": 1.0,
+ "content": "Neural Information Processing Systems, 33:21321–21333, 2020a.",
+ "type": "text"
+ }
+ ],
+ "index": 15,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 305,
+ 505,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 305,
+ 505,
+ 317
+ ],
+ "score": 1.0,
+ "content": "Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive",
+ "type": "text"
+ }
+ ],
+ "index": 16,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 316,
+ 302,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 316,
+ 302,
+ 328
+ ],
+ "score": 1.0,
+ "content": "learning. arXiv preprint arXiv:2003.04297, 2020b.",
+ "type": "text"
+ }
+ ],
+ "index": 17,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 346
+ ],
+ "score": 1.0,
+ "content": "Xiaodong Cui, Vaibhava Goel, and Brian Kingsbury. Data augmentation for deep neural network acoustic",
+ "type": "text"
+ }
+ ],
+ "index": 18,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 343,
+ 502,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 343,
+ 502,
+ 357
+ ],
+ "score": 1.0,
+ "content": "modeling. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 23(9):1469–1477, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 374
+ ],
+ "score": 1.0,
+ "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical",
+ "type": "text"
+ }
+ ],
+ "index": 20,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 373,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 373,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 116,
+ 383,
+ 140,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 383,
+ 140,
+ 393
+ ],
+ "score": 1.0,
+ "content": "2009.",
+ "type": "text"
+ }
+ ],
+ "index": 22,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 401,
+ 505,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 505,
+ 412
+ ],
+ "score": 1.0,
+ "content": "Mengnan Du, Fan Yang, Na Zou, and Xia Hu. Fairness in deep learning: A computational perspective. IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 23,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 411,
+ 260,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 411,
+ 260,
+ 423
+ ],
+ "score": 1.0,
+ "content": "Intelligent Systems, 36(4):25–34, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 24,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Analysis of classifiers’ robustness to adversarial perturba-",
+ "type": "text"
+ }
+ ],
+ "index": 25,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 439,
+ 293,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 439,
+ 293,
+ 451
+ ],
+ "score": 1.0,
+ "content": "tions. Machine learning, 107(3):481–508, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 26,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 506,
+ 469
+ ],
+ "score": 1.0,
+ "content": "Ronald Fisher. Statistical methods and scientific induction. Journal of the Royal Statistical Society: Series B",
+ "type": "text"
+ }
+ ],
+ "index": 27,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 468,
+ 254,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 468,
+ 254,
+ 479
+ ],
+ "score": 1.0,
+ "content": "(Methodological), 17(1):69–78, 1955.",
+ "type": "text"
+ }
+ ],
+ "index": 28,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 485,
+ 426,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 426,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 29,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 502,
+ 507,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 507,
+ 517
+ ],
+ "score": 1.0,
+ "content": "Marvin HJ Gruber. Improving efficiency by shrinkage: The James-Stein and ridge regression estimators.",
+ "type": "text"
+ }
+ ],
+ "index": 30,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 514,
+ 181,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 514,
+ 181,
+ 525
+ ],
+ "score": 1.0,
+ "content": "Routledge, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 31,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 532,
+ 506,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 532,
+ 506,
+ 544
+ ],
+ "score": 1.0,
+ "content": "Xifeng Guo, En Zhu, Xinwang Liu, and Jianping Yin. Deep embedded clustering with data augmentation. In",
+ "type": "text"
+ }
+ ],
+ "index": 32,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 113,
+ 542,
+ 370,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 542,
+ 370,
+ 556
+ ],
+ "score": 1.0,
+ "content": "Asian conference on machine learning, pages 550–565. PMLR, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 560,
+ 506,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 506,
+ 572
+ ],
+ "score": 1.0,
+ "content": "Trevor Hastie, Robert Tibshirani, Jerome H Friedman, and Jerome H Friedman. The elements of statistical",
+ "type": "text"
+ }
+ ],
+ "index": 34,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 571,
+ 392,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 571,
+ 392,
+ 583
+ ],
+ "score": 1.0,
+ "content": "learning: data mining, inference, and prediction, volume 2. Springer, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 35,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 600
+ ],
+ "score": 1.0,
+ "content": "Xin He, Kaiyong Zhao, and Xiaowen Chu. Automl: A survey of the state-of-the-art. Knowledge-Based Systems,",
+ "type": "text"
+ }
+ ],
+ "index": 36,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 597,
+ 187,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 597,
+ 187,
+ 611
+ ],
+ "score": 1.0,
+ "content": "212:106622, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 37,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 630
+ ],
+ "score": 1.0,
+ "content": "Alex Hernández-García and Peter König. Data augmentation instead of explicit regularization. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 38,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 627,
+ 209,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 627,
+ 209,
+ 639
+ ],
+ "score": 1.0,
+ "content": "arXiv:1806.03852, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 39,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": "Sara Hooker, Aaron Courville, Gregory Clark, Yann Dauphin, and Andrea Frome. What do compressed deep",
+ "type": "text"
+ }
+ ],
+ "index": 40,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 655,
+ 352,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 655,
+ 352,
+ 667
+ ],
+ "score": 1.0,
+ "content": "neural networks forget? arXiv preprint arXiv:1911.05248, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 41,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 673,
+ 506,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 673,
+ 506,
+ 685
+ ],
+ "score": 1.0,
+ "content": "Sara Hooker, Nyalleng Moorosi, Gregory Clark, Samy Bengio, and Emily Denton. Characterising bias in",
+ "type": "text"
+ }
+ ],
+ "index": 42,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 682,
+ 339,
+ 696
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 682,
+ 339,
+ 696
+ ],
+ "score": 1.0,
+ "content": "compressed models. arXiv preprint arXiv:2010.03058, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 43,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 702,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 702,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Mengxiao Hu and Jinlong Li. Exploring bias in gan-based data augmentation for small samples. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 44,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 711,
+ 209,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 711,
+ 209,
+ 724
+ ],
+ "score": 1.0,
+ "content": "arXiv:1905.08495, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 45,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 104,
+ 70,
+ 507,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 73,
+ 504,
+ 104
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An efficient",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 83,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 83,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "densenet using learned group convolutions. In Proceedings of the IEEE conference on computer vision and",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 94,
+ 280,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 94,
+ 280,
+ 105
+ ],
+ "score": 1.0,
+ "content": "pattern recognition, pages 2752–2761, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 111,
+ 504,
+ 133
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "Like Hui and Mikhail Belkin. Evaluation of neural architectures trained with square loss vs cross-entropy in",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 116,
+ 122,
+ 335,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 122,
+ 335,
+ 132
+ ],
+ "score": 1.0,
+ "content": "classification tasks. arXiv preprint arXiv:2006.07322, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 140,
+ 504,
+ 161
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 140,
+ 505,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 140,
+ 505,
+ 151
+ ],
+ "score": 1.0,
+ "content": "Maximilian Ilse, Jakub M Tomczak, and Patrick Forré. Selecting data augmentation for simulating interventions.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 149,
+ 420,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 149,
+ 420,
+ 162
+ ],
+ "score": 1.0,
+ "content": "In International Conference on Machine Learning, pages 4555–4562. PMLR, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 168,
+ 504,
+ 189
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 167,
+ 506,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 167,
+ 506,
+ 180
+ ],
+ "score": 1.0,
+ "content": "Vasileios Iosifidis and Eirini Ntoutsi. Dealing with bias via data augmentation in supervised learning scenarios.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 178,
+ 303,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 178,
+ 303,
+ 190
+ ],
+ "score": 1.0,
+ "content": "Jo Bates Paul D. Clough Robert Jäschke, 24, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 196,
+ 505,
+ 228
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 196,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 196,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Nikita Jaipuria, Xianling Zhang, Rohan Bhasin, Mayar Arafa, Punarjay Chakravarty, Shubham Shrivastava, Sagar",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 206,
+ 506,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 206,
+ 506,
+ 219
+ ],
+ "score": 1.0,
+ "content": "Manglani, and Vidya N Murali. Deflating dataset bias using synthetic data augmentation. In Proceedings of",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 217,
+ 502,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 217,
+ 502,
+ 228
+ ],
+ "score": 1.0,
+ "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 772–773, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 234,
+ 504,
+ 256
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 504,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 504,
+ 247
+ ],
+ "score": 1.0,
+ "content": "Michael I Jordan and Tom M Mitchell. Machine learning: Trends, perspectives, and prospects. Science, 349",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 245,
+ 201,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 245,
+ 201,
+ 255
+ ],
+ "score": 1.0,
+ "content": "(6245):255–260, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 262,
+ 506,
+ 314
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 506,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 506,
+ 275
+ ],
+ "score": 1.0,
+ "content": "Marius Kloft, Ulf Brefeld, Pavel Laskov, Klaus-Robert Müller, Alexander Zien, and Sören Sonnen-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 116,
+ 273,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 273,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "burg. Efficient and accurate lp-norm multiple kernel learning. In Y. Bengio, D. Schuurmans, J. Laf-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 284,
+ 506,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 284,
+ 506,
+ 295
+ ],
+ "score": 1.0,
+ "content": "ferty, C. Williams, and A. Culotta, editors, Advances in Neural Information Processing Systems, vol-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 293,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 293,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": "ume 22. Curran Associates, Inc., 2009. URL https://proceedings.neurips.cc/paper/2009/file/",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 302,
+ 320,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 302,
+ 320,
+ 315
+ ],
+ "score": 1.0,
+ "content": "d516b13671a4179d9b7b458a6ebdeb92-Paper.pdf.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 321,
+ 505,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 333
+ ],
+ "score": 1.0,
+ "content": "Ron Kohavi, David H Wolpert, et al. Bias plus variance decomposition for zero-one loss functions. In ICML,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 116,
+ 331,
+ 236,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 331,
+ 236,
+ 343
+ ],
+ "score": 1.0,
+ "content": "volume 96, pages 275–83, 1996.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 349,
+ 504,
+ 371
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 349,
+ 505,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 349,
+ 505,
+ 361
+ ],
+ "score": 1.0,
+ "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 359,
+ 380,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 359,
+ 380,
+ 371
+ ],
+ "score": 1.0,
+ "content": "networks. Advances in neural information processing systems, 25, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 378,
+ 505,
+ 399
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 506,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 506,
+ 390
+ ],
+ "score": 1.0,
+ "content": "Anders Krogh and John Hertz. A simple weight decay can improve generalization. Advances in neural",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 116,
+ 388,
+ 266,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 388,
+ 266,
+ 399
+ ],
+ "score": 1.0,
+ "content": "information processing systems, 4, 1991.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 406,
+ 505,
+ 428
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 506,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 506,
+ 417
+ ],
+ "score": 1.0,
+ "content": "Guillaume Leclerc, Andrew Ilyas, Logan Engstrom, Sung Min Park, Hadi Salman, and Aleksander Madry. ffcv.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 417,
+ 294,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 417,
+ 294,
+ 427
+ ],
+ "score": 1.0,
+ "content": "https://github.com/libffcv/ffcv/, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 434,
+ 505,
+ 456
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 433,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 433,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 115,
+ 445,
+ 350,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 445,
+ 350,
+ 456
+ ],
+ "score": 1.0,
+ "content": "recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 463,
+ 504,
+ 484
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 476
+ ],
+ "score": 1.0,
+ "content": "Daniel LeJeune, Randall Balestriero, Hamid Javadi, and Richard G Baraniuk. Implicit rugosity regularization",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 116,
+ 473,
+ 347,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 473,
+ 347,
+ 484
+ ],
+ "score": 1.0,
+ "content": "via data augmentation. arXiv preprint arXiv:1905.11639, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 491,
+ 505,
+ 523
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 490,
+ 506,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 490,
+ 506,
+ 504
+ ],
+ "score": 1.0,
+ "content": "Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 501,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 501,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 511,
+ 385,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 511,
+ 385,
+ 524
+ ],
+ "score": 1.0,
+ "content": "International Conference on Computer Vision, pages 10012–10022, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 529,
+ 504,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 529,
+ 505,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 529,
+ 505,
+ 542
+ ],
+ "score": 1.0,
+ "content": "Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 116,
+ 540,
+ 315,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 540,
+ 315,
+ 551
+ ],
+ "score": 1.0,
+ "content": "for the 2020s. arXiv preprint arXiv:2201.03545, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 558,
+ 504,
+ 589
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 557,
+ 506,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 557,
+ 506,
+ 569
+ ],
+ "score": 1.0,
+ "content": "Niall McLaughlin, Jesus Martinez Del Rincon, and Paul Miller. Data-augmentation for reducing dataset bias in",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 568,
+ 506,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 568,
+ 506,
+ 580
+ ],
+ "score": 1.0,
+ "content": "person re-identification. In 2015 12th IEEE International conference on advanced video and signal based",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 115,
+ 578,
+ 280,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 578,
+ 280,
+ 590
+ ],
+ "score": 1.0,
+ "content": "surveillance (AVSS), pages 1–6. IEEE, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 596,
+ 504,
+ 618
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 506,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 506,
+ 609
+ ],
+ "score": 1.0,
+ "content": "Yifei Min, Lin Chen, and Amin Karbasi. The curious case of adversarially robust models: More data can help,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 117,
+ 607,
+ 506,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 607,
+ 506,
+ 618
+ ],
+ "score": 1.0,
+ "content": "double descend, or hurt generalization. In Uncertainty in Artificial Intelligence, pages 129–139. PMLR, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 39.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 624,
+ 505,
+ 655
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 625,
+ 505,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 625,
+ 505,
+ 636
+ ],
+ "score": 1.0,
+ "content": "Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 634,
+ 507,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 634,
+ 507,
+ 646
+ ],
+ "score": 1.0,
+ "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707–6717,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 643,
+ 140,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 643,
+ 140,
+ 656
+ ],
+ "score": 1.0,
+ "content": "2020.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 663,
+ 504,
+ 694
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 505,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 505,
+ 675
+ ],
+ "score": 1.0,
+ "content": "Jerzy Neyman and Egon Sharpe Pearson. Ix. on the problem of the most efficient tests of statistical hypotheses.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 116,
+ 673,
+ 505,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 673,
+ 505,
+ 685
+ ],
+ "score": 1.0,
+ "content": "Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 115,
+ 683,
+ 300,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 683,
+ 300,
+ 694
+ ],
+ "score": 1.0,
+ "content": "Physical Character, 231(694-706):289–337, 1933.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 701,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 702,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 702,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. In search of the real inductive bias: On the role of",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 115,
+ 711,
+ 403,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 711,
+ 403,
+ 723
+ ],
+ "score": 1.0,
+ "content": "implicit regularization in deep learning. arXiv preprint arXiv:1412.6614, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 47.5
+ }
+ ],
+ "page_idx": 10,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 755
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 755
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 15,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 73,
+ 504,
+ 104
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An efficient",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 83,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 83,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "densenet using learned group convolutions. In Proceedings of the IEEE conference on computer vision and",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 94,
+ 280,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 94,
+ 280,
+ 105
+ ],
+ "score": 1.0,
+ "content": "pattern recognition, pages 2752–2761, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 106,
+ 73,
+ 506,
+ 105
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 111,
+ 504,
+ 133
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "Like Hui and Mikhail Belkin. Evaluation of neural architectures trained with square loss vs cross-entropy in",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 116,
+ 122,
+ 335,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 122,
+ 335,
+ 132
+ ],
+ "score": 1.0,
+ "content": "classification tasks. arXiv preprint arXiv:2006.07322, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 105,
+ 110,
+ 505,
+ 132
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 140,
+ 504,
+ 161
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 140,
+ 505,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 140,
+ 505,
+ 151
+ ],
+ "score": 1.0,
+ "content": "Maximilian Ilse, Jakub M Tomczak, and Patrick Forré. Selecting data augmentation for simulating interventions.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 149,
+ 420,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 149,
+ 420,
+ 162
+ ],
+ "score": 1.0,
+ "content": "In International Conference on Machine Learning, pages 4555–4562. PMLR, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 106,
+ 140,
+ 505,
+ 162
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 168,
+ 504,
+ 189
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 167,
+ 506,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 167,
+ 506,
+ 180
+ ],
+ "score": 1.0,
+ "content": "Vasileios Iosifidis and Eirini Ntoutsi. Dealing with bias via data augmentation in supervised learning scenarios.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 178,
+ 303,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 178,
+ 303,
+ 190
+ ],
+ "score": 1.0,
+ "content": "Jo Bates Paul D. Clough Robert Jäschke, 24, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 106,
+ 167,
+ 506,
+ 190
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 196,
+ 505,
+ 228
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 196,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 196,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Nikita Jaipuria, Xianling Zhang, Rohan Bhasin, Mayar Arafa, Punarjay Chakravarty, Shubham Shrivastava, Sagar",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 206,
+ 506,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 206,
+ 506,
+ 219
+ ],
+ "score": 1.0,
+ "content": "Manglani, and Vidya N Murali. Deflating dataset bias using synthetic data augmentation. In Proceedings of",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 217,
+ 502,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 217,
+ 502,
+ 228
+ ],
+ "score": 1.0,
+ "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 772–773, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 106,
+ 196,
+ 506,
+ 228
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 234,
+ 504,
+ 256
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 504,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 504,
+ 247
+ ],
+ "score": 1.0,
+ "content": "Michael I Jordan and Tom M Mitchell. Machine learning: Trends, perspectives, and prospects. Science, 349",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 245,
+ 201,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 245,
+ 201,
+ 255
+ ],
+ "score": 1.0,
+ "content": "(6245):255–260, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 105,
+ 234,
+ 504,
+ 255
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 262,
+ 506,
+ 314
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 506,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 506,
+ 275
+ ],
+ "score": 1.0,
+ "content": "Marius Kloft, Ulf Brefeld, Pavel Laskov, Klaus-Robert Müller, Alexander Zien, and Sören Sonnen-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 116,
+ 273,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 273,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "burg. Efficient and accurate lp-norm multiple kernel learning. In Y. Bengio, D. Schuurmans, J. Laf-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 284,
+ 506,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 284,
+ 506,
+ 295
+ ],
+ "score": 1.0,
+ "content": "ferty, C. Williams, and A. Culotta, editors, Advances in Neural Information Processing Systems, vol-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 293,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 293,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": "ume 22. Curran Associates, Inc., 2009. URL https://proceedings.neurips.cc/paper/2009/file/",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 302,
+ 320,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 302,
+ 320,
+ 315
+ ],
+ "score": 1.0,
+ "content": "d516b13671a4179d9b7b458a6ebdeb92-Paper.pdf.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 105,
+ 263,
+ 506,
+ 315
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 321,
+ 505,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 333
+ ],
+ "score": 1.0,
+ "content": "Ron Kohavi, David H Wolpert, et al. Bias plus variance decomposition for zero-one loss functions. In ICML,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 116,
+ 331,
+ 236,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 331,
+ 236,
+ 343
+ ],
+ "score": 1.0,
+ "content": "volume 96, pages 275–83, 1996.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19.5,
+ "bbox_fs": [
+ 105,
+ 321,
+ 506,
+ 343
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 349,
+ 504,
+ 371
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 349,
+ 505,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 349,
+ 505,
+ 361
+ ],
+ "score": 1.0,
+ "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 359,
+ 380,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 359,
+ 380,
+ 371
+ ],
+ "score": 1.0,
+ "content": "networks. Advances in neural information processing systems, 25, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21.5,
+ "bbox_fs": [
+ 106,
+ 349,
+ 505,
+ 371
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 378,
+ 505,
+ 399
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 506,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 506,
+ 390
+ ],
+ "score": 1.0,
+ "content": "Anders Krogh and John Hertz. A simple weight decay can improve generalization. Advances in neural",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 116,
+ 388,
+ 266,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 388,
+ 266,
+ 399
+ ],
+ "score": 1.0,
+ "content": "information processing systems, 4, 1991.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 105,
+ 376,
+ 506,
+ 399
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 406,
+ 505,
+ 428
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 506,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 506,
+ 417
+ ],
+ "score": 1.0,
+ "content": "Guillaume Leclerc, Andrew Ilyas, Logan Engstrom, Sung Min Park, Hadi Salman, and Aleksander Madry. ffcv.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 417,
+ 294,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 417,
+ 294,
+ 427
+ ],
+ "score": 1.0,
+ "content": "https://github.com/libffcv/ffcv/, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 106,
+ 405,
+ 506,
+ 427
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 434,
+ 505,
+ 456
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 433,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 433,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 115,
+ 445,
+ 350,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 445,
+ 350,
+ 456
+ ],
+ "score": 1.0,
+ "content": "recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 106,
+ 433,
+ 505,
+ 456
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 463,
+ 504,
+ 484
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 476
+ ],
+ "score": 1.0,
+ "content": "Daniel LeJeune, Randall Balestriero, Hamid Javadi, and Richard G Baraniuk. Implicit rugosity regularization",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 116,
+ 473,
+ 347,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 473,
+ 347,
+ 484
+ ],
+ "score": 1.0,
+ "content": "via data augmentation. arXiv preprint arXiv:1905.11639, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 105,
+ 462,
+ 505,
+ 484
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 491,
+ 505,
+ 523
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 490,
+ 506,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 490,
+ 506,
+ 504
+ ],
+ "score": 1.0,
+ "content": "Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 501,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 501,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 511,
+ 385,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 511,
+ 385,
+ 524
+ ],
+ "score": 1.0,
+ "content": "International Conference on Computer Vision, pages 10012–10022, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32,
+ "bbox_fs": [
+ 104,
+ 490,
+ 506,
+ 524
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 529,
+ 504,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 529,
+ 505,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 529,
+ 505,
+ 542
+ ],
+ "score": 1.0,
+ "content": "Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 116,
+ 540,
+ 315,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 540,
+ 315,
+ 551
+ ],
+ "score": 1.0,
+ "content": "for the 2020s. arXiv preprint arXiv:2201.03545, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34.5,
+ "bbox_fs": [
+ 105,
+ 529,
+ 505,
+ 551
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 558,
+ 504,
+ 589
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 557,
+ 506,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 557,
+ 506,
+ 569
+ ],
+ "score": 1.0,
+ "content": "Niall McLaughlin, Jesus Martinez Del Rincon, and Paul Miller. Data-augmentation for reducing dataset bias in",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 568,
+ 506,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 568,
+ 506,
+ 580
+ ],
+ "score": 1.0,
+ "content": "person re-identification. In 2015 12th IEEE International conference on advanced video and signal based",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 115,
+ 578,
+ 280,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 578,
+ 280,
+ 590
+ ],
+ "score": 1.0,
+ "content": "surveillance (AVSS), pages 1–6. IEEE, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 105,
+ 557,
+ 506,
+ 590
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 596,
+ 504,
+ 618
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 506,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 506,
+ 609
+ ],
+ "score": 1.0,
+ "content": "Yifei Min, Lin Chen, and Amin Karbasi. The curious case of adversarially robust models: More data can help,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 117,
+ 607,
+ 506,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 607,
+ 506,
+ 618
+ ],
+ "score": 1.0,
+ "content": "double descend, or hurt generalization. In Uncertainty in Artificial Intelligence, pages 129–139. PMLR, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 39.5,
+ "bbox_fs": [
+ 106,
+ 595,
+ 506,
+ 618
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 624,
+ 505,
+ 655
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 625,
+ 505,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 625,
+ 505,
+ 636
+ ],
+ "score": 1.0,
+ "content": "Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 634,
+ 507,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 634,
+ 507,
+ 646
+ ],
+ "score": 1.0,
+ "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707–6717,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 643,
+ 140,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 643,
+ 140,
+ 656
+ ],
+ "score": 1.0,
+ "content": "2020.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 106,
+ 625,
+ 507,
+ 656
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 663,
+ 504,
+ 694
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 505,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 505,
+ 675
+ ],
+ "score": 1.0,
+ "content": "Jerzy Neyman and Egon Sharpe Pearson. Ix. on the problem of the most efficient tests of statistical hypotheses.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 116,
+ 673,
+ 505,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 673,
+ 505,
+ 685
+ ],
+ "score": 1.0,
+ "content": "Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 115,
+ 683,
+ 300,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 683,
+ 300,
+ 694
+ ],
+ "score": 1.0,
+ "content": "Physical Character, 231(694-706):289–337, 1933.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45,
+ "bbox_fs": [
+ 105,
+ 663,
+ 505,
+ 694
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 701,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 702,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 702,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. In search of the real inductive bias: On the role of",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 115,
+ 711,
+ 403,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 711,
+ 403,
+ 723
+ ],
+ "score": 1.0,
+ "content": "implicit regularization in deep learning. arXiv preprint arXiv:1412.6614, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 47.5,
+ "bbox_fs": [
+ 107,
+ 702,
+ 505,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 38,
+ 507,
+ 730
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 87
+ ],
+ "score": 1.0,
+ "content": "Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 114,
+ 83,
+ 259,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 83,
+ 259,
+ 95
+ ],
+ "score": 1.0,
+ "content": "engineering, 22(10):1345–1359, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 100,
+ 507,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 507,
+ 116
+ ],
+ "score": 1.0,
+ "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 114,
+ 111,
+ 506,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 111,
+ 506,
+ 126
+ ],
+ "score": 1.0,
+ "content": "Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 116,
+ 122,
+ 403,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 122,
+ 403,
+ 135
+ ],
+ "score": 1.0,
+ "content": "learning library. Advances in neural information processing systems, 32, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 154
+ ],
+ "score": 1.0,
+ "content": "Ekachai Phaisangittisagul. An analysis of the regularization between l2 and dropout in single hidden layer neural",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 151,
+ 506,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 151,
+ 506,
+ 163
+ ],
+ "score": 1.0,
+ "content": "network. In 2016 7th International Conference on Intelligent Systems, Modelling and Simulation (ISMS),",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 114,
+ 160,
+ 222,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 160,
+ 222,
+ 173
+ ],
+ "score": 1.0,
+ "content": "pages 174–179. IEEE, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 506,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 506,
+ 194
+ ],
+ "score": 1.0,
+ "content": "Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Understanding and mitigating",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 114,
+ 189,
+ 430,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 189,
+ 430,
+ 203
+ ],
+ "score": 1.0,
+ "content": "the tradeoff between robustness and accuracy. arXiv preprint arXiv:2002.10716, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 209,
+ 506,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 209,
+ 506,
+ 221
+ ],
+ "score": 1.0,
+ "content": "Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 114,
+ 218,
+ 214,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 218,
+ 214,
+ 232
+ ],
+ "score": 1.0,
+ "content": "big data, 6(1):1–48, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 104,
+ 236,
+ 507,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 236,
+ 507,
+ 252
+ ],
+ "score": 1.0,
+ "content": "Patrice Simard, Bernard Victorri, Yann LeCun, and John Denker. Tangent prop-a formalism for specifying",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 114,
+ 247,
+ 498,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 247,
+ 498,
+ 261
+ ],
+ "score": 1.0,
+ "content": "selected invariances in an adaptive network. Advances in neural information processing systems, 4, 1991.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 265,
+ 506,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 265,
+ 506,
+ 280
+ ],
+ "score": 1.0,
+ "content": "Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 114,
+ 277,
+ 316,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 277,
+ 316,
+ 289
+ ],
+ "score": 1.0,
+ "content": "Machine Learning, pages 10096–10106. PMLR, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 104,
+ 294,
+ 507,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 294,
+ 507,
+ 309
+ ],
+ "score": 1.0,
+ "content": "Luke Taylor and Geoff Nitschke. Improving deep learning with generic data augmentation. In 2018 IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 306,
+ 438,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 306,
+ 438,
+ 318
+ ],
+ "score": 1.0,
+ "content": "Symposium Series on Computational Intelligence (SSCI), pages 1542–1547. IEEE, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 325,
+ 506,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 506,
+ 337
+ ],
+ "score": 1.0,
+ "content": "Andrei Nikolajevits Tihonov. Solution of incorrectly formulated problems and the regularization method. Soviet",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 114,
+ 336,
+ 216,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 336,
+ 216,
+ 345
+ ],
+ "score": 1.0,
+ "content": "Math., 4:1035–1038, 1963.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 354,
+ 506,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 354,
+ 506,
+ 366
+ ],
+ "score": 1.0,
+ "content": "Andrey Nikolayevich Tikhonov. On the stability of inverse problems. In Dokl. Akad. Nauk SSSR, volume 39,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 114,
+ 363,
+ 199,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 363,
+ 199,
+ 376
+ ],
+ "score": 1.0,
+ "content": "pages 195–198, 1943.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 380,
+ 506,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 506,
+ 396
+ ],
+ "score": 1.0,
+ "content": "Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 115,
+ 393,
+ 374,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 393,
+ 374,
+ 405
+ ],
+ "score": 1.0,
+ "content": "may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 409,
+ 507,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 409,
+ 507,
+ 425
+ ],
+ "score": 1.0,
+ "content": "Giorgio Valentini and Thomas G Dietterich. Bias-variance analysis of support vector machines for the develop-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 115,
+ 421,
+ 485,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 421,
+ 485,
+ 433
+ ],
+ "score": 1.0,
+ "content": "ment of svm-based ensemble methods. Journal of Machine Learning Research, 5(Jul):725–775, 2004.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 104,
+ 439,
+ 507,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 439,
+ 507,
+ 454
+ ],
+ "score": 1.0,
+ "content": "Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 114,
+ 449,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 449,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 114,
+ 459,
+ 409,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 459,
+ 409,
+ 473
+ ],
+ "score": 1.0,
+ "content": "conference on computer vision and pattern recognition, pages 8769–8778, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 493
+ ],
+ "score": 1.0,
+ "content": "Vladimir N Vapnik and A Ya Chervonenkis. On the uniform convergence of relative frequencies of events to",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 114,
+ 488,
+ 392,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 488,
+ 392,
+ 502
+ ],
+ "score": 1.0,
+ "content": "their probabilities. In Measures of complexity, pages 11–30. Springer, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 507,
+ 507,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 507,
+ 521
+ ],
+ "score": 1.0,
+ "content": "VN Vapnik and A Ya Chervonenkis. The method of ordered risk minimization, i. Avtomatika i Telemekhanika,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 519,
+ 173,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 519,
+ 173,
+ 529
+ ],
+ "score": 1.0,
+ "content": "8:21–30, 1974.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 107,
+ 537,
+ 505,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 537,
+ 505,
+ 549
+ ],
+ "score": 1.0,
+ "content": "Victor Veitch, Alexander D’Amour, Steve Yadlowsky, and Jacob Eisenstein. Counterfactual invariance to spurious",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 115,
+ 547,
+ 506,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 547,
+ 506,
+ 559
+ ],
+ "score": 1.0,
+ "content": "correlations in text classification. Advances in Neural Information Processing Systems, 34:16196–16208,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 556,
+ 140,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 556,
+ 140,
+ 569
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 588
+ ],
+ "score": 1.0,
+ "content": "Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 114,
+ 585,
+ 506,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 585,
+ 506,
+ 600
+ ],
+ "score": 1.0,
+ "content": "Bengio. Manifold mixup: Better representations by interpolating hidden states. In International Conference",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 114,
+ 595,
+ 319,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 595,
+ 319,
+ 609
+ ],
+ "score": 1.0,
+ "content": "on Machine Learning, pages 6438–6447. PMLR, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 628
+ ],
+ "score": 1.0,
+ "content": "Ulrike Von Luxburg and Bernhard Schölkopf. Statistical learning theory: Models, concepts, and results. In",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 114,
+ 625,
+ 399,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 625,
+ 399,
+ 637
+ ],
+ "score": 1.0,
+ "content": "Handbook of the History of Logic, volume 10, pages 651–706. Elsevier, 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 642,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "Bernard L Welch. The generalization of ‘student’s’problem when several different population varlances are",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 654,
+ 277,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 654,
+ 277,
+ 666
+ ],
+ "score": 1.0,
+ "content": "involved. Biometrika, 34(1-2):28–35, 1947.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 672,
+ 506,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 672,
+ 506,
+ 685
+ ],
+ "score": 1.0,
+ "content": "Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 114,
+ 682,
+ 466,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 682,
+ 466,
+ 695
+ ],
+ "score": 1.0,
+ "content": "consistency training. Advances in Neural Information Processing Systems, 33:6256–6268, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 507,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 507,
+ 715
+ ],
+ "score": 1.0,
+ "content": "Yi Xu, Asaf Noy, Ming Lin, Qi Qian, Hao Li, and Rong Jin. Wemix: How to better utilize data augmentation.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 114,
+ 711,
+ 264,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 711,
+ 264,
+ 724
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2010.01267, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "page_idx": 11,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "12",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 105,
+ 38,
+ 507,
+ 730
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 87
+ ],
+ "score": 1.0,
+ "content": "Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data",
+ "type": "text"
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 83,
+ 259,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 83,
+ 259,
+ 95
+ ],
+ "score": 1.0,
+ "content": "engineering, 22(10):1345–1359, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 1,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 100,
+ 507,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 507,
+ 116
+ ],
+ "score": 1.0,
+ "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen,",
+ "type": "text"
+ }
+ ],
+ "index": 2,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 111,
+ 506,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 111,
+ 506,
+ 126
+ ],
+ "score": 1.0,
+ "content": "Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 116,
+ 122,
+ 403,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 122,
+ 403,
+ 135
+ ],
+ "score": 1.0,
+ "content": "learning library. Advances in neural information processing systems, 32, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 4,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 139,
+ 505,
+ 154
+ ],
+ "score": 1.0,
+ "content": "Ekachai Phaisangittisagul. An analysis of the regularization between l2 and dropout in single hidden layer neural",
+ "type": "text"
+ }
+ ],
+ "index": 5,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 151,
+ 506,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 151,
+ 506,
+ 163
+ ],
+ "score": 1.0,
+ "content": "network. In 2016 7th International Conference on Intelligent Systems, Modelling and Simulation (ISMS),",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 114,
+ 160,
+ 222,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 160,
+ 222,
+ 173
+ ],
+ "score": 1.0,
+ "content": "pages 174–179. IEEE, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 7,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 506,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 506,
+ 194
+ ],
+ "score": 1.0,
+ "content": "Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Understanding and mitigating",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 189,
+ 430,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 189,
+ 430,
+ 203
+ ],
+ "score": 1.0,
+ "content": "the tradeoff between robustness and accuracy. arXiv preprint arXiv:2002.10716, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 209,
+ 506,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 209,
+ 506,
+ 221
+ ],
+ "score": 1.0,
+ "content": "Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 218,
+ 214,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 218,
+ 214,
+ 232
+ ],
+ "score": 1.0,
+ "content": "big data, 6(1):1–48, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 236,
+ 507,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 236,
+ 507,
+ 252
+ ],
+ "score": 1.0,
+ "content": "Patrice Simard, Bernard Victorri, Yann LeCun, and John Denker. Tangent prop-a formalism for specifying",
+ "type": "text"
+ }
+ ],
+ "index": 12,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 247,
+ 498,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 247,
+ 498,
+ 261
+ ],
+ "score": 1.0,
+ "content": "selected invariances in an adaptive network. Advances in neural information processing systems, 4, 1991.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 265,
+ 506,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 265,
+ 506,
+ 280
+ ],
+ "score": 1.0,
+ "content": "Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 277,
+ 316,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 277,
+ 316,
+ 289
+ ],
+ "score": 1.0,
+ "content": "Machine Learning, pages 10096–10106. PMLR, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 15,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 294,
+ 507,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 294,
+ 507,
+ 309
+ ],
+ "score": 1.0,
+ "content": "Luke Taylor and Geoff Nitschke. Improving deep learning with generic data augmentation. In 2018 IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 16,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 306,
+ 438,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 306,
+ 438,
+ 318
+ ],
+ "score": 1.0,
+ "content": "Symposium Series on Computational Intelligence (SSCI), pages 1542–1547. IEEE, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 17,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 325,
+ 506,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 506,
+ 337
+ ],
+ "score": 1.0,
+ "content": "Andrei Nikolajevits Tihonov. Solution of incorrectly formulated problems and the regularization method. Soviet",
+ "type": "text"
+ }
+ ],
+ "index": 18,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 336,
+ 216,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 336,
+ 216,
+ 345
+ ],
+ "score": 1.0,
+ "content": "Math., 4:1035–1038, 1963.",
+ "type": "text"
+ }
+ ],
+ "index": 19,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 354,
+ 506,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 354,
+ 506,
+ 366
+ ],
+ "score": 1.0,
+ "content": "Andrey Nikolayevich Tikhonov. On the stability of inverse problems. In Dokl. Akad. Nauk SSSR, volume 39,",
+ "type": "text"
+ }
+ ],
+ "index": 20,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 363,
+ 199,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 363,
+ 199,
+ 376
+ ],
+ "score": 1.0,
+ "content": "pages 195–198, 1943.",
+ "type": "text"
+ }
+ ],
+ "index": 21,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 380,
+ 506,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 506,
+ 396
+ ],
+ "score": 1.0,
+ "content": "Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness",
+ "type": "text"
+ }
+ ],
+ "index": 22,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 393,
+ 374,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 393,
+ 374,
+ 405
+ ],
+ "score": 1.0,
+ "content": "may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 23,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 409,
+ 507,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 409,
+ 507,
+ 425
+ ],
+ "score": 1.0,
+ "content": "Giorgio Valentini and Thomas G Dietterich. Bias-variance analysis of support vector machines for the develop-",
+ "type": "text"
+ }
+ ],
+ "index": 24,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 421,
+ 485,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 421,
+ 485,
+ 433
+ ],
+ "score": 1.0,
+ "content": "ment of svm-based ensemble methods. Journal of Machine Learning Research, 5(Jul):725–775, 2004.",
+ "type": "text"
+ }
+ ],
+ "index": 25,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 439,
+ 507,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 439,
+ 507,
+ 454
+ ],
+ "score": 1.0,
+ "content": "Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona,",
+ "type": "text"
+ }
+ ],
+ "index": 26,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 449,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 449,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 114,
+ 459,
+ 409,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 459,
+ 409,
+ 473
+ ],
+ "score": 1.0,
+ "content": "conference on computer vision and pattern recognition, pages 8769–8778, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 28,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 493
+ ],
+ "score": 1.0,
+ "content": "Vladimir N Vapnik and A Ya Chervonenkis. On the uniform convergence of relative frequencies of events to",
+ "type": "text"
+ }
+ ],
+ "index": 29,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 488,
+ 392,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 488,
+ 392,
+ 502
+ ],
+ "score": 1.0,
+ "content": "their probabilities. In Measures of complexity, pages 11–30. Springer, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 30,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 507,
+ 507,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 507,
+ 521
+ ],
+ "score": 1.0,
+ "content": "VN Vapnik and A Ya Chervonenkis. The method of ordered risk minimization, i. Avtomatika i Telemekhanika,",
+ "type": "text"
+ }
+ ],
+ "index": 31,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 519,
+ 173,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 519,
+ 173,
+ 529
+ ],
+ "score": 1.0,
+ "content": "8:21–30, 1974.",
+ "type": "text"
+ }
+ ],
+ "index": 32,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 107,
+ 537,
+ 505,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 537,
+ 505,
+ 549
+ ],
+ "score": 1.0,
+ "content": "Victor Veitch, Alexander D’Amour, Steve Yadlowsky, and Jacob Eisenstein. Counterfactual invariance to spurious",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 547,
+ 506,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 547,
+ 506,
+ 559
+ ],
+ "score": 1.0,
+ "content": "correlations in text classification. Advances in Neural Information Processing Systems, 34:16196–16208,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 556,
+ 140,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 556,
+ 140,
+ 569
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 35,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 588
+ ],
+ "score": 1.0,
+ "content": "Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua",
+ "type": "text"
+ }
+ ],
+ "index": 36,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 585,
+ 506,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 585,
+ 506,
+ 600
+ ],
+ "score": 1.0,
+ "content": "Bengio. Manifold mixup: Better representations by interpolating hidden states. In International Conference",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 114,
+ 595,
+ 319,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 595,
+ 319,
+ 609
+ ],
+ "score": 1.0,
+ "content": "on Machine Learning, pages 6438–6447. PMLR, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 38,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 506,
+ 628
+ ],
+ "score": 1.0,
+ "content": "Ulrike Von Luxburg and Bernhard Schölkopf. Statistical learning theory: Models, concepts, and results. In",
+ "type": "text"
+ }
+ ],
+ "index": 39,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 625,
+ 399,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 625,
+ 399,
+ 637
+ ],
+ "score": 1.0,
+ "content": "Handbook of the History of Logic, volume 10, pages 651–706. Elsevier, 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 40,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 642,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "Bernard L Welch. The generalization of ‘student’s’problem when several different population varlances are",
+ "type": "text"
+ }
+ ],
+ "index": 41,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 654,
+ 277,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 654,
+ 277,
+ 666
+ ],
+ "score": 1.0,
+ "content": "involved. Biometrika, 34(1-2):28–35, 1947.",
+ "type": "text"
+ }
+ ],
+ "index": 42,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 672,
+ 506,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 672,
+ 506,
+ 685
+ ],
+ "score": 1.0,
+ "content": "Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for",
+ "type": "text"
+ }
+ ],
+ "index": 43,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 682,
+ 466,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 682,
+ 466,
+ 695
+ ],
+ "score": 1.0,
+ "content": "consistency training. Advances in Neural Information Processing Systems, 33:6256–6268, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 44,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 507,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 507,
+ 715
+ ],
+ "score": 1.0,
+ "content": "Yi Xu, Asaf Noy, Ming Lin, Qi Qian, Hao Li, and Rong Jin. Wemix: How to better utilize data augmentation.",
+ "type": "text"
+ }
+ ],
+ "index": 45,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 711,
+ 264,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 711,
+ 264,
+ 724
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:2010.01267, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 46,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 104,
+ 72,
+ 507,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 73,
+ 504,
+ 104
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 86
+ ],
+ "score": 1.0,
+ "content": "Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix:",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 83,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 83,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 93,
+ 370,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 93,
+ 370,
+ 105
+ ],
+ "score": 1.0,
+ "content": "international conference on computer vision, pages 6023–6032, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 503,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 123
+ ],
+ "score": 1.0,
+ "content": "Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 120,
+ 506,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 120,
+ 506,
+ 133
+ ],
+ "score": 1.0,
+ "content": "principled trade-off between robustness and accuracy. In International conference on machine learning, pages",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 130,
+ 213,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 130,
+ 213,
+ 141
+ ],
+ "score": 1.0,
+ "content": "7472–7482. PMLR, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 147,
+ 504,
+ 169
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 506,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 506,
+ 160
+ ],
+ "score": 1.0,
+ "content": "Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 115,
+ 158,
+ 316,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 158,
+ 316,
+ 169
+ ],
+ "score": 1.0,
+ "content": "minimization. arXiv preprint arXiv:1710.09412, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 188,
+ 156,
+ 201
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 158,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 158,
+ 203
+ ],
+ "score": 1.0,
+ "content": "Checklist",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 209,
+ 208,
+ 220
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 208,
+ 210,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 208,
+ 210,
+ 222
+ ],
+ "score": 1.0,
+ "content": "1. For all authors...",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 224,
+ 505,
+ 328
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 223,
+ 505,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 223,
+ 505,
+ 236
+ ],
+ "score": 1.0,
+ "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 162,
+ 235,
+ 288,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 235,
+ 288,
+ 247
+ ],
+ "score": 1.0,
+ "content": "contributions and scope? [Yes]",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 144,
+ 247,
+ 506,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 247,
+ 506,
+ 261
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe the limitations of your work? [N/A] We do not propose a novel",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 162,
+ 258,
+ 312,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 258,
+ 312,
+ 270
+ ],
+ "score": 1.0,
+ "content": "method that requires such discussion",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 146,
+ 270,
+ 506,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 270,
+ 506,
+ 283
+ ],
+ "score": 1.0,
+ "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes] We",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 162,
+ 282,
+ 505,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 282,
+ 505,
+ 294
+ ],
+ "score": 1.0,
+ "content": "carefully described why the observations we have made can be dangerous for real",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 162,
+ 294,
+ 237,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 294,
+ 237,
+ 304
+ ],
+ "score": 1.0,
+ "content": "world applications",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 145,
+ 305,
+ 505,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 305,
+ 505,
+ 318
+ ],
+ "score": 1.0,
+ "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 161,
+ 316,
+ 214,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 316,
+ 214,
+ 329
+ ],
+ "score": 1.0,
+ "content": "them? [Yes]",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 331,
+ 302,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 330,
+ 304,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 330,
+ 304,
+ 344
+ ],
+ "score": 1.0,
+ "content": "2. If you are including theoretical results...",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 345,
+ 505,
+ 402
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 345,
+ 505,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 345,
+ 505,
+ 358
+ ],
+ "score": 1.0,
+ "content": "(a) Did you state the full set of assumptions of all theoretical results? [Yes] Our only",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 161,
+ 355,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 355,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "“theoretical result” consists in a formal statistical test for which we precisely describe",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 162,
+ 368,
+ 265,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 368,
+ 265,
+ 379
+ ],
+ "score": 1.0,
+ "content": "our settings and statistics",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 146,
+ 379,
+ 506,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 379,
+ 506,
+ 393
+ ],
+ "score": 1.0,
+ "content": "(b) Did you include complete proofs of all theoretical results? [N/A] No theoretical result",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 161,
+ 392,
+ 285,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 392,
+ 285,
+ 403
+ ],
+ "score": 1.0,
+ "content": "requiring proofs was provided",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 406,
+ 241,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 403,
+ 243,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 403,
+ 243,
+ 419
+ ],
+ "score": 1.0,
+ "content": "3. If you ran experiments...",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 145,
+ 420,
+ 505,
+ 589
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 147,
+ 419,
+ 506,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 419,
+ 506,
+ 433
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the code, data, and instructions needed to reproduce the main exper-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 161,
+ 431,
+ 506,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 431,
+ 506,
+ 443
+ ],
+ "score": 1.0,
+ "content": "imental results (either in the supplemental material or as a URL)? [Yes] We include",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 162,
+ 442,
+ 505,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 442,
+ 505,
+ 454
+ ],
+ "score": 1.0,
+ "content": "summary statistics in the supplementary material which is enough to validate our",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 162,
+ 453,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 453,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": "claims. The full codebase and all the saved models will be released upon completion",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 162,
+ 464,
+ 469,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 464,
+ 469,
+ 476
+ ],
+ "score": 1.0,
+ "content": "of the review process (this includes almost a thousand pre-trained resnet50s)",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 146,
+ 476,
+ 505,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 476,
+ 505,
+ 490
+ ],
+ "score": 1.0,
+ "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 161,
+ 486,
+ 506,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 486,
+ 506,
+ 501
+ ],
+ "score": 1.0,
+ "content": "were chosen)? [Yes] We specify within each caption the key experimental setups,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 161,
+ 498,
+ 383,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 498,
+ 383,
+ 511
+ ],
+ "score": 1.0,
+ "content": "additional training details are provided in the appendix",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 146,
+ 510,
+ 507,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 510,
+ 507,
+ 525
+ ],
+ "score": 1.0,
+ "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 162,
+ 522,
+ 506,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 522,
+ 506,
+ 534
+ ],
+ "score": 1.0,
+ "content": "ments multiple times)? [Yes] we averaged all of our runs over a significant amount of",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 162,
+ 533,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 533,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "realizations (20) and perform a formal statistical test on the significance of our results",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 146,
+ 545,
+ 506,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 545,
+ 506,
+ 559
+ ],
+ "score": 1.0,
+ "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 161,
+ 555,
+ 506,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 555,
+ 506,
+ 569
+ ],
+ "score": 1.0,
+ "content": "of GPUs, internal cluster, or cloud provider)? [Yes] We provide this in the beginning of",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 162,
+ 568,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 568,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "the appendix. Note that we will be providing the saved models and summary statistics",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 161,
+ 577,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 577,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "to make the results easily reproducible even without GPU or computation ressources.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 128,
+ 593,
+ 504,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 591,
+ 507,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 591,
+ 507,
+ 606
+ ],
+ "score": 1.0,
+ "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 608,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 607,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 607,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] For Imagenet and",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 162,
+ 618,
+ 208,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 618,
+ 208,
+ 630
+ ],
+ "score": 1.0,
+ "content": "INaturalist",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 145,
+ 630,
+ 506,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 630,
+ 506,
+ 644
+ ],
+ "score": 1.0,
+ "content": "(b) Did you mention the license of the assets? [Yes] the full code will be released with a",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 162,
+ 642,
+ 283,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 642,
+ 283,
+ 654
+ ],
+ "score": 1.0,
+ "content": "NC by NC license on GitHub",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 146,
+ 654,
+ 506,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 654,
+ 506,
+ 667
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include any new assets either in the supplemental material or as a URL?",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 161,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "[Yes] we include summary result files that contain all the statistics to reproduce the",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 160,
+ 676,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 676,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "main claims of the paper in the supplementary material. The full codebase along with",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 160,
+ 687,
+ 464,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 687,
+ 464,
+ 700
+ ],
+ "score": 1.0,
+ "content": "pre-trained models will be released upon completion of the review process.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 146,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 162,
+ 711,
+ 478,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 711,
+ 478,
+ 723
+ ],
+ "score": 1.0,
+ "content": "using/curating? [N/A] We are only using the Imagenet and INaturalist datasets",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 46.5
+ }
+ ],
+ "page_idx": 12,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "score": 1.0,
+ "content": "13",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 73,
+ 504,
+ 104
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 72,
+ 506,
+ 86
+ ],
+ "score": 1.0,
+ "content": "Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix:",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 83,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 83,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 93,
+ 370,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 93,
+ 370,
+ 105
+ ],
+ "score": 1.0,
+ "content": "international conference on computer vision, pages 6023–6032, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 105,
+ 72,
+ 506,
+ 105
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 503,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 123
+ ],
+ "score": 1.0,
+ "content": "Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 120,
+ 506,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 120,
+ 506,
+ 133
+ ],
+ "score": 1.0,
+ "content": "principled trade-off between robustness and accuracy. In International conference on machine learning, pages",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 130,
+ 213,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 130,
+ 213,
+ 141
+ ],
+ "score": 1.0,
+ "content": "7472–7482. PMLR, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 105,
+ 110,
+ 506,
+ 141
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 147,
+ 504,
+ 169
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 506,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 506,
+ 160
+ ],
+ "score": 1.0,
+ "content": "Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 115,
+ 158,
+ 316,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 158,
+ 316,
+ 169
+ ],
+ "score": 1.0,
+ "content": "minimization. arXiv preprint arXiv:1710.09412, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 105,
+ 147,
+ 506,
+ 169
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 188,
+ 156,
+ 201
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 158,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 158,
+ 203
+ ],
+ "score": 1.0,
+ "content": "Checklist",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 209,
+ 208,
+ 220
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 208,
+ 210,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 208,
+ 210,
+ 222
+ ],
+ "score": 1.0,
+ "content": "1. For all authors...",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9,
+ "bbox_fs": [
+ 129,
+ 208,
+ 210,
+ 222
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 224,
+ 505,
+ 328
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 223,
+ 505,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 223,
+ 505,
+ 236
+ ],
+ "score": 1.0,
+ "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 235,
+ 288,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 235,
+ 288,
+ 247
+ ],
+ "score": 1.0,
+ "content": "contributions and scope? [Yes]",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 144,
+ 247,
+ 506,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 247,
+ 506,
+ 261
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe the limitations of your work? [N/A] We do not propose a novel",
+ "type": "text"
+ }
+ ],
+ "index": 12,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 258,
+ 312,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 258,
+ 312,
+ 270
+ ],
+ "score": 1.0,
+ "content": "method that requires such discussion",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 270,
+ 506,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 270,
+ 506,
+ 283
+ ],
+ "score": 1.0,
+ "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes] We",
+ "type": "text"
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 282,
+ 505,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 282,
+ 505,
+ 294
+ ],
+ "score": 1.0,
+ "content": "carefully described why the observations we have made can be dangerous for real",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 162,
+ 294,
+ 237,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 294,
+ 237,
+ 304
+ ],
+ "score": 1.0,
+ "content": "world applications",
+ "type": "text"
+ }
+ ],
+ "index": 16,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 305,
+ 505,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 305,
+ 505,
+ 318
+ ],
+ "score": 1.0,
+ "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to",
+ "type": "text"
+ }
+ ],
+ "index": 17,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 316,
+ 214,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 316,
+ 214,
+ 329
+ ],
+ "score": 1.0,
+ "content": "them? [Yes]",
+ "type": "text"
+ }
+ ],
+ "index": 18,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 14,
+ "bbox_fs": [
+ 144,
+ 223,
+ 506,
+ 329
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 331,
+ 302,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 330,
+ 304,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 330,
+ 304,
+ 344
+ ],
+ "score": 1.0,
+ "content": "2. If you are including theoretical results...",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 129,
+ 330,
+ 304,
+ 344
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 345,
+ 505,
+ 402
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 345,
+ 505,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 345,
+ 505,
+ 358
+ ],
+ "score": 1.0,
+ "content": "(a) Did you state the full set of assumptions of all theoretical results? [Yes] Our only",
+ "type": "text"
+ }
+ ],
+ "index": 20,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 355,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 355,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "“theoretical result” consists in a formal statistical test for which we precisely describe",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 162,
+ 368,
+ 265,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 368,
+ 265,
+ 379
+ ],
+ "score": 1.0,
+ "content": "our settings and statistics",
+ "type": "text"
+ }
+ ],
+ "index": 22,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 379,
+ 506,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 379,
+ 506,
+ 393
+ ],
+ "score": 1.0,
+ "content": "(b) Did you include complete proofs of all theoretical results? [N/A] No theoretical result",
+ "type": "text"
+ }
+ ],
+ "index": 23,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 392,
+ 285,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 392,
+ 285,
+ 403
+ ],
+ "score": 1.0,
+ "content": "requiring proofs was provided",
+ "type": "text"
+ }
+ ],
+ "index": 24,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 146,
+ 345,
+ 506,
+ 403
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 406,
+ 241,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 403,
+ 243,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 403,
+ 243,
+ 419
+ ],
+ "score": 1.0,
+ "content": "3. If you ran experiments...",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 128,
+ 403,
+ 243,
+ 419
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 145,
+ 420,
+ 505,
+ 589
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 147,
+ 419,
+ 506,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 419,
+ 506,
+ 433
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the code, data, and instructions needed to reproduce the main exper-",
+ "type": "text"
+ }
+ ],
+ "index": 26,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 431,
+ 506,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 431,
+ 506,
+ 443
+ ],
+ "score": 1.0,
+ "content": "imental results (either in the supplemental material or as a URL)? [Yes] We include",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 162,
+ 442,
+ 505,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 442,
+ 505,
+ 454
+ ],
+ "score": 1.0,
+ "content": "summary statistics in the supplementary material which is enough to validate our",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 162,
+ 453,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 453,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": "claims. The full codebase and all the saved models will be released upon completion",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 162,
+ 464,
+ 469,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 464,
+ 469,
+ 476
+ ],
+ "score": 1.0,
+ "content": "of the review process (this includes almost a thousand pre-trained resnet50s)",
+ "type": "text"
+ }
+ ],
+ "index": 30,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 476,
+ 505,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 476,
+ 505,
+ 490
+ ],
+ "score": 1.0,
+ "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they",
+ "type": "text"
+ }
+ ],
+ "index": 31,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 486,
+ 506,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 486,
+ 506,
+ 501
+ ],
+ "score": 1.0,
+ "content": "were chosen)? [Yes] We specify within each caption the key experimental setups,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 161,
+ 498,
+ 383,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 498,
+ 383,
+ 511
+ ],
+ "score": 1.0,
+ "content": "additional training details are provided in the appendix",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 510,
+ 507,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 510,
+ 507,
+ 525
+ ],
+ "score": 1.0,
+ "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-",
+ "type": "text"
+ }
+ ],
+ "index": 34,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 522,
+ 506,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 522,
+ 506,
+ 534
+ ],
+ "score": 1.0,
+ "content": "ments multiple times)? [Yes] we averaged all of our runs over a significant amount of",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 162,
+ 533,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 533,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "realizations (20) and perform a formal statistical test on the significance of our results",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 146,
+ 545,
+ 506,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 545,
+ 506,
+ 559
+ ],
+ "score": 1.0,
+ "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type",
+ "type": "text"
+ }
+ ],
+ "index": 37,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 555,
+ 506,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 555,
+ 506,
+ 569
+ ],
+ "score": 1.0,
+ "content": "of GPUs, internal cluster, or cloud provider)? [Yes] We provide this in the beginning of",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 162,
+ 568,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 568,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "the appendix. Note that we will be providing the saved models and summary statistics",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 161,
+ 577,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 577,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "to make the results easily reproducible even without GPU or computation ressources.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 146,
+ 419,
+ 507,
+ 591
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 128,
+ 593,
+ 504,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 591,
+ 507,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 591,
+ 507,
+ 606
+ ],
+ "score": 1.0,
+ "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 41,
+ "bbox_fs": [
+ 129,
+ 591,
+ 507,
+ 606
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 608,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 607,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 607,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] For Imagenet and",
+ "type": "text"
+ }
+ ],
+ "index": 42,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 618,
+ 208,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 618,
+ 208,
+ 630
+ ],
+ "score": 1.0,
+ "content": "INaturalist",
+ "type": "text"
+ }
+ ],
+ "index": 43,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 630,
+ 506,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 630,
+ 506,
+ 644
+ ],
+ "score": 1.0,
+ "content": "(b) Did you mention the license of the assets? [Yes] the full code will be released with a",
+ "type": "text"
+ }
+ ],
+ "index": 44,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 642,
+ 283,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 642,
+ 283,
+ 654
+ ],
+ "score": 1.0,
+ "content": "NC by NC license on GitHub",
+ "type": "text"
+ }
+ ],
+ "index": 45,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 654,
+ 506,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 654,
+ 506,
+ 667
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include any new assets either in the supplemental material or as a URL?",
+ "type": "text"
+ }
+ ],
+ "index": 46,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "[Yes] we include summary result files that contain all the statistics to reproduce the",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 160,
+ 676,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 676,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "main claims of the paper in the supplementary material. The full codebase along with",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 160,
+ 687,
+ 464,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 687,
+ 464,
+ 700
+ ],
+ "score": 1.0,
+ "content": "pre-trained models will be released upon completion of the review process.",
+ "type": "text"
+ }
+ ],
+ "index": 49,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re",
+ "type": "text"
+ }
+ ],
+ "index": 50,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 711,
+ 478,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 711,
+ 478,
+ 723
+ ],
+ "score": 1.0,
+ "content": "using/curating? [N/A] We are only using the Imagenet and INaturalist datasets",
+ "type": "text"
+ }
+ ],
+ "index": 51,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 46.5,
+ "bbox_fs": [
+ 145,
+ 607,
+ 506,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 145,
+ 73,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 72,
+ 506,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 72,
+ 506,
+ 86
+ ],
+ "score": 1.0,
+ "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 161,
+ 83,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 83,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "information or offensive content? [N/A] We are only using the Imagenet and INaturalist",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 162,
+ 93,
+ 196,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 93,
+ 196,
+ 106
+ ],
+ "score": 1.0,
+ "content": "datasets",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 109,
+ 432,
+ 121
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 107,
+ 433,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 107,
+ 433,
+ 124
+ ],
+ "score": 1.0,
+ "content": "5. If you used crowdsourcing or conducted research with human subjects...",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 124,
+ 506,
+ 195
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 123,
+ 506,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 123,
+ 506,
+ 137
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the full text of instructions given to participants and screenshots, if",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 161,
+ 135,
+ 237,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 135,
+ 237,
+ 147
+ ],
+ "score": 1.0,
+ "content": "applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 145,
+ 148,
+ 505,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 148,
+ 505,
+ 161
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe any potential participant risks, with links to Institutional Review",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 162,
+ 159,
+ 342,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 159,
+ 342,
+ 172
+ ],
+ "score": 1.0,
+ "content": "Board (IRB) approvals, if applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 147,
+ 172,
+ 505,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 172,
+ 505,
+ 185
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 161,
+ 183,
+ 333,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 183,
+ 333,
+ 196
+ ],
+ "score": 1.0,
+ "content": "spent on participant compensation? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 6.5
+ }
+ ],
+ "page_idx": 13,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "14",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 145,
+ 73,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 72,
+ 506,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 72,
+ 506,
+ 86
+ ],
+ "score": 1.0,
+ "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 161,
+ 83,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 83,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "information or offensive content? [N/A] We are only using the Imagenet and INaturalist",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 162,
+ 93,
+ 196,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 93,
+ 196,
+ 106
+ ],
+ "score": 1.0,
+ "content": "datasets",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 145,
+ 72,
+ 506,
+ 106
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 109,
+ 432,
+ 121
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 107,
+ 433,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 107,
+ 433,
+ 124
+ ],
+ "score": 1.0,
+ "content": "5. If you used crowdsourcing or conducted research with human subjects...",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 128,
+ 107,
+ 433,
+ 124
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 124,
+ 506,
+ 195
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 123,
+ 506,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 123,
+ 506,
+ 137
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the full text of instructions given to participants and screenshots, if",
+ "type": "text"
+ }
+ ],
+ "index": 4,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 135,
+ 237,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 135,
+ 237,
+ 147
+ ],
+ "score": 1.0,
+ "content": "applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 5,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 148,
+ 505,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 148,
+ 505,
+ 161
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe any potential participant risks, with links to Institutional Review",
+ "type": "text"
+ }
+ ],
+ "index": 6,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 159,
+ 342,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 159,
+ 342,
+ 172
+ ],
+ "score": 1.0,
+ "content": "Board (IRB) approvals, if applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 7,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 147,
+ 172,
+ 505,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 172,
+ 505,
+ 185
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 183,
+ 333,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 183,
+ 333,
+ 196
+ ],
+ "score": 1.0,
+ "content": "spent on participant compensation? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 145,
+ 123,
+ 506,
+ 196
+ ]
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file