{
"pdf_info": [
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
111,
97,
500,
118
],
"lines": [
{
"bbox": [
110,
96,
500,
120
],
"spans": [
{
"bbox": [
110,
96,
500,
120
],
"score": 1.0,
"content": "Dataset Distillation using Neural Feature Regression",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
142,
159,
301,
204
],
"lines": [
{
"bbox": [
188,
159,
258,
172
],
"spans": [
{
"bbox": [
188,
159,
258,
172
],
"score": 1.0,
"content": "Yongchao Zhou",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
153,
169,
290,
183
],
"spans": [
{
"bbox": [
153,
169,
290,
183
],
"score": 1.0,
"content": "Department of Computer Science",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
177,
180,
267,
194
],
"spans": [
{
"bbox": [
177,
180,
267,
194
],
"score": 1.0,
"content": "University of Toronto",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
141,
193,
302,
204
],
"spans": [
{
"bbox": [
141,
193,
302,
204
],
"score": 1.0,
"content": "yongchao.zhou@mail.utoronto.ca",
"type": "text"
}
],
"index": 7
}
],
"index": 4.0
},
{
"type": "text",
"bbox": [
341,
159,
468,
204
],
"lines": [
{
"bbox": [
363,
158,
447,
172
],
"spans": [
{
"bbox": [
363,
158,
447,
172
],
"score": 1.0,
"content": "Ehsan Nezhadarya",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
372,
170,
438,
182
],
"spans": [
{
"bbox": [
372,
170,
438,
182
],
"score": 1.0,
"content": "Toronto AI Lab",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
357,
182,
452,
192
],
"spans": [
{
"bbox": [
357,
182,
452,
192
],
"score": 1.0,
"content": "LG Electronics Canada",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
341,
192,
469,
205
],
"spans": [
{
"bbox": [
341,
192,
469,
205
],
"score": 1.0,
"content": "ehsan.nezhadarya@lge.com",
"type": "text"
}
],
"index": 8
}
],
"index": 5.0
},
{
"type": "text",
"bbox": [
237,
220,
373,
264
],
"lines": [
{
"bbox": [
282,
220,
329,
233
],
"spans": [
{
"bbox": [
282,
220,
329,
233
],
"score": 1.0,
"content": "Jimmy Ba",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
236,
230,
374,
244
],
"spans": [
{
"bbox": [
236,
230,
374,
244
],
"score": 1.0,
"content": "Department of Computer Science",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
260,
241,
350,
255
],
"spans": [
{
"bbox": [
260,
241,
350,
255
],
"score": 1.0,
"content": "University of Toronto",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
257,
254,
354,
265
],
"spans": [
{
"bbox": [
257,
254,
354,
265
],
"score": 1.0,
"content": "jba@cs.toronto.edu",
"type": "text"
}
],
"index": 12
}
],
"index": 10.5
},
{
"type": "title",
"bbox": [
283,
292,
328,
306
],
"lines": [
{
"bbox": [
282,
292,
330,
306
],
"spans": [
{
"bbox": [
282,
292,
330,
306
],
"score": 1.0,
"content": "Abstract",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
143,
318,
469,
493
],
"lines": [
{
"bbox": [
142,
319,
469,
330
],
"spans": [
{
"bbox": [
142,
319,
469,
330
],
"score": 1.0,
"content": "Dataset distillation aims to learn a small synthetic dataset that preserves most of",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
141,
329,
470,
341
],
"spans": [
{
"bbox": [
141,
329,
470,
341
],
"score": 1.0,
"content": "the information from the original dataset. Dataset distillation can be formulated",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
141,
340,
470,
352
],
"spans": [
{
"bbox": [
141,
340,
470,
352
],
"score": 1.0,
"content": "as a bi-level meta-learning problem where the outer loop optimizes the meta-",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
142,
351,
470,
363
],
"spans": [
{
"bbox": [
142,
351,
470,
363
],
"score": 1.0,
"content": "dataset and the inner loop trains a model on the distilled data. Meta-gradient",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
141,
362,
469,
375
],
"spans": [
{
"bbox": [
141,
362,
469,
375
],
"score": 1.0,
"content": "computation is one of the key challenges in this formulation, as differentiating",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
142,
373,
469,
386
],
"spans": [
{
"bbox": [
142,
373,
469,
386
],
"score": 1.0,
"content": "through the inner loop learning procedure introduces significant computation and",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
141,
384,
470,
397
],
"spans": [
{
"bbox": [
141,
384,
470,
397
],
"score": 1.0,
"content": "memory costs. In this paper, we address these challenges using neural Feature",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
142,
395,
469,
407
],
"spans": [
{
"bbox": [
142,
395,
469,
407
],
"score": 1.0,
"content": "Regression with Pooling (FRePo), achieving the state-of-the-art performance with",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
141,
406,
469,
417
],
"spans": [
{
"bbox": [
141,
406,
469,
417
],
"score": 1.0,
"content": "an order of magnitude less memory requirement and two orders of magnitude",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
141,
416,
470,
429
],
"spans": [
{
"bbox": [
141,
416,
470,
429
],
"score": 1.0,
"content": "faster training than previous methods. The proposed algorithm is analogous to",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
141,
427,
470,
439
],
"spans": [
{
"bbox": [
141,
427,
470,
439
],
"score": 1.0,
"content": "truncated backpropagation through time with a pool of models to alleviate various",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
141,
439,
469,
450
],
"spans": [
{
"bbox": [
141,
439,
469,
450
],
"score": 1.0,
"content": "types of overfitting in dataset distillation. FRePo significantly outperforms the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
141,
450,
470,
461
],
"spans": [
{
"bbox": [
141,
450,
470,
461
],
"score": 1.0,
"content": "previous methods on CIFAR100, Tiny ImageNet, and ImageNet-1K. Furthermore,",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
141,
460,
469,
472
],
"spans": [
{
"bbox": [
141,
460,
469,
472
],
"score": 1.0,
"content": "we show that high-quality distilled data can greatly improve various downstream",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
142,
471,
469,
483
],
"spans": [
{
"bbox": [
142,
471,
469,
483
],
"score": 1.0,
"content": "applications, such as continual learning and membership inference defense. Please",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
141,
481,
435,
495
],
"spans": [
{
"bbox": [
141,
481,
435,
495
],
"score": 1.0,
"content": "check out our webpage at https://sites.google.com/view/frepo.",
"type": "text"
}
],
"index": 29
}
],
"index": 21.5
},
{
"type": "title",
"bbox": [
107,
513,
190,
527
],
"lines": [
{
"bbox": [
105,
511,
192,
529
],
"spans": [
{
"bbox": [
105,
511,
192,
529
],
"score": 1.0,
"content": "1 Introduction",
"type": "text"
}
],
"index": 30
}
],
"index": 30
},
{
"type": "text",
"bbox": [
107,
538,
505,
626
],
"lines": [
{
"bbox": [
105,
538,
506,
551
],
"spans": [
{
"bbox": [
105,
538,
506,
551
],
"score": 1.0,
"content": "Knowledge distillation [1] is a technique in deep learning to compress knowledge for easy deployment.",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
549,
505,
562
],
"spans": [
{
"bbox": [
105,
549,
505,
562
],
"score": 1.0,
"content": "Most previous works focus on model distillation [2, 3] where the knowledge acquired by a large",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
106,
561,
505,
572
],
"spans": [
{
"bbox": [
106,
561,
505,
572
],
"score": 1.0,
"content": "teacher model is transferred to a small student model. In contrast, dataset distillation [4, 5] aims to",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
572,
505,
583
],
"spans": [
{
"bbox": [
106,
572,
505,
583
],
"score": 1.0,
"content": "learn a small set of synthetic examples preserving most of the information from a large dataset such",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
582,
506,
595
],
"spans": [
{
"bbox": [
105,
582,
506,
595
],
"score": 1.0,
"content": "that a model trained on it can achieve similar test performance as one trained on the original dataset.",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
592,
506,
605
],
"spans": [
{
"bbox": [
105,
592,
506,
605
],
"score": 1.0,
"content": "Distilled data can accelerate model training and reduce the cost of storing and sharing a dataset.",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
604,
506,
617
],
"spans": [
{
"bbox": [
105,
604,
506,
617
],
"score": 1.0,
"content": "Moreover, its highly condensed and synthetic nature can also benefit various applications, such as",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
615,
487,
627
],
"spans": [
{
"bbox": [
106,
615,
487,
627
],
"score": 1.0,
"content": "continual learning [5–8], neural architecture search [5, 7], and privacy-preserving tasks [9, 10].",
"type": "text"
}
],
"index": 38
}
],
"index": 34.5
},
{
"type": "text",
"bbox": [
107,
631,
505,
686
],
"lines": [
{
"bbox": [
105,
631,
506,
645
],
"spans": [
{
"bbox": [
105,
631,
506,
645
],
"score": 1.0,
"content": "Dataset distillation was first studied by Maclaurin et al. [11] in the context of gradient-based hyperpa-",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
642,
505,
654
],
"spans": [
{
"bbox": [
106,
642,
505,
654
],
"score": 1.0,
"content": "rameter optimization and subsequently Wang et al. [4] formally proposed dataset distillation as a new",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
653,
505,
665
],
"spans": [
{
"bbox": [
106,
653,
505,
665
],
"score": 1.0,
"content": "task. Dataset distillation can be naturally formulated as a bi-level meta-learning problem. The inner",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
662,
505,
678
],
"spans": [
{
"bbox": [
105,
662,
505,
678
],
"score": 1.0,
"content": "loop optimizes the model parameters on the distilled data (meta-parameters), while the outer loop",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
675,
317,
687
],
"spans": [
{
"bbox": [
105,
675,
317,
687
],
"score": 1.0,
"content": "refines the distilled data with meta-gradient updates.",
"type": "text"
}
],
"index": 43
}
],
"index": 41
},
{
"type": "text",
"bbox": [
108,
691,
504,
714
],
"lines": [
{
"bbox": [
106,
690,
505,
703
],
"spans": [
{
"bbox": [
106,
690,
505,
703
],
"score": 1.0,
"content": "One key challenge in dataset distillation is computing the meta-gradient. Several methods [4, 11–13]",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
703,
505,
715
],
"spans": [
{
"bbox": [
106,
703,
505,
715
],
"score": 1.0,
"content": "compute it by back-propagating through the unrolled computation graph, but they often suffer from",
"type": "text"
}
],
"index": 45
}
],
"index": 44.5
}
],
"page_idx": 0,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
731,
385,
741
],
"lines": [
{
"bbox": [
105,
730,
387,
744
],
"spans": [
{
"bbox": [
105,
730,
387,
744
],
"score": 1.0,
"content": "36th Conference on Neural Information Processing Systems (NeurIPS 2022).",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
111,
97,
500,
118
],
"lines": [
{
"bbox": [
110,
96,
500,
120
],
"spans": [
{
"bbox": [
110,
96,
500,
120
],
"score": 1.0,
"content": "Dataset Distillation using Neural Feature Regression",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
142,
159,
301,
204
],
"lines": [
{
"bbox": [
188,
159,
258,
172
],
"spans": [
{
"bbox": [
188,
159,
258,
172
],
"score": 1.0,
"content": "Yongchao Zhou",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
153,
169,
290,
183
],
"spans": [
{
"bbox": [
153,
169,
290,
183
],
"score": 1.0,
"content": "Department of Computer Science",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
177,
180,
267,
194
],
"spans": [
{
"bbox": [
177,
180,
267,
194
],
"score": 1.0,
"content": "University of Toronto",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
141,
193,
302,
204
],
"spans": [
{
"bbox": [
141,
193,
302,
204
],
"score": 1.0,
"content": "yongchao.zhou@mail.utoronto.ca",
"type": "text"
}
],
"index": 7
}
],
"index": 4.0,
"bbox_fs": [
141,
159,
302,
204
]
},
{
"type": "text",
"bbox": [
341,
159,
468,
204
],
"lines": [
{
"bbox": [
363,
158,
447,
172
],
"spans": [
{
"bbox": [
363,
158,
447,
172
],
"score": 1.0,
"content": "Ehsan Nezhadarya",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
372,
170,
438,
182
],
"spans": [
{
"bbox": [
372,
170,
438,
182
],
"score": 1.0,
"content": "Toronto AI Lab",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
357,
182,
452,
192
],
"spans": [
{
"bbox": [
357,
182,
452,
192
],
"score": 1.0,
"content": "LG Electronics Canada",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
341,
192,
469,
205
],
"spans": [
{
"bbox": [
341,
192,
469,
205
],
"score": 1.0,
"content": "ehsan.nezhadarya@lge.com",
"type": "text"
}
],
"index": 8
}
],
"index": 5.0,
"bbox_fs": [
341,
158,
469,
205
]
},
{
"type": "text",
"bbox": [
237,
220,
373,
264
],
"lines": [
{
"bbox": [
282,
220,
329,
233
],
"spans": [
{
"bbox": [
282,
220,
329,
233
],
"score": 1.0,
"content": "Jimmy Ba",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
236,
230,
374,
244
],
"spans": [
{
"bbox": [
236,
230,
374,
244
],
"score": 1.0,
"content": "Department of Computer Science",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
260,
241,
350,
255
],
"spans": [
{
"bbox": [
260,
241,
350,
255
],
"score": 1.0,
"content": "University of Toronto",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
257,
254,
354,
265
],
"spans": [
{
"bbox": [
257,
254,
354,
265
],
"score": 1.0,
"content": "jba@cs.toronto.edu",
"type": "text"
}
],
"index": 12
}
],
"index": 10.5,
"bbox_fs": [
236,
220,
374,
265
]
},
{
"type": "title",
"bbox": [
283,
292,
328,
306
],
"lines": [
{
"bbox": [
282,
292,
330,
306
],
"spans": [
{
"bbox": [
282,
292,
330,
306
],
"score": 1.0,
"content": "Abstract",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
143,
318,
469,
493
],
"lines": [
{
"bbox": [
142,
319,
469,
330
],
"spans": [
{
"bbox": [
142,
319,
469,
330
],
"score": 1.0,
"content": "Dataset distillation aims to learn a small synthetic dataset that preserves most of",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
141,
329,
470,
341
],
"spans": [
{
"bbox": [
141,
329,
470,
341
],
"score": 1.0,
"content": "the information from the original dataset. Dataset distillation can be formulated",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
141,
340,
470,
352
],
"spans": [
{
"bbox": [
141,
340,
470,
352
],
"score": 1.0,
"content": "as a bi-level meta-learning problem where the outer loop optimizes the meta-",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
142,
351,
470,
363
],
"spans": [
{
"bbox": [
142,
351,
470,
363
],
"score": 1.0,
"content": "dataset and the inner loop trains a model on the distilled data. Meta-gradient",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
141,
362,
469,
375
],
"spans": [
{
"bbox": [
141,
362,
469,
375
],
"score": 1.0,
"content": "computation is one of the key challenges in this formulation, as differentiating",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
142,
373,
469,
386
],
"spans": [
{
"bbox": [
142,
373,
469,
386
],
"score": 1.0,
"content": "through the inner loop learning procedure introduces significant computation and",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
141,
384,
470,
397
],
"spans": [
{
"bbox": [
141,
384,
470,
397
],
"score": 1.0,
"content": "memory costs. In this paper, we address these challenges using neural Feature",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
142,
395,
469,
407
],
"spans": [
{
"bbox": [
142,
395,
469,
407
],
"score": 1.0,
"content": "Regression with Pooling (FRePo), achieving the state-of-the-art performance with",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
141,
406,
469,
417
],
"spans": [
{
"bbox": [
141,
406,
469,
417
],
"score": 1.0,
"content": "an order of magnitude less memory requirement and two orders of magnitude",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
141,
416,
470,
429
],
"spans": [
{
"bbox": [
141,
416,
470,
429
],
"score": 1.0,
"content": "faster training than previous methods. The proposed algorithm is analogous to",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
141,
427,
470,
439
],
"spans": [
{
"bbox": [
141,
427,
470,
439
],
"score": 1.0,
"content": "truncated backpropagation through time with a pool of models to alleviate various",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
141,
439,
469,
450
],
"spans": [
{
"bbox": [
141,
439,
469,
450
],
"score": 1.0,
"content": "types of overfitting in dataset distillation. FRePo significantly outperforms the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
141,
450,
470,
461
],
"spans": [
{
"bbox": [
141,
450,
470,
461
],
"score": 1.0,
"content": "previous methods on CIFAR100, Tiny ImageNet, and ImageNet-1K. Furthermore,",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
141,
460,
469,
472
],
"spans": [
{
"bbox": [
141,
460,
469,
472
],
"score": 1.0,
"content": "we show that high-quality distilled data can greatly improve various downstream",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
142,
471,
469,
483
],
"spans": [
{
"bbox": [
142,
471,
469,
483
],
"score": 1.0,
"content": "applications, such as continual learning and membership inference defense. Please",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
141,
481,
435,
495
],
"spans": [
{
"bbox": [
141,
481,
435,
495
],
"score": 1.0,
"content": "check out our webpage at https://sites.google.com/view/frepo.",
"type": "text"
}
],
"index": 29
}
],
"index": 21.5,
"bbox_fs": [
141,
319,
470,
495
]
},
{
"type": "title",
"bbox": [
107,
513,
190,
527
],
"lines": [
{
"bbox": [
105,
511,
192,
529
],
"spans": [
{
"bbox": [
105,
511,
192,
529
],
"score": 1.0,
"content": "1 Introduction",
"type": "text"
}
],
"index": 30
}
],
"index": 30
},
{
"type": "text",
"bbox": [
107,
538,
505,
626
],
"lines": [
{
"bbox": [
105,
538,
506,
551
],
"spans": [
{
"bbox": [
105,
538,
506,
551
],
"score": 1.0,
"content": "Knowledge distillation [1] is a technique in deep learning to compress knowledge for easy deployment.",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
549,
505,
562
],
"spans": [
{
"bbox": [
105,
549,
505,
562
],
"score": 1.0,
"content": "Most previous works focus on model distillation [2, 3] where the knowledge acquired by a large",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
106,
561,
505,
572
],
"spans": [
{
"bbox": [
106,
561,
505,
572
],
"score": 1.0,
"content": "teacher model is transferred to a small student model. In contrast, dataset distillation [4, 5] aims to",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
572,
505,
583
],
"spans": [
{
"bbox": [
106,
572,
505,
583
],
"score": 1.0,
"content": "learn a small set of synthetic examples preserving most of the information from a large dataset such",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
582,
506,
595
],
"spans": [
{
"bbox": [
105,
582,
506,
595
],
"score": 1.0,
"content": "that a model trained on it can achieve similar test performance as one trained on the original dataset.",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
592,
506,
605
],
"spans": [
{
"bbox": [
105,
592,
506,
605
],
"score": 1.0,
"content": "Distilled data can accelerate model training and reduce the cost of storing and sharing a dataset.",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
604,
506,
617
],
"spans": [
{
"bbox": [
105,
604,
506,
617
],
"score": 1.0,
"content": "Moreover, its highly condensed and synthetic nature can also benefit various applications, such as",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
615,
487,
627
],
"spans": [
{
"bbox": [
106,
615,
487,
627
],
"score": 1.0,
"content": "continual learning [5–8], neural architecture search [5, 7], and privacy-preserving tasks [9, 10].",
"type": "text"
}
],
"index": 38
}
],
"index": 34.5,
"bbox_fs": [
105,
538,
506,
627
]
},
{
"type": "text",
"bbox": [
107,
631,
505,
686
],
"lines": [
{
"bbox": [
105,
631,
506,
645
],
"spans": [
{
"bbox": [
105,
631,
506,
645
],
"score": 1.0,
"content": "Dataset distillation was first studied by Maclaurin et al. [11] in the context of gradient-based hyperpa-",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
642,
505,
654
],
"spans": [
{
"bbox": [
106,
642,
505,
654
],
"score": 1.0,
"content": "rameter optimization and subsequently Wang et al. [4] formally proposed dataset distillation as a new",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
653,
505,
665
],
"spans": [
{
"bbox": [
106,
653,
505,
665
],
"score": 1.0,
"content": "task. Dataset distillation can be naturally formulated as a bi-level meta-learning problem. The inner",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
662,
505,
678
],
"spans": [
{
"bbox": [
105,
662,
505,
678
],
"score": 1.0,
"content": "loop optimizes the model parameters on the distilled data (meta-parameters), while the outer loop",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
675,
317,
687
],
"spans": [
{
"bbox": [
105,
675,
317,
687
],
"score": 1.0,
"content": "refines the distilled data with meta-gradient updates.",
"type": "text"
}
],
"index": 43
}
],
"index": 41,
"bbox_fs": [
105,
631,
506,
687
]
},
{
"type": "text",
"bbox": [
108,
691,
504,
714
],
"lines": [
{
"bbox": [
106,
690,
505,
703
],
"spans": [
{
"bbox": [
106,
690,
505,
703
],
"score": 1.0,
"content": "One key challenge in dataset distillation is computing the meta-gradient. Several methods [4, 11–13]",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
703,
505,
715
],
"spans": [
{
"bbox": [
106,
703,
505,
715
],
"score": 1.0,
"content": "compute it by back-propagating through the unrolled computation graph, but they often suffer from",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
279,
507,
293
],
"spans": [
{
"bbox": [
105,
279,
507,
293
],
"score": 1.0,
"content": "huge compute and memory requirement [14], training instability [15, 16], and truncation bias [17].",
"type": "text",
"cross_page": true
}
],
"index": 6
},
{
"bbox": [
105,
290,
506,
303
],
"spans": [
{
"bbox": [
105,
290,
506,
303
],
"score": 1.0,
"content": "To avoid unrolled optimization, surrogate objectives are used to derive the meta-gradient, such",
"type": "text",
"cross_page": true
}
],
"index": 7
},
{
"bbox": [
105,
301,
506,
315
],
"spans": [
{
"bbox": [
105,
301,
506,
315
],
"score": 1.0,
"content": "as gradient matching [5, 7, 18], feature alignment [8, 19], and training trajectory matching [20].",
"type": "text",
"cross_page": true
}
],
"index": 8
},
{
"bbox": [
105,
312,
505,
325
],
"spans": [
{
"bbox": [
105,
312,
505,
325
],
"score": 1.0,
"content": "Nevertheless, a surrogate objective may introduce its own bias [19], and thus, may not accurately",
"type": "text",
"cross_page": true
}
],
"index": 9
},
{
"bbox": [
105,
323,
505,
336
],
"spans": [
{
"bbox": [
105,
323,
505,
336
],
"score": 1.0,
"content": "reflect the true objective. An alternative is using kernel methods, such as Neural Tangent Kernel",
"type": "text",
"cross_page": true
}
],
"index": 10
},
{
"bbox": [
105,
334,
506,
348
],
"spans": [
{
"bbox": [
105,
334,
506,
348
],
"score": 1.0,
"content": "(NTK) [21], to approximate the inner optimization [22, 23]. However, computing analytical NTK for",
"type": "text",
"cross_page": true
}
],
"index": 11
},
{
"bbox": [
105,
345,
351,
358
],
"spans": [
{
"bbox": [
105,
345,
351,
358
],
"score": 1.0,
"content": "modern neural network can be extremely expensive [22, 23].",
"type": "text",
"cross_page": true
}
],
"index": 12
}
],
"index": 44.5,
"bbox_fs": [
106,
690,
505,
715
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
109,
70,
501,
216
],
"blocks": [
{
"type": "image_body",
"bbox": [
109,
70,
501,
216
],
"group_id": 0,
"lines": [
{
"bbox": [
109,
70,
501,
216
],
"spans": [
{
"bbox": [
109,
70,
501,
216
],
"score": 0.972,
"type": "image",
"image_path": "99c51e47398cb446d8e74bd29e225e4ce9dc4051310f93f7693b478569789644.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
109,
70,
501,
118.66666666666666
],
"spans": [],
"index": 0
},
{
"bbox": [
109,
118.66666666666666,
501,
167.33333333333331
],
"spans": [],
"index": 1
},
{
"bbox": [
109,
167.33333333333331,
501,
215.99999999999997
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
222,
504,
255
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
221,
504,
235
],
"spans": [
{
"bbox": [
106,
221,
276,
235
],
"score": 1.0,
"content": "Figure 1: Example distilled images from",
"type": "text"
},
{
"bbox": [
276,
222,
303,
232
],
"score": 0.67,
"content": "3 2 \\mathbf { x } 3 2",
"type": "inline_equation"
},
{
"bbox": [
304,
221,
467,
235
],
"score": 1.0,
"content": "CIFAR100, 64x64 Tiny ImageNet, and",
"type": "text"
},
{
"bbox": [
468,
222,
504,
232
],
"score": 0.47,
"content": "1 2 8 \\mathrm { x } 1 2 8",
"type": "inline_equation"
}
],
"index": 3
},
{
"bbox": [
106,
233,
505,
245
],
"spans": [
{
"bbox": [
106,
233,
505,
245
],
"score": 1.0,
"content": "ImageNet Subset. The images look real and transfer well to different architectures. They can be used",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
243,
506,
257
],
"spans": [
{
"bbox": [
105,
243,
506,
257
],
"score": 1.0,
"content": "for various downstream applications, such as continual learning and membership inference defense.",
"type": "text"
}
],
"index": 5
}
],
"index": 4
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
107,
279,
505,
357
],
"lines": [
{
"bbox": [
105,
279,
507,
293
],
"spans": [
{
"bbox": [
105,
279,
507,
293
],
"score": 1.0,
"content": "huge compute and memory requirement [14], training instability [15, 16], and truncation bias [17].",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
290,
506,
303
],
"spans": [
{
"bbox": [
105,
290,
506,
303
],
"score": 1.0,
"content": "To avoid unrolled optimization, surrogate objectives are used to derive the meta-gradient, such",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
301,
506,
315
],
"spans": [
{
"bbox": [
105,
301,
506,
315
],
"score": 1.0,
"content": "as gradient matching [5, 7, 18], feature alignment [8, 19], and training trajectory matching [20].",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
312,
505,
325
],
"spans": [
{
"bbox": [
105,
312,
505,
325
],
"score": 1.0,
"content": "Nevertheless, a surrogate objective may introduce its own bias [19], and thus, may not accurately",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
323,
505,
336
],
"spans": [
{
"bbox": [
105,
323,
505,
336
],
"score": 1.0,
"content": "reflect the true objective. An alternative is using kernel methods, such as Neural Tangent Kernel",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
334,
506,
348
],
"spans": [
{
"bbox": [
105,
334,
506,
348
],
"score": 1.0,
"content": "(NTK) [21], to approximate the inner optimization [22, 23]. However, computing analytical NTK for",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
345,
351,
358
],
"spans": [
{
"bbox": [
105,
345,
351,
358
],
"score": 1.0,
"content": "modern neural network can be extremely expensive [22, 23].",
"type": "text"
}
],
"index": 12
}
],
"index": 9
},
{
"type": "text",
"bbox": [
107,
361,
505,
428
],
"lines": [
{
"bbox": [
105,
360,
507,
375
],
"spans": [
{
"bbox": [
105,
360,
507,
375
],
"score": 1.0,
"content": "Even with an accurate meta-gradient, dataset distillation still suffers from various types of overfitting.",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
372,
506,
385
],
"spans": [
{
"bbox": [
105,
372,
506,
385
],
"score": 1.0,
"content": "For instance, the distilled data can easily overfit to a particular learning algorithm [4, 13, 20], a",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
384,
506,
396
],
"spans": [
{
"bbox": [
105,
384,
506,
396
],
"score": 1.0,
"content": "certain stage of optimization [13, 19], or a certain network architecture [5, 7, 20, 22, 23]. Meanwhile,",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
393,
506,
407
],
"spans": [
{
"bbox": [
106,
393,
506,
407
],
"score": 1.0,
"content": "the model can also overfit the distilled data during training, which is the most common cause of",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
405,
505,
416
],
"spans": [
{
"bbox": [
106,
405,
505,
416
],
"score": 1.0,
"content": "overfitting when we train on a small dataset. All these kinds of overfitting impose difficulties on the",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
417,
321,
428
],
"spans": [
{
"bbox": [
106,
417,
321,
428
],
"score": 1.0,
"content": "training and general-purpose use of the distilled data.",
"type": "text"
}
],
"index": 18
}
],
"index": 15.5
},
{
"type": "text",
"bbox": [
107,
433,
505,
564
],
"lines": [
{
"bbox": [
106,
433,
505,
444
],
"spans": [
{
"bbox": [
106,
433,
505,
444
],
"score": 1.0,
"content": "We propose an efficient meta-gradient computation method and a “model pool” to address the",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
443,
505,
455
],
"spans": [
{
"bbox": [
106,
443,
505,
455
],
"score": 1.0,
"content": "overfitting problems. The bottleneck in meta-gradient computation arises due to the complexity of",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
455,
506,
467
],
"spans": [
{
"bbox": [
106,
455,
506,
467
],
"score": 1.0,
"content": "inner optimization, as we need to know how the inner parameters vary with the outer parameters [24].",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
465,
506,
478
],
"spans": [
{
"bbox": [
106,
465,
506,
478
],
"score": 1.0,
"content": "However, the inner optimization can be pretty simple if we only train the last layer of a neural network",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
477,
505,
488
],
"spans": [
{
"bbox": [
106,
477,
505,
488
],
"score": 1.0,
"content": "to convergence while keeping the feature extractor fixed. In this case, computing the prediction on the",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
487,
505,
500
],
"spans": [
{
"bbox": [
105,
487,
505,
500
],
"score": 1.0,
"content": "real data using the model trained on the distilled data can be expressed as a kernel ridge regression",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
497,
505,
511
],
"spans": [
{
"bbox": [
105,
497,
505,
511
],
"score": 1.0,
"content": "(KRR) with respect to the conjugate kernel [25]. Hence, computing the meta-gradient is simply",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
508,
506,
523
],
"spans": [
{
"bbox": [
105,
508,
506,
523
],
"score": 1.0,
"content": "back-propagating through the kernel and a fixed feature extractor. To alleviate overfitting, we propose",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
519,
506,
533
],
"spans": [
{
"bbox": [
105,
519,
506,
533
],
"score": 1.0,
"content": "to maintain a diverse pool of models instead of periodically training and resetting a single model as in",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
530,
506,
543
],
"spans": [
{
"bbox": [
105,
530,
506,
543
],
"score": 1.0,
"content": "prior work [7, 13, 18]. Intuitively, our algorithm targets the following question: what is the best data",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
542,
505,
553
],
"spans": [
{
"bbox": [
106,
542,
505,
553
],
"score": 1.0,
"content": "to train the linear classifier given the current feature extractor? Due to the diverse feature extractors",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
552,
426,
565
],
"spans": [
{
"bbox": [
105,
552,
426,
565
],
"score": 1.0,
"content": "we use, the distilled data generalize well to a wide range of model distributions.",
"type": "text"
}
],
"index": 30
}
],
"index": 24.5
},
{
"type": "title",
"bbox": [
107,
569,
225,
580
],
"lines": [
{
"bbox": [
106,
568,
227,
583
],
"spans": [
{
"bbox": [
106,
568,
227,
583
],
"score": 1.0,
"content": "Summary of Contributions:",
"type": "text"
}
],
"index": 31
}
],
"index": 31
},
{
"type": "text",
"bbox": [
132,
591,
506,
715
],
"lines": [
{
"bbox": [
133,
590,
505,
604
],
"spans": [
{
"bbox": [
133,
590,
505,
604
],
"score": 1.0,
"content": "• We propose an effective method for dataset distillation. Our method, named neural Feature",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
141,
601,
506,
614
],
"spans": [
{
"bbox": [
141,
601,
506,
614
],
"score": 1.0,
"content": "Regression with Pooling (FRePo), achieves state-of-the-art results on various benchmark",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
142,
613,
506,
625
],
"spans": [
{
"bbox": [
142,
613,
202,
625
],
"score": 1.0,
"content": "datasets with a",
"type": "text"
},
{
"bbox": [
203,
613,
223,
623
],
"score": 0.64,
"content": "1 0 0 \\mathrm { x }",
"type": "inline_equation"
},
{
"bbox": [
224,
613,
350,
625
],
"score": 1.0,
"content": "reduction in training time and a",
"type": "text"
},
{
"bbox": [
350,
613,
366,
623
],
"score": 0.39,
"content": "1 0 \\mathrm { x }",
"type": "inline_equation"
},
{
"bbox": [
366,
613,
506,
625
],
"score": 1.0,
"content": "reduction in GPU memory require-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
142,
624,
499,
636
],
"spans": [
{
"bbox": [
142,
624,
499,
636
],
"score": 1.0,
"content": "ment. Our distilled data looks real (Figure 1) and transfers well to different architectures.",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
140,
641,
505,
654
],
"spans": [
{
"bbox": [
140,
641,
505,
654
],
"score": 1.0,
"content": "We show that FRePo scales well to datasets with high-resolution images or complex label",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
141,
652,
507,
666
],
"spans": [
{
"bbox": [
141,
652,
216,
666
],
"score": 1.0,
"content": "space. We achieve",
"type": "text"
},
{
"bbox": [
216,
653,
238,
663
],
"score": 0.86,
"content": "7 . 5 \\%",
"type": "inline_equation"
},
{
"bbox": [
238,
652,
507,
666
],
"score": 1.0,
"content": "top1 accuracy on ImageNet-1K [26] using only one image per class.",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
141,
663,
505,
676
],
"spans": [
{
"bbox": [
141,
663,
273,
676
],
"score": 1.0,
"content": "The same classifier obtains only",
"type": "text"
},
{
"bbox": [
274,
663,
296,
674
],
"score": 0.86,
"content": "1 . 1 \\%",
"type": "inline_equation"
},
{
"bbox": [
296,
663,
505,
676
],
"score": 1.0,
"content": "accuracy from a random subset of real images. The",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
141,
674,
492,
688
],
"spans": [
{
"bbox": [
141,
674,
492,
688
],
"score": 1.0,
"content": "previous methods struggle in this task due to large memory and compute requirements.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
136,
692,
505,
704
],
"spans": [
{
"bbox": [
136,
692,
505,
704
],
"score": 1.0,
"content": "• We demonstrate that high-quality distilled data can significantly improve various downstream",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
142,
703,
446,
716
],
"spans": [
{
"bbox": [
142,
703,
446,
716
],
"score": 1.0,
"content": "applications, such as continual learning and membership inference defense.",
"type": "text"
}
],
"index": 41
}
],
"index": 36.5
}
],
"page_idx": 1,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
742,
308,
750
],
"lines": [
{
"bbox": [
302,
741,
310,
753
],
"spans": [
{
"bbox": [
302,
741,
310,
753
],
"score": 1.0,
"content": "",
"type": "text",
"height": 12,
"width": 8
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
109,
70,
501,
216
],
"blocks": [
{
"type": "image_body",
"bbox": [
109,
70,
501,
216
],
"group_id": 0,
"lines": [
{
"bbox": [
109,
70,
501,
216
],
"spans": [
{
"bbox": [
109,
70,
501,
216
],
"score": 0.972,
"type": "image",
"image_path": "99c51e47398cb446d8e74bd29e225e4ce9dc4051310f93f7693b478569789644.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
109,
70,
501,
118.66666666666666
],
"spans": [],
"index": 0
},
{
"bbox": [
109,
118.66666666666666,
501,
167.33333333333331
],
"spans": [],
"index": 1
},
{
"bbox": [
109,
167.33333333333331,
501,
215.99999999999997
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
222,
504,
255
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
221,
504,
235
],
"spans": [
{
"bbox": [
106,
221,
276,
235
],
"score": 1.0,
"content": "Figure 1: Example distilled images from",
"type": "text"
},
{
"bbox": [
276,
222,
303,
232
],
"score": 0.67,
"content": "3 2 \\mathbf { x } 3 2",
"type": "inline_equation"
},
{
"bbox": [
304,
221,
467,
235
],
"score": 1.0,
"content": "CIFAR100, 64x64 Tiny ImageNet, and",
"type": "text"
},
{
"bbox": [
468,
222,
504,
232
],
"score": 0.47,
"content": "1 2 8 \\mathrm { x } 1 2 8",
"type": "inline_equation"
}
],
"index": 3
},
{
"bbox": [
106,
233,
505,
245
],
"spans": [
{
"bbox": [
106,
233,
505,
245
],
"score": 1.0,
"content": "ImageNet Subset. The images look real and transfer well to different architectures. They can be used",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
243,
506,
257
],
"spans": [
{
"bbox": [
105,
243,
506,
257
],
"score": 1.0,
"content": "for various downstream applications, such as continual learning and membership inference defense.",
"type": "text"
}
],
"index": 5
}
],
"index": 4
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
107,
279,
505,
357
],
"lines": [],
"index": 9,
"bbox_fs": [
105,
279,
507,
358
],
"lines_deleted": true
},
{
"type": "text",
"bbox": [
107,
361,
505,
428
],
"lines": [
{
"bbox": [
105,
360,
507,
375
],
"spans": [
{
"bbox": [
105,
360,
507,
375
],
"score": 1.0,
"content": "Even with an accurate meta-gradient, dataset distillation still suffers from various types of overfitting.",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
372,
506,
385
],
"spans": [
{
"bbox": [
105,
372,
506,
385
],
"score": 1.0,
"content": "For instance, the distilled data can easily overfit to a particular learning algorithm [4, 13, 20], a",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
384,
506,
396
],
"spans": [
{
"bbox": [
105,
384,
506,
396
],
"score": 1.0,
"content": "certain stage of optimization [13, 19], or a certain network architecture [5, 7, 20, 22, 23]. Meanwhile,",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
393,
506,
407
],
"spans": [
{
"bbox": [
106,
393,
506,
407
],
"score": 1.0,
"content": "the model can also overfit the distilled data during training, which is the most common cause of",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
405,
505,
416
],
"spans": [
{
"bbox": [
106,
405,
505,
416
],
"score": 1.0,
"content": "overfitting when we train on a small dataset. All these kinds of overfitting impose difficulties on the",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
417,
321,
428
],
"spans": [
{
"bbox": [
106,
417,
321,
428
],
"score": 1.0,
"content": "training and general-purpose use of the distilled data.",
"type": "text"
}
],
"index": 18
}
],
"index": 15.5,
"bbox_fs": [
105,
360,
507,
428
]
},
{
"type": "text",
"bbox": [
107,
433,
505,
564
],
"lines": [
{
"bbox": [
106,
433,
505,
444
],
"spans": [
{
"bbox": [
106,
433,
505,
444
],
"score": 1.0,
"content": "We propose an efficient meta-gradient computation method and a “model pool” to address the",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
443,
505,
455
],
"spans": [
{
"bbox": [
106,
443,
505,
455
],
"score": 1.0,
"content": "overfitting problems. The bottleneck in meta-gradient computation arises due to the complexity of",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
455,
506,
467
],
"spans": [
{
"bbox": [
106,
455,
506,
467
],
"score": 1.0,
"content": "inner optimization, as we need to know how the inner parameters vary with the outer parameters [24].",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
465,
506,
478
],
"spans": [
{
"bbox": [
106,
465,
506,
478
],
"score": 1.0,
"content": "However, the inner optimization can be pretty simple if we only train the last layer of a neural network",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
477,
505,
488
],
"spans": [
{
"bbox": [
106,
477,
505,
488
],
"score": 1.0,
"content": "to convergence while keeping the feature extractor fixed. In this case, computing the prediction on the",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
487,
505,
500
],
"spans": [
{
"bbox": [
105,
487,
505,
500
],
"score": 1.0,
"content": "real data using the model trained on the distilled data can be expressed as a kernel ridge regression",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
497,
505,
511
],
"spans": [
{
"bbox": [
105,
497,
505,
511
],
"score": 1.0,
"content": "(KRR) with respect to the conjugate kernel [25]. Hence, computing the meta-gradient is simply",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
508,
506,
523
],
"spans": [
{
"bbox": [
105,
508,
506,
523
],
"score": 1.0,
"content": "back-propagating through the kernel and a fixed feature extractor. To alleviate overfitting, we propose",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
519,
506,
533
],
"spans": [
{
"bbox": [
105,
519,
506,
533
],
"score": 1.0,
"content": "to maintain a diverse pool of models instead of periodically training and resetting a single model as in",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
530,
506,
543
],
"spans": [
{
"bbox": [
105,
530,
506,
543
],
"score": 1.0,
"content": "prior work [7, 13, 18]. Intuitively, our algorithm targets the following question: what is the best data",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
542,
505,
553
],
"spans": [
{
"bbox": [
106,
542,
505,
553
],
"score": 1.0,
"content": "to train the linear classifier given the current feature extractor? Due to the diverse feature extractors",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
552,
426,
565
],
"spans": [
{
"bbox": [
105,
552,
426,
565
],
"score": 1.0,
"content": "we use, the distilled data generalize well to a wide range of model distributions.",
"type": "text"
}
],
"index": 30
}
],
"index": 24.5,
"bbox_fs": [
105,
433,
506,
565
]
},
{
"type": "title",
"bbox": [
107,
569,
225,
580
],
"lines": [
{
"bbox": [
106,
568,
227,
583
],
"spans": [
{
"bbox": [
106,
568,
227,
583
],
"score": 1.0,
"content": "Summary of Contributions:",
"type": "text"
}
],
"index": 31
}
],
"index": 31
},
{
"type": "text",
"bbox": [
132,
591,
506,
715
],
"lines": [
{
"bbox": [
133,
590,
505,
604
],
"spans": [
{
"bbox": [
133,
590,
505,
604
],
"score": 1.0,
"content": "• We propose an effective method for dataset distillation. Our method, named neural Feature",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
141,
601,
506,
614
],
"spans": [
{
"bbox": [
141,
601,
506,
614
],
"score": 1.0,
"content": "Regression with Pooling (FRePo), achieves state-of-the-art results on various benchmark",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
142,
613,
506,
625
],
"spans": [
{
"bbox": [
142,
613,
202,
625
],
"score": 1.0,
"content": "datasets with a",
"type": "text"
},
{
"bbox": [
203,
613,
223,
623
],
"score": 0.64,
"content": "1 0 0 \\mathrm { x }",
"type": "inline_equation"
},
{
"bbox": [
224,
613,
350,
625
],
"score": 1.0,
"content": "reduction in training time and a",
"type": "text"
},
{
"bbox": [
350,
613,
366,
623
],
"score": 0.39,
"content": "1 0 \\mathrm { x }",
"type": "inline_equation"
},
{
"bbox": [
366,
613,
506,
625
],
"score": 1.0,
"content": "reduction in GPU memory require-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
142,
624,
499,
636
],
"spans": [
{
"bbox": [
142,
624,
499,
636
],
"score": 1.0,
"content": "ment. Our distilled data looks real (Figure 1) and transfers well to different architectures.",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
140,
641,
505,
654
],
"spans": [
{
"bbox": [
140,
641,
505,
654
],
"score": 1.0,
"content": "We show that FRePo scales well to datasets with high-resolution images or complex label",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
141,
652,
507,
666
],
"spans": [
{
"bbox": [
141,
652,
216,
666
],
"score": 1.0,
"content": "space. We achieve",
"type": "text"
},
{
"bbox": [
216,
653,
238,
663
],
"score": 0.86,
"content": "7 . 5 \\%",
"type": "inline_equation"
},
{
"bbox": [
238,
652,
507,
666
],
"score": 1.0,
"content": "top1 accuracy on ImageNet-1K [26] using only one image per class.",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
141,
663,
505,
676
],
"spans": [
{
"bbox": [
141,
663,
273,
676
],
"score": 1.0,
"content": "The same classifier obtains only",
"type": "text"
},
{
"bbox": [
274,
663,
296,
674
],
"score": 0.86,
"content": "1 . 1 \\%",
"type": "inline_equation"
},
{
"bbox": [
296,
663,
505,
676
],
"score": 1.0,
"content": "accuracy from a random subset of real images. The",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
141,
674,
492,
688
],
"spans": [
{
"bbox": [
141,
674,
492,
688
],
"score": 1.0,
"content": "previous methods struggle in this task due to large memory and compute requirements.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
136,
692,
505,
704
],
"spans": [
{
"bbox": [
136,
692,
505,
704
],
"score": 1.0,
"content": "• We demonstrate that high-quality distilled data can significantly improve various downstream",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
142,
703,
446,
716
],
"spans": [
{
"bbox": [
142,
703,
446,
716
],
"score": 1.0,
"content": "applications, such as continual learning and membership inference defense.",
"type": "text"
}
],
"index": 41
}
],
"index": 36.5,
"bbox_fs": [
133,
590,
507,
716
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
106,
70,
506,
160
],
"blocks": [
{
"type": "image_body",
"bbox": [
106,
70,
506,
160
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
70,
506,
160
],
"spans": [
{
"bbox": [
106,
70,
506,
160
],
"score": 0.964,
"type": "image",
"image_path": "bd356bac9e8a26f60ac46a3e8c2299fdbc56eeb9580bd6cdb628ffbec6e98257.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
106,
70,
506,
100.0
],
"spans": [],
"index": 0
},
{
"bbox": [
106,
100.0,
506,
130.0
],
"spans": [],
"index": 1
},
{
"bbox": [
106,
130.0,
506,
160.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
167,
506,
235
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
166,
506,
181
],
"spans": [
{
"bbox": [
105,
166,
362,
181
],
"score": 1.0,
"content": "Figure 2: Comparison of FRePo and Unrolled Optimization.",
"type": "text"
},
{
"bbox": [
362,
168,
370,
178
],
"score": 0.44,
"content": "S",
"type": "inline_equation"
},
{
"bbox": [
370,
166,
374,
181
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
375,
168,
388,
178
],
"score": 0.73,
"content": "X _ { s }",
"type": "inline_equation"
},
{
"bbox": [
388,
166,
392,
181
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
393,
168,
404,
178
],
"score": 0.78,
"content": "Y _ { s }",
"type": "inline_equation"
},
{
"bbox": [
404,
166,
506,
181
],
"score": 1.0,
"content": "are the distilled dataset,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
177,
506,
193
],
"spans": [
{
"bbox": [
105,
177,
184,
193
],
"score": 1.0,
"content": "images and labels.",
"type": "text"
},
{
"bbox": [
185,
180,
193,
190
],
"score": 0.74,
"content": "\\mathcal { L }",
"type": "inline_equation"
},
{
"bbox": [
193,
177,
311,
193
],
"score": 1.0,
"content": "is the meta-training loss and",
"type": "text"
},
{
"bbox": [
311,
178,
328,
190
],
"score": 0.85,
"content": "\\dot { \\theta } ^ { ( k ) }",
"type": "inline_equation"
},
{
"bbox": [
328,
177,
331,
193
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
332,
178,
349,
191
],
"score": 0.87,
"content": "g ^ { ( k ) }",
"type": "inline_equation"
},
{
"bbox": [
350,
177,
506,
193
],
"score": 1.0,
"content": "are the model parameter and gradient",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
103,
185,
508,
208
],
"spans": [
{
"bbox": [
103,
185,
135,
208
],
"score": 1.0,
"content": "at step",
"type": "text"
},
{
"bbox": [
136,
191,
142,
201
],
"score": 0.58,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
142,
185,
146,
208
],
"score": 1.0,
"content": ".",
"type": "text"
},
{
"bbox": [
147,
190,
171,
203
],
"score": 0.89,
"content": "f ( X )",
"type": "inline_equation"
},
{
"bbox": [
171,
185,
265,
208
],
"score": 1.0,
"content": "is the feature for input",
"type": "text"
},
{
"bbox": [
265,
191,
275,
201
],
"score": 0.8,
"content": "X",
"type": "inline_equation"
},
{
"bbox": [
276,
185,
294,
208
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
294,
190,
324,
204
],
"score": 0.93,
"content": "K _ { X _ { t } X _ { s } } ^ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
324,
185,
415,
208
],
"score": 1.0,
"content": "is the Gram matrix of",
"type": "text"
},
{
"bbox": [
416,
191,
429,
201
],
"score": 0.89,
"content": "X _ { t }",
"type": "inline_equation"
},
{
"bbox": [
429,
185,
447,
208
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
447,
191,
461,
201
],
"score": 0.88,
"content": "X _ { s }",
"type": "inline_equation"
},
{
"bbox": [
461,
185,
508,
208
],
"score": 1.0,
"content": ". FRePo is",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
201,
505,
214
],
"spans": [
{
"bbox": [
106,
201,
505,
214
],
"score": 1.0,
"content": "analogous to 1-step TBPTT as it computes the meta-gradient at each step while performing the online",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
211,
506,
225
],
"spans": [
{
"bbox": [
105,
211,
506,
225
],
"score": 1.0,
"content": "model update. However, instead of backpropagating through the inner optimization, FRePo computes",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
223,
333,
236
],
"spans": [
{
"bbox": [
106,
223,
333,
236
],
"score": 1.0,
"content": "the meta-gradient through a kernel and feature extractor.",
"type": "text"
}
],
"index": 8
}
],
"index": 5.5
}
],
"index": 3.25
},
{
"type": "title",
"bbox": [
107,
253,
166,
266
],
"lines": [
{
"bbox": [
104,
251,
168,
269
],
"spans": [
{
"bbox": [
104,
251,
168,
269
],
"score": 1.0,
"content": "2 Method",
"type": "text"
}
],
"index": 9
}
],
"index": 9
},
{
"type": "title",
"bbox": [
107,
277,
317,
289
],
"lines": [
{
"bbox": [
105,
275,
318,
292
],
"spans": [
{
"bbox": [
105,
275,
318,
292
],
"score": 1.0,
"content": "2.1 Dataset Distillation as Bi-level Optimization",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
106,
296,
506,
399
],
"lines": [
{
"bbox": [
105,
295,
506,
312
],
"spans": [
{
"bbox": [
105,
295,
266,
312
],
"score": 1.0,
"content": "Suppose we have a large labeled dataset",
"type": "text"
},
{
"bbox": [
266,
296,
403,
311
],
"score": 0.93,
"content": "\\mathcal { T } = \\left\\{ \\left( \\mathbf { x } _ { 1 } , \\mathbf { y } _ { 1 } \\right) , \\dotsc , \\left( \\mathbf { x } _ { | T | } , \\mathbf { y } _ { | T | } \\right) \\right\\}",
"type": "inline_equation"
},
{
"bbox": [
404,
295,
425,
312
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
425,
298,
439,
309
],
"score": 0.88,
"content": "| \\tau |",
"type": "inline_equation"
},
{
"bbox": [
440,
295,
506,
312
],
"score": 1.0,
"content": "image and label` ˘(",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
104,
309,
504,
326
],
"spans": [
{
"bbox": [
104,
309,
368,
326
],
"score": 1.0,
"content": "pairs. Dataset distillation aims to learn a small synthetic dataset",
"type": "text"
},
{
"bbox": [
368,
309,
504,
324
],
"score": 0.9,
"content": "\\mathcal { S } = \\left\\{ ( \\mathbf { x } _ { 1 } , \\mathbf { y } _ { 1 } ) , \\dotsc , \\left( \\mathbf { x } _ { | S | } , \\mathbf { y } _ { | S | } \\right) \\right\\}",
"type": "inline_equation"
}
],
"index": 12
},
{
"bbox": [
105,
322,
506,
334
],
"spans": [
{
"bbox": [
105,
322,
278,
334
],
"score": 1.0,
"content": "that preserves most of the information in",
"type": "text"
},
{
"bbox": [
278,
322,
288,
332
],
"score": 0.8,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
288,
322,
506,
334
],
"score": 1.0,
"content": ". We train several neural networks parameterized by",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
107,
332,
506,
345
],
"spans": [
{
"bbox": [
107,
333,
113,
342
],
"score": 0.78,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
113,
332,
175,
345
],
"score": 1.0,
"content": "on the dataset",
"type": "text"
},
{
"bbox": [
175,
333,
183,
342
],
"score": 0.8,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
184,
332,
341,
345
],
"score": 1.0,
"content": "and then compute the validation loss",
"type": "text"
},
{
"bbox": [
341,
332,
412,
344
],
"score": 0.92,
"content": "\\mathcal { L } ( \\mathcal { A } l g \\left( \\theta , \\mathcal { S } \\right) , \\mathcal { T } )",
"type": "inline_equation"
},
{
"bbox": [
412,
332,
493,
345
],
"score": 1.0,
"content": "on the real dataset",
"type": "text"
},
{
"bbox": [
493,
333,
502,
343
],
"score": 0.8,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
503,
332,
506,
345
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
343,
506,
356
],
"spans": [
{
"bbox": [
105,
343,
133,
356
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
343,
176,
355
],
"score": 0.93,
"content": "{ \\mathcal { A } } l g \\left( \\theta , S \\right)",
"type": "inline_equation"
},
{
"bbox": [
176,
343,
451,
356
],
"score": 1.0,
"content": "is the neural network parameters optimized by a learning algorithm",
"type": "text"
},
{
"bbox": [
452,
344,
468,
354
],
"score": 0.85,
"content": "\\mathcal { A } g",
"type": "inline_equation"
},
{
"bbox": [
469,
343,
506,
356
],
"score": 1.0,
"content": "with the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
354,
506,
367
],
"spans": [
{
"bbox": [
105,
354,
187,
367
],
"score": 1.0,
"content": "model initialization",
"type": "text"
},
{
"bbox": [
187,
355,
193,
364
],
"score": 0.77,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
194,
354,
276,
367
],
"score": 1.0,
"content": "and distilled dataset",
"type": "text"
},
{
"bbox": [
276,
355,
284,
364
],
"score": 0.79,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
285,
354,
416,
367
],
"score": 1.0,
"content": "as its inputs. The validation loss",
"type": "text"
},
{
"bbox": [
416,
355,
487,
366
],
"score": 0.91,
"content": "\\mathcal { L } ( \\mathcal { A } l g \\left( \\theta , S \\right) , \\mathcal { T } )",
"type": "inline_equation"
},
{
"bbox": [
488,
354,
506,
367
],
"score": 1.0,
"content": "is a",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
104,
363,
506,
379
],
"spans": [
{
"bbox": [
104,
363,
506,
379
],
"score": 1.0,
"content": "noisy objective with the stochasticity coming from random model initialization and inner learning",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
375,
506,
389
],
"spans": [
{
"bbox": [
105,
375,
506,
389
],
"score": 1.0,
"content": "algorithm. Thus, we are interested in minimizing the expected value of this loss, which we denote it",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
387,
482,
400
],
"spans": [
{
"bbox": [
105,
387,
117,
400
],
"score": 1.0,
"content": "as",
"type": "text"
},
{
"bbox": [
118,
387,
140,
399
],
"score": 0.92,
"content": "F ( S )",
"type": "inline_equation"
},
{
"bbox": [
141,
387,
482,
400
],
"score": 1.0,
"content": ". We formulate the dataset distillation as the following bi-level optimization problem.",
"type": "text"
}
],
"index": 19
}
],
"index": 15
},
{
"type": "interline_equation",
"bbox": [
165,
403,
445,
440
],
"lines": [
{
"bbox": [
165,
403,
445,
440
],
"spans": [
{
"bbox": [
165,
403,
445,
440
],
"score": 0.93,
"content": "\\overbrace { \\mathcal { S } ^ { * } : = \\mathop { \\mathrm { a r g m i n } } _ { \\mathcal { S } } F ( \\mathcal { S } ) } ^ { o u t e r - l e v e l } , \\mathrm { w h e r e } F ( \\mathcal { S } ) = \\mathbb { E } _ { \\theta \\sim P _ { \\theta } } \\biggl [ \\mathcal { L } \\Bigl ( \\overbrace { \\mathcal { A } l g \\left( \\theta , \\mathcal { S } \\right) } ^ { i n n e r - l e v e l } , \\ T \\Bigr ) \\biggr ] .",
"type": "interline_equation",
"image_path": "5d90a6e2a97898bb43c032413b0dda3e03ac154d8a7d9d7a921494e9f9e4b013.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
165,
403,
445,
415.3333333333333
],
"spans": [],
"index": 20
},
{
"bbox": [
165,
415.3333333333333,
445,
427.66666666666663
],
"spans": [],
"index": 21
},
{
"bbox": [
165,
427.66666666666663,
445,
439.99999999999994
],
"spans": [],
"index": 22
}
]
},
{
"type": "text",
"bbox": [
106,
442,
505,
531
],
"lines": [
{
"bbox": [
106,
443,
506,
455
],
"spans": [
{
"bbox": [
106,
443,
415,
455
],
"score": 1.0,
"content": "In this bi-level setup, the outer loop optimizes the distilled data to minimize",
"type": "text"
},
{
"bbox": [
415,
443,
438,
455
],
"score": 0.91,
"content": "F ( S )",
"type": "inline_equation"
},
{
"bbox": [
438,
443,
506,
455
],
"score": 1.0,
"content": ", while the inner",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
453,
505,
466
],
"spans": [
{
"bbox": [
106,
453,
339,
466
],
"score": 1.0,
"content": "loop trains a neural network using the learning algorithm,",
"type": "text"
},
{
"bbox": [
340,
454,
357,
465
],
"score": 0.88,
"content": "\\mathcal { A } g",
"type": "inline_equation"
},
{
"bbox": [
357,
453,
505,
466
],
"score": 1.0,
"content": ", to minimize the training loss on the",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
464,
506,
478
],
"spans": [
{
"bbox": [
105,
464,
160,
478
],
"score": 1.0,
"content": "distilled data",
"type": "text"
},
{
"bbox": [
161,
465,
169,
475
],
"score": 0.77,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
169,
464,
506,
478
],
"score": 1.0,
"content": ". From the meta-learning perspective, the task is defined by the model initialization",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
107,
476,
505,
488
],
"spans": [
{
"bbox": [
107,
477,
113,
486
],
"score": 0.76,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
113,
476,
274,
488
],
"score": 1.0,
"content": ", and we want to learn a meta-parameter",
"type": "text"
},
{
"bbox": [
275,
476,
283,
486
],
"score": 0.79,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
283,
476,
505,
488
],
"score": 1.0,
"content": "that generalizes well to different models sampled from",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
486,
506,
499
],
"spans": [
{
"bbox": [
105,
486,
201,
499
],
"score": 1.0,
"content": "the model distributions",
"type": "text"
},
{
"bbox": [
202,
487,
213,
498
],
"score": 0.88,
"content": "P _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
214,
486,
419,
499
],
"score": 1.0,
"content": ". During learning, we optimize the meta-parameter",
"type": "text"
},
{
"bbox": [
420,
487,
428,
497
],
"score": 0.8,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
428,
486,
506,
499
],
"score": 1.0,
"content": "by minimizing the",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
498,
505,
510
],
"spans": [
{
"bbox": [
105,
498,
180,
509
],
"score": 1.0,
"content": "meta-training loss",
"type": "text"
},
{
"bbox": [
181,
498,
204,
510
],
"score": 0.92,
"content": "F ( S )",
"type": "inline_equation"
},
{
"bbox": [
204,
498,
479,
509
],
"score": 1.0,
"content": ". In contrast, at meta-test time, we train a new model from scratch on",
"type": "text"
},
{
"bbox": [
479,
498,
487,
507
],
"score": 0.8,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
487,
498,
505,
509
],
"score": 1.0,
"content": "and",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
508,
505,
521
],
"spans": [
{
"bbox": [
105,
508,
505,
521
],
"score": 1.0,
"content": "evaluate the trained model on a held-out real dataset. This meta-test performance reflects the quality",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
519,
188,
531
],
"spans": [
{
"bbox": [
106,
519,
188,
531
],
"score": 1.0,
"content": "of the distilled data.",
"type": "text"
}
],
"index": 30
}
],
"index": 26.5
},
{
"type": "title",
"bbox": [
108,
542,
447,
555
],
"lines": [
{
"bbox": [
104,
541,
445,
558
],
"spans": [
{
"bbox": [
104,
541,
411,
558
],
"score": 1.0,
"content": "2.2 Dataset Distillation using Neural Feature Regression with Pooling",
"type": "text"
},
{
"bbox": [
411,
543,
445,
554
],
"score": 0.25,
"content": "\\mathbf { ( F R e P 0 ) }",
"type": "inline_equation"
}
],
"index": 31
}
],
"index": 31
},
{
"type": "text",
"bbox": [
107,
563,
505,
641
],
"lines": [
{
"bbox": [
105,
563,
505,
577
],
"spans": [
{
"bbox": [
105,
563,
417,
577
],
"score": 1.0,
"content": "The outer-level problem can be solved using gradient-based methods of the form",
"type": "text"
},
{
"bbox": [
418,
563,
501,
576
],
"score": 0.89,
"content": "\\mathcal { S } \\gets \\mathcal { S } \\ – \\alpha \\nabla _ { \\mathcal { S } } F ( \\mathcal { S } )",
"type": "inline_equation"
},
{
"bbox": [
502,
563,
505,
577
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
574,
506,
587
],
"spans": [
{
"bbox": [
105,
574,
134,
587
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
576,
142,
584
],
"score": 0.77,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
142,
574,
330,
587
],
"score": 1.0,
"content": "is the learning rate for the distilled data and",
"type": "text"
},
{
"bbox": [
330,
574,
368,
586
],
"score": 0.91,
"content": "\\nabla _ { S } F ( S )",
"type": "inline_equation"
},
{
"bbox": [
369,
574,
506,
587
],
"score": 1.0,
"content": "is the meta-gradient [27]. For a",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
585,
505,
598
],
"spans": [
{
"bbox": [
105,
585,
176,
598
],
"score": 1.0,
"content": "particular model",
"type": "text"
},
{
"bbox": [
177,
586,
183,
595
],
"score": 0.71,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
183,
585,
346,
598
],
"score": 1.0,
"content": ", the meta-gradient can be expressed as",
"type": "text"
},
{
"bbox": [
346,
585,
433,
597
],
"score": 0.93,
"content": "\\nabla _ { \\mathcal { S } } \\hat { \\mathcal { L } } \\left( \\mathcal { A } l g \\left( \\theta , \\mathcal { S } \\right) , \\mathcal { T } \\right)",
"type": "inline_equation"
},
{
"bbox": [
433,
585,
505,
598
],
"score": 1.0,
"content": ". Computing this",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
596,
505,
609
],
"spans": [
{
"bbox": [
105,
596,
387,
609
],
"score": 1.0,
"content": "meta-gradient requires differentiating through inner optimization. If",
"type": "text"
},
{
"bbox": [
387,
597,
404,
608
],
"score": 0.83,
"content": "\\mathcal { A } \\boldsymbol { { l } } _ { g }",
"type": "inline_equation"
},
{
"bbox": [
405,
596,
505,
609
],
"score": 1.0,
"content": "is an iterative algorithm",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
606,
506,
621
],
"spans": [
{
"bbox": [
105,
606,
506,
621
],
"score": 1.0,
"content": "like gradient descent, then backpropagating through the unrolled computation graph [14] can be a",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
617,
505,
632
],
"spans": [
{
"bbox": [
105,
617,
505,
632
],
"score": 1.0,
"content": "solution. However, this type of unrolled optimization introduces significant computation and memory",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
628,
451,
642
],
"spans": [
{
"bbox": [
105,
628,
451,
642
],
"score": 1.0,
"content": "overhead, as the whole training trajectory needs to be stored in memory (Figure 2(b)).",
"type": "text"
}
],
"index": 38
}
],
"index": 35
},
{
"type": "text",
"bbox": [
107,
645,
505,
722
],
"lines": [
{
"bbox": [
105,
644,
506,
658
],
"spans": [
{
"bbox": [
105,
644,
506,
658
],
"score": 1.0,
"content": "Traditionally, these issues are alleviated with truncated backpropagation through time (TBPTT)",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
655,
507,
670
],
"spans": [
{
"bbox": [
105,
655,
507,
670
],
"score": 1.0,
"content": "[28–30]. Instead of backpropagating through an entire unrolled sequence, TBPTT performs backprop-",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
667,
506,
680
],
"spans": [
{
"bbox": [
105,
667,
506,
680
],
"score": 1.0,
"content": "agation for each subsequence separately. It is efficient because its time and memory complexity scale",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
678,
505,
691
],
"spans": [
{
"bbox": [
105,
678,
505,
691
],
"score": 1.0,
"content": "linearly with respect to the truncation steps. However, truncation may yield highly biased gradients",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
687,
505,
703
],
"spans": [
{
"bbox": [
105,
687,
505,
703
],
"score": 1.0,
"content": "that severely impact training. To mitigate this truncation bias [14], we consider training only the top",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
700,
506,
712
],
"spans": [
{
"bbox": [
105,
700,
506,
712
],
"score": 1.0,
"content": "layer of a network to convergence. The key insight is that the data helpful for training the output",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
711,
505,
723
],
"spans": [
{
"bbox": [
105,
711,
505,
723
],
"score": 1.0,
"content": "layer can also help train the whole network. Thus, we decompose the neural network into a feature",
"type": "text"
}
],
"index": 45
}
],
"index": 42
}
],
"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": [
106,
70,
506,
160
],
"blocks": [
{
"type": "image_body",
"bbox": [
106,
70,
506,
160
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
70,
506,
160
],
"spans": [
{
"bbox": [
106,
70,
506,
160
],
"score": 0.964,
"type": "image",
"image_path": "bd356bac9e8a26f60ac46a3e8c2299fdbc56eeb9580bd6cdb628ffbec6e98257.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
106,
70,
506,
100.0
],
"spans": [],
"index": 0
},
{
"bbox": [
106,
100.0,
506,
130.0
],
"spans": [],
"index": 1
},
{
"bbox": [
106,
130.0,
506,
160.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
167,
506,
235
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
166,
506,
181
],
"spans": [
{
"bbox": [
105,
166,
362,
181
],
"score": 1.0,
"content": "Figure 2: Comparison of FRePo and Unrolled Optimization.",
"type": "text"
},
{
"bbox": [
362,
168,
370,
178
],
"score": 0.44,
"content": "S",
"type": "inline_equation"
},
{
"bbox": [
370,
166,
374,
181
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
375,
168,
388,
178
],
"score": 0.73,
"content": "X _ { s }",
"type": "inline_equation"
},
{
"bbox": [
388,
166,
392,
181
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
393,
168,
404,
178
],
"score": 0.78,
"content": "Y _ { s }",
"type": "inline_equation"
},
{
"bbox": [
404,
166,
506,
181
],
"score": 1.0,
"content": "are the distilled dataset,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
177,
506,
193
],
"spans": [
{
"bbox": [
105,
177,
184,
193
],
"score": 1.0,
"content": "images and labels.",
"type": "text"
},
{
"bbox": [
185,
180,
193,
190
],
"score": 0.74,
"content": "\\mathcal { L }",
"type": "inline_equation"
},
{
"bbox": [
193,
177,
311,
193
],
"score": 1.0,
"content": "is the meta-training loss and",
"type": "text"
},
{
"bbox": [
311,
178,
328,
190
],
"score": 0.85,
"content": "\\dot { \\theta } ^ { ( k ) }",
"type": "inline_equation"
},
{
"bbox": [
328,
177,
331,
193
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
332,
178,
349,
191
],
"score": 0.87,
"content": "g ^ { ( k ) }",
"type": "inline_equation"
},
{
"bbox": [
350,
177,
506,
193
],
"score": 1.0,
"content": "are the model parameter and gradient",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
103,
185,
508,
208
],
"spans": [
{
"bbox": [
103,
185,
135,
208
],
"score": 1.0,
"content": "at step",
"type": "text"
},
{
"bbox": [
136,
191,
142,
201
],
"score": 0.58,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
142,
185,
146,
208
],
"score": 1.0,
"content": ".",
"type": "text"
},
{
"bbox": [
147,
190,
171,
203
],
"score": 0.89,
"content": "f ( X )",
"type": "inline_equation"
},
{
"bbox": [
171,
185,
265,
208
],
"score": 1.0,
"content": "is the feature for input",
"type": "text"
},
{
"bbox": [
265,
191,
275,
201
],
"score": 0.8,
"content": "X",
"type": "inline_equation"
},
{
"bbox": [
276,
185,
294,
208
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
294,
190,
324,
204
],
"score": 0.93,
"content": "K _ { X _ { t } X _ { s } } ^ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
324,
185,
415,
208
],
"score": 1.0,
"content": "is the Gram matrix of",
"type": "text"
},
{
"bbox": [
416,
191,
429,
201
],
"score": 0.89,
"content": "X _ { t }",
"type": "inline_equation"
},
{
"bbox": [
429,
185,
447,
208
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
447,
191,
461,
201
],
"score": 0.88,
"content": "X _ { s }",
"type": "inline_equation"
},
{
"bbox": [
461,
185,
508,
208
],
"score": 1.0,
"content": ". FRePo is",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
201,
505,
214
],
"spans": [
{
"bbox": [
106,
201,
505,
214
],
"score": 1.0,
"content": "analogous to 1-step TBPTT as it computes the meta-gradient at each step while performing the online",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
211,
506,
225
],
"spans": [
{
"bbox": [
105,
211,
506,
225
],
"score": 1.0,
"content": "model update. However, instead of backpropagating through the inner optimization, FRePo computes",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
223,
333,
236
],
"spans": [
{
"bbox": [
106,
223,
333,
236
],
"score": 1.0,
"content": "the meta-gradient through a kernel and feature extractor.",
"type": "text"
}
],
"index": 8
}
],
"index": 5.5
}
],
"index": 3.25
},
{
"type": "title",
"bbox": [
107,
253,
166,
266
],
"lines": [
{
"bbox": [
104,
251,
168,
269
],
"spans": [
{
"bbox": [
104,
251,
168,
269
],
"score": 1.0,
"content": "2 Method",
"type": "text"
}
],
"index": 9
}
],
"index": 9
},
{
"type": "title",
"bbox": [
107,
277,
317,
289
],
"lines": [
{
"bbox": [
105,
275,
318,
292
],
"spans": [
{
"bbox": [
105,
275,
318,
292
],
"score": 1.0,
"content": "2.1 Dataset Distillation as Bi-level Optimization",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
106,
296,
506,
399
],
"lines": [
{
"bbox": [
105,
295,
506,
312
],
"spans": [
{
"bbox": [
105,
295,
266,
312
],
"score": 1.0,
"content": "Suppose we have a large labeled dataset",
"type": "text"
},
{
"bbox": [
266,
296,
403,
311
],
"score": 0.93,
"content": "\\mathcal { T } = \\left\\{ \\left( \\mathbf { x } _ { 1 } , \\mathbf { y } _ { 1 } \\right) , \\dotsc , \\left( \\mathbf { x } _ { | T | } , \\mathbf { y } _ { | T | } \\right) \\right\\}",
"type": "inline_equation"
},
{
"bbox": [
404,
295,
425,
312
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
425,
298,
439,
309
],
"score": 0.88,
"content": "| \\tau |",
"type": "inline_equation"
},
{
"bbox": [
440,
295,
506,
312
],
"score": 1.0,
"content": "image and label` ˘(",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
104,
309,
504,
326
],
"spans": [
{
"bbox": [
104,
309,
368,
326
],
"score": 1.0,
"content": "pairs. Dataset distillation aims to learn a small synthetic dataset",
"type": "text"
},
{
"bbox": [
368,
309,
504,
324
],
"score": 0.9,
"content": "\\mathcal { S } = \\left\\{ ( \\mathbf { x } _ { 1 } , \\mathbf { y } _ { 1 } ) , \\dotsc , \\left( \\mathbf { x } _ { | S | } , \\mathbf { y } _ { | S | } \\right) \\right\\}",
"type": "inline_equation"
}
],
"index": 12
},
{
"bbox": [
105,
322,
506,
334
],
"spans": [
{
"bbox": [
105,
322,
278,
334
],
"score": 1.0,
"content": "that preserves most of the information in",
"type": "text"
},
{
"bbox": [
278,
322,
288,
332
],
"score": 0.8,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
288,
322,
506,
334
],
"score": 1.0,
"content": ". We train several neural networks parameterized by",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
107,
332,
506,
345
],
"spans": [
{
"bbox": [
107,
333,
113,
342
],
"score": 0.78,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
113,
332,
175,
345
],
"score": 1.0,
"content": "on the dataset",
"type": "text"
},
{
"bbox": [
175,
333,
183,
342
],
"score": 0.8,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
184,
332,
341,
345
],
"score": 1.0,
"content": "and then compute the validation loss",
"type": "text"
},
{
"bbox": [
341,
332,
412,
344
],
"score": 0.92,
"content": "\\mathcal { L } ( \\mathcal { A } l g \\left( \\theta , \\mathcal { S } \\right) , \\mathcal { T } )",
"type": "inline_equation"
},
{
"bbox": [
412,
332,
493,
345
],
"score": 1.0,
"content": "on the real dataset",
"type": "text"
},
{
"bbox": [
493,
333,
502,
343
],
"score": 0.8,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
503,
332,
506,
345
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
343,
506,
356
],
"spans": [
{
"bbox": [
105,
343,
133,
356
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
343,
176,
355
],
"score": 0.93,
"content": "{ \\mathcal { A } } l g \\left( \\theta , S \\right)",
"type": "inline_equation"
},
{
"bbox": [
176,
343,
451,
356
],
"score": 1.0,
"content": "is the neural network parameters optimized by a learning algorithm",
"type": "text"
},
{
"bbox": [
452,
344,
468,
354
],
"score": 0.85,
"content": "\\mathcal { A } g",
"type": "inline_equation"
},
{
"bbox": [
469,
343,
506,
356
],
"score": 1.0,
"content": "with the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
354,
506,
367
],
"spans": [
{
"bbox": [
105,
354,
187,
367
],
"score": 1.0,
"content": "model initialization",
"type": "text"
},
{
"bbox": [
187,
355,
193,
364
],
"score": 0.77,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
194,
354,
276,
367
],
"score": 1.0,
"content": "and distilled dataset",
"type": "text"
},
{
"bbox": [
276,
355,
284,
364
],
"score": 0.79,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
285,
354,
416,
367
],
"score": 1.0,
"content": "as its inputs. The validation loss",
"type": "text"
},
{
"bbox": [
416,
355,
487,
366
],
"score": 0.91,
"content": "\\mathcal { L } ( \\mathcal { A } l g \\left( \\theta , S \\right) , \\mathcal { T } )",
"type": "inline_equation"
},
{
"bbox": [
488,
354,
506,
367
],
"score": 1.0,
"content": "is a",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
104,
363,
506,
379
],
"spans": [
{
"bbox": [
104,
363,
506,
379
],
"score": 1.0,
"content": "noisy objective with the stochasticity coming from random model initialization and inner learning",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
375,
506,
389
],
"spans": [
{
"bbox": [
105,
375,
506,
389
],
"score": 1.0,
"content": "algorithm. Thus, we are interested in minimizing the expected value of this loss, which we denote it",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
387,
482,
400
],
"spans": [
{
"bbox": [
105,
387,
117,
400
],
"score": 1.0,
"content": "as",
"type": "text"
},
{
"bbox": [
118,
387,
140,
399
],
"score": 0.92,
"content": "F ( S )",
"type": "inline_equation"
},
{
"bbox": [
141,
387,
482,
400
],
"score": 1.0,
"content": ". We formulate the dataset distillation as the following bi-level optimization problem.",
"type": "text"
}
],
"index": 19
}
],
"index": 15,
"bbox_fs": [
104,
295,
506,
400
]
},
{
"type": "interline_equation",
"bbox": [
165,
403,
445,
440
],
"lines": [
{
"bbox": [
165,
403,
445,
440
],
"spans": [
{
"bbox": [
165,
403,
445,
440
],
"score": 0.93,
"content": "\\overbrace { \\mathcal { S } ^ { * } : = \\mathop { \\mathrm { a r g m i n } } _ { \\mathcal { S } } F ( \\mathcal { S } ) } ^ { o u t e r - l e v e l } , \\mathrm { w h e r e } F ( \\mathcal { S } ) = \\mathbb { E } _ { \\theta \\sim P _ { \\theta } } \\biggl [ \\mathcal { L } \\Bigl ( \\overbrace { \\mathcal { A } l g \\left( \\theta , \\mathcal { S } \\right) } ^ { i n n e r - l e v e l } , \\ T \\Bigr ) \\biggr ] .",
"type": "interline_equation",
"image_path": "5d90a6e2a97898bb43c032413b0dda3e03ac154d8a7d9d7a921494e9f9e4b013.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
165,
403,
445,
415.3333333333333
],
"spans": [],
"index": 20
},
{
"bbox": [
165,
415.3333333333333,
445,
427.66666666666663
],
"spans": [],
"index": 21
},
{
"bbox": [
165,
427.66666666666663,
445,
439.99999999999994
],
"spans": [],
"index": 22
}
]
},
{
"type": "text",
"bbox": [
106,
442,
505,
531
],
"lines": [
{
"bbox": [
106,
443,
506,
455
],
"spans": [
{
"bbox": [
106,
443,
415,
455
],
"score": 1.0,
"content": "In this bi-level setup, the outer loop optimizes the distilled data to minimize",
"type": "text"
},
{
"bbox": [
415,
443,
438,
455
],
"score": 0.91,
"content": "F ( S )",
"type": "inline_equation"
},
{
"bbox": [
438,
443,
506,
455
],
"score": 1.0,
"content": ", while the inner",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
453,
505,
466
],
"spans": [
{
"bbox": [
106,
453,
339,
466
],
"score": 1.0,
"content": "loop trains a neural network using the learning algorithm,",
"type": "text"
},
{
"bbox": [
340,
454,
357,
465
],
"score": 0.88,
"content": "\\mathcal { A } g",
"type": "inline_equation"
},
{
"bbox": [
357,
453,
505,
466
],
"score": 1.0,
"content": ", to minimize the training loss on the",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
464,
506,
478
],
"spans": [
{
"bbox": [
105,
464,
160,
478
],
"score": 1.0,
"content": "distilled data",
"type": "text"
},
{
"bbox": [
161,
465,
169,
475
],
"score": 0.77,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
169,
464,
506,
478
],
"score": 1.0,
"content": ". From the meta-learning perspective, the task is defined by the model initialization",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
107,
476,
505,
488
],
"spans": [
{
"bbox": [
107,
477,
113,
486
],
"score": 0.76,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
113,
476,
274,
488
],
"score": 1.0,
"content": ", and we want to learn a meta-parameter",
"type": "text"
},
{
"bbox": [
275,
476,
283,
486
],
"score": 0.79,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
283,
476,
505,
488
],
"score": 1.0,
"content": "that generalizes well to different models sampled from",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
486,
506,
499
],
"spans": [
{
"bbox": [
105,
486,
201,
499
],
"score": 1.0,
"content": "the model distributions",
"type": "text"
},
{
"bbox": [
202,
487,
213,
498
],
"score": 0.88,
"content": "P _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
214,
486,
419,
499
],
"score": 1.0,
"content": ". During learning, we optimize the meta-parameter",
"type": "text"
},
{
"bbox": [
420,
487,
428,
497
],
"score": 0.8,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
428,
486,
506,
499
],
"score": 1.0,
"content": "by minimizing the",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
498,
505,
510
],
"spans": [
{
"bbox": [
105,
498,
180,
509
],
"score": 1.0,
"content": "meta-training loss",
"type": "text"
},
{
"bbox": [
181,
498,
204,
510
],
"score": 0.92,
"content": "F ( S )",
"type": "inline_equation"
},
{
"bbox": [
204,
498,
479,
509
],
"score": 1.0,
"content": ". In contrast, at meta-test time, we train a new model from scratch on",
"type": "text"
},
{
"bbox": [
479,
498,
487,
507
],
"score": 0.8,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
487,
498,
505,
509
],
"score": 1.0,
"content": "and",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
508,
505,
521
],
"spans": [
{
"bbox": [
105,
508,
505,
521
],
"score": 1.0,
"content": "evaluate the trained model on a held-out real dataset. This meta-test performance reflects the quality",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
519,
188,
531
],
"spans": [
{
"bbox": [
106,
519,
188,
531
],
"score": 1.0,
"content": "of the distilled data.",
"type": "text"
}
],
"index": 30
}
],
"index": 26.5,
"bbox_fs": [
105,
443,
506,
531
]
},
{
"type": "title",
"bbox": [
108,
542,
447,
555
],
"lines": [
{
"bbox": [
104,
541,
445,
558
],
"spans": [
{
"bbox": [
104,
541,
411,
558
],
"score": 1.0,
"content": "2.2 Dataset Distillation using Neural Feature Regression with Pooling",
"type": "text"
},
{
"bbox": [
411,
543,
445,
554
],
"score": 0.25,
"content": "\\mathbf { ( F R e P 0 ) }",
"type": "inline_equation"
}
],
"index": 31
}
],
"index": 31
},
{
"type": "text",
"bbox": [
107,
563,
505,
641
],
"lines": [
{
"bbox": [
105,
563,
505,
577
],
"spans": [
{
"bbox": [
105,
563,
417,
577
],
"score": 1.0,
"content": "The outer-level problem can be solved using gradient-based methods of the form",
"type": "text"
},
{
"bbox": [
418,
563,
501,
576
],
"score": 0.89,
"content": "\\mathcal { S } \\gets \\mathcal { S } \\ – \\alpha \\nabla _ { \\mathcal { S } } F ( \\mathcal { S } )",
"type": "inline_equation"
},
{
"bbox": [
502,
563,
505,
577
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
574,
506,
587
],
"spans": [
{
"bbox": [
105,
574,
134,
587
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
576,
142,
584
],
"score": 0.77,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
142,
574,
330,
587
],
"score": 1.0,
"content": "is the learning rate for the distilled data and",
"type": "text"
},
{
"bbox": [
330,
574,
368,
586
],
"score": 0.91,
"content": "\\nabla _ { S } F ( S )",
"type": "inline_equation"
},
{
"bbox": [
369,
574,
506,
587
],
"score": 1.0,
"content": "is the meta-gradient [27]. For a",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
585,
505,
598
],
"spans": [
{
"bbox": [
105,
585,
176,
598
],
"score": 1.0,
"content": "particular model",
"type": "text"
},
{
"bbox": [
177,
586,
183,
595
],
"score": 0.71,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
183,
585,
346,
598
],
"score": 1.0,
"content": ", the meta-gradient can be expressed as",
"type": "text"
},
{
"bbox": [
346,
585,
433,
597
],
"score": 0.93,
"content": "\\nabla _ { \\mathcal { S } } \\hat { \\mathcal { L } } \\left( \\mathcal { A } l g \\left( \\theta , \\mathcal { S } \\right) , \\mathcal { T } \\right)",
"type": "inline_equation"
},
{
"bbox": [
433,
585,
505,
598
],
"score": 1.0,
"content": ". Computing this",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
596,
505,
609
],
"spans": [
{
"bbox": [
105,
596,
387,
609
],
"score": 1.0,
"content": "meta-gradient requires differentiating through inner optimization. If",
"type": "text"
},
{
"bbox": [
387,
597,
404,
608
],
"score": 0.83,
"content": "\\mathcal { A } \\boldsymbol { { l } } _ { g }",
"type": "inline_equation"
},
{
"bbox": [
405,
596,
505,
609
],
"score": 1.0,
"content": "is an iterative algorithm",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
606,
506,
621
],
"spans": [
{
"bbox": [
105,
606,
506,
621
],
"score": 1.0,
"content": "like gradient descent, then backpropagating through the unrolled computation graph [14] can be a",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
617,
505,
632
],
"spans": [
{
"bbox": [
105,
617,
505,
632
],
"score": 1.0,
"content": "solution. However, this type of unrolled optimization introduces significant computation and memory",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
628,
451,
642
],
"spans": [
{
"bbox": [
105,
628,
451,
642
],
"score": 1.0,
"content": "overhead, as the whole training trajectory needs to be stored in memory (Figure 2(b)).",
"type": "text"
}
],
"index": 38
}
],
"index": 35,
"bbox_fs": [
105,
563,
506,
642
]
},
{
"type": "text",
"bbox": [
107,
645,
505,
722
],
"lines": [
{
"bbox": [
105,
644,
506,
658
],
"spans": [
{
"bbox": [
105,
644,
506,
658
],
"score": 1.0,
"content": "Traditionally, these issues are alleviated with truncated backpropagation through time (TBPTT)",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
655,
507,
670
],
"spans": [
{
"bbox": [
105,
655,
507,
670
],
"score": 1.0,
"content": "[28–30]. Instead of backpropagating through an entire unrolled sequence, TBPTT performs backprop-",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
667,
506,
680
],
"spans": [
{
"bbox": [
105,
667,
506,
680
],
"score": 1.0,
"content": "agation for each subsequence separately. It is efficient because its time and memory complexity scale",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
678,
505,
691
],
"spans": [
{
"bbox": [
105,
678,
505,
691
],
"score": 1.0,
"content": "linearly with respect to the truncation steps. However, truncation may yield highly biased gradients",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
687,
505,
703
],
"spans": [
{
"bbox": [
105,
687,
505,
703
],
"score": 1.0,
"content": "that severely impact training. To mitigate this truncation bias [14], we consider training only the top",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
700,
506,
712
],
"spans": [
{
"bbox": [
105,
700,
506,
712
],
"score": 1.0,
"content": "layer of a network to convergence. The key insight is that the data helpful for training the output",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
711,
505,
723
],
"spans": [
{
"bbox": [
105,
711,
505,
723
],
"score": 1.0,
"content": "layer can also help train the whole network. Thus, we decompose the neural network into a feature",
"type": "text"
}
],
"index": 45
}
],
"index": 42,
"bbox_fs": [
105,
644,
507,
723
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
113,
88,
404,
99
],
"lines": [
{
"bbox": [
116,
87,
404,
100
],
"spans": [
{
"bbox": [
116,
87,
156,
100
],
"score": 1.0,
"content": "Require:",
"type": "text"
},
{
"bbox": [
157,
88,
166,
99
],
"score": 0.57,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
167,
87,
241,
100
],
"score": 1.0,
"content": ": a labeled dataset;",
"type": "text"
},
{
"bbox": [
242,
90,
250,
98
],
"score": 0.61,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
250,
87,
404,
100
],
"score": 1.0,
"content": ": the learning rate for the distilled data",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
124,
99,
381,
109
],
"lines": [
{
"bbox": [
122,
97,
378,
111
],
"spans": [
{
"bbox": [
122,
97,
318,
111
],
"score": 1.0,
"content": "nitialization: Initialize a labeled distilled dataset",
"type": "text"
},
{
"bbox": [
319,
99,
374,
110
],
"score": 0.91,
"content": "\\boldsymbol { \\mathcal { S } } = \\left( \\boldsymbol { X _ { s } } , \\boldsymbol { Y _ { s } } \\right)",
"type": "inline_equation"
},
{
"bbox": [
375,
97,
378,
111
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 1
}
],
"index": 1
},
{
"type": "text",
"bbox": [
114,
110,
500,
122
],
"lines": [
{
"bbox": [
115,
105,
497,
124
],
"spans": [
{
"bbox": [
115,
105,
270,
124
],
"score": 1.0,
"content": "Initialization: Initialize a model pool",
"type": "text"
},
{
"bbox": [
270,
110,
284,
120
],
"score": 0.79,
"content": "\\mathcal { M }",
"type": "inline_equation"
},
{
"bbox": [
284,
105,
304,
124
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
305,
111,
315,
119
],
"score": 0.74,
"content": "m",
"type": "inline_equation"
},
{
"bbox": [
315,
105,
347,
124
],
"score": 1.0,
"content": "models",
"type": "text"
},
{
"bbox": [
348,
110,
379,
122
],
"score": 0.89,
"content": "\\left\\{ \\boldsymbol { \\theta } _ { i } \\right\\} _ { i = 1 } ^ { m }",
"type": "inline_equation"
},
{
"bbox": [
379,
105,
484,
124
],
"score": 1.0,
"content": "randomly initialized from",
"type": "text"
},
{
"bbox": [
485,
110,
497,
120
],
"score": 0.86,
"content": "P _ { \\theta }",
"type": "inline_equation"
}
],
"index": 2
}
],
"index": 2
},
{
"type": "text",
"bbox": [
111,
123,
440,
199
],
"lines": [
{
"bbox": [
110,
123,
220,
134
],
"spans": [
{
"bbox": [
110,
123,
220,
134
],
"score": 1.0,
"content": "1: while not converged do",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
111,
133,
375,
145
],
"spans": [
{
"bbox": [
111,
133,
121,
145
],
"score": 1.0,
"content": "2:",
"type": "text"
},
{
"bbox": [
133,
133,
337,
145
],
"score": 1.0,
"content": "Ż Sample a model uniformly from the model pool:",
"type": "text"
},
{
"bbox": [
337,
133,
372,
144
],
"score": 0.91,
"content": "\\theta _ { i } \\sim \\mathcal { M }",
"type": "inline_equation"
},
{
"bbox": [
372,
133,
375,
145
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
110,
144,
432,
157
],
"spans": [
{
"bbox": [
110,
145,
121,
156
],
"score": 1.0,
"content": "3:",
"type": "text"
},
{
"bbox": [
132,
144,
373,
157
],
"score": 1.0,
"content": "Ż Sample a target batch uniformly from the labeled dataset:",
"type": "text"
},
{
"bbox": [
374,
144,
428,
156
],
"score": 0.9,
"content": "( X _ { t } , Y _ { t } ) \\sim \\tau",
"type": "inline_equation"
},
{
"bbox": [
429,
144,
432,
157
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
110,
155,
326,
167
],
"spans": [
{
"bbox": [
110,
155,
121,
167
],
"score": 1.0,
"content": "4:",
"type": "text"
},
{
"bbox": [
133,
157,
140,
165
],
"score": 0.27,
"content": "\\triangleright",
"type": "inline_equation"
},
{
"bbox": [
141,
155,
268,
167
],
"score": 1.0,
"content": "Compute the meta-training loss",
"type": "text"
},
{
"bbox": [
269,
156,
277,
165
],
"score": 0.79,
"content": "\\mathcal { L }",
"type": "inline_equation"
},
{
"bbox": [
277,
155,
326,
167
],
"score": 1.0,
"content": "using Eq. 2",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
110,
165,
440,
179
],
"spans": [
{
"bbox": [
110,
167,
121,
178
],
"score": 1.0,
"content": "5:",
"type": "text"
},
{
"bbox": [
133,
165,
240,
179
],
"score": 1.0,
"content": "Ż Update the distilled data",
"type": "text"
},
{
"bbox": [
241,
166,
249,
176
],
"score": 0.37,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
250,
166,
340,
178
],
"score": 0.86,
"content": "\\mathrm { : } ~ X _ { s } \\gets X _ { s } - \\alpha \\nabla _ { X _ { s } } \\mathcal { L }",
"type": "inline_equation"
},
{
"bbox": [
340,
165,
360,
179
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
360,
166,
440,
178
],
"score": 0.9,
"content": "Y _ { s } \\gets Y _ { s } - \\alpha \\nabla _ { Y _ { s } } \\mathcal { L }",
"type": "inline_equation"
}
],
"index": 7
},
{
"bbox": [
110,
175,
390,
191
],
"spans": [
{
"bbox": [
110,
177,
121,
189
],
"score": 1.0,
"content": "6:",
"type": "text"
},
{
"bbox": [
133,
179,
140,
187
],
"score": 0.35,
"content": "\\triangleright",
"type": "inline_equation"
},
{
"bbox": [
141,
175,
206,
191
],
"score": 1.0,
"content": "Train the model",
"type": "text"
},
{
"bbox": [
207,
177,
216,
188
],
"score": 0.88,
"content": "\\theta _ { i }",
"type": "inline_equation"
},
{
"bbox": [
216,
175,
329,
191
],
"score": 1.0,
"content": "on the current distilled data",
"type": "text"
},
{
"bbox": [
329,
178,
337,
187
],
"score": 0.8,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
337,
175,
390,
191
],
"score": 1.0,
"content": "for one step.",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
111,
187,
435,
201
],
"spans": [
{
"bbox": [
111,
188,
121,
199
],
"score": 1.0,
"content": "7:",
"type": "text"
},
{
"bbox": [
133,
190,
141,
198
],
"score": 0.27,
"content": "\\triangleright",
"type": "inline_equation"
},
{
"bbox": [
141,
187,
231,
201
],
"score": 1.0,
"content": "Reinitialize the model",
"type": "text"
},
{
"bbox": [
231,
188,
264,
199
],
"score": 0.92,
"content": "\\theta _ { i } \\sim P _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
265,
187,
274,
201
],
"score": 1.0,
"content": "if",
"type": "text"
},
{
"bbox": [
275,
188,
284,
199
],
"score": 0.88,
"content": "\\theta _ { i }",
"type": "inline_equation"
},
{
"bbox": [
284,
187,
398,
201
],
"score": 1.0,
"content": "has been updated more than",
"type": "text"
},
{
"bbox": [
399,
189,
410,
198
],
"score": 0.81,
"content": "K",
"type": "inline_equation"
},
{
"bbox": [
410,
187,
435,
201
],
"score": 1.0,
"content": "steps.",
"type": "text"
}
],
"index": 9
}
],
"index": 6
},
{
"type": "text",
"bbox": [
111,
199,
166,
209
],
"lines": [
{
"bbox": [
110,
198,
167,
211
],
"spans": [
{
"bbox": [
110,
198,
167,
211
],
"score": 1.0,
"content": "8: end while",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
115,
212,
310,
223
],
"lines": [
{
"bbox": [
116,
210,
309,
226
],
"spans": [
{
"bbox": [
116,
210,
254,
226
],
"score": 1.0,
"content": "Output: Learned distilled dataset",
"type": "text"
},
{
"bbox": [
254,
212,
309,
224
],
"score": 0.93,
"content": "\\boldsymbol { \\mathcal { S } } = \\left( \\boldsymbol { X _ { s } } , \\boldsymbol { Y _ { s } } \\right)",
"type": "inline_equation"
}
],
"index": 11
}
],
"index": 11
},
{
"type": "text",
"bbox": [
107,
249,
505,
294
],
"lines": [
{
"bbox": [
105,
250,
505,
262
],
"spans": [
{
"bbox": [
105,
250,
505,
262
],
"score": 1.0,
"content": "extractor and a linear classifier. We fix the feature extractor at each meta-gradient computation and",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
261,
505,
273
],
"spans": [
{
"bbox": [
105,
261,
329,
273
],
"score": 1.0,
"content": "train the linear classifier to convergence before updating",
"type": "text"
},
{
"bbox": [
329,
261,
337,
271
],
"score": 0.67,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
338,
261,
505,
273
],
"score": 1.0,
"content": ". After that, we adjust the feature extractor",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
272,
504,
284
],
"spans": [
{
"bbox": [
106,
272,
504,
284
],
"score": 1.0,
"content": "by training the whole network on the updated distilled data. We note that similar two-phase procedure",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
281,
358,
295
],
"spans": [
{
"bbox": [
105,
281,
358,
295
],
"score": 1.0,
"content": "has been studied in the context of representation learning [31].",
"type": "text"
}
],
"index": 15
}
],
"index": 13.5
},
{
"type": "text",
"bbox": [
106,
298,
505,
354
],
"lines": [
{
"bbox": [
106,
298,
505,
312
],
"spans": [
{
"bbox": [
106,
298,
505,
312
],
"score": 1.0,
"content": "Meta-Gradient Computation: If we consider the mean square error loss, then the optimal weights",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
308,
504,
323
],
"spans": [
{
"bbox": [
106,
308,
504,
323
],
"score": 1.0,
"content": "for the linear classifier have a closed-form solution. Moreover, since the feature dimension is typically",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
319,
506,
334
],
"spans": [
{
"bbox": [
105,
319,
506,
334
],
"score": 1.0,
"content": "larger than the number of distilled data, we can use kernel ridge regression (KRR) with a conjugate",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
331,
506,
345
],
"spans": [
{
"bbox": [
105,
331,
506,
345
],
"score": 1.0,
"content": "kernel [25] rather than solving the weights explicitly [12]. The resulting meta-training loss (Eq. 2) is",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
341,
455,
354
],
"spans": [
{
"bbox": [
105,
341,
455,
354
],
"score": 1.0,
"content": "similar to that used in KIP [22, 23], but we use a more flexible kernel rather than NTK.",
"type": "text"
}
],
"index": 20
}
],
"index": 18
},
{
"type": "interline_equation",
"bbox": [
182,
359,
427,
383
],
"lines": [
{
"bbox": [
182,
359,
427,
383
],
"spans": [
{
"bbox": [
182,
359,
427,
383
],
"score": 0.93,
"content": "\\mathcal { L } \\left( \\mathcal { A } l g \\left( \\theta , \\mathcal { S } \\right) , \\mathcal { T } \\right) = \\frac { 1 } { 2 } | | Y _ { t } - K _ { X _ { t } X _ { s } } ^ { \\theta } ( K _ { X _ { s } X _ { s } } ^ { \\theta } + \\lambda I ) ^ { - 1 } Y _ { s } | | _ { 2 } ^ { 2 } ,",
"type": "interline_equation",
"image_path": "0cd9f2dca22b293f4e069930dff443abaa6142b03906a3c62fda7f6203ea2b36.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
182,
359,
427,
383
],
"spans": [],
"index": 21
}
]
},
{
"type": "text",
"bbox": [
106,
389,
506,
471
],
"lines": [
{
"bbox": [
105,
388,
507,
402
],
"spans": [
{
"bbox": [
105,
388,
133,
402
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
389,
167,
402
],
"score": 0.92,
"content": "( X _ { t } , Y _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
168,
388,
186,
402
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
186,
389,
221,
402
],
"score": 0.92,
"content": "( X _ { s } , Y _ { s } )",
"type": "inline_equation"
},
{
"bbox": [
222,
388,
384,
402
],
"score": 1.0,
"content": "are the inputs and labels of the real data",
"type": "text"
},
{
"bbox": [
385,
390,
394,
400
],
"score": 0.82,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
394,
388,
466,
402
],
"score": 1.0,
"content": "and distilled data",
"type": "text"
},
{
"bbox": [
466,
390,
474,
399
],
"score": 0.84,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
474,
388,
507,
402
],
"score": 1.0,
"content": "respec-",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
101,
397,
502,
420
],
"spans": [
{
"bbox": [
101,
397,
426,
420
],
"score": 1.0,
"content": "tively. The Gram matrix between real inputs and distilled inputs is denoted as",
"type": "text"
},
{
"bbox": [
426,
401,
502,
415
],
"score": 0.92,
"content": "K _ { X _ { t } X _ { s } } ^ { \\theta } \\in \\mathbb { R } ^ { | T | \\times | S | }",
"type": "inline_equation"
}
],
"index": 23
},
{
"bbox": [
102,
409,
510,
433
],
"spans": [
{
"bbox": [
102,
409,
347,
433
],
"score": 1.0,
"content": "while the Gram matrix between distilled inputs is denoted as",
"type": "text"
},
{
"bbox": [
348,
414,
422,
429
],
"score": 0.9,
"content": "K _ { X _ { s } X _ { s } } ^ { \\theta } \\in \\mathbb { R } ^ { | S | \\times | S | }",
"type": "inline_equation"
},
{
"bbox": [
423,
409,
426,
433
],
"score": 1.0,
"content": ".",
"type": "text"
},
{
"bbox": [
427,
416,
434,
425
],
"score": 0.66,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
434,
409,
510,
433
],
"score": 1.0,
"content": "t scontrols the regu-",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
426,
506,
439
],
"spans": [
{
"bbox": [
105,
426,
450,
439
],
"score": 1.0,
"content": "larization strength for KRR. Let us denote the neural network feature for a given input",
"type": "text"
},
{
"bbox": [
450,
427,
460,
437
],
"score": 0.83,
"content": "X",
"type": "inline_equation"
},
{
"bbox": [
460,
426,
506,
439
],
"score": 1.0,
"content": "and model",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
104,
434,
507,
451
],
"spans": [
{
"bbox": [
104,
434,
150,
451
],
"score": 1.0,
"content": "parameter",
"type": "text"
},
{
"bbox": [
150,
438,
156,
447
],
"score": 0.81,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
156,
434,
168,
451
],
"score": 1.0,
"content": "as",
"type": "text"
},
{
"bbox": [
169,
437,
237,
449
],
"score": 0.92,
"content": "\\mathbf { \\bar { \\chi } } _ { f ( X , \\theta ) } \\in \\mathbb { R } ^ { N \\times d }",
"type": "inline_equation"
},
{
"bbox": [
237,
434,
268,
451
],
"score": 1.0,
"content": ", where",
"type": "text"
},
{
"bbox": [
269,
439,
279,
447
],
"score": 0.83,
"content": "N",
"type": "inline_equation"
},
{
"bbox": [
279,
434,
390,
451
],
"score": 1.0,
"content": "is the number of input and",
"type": "text"
},
{
"bbox": [
390,
438,
397,
447
],
"score": 0.82,
"content": "d",
"type": "inline_equation"
},
{
"bbox": [
397,
434,
507,
451
],
"score": 1.0,
"content": "is the feature dimension 1.",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
448,
506,
461
],
"spans": [
{
"bbox": [
106,
448,
506,
461
],
"score": 1.0,
"content": "The conjugate kernel is defined by the inner product of the neural network features. Thus, the two",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
460,
270,
471
],
"spans": [
{
"bbox": [
106,
460,
270,
471
],
"score": 1.0,
"content": "Gram matrices are computed as follows:",
"type": "text"
}
],
"index": 28
}
],
"index": 25
},
{
"type": "interline_equation",
"bbox": [
174,
477,
434,
493
],
"lines": [
{
"bbox": [
174,
477,
434,
493
],
"spans": [
{
"bbox": [
174,
477,
434,
493
],
"score": 0.91,
"content": "K _ { X _ { t } X _ { s } } ^ { \\theta } = f ( X _ { t } , \\theta ) f ( X _ { s } , \\theta ) ^ { \\top } , \\quad K _ { X _ { s } X _ { s } } ^ { \\theta } = f ( X _ { s } , \\theta ) f ( X _ { s } , \\theta ) ^ { \\top } ,",
"type": "interline_equation",
"image_path": "87a4499ed59b016b53bcda7f7527fcb6ea8d32afd3493c3f5288d6937be23d26.jpg"
}
]
}
],
"index": 29,
"virtual_lines": [
{
"bbox": [
174,
477,
434,
493
],
"spans": [],
"index": 29
}
]
},
{
"type": "text",
"bbox": [
106,
498,
505,
587
],
"lines": [
{
"bbox": [
105,
499,
506,
512
],
"spans": [
{
"bbox": [
105,
499,
249,
512
],
"score": 1.0,
"content": "Now, computing the meta-gradient",
"type": "text"
},
{
"bbox": [
250,
499,
336,
511
],
"score": 0.93,
"content": "\\nabla _ { \\mathcal { S } } \\mathcal { L } \\left( \\mathcal { A } l g \\left( \\theta , \\mathcal { S } \\right) , \\mathcal { T } \\right)",
"type": "inline_equation"
},
{
"bbox": [
337,
499,
506,
512
],
"score": 1.0,
"content": "is just back-propagating through the con-",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
104,
511,
506,
523
],
"spans": [
{
"bbox": [
104,
511,
506,
523
],
"score": 1.0,
"content": "jugate kernel and a fixed feature extractor, which is very efficient and takes even fewer operations",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
522,
505,
534
],
"spans": [
{
"bbox": [
105,
522,
505,
534
],
"score": 1.0,
"content": "than computing the gradient for the network’s weights. Moreover, we decouple the meta-gradient",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
532,
506,
545
],
"spans": [
{
"bbox": [
105,
532,
506,
545
],
"score": 1.0,
"content": "computation from the model online update. Hence, we can train the online model using any optimizer,",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
543,
506,
556
],
"spans": [
{
"bbox": [
106,
543,
506,
556
],
"score": 1.0,
"content": "and the distilled data will be agnostic to the specific learning algorithm choice. Our proposed method",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
555,
505,
567
],
"spans": [
{
"bbox": [
105,
555,
505,
567
],
"score": 1.0,
"content": "is similar to 1-step TBPTT in that we compute the meta-gradient at each step while performing the",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
104,
564,
506,
578
],
"spans": [
{
"bbox": [
104,
564,
506,
578
],
"score": 1.0,
"content": "online model update. Unlike the conventional 1-step TBPTT, we compute the meta-gradient using a",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
576,
391,
588
],
"spans": [
{
"bbox": [
106,
576,
391,
588
],
"score": 1.0,
"content": "KRR output layer to mitigate truncation bias, illustrated in Figure 2(a).",
"type": "text"
}
],
"index": 37
}
],
"index": 33.5
},
{
"type": "text",
"bbox": [
106,
591,
506,
691
],
"lines": [
{
"bbox": [
105,
591,
507,
605
],
"spans": [
{
"bbox": [
105,
591,
507,
605
],
"score": 1.0,
"content": "Model Pool: As discussed in Section 1, there are various types of overfitting in dataset distillation.",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
602,
506,
616
],
"spans": [
{
"bbox": [
105,
602,
506,
616
],
"score": 1.0,
"content": "Several techniques have been proposed to alleviate such problem, such as random initialization [4],",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
104,
613,
507,
627
],
"spans": [
{
"bbox": [
104,
613,
507,
627
],
"score": 1.0,
"content": "periodic reset [13, 5, 7], and dynamic bi-level optimization [19]. These techniques share the same",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
624,
507,
637
],
"spans": [
{
"bbox": [
105,
624,
507,
637
],
"score": 1.0,
"content": "underlying principle: the model diversity matters. Thus, we propose to maintain a “model pool”",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
636,
506,
648
],
"spans": [
{
"bbox": [
105,
636,
506,
648
],
"score": 1.0,
"content": "filled with diverse set of parameters obtained from different number of training steps and different",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
104,
646,
506,
660
],
"spans": [
{
"bbox": [
104,
646,
506,
660
],
"score": 1.0,
"content": "random initializations. Unlike the previous methods that periodically training and resetting a single",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
658,
505,
670
],
"spans": [
{
"bbox": [
105,
658,
505,
670
],
"score": 1.0,
"content": "model, FRePo randomly sample a model from the pool at each meta-gradient computation and update",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
104,
667,
506,
682
],
"spans": [
{
"bbox": [
104,
667,
452,
682
],
"score": 1.0,
"content": "it using the current distilled data. However, if a model has been updated more than",
"type": "text"
},
{
"bbox": [
453,
669,
463,
678
],
"score": 0.81,
"content": "K",
"type": "inline_equation"
},
{
"bbox": [
464,
667,
506,
682
],
"score": 1.0,
"content": "steps, we",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
679,
505,
692
],
"spans": [
{
"bbox": [
105,
679,
505,
692
],
"score": 1.0,
"content": "reinitialize it with a new random seed. From the meta-learning perspective, we maintain a diverse",
"type": "text"
}
],
"index": 46
}
],
"index": 42
}
],
"page_idx": 3,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
701,
505,
722
],
"lines": [
{
"bbox": [
120,
700,
505,
712
],
"spans": [
{
"bbox": [
120,
700,
505,
712
],
"score": 1.0,
"content": "1In practice, we use all the synthetic data and sample a minibatch from the real dataset to compute the",
"type": "text"
}
]
},
{
"bbox": [
106,
712,
213,
722
],
"spans": [
{
"bbox": [
106,
712,
213,
722
],
"score": 1.0,
"content": "meta-gradient (Algorithm 1).",
"type": "text"
}
]
}
]
},
{
"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
}
]
}
]
},
{
"type": "discarded",
"bbox": [
107,
72,
463,
84
],
"lines": [
{
"bbox": [
105,
70,
463,
86
],
"spans": [
{
"bbox": [
105,
70,
463,
86
],
"score": 1.0,
"content": "Algorithm 1 Dataset Distillation using Neural Feature Regression with Pooling (FRePo)",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
113,
88,
404,
99
],
"lines": [
{
"bbox": [
116,
87,
404,
100
],
"spans": [
{
"bbox": [
116,
87,
156,
100
],
"score": 1.0,
"content": "Require:",
"type": "text"
},
{
"bbox": [
157,
88,
166,
99
],
"score": 0.57,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
167,
87,
241,
100
],
"score": 1.0,
"content": ": a labeled dataset;",
"type": "text"
},
{
"bbox": [
242,
90,
250,
98
],
"score": 0.61,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
250,
87,
404,
100
],
"score": 1.0,
"content": ": the learning rate for the distilled data",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
122,
97,
378,
111
],
"spans": [
{
"bbox": [
122,
97,
318,
111
],
"score": 1.0,
"content": "nitialization: Initialize a labeled distilled dataset",
"type": "text"
},
{
"bbox": [
319,
99,
374,
110
],
"score": 0.91,
"content": "\\boldsymbol { \\mathcal { S } } = \\left( \\boldsymbol { X _ { s } } , \\boldsymbol { Y _ { s } } \\right)",
"type": "inline_equation"
},
{
"bbox": [
375,
97,
378,
111
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 1
}
],
"index": 0,
"bbox_fs": [
116,
87,
404,
100
]
},
{
"type": "text",
"bbox": [
124,
99,
381,
109
],
"lines": [],
"index": 1,
"bbox_fs": [
122,
97,
378,
111
],
"lines_deleted": true
},
{
"type": "text",
"bbox": [
114,
110,
500,
122
],
"lines": [
{
"bbox": [
115,
105,
497,
124
],
"spans": [
{
"bbox": [
115,
105,
270,
124
],
"score": 1.0,
"content": "Initialization: Initialize a model pool",
"type": "text"
},
{
"bbox": [
270,
110,
284,
120
],
"score": 0.79,
"content": "\\mathcal { M }",
"type": "inline_equation"
},
{
"bbox": [
284,
105,
304,
124
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
305,
111,
315,
119
],
"score": 0.74,
"content": "m",
"type": "inline_equation"
},
{
"bbox": [
315,
105,
347,
124
],
"score": 1.0,
"content": "models",
"type": "text"
},
{
"bbox": [
348,
110,
379,
122
],
"score": 0.89,
"content": "\\left\\{ \\boldsymbol { \\theta } _ { i } \\right\\} _ { i = 1 } ^ { m }",
"type": "inline_equation"
},
{
"bbox": [
379,
105,
484,
124
],
"score": 1.0,
"content": "randomly initialized from",
"type": "text"
},
{
"bbox": [
485,
110,
497,
120
],
"score": 0.86,
"content": "P _ { \\theta }",
"type": "inline_equation"
}
],
"index": 2
}
],
"index": 2,
"bbox_fs": [
115,
105,
497,
124
]
},
{
"type": "index",
"bbox": [
111,
123,
440,
199
],
"lines": [
{
"bbox": [
110,
123,
220,
134
],
"spans": [
{
"bbox": [
110,
123,
220,
134
],
"score": 1.0,
"content": "1: while not converged do",
"type": "text"
}
],
"index": 3,
"is_list_start_line": true
},
{
"bbox": [
111,
133,
375,
145
],
"spans": [
{
"bbox": [
111,
133,
121,
145
],
"score": 1.0,
"content": "2:",
"type": "text"
},
{
"bbox": [
133,
133,
337,
145
],
"score": 1.0,
"content": "Ż Sample a model uniformly from the model pool:",
"type": "text"
},
{
"bbox": [
337,
133,
372,
144
],
"score": 0.91,
"content": "\\theta _ { i } \\sim \\mathcal { M }",
"type": "inline_equation"
},
{
"bbox": [
372,
133,
375,
145
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 4,
"is_list_start_line": true
},
{
"bbox": [
110,
144,
432,
157
],
"spans": [
{
"bbox": [
110,
145,
121,
156
],
"score": 1.0,
"content": "3:",
"type": "text"
},
{
"bbox": [
132,
144,
373,
157
],
"score": 1.0,
"content": "Ż Sample a target batch uniformly from the labeled dataset:",
"type": "text"
},
{
"bbox": [
374,
144,
428,
156
],
"score": 0.9,
"content": "( X _ { t } , Y _ { t } ) \\sim \\tau",
"type": "inline_equation"
},
{
"bbox": [
429,
144,
432,
157
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 5,
"is_list_start_line": true
},
{
"bbox": [
110,
155,
326,
167
],
"spans": [
{
"bbox": [
110,
155,
121,
167
],
"score": 1.0,
"content": "4:",
"type": "text"
},
{
"bbox": [
133,
157,
140,
165
],
"score": 0.27,
"content": "\\triangleright",
"type": "inline_equation"
},
{
"bbox": [
141,
155,
268,
167
],
"score": 1.0,
"content": "Compute the meta-training loss",
"type": "text"
},
{
"bbox": [
269,
156,
277,
165
],
"score": 0.79,
"content": "\\mathcal { L }",
"type": "inline_equation"
},
{
"bbox": [
277,
155,
326,
167
],
"score": 1.0,
"content": "using Eq. 2",
"type": "text"
}
],
"index": 6,
"is_list_start_line": true
},
{
"bbox": [
110,
165,
440,
179
],
"spans": [
{
"bbox": [
110,
167,
121,
178
],
"score": 1.0,
"content": "5:",
"type": "text"
},
{
"bbox": [
133,
165,
240,
179
],
"score": 1.0,
"content": "Ż Update the distilled data",
"type": "text"
},
{
"bbox": [
241,
166,
249,
176
],
"score": 0.37,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
250,
166,
340,
178
],
"score": 0.86,
"content": "\\mathrm { : } ~ X _ { s } \\gets X _ { s } - \\alpha \\nabla _ { X _ { s } } \\mathcal { L }",
"type": "inline_equation"
},
{
"bbox": [
340,
165,
360,
179
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
360,
166,
440,
178
],
"score": 0.9,
"content": "Y _ { s } \\gets Y _ { s } - \\alpha \\nabla _ { Y _ { s } } \\mathcal { L }",
"type": "inline_equation"
}
],
"index": 7,
"is_list_start_line": true
},
{
"bbox": [
110,
175,
390,
191
],
"spans": [
{
"bbox": [
110,
177,
121,
189
],
"score": 1.0,
"content": "6:",
"type": "text"
},
{
"bbox": [
133,
179,
140,
187
],
"score": 0.35,
"content": "\\triangleright",
"type": "inline_equation"
},
{
"bbox": [
141,
175,
206,
191
],
"score": 1.0,
"content": "Train the model",
"type": "text"
},
{
"bbox": [
207,
177,
216,
188
],
"score": 0.88,
"content": "\\theta _ { i }",
"type": "inline_equation"
},
{
"bbox": [
216,
175,
329,
191
],
"score": 1.0,
"content": "on the current distilled data",
"type": "text"
},
{
"bbox": [
329,
178,
337,
187
],
"score": 0.8,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
337,
175,
390,
191
],
"score": 1.0,
"content": "for one step.",
"type": "text"
}
],
"index": 8,
"is_list_start_line": true
},
{
"bbox": [
111,
187,
435,
201
],
"spans": [
{
"bbox": [
111,
188,
121,
199
],
"score": 1.0,
"content": "7:",
"type": "text"
},
{
"bbox": [
133,
190,
141,
198
],
"score": 0.27,
"content": "\\triangleright",
"type": "inline_equation"
},
{
"bbox": [
141,
187,
231,
201
],
"score": 1.0,
"content": "Reinitialize the model",
"type": "text"
},
{
"bbox": [
231,
188,
264,
199
],
"score": 0.92,
"content": "\\theta _ { i } \\sim P _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
265,
187,
274,
201
],
"score": 1.0,
"content": "if",
"type": "text"
},
{
"bbox": [
275,
188,
284,
199
],
"score": 0.88,
"content": "\\theta _ { i }",
"type": "inline_equation"
},
{
"bbox": [
284,
187,
398,
201
],
"score": 1.0,
"content": "has been updated more than",
"type": "text"
},
{
"bbox": [
399,
189,
410,
198
],
"score": 0.81,
"content": "K",
"type": "inline_equation"
},
{
"bbox": [
410,
187,
435,
201
],
"score": 1.0,
"content": "steps.",
"type": "text"
}
],
"index": 9,
"is_list_start_line": true
}
],
"index": 6,
"bbox_fs": [
110,
123,
440,
201
]
},
{
"type": "text",
"bbox": [
111,
199,
166,
209
],
"lines": [
{
"bbox": [
110,
198,
167,
211
],
"spans": [
{
"bbox": [
110,
198,
167,
211
],
"score": 1.0,
"content": "8: end while",
"type": "text"
}
],
"index": 10
}
],
"index": 10,
"bbox_fs": [
110,
198,
167,
211
]
},
{
"type": "text",
"bbox": [
115,
212,
310,
223
],
"lines": [
{
"bbox": [
116,
210,
309,
226
],
"spans": [
{
"bbox": [
116,
210,
254,
226
],
"score": 1.0,
"content": "Output: Learned distilled dataset",
"type": "text"
},
{
"bbox": [
254,
212,
309,
224
],
"score": 0.93,
"content": "\\boldsymbol { \\mathcal { S } } = \\left( \\boldsymbol { X _ { s } } , \\boldsymbol { Y _ { s } } \\right)",
"type": "inline_equation"
}
],
"index": 11
}
],
"index": 11,
"bbox_fs": [
116,
210,
309,
226
]
},
{
"type": "text",
"bbox": [
107,
249,
505,
294
],
"lines": [
{
"bbox": [
105,
250,
505,
262
],
"spans": [
{
"bbox": [
105,
250,
505,
262
],
"score": 1.0,
"content": "extractor and a linear classifier. We fix the feature extractor at each meta-gradient computation and",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
261,
505,
273
],
"spans": [
{
"bbox": [
105,
261,
329,
273
],
"score": 1.0,
"content": "train the linear classifier to convergence before updating",
"type": "text"
},
{
"bbox": [
329,
261,
337,
271
],
"score": 0.67,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
338,
261,
505,
273
],
"score": 1.0,
"content": ". After that, we adjust the feature extractor",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
272,
504,
284
],
"spans": [
{
"bbox": [
106,
272,
504,
284
],
"score": 1.0,
"content": "by training the whole network on the updated distilled data. We note that similar two-phase procedure",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
281,
358,
295
],
"spans": [
{
"bbox": [
105,
281,
358,
295
],
"score": 1.0,
"content": "has been studied in the context of representation learning [31].",
"type": "text"
}
],
"index": 15
}
],
"index": 13.5,
"bbox_fs": [
105,
250,
505,
295
]
},
{
"type": "text",
"bbox": [
106,
298,
505,
354
],
"lines": [
{
"bbox": [
106,
298,
505,
312
],
"spans": [
{
"bbox": [
106,
298,
505,
312
],
"score": 1.0,
"content": "Meta-Gradient Computation: If we consider the mean square error loss, then the optimal weights",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
308,
504,
323
],
"spans": [
{
"bbox": [
106,
308,
504,
323
],
"score": 1.0,
"content": "for the linear classifier have a closed-form solution. Moreover, since the feature dimension is typically",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
319,
506,
334
],
"spans": [
{
"bbox": [
105,
319,
506,
334
],
"score": 1.0,
"content": "larger than the number of distilled data, we can use kernel ridge regression (KRR) with a conjugate",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
331,
506,
345
],
"spans": [
{
"bbox": [
105,
331,
506,
345
],
"score": 1.0,
"content": "kernel [25] rather than solving the weights explicitly [12]. The resulting meta-training loss (Eq. 2) is",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
341,
455,
354
],
"spans": [
{
"bbox": [
105,
341,
455,
354
],
"score": 1.0,
"content": "similar to that used in KIP [22, 23], but we use a more flexible kernel rather than NTK.",
"type": "text"
}
],
"index": 20
}
],
"index": 18,
"bbox_fs": [
105,
298,
506,
354
]
},
{
"type": "interline_equation",
"bbox": [
182,
359,
427,
383
],
"lines": [
{
"bbox": [
182,
359,
427,
383
],
"spans": [
{
"bbox": [
182,
359,
427,
383
],
"score": 0.93,
"content": "\\mathcal { L } \\left( \\mathcal { A } l g \\left( \\theta , \\mathcal { S } \\right) , \\mathcal { T } \\right) = \\frac { 1 } { 2 } | | Y _ { t } - K _ { X _ { t } X _ { s } } ^ { \\theta } ( K _ { X _ { s } X _ { s } } ^ { \\theta } + \\lambda I ) ^ { - 1 } Y _ { s } | | _ { 2 } ^ { 2 } ,",
"type": "interline_equation",
"image_path": "0cd9f2dca22b293f4e069930dff443abaa6142b03906a3c62fda7f6203ea2b36.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
182,
359,
427,
383
],
"spans": [],
"index": 21
}
]
},
{
"type": "text",
"bbox": [
106,
389,
506,
471
],
"lines": [
{
"bbox": [
105,
388,
507,
402
],
"spans": [
{
"bbox": [
105,
388,
133,
402
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
389,
167,
402
],
"score": 0.92,
"content": "( X _ { t } , Y _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
168,
388,
186,
402
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
186,
389,
221,
402
],
"score": 0.92,
"content": "( X _ { s } , Y _ { s } )",
"type": "inline_equation"
},
{
"bbox": [
222,
388,
384,
402
],
"score": 1.0,
"content": "are the inputs and labels of the real data",
"type": "text"
},
{
"bbox": [
385,
390,
394,
400
],
"score": 0.82,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
394,
388,
466,
402
],
"score": 1.0,
"content": "and distilled data",
"type": "text"
},
{
"bbox": [
466,
390,
474,
399
],
"score": 0.84,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
474,
388,
507,
402
],
"score": 1.0,
"content": "respec-",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
101,
397,
502,
420
],
"spans": [
{
"bbox": [
101,
397,
426,
420
],
"score": 1.0,
"content": "tively. The Gram matrix between real inputs and distilled inputs is denoted as",
"type": "text"
},
{
"bbox": [
426,
401,
502,
415
],
"score": 0.92,
"content": "K _ { X _ { t } X _ { s } } ^ { \\theta } \\in \\mathbb { R } ^ { | T | \\times | S | }",
"type": "inline_equation"
}
],
"index": 23
},
{
"bbox": [
102,
409,
510,
433
],
"spans": [
{
"bbox": [
102,
409,
347,
433
],
"score": 1.0,
"content": "while the Gram matrix between distilled inputs is denoted as",
"type": "text"
},
{
"bbox": [
348,
414,
422,
429
],
"score": 0.9,
"content": "K _ { X _ { s } X _ { s } } ^ { \\theta } \\in \\mathbb { R } ^ { | S | \\times | S | }",
"type": "inline_equation"
},
{
"bbox": [
423,
409,
426,
433
],
"score": 1.0,
"content": ".",
"type": "text"
},
{
"bbox": [
427,
416,
434,
425
],
"score": 0.66,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
434,
409,
510,
433
],
"score": 1.0,
"content": "t scontrols the regu-",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
426,
506,
439
],
"spans": [
{
"bbox": [
105,
426,
450,
439
],
"score": 1.0,
"content": "larization strength for KRR. Let us denote the neural network feature for a given input",
"type": "text"
},
{
"bbox": [
450,
427,
460,
437
],
"score": 0.83,
"content": "X",
"type": "inline_equation"
},
{
"bbox": [
460,
426,
506,
439
],
"score": 1.0,
"content": "and model",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
104,
434,
507,
451
],
"spans": [
{
"bbox": [
104,
434,
150,
451
],
"score": 1.0,
"content": "parameter",
"type": "text"
},
{
"bbox": [
150,
438,
156,
447
],
"score": 0.81,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
156,
434,
168,
451
],
"score": 1.0,
"content": "as",
"type": "text"
},
{
"bbox": [
169,
437,
237,
449
],
"score": 0.92,
"content": "\\mathbf { \\bar { \\chi } } _ { f ( X , \\theta ) } \\in \\mathbb { R } ^ { N \\times d }",
"type": "inline_equation"
},
{
"bbox": [
237,
434,
268,
451
],
"score": 1.0,
"content": ", where",
"type": "text"
},
{
"bbox": [
269,
439,
279,
447
],
"score": 0.83,
"content": "N",
"type": "inline_equation"
},
{
"bbox": [
279,
434,
390,
451
],
"score": 1.0,
"content": "is the number of input and",
"type": "text"
},
{
"bbox": [
390,
438,
397,
447
],
"score": 0.82,
"content": "d",
"type": "inline_equation"
},
{
"bbox": [
397,
434,
507,
451
],
"score": 1.0,
"content": "is the feature dimension 1.",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
448,
506,
461
],
"spans": [
{
"bbox": [
106,
448,
506,
461
],
"score": 1.0,
"content": "The conjugate kernel is defined by the inner product of the neural network features. Thus, the two",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
460,
270,
471
],
"spans": [
{
"bbox": [
106,
460,
270,
471
],
"score": 1.0,
"content": "Gram matrices are computed as follows:",
"type": "text"
}
],
"index": 28
}
],
"index": 25,
"bbox_fs": [
101,
388,
510,
471
]
},
{
"type": "interline_equation",
"bbox": [
174,
477,
434,
493
],
"lines": [
{
"bbox": [
174,
477,
434,
493
],
"spans": [
{
"bbox": [
174,
477,
434,
493
],
"score": 0.91,
"content": "K _ { X _ { t } X _ { s } } ^ { \\theta } = f ( X _ { t } , \\theta ) f ( X _ { s } , \\theta ) ^ { \\top } , \\quad K _ { X _ { s } X _ { s } } ^ { \\theta } = f ( X _ { s } , \\theta ) f ( X _ { s } , \\theta ) ^ { \\top } ,",
"type": "interline_equation",
"image_path": "87a4499ed59b016b53bcda7f7527fcb6ea8d32afd3493c3f5288d6937be23d26.jpg"
}
]
}
],
"index": 29,
"virtual_lines": [
{
"bbox": [
174,
477,
434,
493
],
"spans": [],
"index": 29
}
]
},
{
"type": "text",
"bbox": [
106,
498,
505,
587
],
"lines": [
{
"bbox": [
105,
499,
506,
512
],
"spans": [
{
"bbox": [
105,
499,
249,
512
],
"score": 1.0,
"content": "Now, computing the meta-gradient",
"type": "text"
},
{
"bbox": [
250,
499,
336,
511
],
"score": 0.93,
"content": "\\nabla _ { \\mathcal { S } } \\mathcal { L } \\left( \\mathcal { A } l g \\left( \\theta , \\mathcal { S } \\right) , \\mathcal { T } \\right)",
"type": "inline_equation"
},
{
"bbox": [
337,
499,
506,
512
],
"score": 1.0,
"content": "is just back-propagating through the con-",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
104,
511,
506,
523
],
"spans": [
{
"bbox": [
104,
511,
506,
523
],
"score": 1.0,
"content": "jugate kernel and a fixed feature extractor, which is very efficient and takes even fewer operations",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
522,
505,
534
],
"spans": [
{
"bbox": [
105,
522,
505,
534
],
"score": 1.0,
"content": "than computing the gradient for the network’s weights. Moreover, we decouple the meta-gradient",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
532,
506,
545
],
"spans": [
{
"bbox": [
105,
532,
506,
545
],
"score": 1.0,
"content": "computation from the model online update. Hence, we can train the online model using any optimizer,",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
543,
506,
556
],
"spans": [
{
"bbox": [
106,
543,
506,
556
],
"score": 1.0,
"content": "and the distilled data will be agnostic to the specific learning algorithm choice. Our proposed method",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
555,
505,
567
],
"spans": [
{
"bbox": [
105,
555,
505,
567
],
"score": 1.0,
"content": "is similar to 1-step TBPTT in that we compute the meta-gradient at each step while performing the",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
104,
564,
506,
578
],
"spans": [
{
"bbox": [
104,
564,
506,
578
],
"score": 1.0,
"content": "online model update. Unlike the conventional 1-step TBPTT, we compute the meta-gradient using a",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
576,
391,
588
],
"spans": [
{
"bbox": [
106,
576,
391,
588
],
"score": 1.0,
"content": "KRR output layer to mitigate truncation bias, illustrated in Figure 2(a).",
"type": "text"
}
],
"index": 37
}
],
"index": 33.5,
"bbox_fs": [
104,
499,
506,
588
]
},
{
"type": "text",
"bbox": [
106,
591,
506,
691
],
"lines": [
{
"bbox": [
105,
591,
507,
605
],
"spans": [
{
"bbox": [
105,
591,
507,
605
],
"score": 1.0,
"content": "Model Pool: As discussed in Section 1, there are various types of overfitting in dataset distillation.",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
602,
506,
616
],
"spans": [
{
"bbox": [
105,
602,
506,
616
],
"score": 1.0,
"content": "Several techniques have been proposed to alleviate such problem, such as random initialization [4],",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
104,
613,
507,
627
],
"spans": [
{
"bbox": [
104,
613,
507,
627
],
"score": 1.0,
"content": "periodic reset [13, 5, 7], and dynamic bi-level optimization [19]. These techniques share the same",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
624,
507,
637
],
"spans": [
{
"bbox": [
105,
624,
507,
637
],
"score": 1.0,
"content": "underlying principle: the model diversity matters. Thus, we propose to maintain a “model pool”",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
636,
506,
648
],
"spans": [
{
"bbox": [
105,
636,
506,
648
],
"score": 1.0,
"content": "filled with diverse set of parameters obtained from different number of training steps and different",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
104,
646,
506,
660
],
"spans": [
{
"bbox": [
104,
646,
506,
660
],
"score": 1.0,
"content": "random initializations. Unlike the previous methods that periodically training and resetting a single",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
658,
505,
670
],
"spans": [
{
"bbox": [
105,
658,
505,
670
],
"score": 1.0,
"content": "model, FRePo randomly sample a model from the pool at each meta-gradient computation and update",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
104,
667,
506,
682
],
"spans": [
{
"bbox": [
104,
667,
452,
682
],
"score": 1.0,
"content": "it using the current distilled data. However, if a model has been updated more than",
"type": "text"
},
{
"bbox": [
453,
669,
463,
678
],
"score": 0.81,
"content": "K",
"type": "inline_equation"
},
{
"bbox": [
464,
667,
506,
682
],
"score": 1.0,
"content": "steps, we",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
679,
505,
692
],
"spans": [
{
"bbox": [
105,
679,
505,
692
],
"score": 1.0,
"content": "reinitialize it with a new random seed. From the meta-learning perspective, we maintain a diverse",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
106,
72,
505,
85
],
"spans": [
{
"bbox": [
106,
72,
505,
85
],
"score": 1.0,
"content": "set of meta-tasks to sample from and avoid sampling very similar tasks at each consecutive gradient",
"type": "text",
"cross_page": true
}
],
"index": 0
},
{
"bbox": [
105,
83,
320,
96
],
"spans": [
{
"bbox": [
105,
83,
320,
96
],
"score": 1.0,
"content": "computation to avoid overfitting to a particular setup.",
"type": "text",
"cross_page": true
}
],
"index": 1
}
],
"index": 42,
"bbox_fs": [
104,
591,
507,
692
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
106,
73,
504,
95
],
"lines": [
{
"bbox": [
106,
72,
505,
85
],
"spans": [
{
"bbox": [
106,
72,
505,
85
],
"score": 1.0,
"content": "set of meta-tasks to sample from and avoid sampling very similar tasks at each consecutive gradient",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
83,
320,
96
],
"spans": [
{
"bbox": [
105,
83,
320,
96
],
"score": 1.0,
"content": "computation to avoid overfitting to a particular setup.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
107,
100,
505,
166
],
"lines": [
{
"bbox": [
105,
100,
505,
113
],
"spans": [
{
"bbox": [
105,
100,
505,
113
],
"score": 1.0,
"content": "Pool Diversity: We can increase the regularization strength by increasing the diversity of the model",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
111,
506,
124
],
"spans": [
{
"bbox": [
105,
111,
200,
124
],
"score": 1.0,
"content": "pool by setting a larger",
"type": "text"
},
{
"bbox": [
200,
111,
210,
121
],
"score": 0.78,
"content": "K",
"type": "inline_equation"
},
{
"bbox": [
210,
111,
506,
124
],
"score": 1.0,
"content": ", using data augmentation when training the model on the distilled data, or",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
122,
505,
134
],
"spans": [
{
"bbox": [
105,
122,
505,
134
],
"score": 1.0,
"content": "using models with different architectures. To keep our method simple, we use the same architecture",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
132,
505,
145
],
"spans": [
{
"bbox": [
105,
132,
505,
145
],
"score": 1.0,
"content": "for all models in the pool and do not use any data augmentation when training the model on the",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
143,
505,
156
],
"spans": [
{
"bbox": [
105,
143,
505,
156
],
"score": 1.0,
"content": "distilled data. Thus, our model pool only contains models with different initialization, at different",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
154,
403,
167
],
"spans": [
{
"bbox": [
105,
154,
403,
167
],
"score": 1.0,
"content": "optimization stages, and trained at different time-step of the distilled data.",
"type": "text"
}
],
"index": 7
}
],
"index": 4.5
},
{
"type": "title",
"bbox": [
107,
184,
197,
198
],
"lines": [
{
"bbox": [
104,
183,
198,
200
],
"spans": [
{
"bbox": [
104,
183,
198,
200
],
"score": 1.0,
"content": "3 Related Work",
"type": "text"
}
],
"index": 8
}
],
"index": 8
},
{
"type": "text",
"bbox": [
107,
212,
505,
278
],
"lines": [
{
"bbox": [
106,
210,
505,
225
],
"spans": [
{
"bbox": [
106,
210,
505,
225
],
"score": 1.0,
"content": "Unrolling in Bi-Level Optimization: One way to compute the meta-gradient is to differentiate",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
223,
505,
236
],
"spans": [
{
"bbox": [
106,
223,
505,
236
],
"score": 1.0,
"content": "through the unrolled inner optimization [4, 11–13]. However, this approach inherits several difficulties",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
234,
504,
246
],
"spans": [
{
"bbox": [
106,
234,
504,
246
],
"score": 1.0,
"content": "of the unrolled optimization, such as: 1) large computation and memory cost [14]; 2) truncation",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
243,
505,
258
],
"spans": [
{
"bbox": [
105,
243,
505,
258
],
"score": 1.0,
"content": "bias with short unrolls [17]; 3) exploding or vanishing gradients with long unrolls [15]; 4) chaotic",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
255,
505,
268
],
"spans": [
{
"bbox": [
105,
255,
505,
268
],
"score": 1.0,
"content": "and poorly conditioned loss landscapes with long unrolls [16]. In contrast, our method considers",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
267,
503,
279
],
"spans": [
{
"bbox": [
105,
267,
503,
279
],
"score": 1.0,
"content": "approximating the inner optimization with kernel ridge regression instead of unrolled optimization.",
"type": "text"
}
],
"index": 14
}
],
"index": 11.5
},
{
"type": "text",
"bbox": [
107,
282,
505,
382
],
"lines": [
{
"bbox": [
105,
282,
507,
295
],
"spans": [
{
"bbox": [
105,
282,
507,
295
],
"score": 1.0,
"content": "Surrogate Objective: To avoid unrolled optimization, several works turn to surrogate objectives.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
294,
505,
306
],
"spans": [
{
"bbox": [
106,
294,
505,
306
],
"score": 1.0,
"content": "DC [5], DSA [7], and DCC [18] formulate the dataset distillation as a gradient matching problem",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
304,
507,
318
],
"spans": [
{
"bbox": [
105,
304,
507,
318
],
"score": 1.0,
"content": "between the gradients of neural network weights computed on the real and distilled data. In contrast,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
315,
506,
328
],
"spans": [
{
"bbox": [
105,
315,
506,
328
],
"score": 1.0,
"content": "DM [8] and CAFE [19] consider the feature distribution alignment between the real and distilled data.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
326,
505,
339
],
"spans": [
{
"bbox": [
105,
326,
505,
339
],
"score": 1.0,
"content": "Moreover, MTT [20] shows that knowledge from many expert training trajectories can be distilled to",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
337,
505,
351
],
"spans": [
{
"bbox": [
105,
337,
505,
351
],
"score": 1.0,
"content": "a dataset by using a training trajectory matching objective. Nevertheless, surrogate objectives may",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
348,
505,
361
],
"spans": [
{
"bbox": [
106,
348,
505,
361
],
"score": 1.0,
"content": "introduce new biases and thus, may not accurately reflect the true objective. For example, gradient",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
358,
506,
371
],
"spans": [
{
"bbox": [
105,
358,
506,
371
],
"score": 1.0,
"content": "matching approaches [5, 7, 18] only focus on short-range behavior and may easily overfit to a biased",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
371,
333,
383
],
"spans": [
{
"bbox": [
106,
371,
333,
383
],
"score": 1.0,
"content": "set of samples that produce dominant gradients [19, 20].",
"type": "text"
}
],
"index": 23
}
],
"index": 19
},
{
"type": "text",
"bbox": [
107,
386,
505,
485
],
"lines": [
{
"bbox": [
106,
387,
506,
399
],
"spans": [
{
"bbox": [
106,
387,
506,
399
],
"score": 1.0,
"content": "Closed-form Approximation: An alternative way to circumvent unrolled optimization is to find a",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
398,
506,
410
],
"spans": [
{
"bbox": [
106,
398,
506,
410
],
"score": 1.0,
"content": "closed-form approximation to the inner optimization. Based on the correspondence between infinitely-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
409,
505,
421
],
"spans": [
{
"bbox": [
106,
409,
505,
421
],
"score": 1.0,
"content": "wide neural networks and kernel methods, KIP [22, 23] approximates the inner optimization with",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
419,
506,
432
],
"spans": [
{
"bbox": [
105,
419,
506,
432
],
"score": 1.0,
"content": "NTK [21]. In this case, the meta-gradient can be computed by back-propagating through the NTK.",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
430,
506,
444
],
"spans": [
{
"bbox": [
105,
430,
506,
444
],
"score": 1.0,
"content": "However, computing NTK for modern neural networks is extremely expensive. Thus, using NTK",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
441,
506,
454
],
"spans": [
{
"bbox": [
106,
441,
506,
454
],
"score": 1.0,
"content": "for dataset distillation requires thousands of GPU hours and sophisticated implementation of the",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
452,
505,
465
],
"spans": [
{
"bbox": [
106,
452,
505,
465
],
"score": 1.0,
"content": "distributed kernel computation framework [23]. Similar to ours, Bohdal et al. [12] also decomposes",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
462,
506,
475
],
"spans": [
{
"bbox": [
106,
462,
506,
475
],
"score": 1.0,
"content": "the neural network as a feature extractor and a linear classifier. However, they only learn the label",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
474,
425,
487
],
"spans": [
{
"bbox": [
106,
474,
425,
487
],
"score": 1.0,
"content": "and explicitly solve for the optimal classifier weights rather than perform KRR.",
"type": "text"
}
],
"index": 32
}
],
"index": 28
},
{
"type": "title",
"bbox": [
108,
504,
223,
517
],
"lines": [
{
"bbox": [
105,
503,
225,
519
],
"spans": [
{
"bbox": [
105,
503,
225,
519
],
"score": 1.0,
"content": "4 Dataset Distillation",
"type": "text"
}
],
"index": 33
}
],
"index": 33
},
{
"type": "title",
"bbox": [
107,
531,
231,
542
],
"lines": [
{
"bbox": [
105,
530,
231,
544
],
"spans": [
{
"bbox": [
105,
530,
231,
544
],
"score": 1.0,
"content": "4.1 Implementation Details",
"type": "text"
}
],
"index": 34
}
],
"index": 34
},
{
"type": "text",
"bbox": [
107,
552,
505,
662
],
"lines": [
{
"bbox": [
106,
553,
505,
564
],
"spans": [
{
"bbox": [
106,
553,
505,
564
],
"score": 1.0,
"content": "We compare our method to four state-of-the-art dataset distillation methods [7, 8, 20, 23] on various",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
563,
506,
576
],
"spans": [
{
"bbox": [
105,
563,
506,
576
],
"score": 1.0,
"content": "benchmark datasets [26, 32–37]. We train the distilled data using Algorithm 1 with the same set of",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
574,
506,
587
],
"spans": [
{
"bbox": [
105,
574,
506,
587
],
"score": 1.0,
"content": "hyperparameters for all experiments except stated otherwise. Unlike prior work [7, 8, 20], we do not",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
585,
506,
598
],
"spans": [
{
"bbox": [
105,
585,
506,
598
],
"score": 1.0,
"content": "apply data augmentation during training. However, we apply the same data augmentation [7, 20]",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
596,
506,
609
],
"spans": [
{
"bbox": [
105,
596,
506,
609
],
"score": 1.0,
"content": "during evaluation for a fair comparison. We preprocess the data in a similar way as in previous works",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
606,
506,
620
],
"spans": [
{
"bbox": [
105,
606,
506,
620
],
"score": 1.0,
"content": "[20, 23] but use a wider architecture than previous works [7, 8, 20] because the KRR component does",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
618,
506,
630
],
"spans": [
{
"bbox": [
105,
618,
506,
630
],
"score": 1.0,
"content": "not behave well when the feature dimension is low, resulting in a significant performance drop for our",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
628,
506,
641
],
"spans": [
{
"bbox": [
105,
628,
506,
641
],
"score": 1.0,
"content": "method. Results on the original architecture are included in Appendix ??. We evaluate each distilled",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
640,
505,
652
],
"spans": [
{
"bbox": [
106,
640,
505,
652
],
"score": 1.0,
"content": "data using five random neural networks and report the mean and standard deviation. For the baseline",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
650,
497,
663
],
"spans": [
{
"bbox": [
105,
650,
497,
663
],
"score": 1.0,
"content": "method, we report the best of the reported value in the original paper and our reproducing results.",
"type": "text"
}
],
"index": 44
}
],
"index": 39.5
},
{
"type": "text",
"bbox": [
107,
667,
504,
722
],
"lines": [
{
"bbox": [
105,
666,
506,
680
],
"spans": [
{
"bbox": [
105,
666,
506,
680
],
"score": 1.0,
"content": "For the sake of brevity, we provide implementation details about data preprocessing, distilled data",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
677,
505,
691
],
"spans": [
{
"bbox": [
105,
677,
505,
691
],
"score": 1.0,
"content": "initialization, and hyperparameters in Appendix ?? and various ablation studies regarding the model",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
688,
506,
702
],
"spans": [
{
"bbox": [
105,
688,
506,
702
],
"score": 1.0,
"content": "pool, batch size, distilled data initialization, label learning, and model architectures in Appendix",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
698,
506,
713
],
"spans": [
{
"bbox": [
105,
698,
506,
713
],
"score": 1.0,
"content": "??. More distilled image visualizations can be found in Appendix ??. Our code is available at",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
105,
710,
294,
724
],
"spans": [
{
"bbox": [
105,
710,
294,
724
],
"score": 1.0,
"content": "https://github.com/yongchao97/FRePo.",
"type": "text"
}
],
"index": 49
}
],
"index": 47
}
],
"page_idx": 4,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
741,
309,
750
],
"lines": [
{
"bbox": [
301,
740,
310,
753
],
"spans": [
{
"bbox": [
301,
740,
310,
753
],
"score": 1.0,
"content": "5",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
106,
73,
504,
95
],
"lines": [],
"index": 0.5,
"bbox_fs": [
105,
72,
505,
96
],
"lines_deleted": true
},
{
"type": "text",
"bbox": [
107,
100,
505,
166
],
"lines": [
{
"bbox": [
105,
100,
505,
113
],
"spans": [
{
"bbox": [
105,
100,
505,
113
],
"score": 1.0,
"content": "Pool Diversity: We can increase the regularization strength by increasing the diversity of the model",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
111,
506,
124
],
"spans": [
{
"bbox": [
105,
111,
200,
124
],
"score": 1.0,
"content": "pool by setting a larger",
"type": "text"
},
{
"bbox": [
200,
111,
210,
121
],
"score": 0.78,
"content": "K",
"type": "inline_equation"
},
{
"bbox": [
210,
111,
506,
124
],
"score": 1.0,
"content": ", using data augmentation when training the model on the distilled data, or",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
122,
505,
134
],
"spans": [
{
"bbox": [
105,
122,
505,
134
],
"score": 1.0,
"content": "using models with different architectures. To keep our method simple, we use the same architecture",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
132,
505,
145
],
"spans": [
{
"bbox": [
105,
132,
505,
145
],
"score": 1.0,
"content": "for all models in the pool and do not use any data augmentation when training the model on the",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
143,
505,
156
],
"spans": [
{
"bbox": [
105,
143,
505,
156
],
"score": 1.0,
"content": "distilled data. Thus, our model pool only contains models with different initialization, at different",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
154,
403,
167
],
"spans": [
{
"bbox": [
105,
154,
403,
167
],
"score": 1.0,
"content": "optimization stages, and trained at different time-step of the distilled data.",
"type": "text"
}
],
"index": 7
}
],
"index": 4.5,
"bbox_fs": [
105,
100,
506,
167
]
},
{
"type": "title",
"bbox": [
107,
184,
197,
198
],
"lines": [
{
"bbox": [
104,
183,
198,
200
],
"spans": [
{
"bbox": [
104,
183,
198,
200
],
"score": 1.0,
"content": "3 Related Work",
"type": "text"
}
],
"index": 8
}
],
"index": 8
},
{
"type": "text",
"bbox": [
107,
212,
505,
278
],
"lines": [
{
"bbox": [
106,
210,
505,
225
],
"spans": [
{
"bbox": [
106,
210,
505,
225
],
"score": 1.0,
"content": "Unrolling in Bi-Level Optimization: One way to compute the meta-gradient is to differentiate",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
223,
505,
236
],
"spans": [
{
"bbox": [
106,
223,
505,
236
],
"score": 1.0,
"content": "through the unrolled inner optimization [4, 11–13]. However, this approach inherits several difficulties",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
234,
504,
246
],
"spans": [
{
"bbox": [
106,
234,
504,
246
],
"score": 1.0,
"content": "of the unrolled optimization, such as: 1) large computation and memory cost [14]; 2) truncation",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
243,
505,
258
],
"spans": [
{
"bbox": [
105,
243,
505,
258
],
"score": 1.0,
"content": "bias with short unrolls [17]; 3) exploding or vanishing gradients with long unrolls [15]; 4) chaotic",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
255,
505,
268
],
"spans": [
{
"bbox": [
105,
255,
505,
268
],
"score": 1.0,
"content": "and poorly conditioned loss landscapes with long unrolls [16]. In contrast, our method considers",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
267,
503,
279
],
"spans": [
{
"bbox": [
105,
267,
503,
279
],
"score": 1.0,
"content": "approximating the inner optimization with kernel ridge regression instead of unrolled optimization.",
"type": "text"
}
],
"index": 14
}
],
"index": 11.5,
"bbox_fs": [
105,
210,
505,
279
]
},
{
"type": "text",
"bbox": [
107,
282,
505,
382
],
"lines": [
{
"bbox": [
105,
282,
507,
295
],
"spans": [
{
"bbox": [
105,
282,
507,
295
],
"score": 1.0,
"content": "Surrogate Objective: To avoid unrolled optimization, several works turn to surrogate objectives.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
294,
505,
306
],
"spans": [
{
"bbox": [
106,
294,
505,
306
],
"score": 1.0,
"content": "DC [5], DSA [7], and DCC [18] formulate the dataset distillation as a gradient matching problem",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
304,
507,
318
],
"spans": [
{
"bbox": [
105,
304,
507,
318
],
"score": 1.0,
"content": "between the gradients of neural network weights computed on the real and distilled data. In contrast,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
315,
506,
328
],
"spans": [
{
"bbox": [
105,
315,
506,
328
],
"score": 1.0,
"content": "DM [8] and CAFE [19] consider the feature distribution alignment between the real and distilled data.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
326,
505,
339
],
"spans": [
{
"bbox": [
105,
326,
505,
339
],
"score": 1.0,
"content": "Moreover, MTT [20] shows that knowledge from many expert training trajectories can be distilled to",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
337,
505,
351
],
"spans": [
{
"bbox": [
105,
337,
505,
351
],
"score": 1.0,
"content": "a dataset by using a training trajectory matching objective. Nevertheless, surrogate objectives may",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
348,
505,
361
],
"spans": [
{
"bbox": [
106,
348,
505,
361
],
"score": 1.0,
"content": "introduce new biases and thus, may not accurately reflect the true objective. For example, gradient",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
358,
506,
371
],
"spans": [
{
"bbox": [
105,
358,
506,
371
],
"score": 1.0,
"content": "matching approaches [5, 7, 18] only focus on short-range behavior and may easily overfit to a biased",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
371,
333,
383
],
"spans": [
{
"bbox": [
106,
371,
333,
383
],
"score": 1.0,
"content": "set of samples that produce dominant gradients [19, 20].",
"type": "text"
}
],
"index": 23
}
],
"index": 19,
"bbox_fs": [
105,
282,
507,
383
]
},
{
"type": "text",
"bbox": [
107,
386,
505,
485
],
"lines": [
{
"bbox": [
106,
387,
506,
399
],
"spans": [
{
"bbox": [
106,
387,
506,
399
],
"score": 1.0,
"content": "Closed-form Approximation: An alternative way to circumvent unrolled optimization is to find a",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
398,
506,
410
],
"spans": [
{
"bbox": [
106,
398,
506,
410
],
"score": 1.0,
"content": "closed-form approximation to the inner optimization. Based on the correspondence between infinitely-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
409,
505,
421
],
"spans": [
{
"bbox": [
106,
409,
505,
421
],
"score": 1.0,
"content": "wide neural networks and kernel methods, KIP [22, 23] approximates the inner optimization with",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
419,
506,
432
],
"spans": [
{
"bbox": [
105,
419,
506,
432
],
"score": 1.0,
"content": "NTK [21]. In this case, the meta-gradient can be computed by back-propagating through the NTK.",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
430,
506,
444
],
"spans": [
{
"bbox": [
105,
430,
506,
444
],
"score": 1.0,
"content": "However, computing NTK for modern neural networks is extremely expensive. Thus, using NTK",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
441,
506,
454
],
"spans": [
{
"bbox": [
106,
441,
506,
454
],
"score": 1.0,
"content": "for dataset distillation requires thousands of GPU hours and sophisticated implementation of the",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
452,
505,
465
],
"spans": [
{
"bbox": [
106,
452,
505,
465
],
"score": 1.0,
"content": "distributed kernel computation framework [23]. Similar to ours, Bohdal et al. [12] also decomposes",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
462,
506,
475
],
"spans": [
{
"bbox": [
106,
462,
506,
475
],
"score": 1.0,
"content": "the neural network as a feature extractor and a linear classifier. However, they only learn the label",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
474,
425,
487
],
"spans": [
{
"bbox": [
106,
474,
425,
487
],
"score": 1.0,
"content": "and explicitly solve for the optimal classifier weights rather than perform KRR.",
"type": "text"
}
],
"index": 32
}
],
"index": 28,
"bbox_fs": [
105,
387,
506,
487
]
},
{
"type": "title",
"bbox": [
108,
504,
223,
517
],
"lines": [
{
"bbox": [
105,
503,
225,
519
],
"spans": [
{
"bbox": [
105,
503,
225,
519
],
"score": 1.0,
"content": "4 Dataset Distillation",
"type": "text"
}
],
"index": 33
}
],
"index": 33
},
{
"type": "title",
"bbox": [
107,
531,
231,
542
],
"lines": [
{
"bbox": [
105,
530,
231,
544
],
"spans": [
{
"bbox": [
105,
530,
231,
544
],
"score": 1.0,
"content": "4.1 Implementation Details",
"type": "text"
}
],
"index": 34
}
],
"index": 34
},
{
"type": "text",
"bbox": [
107,
552,
505,
662
],
"lines": [
{
"bbox": [
106,
553,
505,
564
],
"spans": [
{
"bbox": [
106,
553,
505,
564
],
"score": 1.0,
"content": "We compare our method to four state-of-the-art dataset distillation methods [7, 8, 20, 23] on various",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
563,
506,
576
],
"spans": [
{
"bbox": [
105,
563,
506,
576
],
"score": 1.0,
"content": "benchmark datasets [26, 32–37]. We train the distilled data using Algorithm 1 with the same set of",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
574,
506,
587
],
"spans": [
{
"bbox": [
105,
574,
506,
587
],
"score": 1.0,
"content": "hyperparameters for all experiments except stated otherwise. Unlike prior work [7, 8, 20], we do not",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
585,
506,
598
],
"spans": [
{
"bbox": [
105,
585,
506,
598
],
"score": 1.0,
"content": "apply data augmentation during training. However, we apply the same data augmentation [7, 20]",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
596,
506,
609
],
"spans": [
{
"bbox": [
105,
596,
506,
609
],
"score": 1.0,
"content": "during evaluation for a fair comparison. We preprocess the data in a similar way as in previous works",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
606,
506,
620
],
"spans": [
{
"bbox": [
105,
606,
506,
620
],
"score": 1.0,
"content": "[20, 23] but use a wider architecture than previous works [7, 8, 20] because the KRR component does",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
618,
506,
630
],
"spans": [
{
"bbox": [
105,
618,
506,
630
],
"score": 1.0,
"content": "not behave well when the feature dimension is low, resulting in a significant performance drop for our",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
628,
506,
641
],
"spans": [
{
"bbox": [
105,
628,
506,
641
],
"score": 1.0,
"content": "method. Results on the original architecture are included in Appendix ??. We evaluate each distilled",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
640,
505,
652
],
"spans": [
{
"bbox": [
106,
640,
505,
652
],
"score": 1.0,
"content": "data using five random neural networks and report the mean and standard deviation. For the baseline",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
650,
497,
663
],
"spans": [
{
"bbox": [
105,
650,
497,
663
],
"score": 1.0,
"content": "method, we report the best of the reported value in the original paper and our reproducing results.",
"type": "text"
}
],
"index": 44
}
],
"index": 39.5,
"bbox_fs": [
105,
553,
506,
663
]
},
{
"type": "text",
"bbox": [
107,
667,
504,
722
],
"lines": [
{
"bbox": [
105,
666,
506,
680
],
"spans": [
{
"bbox": [
105,
666,
506,
680
],
"score": 1.0,
"content": "For the sake of brevity, we provide implementation details about data preprocessing, distilled data",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
677,
505,
691
],
"spans": [
{
"bbox": [
105,
677,
505,
691
],
"score": 1.0,
"content": "initialization, and hyperparameters in Appendix ?? and various ablation studies regarding the model",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
688,
506,
702
],
"spans": [
{
"bbox": [
105,
688,
506,
702
],
"score": 1.0,
"content": "pool, batch size, distilled data initialization, label learning, and model architectures in Appendix",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
698,
506,
713
],
"spans": [
{
"bbox": [
105,
698,
506,
713
],
"score": 1.0,
"content": "??. More distilled image visualizations can be found in Appendix ??. Our code is available at",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
105,
710,
294,
724
],
"spans": [
{
"bbox": [
105,
710,
294,
724
],
"score": 1.0,
"content": "https://github.com/yongchao97/FRePo.",
"type": "text"
}
],
"index": 49
}
],
"index": 47,
"bbox_fs": [
105,
666,
506,
724
]
}
]
},
{
"preproc_blocks": [
{
"type": "table",
"bbox": [
106,
111,
512,
333
],
"blocks": [
{
"type": "table_caption",
"bbox": [
106,
78,
506,
111
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
77,
506,
91
],
"spans": [
{
"bbox": [
105,
77,
506,
91
],
"score": 1.0,
"content": "Table 1: Test accuracies of models trained on the distilled data from scratch. : denotes performance",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
88,
504,
101
],
"spans": [
{
"bbox": [
106,
88,
504,
101
],
"score": 1.0,
"content": "better than the original reported performance. KRR preformance is shown in bracket. FRePo performs",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
100,
476,
112
],
"spans": [
{
"bbox": [
106,
100,
476,
112
],
"score": 1.0,
"content": "extremely well for one image per class setting on CIFAR100, Tiny ImageNet and CUB-200.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "table_body",
"bbox": [
106,
111,
512,
333
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
111,
512,
333
],
"spans": [
{
"bbox": [
106,
111,
512,
333
],
"score": 0.983,
"html": " | Img/Cls | DSA [7] | DM[8] | KIP [23] | MTT [20] | FRePo |
| MNIST | 1 | 88.7±0.6 | 89.9 ± 0.8† | 90.1± 0.1 | 91.4 ± 0.9† | 93.0 ± 0.4 (92.6 ± 0.4) |
| 10 | 97.9 ±0.1† | 97.6 ± 0.1† | 97.5 ± 0.0 | 97.3 ± 0.1+ | 98.6 ± 0.1 (98.6 ± 0.1) |
| 50 | 99.2 ± 0.1 | 98.6 ± 0.1 | 98.3 ± 0.1 | 98.5±0.1+ | 99.2 ± 0.0 (99.2± 0.1) |
| F-MNIST | 1 | 70.6 ± 0.6 | 71.5 ± 0.5† | 73.5± 0.5 | 75.1 ± 0.9† | 75.6 ± 0.3 (77.1 ± 0.2) |
| 10 | 84.8±0.3t | 83.6±0.2t | 86.8±0.1 | 87.2± 0.3+ | 86.2 ± 0.2 (86.8 ± 0.1) |
| 50 | 88.8±0.2t | 88.2±0.1† | 88.0±0.1 | 88.3± 0.1+ | 89.6 ± 0.1 (89.9 ± 0.1) |
| CIFAR10 | 1 | 36.7± 0.8† | 31.0 ± 0.6† | 49.9 ± 0.2 | 46.3 ± 0.8 | 46.8 ± 0.7 (47.9 ± 0.6) |
| 10 | 53.2 ± 0.8† | 49.2 ± 0.8† | 62.7± 0.3 | 65.3 ± 0.7 | 65.5 ± 0.4 (68.0 ± 0.2) |
| 50 | 66.8± 0.4† | 63.7±0.5t | 68.6± 0.2 | 71.6 ± 0.2 | 71.7 ± 0.2 (74.4 ± 0.1) |
| CIFAR100 | 1 | 16.8± 0.2† | 12.2 ± 0.4† | 15.7 ± 0.2 | 24.3 ± 0.3 | 28.7 ± 0.1 (32.3 ± 0.1) |
| 10 | 32.3 ±0.3 | 29.7 ±0.3 | 28.3 ± 0.1 | 40.1 ± 0.4 | 42.5 ± 0.2 (44.9 ± 0.2) |
| 50 | 42.8± 0.4 | 43.6 ± 0.4 | 1 | 47.7 ± 0.2 | 44.3 ± 0.2 (43.0 ± 0.3) |
| T-ImageNet | 1 | 6.6± 0.2t | 3.9± 0.2 | | 8.8 ±0.3 | 15.4 ± 0.3 (19.1 ± 0.3) |
| 10 | 一 | 12.9 ± 0.4 | | 23.2 ± 0.2 | 25.4 ± 0.2 (26.5± 0.1) |
| CUB-200 | 1 | 1.3 ± 0.1† | 1.6 ± 0.1† | | 2.2± 0.1† | 12.4 ± 0.2 (13.7 ± 0.2) |
| 10 | 4.5 ± 0.3† | 4.4 ± 0.2† | | 1 | 16.8 ± 0.1 (16.1 ± 0.3) |
",
"type": "table",
"image_path": "a5d293177e310c0c210335bf5ab5229cfb17e497a341c9da5f6d75598dd1f5dd.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
106,
111,
512,
185.0
],
"spans": [],
"index": 3
},
{
"bbox": [
106,
185.0,
512,
259.0
],
"spans": [],
"index": 4
},
{
"bbox": [
106,
259.0,
512,
333.0
],
"spans": [],
"index": 5
}
]
}
],
"index": 2.5
},
{
"type": "image",
"bbox": [
106,
351,
504,
439
],
"blocks": [
{
"type": "image_body",
"bbox": [
106,
351,
504,
439
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
351,
504,
439
],
"spans": [
{
"bbox": [
106,
351,
504,
439
],
"score": 0.96,
"type": "image",
"image_path": "f1494e6fda90eb5674ceeec703ad75133081347b4f6471179e08a6c207b5ecd8.jpg"
}
]
}
],
"index": 7,
"virtual_lines": [
{
"bbox": [
106,
351,
504,
380.3333333333333
],
"spans": [],
"index": 6
},
{
"bbox": [
106,
380.3333333333333,
504,
409.66666666666663
],
"spans": [],
"index": 7
},
{
"bbox": [
106,
409.66666666666663,
504,
438.99999999999994
],
"spans": [],
"index": 8
}
]
},
{
"type": "image_caption",
"bbox": [
106,
445,
506,
490
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
444,
506,
459
],
"spans": [
{
"bbox": [
105,
444,
347,
459
],
"score": 1.0,
"content": "Figure 3: (a,b) Training efficiency comparison when learning",
"type": "text"
},
{
"bbox": [
347,
446,
387,
456
],
"score": 0.45,
"content": "1 \\mathrm { I m g / C l s }",
"type": "inline_equation"
},
{
"bbox": [
387,
444,
506,
459
],
"score": 1.0,
"content": "on CIFAR100. (c,d) Time per",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
456,
506,
469
],
"spans": [
{
"bbox": [
105,
456,
506,
469
],
"score": 1.0,
"content": "iteration and peak memory usage as we increase the model size. FRePo is significantly more efficient",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
467,
506,
480
],
"spans": [
{
"bbox": [
105,
467,
506,
480
],
"score": 1.0,
"content": "than the previous methods, almost two orders of magnitude faster than the second-best method (i.e.,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
478,
334,
491
],
"spans": [
{
"bbox": [
105,
478,
334,
491
],
"score": 1.0,
"content": "MTT), with only 1/10 of the GPU memory requirement.",
"type": "text"
}
],
"index": 12
}
],
"index": 10.5
}
],
"index": 8.75
},
{
"type": "title",
"bbox": [
108,
514,
226,
526
],
"lines": [
{
"bbox": [
106,
514,
227,
527
],
"spans": [
{
"bbox": [
106,
514,
227,
527
],
"score": 1.0,
"content": "4.2 Standard Benchmarks",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
106,
536,
505,
722
],
"lines": [
{
"bbox": [
106,
536,
505,
547
],
"spans": [
{
"bbox": [
106,
536,
505,
547
],
"score": 1.0,
"content": "Distillation Performance: We first evaluate our method on six standard benchmark datasets. We",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
547,
506,
560
],
"spans": [
{
"bbox": [
105,
547,
506,
560
],
"score": 1.0,
"content": "learn 1, 10, and 50 images per class for datasets with only ten classes, while we learn 1 and 10",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
104,
557,
506,
571
],
"spans": [
{
"bbox": [
104,
557,
506,
571
],
"score": 1.0,
"content": "images per class for CIFAR100 [34] with 100 classes, Tiny ImageNet [35] with 200 classes, and",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
568,
506,
581
],
"spans": [
{
"bbox": [
106,
568,
506,
581
],
"score": 1.0,
"content": "CUB-200 [37] with 200 fine-grained classes. As shown in Table in 1, we achieve the state-of-the-art",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
579,
505,
592
],
"spans": [
{
"bbox": [
105,
579,
505,
592
],
"score": 1.0,
"content": "performance in most settings despite the hyperparameter may be suboptimal. Our method performs",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
591,
505,
604
],
"spans": [
{
"bbox": [
105,
591,
505,
604
],
"score": 1.0,
"content": "exceptionally well on datasets with a complex label space when learning few images per class. For",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
600,
506,
616
],
"spans": [
{
"bbox": [
105,
600,
506,
616
],
"score": 1.0,
"content": "example, we improve the CIFAR100, Tiny ImageNet, and CUB-200 in one image per class setting",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
610,
506,
626
],
"spans": [
{
"bbox": [
105,
610,
129,
626
],
"score": 1.0,
"content": "from",
"type": "text"
},
{
"bbox": [
129,
613,
156,
623
],
"score": 0.84,
"content": "2 4 . 3 \\%",
"type": "inline_equation"
},
{
"bbox": [
157,
610,
160,
626
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
160,
613,
183,
623
],
"score": 0.85,
"content": "8 . 8 \\%",
"type": "inline_equation"
},
{
"bbox": [
183,
610,
204,
626
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
204,
613,
227,
623
],
"score": 0.87,
"content": "2 . 2 \\%",
"type": "inline_equation"
},
{
"bbox": [
228,
610,
239,
626
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
239,
613,
266,
623
],
"score": 0.84,
"content": "2 8 . 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
267,
610,
271,
626
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
271,
613,
298,
623
],
"score": 0.85,
"content": "1 5 . 4 \\%",
"type": "inline_equation"
},
{
"bbox": [
298,
610,
320,
626
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
321,
613,
347,
623
],
"score": 0.87,
"content": "1 2 . 4 \\%",
"type": "inline_equation"
},
{
"bbox": [
348,
610,
506,
626
],
"score": 1.0,
"content": ", respectively. Figure 4 shows that our",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
622,
506,
637
],
"spans": [
{
"bbox": [
105,
622,
506,
637
],
"score": 1.0,
"content": "distilled images look real and natural though we do not directly optimize for this objective. We",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
634,
506,
648
],
"spans": [
{
"bbox": [
105,
634,
506,
648
],
"score": 1.0,
"content": "observe a strong correlation between the test accuracy and image quality: the better the image quality,",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
646,
505,
658
],
"spans": [
{
"bbox": [
106,
646,
505,
658
],
"score": 1.0,
"content": "the higher the test accuracy. Our results suggest that a highly condensed dataset does not need to be",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
655,
506,
669
],
"spans": [
{
"bbox": [
105,
655,
506,
669
],
"score": 1.0,
"content": "very different from the real dataset as it may just reflect the most common pattern in a dataset. We",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
667,
507,
681
],
"spans": [
{
"bbox": [
105,
667,
507,
681
],
"score": 1.0,
"content": "also report the KRR predictor’s test accuracy using the feature extractor trained on the distilled data.",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
677,
506,
691
],
"spans": [
{
"bbox": [
105,
677,
506,
691
],
"score": 1.0,
"content": "When the dataset is as simple as MNIST [32], the KRR predictor achieves similar performance as",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
689,
505,
702
],
"spans": [
{
"bbox": [
106,
689,
505,
702
],
"score": 1.0,
"content": "the neural network predictor. In contrast, for more complex datasets, the KRR predictor consistently",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
700,
506,
713
],
"spans": [
{
"bbox": [
106,
700,
408,
713
],
"score": 1.0,
"content": "outperforms the neural network predictor, with the most significant gap being",
"type": "text"
},
{
"bbox": [
408,
700,
430,
711
],
"score": 0.84,
"content": "3 . 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
430,
700,
506,
713
],
"score": 1.0,
"content": "for Tiny ImageNet",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
710,
243,
724
],
"spans": [
{
"bbox": [
105,
710,
243,
724
],
"score": 1.0,
"content": "in the one image per class setting.",
"type": "text"
}
],
"index": 30
}
],
"index": 22
}
],
"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": "table",
"bbox": [
106,
111,
512,
333
],
"blocks": [
{
"type": "table_caption",
"bbox": [
106,
78,
506,
111
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
77,
506,
91
],
"spans": [
{
"bbox": [
105,
77,
506,
91
],
"score": 1.0,
"content": "Table 1: Test accuracies of models trained on the distilled data from scratch. : denotes performance",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
88,
504,
101
],
"spans": [
{
"bbox": [
106,
88,
504,
101
],
"score": 1.0,
"content": "better than the original reported performance. KRR preformance is shown in bracket. FRePo performs",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
100,
476,
112
],
"spans": [
{
"bbox": [
106,
100,
476,
112
],
"score": 1.0,
"content": "extremely well for one image per class setting on CIFAR100, Tiny ImageNet and CUB-200.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "table_body",
"bbox": [
106,
111,
512,
333
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
111,
512,
333
],
"spans": [
{
"bbox": [
106,
111,
512,
333
],
"score": 0.983,
"html": " | Img/Cls | DSA [7] | DM[8] | KIP [23] | MTT [20] | FRePo |
| MNIST | 1 | 88.7±0.6 | 89.9 ± 0.8† | 90.1± 0.1 | 91.4 ± 0.9† | 93.0 ± 0.4 (92.6 ± 0.4) |
| 10 | 97.9 ±0.1† | 97.6 ± 0.1† | 97.5 ± 0.0 | 97.3 ± 0.1+ | 98.6 ± 0.1 (98.6 ± 0.1) |
| 50 | 99.2 ± 0.1 | 98.6 ± 0.1 | 98.3 ± 0.1 | 98.5±0.1+ | 99.2 ± 0.0 (99.2± 0.1) |
| F-MNIST | 1 | 70.6 ± 0.6 | 71.5 ± 0.5† | 73.5± 0.5 | 75.1 ± 0.9† | 75.6 ± 0.3 (77.1 ± 0.2) |
| 10 | 84.8±0.3t | 83.6±0.2t | 86.8±0.1 | 87.2± 0.3+ | 86.2 ± 0.2 (86.8 ± 0.1) |
| 50 | 88.8±0.2t | 88.2±0.1† | 88.0±0.1 | 88.3± 0.1+ | 89.6 ± 0.1 (89.9 ± 0.1) |
| CIFAR10 | 1 | 36.7± 0.8† | 31.0 ± 0.6† | 49.9 ± 0.2 | 46.3 ± 0.8 | 46.8 ± 0.7 (47.9 ± 0.6) |
| 10 | 53.2 ± 0.8† | 49.2 ± 0.8† | 62.7± 0.3 | 65.3 ± 0.7 | 65.5 ± 0.4 (68.0 ± 0.2) |
| 50 | 66.8± 0.4† | 63.7±0.5t | 68.6± 0.2 | 71.6 ± 0.2 | 71.7 ± 0.2 (74.4 ± 0.1) |
| CIFAR100 | 1 | 16.8± 0.2† | 12.2 ± 0.4† | 15.7 ± 0.2 | 24.3 ± 0.3 | 28.7 ± 0.1 (32.3 ± 0.1) |
| 10 | 32.3 ±0.3 | 29.7 ±0.3 | 28.3 ± 0.1 | 40.1 ± 0.4 | 42.5 ± 0.2 (44.9 ± 0.2) |
| 50 | 42.8± 0.4 | 43.6 ± 0.4 | 1 | 47.7 ± 0.2 | 44.3 ± 0.2 (43.0 ± 0.3) |
| T-ImageNet | 1 | 6.6± 0.2t | 3.9± 0.2 | | 8.8 ±0.3 | 15.4 ± 0.3 (19.1 ± 0.3) |
| 10 | 一 | 12.9 ± 0.4 | | 23.2 ± 0.2 | 25.4 ± 0.2 (26.5± 0.1) |
| CUB-200 | 1 | 1.3 ± 0.1† | 1.6 ± 0.1† | | 2.2± 0.1† | 12.4 ± 0.2 (13.7 ± 0.2) |
| 10 | 4.5 ± 0.3† | 4.4 ± 0.2† | | 1 | 16.8 ± 0.1 (16.1 ± 0.3) |
",
"type": "table",
"image_path": "a5d293177e310c0c210335bf5ab5229cfb17e497a341c9da5f6d75598dd1f5dd.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
106,
111,
512,
185.0
],
"spans": [],
"index": 3
},
{
"bbox": [
106,
185.0,
512,
259.0
],
"spans": [],
"index": 4
},
{
"bbox": [
106,
259.0,
512,
333.0
],
"spans": [],
"index": 5
}
]
}
],
"index": 2.5
},
{
"type": "image",
"bbox": [
106,
351,
504,
439
],
"blocks": [
{
"type": "image_body",
"bbox": [
106,
351,
504,
439
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
351,
504,
439
],
"spans": [
{
"bbox": [
106,
351,
504,
439
],
"score": 0.96,
"type": "image",
"image_path": "f1494e6fda90eb5674ceeec703ad75133081347b4f6471179e08a6c207b5ecd8.jpg"
}
]
}
],
"index": 7,
"virtual_lines": [
{
"bbox": [
106,
351,
504,
380.3333333333333
],
"spans": [],
"index": 6
},
{
"bbox": [
106,
380.3333333333333,
504,
409.66666666666663
],
"spans": [],
"index": 7
},
{
"bbox": [
106,
409.66666666666663,
504,
438.99999999999994
],
"spans": [],
"index": 8
}
]
},
{
"type": "image_caption",
"bbox": [
106,
445,
506,
490
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
444,
506,
459
],
"spans": [
{
"bbox": [
105,
444,
347,
459
],
"score": 1.0,
"content": "Figure 3: (a,b) Training efficiency comparison when learning",
"type": "text"
},
{
"bbox": [
347,
446,
387,
456
],
"score": 0.45,
"content": "1 \\mathrm { I m g / C l s }",
"type": "inline_equation"
},
{
"bbox": [
387,
444,
506,
459
],
"score": 1.0,
"content": "on CIFAR100. (c,d) Time per",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
456,
506,
469
],
"spans": [
{
"bbox": [
105,
456,
506,
469
],
"score": 1.0,
"content": "iteration and peak memory usage as we increase the model size. FRePo is significantly more efficient",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
467,
506,
480
],
"spans": [
{
"bbox": [
105,
467,
506,
480
],
"score": 1.0,
"content": "than the previous methods, almost two orders of magnitude faster than the second-best method (i.e.,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
478,
334,
491
],
"spans": [
{
"bbox": [
105,
478,
334,
491
],
"score": 1.0,
"content": "MTT), with only 1/10 of the GPU memory requirement.",
"type": "text"
}
],
"index": 12
}
],
"index": 10.5
}
],
"index": 8.75
},
{
"type": "title",
"bbox": [
108,
514,
226,
526
],
"lines": [
{
"bbox": [
106,
514,
227,
527
],
"spans": [
{
"bbox": [
106,
514,
227,
527
],
"score": 1.0,
"content": "4.2 Standard Benchmarks",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
106,
536,
505,
722
],
"lines": [
{
"bbox": [
106,
536,
505,
547
],
"spans": [
{
"bbox": [
106,
536,
505,
547
],
"score": 1.0,
"content": "Distillation Performance: We first evaluate our method on six standard benchmark datasets. We",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
547,
506,
560
],
"spans": [
{
"bbox": [
105,
547,
506,
560
],
"score": 1.0,
"content": "learn 1, 10, and 50 images per class for datasets with only ten classes, while we learn 1 and 10",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
104,
557,
506,
571
],
"spans": [
{
"bbox": [
104,
557,
506,
571
],
"score": 1.0,
"content": "images per class for CIFAR100 [34] with 100 classes, Tiny ImageNet [35] with 200 classes, and",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
568,
506,
581
],
"spans": [
{
"bbox": [
106,
568,
506,
581
],
"score": 1.0,
"content": "CUB-200 [37] with 200 fine-grained classes. As shown in Table in 1, we achieve the state-of-the-art",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
579,
505,
592
],
"spans": [
{
"bbox": [
105,
579,
505,
592
],
"score": 1.0,
"content": "performance in most settings despite the hyperparameter may be suboptimal. Our method performs",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
591,
505,
604
],
"spans": [
{
"bbox": [
105,
591,
505,
604
],
"score": 1.0,
"content": "exceptionally well on datasets with a complex label space when learning few images per class. For",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
600,
506,
616
],
"spans": [
{
"bbox": [
105,
600,
506,
616
],
"score": 1.0,
"content": "example, we improve the CIFAR100, Tiny ImageNet, and CUB-200 in one image per class setting",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
610,
506,
626
],
"spans": [
{
"bbox": [
105,
610,
129,
626
],
"score": 1.0,
"content": "from",
"type": "text"
},
{
"bbox": [
129,
613,
156,
623
],
"score": 0.84,
"content": "2 4 . 3 \\%",
"type": "inline_equation"
},
{
"bbox": [
157,
610,
160,
626
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
160,
613,
183,
623
],
"score": 0.85,
"content": "8 . 8 \\%",
"type": "inline_equation"
},
{
"bbox": [
183,
610,
204,
626
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
204,
613,
227,
623
],
"score": 0.87,
"content": "2 . 2 \\%",
"type": "inline_equation"
},
{
"bbox": [
228,
610,
239,
626
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
239,
613,
266,
623
],
"score": 0.84,
"content": "2 8 . 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
267,
610,
271,
626
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
271,
613,
298,
623
],
"score": 0.85,
"content": "1 5 . 4 \\%",
"type": "inline_equation"
},
{
"bbox": [
298,
610,
320,
626
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
321,
613,
347,
623
],
"score": 0.87,
"content": "1 2 . 4 \\%",
"type": "inline_equation"
},
{
"bbox": [
348,
610,
506,
626
],
"score": 1.0,
"content": ", respectively. Figure 4 shows that our",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
622,
506,
637
],
"spans": [
{
"bbox": [
105,
622,
506,
637
],
"score": 1.0,
"content": "distilled images look real and natural though we do not directly optimize for this objective. We",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
634,
506,
648
],
"spans": [
{
"bbox": [
105,
634,
506,
648
],
"score": 1.0,
"content": "observe a strong correlation between the test accuracy and image quality: the better the image quality,",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
646,
505,
658
],
"spans": [
{
"bbox": [
106,
646,
505,
658
],
"score": 1.0,
"content": "the higher the test accuracy. Our results suggest that a highly condensed dataset does not need to be",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
655,
506,
669
],
"spans": [
{
"bbox": [
105,
655,
506,
669
],
"score": 1.0,
"content": "very different from the real dataset as it may just reflect the most common pattern in a dataset. We",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
667,
507,
681
],
"spans": [
{
"bbox": [
105,
667,
507,
681
],
"score": 1.0,
"content": "also report the KRR predictor’s test accuracy using the feature extractor trained on the distilled data.",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
677,
506,
691
],
"spans": [
{
"bbox": [
105,
677,
506,
691
],
"score": 1.0,
"content": "When the dataset is as simple as MNIST [32], the KRR predictor achieves similar performance as",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
689,
505,
702
],
"spans": [
{
"bbox": [
106,
689,
505,
702
],
"score": 1.0,
"content": "the neural network predictor. In contrast, for more complex datasets, the KRR predictor consistently",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
700,
506,
713
],
"spans": [
{
"bbox": [
106,
700,
408,
713
],
"score": 1.0,
"content": "outperforms the neural network predictor, with the most significant gap being",
"type": "text"
},
{
"bbox": [
408,
700,
430,
711
],
"score": 0.84,
"content": "3 . 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
430,
700,
506,
713
],
"score": 1.0,
"content": "for Tiny ImageNet",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
710,
243,
724
],
"spans": [
{
"bbox": [
105,
710,
243,
724
],
"score": 1.0,
"content": "in the one image per class setting.",
"type": "text"
}
],
"index": 30
}
],
"index": 22,
"bbox_fs": [
104,
536,
507,
724
]
}
]
},
{
"preproc_blocks": [
{
"type": "table",
"bbox": [
106,
122,
522,
209
],
"blocks": [
{
"type": "table_caption",
"bbox": [
106,
77,
506,
121
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
77,
505,
90
],
"spans": [
{
"bbox": [
105,
77,
370,
90
],
"score": 1.0,
"content": "Table 2: Cross-architecture transfer performance on CIFAR10 with",
"type": "text"
},
{
"bbox": [
370,
77,
415,
89
],
"score": 0.56,
"content": "1 0 \\mathrm { I m g / C l s }",
"type": "inline_equation"
},
{
"bbox": [
416,
77,
505,
90
],
"score": 1.0,
"content": ". Despite being trained",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
87,
506,
102
],
"spans": [
{
"bbox": [
105,
87,
506,
102
],
"score": 1.0,
"content": "for a specific architecture, our distilled data transfer well to various architectures unseen during",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
100,
506,
111
],
"spans": [
{
"bbox": [
106,
100,
506,
111
],
"score": 1.0,
"content": "training. Conv is the default evaluation model used for each method. NN, DN, IN, and BN stand for",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
110,
505,
123
],
"spans": [
{
"bbox": [
105,
110,
505,
123
],
"score": 1.0,
"content": "no normalization, default normalization, Instance Normalization, Batch Normalization respectively.",
"type": "text"
}
],
"index": 3
}
],
"index": 1.5
},
{
"type": "table_body",
"bbox": [
106,
122,
522,
209
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
122,
522,
209
],
"spans": [
{
"bbox": [
106,
122,
522,
209
],
"score": 0.97,
"html": " | Train Arch | Evaluation Architecture |
| Conv | Conv-NN | ResNet-DN | ResNet-BN | VGG-BN | AlexNet |
| DSA [7] | Conv-IN | 53.2 ± 0.8 | 36.4 ± 1.5 | 42.1 ± 0.7 | 34.1 ± 1.4 | 46.3 ± 1.3 | 34.0 ± 2.3 |
| DM[8] | Conv-IN | 49.2 ± 0.8 | 35.2 ± 0.5 | 36.8 ± 1.2 | 35.5 ± 1.3 | 41.2 ± 1.8 | 34.9 ± 1.1 |
| MTT[20] | Conv-IN | 64.4 ± 0.9 | 41.6 ± 1.3 | 49.2 ± 1.1 | 42.9 ± 1.5 | 46.6 ± 2.0 | 34.2 ± 2.6 |
| KIP [23] | Conv-NTK | 62.7 ± 0.3 | 58.2 ±0.4 | 49.0 ± 1.2 | 45.8 ± 1.4 | 30.1 ± 1.5 | 57.2 ± 0.4 |
| FRePo | Conv-BN | 65.5 ± 0.4 | 65.5 ± 0.4 | 58.1 ± 0.6 | 57.7 ± 0.7 | 59.4 ± 0.7 | 61.9 ± 0.7 |
",
"type": "table",
"image_path": "4e44b3dde3c2a554e1627b4447e7df3e28b023aaff4e028f17f27dce37cbd8da.jpg"
}
]
}
],
"index": 5,
"virtual_lines": [
{
"bbox": [
106,
122,
522,
151.0
],
"spans": [],
"index": 4
},
{
"bbox": [
106,
151.0,
522,
180.0
],
"spans": [],
"index": 5
},
{
"bbox": [
106,
180.0,
522,
209.0
],
"spans": [],
"index": 6
}
]
}
],
"index": 3.25
},
{
"type": "image",
"bbox": [
108,
226,
502,
336
],
"blocks": [
{
"type": "image_body",
"bbox": [
108,
226,
502,
336
],
"group_id": 0,
"lines": [
{
"bbox": [
108,
226,
502,
336
],
"spans": [
{
"bbox": [
108,
226,
502,
336
],
"score": 0.967,
"type": "image",
"image_path": "06d156480cb24f7de4d97dc51031b5fb9e76281f33da496484aeb238e841000d.jpg"
}
]
}
],
"index": 8,
"virtual_lines": [
{
"bbox": [
108,
226,
502,
262.6666666666667
],
"spans": [],
"index": 7
},
{
"bbox": [
108,
262.6666666666667,
502,
299.33333333333337
],
"spans": [],
"index": 8
},
{
"bbox": [
108,
299.33333333333337,
502,
336.00000000000006
],
"spans": [],
"index": 9
}
]
},
{
"type": "image_caption",
"bbox": [
106,
345,
506,
379
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
344,
505,
360
],
"spans": [
{
"bbox": [
105,
344,
505,
360
],
"score": 1.0,
"content": "Figure 4: (a,b,c) Distilled 1 img/cls from CIFAR100 using FRePo, MTT, and DSA. High quality",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
356,
507,
370
],
"spans": [
{
"bbox": [
105,
356,
507,
370
],
"score": 1.0,
"content": "images also produce high test accuracy. (d) Three categories of learned labels. (Top) High confidence,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
368,
491,
380
],
"spans": [
{
"bbox": [
106,
368,
491,
380
],
"score": 1.0,
"content": "large margin; (Middle) High confidence, small margin; (Bottom) Low confidence, small margin.",
"type": "text"
}
],
"index": 12
}
],
"index": 11
}
],
"index": 9.5
},
{
"type": "text",
"bbox": [
106,
405,
505,
547
],
"lines": [
{
"bbox": [
105,
404,
505,
418
],
"spans": [
{
"bbox": [
105,
404,
505,
418
],
"score": 1.0,
"content": "Label Learning: A similar trend can also be observed for label learning. When the dataset is simple",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
416,
506,
429
],
"spans": [
{
"bbox": [
106,
416,
506,
429
],
"score": 1.0,
"content": "and has only a few classes, label learning may not be necessary. However, it becomes crucial for",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
427,
506,
440
],
"spans": [
{
"bbox": [
105,
427,
506,
440
],
"score": 1.0,
"content": "complex datasets with many labels, such as CIFAR100 and Tiny-ImageNet (See more details in",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
437,
505,
450
],
"spans": [
{
"bbox": [
105,
437,
505,
450
],
"score": 1.0,
"content": "Appendix ??). Similar to the teacher label in the knowledge distillation [1], we observe that the",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
448,
506,
461
],
"spans": [
{
"bbox": [
105,
448,
506,
461
],
"score": 1.0,
"content": "distilled label also encodes the class similarity. We identify three typical cases in Figure 4d. The",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
460,
505,
472
],
"spans": [
{
"bbox": [
106,
460,
505,
472
],
"score": 1.0,
"content": "first group consists of highly confident labels with a much higher value for one class than other",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
470,
506,
483
],
"spans": [
{
"bbox": [
106,
470,
506,
483
],
"score": 1.0,
"content": "classes (large margin), such as sunflower, bicycle, and chair. In contrast, the distilled labels in the",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
481,
506,
494
],
"spans": [
{
"bbox": [
105,
481,
506,
494
],
"score": 1.0,
"content": "second group are confident but may get confused with some closely-related classes (small margin).",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
492,
507,
505
],
"spans": [
{
"bbox": [
105,
492,
507,
505
],
"score": 1.0,
"content": "For instance, the learned label for \"girl\" has almost equally high values for the girl, woman, man,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
502,
506,
517
],
"spans": [
{
"bbox": [
105,
502,
506,
517
],
"score": 1.0,
"content": "boy, and baby, suggesting that these classes are very similar and may be difficult for the model to",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
514,
506,
527
],
"spans": [
{
"bbox": [
105,
514,
506,
527
],
"score": 1.0,
"content": "distinguish them apart. The last group contains distilled labels with low values for all classes, such",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
525,
505,
537
],
"spans": [
{
"bbox": [
106,
525,
505,
537
],
"score": 1.0,
"content": "as bear, beaver, and squirrel. It is often hard for humans to recognize the distilled images in such a",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
537,
392,
549
],
"spans": [
{
"bbox": [
105,
537,
392,
549
],
"score": 1.0,
"content": "group, suggesting that they may be the challenging classes in a dataset.",
"type": "text"
}
],
"index": 25
}
],
"index": 19
},
{
"type": "text",
"bbox": [
106,
552,
505,
651
],
"lines": [
{
"bbox": [
105,
552,
506,
565
],
"spans": [
{
"bbox": [
105,
552,
506,
565
],
"score": 1.0,
"content": "Training Cost Analysis: Figure 3a, 3b shows that our method is significantly more time-efficient",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
563,
506,
576
],
"spans": [
{
"bbox": [
106,
563,
506,
576
],
"score": 1.0,
"content": "than the previous methods. When learning one image per class on CIFAR100, FRePo reaches a",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
574,
505,
587
],
"spans": [
{
"bbox": [
106,
574,
192,
587
],
"score": 1.0,
"content": "similar test accuracy",
"type": "text"
},
{
"bbox": [
192,
574,
226,
586
],
"score": 0.87,
"content": "( 2 3 . 4 \\% )",
"type": "inline_equation"
},
{
"bbox": [
226,
574,
335,
587
],
"score": 1.0,
"content": "to the second-best method",
"type": "text"
},
{
"bbox": [
335,
574,
369,
586
],
"score": 0.86,
"content": "( 2 4 . 0 \\% )",
"type": "inline_equation"
},
{
"bbox": [
369,
574,
505,
587
],
"score": 1.0,
"content": "in 38 seconds, compared to 3805",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
584,
505,
599
],
"spans": [
{
"bbox": [
105,
584,
484,
599
],
"score": 1.0,
"content": "seconds for MTT, which is roughly two orders of magnitude faster. Moreover, FRePo achieves",
"type": "text"
},
{
"bbox": [
485,
585,
505,
596
],
"score": 0.87,
"content": "92 \\%",
"type": "inline_equation"
}
],
"index": 29
},
{
"bbox": [
105,
596,
506,
609
],
"spans": [
{
"bbox": [
105,
596,
211,
609
],
"score": 1.0,
"content": "of its final test accuracy (",
"type": "text"
},
{
"bbox": [
211,
596,
239,
607
],
"score": 0.84,
"content": "2 6 . 4 \\%",
"type": "inline_equation"
},
{
"bbox": [
239,
596,
267,
609
],
"score": 1.0,
"content": "out of",
"type": "text"
},
{
"bbox": [
268,
596,
296,
607
],
"score": 0.84,
"content": "2 8 . 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
297,
596,
506,
609
],
"score": 1.0,
"content": ") in only 385 seconds. As shown in Figure 3c, our",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
607,
505,
620
],
"spans": [
{
"bbox": [
106,
607,
505,
620
],
"score": 1.0,
"content": "algorithm takes much less time to perform one gradient step on the distilled data. Thus, we can",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
618,
505,
631
],
"spans": [
{
"bbox": [
105,
618,
505,
631
],
"score": 1.0,
"content": "perform more gradient steps in a fixed time. Furthermore, Figure 3d suggests that our algorithm has",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
627,
506,
642
],
"spans": [
{
"bbox": [
105,
627,
506,
642
],
"score": 1.0,
"content": "much less GPU memory requirement. Therefore, we can potentially use a much larger and more",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
640,
447,
653
],
"spans": [
{
"bbox": [
106,
640,
447,
653
],
"score": 1.0,
"content": "complex model to take advantage of the advancement in neural network architecture.",
"type": "text"
}
],
"index": 34
}
],
"index": 30
},
{
"type": "text",
"bbox": [
107,
656,
505,
722
],
"lines": [
{
"bbox": [
106,
655,
505,
668
],
"spans": [
{
"bbox": [
106,
655,
505,
668
],
"score": 1.0,
"content": "Cross-Architecture Generalization: One desired property of our distilled data is that it generalizes",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
666,
505,
680
],
"spans": [
{
"bbox": [
105,
666,
505,
680
],
"score": 1.0,
"content": "well to architecture it has not seen during the training. Similar to previous works [5, 20], we evaluate",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
676,
506,
692
],
"spans": [
{
"bbox": [
105,
676,
506,
692
],
"score": 1.0,
"content": "the distilled data from CIFAR10 on a wide range of architectures which it has not seen during training,",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
689,
506,
701
],
"spans": [
{
"bbox": [
106,
689,
506,
701
],
"score": 1.0,
"content": "including AlexNet [38], VGG [39], and ResNet [40]. Table 2 shows that our method outperforms",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
699,
506,
713
],
"spans": [
{
"bbox": [
105,
699,
506,
713
],
"score": 1.0,
"content": "previous methods on all unseen architectures. Instance Normalization (IN) [41], as the vital ingredient",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
711,
505,
723
],
"spans": [
{
"bbox": [
105,
711,
505,
723
],
"score": 1.0,
"content": "in several methods (DSA, DM, MTT), seems to hurt the cross-architecture transfer. The performance",
"type": "text"
}
],
"index": 40
}
],
"index": 37.5
}
],
"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": "table",
"bbox": [
106,
122,
522,
209
],
"blocks": [
{
"type": "table_caption",
"bbox": [
106,
77,
506,
121
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
77,
505,
90
],
"spans": [
{
"bbox": [
105,
77,
370,
90
],
"score": 1.0,
"content": "Table 2: Cross-architecture transfer performance on CIFAR10 with",
"type": "text"
},
{
"bbox": [
370,
77,
415,
89
],
"score": 0.56,
"content": "1 0 \\mathrm { I m g / C l s }",
"type": "inline_equation"
},
{
"bbox": [
416,
77,
505,
90
],
"score": 1.0,
"content": ". Despite being trained",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
87,
506,
102
],
"spans": [
{
"bbox": [
105,
87,
506,
102
],
"score": 1.0,
"content": "for a specific architecture, our distilled data transfer well to various architectures unseen during",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
100,
506,
111
],
"spans": [
{
"bbox": [
106,
100,
506,
111
],
"score": 1.0,
"content": "training. Conv is the default evaluation model used for each method. NN, DN, IN, and BN stand for",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
110,
505,
123
],
"spans": [
{
"bbox": [
105,
110,
505,
123
],
"score": 1.0,
"content": "no normalization, default normalization, Instance Normalization, Batch Normalization respectively.",
"type": "text"
}
],
"index": 3
}
],
"index": 1.5
},
{
"type": "table_body",
"bbox": [
106,
122,
522,
209
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
122,
522,
209
],
"spans": [
{
"bbox": [
106,
122,
522,
209
],
"score": 0.97,
"html": " | Train Arch | Evaluation Architecture |
| Conv | Conv-NN | ResNet-DN | ResNet-BN | VGG-BN | AlexNet |
| DSA [7] | Conv-IN | 53.2 ± 0.8 | 36.4 ± 1.5 | 42.1 ± 0.7 | 34.1 ± 1.4 | 46.3 ± 1.3 | 34.0 ± 2.3 |
| DM[8] | Conv-IN | 49.2 ± 0.8 | 35.2 ± 0.5 | 36.8 ± 1.2 | 35.5 ± 1.3 | 41.2 ± 1.8 | 34.9 ± 1.1 |
| MTT[20] | Conv-IN | 64.4 ± 0.9 | 41.6 ± 1.3 | 49.2 ± 1.1 | 42.9 ± 1.5 | 46.6 ± 2.0 | 34.2 ± 2.6 |
| KIP [23] | Conv-NTK | 62.7 ± 0.3 | 58.2 ±0.4 | 49.0 ± 1.2 | 45.8 ± 1.4 | 30.1 ± 1.5 | 57.2 ± 0.4 |
| FRePo | Conv-BN | 65.5 ± 0.4 | 65.5 ± 0.4 | 58.1 ± 0.6 | 57.7 ± 0.7 | 59.4 ± 0.7 | 61.9 ± 0.7 |
",
"type": "table",
"image_path": "4e44b3dde3c2a554e1627b4447e7df3e28b023aaff4e028f17f27dce37cbd8da.jpg"
}
]
}
],
"index": 5,
"virtual_lines": [
{
"bbox": [
106,
122,
522,
151.0
],
"spans": [],
"index": 4
},
{
"bbox": [
106,
151.0,
522,
180.0
],
"spans": [],
"index": 5
},
{
"bbox": [
106,
180.0,
522,
209.0
],
"spans": [],
"index": 6
}
]
}
],
"index": 3.25
},
{
"type": "image",
"bbox": [
108,
226,
502,
336
],
"blocks": [
{
"type": "image_body",
"bbox": [
108,
226,
502,
336
],
"group_id": 0,
"lines": [
{
"bbox": [
108,
226,
502,
336
],
"spans": [
{
"bbox": [
108,
226,
502,
336
],
"score": 0.967,
"type": "image",
"image_path": "06d156480cb24f7de4d97dc51031b5fb9e76281f33da496484aeb238e841000d.jpg"
}
]
}
],
"index": 8,
"virtual_lines": [
{
"bbox": [
108,
226,
502,
262.6666666666667
],
"spans": [],
"index": 7
},
{
"bbox": [
108,
262.6666666666667,
502,
299.33333333333337
],
"spans": [],
"index": 8
},
{
"bbox": [
108,
299.33333333333337,
502,
336.00000000000006
],
"spans": [],
"index": 9
}
]
},
{
"type": "image_caption",
"bbox": [
106,
345,
506,
379
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
344,
505,
360
],
"spans": [
{
"bbox": [
105,
344,
505,
360
],
"score": 1.0,
"content": "Figure 4: (a,b,c) Distilled 1 img/cls from CIFAR100 using FRePo, MTT, and DSA. High quality",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
356,
507,
370
],
"spans": [
{
"bbox": [
105,
356,
507,
370
],
"score": 1.0,
"content": "images also produce high test accuracy. (d) Three categories of learned labels. (Top) High confidence,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
368,
491,
380
],
"spans": [
{
"bbox": [
106,
368,
491,
380
],
"score": 1.0,
"content": "large margin; (Middle) High confidence, small margin; (Bottom) Low confidence, small margin.",
"type": "text"
}
],
"index": 12
}
],
"index": 11
}
],
"index": 9.5
},
{
"type": "text",
"bbox": [
106,
405,
505,
547
],
"lines": [
{
"bbox": [
105,
404,
505,
418
],
"spans": [
{
"bbox": [
105,
404,
505,
418
],
"score": 1.0,
"content": "Label Learning: A similar trend can also be observed for label learning. When the dataset is simple",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
416,
506,
429
],
"spans": [
{
"bbox": [
106,
416,
506,
429
],
"score": 1.0,
"content": "and has only a few classes, label learning may not be necessary. However, it becomes crucial for",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
427,
506,
440
],
"spans": [
{
"bbox": [
105,
427,
506,
440
],
"score": 1.0,
"content": "complex datasets with many labels, such as CIFAR100 and Tiny-ImageNet (See more details in",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
437,
505,
450
],
"spans": [
{
"bbox": [
105,
437,
505,
450
],
"score": 1.0,
"content": "Appendix ??). Similar to the teacher label in the knowledge distillation [1], we observe that the",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
448,
506,
461
],
"spans": [
{
"bbox": [
105,
448,
506,
461
],
"score": 1.0,
"content": "distilled label also encodes the class similarity. We identify three typical cases in Figure 4d. The",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
460,
505,
472
],
"spans": [
{
"bbox": [
106,
460,
505,
472
],
"score": 1.0,
"content": "first group consists of highly confident labels with a much higher value for one class than other",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
470,
506,
483
],
"spans": [
{
"bbox": [
106,
470,
506,
483
],
"score": 1.0,
"content": "classes (large margin), such as sunflower, bicycle, and chair. In contrast, the distilled labels in the",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
481,
506,
494
],
"spans": [
{
"bbox": [
105,
481,
506,
494
],
"score": 1.0,
"content": "second group are confident but may get confused with some closely-related classes (small margin).",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
492,
507,
505
],
"spans": [
{
"bbox": [
105,
492,
507,
505
],
"score": 1.0,
"content": "For instance, the learned label for \"girl\" has almost equally high values for the girl, woman, man,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
502,
506,
517
],
"spans": [
{
"bbox": [
105,
502,
506,
517
],
"score": 1.0,
"content": "boy, and baby, suggesting that these classes are very similar and may be difficult for the model to",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
514,
506,
527
],
"spans": [
{
"bbox": [
105,
514,
506,
527
],
"score": 1.0,
"content": "distinguish them apart. The last group contains distilled labels with low values for all classes, such",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
525,
505,
537
],
"spans": [
{
"bbox": [
106,
525,
505,
537
],
"score": 1.0,
"content": "as bear, beaver, and squirrel. It is often hard for humans to recognize the distilled images in such a",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
537,
392,
549
],
"spans": [
{
"bbox": [
105,
537,
392,
549
],
"score": 1.0,
"content": "group, suggesting that they may be the challenging classes in a dataset.",
"type": "text"
}
],
"index": 25
}
],
"index": 19,
"bbox_fs": [
105,
404,
507,
549
]
},
{
"type": "text",
"bbox": [
106,
552,
505,
651
],
"lines": [
{
"bbox": [
105,
552,
506,
565
],
"spans": [
{
"bbox": [
105,
552,
506,
565
],
"score": 1.0,
"content": "Training Cost Analysis: Figure 3a, 3b shows that our method is significantly more time-efficient",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
563,
506,
576
],
"spans": [
{
"bbox": [
106,
563,
506,
576
],
"score": 1.0,
"content": "than the previous methods. When learning one image per class on CIFAR100, FRePo reaches a",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
574,
505,
587
],
"spans": [
{
"bbox": [
106,
574,
192,
587
],
"score": 1.0,
"content": "similar test accuracy",
"type": "text"
},
{
"bbox": [
192,
574,
226,
586
],
"score": 0.87,
"content": "( 2 3 . 4 \\% )",
"type": "inline_equation"
},
{
"bbox": [
226,
574,
335,
587
],
"score": 1.0,
"content": "to the second-best method",
"type": "text"
},
{
"bbox": [
335,
574,
369,
586
],
"score": 0.86,
"content": "( 2 4 . 0 \\% )",
"type": "inline_equation"
},
{
"bbox": [
369,
574,
505,
587
],
"score": 1.0,
"content": "in 38 seconds, compared to 3805",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
584,
505,
599
],
"spans": [
{
"bbox": [
105,
584,
484,
599
],
"score": 1.0,
"content": "seconds for MTT, which is roughly two orders of magnitude faster. Moreover, FRePo achieves",
"type": "text"
},
{
"bbox": [
485,
585,
505,
596
],
"score": 0.87,
"content": "92 \\%",
"type": "inline_equation"
}
],
"index": 29
},
{
"bbox": [
105,
596,
506,
609
],
"spans": [
{
"bbox": [
105,
596,
211,
609
],
"score": 1.0,
"content": "of its final test accuracy (",
"type": "text"
},
{
"bbox": [
211,
596,
239,
607
],
"score": 0.84,
"content": "2 6 . 4 \\%",
"type": "inline_equation"
},
{
"bbox": [
239,
596,
267,
609
],
"score": 1.0,
"content": "out of",
"type": "text"
},
{
"bbox": [
268,
596,
296,
607
],
"score": 0.84,
"content": "2 8 . 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
297,
596,
506,
609
],
"score": 1.0,
"content": ") in only 385 seconds. As shown in Figure 3c, our",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
607,
505,
620
],
"spans": [
{
"bbox": [
106,
607,
505,
620
],
"score": 1.0,
"content": "algorithm takes much less time to perform one gradient step on the distilled data. Thus, we can",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
618,
505,
631
],
"spans": [
{
"bbox": [
105,
618,
505,
631
],
"score": 1.0,
"content": "perform more gradient steps in a fixed time. Furthermore, Figure 3d suggests that our algorithm has",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
627,
506,
642
],
"spans": [
{
"bbox": [
105,
627,
506,
642
],
"score": 1.0,
"content": "much less GPU memory requirement. Therefore, we can potentially use a much larger and more",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
640,
447,
653
],
"spans": [
{
"bbox": [
106,
640,
447,
653
],
"score": 1.0,
"content": "complex model to take advantage of the advancement in neural network architecture.",
"type": "text"
}
],
"index": 34
}
],
"index": 30,
"bbox_fs": [
105,
552,
506,
653
]
},
{
"type": "text",
"bbox": [
107,
656,
505,
722
],
"lines": [
{
"bbox": [
106,
655,
505,
668
],
"spans": [
{
"bbox": [
106,
655,
505,
668
],
"score": 1.0,
"content": "Cross-Architecture Generalization: One desired property of our distilled data is that it generalizes",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
666,
505,
680
],
"spans": [
{
"bbox": [
105,
666,
505,
680
],
"score": 1.0,
"content": "well to architecture it has not seen during the training. Similar to previous works [5, 20], we evaluate",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
676,
506,
692
],
"spans": [
{
"bbox": [
105,
676,
506,
692
],
"score": 1.0,
"content": "the distilled data from CIFAR10 on a wide range of architectures which it has not seen during training,",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
689,
506,
701
],
"spans": [
{
"bbox": [
106,
689,
506,
701
],
"score": 1.0,
"content": "including AlexNet [38], VGG [39], and ResNet [40]. Table 2 shows that our method outperforms",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
699,
506,
713
],
"spans": [
{
"bbox": [
105,
699,
506,
713
],
"score": 1.0,
"content": "previous methods on all unseen architectures. Instance Normalization (IN) [41], as the vital ingredient",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
711,
505,
723
],
"spans": [
{
"bbox": [
105,
711,
505,
723
],
"score": 1.0,
"content": "in several methods (DSA, DM, MTT), seems to hurt the cross-architecture transfer. The performance",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
194,
506,
206
],
"spans": [
{
"bbox": [
106,
194,
506,
206
],
"score": 1.0,
"content": "degrades a lot when no normalization (NN) is applied (Conv-NN, AlexNet) or using a different",
"type": "text",
"cross_page": true
}
],
"index": 6
},
{
"bbox": [
105,
205,
505,
218
],
"spans": [
{
"bbox": [
105,
205,
505,
218
],
"score": 1.0,
"content": "normalization, like Batch Normalization (BN) [42]. It suggests that the distilled data generated by",
"type": "text",
"cross_page": true
}
],
"index": 7
},
{
"bbox": [
105,
216,
506,
228
],
"spans": [
{
"bbox": [
105,
216,
506,
228
],
"score": 1.0,
"content": "those methods encode the inductive bias of a particular training architecture. In contrast, our distilled",
"type": "text",
"cross_page": true
}
],
"index": 8
},
{
"bbox": [
106,
227,
506,
239
],
"spans": [
{
"bbox": [
106,
227,
506,
239
],
"score": 1.0,
"content": "data generalize well to various architectures, including those without normalization (Conv-NN,",
"type": "text",
"cross_page": true
}
],
"index": 9
},
{
"bbox": [
105,
237,
506,
250
],
"spans": [
{
"bbox": [
105,
237,
506,
250
],
"score": 1.0,
"content": "AlexNet). Note that Figure 1, 4 also indicate that our distilled data encode less architectural bias as",
"type": "text",
"cross_page": true
}
],
"index": 10
},
{
"bbox": [
105,
249,
506,
261
],
"spans": [
{
"bbox": [
105,
249,
506,
261
],
"score": 1.0,
"content": "the distilled images look natural and authentic. A simple idea to further alleviate the overfitting of a",
"type": "text",
"cross_page": true
}
],
"index": 11
},
{
"bbox": [
105,
259,
506,
273
],
"spans": [
{
"bbox": [
105,
259,
506,
273
],
"score": 1.0,
"content": "particular architecture is to include more architectures in the model pool. However, the training may",
"type": "text",
"cross_page": true
}
],
"index": 12
},
{
"bbox": [
105,
271,
474,
283
],
"spans": [
{
"bbox": [
105,
271,
474,
283
],
"score": 1.0,
"content": "not be stable as the meta-gradient computed by different architectures can be very different.",
"type": "text",
"cross_page": true
}
],
"index": 13
}
],
"index": 37.5,
"bbox_fs": [
105,
655,
506,
723
]
}
]
},
{
"preproc_blocks": [
{
"type": "table",
"bbox": [
119,
111,
492,
183
],
"blocks": [
{
"type": "table_caption",
"bbox": [
107,
77,
505,
110
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
75,
506,
91
],
"spans": [
{
"bbox": [
105,
75,
506,
91
],
"score": 1.0,
"content": "Table 3: Distillation performance on higher resolution (128x128) dataset (i.e. ImageNette, Image-",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
88,
505,
101
],
"spans": [
{
"bbox": [
105,
88,
505,
101
],
"score": 1.0,
"content": "Woof) and medium resolution (64x64) dataset with a complex label space (i.e. ImageNet-1K). FRePo",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
99,
461,
111
],
"spans": [
{
"bbox": [
106,
99,
461,
111
],
"score": 1.0,
"content": "scales to high-resolution images and learns the discriminate feature of complex datasets.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "table_body",
"bbox": [
119,
111,
492,
183
],
"group_id": 0,
"lines": [
{
"bbox": [
119,
111,
492,
183
],
"spans": [
{
"bbox": [
119,
111,
492,
183
],
"score": 0.977,
"html": " | ImageNette (128x128) | ImageWoof (128x128) | ImageNet (64x64) |
| Img/Cls | 1 | 10 | 1 | 10 | 1 | 2 |
| Random Subset | 23.5± 4.8 | 47.7 ± 2.4 | 14.2 ± 0.9 | 27.0± 1.9 | 1.1 ± 0.1 | 1.4 ± 0.1 |
| MTT[20] | 47.7± 0.9 | 63.0 ± 1.3 | 28.6 ± 0.8 | 35.8 ± 1.8 | 1 | 1 |
| FRePo | 48.1 ± 0.7 | 66.5 ± 0.8 | 29.7 ± 0.6 | 42.2 ± 0.9 | 7.5 ± 0.3 | 9.7 ± 0.2 |
",
"type": "table",
"image_path": "9e5f726c177726a5d4ac70778034da91123ecdb9239c20f47687477b0608342a.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
119,
111,
492,
135.0
],
"spans": [],
"index": 3
},
{
"bbox": [
119,
135.0,
492,
159.0
],
"spans": [],
"index": 4
},
{
"bbox": [
119,
159.0,
492,
183.0
],
"spans": [],
"index": 5
}
]
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
106,
194,
506,
282
],
"lines": [
{
"bbox": [
106,
194,
506,
206
],
"spans": [
{
"bbox": [
106,
194,
506,
206
],
"score": 1.0,
"content": "degrades a lot when no normalization (NN) is applied (Conv-NN, AlexNet) or using a different",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
205,
505,
218
],
"spans": [
{
"bbox": [
105,
205,
505,
218
],
"score": 1.0,
"content": "normalization, like Batch Normalization (BN) [42]. It suggests that the distilled data generated by",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
216,
506,
228
],
"spans": [
{
"bbox": [
105,
216,
506,
228
],
"score": 1.0,
"content": "those methods encode the inductive bias of a particular training architecture. In contrast, our distilled",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
227,
506,
239
],
"spans": [
{
"bbox": [
106,
227,
506,
239
],
"score": 1.0,
"content": "data generalize well to various architectures, including those without normalization (Conv-NN,",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
237,
506,
250
],
"spans": [
{
"bbox": [
105,
237,
506,
250
],
"score": 1.0,
"content": "AlexNet). Note that Figure 1, 4 also indicate that our distilled data encode less architectural bias as",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
249,
506,
261
],
"spans": [
{
"bbox": [
105,
249,
506,
261
],
"score": 1.0,
"content": "the distilled images look natural and authentic. A simple idea to further alleviate the overfitting of a",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
259,
506,
273
],
"spans": [
{
"bbox": [
105,
259,
506,
273
],
"score": 1.0,
"content": "particular architecture is to include more architectures in the model pool. However, the training may",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
271,
474,
283
],
"spans": [
{
"bbox": [
105,
271,
474,
283
],
"score": 1.0,
"content": "not be stable as the meta-gradient computed by different architectures can be very different.",
"type": "text"
}
],
"index": 13
}
],
"index": 9.5
},
{
"type": "title",
"bbox": [
107,
297,
172,
309
],
"lines": [
{
"bbox": [
105,
295,
174,
312
],
"spans": [
{
"bbox": [
105,
295,
174,
312
],
"score": 1.0,
"content": "4.3 ImageNet",
"type": "text"
}
],
"index": 14
}
],
"index": 14
},
{
"type": "text",
"bbox": [
107,
318,
505,
417
],
"lines": [
{
"bbox": [
106,
318,
505,
330
],
"spans": [
{
"bbox": [
106,
318,
505,
330
],
"score": 1.0,
"content": "High Resolution ImageNet Subset To understand how well our method performs on high-resolution",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
329,
506,
342
],
"spans": [
{
"bbox": [
105,
329,
506,
342
],
"score": 1.0,
"content": "images, we evaluate it on ImageNette and ImageWoof datasets [36] with a resolution of 128x128. We",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
340,
505,
352
],
"spans": [
{
"bbox": [
106,
340,
505,
352
],
"score": 1.0,
"content": "learn 1 and 10 images per class on both datasets and report the performance in Table 3 and visualize",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
351,
505,
364
],
"spans": [
{
"bbox": [
105,
351,
505,
364
],
"score": 1.0,
"content": "some distilled images in Figure 1. As shown in Table 3, we outperform MTT on all settings and",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
104,
361,
507,
376
],
"spans": [
{
"bbox": [
104,
361,
507,
376
],
"score": 1.0,
"content": "achieve much better performance when we distill ten images per class on a more difficult dataset",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
372,
505,
385
],
"spans": [
{
"bbox": [
106,
372,
505,
385
],
"score": 1.0,
"content": "ImageWoof. It suggests that our distilled data is better at capturing the discriminative features for",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
383,
506,
397
],
"spans": [
{
"bbox": [
105,
383,
506,
397
],
"score": 1.0,
"content": "each class. Figure 1 shows that our distilled images look real and capture the distinguishable feature",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
393,
507,
408
],
"spans": [
{
"bbox": [
105,
393,
507,
408
],
"score": 1.0,
"content": "of different classes. For the easy dataset (i.e., ImageNette), all images have clear different structures,",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
405,
371,
418
],
"spans": [
{
"bbox": [
105,
405,
371,
418
],
"score": 1.0,
"content": "while for ImageWoof, the texture of each dog seems to be crucial.",
"type": "text"
}
],
"index": 23
}
],
"index": 19
},
{
"type": "text",
"bbox": [
107,
422,
505,
466
],
"lines": [
{
"bbox": [
106,
422,
505,
434
],
"spans": [
{
"bbox": [
106,
422,
505,
434
],
"score": 1.0,
"content": "Resized ImageNet-1K: We also evaluate our method on a resized version of ILSVRC2012 [26] with",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
433,
505,
445
],
"spans": [
{
"bbox": [
105,
433,
505,
445
],
"score": 1.0,
"content": "a resolution of 64x64 to see how it performs on a complex label space. Surprisingly, we can achieve",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
443,
505,
457
],
"spans": [
{
"bbox": [
106,
444,
128,
455
],
"score": 0.86,
"content": "7 . 5 \\%",
"type": "inline_equation"
},
{
"bbox": [
129,
443,
146,
457
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
146,
444,
168,
455
],
"score": 0.84,
"content": "9 . 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
169,
443,
302,
457
],
"score": 1.0,
"content": "Top1 accuracy using only 1k and",
"type": "text"
},
{
"bbox": [
303,
444,
315,
454
],
"score": 0.38,
"content": "2 \\mathrm { k }",
"type": "inline_equation"
},
{
"bbox": [
315,
443,
443,
457
],
"score": 1.0,
"content": "training examples, compared to",
"type": "text"
},
{
"bbox": [
443,
444,
465,
454
],
"score": 0.85,
"content": "1 . 1 \\%",
"type": "inline_equation"
},
{
"bbox": [
465,
443,
483,
457
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
483,
444,
505,
454
],
"score": 0.84,
"content": "1 . 4 \\%",
"type": "inline_equation"
}
],
"index": 26
},
{
"bbox": [
105,
455,
245,
467
],
"spans": [
{
"bbox": [
105,
455,
245,
467
],
"score": 1.0,
"content": "using an equally-sized real subset.",
"type": "text"
}
],
"index": 27
}
],
"index": 25.5
},
{
"type": "title",
"bbox": [
107,
483,
185,
497
],
"lines": [
{
"bbox": [
104,
481,
187,
501
],
"spans": [
{
"bbox": [
104,
481,
187,
501
],
"score": 1.0,
"content": "5 Application",
"type": "text"
}
],
"index": 28
}
],
"index": 28
},
{
"type": "title",
"bbox": [
107,
509,
215,
522
],
"lines": [
{
"bbox": [
105,
507,
216,
524
],
"spans": [
{
"bbox": [
105,
507,
216,
524
],
"score": 1.0,
"content": "5.1 Continual Learning",
"type": "text"
}
],
"index": 29
}
],
"index": 29
},
{
"type": "text",
"bbox": [
107,
530,
505,
618
],
"lines": [
{
"bbox": [
106,
531,
506,
543
],
"spans": [
{
"bbox": [
106,
531,
506,
543
],
"score": 1.0,
"content": "Continual learning (CL) [43] aims to address the catastrophic forgetting problem [43–45] when a",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
542,
505,
554
],
"spans": [
{
"bbox": [
105,
542,
505,
554
],
"score": 1.0,
"content": "model learns sequentially from a stream of tasks. A commonly used strategy to recall past knowledge",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
553,
504,
565
],
"spans": [
{
"bbox": [
105,
553,
504,
565
],
"score": 1.0,
"content": "is based on a replay buffer, which stores representative samples from previous tasks [46–49]. Since",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
564,
505,
576
],
"spans": [
{
"bbox": [
105,
564,
505,
576
],
"score": 1.0,
"content": "sample selection is an important component of constructing an effective buffer [48–51], we believe",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
573,
506,
587
],
"spans": [
{
"bbox": [
105,
573,
506,
587
],
"score": 1.0,
"content": "distilled data can be a key ingredient for a continual learning algorithm due to its highly condensed",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
585,
505,
597
],
"spans": [
{
"bbox": [
105,
585,
505,
597
],
"score": 1.0,
"content": "nature. Several works [6–8, 52] have successfully applied the dataset distillation to the continual",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
597,
506,
609
],
"spans": [
{
"bbox": [
105,
597,
506,
609
],
"score": 1.0,
"content": "learning scenario. Our work shows that we can achieve much better results by using a better dataset",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
606,
195,
620
],
"spans": [
{
"bbox": [
105,
606,
195,
620
],
"score": 1.0,
"content": "distillation technique.",
"type": "text"
}
],
"index": 37
}
],
"index": 33.5
},
{
"type": "text",
"bbox": [
107,
623,
505,
722
],
"lines": [
{
"bbox": [
105,
623,
506,
636
],
"spans": [
{
"bbox": [
105,
623,
506,
636
],
"score": 1.0,
"content": "We follow Zhao and Bilen [8] that sets up the baseline based on GDumb [49] which greedily stores",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
633,
507,
647
],
"spans": [
{
"bbox": [
105,
633,
507,
647
],
"score": 1.0,
"content": "class-balanced training examples in memory and train model from scratch on the latest memory only.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
646,
505,
658
],
"spans": [
{
"bbox": [
105,
646,
505,
658
],
"score": 1.0,
"content": "In that case, the continual learning performance only depends on the quality of the replay buffer. We",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
655,
506,
670
],
"spans": [
{
"bbox": [
105,
655,
506,
670
],
"score": 1.0,
"content": "perform 5 and 10 step class-incremental learning [53] on CIFAR100 with an increasing buffer size",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
667,
506,
681
],
"spans": [
{
"bbox": [
105,
667,
506,
681
],
"score": 1.0,
"content": "of 20 images per class. Specifically, we distill 400 and 200 images at each step and put them into",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
677,
506,
691
],
"spans": [
{
"bbox": [
105,
677,
506,
691
],
"score": 1.0,
"content": "the replay buffer. We follow the same class split as Zhao and Bilen [8] and compare our method to",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
687,
505,
702
],
"spans": [
{
"bbox": [
105,
687,
505,
702
],
"score": 1.0,
"content": "random [49], herding [54, 55], DSA [7], and DM [8]. We use the default data preprocessing and",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
699,
505,
712
],
"spans": [
{
"bbox": [
105,
699,
505,
712
],
"score": 1.0,
"content": "default model for each method in this experiment as we find it gives the best performance for each",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
710,
481,
723
],
"spans": [
{
"bbox": [
105,
710,
481,
723
],
"score": 1.0,
"content": "method. We use the test accuracy on all observed classes as the performance measure [8, 48].",
"type": "text"
}
],
"index": 46
}
],
"index": 42
}
],
"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": "table",
"bbox": [
119,
111,
492,
183
],
"blocks": [
{
"type": "table_caption",
"bbox": [
107,
77,
505,
110
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
75,
506,
91
],
"spans": [
{
"bbox": [
105,
75,
506,
91
],
"score": 1.0,
"content": "Table 3: Distillation performance on higher resolution (128x128) dataset (i.e. ImageNette, Image-",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
88,
505,
101
],
"spans": [
{
"bbox": [
105,
88,
505,
101
],
"score": 1.0,
"content": "Woof) and medium resolution (64x64) dataset with a complex label space (i.e. ImageNet-1K). FRePo",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
99,
461,
111
],
"spans": [
{
"bbox": [
106,
99,
461,
111
],
"score": 1.0,
"content": "scales to high-resolution images and learns the discriminate feature of complex datasets.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "table_body",
"bbox": [
119,
111,
492,
183
],
"group_id": 0,
"lines": [
{
"bbox": [
119,
111,
492,
183
],
"spans": [
{
"bbox": [
119,
111,
492,
183
],
"score": 0.977,
"html": " | ImageNette (128x128) | ImageWoof (128x128) | ImageNet (64x64) |
| Img/Cls | 1 | 10 | 1 | 10 | 1 | 2 |
| Random Subset | 23.5± 4.8 | 47.7 ± 2.4 | 14.2 ± 0.9 | 27.0± 1.9 | 1.1 ± 0.1 | 1.4 ± 0.1 |
| MTT[20] | 47.7± 0.9 | 63.0 ± 1.3 | 28.6 ± 0.8 | 35.8 ± 1.8 | 1 | 1 |
| FRePo | 48.1 ± 0.7 | 66.5 ± 0.8 | 29.7 ± 0.6 | 42.2 ± 0.9 | 7.5 ± 0.3 | 9.7 ± 0.2 |
",
"type": "table",
"image_path": "9e5f726c177726a5d4ac70778034da91123ecdb9239c20f47687477b0608342a.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
119,
111,
492,
135.0
],
"spans": [],
"index": 3
},
{
"bbox": [
119,
135.0,
492,
159.0
],
"spans": [],
"index": 4
},
{
"bbox": [
119,
159.0,
492,
183.0
],
"spans": [],
"index": 5
}
]
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
106,
194,
506,
282
],
"lines": [],
"index": 9.5,
"bbox_fs": [
105,
194,
506,
283
],
"lines_deleted": true
},
{
"type": "title",
"bbox": [
107,
297,
172,
309
],
"lines": [
{
"bbox": [
105,
295,
174,
312
],
"spans": [
{
"bbox": [
105,
295,
174,
312
],
"score": 1.0,
"content": "4.3 ImageNet",
"type": "text"
}
],
"index": 14
}
],
"index": 14
},
{
"type": "text",
"bbox": [
107,
318,
505,
417
],
"lines": [
{
"bbox": [
106,
318,
505,
330
],
"spans": [
{
"bbox": [
106,
318,
505,
330
],
"score": 1.0,
"content": "High Resolution ImageNet Subset To understand how well our method performs on high-resolution",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
329,
506,
342
],
"spans": [
{
"bbox": [
105,
329,
506,
342
],
"score": 1.0,
"content": "images, we evaluate it on ImageNette and ImageWoof datasets [36] with a resolution of 128x128. We",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
340,
505,
352
],
"spans": [
{
"bbox": [
106,
340,
505,
352
],
"score": 1.0,
"content": "learn 1 and 10 images per class on both datasets and report the performance in Table 3 and visualize",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
351,
505,
364
],
"spans": [
{
"bbox": [
105,
351,
505,
364
],
"score": 1.0,
"content": "some distilled images in Figure 1. As shown in Table 3, we outperform MTT on all settings and",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
104,
361,
507,
376
],
"spans": [
{
"bbox": [
104,
361,
507,
376
],
"score": 1.0,
"content": "achieve much better performance when we distill ten images per class on a more difficult dataset",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
372,
505,
385
],
"spans": [
{
"bbox": [
106,
372,
505,
385
],
"score": 1.0,
"content": "ImageWoof. It suggests that our distilled data is better at capturing the discriminative features for",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
383,
506,
397
],
"spans": [
{
"bbox": [
105,
383,
506,
397
],
"score": 1.0,
"content": "each class. Figure 1 shows that our distilled images look real and capture the distinguishable feature",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
393,
507,
408
],
"spans": [
{
"bbox": [
105,
393,
507,
408
],
"score": 1.0,
"content": "of different classes. For the easy dataset (i.e., ImageNette), all images have clear different structures,",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
405,
371,
418
],
"spans": [
{
"bbox": [
105,
405,
371,
418
],
"score": 1.0,
"content": "while for ImageWoof, the texture of each dog seems to be crucial.",
"type": "text"
}
],
"index": 23
}
],
"index": 19,
"bbox_fs": [
104,
318,
507,
418
]
},
{
"type": "text",
"bbox": [
107,
422,
505,
466
],
"lines": [
{
"bbox": [
106,
422,
505,
434
],
"spans": [
{
"bbox": [
106,
422,
505,
434
],
"score": 1.0,
"content": "Resized ImageNet-1K: We also evaluate our method on a resized version of ILSVRC2012 [26] with",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
433,
505,
445
],
"spans": [
{
"bbox": [
105,
433,
505,
445
],
"score": 1.0,
"content": "a resolution of 64x64 to see how it performs on a complex label space. Surprisingly, we can achieve",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
443,
505,
457
],
"spans": [
{
"bbox": [
106,
444,
128,
455
],
"score": 0.86,
"content": "7 . 5 \\%",
"type": "inline_equation"
},
{
"bbox": [
129,
443,
146,
457
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
146,
444,
168,
455
],
"score": 0.84,
"content": "9 . 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
169,
443,
302,
457
],
"score": 1.0,
"content": "Top1 accuracy using only 1k and",
"type": "text"
},
{
"bbox": [
303,
444,
315,
454
],
"score": 0.38,
"content": "2 \\mathrm { k }",
"type": "inline_equation"
},
{
"bbox": [
315,
443,
443,
457
],
"score": 1.0,
"content": "training examples, compared to",
"type": "text"
},
{
"bbox": [
443,
444,
465,
454
],
"score": 0.85,
"content": "1 . 1 \\%",
"type": "inline_equation"
},
{
"bbox": [
465,
443,
483,
457
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
483,
444,
505,
454
],
"score": 0.84,
"content": "1 . 4 \\%",
"type": "inline_equation"
}
],
"index": 26
},
{
"bbox": [
105,
455,
245,
467
],
"spans": [
{
"bbox": [
105,
455,
245,
467
],
"score": 1.0,
"content": "using an equally-sized real subset.",
"type": "text"
}
],
"index": 27
}
],
"index": 25.5,
"bbox_fs": [
105,
422,
505,
467
]
},
{
"type": "title",
"bbox": [
107,
483,
185,
497
],
"lines": [
{
"bbox": [
104,
481,
187,
501
],
"spans": [
{
"bbox": [
104,
481,
187,
501
],
"score": 1.0,
"content": "5 Application",
"type": "text"
}
],
"index": 28
}
],
"index": 28
},
{
"type": "title",
"bbox": [
107,
509,
215,
522
],
"lines": [
{
"bbox": [
105,
507,
216,
524
],
"spans": [
{
"bbox": [
105,
507,
216,
524
],
"score": 1.0,
"content": "5.1 Continual Learning",
"type": "text"
}
],
"index": 29
}
],
"index": 29
},
{
"type": "text",
"bbox": [
107,
530,
505,
618
],
"lines": [
{
"bbox": [
106,
531,
506,
543
],
"spans": [
{
"bbox": [
106,
531,
506,
543
],
"score": 1.0,
"content": "Continual learning (CL) [43] aims to address the catastrophic forgetting problem [43–45] when a",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
542,
505,
554
],
"spans": [
{
"bbox": [
105,
542,
505,
554
],
"score": 1.0,
"content": "model learns sequentially from a stream of tasks. A commonly used strategy to recall past knowledge",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
553,
504,
565
],
"spans": [
{
"bbox": [
105,
553,
504,
565
],
"score": 1.0,
"content": "is based on a replay buffer, which stores representative samples from previous tasks [46–49]. Since",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
564,
505,
576
],
"spans": [
{
"bbox": [
105,
564,
505,
576
],
"score": 1.0,
"content": "sample selection is an important component of constructing an effective buffer [48–51], we believe",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
573,
506,
587
],
"spans": [
{
"bbox": [
105,
573,
506,
587
],
"score": 1.0,
"content": "distilled data can be a key ingredient for a continual learning algorithm due to its highly condensed",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
585,
505,
597
],
"spans": [
{
"bbox": [
105,
585,
505,
597
],
"score": 1.0,
"content": "nature. Several works [6–8, 52] have successfully applied the dataset distillation to the continual",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
597,
506,
609
],
"spans": [
{
"bbox": [
105,
597,
506,
609
],
"score": 1.0,
"content": "learning scenario. Our work shows that we can achieve much better results by using a better dataset",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
606,
195,
620
],
"spans": [
{
"bbox": [
105,
606,
195,
620
],
"score": 1.0,
"content": "distillation technique.",
"type": "text"
}
],
"index": 37
}
],
"index": 33.5,
"bbox_fs": [
105,
531,
506,
620
]
},
{
"type": "text",
"bbox": [
107,
623,
505,
722
],
"lines": [
{
"bbox": [
105,
623,
506,
636
],
"spans": [
{
"bbox": [
105,
623,
506,
636
],
"score": 1.0,
"content": "We follow Zhao and Bilen [8] that sets up the baseline based on GDumb [49] which greedily stores",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
633,
507,
647
],
"spans": [
{
"bbox": [
105,
633,
507,
647
],
"score": 1.0,
"content": "class-balanced training examples in memory and train model from scratch on the latest memory only.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
646,
505,
658
],
"spans": [
{
"bbox": [
105,
646,
505,
658
],
"score": 1.0,
"content": "In that case, the continual learning performance only depends on the quality of the replay buffer. We",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
655,
506,
670
],
"spans": [
{
"bbox": [
105,
655,
506,
670
],
"score": 1.0,
"content": "perform 5 and 10 step class-incremental learning [53] on CIFAR100 with an increasing buffer size",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
667,
506,
681
],
"spans": [
{
"bbox": [
105,
667,
506,
681
],
"score": 1.0,
"content": "of 20 images per class. Specifically, we distill 400 and 200 images at each step and put them into",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
677,
506,
691
],
"spans": [
{
"bbox": [
105,
677,
506,
691
],
"score": 1.0,
"content": "the replay buffer. We follow the same class split as Zhao and Bilen [8] and compare our method to",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
687,
505,
702
],
"spans": [
{
"bbox": [
105,
687,
505,
702
],
"score": 1.0,
"content": "random [49], herding [54, 55], DSA [7], and DM [8]. We use the default data preprocessing and",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
699,
505,
712
],
"spans": [
{
"bbox": [
105,
699,
505,
712
],
"score": 1.0,
"content": "default model for each method in this experiment as we find it gives the best performance for each",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
710,
481,
723
],
"spans": [
{
"bbox": [
105,
710,
481,
723
],
"score": 1.0,
"content": "method. We use the test accuracy on all observed classes as the performance measure [8, 48].",
"type": "text"
}
],
"index": 46
}
],
"index": 42,
"bbox_fs": [
105,
623,
507,
723
]
}
]
},
{
"preproc_blocks": [
{
"type": "table",
"bbox": [
115,
111,
495,
198
],
"blocks": [
{
"type": "table_caption",
"bbox": [
107,
77,
505,
110
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
77,
505,
89
],
"spans": [
{
"bbox": [
105,
77,
505,
89
],
"score": 1.0,
"content": "Table 4: AUC of five attackers on models trained on the real and distilled MNIST data. The model",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
89,
505,
100
],
"spans": [
{
"bbox": [
106,
89,
505,
100
],
"score": 1.0,
"content": "trained on the real data is vulnerable to MIAs, while the model trained on the distilled data is robust",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
98,
505,
113
],
"spans": [
{
"bbox": [
105,
98,
505,
113
],
"score": 1.0,
"content": "to MIAs. Training on distilled data allows privacy preservation while retaining model performance.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "table_body",
"bbox": [
115,
111,
495,
198
],
"group_id": 0,
"lines": [
{
"bbox": [
115,
111,
495,
198
],
"spans": [
{
"bbox": [
115,
111,
495,
198
],
"score": 0.944,
"html": " | Test Acc (%) | Attack AUC |
| Threshold | LR | MLP | RF | KNN |
| Real | 99.2 ± 0.1 | 0.99 ± 0.01 | 0.99 ± 0.00 | 1.00 ±0.00 | 1.00 ± 0.00 | 0.97 ±0.00 |
| Subset | 96.8± 0.2 | 0.52 ±0.00 | 0.50 ± 0.01 | 0.53 ± 0.01 | 0.55 ± 0.00 | 0.54 ±0.00 |
| DSA | 98.5 ± 0.1 | 0.50 ± 0.00 | 0.51 ± 0.00 | 0.54 ± 0.00 | 0.54 ± 0.01 | 0.54 ± 0.01 |
| DM | 98.3 ± 0.0 | 0.50 ± 0.00 | 0.51 ± 0.01 | 0.54 ± 0.01 | 0.54 ± 0.01 | 0.53 ± 0.01 |
| FRePo | 98.5± 0.1 | 0.52 ±0.00 | 0.51 ± 0.00 | 0.53 ± 0.01 | 0.52 ± 0.01 | 0.51 ± 0.01 |
",
"type": "table",
"image_path": "8d5bbd95209fde91ac8ff75261cfe03055c38b7bc922e6db2fdbfc338122aed9.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
115,
111,
495,
140.0
],
"spans": [],
"index": 3
},
{
"bbox": [
115,
140.0,
495,
169.0
],
"spans": [],
"index": 4
},
{
"bbox": [
115,
169.0,
495,
198.0
],
"spans": [],
"index": 5
}
]
}
],
"index": 2.5
},
{
"type": "image",
"bbox": [
111,
215,
500,
302
],
"blocks": [
{
"type": "image_body",
"bbox": [
111,
215,
500,
302
],
"group_id": 0,
"lines": [
{
"bbox": [
111,
215,
500,
302
],
"spans": [
{
"bbox": [
111,
215,
500,
302
],
"score": 0.953,
"type": "image",
"image_path": "3425d57f4bfa2229b9a55a067323e25dd62e8fcc52fb803fdd882c3fcc7566d9.jpg"
}
]
}
],
"index": 7,
"virtual_lines": [
{
"bbox": [
111,
215,
500,
244.0
],
"spans": [],
"index": 6
},
{
"bbox": [
111,
244.0,
500,
273.0
],
"spans": [],
"index": 7
},
{
"bbox": [
111,
273.0,
500,
302.0
],
"spans": [],
"index": 8
}
]
},
{
"type": "image_caption",
"bbox": [
106,
308,
505,
363
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
308,
505,
320
],
"spans": [
{
"bbox": [
106,
308,
505,
320
],
"score": 1.0,
"content": "Figure 5: (a,b) Multi-class accuracies across all classes observed up to a certain time point. We",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
318,
505,
331
],
"spans": [
{
"bbox": [
105,
318,
505,
331
],
"score": 1.0,
"content": "perform significantly better than other methods in both 5 and 10 step class-incremental continual",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
329,
506,
344
],
"spans": [
{
"bbox": [
105,
329,
506,
344
],
"score": 1.0,
"content": "learning. (c,d) Test accuracy and attack AUC as we increase the number of training steps. AUC keeps",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
340,
505,
353
],
"spans": [
{
"bbox": [
105,
340,
505,
353
],
"score": 1.0,
"content": "increasing when training a model on the real data for more steps. In contrast, AUC keeps low when",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
106,
352,
209,
365
],
"spans": [
{
"bbox": [
106,
352,
209,
365
],
"score": 1.0,
"content": "training on distilled data.",
"type": "text"
}
],
"index": 13
}
],
"index": 11
}
],
"index": 9.0
},
{
"type": "text",
"bbox": [
106,
387,
505,
475
],
"lines": [
{
"bbox": [
106,
388,
505,
399
],
"spans": [
{
"bbox": [
106,
388,
505,
399
],
"score": 1.0,
"content": "Figure 5 shows that our method performs significantly better than all previous methods. The final",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
398,
506,
410
],
"spans": [
{
"bbox": [
105,
398,
474,
410
],
"score": 1.0,
"content": "test accuracy for all classes for our method (FRePo) and the second-best method (DM) are",
"type": "text"
},
{
"bbox": [
475,
398,
502,
409
],
"score": 0.86,
"content": "4 1 . 6 \\%",
"type": "inline_equation"
},
{
"bbox": [
502,
398,
506,
410
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
409,
506,
423
],
"spans": [
{
"bbox": [
106,
409,
133,
420
],
"score": 0.87,
"content": "3 3 . 9 \\%",
"type": "inline_equation"
},
{
"bbox": [
134,
409,
225,
423
],
"score": 1.0,
"content": "in 5-step learning, and",
"type": "text"
},
{
"bbox": [
225,
410,
252,
420
],
"score": 0.83,
"content": "3 8 . 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
252,
409,
255,
423
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
255,
410,
282,
420
],
"score": 0.82,
"content": "3 4 . 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
282,
409,
506,
423
],
"score": 1.0,
"content": "in 10-step learning. However, we notice that for FRePo,",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
419,
506,
433
],
"spans": [
{
"bbox": [
105,
419,
458,
433
],
"score": 1.0,
"content": "distilling 2000 images in a continual learning setup achieves a similar test accuracy",
"type": "text"
},
{
"bbox": [
458,
420,
492,
432
],
"score": 0.87,
"content": "( 4 1 . 6 \\% )",
"type": "inline_equation"
},
{
"bbox": [
492,
419,
506,
433
],
"score": 1.0,
"content": "as",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
430,
506,
444
],
"spans": [
{
"bbox": [
105,
430,
314,
444
],
"score": 1.0,
"content": "distilling only 1000 images from the whole dataset",
"type": "text"
},
{
"bbox": [
314,
431,
342,
442
],
"score": 0.83,
"content": "( 4 1 . 3 \\%",
"type": "inline_equation"
},
{
"bbox": [
343,
430,
506,
444
],
"score": 1.0,
"content": "from Table 1). In addition, performance",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
442,
506,
455
],
"spans": [
{
"bbox": [
105,
442,
506,
455
],
"score": 1.0,
"content": "drops as we perform more steps. It suggests that FRePo considers all available classes to derive",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
452,
506,
465
],
"spans": [
{
"bbox": [
105,
452,
506,
465
],
"score": 1.0,
"content": "the most condensed dataset. Splitting the data into multiple groups and performing independent",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
464,
482,
476
],
"spans": [
{
"bbox": [
106,
464,
482,
476
],
"score": 1.0,
"content": "distillation may generate redundant information or fail to capture the distinguishable features.",
"type": "text"
}
],
"index": 21
}
],
"index": 17.5
},
{
"type": "title",
"bbox": [
108,
491,
263,
503
],
"lines": [
{
"bbox": [
105,
490,
264,
505
],
"spans": [
{
"bbox": [
105,
490,
264,
505
],
"score": 1.0,
"content": "5.2 Membership Inference Defense",
"type": "text"
}
],
"index": 22
}
],
"index": 22
},
{
"type": "text",
"bbox": [
106,
508,
505,
597
],
"lines": [
{
"bbox": [
106,
509,
505,
521
],
"spans": [
{
"bbox": [
106,
509,
505,
521
],
"score": 1.0,
"content": "Membership inference attacks (MIA) aim to infer whether a given data point has been used to train",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
520,
506,
532
],
"spans": [
{
"bbox": [
106,
520,
506,
532
],
"score": 1.0,
"content": "the model or not [56–58]. Ideally, we want a model to learn from the data but not memorize it to",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
531,
505,
542
],
"spans": [
{
"bbox": [
105,
531,
505,
542
],
"score": 1.0,
"content": "preserve privacy. However, deep neural networks are well-known for their ability to memorize all",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
542,
505,
553
],
"spans": [
{
"bbox": [
105,
542,
505,
553
],
"score": 1.0,
"content": "the training examples, even on large and randomly labeled datasets [59]. Several methods have been",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
552,
506,
567
],
"spans": [
{
"bbox": [
105,
552,
506,
567
],
"score": 1.0,
"content": "proposed to defend against such attacks by either modifying the training procedure [60] or changing",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
563,
505,
576
],
"spans": [
{
"bbox": [
106,
563,
505,
576
],
"score": 1.0,
"content": "the inference workflow [61]. This section shows that the distilled data contain little information",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
574,
507,
587
],
"spans": [
{
"bbox": [
105,
574,
507,
587
],
"score": 1.0,
"content": "regarding sample presence in the original dataset. Thus, instead of training on the original datasets,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
586,
463,
598
],
"spans": [
{
"bbox": [
106,
586,
463,
598
],
"score": 1.0,
"content": "training on distilled data allows privacy preservation while retaining model performance.",
"type": "text"
}
],
"index": 30
}
],
"index": 26.5
},
{
"type": "text",
"bbox": [
106,
601,
505,
722
],
"lines": [
{
"bbox": [
105,
601,
505,
614
],
"spans": [
{
"bbox": [
105,
601,
505,
614
],
"score": 1.0,
"content": "We consider three distilled data generated by DSA [7], DM [8] and FRePo. We perform five popular",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
612,
505,
625
],
"spans": [
{
"bbox": [
106,
612,
505,
625
],
"score": 1.0,
"content": "\"black box\" MIA provided by Tensorflow Privacy [62] on models trained on the real data or the data",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
622,
506,
637
],
"spans": [
{
"bbox": [
105,
622,
506,
637
],
"score": 1.0,
"content": "distilled from it. The attack methods include a threshold attack and four model-based attacks using",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
635,
505,
647
],
"spans": [
{
"bbox": [
105,
635,
505,
647
],
"score": 1.0,
"content": "logistic regression (LR), multi-layer perceptron (MLP), random forest (RF) and K-nearest neighbor",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
644,
506,
658
],
"spans": [
{
"bbox": [
105,
644,
506,
658
],
"score": 1.0,
"content": "(KNN). The inputs to those attack methods are ground-truth labels, model predictions, and losses. To",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
657,
506,
669
],
"spans": [
{
"bbox": [
105,
657,
506,
669
],
"score": 1.0,
"content": "measure the privacy vulnerability of the trained model, we compute the area under the ROC curve",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
667,
505,
680
],
"spans": [
{
"bbox": [
105,
667,
505,
680
],
"score": 1.0,
"content": "(AUC) of an attack classifier. Following prior work, [56, 63], we keep a balanced set of training",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
678,
505,
691
],
"spans": [
{
"bbox": [
105,
678,
505,
691
],
"score": 1.0,
"content": "examples (member) and test examples (non-member) with 10K each to maximize the uncertainty of",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
689,
505,
701
],
"spans": [
{
"bbox": [
105,
689,
321,
701
],
"score": 1.0,
"content": "MIA. Thus, the random guessing strategy results in a",
"type": "text"
},
{
"bbox": [
321,
689,
340,
699
],
"score": 0.87,
"content": "50 \\%",
"type": "inline_equation"
},
{
"bbox": [
341,
689,
505,
701
],
"score": 1.0,
"content": "MIA accuracy. We conduct experiments",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
699,
505,
713
],
"spans": [
{
"bbox": [
105,
699,
505,
713
],
"score": 1.0,
"content": "on MNIST and FashionMNIST with a distillation size of 500. For space reasons, we provide more",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
711,
297,
723
],
"spans": [
{
"bbox": [
106,
711,
297,
723
],
"score": 1.0,
"content": "implementation details and results in appendix.",
"type": "text"
}
],
"index": 41
}
],
"index": 36
}
],
"page_idx": 8,
"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": "9",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "table",
"bbox": [
115,
111,
495,
198
],
"blocks": [
{
"type": "table_caption",
"bbox": [
107,
77,
505,
110
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
77,
505,
89
],
"spans": [
{
"bbox": [
105,
77,
505,
89
],
"score": 1.0,
"content": "Table 4: AUC of five attackers on models trained on the real and distilled MNIST data. The model",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
89,
505,
100
],
"spans": [
{
"bbox": [
106,
89,
505,
100
],
"score": 1.0,
"content": "trained on the real data is vulnerable to MIAs, while the model trained on the distilled data is robust",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
98,
505,
113
],
"spans": [
{
"bbox": [
105,
98,
505,
113
],
"score": 1.0,
"content": "to MIAs. Training on distilled data allows privacy preservation while retaining model performance.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "table_body",
"bbox": [
115,
111,
495,
198
],
"group_id": 0,
"lines": [
{
"bbox": [
115,
111,
495,
198
],
"spans": [
{
"bbox": [
115,
111,
495,
198
],
"score": 0.944,
"html": " | Test Acc (%) | Attack AUC |
| Threshold | LR | MLP | RF | KNN |
| Real | 99.2 ± 0.1 | 0.99 ± 0.01 | 0.99 ± 0.00 | 1.00 ±0.00 | 1.00 ± 0.00 | 0.97 ±0.00 |
| Subset | 96.8± 0.2 | 0.52 ±0.00 | 0.50 ± 0.01 | 0.53 ± 0.01 | 0.55 ± 0.00 | 0.54 ±0.00 |
| DSA | 98.5 ± 0.1 | 0.50 ± 0.00 | 0.51 ± 0.00 | 0.54 ± 0.00 | 0.54 ± 0.01 | 0.54 ± 0.01 |
| DM | 98.3 ± 0.0 | 0.50 ± 0.00 | 0.51 ± 0.01 | 0.54 ± 0.01 | 0.54 ± 0.01 | 0.53 ± 0.01 |
| FRePo | 98.5± 0.1 | 0.52 ±0.00 | 0.51 ± 0.00 | 0.53 ± 0.01 | 0.52 ± 0.01 | 0.51 ± 0.01 |
",
"type": "table",
"image_path": "8d5bbd95209fde91ac8ff75261cfe03055c38b7bc922e6db2fdbfc338122aed9.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
115,
111,
495,
140.0
],
"spans": [],
"index": 3
},
{
"bbox": [
115,
140.0,
495,
169.0
],
"spans": [],
"index": 4
},
{
"bbox": [
115,
169.0,
495,
198.0
],
"spans": [],
"index": 5
}
]
}
],
"index": 2.5
},
{
"type": "image",
"bbox": [
111,
215,
500,
302
],
"blocks": [
{
"type": "image_body",
"bbox": [
111,
215,
500,
302
],
"group_id": 0,
"lines": [
{
"bbox": [
111,
215,
500,
302
],
"spans": [
{
"bbox": [
111,
215,
500,
302
],
"score": 0.953,
"type": "image",
"image_path": "3425d57f4bfa2229b9a55a067323e25dd62e8fcc52fb803fdd882c3fcc7566d9.jpg"
}
]
}
],
"index": 7,
"virtual_lines": [
{
"bbox": [
111,
215,
500,
244.0
],
"spans": [],
"index": 6
},
{
"bbox": [
111,
244.0,
500,
273.0
],
"spans": [],
"index": 7
},
{
"bbox": [
111,
273.0,
500,
302.0
],
"spans": [],
"index": 8
}
]
},
{
"type": "image_caption",
"bbox": [
106,
308,
505,
363
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
308,
505,
320
],
"spans": [
{
"bbox": [
106,
308,
505,
320
],
"score": 1.0,
"content": "Figure 5: (a,b) Multi-class accuracies across all classes observed up to a certain time point. We",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
318,
505,
331
],
"spans": [
{
"bbox": [
105,
318,
505,
331
],
"score": 1.0,
"content": "perform significantly better than other methods in both 5 and 10 step class-incremental continual",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
329,
506,
344
],
"spans": [
{
"bbox": [
105,
329,
506,
344
],
"score": 1.0,
"content": "learning. (c,d) Test accuracy and attack AUC as we increase the number of training steps. AUC keeps",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
340,
505,
353
],
"spans": [
{
"bbox": [
105,
340,
505,
353
],
"score": 1.0,
"content": "increasing when training a model on the real data for more steps. In contrast, AUC keeps low when",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
106,
352,
209,
365
],
"spans": [
{
"bbox": [
106,
352,
209,
365
],
"score": 1.0,
"content": "training on distilled data.",
"type": "text"
}
],
"index": 13
}
],
"index": 11
}
],
"index": 9.0
},
{
"type": "text",
"bbox": [
106,
387,
505,
475
],
"lines": [
{
"bbox": [
106,
388,
505,
399
],
"spans": [
{
"bbox": [
106,
388,
505,
399
],
"score": 1.0,
"content": "Figure 5 shows that our method performs significantly better than all previous methods. The final",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
398,
506,
410
],
"spans": [
{
"bbox": [
105,
398,
474,
410
],
"score": 1.0,
"content": "test accuracy for all classes for our method (FRePo) and the second-best method (DM) are",
"type": "text"
},
{
"bbox": [
475,
398,
502,
409
],
"score": 0.86,
"content": "4 1 . 6 \\%",
"type": "inline_equation"
},
{
"bbox": [
502,
398,
506,
410
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
409,
506,
423
],
"spans": [
{
"bbox": [
106,
409,
133,
420
],
"score": 0.87,
"content": "3 3 . 9 \\%",
"type": "inline_equation"
},
{
"bbox": [
134,
409,
225,
423
],
"score": 1.0,
"content": "in 5-step learning, and",
"type": "text"
},
{
"bbox": [
225,
410,
252,
420
],
"score": 0.83,
"content": "3 8 . 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
252,
409,
255,
423
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
255,
410,
282,
420
],
"score": 0.82,
"content": "3 4 . 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
282,
409,
506,
423
],
"score": 1.0,
"content": "in 10-step learning. However, we notice that for FRePo,",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
419,
506,
433
],
"spans": [
{
"bbox": [
105,
419,
458,
433
],
"score": 1.0,
"content": "distilling 2000 images in a continual learning setup achieves a similar test accuracy",
"type": "text"
},
{
"bbox": [
458,
420,
492,
432
],
"score": 0.87,
"content": "( 4 1 . 6 \\% )",
"type": "inline_equation"
},
{
"bbox": [
492,
419,
506,
433
],
"score": 1.0,
"content": "as",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
430,
506,
444
],
"spans": [
{
"bbox": [
105,
430,
314,
444
],
"score": 1.0,
"content": "distilling only 1000 images from the whole dataset",
"type": "text"
},
{
"bbox": [
314,
431,
342,
442
],
"score": 0.83,
"content": "( 4 1 . 3 \\%",
"type": "inline_equation"
},
{
"bbox": [
343,
430,
506,
444
],
"score": 1.0,
"content": "from Table 1). In addition, performance",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
442,
506,
455
],
"spans": [
{
"bbox": [
105,
442,
506,
455
],
"score": 1.0,
"content": "drops as we perform more steps. It suggests that FRePo considers all available classes to derive",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
452,
506,
465
],
"spans": [
{
"bbox": [
105,
452,
506,
465
],
"score": 1.0,
"content": "the most condensed dataset. Splitting the data into multiple groups and performing independent",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
464,
482,
476
],
"spans": [
{
"bbox": [
106,
464,
482,
476
],
"score": 1.0,
"content": "distillation may generate redundant information or fail to capture the distinguishable features.",
"type": "text"
}
],
"index": 21
}
],
"index": 17.5,
"bbox_fs": [
105,
388,
506,
476
]
},
{
"type": "title",
"bbox": [
108,
491,
263,
503
],
"lines": [
{
"bbox": [
105,
490,
264,
505
],
"spans": [
{
"bbox": [
105,
490,
264,
505
],
"score": 1.0,
"content": "5.2 Membership Inference Defense",
"type": "text"
}
],
"index": 22
}
],
"index": 22
},
{
"type": "text",
"bbox": [
106,
508,
505,
597
],
"lines": [
{
"bbox": [
106,
509,
505,
521
],
"spans": [
{
"bbox": [
106,
509,
505,
521
],
"score": 1.0,
"content": "Membership inference attacks (MIA) aim to infer whether a given data point has been used to train",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
520,
506,
532
],
"spans": [
{
"bbox": [
106,
520,
506,
532
],
"score": 1.0,
"content": "the model or not [56–58]. Ideally, we want a model to learn from the data but not memorize it to",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
531,
505,
542
],
"spans": [
{
"bbox": [
105,
531,
505,
542
],
"score": 1.0,
"content": "preserve privacy. However, deep neural networks are well-known for their ability to memorize all",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
542,
505,
553
],
"spans": [
{
"bbox": [
105,
542,
505,
553
],
"score": 1.0,
"content": "the training examples, even on large and randomly labeled datasets [59]. Several methods have been",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
552,
506,
567
],
"spans": [
{
"bbox": [
105,
552,
506,
567
],
"score": 1.0,
"content": "proposed to defend against such attacks by either modifying the training procedure [60] or changing",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
563,
505,
576
],
"spans": [
{
"bbox": [
106,
563,
505,
576
],
"score": 1.0,
"content": "the inference workflow [61]. This section shows that the distilled data contain little information",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
574,
507,
587
],
"spans": [
{
"bbox": [
105,
574,
507,
587
],
"score": 1.0,
"content": "regarding sample presence in the original dataset. Thus, instead of training on the original datasets,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
586,
463,
598
],
"spans": [
{
"bbox": [
106,
586,
463,
598
],
"score": 1.0,
"content": "training on distilled data allows privacy preservation while retaining model performance.",
"type": "text"
}
],
"index": 30
}
],
"index": 26.5,
"bbox_fs": [
105,
509,
507,
598
]
},
{
"type": "text",
"bbox": [
106,
601,
505,
722
],
"lines": [
{
"bbox": [
105,
601,
505,
614
],
"spans": [
{
"bbox": [
105,
601,
505,
614
],
"score": 1.0,
"content": "We consider three distilled data generated by DSA [7], DM [8] and FRePo. We perform five popular",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
612,
505,
625
],
"spans": [
{
"bbox": [
106,
612,
505,
625
],
"score": 1.0,
"content": "\"black box\" MIA provided by Tensorflow Privacy [62] on models trained on the real data or the data",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
622,
506,
637
],
"spans": [
{
"bbox": [
105,
622,
506,
637
],
"score": 1.0,
"content": "distilled from it. The attack methods include a threshold attack and four model-based attacks using",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
635,
505,
647
],
"spans": [
{
"bbox": [
105,
635,
505,
647
],
"score": 1.0,
"content": "logistic regression (LR), multi-layer perceptron (MLP), random forest (RF) and K-nearest neighbor",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
644,
506,
658
],
"spans": [
{
"bbox": [
105,
644,
506,
658
],
"score": 1.0,
"content": "(KNN). The inputs to those attack methods are ground-truth labels, model predictions, and losses. To",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
657,
506,
669
],
"spans": [
{
"bbox": [
105,
657,
506,
669
],
"score": 1.0,
"content": "measure the privacy vulnerability of the trained model, we compute the area under the ROC curve",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
667,
505,
680
],
"spans": [
{
"bbox": [
105,
667,
505,
680
],
"score": 1.0,
"content": "(AUC) of an attack classifier. Following prior work, [56, 63], we keep a balanced set of training",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
678,
505,
691
],
"spans": [
{
"bbox": [
105,
678,
505,
691
],
"score": 1.0,
"content": "examples (member) and test examples (non-member) with 10K each to maximize the uncertainty of",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
689,
505,
701
],
"spans": [
{
"bbox": [
105,
689,
321,
701
],
"score": 1.0,
"content": "MIA. Thus, the random guessing strategy results in a",
"type": "text"
},
{
"bbox": [
321,
689,
340,
699
],
"score": 0.87,
"content": "50 \\%",
"type": "inline_equation"
},
{
"bbox": [
341,
689,
505,
701
],
"score": 1.0,
"content": "MIA accuracy. We conduct experiments",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
699,
505,
713
],
"spans": [
{
"bbox": [
105,
699,
505,
713
],
"score": 1.0,
"content": "on MNIST and FashionMNIST with a distillation size of 500. For space reasons, we provide more",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
711,
297,
723
],
"spans": [
{
"bbox": [
106,
711,
297,
723
],
"score": 1.0,
"content": "implementation details and results in appendix.",
"type": "text"
}
],
"index": 41
}
],
"index": 36,
"bbox_fs": [
105,
601,
506,
723
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
106,
72,
505,
182
],
"lines": [
{
"bbox": [
105,
73,
505,
85
],
"spans": [
{
"bbox": [
105,
73,
505,
85
],
"score": 1.0,
"content": "As shown in Table 4, all models trained on the distilled data preserve privacy as their attack AUCs",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
83,
506,
96
],
"spans": [
{
"bbox": [
105,
83,
506,
96
],
"score": 1.0,
"content": "are closed to random guessing. However, we observe a small drop in test accuracy compared to",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
93,
506,
107
],
"spans": [
{
"bbox": [
105,
93,
506,
107
],
"score": 1.0,
"content": "the model trained on the full dataset, which is expected as we only distill 500 examples instead of",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
105,
506,
118
],
"spans": [
{
"bbox": [
105,
105,
506,
118
],
"score": 1.0,
"content": "10,000 examples. Compared to the model trained on an equally sized subset of the original data, the",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
116,
506,
129
],
"spans": [
{
"bbox": [
105,
116,
506,
129
],
"score": 1.0,
"content": "model trained on distilled data results in much better test performance. Figure 5c, 5d demonstrate the",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
128,
505,
139
],
"spans": [
{
"bbox": [
106,
128,
505,
139
],
"score": 1.0,
"content": "trade-off between test accuracy and attack effectiveness as measured by ROC AUC. It shows that",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
138,
505,
150
],
"spans": [
{
"bbox": [
106,
138,
505,
150
],
"score": 1.0,
"content": "early stopping can be an effective technique to preserve privacy. However, we will still be under high",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
104,
147,
506,
162
],
"spans": [
{
"bbox": [
104,
147,
506,
162
],
"score": 1.0,
"content": "MIA risk if we perform early stopping by monitoring the validation loss. In contrast, training a model",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
159,
505,
173
],
"spans": [
{
"bbox": [
105,
159,
505,
173
],
"score": 1.0,
"content": "on the distilled data does not have this problem as the attack AUCs keep at a very low level regardless",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
171,
176,
184
],
"spans": [
{
"bbox": [
105,
171,
176,
184
],
"score": 1.0,
"content": "of training steps.",
"type": "text"
}
],
"index": 9
}
],
"index": 4.5
},
{
"type": "title",
"bbox": [
107,
198,
183,
211
],
"lines": [
{
"bbox": [
105,
196,
185,
214
],
"spans": [
{
"bbox": [
105,
196,
185,
214
],
"score": 1.0,
"content": "6 Conclusion",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
107,
223,
505,
299
],
"lines": [
{
"bbox": [
106,
222,
506,
236
],
"spans": [
{
"bbox": [
106,
222,
506,
236
],
"score": 1.0,
"content": "We propose neural Feature Regression with Pooling (FRePo) to overcome two challenges in dataset",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
233,
506,
247
],
"spans": [
{
"bbox": [
105,
233,
506,
247
],
"score": 1.0,
"content": "distillation: meta-gradient computation and various types of overfitting in dataset distillation. We",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
245,
505,
257
],
"spans": [
{
"bbox": [
105,
245,
358,
257
],
"score": 1.0,
"content": "obtain state-of-the-art performance on various datasets with a",
"type": "text"
},
{
"bbox": [
359,
245,
381,
255
],
"score": 0.32,
"content": "1 0 0 \\mathrm { x }",
"type": "inline_equation"
},
{
"bbox": [
381,
245,
505,
257
],
"score": 1.0,
"content": "reduction in training time and",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
256,
505,
268
],
"spans": [
{
"bbox": [
105,
256,
505,
268
],
"score": 1.0,
"content": "a 10x reduction in GPU memory requirement. The distilled data generated by FRePo looks real",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
266,
505,
279
],
"spans": [
{
"bbox": [
105,
266,
505,
279
],
"score": 1.0,
"content": "and natural and generalizes well to a wide range of architectures. Furthermore, we demonstrate two",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
277,
506,
291
],
"spans": [
{
"bbox": [
105,
277,
506,
291
],
"score": 1.0,
"content": "applications that take advantage of the high-quality distilled data, namely, continual learning and",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
288,
234,
301
],
"spans": [
{
"bbox": [
105,
288,
234,
301
],
"score": 1.0,
"content": "membership inference defense.",
"type": "text"
}
],
"index": 17
}
],
"index": 14
},
{
"type": "text",
"bbox": [
107,
304,
505,
371
],
"lines": [
{
"bbox": [
105,
304,
506,
317
],
"spans": [
{
"bbox": [
105,
304,
506,
317
],
"score": 1.0,
"content": "Broader Impact “Synthetic data”, in the broader sense of artificial data created by generative models,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
316,
505,
329
],
"spans": [
{
"bbox": [
105,
316,
505,
329
],
"score": 1.0,
"content": "can help researchers understand how an otherwise opaque learning machine “sees” the world. There",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
326,
505,
339
],
"spans": [
{
"bbox": [
105,
326,
505,
339
],
"score": 1.0,
"content": "have been concerns regarding the risk of fake data. This paper explores a new research direction",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
338,
505,
349
],
"spans": [
{
"bbox": [
105,
338,
505,
349
],
"score": 1.0,
"content": "in generating synthetic data only for downstream classification tasks. We believe this work can",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
348,
506,
362
],
"spans": [
{
"bbox": [
105,
348,
506,
362
],
"score": 1.0,
"content": "provide additional interpretability and potentially address the common concerns in machine learning",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
358,
234,
374
],
"spans": [
{
"bbox": [
105,
358,
234,
374
],
"score": 1.0,
"content": "regarding training data privacy.",
"type": "text"
}
],
"index": 23
}
],
"index": 20.5
},
{
"type": "title",
"bbox": [
107,
386,
339,
400
],
"lines": [
{
"bbox": [
105,
384,
341,
404
],
"spans": [
{
"bbox": [
105,
384,
341,
404
],
"score": 1.0,
"content": "Acknowledgments and Disclosure of Funding",
"type": "text"
}
],
"index": 24
}
],
"index": 24
},
{
"type": "text",
"bbox": [
107,
411,
506,
488
],
"lines": [
{
"bbox": [
105,
409,
506,
425
],
"spans": [
{
"bbox": [
105,
409,
506,
425
],
"score": 1.0,
"content": "We would like to thank Harris Chan, Andrew Jung, Michael Zhang, Philip Fradkin, Denny Wu, Chong",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
421,
505,
436
],
"spans": [
{
"bbox": [
105,
421,
505,
436
],
"score": 1.0,
"content": "Shao, Leo Lee, Alice Gao, Keiran Paster, and Lazar Atanackovic for their valuable feedback. Jimmy",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
433,
506,
446
],
"spans": [
{
"bbox": [
105,
433,
506,
446
],
"score": 1.0,
"content": "Ba was supported by NSERC Grant [2020-06904], CIFAR AI Chairs program, Google Research",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
444,
505,
457
],
"spans": [
{
"bbox": [
105,
444,
505,
457
],
"score": 1.0,
"content": "Scholar Program and Amazon Research Award. This project was supported by LG Electronics",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
454,
506,
468
],
"spans": [
{
"bbox": [
105,
454,
506,
468
],
"score": 1.0,
"content": "Canada. Resources used in preparing this research were provided, in part, by the Province of Ontario,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
104,
465,
506,
479
],
"spans": [
{
"bbox": [
104,
465,
506,
479
],
"score": 1.0,
"content": "the Government of Canada through CIFAR, and companies sponsoring the Vector Institute for",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
476,
197,
490
],
"spans": [
{
"bbox": [
105,
476,
197,
490
],
"score": 1.0,
"content": "Artificial Intelligence.",
"type": "text"
}
],
"index": 31
}
],
"index": 28
},
{
"type": "title",
"bbox": [
107,
504,
163,
517
],
"lines": [
{
"bbox": [
106,
502,
165,
519
],
"spans": [
{
"bbox": [
106,
502,
165,
519
],
"score": 1.0,
"content": "References",
"type": "text"
}
],
"index": 32
}
],
"index": 32
},
{
"type": "text",
"bbox": [
109,
523,
506,
722
],
"lines": [
{
"bbox": [
110,
523,
505,
538
],
"spans": [
{
"bbox": [
110,
523,
505,
538
],
"score": 1.0,
"content": "[1] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
127,
534,
480,
549
],
"spans": [
{
"bbox": [
127,
534,
480,
549
],
"score": 1.0,
"content": "network. CoRR, abs/1503.02531, 2015. URL http://arxiv.org/abs/1503.02531.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
111,
555,
506,
568
],
"spans": [
{
"bbox": [
111,
555,
506,
568
],
"score": 1.0,
"content": "[2] Takashi Fukuda, Masayuki Suzuki, Gakuto Kurata, Samuel Thomas, Jia Cui, and Bhuvana Ram-",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
127,
565,
506,
579
],
"spans": [
{
"bbox": [
127,
565,
506,
579
],
"score": 1.0,
"content": "abhadran. Efficient knowledge distillation from an ensemble of teachers. In Francisco Lacerda,",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
127,
576,
505,
590
],
"spans": [
{
"bbox": [
127,
576,
505,
590
],
"score": 1.0,
"content": "editor, Interspeech 2017, 18th Annual Conference of the International Speech Communication",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
126,
586,
506,
601
],
"spans": [
{
"bbox": [
126,
586,
506,
601
],
"score": 1.0,
"content": "Association, Stockholm, Sweden, August 20-24, 2017, pages 3697–3701. ISCA, 2017. URL",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
127,
598,
505,
612
],
"spans": [
{
"bbox": [
127,
598,
505,
612
],
"score": 1.0,
"content": "http://www.isca-speech.org/archive/Interspeech_2017/abstracts/0614.html.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
110,
617,
505,
631
],
"spans": [
{
"bbox": [
110,
617,
505,
631
],
"score": 1.0,
"content": "[3] Antonio Polino, Razvan Pascanu, and Dan Alistarh. Model compression via distillation and",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
127,
628,
507,
642
],
"spans": [
{
"bbox": [
127,
628,
507,
642
],
"score": 1.0,
"content": "quantization. In 6th International Conference on Learning Representations, ICLR 2018, Van-",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
126,
639,
507,
654
],
"spans": [
{
"bbox": [
126,
639,
507,
654
],
"score": 1.0,
"content": "couver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net,",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
126,
649,
397,
664
],
"spans": [
{
"bbox": [
126,
649,
338,
664
],
"score": 1.0,
"content": "2018. URL https://openreview.net/forum?id",
"type": "text"
},
{
"bbox": [
339,
651,
345,
659
],
"score": 0.32,
"content": "\\cdot ^ { = }",
"type": "inline_equation"
},
{
"bbox": [
345,
649,
397,
664
],
"score": 1.0,
"content": "S1XolQbRW.",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
110,
669,
506,
682
],
"spans": [
{
"bbox": [
110,
669,
506,
682
],
"score": 1.0,
"content": "[4] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A. Efros. Dataset distillation.",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
128,
681,
440,
694
],
"spans": [
{
"bbox": [
128,
681,
440,
694
],
"score": 1.0,
"content": "CoRR, abs/1811.10959, 2018. URL http://arxiv.org/abs/1811.10959.",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
110,
700,
507,
714
],
"spans": [
{
"bbox": [
110,
700,
507,
714
],
"score": 1.0,
"content": "[5] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient match-",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
127,
712,
506,
724
],
"spans": [
{
"bbox": [
127,
712,
506,
724
],
"score": 1.0,
"content": "ing. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event,",
"type": "text"
}
],
"index": 47
}
],
"index": 40
}
],
"page_idx": 9,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
301,
742,
311,
750
],
"lines": [
{
"bbox": [
299,
740,
313,
754
],
"spans": [
{
"bbox": [
299,
740,
313,
754
],
"score": 1.0,
"content": "",
"type": "text",
"height": 14,
"width": 14
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
106,
72,
505,
182
],
"lines": [
{
"bbox": [
105,
73,
505,
85
],
"spans": [
{
"bbox": [
105,
73,
505,
85
],
"score": 1.0,
"content": "As shown in Table 4, all models trained on the distilled data preserve privacy as their attack AUCs",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
83,
506,
96
],
"spans": [
{
"bbox": [
105,
83,
506,
96
],
"score": 1.0,
"content": "are closed to random guessing. However, we observe a small drop in test accuracy compared to",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
93,
506,
107
],
"spans": [
{
"bbox": [
105,
93,
506,
107
],
"score": 1.0,
"content": "the model trained on the full dataset, which is expected as we only distill 500 examples instead of",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
105,
506,
118
],
"spans": [
{
"bbox": [
105,
105,
506,
118
],
"score": 1.0,
"content": "10,000 examples. Compared to the model trained on an equally sized subset of the original data, the",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
116,
506,
129
],
"spans": [
{
"bbox": [
105,
116,
506,
129
],
"score": 1.0,
"content": "model trained on distilled data results in much better test performance. Figure 5c, 5d demonstrate the",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
128,
505,
139
],
"spans": [
{
"bbox": [
106,
128,
505,
139
],
"score": 1.0,
"content": "trade-off between test accuracy and attack effectiveness as measured by ROC AUC. It shows that",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
138,
505,
150
],
"spans": [
{
"bbox": [
106,
138,
505,
150
],
"score": 1.0,
"content": "early stopping can be an effective technique to preserve privacy. However, we will still be under high",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
104,
147,
506,
162
],
"spans": [
{
"bbox": [
104,
147,
506,
162
],
"score": 1.0,
"content": "MIA risk if we perform early stopping by monitoring the validation loss. In contrast, training a model",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
159,
505,
173
],
"spans": [
{
"bbox": [
105,
159,
505,
173
],
"score": 1.0,
"content": "on the distilled data does not have this problem as the attack AUCs keep at a very low level regardless",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
171,
176,
184
],
"spans": [
{
"bbox": [
105,
171,
176,
184
],
"score": 1.0,
"content": "of training steps.",
"type": "text"
}
],
"index": 9
}
],
"index": 4.5,
"bbox_fs": [
104,
73,
506,
184
]
},
{
"type": "title",
"bbox": [
107,
198,
183,
211
],
"lines": [
{
"bbox": [
105,
196,
185,
214
],
"spans": [
{
"bbox": [
105,
196,
185,
214
],
"score": 1.0,
"content": "6 Conclusion",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
107,
223,
505,
299
],
"lines": [
{
"bbox": [
106,
222,
506,
236
],
"spans": [
{
"bbox": [
106,
222,
506,
236
],
"score": 1.0,
"content": "We propose neural Feature Regression with Pooling (FRePo) to overcome two challenges in dataset",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
233,
506,
247
],
"spans": [
{
"bbox": [
105,
233,
506,
247
],
"score": 1.0,
"content": "distillation: meta-gradient computation and various types of overfitting in dataset distillation. We",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
245,
505,
257
],
"spans": [
{
"bbox": [
105,
245,
358,
257
],
"score": 1.0,
"content": "obtain state-of-the-art performance on various datasets with a",
"type": "text"
},
{
"bbox": [
359,
245,
381,
255
],
"score": 0.32,
"content": "1 0 0 \\mathrm { x }",
"type": "inline_equation"
},
{
"bbox": [
381,
245,
505,
257
],
"score": 1.0,
"content": "reduction in training time and",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
256,
505,
268
],
"spans": [
{
"bbox": [
105,
256,
505,
268
],
"score": 1.0,
"content": "a 10x reduction in GPU memory requirement. The distilled data generated by FRePo looks real",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
266,
505,
279
],
"spans": [
{
"bbox": [
105,
266,
505,
279
],
"score": 1.0,
"content": "and natural and generalizes well to a wide range of architectures. Furthermore, we demonstrate two",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
277,
506,
291
],
"spans": [
{
"bbox": [
105,
277,
506,
291
],
"score": 1.0,
"content": "applications that take advantage of the high-quality distilled data, namely, continual learning and",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
288,
234,
301
],
"spans": [
{
"bbox": [
105,
288,
234,
301
],
"score": 1.0,
"content": "membership inference defense.",
"type": "text"
}
],
"index": 17
}
],
"index": 14,
"bbox_fs": [
105,
222,
506,
301
]
},
{
"type": "text",
"bbox": [
107,
304,
505,
371
],
"lines": [
{
"bbox": [
105,
304,
506,
317
],
"spans": [
{
"bbox": [
105,
304,
506,
317
],
"score": 1.0,
"content": "Broader Impact “Synthetic data”, in the broader sense of artificial data created by generative models,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
316,
505,
329
],
"spans": [
{
"bbox": [
105,
316,
505,
329
],
"score": 1.0,
"content": "can help researchers understand how an otherwise opaque learning machine “sees” the world. There",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
326,
505,
339
],
"spans": [
{
"bbox": [
105,
326,
505,
339
],
"score": 1.0,
"content": "have been concerns regarding the risk of fake data. This paper explores a new research direction",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
338,
505,
349
],
"spans": [
{
"bbox": [
105,
338,
505,
349
],
"score": 1.0,
"content": "in generating synthetic data only for downstream classification tasks. We believe this work can",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
348,
506,
362
],
"spans": [
{
"bbox": [
105,
348,
506,
362
],
"score": 1.0,
"content": "provide additional interpretability and potentially address the common concerns in machine learning",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
358,
234,
374
],
"spans": [
{
"bbox": [
105,
358,
234,
374
],
"score": 1.0,
"content": "regarding training data privacy.",
"type": "text"
}
],
"index": 23
}
],
"index": 20.5,
"bbox_fs": [
105,
304,
506,
374
]
},
{
"type": "title",
"bbox": [
107,
386,
339,
400
],
"lines": [
{
"bbox": [
105,
384,
341,
404
],
"spans": [
{
"bbox": [
105,
384,
341,
404
],
"score": 1.0,
"content": "Acknowledgments and Disclosure of Funding",
"type": "text"
}
],
"index": 24
}
],
"index": 24
},
{
"type": "text",
"bbox": [
107,
411,
506,
488
],
"lines": [
{
"bbox": [
105,
409,
506,
425
],
"spans": [
{
"bbox": [
105,
409,
506,
425
],
"score": 1.0,
"content": "We would like to thank Harris Chan, Andrew Jung, Michael Zhang, Philip Fradkin, Denny Wu, Chong",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
421,
505,
436
],
"spans": [
{
"bbox": [
105,
421,
505,
436
],
"score": 1.0,
"content": "Shao, Leo Lee, Alice Gao, Keiran Paster, and Lazar Atanackovic for their valuable feedback. Jimmy",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
433,
506,
446
],
"spans": [
{
"bbox": [
105,
433,
506,
446
],
"score": 1.0,
"content": "Ba was supported by NSERC Grant [2020-06904], CIFAR AI Chairs program, Google Research",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
444,
505,
457
],
"spans": [
{
"bbox": [
105,
444,
505,
457
],
"score": 1.0,
"content": "Scholar Program and Amazon Research Award. This project was supported by LG Electronics",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
454,
506,
468
],
"spans": [
{
"bbox": [
105,
454,
506,
468
],
"score": 1.0,
"content": "Canada. Resources used in preparing this research were provided, in part, by the Province of Ontario,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
104,
465,
506,
479
],
"spans": [
{
"bbox": [
104,
465,
506,
479
],
"score": 1.0,
"content": "the Government of Canada through CIFAR, and companies sponsoring the Vector Institute for",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
476,
197,
490
],
"spans": [
{
"bbox": [
105,
476,
197,
490
],
"score": 1.0,
"content": "Artificial Intelligence.",
"type": "text"
}
],
"index": 31
}
],
"index": 28,
"bbox_fs": [
104,
409,
506,
490
]
},
{
"type": "title",
"bbox": [
107,
504,
163,
517
],
"lines": [
{
"bbox": [
106,
502,
165,
519
],
"spans": [
{
"bbox": [
106,
502,
165,
519
],
"score": 1.0,
"content": "References",
"type": "text"
}
],
"index": 32
}
],
"index": 32
},
{
"type": "list",
"bbox": [
109,
523,
506,
722
],
"lines": [
{
"bbox": [
110,
523,
505,
538
],
"spans": [
{
"bbox": [
110,
523,
505,
538
],
"score": 1.0,
"content": "[1] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural",
"type": "text"
}
],
"index": 33,
"is_list_start_line": true
},
{
"bbox": [
127,
534,
480,
549
],
"spans": [
{
"bbox": [
127,
534,
480,
549
],
"score": 1.0,
"content": "network. CoRR, abs/1503.02531, 2015. URL http://arxiv.org/abs/1503.02531.",
"type": "text"
}
],
"index": 34,
"is_list_end_line": true
},
{
"bbox": [
111,
555,
506,
568
],
"spans": [
{
"bbox": [
111,
555,
506,
568
],
"score": 1.0,
"content": "[2] Takashi Fukuda, Masayuki Suzuki, Gakuto Kurata, Samuel Thomas, Jia Cui, and Bhuvana Ram-",
"type": "text"
}
],
"index": 35,
"is_list_start_line": true
},
{
"bbox": [
127,
565,
506,
579
],
"spans": [
{
"bbox": [
127,
565,
506,
579
],
"score": 1.0,
"content": "abhadran. Efficient knowledge distillation from an ensemble of teachers. In Francisco Lacerda,",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
127,
576,
505,
590
],
"spans": [
{
"bbox": [
127,
576,
505,
590
],
"score": 1.0,
"content": "editor, Interspeech 2017, 18th Annual Conference of the International Speech Communication",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
126,
586,
506,
601
],
"spans": [
{
"bbox": [
126,
586,
506,
601
],
"score": 1.0,
"content": "Association, Stockholm, Sweden, August 20-24, 2017, pages 3697–3701. ISCA, 2017. URL",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
127,
598,
505,
612
],
"spans": [
{
"bbox": [
127,
598,
505,
612
],
"score": 1.0,
"content": "http://www.isca-speech.org/archive/Interspeech_2017/abstracts/0614.html.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
110,
617,
505,
631
],
"spans": [
{
"bbox": [
110,
617,
505,
631
],
"score": 1.0,
"content": "[3] Antonio Polino, Razvan Pascanu, and Dan Alistarh. Model compression via distillation and",
"type": "text"
}
],
"index": 40,
"is_list_start_line": true
},
{
"bbox": [
127,
628,
507,
642
],
"spans": [
{
"bbox": [
127,
628,
507,
642
],
"score": 1.0,
"content": "quantization. In 6th International Conference on Learning Representations, ICLR 2018, Van-",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
126,
639,
507,
654
],
"spans": [
{
"bbox": [
126,
639,
507,
654
],
"score": 1.0,
"content": "couver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net,",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
126,
649,
397,
664
],
"spans": [
{
"bbox": [
126,
649,
338,
664
],
"score": 1.0,
"content": "2018. URL https://openreview.net/forum?id",
"type": "text"
},
{
"bbox": [
339,
651,
345,
659
],
"score": 0.32,
"content": "\\cdot ^ { = }",
"type": "inline_equation"
},
{
"bbox": [
345,
649,
397,
664
],
"score": 1.0,
"content": "S1XolQbRW.",
"type": "text"
}
],
"index": 43,
"is_list_end_line": true
},
{
"bbox": [
110,
669,
506,
682
],
"spans": [
{
"bbox": [
110,
669,
506,
682
],
"score": 1.0,
"content": "[4] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A. Efros. Dataset distillation.",
"type": "text"
}
],
"index": 44,
"is_list_start_line": true
},
{
"bbox": [
128,
681,
440,
694
],
"spans": [
{
"bbox": [
128,
681,
440,
694
],
"score": 1.0,
"content": "CoRR, abs/1811.10959, 2018. URL http://arxiv.org/abs/1811.10959.",
"type": "text"
}
],
"index": 45,
"is_list_end_line": true
},
{
"bbox": [
110,
700,
507,
714
],
"spans": [
{
"bbox": [
110,
700,
507,
714
],
"score": 1.0,
"content": "[5] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient match-",
"type": "text"
}
],
"index": 46,
"is_list_start_line": true
},
{
"bbox": [
127,
712,
506,
724
],
"spans": [
{
"bbox": [
127,
712,
506,
724
],
"score": 1.0,
"content": "ing. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event,",
"type": "text"
}
],
"index": 47
}
],
"index": 40,
"bbox_fs": [
110,
523,
507,
724
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
124,
72,
503,
95
],
"lines": [
{
"bbox": [
126,
72,
505,
86
],
"spans": [
{
"bbox": [
126,
72,
505,
86
],
"score": 1.0,
"content": "Austria, May 3-7, 2021. OpenReview.net, 2021. URL https://openreview.net/forum?",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
127,
84,
207,
95
],
"spans": [
{
"bbox": [
127,
84,
207,
95
],
"score": 1.0,
"content": "id=mSAKhLYLSsl.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
109,
101,
504,
136
],
"lines": [
{
"bbox": [
110,
101,
506,
114
],
"spans": [
{
"bbox": [
110,
101,
506,
114
],
"score": 1.0,
"content": "[6] Andrea Rosasco, Antonio Carta, Andrea Cossu, Vincenzo Lomonaco, and Davide Bacciu.",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
127,
112,
507,
126
],
"spans": [
{
"bbox": [
127,
112,
507,
126
],
"score": 1.0,
"content": "Distilled replay: Overcoming forgetting through synthetic samples. CoRR, abs/2103.15851,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
127,
123,
349,
136
],
"spans": [
{
"bbox": [
127,
123,
349,
136
],
"score": 1.0,
"content": "2021. URL https://arxiv.org/abs/2103.15851.",
"type": "text"
}
],
"index": 4
}
],
"index": 3
},
{
"type": "text",
"bbox": [
111,
141,
506,
196
],
"lines": [
{
"bbox": [
110,
141,
505,
154
],
"spans": [
{
"bbox": [
110,
141,
505,
154
],
"score": 1.0,
"content": "[7] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
127,
152,
505,
165
],
"spans": [
{
"bbox": [
127,
152,
505,
165
],
"score": 1.0,
"content": "Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
126,
162,
506,
177
],
"spans": [
{
"bbox": [
126,
162,
506,
177
],
"score": 1.0,
"content": "Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
126,
172,
507,
189
],
"spans": [
{
"bbox": [
126,
172,
507,
189
],
"score": 1.0,
"content": "Machine Learning Research, pages 12674–12685. PMLR, 2021. URL http://proceedings.",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
127,
185,
274,
197
],
"spans": [
{
"bbox": [
127,
185,
274,
197
],
"score": 1.0,
"content": "mlr.press/v139/zhao21a.html.",
"type": "text"
}
],
"index": 9
}
],
"index": 7
},
{
"type": "text",
"bbox": [
108,
203,
506,
226
],
"lines": [
{
"bbox": [
110,
202,
507,
216
],
"spans": [
{
"bbox": [
110,
202,
507,
216
],
"score": 1.0,
"content": "[8] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. CoRR,",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
128,
214,
417,
226
],
"spans": [
{
"bbox": [
128,
214,
417,
226
],
"score": 1.0,
"content": "abs/2110.04181, 2021. URL https://arxiv.org/abs/2110.04181.",
"type": "text"
}
],
"index": 11
}
],
"index": 10.5
},
{
"type": "text",
"bbox": [
110,
231,
506,
265
],
"lines": [
{
"bbox": [
110,
231,
505,
245
],
"spans": [
{
"bbox": [
110,
231,
505,
245
],
"score": 1.0,
"content": "[9] Guang Li, Ren Togo, Takahiro Ogawa, and Miki Haseyama. Soft-label anonymous gastric",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
126,
243,
507,
255
],
"spans": [
{
"bbox": [
126,
243,
507,
255
],
"score": 1.0,
"content": "x-ray image distillation. CoRR, abs/2104.02857, 2021. URL https://arxiv.org/abs/2104.",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
127,
254,
160,
266
],
"spans": [
{
"bbox": [
127,
254,
160,
266
],
"score": 1.0,
"content": "02857.",
"type": "text"
}
],
"index": 14
}
],
"index": 13
},
{
"type": "text",
"bbox": [
105,
271,
506,
294
],
"lines": [
{
"bbox": [
106,
271,
506,
284
],
"spans": [
{
"bbox": [
106,
271,
506,
284
],
"score": 1.0,
"content": "[10] Jack Goetz and Ambuj Tewari. Federated learning via synthetic data. CoRR, abs/2008.04489,",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
128,
283,
349,
294
],
"spans": [
{
"bbox": [
128,
283,
349,
294
],
"score": 1.0,
"content": "2020. URL https://arxiv.org/abs/2008.04489.",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5
},
{
"type": "text",
"bbox": [
106,
300,
505,
357
],
"lines": [
{
"bbox": [
107,
301,
506,
313
],
"spans": [
{
"bbox": [
107,
301,
506,
313
],
"score": 1.0,
"content": "[11] Dougal Maclaurin, David Duvenaud, and Ryan P. Adams. Gradient-based hyperparameter",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
127,
311,
506,
325
],
"spans": [
{
"bbox": [
127,
311,
506,
325
],
"score": 1.0,
"content": "optimization through reversible learning. In Francis R. Bach and David M. Blei, editors,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
126,
321,
506,
336
],
"spans": [
{
"bbox": [
126,
321,
506,
336
],
"score": 1.0,
"content": "Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille,",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
127,
333,
507,
346
],
"spans": [
{
"bbox": [
127,
333,
507,
346
],
"score": 1.0,
"content": "France, 6-11 July 2015, volume 37 of JMLR Workshop and Conference Proceedings, pages 2113–",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
127,
345,
506,
357
],
"spans": [
{
"bbox": [
127,
345,
506,
357
],
"score": 1.0,
"content": "2122. JMLR.org, 2015. URL http://proceedings.mlr.press/v37/maclaurin15.html.",
"type": "text"
}
],
"index": 21
}
],
"index": 19
},
{
"type": "text",
"bbox": [
106,
362,
504,
395
],
"lines": [
{
"bbox": [
105,
361,
506,
376
],
"spans": [
{
"bbox": [
105,
361,
506,
376
],
"score": 1.0,
"content": "[12] Ondrej Bohdal, Yongxin Yang, and Timothy M. Hospedales. Flexible dataset distillation: Learn",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
126,
372,
506,
386
],
"spans": [
{
"bbox": [
126,
372,
506,
386
],
"score": 1.0,
"content": "labels instead of images. CoRR, abs/2006.08572, 2020. URL https://arxiv.org/abs/",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
127,
384,
186,
396
],
"spans": [
{
"bbox": [
127,
384,
186,
396
],
"score": 1.0,
"content": "2006.08572.",
"type": "text"
}
],
"index": 24
}
],
"index": 23
},
{
"type": "text",
"bbox": [
106,
402,
503,
425
],
"lines": [
{
"bbox": [
105,
402,
505,
415
],
"spans": [
{
"bbox": [
105,
402,
505,
415
],
"score": 1.0,
"content": "[13] Ilia Sucholutsky and Matthias Schonlau. Improving dataset distillation. CoRR, abs/1910.02551,",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
128,
414,
343,
425
],
"spans": [
{
"bbox": [
128,
414,
343,
425
],
"score": 1.0,
"content": "2019. URL http://arxiv.org/abs/1910.02551.",
"type": "text"
}
],
"index": 26
}
],
"index": 25.5
},
{
"type": "text",
"bbox": [
106,
431,
506,
497
],
"lines": [
{
"bbox": [
106,
431,
506,
444
],
"spans": [
{
"bbox": [
106,
431,
506,
444
],
"score": 1.0,
"content": "[14] Paul Vicol, Luke Metz, and Jascha Sohl-Dickstein. Unbiased gradient estimation in un-",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
127,
442,
505,
455
],
"spans": [
{
"bbox": [
127,
442,
505,
455
],
"score": 1.0,
"content": "rolled computation graphs with persistent evolution strategies. In Marina Meila and Tong",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
127,
453,
506,
466
],
"spans": [
{
"bbox": [
127,
453,
506,
466
],
"score": 1.0,
"content": "Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
127,
464,
506,
477
],
"spans": [
{
"bbox": [
127,
464,
506,
477
],
"score": 1.0,
"content": "2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Re-",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
127,
474,
506,
488
],
"spans": [
{
"bbox": [
127,
474,
506,
488
],
"score": 1.0,
"content": "search, pages 10553–10563. PMLR, 2021. URL http://proceedings.mlr.press/v139/",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
127,
486,
200,
497
],
"spans": [
{
"bbox": [
127,
486,
200,
497
],
"score": 1.0,
"content": "vicol21a.html.",
"type": "text"
}
],
"index": 32
}
],
"index": 29.5
},
{
"type": "text",
"bbox": [
107,
504,
506,
559
],
"lines": [
{
"bbox": [
106,
504,
506,
517
],
"spans": [
{
"bbox": [
106,
504,
506,
517
],
"score": 1.0,
"content": "[15] Razvan Pascanu, Tomás Mikolov, and Yoshua Bengio. On the difficulty of training recurrent",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
127,
514,
506,
528
],
"spans": [
{
"bbox": [
127,
514,
506,
528
],
"score": 1.0,
"content": "neural networks. In Proceedings of the 30th International Conference on Machine Learning,",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
127,
524,
505,
539
],
"spans": [
{
"bbox": [
127,
524,
505,
539
],
"score": 1.0,
"content": "ICML 2013, Atlanta, GA, USA, 16-21 June 2013, volume 28 of JMLR Workshop and Conference",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
127,
536,
505,
550
],
"spans": [
{
"bbox": [
127,
536,
505,
550
],
"score": 1.0,
"content": "Proceedings, pages 1310–1318. JMLR.org, 2013. URL http://proceedings.mlr.press/",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
127,
546,
227,
561
],
"spans": [
{
"bbox": [
127,
546,
227,
561
],
"score": 1.0,
"content": "v28/pascanu13.html.",
"type": "text"
}
],
"index": 37
}
],
"index": 35
},
{
"type": "text",
"bbox": [
106,
565,
506,
632
],
"lines": [
{
"bbox": [
105,
565,
507,
578
],
"spans": [
{
"bbox": [
105,
565,
507,
578
],
"score": 1.0,
"content": "[16] Luke Metz, Niru Maheswaranathan, Jeremy Nixon, C. Daniel Freeman, and Jascha Sohl-",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
126,
576,
506,
589
],
"spans": [
{
"bbox": [
126,
576,
506,
589
],
"score": 1.0,
"content": "Dickstein. Understanding and correcting pathologies in the training of learned optimizers. In",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
127,
587,
506,
600
],
"spans": [
{
"bbox": [
127,
587,
506,
600
],
"score": 1.0,
"content": "Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
127,
597,
507,
612
],
"spans": [
{
"bbox": [
127,
597,
507,
612
],
"score": 1.0,
"content": "Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
127,
608,
507,
623
],
"spans": [
{
"bbox": [
127,
608,
507,
623
],
"score": 1.0,
"content": "volume 97 of Proceedings of Machine Learning Research, pages 4556–4565. PMLR, 2019.",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
127,
620,
391,
633
],
"spans": [
{
"bbox": [
127,
620,
391,
633
],
"score": 1.0,
"content": "URL http://proceedings.mlr.press/v97/metz19a.html.",
"type": "text"
}
],
"index": 43
}
],
"index": 40.5
},
{
"type": "text",
"bbox": [
107,
637,
506,
683
],
"lines": [
{
"bbox": [
106,
638,
505,
651
],
"spans": [
{
"bbox": [
106,
638,
505,
651
],
"score": 1.0,
"content": "[17] Yuhuai Wu, Mengye Ren, Renjie Liao, and Roger B. Grosse. Understanding short-horizon bias",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
127,
650,
506,
662
],
"spans": [
{
"bbox": [
127,
650,
506,
662
],
"score": 1.0,
"content": "in stochastic meta-optimization. In 6th International Conference on Learning Representations,",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
127,
659,
507,
673
],
"spans": [
{
"bbox": [
127,
659,
507,
673
],
"score": 1.0,
"content": "ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings.",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
127,
671,
468,
684
],
"spans": [
{
"bbox": [
127,
671,
468,
684
],
"score": 1.0,
"content": "OpenReview.net, 2018. URL https://openreview.net/forum?id=H1MczcgR-.",
"type": "text"
}
],
"index": 47
}
],
"index": 45.5
},
{
"type": "text",
"bbox": [
107,
689,
507,
722
],
"lines": [
{
"bbox": [
105,
688,
506,
702
],
"spans": [
{
"bbox": [
105,
688,
506,
702
],
"score": 1.0,
"content": "[18] Saehyung Lee, Sanghyuk Chun, Sangwon Jung, Sangdoo Yun, and Sungroh Yoon. Dataset",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
127,
700,
507,
712
],
"spans": [
{
"bbox": [
127,
700,
507,
712
],
"score": 1.0,
"content": "condensation with contrastive signals. CoRR, abs/2202.02916, 2022. URL https://arxiv.",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
127,
711,
228,
723
],
"spans": [
{
"bbox": [
127,
711,
228,
723
],
"score": 1.0,
"content": "org/abs/2202.02916.",
"type": "text"
}
],
"index": 50
}
],
"index": 49
}
],
"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": [
124,
72,
503,
95
],
"lines": [
{
"bbox": [
126,
72,
505,
86
],
"spans": [
{
"bbox": [
126,
72,
505,
86
],
"score": 1.0,
"content": "Austria, May 3-7, 2021. OpenReview.net, 2021. URL https://openreview.net/forum?",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
127,
84,
207,
95
],
"spans": [
{
"bbox": [
127,
84,
207,
95
],
"score": 1.0,
"content": "id=mSAKhLYLSsl.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5,
"bbox_fs": [
126,
72,
505,
95
]
},
{
"type": "text",
"bbox": [
109,
101,
504,
136
],
"lines": [
{
"bbox": [
110,
101,
506,
114
],
"spans": [
{
"bbox": [
110,
101,
506,
114
],
"score": 1.0,
"content": "[6] Andrea Rosasco, Antonio Carta, Andrea Cossu, Vincenzo Lomonaco, and Davide Bacciu.",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
127,
112,
507,
126
],
"spans": [
{
"bbox": [
127,
112,
507,
126
],
"score": 1.0,
"content": "Distilled replay: Overcoming forgetting through synthetic samples. CoRR, abs/2103.15851,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
127,
123,
349,
136
],
"spans": [
{
"bbox": [
127,
123,
349,
136
],
"score": 1.0,
"content": "2021. URL https://arxiv.org/abs/2103.15851.",
"type": "text"
}
],
"index": 4
}
],
"index": 3,
"bbox_fs": [
110,
101,
507,
136
]
},
{
"type": "text",
"bbox": [
111,
141,
506,
196
],
"lines": [
{
"bbox": [
110,
141,
505,
154
],
"spans": [
{
"bbox": [
110,
141,
505,
154
],
"score": 1.0,
"content": "[7] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
127,
152,
505,
165
],
"spans": [
{
"bbox": [
127,
152,
505,
165
],
"score": 1.0,
"content": "Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
126,
162,
506,
177
],
"spans": [
{
"bbox": [
126,
162,
506,
177
],
"score": 1.0,
"content": "Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
126,
172,
507,
189
],
"spans": [
{
"bbox": [
126,
172,
507,
189
],
"score": 1.0,
"content": "Machine Learning Research, pages 12674–12685. PMLR, 2021. URL http://proceedings.",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
127,
185,
274,
197
],
"spans": [
{
"bbox": [
127,
185,
274,
197
],
"score": 1.0,
"content": "mlr.press/v139/zhao21a.html.",
"type": "text"
}
],
"index": 9
}
],
"index": 7,
"bbox_fs": [
110,
141,
507,
197
]
},
{
"type": "text",
"bbox": [
108,
203,
506,
226
],
"lines": [
{
"bbox": [
110,
202,
507,
216
],
"spans": [
{
"bbox": [
110,
202,
507,
216
],
"score": 1.0,
"content": "[8] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. CoRR,",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
128,
214,
417,
226
],
"spans": [
{
"bbox": [
128,
214,
417,
226
],
"score": 1.0,
"content": "abs/2110.04181, 2021. URL https://arxiv.org/abs/2110.04181.",
"type": "text"
}
],
"index": 11
}
],
"index": 10.5,
"bbox_fs": [
110,
202,
507,
226
]
},
{
"type": "text",
"bbox": [
110,
231,
506,
265
],
"lines": [
{
"bbox": [
110,
231,
505,
245
],
"spans": [
{
"bbox": [
110,
231,
505,
245
],
"score": 1.0,
"content": "[9] Guang Li, Ren Togo, Takahiro Ogawa, and Miki Haseyama. Soft-label anonymous gastric",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
126,
243,
507,
255
],
"spans": [
{
"bbox": [
126,
243,
507,
255
],
"score": 1.0,
"content": "x-ray image distillation. CoRR, abs/2104.02857, 2021. URL https://arxiv.org/abs/2104.",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
127,
254,
160,
266
],
"spans": [
{
"bbox": [
127,
254,
160,
266
],
"score": 1.0,
"content": "02857.",
"type": "text"
}
],
"index": 14
}
],
"index": 13,
"bbox_fs": [
110,
231,
507,
266
]
},
{
"type": "text",
"bbox": [
105,
271,
506,
294
],
"lines": [
{
"bbox": [
106,
271,
506,
284
],
"spans": [
{
"bbox": [
106,
271,
506,
284
],
"score": 1.0,
"content": "[10] Jack Goetz and Ambuj Tewari. Federated learning via synthetic data. CoRR, abs/2008.04489,",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
128,
283,
349,
294
],
"spans": [
{
"bbox": [
128,
283,
349,
294
],
"score": 1.0,
"content": "2020. URL https://arxiv.org/abs/2008.04489.",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5,
"bbox_fs": [
106,
271,
506,
294
]
},
{
"type": "text",
"bbox": [
106,
300,
505,
357
],
"lines": [
{
"bbox": [
107,
301,
506,
313
],
"spans": [
{
"bbox": [
107,
301,
506,
313
],
"score": 1.0,
"content": "[11] Dougal Maclaurin, David Duvenaud, and Ryan P. Adams. Gradient-based hyperparameter",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
127,
311,
506,
325
],
"spans": [
{
"bbox": [
127,
311,
506,
325
],
"score": 1.0,
"content": "optimization through reversible learning. In Francis R. Bach and David M. Blei, editors,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
126,
321,
506,
336
],
"spans": [
{
"bbox": [
126,
321,
506,
336
],
"score": 1.0,
"content": "Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille,",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
127,
333,
507,
346
],
"spans": [
{
"bbox": [
127,
333,
507,
346
],
"score": 1.0,
"content": "France, 6-11 July 2015, volume 37 of JMLR Workshop and Conference Proceedings, pages 2113–",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
127,
345,
506,
357
],
"spans": [
{
"bbox": [
127,
345,
506,
357
],
"score": 1.0,
"content": "2122. JMLR.org, 2015. URL http://proceedings.mlr.press/v37/maclaurin15.html.",
"type": "text"
}
],
"index": 21
}
],
"index": 19,
"bbox_fs": [
107,
301,
507,
357
]
},
{
"type": "text",
"bbox": [
106,
362,
504,
395
],
"lines": [
{
"bbox": [
105,
361,
506,
376
],
"spans": [
{
"bbox": [
105,
361,
506,
376
],
"score": 1.0,
"content": "[12] Ondrej Bohdal, Yongxin Yang, and Timothy M. Hospedales. Flexible dataset distillation: Learn",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
126,
372,
506,
386
],
"spans": [
{
"bbox": [
126,
372,
506,
386
],
"score": 1.0,
"content": "labels instead of images. CoRR, abs/2006.08572, 2020. URL https://arxiv.org/abs/",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
127,
384,
186,
396
],
"spans": [
{
"bbox": [
127,
384,
186,
396
],
"score": 1.0,
"content": "2006.08572.",
"type": "text"
}
],
"index": 24
}
],
"index": 23,
"bbox_fs": [
105,
361,
506,
396
]
},
{
"type": "text",
"bbox": [
106,
402,
503,
425
],
"lines": [
{
"bbox": [
105,
402,
505,
415
],
"spans": [
{
"bbox": [
105,
402,
505,
415
],
"score": 1.0,
"content": "[13] Ilia Sucholutsky and Matthias Schonlau. Improving dataset distillation. CoRR, abs/1910.02551,",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
128,
414,
343,
425
],
"spans": [
{
"bbox": [
128,
414,
343,
425
],
"score": 1.0,
"content": "2019. URL http://arxiv.org/abs/1910.02551.",
"type": "text"
}
],
"index": 26
}
],
"index": 25.5,
"bbox_fs": [
105,
402,
505,
425
]
},
{
"type": "text",
"bbox": [
106,
431,
506,
497
],
"lines": [
{
"bbox": [
106,
431,
506,
444
],
"spans": [
{
"bbox": [
106,
431,
506,
444
],
"score": 1.0,
"content": "[14] Paul Vicol, Luke Metz, and Jascha Sohl-Dickstein. Unbiased gradient estimation in un-",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
127,
442,
505,
455
],
"spans": [
{
"bbox": [
127,
442,
505,
455
],
"score": 1.0,
"content": "rolled computation graphs with persistent evolution strategies. In Marina Meila and Tong",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
127,
453,
506,
466
],
"spans": [
{
"bbox": [
127,
453,
506,
466
],
"score": 1.0,
"content": "Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
127,
464,
506,
477
],
"spans": [
{
"bbox": [
127,
464,
506,
477
],
"score": 1.0,
"content": "2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Re-",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
127,
474,
506,
488
],
"spans": [
{
"bbox": [
127,
474,
506,
488
],
"score": 1.0,
"content": "search, pages 10553–10563. PMLR, 2021. URL http://proceedings.mlr.press/v139/",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
127,
486,
200,
497
],
"spans": [
{
"bbox": [
127,
486,
200,
497
],
"score": 1.0,
"content": "vicol21a.html.",
"type": "text"
}
],
"index": 32
}
],
"index": 29.5,
"bbox_fs": [
106,
431,
506,
497
]
},
{
"type": "text",
"bbox": [
107,
504,
506,
559
],
"lines": [
{
"bbox": [
106,
504,
506,
517
],
"spans": [
{
"bbox": [
106,
504,
506,
517
],
"score": 1.0,
"content": "[15] Razvan Pascanu, Tomás Mikolov, and Yoshua Bengio. On the difficulty of training recurrent",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
127,
514,
506,
528
],
"spans": [
{
"bbox": [
127,
514,
506,
528
],
"score": 1.0,
"content": "neural networks. In Proceedings of the 30th International Conference on Machine Learning,",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
127,
524,
505,
539
],
"spans": [
{
"bbox": [
127,
524,
505,
539
],
"score": 1.0,
"content": "ICML 2013, Atlanta, GA, USA, 16-21 June 2013, volume 28 of JMLR Workshop and Conference",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
127,
536,
505,
550
],
"spans": [
{
"bbox": [
127,
536,
505,
550
],
"score": 1.0,
"content": "Proceedings, pages 1310–1318. JMLR.org, 2013. URL http://proceedings.mlr.press/",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
127,
546,
227,
561
],
"spans": [
{
"bbox": [
127,
546,
227,
561
],
"score": 1.0,
"content": "v28/pascanu13.html.",
"type": "text"
}
],
"index": 37
}
],
"index": 35,
"bbox_fs": [
106,
504,
506,
561
]
},
{
"type": "text",
"bbox": [
106,
565,
506,
632
],
"lines": [
{
"bbox": [
105,
565,
507,
578
],
"spans": [
{
"bbox": [
105,
565,
507,
578
],
"score": 1.0,
"content": "[16] Luke Metz, Niru Maheswaranathan, Jeremy Nixon, C. Daniel Freeman, and Jascha Sohl-",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
126,
576,
506,
589
],
"spans": [
{
"bbox": [
126,
576,
506,
589
],
"score": 1.0,
"content": "Dickstein. Understanding and correcting pathologies in the training of learned optimizers. In",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
127,
587,
506,
600
],
"spans": [
{
"bbox": [
127,
587,
506,
600
],
"score": 1.0,
"content": "Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
127,
597,
507,
612
],
"spans": [
{
"bbox": [
127,
597,
507,
612
],
"score": 1.0,
"content": "Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
127,
608,
507,
623
],
"spans": [
{
"bbox": [
127,
608,
507,
623
],
"score": 1.0,
"content": "volume 97 of Proceedings of Machine Learning Research, pages 4556–4565. PMLR, 2019.",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
127,
620,
391,
633
],
"spans": [
{
"bbox": [
127,
620,
391,
633
],
"score": 1.0,
"content": "URL http://proceedings.mlr.press/v97/metz19a.html.",
"type": "text"
}
],
"index": 43
}
],
"index": 40.5,
"bbox_fs": [
105,
565,
507,
633
]
},
{
"type": "text",
"bbox": [
107,
637,
506,
683
],
"lines": [
{
"bbox": [
106,
638,
505,
651
],
"spans": [
{
"bbox": [
106,
638,
505,
651
],
"score": 1.0,
"content": "[17] Yuhuai Wu, Mengye Ren, Renjie Liao, and Roger B. Grosse. Understanding short-horizon bias",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
127,
650,
506,
662
],
"spans": [
{
"bbox": [
127,
650,
506,
662
],
"score": 1.0,
"content": "in stochastic meta-optimization. In 6th International Conference on Learning Representations,",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
127,
659,
507,
673
],
"spans": [
{
"bbox": [
127,
659,
507,
673
],
"score": 1.0,
"content": "ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings.",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
127,
671,
468,
684
],
"spans": [
{
"bbox": [
127,
671,
468,
684
],
"score": 1.0,
"content": "OpenReview.net, 2018. URL https://openreview.net/forum?id=H1MczcgR-.",
"type": "text"
}
],
"index": 47
}
],
"index": 45.5,
"bbox_fs": [
106,
638,
507,
684
]
},
{
"type": "text",
"bbox": [
107,
689,
507,
722
],
"lines": [
{
"bbox": [
105,
688,
506,
702
],
"spans": [
{
"bbox": [
105,
688,
506,
702
],
"score": 1.0,
"content": "[18] Saehyung Lee, Sanghyuk Chun, Sangwon Jung, Sangdoo Yun, and Sungroh Yoon. Dataset",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
127,
700,
507,
712
],
"spans": [
{
"bbox": [
127,
700,
507,
712
],
"score": 1.0,
"content": "condensation with contrastive signals. CoRR, abs/2202.02916, 2022. URL https://arxiv.",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
127,
711,
228,
723
],
"spans": [
{
"bbox": [
127,
711,
228,
723
],
"score": 1.0,
"content": "org/abs/2202.02916.",
"type": "text"
}
],
"index": 50
}
],
"index": 49,
"bbox_fs": [
105,
688,
507,
723
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
106,
73,
506,
117
],
"lines": [
{
"bbox": [
105,
72,
507,
86
],
"spans": [
{
"bbox": [
105,
72,
507,
86
],
"score": 1.0,
"content": "[19] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang,",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
126,
83,
506,
97
],
"spans": [
{
"bbox": [
126,
83,
506,
97
],
"score": 1.0,
"content": "Hakan Bilen, Xinchao Wang, and Yang You. CAFE: learning to condense dataset by aligning",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
126,
93,
506,
108
],
"spans": [
{
"bbox": [
126,
93,
506,
108
],
"score": 1.0,
"content": "features. CoRR, abs/2203.01531, 2022. doi: 10.48550/arXiv.2203.01531. URL https:",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
127,
106,
316,
118
],
"spans": [
{
"bbox": [
127,
106,
316,
118
],
"score": 1.0,
"content": "//doi.org/10.48550/arXiv.2203.01531.",
"type": "text"
}
],
"index": 3
}
],
"index": 1.5
},
{
"type": "text",
"bbox": [
106,
123,
505,
156
],
"lines": [
{
"bbox": [
105,
123,
507,
136
],
"spans": [
{
"bbox": [
105,
123,
507,
136
],
"score": 1.0,
"content": "[20] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A. Efros, and Jun-Yan Zhu.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
127,
134,
507,
146
],
"spans": [
{
"bbox": [
127,
134,
507,
146
],
"score": 1.0,
"content": "Dataset distillation by matching training trajectories. CoRR, abs/2203.11932, 2022. doi:",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
127,
144,
487,
158
],
"spans": [
{
"bbox": [
127,
144,
487,
158
],
"score": 1.0,
"content": "10.48550/arXiv.2203.11932. URL https://doi.org/10.48550/arXiv.2203.11932.",
"type": "text"
}
],
"index": 6
}
],
"index": 5
},
{
"type": "text",
"bbox": [
107,
162,
505,
250
],
"lines": [
{
"bbox": [
106,
162,
506,
176
],
"spans": [
{
"bbox": [
106,
162,
506,
176
],
"score": 1.0,
"content": "[21] Jaehoon Lee, Lechao Xiao, Samuel S. Schoenholz, Yasaman Bahri, Roman Novak, Jascha",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
127,
173,
505,
186
],
"spans": [
{
"bbox": [
127,
173,
505,
186
],
"score": 1.0,
"content": "Sohl-Dickstein, and Jeffrey Pennington. Wide neural networks of any depth evolve as",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
127,
184,
506,
198
],
"spans": [
{
"bbox": [
127,
184,
506,
198
],
"score": 1.0,
"content": "linear models under gradient descent. In Hanna M. Wallach, Hugo Larochelle, Alina",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
127,
196,
505,
208
],
"spans": [
{
"bbox": [
127,
196,
505,
208
],
"score": 1.0,
"content": "Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
127,
206,
505,
219
],
"spans": [
{
"bbox": [
127,
206,
505,
219
],
"score": 1.0,
"content": "in Neural Information Processing Systems 32: Annual Conference on Neural Information",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
127,
217,
505,
229
],
"spans": [
{
"bbox": [
127,
217,
505,
229
],
"score": 1.0,
"content": "Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada,",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
126,
227,
506,
241
],
"spans": [
{
"bbox": [
126,
227,
506,
241
],
"score": 1.0,
"content": "pages 8570–8581, 2019. URL https://proceedings.neurips.cc/paper/2019/hash/",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
128,
240,
374,
250
],
"spans": [
{
"bbox": [
128,
240,
374,
250
],
"score": 1.0,
"content": "0d1a9651497a38d8b1c3871c84528bd4-Abstract.html.",
"type": "text"
}
],
"index": 14
}
],
"index": 10.5
},
{
"type": "text",
"bbox": [
106,
256,
506,
300
],
"lines": [
{
"bbox": [
106,
257,
506,
270
],
"spans": [
{
"bbox": [
106,
257,
506,
270
],
"score": 1.0,
"content": "[22] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
127,
267,
506,
280
],
"spans": [
{
"bbox": [
127,
267,
506,
280
],
"score": 1.0,
"content": "regression. In 9th International Conference on Learning Representations, ICLR 2021, Virtual",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
127,
279,
506,
292
],
"spans": [
{
"bbox": [
127,
279,
506,
292
],
"score": 1.0,
"content": "Event, Austria, May 3-7, 2021. OpenReview.net, 2021. URL https://openreview.net/",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
127,
290,
238,
302
],
"spans": [
{
"bbox": [
127,
290,
170,
302
],
"score": 1.0,
"content": "forum?id",
"type": "text"
},
{
"bbox": [
171,
290,
180,
299
],
"score": 0.33,
"content": "= 1",
"type": "inline_equation"
},
{
"bbox": [
180,
290,
238,
302
],
"score": 1.0,
"content": "-PrrQrK0QR.",
"type": "text"
}
],
"index": 18
}
],
"index": 16.5
},
{
"type": "text",
"bbox": [
107,
307,
506,
384
],
"lines": [
{
"bbox": [
106,
306,
506,
320
],
"spans": [
{
"bbox": [
106,
306,
506,
320
],
"score": 1.0,
"content": "[23] Timothy Nguyen, Roman Novak, Lechao Xiao, and Jaehoon Lee. Dataset distillation",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
127,
318,
506,
331
],
"spans": [
{
"bbox": [
127,
318,
506,
331
],
"score": 1.0,
"content": "with infinitely wide convolutional networks. In Marc’Aurelio Ranzato, Alina Beygelz-",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
127,
328,
506,
342
],
"spans": [
{
"bbox": [
127,
328,
506,
342
],
"score": 1.0,
"content": "imer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan, editors, Ad-",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
127,
340,
506,
352
],
"spans": [
{
"bbox": [
127,
340,
506,
352
],
"score": 1.0,
"content": "vances in Neural Information Processing Systems 34: Annual Conference on Neural",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
127,
351,
507,
364
],
"spans": [
{
"bbox": [
127,
351,
507,
364
],
"score": 1.0,
"content": "Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual,",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
126,
361,
506,
375
],
"spans": [
{
"bbox": [
126,
361,
506,
375
],
"score": 1.0,
"content": "pages 5186–5198, 2021. URL https://proceedings.neurips.cc/paper/2021/hash/",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
128,
374,
374,
384
],
"spans": [
{
"bbox": [
128,
374,
374,
384
],
"score": 1.0,
"content": "299a23a2291e2126b91d54f3601ec162-Abstract.html.",
"type": "text"
}
],
"index": 25
}
],
"index": 22
},
{
"type": "text",
"bbox": [
107,
390,
506,
446
],
"lines": [
{
"bbox": [
106,
389,
505,
403
],
"spans": [
{
"bbox": [
106,
389,
505,
403
],
"score": 1.0,
"content": "[24] Jonathan Lorraine, Paul Vicol, and David Duvenaud. Optimizing millions of hyperparameters by",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
128,
401,
505,
413
],
"spans": [
{
"bbox": [
128,
401,
505,
413
],
"score": 1.0,
"content": "implicit differentiation. In Silvia Chiappa and Roberto Calandra, editors, The 23rd International",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
127,
411,
506,
426
],
"spans": [
{
"bbox": [
127,
411,
506,
426
],
"score": 1.0,
"content": "Conference on Artificial Intelligence and Statistics, AISTATS 2020, 26-28 August 2020, Online",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
127,
423,
507,
436
],
"spans": [
{
"bbox": [
127,
423,
507,
436
],
"score": 1.0,
"content": "[Palermo, Sicily, Italy], volume 108 of Proceedings of Machine Learning Research, pages 1540–",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
127,
433,
500,
447
],
"spans": [
{
"bbox": [
127,
433,
500,
447
],
"score": 1.0,
"content": "1552. PMLR, 2020. URL http://proceedings.mlr.press/v108/lorraine20a.html.",
"type": "text"
}
],
"index": 30
}
],
"index": 28
},
{
"type": "text",
"bbox": [
107,
452,
384,
464
],
"lines": [
{
"bbox": [
106,
451,
385,
464
],
"spans": [
{
"bbox": [
106,
451,
385,
464
],
"score": 1.0,
"content": "[25] Radford M. Neal. Bayesian learning for neural networks. 1995.",
"type": "text"
}
],
"index": 31
}
],
"index": 31
},
{
"type": "text",
"bbox": [
107,
469,
505,
514
],
"lines": [
{
"bbox": [
105,
469,
506,
483
],
"spans": [
{
"bbox": [
105,
469,
506,
483
],
"score": 1.0,
"content": "[26] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
127,
480,
507,
493
],
"spans": [
{
"bbox": [
127,
480,
507,
493
],
"score": 1.0,
"content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei.",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
126,
491,
505,
505
],
"spans": [
{
"bbox": [
126,
491,
505,
505
],
"score": 1.0,
"content": "ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
126,
502,
394,
516
],
"spans": [
{
"bbox": [
126,
502,
394,
516
],
"score": 1.0,
"content": "(IJCV), 115(3):211–252, 2015. doi: 10.1007/s11263-015-0816-y.",
"type": "text"
}
],
"index": 35
}
],
"index": 33.5
},
{
"type": "text",
"bbox": [
106,
520,
506,
596
],
"lines": [
{
"bbox": [
105,
519,
506,
533
],
"spans": [
{
"bbox": [
105,
519,
506,
533
],
"score": 1.0,
"content": "[27] Aravind Rajeswaran, Chelsea Finn, Sham M. Kakade, and Sergey Levine. Meta-learning",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
127,
530,
507,
544
],
"spans": [
{
"bbox": [
127,
530,
507,
544
],
"score": 1.0,
"content": "with implicit gradients. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer,",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
127,
541,
506,
554
],
"spans": [
{
"bbox": [
127,
541,
506,
554
],
"score": 1.0,
"content": "Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neu-",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
127,
552,
507,
566
],
"spans": [
{
"bbox": [
127,
552,
507,
566
],
"score": 1.0,
"content": "ral Information Processing Systems 32: Annual Conference on Neural Information Pro-",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
127,
563,
506,
576
],
"spans": [
{
"bbox": [
127,
563,
506,
576
],
"score": 1.0,
"content": "cessing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
127,
574,
506,
587
],
"spans": [
{
"bbox": [
127,
574,
506,
587
],
"score": 1.0,
"content": "pages 113–124, 2019. URL https://proceedings.neurips.cc/paper/2019/hash/",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
128,
586,
374,
597
],
"spans": [
{
"bbox": [
128,
586,
374,
597
],
"score": 1.0,
"content": "072b030ba126b2f4b2374f342be9ed44-Abstract.html.",
"type": "text"
}
],
"index": 42
}
],
"index": 39
},
{
"type": "text",
"bbox": [
105,
603,
504,
625
],
"lines": [
{
"bbox": [
106,
603,
505,
615
],
"spans": [
{
"bbox": [
106,
603,
505,
615
],
"score": 1.0,
"content": "[28] P.J. Werbos. Backpropagation through time: what it does and how to do it. Proceedings of the",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
128,
614,
351,
626
],
"spans": [
{
"bbox": [
128,
614,
351,
626
],
"score": 1.0,
"content": "IEEE, 78(10):1550–1560, 1990. doi: 10.1109/5.58337.",
"type": "text"
}
],
"index": 44
}
],
"index": 43.5
},
{
"type": "text",
"bbox": [
106,
631,
504,
654
],
"lines": [
{
"bbox": [
105,
630,
507,
645
],
"spans": [
{
"bbox": [
105,
630,
507,
645
],
"score": 1.0,
"content": "[29] Ilya Sutskever. Training recurrent neural networks. University of Toronto Toronto, ON, Canada,",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
125,
640,
155,
654
],
"spans": [
{
"bbox": [
125,
640,
155,
654
],
"score": 1.0,
"content": "2013.",
"type": "text"
}
],
"index": 46
}
],
"index": 45.5
},
{
"type": "text",
"bbox": [
106,
660,
505,
683
],
"lines": [
{
"bbox": [
106,
659,
507,
674
],
"spans": [
{
"bbox": [
106,
659,
507,
674
],
"score": 1.0,
"content": "[30] Corentin Tallec and Yann Ollivier. Unbiasing truncated backpropagation through time. CoRR,",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
127,
670,
412,
684
],
"spans": [
{
"bbox": [
127,
670,
412,
684
],
"score": 1.0,
"content": "abs/1705.08209, 2017. URL http://arxiv.org/abs/1705.08209.",
"type": "text"
}
],
"index": 48
}
],
"index": 47.5
},
{
"type": "text",
"bbox": [
107,
689,
506,
722
],
"lines": [
{
"bbox": [
106,
688,
506,
702
],
"spans": [
{
"bbox": [
106,
688,
506,
702
],
"score": 1.0,
"content": "[31] Jimmy Ba, Murat A Erdogdu, Taiji Suzuki, Zhichao Wang, Denny Wu, and Greg Yang. High-",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
127,
699,
506,
713
],
"spans": [
{
"bbox": [
127,
699,
506,
713
],
"score": 1.0,
"content": "dimensional asymptotics of feature learning: How one gradient step improves the representation.",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
128,
711,
290,
723
],
"spans": [
{
"bbox": [
128,
711,
290,
723
],
"score": 1.0,
"content": "arXiv preprint arXiv:2205.01445, 2022.",
"type": "text"
}
],
"index": 51
}
],
"index": 50
}
],
"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": "text",
"bbox": [
106,
73,
506,
117
],
"lines": [
{
"bbox": [
105,
72,
507,
86
],
"spans": [
{
"bbox": [
105,
72,
507,
86
],
"score": 1.0,
"content": "[19] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang,",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
126,
83,
506,
97
],
"spans": [
{
"bbox": [
126,
83,
506,
97
],
"score": 1.0,
"content": "Hakan Bilen, Xinchao Wang, and Yang You. CAFE: learning to condense dataset by aligning",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
126,
93,
506,
108
],
"spans": [
{
"bbox": [
126,
93,
506,
108
],
"score": 1.0,
"content": "features. CoRR, abs/2203.01531, 2022. doi: 10.48550/arXiv.2203.01531. URL https:",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
127,
106,
316,
118
],
"spans": [
{
"bbox": [
127,
106,
316,
118
],
"score": 1.0,
"content": "//doi.org/10.48550/arXiv.2203.01531.",
"type": "text"
}
],
"index": 3
}
],
"index": 1.5,
"bbox_fs": [
105,
72,
507,
118
]
},
{
"type": "text",
"bbox": [
106,
123,
505,
156
],
"lines": [
{
"bbox": [
105,
123,
507,
136
],
"spans": [
{
"bbox": [
105,
123,
507,
136
],
"score": 1.0,
"content": "[20] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A. Efros, and Jun-Yan Zhu.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
127,
134,
507,
146
],
"spans": [
{
"bbox": [
127,
134,
507,
146
],
"score": 1.0,
"content": "Dataset distillation by matching training trajectories. CoRR, abs/2203.11932, 2022. doi:",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
127,
144,
487,
158
],
"spans": [
{
"bbox": [
127,
144,
487,
158
],
"score": 1.0,
"content": "10.48550/arXiv.2203.11932. URL https://doi.org/10.48550/arXiv.2203.11932.",
"type": "text"
}
],
"index": 6
}
],
"index": 5,
"bbox_fs": [
105,
123,
507,
158
]
},
{
"type": "text",
"bbox": [
107,
162,
505,
250
],
"lines": [
{
"bbox": [
106,
162,
506,
176
],
"spans": [
{
"bbox": [
106,
162,
506,
176
],
"score": 1.0,
"content": "[21] Jaehoon Lee, Lechao Xiao, Samuel S. Schoenholz, Yasaman Bahri, Roman Novak, Jascha",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
127,
173,
505,
186
],
"spans": [
{
"bbox": [
127,
173,
505,
186
],
"score": 1.0,
"content": "Sohl-Dickstein, and Jeffrey Pennington. Wide neural networks of any depth evolve as",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
127,
184,
506,
198
],
"spans": [
{
"bbox": [
127,
184,
506,
198
],
"score": 1.0,
"content": "linear models under gradient descent. In Hanna M. Wallach, Hugo Larochelle, Alina",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
127,
196,
505,
208
],
"spans": [
{
"bbox": [
127,
196,
505,
208
],
"score": 1.0,
"content": "Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
127,
206,
505,
219
],
"spans": [
{
"bbox": [
127,
206,
505,
219
],
"score": 1.0,
"content": "in Neural Information Processing Systems 32: Annual Conference on Neural Information",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
127,
217,
505,
229
],
"spans": [
{
"bbox": [
127,
217,
505,
229
],
"score": 1.0,
"content": "Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada,",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
126,
227,
506,
241
],
"spans": [
{
"bbox": [
126,
227,
506,
241
],
"score": 1.0,
"content": "pages 8570–8581, 2019. URL https://proceedings.neurips.cc/paper/2019/hash/",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
128,
240,
374,
250
],
"spans": [
{
"bbox": [
128,
240,
374,
250
],
"score": 1.0,
"content": "0d1a9651497a38d8b1c3871c84528bd4-Abstract.html.",
"type": "text"
}
],
"index": 14
}
],
"index": 10.5,
"bbox_fs": [
106,
162,
506,
250
]
},
{
"type": "text",
"bbox": [
106,
256,
506,
300
],
"lines": [
{
"bbox": [
106,
257,
506,
270
],
"spans": [
{
"bbox": [
106,
257,
506,
270
],
"score": 1.0,
"content": "[22] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
127,
267,
506,
280
],
"spans": [
{
"bbox": [
127,
267,
506,
280
],
"score": 1.0,
"content": "regression. In 9th International Conference on Learning Representations, ICLR 2021, Virtual",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
127,
279,
506,
292
],
"spans": [
{
"bbox": [
127,
279,
506,
292
],
"score": 1.0,
"content": "Event, Austria, May 3-7, 2021. OpenReview.net, 2021. URL https://openreview.net/",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
127,
290,
238,
302
],
"spans": [
{
"bbox": [
127,
290,
170,
302
],
"score": 1.0,
"content": "forum?id",
"type": "text"
},
{
"bbox": [
171,
290,
180,
299
],
"score": 0.33,
"content": "= 1",
"type": "inline_equation"
},
{
"bbox": [
180,
290,
238,
302
],
"score": 1.0,
"content": "-PrrQrK0QR.",
"type": "text"
}
],
"index": 18
}
],
"index": 16.5,
"bbox_fs": [
106,
257,
506,
302
]
},
{
"type": "text",
"bbox": [
107,
307,
506,
384
],
"lines": [
{
"bbox": [
106,
306,
506,
320
],
"spans": [
{
"bbox": [
106,
306,
506,
320
],
"score": 1.0,
"content": "[23] Timothy Nguyen, Roman Novak, Lechao Xiao, and Jaehoon Lee. Dataset distillation",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
127,
318,
506,
331
],
"spans": [
{
"bbox": [
127,
318,
506,
331
],
"score": 1.0,
"content": "with infinitely wide convolutional networks. In Marc’Aurelio Ranzato, Alina Beygelz-",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
127,
328,
506,
342
],
"spans": [
{
"bbox": [
127,
328,
506,
342
],
"score": 1.0,
"content": "imer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan, editors, Ad-",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
127,
340,
506,
352
],
"spans": [
{
"bbox": [
127,
340,
506,
352
],
"score": 1.0,
"content": "vances in Neural Information Processing Systems 34: Annual Conference on Neural",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
127,
351,
507,
364
],
"spans": [
{
"bbox": [
127,
351,
507,
364
],
"score": 1.0,
"content": "Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual,",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
126,
361,
506,
375
],
"spans": [
{
"bbox": [
126,
361,
506,
375
],
"score": 1.0,
"content": "pages 5186–5198, 2021. URL https://proceedings.neurips.cc/paper/2021/hash/",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
128,
374,
374,
384
],
"spans": [
{
"bbox": [
128,
374,
374,
384
],
"score": 1.0,
"content": "299a23a2291e2126b91d54f3601ec162-Abstract.html.",
"type": "text"
}
],
"index": 25
}
],
"index": 22,
"bbox_fs": [
106,
306,
507,
384
]
},
{
"type": "text",
"bbox": [
107,
390,
506,
446
],
"lines": [
{
"bbox": [
106,
389,
505,
403
],
"spans": [
{
"bbox": [
106,
389,
505,
403
],
"score": 1.0,
"content": "[24] Jonathan Lorraine, Paul Vicol, and David Duvenaud. Optimizing millions of hyperparameters by",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
128,
401,
505,
413
],
"spans": [
{
"bbox": [
128,
401,
505,
413
],
"score": 1.0,
"content": "implicit differentiation. In Silvia Chiappa and Roberto Calandra, editors, The 23rd International",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
127,
411,
506,
426
],
"spans": [
{
"bbox": [
127,
411,
506,
426
],
"score": 1.0,
"content": "Conference on Artificial Intelligence and Statistics, AISTATS 2020, 26-28 August 2020, Online",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
127,
423,
507,
436
],
"spans": [
{
"bbox": [
127,
423,
507,
436
],
"score": 1.0,
"content": "[Palermo, Sicily, Italy], volume 108 of Proceedings of Machine Learning Research, pages 1540–",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
127,
433,
500,
447
],
"spans": [
{
"bbox": [
127,
433,
500,
447
],
"score": 1.0,
"content": "1552. PMLR, 2020. URL http://proceedings.mlr.press/v108/lorraine20a.html.",
"type": "text"
}
],
"index": 30
}
],
"index": 28,
"bbox_fs": [
106,
389,
507,
447
]
},
{
"type": "text",
"bbox": [
107,
452,
384,
464
],
"lines": [
{
"bbox": [
106,
451,
385,
464
],
"spans": [
{
"bbox": [
106,
451,
385,
464
],
"score": 1.0,
"content": "[25] Radford M. Neal. Bayesian learning for neural networks. 1995.",
"type": "text"
}
],
"index": 31
}
],
"index": 31,
"bbox_fs": [
106,
451,
385,
464
]
},
{
"type": "text",
"bbox": [
107,
469,
505,
514
],
"lines": [
{
"bbox": [
105,
469,
506,
483
],
"spans": [
{
"bbox": [
105,
469,
506,
483
],
"score": 1.0,
"content": "[26] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
127,
480,
507,
493
],
"spans": [
{
"bbox": [
127,
480,
507,
493
],
"score": 1.0,
"content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei.",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
126,
491,
505,
505
],
"spans": [
{
"bbox": [
126,
491,
505,
505
],
"score": 1.0,
"content": "ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
126,
502,
394,
516
],
"spans": [
{
"bbox": [
126,
502,
394,
516
],
"score": 1.0,
"content": "(IJCV), 115(3):211–252, 2015. doi: 10.1007/s11263-015-0816-y.",
"type": "text"
}
],
"index": 35
}
],
"index": 33.5,
"bbox_fs": [
105,
469,
507,
516
]
},
{
"type": "text",
"bbox": [
106,
520,
506,
596
],
"lines": [
{
"bbox": [
105,
519,
506,
533
],
"spans": [
{
"bbox": [
105,
519,
506,
533
],
"score": 1.0,
"content": "[27] Aravind Rajeswaran, Chelsea Finn, Sham M. Kakade, and Sergey Levine. Meta-learning",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
127,
530,
507,
544
],
"spans": [
{
"bbox": [
127,
530,
507,
544
],
"score": 1.0,
"content": "with implicit gradients. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer,",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
127,
541,
506,
554
],
"spans": [
{
"bbox": [
127,
541,
506,
554
],
"score": 1.0,
"content": "Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neu-",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
127,
552,
507,
566
],
"spans": [
{
"bbox": [
127,
552,
507,
566
],
"score": 1.0,
"content": "ral Information Processing Systems 32: Annual Conference on Neural Information Pro-",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
127,
563,
506,
576
],
"spans": [
{
"bbox": [
127,
563,
506,
576
],
"score": 1.0,
"content": "cessing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
127,
574,
506,
587
],
"spans": [
{
"bbox": [
127,
574,
506,
587
],
"score": 1.0,
"content": "pages 113–124, 2019. URL https://proceedings.neurips.cc/paper/2019/hash/",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
128,
586,
374,
597
],
"spans": [
{
"bbox": [
128,
586,
374,
597
],
"score": 1.0,
"content": "072b030ba126b2f4b2374f342be9ed44-Abstract.html.",
"type": "text"
}
],
"index": 42
}
],
"index": 39,
"bbox_fs": [
105,
519,
507,
597
]
},
{
"type": "text",
"bbox": [
105,
603,
504,
625
],
"lines": [
{
"bbox": [
106,
603,
505,
615
],
"spans": [
{
"bbox": [
106,
603,
505,
615
],
"score": 1.0,
"content": "[28] P.J. Werbos. Backpropagation through time: what it does and how to do it. Proceedings of the",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
128,
614,
351,
626
],
"spans": [
{
"bbox": [
128,
614,
351,
626
],
"score": 1.0,
"content": "IEEE, 78(10):1550–1560, 1990. doi: 10.1109/5.58337.",
"type": "text"
}
],
"index": 44
}
],
"index": 43.5,
"bbox_fs": [
106,
603,
505,
626
]
},
{
"type": "text",
"bbox": [
106,
631,
504,
654
],
"lines": [
{
"bbox": [
105,
630,
507,
645
],
"spans": [
{
"bbox": [
105,
630,
507,
645
],
"score": 1.0,
"content": "[29] Ilya Sutskever. Training recurrent neural networks. University of Toronto Toronto, ON, Canada,",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
125,
640,
155,
654
],
"spans": [
{
"bbox": [
125,
640,
155,
654
],
"score": 1.0,
"content": "2013.",
"type": "text"
}
],
"index": 46
}
],
"index": 45.5,
"bbox_fs": [
105,
630,
507,
654
]
},
{
"type": "text",
"bbox": [
106,
660,
505,
683
],
"lines": [
{
"bbox": [
106,
659,
507,
674
],
"spans": [
{
"bbox": [
106,
659,
507,
674
],
"score": 1.0,
"content": "[30] Corentin Tallec and Yann Ollivier. Unbiasing truncated backpropagation through time. CoRR,",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
127,
670,
412,
684
],
"spans": [
{
"bbox": [
127,
670,
412,
684
],
"score": 1.0,
"content": "abs/1705.08209, 2017. URL http://arxiv.org/abs/1705.08209.",
"type": "text"
}
],
"index": 48
}
],
"index": 47.5,
"bbox_fs": [
106,
659,
507,
684
]
},
{
"type": "text",
"bbox": [
107,
689,
506,
722
],
"lines": [
{
"bbox": [
106,
688,
506,
702
],
"spans": [
{
"bbox": [
106,
688,
506,
702
],
"score": 1.0,
"content": "[31] Jimmy Ba, Murat A Erdogdu, Taiji Suzuki, Zhichao Wang, Denny Wu, and Greg Yang. High-",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
127,
699,
506,
713
],
"spans": [
{
"bbox": [
127,
699,
506,
713
],
"score": 1.0,
"content": "dimensional asymptotics of feature learning: How one gradient step improves the representation.",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
128,
711,
290,
723
],
"spans": [
{
"bbox": [
128,
711,
290,
723
],
"score": 1.0,
"content": "arXiv preprint arXiv:2205.01445, 2022.",
"type": "text"
}
],
"index": 51
}
],
"index": 50,
"bbox_fs": [
106,
688,
506,
723
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
106,
72,
505,
106
],
"lines": [
{
"bbox": [
105,
72,
505,
86
],
"spans": [
{
"bbox": [
105,
72,
505,
86
],
"score": 1.0,
"content": "[32] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
127,
83,
507,
96
],
"spans": [
{
"bbox": [
127,
83,
507,
96
],
"score": 1.0,
"content": "applied to document recognition. Proc. IEEE, 86(11):2278–2324, 1998. doi: 10.1109/5.726791.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
128,
94,
323,
106
],
"spans": [
{
"bbox": [
128,
94,
323,
106
],
"score": 1.0,
"content": "URL https://doi.org/10.1109/5.726791.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "text",
"bbox": [
106,
113,
504,
147
],
"lines": [
{
"bbox": [
105,
112,
506,
127
],
"spans": [
{
"bbox": [
105,
112,
506,
127
],
"score": 1.0,
"content": "[33] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
127,
124,
505,
137
],
"spans": [
{
"bbox": [
127,
124,
505,
137
],
"score": 1.0,
"content": "benchmarking machine learning algorithms. CoRR, abs/1708.07747, 2017. URL http://",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
127,
136,
259,
148
],
"spans": [
{
"bbox": [
127,
136,
259,
148
],
"score": 1.0,
"content": "arxiv.org/abs/1708.07747.",
"type": "text"
}
],
"index": 5
}
],
"index": 4
},
{
"type": "text",
"bbox": [
105,
154,
504,
167
],
"lines": [
{
"bbox": [
105,
153,
506,
168
],
"spans": [
{
"bbox": [
105,
153,
506,
168
],
"score": 1.0,
"content": "[34] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009.",
"type": "text"
}
],
"index": 6
}
],
"index": 6
},
{
"type": "text",
"bbox": [
107,
173,
437,
186
],
"lines": [
{
"bbox": [
106,
173,
435,
187
],
"spans": [
{
"bbox": [
106,
173,
435,
187
],
"score": 1.0,
"content": "[35] Ya Le and Xuan S. Yang. Tiny imagenet visual recognition challenge. 2015.",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "text",
"bbox": [
106,
192,
504,
215
],
"lines": [
{
"bbox": [
105,
191,
506,
206
],
"spans": [
{
"bbox": [
105,
191,
506,
206
],
"score": 1.0,
"content": "[36] Jeremy Howard. A smaller subset of 10 easily classified classes from imagenet, and a little",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
127,
203,
404,
216
],
"spans": [
{
"bbox": [
127,
203,
404,
216
],
"score": 1.0,
"content": "more french. URL https://github.com/fastai/imagenette/.",
"type": "text"
}
],
"index": 9
}
],
"index": 8.5
},
{
"type": "text",
"bbox": [
107,
222,
196,
234
],
"lines": [
{
"bbox": [
106,
222,
197,
236
],
"spans": [
{
"bbox": [
106,
222,
197,
236
],
"score": 1.0,
"content": "[37] Technical report.",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
107,
241,
506,
318
],
"lines": [
{
"bbox": [
105,
240,
506,
255
],
"spans": [
{
"bbox": [
105,
240,
506,
255
],
"score": 1.0,
"content": "[38] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
127,
252,
506,
265
],
"spans": [
{
"bbox": [
127,
252,
506,
265
],
"score": 1.0,
"content": "convolutional neural networks. In Peter L. Bartlett, Fernando C. N. Pereira, Christopher J. C.",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
127,
263,
505,
276
],
"spans": [
{
"bbox": [
127,
263,
505,
276
],
"score": 1.0,
"content": "Burges, Léon Bottou, and Kilian Q. Weinberger, editors, Advances in Neural Information",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
127,
274,
505,
287
],
"spans": [
{
"bbox": [
127,
274,
505,
287
],
"score": 1.0,
"content": "Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
127,
285,
506,
298
],
"spans": [
{
"bbox": [
127,
285,
506,
298
],
"score": 1.0,
"content": "2012. Proceedings of a meeting held December 3-6, 2012, Lake Tahoe, Nevada, United States,",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
126,
296,
506,
309
],
"spans": [
{
"bbox": [
126,
296,
506,
309
],
"score": 1.0,
"content": "pages 1106–1114, 2012. URL https://proceedings.neurips.cc/paper/2012/hash/",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
128,
308,
374,
319
],
"spans": [
{
"bbox": [
128,
308,
374,
319
],
"score": 1.0,
"content": "c399862d3b9d6b76c8436e924a68c45b-Abstract.html.",
"type": "text"
}
],
"index": 17
}
],
"index": 14
},
{
"type": "text",
"bbox": [
107,
326,
505,
371
],
"lines": [
{
"bbox": [
106,
326,
505,
339
],
"spans": [
{
"bbox": [
106,
326,
505,
339
],
"score": 1.0,
"content": "[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
126,
337,
505,
350
],
"spans": [
{
"bbox": [
126,
337,
505,
350
],
"score": 1.0,
"content": "image recognition. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
127,
349,
505,
361
],
"spans": [
{
"bbox": [
127,
349,
505,
361
],
"score": 1.0,
"content": "on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
127,
360,
417,
371
],
"spans": [
{
"bbox": [
127,
360,
417,
371
],
"score": 1.0,
"content": "Track Proceedings, 2015. URL http://arxiv.org/abs/1409.1556.",
"type": "text"
}
],
"index": 21
}
],
"index": 19.5
},
{
"type": "text",
"bbox": [
107,
378,
505,
423
],
"lines": [
{
"bbox": [
105,
377,
506,
392
],
"spans": [
{
"bbox": [
105,
377,
506,
392
],
"score": 1.0,
"content": "[40] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
126,
389,
506,
401
],
"spans": [
{
"bbox": [
126,
389,
506,
401
],
"score": 1.0,
"content": "recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
126,
399,
507,
414
],
"spans": [
{
"bbox": [
126,
399,
507,
414
],
"score": 1.0,
"content": "2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770–778. IEEE Computer Society, 2016.",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
126,
410,
465,
424
],
"spans": [
{
"bbox": [
126,
410,
465,
424
],
"score": 1.0,
"content": "doi: 10.1109/CVPR.2016.90. URL https://doi.org/10.1109/CVPR.2016.90.",
"type": "text"
}
],
"index": 25
}
],
"index": 23.5
},
{
"type": "text",
"bbox": [
105,
429,
505,
462
],
"lines": [
{
"bbox": [
105,
429,
506,
443
],
"spans": [
{
"bbox": [
105,
429,
506,
443
],
"score": 1.0,
"content": "[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
127,
441,
505,
453
],
"spans": [
{
"bbox": [
127,
441,
505,
453
],
"score": 1.0,
"content": "ingredient for fast stylization. CoRR, abs/1607.08022, 2016. URL http://arxiv.org/abs/",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
128,
452,
185,
463
],
"spans": [
{
"bbox": [
128,
452,
185,
463
],
"score": 1.0,
"content": "1607.08022.",
"type": "text"
}
],
"index": 28
}
],
"index": 27
},
{
"type": "text",
"bbox": [
107,
470,
505,
526
],
"lines": [
{
"bbox": [
105,
469,
506,
485
],
"spans": [
{
"bbox": [
105,
469,
506,
485
],
"score": 1.0,
"content": "[42] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
127,
481,
505,
495
],
"spans": [
{
"bbox": [
127,
481,
505,
495
],
"score": 1.0,
"content": "by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors, Proceedings",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
128,
493,
505,
505
],
"spans": [
{
"bbox": [
128,
493,
505,
505
],
"score": 1.0,
"content": "of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
126,
500,
507,
519
],
"spans": [
{
"bbox": [
126,
500,
507,
519
],
"score": 1.0,
"content": "2015, volume 37 of JMLR Workshop and Conference Proceedings, pages 448–456. JMLR.org,",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
127,
514,
417,
527
],
"spans": [
{
"bbox": [
127,
514,
417,
527
],
"score": 1.0,
"content": "2015. URL http://proceedings.mlr.press/v37/ioffe15.html.",
"type": "text"
}
],
"index": 33
}
],
"index": 31
},
{
"type": "text",
"bbox": [
106,
533,
506,
589
],
"lines": [
{
"bbox": [
106,
533,
507,
546
],
"spans": [
{
"bbox": [
106,
533,
507,
546
],
"score": 1.0,
"content": "[43] James Kirkpatrick, Razvan Pascanu, Neil C. Rabinowitz, Joel Veness, Guillaume Des-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
127,
545,
506,
557
],
"spans": [
{
"bbox": [
127,
545,
506,
557
],
"score": 1.0,
"content": "jardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
128,
555,
506,
567
],
"spans": [
{
"bbox": [
128,
555,
506,
567
],
"score": 1.0,
"content": "Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Over-",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
127,
567,
506,
578
],
"spans": [
{
"bbox": [
127,
567,
506,
578
],
"score": 1.0,
"content": "coming catastrophic forgetting in neural networks. CoRR, abs/1612.00796, 2016. URL",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
127,
578,
295,
590
],
"spans": [
{
"bbox": [
127,
578,
295,
590
],
"score": 1.0,
"content": "http://arxiv.org/abs/1612.00796.",
"type": "text"
}
],
"index": 38
}
],
"index": 36
},
{
"type": "text",
"bbox": [
102,
596,
505,
619
],
"lines": [
{
"bbox": [
105,
595,
506,
610
],
"spans": [
{
"bbox": [
105,
595,
506,
610
],
"score": 1.0,
"content": "[44] Robert French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences,",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
128,
607,
372,
619
],
"spans": [
{
"bbox": [
128,
607,
372,
619
],
"score": 1.0,
"content": "3:128–135, 05 1999. doi: 10.1016/S1364-6613(99)01294-2.",
"type": "text"
}
],
"index": 40
}
],
"index": 39.5
},
{
"type": "text",
"bbox": [
106,
626,
505,
659
],
"lines": [
{
"bbox": [
105,
625,
506,
640
],
"spans": [
{
"bbox": [
105,
625,
506,
640
],
"score": 1.0,
"content": "[45] Anthony V. Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connect. Sci., 7",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
127,
636,
506,
650
],
"spans": [
{
"bbox": [
127,
636,
506,
650
],
"score": 1.0,
"content": "(2):123–146, 1995. doi: 10.1080/09540099550039318. URL https://doi.org/10.1080/",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
128,
648,
222,
659
],
"spans": [
{
"bbox": [
128,
648,
222,
659
],
"score": 1.0,
"content": "09540099550039318.",
"type": "text"
}
],
"index": 43
}
],
"index": 42
},
{
"type": "text",
"bbox": [
107,
667,
506,
722
],
"lines": [
{
"bbox": [
106,
667,
506,
680
],
"spans": [
{
"bbox": [
106,
667,
506,
680
],
"score": 1.0,
"content": "[46] Pietro Buzzega, Matteo Boschini, Angelo Porrello, and Simone Calderara. Rethinking ex-",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
127,
678,
505,
691
],
"spans": [
{
"bbox": [
127,
678,
505,
691
],
"score": 1.0,
"content": "perience replay: a bag of tricks for continual learning. In 25th International Conference",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
126,
689,
506,
701
],
"spans": [
{
"bbox": [
126,
689,
506,
701
],
"score": 1.0,
"content": "on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021,",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
126,
698,
508,
715
],
"spans": [
{
"bbox": [
126,
698,
508,
715
],
"score": 1.0,
"content": "pages 2180–2187. IEEE, 2020. doi: 10.1109/ICPR48806.2021.9412614. URL https:",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
127,
711,
342,
723
],
"spans": [
{
"bbox": [
127,
711,
342,
723
],
"score": 1.0,
"content": "//doi.org/10.1109/ICPR48806.2021.9412614.",
"type": "text"
}
],
"index": 48
}
],
"index": 46
}
],
"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": [
106,
72,
505,
106
],
"lines": [
{
"bbox": [
105,
72,
505,
86
],
"spans": [
{
"bbox": [
105,
72,
505,
86
],
"score": 1.0,
"content": "[32] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
127,
83,
507,
96
],
"spans": [
{
"bbox": [
127,
83,
507,
96
],
"score": 1.0,
"content": "applied to document recognition. Proc. IEEE, 86(11):2278–2324, 1998. doi: 10.1109/5.726791.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
128,
94,
323,
106
],
"spans": [
{
"bbox": [
128,
94,
323,
106
],
"score": 1.0,
"content": "URL https://doi.org/10.1109/5.726791.",
"type": "text"
}
],
"index": 2
}
],
"index": 1,
"bbox_fs": [
105,
72,
507,
106
]
},
{
"type": "text",
"bbox": [
106,
113,
504,
147
],
"lines": [
{
"bbox": [
105,
112,
506,
127
],
"spans": [
{
"bbox": [
105,
112,
506,
127
],
"score": 1.0,
"content": "[33] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
127,
124,
505,
137
],
"spans": [
{
"bbox": [
127,
124,
505,
137
],
"score": 1.0,
"content": "benchmarking machine learning algorithms. CoRR, abs/1708.07747, 2017. URL http://",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
127,
136,
259,
148
],
"spans": [
{
"bbox": [
127,
136,
259,
148
],
"score": 1.0,
"content": "arxiv.org/abs/1708.07747.",
"type": "text"
}
],
"index": 5
}
],
"index": 4,
"bbox_fs": [
105,
112,
506,
148
]
},
{
"type": "text",
"bbox": [
105,
154,
504,
167
],
"lines": [
{
"bbox": [
105,
153,
506,
168
],
"spans": [
{
"bbox": [
105,
153,
506,
168
],
"score": 1.0,
"content": "[34] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009.",
"type": "text"
}
],
"index": 6
}
],
"index": 6,
"bbox_fs": [
105,
153,
506,
168
]
},
{
"type": "text",
"bbox": [
107,
173,
437,
186
],
"lines": [
{
"bbox": [
106,
173,
435,
187
],
"spans": [
{
"bbox": [
106,
173,
435,
187
],
"score": 1.0,
"content": "[35] Ya Le and Xuan S. Yang. Tiny imagenet visual recognition challenge. 2015.",
"type": "text"
}
],
"index": 7
}
],
"index": 7,
"bbox_fs": [
106,
173,
435,
187
]
},
{
"type": "text",
"bbox": [
106,
192,
504,
215
],
"lines": [
{
"bbox": [
105,
191,
506,
206
],
"spans": [
{
"bbox": [
105,
191,
506,
206
],
"score": 1.0,
"content": "[36] Jeremy Howard. A smaller subset of 10 easily classified classes from imagenet, and a little",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
127,
203,
404,
216
],
"spans": [
{
"bbox": [
127,
203,
404,
216
],
"score": 1.0,
"content": "more french. URL https://github.com/fastai/imagenette/.",
"type": "text"
}
],
"index": 9
}
],
"index": 8.5,
"bbox_fs": [
105,
191,
506,
216
]
},
{
"type": "text",
"bbox": [
107,
222,
196,
234
],
"lines": [
{
"bbox": [
106,
222,
197,
236
],
"spans": [
{
"bbox": [
106,
222,
197,
236
],
"score": 1.0,
"content": "[37] Technical report.",
"type": "text"
}
],
"index": 10
}
],
"index": 10,
"bbox_fs": [
106,
222,
197,
236
]
},
{
"type": "text",
"bbox": [
107,
241,
506,
318
],
"lines": [
{
"bbox": [
105,
240,
506,
255
],
"spans": [
{
"bbox": [
105,
240,
506,
255
],
"score": 1.0,
"content": "[38] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
127,
252,
506,
265
],
"spans": [
{
"bbox": [
127,
252,
506,
265
],
"score": 1.0,
"content": "convolutional neural networks. In Peter L. Bartlett, Fernando C. N. Pereira, Christopher J. C.",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
127,
263,
505,
276
],
"spans": [
{
"bbox": [
127,
263,
505,
276
],
"score": 1.0,
"content": "Burges, Léon Bottou, and Kilian Q. Weinberger, editors, Advances in Neural Information",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
127,
274,
505,
287
],
"spans": [
{
"bbox": [
127,
274,
505,
287
],
"score": 1.0,
"content": "Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
127,
285,
506,
298
],
"spans": [
{
"bbox": [
127,
285,
506,
298
],
"score": 1.0,
"content": "2012. Proceedings of a meeting held December 3-6, 2012, Lake Tahoe, Nevada, United States,",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
126,
296,
506,
309
],
"spans": [
{
"bbox": [
126,
296,
506,
309
],
"score": 1.0,
"content": "pages 1106–1114, 2012. URL https://proceedings.neurips.cc/paper/2012/hash/",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
128,
308,
374,
319
],
"spans": [
{
"bbox": [
128,
308,
374,
319
],
"score": 1.0,
"content": "c399862d3b9d6b76c8436e924a68c45b-Abstract.html.",
"type": "text"
}
],
"index": 17
}
],
"index": 14,
"bbox_fs": [
105,
240,
506,
319
]
},
{
"type": "text",
"bbox": [
107,
326,
505,
371
],
"lines": [
{
"bbox": [
106,
326,
505,
339
],
"spans": [
{
"bbox": [
106,
326,
505,
339
],
"score": 1.0,
"content": "[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
126,
337,
505,
350
],
"spans": [
{
"bbox": [
126,
337,
505,
350
],
"score": 1.0,
"content": "image recognition. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
127,
349,
505,
361
],
"spans": [
{
"bbox": [
127,
349,
505,
361
],
"score": 1.0,
"content": "on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
127,
360,
417,
371
],
"spans": [
{
"bbox": [
127,
360,
417,
371
],
"score": 1.0,
"content": "Track Proceedings, 2015. URL http://arxiv.org/abs/1409.1556.",
"type": "text"
}
],
"index": 21
}
],
"index": 19.5,
"bbox_fs": [
106,
326,
505,
371
]
},
{
"type": "text",
"bbox": [
107,
378,
505,
423
],
"lines": [
{
"bbox": [
105,
377,
506,
392
],
"spans": [
{
"bbox": [
105,
377,
506,
392
],
"score": 1.0,
"content": "[40] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
126,
389,
506,
401
],
"spans": [
{
"bbox": [
126,
389,
506,
401
],
"score": 1.0,
"content": "recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
126,
399,
507,
414
],
"spans": [
{
"bbox": [
126,
399,
507,
414
],
"score": 1.0,
"content": "2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770–778. IEEE Computer Society, 2016.",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
126,
410,
465,
424
],
"spans": [
{
"bbox": [
126,
410,
465,
424
],
"score": 1.0,
"content": "doi: 10.1109/CVPR.2016.90. URL https://doi.org/10.1109/CVPR.2016.90.",
"type": "text"
}
],
"index": 25
}
],
"index": 23.5,
"bbox_fs": [
105,
377,
507,
424
]
},
{
"type": "text",
"bbox": [
105,
429,
505,
462
],
"lines": [
{
"bbox": [
105,
429,
506,
443
],
"spans": [
{
"bbox": [
105,
429,
506,
443
],
"score": 1.0,
"content": "[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
127,
441,
505,
453
],
"spans": [
{
"bbox": [
127,
441,
505,
453
],
"score": 1.0,
"content": "ingredient for fast stylization. CoRR, abs/1607.08022, 2016. URL http://arxiv.org/abs/",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
128,
452,
185,
463
],
"spans": [
{
"bbox": [
128,
452,
185,
463
],
"score": 1.0,
"content": "1607.08022.",
"type": "text"
}
],
"index": 28
}
],
"index": 27,
"bbox_fs": [
105,
429,
506,
463
]
},
{
"type": "text",
"bbox": [
107,
470,
505,
526
],
"lines": [
{
"bbox": [
105,
469,
506,
485
],
"spans": [
{
"bbox": [
105,
469,
506,
485
],
"score": 1.0,
"content": "[42] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
127,
481,
505,
495
],
"spans": [
{
"bbox": [
127,
481,
505,
495
],
"score": 1.0,
"content": "by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors, Proceedings",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
128,
493,
505,
505
],
"spans": [
{
"bbox": [
128,
493,
505,
505
],
"score": 1.0,
"content": "of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
126,
500,
507,
519
],
"spans": [
{
"bbox": [
126,
500,
507,
519
],
"score": 1.0,
"content": "2015, volume 37 of JMLR Workshop and Conference Proceedings, pages 448–456. JMLR.org,",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
127,
514,
417,
527
],
"spans": [
{
"bbox": [
127,
514,
417,
527
],
"score": 1.0,
"content": "2015. URL http://proceedings.mlr.press/v37/ioffe15.html.",
"type": "text"
}
],
"index": 33
}
],
"index": 31,
"bbox_fs": [
105,
469,
507,
527
]
},
{
"type": "text",
"bbox": [
106,
533,
506,
589
],
"lines": [
{
"bbox": [
106,
533,
507,
546
],
"spans": [
{
"bbox": [
106,
533,
507,
546
],
"score": 1.0,
"content": "[43] James Kirkpatrick, Razvan Pascanu, Neil C. Rabinowitz, Joel Veness, Guillaume Des-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
127,
545,
506,
557
],
"spans": [
{
"bbox": [
127,
545,
506,
557
],
"score": 1.0,
"content": "jardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
128,
555,
506,
567
],
"spans": [
{
"bbox": [
128,
555,
506,
567
],
"score": 1.0,
"content": "Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Over-",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
127,
567,
506,
578
],
"spans": [
{
"bbox": [
127,
567,
506,
578
],
"score": 1.0,
"content": "coming catastrophic forgetting in neural networks. CoRR, abs/1612.00796, 2016. URL",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
127,
578,
295,
590
],
"spans": [
{
"bbox": [
127,
578,
295,
590
],
"score": 1.0,
"content": "http://arxiv.org/abs/1612.00796.",
"type": "text"
}
],
"index": 38
}
],
"index": 36,
"bbox_fs": [
106,
533,
507,
590
]
},
{
"type": "text",
"bbox": [
102,
596,
505,
619
],
"lines": [
{
"bbox": [
105,
595,
506,
610
],
"spans": [
{
"bbox": [
105,
595,
506,
610
],
"score": 1.0,
"content": "[44] Robert French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences,",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
128,
607,
372,
619
],
"spans": [
{
"bbox": [
128,
607,
372,
619
],
"score": 1.0,
"content": "3:128–135, 05 1999. doi: 10.1016/S1364-6613(99)01294-2.",
"type": "text"
}
],
"index": 40
}
],
"index": 39.5,
"bbox_fs": [
105,
595,
506,
619
]
},
{
"type": "text",
"bbox": [
106,
626,
505,
659
],
"lines": [
{
"bbox": [
105,
625,
506,
640
],
"spans": [
{
"bbox": [
105,
625,
506,
640
],
"score": 1.0,
"content": "[45] Anthony V. Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connect. Sci., 7",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
127,
636,
506,
650
],
"spans": [
{
"bbox": [
127,
636,
506,
650
],
"score": 1.0,
"content": "(2):123–146, 1995. doi: 10.1080/09540099550039318. URL https://doi.org/10.1080/",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
128,
648,
222,
659
],
"spans": [
{
"bbox": [
128,
648,
222,
659
],
"score": 1.0,
"content": "09540099550039318.",
"type": "text"
}
],
"index": 43
}
],
"index": 42,
"bbox_fs": [
105,
625,
506,
659
]
},
{
"type": "text",
"bbox": [
107,
667,
506,
722
],
"lines": [
{
"bbox": [
106,
667,
506,
680
],
"spans": [
{
"bbox": [
106,
667,
506,
680
],
"score": 1.0,
"content": "[46] Pietro Buzzega, Matteo Boschini, Angelo Porrello, and Simone Calderara. Rethinking ex-",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
127,
678,
505,
691
],
"spans": [
{
"bbox": [
127,
678,
505,
691
],
"score": 1.0,
"content": "perience replay: a bag of tricks for continual learning. In 25th International Conference",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
126,
689,
506,
701
],
"spans": [
{
"bbox": [
126,
689,
506,
701
],
"score": 1.0,
"content": "on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021,",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
126,
698,
508,
715
],
"spans": [
{
"bbox": [
126,
698,
508,
715
],
"score": 1.0,
"content": "pages 2180–2187. IEEE, 2020. doi: 10.1109/ICPR48806.2021.9412614. URL https:",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
127,
711,
342,
723
],
"spans": [
{
"bbox": [
127,
711,
342,
723
],
"score": 1.0,
"content": "//doi.org/10.1109/ICPR48806.2021.9412614.",
"type": "text"
}
],
"index": 48
}
],
"index": 46,
"bbox_fs": [
106,
667,
508,
723
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
106,
72,
506,
139
],
"lines": [
{
"bbox": [
106,
73,
507,
86
],
"spans": [
{
"bbox": [
106,
73,
507,
86
],
"score": 1.0,
"content": "[47] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
127,
83,
506,
96
],
"spans": [
{
"bbox": [
127,
83,
506,
96
],
"score": 1.0,
"content": "incremental learning. In IEEE Conference on Computer Vision and Pattern Recognition,",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
127,
94,
506,
107
],
"spans": [
{
"bbox": [
127,
94,
506,
107
],
"score": 1.0,
"content": "CVPR 2021, virtual, June 19-25, 2021, pages 2544–2553. Computer Vision Foundation /",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
127,
105,
505,
118
],
"spans": [
{
"bbox": [
127,
105,
505,
118
],
"score": 1.0,
"content": "IEEE, 2021. URL https://openaccess.thecvf.com/content/CVPR2021/html/Liu_",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
127,
117,
502,
129
],
"spans": [
{
"bbox": [
127,
117,
502,
129
],
"score": 1.0,
"content": "Adaptive_Aggregation_Networks_for_Class-Incremental_Learning_CVPR_2021_",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
126,
128,
186,
139
],
"spans": [
{
"bbox": [
126,
128,
186,
139
],
"score": 1.0,
"content": "paper.html.",
"type": "text"
}
],
"index": 5
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
107,
145,
506,
179
],
"lines": [
{
"bbox": [
105,
144,
506,
158
],
"spans": [
{
"bbox": [
105,
144,
506,
158
],
"score": 1.0,
"content": "[48] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, and Christoph H. Lampert. icarl: Incremental",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
127,
157,
507,
169
],
"spans": [
{
"bbox": [
127,
157,
507,
169
],
"score": 1.0,
"content": "classifier and representation learning. CoRR, abs/1611.07725, 2016. URL http://arxiv.",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
127,
168,
228,
180
],
"spans": [
{
"bbox": [
127,
168,
228,
180
],
"score": 1.0,
"content": "org/abs/1611.07725.",
"type": "text"
}
],
"index": 8
}
],
"index": 7
},
{
"type": "text",
"bbox": [
107,
186,
506,
253
],
"lines": [
{
"bbox": [
106,
186,
506,
199
],
"spans": [
{
"bbox": [
106,
186,
506,
199
],
"score": 1.0,
"content": "[49] Ameya Prabhu, Philip H. S. Torr, and Puneet K. Dokania. Gdumb: A simple approach that",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
126,
197,
507,
211
],
"spans": [
{
"bbox": [
126,
197,
507,
211
],
"score": 1.0,
"content": "questions our progress in continual learning. In Andrea Vedaldi, Horst Bischof, Thomas Brox,",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
127,
208,
506,
222
],
"spans": [
{
"bbox": [
127,
208,
506,
222
],
"score": 1.0,
"content": "and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
127,
219,
506,
232
],
"spans": [
{
"bbox": [
127,
219,
506,
232
],
"score": 1.0,
"content": "Glasgow, UK, August 23-28, 2020, Proceedings, Part II, volume 12347 of Lecture Notes in",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
126,
229,
507,
244
],
"spans": [
{
"bbox": [
126,
229,
507,
244
],
"score": 1.0,
"content": "Computer Science, pages 524–540. Springer, 2020. doi: 10.1007/978-3-030-58536-5\\_31. URL",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
127,
241,
366,
254
],
"spans": [
{
"bbox": [
127,
241,
366,
254
],
"score": 1.0,
"content": "https://doi.org/10.1007/978-3-030-58536-5_31.",
"type": "text"
}
],
"index": 14
}
],
"index": 11.5
},
{
"type": "text",
"bbox": [
107,
259,
506,
336
],
"lines": [
{
"bbox": [
106,
260,
505,
272
],
"spans": [
{
"bbox": [
106,
260,
505,
272
],
"score": 1.0,
"content": "[50] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
127,
271,
505,
282
],
"spans": [
{
"bbox": [
127,
271,
505,
282
],
"score": 1.0,
"content": "selection for online continual learning. In Hanna M. Wallach, Hugo Larochelle, Alina",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
126,
281,
506,
295
],
"spans": [
{
"bbox": [
126,
281,
506,
295
],
"score": 1.0,
"content": "Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
126,
292,
506,
306
],
"spans": [
{
"bbox": [
126,
292,
506,
306
],
"score": 1.0,
"content": "in Neural Information Processing Systems 32: Annual Conference on Neural Information",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
127,
304,
506,
316
],
"spans": [
{
"bbox": [
127,
304,
506,
316
],
"score": 1.0,
"content": "Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada,",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
126,
314,
506,
327
],
"spans": [
{
"bbox": [
126,
314,
506,
327
],
"score": 1.0,
"content": "pages 11816–11825, 2019. URL https://proceedings.neurips.cc/paper/2019/hash/",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
127,
325,
375,
337
],
"spans": [
{
"bbox": [
127,
325,
375,
337
],
"score": 1.0,
"content": "e562cd9c0768d5464b64cf61da7fc6bb-Abstract.html.",
"type": "text"
}
],
"index": 21
}
],
"index": 18
},
{
"type": "text",
"bbox": [
107,
344,
506,
421
],
"lines": [
{
"bbox": [
105,
344,
506,
357
],
"spans": [
{
"bbox": [
105,
344,
506,
357
],
"score": 1.0,
"content": "[51] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin,",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
127,
355,
506,
367
],
"spans": [
{
"bbox": [
127,
355,
506,
367
],
"score": 1.0,
"content": "and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. In Hanna M.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
127,
366,
505,
378
],
"spans": [
{
"bbox": [
127,
366,
505,
378
],
"score": 1.0,
"content": "Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
127,
376,
505,
390
],
"spans": [
{
"bbox": [
127,
376,
505,
390
],
"score": 1.0,
"content": "Garnett, editors, Advances in Neural Information Processing Systems 32: Annual Conference on",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
127,
388,
506,
401
],
"spans": [
{
"bbox": [
127,
388,
506,
401
],
"score": 1.0,
"content": "Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver,",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
127,
398,
505,
411
],
"spans": [
{
"bbox": [
127,
398,
505,
411
],
"score": 1.0,
"content": "BC, Canada, pages 11849–11860, 2019. URL https://proceedings.neurips.cc/paper/",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
128,
410,
427,
421
],
"spans": [
{
"bbox": [
128,
410,
427,
421
],
"score": 1.0,
"content": "2019/hash/15825aee15eb335cc13f9b559f166ee8-Abstract.html.",
"type": "text"
}
],
"index": 28
}
],
"index": 25
},
{
"type": "text",
"bbox": [
107,
428,
506,
505
],
"lines": [
{
"bbox": [
105,
428,
507,
441
],
"spans": [
{
"bbox": [
105,
428,
507,
441
],
"score": 1.0,
"content": "[52] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
126,
439,
506,
453
],
"spans": [
{
"bbox": [
126,
439,
506,
453
],
"score": 1.0,
"content": "class incremental learning without forgetting. In 2020 IEEE/CVF Conference on Computer",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
128,
450,
507,
462
],
"spans": [
{
"bbox": [
128,
450,
507,
462
],
"score": 1.0,
"content": "Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 12242–",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
128,
461,
505,
473
],
"spans": [
{
"bbox": [
128,
461,
505,
473
],
"score": 1.0,
"content": "12251. Computer Vision Foundation / IEEE, 2020. doi: 10.1109/CVPR42600.2020.01226.",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
126,
471,
505,
486
],
"spans": [
{
"bbox": [
126,
471,
505,
486
],
"score": 1.0,
"content": "URL https://openaccess.thecvf.com/content_CVPR_2020/html/Liu_Mnemonics_",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
127,
483,
502,
496
],
"spans": [
{
"bbox": [
127,
483,
502,
496
],
"score": 1.0,
"content": "Training_Multi-Class_Incremental_Learning_Without_Forgetting_CVPR_2020_",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
126,
494,
186,
507
],
"spans": [
{
"bbox": [
126,
494,
186,
507
],
"score": 1.0,
"content": "paper.html.",
"type": "text"
}
],
"index": 35
}
],
"index": 32
},
{
"type": "text",
"bbox": [
105,
512,
505,
535
],
"lines": [
{
"bbox": [
106,
512,
506,
525
],
"spans": [
{
"bbox": [
106,
512,
506,
525
],
"score": 1.0,
"content": "[53] Gido M. van de Ven and Andreas S. Tolias. Three scenarios for continual learning. CoRR,",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
127,
523,
411,
536
],
"spans": [
{
"bbox": [
127,
523,
411,
536
],
"score": 1.0,
"content": "abs/1904.07734, 2019. URL http://arxiv.org/abs/1904.07734.",
"type": "text"
}
],
"index": 37
}
],
"index": 36.5
},
{
"type": "text",
"bbox": [
107,
541,
506,
609
],
"lines": [
{
"bbox": [
105,
542,
506,
554
],
"spans": [
{
"bbox": [
105,
542,
506,
554
],
"score": 1.0,
"content": "[54] Francisco M. Castro, Manuel J. Marín-Jiménez, Nicolás Guil, Cordelia Schmid, and Karteek",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
126,
552,
506,
566
],
"spans": [
{
"bbox": [
126,
552,
506,
566
],
"score": 1.0,
"content": "Alahari. End-to-end incremental learning. In Vittorio Ferrari, Martial Hebert, Cristian Smin-",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
126,
562,
507,
578
],
"spans": [
{
"bbox": [
126,
562,
507,
578
],
"score": 1.0,
"content": "chisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018 - 15th European Conference,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
127,
574,
506,
587
],
"spans": [
{
"bbox": [
127,
574,
506,
587
],
"score": 1.0,
"content": "Munich, Germany, September 8-14, 2018, Proceedings, Part XII, volume 11216 of Lecture Notes",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
127,
586,
507,
599
],
"spans": [
{
"bbox": [
127,
586,
507,
599
],
"score": 1.0,
"content": "in Computer Science, pages 241–257. Springer, 2018. doi: 10.1007/978-3-030-01258-8\\_15.",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
128,
597,
388,
610
],
"spans": [
{
"bbox": [
128,
597,
388,
610
],
"score": 1.0,
"content": "URL https://doi.org/10.1007/978-3-030-01258-8_15.",
"type": "text"
}
],
"index": 43
}
],
"index": 40.5
},
{
"type": "text",
"bbox": [
107,
615,
506,
671
],
"lines": [
{
"bbox": [
106,
615,
505,
628
],
"spans": [
{
"bbox": [
106,
615,
505,
628
],
"score": 1.0,
"content": "[55] Yutian Chen, Max Welling, and Alexander J. Smola. Super-samples from kernel herding. In Peter",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
127,
626,
505,
640
],
"spans": [
{
"bbox": [
127,
626,
505,
640
],
"score": 1.0,
"content": "Grünwald and Peter Spirtes, editors, UAI 2010, Proceedings of the Twenty-Sixth Conference",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
126,
635,
506,
651
],
"spans": [
{
"bbox": [
126,
635,
506,
651
],
"score": 1.0,
"content": "on Uncertainty in Artificial Intelligence, Catalina Island, CA, USA, July 8-11, 2010, pages",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
127,
647,
507,
662
],
"spans": [
{
"bbox": [
127,
647,
507,
662
],
"score": 1.0,
"content": "109–116. AUAI Press, 2010. URL https://dslpitt.org/uai/displayArticleDetails.",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
127,
659,
394,
672
],
"spans": [
{
"bbox": [
127,
659,
170,
672
],
"score": 1.0,
"content": "jsp?mmnu",
"type": "text"
},
{
"bbox": [
171,
661,
177,
668
],
"score": 0.46,
"content": "\\equiv",
"type": "inline_equation"
},
{
"bbox": [
177,
659,
207,
672
],
"score": 1.0,
"content": "1&smnu",
"type": "text"
},
{
"bbox": [
207,
661,
213,
668
],
"score": 0.52,
"content": "\\underset { . } { = }",
"type": "inline_equation"
},
{
"bbox": [
213,
659,
275,
672
],
"score": 1.0,
"content": "2&article_id",
"type": "text"
},
{
"bbox": [
275,
660,
281,
669
],
"score": 0.33,
"content": "\\equiv",
"type": "inline_equation"
},
{
"bbox": [
281,
659,
394,
672
],
"score": 1.0,
"content": "2148&proceeding_id=26.",
"type": "text"
}
],
"index": 48
}
],
"index": 46
},
{
"type": "text",
"bbox": [
108,
678,
505,
722
],
"lines": [
{
"bbox": [
106,
678,
505,
691
],
"spans": [
{
"bbox": [
106,
678,
505,
691
],
"score": 1.0,
"content": "[56] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
126,
688,
507,
702
],
"spans": [
{
"bbox": [
126,
688,
507,
702
],
"score": 1.0,
"content": "attacks against machine learning models. In 2017 IEEE Symposium on Security and Privacy,",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
126,
699,
507,
713
],
"spans": [
{
"bbox": [
126,
699,
507,
713
],
"score": 1.0,
"content": "SP 2017, San Jose, CA, USA, May 22-26, 2017, pages 3–18. IEEE Computer Society, 2017. doi:",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
127,
710,
419,
723
],
"spans": [
{
"bbox": [
127,
710,
419,
723
],
"score": 1.0,
"content": "10.1109/SP.2017.41. URL https://doi.org/10.1109/SP.2017.41.",
"type": "text"
}
],
"index": 52
}
],
"index": 50.5
}
],
"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": [
106,
72,
506,
139
],
"lines": [
{
"bbox": [
106,
73,
507,
86
],
"spans": [
{
"bbox": [
106,
73,
507,
86
],
"score": 1.0,
"content": "[47] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
127,
83,
506,
96
],
"spans": [
{
"bbox": [
127,
83,
506,
96
],
"score": 1.0,
"content": "incremental learning. In IEEE Conference on Computer Vision and Pattern Recognition,",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
127,
94,
506,
107
],
"spans": [
{
"bbox": [
127,
94,
506,
107
],
"score": 1.0,
"content": "CVPR 2021, virtual, June 19-25, 2021, pages 2544–2553. Computer Vision Foundation /",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
127,
105,
505,
118
],
"spans": [
{
"bbox": [
127,
105,
505,
118
],
"score": 1.0,
"content": "IEEE, 2021. URL https://openaccess.thecvf.com/content/CVPR2021/html/Liu_",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
127,
117,
502,
129
],
"spans": [
{
"bbox": [
127,
117,
502,
129
],
"score": 1.0,
"content": "Adaptive_Aggregation_Networks_for_Class-Incremental_Learning_CVPR_2021_",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
126,
128,
186,
139
],
"spans": [
{
"bbox": [
126,
128,
186,
139
],
"score": 1.0,
"content": "paper.html.",
"type": "text"
}
],
"index": 5
}
],
"index": 2.5,
"bbox_fs": [
106,
73,
507,
139
]
},
{
"type": "text",
"bbox": [
107,
145,
506,
179
],
"lines": [
{
"bbox": [
105,
144,
506,
158
],
"spans": [
{
"bbox": [
105,
144,
506,
158
],
"score": 1.0,
"content": "[48] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, and Christoph H. Lampert. icarl: Incremental",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
127,
157,
507,
169
],
"spans": [
{
"bbox": [
127,
157,
507,
169
],
"score": 1.0,
"content": "classifier and representation learning. CoRR, abs/1611.07725, 2016. URL http://arxiv.",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
127,
168,
228,
180
],
"spans": [
{
"bbox": [
127,
168,
228,
180
],
"score": 1.0,
"content": "org/abs/1611.07725.",
"type": "text"
}
],
"index": 8
}
],
"index": 7,
"bbox_fs": [
105,
144,
507,
180
]
},
{
"type": "text",
"bbox": [
107,
186,
506,
253
],
"lines": [
{
"bbox": [
106,
186,
506,
199
],
"spans": [
{
"bbox": [
106,
186,
506,
199
],
"score": 1.0,
"content": "[49] Ameya Prabhu, Philip H. S. Torr, and Puneet K. Dokania. Gdumb: A simple approach that",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
126,
197,
507,
211
],
"spans": [
{
"bbox": [
126,
197,
507,
211
],
"score": 1.0,
"content": "questions our progress in continual learning. In Andrea Vedaldi, Horst Bischof, Thomas Brox,",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
127,
208,
506,
222
],
"spans": [
{
"bbox": [
127,
208,
506,
222
],
"score": 1.0,
"content": "and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
127,
219,
506,
232
],
"spans": [
{
"bbox": [
127,
219,
506,
232
],
"score": 1.0,
"content": "Glasgow, UK, August 23-28, 2020, Proceedings, Part II, volume 12347 of Lecture Notes in",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
126,
229,
507,
244
],
"spans": [
{
"bbox": [
126,
229,
507,
244
],
"score": 1.0,
"content": "Computer Science, pages 524–540. Springer, 2020. doi: 10.1007/978-3-030-58536-5\\_31. URL",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
127,
241,
366,
254
],
"spans": [
{
"bbox": [
127,
241,
366,
254
],
"score": 1.0,
"content": "https://doi.org/10.1007/978-3-030-58536-5_31.",
"type": "text"
}
],
"index": 14
}
],
"index": 11.5,
"bbox_fs": [
106,
186,
507,
254
]
},
{
"type": "text",
"bbox": [
107,
259,
506,
336
],
"lines": [
{
"bbox": [
106,
260,
505,
272
],
"spans": [
{
"bbox": [
106,
260,
505,
272
],
"score": 1.0,
"content": "[50] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
127,
271,
505,
282
],
"spans": [
{
"bbox": [
127,
271,
505,
282
],
"score": 1.0,
"content": "selection for online continual learning. In Hanna M. Wallach, Hugo Larochelle, Alina",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
126,
281,
506,
295
],
"spans": [
{
"bbox": [
126,
281,
506,
295
],
"score": 1.0,
"content": "Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
126,
292,
506,
306
],
"spans": [
{
"bbox": [
126,
292,
506,
306
],
"score": 1.0,
"content": "in Neural Information Processing Systems 32: Annual Conference on Neural Information",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
127,
304,
506,
316
],
"spans": [
{
"bbox": [
127,
304,
506,
316
],
"score": 1.0,
"content": "Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada,",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
126,
314,
506,
327
],
"spans": [
{
"bbox": [
126,
314,
506,
327
],
"score": 1.0,
"content": "pages 11816–11825, 2019. URL https://proceedings.neurips.cc/paper/2019/hash/",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
127,
325,
375,
337
],
"spans": [
{
"bbox": [
127,
325,
375,
337
],
"score": 1.0,
"content": "e562cd9c0768d5464b64cf61da7fc6bb-Abstract.html.",
"type": "text"
}
],
"index": 21
}
],
"index": 18,
"bbox_fs": [
106,
260,
506,
337
]
},
{
"type": "text",
"bbox": [
107,
344,
506,
421
],
"lines": [
{
"bbox": [
105,
344,
506,
357
],
"spans": [
{
"bbox": [
105,
344,
506,
357
],
"score": 1.0,
"content": "[51] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin,",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
127,
355,
506,
367
],
"spans": [
{
"bbox": [
127,
355,
506,
367
],
"score": 1.0,
"content": "and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. In Hanna M.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
127,
366,
505,
378
],
"spans": [
{
"bbox": [
127,
366,
505,
378
],
"score": 1.0,
"content": "Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
127,
376,
505,
390
],
"spans": [
{
"bbox": [
127,
376,
505,
390
],
"score": 1.0,
"content": "Garnett, editors, Advances in Neural Information Processing Systems 32: Annual Conference on",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
127,
388,
506,
401
],
"spans": [
{
"bbox": [
127,
388,
506,
401
],
"score": 1.0,
"content": "Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver,",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
127,
398,
505,
411
],
"spans": [
{
"bbox": [
127,
398,
505,
411
],
"score": 1.0,
"content": "BC, Canada, pages 11849–11860, 2019. URL https://proceedings.neurips.cc/paper/",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
128,
410,
427,
421
],
"spans": [
{
"bbox": [
128,
410,
427,
421
],
"score": 1.0,
"content": "2019/hash/15825aee15eb335cc13f9b559f166ee8-Abstract.html.",
"type": "text"
}
],
"index": 28
}
],
"index": 25,
"bbox_fs": [
105,
344,
506,
421
]
},
{
"type": "text",
"bbox": [
107,
428,
506,
505
],
"lines": [
{
"bbox": [
105,
428,
507,
441
],
"spans": [
{
"bbox": [
105,
428,
507,
441
],
"score": 1.0,
"content": "[52] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
126,
439,
506,
453
],
"spans": [
{
"bbox": [
126,
439,
506,
453
],
"score": 1.0,
"content": "class incremental learning without forgetting. In 2020 IEEE/CVF Conference on Computer",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
128,
450,
507,
462
],
"spans": [
{
"bbox": [
128,
450,
507,
462
],
"score": 1.0,
"content": "Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 12242–",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
128,
461,
505,
473
],
"spans": [
{
"bbox": [
128,
461,
505,
473
],
"score": 1.0,
"content": "12251. Computer Vision Foundation / IEEE, 2020. doi: 10.1109/CVPR42600.2020.01226.",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
126,
471,
505,
486
],
"spans": [
{
"bbox": [
126,
471,
505,
486
],
"score": 1.0,
"content": "URL https://openaccess.thecvf.com/content_CVPR_2020/html/Liu_Mnemonics_",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
127,
483,
502,
496
],
"spans": [
{
"bbox": [
127,
483,
502,
496
],
"score": 1.0,
"content": "Training_Multi-Class_Incremental_Learning_Without_Forgetting_CVPR_2020_",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
126,
494,
186,
507
],
"spans": [
{
"bbox": [
126,
494,
186,
507
],
"score": 1.0,
"content": "paper.html.",
"type": "text"
}
],
"index": 35
}
],
"index": 32,
"bbox_fs": [
105,
428,
507,
507
]
},
{
"type": "text",
"bbox": [
105,
512,
505,
535
],
"lines": [
{
"bbox": [
106,
512,
506,
525
],
"spans": [
{
"bbox": [
106,
512,
506,
525
],
"score": 1.0,
"content": "[53] Gido M. van de Ven and Andreas S. Tolias. Three scenarios for continual learning. CoRR,",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
127,
523,
411,
536
],
"spans": [
{
"bbox": [
127,
523,
411,
536
],
"score": 1.0,
"content": "abs/1904.07734, 2019. URL http://arxiv.org/abs/1904.07734.",
"type": "text"
}
],
"index": 37
}
],
"index": 36.5,
"bbox_fs": [
106,
512,
506,
536
]
},
{
"type": "text",
"bbox": [
107,
541,
506,
609
],
"lines": [
{
"bbox": [
105,
542,
506,
554
],
"spans": [
{
"bbox": [
105,
542,
506,
554
],
"score": 1.0,
"content": "[54] Francisco M. Castro, Manuel J. Marín-Jiménez, Nicolás Guil, Cordelia Schmid, and Karteek",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
126,
552,
506,
566
],
"spans": [
{
"bbox": [
126,
552,
506,
566
],
"score": 1.0,
"content": "Alahari. End-to-end incremental learning. In Vittorio Ferrari, Martial Hebert, Cristian Smin-",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
126,
562,
507,
578
],
"spans": [
{
"bbox": [
126,
562,
507,
578
],
"score": 1.0,
"content": "chisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018 - 15th European Conference,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
127,
574,
506,
587
],
"spans": [
{
"bbox": [
127,
574,
506,
587
],
"score": 1.0,
"content": "Munich, Germany, September 8-14, 2018, Proceedings, Part XII, volume 11216 of Lecture Notes",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
127,
586,
507,
599
],
"spans": [
{
"bbox": [
127,
586,
507,
599
],
"score": 1.0,
"content": "in Computer Science, pages 241–257. Springer, 2018. doi: 10.1007/978-3-030-01258-8\\_15.",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
128,
597,
388,
610
],
"spans": [
{
"bbox": [
128,
597,
388,
610
],
"score": 1.0,
"content": "URL https://doi.org/10.1007/978-3-030-01258-8_15.",
"type": "text"
}
],
"index": 43
}
],
"index": 40.5,
"bbox_fs": [
105,
542,
507,
610
]
},
{
"type": "text",
"bbox": [
107,
615,
506,
671
],
"lines": [
{
"bbox": [
106,
615,
505,
628
],
"spans": [
{
"bbox": [
106,
615,
505,
628
],
"score": 1.0,
"content": "[55] Yutian Chen, Max Welling, and Alexander J. Smola. Super-samples from kernel herding. In Peter",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
127,
626,
505,
640
],
"spans": [
{
"bbox": [
127,
626,
505,
640
],
"score": 1.0,
"content": "Grünwald and Peter Spirtes, editors, UAI 2010, Proceedings of the Twenty-Sixth Conference",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
126,
635,
506,
651
],
"spans": [
{
"bbox": [
126,
635,
506,
651
],
"score": 1.0,
"content": "on Uncertainty in Artificial Intelligence, Catalina Island, CA, USA, July 8-11, 2010, pages",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
127,
647,
507,
662
],
"spans": [
{
"bbox": [
127,
647,
507,
662
],
"score": 1.0,
"content": "109–116. AUAI Press, 2010. URL https://dslpitt.org/uai/displayArticleDetails.",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
127,
659,
394,
672
],
"spans": [
{
"bbox": [
127,
659,
170,
672
],
"score": 1.0,
"content": "jsp?mmnu",
"type": "text"
},
{
"bbox": [
171,
661,
177,
668
],
"score": 0.46,
"content": "\\equiv",
"type": "inline_equation"
},
{
"bbox": [
177,
659,
207,
672
],
"score": 1.0,
"content": "1&smnu",
"type": "text"
},
{
"bbox": [
207,
661,
213,
668
],
"score": 0.52,
"content": "\\underset { . } { = }",
"type": "inline_equation"
},
{
"bbox": [
213,
659,
275,
672
],
"score": 1.0,
"content": "2&article_id",
"type": "text"
},
{
"bbox": [
275,
660,
281,
669
],
"score": 0.33,
"content": "\\equiv",
"type": "inline_equation"
},
{
"bbox": [
281,
659,
394,
672
],
"score": 1.0,
"content": "2148&proceeding_id=26.",
"type": "text"
}
],
"index": 48
}
],
"index": 46,
"bbox_fs": [
106,
615,
507,
672
]
},
{
"type": "text",
"bbox": [
108,
678,
505,
722
],
"lines": [
{
"bbox": [
106,
678,
505,
691
],
"spans": [
{
"bbox": [
106,
678,
505,
691
],
"score": 1.0,
"content": "[56] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
126,
688,
507,
702
],
"spans": [
{
"bbox": [
126,
688,
507,
702
],
"score": 1.0,
"content": "attacks against machine learning models. In 2017 IEEE Symposium on Security and Privacy,",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
126,
699,
507,
713
],
"spans": [
{
"bbox": [
126,
699,
507,
713
],
"score": 1.0,
"content": "SP 2017, San Jose, CA, USA, May 22-26, 2017, pages 3–18. IEEE Computer Society, 2017. doi:",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
127,
710,
419,
723
],
"spans": [
{
"bbox": [
127,
710,
419,
723
],
"score": 1.0,
"content": "10.1109/SP.2017.41. URL https://doi.org/10.1109/SP.2017.41.",
"type": "text"
}
],
"index": 52
}
],
"index": 50.5,
"bbox_fs": [
106,
678,
507,
723
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
105,
72,
504,
106
],
"lines": [
{
"bbox": [
106,
72,
505,
86
],
"spans": [
{
"bbox": [
106,
72,
505,
86
],
"score": 1.0,
"content": "[57] Yunhui Long, Vincent Bindschaedler, Lei Wang, Diyue Bu, Xiaofeng Wang, Haixu Tang, Carl A.",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
127,
82,
505,
97
],
"spans": [
{
"bbox": [
127,
82,
505,
97
],
"score": 1.0,
"content": "Gunter, and Kai Chen. Understanding membership inferences on well-generalized learning",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
127,
95,
475,
107
],
"spans": [
{
"bbox": [
127,
95,
475,
107
],
"score": 1.0,
"content": "models. CoRR, abs/1802.04889, 2018. URL http://arxiv.org/abs/1802.04889.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "text",
"bbox": [
107,
113,
506,
158
],
"lines": [
{
"bbox": [
106,
113,
506,
126
],
"spans": [
{
"bbox": [
106,
113,
506,
126
],
"score": 1.0,
"content": "[58] Ahmed Salem, Yang Zhang, Mathias Humbert, Pascal Berrang, Mario Fritz, and Michael",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
127,
123,
506,
137
],
"spans": [
{
"bbox": [
127,
123,
506,
137
],
"score": 1.0,
"content": "Backes. Ml-leaks: Model and data independent membership inference attacks and defenses on",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
127,
135,
506,
148
],
"spans": [
{
"bbox": [
127,
135,
506,
148
],
"score": 1.0,
"content": "machine learning models. In 26th Annual Network and Distributed System Security Symposium,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
127,
147,
498,
158
],
"spans": [
{
"bbox": [
127,
147,
498,
158
],
"score": 1.0,
"content": "NDSS 2019, San Diego, California, USA, February 24-27, 2019. The Internet Society, 2019.",
"type": "text"
}
],
"index": 6
}
],
"index": 4.5
},
{
"type": "text",
"bbox": [
107,
164,
505,
209
],
"lines": [
{
"bbox": [
106,
164,
505,
178
],
"spans": [
{
"bbox": [
106,
164,
505,
178
],
"score": 1.0,
"content": "[59] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
126,
174,
506,
190
],
"spans": [
{
"bbox": [
126,
174,
506,
190
],
"score": 1.0,
"content": "deep learning requires rethinking generalization. In 5th International Conference on Learning",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
126,
186,
507,
200
],
"spans": [
{
"bbox": [
126,
186,
507,
200
],
"score": 1.0,
"content": "Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings.",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
127,
197,
468,
211
],
"spans": [
{
"bbox": [
127,
197,
409,
211
],
"score": 1.0,
"content": "OpenReview.net, 2017. URL https://openreview.net/forum?id",
"type": "text"
},
{
"bbox": [
410,
199,
415,
207
],
"score": 0.52,
"content": "=",
"type": "inline_equation"
},
{
"bbox": [
416,
197,
468,
211
],
"score": 1.0,
"content": "Sy8gdB9xx.",
"type": "text"
}
],
"index": 10
}
],
"index": 8.5
},
{
"type": "text",
"bbox": [
107,
216,
506,
282
],
"lines": [
{
"bbox": [
105,
215,
506,
230
],
"spans": [
{
"bbox": [
105,
215,
506,
230
],
"score": 1.0,
"content": "[60] Milad Nasr, Reza Shokri, and Amir Houmansadr. Machine learning with membership privacy",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
127,
227,
505,
240
],
"spans": [
{
"bbox": [
127,
227,
505,
240
],
"score": 1.0,
"content": "using adversarial regularization. In David Lie, Mohammad Mannan, Michael Backes, and",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
127,
238,
506,
251
],
"spans": [
{
"bbox": [
127,
238,
506,
251
],
"score": 1.0,
"content": "XiaoFeng Wang, editors, Proceedings of the 2018 ACM SIGSAC Conference on Computer",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
127,
248,
506,
263
],
"spans": [
{
"bbox": [
127,
248,
506,
263
],
"score": 1.0,
"content": "and Communications Security, CCS 2018, Toronto, ON, Canada, October 15-19, 2018, pages",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
126,
259,
506,
273
],
"spans": [
{
"bbox": [
126,
259,
506,
273
],
"score": 1.0,
"content": "634–646. ACM, 2018. doi: 10.1145/3243734.3243855. URL https://doi.org/10.1145/",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
128,
272,
211,
282
],
"spans": [
{
"bbox": [
128,
272,
211,
282
],
"score": 1.0,
"content": "3243734.3243855.",
"type": "text"
}
],
"index": 16
}
],
"index": 13.5
},
{
"type": "text",
"bbox": [
107,
289,
506,
356
],
"lines": [
{
"bbox": [
106,
289,
506,
303
],
"spans": [
{
"bbox": [
106,
289,
506,
303
],
"score": 1.0,
"content": "[61] Jinyuan Jia, Ahmed Salem, Michael Backes, Yang Zhang, and Neil Zhenqiang Gong. Memguard:",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
127,
301,
506,
314
],
"spans": [
{
"bbox": [
127,
301,
506,
314
],
"score": 1.0,
"content": "Defending against black-box membership inference attacks via adversarial examples. In Lorenzo",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
127,
311,
506,
325
],
"spans": [
{
"bbox": [
127,
311,
506,
325
],
"score": 1.0,
"content": "Cavallaro, Johannes Kinder, XiaoFeng Wang, and Jonathan Katz, editors, Proceedings of the",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
126,
322,
506,
335
],
"spans": [
{
"bbox": [
126,
322,
506,
335
],
"score": 1.0,
"content": "2019 ACM SIGSAC Conference on Computer and Communications Security, CCS 2019, London,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
127,
333,
507,
346
],
"spans": [
{
"bbox": [
127,
333,
507,
346
],
"score": 1.0,
"content": "UK, November 11-15, 2019, pages 259–274. ACM, 2019. doi: 10.1145/3319535.3363201.",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
128,
345,
360,
357
],
"spans": [
{
"bbox": [
128,
345,
360,
357
],
"score": 1.0,
"content": "URL https://doi.org/10.1145/3319535.3363201.",
"type": "text"
}
],
"index": 22
}
],
"index": 19.5
},
{
"type": "text",
"bbox": [
105,
363,
503,
387
],
"lines": [
{
"bbox": [
105,
362,
505,
376
],
"spans": [
{
"bbox": [
105,
362,
505,
376
],
"score": 1.0,
"content": "[62] tensorflow/privacy: library for training machine learning models with privacy for training data,",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
127,
373,
375,
388
],
"spans": [
{
"bbox": [
127,
373,
375,
388
],
"score": 1.0,
"content": "2022. URL https://github.com/tensorflow/privacy.",
"type": "text"
}
],
"index": 24
}
],
"index": 23.5
},
{
"type": "text",
"bbox": [
107,
393,
505,
448
],
"lines": [
{
"bbox": [
106,
392,
505,
406
],
"spans": [
{
"bbox": [
106,
392,
505,
406
],
"score": 1.0,
"content": "[63] Samuel Yeom, Irene Giacomelli, Matt Fredrikson, and Somesh Jha. Privacy risk in machine",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
127,
404,
505,
417
],
"spans": [
{
"bbox": [
127,
404,
505,
417
],
"score": 1.0,
"content": "learning: Analyzing the connection to overfitting. In 31st IEEE Computer Security Foundations",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
127,
415,
506,
428
],
"spans": [
{
"bbox": [
127,
415,
506,
428
],
"score": 1.0,
"content": "Symposium, CSF 2018, Oxford, United Kingdom, July 9-12, 2018, pages 268–282. IEEE",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
127,
426,
505,
439
],
"spans": [
{
"bbox": [
127,
426,
505,
439
],
"score": 1.0,
"content": "Computer Society, 2018. doi: 10.1109/CSF.2018.00027. URL https://doi.org/10.1109/",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
128,
437,
206,
449
],
"spans": [
{
"bbox": [
128,
437,
206,
449
],
"score": 1.0,
"content": "CSF.2018.00027.",
"type": "text"
}
],
"index": 29
}
],
"index": 27
}
],
"page_idx": 14,
"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": "15",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
105,
72,
504,
106
],
"lines": [
{
"bbox": [
106,
72,
505,
86
],
"spans": [
{
"bbox": [
106,
72,
505,
86
],
"score": 1.0,
"content": "[57] Yunhui Long, Vincent Bindschaedler, Lei Wang, Diyue Bu, Xiaofeng Wang, Haixu Tang, Carl A.",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
127,
82,
505,
97
],
"spans": [
{
"bbox": [
127,
82,
505,
97
],
"score": 1.0,
"content": "Gunter, and Kai Chen. Understanding membership inferences on well-generalized learning",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
127,
95,
475,
107
],
"spans": [
{
"bbox": [
127,
95,
475,
107
],
"score": 1.0,
"content": "models. CoRR, abs/1802.04889, 2018. URL http://arxiv.org/abs/1802.04889.",
"type": "text"
}
],
"index": 2
}
],
"index": 1,
"bbox_fs": [
106,
72,
505,
107
]
},
{
"type": "text",
"bbox": [
107,
113,
506,
158
],
"lines": [
{
"bbox": [
106,
113,
506,
126
],
"spans": [
{
"bbox": [
106,
113,
506,
126
],
"score": 1.0,
"content": "[58] Ahmed Salem, Yang Zhang, Mathias Humbert, Pascal Berrang, Mario Fritz, and Michael",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
127,
123,
506,
137
],
"spans": [
{
"bbox": [
127,
123,
506,
137
],
"score": 1.0,
"content": "Backes. Ml-leaks: Model and data independent membership inference attacks and defenses on",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
127,
135,
506,
148
],
"spans": [
{
"bbox": [
127,
135,
506,
148
],
"score": 1.0,
"content": "machine learning models. In 26th Annual Network and Distributed System Security Symposium,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
127,
147,
498,
158
],
"spans": [
{
"bbox": [
127,
147,
498,
158
],
"score": 1.0,
"content": "NDSS 2019, San Diego, California, USA, February 24-27, 2019. The Internet Society, 2019.",
"type": "text"
}
],
"index": 6
}
],
"index": 4.5,
"bbox_fs": [
106,
113,
506,
158
]
},
{
"type": "text",
"bbox": [
107,
164,
505,
209
],
"lines": [
{
"bbox": [
106,
164,
505,
178
],
"spans": [
{
"bbox": [
106,
164,
505,
178
],
"score": 1.0,
"content": "[59] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
126,
174,
506,
190
],
"spans": [
{
"bbox": [
126,
174,
506,
190
],
"score": 1.0,
"content": "deep learning requires rethinking generalization. In 5th International Conference on Learning",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
126,
186,
507,
200
],
"spans": [
{
"bbox": [
126,
186,
507,
200
],
"score": 1.0,
"content": "Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings.",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
127,
197,
468,
211
],
"spans": [
{
"bbox": [
127,
197,
409,
211
],
"score": 1.0,
"content": "OpenReview.net, 2017. URL https://openreview.net/forum?id",
"type": "text"
},
{
"bbox": [
410,
199,
415,
207
],
"score": 0.52,
"content": "=",
"type": "inline_equation"
},
{
"bbox": [
416,
197,
468,
211
],
"score": 1.0,
"content": "Sy8gdB9xx.",
"type": "text"
}
],
"index": 10
}
],
"index": 8.5,
"bbox_fs": [
106,
164,
507,
211
]
},
{
"type": "text",
"bbox": [
107,
216,
506,
282
],
"lines": [
{
"bbox": [
105,
215,
506,
230
],
"spans": [
{
"bbox": [
105,
215,
506,
230
],
"score": 1.0,
"content": "[60] Milad Nasr, Reza Shokri, and Amir Houmansadr. Machine learning with membership privacy",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
127,
227,
505,
240
],
"spans": [
{
"bbox": [
127,
227,
505,
240
],
"score": 1.0,
"content": "using adversarial regularization. In David Lie, Mohammad Mannan, Michael Backes, and",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
127,
238,
506,
251
],
"spans": [
{
"bbox": [
127,
238,
506,
251
],
"score": 1.0,
"content": "XiaoFeng Wang, editors, Proceedings of the 2018 ACM SIGSAC Conference on Computer",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
127,
248,
506,
263
],
"spans": [
{
"bbox": [
127,
248,
506,
263
],
"score": 1.0,
"content": "and Communications Security, CCS 2018, Toronto, ON, Canada, October 15-19, 2018, pages",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
126,
259,
506,
273
],
"spans": [
{
"bbox": [
126,
259,
506,
273
],
"score": 1.0,
"content": "634–646. ACM, 2018. doi: 10.1145/3243734.3243855. URL https://doi.org/10.1145/",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
128,
272,
211,
282
],
"spans": [
{
"bbox": [
128,
272,
211,
282
],
"score": 1.0,
"content": "3243734.3243855.",
"type": "text"
}
],
"index": 16
}
],
"index": 13.5,
"bbox_fs": [
105,
215,
506,
282
]
},
{
"type": "text",
"bbox": [
107,
289,
506,
356
],
"lines": [
{
"bbox": [
106,
289,
506,
303
],
"spans": [
{
"bbox": [
106,
289,
506,
303
],
"score": 1.0,
"content": "[61] Jinyuan Jia, Ahmed Salem, Michael Backes, Yang Zhang, and Neil Zhenqiang Gong. Memguard:",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
127,
301,
506,
314
],
"spans": [
{
"bbox": [
127,
301,
506,
314
],
"score": 1.0,
"content": "Defending against black-box membership inference attacks via adversarial examples. In Lorenzo",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
127,
311,
506,
325
],
"spans": [
{
"bbox": [
127,
311,
506,
325
],
"score": 1.0,
"content": "Cavallaro, Johannes Kinder, XiaoFeng Wang, and Jonathan Katz, editors, Proceedings of the",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
126,
322,
506,
335
],
"spans": [
{
"bbox": [
126,
322,
506,
335
],
"score": 1.0,
"content": "2019 ACM SIGSAC Conference on Computer and Communications Security, CCS 2019, London,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
127,
333,
507,
346
],
"spans": [
{
"bbox": [
127,
333,
507,
346
],
"score": 1.0,
"content": "UK, November 11-15, 2019, pages 259–274. ACM, 2019. doi: 10.1145/3319535.3363201.",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
128,
345,
360,
357
],
"spans": [
{
"bbox": [
128,
345,
360,
357
],
"score": 1.0,
"content": "URL https://doi.org/10.1145/3319535.3363201.",
"type": "text"
}
],
"index": 22
}
],
"index": 19.5,
"bbox_fs": [
106,
289,
507,
357
]
},
{
"type": "text",
"bbox": [
105,
363,
503,
387
],
"lines": [
{
"bbox": [
105,
362,
505,
376
],
"spans": [
{
"bbox": [
105,
362,
505,
376
],
"score": 1.0,
"content": "[62] tensorflow/privacy: library for training machine learning models with privacy for training data,",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
127,
373,
375,
388
],
"spans": [
{
"bbox": [
127,
373,
375,
388
],
"score": 1.0,
"content": "2022. URL https://github.com/tensorflow/privacy.",
"type": "text"
}
],
"index": 24
}
],
"index": 23.5,
"bbox_fs": [
105,
362,
505,
388
]
},
{
"type": "text",
"bbox": [
107,
393,
505,
448
],
"lines": [
{
"bbox": [
106,
392,
505,
406
],
"spans": [
{
"bbox": [
106,
392,
505,
406
],
"score": 1.0,
"content": "[63] Samuel Yeom, Irene Giacomelli, Matt Fredrikson, and Somesh Jha. Privacy risk in machine",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
127,
404,
505,
417
],
"spans": [
{
"bbox": [
127,
404,
505,
417
],
"score": 1.0,
"content": "learning: Analyzing the connection to overfitting. In 31st IEEE Computer Security Foundations",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
127,
415,
506,
428
],
"spans": [
{
"bbox": [
127,
415,
506,
428
],
"score": 1.0,
"content": "Symposium, CSF 2018, Oxford, United Kingdom, July 9-12, 2018, pages 268–282. IEEE",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
127,
426,
505,
439
],
"spans": [
{
"bbox": [
127,
426,
505,
439
],
"score": 1.0,
"content": "Computer Society, 2018. doi: 10.1109/CSF.2018.00027. URL https://doi.org/10.1109/",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
128,
437,
206,
449
],
"spans": [
{
"bbox": [
128,
437,
206,
449
],
"score": 1.0,
"content": "CSF.2018.00027.",
"type": "text"
}
],
"index": 29
}
],
"index": 27,
"bbox_fs": [
106,
392,
506,
449
]
}
]
}
],
"_backend": "pipeline",
"_version_name": "2.2.2"
}