diff --git "a/parse/train/SJTQLdqlg/SJTQLdqlg_middle.json" "b/parse/train/SJTQLdqlg/SJTQLdqlg_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/train/SJTQLdqlg/SJTQLdqlg_middle.json"
@@ -0,0 +1,28695 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 78,
+ 412,
+ 96
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 77,
+ 414,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 77,
+ 414,
+ 97
+ ],
+ "score": 1.0,
+ "content": "LEARNING TO REMEMBER RARE EVENTS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 115,
+ 252,
+ 148
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 114,
+ 181,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 114,
+ 181,
+ 127
+ ],
+ "score": 1.0,
+ "content": "Łukasz Kaiser∗",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 111,
+ 125,
+ 169,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 125,
+ 169,
+ 138
+ ],
+ "score": 1.0,
+ "content": "Google Brain",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 112,
+ 136,
+ 253,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 136,
+ 253,
+ 149
+ ],
+ "score": 1.0,
+ "content": "lukaszkaiser@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 371,
+ 115,
+ 498,
+ 148
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 369,
+ 114,
+ 438,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 369,
+ 114,
+ 438,
+ 126
+ ],
+ "score": 1.0,
+ "content": "Ofir Nachum∗†",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 371,
+ 126,
+ 428,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 126,
+ 428,
+ 137
+ ],
+ "score": 1.0,
+ "content": "Google Brain",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 371,
+ 137,
+ 499,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 137,
+ 499,
+ 149
+ ],
+ "score": 1.0,
+ "content": "ofirnachum@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 165,
+ 210,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 163,
+ 166,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 163,
+ 166,
+ 177
+ ],
+ "score": 1.0,
+ "content": "Aurko Roy‡",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 111,
+ 176,
+ 169,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 176,
+ 169,
+ 187
+ ],
+ "score": 1.0,
+ "content": "Georgia Tech",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 111,
+ 187,
+ 210,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 187,
+ 210,
+ 199
+ ],
+ "score": 1.0,
+ "content": "aurko@gatech.edu",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 369,
+ 165,
+ 473,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 368,
+ 164,
+ 428,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 368,
+ 164,
+ 428,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Samy Bengio",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 368,
+ 176,
+ 426,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 368,
+ 176,
+ 426,
+ 188
+ ],
+ "score": 1.0,
+ "content": "Google Brain",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 368,
+ 187,
+ 474,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 368,
+ 187,
+ 474,
+ 200
+ ],
+ "score": 1.0,
+ "content": "bengio@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 227,
+ 333,
+ 239
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 226,
+ 336,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 226,
+ 336,
+ 241
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 253,
+ 468,
+ 330
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 253,
+ 470,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 253,
+ 470,
+ 267
+ ],
+ "score": 1.0,
+ "content": "Despite recent advances, memory-augmented deep neural networks are still lim-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 264,
+ 469,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 264,
+ 469,
+ 277
+ ],
+ "score": 1.0,
+ "content": "ited when it comes to life-long and one-shot learning, especially in remembering",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 140,
+ 275,
+ 470,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 275,
+ 470,
+ 289
+ ],
+ "score": 1.0,
+ "content": "rare events. We present a large-scale life-long memory module for use in deep",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 286,
+ 470,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 286,
+ 470,
+ 299
+ ],
+ "score": 1.0,
+ "content": "learning. The module exploits fast nearest-neighbor algorithms for efficiency and",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 298,
+ 469,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 298,
+ 469,
+ 309
+ ],
+ "score": 1.0,
+ "content": "thus scales to large memory sizes. Except for the nearest-neighbor query, the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 308,
+ 470,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 308,
+ 470,
+ 321
+ ],
+ "score": 1.0,
+ "content": "module is fully differentiable and trained end-to-end with no extra supervision. It",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 318,
+ 457,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 318,
+ 457,
+ 333
+ ],
+ "score": 1.0,
+ "content": "operates in a life-long manner, i.e., without the need to reset it during training.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 333,
+ 468,
+ 431
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 333,
+ 469,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 333,
+ 469,
+ 345
+ ],
+ "score": 1.0,
+ "content": "Our memory module can be easily added to any part of a supervised neural net-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 343,
+ 469,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 343,
+ 469,
+ 356
+ ],
+ "score": 1.0,
+ "content": "work. To show its versatility we add it to a number of networks, from simple",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 367
+ ],
+ "score": 1.0,
+ "content": "convolutional ones tested on image classification to deep sequence-to-sequence",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 365,
+ 470,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 365,
+ 470,
+ 378
+ ],
+ "score": 1.0,
+ "content": "and recurrent-convolutional models. In all cases, the enhanced network gains the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 389
+ ],
+ "score": 1.0,
+ "content": "ability to remember and do life-long one-shot learning. Our module remembers",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 388,
+ 469,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 388,
+ 469,
+ 400
+ ],
+ "score": 1.0,
+ "content": "training examples shown many thousands of steps in the past and it can success-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 141,
+ 398,
+ 470,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 398,
+ 470,
+ 411
+ ],
+ "score": 1.0,
+ "content": "fully generalize from them. We set new state-of-the-art for one-shot learning on",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 141,
+ 410,
+ 469,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 410,
+ 469,
+ 422
+ ],
+ "score": 1.0,
+ "content": "the Omniglot dataset and demonstrate, for the first time, life-long one-shot learn-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 141,
+ 420,
+ 439,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 420,
+ 439,
+ 432
+ ],
+ "score": 1.0,
+ "content": "ing in recurrent neural networks on a large-scale machine translation task.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 455,
+ 205,
+ 467
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 208,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 208,
+ 470
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 480,
+ 505,
+ 579
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 506,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 506,
+ 493
+ ],
+ "score": 1.0,
+ "content": "Machine learning systems have been successful in many domains, from computer vision",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 503
+ ],
+ "score": 1.0,
+ "content": "(Krizhevsky et al., 2012) to speech recognition (Hinton et al., 2012) and machine translation",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 501,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "(Sutskever et al., 2014; Bahdanau et al., 2014; Cho et al., 2014). Neural machine translation (NMT)",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 104,
+ 513,
+ 506,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 513,
+ 506,
+ 526
+ ],
+ "score": 1.0,
+ "content": "is so successful that for some language pairs it approaches, on average, the quality of human trans-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 523,
+ 506,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 506,
+ 537
+ ],
+ "score": 1.0,
+ "content": "lators (Wu et al., 2016). The words on average are crucial though. When a sentence resembles one",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 534,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 534,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "from the abundant training data, the translation will be accurate. However, when encountering a",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 546,
+ 505,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 505,
+ 559
+ ],
+ "score": 1.0,
+ "content": "rare word such as Dostoevsky (in German, Dostojewski), many models will fail. The correct Ger-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 557,
+ 506,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 557,
+ 506,
+ 570
+ ],
+ "score": 1.0,
+ "content": "man translation of Dostoevsky does not appear enough times in the training data for the model to",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 568,
+ 234,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 568,
+ 234,
+ 581
+ ],
+ "score": 1.0,
+ "content": "sufficiently learn its translation.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 585,
+ 505,
+ 640
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 585,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 585,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "While more example sentences concerning the famous Russian author might eventually be added to",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 596,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 596,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "the training data, there are many other rare words or rare events of other kinds. This illustrates a",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "general problem with current deep learning models: it is necessary to extend the training data and",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "re-train them to handle such rare or new events. Humans, on the other hand, learn in a life-long",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 629,
+ 254,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 629,
+ 254,
+ 641
+ ],
+ "score": 1.0,
+ "content": "fashion, often from single examples.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 646,
+ 502,
+ 690
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 646,
+ 504,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 646,
+ 504,
+ 658
+ ],
+ "score": 1.0,
+ "content": "We present a life-long memory module that enables one-shot learning in a variety of neural networks.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 656,
+ 505,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 656,
+ 505,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Our memory module consists of key-value pairs. Keys are activations of a chosen layer of a neural",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 668,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 668,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "network, and values are the ground-truth targets for the given example. This way, as the network",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 679,
+ 504,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 679,
+ 504,
+ 691
+ ],
+ "score": 1.0,
+ "content": "is trained, its memory increases and becomes more useful. Eventually it can give predictions that",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 46.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 118,
+ 699,
+ 469,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 698,
+ 258,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 698,
+ 258,
+ 712
+ ],
+ "score": 1.0,
+ "content": "∗First two authors contributed equally.",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 118,
+ 708,
+ 470,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 708,
+ 470,
+ 723
+ ],
+ "score": 1.0,
+ "content": "†Work done as a member of the Google Brain Residency program (g.co/brainresidency).",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 117,
+ 718,
+ 290,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 718,
+ 290,
+ 734
+ ],
+ "score": 1.0,
+ "content": "‡Work done during internship at Google Brain.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "1",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 78,
+ 412,
+ 96
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 77,
+ 414,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 77,
+ 414,
+ 97
+ ],
+ "score": 1.0,
+ "content": "LEARNING TO REMEMBER RARE EVENTS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 113,
+ 115,
+ 252,
+ 148
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 114,
+ 181,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 114,
+ 181,
+ 127
+ ],
+ "score": 1.0,
+ "content": "Łukasz Kaiser∗",
+ "type": "text"
+ }
+ ],
+ "index": 1,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 125,
+ 169,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 125,
+ 169,
+ 138
+ ],
+ "score": 1.0,
+ "content": "Google Brain",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 112,
+ 136,
+ 253,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 136,
+ 253,
+ 149
+ ],
+ "score": 1.0,
+ "content": "lukaszkaiser@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 5,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 369,
+ 114,
+ 438,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 369,
+ 114,
+ 438,
+ 126
+ ],
+ "score": 1.0,
+ "content": "Ofir Nachum∗†",
+ "type": "text"
+ }
+ ],
+ "index": 2,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 371,
+ 126,
+ 428,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 126,
+ 428,
+ 137
+ ],
+ "score": 1.0,
+ "content": "Google Brain",
+ "type": "text"
+ }
+ ],
+ "index": 4,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 371,
+ 137,
+ 499,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 137,
+ 499,
+ 149
+ ],
+ "score": 1.0,
+ "content": "ofirnachum@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 6,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 163,
+ 166,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 163,
+ 166,
+ 177
+ ],
+ "score": 1.0,
+ "content": "Aurko Roy‡",
+ "type": "text"
+ }
+ ],
+ "index": 7,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 176,
+ 169,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 176,
+ 169,
+ 187
+ ],
+ "score": 1.0,
+ "content": "Georgia Tech",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 187,
+ 210,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 187,
+ 210,
+ 199
+ ],
+ "score": 1.0,
+ "content": "aurko@gatech.edu",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 368,
+ 164,
+ 428,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 368,
+ 164,
+ 428,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Samy Bengio",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 368,
+ 176,
+ 426,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 368,
+ 176,
+ 426,
+ 188
+ ],
+ "score": 1.0,
+ "content": "Google Brain",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 368,
+ 187,
+ 474,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 368,
+ 187,
+ 474,
+ 200
+ ],
+ "score": 1.0,
+ "content": "bengio@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 12,
+ "is_list_start_line": true
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 111,
+ 114,
+ 253,
+ 149
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 371,
+ 115,
+ 498,
+ 148
+ ],
+ "lines": [],
+ "index": 4,
+ "bbox_fs": [
+ 369,
+ 114,
+ 499,
+ 149
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 113,
+ 165,
+ 210,
+ 198
+ ],
+ "lines": [],
+ "index": 9,
+ "bbox_fs": [
+ 111,
+ 163,
+ 210,
+ 199
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 369,
+ 165,
+ 473,
+ 199
+ ],
+ "lines": [],
+ "index": 10,
+ "bbox_fs": [
+ 368,
+ 164,
+ 474,
+ 200
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 227,
+ 333,
+ 239
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 226,
+ 336,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 226,
+ 336,
+ 241
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 253,
+ 468,
+ 330
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 253,
+ 470,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 253,
+ 470,
+ 267
+ ],
+ "score": 1.0,
+ "content": "Despite recent advances, memory-augmented deep neural networks are still lim-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 264,
+ 469,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 264,
+ 469,
+ 277
+ ],
+ "score": 1.0,
+ "content": "ited when it comes to life-long and one-shot learning, especially in remembering",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 140,
+ 275,
+ 470,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 275,
+ 470,
+ 289
+ ],
+ "score": 1.0,
+ "content": "rare events. We present a large-scale life-long memory module for use in deep",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 286,
+ 470,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 286,
+ 470,
+ 299
+ ],
+ "score": 1.0,
+ "content": "learning. The module exploits fast nearest-neighbor algorithms for efficiency and",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 298,
+ 469,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 298,
+ 469,
+ 309
+ ],
+ "score": 1.0,
+ "content": "thus scales to large memory sizes. Except for the nearest-neighbor query, the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 308,
+ 470,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 308,
+ 470,
+ 321
+ ],
+ "score": 1.0,
+ "content": "module is fully differentiable and trained end-to-end with no extra supervision. It",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 318,
+ 457,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 318,
+ 457,
+ 333
+ ],
+ "score": 1.0,
+ "content": "operates in a life-long manner, i.e., without the need to reset it during training.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 140,
+ 253,
+ 470,
+ 333
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 333,
+ 468,
+ 431
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 333,
+ 469,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 333,
+ 469,
+ 345
+ ],
+ "score": 1.0,
+ "content": "Our memory module can be easily added to any part of a supervised neural net-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 343,
+ 469,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 343,
+ 469,
+ 356
+ ],
+ "score": 1.0,
+ "content": "work. To show its versatility we add it to a number of networks, from simple",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 367
+ ],
+ "score": 1.0,
+ "content": "convolutional ones tested on image classification to deep sequence-to-sequence",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 365,
+ 470,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 365,
+ 470,
+ 378
+ ],
+ "score": 1.0,
+ "content": "and recurrent-convolutional models. In all cases, the enhanced network gains the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 389
+ ],
+ "score": 1.0,
+ "content": "ability to remember and do life-long one-shot learning. Our module remembers",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 388,
+ 469,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 388,
+ 469,
+ 400
+ ],
+ "score": 1.0,
+ "content": "training examples shown many thousands of steps in the past and it can success-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 141,
+ 398,
+ 470,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 398,
+ 470,
+ 411
+ ],
+ "score": 1.0,
+ "content": "fully generalize from them. We set new state-of-the-art for one-shot learning on",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 141,
+ 410,
+ 469,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 410,
+ 469,
+ 422
+ ],
+ "score": 1.0,
+ "content": "the Omniglot dataset and demonstrate, for the first time, life-long one-shot learn-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 141,
+ 420,
+ 439,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 420,
+ 439,
+ 432
+ ],
+ "score": 1.0,
+ "content": "ing in recurrent neural networks on a large-scale machine translation task.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 141,
+ 333,
+ 470,
+ 432
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 455,
+ 205,
+ 467
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 208,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 208,
+ 470
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 480,
+ 505,
+ 579
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 506,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 506,
+ 493
+ ],
+ "score": 1.0,
+ "content": "Machine learning systems have been successful in many domains, from computer vision",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 503
+ ],
+ "score": 1.0,
+ "content": "(Krizhevsky et al., 2012) to speech recognition (Hinton et al., 2012) and machine translation",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 501,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "(Sutskever et al., 2014; Bahdanau et al., 2014; Cho et al., 2014). Neural machine translation (NMT)",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 104,
+ 513,
+ 506,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 513,
+ 506,
+ 526
+ ],
+ "score": 1.0,
+ "content": "is so successful that for some language pairs it approaches, on average, the quality of human trans-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 523,
+ 506,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 506,
+ 537
+ ],
+ "score": 1.0,
+ "content": "lators (Wu et al., 2016). The words on average are crucial though. When a sentence resembles one",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 534,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 534,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "from the abundant training data, the translation will be accurate. However, when encountering a",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 546,
+ 505,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 505,
+ 559
+ ],
+ "score": 1.0,
+ "content": "rare word such as Dostoevsky (in German, Dostojewski), many models will fail. The correct Ger-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 557,
+ 506,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 557,
+ 506,
+ 570
+ ],
+ "score": 1.0,
+ "content": "man translation of Dostoevsky does not appear enough times in the training data for the model to",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 568,
+ 234,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 568,
+ 234,
+ 581
+ ],
+ "score": 1.0,
+ "content": "sufficiently learn its translation.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 104,
+ 480,
+ 506,
+ 581
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 585,
+ 505,
+ 640
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 585,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 585,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "While more example sentences concerning the famous Russian author might eventually be added to",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 596,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 596,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "the training data, there are many other rare words or rare events of other kinds. This illustrates a",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "general problem with current deep learning models: it is necessary to extend the training data and",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "re-train them to handle such rare or new events. Humans, on the other hand, learn in a life-long",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 629,
+ 254,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 629,
+ 254,
+ 641
+ ],
+ "score": 1.0,
+ "content": "fashion, often from single examples.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 105,
+ 585,
+ 506,
+ 641
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 646,
+ 502,
+ 690
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 646,
+ 504,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 646,
+ 504,
+ 658
+ ],
+ "score": 1.0,
+ "content": "We present a life-long memory module that enables one-shot learning in a variety of neural networks.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 656,
+ 505,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 656,
+ 505,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Our memory module consists of key-value pairs. Keys are activations of a chosen layer of a neural",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 668,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 668,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "network, and values are the ground-truth targets for the given example. This way, as the network",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 679,
+ 504,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 679,
+ 504,
+ 691
+ ],
+ "score": 1.0,
+ "content": "is trained, its memory increases and becomes more useful. Eventually it can give predictions that",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "leverage on knowledge from past data with similar activations. Given a new example, the network",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 493,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 493,
+ 107
+ ],
+ "score": 1.0,
+ "content": "writes it to memory and is able to use it afterwards, even if the example was presented just once.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 46.5,
+ "bbox_fs": [
+ 105,
+ 646,
+ 505,
+ 691
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "leverage on knowledge from past data with similar activations. Given a new example, the network",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 493,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 493,
+ 107
+ ],
+ "score": 1.0,
+ "content": "writes it to memory and is able to use it afterwards, even if the example was presented just once.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 504,
+ 166
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 109,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 109,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "There are many advantages of having a long-term memory. One-shot learning is a desirable property",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": "in its own right, and some tasks, as we will show below, are simply not solvable without it. Even",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "real-world tasks where we have large training sets, such as translation, can benefit from long-term",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "memory. Finally, since the memory can be traced back to training examples, it might help explain",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 460,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 460,
+ 167
+ ],
+ "score": 1.0,
+ "content": "the decisions that the model is making and thus improve understandability of the model.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 171,
+ 505,
+ 204
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 505,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 505,
+ 184
+ ],
+ "score": 1.0,
+ "content": "It is not immediately clear how to measure the performance of a life-long one-shot learning model,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 182,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 182,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": "since most deep learning evaluations focus on the average performance and do not have a one-shot",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 193,
+ 499,
+ 206
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 193,
+ 499,
+ 206
+ ],
+ "score": 1.0,
+ "content": "component. We therefore evaluate in a few ways, to show that our memory module indeed works:",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 213,
+ 506,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 214,
+ 505,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 214,
+ 505,
+ 226
+ ],
+ "score": 1.0,
+ "content": "(1) We evaluate on the well-known one-shot learning task Omniglot, which is the only dataset with",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 122,
+ 223,
+ 506,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 223,
+ 506,
+ 239
+ ],
+ "score": 1.0,
+ "content": "explicit one-shot learning evaluation. This dataset is small and does not benefit from life-long",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 122,
+ 236,
+ 505,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 236,
+ 505,
+ 249
+ ],
+ "score": 1.0,
+ "content": "learning capability of our module, but we still exceed the best previous results and set new",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 123,
+ 247,
+ 187,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 247,
+ 187,
+ 259
+ ],
+ "score": 1.0,
+ "content": "state-of-the-art.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 107,
+ 261,
+ 505,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 261,
+ 505,
+ 274
+ ],
+ "score": 1.0,
+ "content": "(2) We devise a synthetic task that requires life-long one-shot learning. On this task, standard",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 122,
+ 272,
+ 450,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 272,
+ 450,
+ 286
+ ],
+ "score": 1.0,
+ "content": "models fare poorly while our model can solve it well, demonstrating its strengths.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 107,
+ 286,
+ 505,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 286,
+ 505,
+ 301
+ ],
+ "score": 1.0,
+ "content": "(3) Finally, we train an English-German translation model that has our life-long one-shot learning",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 121,
+ 297,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 297,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "module. It retains very good performance on average and is also capable of one-shot learning.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 122,
+ 309,
+ 505,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 309,
+ 505,
+ 323
+ ],
+ "score": 1.0,
+ "content": "On the qualitative side, we find that it can translate rarely-occurring words like Dostoevsky.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 122,
+ 320,
+ 506,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 320,
+ 506,
+ 333
+ ],
+ "score": 1.0,
+ "content": "On the quantitative side, we see that the BLEU score for the generated translations can be",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 123,
+ 331,
+ 423,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 331,
+ 423,
+ 344
+ ],
+ "score": 1.0,
+ "content": "significantly increased by showing it related translations before evaluating.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 359,
+ 224,
+ 372
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 357,
+ 226,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 357,
+ 226,
+ 374
+ ],
+ "score": 1.0,
+ "content": "2 MEMORY MODULE",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 383,
+ 505,
+ 428
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 244,
+ 396
+ ],
+ "score": 1.0,
+ "content": "Our memory consists of a matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 244,
+ 384,
+ 254,
+ 394
+ ],
+ "score": 0.82,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 383,
+ 362,
+ 396
+ ],
+ "score": 1.0,
+ "content": "of memory keys, a vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 363,
+ 384,
+ 372,
+ 394
+ ],
+ "score": 0.78,
+ "content": "V",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 372,
+ 383,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "of memory values, and an addi-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 159,
+ 408
+ ],
+ "score": 1.0,
+ "content": "tional vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 396,
+ 168,
+ 405
+ ],
+ "score": 0.79,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 394,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "that tracks the age of items stored in memory. Keys can be arbitrary vectors of size",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 107,
+ 406,
+ 505,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 407,
+ 155,
+ 417
+ ],
+ "score": 0.54,
+ "content": "{ \\mathrm { k e y } } - s { \\mathrm { i } } z { \\mathrm { e } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 406,
+ 505,
+ 418
+ ],
+ "score": 1.0,
+ "content": ", and we assume that the memory values are single integers representing a class or token",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 348,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 348,
+ 429
+ ],
+ "score": 1.0,
+ "content": "ID. We define a memory of size memory-size as a triple:",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 179,
+ 433,
+ 431,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 179,
+ 433,
+ 431,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 179,
+ 433,
+ 431,
+ 447
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { M } = ( K _ { \\mathrm { m e m o r y - s i z e } \\times \\mathrm { k e y - s i z e } } , ~ V _ { \\mathrm { m e m o r y - s i z e } } , ~ A _ { \\mathrm { m e m o r y - s i z e } } ) .",
+ "type": "interline_equation",
+ "image_path": "4d094333e0364c1022d04cb764b0856863c3efe2bc3daec41d71b71044517129.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 26,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 179,
+ 433,
+ 431,
+ 447
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 458,
+ 504,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 505,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 466,
+ 471
+ ],
+ "score": 1.0,
+ "content": "A memory query is a vector of size key-size which we assume to be normalized, i.e.,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 458,
+ 501,
+ 470
+ ],
+ "score": 0.93,
+ "content": "\\| q \\| = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 457,
+ 505,
+ 471
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 468,
+ 506,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 165,
+ 482
+ ],
+ "score": 1.0,
+ "content": "Given a query",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 471,
+ 171,
+ 480
+ ],
+ "score": 0.79,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 172,
+ 468,
+ 310,
+ 482
+ ],
+ "score": 1.0,
+ "content": ", we define the nearest neighbor of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 310,
+ 471,
+ 316,
+ 481
+ ],
+ "score": 0.82,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 468,
+ 328,
+ 482
+ ],
+ "score": 1.0,
+ "content": "in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 470,
+ 341,
+ 479
+ ],
+ "score": 0.83,
+ "content": "\\mathcal { M }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 341,
+ 468,
+ 506,
+ 482
+ ],
+ "score": 1.0,
+ "content": "as any of the keys that maximize the dot",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 170,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 159,
+ 493
+ ],
+ "score": 1.0,
+ "content": "product with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 482,
+ 166,
+ 492
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 480,
+ 170,
+ 493
+ ],
+ "score": 1.0,
+ "content": ":",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 240,
+ 490,
+ 371,
+ 504
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 240,
+ 490,
+ 371,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 240,
+ 490,
+ 371,
+ 504
+ ],
+ "score": 0.91,
+ "content": "\\operatorname { N N } ( q , \\mathcal { M } ) = \\operatorname { a r g m a x } _ { i } q \\cdot K [ i ] .",
+ "type": "interline_equation",
+ "image_path": "9d429d010560f5484bdf780ca6e9411504c7461104f9b6f874434c579802e267.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 30,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 240,
+ 490,
+ 371,
+ 504
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 506,
+ 505,
+ 540
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 506,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 506,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "Since the keys are normalized, the above notion corresponds to the nearest neighbor with respect",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 378,
+ 529
+ ],
+ "score": 1.0,
+ "content": "to cosine similarity. We will also use the natural extension of it to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 517,
+ 385,
+ 527
+ ],
+ "score": 0.83,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 386,
+ 516,
+ 506,
+ 529
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors, which we",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 482,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 135,
+ 541
+ ],
+ "score": 1.0,
+ "content": "denote",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 136,
+ 528,
+ 185,
+ 540
+ ],
+ "score": 0.92,
+ "content": "\\mathrm { N N } _ { k } ( q , \\mathcal { M } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 527,
+ 371,
+ 541
+ ],
+ "score": 1.0,
+ "content": ". In our experiments we always used the set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 528,
+ 406,
+ 538
+ ],
+ "score": 0.89,
+ "content": "k = 2 5 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 406,
+ 527,
+ 482,
+ 541
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 544,
+ 504,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 505,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 191,
+ 558
+ ],
+ "score": 1.0,
+ "content": "When given a query",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 547,
+ 198,
+ 556
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 198,
+ 543,
+ 254,
+ 558
+ ],
+ "score": 1.0,
+ "content": ", the memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 544,
+ 321,
+ 557
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { M } = ( K , V , A )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 543,
+ 379,
+ 558
+ ],
+ "score": 1.0,
+ "content": "will compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 545,
+ 386,
+ 555
+ ],
+ "score": 0.83,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 387,
+ 543,
+ 505,
+ 558
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors (sorted by",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 556,
+ 226,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 556,
+ 226,
+ 568
+ ],
+ "score": 1.0,
+ "content": "decreasing cosine similarity):",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 248,
+ 565,
+ 363,
+ 579
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 248,
+ 565,
+ 363,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 248,
+ 565,
+ 363,
+ 579
+ ],
+ "score": 0.92,
+ "content": "( n _ { 1 } , \\dots , n _ { k } ) = \\Nu \\Nu _ { k } ( q , { \\mathcal { M } } )",
+ "type": "interline_equation",
+ "image_path": "40ce1960e7101f9b3d4fbc83b8ab1967ce2f34b96b6a44956a361874e8718032.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 36,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 248,
+ 565,
+ 363,
+ 579
+ ],
+ "spans": [],
+ "index": 36
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 581,
+ 505,
+ 637
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 264,
+ 594
+ ],
+ "score": 1.0,
+ "content": "and return, as the main result, the value",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 581,
+ 289,
+ 593
+ ],
+ "score": 0.89,
+ "content": "V [ n _ { 1 } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 289,
+ 582,
+ 505,
+ 594
+ ],
+ "score": 1.0,
+ "content": ". Additionally, we will compute the cosine similarities",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 107,
+ 592,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 592,
+ 165,
+ 604
+ ],
+ "score": 0.93,
+ "content": "d _ { i } = \\boldsymbol { q } \\cdot \\boldsymbol { K } [ n _ { i } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 165,
+ 592,
+ 242,
+ 605
+ ],
+ "score": 1.0,
+ "content": "and return softmax",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 593,
+ 312,
+ 604
+ ],
+ "score": 0.72,
+ "content": "( d _ { 1 } \\cdot t , \\ldots , d _ { k } \\cdot t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 592,
+ 375,
+ 605
+ ],
+ "score": 1.0,
+ "content": ". The parameter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 594,
+ 381,
+ 602
+ ],
+ "score": 0.7,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 592,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "denotes the inverse of softmax",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 218,
+ 617
+ ],
+ "score": 1.0,
+ "content": "temperature and we set it to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 604,
+ 246,
+ 614
+ ],
+ "score": 0.9,
+ "content": "t = 4 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 603,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "in our experiments. In models where the memory output is again",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 627
+ ],
+ "score": 1.0,
+ "content": "embedded into a dense vector, we multiply the embedded output by the corresponding softmax",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 625,
+ 385,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 625,
+ 385,
+ 639
+ ],
+ "score": 1.0,
+ "content": "component so as to provide a signal about confidence of the memory.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 39
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 642,
+ 504,
+ 675
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 506,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 506,
+ 655
+ ],
+ "score": 1.0,
+ "content": "The forward computation of the memory module is thus very simple, the only interesting part being",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 506,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 506,
+ 666
+ ],
+ "score": 1.0,
+ "content": "how to compute nearest neighbors efficiently, which we discuss below. But we must also answer the",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 664,
+ 254,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 254,
+ 676
+ ],
+ "score": 1.0,
+ "content": "question how the memory is trained.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 344,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Memory Loss. Assume now that in addition to a query",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 690,
+ 351,
+ 699
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "we are also given the correct desired",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 183,
+ 711
+ ],
+ "score": 1.0,
+ "content": "(supervised) value",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 701,
+ 190,
+ 709
+ ],
+ "score": 0.72,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 699,
+ 330,
+ 711
+ ],
+ "score": 1.0,
+ "content": ". In the case of classification, this",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 701,
+ 336,
+ 709
+ ],
+ "score": 0.77,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 337,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "would be the class label. In a sequence-",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 179,
+ 723
+ ],
+ "score": 1.0,
+ "content": "to-sequence task,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 712,
+ 186,
+ 720
+ ],
+ "score": 0.65,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "would be the desired output token of the current time step. After computing",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 121,
+ 733
+ ],
+ "score": 1.0,
+ "content": "the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 121,
+ 721,
+ 128,
+ 730
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 129,
+ 720,
+ 203,
+ 733
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 203,
+ 721,
+ 255,
+ 732
+ ],
+ "score": 0.92,
+ "content": "( n _ { 1 } , \\ldots , n _ { k } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 720,
+ 308,
+ 733
+ ],
+ "score": 1.0,
+ "content": "as above, let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 722,
+ 315,
+ 732
+ ],
+ "score": 0.81,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 720,
+ 442,
+ 733
+ ],
+ "score": 1.0,
+ "content": "be the smallest index such that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 442,
+ 720,
+ 487,
+ 733
+ ],
+ "score": 0.94,
+ "content": "V [ n _ { p } ] = { \\bar { v } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 46.5
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "score": 1.0,
+ "content": "2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [],
+ "index": 0.5,
+ "bbox_fs": [
+ 106,
+ 82,
+ 505,
+ 107
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 504,
+ 166
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 109,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 109,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "There are many advantages of having a long-term memory. One-shot learning is a desirable property",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": "in its own right, and some tasks, as we will show below, are simply not solvable without it. Even",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "real-world tasks where we have large training sets, such as translation, can benefit from long-term",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "memory. Finally, since the memory can be traced back to training examples, it might help explain",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 460,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 460,
+ 167
+ ],
+ "score": 1.0,
+ "content": "the decisions that the model is making and thus improve understandability of the model.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 105,
+ 109,
+ 506,
+ 167
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 171,
+ 505,
+ 204
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 505,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 505,
+ 184
+ ],
+ "score": 1.0,
+ "content": "It is not immediately clear how to measure the performance of a life-long one-shot learning model,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 182,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 182,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": "since most deep learning evaluations focus on the average performance and do not have a one-shot",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 193,
+ 499,
+ 206
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 193,
+ 499,
+ 206
+ ],
+ "score": 1.0,
+ "content": "component. We therefore evaluate in a few ways, to show that our memory module indeed works:",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8,
+ "bbox_fs": [
+ 105,
+ 171,
+ 505,
+ 206
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 106,
+ 213,
+ 506,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 214,
+ 505,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 214,
+ 505,
+ 226
+ ],
+ "score": 1.0,
+ "content": "(1) We evaluate on the well-known one-shot learning task Omniglot, which is the only dataset with",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 122,
+ 223,
+ 506,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 223,
+ 506,
+ 239
+ ],
+ "score": 1.0,
+ "content": "explicit one-shot learning evaluation. This dataset is small and does not benefit from life-long",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 122,
+ 236,
+ 505,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 236,
+ 505,
+ 249
+ ],
+ "score": 1.0,
+ "content": "learning capability of our module, but we still exceed the best previous results and set new",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 123,
+ 247,
+ 187,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 247,
+ 187,
+ 259
+ ],
+ "score": 1.0,
+ "content": "state-of-the-art.",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 107,
+ 261,
+ 505,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 261,
+ 505,
+ 274
+ ],
+ "score": 1.0,
+ "content": "(2) We devise a synthetic task that requires life-long one-shot learning. On this task, standard",
+ "type": "text"
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 122,
+ 272,
+ 450,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 272,
+ 450,
+ 286
+ ],
+ "score": 1.0,
+ "content": "models fare poorly while our model can solve it well, demonstrating its strengths.",
+ "type": "text"
+ }
+ ],
+ "index": 15,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 107,
+ 286,
+ 505,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 286,
+ 505,
+ 301
+ ],
+ "score": 1.0,
+ "content": "(3) Finally, we train an English-German translation model that has our life-long one-shot learning",
+ "type": "text"
+ }
+ ],
+ "index": 16,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 121,
+ 297,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 297,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "module. It retains very good performance on average and is also capable of one-shot learning.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 122,
+ 309,
+ 505,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 309,
+ 505,
+ 323
+ ],
+ "score": 1.0,
+ "content": "On the qualitative side, we find that it can translate rarely-occurring words like Dostoevsky.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 122,
+ 320,
+ 506,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 320,
+ 506,
+ 333
+ ],
+ "score": 1.0,
+ "content": "On the quantitative side, we see that the BLEU score for the generated translations can be",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 123,
+ 331,
+ 423,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 331,
+ 423,
+ 344
+ ],
+ "score": 1.0,
+ "content": "significantly increased by showing it related translations before evaluating.",
+ "type": "text"
+ }
+ ],
+ "index": 20,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 107,
+ 214,
+ 506,
+ 344
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 359,
+ 224,
+ 372
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 357,
+ 226,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 357,
+ 226,
+ 374
+ ],
+ "score": 1.0,
+ "content": "2 MEMORY MODULE",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 383,
+ 505,
+ 428
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 244,
+ 396
+ ],
+ "score": 1.0,
+ "content": "Our memory consists of a matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 244,
+ 384,
+ 254,
+ 394
+ ],
+ "score": 0.82,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 383,
+ 362,
+ 396
+ ],
+ "score": 1.0,
+ "content": "of memory keys, a vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 363,
+ 384,
+ 372,
+ 394
+ ],
+ "score": 0.78,
+ "content": "V",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 372,
+ 383,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "of memory values, and an addi-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 159,
+ 408
+ ],
+ "score": 1.0,
+ "content": "tional vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 396,
+ 168,
+ 405
+ ],
+ "score": 0.79,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 394,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "that tracks the age of items stored in memory. Keys can be arbitrary vectors of size",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 107,
+ 406,
+ 505,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 407,
+ 155,
+ 417
+ ],
+ "score": 0.54,
+ "content": "{ \\mathrm { k e y } } - s { \\mathrm { i } } z { \\mathrm { e } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 406,
+ 505,
+ 418
+ ],
+ "score": 1.0,
+ "content": ", and we assume that the memory values are single integers representing a class or token",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 348,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 348,
+ 429
+ ],
+ "score": 1.0,
+ "content": "ID. We define a memory of size memory-size as a triple:",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 105,
+ 383,
+ 505,
+ 429
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 179,
+ 433,
+ 431,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 179,
+ 433,
+ 431,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 179,
+ 433,
+ 431,
+ 447
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { M } = ( K _ { \\mathrm { m e m o r y - s i z e } \\times \\mathrm { k e y - s i z e } } , ~ V _ { \\mathrm { m e m o r y - s i z e } } , ~ A _ { \\mathrm { m e m o r y - s i z e } } ) .",
+ "type": "interline_equation",
+ "image_path": "4d094333e0364c1022d04cb764b0856863c3efe2bc3daec41d71b71044517129.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 26,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 179,
+ 433,
+ 431,
+ 447
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 458,
+ 504,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 505,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 466,
+ 471
+ ],
+ "score": 1.0,
+ "content": "A memory query is a vector of size key-size which we assume to be normalized, i.e.,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 458,
+ 501,
+ 470
+ ],
+ "score": 0.93,
+ "content": "\\| q \\| = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 457,
+ 505,
+ 471
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 468,
+ 506,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 165,
+ 482
+ ],
+ "score": 1.0,
+ "content": "Given a query",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 471,
+ 171,
+ 480
+ ],
+ "score": 0.79,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 172,
+ 468,
+ 310,
+ 482
+ ],
+ "score": 1.0,
+ "content": ", we define the nearest neighbor of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 310,
+ 471,
+ 316,
+ 481
+ ],
+ "score": 0.82,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 468,
+ 328,
+ 482
+ ],
+ "score": 1.0,
+ "content": "in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 470,
+ 341,
+ 479
+ ],
+ "score": 0.83,
+ "content": "\\mathcal { M }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 341,
+ 468,
+ 506,
+ 482
+ ],
+ "score": 1.0,
+ "content": "as any of the keys that maximize the dot",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 170,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 159,
+ 493
+ ],
+ "score": 1.0,
+ "content": "product with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 482,
+ 166,
+ 492
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 480,
+ 170,
+ 493
+ ],
+ "score": 1.0,
+ "content": ":",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28,
+ "bbox_fs": [
+ 105,
+ 457,
+ 506,
+ 493
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 240,
+ 490,
+ 371,
+ 504
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 240,
+ 490,
+ 371,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 240,
+ 490,
+ 371,
+ 504
+ ],
+ "score": 0.91,
+ "content": "\\operatorname { N N } ( q , \\mathcal { M } ) = \\operatorname { a r g m a x } _ { i } q \\cdot K [ i ] .",
+ "type": "interline_equation",
+ "image_path": "9d429d010560f5484bdf780ca6e9411504c7461104f9b6f874434c579802e267.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 30,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 240,
+ 490,
+ 371,
+ 504
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 506,
+ 505,
+ 540
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 506,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 506,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "Since the keys are normalized, the above notion corresponds to the nearest neighbor with respect",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 378,
+ 529
+ ],
+ "score": 1.0,
+ "content": "to cosine similarity. We will also use the natural extension of it to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 517,
+ 385,
+ 527
+ ],
+ "score": 0.83,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 386,
+ 516,
+ 506,
+ 529
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors, which we",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 482,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 135,
+ 541
+ ],
+ "score": 1.0,
+ "content": "denote",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 136,
+ 528,
+ 185,
+ 540
+ ],
+ "score": 0.92,
+ "content": "\\mathrm { N N } _ { k } ( q , \\mathcal { M } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 527,
+ 371,
+ 541
+ ],
+ "score": 1.0,
+ "content": ". In our experiments we always used the set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 528,
+ 406,
+ 538
+ ],
+ "score": 0.89,
+ "content": "k = 2 5 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 406,
+ 527,
+ 482,
+ 541
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32,
+ "bbox_fs": [
+ 105,
+ 506,
+ 506,
+ 541
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 544,
+ 504,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 505,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 191,
+ 558
+ ],
+ "score": 1.0,
+ "content": "When given a query",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 547,
+ 198,
+ 556
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 198,
+ 543,
+ 254,
+ 558
+ ],
+ "score": 1.0,
+ "content": ", the memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 544,
+ 321,
+ 557
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { M } = ( K , V , A )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 543,
+ 379,
+ 558
+ ],
+ "score": 1.0,
+ "content": "will compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 545,
+ 386,
+ 555
+ ],
+ "score": 0.83,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 387,
+ 543,
+ 505,
+ 558
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors (sorted by",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 556,
+ 226,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 556,
+ 226,
+ 568
+ ],
+ "score": 1.0,
+ "content": "decreasing cosine similarity):",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34.5,
+ "bbox_fs": [
+ 105,
+ 543,
+ 505,
+ 568
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 248,
+ 565,
+ 363,
+ 579
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 248,
+ 565,
+ 363,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 248,
+ 565,
+ 363,
+ 579
+ ],
+ "score": 0.92,
+ "content": "( n _ { 1 } , \\dots , n _ { k } ) = \\Nu \\Nu _ { k } ( q , { \\mathcal { M } } )",
+ "type": "interline_equation",
+ "image_path": "40ce1960e7101f9b3d4fbc83b8ab1967ce2f34b96b6a44956a361874e8718032.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 36,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 248,
+ 565,
+ 363,
+ 579
+ ],
+ "spans": [],
+ "index": 36
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 581,
+ 505,
+ 637
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 264,
+ 594
+ ],
+ "score": 1.0,
+ "content": "and return, as the main result, the value",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 581,
+ 289,
+ 593
+ ],
+ "score": 0.89,
+ "content": "V [ n _ { 1 } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 289,
+ 582,
+ 505,
+ 594
+ ],
+ "score": 1.0,
+ "content": ". Additionally, we will compute the cosine similarities",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 107,
+ 592,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 592,
+ 165,
+ 604
+ ],
+ "score": 0.93,
+ "content": "d _ { i } = \\boldsymbol { q } \\cdot \\boldsymbol { K } [ n _ { i } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 165,
+ 592,
+ 242,
+ 605
+ ],
+ "score": 1.0,
+ "content": "and return softmax",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 593,
+ 312,
+ 604
+ ],
+ "score": 0.72,
+ "content": "( d _ { 1 } \\cdot t , \\ldots , d _ { k } \\cdot t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 592,
+ 375,
+ 605
+ ],
+ "score": 1.0,
+ "content": ". The parameter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 594,
+ 381,
+ 602
+ ],
+ "score": 0.7,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 592,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "denotes the inverse of softmax",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 218,
+ 617
+ ],
+ "score": 1.0,
+ "content": "temperature and we set it to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 604,
+ 246,
+ 614
+ ],
+ "score": 0.9,
+ "content": "t = 4 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 603,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "in our experiments. In models where the memory output is again",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 627
+ ],
+ "score": 1.0,
+ "content": "embedded into a dense vector, we multiply the embedded output by the corresponding softmax",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 625,
+ 385,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 625,
+ 385,
+ 639
+ ],
+ "score": 1.0,
+ "content": "component so as to provide a signal about confidence of the memory.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 39,
+ "bbox_fs": [
+ 105,
+ 581,
+ 505,
+ 639
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 642,
+ 504,
+ 675
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 506,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 506,
+ 655
+ ],
+ "score": 1.0,
+ "content": "The forward computation of the memory module is thus very simple, the only interesting part being",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 506,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 506,
+ 666
+ ],
+ "score": 1.0,
+ "content": "how to compute nearest neighbors efficiently, which we discuss below. But we must also answer the",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 664,
+ 254,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 254,
+ 676
+ ],
+ "score": 1.0,
+ "content": "question how the memory is trained.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43,
+ "bbox_fs": [
+ 105,
+ 641,
+ 506,
+ 676
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 344,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Memory Loss. Assume now that in addition to a query",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 690,
+ 351,
+ 699
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "we are also given the correct desired",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 183,
+ 711
+ ],
+ "score": 1.0,
+ "content": "(supervised) value",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 701,
+ 190,
+ 709
+ ],
+ "score": 0.72,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 699,
+ 330,
+ 711
+ ],
+ "score": 1.0,
+ "content": ". In the case of classification, this",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 701,
+ 336,
+ 709
+ ],
+ "score": 0.77,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 337,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "would be the class label. In a sequence-",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 179,
+ 723
+ ],
+ "score": 1.0,
+ "content": "to-sequence task,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 712,
+ 186,
+ 720
+ ],
+ "score": 0.65,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "would be the desired output token of the current time step. After computing",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 121,
+ 733
+ ],
+ "score": 1.0,
+ "content": "the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 121,
+ 721,
+ 128,
+ 730
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 129,
+ 720,
+ 203,
+ 733
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 203,
+ 721,
+ 255,
+ 732
+ ],
+ "score": 0.92,
+ "content": "( n _ { 1 } , \\ldots , n _ { k } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 720,
+ 308,
+ 733
+ ],
+ "score": 1.0,
+ "content": "as above, let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 722,
+ 315,
+ 732
+ ],
+ "score": 0.81,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 720,
+ 442,
+ 733
+ ],
+ "score": 1.0,
+ "content": "be the smallest index such that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 442,
+ 720,
+ 487,
+ 733
+ ],
+ "score": 0.94,
+ "content": "V [ n _ { p } ] = { \\bar { v } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 46.5,
+ "bbox_fs": [
+ 105,
+ 687,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 83,
+ 302,
+ 112
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 81,
+ 302,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 81,
+ 131,
+ 97
+ ],
+ "score": 1.0,
+ "content": "Case",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 83,
+ 302,
+ 96
+ ],
+ "score": 0.72,
+ "content": "1 \\colon V [ n _ { 1 } ] = v ; \\quad \\operatorname { L o s s } = [ q \\cdot k _ { b } - q \\cdot k _ { 1 } + \\alpha ] _ { + }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 118,
+ 96,
+ 269,
+ 112
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 96,
+ 150,
+ 112
+ ],
+ "score": 1.0,
+ "content": "Update:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 97,
+ 269,
+ 112
+ ],
+ "score": 0.63,
+ "content": "\\begin{array} { r l } { K [ n _ { 1 } ] \\gets \\frac { q + k _ { 1 } } { \\| q + k _ { 1 } \\| } } & { { } A [ n _ { 1 } ] \\gets 0 } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 318,
+ 83,
+ 514,
+ 111
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 318,
+ 83,
+ 514,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 318,
+ 83,
+ 514,
+ 96
+ ],
+ "score": 0.52,
+ "content": "{ \\mathrm { C a s e ~ } } 2 \\colon V [ n _ { 1 } ] \\neq v ; \\quad { \\mathrm { L o s s } } = [ q \\cdot k _ { 1 } - q \\cdot k _ { p } + \\alpha ] _ { + }",
+ "type": "inline_equation",
+ "image_path": "547fd45890bcca9766874b0ee533f043055fea5cfb54837b05da3ea0d3e33b0c.jpg"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 331,
+ 98,
+ 506,
+ 111
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 331,
+ 98,
+ 362,
+ 111
+ ],
+ "score": 1.0,
+ "content": "Update:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 362,
+ 98,
+ 405,
+ 110
+ ],
+ "score": 0.39,
+ "content": "K [ n ^ { \\prime } ] \\gets q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 384,
+ 99,
+ 506,
+ 110
+ ],
+ "score": 0.31,
+ "content": "\\mid q \\quad V [ n ^ { \\prime } ] v \\quad A [ n ^ { \\prime } ] 0",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 111,
+ 125,
+ 284,
+ 237
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 111,
+ 125,
+ 284,
+ 237
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 125,
+ 284,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 125,
+ 284,
+ 237
+ ],
+ "score": 0.914,
+ "type": "image",
+ "image_path": "9a0e9979f224838b6739ede56f86aa3a60d5f2bde94e10bf2bf90a017be31af0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 111,
+ 125,
+ 284,
+ 139.0
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 111,
+ 139.0,
+ 284,
+ 153.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 111,
+ 153.0,
+ 284,
+ 167.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 111,
+ 167.0,
+ 284,
+ 181.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 111,
+ 181.0,
+ 284,
+ 195.0
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 111,
+ 195.0,
+ 284,
+ 209.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 111,
+ 209.0,
+ 284,
+ 223.0
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 111,
+ 223.0,
+ 284,
+ 237.0
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 255,
+ 503,
+ 267
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 338,
+ 268
+ ],
+ "score": 1.0,
+ "content": "Figure 1: The operation of the memory module on a query",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 257,
+ 344,
+ 267
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 253,
+ 418,
+ 268
+ ],
+ "score": 1.0,
+ "content": "with correct value",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 257,
+ 425,
+ 265
+ ],
+ "score": 0.71,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 425,
+ 253,
+ 505,
+ 268
+ ],
+ "score": 1.0,
+ "content": "; see text for details.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 13.75
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 322,
+ 125,
+ 496,
+ 237
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 322,
+ 125,
+ 496,
+ 237
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 322,
+ 125,
+ 496,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 125,
+ 496,
+ 237
+ ],
+ "score": 0.877,
+ "type": "image",
+ "image_path": "439659069ba27ddf6723d23d463db67cb813a1033d57af231b411cfaded4e734.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 322,
+ 125,
+ 496,
+ 139.0
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 322,
+ 139.0,
+ 496,
+ 153.0
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 322,
+ 153.0,
+ 496,
+ 167.0
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 322,
+ 167.0,
+ 496,
+ 181.0
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 322,
+ 181.0,
+ 496,
+ 195.0
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 322,
+ 195.0,
+ 496,
+ 209.0
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 322,
+ 209.0,
+ 496,
+ 223.0
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 322,
+ 223.0,
+ 496,
+ 237.0
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 320
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 288,
+ 113,
+ 297
+ ],
+ "score": 0.55,
+ "content": "b",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 113,
+ 286,
+ 231,
+ 300
+ ],
+ "score": 1.0,
+ "content": "the smallest index such that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 287,
+ 278,
+ 299
+ ],
+ "score": 0.93,
+ "content": "V [ n _ { b } ] \\ne v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 278,
+ 286,
+ 318,
+ 300
+ ],
+ "score": 1.0,
+ "content": ". We call",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 319,
+ 289,
+ 331,
+ 299
+ ],
+ "score": 0.87,
+ "content": "n _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 286,
+ 440,
+ 300
+ ],
+ "score": 1.0,
+ "content": "the positive neighbor and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 440,
+ 289,
+ 451,
+ 298
+ ],
+ "score": 0.85,
+ "content": "n _ { b }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 286,
+ 505,
+ 300
+ ],
+ "score": 1.0,
+ "content": "the negative",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 334,
+ 311
+ ],
+ "score": 1.0,
+ "content": "neighbor. When no positive neighbor is among the top-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 335,
+ 299,
+ 341,
+ 308
+ ],
+ "score": 0.81,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 341,
+ 298,
+ 505,
+ 311
+ ],
+ "score": 1.0,
+ "content": ", we pick any vector from memory with",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 309,
+ 335,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 130,
+ 321
+ ],
+ "score": 1.0,
+ "content": "value",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 131,
+ 311,
+ 137,
+ 319
+ ],
+ "score": 0.75,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 137,
+ 309,
+ 180,
+ 321
+ ],
+ "score": 1.0,
+ "content": "instead of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 309,
+ 206,
+ 321
+ ],
+ "score": 0.93,
+ "content": "K [ n _ { p } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 206,
+ 309,
+ 335,
+ 321
+ ],
+ "score": 1.0,
+ "content": ". We define the memory loss as:",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 209,
+ 325,
+ 401,
+ 341
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 209,
+ 325,
+ 401,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 209,
+ 325,
+ 401,
+ 341
+ ],
+ "score": 0.91,
+ "content": "\\mathrm { l o s s } ( q , v , { \\cal M } ) = \\left[ q \\cdot K [ n _ { b } ] - q \\cdot K [ n _ { p } ] + \\alpha \\right] _ { + } .",
+ "type": "interline_equation",
+ "image_path": "5e66fbc0762a10a6278bd0d001ce44f3d279547f91ceee348b92a7a76deeec8b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 24,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 209,
+ 325,
+ 401,
+ 341
+ ],
+ "spans": [],
+ "index": 24
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 429
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 173,
+ 364
+ ],
+ "score": 1.0,
+ "content": "Recall that both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 354,
+ 180,
+ 363
+ ],
+ "score": 0.82,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 351,
+ 505,
+ 364
+ ],
+ "score": 1.0,
+ "content": "and the keys in memory are normalized, so the products in the above loss term",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 278,
+ 375
+ ],
+ "score": 1.0,
+ "content": "correspond to cosine similarities between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 279,
+ 365,
+ 285,
+ 374
+ ],
+ "score": 0.76,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 363,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": ", the positive key, and the negative key. Since cosine",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 386
+ ],
+ "score": 1.0,
+ "content": "similarity is maximal for equal terms, we want to maximize the similarity to the positive key and",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 493,
+ 398
+ ],
+ "score": 1.0,
+ "content": "minimize the similarity to the negative one. But once they are far enough apart (by the margin",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 494,
+ 388,
+ 501,
+ 394
+ ],
+ "score": 0.72,
+ "content": "\\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 384,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "0.1 in all our experiments), we do not propagate any loss. This definition and reasoning behind it are",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "almost identical to the one in Schroff et al. (2015) and similar to many other distance metric learning",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 417,
+ 328,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 328,
+ 430
+ ],
+ "score": 1.0,
+ "content": "works (Weinberger & Saul, 2009; Weston et al., 2011).",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 441,
+ 505,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 448,
+ 453
+ ],
+ "score": 1.0,
+ "content": "Memory Update. In addition to computing the loss, we will also update the memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 448,
+ 442,
+ 461,
+ 451
+ ],
+ "score": 0.82,
+ "content": "\\mathcal { M }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 461,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "to account",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 276,
+ 465
+ ],
+ "score": 1.0,
+ "content": "for the fact that the newly presented query",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 276,
+ 454,
+ 283,
+ 464
+ ],
+ "score": 0.76,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 283,
+ 451,
+ 344,
+ 465
+ ],
+ "score": 1.0,
+ "content": "corresponds to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 454,
+ 351,
+ 462
+ ],
+ "score": 0.68,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 451,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": ". The update is done in a different way",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 464,
+ 504,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 464,
+ 497,
+ 475
+ ],
+ "score": 1.0,
+ "content": "depending on whether the main value returned by the memory module already is the correct value",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 497,
+ 465,
+ 504,
+ 473
+ ],
+ "score": 0.71,
+ "content": "v",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 473,
+ 377,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 192,
+ 488
+ ],
+ "score": 1.0,
+ "content": "or not. As before, let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 474,
+ 260,
+ 486
+ ],
+ "score": 0.93,
+ "content": "n _ { 1 } = \\mathrm { N N } ( q , { \\mathcal { M } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 473,
+ 367,
+ 488
+ ],
+ "score": 1.0,
+ "content": "be the nearest neighbor to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 367,
+ 476,
+ 373,
+ 486
+ ],
+ "score": 0.76,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 373,
+ 473,
+ 377,
+ 488
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 33.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 490,
+ 504,
+ 514
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 327,
+ 504
+ ],
+ "score": 1.0,
+ "content": "If the memory already returns the correct value, i.e., if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 490,
+ 371,
+ 503
+ ],
+ "score": 0.93,
+ "content": "V [ n _ { 1 } ] = v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 372,
+ 489,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": ", then we only update the key for",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 107,
+ 502,
+ 383,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 503,
+ 118,
+ 513
+ ],
+ "score": 0.83,
+ "content": "n _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 118,
+ 502,
+ 297,
+ 515
+ ],
+ "score": 1.0,
+ "content": "by taking the average of the current key and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 504,
+ 303,
+ 514
+ ],
+ "score": 0.81,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 502,
+ 383,
+ 515
+ ],
+ "score": 1.0,
+ "content": "and normalizing it:",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 255,
+ 518,
+ 356,
+ 546
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 255,
+ 518,
+ 356,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 255,
+ 518,
+ 356,
+ 546
+ ],
+ "score": 0.94,
+ "content": "K [ n _ { 1 } ] \\gets \\frac { q + K [ n _ { 1 } ] } { \\lVert q + K [ n _ { 1 } ] \\rVert } .",
+ "type": "interline_equation",
+ "image_path": "3fa7aa1f9cda9c5a5982ec2c4a5782e58eabac60bb96dea28f0da76a72647b1c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 38.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 255,
+ 518,
+ 356,
+ 532.0
+ ],
+ "spans": [],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 255,
+ 532.0,
+ 356,
+ 546.0
+ ],
+ "spans": [],
+ "index": 39
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 550,
+ 317,
+ 563
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 318,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 269,
+ 564
+ ],
+ "score": 1.0,
+ "content": "When doing this, we also re-set the age:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 550,
+ 314,
+ 563
+ ],
+ "score": 0.93,
+ "content": "A [ n _ { 1 } ] 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 315,
+ 549,
+ 318,
+ 564
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 567,
+ 505,
+ 623
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 506,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 178,
+ 581
+ ],
+ "score": 1.0,
+ "content": "Otherwise, when",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 567,
+ 225,
+ 580
+ ],
+ "score": 0.93,
+ "content": "V [ n _ { 1 } ] \\neq v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 226,
+ 566,
+ 455,
+ 581
+ ],
+ "score": 1.0,
+ "content": ", we find a new place in the memory and write the pair",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 568,
+ 478,
+ 580
+ ],
+ "score": 0.92,
+ "content": "( q , v )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 566,
+ 506,
+ 581
+ ],
+ "score": 1.0,
+ "content": "there.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 591
+ ],
+ "score": 1.0,
+ "content": "Which place should we choose? We find memory items with maximum age, and write to one of",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 589,
+ 505,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 590,
+ 314,
+ 602
+ ],
+ "score": 1.0,
+ "content": "those (randomly chosen). More formally, we pick",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 314,
+ 589,
+ 414,
+ 601
+ ],
+ "score": 0.91,
+ "content": "n ^ { \\prime } = \\mathrm { a r g m a x } _ { i } A [ i ] + r _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 414,
+ 590,
+ 443,
+ 602
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 589,
+ 494,
+ 601
+ ],
+ "score": 0.92,
+ "content": "| r _ { i } | \\ll | \\mathcal { M } |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 494,
+ 590,
+ 505,
+ 602
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 613
+ ],
+ "score": 1.0,
+ "content": "a random number that introduces some randomness in the choice so as to avoid race conditions in",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 611,
+ 306,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 611,
+ 306,
+ 624
+ ],
+ "score": 1.0,
+ "content": "asynchronous multi-replica training. We then set:",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 43
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 223,
+ 627,
+ 387,
+ 642
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 223,
+ 627,
+ 387,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 223,
+ 627,
+ 387,
+ 642
+ ],
+ "score": 0.92,
+ "content": "K [ n ^ { \\prime } ] q , \\quad V [ n ^ { \\prime } ] v , \\quad A [ n ^ { \\prime } ] 0 .",
+ "type": "interline_equation",
+ "image_path": "9a5e7f44b6243f2c40fd82bf4fa870633978ae523232ca81e6765dd8dbd03591.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 46,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 223,
+ 627,
+ 387,
+ 642
+ ],
+ "spans": [],
+ "index": 46
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 646,
+ 505,
+ 670
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 505,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 505,
+ 660
+ ],
+ "score": 1.0,
+ "content": "With every memory update we also increment the age of all non-updated indices by 1. The full",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 335,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 335,
+ 671
+ ],
+ "score": 1.0,
+ "content": "operation of the memory module is depicted in Figure 1.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 47.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 681,
+ 503,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 694
+ ],
+ "score": 1.0,
+ "content": "Efficient nearest neighbor computation. The most expensive operation in our memory module",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 692,
+ 484,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 692,
+ 193,
+ 706
+ ],
+ "score": 1.0,
+ "content": "is the computation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 693,
+ 200,
+ 703
+ ],
+ "score": 0.81,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 692,
+ 484,
+ 706
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors. This can be done exactly or in an approximate way.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 49.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 709,
+ 503,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 350,
+ 723
+ ],
+ "score": 1.0,
+ "content": "In the exact mode, to calculate the nearest neighbors in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 351,
+ 710,
+ 361,
+ 720
+ ],
+ "score": 0.84,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 361,
+ 708,
+ 479,
+ 723
+ ],
+ "score": 1.0,
+ "content": "to a mini-batch of queries",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 480,
+ 709,
+ 504,
+ 721
+ ],
+ "score": 0.88,
+ "content": "Q \\ =",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 107,
+ 719,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 721,
+ 154,
+ 733
+ ],
+ "score": 0.91,
+ "content": "( q _ { 1 } , \\dots , q _ { b } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 719,
+ 329,
+ 734
+ ],
+ "score": 1.0,
+ "content": ", we perform a single matrix multiplication:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 721,
+ 362,
+ 732
+ ],
+ "score": 0.92,
+ "content": "Q \\times K ^ { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 362,
+ 719,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": ". This multiplies the batch-size",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 51.5
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 83,
+ 302,
+ 112
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 81,
+ 302,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 81,
+ 131,
+ 97
+ ],
+ "score": 1.0,
+ "content": "Case",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 83,
+ 302,
+ 96
+ ],
+ "score": 0.72,
+ "content": "1 \\colon V [ n _ { 1 } ] = v ; \\quad \\operatorname { L o s s } = [ q \\cdot k _ { b } - q \\cdot k _ { 1 } + \\alpha ] _ { + }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 118,
+ 96,
+ 269,
+ 112
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 96,
+ 150,
+ 112
+ ],
+ "score": 1.0,
+ "content": "Update:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 97,
+ 269,
+ 112
+ ],
+ "score": 0.63,
+ "content": "\\begin{array} { r l } { K [ n _ { 1 } ] \\gets \\frac { q + k _ { 1 } } { \\| q + k _ { 1 } \\| } } & { { } A [ n _ { 1 } ] \\gets 0 } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 108,
+ 81,
+ 302,
+ 112
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 318,
+ 83,
+ 514,
+ 111
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 318,
+ 83,
+ 514,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 318,
+ 83,
+ 514,
+ 96
+ ],
+ "score": 0.52,
+ "content": "{ \\mathrm { C a s e ~ } } 2 \\colon V [ n _ { 1 } ] \\neq v ; \\quad { \\mathrm { L o s s } } = [ q \\cdot k _ { 1 } - q \\cdot k _ { p } + \\alpha ] _ { + }",
+ "type": "inline_equation",
+ "image_path": "547fd45890bcca9766874b0ee533f043055fea5cfb54837b05da3ea0d3e33b0c.jpg"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 331,
+ 98,
+ 506,
+ 111
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 331,
+ 98,
+ 362,
+ 111
+ ],
+ "score": 1.0,
+ "content": "Update:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 362,
+ 98,
+ 405,
+ 110
+ ],
+ "score": 0.39,
+ "content": "K [ n ^ { \\prime } ] \\gets q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 384,
+ 99,
+ 506,
+ 110
+ ],
+ "score": 0.31,
+ "content": "\\mid q \\quad V [ n ^ { \\prime } ] v \\quad A [ n ^ { \\prime } ] 0",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 318,
+ 83,
+ 514,
+ 111
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 111,
+ 125,
+ 284,
+ 237
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 111,
+ 125,
+ 284,
+ 237
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 125,
+ 284,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 125,
+ 284,
+ 237
+ ],
+ "score": 0.914,
+ "type": "image",
+ "image_path": "9a0e9979f224838b6739ede56f86aa3a60d5f2bde94e10bf2bf90a017be31af0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 111,
+ 125,
+ 284,
+ 139.0
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 111,
+ 139.0,
+ 284,
+ 153.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 111,
+ 153.0,
+ 284,
+ 167.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 111,
+ 167.0,
+ 284,
+ 181.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 111,
+ 181.0,
+ 284,
+ 195.0
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 111,
+ 195.0,
+ 284,
+ 209.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 111,
+ 209.0,
+ 284,
+ 223.0
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 111,
+ 223.0,
+ 284,
+ 237.0
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 255,
+ 503,
+ 267
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 338,
+ 268
+ ],
+ "score": 1.0,
+ "content": "Figure 1: The operation of the memory module on a query",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 257,
+ 344,
+ 267
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 253,
+ 418,
+ 268
+ ],
+ "score": 1.0,
+ "content": "with correct value",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 257,
+ 425,
+ 265
+ ],
+ "score": 0.71,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 425,
+ 253,
+ 505,
+ 268
+ ],
+ "score": 1.0,
+ "content": "; see text for details.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 13.75
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 322,
+ 125,
+ 496,
+ 237
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 322,
+ 125,
+ 496,
+ 237
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 322,
+ 125,
+ 496,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 125,
+ 496,
+ 237
+ ],
+ "score": 0.877,
+ "type": "image",
+ "image_path": "439659069ba27ddf6723d23d463db67cb813a1033d57af231b411cfaded4e734.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 322,
+ 125,
+ 496,
+ 139.0
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 322,
+ 139.0,
+ 496,
+ 153.0
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 322,
+ 153.0,
+ 496,
+ 167.0
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 322,
+ 167.0,
+ 496,
+ 181.0
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 322,
+ 181.0,
+ 496,
+ 195.0
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 322,
+ 195.0,
+ 496,
+ 209.0
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 322,
+ 209.0,
+ 496,
+ 223.0
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 322,
+ 223.0,
+ 496,
+ 237.0
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 320
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 288,
+ 113,
+ 297
+ ],
+ "score": 0.55,
+ "content": "b",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 113,
+ 286,
+ 231,
+ 300
+ ],
+ "score": 1.0,
+ "content": "the smallest index such that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 287,
+ 278,
+ 299
+ ],
+ "score": 0.93,
+ "content": "V [ n _ { b } ] \\ne v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 278,
+ 286,
+ 318,
+ 300
+ ],
+ "score": 1.0,
+ "content": ". We call",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 319,
+ 289,
+ 331,
+ 299
+ ],
+ "score": 0.87,
+ "content": "n _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 286,
+ 440,
+ 300
+ ],
+ "score": 1.0,
+ "content": "the positive neighbor and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 440,
+ 289,
+ 451,
+ 298
+ ],
+ "score": 0.85,
+ "content": "n _ { b }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 286,
+ 505,
+ 300
+ ],
+ "score": 1.0,
+ "content": "the negative",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 334,
+ 311
+ ],
+ "score": 1.0,
+ "content": "neighbor. When no positive neighbor is among the top-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 335,
+ 299,
+ 341,
+ 308
+ ],
+ "score": 0.81,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 341,
+ 298,
+ 505,
+ 311
+ ],
+ "score": 1.0,
+ "content": ", we pick any vector from memory with",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 309,
+ 335,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 130,
+ 321
+ ],
+ "score": 1.0,
+ "content": "value",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 131,
+ 311,
+ 137,
+ 319
+ ],
+ "score": 0.75,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 137,
+ 309,
+ 180,
+ 321
+ ],
+ "score": 1.0,
+ "content": "instead of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 309,
+ 206,
+ 321
+ ],
+ "score": 0.93,
+ "content": "K [ n _ { p } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 206,
+ 309,
+ 335,
+ 321
+ ],
+ "score": 1.0,
+ "content": ". We define the memory loss as:",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 286,
+ 505,
+ 321
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 209,
+ 325,
+ 401,
+ 341
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 209,
+ 325,
+ 401,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 209,
+ 325,
+ 401,
+ 341
+ ],
+ "score": 0.91,
+ "content": "\\mathrm { l o s s } ( q , v , { \\cal M } ) = \\left[ q \\cdot K [ n _ { b } ] - q \\cdot K [ n _ { p } ] + \\alpha \\right] _ { + } .",
+ "type": "interline_equation",
+ "image_path": "5e66fbc0762a10a6278bd0d001ce44f3d279547f91ceee348b92a7a76deeec8b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 24,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 209,
+ 325,
+ 401,
+ 341
+ ],
+ "spans": [],
+ "index": 24
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 429
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 173,
+ 364
+ ],
+ "score": 1.0,
+ "content": "Recall that both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 354,
+ 180,
+ 363
+ ],
+ "score": 0.82,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 351,
+ 505,
+ 364
+ ],
+ "score": 1.0,
+ "content": "and the keys in memory are normalized, so the products in the above loss term",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 278,
+ 375
+ ],
+ "score": 1.0,
+ "content": "correspond to cosine similarities between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 279,
+ 365,
+ 285,
+ 374
+ ],
+ "score": 0.76,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 363,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": ", the positive key, and the negative key. Since cosine",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 386
+ ],
+ "score": 1.0,
+ "content": "similarity is maximal for equal terms, we want to maximize the similarity to the positive key and",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 493,
+ 398
+ ],
+ "score": 1.0,
+ "content": "minimize the similarity to the negative one. But once they are far enough apart (by the margin",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 494,
+ 388,
+ 501,
+ 394
+ ],
+ "score": 0.72,
+ "content": "\\alpha",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 384,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "0.1 in all our experiments), we do not propagate any loss. This definition and reasoning behind it are",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "almost identical to the one in Schroff et al. (2015) and similar to many other distance metric learning",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 417,
+ 328,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 328,
+ 430
+ ],
+ "score": 1.0,
+ "content": "works (Weinberger & Saul, 2009; Weston et al., 2011).",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28,
+ "bbox_fs": [
+ 105,
+ 351,
+ 505,
+ 430
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 441,
+ 505,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 448,
+ 453
+ ],
+ "score": 1.0,
+ "content": "Memory Update. In addition to computing the loss, we will also update the memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 448,
+ 442,
+ 461,
+ 451
+ ],
+ "score": 0.82,
+ "content": "\\mathcal { M }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 461,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "to account",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 276,
+ 465
+ ],
+ "score": 1.0,
+ "content": "for the fact that the newly presented query",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 276,
+ 454,
+ 283,
+ 464
+ ],
+ "score": 0.76,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 283,
+ 451,
+ 344,
+ 465
+ ],
+ "score": 1.0,
+ "content": "corresponds to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 454,
+ 351,
+ 462
+ ],
+ "score": 0.68,
+ "content": "v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 451,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": ". The update is done in a different way",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 464,
+ 504,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 464,
+ 497,
+ 475
+ ],
+ "score": 1.0,
+ "content": "depending on whether the main value returned by the memory module already is the correct value",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 497,
+ 465,
+ 504,
+ 473
+ ],
+ "score": 0.71,
+ "content": "v",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 473,
+ 377,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 192,
+ 488
+ ],
+ "score": 1.0,
+ "content": "or not. As before, let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 474,
+ 260,
+ 486
+ ],
+ "score": 0.93,
+ "content": "n _ { 1 } = \\mathrm { N N } ( q , { \\mathcal { M } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 473,
+ 367,
+ 488
+ ],
+ "score": 1.0,
+ "content": "be the nearest neighbor to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 367,
+ 476,
+ 373,
+ 486
+ ],
+ "score": 0.76,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 373,
+ 473,
+ 377,
+ 488
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 33.5,
+ "bbox_fs": [
+ 105,
+ 441,
+ 505,
+ 488
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 490,
+ 504,
+ 514
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 327,
+ 504
+ ],
+ "score": 1.0,
+ "content": "If the memory already returns the correct value, i.e., if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 490,
+ 371,
+ 503
+ ],
+ "score": 0.93,
+ "content": "V [ n _ { 1 } ] = v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 372,
+ 489,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": ", then we only update the key for",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 107,
+ 502,
+ 383,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 503,
+ 118,
+ 513
+ ],
+ "score": 0.83,
+ "content": "n _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 118,
+ 502,
+ 297,
+ 515
+ ],
+ "score": 1.0,
+ "content": "by taking the average of the current key and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 504,
+ 303,
+ 514
+ ],
+ "score": 0.81,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 502,
+ 383,
+ 515
+ ],
+ "score": 1.0,
+ "content": "and normalizing it:",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 105,
+ 489,
+ 505,
+ 515
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 255,
+ 518,
+ 356,
+ 546
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 255,
+ 518,
+ 356,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 255,
+ 518,
+ 356,
+ 546
+ ],
+ "score": 0.94,
+ "content": "K [ n _ { 1 } ] \\gets \\frac { q + K [ n _ { 1 } ] } { \\lVert q + K [ n _ { 1 } ] \\rVert } .",
+ "type": "interline_equation",
+ "image_path": "3fa7aa1f9cda9c5a5982ec2c4a5782e58eabac60bb96dea28f0da76a72647b1c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 38.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 255,
+ 518,
+ 356,
+ 532.0
+ ],
+ "spans": [],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 255,
+ 532.0,
+ 356,
+ 546.0
+ ],
+ "spans": [],
+ "index": 39
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 550,
+ 317,
+ 563
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 318,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 269,
+ 564
+ ],
+ "score": 1.0,
+ "content": "When doing this, we also re-set the age:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 550,
+ 314,
+ 563
+ ],
+ "score": 0.93,
+ "content": "A [ n _ { 1 } ] 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 315,
+ 549,
+ 318,
+ 564
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 106,
+ 549,
+ 318,
+ 564
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 567,
+ 505,
+ 623
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 506,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 178,
+ 581
+ ],
+ "score": 1.0,
+ "content": "Otherwise, when",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 567,
+ 225,
+ 580
+ ],
+ "score": 0.93,
+ "content": "V [ n _ { 1 } ] \\neq v",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 226,
+ 566,
+ 455,
+ 581
+ ],
+ "score": 1.0,
+ "content": ", we find a new place in the memory and write the pair",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 568,
+ 478,
+ 580
+ ],
+ "score": 0.92,
+ "content": "( q , v )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 566,
+ 506,
+ 581
+ ],
+ "score": 1.0,
+ "content": "there.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 591
+ ],
+ "score": 1.0,
+ "content": "Which place should we choose? We find memory items with maximum age, and write to one of",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 589,
+ 505,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 590,
+ 314,
+ 602
+ ],
+ "score": 1.0,
+ "content": "those (randomly chosen). More formally, we pick",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 314,
+ 589,
+ 414,
+ 601
+ ],
+ "score": 0.91,
+ "content": "n ^ { \\prime } = \\mathrm { a r g m a x } _ { i } A [ i ] + r _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 414,
+ 590,
+ 443,
+ 602
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 589,
+ 494,
+ 601
+ ],
+ "score": 0.92,
+ "content": "| r _ { i } | \\ll | \\mathcal { M } |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 494,
+ 590,
+ 505,
+ 602
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 613
+ ],
+ "score": 1.0,
+ "content": "a random number that introduces some randomness in the choice so as to avoid race conditions in",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 611,
+ 306,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 611,
+ 306,
+ 624
+ ],
+ "score": 1.0,
+ "content": "asynchronous multi-replica training. We then set:",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 43,
+ "bbox_fs": [
+ 105,
+ 566,
+ 506,
+ 624
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 223,
+ 627,
+ 387,
+ 642
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 223,
+ 627,
+ 387,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 223,
+ 627,
+ 387,
+ 642
+ ],
+ "score": 0.92,
+ "content": "K [ n ^ { \\prime } ] q , \\quad V [ n ^ { \\prime } ] v , \\quad A [ n ^ { \\prime } ] 0 .",
+ "type": "interline_equation",
+ "image_path": "9a5e7f44b6243f2c40fd82bf4fa870633978ae523232ca81e6765dd8dbd03591.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 46,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 223,
+ 627,
+ 387,
+ 642
+ ],
+ "spans": [],
+ "index": 46
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 646,
+ 505,
+ 670
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 505,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 505,
+ 660
+ ],
+ "score": 1.0,
+ "content": "With every memory update we also increment the age of all non-updated indices by 1. The full",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 335,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 335,
+ 671
+ ],
+ "score": 1.0,
+ "content": "operation of the memory module is depicted in Figure 1.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 47.5,
+ "bbox_fs": [
+ 105,
+ 646,
+ 505,
+ 671
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 681,
+ 503,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 694
+ ],
+ "score": 1.0,
+ "content": "Efficient nearest neighbor computation. The most expensive operation in our memory module",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 692,
+ 484,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 692,
+ 193,
+ 706
+ ],
+ "score": 1.0,
+ "content": "is the computation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 693,
+ 200,
+ 703
+ ],
+ "score": 0.81,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 692,
+ 484,
+ 706
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors. This can be done exactly or in an approximate way.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 49.5,
+ "bbox_fs": [
+ 105,
+ 681,
+ 505,
+ 706
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 709,
+ 503,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 350,
+ 723
+ ],
+ "score": 1.0,
+ "content": "In the exact mode, to calculate the nearest neighbors in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 351,
+ 710,
+ 361,
+ 720
+ ],
+ "score": 0.84,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 361,
+ 708,
+ 479,
+ 723
+ ],
+ "score": 1.0,
+ "content": "to a mini-batch of queries",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 480,
+ 709,
+ 504,
+ 721
+ ],
+ "score": 0.88,
+ "content": "Q \\ =",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 107,
+ 719,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 721,
+ 154,
+ 733
+ ],
+ "score": 0.91,
+ "content": "( q _ { 1 } , \\dots , q _ { b } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 719,
+ 329,
+ 734
+ ],
+ "score": 1.0,
+ "content": ", we perform a single matrix multiplication:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 721,
+ 362,
+ 732
+ ],
+ "score": 0.92,
+ "content": "Q \\times K ^ { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 362,
+ 719,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": ". This multiplies the batch-size",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 106,
+ 299,
+ 506,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 167,
+ 311
+ ],
+ "score": 0.28,
+ "content": "\\times \\mathrm { \\ k e y - s i z e }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 167,
+ 299,
+ 198,
+ 312
+ ],
+ "score": 1.0,
+ "content": "matrix",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 198,
+ 300,
+ 208,
+ 312
+ ],
+ "score": 0.83,
+ "content": "Q",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 208,
+ 299,
+ 238,
+ 312
+ ],
+ "score": 1.0,
+ "content": "by the",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 239,
+ 301,
+ 368,
+ 311
+ ],
+ "score": 0.56,
+ "content": "{ \\bf k e y - s i z e } \\times { \\bf m e m o r y - s i z e }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 368,
+ 299,
+ 399,
+ 312
+ ],
+ "score": 1.0,
+ "content": "matrix",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 400,
+ 299,
+ 416,
+ 310
+ ],
+ "score": 0.89,
+ "content": "K ^ { T }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 416,
+ 299,
+ 506,
+ 312
+ ],
+ "score": 1.0,
+ "content": ", and the result is the",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 505,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 165,
+ 323
+ ],
+ "score": 1.0,
+ "content": "batch-size",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 165,
+ 312,
+ 191,
+ 322
+ ],
+ "score": 0.35,
+ "content": "\\div \\times \\mathrm { ~ m } \\in",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 191,
+ 310,
+ 495,
+ 323
+ ],
+ "score": 1.0,
+ "content": "emory-size matrix of all distances, from which we can choose the top-",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 495,
+ 312,
+ 501,
+ 321
+ ],
+ "score": 0.77,
+ "content": "k",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 501,
+ 310,
+ 505,
+ 323
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "score": 1.0,
+ "content": "This procedure is linear in memory-size, so it can be expensive for very large memory sizes. But",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 345
+ ],
+ "score": 1.0,
+ "content": "matrix multiplication is very heavily optimized, so in our experiments on GPUs we find that this",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 344,
+ 375,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 375,
+ 357
+ ],
+ "score": 1.0,
+ "content": "operation is not a bottleneck for memory sizes up to half a million.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 51.5,
+ "bbox_fs": [
+ 105,
+ 708,
+ 505,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 181,
+ 84,
+ 425,
+ 217
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 181,
+ 84,
+ 425,
+ 217
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 181,
+ 84,
+ 425,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 181,
+ 84,
+ 425,
+ 217
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "b8f9c043a8b0838456ee094f0909ad7a399d67f149e328b2e34a7d07e1152410.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 181,
+ 84,
+ 425,
+ 97.3
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 181,
+ 97.3,
+ 425,
+ 110.6
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 181,
+ 110.6,
+ 425,
+ 123.89999999999999
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 181,
+ 123.89999999999999,
+ 425,
+ 137.2
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 181,
+ 137.2,
+ 425,
+ 150.5
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 181,
+ 150.5,
+ 425,
+ 163.8
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 181,
+ 163.8,
+ 425,
+ 177.10000000000002
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 181,
+ 177.10000000000002,
+ 425,
+ 190.40000000000003
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 181,
+ 190.40000000000003,
+ 425,
+ 203.70000000000005
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 181,
+ 203.70000000000005,
+ 425,
+ 217.00000000000006
+ ],
+ "spans": [],
+ "index": 9
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 240,
+ 503,
+ 274
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 444,
+ 252
+ ],
+ "score": 1.0,
+ "content": "Figure 2: The GNMT model with added memory module. On each decoding step",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 241,
+ 449,
+ 250
+ ],
+ "score": 0.65,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 240,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": ", the result of",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 252,
+ 505,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 252,
+ 158,
+ 264
+ ],
+ "score": 1.0,
+ "content": "the attention",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 253,
+ 168,
+ 262
+ ],
+ "score": 0.85,
+ "content": "a _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 169,
+ 252,
+ 505,
+ 264
+ ],
+ "score": 1.0,
+ "content": "is used to query the memory. The resulting value is combined with the output of the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 262,
+ 425,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 262,
+ 301,
+ 274
+ ],
+ "score": 1.0,
+ "content": "final LSTM layer to produce the predicted logits",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 302,
+ 263,
+ 311,
+ 273
+ ],
+ "score": 0.87,
+ "content": "\\hat { y } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 262,
+ 425,
+ 274
+ ],
+ "score": 1.0,
+ "content": ". See text for further details.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 7.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 299,
+ 504,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 506,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 167,
+ 311
+ ],
+ "score": 0.28,
+ "content": "\\times \\mathrm { \\ k e y - s i z e }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 299,
+ 198,
+ 312
+ ],
+ "score": 1.0,
+ "content": "matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 300,
+ 208,
+ 312
+ ],
+ "score": 0.83,
+ "content": "Q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 299,
+ 238,
+ 312
+ ],
+ "score": 1.0,
+ "content": "by the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 239,
+ 301,
+ 368,
+ 311
+ ],
+ "score": 0.56,
+ "content": "{ \\bf k e y - s i z e } \\times { \\bf m e m o r y - s i z e }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 368,
+ 299,
+ 399,
+ 312
+ ],
+ "score": 1.0,
+ "content": "matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 400,
+ 299,
+ 416,
+ 310
+ ],
+ "score": 0.89,
+ "content": "K ^ { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 416,
+ 299,
+ 506,
+ 312
+ ],
+ "score": 1.0,
+ "content": ", and the result is the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 505,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 165,
+ 323
+ ],
+ "score": 1.0,
+ "content": "batch-size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 312,
+ 191,
+ 322
+ ],
+ "score": 0.35,
+ "content": "\\div \\times \\mathrm { ~ m } \\in",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 191,
+ 310,
+ 495,
+ 323
+ ],
+ "score": 1.0,
+ "content": "emory-size matrix of all distances, from which we can choose the top-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 312,
+ 501,
+ 321
+ ],
+ "score": 0.77,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 310,
+ 505,
+ 323
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "score": 1.0,
+ "content": "This procedure is linear in memory-size, so it can be expensive for very large memory sizes. But",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 506,
+ 345
+ ],
+ "score": 1.0,
+ "content": "matrix multiplication is very heavily optimized, so in our experiments on GPUs we find that this",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 344,
+ 375,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 375,
+ 357
+ ],
+ "score": 1.0,
+ "content": "operation is not a bottleneck for memory sizes up to half a million.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 361,
+ 505,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 361,
+ 505,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 361,
+ 239,
+ 373
+ ],
+ "score": 1.0,
+ "content": "If the exact mode is too slow, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 239,
+ 361,
+ 246,
+ 371
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 361,
+ 505,
+ 373
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors can be computed approximately using locality",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "sensitive hashing (LSH). LSH is a hashing scheme so that near neighbors get similar hashes (Indyk",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 218,
+ 395
+ ],
+ "score": 1.0,
+ "content": "& Motwani, 1998; Andoni",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 383,
+ 228,
+ 393
+ ],
+ "score": 0.34,
+ "content": "\\&",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 228,
+ 383,
+ 505,
+ 395
+ ],
+ "score": 1.0,
+ "content": "Indyk, 2006). For cosine similarity, the computation of an LSH is",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 393,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 393,
+ 376,
+ 407
+ ],
+ "score": 1.0,
+ "content": "very simple. We pick a number of random normalized hash vectors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 377,
+ 394,
+ 419,
+ 405
+ ],
+ "score": 0.92,
+ "content": "h _ { 1 } , \\ldots , h _ { l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 393,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": ". The hash of a query",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 407,
+ 113,
+ 416
+ ],
+ "score": 0.73,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 113,
+ 404,
+ 183,
+ 419
+ ],
+ "score": 1.0,
+ "content": "is a sequence of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 405,
+ 188,
+ 415
+ ],
+ "score": 0.68,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 189,
+ 404,
+ 210,
+ 419
+ ],
+ "score": 1.0,
+ "content": "bits,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 210,
+ 405,
+ 249,
+ 416
+ ],
+ "score": 0.91,
+ "content": "b _ { 1 } , \\ldots , b _ { l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 404,
+ 293,
+ 419
+ ],
+ "score": 1.0,
+ "content": ", such that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 294,
+ 405,
+ 324,
+ 416
+ ],
+ "score": 0.91,
+ "content": "b _ { i } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 324,
+ 404,
+ 388,
+ 419
+ ],
+ "score": 1.0,
+ "content": "if, and only if,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 405,
+ 433,
+ 416
+ ],
+ "score": 0.91,
+ "content": "q \\cdot h _ { i } > 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 433,
+ 404,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": ". It turns out that",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 415,
+ 506,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 415,
+ 506,
+ 428
+ ],
+ "score": 1.0,
+ "content": "near neighbors will, with high probability, have a large number of identical bits in their hash. To",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "compute the nearest neighbors it is therefore sufficient to only look into parts of the memory with",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 437,
+ 505,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 505,
+ 450
+ ],
+ "score": 1.0,
+ "content": "similar hashes. This makes the nearest neighbor computation work in approximately constant time",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 449,
+ 490,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 490,
+ 461
+ ],
+ "score": 1.0,
+ "content": "– we only need to multiply the query by the hash vectors, and then only use the nearest buckets.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 480,
+ 263,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 265,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 265,
+ 492
+ ],
+ "score": 1.0,
+ "content": "2.1 USING THE MEMORY MODULE",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 502,
+ 504,
+ 525
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "The memory module presented above can be added to any classification network. There are two",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 514,
+ 486,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 514,
+ 486,
+ 527
+ ],
+ "score": 1.0,
+ "content": "main choices: which layer to use to generate queries, and how to use the output of the module.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 531,
+ 505,
+ 564
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 505,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 505,
+ 544
+ ],
+ "score": 1.0,
+ "content": "In the simplest case, we use the final layer of a network as query and the output of the module is",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "directly used for classification. This simplest case is similar to matching networks (Oriol Vinyals,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 552,
+ 454,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 454,
+ 566
+ ],
+ "score": 1.0,
+ "content": "2016b) and our memory module yields good results already in this setting (see below).",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 570,
+ 505,
+ 625
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 570,
+ 505,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 570,
+ 505,
+ 582
+ ],
+ "score": 1.0,
+ "content": "Instead of using the output of the module directly, it is possible to embed it again into a dense",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 581,
+ 505,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 505,
+ 594
+ ],
+ "score": 1.0,
+ "content": "representation and mix it with other predictions made by the network. To study this setting, we",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 592,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 592,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "add the memory module to sequence-to-sequence recurrent neural networks. As described in detail",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 616
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 616
+ ],
+ "score": 1.0,
+ "content": "below, a query to memory is made in every step of the decoder network. Memory output is embedded",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 479,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 479,
+ 626
+ ],
+ "score": 1.0,
+ "content": "again into a dense representation and combined with inputs from other layers of the network.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 643,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 656
+ ],
+ "score": 1.0,
+ "content": "Convolutional Network with Memory. To test our memory module in a simple setting, we first",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "add it to a basic convolutional network network for image classification. Our network consists of",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "two convolutional layers with ReLU non-linearity, followed by a max-pooling layer, another two",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "convolutional-ReLU layers, another max-pooling, and two fully connected layers. All convolutions",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 122,
+ 700
+ ],
+ "score": 1.0,
+ "content": "use",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 122,
+ 688,
+ 145,
+ 698
+ ],
+ "score": 0.89,
+ "content": "3 \\times 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "filters with 64 channels in the first pair, and 128 in the second. The fully connected layers",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "have dimension 256 and dropout applied between them. The output of the final layer is used as query",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "to our memory module and the nearest neighbor returned by the memory is used as the final network",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "prediction. Even this basic architecture yields good results in one-shot learning, as discussed below.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 41.5
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 759
+ ],
+ "lines": []
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 181,
+ 84,
+ 425,
+ 217
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 181,
+ 84,
+ 425,
+ 217
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 181,
+ 84,
+ 425,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 181,
+ 84,
+ 425,
+ 217
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "b8f9c043a8b0838456ee094f0909ad7a399d67f149e328b2e34a7d07e1152410.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 181,
+ 84,
+ 425,
+ 97.3
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 181,
+ 97.3,
+ 425,
+ 110.6
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 181,
+ 110.6,
+ 425,
+ 123.89999999999999
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 181,
+ 123.89999999999999,
+ 425,
+ 137.2
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 181,
+ 137.2,
+ 425,
+ 150.5
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 181,
+ 150.5,
+ 425,
+ 163.8
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 181,
+ 163.8,
+ 425,
+ 177.10000000000002
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 181,
+ 177.10000000000002,
+ 425,
+ 190.40000000000003
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 181,
+ 190.40000000000003,
+ 425,
+ 203.70000000000005
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 181,
+ 203.70000000000005,
+ 425,
+ 217.00000000000006
+ ],
+ "spans": [],
+ "index": 9
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 240,
+ 503,
+ 274
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 444,
+ 252
+ ],
+ "score": 1.0,
+ "content": "Figure 2: The GNMT model with added memory module. On each decoding step",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 241,
+ 449,
+ 250
+ ],
+ "score": 0.65,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 240,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": ", the result of",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 252,
+ 505,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 252,
+ 158,
+ 264
+ ],
+ "score": 1.0,
+ "content": "the attention",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 253,
+ 168,
+ 262
+ ],
+ "score": 0.85,
+ "content": "a _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 169,
+ 252,
+ 505,
+ 264
+ ],
+ "score": 1.0,
+ "content": "is used to query the memory. The resulting value is combined with the output of the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 262,
+ 425,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 262,
+ 301,
+ 274
+ ],
+ "score": 1.0,
+ "content": "final LSTM layer to produce the predicted logits",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 302,
+ 263,
+ 311,
+ 273
+ ],
+ "score": 0.87,
+ "content": "\\hat { y } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 262,
+ 425,
+ 274
+ ],
+ "score": 1.0,
+ "content": ". See text for further details.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 7.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 299,
+ 504,
+ 355
+ ],
+ "lines": [],
+ "index": 15,
+ "bbox_fs": [
+ 105,
+ 299,
+ 506,
+ 357
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 361,
+ 505,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 361,
+ 505,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 361,
+ 239,
+ 373
+ ],
+ "score": 1.0,
+ "content": "If the exact mode is too slow, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 239,
+ 361,
+ 246,
+ 371
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 361,
+ 505,
+ 373
+ ],
+ "score": 1.0,
+ "content": "nearest neighbors can be computed approximately using locality",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "sensitive hashing (LSH). LSH is a hashing scheme so that near neighbors get similar hashes (Indyk",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 218,
+ 395
+ ],
+ "score": 1.0,
+ "content": "& Motwani, 1998; Andoni",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 383,
+ 228,
+ 393
+ ],
+ "score": 0.34,
+ "content": "\\&",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 228,
+ 383,
+ 505,
+ 395
+ ],
+ "score": 1.0,
+ "content": "Indyk, 2006). For cosine similarity, the computation of an LSH is",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 393,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 393,
+ 376,
+ 407
+ ],
+ "score": 1.0,
+ "content": "very simple. We pick a number of random normalized hash vectors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 377,
+ 394,
+ 419,
+ 405
+ ],
+ "score": 0.92,
+ "content": "h _ { 1 } , \\ldots , h _ { l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 393,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": ". The hash of a query",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 407,
+ 113,
+ 416
+ ],
+ "score": 0.73,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 113,
+ 404,
+ 183,
+ 419
+ ],
+ "score": 1.0,
+ "content": "is a sequence of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 405,
+ 188,
+ 415
+ ],
+ "score": 0.68,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 189,
+ 404,
+ 210,
+ 419
+ ],
+ "score": 1.0,
+ "content": "bits,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 210,
+ 405,
+ 249,
+ 416
+ ],
+ "score": 0.91,
+ "content": "b _ { 1 } , \\ldots , b _ { l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 404,
+ 293,
+ 419
+ ],
+ "score": 1.0,
+ "content": ", such that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 294,
+ 405,
+ 324,
+ 416
+ ],
+ "score": 0.91,
+ "content": "b _ { i } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 324,
+ 404,
+ 388,
+ 419
+ ],
+ "score": 1.0,
+ "content": "if, and only if,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 405,
+ 433,
+ 416
+ ],
+ "score": 0.91,
+ "content": "q \\cdot h _ { i } > 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 433,
+ 404,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": ". It turns out that",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 415,
+ 506,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 415,
+ 506,
+ 428
+ ],
+ "score": 1.0,
+ "content": "near neighbors will, with high probability, have a large number of identical bits in their hash. To",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "compute the nearest neighbors it is therefore sufficient to only look into parts of the memory with",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 437,
+ 505,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 505,
+ 450
+ ],
+ "score": 1.0,
+ "content": "similar hashes. This makes the nearest neighbor computation work in approximately constant time",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 449,
+ 490,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 490,
+ 461
+ ],
+ "score": 1.0,
+ "content": "– we only need to multiply the query by the hash vectors, and then only use the nearest buckets.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 361,
+ 506,
+ 461
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 480,
+ 263,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 265,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 265,
+ 492
+ ],
+ "score": 1.0,
+ "content": "2.1 USING THE MEMORY MODULE",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 502,
+ 504,
+ 525
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "The memory module presented above can be added to any classification network. There are two",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 514,
+ 486,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 514,
+ 486,
+ 527
+ ],
+ "score": 1.0,
+ "content": "main choices: which layer to use to generate queries, and how to use the output of the module.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 106,
+ 502,
+ 505,
+ 527
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 531,
+ 505,
+ 564
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 505,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 505,
+ 544
+ ],
+ "score": 1.0,
+ "content": "In the simplest case, we use the final layer of a network as query and the output of the module is",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "directly used for classification. This simplest case is similar to matching networks (Oriol Vinyals,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 552,
+ 454,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 454,
+ 566
+ ],
+ "score": 1.0,
+ "content": "2016b) and our memory module yields good results already in this setting (see below).",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 105,
+ 531,
+ 505,
+ 566
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 570,
+ 505,
+ 625
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 570,
+ 505,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 570,
+ 505,
+ 582
+ ],
+ "score": 1.0,
+ "content": "Instead of using the output of the module directly, it is possible to embed it again into a dense",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 581,
+ 505,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 505,
+ 594
+ ],
+ "score": 1.0,
+ "content": "representation and mix it with other predictions made by the network. To study this setting, we",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 592,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 592,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "add the memory module to sequence-to-sequence recurrent neural networks. As described in detail",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 616
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 505,
+ 616
+ ],
+ "score": 1.0,
+ "content": "below, a query to memory is made in every step of the decoder network. Memory output is embedded",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 479,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 479,
+ 626
+ ],
+ "score": 1.0,
+ "content": "again into a dense representation and combined with inputs from other layers of the network.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 570,
+ 505,
+ 626
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 643,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 656
+ ],
+ "score": 1.0,
+ "content": "Convolutional Network with Memory. To test our memory module in a simple setting, we first",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "add it to a basic convolutional network network for image classification. Our network consists of",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "two convolutional layers with ReLU non-linearity, followed by a max-pooling layer, another two",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "convolutional-ReLU layers, another max-pooling, and two fully connected layers. All convolutions",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 122,
+ 700
+ ],
+ "score": 1.0,
+ "content": "use",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 122,
+ 688,
+ 145,
+ 698
+ ],
+ "score": 0.89,
+ "content": "3 \\times 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "filters with 64 channels in the first pair, and 128 in the second. The fully connected layers",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "have dimension 256 and dropout applied between them. The output of the final layer is used as query",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "to our memory module and the nearest neighbor returned by the memory is used as the final network",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "prediction. Even this basic architecture yields good results in one-shot learning, as discussed below.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 41.5,
+ "bbox_fs": [
+ 105,
+ 644,
+ 505,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 127,
+ 78,
+ 488,
+ 216
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 127,
+ 78,
+ 488,
+ 216
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 127,
+ 78,
+ 488,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 78,
+ 488,
+ 216
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "7d63c9c6459cd8c2b28898bc12351be72bca2ac04e812c6be2789ac05c86e989.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 127,
+ 78,
+ 488,
+ 124.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 127,
+ 124.0,
+ 488,
+ 170.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 170.0,
+ 488,
+ 216.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 267
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 233,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 233,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Extended Neural GPU with memory module. Memory query is read from the position one",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": "below the current output logit, and the embedded memory value is put at the same position of the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 474,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 153,
+ 268
+ ],
+ "score": 1.0,
+ "content": "output tape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 153,
+ 257,
+ 160,
+ 266
+ ],
+ "score": 0.75,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 254,
+ 474,
+ 268
+ ],
+ "score": 1.0,
+ "content": ". The network learns to use these values to produce the output in the next step.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 287,
+ 505,
+ 430
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 288,
+ 504,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 288,
+ 504,
+ 299
+ ],
+ "score": 1.0,
+ "content": "Sequence-to-sequence with Memory. For large-scale experiments, we add the memory mod-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 310
+ ],
+ "score": 1.0,
+ "content": "ule into a large sequence-to-sequence model. Such sequence-to-sequence recurrent neural networks",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 308,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "(RNNs) with long short-term memory (LSTM) cells (Hochreiter & Schmidhuber, 1997) have proven",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "especially successful at natural language processing (NLP) tasks, including machine translation",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 331,
+ 505,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 331,
+ 505,
+ 343
+ ],
+ "score": 1.0,
+ "content": "(Sutskever et al., 2014; Bahdanau et al., 2014; Cho et al., 2014). We add the memory module to",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 342,
+ 506,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 506,
+ 355
+ ],
+ "score": 1.0,
+ "content": "the Google Neural Machine Translation (GNMT) model (Wu et al., 2016). This model consists of",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 365
+ ],
+ "score": 1.0,
+ "content": "an encoder RNN, which creates a representation of the source language sentence, and a decoder",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 504,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 504,
+ 377
+ ],
+ "score": 1.0,
+ "content": "RNN that outputs the target language sentence. We left the encoder RNN unmodified. In the de-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "score": 1.0,
+ "content": "coder RNN, we use the vector retrieved by the attention mechanism as query to the memory module.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": "In the GNMT model, the attention vector is used in all LSTM layers beyond the second one, so the",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "computation of the other layers and the memory can happen in parallel. Before the final softmax",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "layer, we combine the embedded memory output with the output of the final LSTM layer using an",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 419,
+ 294,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 419,
+ 294,
+ 431
+ ],
+ "score": 1.0,
+ "content": "additional linear layer, as depicted in Figure 2.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 442,
+ 505,
+ 552
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "score": 1.0,
+ "content": "Extended Neural GPU with Memory. To test versatility of our memory module, we also add it to",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 466
+ ],
+ "score": 1.0,
+ "content": "the Extended Neural GPU, a convolutional-recurrent model introduced by Kaiser & Bengio (2016).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 478
+ ],
+ "score": 1.0,
+ "content": "The Extended Neural GPU is a sequence-to-sequence model too, but its decoder is convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 506,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 506,
+ 488
+ ],
+ "score": 1.0,
+ "content": "and the size of its state changes depending on the size of the input. Again, we leave the encoder",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 486,
+ 506,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 486,
+ 506,
+ 499
+ ],
+ "score": 1.0,
+ "content": "part of the model intact, and extend the decoder part by a memory query. This time, we use the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 496,
+ 506,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 496,
+ 506,
+ 511
+ ],
+ "score": 1.0,
+ "content": "position one step ahead to query memory, and we put the embedded result to the output tape, as",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "shown in Figure 3. Note that in this model the result of the memory will be processed by two",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 104,
+ 519,
+ 506,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 519,
+ 506,
+ 531
+ ],
+ "score": 1.0,
+ "content": "recurrent-convolutional cells before the corresponding output is produced. The fact that this model",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 530,
+ 505,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 530,
+ 505,
+ 543
+ ],
+ "score": 1.0,
+ "content": "still does one-shot learning confirms that the output of our memory module can be used deep inside",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 270,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 270,
+ 554
+ ],
+ "score": 1.0,
+ "content": "a network, not just near the output layer.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 569,
+ 211,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 213,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 213,
+ 583
+ ],
+ "score": 1.0,
+ "content": "3 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 594,
+ 504,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "Memory in Neural Networks. Augmenting neural networks with memory has been heavily stud-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "ied recently. Many of these approaches design a memory component that is intended as a general-",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 617,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 617,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "ization of the memory in standard recurrent neural networks. In recurrent networks, the state passed",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "from one time step to the next can be interpreted as the network’s memory representation of the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "current example. Moving away from this fixed-length vector representation of memory to a larger",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 329,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 329,
+ 660
+ ],
+ "score": 1.0,
+ "content": "and more versatile form is at the core of these methods.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 665,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 665,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "Augmenting recurrent neural networks with attention (Bahdanau et al., 2014) can be interpreted as",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "creating a large memory component that allows content-based addressing. More generally, Graves",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "et al. (2014) augmented a recurrent neural network with a computing-inspired memory component",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "that can be addressed via both content- and address-based queries. Sukhbaatar et al. (2015) present",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 104,
+ 708,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 708,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": "a similar augmentation and show the importance of allowing multiple reads and writes to memory",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "between inputs. These approaches excel at tasks where it is necessary to store large parts of a se-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 38.5
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 127,
+ 78,
+ 488,
+ 216
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 127,
+ 78,
+ 488,
+ 216
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 127,
+ 78,
+ 488,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 78,
+ 488,
+ 216
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "7d63c9c6459cd8c2b28898bc12351be72bca2ac04e812c6be2789ac05c86e989.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 127,
+ 78,
+ 488,
+ 124.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 127,
+ 124.0,
+ 488,
+ 170.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 170.0,
+ 488,
+ 216.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 267
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 233,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 233,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Extended Neural GPU with memory module. Memory query is read from the position one",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": "below the current output logit, and the embedded memory value is put at the same position of the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 474,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 153,
+ 268
+ ],
+ "score": 1.0,
+ "content": "output tape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 153,
+ 257,
+ 160,
+ 266
+ ],
+ "score": 0.75,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 254,
+ 474,
+ 268
+ ],
+ "score": 1.0,
+ "content": ". The network learns to use these values to produce the output in the next step.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 287,
+ 505,
+ 430
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 288,
+ 504,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 288,
+ 504,
+ 299
+ ],
+ "score": 1.0,
+ "content": "Sequence-to-sequence with Memory. For large-scale experiments, we add the memory mod-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 310
+ ],
+ "score": 1.0,
+ "content": "ule into a large sequence-to-sequence model. Such sequence-to-sequence recurrent neural networks",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 308,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "(RNNs) with long short-term memory (LSTM) cells (Hochreiter & Schmidhuber, 1997) have proven",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "especially successful at natural language processing (NLP) tasks, including machine translation",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 331,
+ 505,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 331,
+ 505,
+ 343
+ ],
+ "score": 1.0,
+ "content": "(Sutskever et al., 2014; Bahdanau et al., 2014; Cho et al., 2014). We add the memory module to",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 342,
+ 506,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 506,
+ 355
+ ],
+ "score": 1.0,
+ "content": "the Google Neural Machine Translation (GNMT) model (Wu et al., 2016). This model consists of",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 365
+ ],
+ "score": 1.0,
+ "content": "an encoder RNN, which creates a representation of the source language sentence, and a decoder",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 504,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 504,
+ 377
+ ],
+ "score": 1.0,
+ "content": "RNN that outputs the target language sentence. We left the encoder RNN unmodified. In the de-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "score": 1.0,
+ "content": "coder RNN, we use the vector retrieved by the attention mechanism as query to the memory module.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": "In the GNMT model, the attention vector is used in all LSTM layers beyond the second one, so the",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "computation of the other layers and the memory can happen in parallel. Before the final softmax",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "layer, we combine the embedded memory output with the output of the final LSTM layer using an",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 419,
+ 294,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 419,
+ 294,
+ 431
+ ],
+ "score": 1.0,
+ "content": "additional linear layer, as depicted in Figure 2.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 105,
+ 288,
+ 506,
+ 431
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 442,
+ 505,
+ 552
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "score": 1.0,
+ "content": "Extended Neural GPU with Memory. To test versatility of our memory module, we also add it to",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 466
+ ],
+ "score": 1.0,
+ "content": "the Extended Neural GPU, a convolutional-recurrent model introduced by Kaiser & Bengio (2016).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 478
+ ],
+ "score": 1.0,
+ "content": "The Extended Neural GPU is a sequence-to-sequence model too, but its decoder is convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 506,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 506,
+ 488
+ ],
+ "score": 1.0,
+ "content": "and the size of its state changes depending on the size of the input. Again, we leave the encoder",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 486,
+ 506,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 486,
+ 506,
+ 499
+ ],
+ "score": 1.0,
+ "content": "part of the model intact, and extend the decoder part by a memory query. This time, we use the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 496,
+ 506,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 496,
+ 506,
+ 511
+ ],
+ "score": 1.0,
+ "content": "position one step ahead to query memory, and we put the embedded result to the output tape, as",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "shown in Figure 3. Note that in this model the result of the memory will be processed by two",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 104,
+ 519,
+ 506,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 519,
+ 506,
+ 531
+ ],
+ "score": 1.0,
+ "content": "recurrent-convolutional cells before the corresponding output is produced. The fact that this model",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 530,
+ 505,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 530,
+ 505,
+ 543
+ ],
+ "score": 1.0,
+ "content": "still does one-shot learning confirms that the output of our memory module can be used deep inside",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 270,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 270,
+ 554
+ ],
+ "score": 1.0,
+ "content": "a network, not just near the output layer.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 104,
+ 442,
+ 506,
+ 554
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 569,
+ 211,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 213,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 213,
+ 583
+ ],
+ "score": 1.0,
+ "content": "3 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 594,
+ 504,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "Memory in Neural Networks. Augmenting neural networks with memory has been heavily stud-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "ied recently. Many of these approaches design a memory component that is intended as a general-",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 617,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 617,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "ization of the memory in standard recurrent neural networks. In recurrent networks, the state passed",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "from one time step to the next can be interpreted as the network’s memory representation of the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "current example. Moving away from this fixed-length vector representation of memory to a larger",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 329,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 329,
+ 660
+ ],
+ "score": 1.0,
+ "content": "and more versatile form is at the core of these methods.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 105,
+ 594,
+ 506,
+ 660
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 665,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 665,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "Augmenting recurrent neural networks with attention (Bahdanau et al., 2014) can be interpreted as",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "creating a large memory component that allows content-based addressing. More generally, Graves",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "et al. (2014) augmented a recurrent neural network with a computing-inspired memory component",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "that can be addressed via both content- and address-based queries. Sukhbaatar et al. (2015) present",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 104,
+ 708,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 708,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": "a similar augmentation and show the importance of allowing multiple reads and writes to memory",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "between inputs. These approaches excel at tasks where it is necessary to store large parts of a se-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "quential input in a representation that can later be precisely queried. Such tasks include algorithmic",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 461,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 461,
+ 106
+ ],
+ "score": 1.0,
+ "content": "sequence manipulation tasks, natural language modelling, and question-answering tasks.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 38.5,
+ "bbox_fs": [
+ 104,
+ 665,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 503,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "quential input in a representation that can later be precisely queried. Such tasks include algorithmic",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 461,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 461,
+ 106
+ ],
+ "score": 1.0,
+ "content": "sequence manipulation tasks, natural language modelling, and question-answering tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 505,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 506,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 506,
+ 124
+ ],
+ "score": 1.0,
+ "content": "The success of these approaches hinges on making the memory component fully differentiable and",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": "backpropagating signal through every access of memory. In this setting, computational requirements",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 145
+ ],
+ "score": 1.0,
+ "content": "necessitate that the memory be small. Some attempts have been made at making hard access queries",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "to memory (Zaremba & Sutskever, 2015; Xu et al., 2015), but it was usually challenging to match",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 167
+ ],
+ "score": 1.0,
+ "content": "the soft version. Recently, more successful training for hard queries was reported (Gulc¸ehre et al. ¨ ,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "2016) that makes use of a curriculum strategy that mixes soft and hard queries at training time. Our",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 190
+ ],
+ "score": 1.0,
+ "content": "approach applies hard access as well, but we encourage the model to make good queries via a special",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 163,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 163,
+ 200
+ ],
+ "score": 1.0,
+ "content": "memory loss.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 204,
+ 505,
+ 281
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 217
+ ],
+ "score": 1.0,
+ "content": "Modifications to allow for large-scale memory in neural networks have been proposed. The original",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 215,
+ 506,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 506,
+ 228
+ ],
+ "score": 1.0,
+ "content": "implementation of memory networks (Weston et al., 2014) and later work on scaling it (Bordes",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 239
+ ],
+ "score": 1.0,
+ "content": "et al., 2015; Chandar et al., 2016) used memory with size in the millions. The cost of doing so is",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "that the memory must be fixed prior to training. Moreover, since during the beginning of training the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 247,
+ 506,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 247,
+ 506,
+ 261
+ ],
+ "score": 1.0,
+ "content": "model is unlikely to query the memory correctly, strong supervision is used to encourage the model",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "to query memory locations that are useful. These hints are either given as additional supervising",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 404,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 404,
+ 282
+ ],
+ "score": 1.0,
+ "content": "information by the task or determined heuristically as in Hill et al. (2015).",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 286,
+ 505,
+ 375
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 285,
+ 505,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 285,
+ 505,
+ 300
+ ],
+ "score": 1.0,
+ "content": "All the work discussed so far has either used a memory that is fixed before training or used a memory",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 311
+ ],
+ "score": 1.0,
+ "content": "that is not persistent between different examples. For one-shot and lifelong learning, a memory must",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 308,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "necessarily be both volatile during training and persistent between examples. To bridge this gap,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 332
+ ],
+ "score": 1.0,
+ "content": "Santoro et al. (2016) propose to partition training into distinct episodes consisting of a sequence",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 330,
+ 506,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 330,
+ 195,
+ 345
+ ],
+ "score": 1.0,
+ "content": "of labelled examples",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 330,
+ 248,
+ 343
+ ],
+ "score": 0.93,
+ "content": "\\bar { \\{ ( x _ { i } , y _ { i } ) \\} } _ { i = 1 } ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 330,
+ 506,
+ 345
+ ],
+ "score": 1.0,
+ "content": ". A network augmented with a fully-differentiable memory is",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 506,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 177,
+ 355
+ ],
+ "score": 1.0,
+ "content": "trained to predict",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 343,
+ 187,
+ 353
+ ],
+ "score": 0.84,
+ "content": "y _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 342,
+ 305,
+ 355
+ ],
+ "score": 1.0,
+ "content": "given the previous sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 305,
+ 342,
+ 379,
+ 354
+ ],
+ "score": 0.9,
+ "content": "( x _ { 1 } , y _ { 1 } , \\dotsc , x _ { i - 1 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 379,
+ 342,
+ 506,
+ 355
+ ],
+ "score": 1.0,
+ "content": ". This way, the model learns to",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 365
+ ],
+ "score": 1.0,
+ "content": "store important examples with their corresponding labels in memory and later re-use this information",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 364,
+ 504,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 504,
+ 376
+ ],
+ "score": 1.0,
+ "content": "to correctly classify new examples. This model successfully exhibits one-shot learning on Omniglot.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 380,
+ 504,
+ 425
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 506,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 506,
+ 392
+ ],
+ "score": 1.0,
+ "content": "However, this approach again requires fully-differentiable memory access and thus limits the size of",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 506,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 506,
+ 403
+ ],
+ "score": 1.0,
+ "content": "the memory as well as the length of an episode. This restriction has recently been alleviated by Rae",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 506,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 506,
+ 415
+ ],
+ "score": 1.0,
+ "content": "et al. (2016). Their model can utilize large memories, but unlike our work does not have an explicit",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 413,
+ 285,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 285,
+ 426
+ ],
+ "score": 1.0,
+ "content": "cost to guide the formation of memory keys.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 430,
+ 505,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "For classification tasks like Omniglot, it is easy to construct short episodes so that they include a few",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "examples from each of several classes. However, this becomes harder as the output becomes richer.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 452,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": "For example, in the difficult sequence-to-sequence tasks which we consider, it is hard to determine",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 462,
+ 505,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 462,
+ 505,
+ 476
+ ],
+ "score": 1.0,
+ "content": "which examples would be helpful for correctly predicting others a priori, and so constructing short",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 474,
+ 484,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 484,
+ 486
+ ],
+ "score": 1.0,
+ "content": "episodes each containing examples that are similar and act as hints to each other is intractable.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 533,
+ 505,
+ 566
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "One-shot Learning. While the recent work of Santoro et al. (2016) succeeded in bridging the",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 507,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 507,
+ 558
+ ],
+ "score": 1.0,
+ "content": "gap between memory-based models and one-shot learning, the field of one-shot learning has seen a",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 273,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 273,
+ 568
+ ],
+ "score": 1.0,
+ "content": "variety of different approaches over time.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 572,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 584
+ ],
+ "score": 1.0,
+ "content": "Early work utilized Bayesian methods to model data generatively (Fei-Fei et al., 2006; Lake et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "score": 1.0,
+ "content": "2011). The paper that introduced the Omniglot dataset (Lake et al., 2011) approached the task with a",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "generative model for strokes. This way, given a single character image, the probability of a different",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "image being of the same character may be approximated via standard techniques. One early neural",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "network approach to one-shot learning was given by Siamese networks (Koch, 2015). When our",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "approach is applied to the Omniglot image classification dataset, the resulting training algorithm is",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "actually similar to that of Siamese networks. The only difference is in the loss function: Siamese",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 442,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 442,
+ 661
+ ],
+ "score": 1.0,
+ "content": "networks utilize a cross-entropy loss whereas our method uses a margin triplet loss.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "A more sophisticated neural network approach is given by Vinyals et al. (2016). The strengths of",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 688
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 688
+ ],
+ "score": 1.0,
+ "content": "this approach are (1) the model architecture utilizes recent advances in attention-augmented neural",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "networks for set-to-set learning (Oriol Vinyals, 2016a), and (2) the training algorithm is designed to",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 264,
+ 712
+ ],
+ "score": 1.0,
+ "content": "exactly match the testing phase (given",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 699,
+ 271,
+ 709
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 272,
+ 699,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "distinct images and an additional image, the model must",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 189,
+ 722
+ ],
+ "score": 1.0,
+ "content": "predict which of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 710,
+ 196,
+ 720
+ ],
+ "score": 0.8,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 196,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "images is of the same class as the additional image). This approach may also",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 388,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 388,
+ 734
+ ],
+ "score": 1.0,
+ "content": "be considered as a generalization of previous work on metric learning.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 47.5
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 503,
+ 105
+ ],
+ "lines": [],
+ "index": 0.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 505,
+ 106
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 505,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 506,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 506,
+ 124
+ ],
+ "score": 1.0,
+ "content": "The success of these approaches hinges on making the memory component fully differentiable and",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": "backpropagating signal through every access of memory. In this setting, computational requirements",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 145
+ ],
+ "score": 1.0,
+ "content": "necessitate that the memory be small. Some attempts have been made at making hard access queries",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "to memory (Zaremba & Sutskever, 2015; Xu et al., 2015), but it was usually challenging to match",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 167
+ ],
+ "score": 1.0,
+ "content": "the soft version. Recently, more successful training for hard queries was reported (Gulc¸ehre et al. ¨ ,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "2016) that makes use of a curriculum strategy that mixes soft and hard queries at training time. Our",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 190
+ ],
+ "score": 1.0,
+ "content": "approach applies hard access as well, but we encourage the model to make good queries via a special",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 163,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 163,
+ 200
+ ],
+ "score": 1.0,
+ "content": "memory loss.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 110,
+ 506,
+ 200
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 204,
+ 505,
+ 281
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 217
+ ],
+ "score": 1.0,
+ "content": "Modifications to allow for large-scale memory in neural networks have been proposed. The original",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 215,
+ 506,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 506,
+ 228
+ ],
+ "score": 1.0,
+ "content": "implementation of memory networks (Weston et al., 2014) and later work on scaling it (Bordes",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 506,
+ 239
+ ],
+ "score": 1.0,
+ "content": "et al., 2015; Chandar et al., 2016) used memory with size in the millions. The cost of doing so is",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "that the memory must be fixed prior to training. Moreover, since during the beginning of training the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 247,
+ 506,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 247,
+ 506,
+ 261
+ ],
+ "score": 1.0,
+ "content": "model is unlikely to query the memory correctly, strong supervision is used to encourage the model",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "to query memory locations that are useful. These hints are either given as additional supervising",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 404,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 404,
+ 282
+ ],
+ "score": 1.0,
+ "content": "information by the task or determined heuristically as in Hill et al. (2015).",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 105,
+ 204,
+ 506,
+ 282
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 286,
+ 505,
+ 375
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 285,
+ 505,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 285,
+ 505,
+ 300
+ ],
+ "score": 1.0,
+ "content": "All the work discussed so far has either used a memory that is fixed before training or used a memory",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 311
+ ],
+ "score": 1.0,
+ "content": "that is not persistent between different examples. For one-shot and lifelong learning, a memory must",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 308,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "necessarily be both volatile during training and persistent between examples. To bridge this gap,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 332
+ ],
+ "score": 1.0,
+ "content": "Santoro et al. (2016) propose to partition training into distinct episodes consisting of a sequence",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 330,
+ 506,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 330,
+ 195,
+ 345
+ ],
+ "score": 1.0,
+ "content": "of labelled examples",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 330,
+ 248,
+ 343
+ ],
+ "score": 0.93,
+ "content": "\\bar { \\{ ( x _ { i } , y _ { i } ) \\} } _ { i = 1 } ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 330,
+ 506,
+ 345
+ ],
+ "score": 1.0,
+ "content": ". A network augmented with a fully-differentiable memory is",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 506,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 177,
+ 355
+ ],
+ "score": 1.0,
+ "content": "trained to predict",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 343,
+ 187,
+ 353
+ ],
+ "score": 0.84,
+ "content": "y _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 342,
+ 305,
+ 355
+ ],
+ "score": 1.0,
+ "content": "given the previous sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 305,
+ 342,
+ 379,
+ 354
+ ],
+ "score": 0.9,
+ "content": "( x _ { 1 } , y _ { 1 } , \\dotsc , x _ { i - 1 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 379,
+ 342,
+ 506,
+ 355
+ ],
+ "score": 1.0,
+ "content": ". This way, the model learns to",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 365
+ ],
+ "score": 1.0,
+ "content": "store important examples with their corresponding labels in memory and later re-use this information",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 364,
+ 504,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 504,
+ 376
+ ],
+ "score": 1.0,
+ "content": "to correctly classify new examples. This model successfully exhibits one-shot learning on Omniglot.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 105,
+ 285,
+ 506,
+ 376
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 380,
+ 504,
+ 425
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 506,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 506,
+ 392
+ ],
+ "score": 1.0,
+ "content": "However, this approach again requires fully-differentiable memory access and thus limits the size of",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 506,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 506,
+ 403
+ ],
+ "score": 1.0,
+ "content": "the memory as well as the length of an episode. This restriction has recently been alleviated by Rae",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 506,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 506,
+ 415
+ ],
+ "score": 1.0,
+ "content": "et al. (2016). Their model can utilize large memories, but unlike our work does not have an explicit",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 413,
+ 285,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 285,
+ 426
+ ],
+ "score": 1.0,
+ "content": "cost to guide the formation of memory keys.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 380,
+ 506,
+ 426
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 430,
+ 505,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "For classification tasks like Omniglot, it is easy to construct short episodes so that they include a few",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "examples from each of several classes. However, this becomes harder as the output becomes richer.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 452,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": "For example, in the difficult sequence-to-sequence tasks which we consider, it is hard to determine",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 462,
+ 505,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 462,
+ 505,
+ 476
+ ],
+ "score": 1.0,
+ "content": "which examples would be helpful for correctly predicting others a priori, and so constructing short",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 474,
+ 484,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 484,
+ 486
+ ],
+ "score": 1.0,
+ "content": "episodes each containing examples that are similar and act as hints to each other is intractable.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 105,
+ 430,
+ 505,
+ 486
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 533,
+ 505,
+ 566
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "One-shot Learning. While the recent work of Santoro et al. (2016) succeeded in bridging the",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 507,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 507,
+ 558
+ ],
+ "score": 1.0,
+ "content": "gap between memory-based models and one-shot learning, the field of one-shot learning has seen a",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 273,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 273,
+ 568
+ ],
+ "score": 1.0,
+ "content": "variety of different approaches over time.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 532,
+ 507,
+ 568
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 572,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 584
+ ],
+ "score": 1.0,
+ "content": "Early work utilized Bayesian methods to model data generatively (Fei-Fei et al., 2006; Lake et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "score": 1.0,
+ "content": "2011). The paper that introduced the Omniglot dataset (Lake et al., 2011) approached the task with a",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "generative model for strokes. This way, given a single character image, the probability of a different",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "image being of the same character may be approximated via standard techniques. One early neural",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "network approach to one-shot learning was given by Siamese networks (Koch, 2015). When our",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "approach is applied to the Omniglot image classification dataset, the resulting training algorithm is",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "actually similar to that of Siamese networks. The only difference is in the loss function: Siamese",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 442,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 442,
+ 661
+ ],
+ "score": 1.0,
+ "content": "networks utilize a cross-entropy loss whereas our method uses a margin triplet loss.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 572,
+ 506,
+ 661
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "A more sophisticated neural network approach is given by Vinyals et al. (2016). The strengths of",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 688
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 688
+ ],
+ "score": 1.0,
+ "content": "this approach are (1) the model architecture utilizes recent advances in attention-augmented neural",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "networks for set-to-set learning (Oriol Vinyals, 2016a), and (2) the training algorithm is designed to",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 264,
+ 712
+ ],
+ "score": 1.0,
+ "content": "exactly match the testing phase (given",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 699,
+ 271,
+ 709
+ ],
+ "score": 0.79,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 272,
+ 699,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "distinct images and an additional image, the model must",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 189,
+ 722
+ ],
+ "score": 1.0,
+ "content": "predict which of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 710,
+ 196,
+ 720
+ ],
+ "score": 0.8,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 196,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "images is of the same class as the additional image). This approach may also",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 388,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 388,
+ 734
+ ],
+ "score": 1.0,
+ "content": "be considered as a generalization of previous work on metric learning.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 47.5,
+ "bbox_fs": [
+ 105,
+ 665,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 136,
+ 514,
+ 205
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 89,
+ 505,
+ 123
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "score": 1.0,
+ "content": "Table 1: Results on the Omniglot dataset. Although our model uses only a simple convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 100,
+ 505,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 505,
+ 113
+ ],
+ "score": 1.0,
+ "content": "neural network, the addition of our memory module allows it to approach much more complex",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 111,
+ 298,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 111,
+ 298,
+ 123
+ ],
+ "score": 1.0,
+ "content": "models on 1-shot and multi-shot learning tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 106,
+ 136,
+ 514,
+ 205
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 514,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 514,
+ 205
+ ],
+ "score": 0.979,
+ "html": "
| Model | 5-way 1-shot | 5-way 5-shot | 20-way 1-shot | 20-way 5-shot |
| Pixels Nearest Neighbor | 41.7% | 63.2% | 26.7% | 42.6% |
| MANN (no convolutions) | 82.8% | 94.9% | 1 | 1 |
| Convolutional Siamese Net | 96.7% | 98.4% | 88.0% | 96.5% |
| Matching Network | 98.1% | 98.9% | 93.8% | 98.5% |
| ConvNet with Memory Module | 98.4% | 99.6% | 95.0% | 98.6% |
",
+ "type": "table",
+ "image_path": "3e84788e380c6c8f31c00e5fa93f09a01f2fb8880153858877a44dfe07b512aa.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 514,
+ 159.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 159.0,
+ 514,
+ 182.0
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 182.0,
+ 514,
+ 205.0
+ ],
+ "spans": [],
+ "index": 5
+ }
+ ]
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 229,
+ 200,
+ 241
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 201,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 201,
+ 243
+ ],
+ "score": 1.0,
+ "content": "4 EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 253,
+ 505,
+ 320
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "We perform experiments using all three architectures described above. We experiment both on real-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 264,
+ 506,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 264,
+ 506,
+ 277
+ ],
+ "score": 1.0,
+ "content": "world data and on synthetic tasks that give us some insight into the performance and limitations of",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": "the memory module. In all our experiments we use the Adam optimizer (Kingma & Ba, 2014) and",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 505,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 344,
+ 299
+ ],
+ "score": 1.0,
+ "content": "the parameters for the memory module remain unchanged",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 286,
+ 421,
+ 298
+ ],
+ "score": 0.33,
+ "content": "( k = 2 5 6 , \\alpha = 0 . 1 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 286,
+ 505,
+ 299
+ ],
+ "score": 1.0,
+ "content": ". Good performance",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 310
+ ],
+ "score": 1.0,
+ "content": "with a single set of parameters shows the versatility of our memory module. The source code for the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 308,
+ 432,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 432,
+ 321
+ ],
+ "score": 1.0,
+ "content": "memory module, together with our settings for Omniglot, is available on github1.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 331,
+ 505,
+ 431
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 343
+ ],
+ "score": 1.0,
+ "content": "Omniglot. The Omniglot dataset (Lake et al., 2011) consists of 1623 characters from 50 different",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 355
+ ],
+ "score": 1.0,
+ "content": "alphabets, each hand-drawn by 20 different people. The large number of classes (characters) with",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 354,
+ 505,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 354,
+ 505,
+ 366
+ ],
+ "score": 1.0,
+ "content": "relatively few data per class (20), makes this an ideal data set for testing one-shot classification. In",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 365,
+ 505,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 365,
+ 121,
+ 377
+ ],
+ "score": 1.0,
+ "content": "the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 122,
+ 365,
+ 132,
+ 375
+ ],
+ "score": 0.78,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 365,
+ 271,
+ 377
+ ],
+ "score": 1.0,
+ "content": "-way Omniglot task setup we pick",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 271,
+ 365,
+ 282,
+ 375
+ ],
+ "score": 0.74,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 282,
+ 365,
+ 505,
+ 377
+ ],
+ "score": 1.0,
+ "content": "unseen character classes, independent of alphabet. We",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 506,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 436,
+ 389
+ ],
+ "score": 1.0,
+ "content": "provide the model with one drawing of each character and measure its accuracy the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 436,
+ 376,
+ 446,
+ 385
+ ],
+ "score": 0.82,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 447,
+ 375,
+ 506,
+ 389
+ ],
+ "score": 1.0,
+ "content": "-th time it sees",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 387,
+ 505,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 387,
+ 505,
+ 399
+ ],
+ "score": 1.0,
+ "content": "the character class. Our setup is identical to Oriol Vinyals (2016b), so we also augmented the data",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "set with random rotations by multiples of 90 degrees and use 1200 characters for training, and the",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 421
+ ],
+ "score": 1.0,
+ "content": "remaining character classes for evaluation. We present the results from Oriol Vinyals (2016b) and",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 496,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 496,
+ 432
+ ],
+ "score": 1.0,
+ "content": "ours in Table 1. Even with a simpler network without batch normalization, we get similar results.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 443,
+ 505,
+ 477
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "Synthetic task. To better understand the memory module operation and to test what it can remem-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 467
+ ],
+ "score": 1.0,
+ "content": "ber, we devise a synthetic task and train the Extended Neural GPU with and without memory (we",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 458,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 458,
+ 478
+ ],
+ "score": 1.0,
+ "content": "use a small Extended Neural GPU with 32 channels and memory of size half a million).",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 482,
+ 505,
+ 515
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 480,
+ 494
+ ],
+ "score": 1.0,
+ "content": "To create training and test data for our synthetic task, we use symbols from the set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 480,
+ 482,
+ 505,
+ 493
+ ],
+ "score": 0.87,
+ "content": "S \\_ =",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 107,
+ 492,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 493,
+ 170,
+ 505
+ ],
+ "score": 0.91,
+ "content": "\\{ 2 , \\ldots , 1 6 0 0 0 \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 170,
+ 492,
+ 294,
+ 506
+ ],
+ "score": 1.0,
+ "content": "and first fix a random function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 493,
+ 339,
+ 504
+ ],
+ "score": 0.92,
+ "content": "f : S S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 492,
+ 397,
+ 506
+ ],
+ "score": 1.0,
+ "content": ". The function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 493,
+ 405,
+ 504
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 405,
+ 492,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "is chosen at random, but",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 466,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 466,
+ 516
+ ],
+ "score": 1.0,
+ "content": "fixed and the same for all training and testing examples (we used 40K training examples).",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 520,
+ 505,
+ 587
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 519,
+ 506,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 519,
+ 506,
+ 534
+ ],
+ "score": 1.0,
+ "content": "In our synthetic task, the input is a sequence consisting of As and Bs with one continuous substring",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 530,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 530,
+ 203,
+ 546
+ ],
+ "score": 1.0,
+ "content": "of 7 digits from the set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 532,
+ 238,
+ 543
+ ],
+ "score": 0.85,
+ "content": "0 , 1 , 2 , 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 530,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": ". The substring is interpreted as a number written in base-4, e.g.,",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 542,
+ 506,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 178,
+ 554
+ ],
+ "score": 0.91,
+ "content": "1 9 8 2 = 1 3 2 3 3 2 _ { 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 179,
+ 542,
+ 506,
+ 555
+ ],
+ "score": 1.0,
+ "content": ", so the string 132332 would be interpreted as 1982. The corresponding output",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 554,
+ 506,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 476,
+ 566
+ ],
+ "score": 1.0,
+ "content": "is created by copying all As and Bs, but mapping the number through the random function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 476,
+ 554,
+ 483,
+ 565
+ ],
+ "score": 0.84,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 483,
+ 554,
+ 506,
+ 566
+ ],
+ "score": 1.0,
+ "content": ". For",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 187,
+ 577
+ ],
+ "score": 1.0,
+ "content": "instance, assuming",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 188,
+ 564,
+ 261,
+ 577
+ ],
+ "score": 0.91,
+ "content": "{ \\bar { f } } ( 1 9 8 2 ) = 3 7 2 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": ", the output corresponding to 132332 would be 322032 as",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 576,
+ 354,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 175,
+ 587
+ ],
+ "score": 0.91,
+ "content": "3 7 2 6 = 3 2 2 0 3 2 _ { 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 576,
+ 354,
+ 588
+ ],
+ "score": 1.0,
+ "content": ". Here is an example of an input-output pair:",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 167,
+ 597,
+ 445,
+ 622
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 167,
+ 597,
+ 445,
+ 622
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 167,
+ 597,
+ 445,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 167,
+ 597,
+ 445,
+ 622
+ ],
+ "score": 0.961,
+ "html": "| Input | A | 0 | 1 | 3 | 2 | 3 | 3 | 2 | B | A | B | A | B |
| Output | A | 0 | 3 | 2 | 2 | 0 | 3 | 2 | B | A | B | A | B |
",
+ "type": "table",
+ "image_path": "3c51d635a68755ef96260617b0c660691d11e3d7759a893cee9f1d7f22ce0f03.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 34,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 167,
+ 597,
+ 445,
+ 622
+ ],
+ "spans": [],
+ "index": 34
+ }
+ ]
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 631,
+ 505,
+ 676
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 630,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 630,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": "This task clearly requires memory to store the fixed random function. Since there are 16K elements",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 642,
+ 505,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 642,
+ 505,
+ 654
+ ],
+ "score": 1.0,
+ "content": "to learn, it is hard to memorize, and each single instance occurs quite rarely. The raw Extended",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "Neural GPU (or any other sequence-to-sequence model) are limited by their size. With long training,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 664,
+ 445,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 445,
+ 676
+ ],
+ "score": 1.0,
+ "content": "the small model can memorize some of the sequences, but it is only a small fraction.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 681,
+ 502,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 504,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 504,
+ 694
+ ],
+ "score": 1.0,
+ "content": "Additionally, there is no direct indication in the data what part of the input should trigger the pro-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 691,
+ 504,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 691,
+ 504,
+ 705
+ ],
+ "score": 1.0,
+ "content": "duction of each output symbol. For example, to produce the first 3 output in the above example, the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 39.5
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 105,
+ 712,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 708,
+ 505,
+ 725
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 708,
+ 505,
+ 725
+ ],
+ "score": 1.0,
+ "content": "1 https://github.com/tensorflow/models/tree/master/learning_to_remember_",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 105,
+ 722,
+ 169,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 722,
+ 169,
+ 732
+ ],
+ "score": 1.0,
+ "content": "rare_events",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "7",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 136,
+ 514,
+ 205
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 89,
+ 505,
+ 123
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "score": 1.0,
+ "content": "Table 1: Results on the Omniglot dataset. Although our model uses only a simple convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 100,
+ 505,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 505,
+ 113
+ ],
+ "score": 1.0,
+ "content": "neural network, the addition of our memory module allows it to approach much more complex",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 111,
+ 298,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 111,
+ 298,
+ 123
+ ],
+ "score": 1.0,
+ "content": "models on 1-shot and multi-shot learning tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 106,
+ 136,
+ 514,
+ 205
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 514,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 514,
+ 205
+ ],
+ "score": 0.979,
+ "html": "| Model | 5-way 1-shot | 5-way 5-shot | 20-way 1-shot | 20-way 5-shot |
| Pixels Nearest Neighbor | 41.7% | 63.2% | 26.7% | 42.6% |
| MANN (no convolutions) | 82.8% | 94.9% | 1 | 1 |
| Convolutional Siamese Net | 96.7% | 98.4% | 88.0% | 96.5% |
| Matching Network | 98.1% | 98.9% | 93.8% | 98.5% |
| ConvNet with Memory Module | 98.4% | 99.6% | 95.0% | 98.6% |
",
+ "type": "table",
+ "image_path": "3e84788e380c6c8f31c00e5fa93f09a01f2fb8880153858877a44dfe07b512aa.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 514,
+ 159.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 159.0,
+ 514,
+ 182.0
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 182.0,
+ 514,
+ 205.0
+ ],
+ "spans": [],
+ "index": 5
+ }
+ ]
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 229,
+ 200,
+ 241
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 201,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 201,
+ 243
+ ],
+ "score": 1.0,
+ "content": "4 EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 253,
+ 505,
+ 320
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "We perform experiments using all three architectures described above. We experiment both on real-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 264,
+ 506,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 264,
+ 506,
+ 277
+ ],
+ "score": 1.0,
+ "content": "world data and on synthetic tasks that give us some insight into the performance and limitations of",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": "the memory module. In all our experiments we use the Adam optimizer (Kingma & Ba, 2014) and",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 505,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 344,
+ 299
+ ],
+ "score": 1.0,
+ "content": "the parameters for the memory module remain unchanged",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 286,
+ 421,
+ 298
+ ],
+ "score": 0.33,
+ "content": "( k = 2 5 6 , \\alpha = 0 . 1 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 286,
+ 505,
+ 299
+ ],
+ "score": 1.0,
+ "content": ". Good performance",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 310
+ ],
+ "score": 1.0,
+ "content": "with a single set of parameters shows the versatility of our memory module. The source code for the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 308,
+ 432,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 432,
+ 321
+ ],
+ "score": 1.0,
+ "content": "memory module, together with our settings for Omniglot, is available on github1.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 105,
+ 253,
+ 506,
+ 321
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 331,
+ 505,
+ 431
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 343
+ ],
+ "score": 1.0,
+ "content": "Omniglot. The Omniglot dataset (Lake et al., 2011) consists of 1623 characters from 50 different",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 505,
+ 355
+ ],
+ "score": 1.0,
+ "content": "alphabets, each hand-drawn by 20 different people. The large number of classes (characters) with",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 354,
+ 505,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 354,
+ 505,
+ 366
+ ],
+ "score": 1.0,
+ "content": "relatively few data per class (20), makes this an ideal data set for testing one-shot classification. In",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 365,
+ 505,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 365,
+ 121,
+ 377
+ ],
+ "score": 1.0,
+ "content": "the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 122,
+ 365,
+ 132,
+ 375
+ ],
+ "score": 0.78,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 365,
+ 271,
+ 377
+ ],
+ "score": 1.0,
+ "content": "-way Omniglot task setup we pick",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 271,
+ 365,
+ 282,
+ 375
+ ],
+ "score": 0.74,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 282,
+ 365,
+ 505,
+ 377
+ ],
+ "score": 1.0,
+ "content": "unseen character classes, independent of alphabet. We",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 506,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 436,
+ 389
+ ],
+ "score": 1.0,
+ "content": "provide the model with one drawing of each character and measure its accuracy the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 436,
+ 376,
+ 446,
+ 385
+ ],
+ "score": 0.82,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 447,
+ 375,
+ 506,
+ 389
+ ],
+ "score": 1.0,
+ "content": "-th time it sees",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 387,
+ 505,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 387,
+ 505,
+ 399
+ ],
+ "score": 1.0,
+ "content": "the character class. Our setup is identical to Oriol Vinyals (2016b), so we also augmented the data",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "set with random rotations by multiples of 90 degrees and use 1200 characters for training, and the",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 421
+ ],
+ "score": 1.0,
+ "content": "remaining character classes for evaluation. We present the results from Oriol Vinyals (2016b) and",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 496,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 496,
+ 432
+ ],
+ "score": 1.0,
+ "content": "ours in Table 1. Even with a simpler network without batch normalization, we get similar results.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 105,
+ 333,
+ 506,
+ 432
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 443,
+ 505,
+ 477
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "Synthetic task. To better understand the memory module operation and to test what it can remem-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 467
+ ],
+ "score": 1.0,
+ "content": "ber, we devise a synthetic task and train the Extended Neural GPU with and without memory (we",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 458,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 458,
+ 478
+ ],
+ "score": 1.0,
+ "content": "use a small Extended Neural GPU with 32 channels and memory of size half a million).",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 105,
+ 442,
+ 505,
+ 478
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 482,
+ 505,
+ 515
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 480,
+ 494
+ ],
+ "score": 1.0,
+ "content": "To create training and test data for our synthetic task, we use symbols from the set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 480,
+ 482,
+ 505,
+ 493
+ ],
+ "score": 0.87,
+ "content": "S \\_ =",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 107,
+ 492,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 493,
+ 170,
+ 505
+ ],
+ "score": 0.91,
+ "content": "\\{ 2 , \\ldots , 1 6 0 0 0 \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 170,
+ 492,
+ 294,
+ 506
+ ],
+ "score": 1.0,
+ "content": "and first fix a random function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 493,
+ 339,
+ 504
+ ],
+ "score": 0.92,
+ "content": "f : S S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 492,
+ 397,
+ 506
+ ],
+ "score": 1.0,
+ "content": ". The function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 493,
+ 405,
+ 504
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 405,
+ 492,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "is chosen at random, but",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 466,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 466,
+ 516
+ ],
+ "score": 1.0,
+ "content": "fixed and the same for all training and testing examples (we used 40K training examples).",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 105,
+ 482,
+ 506,
+ 516
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 520,
+ 505,
+ 587
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 519,
+ 506,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 519,
+ 506,
+ 534
+ ],
+ "score": 1.0,
+ "content": "In our synthetic task, the input is a sequence consisting of As and Bs with one continuous substring",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 530,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 530,
+ 203,
+ 546
+ ],
+ "score": 1.0,
+ "content": "of 7 digits from the set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 532,
+ 238,
+ 543
+ ],
+ "score": 0.85,
+ "content": "0 , 1 , 2 , 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 530,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": ". The substring is interpreted as a number written in base-4, e.g.,",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 542,
+ 506,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 178,
+ 554
+ ],
+ "score": 0.91,
+ "content": "1 9 8 2 = 1 3 2 3 3 2 _ { 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 179,
+ 542,
+ 506,
+ 555
+ ],
+ "score": 1.0,
+ "content": ", so the string 132332 would be interpreted as 1982. The corresponding output",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 554,
+ 506,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 476,
+ 566
+ ],
+ "score": 1.0,
+ "content": "is created by copying all As and Bs, but mapping the number through the random function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 476,
+ 554,
+ 483,
+ 565
+ ],
+ "score": 0.84,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 483,
+ 554,
+ 506,
+ 566
+ ],
+ "score": 1.0,
+ "content": ". For",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 187,
+ 577
+ ],
+ "score": 1.0,
+ "content": "instance, assuming",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 188,
+ 564,
+ 261,
+ 577
+ ],
+ "score": 0.91,
+ "content": "{ \\bar { f } } ( 1 9 8 2 ) = 3 7 2 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": ", the output corresponding to 132332 would be 322032 as",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 576,
+ 354,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 175,
+ 587
+ ],
+ "score": 0.91,
+ "content": "3 7 2 6 = 3 2 2 0 3 2 _ { 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 576,
+ 354,
+ 588
+ ],
+ "score": 1.0,
+ "content": ". Here is an example of an input-output pair:",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 104,
+ 519,
+ 506,
+ 588
+ ]
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 167,
+ 597,
+ 445,
+ 622
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 167,
+ 597,
+ 445,
+ 622
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 167,
+ 597,
+ 445,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 167,
+ 597,
+ 445,
+ 622
+ ],
+ "score": 0.961,
+ "html": "| Input | A | 0 | 1 | 3 | 2 | 3 | 3 | 2 | B | A | B | A | B |
| Output | A | 0 | 3 | 2 | 2 | 0 | 3 | 2 | B | A | B | A | B |
",
+ "type": "table",
+ "image_path": "3c51d635a68755ef96260617b0c660691d11e3d7759a893cee9f1d7f22ce0f03.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 34,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 167,
+ 597,
+ 445,
+ 622
+ ],
+ "spans": [],
+ "index": 34
+ }
+ ]
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 631,
+ 505,
+ 676
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 630,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 630,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": "This task clearly requires memory to store the fixed random function. Since there are 16K elements",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 642,
+ 505,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 642,
+ 505,
+ 654
+ ],
+ "score": 1.0,
+ "content": "to learn, it is hard to memorize, and each single instance occurs quite rarely. The raw Extended",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "Neural GPU (or any other sequence-to-sequence model) are limited by their size. With long training,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 664,
+ 445,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 445,
+ 676
+ ],
+ "score": 1.0,
+ "content": "the small model can memorize some of the sequences, but it is only a small fraction.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 105,
+ 630,
+ 505,
+ 676
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 681,
+ 502,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 504,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 504,
+ 694
+ ],
+ "score": 1.0,
+ "content": "Additionally, there is no direct indication in the data what part of the input should trigger the pro-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 691,
+ 504,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 691,
+ 504,
+ 705
+ ],
+ "score": 1.0,
+ "content": "duction of each output symbol. For example, to produce the first 3 output in the above example, the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 39.5,
+ "bbox_fs": [
+ 106,
+ 680,
+ 504,
+ 705
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 169,
+ 115,
+ 441,
+ 186
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 89,
+ 503,
+ 112
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "score": 1.0,
+ "content": "Table 2: Results on the synthetic task. We report the percentage of fully correct sequences from the",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 100,
+ 384,
+ 112
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 384,
+ 112
+ ],
+ "score": 1.0,
+ "content": "test set, which contains 10000 random examples. See text for details.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 169,
+ 115,
+ 441,
+ 186
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 169,
+ 115,
+ 441,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 169,
+ 115,
+ 441,
+ 186
+ ],
+ "score": 0.98,
+ "html": "| Model | Accuracy |
| HammingNearestNeighborBaseline Sequence-to-Sequence with AttentionBaseline Extended Neural GPU | 0.1%0.9%12.2% |
| Sequence-to-Sequence with Attention and MemoryExtended Neural GPU with Memory Module | 35.2%71.3% |
",
+ "type": "table",
+ "image_path": "18a08bfd3c22f4560cc91e2fa7269939fec5786838ff6eb450649973f936c0e8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 169,
+ 115,
+ 441,
+ 138.66666666666666
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 169,
+ 138.66666666666666,
+ 441,
+ 162.33333333333331
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 169,
+ 162.33333333333331,
+ 441,
+ 185.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ }
+ ],
+ "index": 1.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 211,
+ 505,
+ 257
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 210,
+ 505,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 210,
+ 505,
+ 224
+ ],
+ "score": 1.0,
+ "content": "Table 3: Results on the WMT En-De task. As described in the text, we split the test set in two",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 223,
+ 504,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 504,
+ 235
+ ],
+ "score": 1.0,
+ "content": "(odd lines and even lines) to evaluate the model on one-shot learning. Given the even test set, the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 234,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "model can perform better on the odd test set. We also see a dramatic improvement when the model",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 245,
+ 484,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 245,
+ 484,
+ 257
+ ],
+ "score": 1.0,
+ "content": "is provided with the whole test set, validating that the memory module is working as intended.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 141,
+ 261,
+ 469,
+ 320
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 141,
+ 261,
+ 469,
+ 320
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 261,
+ 469,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 261,
+ 469,
+ 320
+ ],
+ "score": 0.977,
+ "html": "| Model | Full Test | Odd Test |
| GNMT | 23.25 | 23.17 |
| GNMT withMemoryModule | 23.29 | 23.16 |
| GNMTwithMemoryModule andEven Testcontext | 1 | 23.60 |
| GNMT with Memory Module and Whole Test context | 31.11* | 1 |
",
+ "type": "table",
+ "image_path": "079f9dac90fea47a116c7b9245994f9ad491792a80f2179ec00f5037178d0b4f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 141,
+ 261,
+ 469,
+ 280.6666666666667
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 280.6666666666667,
+ 469,
+ 300.33333333333337
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 300.33333333333337,
+ 469,
+ 320.00000000000006
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 347,
+ 505,
+ 468
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 348,
+ 505,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 348,
+ 505,
+ 360
+ ],
+ "score": 1.0,
+ "content": "memory key needs to encode all base-4 symbols from the input. Not just one or two aligned sym-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 358,
+ 506,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 506,
+ 371
+ ],
+ "score": 1.0,
+ "content": "bols, but a number of them. Moreover, it should not encode more symbols or it will not generalize",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 506,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 506,
+ 382
+ ],
+ "score": 1.0,
+ "content": "to the test set. Similarly, a basic nearest neighbor classifier fails on this task. We use sequences of",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 104,
+ 379,
+ 506,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 379,
+ 506,
+ 394
+ ],
+ "score": 1.0,
+ "content": "length up to 40 during training, but there are only 7 relevant symbols. The simple nearest neighbor",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 104,
+ 390,
+ 506,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 390,
+ 506,
+ 404
+ ],
+ "score": 1.0,
+ "content": "by Hamming distance will most probably select some sequence with similar prefix or suffix of As",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 415
+ ],
+ "score": 1.0,
+ "content": "and Bs, and not the one with the corresponding base-4 part. We also trained a large sequence-to-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 413,
+ 505,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 505,
+ 426
+ ],
+ "score": 1.0,
+ "content": "sequence model with attention on this task (a 2-layer LSTM model with 256 units in each layer).",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "This model can memorize the whole training set, but it suffers from a similar problem as the Ham-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 486,
+ 447
+ ],
+ "score": 1.0,
+ "content": "ming nearest neighbor – it almost doesn’t generalize, its accuracy on the test set is only about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 487,
+ 435,
+ 501,
+ 446
+ ],
+ "score": 0.86,
+ "content": "1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 435,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 426,
+ 459
+ ],
+ "score": 1.0,
+ "content": "The same model with a memory module generalizes much better, reaching over",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 426,
+ 446,
+ 446,
+ 457
+ ],
+ "score": 0.89,
+ "content": "3 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 446,
+ 446,
+ 505,
+ 459
+ ],
+ "score": 1.0,
+ "content": "accuracy. The",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 456,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 456,
+ 469
+ ],
+ "score": 1.0,
+ "content": "Extended Neural GPU with our memory module yields even better results, see Table 2.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 483,
+ 505,
+ 516
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 496
+ ],
+ "score": 1.0,
+ "content": "Translation. To evaluate the memory module in a large-scale setting we use the GNMT",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "model (Wu et al., 2016) extended with our memory module on the WMT14 English-to-German",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 413,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 413,
+ 518
+ ],
+ "score": 1.0,
+ "content": "translation task. We evaluate the model both qualitatively and quantitatively.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 523,
+ 505,
+ 555
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "On the qualitative side, we note that our memory-augmented model can successfully translate rare",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 533,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "words like Dostoevsky, unlike the baseline model which predicts an identity-mapped Dostoevsky for",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 544,
+ 262,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 262,
+ 557
+ ],
+ "score": 1.0,
+ "content": "the German translation of Dostoevsky.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 561,
+ 503,
+ 583
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 505,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 505,
+ 575
+ ],
+ "score": 1.0,
+ "content": "On the quantitative side, we use the WMT test set. We find that in terms of BLEU score, an aggregate",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 572,
+ 465,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 465,
+ 584
+ ],
+ "score": 1.0,
+ "content": "measure, the memory-augmented GNMT is on par with the baseline GNMT, see Table 3.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 589,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "To evaluate our memory-augmented model for one-shot capabilities we split the test set in two. We",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 600,
+ 506,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 506,
+ 613
+ ],
+ "score": 1.0,
+ "content": "take the even lines of the test set (index starting at 0) as a context set and the odd lines of the test set",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 609,
+ 507,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 609,
+ 507,
+ 626
+ ],
+ "score": 1.0,
+ "content": "as the one-shot evaluation set. While showing the context set to the model, no additional training",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "occurs, only memory updates are allowed. So the weights of the model do not change, but the",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "score": 1.0,
+ "content": "memory does. Since the sentences in the test set are highly-correlated to each other (they come from",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 505,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 505,
+ 656
+ ],
+ "score": 1.0,
+ "content": "paragraphs with preserved order), we expect that if we allow a one-shot capable model to use the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 653,
+ 505,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 653,
+ 505,
+ 669
+ ],
+ "score": 1.0,
+ "content": "context set to update its memory and then evaluate it on the other half of the test set, its accuracy",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 664,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "will increase. For our GNMT with memory model, we passed the context set through the memory",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "update operations 3 times. As seen in Table 3, the context set indeed helps when evaluating on the",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "odd lines, increasing the BLEU score by almost 0.5. As further indication that our memory module",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "works properly, we also evaluate the model after showing the whole test set as a context set. Note",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "that this is essentially an oracle: the memory module gets to see all the correct answers, we do this",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 484,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 484,
+ 734
+ ],
+ "score": 1.0,
+ "content": "only to test and debug. As expected, this increases BLEU score dramatically, by over 8 points.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 750,
+ 309,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 750,
+ 309,
+ 761
+ ],
+ "score": 1.0,
+ "content": "8",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 169,
+ 115,
+ 441,
+ 186
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 89,
+ 503,
+ 112
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "score": 1.0,
+ "content": "Table 2: Results on the synthetic task. We report the percentage of fully correct sequences from the",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 100,
+ 384,
+ 112
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 384,
+ 112
+ ],
+ "score": 1.0,
+ "content": "test set, which contains 10000 random examples. See text for details.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 169,
+ 115,
+ 441,
+ 186
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 169,
+ 115,
+ 441,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 169,
+ 115,
+ 441,
+ 186
+ ],
+ "score": 0.98,
+ "html": "| Model | Accuracy |
| HammingNearestNeighborBaseline Sequence-to-Sequence with AttentionBaseline Extended Neural GPU | 0.1%0.9%12.2% |
| Sequence-to-Sequence with Attention and MemoryExtended Neural GPU with Memory Module | 35.2%71.3% |
",
+ "type": "table",
+ "image_path": "18a08bfd3c22f4560cc91e2fa7269939fec5786838ff6eb450649973f936c0e8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 169,
+ 115,
+ 441,
+ 138.66666666666666
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 169,
+ 138.66666666666666,
+ 441,
+ 162.33333333333331
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 169,
+ 162.33333333333331,
+ 441,
+ 185.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ }
+ ],
+ "index": 1.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 211,
+ 505,
+ 257
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 210,
+ 505,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 210,
+ 505,
+ 224
+ ],
+ "score": 1.0,
+ "content": "Table 3: Results on the WMT En-De task. As described in the text, we split the test set in two",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 223,
+ 504,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 504,
+ 235
+ ],
+ "score": 1.0,
+ "content": "(odd lines and even lines) to evaluate the model on one-shot learning. Given the even test set, the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 234,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "model can perform better on the odd test set. We also see a dramatic improvement when the model",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 245,
+ 484,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 245,
+ 484,
+ 257
+ ],
+ "score": 1.0,
+ "content": "is provided with the whole test set, validating that the memory module is working as intended.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 105,
+ 210,
+ 505,
+ 257
+ ]
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 141,
+ 261,
+ 469,
+ 320
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 141,
+ 261,
+ 469,
+ 320
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 261,
+ 469,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 261,
+ 469,
+ 320
+ ],
+ "score": 0.977,
+ "html": "| Model | Full Test | Odd Test |
| GNMT | 23.25 | 23.17 |
| GNMT withMemoryModule | 23.29 | 23.16 |
| GNMTwithMemoryModule andEven Testcontext | 1 | 23.60 |
| GNMT with Memory Module and Whole Test context | 31.11* | 1 |
",
+ "type": "table",
+ "image_path": "079f9dac90fea47a116c7b9245994f9ad491792a80f2179ec00f5037178d0b4f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 141,
+ 261,
+ 469,
+ 280.6666666666667
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 280.6666666666667,
+ 469,
+ 300.33333333333337
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 300.33333333333337,
+ 469,
+ 320.00000000000006
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 347,
+ 505,
+ 468
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 348,
+ 505,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 348,
+ 505,
+ 360
+ ],
+ "score": 1.0,
+ "content": "memory key needs to encode all base-4 symbols from the input. Not just one or two aligned sym-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 358,
+ 506,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 506,
+ 371
+ ],
+ "score": 1.0,
+ "content": "bols, but a number of them. Moreover, it should not encode more symbols or it will not generalize",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 506,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 506,
+ 382
+ ],
+ "score": 1.0,
+ "content": "to the test set. Similarly, a basic nearest neighbor classifier fails on this task. We use sequences of",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 104,
+ 379,
+ 506,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 379,
+ 506,
+ 394
+ ],
+ "score": 1.0,
+ "content": "length up to 40 during training, but there are only 7 relevant symbols. The simple nearest neighbor",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 104,
+ 390,
+ 506,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 390,
+ 506,
+ 404
+ ],
+ "score": 1.0,
+ "content": "by Hamming distance will most probably select some sequence with similar prefix or suffix of As",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 415
+ ],
+ "score": 1.0,
+ "content": "and Bs, and not the one with the corresponding base-4 part. We also trained a large sequence-to-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 413,
+ 505,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 505,
+ 426
+ ],
+ "score": 1.0,
+ "content": "sequence model with attention on this task (a 2-layer LSTM model with 256 units in each layer).",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "This model can memorize the whole training set, but it suffers from a similar problem as the Ham-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 486,
+ 447
+ ],
+ "score": 1.0,
+ "content": "ming nearest neighbor – it almost doesn’t generalize, its accuracy on the test set is only about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 487,
+ 435,
+ 501,
+ 446
+ ],
+ "score": 0.86,
+ "content": "1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 435,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 426,
+ 459
+ ],
+ "score": 1.0,
+ "content": "The same model with a memory module generalizes much better, reaching over",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 426,
+ 446,
+ 446,
+ 457
+ ],
+ "score": 0.89,
+ "content": "3 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 446,
+ 446,
+ 505,
+ 459
+ ],
+ "score": 1.0,
+ "content": "accuracy. The",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 456,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 456,
+ 469
+ ],
+ "score": 1.0,
+ "content": "Extended Neural GPU with our memory module yields even better results, see Table 2.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 104,
+ 348,
+ 506,
+ 469
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 483,
+ 505,
+ 516
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 496
+ ],
+ "score": 1.0,
+ "content": "Translation. To evaluate the memory module in a large-scale setting we use the GNMT",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "model (Wu et al., 2016) extended with our memory module on the WMT14 English-to-German",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 413,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 413,
+ 518
+ ],
+ "score": 1.0,
+ "content": "translation task. We evaluate the model both qualitatively and quantitatively.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 105,
+ 483,
+ 505,
+ 518
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 523,
+ 505,
+ 555
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "On the qualitative side, we note that our memory-augmented model can successfully translate rare",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 533,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "words like Dostoevsky, unlike the baseline model which predicts an identity-mapped Dostoevsky for",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 544,
+ 262,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 262,
+ 557
+ ],
+ "score": 1.0,
+ "content": "the German translation of Dostoevsky.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 105,
+ 522,
+ 505,
+ 557
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 561,
+ 503,
+ 583
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 505,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 505,
+ 575
+ ],
+ "score": 1.0,
+ "content": "On the quantitative side, we use the WMT test set. We find that in terms of BLEU score, an aggregate",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 572,
+ 465,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 465,
+ 584
+ ],
+ "score": 1.0,
+ "content": "measure, the memory-augmented GNMT is on par with the baseline GNMT, see Table 3.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 106,
+ 560,
+ 505,
+ 584
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 589,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "To evaluate our memory-augmented model for one-shot capabilities we split the test set in two. We",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 600,
+ 506,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 506,
+ 613
+ ],
+ "score": 1.0,
+ "content": "take the even lines of the test set (index starting at 0) as a context set and the odd lines of the test set",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 609,
+ 507,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 609,
+ 507,
+ 626
+ ],
+ "score": 1.0,
+ "content": "as the one-shot evaluation set. While showing the context set to the model, no additional training",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "occurs, only memory updates are allowed. So the weights of the model do not change, but the",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "score": 1.0,
+ "content": "memory does. Since the sentences in the test set are highly-correlated to each other (they come from",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 505,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 505,
+ 656
+ ],
+ "score": 1.0,
+ "content": "paragraphs with preserved order), we expect that if we allow a one-shot capable model to use the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 653,
+ 505,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 653,
+ 505,
+ 669
+ ],
+ "score": 1.0,
+ "content": "context set to update its memory and then evaluate it on the other half of the test set, its accuracy",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 664,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "will increase. For our GNMT with memory model, we passed the context set through the memory",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "update operations 3 times. As seen in Table 3, the context set indeed helps when evaluating on the",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "odd lines, increasing the BLEU score by almost 0.5. As further indication that our memory module",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "works properly, we also evaluate the model after showing the whole test set as a context set. Note",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "that this is essentially an oracle: the memory module gets to see all the correct answers, we do this",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 484,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 484,
+ 734
+ ],
+ "score": 1.0,
+ "content": "only to test and debug. As expected, this increases BLEU score dramatically, by over 8 points.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 104,
+ 588,
+ 507,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 81,
+ 190,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 192,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 192,
+ 96
+ ],
+ "score": 1.0,
+ "content": "5 DISCUSSION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 105,
+ 505,
+ 172
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "We presented a long-term memory module that can be used for life-long learning. It is versatile, so it",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "can be added to different deep learning models and at different layers to give the networks one-shot",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 505,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 505,
+ 140
+ ],
+ "score": 1.0,
+ "content": "learning capability. Several parts of the presented memory module could be tuned and studied in",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 138,
+ 505,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 505,
+ 151
+ ],
+ "score": 1.0,
+ "content": "more detail. The update rule that averages the query with the correct key could be parametrized.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 150,
+ 505,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 150,
+ 505,
+ 161
+ ],
+ "score": 1.0,
+ "content": "Instead of returning only the single nearest neighbor we could also return a number of them to be",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 453,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 453,
+ 173
+ ],
+ "score": 1.0,
+ "content": "processed by other layers of the network. We leave these questions for future research.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 190
+ ],
+ "score": 1.0,
+ "content": "The main issue we encountered, though, is that evaluating one-shot learning is difficult, as standard",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 188,
+ 505,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": "metrics do not focus on this scenario. In this work, we adapted the standard metrics to investigate",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 212
+ ],
+ "score": 1.0,
+ "content": "our approach. For example, in the translation task we used half of the test set as context for the other",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 211,
+ 504,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 211,
+ 504,
+ 222
+ ],
+ "score": 1.0,
+ "content": "half, and we still report the standard BLEU score. This allows us to show that our module works,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 221,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 221,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "but it is only a temporary solution. Better metrics are needed to accelerate progress of one-shot and",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 104,
+ 231,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 231,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "life-long learning. Thus, we consider the present work as just a first step on the way to making deep",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 349,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 349,
+ 255
+ ],
+ "score": 1.0,
+ "content": "models learn to remember rare events through their lifetime.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 271,
+ 175,
+ 283
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 270,
+ 176,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 270,
+ 176,
+ 285
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 282,
+ 506,
+ 738
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "A. Andoni and P. Indyk. Near-optimal hashing algorithms for approximate nearest neighbor in",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 300,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 300,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "high dimensions. In 2006 47th Annual IEEE Symposium on Foundations of Computer Science",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 312,
+ 382,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 312,
+ 382,
+ 324
+ ],
+ "score": 1.0,
+ "content": "(FOCS’06), pp. 459–468, Oct 2006. doi: 10.1109/FOCS.2006.49.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 104,
+ 329,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 329,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 114,
+ 340,
+ 504,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 340,
+ 504,
+ 354
+ ],
+ "score": 1.0,
+ "content": "learning to align and translate. CoRR, abs/1409.0473, 2014. URL http://arxiv.org/",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 117,
+ 353,
+ 199,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 353,
+ 199,
+ 364
+ ],
+ "score": 1.0,
+ "content": "abs/1409.0473.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 104,
+ 369,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 369,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "Antoine Bordes, Nicolas Usunier, Sumit Chopra, and Jason Weston. Large-scale simple question",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 114,
+ 381,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 381,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": "answering with memory networks. CoRR, abs/1506.02075, 2015. URL http://arxiv.org/",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 117,
+ 393,
+ 205,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 393,
+ 205,
+ 404
+ ],
+ "score": 1.0,
+ "content": "abs/1506.02075.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 410,
+ 504,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 504,
+ 423
+ ],
+ "score": 1.0,
+ "content": "Sarath Chandar, Sungjin Ahn, Hugo Larochelle, Pascal Vincent, Gerald Tesauro, and Yoshua Ben-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 115,
+ 421,
+ 428,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 421,
+ 428,
+ 434
+ ],
+ "score": 1.0,
+ "content": "gio. Hierarchical memory networks. arXiv preprint arXiv:1605.07427, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 449,
+ 505,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 449,
+ 505,
+ 464
+ ],
+ "score": 1.0,
+ "content": "Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical ma-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 114,
+ 460,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 460,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "chine translation. CoRR, abs/1406.1078, 2014. URL http://arxiv.org/abs/1406.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 473,
+ 146,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 473,
+ 146,
+ 484
+ ],
+ "score": 1.0,
+ "content": "1078.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 489,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 489,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Trans.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 501,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 501,
+ 505,
+ 514
+ ],
+ "score": 1.0,
+ "content": "Pattern Anal. Mach. Intell., 28(4):594–611, April 2006. ISSN 0162-8828. doi: 10.1109/TPAMI.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 116,
+ 513,
+ 416,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 513,
+ 416,
+ 525
+ ],
+ "score": 1.0,
+ "content": "2006.79. URL http://dx.doi.org/10.1109/TPAMI.2006.79.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 530,
+ 506,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 530,
+ 506,
+ 545
+ ],
+ "score": 1.0,
+ "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. CoRR, abs/1410.5401,",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 115,
+ 541,
+ 349,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 541,
+ 349,
+ 555
+ ],
+ "score": 1.0,
+ "content": "2014. URL http://arxiv.org/abs/1410.5401.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "C¸ aglar Gulc¸ehre, Sarath Chandar, Kyunghyun Cho, and Yoshua Bengio. Dynamic neural turing ¨",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 115,
+ 571,
+ 434,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 571,
+ 434,
+ 584
+ ],
+ "score": 1.0,
+ "content": "machine with soft and hard addressing schemes. CoRR, abs/1607.00036, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 103,
+ 587,
+ 506,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 587,
+ 506,
+ 604
+ ],
+ "score": 1.0,
+ "content": "Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The goldilocks principle: Reading",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 114,
+ 599,
+ 462,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 599,
+ 462,
+ 614
+ ],
+ "score": 1.0,
+ "content": "children’s books with explicit memory representations. CoRR, abs/1511.02301, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 632
+ ],
+ "score": 1.0,
+ "content": "Geoffrey Hinton, Li Deng, Dong Yu, George Dahl, Abdelrahman Mohamed, Navdeep Jaitly, An-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 114,
+ 628,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 628,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": "drew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara Sainath, and Brian Kingsbury. Deep",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 114,
+ 640,
+ 505,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 640,
+ 505,
+ 654
+ ],
+ "score": 1.0,
+ "content": "neural networks for acoustic modeling in speech recognition: The shared views of four research",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 651,
+ 372,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 651,
+ 372,
+ 665
+ ],
+ "score": 1.0,
+ "content": "groups. IEEE Signal Processing Magazine, 29(6):82–97, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 669,
+ 505,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 669,
+ 505,
+ 684
+ ],
+ "score": 1.0,
+ "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 116,
+ 680,
+ 191,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 680,
+ 191,
+ 693
+ ],
+ "score": 1.0,
+ "content": "1735–1780, 1997.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Piotr Indyk and Rajeev Motwani. Approximate nearest neighbors: towards removing the curse of",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 114,
+ 708,
+ 505,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 708,
+ 505,
+ 724
+ ],
+ "score": 1.0,
+ "content": "dimensionality. In Proceedings of the thirtieth annual ACM symposium on Theory of computing,",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 223,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 223,
+ 732
+ ],
+ "score": 1.0,
+ "content": "pp. 604–613. ACM, 1998.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "page_idx": 8,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "9",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 81,
+ 190,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 192,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 192,
+ 96
+ ],
+ "score": 1.0,
+ "content": "5 DISCUSSION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 105,
+ 505,
+ 172
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "We presented a long-term memory module that can be used for life-long learning. It is versatile, so it",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "can be added to different deep learning models and at different layers to give the networks one-shot",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 505,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 505,
+ 140
+ ],
+ "score": 1.0,
+ "content": "learning capability. Several parts of the presented memory module could be tuned and studied in",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 138,
+ 505,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 505,
+ 151
+ ],
+ "score": 1.0,
+ "content": "more detail. The update rule that averages the query with the correct key could be parametrized.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 150,
+ 505,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 150,
+ 505,
+ 161
+ ],
+ "score": 1.0,
+ "content": "Instead of returning only the single nearest neighbor we could also return a number of them to be",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 453,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 453,
+ 173
+ ],
+ "score": 1.0,
+ "content": "processed by other layers of the network. We leave these questions for future research.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 105,
+ 106,
+ 506,
+ 173
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 190
+ ],
+ "score": 1.0,
+ "content": "The main issue we encountered, though, is that evaluating one-shot learning is difficult, as standard",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 188,
+ 505,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": "metrics do not focus on this scenario. In this work, we adapted the standard metrics to investigate",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 212
+ ],
+ "score": 1.0,
+ "content": "our approach. For example, in the translation task we used half of the test set as context for the other",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 211,
+ 504,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 211,
+ 504,
+ 222
+ ],
+ "score": 1.0,
+ "content": "half, and we still report the standard BLEU score. This allows us to show that our module works,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 221,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 221,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "but it is only a temporary solution. Better metrics are needed to accelerate progress of one-shot and",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 104,
+ 231,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 231,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "life-long learning. Thus, we consider the present work as just a first step on the way to making deep",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 349,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 349,
+ 255
+ ],
+ "score": 1.0,
+ "content": "models learn to remember rare events through their lifetime.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 104,
+ 177,
+ 505,
+ 255
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 271,
+ 175,
+ 283
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 270,
+ 176,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 270,
+ 176,
+ 285
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 105,
+ 282,
+ 506,
+ 738
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "A. Andoni and P. Indyk. Near-optimal hashing algorithms for approximate nearest neighbor in",
+ "type": "text"
+ }
+ ],
+ "index": 15,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 300,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 300,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "high dimensions. In 2006 47th Annual IEEE Symposium on Foundations of Computer Science",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 312,
+ 382,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 312,
+ 382,
+ 324
+ ],
+ "score": 1.0,
+ "content": "(FOCS’06), pp. 459–468, Oct 2006. doi: 10.1109/FOCS.2006.49.",
+ "type": "text"
+ }
+ ],
+ "index": 17,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 329,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 329,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly",
+ "type": "text"
+ }
+ ],
+ "index": 18,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 340,
+ 504,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 340,
+ 504,
+ 354
+ ],
+ "score": 1.0,
+ "content": "learning to align and translate. CoRR, abs/1409.0473, 2014. URL http://arxiv.org/",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 117,
+ 353,
+ 199,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 353,
+ 199,
+ 364
+ ],
+ "score": 1.0,
+ "content": "abs/1409.0473.",
+ "type": "text"
+ }
+ ],
+ "index": 20,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 369,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 369,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "Antoine Bordes, Nicolas Usunier, Sumit Chopra, and Jason Weston. Large-scale simple question",
+ "type": "text"
+ }
+ ],
+ "index": 21,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 381,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 381,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": "answering with memory networks. CoRR, abs/1506.02075, 2015. URL http://arxiv.org/",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 117,
+ 393,
+ 205,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 393,
+ 205,
+ 404
+ ],
+ "score": 1.0,
+ "content": "abs/1506.02075.",
+ "type": "text"
+ }
+ ],
+ "index": 23,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 410,
+ 504,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 504,
+ 423
+ ],
+ "score": 1.0,
+ "content": "Sarath Chandar, Sungjin Ahn, Hugo Larochelle, Pascal Vincent, Gerald Tesauro, and Yoshua Ben-",
+ "type": "text"
+ }
+ ],
+ "index": 24,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 421,
+ 428,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 421,
+ 428,
+ 434
+ ],
+ "score": 1.0,
+ "content": "gio. Hierarchical memory networks. arXiv preprint arXiv:1605.07427, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 25,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and",
+ "type": "text"
+ }
+ ],
+ "index": 26,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 449,
+ 505,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 449,
+ 505,
+ 464
+ ],
+ "score": 1.0,
+ "content": "Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical ma-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 114,
+ 460,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 460,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "chine translation. CoRR, abs/1406.1078, 2014. URL http://arxiv.org/abs/1406.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 473,
+ 146,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 473,
+ 146,
+ 484
+ ],
+ "score": 1.0,
+ "content": "1078.",
+ "type": "text"
+ }
+ ],
+ "index": 29,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 104,
+ 489,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 489,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Trans.",
+ "type": "text"
+ }
+ ],
+ "index": 30,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 501,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 501,
+ 505,
+ 514
+ ],
+ "score": 1.0,
+ "content": "Pattern Anal. Mach. Intell., 28(4):594–611, April 2006. ISSN 0162-8828. doi: 10.1109/TPAMI.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 116,
+ 513,
+ 416,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 513,
+ 416,
+ 525
+ ],
+ "score": 1.0,
+ "content": "2006.79. URL http://dx.doi.org/10.1109/TPAMI.2006.79.",
+ "type": "text"
+ }
+ ],
+ "index": 32,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 530,
+ 506,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 530,
+ 506,
+ 545
+ ],
+ "score": 1.0,
+ "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. CoRR, abs/1410.5401,",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 541,
+ 349,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 541,
+ 349,
+ 555
+ ],
+ "score": 1.0,
+ "content": "2014. URL http://arxiv.org/abs/1410.5401.",
+ "type": "text"
+ }
+ ],
+ "index": 34,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "C¸ aglar Gulc¸ehre, Sarath Chandar, Kyunghyun Cho, and Yoshua Bengio. Dynamic neural turing ¨",
+ "type": "text"
+ }
+ ],
+ "index": 35,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 571,
+ 434,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 571,
+ 434,
+ 584
+ ],
+ "score": 1.0,
+ "content": "machine with soft and hard addressing schemes. CoRR, abs/1607.00036, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 36,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 103,
+ 587,
+ 506,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 587,
+ 506,
+ 604
+ ],
+ "score": 1.0,
+ "content": "Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The goldilocks principle: Reading",
+ "type": "text"
+ }
+ ],
+ "index": 37,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 599,
+ 462,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 599,
+ 462,
+ 614
+ ],
+ "score": 1.0,
+ "content": "children’s books with explicit memory representations. CoRR, abs/1511.02301, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 38,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 632
+ ],
+ "score": 1.0,
+ "content": "Geoffrey Hinton, Li Deng, Dong Yu, George Dahl, Abdelrahman Mohamed, Navdeep Jaitly, An-",
+ "type": "text"
+ }
+ ],
+ "index": 39,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 628,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 628,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": "drew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara Sainath, and Brian Kingsbury. Deep",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 114,
+ 640,
+ 505,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 640,
+ 505,
+ 654
+ ],
+ "score": 1.0,
+ "content": "neural networks for acoustic modeling in speech recognition: The shared views of four research",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 651,
+ 372,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 651,
+ 372,
+ 665
+ ],
+ "score": 1.0,
+ "content": "groups. IEEE Signal Processing Magazine, 29(6):82–97, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 42,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 669,
+ 505,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 669,
+ 505,
+ 684
+ ],
+ "score": 1.0,
+ "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):",
+ "type": "text"
+ }
+ ],
+ "index": 43,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 116,
+ 680,
+ 191,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 680,
+ 191,
+ 693
+ ],
+ "score": 1.0,
+ "content": "1735–1780, 1997.",
+ "type": "text"
+ }
+ ],
+ "index": 44,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Piotr Indyk and Rajeev Motwani. Approximate nearest neighbors: towards removing the curse of",
+ "type": "text"
+ }
+ ],
+ "index": 45,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 708,
+ 505,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 708,
+ 505,
+ 724
+ ],
+ "score": 1.0,
+ "content": "dimensionality. In Proceedings of the thirtieth annual ACM symposium on Theory of computing,",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 223,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 223,
+ 732
+ ],
+ "score": 1.0,
+ "content": "pp. 604–613. ACM, 1998.",
+ "type": "text"
+ }
+ ],
+ "index": 47,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 103,
+ 289,
+ 506,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Łukasz Kaiser and Samy Bengio. Can active memory replace attention? In Advances in Neural",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 93,
+ 307,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 93,
+ 307,
+ 106
+ ],
+ "score": 1.0,
+ "content": "Information Processing Systems, (NIPS), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 112,
+ 504,
+ 136
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "score": 1.0,
+ "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 117,
+ 124,
+ 412,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 124,
+ 412,
+ 135
+ ],
+ "score": 1.0,
+ "content": "abs/1412.6980, 2014. URL http://arxiv.org/abs/1412.6980.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 142,
+ 504,
+ 165
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 142,
+ 506,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 142,
+ 506,
+ 155
+ ],
+ "score": 1.0,
+ "content": "Gregory Koch. Siamese neural networks for one-shot image recognition. PhD thesis, University of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 153,
+ 178,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 153,
+ 178,
+ 165
+ ],
+ "score": 1.0,
+ "content": "Toronto, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 171,
+ 504,
+ 195
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 504,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 504,
+ 185
+ ],
+ "score": 1.0,
+ "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton. Imagenet classification with deep convolu-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 116,
+ 183,
+ 457,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 183,
+ 457,
+ 196
+ ],
+ "score": 1.0,
+ "content": "tional neural network. In Advances in Neural Information Processing Systems, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 225
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 216
+ ],
+ "score": 1.0,
+ "content": "Brenden M Lake, Ruslan Salakhutdinov, Jason Gross, and Joshua B Tenenbaum. One shot learning",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 213,
+ 249,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 213,
+ 249,
+ 225
+ ],
+ "score": 1.0,
+ "content": "of simple visual concepts. 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 231,
+ 504,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "Manjunath Kudlur Oriol Vinyals, Samy Bengio. Order matters: Sequence to sequence for sets. In",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 243,
+ 400,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 243,
+ 400,
+ 255
+ ],
+ "score": 1.0,
+ "content": "International Conference on Learning Representations (ICLR), 2016a.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 261,
+ 504,
+ 285
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 275
+ ],
+ "score": 1.0,
+ "content": "Timothy Lillicrap Koray Kavukcuoglu Daan Wierstra Oriol Vinyals, Charles Blundell. Matching",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 273,
+ 371,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 273,
+ 371,
+ 285
+ ],
+ "score": 1.0,
+ "content": "networks for one shot learning. CoRR, abs/1606.04080, 2016b.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 291,
+ 503,
+ 325
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "Jack W Rae, Jonathan J Hunt, Tim Harley, Ivo Danihelka, Andrew Senior, Greg Wayne, Alex",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 115,
+ 302,
+ 505,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 302,
+ 505,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Graves, and Timothy P Lillicrap. Scaling memory-augmented neural networks with sparse reads",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 314,
+ 445,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 314,
+ 445,
+ 326
+ ],
+ "score": 1.0,
+ "content": "and writes. In Advances in Neural Information Processing Systems, (NIPS), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 332,
+ 503,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 345
+ ],
+ "score": 1.0,
+ "content": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy P. Lillicrap. One-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 344,
+ 465,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 344,
+ 465,
+ 356
+ ],
+ "score": 1.0,
+ "content": "shot learning with memory-augmented neural networks. CoRR, abs/1605.06065, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 362,
+ 502,
+ 385
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 504,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 504,
+ 375
+ ],
+ "score": 1.0,
+ "content": "Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 116,
+ 373,
+ 347,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 373,
+ 347,
+ 385
+ ],
+ "score": 1.0,
+ "content": "recognition and clustering. In CVPR, pp. 815–823, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 392,
+ 502,
+ 415
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 504,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 504,
+ 406
+ ],
+ "score": 1.0,
+ "content": "Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. Weakly supervised memory",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 116,
+ 403,
+ 495,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 403,
+ 495,
+ 416
+ ],
+ "score": 1.0,
+ "content": "networks. CoRR, abs/1503.08895, 2015. URL http://arxiv.org/abs/1503.08895.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 421,
+ 502,
+ 445
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 504,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 504,
+ 434
+ ],
+ "score": 1.0,
+ "content": "Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. Sequence to sequence learning with neural net-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 116,
+ 433,
+ 461,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 433,
+ 461,
+ 446
+ ],
+ "score": 1.0,
+ "content": "works. In Advances in Neural Information Processing Systems, pp. 3104–3112, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 451,
+ 503,
+ 475
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": "Oriol Vinyals, Charles Blundell, Timothy P. Lillicrap, Koray Kavukcuoglu, and Daan Wierstra.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 462,
+ 407,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 462,
+ 407,
+ 475
+ ],
+ "score": 1.0,
+ "content": "Matching networks for one shot learning. CoRR, abs/1606.04080, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 481,
+ 503,
+ 505
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 504,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 504,
+ 495
+ ],
+ "score": 1.0,
+ "content": "Kilian Q Weinberger and Lawrence K Saul. Distance metric learning for large margin nearest neigh-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 492,
+ 450,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 492,
+ 450,
+ 505
+ ],
+ "score": 1.0,
+ "content": "bor classification. Journal of Machine Learning Research, 10(Feb):207–244, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 511,
+ 504,
+ 545
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 510,
+ 505,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 510,
+ 505,
+ 525
+ ],
+ "score": 1.0,
+ "content": "Jason Weston, Samy Bengio, and Nicolas Usunier. Wsabie: Scaling up to large vocabulary image",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 115,
+ 523,
+ 505,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 523,
+ 505,
+ 535
+ ],
+ "score": 1.0,
+ "content": "annotation. In Proceedings of the International Joint Conference on Artificial Intelligence, IJCAI,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 533,
+ 144,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 533,
+ 144,
+ 546
+ ],
+ "score": 1.0,
+ "content": "2011.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 552,
+ 503,
+ 575
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. CoRR, abs/1410.3916, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 117,
+ 564,
+ 323,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 564,
+ 323,
+ 575
+ ],
+ "score": 1.0,
+ "content": "URL http://arxiv.org/abs/1410.3916.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 582,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 596
+ ],
+ "score": 1.0,
+ "content": "Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 593,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 593,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner, Apurva Shah, Melvin John-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 114,
+ 603,
+ 506,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 603,
+ 506,
+ 618
+ ],
+ "score": 1.0,
+ "content": "son, Xiaobing Liu, Łukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa,",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 116,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "Keith Stevens, George Kurian, Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 116,
+ 627,
+ 505,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 627,
+ 505,
+ 638
+ ],
+ "score": 1.0,
+ "content": "Alex Rudnick, Oriol Vinyals, Greg Corrado, Macduff Hughes, and Jeffrey Dean. Google’s neural",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 637,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 637,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "machine translation system: Bridging the gap between human and machine translation. CoRR,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 648,
+ 424,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 648,
+ 424,
+ 661
+ ],
+ "score": 1.0,
+ "content": "abs/1609.08144, 2016. URL http://arxiv.org/abs/1609.08144.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 667,
+ 504,
+ 701
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "Kelvin Xu, Jimmy Lei Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 677,
+ 505,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 677,
+ 505,
+ 691
+ ],
+ "score": 1.0,
+ "content": "Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 117,
+ 689,
+ 268,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 689,
+ 268,
+ 700
+ ],
+ "score": 1.0,
+ "content": "with visual attention. In ICML, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 708,
+ 504,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 505,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 505,
+ 721
+ ],
+ "score": 1.0,
+ "content": "Wojciech Zaremba and Ilya Sutskever. Reinforcement learning neural turing machines. CoRR,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 115,
+ 718,
+ 425,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 718,
+ 425,
+ 731
+ ],
+ "score": 1.0,
+ "content": "abs/1505.00521, 2015. URL http://arxiv.org/abs/1505.00521.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 44.5
+ }
+ ],
+ "page_idx": 9,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "10",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Łukasz Kaiser and Samy Bengio. Can active memory replace attention? In Advances in Neural",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 93,
+ 307,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 93,
+ 307,
+ 106
+ ],
+ "score": 1.0,
+ "content": "Information Processing Systems, (NIPS), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 104,
+ 81,
+ 505,
+ 106
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 112,
+ 504,
+ 136
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "score": 1.0,
+ "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 117,
+ 124,
+ 412,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 124,
+ 412,
+ 135
+ ],
+ "score": 1.0,
+ "content": "abs/1412.6980, 2014. URL http://arxiv.org/abs/1412.6980.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 106,
+ 112,
+ 505,
+ 135
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 142,
+ 504,
+ 165
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 142,
+ 506,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 142,
+ 506,
+ 155
+ ],
+ "score": 1.0,
+ "content": "Gregory Koch. Siamese neural networks for one-shot image recognition. PhD thesis, University of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 153,
+ 178,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 153,
+ 178,
+ 165
+ ],
+ "score": 1.0,
+ "content": "Toronto, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 106,
+ 142,
+ 506,
+ 165
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 171,
+ 504,
+ 195
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 504,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 504,
+ 185
+ ],
+ "score": 1.0,
+ "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton. Imagenet classification with deep convolu-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 116,
+ 183,
+ 457,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 183,
+ 457,
+ 196
+ ],
+ "score": 1.0,
+ "content": "tional neural network. In Advances in Neural Information Processing Systems, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 106,
+ 171,
+ 504,
+ 196
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 225
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 216
+ ],
+ "score": 1.0,
+ "content": "Brenden M Lake, Ruslan Salakhutdinov, Jason Gross, and Joshua B Tenenbaum. One shot learning",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 213,
+ 249,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 213,
+ 249,
+ 225
+ ],
+ "score": 1.0,
+ "content": "of simple visual concepts. 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 105,
+ 199,
+ 505,
+ 225
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 231,
+ 504,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "Manjunath Kudlur Oriol Vinyals, Samy Bengio. Order matters: Sequence to sequence for sets. In",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 243,
+ 400,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 243,
+ 400,
+ 255
+ ],
+ "score": 1.0,
+ "content": "International Conference on Learning Representations (ICLR), 2016a.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 105,
+ 231,
+ 505,
+ 255
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 261,
+ 504,
+ 285
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 275
+ ],
+ "score": 1.0,
+ "content": "Timothy Lillicrap Koray Kavukcuoglu Daan Wierstra Oriol Vinyals, Charles Blundell. Matching",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 273,
+ 371,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 273,
+ 371,
+ 285
+ ],
+ "score": 1.0,
+ "content": "networks for one shot learning. CoRR, abs/1606.04080, 2016b.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 105,
+ 260,
+ 506,
+ 285
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 291,
+ 503,
+ 325
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "Jack W Rae, Jonathan J Hunt, Tim Harley, Ivo Danihelka, Andrew Senior, Greg Wayne, Alex",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 115,
+ 302,
+ 505,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 302,
+ 505,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Graves, and Timothy P Lillicrap. Scaling memory-augmented neural networks with sparse reads",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 314,
+ 445,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 314,
+ 445,
+ 326
+ ],
+ "score": 1.0,
+ "content": "and writes. In Advances in Neural Information Processing Systems, (NIPS), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 105,
+ 291,
+ 505,
+ 326
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 332,
+ 503,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 345
+ ],
+ "score": 1.0,
+ "content": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy P. Lillicrap. One-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 344,
+ 465,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 344,
+ 465,
+ 356
+ ],
+ "score": 1.0,
+ "content": "shot learning with memory-augmented neural networks. CoRR, abs/1605.06065, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 106,
+ 333,
+ 505,
+ 356
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 362,
+ 502,
+ 385
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 504,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 504,
+ 375
+ ],
+ "score": 1.0,
+ "content": "Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 116,
+ 373,
+ 347,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 373,
+ 347,
+ 385
+ ],
+ "score": 1.0,
+ "content": "recognition and clustering. In CVPR, pp. 815–823, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19.5,
+ "bbox_fs": [
+ 106,
+ 362,
+ 504,
+ 385
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 392,
+ 502,
+ 415
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 504,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 504,
+ 406
+ ],
+ "score": 1.0,
+ "content": "Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. Weakly supervised memory",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 116,
+ 403,
+ 495,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 403,
+ 495,
+ 416
+ ],
+ "score": 1.0,
+ "content": "networks. CoRR, abs/1503.08895, 2015. URL http://arxiv.org/abs/1503.08895.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21.5,
+ "bbox_fs": [
+ 105,
+ 391,
+ 504,
+ 416
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 421,
+ 502,
+ 445
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 504,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 504,
+ 434
+ ],
+ "score": 1.0,
+ "content": "Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. Sequence to sequence learning with neural net-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 116,
+ 433,
+ 461,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 433,
+ 461,
+ 446
+ ],
+ "score": 1.0,
+ "content": "works. In Advances in Neural Information Processing Systems, pp. 3104–3112, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 106,
+ 423,
+ 504,
+ 446
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 451,
+ 503,
+ 475
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": "Oriol Vinyals, Charles Blundell, Timothy P. Lillicrap, Koray Kavukcuoglu, and Daan Wierstra.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 462,
+ 407,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 462,
+ 407,
+ 475
+ ],
+ "score": 1.0,
+ "content": "Matching networks for one shot learning. CoRR, abs/1606.04080, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 106,
+ 451,
+ 505,
+ 475
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 481,
+ 503,
+ 505
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 504,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 504,
+ 495
+ ],
+ "score": 1.0,
+ "content": "Kilian Q Weinberger and Lawrence K Saul. Distance metric learning for large margin nearest neigh-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 492,
+ 450,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 492,
+ 450,
+ 505
+ ],
+ "score": 1.0,
+ "content": "bor classification. Journal of Machine Learning Research, 10(Feb):207–244, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 481,
+ 504,
+ 505
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 511,
+ 504,
+ 545
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 510,
+ 505,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 510,
+ 505,
+ 525
+ ],
+ "score": 1.0,
+ "content": "Jason Weston, Samy Bengio, and Nicolas Usunier. Wsabie: Scaling up to large vocabulary image",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 115,
+ 523,
+ 505,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 523,
+ 505,
+ 535
+ ],
+ "score": 1.0,
+ "content": "annotation. In Proceedings of the International Joint Conference on Artificial Intelligence, IJCAI,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 533,
+ 144,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 533,
+ 144,
+ 546
+ ],
+ "score": 1.0,
+ "content": "2011.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 510,
+ 505,
+ 546
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 552,
+ 503,
+ 575
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. CoRR, abs/1410.3916, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 117,
+ 564,
+ 323,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 564,
+ 323,
+ 575
+ ],
+ "score": 1.0,
+ "content": "URL http://arxiv.org/abs/1410.3916.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 105,
+ 552,
+ 505,
+ 575
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 582,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 596
+ ],
+ "score": 1.0,
+ "content": "Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 593,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 593,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner, Apurva Shah, Melvin John-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 114,
+ 603,
+ 506,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 603,
+ 506,
+ 618
+ ],
+ "score": 1.0,
+ "content": "son, Xiaobing Liu, Łukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa,",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 116,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "Keith Stevens, George Kurian, Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 116,
+ 627,
+ 505,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 627,
+ 505,
+ 638
+ ],
+ "score": 1.0,
+ "content": "Alex Rudnick, Oriol Vinyals, Greg Corrado, Macduff Hughes, and Jeffrey Dean. Google’s neural",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 637,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 637,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "machine translation system: Bridging the gap between human and machine translation. CoRR,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 648,
+ 424,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 648,
+ 424,
+ 661
+ ],
+ "score": 1.0,
+ "content": "abs/1609.08144, 2016. URL http://arxiv.org/abs/1609.08144.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 106,
+ 581,
+ 506,
+ 661
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 667,
+ 504,
+ 701
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "Kelvin Xu, Jimmy Lei Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 677,
+ 505,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 677,
+ 505,
+ 691
+ ],
+ "score": 1.0,
+ "content": "Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 117,
+ 689,
+ 268,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 689,
+ 268,
+ 700
+ ],
+ "score": 1.0,
+ "content": "with visual attention. In ICML, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 105,
+ 666,
+ 505,
+ 700
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 708,
+ 504,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 505,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 505,
+ 721
+ ],
+ "score": 1.0,
+ "content": "Wojciech Zaremba and Ilya Sutskever. Reinforcement learning neural turing machines. CoRR,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 115,
+ 718,
+ 425,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 718,
+ 425,
+ 731
+ ],
+ "score": 1.0,
+ "content": "abs/1505.00521, 2015. URL http://arxiv.org/abs/1505.00521.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 44.5,
+ "bbox_fs": [
+ 106,
+ 708,
+ 505,
+ 731
+ ]
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file