diff --git "a/parse/train/S1Euwz-Rb/S1Euwz-Rb_middle.json" "b/parse/train/S1Euwz-Rb/S1Euwz-Rb_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/train/S1Euwz-Rb/S1Euwz-Rb_middle.json"
@@ -0,0 +1,54670 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 79,
+ 414,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 415,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 415,
+ 97
+ ],
+ "score": 1.0,
+ "content": "COMPOSITIONAL ATTENTION NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 99,
+ 309,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 99,
+ 309,
+ 117
+ ],
+ "score": 1.0,
+ "content": "FOR MACHINE REASONING",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 135,
+ 252,
+ 179
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 136,
+ 186,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 136,
+ 186,
+ 146
+ ],
+ "score": 1.0,
+ "content": "Drew A. Hudson",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 111,
+ 145,
+ 249,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 145,
+ 249,
+ 159
+ ],
+ "score": 1.0,
+ "content": "Department of Computer Science",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 112,
+ 156,
+ 194,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 156,
+ 194,
+ 169
+ ],
+ "score": 1.0,
+ "content": "Stanford University",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 113,
+ 169,
+ 252,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 169,
+ 252,
+ 180
+ ],
+ "score": 1.0,
+ "content": "dorarad@cs.stanford.edu",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 310,
+ 136,
+ 449,
+ 180
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 310,
+ 134,
+ 418,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 134,
+ 418,
+ 148
+ ],
+ "score": 1.0,
+ "content": "Christopher D. Manning",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 309,
+ 145,
+ 447,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 145,
+ 447,
+ 159
+ ],
+ "score": 1.0,
+ "content": "Department of Computer Science",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 309,
+ 156,
+ 392,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 156,
+ 392,
+ 170
+ ],
+ "score": 1.0,
+ "content": "Stanford University",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 309,
+ 169,
+ 450,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 169,
+ 450,
+ 180
+ ],
+ "score": 1.0,
+ "content": "manning@cs.stanford.edu",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 208,
+ 333,
+ 220
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 207,
+ 336,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 207,
+ 336,
+ 222
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 235,
+ 468,
+ 411
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 235,
+ 469,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 235,
+ 469,
+ 248
+ ],
+ "score": 1.0,
+ "content": "We present Compositional Attention Networks, a novel fully differentiable neu-",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 245,
+ 469,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 245,
+ 469,
+ 260
+ ],
+ "score": 1.0,
+ "content": "ral network architecture, designed to facilitate explicit and expressive reasoning.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 256,
+ 470,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 256,
+ 470,
+ 271
+ ],
+ "score": 1.0,
+ "content": "While many types of neural networks are effective at learning and generalizing",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 268,
+ 469,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 268,
+ 469,
+ 280
+ ],
+ "score": 1.0,
+ "content": "from massive quantities of data, this model moves away from monolithic black-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 279,
+ 469,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 279,
+ 469,
+ 291
+ ],
+ "score": 1.0,
+ "content": "box architectures towards a design that provides a strong prior for iterative rea-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 142,
+ 291,
+ 469,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 291,
+ 469,
+ 302
+ ],
+ "score": 1.0,
+ "content": "soning, allowing it to support explainable and structured learning, as well as",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 301,
+ 470,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 301,
+ 470,
+ 313
+ ],
+ "score": 1.0,
+ "content": "generalization from a modest amount of data. The model builds on the great",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 312,
+ 469,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 312,
+ 469,
+ 324
+ ],
+ "score": 1.0,
+ "content": "success of existing recurrent cells such as LSTMs: It sequences a single recur-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 323,
+ 469,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 323,
+ 469,
+ 335
+ ],
+ "score": 1.0,
+ "content": "rent Memory, Attention, and Control (MAC) cell, and by careful design imposes",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 334,
+ 469,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 334,
+ 469,
+ 346
+ ],
+ "score": 1.0,
+ "content": "structural constraints on the operation of each cell and the interactions between",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 142,
+ 345,
+ 469,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 345,
+ 469,
+ 357
+ ],
+ "score": 1.0,
+ "content": "them, incorporating explicit control and soft attention mechanisms into their in-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 369
+ ],
+ "score": 1.0,
+ "content": "terfaces. We demonstrate the model’s strength and robustness on the challenging",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 142,
+ 366,
+ 469,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 366,
+ 417,
+ 379
+ ],
+ "score": 1.0,
+ "content": "CLEVR dataset for visual reasoning, achieving a new state-of-the-art",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 417,
+ 367,
+ 444,
+ 377
+ ],
+ "score": 0.87,
+ "content": "9 8 . 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 445,
+ 366,
+ 469,
+ 379
+ ],
+ "score": 1.0,
+ "content": "accu-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 378,
+ 469,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 378,
+ 469,
+ 390
+ ],
+ "score": 1.0,
+ "content": "racy, halving the error rate of the previous best model. More importantly, we show",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 387,
+ 470,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 387,
+ 470,
+ 402
+ ],
+ "score": 1.0,
+ "content": "that the new model is more computationally efficient and data-efficient, requiring",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 400,
+ 414,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 400,
+ 414,
+ 412
+ ],
+ "score": 1.0,
+ "content": "an order of magnitude less time and/or data to achieve good results.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 436,
+ 205,
+ 448
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 208,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 208,
+ 452
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 462,
+ 504,
+ 517
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 476
+ ],
+ "score": 1.0,
+ "content": "This paper considers how best to design neural networks to perform the iterative reasoning necessary",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 474,
+ 506,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 506,
+ 486
+ ],
+ "score": 1.0,
+ "content": "for complex problem solving. Putting facts and observations together to arrive at conclusions is a",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 485,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "central necessary ability as we work to move neural networks beyond their current great success with",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 508
+ ],
+ "score": 1.0,
+ "content": "sensory perception tasks (LeCun et al., 1998; Krizhevsky et al., 2012) towards displaying Artificial",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 192,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 192,
+ 520
+ ],
+ "score": 1.0,
+ "content": "General Intelligence.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 523,
+ 375,
+ 698
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 376,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 376,
+ 535
+ ],
+ "score": 1.0,
+ "content": "Concretely, we develop a novel model that we apply to the CLEVR",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 533,
+ 376,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 376,
+ 547
+ ],
+ "score": 1.0,
+ "content": "dataset (Johnson et al., 2016) for visual question answering (VQA).",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 545,
+ 376,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 376,
+ 558
+ ],
+ "score": 1.0,
+ "content": "VQA (Antol et al., 2015; Gupta, 2017) is a challenging multimodal",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 556,
+ 376,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 556,
+ 376,
+ 569
+ ],
+ "score": 1.0,
+ "content": "task that requires responding to natural language questions about",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 567,
+ 376,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 567,
+ 376,
+ 580
+ ],
+ "score": 1.0,
+ "content": "images. However, Agrawal et al. (2016) show how the first gen-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 376,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 376,
+ 591
+ ],
+ "score": 1.0,
+ "content": "eration of successful models on VQA tasks tend to acquire only",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 589,
+ 377,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 589,
+ 377,
+ 602
+ ],
+ "score": 1.0,
+ "content": "superficial comprehension of both the image and the question, ex-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 600,
+ 377,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 377,
+ 613
+ ],
+ "score": 1.0,
+ "content": "ploiting dataset biases rather than capturing a sound perception and",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 612,
+ 376,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 376,
+ 623
+ ],
+ "score": 1.0,
+ "content": "reasoning process that would lead to the correct answer (Sturm,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 621,
+ 376,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 376,
+ 633
+ ],
+ "score": 1.0,
+ "content": "2014). CLEVR was created to address this problem. As illus-",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 376,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 376,
+ 645
+ ],
+ "score": 1.0,
+ "content": "trated in figure 1, instances in the dataset consist of rendered im-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 376,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 376,
+ 656
+ ],
+ "score": 1.0,
+ "content": "ages featuring 3D objects of several shapes, colors, materials and",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 376,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 376,
+ 667
+ ],
+ "score": 1.0,
+ "content": "sizes, coupled with unbiased, compositional questions that require",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 666,
+ 376,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 376,
+ 678
+ ],
+ "score": 1.0,
+ "content": "an array of challenging reasoning skills such as following transitive",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 376,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 376,
+ 690
+ ],
+ "score": 1.0,
+ "content": "relations, counting objects and comparing their properties, without",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 376,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 376,
+ 700
+ ],
+ "score": 1.0,
+ "content": "allowing any shortcuts around such reasoning. Notably, each in-",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 387,
+ 537,
+ 501,
+ 613
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 387,
+ 537,
+ 501,
+ 613
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 387,
+ 537,
+ 501,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 537,
+ 501,
+ 613
+ ],
+ "score": 0.962,
+ "type": "image",
+ "image_path": "674dcd95919d901cb7279f9b0552761c7620c6d163165031e5b1cd19087e50a7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 49.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 387,
+ 537,
+ 501,
+ 575.0
+ ],
+ "spans": [],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 387,
+ 575.0,
+ 501,
+ 613.0
+ ],
+ "spans": [],
+ "index": 50
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 383,
+ 622,
+ 504,
+ 687
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 383,
+ 621,
+ 504,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 621,
+ 504,
+ 633
+ ],
+ "score": 1.0,
+ "content": "Figure 1: A sample im-",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 382,
+ 632,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 632,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "age from the CLEVR dataset,",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 383,
+ 643,
+ 506,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 643,
+ 506,
+ 655
+ ],
+ "score": 1.0,
+ "content": "with a question: “There is a",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 382,
+ 655,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 655,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "purple cube behind a metal",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 383,
+ 666,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 666,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": "object left to a large ball; what",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 383,
+ 676,
+ 444,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 676,
+ 444,
+ 687
+ ],
+ "score": 1.0,
+ "content": "material is it?”",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ }
+ ],
+ "index": 53.5
+ }
+ ],
+ "index": 51.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 700,
+ 504,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "stance in CLEVR is also accompanied by a tree-structured functional program that was both used to",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "construct the question and reflects its reasoning procedure – a series of predefined operations – that",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 264,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 264,
+ 733
+ ],
+ "score": 1.0,
+ "content": "can be composed together to answer it.",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ }
+ ],
+ "index": 58
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "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 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "1",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 79,
+ 414,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 415,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 415,
+ 97
+ ],
+ "score": 1.0,
+ "content": "COMPOSITIONAL ATTENTION NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 99,
+ 309,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 99,
+ 309,
+ 117
+ ],
+ "score": 1.0,
+ "content": "FOR MACHINE REASONING",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 112,
+ 135,
+ 252,
+ 179
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 136,
+ 186,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 136,
+ 186,
+ 146
+ ],
+ "score": 1.0,
+ "content": "Drew A. Hudson",
+ "type": "text"
+ }
+ ],
+ "index": 2,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 145,
+ 249,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 145,
+ 249,
+ 159
+ ],
+ "score": 1.0,
+ "content": "Department of Computer Science",
+ "type": "text"
+ }
+ ],
+ "index": 5,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 112,
+ 156,
+ 194,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 156,
+ 194,
+ 169
+ ],
+ "score": 1.0,
+ "content": "Stanford University",
+ "type": "text"
+ }
+ ],
+ "index": 4,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 113,
+ 169,
+ 252,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 169,
+ 252,
+ 180
+ ],
+ "score": 1.0,
+ "content": "dorarad@cs.stanford.edu",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 111,
+ 136,
+ 252,
+ 180
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 310,
+ 136,
+ 449,
+ 180
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 310,
+ 134,
+ 418,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 134,
+ 418,
+ 148
+ ],
+ "score": 1.0,
+ "content": "Christopher D. Manning",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 309,
+ 145,
+ 447,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 145,
+ 447,
+ 159
+ ],
+ "score": 1.0,
+ "content": "Department of Computer Science",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 309,
+ 156,
+ 392,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 156,
+ 392,
+ 170
+ ],
+ "score": 1.0,
+ "content": "Stanford University",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 309,
+ 169,
+ 450,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 169,
+ 450,
+ 180
+ ],
+ "score": 1.0,
+ "content": "manning@cs.stanford.edu",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 309,
+ 134,
+ 450,
+ 180
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 208,
+ 333,
+ 220
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 207,
+ 336,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 207,
+ 336,
+ 222
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 235,
+ 468,
+ 411
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 235,
+ 469,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 235,
+ 469,
+ 248
+ ],
+ "score": 1.0,
+ "content": "We present Compositional Attention Networks, a novel fully differentiable neu-",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 245,
+ 469,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 245,
+ 469,
+ 260
+ ],
+ "score": 1.0,
+ "content": "ral network architecture, designed to facilitate explicit and expressive reasoning.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 256,
+ 470,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 256,
+ 470,
+ 271
+ ],
+ "score": 1.0,
+ "content": "While many types of neural networks are effective at learning and generalizing",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 268,
+ 469,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 268,
+ 469,
+ 280
+ ],
+ "score": 1.0,
+ "content": "from massive quantities of data, this model moves away from monolithic black-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 279,
+ 469,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 279,
+ 469,
+ 291
+ ],
+ "score": 1.0,
+ "content": "box architectures towards a design that provides a strong prior for iterative rea-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 142,
+ 291,
+ 469,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 291,
+ 469,
+ 302
+ ],
+ "score": 1.0,
+ "content": "soning, allowing it to support explainable and structured learning, as well as",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 301,
+ 470,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 301,
+ 470,
+ 313
+ ],
+ "score": 1.0,
+ "content": "generalization from a modest amount of data. The model builds on the great",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 312,
+ 469,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 312,
+ 469,
+ 324
+ ],
+ "score": 1.0,
+ "content": "success of existing recurrent cells such as LSTMs: It sequences a single recur-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 323,
+ 469,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 323,
+ 469,
+ 335
+ ],
+ "score": 1.0,
+ "content": "rent Memory, Attention, and Control (MAC) cell, and by careful design imposes",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 334,
+ 469,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 334,
+ 469,
+ 346
+ ],
+ "score": 1.0,
+ "content": "structural constraints on the operation of each cell and the interactions between",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 142,
+ 345,
+ 469,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 345,
+ 469,
+ 357
+ ],
+ "score": 1.0,
+ "content": "them, incorporating explicit control and soft attention mechanisms into their in-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 369
+ ],
+ "score": 1.0,
+ "content": "terfaces. We demonstrate the model’s strength and robustness on the challenging",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 142,
+ 366,
+ 469,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 366,
+ 417,
+ 379
+ ],
+ "score": 1.0,
+ "content": "CLEVR dataset for visual reasoning, achieving a new state-of-the-art",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 417,
+ 367,
+ 444,
+ 377
+ ],
+ "score": 0.87,
+ "content": "9 8 . 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 445,
+ 366,
+ 469,
+ 379
+ ],
+ "score": 1.0,
+ "content": "accu-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 378,
+ 469,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 378,
+ 469,
+ 390
+ ],
+ "score": 1.0,
+ "content": "racy, halving the error rate of the previous best model. More importantly, we show",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 387,
+ 470,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 387,
+ 470,
+ 402
+ ],
+ "score": 1.0,
+ "content": "that the new model is more computationally efficient and data-efficient, requiring",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 400,
+ 414,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 400,
+ 414,
+ 412
+ ],
+ "score": 1.0,
+ "content": "an order of magnitude less time and/or data to achieve good results.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 141,
+ 235,
+ 470,
+ 412
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 436,
+ 205,
+ 448
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 208,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 208,
+ 452
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 462,
+ 504,
+ 517
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 505,
+ 476
+ ],
+ "score": 1.0,
+ "content": "This paper considers how best to design neural networks to perform the iterative reasoning necessary",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 474,
+ 506,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 506,
+ 486
+ ],
+ "score": 1.0,
+ "content": "for complex problem solving. Putting facts and observations together to arrive at conclusions is a",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 485,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "central necessary ability as we work to move neural networks beyond their current great success with",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 508
+ ],
+ "score": 1.0,
+ "content": "sensory perception tasks (LeCun et al., 1998; Krizhevsky et al., 2012) towards displaying Artificial",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 192,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 192,
+ 520
+ ],
+ "score": 1.0,
+ "content": "General Intelligence.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 462,
+ 506,
+ 520
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 523,
+ 375,
+ 698
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 376,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 376,
+ 535
+ ],
+ "score": 1.0,
+ "content": "Concretely, we develop a novel model that we apply to the CLEVR",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 533,
+ 376,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 376,
+ 547
+ ],
+ "score": 1.0,
+ "content": "dataset (Johnson et al., 2016) for visual question answering (VQA).",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 545,
+ 376,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 376,
+ 558
+ ],
+ "score": 1.0,
+ "content": "VQA (Antol et al., 2015; Gupta, 2017) is a challenging multimodal",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 556,
+ 376,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 556,
+ 376,
+ 569
+ ],
+ "score": 1.0,
+ "content": "task that requires responding to natural language questions about",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 567,
+ 376,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 567,
+ 376,
+ 580
+ ],
+ "score": 1.0,
+ "content": "images. However, Agrawal et al. (2016) show how the first gen-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 376,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 376,
+ 591
+ ],
+ "score": 1.0,
+ "content": "eration of successful models on VQA tasks tend to acquire only",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 589,
+ 377,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 589,
+ 377,
+ 602
+ ],
+ "score": 1.0,
+ "content": "superficial comprehension of both the image and the question, ex-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 600,
+ 377,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 377,
+ 613
+ ],
+ "score": 1.0,
+ "content": "ploiting dataset biases rather than capturing a sound perception and",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 612,
+ 376,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 376,
+ 623
+ ],
+ "score": 1.0,
+ "content": "reasoning process that would lead to the correct answer (Sturm,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 621,
+ 376,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 376,
+ 633
+ ],
+ "score": 1.0,
+ "content": "2014). CLEVR was created to address this problem. As illus-",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 376,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 376,
+ 645
+ ],
+ "score": 1.0,
+ "content": "trated in figure 1, instances in the dataset consist of rendered im-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 376,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 376,
+ 656
+ ],
+ "score": 1.0,
+ "content": "ages featuring 3D objects of several shapes, colors, materials and",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 376,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 376,
+ 667
+ ],
+ "score": 1.0,
+ "content": "sizes, coupled with unbiased, compositional questions that require",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 666,
+ 376,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 376,
+ 678
+ ],
+ "score": 1.0,
+ "content": "an array of challenging reasoning skills such as following transitive",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 376,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 376,
+ 690
+ ],
+ "score": 1.0,
+ "content": "relations, counting objects and comparing their properties, without",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 376,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 376,
+ 700
+ ],
+ "score": 1.0,
+ "content": "allowing any shortcuts around such reasoning. Notably, each in-",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "stance in CLEVR is also accompanied by a tree-structured functional program that was both used to",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "construct the question and reflects its reasoning procedure – a series of predefined operations – that",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 264,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 264,
+ 733
+ ],
+ "score": 1.0,
+ "content": "can be composed together to answer it.",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 523,
+ 377,
+ 700
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 387,
+ 537,
+ 501,
+ 613
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 387,
+ 537,
+ 501,
+ 613
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 387,
+ 537,
+ 501,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 537,
+ 501,
+ 613
+ ],
+ "score": 0.962,
+ "type": "image",
+ "image_path": "674dcd95919d901cb7279f9b0552761c7620c6d163165031e5b1cd19087e50a7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 49.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 387,
+ 537,
+ 501,
+ 575.0
+ ],
+ "spans": [],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 387,
+ 575.0,
+ 501,
+ 613.0
+ ],
+ "spans": [],
+ "index": 50
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 383,
+ 622,
+ 504,
+ 687
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 383,
+ 621,
+ 504,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 621,
+ 504,
+ 633
+ ],
+ "score": 1.0,
+ "content": "Figure 1: A sample im-",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 382,
+ 632,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 632,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "age from the CLEVR dataset,",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 383,
+ 643,
+ 506,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 643,
+ 506,
+ 655
+ ],
+ "score": 1.0,
+ "content": "with a question: “There is a",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 382,
+ 655,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 655,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "purple cube behind a metal",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 383,
+ 666,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 666,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": "object left to a large ball; what",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 383,
+ 676,
+ 444,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 676,
+ 444,
+ 687
+ ],
+ "score": 1.0,
+ "content": "material is it?”",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ }
+ ],
+ "index": 53.5
+ }
+ ],
+ "index": 51.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 700,
+ 504,
+ 731
+ ],
+ "lines": [],
+ "index": 58,
+ "bbox_fs": [
+ 106,
+ 699,
+ 505,
+ 733
+ ],
+ "lines_deleted": true
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 192
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Most neural networks are essentially very large correlation engines that will hone in on any statis-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "tical, potentially spurious pattern that allows them to model the observed data more accurately. In",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "contrast, we seek to create a model structure that requires combining sound inference steps to solve",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "a problem instance. At the other extreme, some approaches adopt symbolic structures that resemble",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "the expression trees of programming languages to perform reasoning (Andreas et al., 2016b; Hu",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "score": 1.0,
+ "content": "et al., 2017). In particular, some approaches to CLEVR use the supplied functional programs for",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 161
+ ],
+ "score": 1.0,
+ "content": "supervised or semi-supervised training (Andreas et al., 2016a; Johnson et al., 2017). Not only do we",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "wish to avoid using such supervision in our work, but we in general suspect that the rigidity of these",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 170,
+ 506,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 170,
+ 506,
+ 183
+ ],
+ "score": 1.0,
+ "content": "structures and the use of an inventory of operation-specific neural modules undermines robustness",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 181,
+ 471,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 181,
+ 471,
+ 194
+ ],
+ "score": 1.0,
+ "content": "and generalization, and at any rate requires more complex reinforcement learning methods.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 198,
+ 505,
+ 396
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 197,
+ 505,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 197,
+ 505,
+ 211
+ ],
+ "score": 1.0,
+ "content": "To address these weaknesses, while still seeking to use a sound and transparent underlying reasoning",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "process, we propose Compositional Attention Networks, a novel, fully differentiable, non-modular",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 233
+ ],
+ "score": 1.0,
+ "content": "architecture for reasoning tasks. Our model is a straightforward recurrent neural network with at-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 231,
+ 504,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 231,
+ 504,
+ 243
+ ],
+ "score": 1.0,
+ "content": "tention; the novelty lies in the use of a new Memory, Attention and Composition (MAC) cell. The",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "constrained and deliberate design of the MAC cell was developed as a kind of strong structural",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 265
+ ],
+ "score": 1.0,
+ "content": "prior that encourages the network to solve problems by stringing together a sequence of transparent",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 506,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 506,
+ 277
+ ],
+ "score": 1.0,
+ "content": "reasoning steps. MAC cells are versatile but constrained neural units. They explicitly separate out",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 287
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 287
+ ],
+ "score": 1.0,
+ "content": "memory from control, both represented recurrently. The unit contains three sub-units: The control",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 298
+ ],
+ "score": 1.0,
+ "content": "unit updates the control representation based on outside instructions (for VQA, the question), learn-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 309
+ ],
+ "score": 1.0,
+ "content": "ing to successively attend to different parts of the instructions; the read unit gets information out of",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 321
+ ],
+ "score": 1.0,
+ "content": "a knowledge base (for VQA, the image) based on the control signal and the previous memory; the",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 319,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 319,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "write unit updates the memory based on soft self-attention to previous memories, controlled by the",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 329,
+ 505,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 329,
+ 505,
+ 343
+ ],
+ "score": 1.0,
+ "content": "retrieved information and the control signal. A universal MAC unit with a single set of parameters",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 340,
+ 505,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 340,
+ 505,
+ 353
+ ],
+ "score": 1.0,
+ "content": "is used throughout the reasoning process, but its behavior can vary widely based on the context in",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 365
+ ],
+ "score": 1.0,
+ "content": "which it is applied – the input to the control unit and the contents of the knowledge base. With atten-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 362,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "tion, our MAC network has the capacity to represent arbitrarily complex acyclic reasoning graphs in",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 373,
+ 506,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 373,
+ 506,
+ 387
+ ],
+ "score": 1.0,
+ "content": "a soft manner, while having physically sequential structure. The result is a continuous counterpart",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 423,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 423,
+ 397
+ ],
+ "score": 1.0,
+ "content": "to module networks that can be trained end-to-end simply by backpropagation.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 402,
+ 505,
+ 566
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 415
+ ],
+ "score": 1.0,
+ "content": "We test the behavior of our new network on CLEVR and its associated datasets. On the primary",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 316,
+ 425
+ ],
+ "score": 1.0,
+ "content": "CLEVR reasoning task, we achieve an accuracy of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 412,
+ 343,
+ 423
+ ],
+ "score": 0.87,
+ "content": "9 8 . 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 344,
+ 412,
+ 505,
+ 425
+ ],
+ "score": 1.0,
+ "content": ", halving the error rate compared to the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 423,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": "previous state-of-the-art FiLM model (Perez et al., 2017). In particular, we show that our architec-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 434,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "ture yields better performance on questions involving counting and aggregation. In supplementary",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 104,
+ 443,
+ 506,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 443,
+ 506,
+ 460
+ ],
+ "score": 1.0,
+ "content": "studies, we show that the MAC network learns more quickly (both in terms of number of training",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 469
+ ],
+ "score": 1.0,
+ "content": "epochs and training time) and more effectively from limited amounts of training data. Moreover, it",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 466,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 314,
+ 480
+ ],
+ "score": 1.0,
+ "content": "also achieves a new state-of-the-art performance of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 467,
+ 342,
+ 478
+ ],
+ "score": 0.88,
+ "content": "8 2 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 466,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "on the more varied and difficult human-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "authored questions of the CLEVR-Humans dataset. The careful design of our cell encourages com-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 502
+ ],
+ "score": 1.0,
+ "content": "positionality, versatility and transparency. We achieve these properties by defining attention-based",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 512
+ ],
+ "score": 1.0,
+ "content": "interfaces that constrict the cell’s input and output spaces, and so constrain the interactions both",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "between and inside cells in order to guide them towards simple reasoning behaviors. Although",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "each cell’s functionality has only a limited range of possible continuous reasoning behaviors, when",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "chained together in a MAC network, the whole system becomes expressive and powerful. In the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 506,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 506,
+ 556
+ ],
+ "score": 1.0,
+ "content": "future, we believe that the architecture will also prove beneficial for other multi-step reasoning and",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 554,
+ 458,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 458,
+ 568
+ ],
+ "score": 1.0,
+ "content": "inference tasks, for instance in machine comprehension and textual question answering.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 591,
+ 211,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 213,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 213,
+ 606
+ ],
+ "score": 1.0,
+ "content": "2 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 43
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 621,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 622,
+ 505,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 622,
+ 505,
+ 634
+ ],
+ "score": 1.0,
+ "content": "There have been several prominent models that address the CLEVR task. By and large they can be",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "score": 1.0,
+ "content": "partitioned into two groups: module networks, which in practice have all used the strong supervision",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 643,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 643,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "provided in the form of tree-structured functional programs that accompany each data instance, and",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 655,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "large, relatively unstructured end-to-end differentiable networks that complement a fairly standard",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "stack of CNNs with components that aid in performing reasoning tasks. In contrast to modular",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 689
+ ],
+ "score": 1.0,
+ "content": "approaches (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017), our model does not",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "require additional supervision and makes use of a single computational cell chained in sequence",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "(like an LSTM) rather than a collection of custom modules deployed in a rigid tree structure. In",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "contrast to augmented CNN approaches (Santoro et al., 2017; Perez et al., 2017), we suggest that our",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "approach provides an ability for relational reasoning with better generalization capacity and higher",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 48.5
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "score": 1.0,
+ "content": "2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 192
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Most neural networks are essentially very large correlation engines that will hone in on any statis-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "tical, potentially spurious pattern that allows them to model the observed data more accurately. In",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "contrast, we seek to create a model structure that requires combining sound inference steps to solve",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "a problem instance. At the other extreme, some approaches adopt symbolic structures that resemble",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "the expression trees of programming languages to perform reasoning (Andreas et al., 2016b; Hu",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "score": 1.0,
+ "content": "et al., 2017). In particular, some approaches to CLEVR use the supplied functional programs for",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 161
+ ],
+ "score": 1.0,
+ "content": "supervised or semi-supervised training (Andreas et al., 2016a; Johnson et al., 2017). Not only do we",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "wish to avoid using such supervision in our work, but we in general suspect that the rigidity of these",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 170,
+ 506,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 170,
+ 506,
+ 183
+ ],
+ "score": 1.0,
+ "content": "structures and the use of an inventory of operation-specific neural modules undermines robustness",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 181,
+ 471,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 181,
+ 471,
+ 194
+ ],
+ "score": 1.0,
+ "content": "and generalization, and at any rate requires more complex reinforcement learning methods.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 194
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 198,
+ 505,
+ 396
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 197,
+ 505,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 197,
+ 505,
+ 211
+ ],
+ "score": 1.0,
+ "content": "To address these weaknesses, while still seeking to use a sound and transparent underlying reasoning",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "process, we propose Compositional Attention Networks, a novel, fully differentiable, non-modular",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 233
+ ],
+ "score": 1.0,
+ "content": "architecture for reasoning tasks. Our model is a straightforward recurrent neural network with at-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 231,
+ 504,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 231,
+ 504,
+ 243
+ ],
+ "score": 1.0,
+ "content": "tention; the novelty lies in the use of a new Memory, Attention and Composition (MAC) cell. The",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "constrained and deliberate design of the MAC cell was developed as a kind of strong structural",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 265
+ ],
+ "score": 1.0,
+ "content": "prior that encourages the network to solve problems by stringing together a sequence of transparent",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 506,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 506,
+ 277
+ ],
+ "score": 1.0,
+ "content": "reasoning steps. MAC cells are versatile but constrained neural units. They explicitly separate out",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 287
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 287
+ ],
+ "score": 1.0,
+ "content": "memory from control, both represented recurrently. The unit contains three sub-units: The control",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 298
+ ],
+ "score": 1.0,
+ "content": "unit updates the control representation based on outside instructions (for VQA, the question), learn-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 309
+ ],
+ "score": 1.0,
+ "content": "ing to successively attend to different parts of the instructions; the read unit gets information out of",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 321
+ ],
+ "score": 1.0,
+ "content": "a knowledge base (for VQA, the image) based on the control signal and the previous memory; the",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 319,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 319,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "write unit updates the memory based on soft self-attention to previous memories, controlled by the",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 329,
+ 505,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 329,
+ 505,
+ 343
+ ],
+ "score": 1.0,
+ "content": "retrieved information and the control signal. A universal MAC unit with a single set of parameters",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 340,
+ 505,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 340,
+ 505,
+ 353
+ ],
+ "score": 1.0,
+ "content": "is used throughout the reasoning process, but its behavior can vary widely based on the context in",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 365
+ ],
+ "score": 1.0,
+ "content": "which it is applied – the input to the control unit and the contents of the knowledge base. With atten-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 362,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "tion, our MAC network has the capacity to represent arbitrarily complex acyclic reasoning graphs in",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 373,
+ 506,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 373,
+ 506,
+ 387
+ ],
+ "score": 1.0,
+ "content": "a soft manner, while having physically sequential structure. The result is a continuous counterpart",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 423,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 423,
+ 397
+ ],
+ "score": 1.0,
+ "content": "to module networks that can be trained end-to-end simply by backpropagation.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 105,
+ 197,
+ 506,
+ 397
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 402,
+ 505,
+ 566
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 415
+ ],
+ "score": 1.0,
+ "content": "We test the behavior of our new network on CLEVR and its associated datasets. On the primary",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 316,
+ 425
+ ],
+ "score": 1.0,
+ "content": "CLEVR reasoning task, we achieve an accuracy of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 412,
+ 343,
+ 423
+ ],
+ "score": 0.87,
+ "content": "9 8 . 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 344,
+ 412,
+ 505,
+ 425
+ ],
+ "score": 1.0,
+ "content": ", halving the error rate compared to the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 423,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": "previous state-of-the-art FiLM model (Perez et al., 2017). In particular, we show that our architec-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 434,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "ture yields better performance on questions involving counting and aggregation. In supplementary",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 104,
+ 443,
+ 506,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 443,
+ 506,
+ 460
+ ],
+ "score": 1.0,
+ "content": "studies, we show that the MAC network learns more quickly (both in terms of number of training",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 469
+ ],
+ "score": 1.0,
+ "content": "epochs and training time) and more effectively from limited amounts of training data. Moreover, it",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 466,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 314,
+ 480
+ ],
+ "score": 1.0,
+ "content": "also achieves a new state-of-the-art performance of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 467,
+ 342,
+ 478
+ ],
+ "score": 0.88,
+ "content": "8 2 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 466,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "on the more varied and difficult human-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "authored questions of the CLEVR-Humans dataset. The careful design of our cell encourages com-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 502
+ ],
+ "score": 1.0,
+ "content": "positionality, versatility and transparency. We achieve these properties by defining attention-based",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 512
+ ],
+ "score": 1.0,
+ "content": "interfaces that constrict the cell’s input and output spaces, and so constrain the interactions both",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "between and inside cells in order to guide them towards simple reasoning behaviors. Although",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "each cell’s functionality has only a limited range of possible continuous reasoning behaviors, when",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "chained together in a MAC network, the whole system becomes expressive and powerful. In the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 506,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 506,
+ 556
+ ],
+ "score": 1.0,
+ "content": "future, we believe that the architecture will also prove beneficial for other multi-step reasoning and",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 554,
+ 458,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 458,
+ 568
+ ],
+ "score": 1.0,
+ "content": "inference tasks, for instance in machine comprehension and textual question answering.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 104,
+ 400,
+ 506,
+ 568
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 591,
+ 211,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 213,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 213,
+ 606
+ ],
+ "score": 1.0,
+ "content": "2 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 43
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 621,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 622,
+ 505,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 622,
+ 505,
+ 634
+ ],
+ "score": 1.0,
+ "content": "There have been several prominent models that address the CLEVR task. By and large they can be",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "score": 1.0,
+ "content": "partitioned into two groups: module networks, which in practice have all used the strong supervision",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 643,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 643,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "provided in the form of tree-structured functional programs that accompany each data instance, and",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 655,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "large, relatively unstructured end-to-end differentiable networks that complement a fairly standard",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "stack of CNNs with components that aid in performing reasoning tasks. In contrast to modular",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 689
+ ],
+ "score": 1.0,
+ "content": "approaches (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017), our model does not",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "require additional supervision and makes use of a single computational cell chained in sequence",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "(like an LSTM) rather than a collection of custom modules deployed in a rigid tree structure. In",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "contrast to augmented CNN approaches (Santoro et al., 2017; Perez et al., 2017), we suggest that our",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "approach provides an ability for relational reasoning with better generalization capacity and higher",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "computational efficiency. These approaches and other related work are discussed and contrasted in",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 327,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 327,
+ 106
+ ],
+ "score": 1.0,
+ "content": "more detail in the supplementary material in section C.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 48.5,
+ "bbox_fs": [
+ 105,
+ 622,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "computational efficiency. These approaches and other related work are discussed and contrasted in",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 327,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 327,
+ 106
+ ],
+ "score": 1.0,
+ "content": "more detail in the supplementary material in section C.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 120,
+ 339,
+ 133
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 119,
+ 340,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 119,
+ 340,
+ 135
+ ],
+ "score": 1.0,
+ "content": "3 COMPOSITIONAL ATTENTION NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 145,
+ 505,
+ 178
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 157
+ ],
+ "score": 1.0,
+ "content": "Compositional Attention Networks is an end-to-end architecture for question-answering tasks that",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 156,
+ 505,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 505,
+ 169
+ ],
+ "score": 1.0,
+ "content": "sequentially performs an explicit reasoning process by stringing together small building blocks,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 399,
+ 181
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 399,
+ 181
+ ],
+ "score": 1.0,
+ "content": "called MAC cells, each is responsible for performing one reasoning step.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 183,
+ 504,
+ 228
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 505,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 505,
+ 196
+ ],
+ "score": 1.0,
+ "content": "We now provide an overview of the model, and a detailed discussion of the MAC cell. The model",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 207
+ ],
+ "score": 1.0,
+ "content": "is composed of three components: an Input unit, the core MAC network, and an output unit. A",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 219
+ ],
+ "score": 1.0,
+ "content": "TensorFlow implementation of the network, along with pretrained models will be made publicly",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 217,
+ 148,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 217,
+ 148,
+ 228
+ ],
+ "score": 1.0,
+ "content": "available.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 233,
+ 505,
+ 289
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "In this paper we explore the model in the context of VQA. However, it should be noted that while the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 505,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 505,
+ 258
+ ],
+ "score": 1.0,
+ "content": "input and output units are naturally domain-specific and should be designed to fit the task at hand,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 269
+ ],
+ "score": 1.0,
+ "content": "the MAC network has been designed to be generic and more broadly applicable, and may prove",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 267,
+ 505,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 267,
+ 505,
+ 280
+ ],
+ "score": 1.0,
+ "content": "useful in contexts beyond those explored in the paper, such as machine comprehension or question",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 278,
+ 475,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 278,
+ 475,
+ 290
+ ],
+ "score": 1.0,
+ "content": "answering over knowledge bases, which in our belief is a promising avenue for future work.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 302,
+ 204,
+ 313
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 205,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 205,
+ 315
+ ],
+ "score": 1.0,
+ "content": "3.1 THE INPUT UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 322,
+ 505,
+ 367
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 335
+ ],
+ "score": 1.0,
+ "content": "The input unit processes the raw inputs given to the system into distributed vector representations. It",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 346
+ ],
+ "score": 1.0,
+ "content": "receives a text question (or in general, a query), and an image (or in general, a Knowledge Base (KB))",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 505,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 505,
+ 357
+ ],
+ "score": 1.0,
+ "content": "and processes each of them with a matching sub-unit, for the query and the KB, here a biLSTM and",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 412,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 412,
+ 367
+ ],
+ "score": 1.0,
+ "content": "a CNN. More details can be found in the supplementary material, section A.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 372,
+ 505,
+ 450
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "At the end of this stage, we get from the query sub-unit a series of biLSTM output states, which we",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 224,
+ 396
+ ],
+ "score": 1.0,
+ "content": "refer to as contextual words,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 384,
+ 281,
+ 396
+ ],
+ "score": 0.92,
+ "content": "[ c w _ { 1 } , . . . , c w _ { S } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 383,
+ 312,
+ 396
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 313,
+ 384,
+ 321,
+ 393
+ ],
+ "score": 0.82,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 383,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "is the length of the question. In addition, we",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 122,
+ 407
+ ],
+ "score": 1.0,
+ "content": "get",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 122,
+ 394,
+ 185,
+ 407
+ ],
+ "score": 0.94,
+ "content": "q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 394,
+ 506,
+ 407
+ ],
+ "score": 1.0,
+ "content": ", the concatenation of the hidden states from the backward and forward LSTM",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 506,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 184,
+ 418
+ ],
+ "score": 1.0,
+ "content": "passes. We refer to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 408,
+ 190,
+ 417
+ ],
+ "score": 0.8,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 191,
+ 405,
+ 506,
+ 418
+ ],
+ "score": 1.0,
+ "content": "as the question representation. Furthermore, we get from the Knowledge-Base",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 428
+ ],
+ "score": 1.0,
+ "content": "sub-unit a static representation of the knowledge base. For the case of VQA, it will be represented",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 201,
+ 440
+ ],
+ "score": 1.0,
+ "content": "by a continuous matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 201,
+ 428,
+ 226,
+ 438
+ ],
+ "score": 0.91,
+ "content": "K B _ { V }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 226,
+ 427,
+ 282,
+ 440
+ ],
+ "score": 1.0,
+ "content": "of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 428,
+ 316,
+ 439
+ ],
+ "score": 0.88,
+ "content": "H , W , d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 427,
+ 346,
+ 440
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 346,
+ 428,
+ 405,
+ 438
+ ],
+ "score": 0.91,
+ "content": "H = W = 1 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 405,
+ 427,
+ 505,
+ 440
+ ],
+ "score": 1.0,
+ "content": "are the height and width",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 359,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 359,
+ 451
+ ],
+ "score": 1.0,
+ "content": "of the transformed image, corresponding to each of its regions.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 462,
+ 200,
+ 474
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 203,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 203,
+ 476
+ ],
+ "score": 1.0,
+ "content": "3.2 THE MAC CELL",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 483,
+ 505,
+ 560
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 505,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 505,
+ 496
+ ],
+ "score": 1.0,
+ "content": "The MAC network, which is the heart of our model, chains a sequence of small building blocks,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "called MAC cells, each responsible for performing one reasoning step. The model is provided access",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 505,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 505,
+ 519
+ ],
+ "score": 1.0,
+ "content": "to a Knowledge Base (KB), which is, for the specific case of VQA, the given image, and then upon",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 517,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 378,
+ 529
+ ],
+ "score": 1.0,
+ "content": "receiving a query, i.e. a question, the model iteratively focuses, in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 518,
+ 385,
+ 528
+ ],
+ "score": 0.78,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 386,
+ 517,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "steps, on the query’s various",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "parts, each reflects in turn the current reasoning step, which we term the control. Consequently,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "score": 1.0,
+ "content": "guided by this control, it retrieves the relevant information from the KB, that is then passed to the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 232,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 232,
+ 561
+ ],
+ "score": 1.0,
+ "content": "next cell in a recurrent fashion.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 566,
+ 504,
+ 610
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "Drawing inspiration from the Model-View-Controller paradigm used in software design and from",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 506,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 506,
+ 589
+ ],
+ "score": 1.0,
+ "content": "the commonly exercised separation between control and data paths in computer architecture, the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "MAC cell is composed of three units: control unit, read unit and write unit. Each has a clearly",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 599,
+ 462,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 462,
+ 611
+ ],
+ "score": 1.0,
+ "content": "defined role and an interface through which it interacts with the other units. See figure 2.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 616,
+ 504,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "The careful design and imposed interfaces that constrain the interaction between the units inside the",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "MAC cell, as described below, serve as structural prior that limits the space of hypotheses it can",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "learn, thereby guiding it towards acquiring the intended reasoning behaviors. As such, this prior",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 354,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 354,
+ 661
+ ],
+ "score": 1.0,
+ "content": "facilitates the learning process and mitigate overfitting issues.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "In particular, and similar in spirit to Perez et al. (2017), we allow the question to interact with the",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "Knowledge Base – the image for the case of VQA, only through indirect means: by guiding the cell",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "to attend to different elements in the KB, as well as controlling its operation through gating mecha-",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "nisms. Thus, in both cases, the interaction between these mediums, visual and textual, or knowledge",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "and query, is mediated through probability distributions, either in the form of attention maps, or as",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "gates, further detailed below. This stands in stark contrast to many common approaches that fuse the",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 45.5
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 303,
+ 751,
+ 309,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [],
+ "index": 0.5,
+ "bbox_fs": [
+ 106,
+ 82,
+ 505,
+ 106
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 120,
+ 339,
+ 133
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 119,
+ 340,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 119,
+ 340,
+ 135
+ ],
+ "score": 1.0,
+ "content": "3 COMPOSITIONAL ATTENTION NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 145,
+ 505,
+ 178
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 157
+ ],
+ "score": 1.0,
+ "content": "Compositional Attention Networks is an end-to-end architecture for question-answering tasks that",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 156,
+ 505,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 505,
+ 169
+ ],
+ "score": 1.0,
+ "content": "sequentially performs an explicit reasoning process by stringing together small building blocks,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 399,
+ 181
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 399,
+ 181
+ ],
+ "score": 1.0,
+ "content": "called MAC cells, each is responsible for performing one reasoning step.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 105,
+ 145,
+ 505,
+ 181
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 183,
+ 504,
+ 228
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 505,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 505,
+ 196
+ ],
+ "score": 1.0,
+ "content": "We now provide an overview of the model, and a detailed discussion of the MAC cell. The model",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 207
+ ],
+ "score": 1.0,
+ "content": "is composed of three components: an Input unit, the core MAC network, and an output unit. A",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 205,
+ 506,
+ 219
+ ],
+ "score": 1.0,
+ "content": "TensorFlow implementation of the network, along with pretrained models will be made publicly",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 217,
+ 148,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 217,
+ 148,
+ 228
+ ],
+ "score": 1.0,
+ "content": "available.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 105,
+ 184,
+ 506,
+ 228
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 233,
+ 505,
+ 289
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "In this paper we explore the model in the context of VQA. However, it should be noted that while the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 505,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 505,
+ 258
+ ],
+ "score": 1.0,
+ "content": "input and output units are naturally domain-specific and should be designed to fit the task at hand,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 269
+ ],
+ "score": 1.0,
+ "content": "the MAC network has been designed to be generic and more broadly applicable, and may prove",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 267,
+ 505,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 267,
+ 505,
+ 280
+ ],
+ "score": 1.0,
+ "content": "useful in contexts beyond those explored in the paper, such as machine comprehension or question",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 278,
+ 475,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 278,
+ 475,
+ 290
+ ],
+ "score": 1.0,
+ "content": "answering over knowledge bases, which in our belief is a promising avenue for future work.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 105,
+ 234,
+ 505,
+ 290
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 302,
+ 204,
+ 313
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 205,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 205,
+ 315
+ ],
+ "score": 1.0,
+ "content": "3.1 THE INPUT UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 322,
+ 505,
+ 367
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 335
+ ],
+ "score": 1.0,
+ "content": "The input unit processes the raw inputs given to the system into distributed vector representations. It",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 346
+ ],
+ "score": 1.0,
+ "content": "receives a text question (or in general, a query), and an image (or in general, a Knowledge Base (KB))",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 505,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 505,
+ 357
+ ],
+ "score": 1.0,
+ "content": "and processes each of them with a matching sub-unit, for the query and the KB, here a biLSTM and",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 412,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 412,
+ 367
+ ],
+ "score": 1.0,
+ "content": "a CNN. More details can be found in the supplementary material, section A.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 105,
+ 322,
+ 506,
+ 367
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 372,
+ 505,
+ 450
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "At the end of this stage, we get from the query sub-unit a series of biLSTM output states, which we",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 224,
+ 396
+ ],
+ "score": 1.0,
+ "content": "refer to as contextual words,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 384,
+ 281,
+ 396
+ ],
+ "score": 0.92,
+ "content": "[ c w _ { 1 } , . . . , c w _ { S } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 383,
+ 312,
+ 396
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 313,
+ 384,
+ 321,
+ 393
+ ],
+ "score": 0.82,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 383,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "is the length of the question. In addition, we",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 506,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 122,
+ 407
+ ],
+ "score": 1.0,
+ "content": "get",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 122,
+ 394,
+ 185,
+ 407
+ ],
+ "score": 0.94,
+ "content": "q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 394,
+ 506,
+ 407
+ ],
+ "score": 1.0,
+ "content": ", the concatenation of the hidden states from the backward and forward LSTM",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 506,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 184,
+ 418
+ ],
+ "score": 1.0,
+ "content": "passes. We refer to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 408,
+ 190,
+ 417
+ ],
+ "score": 0.8,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 191,
+ 405,
+ 506,
+ 418
+ ],
+ "score": 1.0,
+ "content": "as the question representation. Furthermore, we get from the Knowledge-Base",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 428
+ ],
+ "score": 1.0,
+ "content": "sub-unit a static representation of the knowledge base. For the case of VQA, it will be represented",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 427,
+ 505,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 201,
+ 440
+ ],
+ "score": 1.0,
+ "content": "by a continuous matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 201,
+ 428,
+ 226,
+ 438
+ ],
+ "score": 0.91,
+ "content": "K B _ { V }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 226,
+ 427,
+ 282,
+ 440
+ ],
+ "score": 1.0,
+ "content": "of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 428,
+ 316,
+ 439
+ ],
+ "score": 0.88,
+ "content": "H , W , d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 427,
+ 346,
+ 440
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 346,
+ 428,
+ 405,
+ 438
+ ],
+ "score": 0.91,
+ "content": "H = W = 1 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 405,
+ 427,
+ 505,
+ 440
+ ],
+ "score": 1.0,
+ "content": "are the height and width",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 359,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 359,
+ 451
+ ],
+ "score": 1.0,
+ "content": "of the transformed image, corresponding to each of its regions.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 105,
+ 372,
+ 506,
+ 451
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 462,
+ 200,
+ 474
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 203,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 203,
+ 476
+ ],
+ "score": 1.0,
+ "content": "3.2 THE MAC CELL",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 483,
+ 505,
+ 560
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 505,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 505,
+ 496
+ ],
+ "score": 1.0,
+ "content": "The MAC network, which is the heart of our model, chains a sequence of small building blocks,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "called MAC cells, each responsible for performing one reasoning step. The model is provided access",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 505,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 505,
+ 519
+ ],
+ "score": 1.0,
+ "content": "to a Knowledge Base (KB), which is, for the specific case of VQA, the given image, and then upon",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 517,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 378,
+ 529
+ ],
+ "score": 1.0,
+ "content": "receiving a query, i.e. a question, the model iteratively focuses, in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 518,
+ 385,
+ 528
+ ],
+ "score": 0.78,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 386,
+ 517,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "steps, on the query’s various",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "parts, each reflects in turn the current reasoning step, which we term the control. Consequently,",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "score": 1.0,
+ "content": "guided by this control, it retrieves the relevant information from the KB, that is then passed to the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 232,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 232,
+ 561
+ ],
+ "score": 1.0,
+ "content": "next cell in a recurrent fashion.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 105,
+ 483,
+ 506,
+ 561
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 566,
+ 504,
+ 610
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "Drawing inspiration from the Model-View-Controller paradigm used in software design and from",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 506,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 506,
+ 589
+ ],
+ "score": 1.0,
+ "content": "the commonly exercised separation between control and data paths in computer architecture, the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "MAC cell is composed of three units: control unit, read unit and write unit. Each has a clearly",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 599,
+ 462,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 462,
+ 611
+ ],
+ "score": 1.0,
+ "content": "defined role and an interface through which it interacts with the other units. See figure 2.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 105,
+ 566,
+ 506,
+ 611
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 616,
+ 504,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "The careful design and imposed interfaces that constrain the interaction between the units inside the",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "MAC cell, as described below, serve as structural prior that limits the space of hypotheses it can",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "learn, thereby guiding it towards acquiring the intended reasoning behaviors. As such, this prior",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 354,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 354,
+ 661
+ ],
+ "score": 1.0,
+ "content": "facilitates the learning process and mitigate overfitting issues.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 615,
+ 505,
+ 661
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "In particular, and similar in spirit to Perez et al. (2017), we allow the question to interact with the",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "Knowledge Base – the image for the case of VQA, only through indirect means: by guiding the cell",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "to attend to different elements in the KB, as well as controlling its operation through gating mecha-",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "nisms. Thus, in both cases, the interaction between these mediums, visual and textual, or knowledge",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "and query, is mediated through probability distributions, either in the form of attention maps, or as",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "gates, further detailed below. This stands in stark contrast to many common approaches that fuse the",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 251,
+ 505,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 251,
+ 505,
+ 265
+ ],
+ "score": 1.0,
+ "content": "question and image together into the same vector space through linear combinations, multiplication,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 262,
+ 505,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 262,
+ 505,
+ 275
+ ],
+ "score": 1.0,
+ "content": "or concatenation. Rather, our controlled interaction distills the influence that the query should have",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 274,
+ 473,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 274,
+ 473,
+ 286
+ ],
+ "score": 1.0,
+ "content": "in processing the Knowledge Base, casting it onto discrete probability distributions instead.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 45.5,
+ "bbox_fs": [
+ 105,
+ 666,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 215,
+ 82,
+ 370,
+ 198
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 215,
+ 82,
+ 370,
+ 198
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 215,
+ 82,
+ 370,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 215,
+ 82,
+ 370,
+ 198
+ ],
+ "score": 0.966,
+ "type": "image",
+ "image_path": "21ef6959e8f5c66e79f26b2aa043aef1bf62c1ad19d7a86079e5194974b914be.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 215,
+ 82,
+ 370,
+ 96.5
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 215,
+ 96.5,
+ 370,
+ 111.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 215,
+ 111.0,
+ 370,
+ 125.5
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 215,
+ 125.5,
+ 370,
+ 140.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 215,
+ 140.0,
+ 370,
+ 154.5
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 215,
+ 154.5,
+ 370,
+ 169.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 215,
+ 169.0,
+ 370,
+ 183.5
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 215,
+ 183.5,
+ 370,
+ 198.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 208,
+ 505,
+ 231
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 209,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 209,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "Figure 2: The MAC cell, which is a recurrent unit comprised of a Control Unit, Read Unit, and",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "Write Unit. Blue shows the control flow and red shows the memory flow. See section 3.2 for details.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5
+ }
+ ],
+ "index": 6.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 252,
+ 504,
+ 285
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 251,
+ 505,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 251,
+ 505,
+ 265
+ ],
+ "score": 1.0,
+ "content": "question and image together into the same vector space through linear combinations, multiplication,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 262,
+ 505,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 262,
+ 505,
+ 275
+ ],
+ "score": 1.0,
+ "content": "or concatenation. Rather, our controlled interaction distills the influence that the query should have",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 274,
+ 473,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 274,
+ 473,
+ 286
+ ],
+ "score": 1.0,
+ "content": "in processing the Knowledge Base, casting it onto discrete probability distributions instead.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 290,
+ 505,
+ 368
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 290,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 290,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "The MAC cell has been designed to replace the discrete and predefined “modules” used in the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "score": 1.0,
+ "content": "modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017). Rather, we",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "create one universal and versatile cell that is applied across all the reasoning steps, sharing both",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 336
+ ],
+ "score": 1.0,
+ "content": "its architecture as well as its parameters, across all of its instantiations. In contrast to the discrete",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": "modules, each trained to specialize to some specific elementary reasoning task, the MAC cell is",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 505,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 505,
+ 357
+ ],
+ "score": 1.0,
+ "content": "capable of demonstrating a continuous range of possible reasoning behaviors conditioned on the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 427,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 427,
+ 369
+ ],
+ "score": 1.0,
+ "content": "context in which it is applied – namely, the inputs it receives from the prior cell.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 373,
+ 505,
+ 472
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 373,
+ 145,
+ 385
+ ],
+ "score": 1.0,
+ "content": "Each cell",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 373,
+ 176,
+ 384
+ ],
+ "score": 0.91,
+ "content": "M A C _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 373,
+ 314,
+ 385
+ ],
+ "score": 1.0,
+ "content": "maintains two dual states: control",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 375,
+ 323,
+ 384
+ ],
+ "score": 0.85,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 324,
+ 373,
+ 375,
+ 385
+ ],
+ "score": 1.0,
+ "content": "and memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 375,
+ 389,
+ 384
+ ],
+ "score": 0.85,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 389,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": ", both are continuous vectors",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 384,
+ 504,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 384,
+ 161,
+ 396
+ ],
+ "score": 1.0,
+ "content": "of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 162,
+ 385,
+ 168,
+ 394
+ ],
+ "score": 0.7,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 384,
+ 221,
+ 396
+ ],
+ "score": 1.0,
+ "content": ". The control",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 386,
+ 231,
+ 395
+ ],
+ "score": 0.84,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 384,
+ 504,
+ 396
+ ],
+ "score": 1.0,
+ "content": "represents the reasoning operation the MAC cell should accomplish",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "in the current step – focusing only on some aspect of the whole question. This is represented by",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 430,
+ 419
+ ],
+ "score": 1.0,
+ "content": "a weighted-average attention-based sum of the question words. The memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 408,
+ 443,
+ 417
+ ],
+ "score": 0.85,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 444,
+ 406,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "represents the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 431
+ ],
+ "score": 1.0,
+ "content": "current context information deemed relevant to respond to the query, or answer the question.This is",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "represented practically by a weighted average over elements from the KB, or for the case of VQA,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 194,
+ 452
+ ],
+ "score": 1.0,
+ "content": "regions in the image.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 441,
+ 209,
+ 450
+ ],
+ "score": 0.85,
+ "content": "m _ { 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 210,
+ 439,
+ 227,
+ 452
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 228,
+ 441,
+ 238,
+ 450
+ ],
+ "score": 0.85,
+ "content": "c _ { 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 239,
+ 439,
+ 495,
+ 452
+ ],
+ "score": 1.0,
+ "content": "are initialized each to a random vector parameter of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 439,
+ 501,
+ 449
+ ],
+ "score": 0.73,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 439,
+ 506,
+ 452
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 449,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 449,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "The memory and control states are passed from one cell to the next in a recurrent fashion, and used",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 461,
+ 484,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 484,
+ 474
+ ],
+ "score": 1.0,
+ "content": "in a way reminiscent of Key-Value memory networks (Miller et al., 2016), as discussed below.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 484,
+ 228,
+ 496
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 230,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 230,
+ 497
+ ],
+ "score": 1.0,
+ "content": "3.2.1 THE CONTROL UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 503,
+ 505,
+ 536
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "The control unit determines the reasoning operation that should be applied at this step. It receives",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 514,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 195,
+ 527
+ ],
+ "score": 1.0,
+ "content": "the contextual words",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 514,
+ 252,
+ 526
+ ],
+ "score": 0.92,
+ "content": "[ c w _ { 1 } , . . . , c w _ { S } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 514,
+ 369,
+ 527
+ ],
+ "score": 1.0,
+ "content": ", the question representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 370,
+ 516,
+ 376,
+ 525
+ ],
+ "score": 0.76,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 514,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": ", and the control state from the",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 369,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 184,
+ 537
+ ],
+ "score": 1.0,
+ "content": "previous MAC cell",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 527,
+ 204,
+ 537
+ ],
+ "score": 0.88,
+ "content": "c _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 525,
+ 358,
+ 537
+ ],
+ "score": 1.0,
+ "content": ", all of which are vectors of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 526,
+ 365,
+ 535
+ ],
+ "score": 0.76,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 525,
+ 369,
+ 537
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 542,
+ 505,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": "We would like to allow our MAC cell to perform continuously varied and adaptive range of behav-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 553,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 553,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "iors, as demanded by the question. Therefore, we define the behavior of each cell to be a function of",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 193,
+ 577
+ ],
+ "score": 1.0,
+ "content": "the contextual words",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 564,
+ 250,
+ 576
+ ],
+ "score": 0.92,
+ "content": "[ c w _ { 1 } , . . . , c w _ { S } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 251,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": ", weighted-averaged according to the attention distribution that",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "the control unit produces at each step. This will allow the cell to adapt its behavior – the reason-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "ing operation it performs – to the question it receives, instead of having a fixed set of predefined",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 500,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 500,
+ 610
+ ],
+ "score": 1.0,
+ "content": "behaviours as is the case in competing approaches Andreas et al. (2016a;b); Johnson et al. (2017).",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 35.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 614,
+ 505,
+ 648
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 428,
+ 627
+ ],
+ "score": 1.0,
+ "content": "The formal specification of the control unit is shown in figure 3. The question",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 616,
+ 435,
+ 626
+ ],
+ "score": 0.8,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 613,
+ 505,
+ 627
+ ],
+ "score": 1.0,
+ "content": "is linearly trans-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 626,
+ 505,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 626,
+ 194,
+ 637
+ ],
+ "score": 1.0,
+ "content": "formed into a vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 627,
+ 204,
+ 636
+ ],
+ "score": 0.84,
+ "content": "q _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 626,
+ 505,
+ 637
+ ],
+ "score": 1.0,
+ "content": "of the same dimension, which in turn is concatenated with the previous",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 635,
+ 425,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 158,
+ 650
+ ],
+ "score": 1.0,
+ "content": "control state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 638,
+ 177,
+ 648
+ ],
+ "score": 0.88,
+ "content": "c _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 635,
+ 320,
+ 650
+ ],
+ "score": 1.0,
+ "content": "and linearly transformed again to a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 320,
+ 636,
+ 326,
+ 646
+ ],
+ "score": 0.79,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 327,
+ 635,
+ 407,
+ 650
+ ],
+ "score": 1.0,
+ "content": "-dimensional vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 407,
+ 638,
+ 420,
+ 648
+ ],
+ "score": 0.88,
+ "content": "c q _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 635,
+ 425,
+ 650
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 250,
+ 665,
+ 362,
+ 700
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 250,
+ 665,
+ 362,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 250,
+ 665,
+ 362,
+ 700
+ ],
+ "score": 0.91,
+ "content": "\\begin{array} { c } { q _ { i } = W _ { i } ^ { d , d } \\cdot q + b _ { i } ^ { d } } \\\\ { c q _ { i } = W ^ { 2 d , d } \\left[ q _ { i } , c _ { i - 1 } \\right] + b ^ { d } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "629b6ac32640eb52fb3021fd3b45ba817455f27666f7d51a163756755ee964e5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 42.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 250,
+ 665,
+ 362,
+ 682.5
+ ],
+ "spans": [],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 250,
+ 682.5,
+ 362,
+ 700.0
+ ],
+ "spans": [],
+ "index": 43
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 707,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 707,
+ 505,
+ 720
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 707,
+ 505,
+ 720
+ ],
+ "score": 1.0,
+ "content": "Note that in contrast to all other parameters of the cell, which are shared across its instantiations",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 103,
+ 717,
+ 507,
+ 735
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 717,
+ 191,
+ 735
+ ],
+ "score": 1.0,
+ "content": "at the different steps",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 721,
+ 237,
+ 732
+ ],
+ "score": 0.91,
+ "content": "i = 1 , . . . , p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 717,
+ 304,
+ 735
+ ],
+ "score": 1.0,
+ "content": ", the parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 720,
+ 327,
+ 733
+ ],
+ "score": 0.93,
+ "content": "W _ { i } ^ { d , d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 327,
+ 717,
+ 346,
+ 735
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 346,
+ 721,
+ 356,
+ 733
+ ],
+ "score": 0.89,
+ "content": "b _ { i } ^ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 356,
+ 717,
+ 507,
+ 735
+ ],
+ "score": 1.0,
+ "content": "are different for each iteration. This",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 44.5
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "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 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 759
+ ],
+ "lines": []
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 215,
+ 82,
+ 370,
+ 198
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 215,
+ 82,
+ 370,
+ 198
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 215,
+ 82,
+ 370,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 215,
+ 82,
+ 370,
+ 198
+ ],
+ "score": 0.966,
+ "type": "image",
+ "image_path": "21ef6959e8f5c66e79f26b2aa043aef1bf62c1ad19d7a86079e5194974b914be.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 215,
+ 82,
+ 370,
+ 96.5
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 215,
+ 96.5,
+ 370,
+ 111.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 215,
+ 111.0,
+ 370,
+ 125.5
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 215,
+ 125.5,
+ 370,
+ 140.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 215,
+ 140.0,
+ 370,
+ 154.5
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 215,
+ 154.5,
+ 370,
+ 169.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 215,
+ 169.0,
+ 370,
+ 183.5
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 215,
+ 183.5,
+ 370,
+ 198.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 208,
+ 505,
+ 231
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 209,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 209,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "Figure 2: The MAC cell, which is a recurrent unit comprised of a Control Unit, Read Unit, and",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "Write Unit. Blue shows the control flow and red shows the memory flow. See section 3.2 for details.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5
+ }
+ ],
+ "index": 6.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 252,
+ 504,
+ 285
+ ],
+ "lines": [],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 251,
+ 505,
+ 286
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 290,
+ 505,
+ 368
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 290,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 290,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "The MAC cell has been designed to replace the discrete and predefined “modules” used in the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "score": 1.0,
+ "content": "modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017). Rather, we",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "create one universal and versatile cell that is applied across all the reasoning steps, sharing both",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 336
+ ],
+ "score": 1.0,
+ "content": "its architecture as well as its parameters, across all of its instantiations. In contrast to the discrete",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": "modules, each trained to specialize to some specific elementary reasoning task, the MAC cell is",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 505,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 505,
+ 357
+ ],
+ "score": 1.0,
+ "content": "capable of demonstrating a continuous range of possible reasoning behaviors conditioned on the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 427,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 427,
+ 369
+ ],
+ "score": 1.0,
+ "content": "context in which it is applied – namely, the inputs it receives from the prior cell.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 105,
+ 290,
+ 505,
+ 369
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 373,
+ 505,
+ 472
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 373,
+ 145,
+ 385
+ ],
+ "score": 1.0,
+ "content": "Each cell",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 373,
+ 176,
+ 384
+ ],
+ "score": 0.91,
+ "content": "M A C _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 373,
+ 314,
+ 385
+ ],
+ "score": 1.0,
+ "content": "maintains two dual states: control",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 375,
+ 323,
+ 384
+ ],
+ "score": 0.85,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 324,
+ 373,
+ 375,
+ 385
+ ],
+ "score": 1.0,
+ "content": "and memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 375,
+ 389,
+ 384
+ ],
+ "score": 0.85,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 389,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": ", both are continuous vectors",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 384,
+ 504,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 384,
+ 161,
+ 396
+ ],
+ "score": 1.0,
+ "content": "of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 162,
+ 385,
+ 168,
+ 394
+ ],
+ "score": 0.7,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 384,
+ 221,
+ 396
+ ],
+ "score": 1.0,
+ "content": ". The control",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 386,
+ 231,
+ 395
+ ],
+ "score": 0.84,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 384,
+ 504,
+ 396
+ ],
+ "score": 1.0,
+ "content": "represents the reasoning operation the MAC cell should accomplish",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "in the current step – focusing only on some aspect of the whole question. This is represented by",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 430,
+ 419
+ ],
+ "score": 1.0,
+ "content": "a weighted-average attention-based sum of the question words. The memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 408,
+ 443,
+ 417
+ ],
+ "score": 0.85,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 444,
+ 406,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "represents the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 431
+ ],
+ "score": 1.0,
+ "content": "current context information deemed relevant to respond to the query, or answer the question.This is",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "represented practically by a weighted average over elements from the KB, or for the case of VQA,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 194,
+ 452
+ ],
+ "score": 1.0,
+ "content": "regions in the image.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 441,
+ 209,
+ 450
+ ],
+ "score": 0.85,
+ "content": "m _ { 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 210,
+ 439,
+ 227,
+ 452
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 228,
+ 441,
+ 238,
+ 450
+ ],
+ "score": 0.85,
+ "content": "c _ { 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 239,
+ 439,
+ 495,
+ 452
+ ],
+ "score": 1.0,
+ "content": "are initialized each to a random vector parameter of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 439,
+ 501,
+ 449
+ ],
+ "score": 0.73,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 439,
+ 506,
+ 452
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 449,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 449,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "The memory and control states are passed from one cell to the next in a recurrent fashion, and used",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 461,
+ 484,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 484,
+ 474
+ ],
+ "score": 1.0,
+ "content": "in a way reminiscent of Key-Value memory networks (Miller et al., 2016), as discussed below.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 105,
+ 373,
+ 506,
+ 474
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 484,
+ 228,
+ 496
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 230,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 230,
+ 497
+ ],
+ "score": 1.0,
+ "content": "3.2.1 THE CONTROL UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 503,
+ 505,
+ 536
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "The control unit determines the reasoning operation that should be applied at this step. It receives",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 514,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 195,
+ 527
+ ],
+ "score": 1.0,
+ "content": "the contextual words",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 514,
+ 252,
+ 526
+ ],
+ "score": 0.92,
+ "content": "[ c w _ { 1 } , . . . , c w _ { S } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 514,
+ 369,
+ 527
+ ],
+ "score": 1.0,
+ "content": ", the question representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 370,
+ 516,
+ 376,
+ 525
+ ],
+ "score": 0.76,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 514,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": ", and the control state from the",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 369,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 184,
+ 537
+ ],
+ "score": 1.0,
+ "content": "previous MAC cell",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 527,
+ 204,
+ 537
+ ],
+ "score": 0.88,
+ "content": "c _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 525,
+ 358,
+ 537
+ ],
+ "score": 1.0,
+ "content": ", all of which are vectors of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 526,
+ 365,
+ 535
+ ],
+ "score": 0.76,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 525,
+ 369,
+ 537
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 105,
+ 502,
+ 505,
+ 537
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 542,
+ 505,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": "We would like to allow our MAC cell to perform continuously varied and adaptive range of behav-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 553,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 553,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "iors, as demanded by the question. Therefore, we define the behavior of each cell to be a function of",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 193,
+ 577
+ ],
+ "score": 1.0,
+ "content": "the contextual words",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 564,
+ 250,
+ 576
+ ],
+ "score": 0.92,
+ "content": "[ c w _ { 1 } , . . . , c w _ { S } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 251,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": ", weighted-averaged according to the attention distribution that",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "the control unit produces at each step. This will allow the cell to adapt its behavior – the reason-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "ing operation it performs – to the question it receives, instead of having a fixed set of predefined",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 500,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 500,
+ 610
+ ],
+ "score": 1.0,
+ "content": "behaviours as is the case in competing approaches Andreas et al. (2016a;b); Johnson et al. (2017).",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 35.5,
+ "bbox_fs": [
+ 105,
+ 541,
+ 506,
+ 610
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 614,
+ 505,
+ 648
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 428,
+ 627
+ ],
+ "score": 1.0,
+ "content": "The formal specification of the control unit is shown in figure 3. The question",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 616,
+ 435,
+ 626
+ ],
+ "score": 0.8,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 613,
+ 505,
+ 627
+ ],
+ "score": 1.0,
+ "content": "is linearly trans-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 626,
+ 505,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 626,
+ 194,
+ 637
+ ],
+ "score": 1.0,
+ "content": "formed into a vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 627,
+ 204,
+ 636
+ ],
+ "score": 0.84,
+ "content": "q _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 626,
+ 505,
+ 637
+ ],
+ "score": 1.0,
+ "content": "of the same dimension, which in turn is concatenated with the previous",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 635,
+ 425,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 158,
+ 650
+ ],
+ "score": 1.0,
+ "content": "control state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 638,
+ 177,
+ 648
+ ],
+ "score": 0.88,
+ "content": "c _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 635,
+ 320,
+ 650
+ ],
+ "score": 1.0,
+ "content": "and linearly transformed again to a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 320,
+ 636,
+ 326,
+ 646
+ ],
+ "score": 0.79,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 327,
+ 635,
+ 407,
+ 650
+ ],
+ "score": 1.0,
+ "content": "-dimensional vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 407,
+ 638,
+ 420,
+ 648
+ ],
+ "score": 0.88,
+ "content": "c q _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 635,
+ 425,
+ 650
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 105,
+ 613,
+ 505,
+ 650
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 250,
+ 665,
+ 362,
+ 700
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 250,
+ 665,
+ 362,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 250,
+ 665,
+ 362,
+ 700
+ ],
+ "score": 0.91,
+ "content": "\\begin{array} { c } { q _ { i } = W _ { i } ^ { d , d } \\cdot q + b _ { i } ^ { d } } \\\\ { c q _ { i } = W ^ { 2 d , d } \\left[ q _ { i } , c _ { i - 1 } \\right] + b ^ { d } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "629b6ac32640eb52fb3021fd3b45ba817455f27666f7d51a163756755ee964e5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 42.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 250,
+ 665,
+ 362,
+ 682.5
+ ],
+ "spans": [],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 250,
+ 682.5,
+ 362,
+ 700.0
+ ],
+ "spans": [],
+ "index": 43
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 707,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 707,
+ 505,
+ 720
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 707,
+ 505,
+ 720
+ ],
+ "score": 1.0,
+ "content": "Note that in contrast to all other parameters of the cell, which are shared across its instantiations",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 103,
+ 717,
+ 507,
+ 735
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 717,
+ 191,
+ 735
+ ],
+ "score": 1.0,
+ "content": "at the different steps",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 721,
+ 237,
+ 732
+ ],
+ "score": 0.91,
+ "content": "i = 1 , . . . , p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 717,
+ 304,
+ 735
+ ],
+ "score": 1.0,
+ "content": ", the parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 720,
+ 327,
+ 733
+ ],
+ "score": 0.93,
+ "content": "W _ { i } ^ { d , d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 327,
+ 717,
+ 346,
+ 735
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 346,
+ 721,
+ 356,
+ 733
+ ],
+ "score": 0.89,
+ "content": "b _ { i } ^ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 356,
+ 717,
+ 507,
+ 735
+ ],
+ "score": 1.0,
+ "content": "are different for each iteration. This",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "is done to allow each cell to attend more readily to different aspects (i.e. parts) of the questions,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "depending on the index of the current step – its relative stage in the context of the whole reasoning",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 142,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 142,
+ 245
+ ],
+ "score": 1.0,
+ "content": "process.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 44.5,
+ "bbox_fs": [
+ 103,
+ 707,
+ 507,
+ 735
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 202,
+ 81,
+ 406,
+ 154
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 202,
+ 81,
+ 406,
+ 154
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 202,
+ 81,
+ 406,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 202,
+ 81,
+ 406,
+ 154
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "432535676eeb69a82d7848d82a8a327d906d019f07db47ee24440c80b3beab34.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 202,
+ 81,
+ 406,
+ 95.6
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 202,
+ 95.6,
+ 406,
+ 110.19999999999999
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 202,
+ 110.19999999999999,
+ 406,
+ 124.79999999999998
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 202,
+ 124.79999999999998,
+ 406,
+ 139.39999999999998
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 202,
+ 139.39999999999998,
+ 406,
+ 153.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 164,
+ 504,
+ 186
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 506,
+ 176
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 506,
+ 176
+ ],
+ "score": 1.0,
+ "content": "Figure 3: The Control Unit (CU) of the MAC cell. See section 3.2.1 for details. Best viewed in",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 173,
+ 133,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 173,
+ 133,
+ 188
+ ],
+ "score": 1.0,
+ "content": "color.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 209,
+ 505,
+ 243
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "is done to allow each cell to attend more readily to different aspects (i.e. parts) of the questions,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "depending on the index of the current step – its relative stage in the context of the whole reasoning",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 142,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 142,
+ 245
+ ],
+ "score": 1.0,
+ "content": "process.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 248,
+ 505,
+ 282
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 248,
+ 505,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 250,
+ 120,
+ 260
+ ],
+ "score": 0.84,
+ "content": "c q _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 120,
+ 248,
+ 505,
+ 261
+ ],
+ "score": 1.0,
+ "content": "represents the current reasoning operation we would like to perform in a continuous way, taking",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 321,
+ 271
+ ],
+ "score": 1.0,
+ "content": "into account both the overall meaning of the question",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 321,
+ 261,
+ 330,
+ 271
+ ],
+ "score": 0.83,
+ "content": "q _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 330,
+ 259,
+ 505,
+ 271
+ ],
+ "score": 1.0,
+ "content": ", as well as the words the model attended to",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 213,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 189,
+ 284
+ ],
+ "score": 1.0,
+ "content": "in the previous step,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 272,
+ 208,
+ 282
+ ],
+ "score": 0.86,
+ "content": "c _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 270,
+ 213,
+ 284
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 364
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 506,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 506,
+ 300
+ ],
+ "score": 1.0,
+ "content": "However, we would like to prevent the cell from diverging in the reasoning operations it tries to",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 311
+ ],
+ "score": 1.0,
+ "content": "perform, and instead anchor it back in the question words, by using them to represent the reasoning",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 308,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 469,
+ 322
+ ],
+ "score": 1.0,
+ "content": "operation of the current step. We can achieve that by computing an attention distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 311,
+ 483,
+ 320
+ ],
+ "score": 0.85,
+ "content": "c v _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 308,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "over",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 320,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 193,
+ 333
+ ],
+ "score": 1.0,
+ "content": "the contextual words",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 320,
+ 250,
+ 332
+ ],
+ "score": 0.93,
+ "content": "[ c w _ { 1 } , . . . , c w _ { S } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 320,
+ 362,
+ 333
+ ],
+ "score": 1.0,
+ "content": "based on their similarity to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 362,
+ 321,
+ 375,
+ 332
+ ],
+ "score": 0.87,
+ "content": "c q _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 320,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": ". Then, summing the contextual",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 331,
+ 505,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 331,
+ 285,
+ 343
+ ],
+ "score": 1.0,
+ "content": "words according to the attention distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 332,
+ 300,
+ 342
+ ],
+ "score": 0.87,
+ "content": "c v _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 331,
+ 465,
+ 343
+ ],
+ "score": 1.0,
+ "content": "will allow us to have a new control state,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 333,
+ 474,
+ 342
+ ],
+ "score": 0.83,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 475,
+ 331,
+ 505,
+ 343
+ ],
+ "score": 1.0,
+ "content": ", which",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 354
+ ],
+ "score": 1.0,
+ "content": "is represented again in terms of words from the question. Intuitively, it is the gist of the question that",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 104,
+ 351,
+ 436,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 351,
+ 436,
+ 367
+ ],
+ "score": 1.0,
+ "content": "is relevant to the reasoning operation we would like to perform in the current step.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 223,
+ 385,
+ 387,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 223,
+ 385,
+ 387,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 223,
+ 385,
+ 387,
+ 437
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { c l c r } { { c v _ { i , s } = \\displaystyle \\mathrm { s o f t m a x } \\big ( W ^ { d , 1 } \\big ( c q _ { i } \\circ c w _ { s } \\big ) + b ^ { 1 } \\big ) } } \\\\ { { \\big . } } \\\\ { { c _ { i } = \\displaystyle \\sum _ { s = 1 } ^ { S } c v _ { i , s } \\cdot c w _ { s } } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "d43f433dbdaf70d728243b740937cc4b7c74be9d6cda3d636612e146da07d190.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 223,
+ 385,
+ 387,
+ 398.0
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 223,
+ 398.0,
+ 387,
+ 411.0
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 223,
+ 411.0,
+ 387,
+ 424.0
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 223,
+ 424.0,
+ 387,
+ 437.0
+ ],
+ "spans": [],
+ "index": 23
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 448,
+ 504,
+ 470
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 448,
+ 505,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 448,
+ 329,
+ 460
+ ],
+ "score": 1.0,
+ "content": "Finally, the control unit returns the current control state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 451,
+ 338,
+ 459
+ ],
+ "score": 0.86,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 448,
+ 455,
+ 460
+ ],
+ "score": 1.0,
+ "content": ", along with an attention map",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 451,
+ 469,
+ 459
+ ],
+ "score": 0.85,
+ "content": "c v _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 448,
+ 505,
+ 460
+ ],
+ "score": 1.0,
+ "content": "over the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 459,
+ 179,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 459,
+ 179,
+ 470
+ ],
+ "score": 1.0,
+ "content": "contextual words.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 483,
+ 211,
+ 495
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 213,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 213,
+ 497
+ ],
+ "score": 1.0,
+ "content": "3.2.2 THE READ UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 503,
+ 505,
+ 547
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 347,
+ 516
+ ],
+ "score": 1.0,
+ "content": "The Read Unit is provided with access to the knowledge base",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 348,
+ 504,
+ 372,
+ 514
+ ],
+ "score": 0.9,
+ "content": "K B _ { V }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 372,
+ 502,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": ", along with the previous memory",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 515,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 128,
+ 527
+ ],
+ "score": 1.0,
+ "content": "state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 516,
+ 152,
+ 526
+ ],
+ "score": 0.91,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 153,
+ 515,
+ 253,
+ 527
+ ],
+ "score": 1.0,
+ "content": "and the current control",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 253,
+ 516,
+ 262,
+ 525
+ ],
+ "score": 0.83,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 515,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": ". It is responsible for retrieving relevant content from the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 178,
+ 538
+ ],
+ "score": 1.0,
+ "content": "Knowledge Base",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 525,
+ 204,
+ 536
+ ],
+ "score": 0.91,
+ "content": "K B _ { V }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 524,
+ 505,
+ 538
+ ],
+ "score": 1.0,
+ "content": "for the reasoning task that the MAC cell should accomplish at this step,",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 502,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 300,
+ 549
+ ],
+ "score": 1.0,
+ "content": "which is represented by the current control state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 300,
+ 538,
+ 309,
+ 547
+ ],
+ "score": 0.84,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 536,
+ 502,
+ 549
+ ],
+ "score": 1.0,
+ "content": ", as explained above. Figure 4 shows a diagram.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 552,
+ 505,
+ 619
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 552,
+ 506,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 552,
+ 506,
+ 565
+ ],
+ "score": 1.0,
+ "content": "The relevance of the new information is judged in two stages by the “relatedness” of each element",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 563,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 446,
+ 577
+ ],
+ "score": 1.0,
+ "content": "in the KB (or for the case of VQA, each region in the image) to either the memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 446,
+ 565,
+ 469,
+ 576
+ ],
+ "score": 0.9,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 563,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "that has",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 575,
+ 505,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 575,
+ 442,
+ 588
+ ],
+ "score": 1.0,
+ "content": "accumulated relevant information from previous iterations, or to the current control",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 577,
+ 451,
+ 586
+ ],
+ "score": 0.84,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 575,
+ 505,
+ 588
+ ],
+ "score": 1.0,
+ "content": ", pointing to-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 586,
+ 505,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 586,
+ 505,
+ 598
+ ],
+ "score": 1.0,
+ "content": "wards the next piece of information that should be taken into account. Here, relatedness is measured",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "by trained linear transformations comparing each element to the previous memory and the current",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 608,
+ 140,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 608,
+ 140,
+ 620
+ ],
+ "score": 1.0,
+ "content": "control.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 33.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 624,
+ 504,
+ 648
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 623,
+ 505,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 623,
+ 414,
+ 638
+ ],
+ "score": 1.0,
+ "content": "More formally, at the first stage, the interaction between each element",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 414,
+ 626,
+ 446,
+ 637
+ ],
+ "score": 0.92,
+ "content": "K B _ { h , w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 446,
+ 623,
+ 482,
+ 638
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 625,
+ 505,
+ 637
+ ],
+ "score": 0.85,
+ "content": "\\begin{array} { r l } { h } & { { } = } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 107,
+ 635,
+ 394,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 636,
+ 196,
+ 648
+ ],
+ "score": 0.91,
+ "content": "1 , . . . , H , w = 1 , . . . , W",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 197,
+ 635,
+ 304,
+ 649
+ ],
+ "score": 1.0,
+ "content": ", and the previous memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 637,
+ 327,
+ 648
+ ],
+ "score": 0.89,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 328,
+ 635,
+ 394,
+ 649
+ ],
+ "score": 1.0,
+ "content": "is computed by:",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 250,
+ 662,
+ 361,
+ 678
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 250,
+ 662,
+ 361,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 250,
+ 662,
+ 361,
+ 678
+ ],
+ "score": 0.91,
+ "content": "m _ { i - 1 } ^ { \\prime } = W ^ { d , d } \\cdot m _ { i - 1 } + b ^ { d }",
+ "type": "interline_equation",
+ "image_path": "92d034ab00623a6201883d47cd297723d599fe90d135f474e1198ca4a4622a1a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 39,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 250,
+ 662,
+ 361,
+ 678
+ ],
+ "spans": [],
+ "index": 39
+ }
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 240,
+ 696,
+ 370,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 240,
+ 696,
+ 370,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 240,
+ 696,
+ 370,
+ 732
+ ],
+ "score": 0.91,
+ "content": "\\begin{array} { r l } { \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } = \\boldsymbol { W } ^ { d , d } \\cdot \\boldsymbol { K } \\boldsymbol { B } _ { h , w } + \\boldsymbol { b } ^ { d } } & { { } } \\\\ { \\big ( \\boldsymbol { I } _ { m - K B } \\big ) _ { h , w } = \\boldsymbol { m } _ { i - 1 } ^ { \\prime } \\circ \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } } & { { } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "ba795ad9270013570f22ffe5663537aa1ee717d610072223ec7afd2128e47f00.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 40.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 240,
+ 696,
+ 370,
+ 714.0
+ ],
+ "spans": [],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 240,
+ 714.0,
+ 370,
+ 732.0
+ ],
+ "spans": [],
+ "index": 41
+ }
+ ]
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "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 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 202,
+ 81,
+ 406,
+ 154
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 202,
+ 81,
+ 406,
+ 154
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 202,
+ 81,
+ 406,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 202,
+ 81,
+ 406,
+ 154
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "432535676eeb69a82d7848d82a8a327d906d019f07db47ee24440c80b3beab34.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 202,
+ 81,
+ 406,
+ 95.6
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 202,
+ 95.6,
+ 406,
+ 110.19999999999999
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 202,
+ 110.19999999999999,
+ 406,
+ 124.79999999999998
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 202,
+ 124.79999999999998,
+ 406,
+ 139.39999999999998
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 202,
+ 139.39999999999998,
+ 406,
+ 153.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 164,
+ 504,
+ 186
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 506,
+ 176
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 506,
+ 176
+ ],
+ "score": 1.0,
+ "content": "Figure 3: The Control Unit (CU) of the MAC cell. See section 3.2.1 for details. Best viewed in",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 173,
+ 133,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 173,
+ 133,
+ 188
+ ],
+ "score": 1.0,
+ "content": "color.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 209,
+ 505,
+ 243
+ ],
+ "lines": [],
+ "index": 8,
+ "bbox_fs": [
+ 105,
+ 208,
+ 505,
+ 245
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 248,
+ 505,
+ 282
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 248,
+ 505,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 250,
+ 120,
+ 260
+ ],
+ "score": 0.84,
+ "content": "c q _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 120,
+ 248,
+ 505,
+ 261
+ ],
+ "score": 1.0,
+ "content": "represents the current reasoning operation we would like to perform in a continuous way, taking",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 321,
+ 271
+ ],
+ "score": 1.0,
+ "content": "into account both the overall meaning of the question",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 321,
+ 261,
+ 330,
+ 271
+ ],
+ "score": 0.83,
+ "content": "q _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 330,
+ 259,
+ 505,
+ 271
+ ],
+ "score": 1.0,
+ "content": ", as well as the words the model attended to",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 213,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 189,
+ 284
+ ],
+ "score": 1.0,
+ "content": "in the previous step,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 272,
+ 208,
+ 282
+ ],
+ "score": 0.86,
+ "content": "c _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 270,
+ 213,
+ 284
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 248,
+ 505,
+ 284
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 364
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 506,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 506,
+ 300
+ ],
+ "score": 1.0,
+ "content": "However, we would like to prevent the cell from diverging in the reasoning operations it tries to",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 311
+ ],
+ "score": 1.0,
+ "content": "perform, and instead anchor it back in the question words, by using them to represent the reasoning",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 308,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 469,
+ 322
+ ],
+ "score": 1.0,
+ "content": "operation of the current step. We can achieve that by computing an attention distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 311,
+ 483,
+ 320
+ ],
+ "score": 0.85,
+ "content": "c v _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 308,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "over",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 320,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 193,
+ 333
+ ],
+ "score": 1.0,
+ "content": "the contextual words",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 320,
+ 250,
+ 332
+ ],
+ "score": 0.93,
+ "content": "[ c w _ { 1 } , . . . , c w _ { S } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 320,
+ 362,
+ 333
+ ],
+ "score": 1.0,
+ "content": "based on their similarity to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 362,
+ 321,
+ 375,
+ 332
+ ],
+ "score": 0.87,
+ "content": "c q _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 320,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": ". Then, summing the contextual",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 331,
+ 505,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 331,
+ 285,
+ 343
+ ],
+ "score": 1.0,
+ "content": "words according to the attention distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 332,
+ 300,
+ 342
+ ],
+ "score": 0.87,
+ "content": "c v _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 331,
+ 465,
+ 343
+ ],
+ "score": 1.0,
+ "content": "will allow us to have a new control state,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 333,
+ 474,
+ 342
+ ],
+ "score": 0.83,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 475,
+ 331,
+ 505,
+ 343
+ ],
+ "score": 1.0,
+ "content": ", which",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 354
+ ],
+ "score": 1.0,
+ "content": "is represented again in terms of words from the question. Intuitively, it is the gist of the question that",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 104,
+ 351,
+ 436,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 351,
+ 436,
+ 367
+ ],
+ "score": 1.0,
+ "content": "is relevant to the reasoning operation we would like to perform in the current step.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 104,
+ 287,
+ 506,
+ 367
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 223,
+ 385,
+ 387,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 223,
+ 385,
+ 387,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 223,
+ 385,
+ 387,
+ 437
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { c l c r } { { c v _ { i , s } = \\displaystyle \\mathrm { s o f t m a x } \\big ( W ^ { d , 1 } \\big ( c q _ { i } \\circ c w _ { s } \\big ) + b ^ { 1 } \\big ) } } \\\\ { { \\big . } } \\\\ { { c _ { i } = \\displaystyle \\sum _ { s = 1 } ^ { S } c v _ { i , s } \\cdot c w _ { s } } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "d43f433dbdaf70d728243b740937cc4b7c74be9d6cda3d636612e146da07d190.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 223,
+ 385,
+ 387,
+ 398.0
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 223,
+ 398.0,
+ 387,
+ 411.0
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 223,
+ 411.0,
+ 387,
+ 424.0
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 223,
+ 424.0,
+ 387,
+ 437.0
+ ],
+ "spans": [],
+ "index": 23
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 448,
+ 504,
+ 470
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 448,
+ 505,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 448,
+ 329,
+ 460
+ ],
+ "score": 1.0,
+ "content": "Finally, the control unit returns the current control state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 451,
+ 338,
+ 459
+ ],
+ "score": 0.86,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 448,
+ 455,
+ 460
+ ],
+ "score": 1.0,
+ "content": ", along with an attention map",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 451,
+ 469,
+ 459
+ ],
+ "score": 0.85,
+ "content": "c v _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 448,
+ 505,
+ 460
+ ],
+ "score": 1.0,
+ "content": "over the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 459,
+ 179,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 459,
+ 179,
+ 470
+ ],
+ "score": 1.0,
+ "content": "contextual words.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 105,
+ 448,
+ 505,
+ 470
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 483,
+ 211,
+ 495
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 213,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 213,
+ 497
+ ],
+ "score": 1.0,
+ "content": "3.2.2 THE READ UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 503,
+ 505,
+ 547
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 347,
+ 516
+ ],
+ "score": 1.0,
+ "content": "The Read Unit is provided with access to the knowledge base",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 348,
+ 504,
+ 372,
+ 514
+ ],
+ "score": 0.9,
+ "content": "K B _ { V }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 372,
+ 502,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": ", along with the previous memory",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 515,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 128,
+ 527
+ ],
+ "score": 1.0,
+ "content": "state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 516,
+ 152,
+ 526
+ ],
+ "score": 0.91,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 153,
+ 515,
+ 253,
+ 527
+ ],
+ "score": 1.0,
+ "content": "and the current control",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 253,
+ 516,
+ 262,
+ 525
+ ],
+ "score": 0.83,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 515,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": ". It is responsible for retrieving relevant content from the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 178,
+ 538
+ ],
+ "score": 1.0,
+ "content": "Knowledge Base",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 525,
+ 204,
+ 536
+ ],
+ "score": 0.91,
+ "content": "K B _ { V }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 524,
+ 505,
+ 538
+ ],
+ "score": 1.0,
+ "content": "for the reasoning task that the MAC cell should accomplish at this step,",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 502,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 300,
+ 549
+ ],
+ "score": 1.0,
+ "content": "which is represented by the current control state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 300,
+ 538,
+ 309,
+ 547
+ ],
+ "score": 0.84,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 536,
+ 502,
+ 549
+ ],
+ "score": 1.0,
+ "content": ", as explained above. Figure 4 shows a diagram.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 105,
+ 502,
+ 505,
+ 549
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 552,
+ 505,
+ 619
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 552,
+ 506,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 552,
+ 506,
+ 565
+ ],
+ "score": 1.0,
+ "content": "The relevance of the new information is judged in two stages by the “relatedness” of each element",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 563,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 446,
+ 577
+ ],
+ "score": 1.0,
+ "content": "in the KB (or for the case of VQA, each region in the image) to either the memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 446,
+ 565,
+ 469,
+ 576
+ ],
+ "score": 0.9,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 563,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "that has",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 575,
+ 505,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 575,
+ 442,
+ 588
+ ],
+ "score": 1.0,
+ "content": "accumulated relevant information from previous iterations, or to the current control",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 577,
+ 451,
+ 586
+ ],
+ "score": 0.84,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 575,
+ 505,
+ 588
+ ],
+ "score": 1.0,
+ "content": ", pointing to-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 586,
+ 505,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 586,
+ 505,
+ 598
+ ],
+ "score": 1.0,
+ "content": "wards the next piece of information that should be taken into account. Here, relatedness is measured",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "by trained linear transformations comparing each element to the previous memory and the current",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 608,
+ 140,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 608,
+ 140,
+ 620
+ ],
+ "score": 1.0,
+ "content": "control.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 33.5,
+ "bbox_fs": [
+ 105,
+ 552,
+ 506,
+ 620
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 624,
+ 504,
+ 648
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 623,
+ 505,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 623,
+ 414,
+ 638
+ ],
+ "score": 1.0,
+ "content": "More formally, at the first stage, the interaction between each element",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 414,
+ 626,
+ 446,
+ 637
+ ],
+ "score": 0.92,
+ "content": "K B _ { h , w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 446,
+ 623,
+ 482,
+ 638
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 625,
+ 505,
+ 637
+ ],
+ "score": 0.85,
+ "content": "\\begin{array} { r l } { h } & { { } = } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 107,
+ 635,
+ 394,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 636,
+ 196,
+ 648
+ ],
+ "score": 0.91,
+ "content": "1 , . . . , H , w = 1 , . . . , W",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 197,
+ 635,
+ 304,
+ 649
+ ],
+ "score": 1.0,
+ "content": ", and the previous memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 637,
+ 327,
+ 648
+ ],
+ "score": 0.89,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 328,
+ 635,
+ 394,
+ 649
+ ],
+ "score": 1.0,
+ "content": "is computed by:",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 105,
+ 623,
+ 505,
+ 649
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 250,
+ 662,
+ 361,
+ 678
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 250,
+ 662,
+ 361,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 250,
+ 662,
+ 361,
+ 678
+ ],
+ "score": 0.91,
+ "content": "m _ { i - 1 } ^ { \\prime } = W ^ { d , d } \\cdot m _ { i - 1 } + b ^ { d }",
+ "type": "interline_equation",
+ "image_path": "92d034ab00623a6201883d47cd297723d599fe90d135f474e1198ca4a4622a1a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 39,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 250,
+ 662,
+ 361,
+ 678
+ ],
+ "spans": [],
+ "index": 39
+ }
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 240,
+ 696,
+ 370,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 240,
+ 696,
+ 370,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 240,
+ 696,
+ 370,
+ 732
+ ],
+ "score": 0.91,
+ "content": "\\begin{array} { r l } { \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } = \\boldsymbol { W } ^ { d , d } \\cdot \\boldsymbol { K } \\boldsymbol { B } _ { h , w } + \\boldsymbol { b } ^ { d } } & { { } } \\\\ { \\big ( \\boldsymbol { I } _ { m - K B } \\big ) _ { h , w } = \\boldsymbol { m } _ { i - 1 } ^ { \\prime } \\circ \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } } & { { } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "ba795ad9270013570f22ffe5663537aa1ee717d610072223ec7afd2128e47f00.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 40.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 240,
+ 696,
+ 370,
+ 714.0
+ ],
+ "spans": [],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 240,
+ 714.0,
+ 370,
+ 732.0
+ ],
+ "spans": [],
+ "index": 41
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 200,
+ 81,
+ 409,
+ 154
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 200,
+ 81,
+ 409,
+ 154
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 200,
+ 81,
+ 409,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 200,
+ 81,
+ 409,
+ 154
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "a538d781a85f8151d41cc83e31d872618ae93e0b206ef08dd299f4020e8135bf.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 200,
+ 81,
+ 409,
+ 95.6
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 200,
+ 95.6,
+ 409,
+ 110.19999999999999
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 200,
+ 110.19999999999999,
+ 409,
+ 124.79999999999998
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 200,
+ 124.79999999999998,
+ 409,
+ 139.39999999999998
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 200,
+ 139.39999999999998,
+ 409,
+ 153.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 165,
+ 504,
+ 188
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 505,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 505,
+ 177
+ ],
+ "score": 1.0,
+ "content": "Figure 4: The Read Unit (RU) diagram. Blue refers to control flow, purple to knowledge flow and",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 176,
+ 323,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 323,
+ 189
+ ],
+ "score": 1.0,
+ "content": "red to memory flow. See section 3.2.2 for description.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 207,
+ 504,
+ 262
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 208,
+ 505,
+ 220
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 208,
+ 505,
+ 220
+ ],
+ "score": 1.0,
+ "content": "These memory-KB interactions measure the relatedness of each element in the KB to the mem-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 231
+ ],
+ "score": 1.0,
+ "content": "ory accumulated so far, which holds information that has been deemed relevant to handle previous",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 230,
+ 505,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 230,
+ 505,
+ 241
+ ],
+ "score": 1.0,
+ "content": "reasoning steps towards addressing the question. They allow the model to perform transitive infer-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 239,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 239,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "ence, retrieving a new piece of information that now seems important in light of the recent memory",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 221,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 221,
+ 264
+ ],
+ "score": 1.0,
+ "content": "retrieved in a prior iteration.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 268,
+ 505,
+ 346
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 281
+ ],
+ "score": 1.0,
+ "content": "However, there are cases which necessitate the model to temporarily ignore current memories, when",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 280,
+ 505,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 505,
+ 291
+ ],
+ "score": 1.0,
+ "content": "choosing the new information to retrieve. Logical OR is a classical example: when the model has",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 290,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 290,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "to look at two different objects at the same time, and assuming it stored one of them at the first",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "score": 1.0,
+ "content": "iteration, it should briefly ignore it, considering new information that is relevant to the question",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "but is unrelated to the memory. In order to achieve such capability, the read unit concatenates the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 323,
+ 505,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 505,
+ 336
+ ],
+ "score": 1.0,
+ "content": "original KB elements to each corresponding memory-KB interaction, which are then projected back",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 259,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 117,
+ 347
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 335,
+ 123,
+ 344
+ ],
+ "score": 0.79,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 124,
+ 334,
+ 259,
+ 347
+ ],
+ "score": 1.0,
+ "content": "-dimensional space (equation 6a):",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 219,
+ 363,
+ 392,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 219,
+ 363,
+ 392,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 219,
+ 363,
+ 392,
+ 398
+ ],
+ "score": 0.9,
+ "content": "\\begin{array} { c } { { I _ { m - K B } , ^ { \\prime } = W ^ { 2 d , d } \\left[ I _ { m - K B } , K B _ { h , w } \\right] + b ^ { d } } } \\\\ { { I _ { c m - K B } = c _ { i } \\circ \\left( I _ { m - K B } \\right) ^ { \\prime } } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "ada1fba84ad0bac03e8e9a8f8cc72cbc8bba767e7414ee8b9b65dab9dee93066.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 219,
+ 363,
+ 392,
+ 380.5
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 219,
+ 380.5,
+ 392,
+ 398.0
+ ],
+ "spans": [],
+ "index": 20
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 405,
+ 504,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 506,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 333,
+ 418
+ ],
+ "score": 1.0,
+ "content": "At the second stage, the read unit compares the current",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 334,
+ 407,
+ 343,
+ 416
+ ],
+ "score": 0.85,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 405,
+ 506,
+ 418
+ ],
+ "score": 1.0,
+ "content": "with these memory-KB interactions, in",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 416,
+ 505,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 416,
+ 505,
+ 428
+ ],
+ "score": 1.0,
+ "content": "order to focus on the information that is relevant to the current reasoning operation that the MAC",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 425,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 425,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "cell seeks to accomplish. The result is then passed to a softmax layer yielding an attention map",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 437,
+ 506,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 439,
+ 125,
+ 449
+ ],
+ "score": 0.85,
+ "content": "m v _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 126,
+ 437,
+ 506,
+ 451
+ ],
+ "score": 1.0,
+ "content": "over the KB, which is used in turn to retrieve the relevant information to perform the current",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 448,
+ 169,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 169,
+ 462
+ ],
+ "score": 1.0,
+ "content": "reasoning step.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 224,
+ 478,
+ 387,
+ 532
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 224,
+ 478,
+ 387,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 224,
+ 478,
+ 387,
+ 532
+ ],
+ "score": 0.91,
+ "content": "\\begin{array} { r } { m v _ { i } = \\operatorname { s o f t m a x } \\left( W ^ { d , d } \\cdot I _ { c m - K B } + b ^ { d } \\right) } \\\\ { m _ { n e w } = \\displaystyle \\sum _ { h , w = 1 , 1 } ^ { H , W } \\left( m v _ { i } \\right) _ { h , w } \\cdot K B _ { h , w } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "1af356ca707634550da691f9e1c5befb2540a27633c9abfa602dcd4cdf1d0ea5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 27.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 224,
+ 478,
+ 387,
+ 491.5
+ ],
+ "spans": [],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 224,
+ 491.5,
+ 387,
+ 505.0
+ ],
+ "spans": [],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 224,
+ 505.0,
+ 387,
+ 518.5
+ ],
+ "spans": [],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 224,
+ 518.5,
+ 387,
+ 532.0
+ ],
+ "spans": [],
+ "index": 29
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 541,
+ 504,
+ 563
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 358,
+ 555
+ ],
+ "score": 1.0,
+ "content": "Finally, the read unit returns the newly retrieved information",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 543,
+ 383,
+ 552
+ ],
+ "score": 0.89,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 538,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": ", along with an attention map",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 552,
+ 233,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 125,
+ 563
+ ],
+ "score": 0.84,
+ "content": "m v _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 125,
+ 552,
+ 233,
+ 564
+ ],
+ "score": 1.0,
+ "content": "over the Knowledge Base.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 568,
+ 505,
+ 678
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 506,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 406,
+ 582
+ ],
+ "score": 1.0,
+ "content": "To give an example of the read unit operation, assume a given question",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 571,
+ 413,
+ 580
+ ],
+ "score": 0.76,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 413,
+ 568,
+ 506,
+ 582
+ ],
+ "score": 1.0,
+ "content": "such as “What object",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "is located left to the blue ball?”, whose associated answer is “cube”. Initially, no cue is provided",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 590,
+ 505,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 505,
+ 603
+ ],
+ "score": 1.0,
+ "content": "to the model to attend to that cube, since no direct information about it presents in the question.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "Instead, based on its comprehension of the question, the model may start by focusing on the blue",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 612,
+ 506,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 324,
+ 625
+ ],
+ "score": 1.0,
+ "content": "ball at the first iteration, such that the memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 614,
+ 339,
+ 623
+ ],
+ "score": 0.85,
+ "content": "m _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 612,
+ 506,
+ 625
+ ],
+ "score": 1.0,
+ "content": "will capture the blue ball. However, in",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 636
+ ],
+ "score": 1.0,
+ "content": "the second iteration, the control unit, after re-examining the question, may realize it should now",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 634,
+ 505,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 634,
+ 250,
+ 647
+ ],
+ "score": 1.0,
+ "content": "look left, storing the word “left” in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 636,
+ 260,
+ 645
+ ],
+ "score": 0.84,
+ "content": "c _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 260,
+ 634,
+ 385,
+ 647
+ ],
+ "score": 1.0,
+ "content": ". Then, when considering both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 386,
+ 635,
+ 401,
+ 645
+ ],
+ "score": 0.86,
+ "content": "m _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 401,
+ 634,
+ 419,
+ 647
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 636,
+ 429,
+ 645
+ ],
+ "score": 0.84,
+ "content": "c _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 430,
+ 634,
+ 505,
+ 647
+ ],
+ "score": 1.0,
+ "content": ", the read unit will",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 645,
+ 505,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 645,
+ 466,
+ 658
+ ],
+ "score": 1.0,
+ "content": "realize it should perform a reasoning operation corresponding to the word “left” (stored in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 646,
+ 476,
+ 656
+ ],
+ "score": 0.79,
+ "content": "c _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 645,
+ 505,
+ 658
+ ],
+ "score": 1.0,
+ "content": ") given",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 104,
+ 657,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 657,
+ 264,
+ 668
+ ],
+ "score": 1.0,
+ "content": "a memory representing the blue ball in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 658,
+ 278,
+ 667
+ ],
+ "score": 0.87,
+ "content": "m _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 279,
+ 657,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": ", thereby allowing it to look left to the blue ball and find",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 146,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 146,
+ 679
+ ],
+ "score": 1.0,
+ "content": "the cube.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 690,
+ 215,
+ 702
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 217,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 217,
+ 703
+ ],
+ "score": 1.0,
+ "content": "3.2.3 THE WRITE UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 504,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 364,
+ 722
+ ],
+ "score": 1.0,
+ "content": "The Write Unit is responsible for creating the new memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 364,
+ 711,
+ 377,
+ 721
+ ],
+ "score": 0.86,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 378,
+ 709,
+ 504,
+ 722
+ ],
+ "score": 1.0,
+ "content": "that will reflect all the informa-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 732
+ ],
+ "score": 1.0,
+ "content": "tion considered to be important to answer the question so far, i.e. up to the current iteration in the",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43.5
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "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 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 200,
+ 81,
+ 409,
+ 154
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 200,
+ 81,
+ 409,
+ 154
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 200,
+ 81,
+ 409,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 200,
+ 81,
+ 409,
+ 154
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "a538d781a85f8151d41cc83e31d872618ae93e0b206ef08dd299f4020e8135bf.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 200,
+ 81,
+ 409,
+ 95.6
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 200,
+ 95.6,
+ 409,
+ 110.19999999999999
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 200,
+ 110.19999999999999,
+ 409,
+ 124.79999999999998
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 200,
+ 124.79999999999998,
+ 409,
+ 139.39999999999998
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 200,
+ 139.39999999999998,
+ 409,
+ 153.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 165,
+ 504,
+ 188
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 505,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 505,
+ 177
+ ],
+ "score": 1.0,
+ "content": "Figure 4: The Read Unit (RU) diagram. Blue refers to control flow, purple to knowledge flow and",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 176,
+ 323,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 323,
+ 189
+ ],
+ "score": 1.0,
+ "content": "red to memory flow. See section 3.2.2 for description.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 207,
+ 504,
+ 262
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 208,
+ 505,
+ 220
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 208,
+ 505,
+ 220
+ ],
+ "score": 1.0,
+ "content": "These memory-KB interactions measure the relatedness of each element in the KB to the mem-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 231
+ ],
+ "score": 1.0,
+ "content": "ory accumulated so far, which holds information that has been deemed relevant to handle previous",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 230,
+ 505,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 230,
+ 505,
+ 241
+ ],
+ "score": 1.0,
+ "content": "reasoning steps towards addressing the question. They allow the model to perform transitive infer-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 239,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 239,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "ence, retrieving a new piece of information that now seems important in light of the recent memory",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 221,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 221,
+ 264
+ ],
+ "score": 1.0,
+ "content": "retrieved in a prior iteration.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9,
+ "bbox_fs": [
+ 104,
+ 208,
+ 506,
+ 264
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 268,
+ 505,
+ 346
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 281
+ ],
+ "score": 1.0,
+ "content": "However, there are cases which necessitate the model to temporarily ignore current memories, when",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 280,
+ 505,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 505,
+ 291
+ ],
+ "score": 1.0,
+ "content": "choosing the new information to retrieve. Logical OR is a classical example: when the model has",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 290,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 290,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "to look at two different objects at the same time, and assuming it stored one of them at the first",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 314
+ ],
+ "score": 1.0,
+ "content": "iteration, it should briefly ignore it, considering new information that is relevant to the question",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "but is unrelated to the memory. In order to achieve such capability, the read unit concatenates the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 323,
+ 505,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 505,
+ 336
+ ],
+ "score": 1.0,
+ "content": "original KB elements to each corresponding memory-KB interaction, which are then projected back",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 259,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 117,
+ 347
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 335,
+ 123,
+ 344
+ ],
+ "score": 0.79,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 124,
+ 334,
+ 259,
+ 347
+ ],
+ "score": 1.0,
+ "content": "-dimensional space (equation 6a):",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 105,
+ 268,
+ 505,
+ 347
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 219,
+ 363,
+ 392,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 219,
+ 363,
+ 392,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 219,
+ 363,
+ 392,
+ 398
+ ],
+ "score": 0.9,
+ "content": "\\begin{array} { c } { { I _ { m - K B } , ^ { \\prime } = W ^ { 2 d , d } \\left[ I _ { m - K B } , K B _ { h , w } \\right] + b ^ { d } } } \\\\ { { I _ { c m - K B } = c _ { i } \\circ \\left( I _ { m - K B } \\right) ^ { \\prime } } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "ada1fba84ad0bac03e8e9a8f8cc72cbc8bba767e7414ee8b9b65dab9dee93066.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 219,
+ 363,
+ 392,
+ 380.5
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 219,
+ 380.5,
+ 392,
+ 398.0
+ ],
+ "spans": [],
+ "index": 20
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 405,
+ 504,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 506,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 333,
+ 418
+ ],
+ "score": 1.0,
+ "content": "At the second stage, the read unit compares the current",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 334,
+ 407,
+ 343,
+ 416
+ ],
+ "score": 0.85,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 405,
+ 506,
+ 418
+ ],
+ "score": 1.0,
+ "content": "with these memory-KB interactions, in",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 416,
+ 505,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 416,
+ 505,
+ 428
+ ],
+ "score": 1.0,
+ "content": "order to focus on the information that is relevant to the current reasoning operation that the MAC",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 425,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 425,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "cell seeks to accomplish. The result is then passed to a softmax layer yielding an attention map",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 437,
+ 506,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 439,
+ 125,
+ 449
+ ],
+ "score": 0.85,
+ "content": "m v _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 126,
+ 437,
+ 506,
+ 451
+ ],
+ "score": 1.0,
+ "content": "over the KB, which is used in turn to retrieve the relevant information to perform the current",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 448,
+ 169,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 169,
+ 462
+ ],
+ "score": 1.0,
+ "content": "reasoning step.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 105,
+ 405,
+ 506,
+ 462
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 224,
+ 478,
+ 387,
+ 532
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 224,
+ 478,
+ 387,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 224,
+ 478,
+ 387,
+ 532
+ ],
+ "score": 0.91,
+ "content": "\\begin{array} { r } { m v _ { i } = \\operatorname { s o f t m a x } \\left( W ^ { d , d } \\cdot I _ { c m - K B } + b ^ { d } \\right) } \\\\ { m _ { n e w } = \\displaystyle \\sum _ { h , w = 1 , 1 } ^ { H , W } \\left( m v _ { i } \\right) _ { h , w } \\cdot K B _ { h , w } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "1af356ca707634550da691f9e1c5befb2540a27633c9abfa602dcd4cdf1d0ea5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 27.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 224,
+ 478,
+ 387,
+ 491.5
+ ],
+ "spans": [],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 224,
+ 491.5,
+ 387,
+ 505.0
+ ],
+ "spans": [],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 224,
+ 505.0,
+ 387,
+ 518.5
+ ],
+ "spans": [],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 224,
+ 518.5,
+ 387,
+ 532.0
+ ],
+ "spans": [],
+ "index": 29
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 541,
+ 504,
+ 563
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 358,
+ 555
+ ],
+ "score": 1.0,
+ "content": "Finally, the read unit returns the newly retrieved information",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 543,
+ 383,
+ 552
+ ],
+ "score": 0.89,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 538,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": ", along with an attention map",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 552,
+ 233,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 125,
+ 563
+ ],
+ "score": 0.84,
+ "content": "m v _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 125,
+ 552,
+ 233,
+ 564
+ ],
+ "score": 1.0,
+ "content": "over the Knowledge Base.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 105,
+ 538,
+ 505,
+ 564
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 568,
+ 505,
+ 678
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 506,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 406,
+ 582
+ ],
+ "score": 1.0,
+ "content": "To give an example of the read unit operation, assume a given question",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 571,
+ 413,
+ 580
+ ],
+ "score": 0.76,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 413,
+ 568,
+ 506,
+ 582
+ ],
+ "score": 1.0,
+ "content": "such as “What object",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "is located left to the blue ball?”, whose associated answer is “cube”. Initially, no cue is provided",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 590,
+ 505,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 590,
+ 505,
+ 603
+ ],
+ "score": 1.0,
+ "content": "to the model to attend to that cube, since no direct information about it presents in the question.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "Instead, based on its comprehension of the question, the model may start by focusing on the blue",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 612,
+ 506,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 324,
+ 625
+ ],
+ "score": 1.0,
+ "content": "ball at the first iteration, such that the memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 614,
+ 339,
+ 623
+ ],
+ "score": 0.85,
+ "content": "m _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 612,
+ 506,
+ 625
+ ],
+ "score": 1.0,
+ "content": "will capture the blue ball. However, in",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 636
+ ],
+ "score": 1.0,
+ "content": "the second iteration, the control unit, after re-examining the question, may realize it should now",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 634,
+ 505,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 634,
+ 250,
+ 647
+ ],
+ "score": 1.0,
+ "content": "look left, storing the word “left” in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 636,
+ 260,
+ 645
+ ],
+ "score": 0.84,
+ "content": "c _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 260,
+ 634,
+ 385,
+ 647
+ ],
+ "score": 1.0,
+ "content": ". Then, when considering both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 386,
+ 635,
+ 401,
+ 645
+ ],
+ "score": 0.86,
+ "content": "m _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 401,
+ 634,
+ 419,
+ 647
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 636,
+ 429,
+ 645
+ ],
+ "score": 0.84,
+ "content": "c _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 430,
+ 634,
+ 505,
+ 647
+ ],
+ "score": 1.0,
+ "content": ", the read unit will",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 645,
+ 505,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 645,
+ 466,
+ 658
+ ],
+ "score": 1.0,
+ "content": "realize it should perform a reasoning operation corresponding to the word “left” (stored in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 646,
+ 476,
+ 656
+ ],
+ "score": 0.79,
+ "content": "c _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 645,
+ 505,
+ 658
+ ],
+ "score": 1.0,
+ "content": ") given",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 104,
+ 657,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 657,
+ 264,
+ 668
+ ],
+ "score": 1.0,
+ "content": "a memory representing the blue ball in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 658,
+ 278,
+ 667
+ ],
+ "score": 0.87,
+ "content": "m _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 279,
+ 657,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": ", thereby allowing it to look left to the blue ball and find",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 146,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 146,
+ 679
+ ],
+ "score": 1.0,
+ "content": "the cube.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 104,
+ 568,
+ 506,
+ 679
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 690,
+ 215,
+ 702
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 217,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 217,
+ 703
+ ],
+ "score": 1.0,
+ "content": "3.2.3 THE WRITE UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 504,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 364,
+ 722
+ ],
+ "score": 1.0,
+ "content": "The Write Unit is responsible for creating the new memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 364,
+ 711,
+ 377,
+ 721
+ ],
+ "score": 0.86,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 378,
+ 709,
+ 504,
+ 722
+ ],
+ "score": 1.0,
+ "content": "that will reflect all the informa-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 732
+ ],
+ "score": 1.0,
+ "content": "tion considered to be important to answer the question so far, i.e. up to the current iteration in the",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43.5,
+ "bbox_fs": [
+ 105,
+ 709,
+ 505,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 225,
+ 81,
+ 387,
+ 157
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 225,
+ 81,
+ 387,
+ 157
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 225,
+ 81,
+ 387,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 225,
+ 81,
+ 387,
+ 157
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "da3df6ec5443731e5ef4aaf8424acd7af4458f638b9bc19f84b7d1ed88b9267e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 225,
+ 81,
+ 387,
+ 96.2
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 225,
+ 96.2,
+ 387,
+ 111.4
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 225,
+ 111.4,
+ 387,
+ 126.60000000000001
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 225,
+ 126.60000000000001,
+ 387,
+ 141.8
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 225,
+ 141.8,
+ 387,
+ 157.0
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 168,
+ 504,
+ 191
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 505,
+ 181
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 505,
+ 181
+ ],
+ "score": 1.0,
+ "content": "Figure 5: The Write Unit (WU) diagram. Blue refers to control flow and red to memory flow. See",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 223,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 223,
+ 192
+ ],
+ "score": 1.0,
+ "content": "section 3.2.3 for description.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 209,
+ 505,
+ 243
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 316,
+ 222
+ ],
+ "score": 1.0,
+ "content": "reasoning process. It receives the last memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 212,
+ 340,
+ 222
+ ],
+ "score": 0.89,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 209,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "from the previous MAC cell, along with",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 403,
+ 234
+ ],
+ "score": 1.0,
+ "content": "the newly retrieved information from the read unit in the current iteration,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 404,
+ 222,
+ 429,
+ 232
+ ],
+ "score": 0.89,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 220,
+ 506,
+ 234
+ ],
+ "score": 1.0,
+ "content": ". See figure 5 for a",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 230,
+ 145,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 230,
+ 145,
+ 245
+ ],
+ "score": 1.0,
+ "content": "diagram.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 248,
+ 504,
+ 271
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 261
+ ],
+ "score": 1.0,
+ "content": "In the main design we have explored, merging the new information with the previous memory state",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 259,
+ 275,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 275,
+ 272
+ ],
+ "score": 1.0,
+ "content": "is done simply by a linear transformation.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 240,
+ 285,
+ 371,
+ 300
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 240,
+ 285,
+ 371,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 240,
+ 285,
+ 371,
+ 300
+ ],
+ "score": 0.93,
+ "content": "m _ { i } ^ { \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i - 1 } ] + b ^ { d }",
+ "type": "interline_equation",
+ "image_path": "fae487b9caae60fb93edbf054820528c6b18061cf0a3ab7e592c6d4fbd281b4a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 240,
+ 285,
+ 371,
+ 300
+ ],
+ "spans": [],
+ "index": 12
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 306,
+ 505,
+ 372
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 306,
+ 505,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 306,
+ 505,
+ 319
+ ],
+ "score": 1.0,
+ "content": "In addition, we have explored two variations of this design. The first, self-attention, allows consid-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 317,
+ 505,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 343,
+ 330
+ ],
+ "score": 1.0,
+ "content": "ering any previous memories rather than just the last one",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 343,
+ 319,
+ 366,
+ 329
+ ],
+ "score": 0.9,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 317,
+ 505,
+ 330
+ ],
+ "score": 1.0,
+ "content": ", thus providing the network with",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 329,
+ 505,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 329,
+ 505,
+ 341
+ ],
+ "score": 1.0,
+ "content": "the capacity to model non-sequential reasoning processes. The second variation is adding gating",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 340,
+ 505,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 340,
+ 505,
+ 351
+ ],
+ "score": 1.0,
+ "content": "mechanisms to the writing unit. These may allow the model to dynamically adjust the practical",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "length of the computation to the question complexity and stabilize the memory content throughout",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 362,
+ 333,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 333,
+ 373
+ ],
+ "score": 1.0,
+ "content": "the sequential network (similarly to GRUs and LSTMs).",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 384,
+ 505,
+ 462
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "Self-Attention. The current architecture that we have presented allows the model to perform rea-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "soning steps in a sequence, passing control and memory states from one cell to the following. How-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "ever, we would like to grant the system with more flexibility. Particularly, we would like to allow it",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "to capture more complicated reasoning processes such as trees and graphs - Directed Acyclic Graph",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 428,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 428,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "(DAG) in particular, where several branches of reasoning sub-processes are merged together in later",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 438,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 438,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "stages. Indeed, the CLEVR dataset includes cases where the questions embody tree-like reasoning",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 471,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 471,
+ 462
+ ],
+ "score": 1.0,
+ "content": "process, rather than just sequences, which we would like to address correctly in our model.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 467,
+ 505,
+ 511
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 505,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 505,
+ 478
+ ],
+ "score": 1.0,
+ "content": "We achieve that by adding self-attention connections between each MAC cell and all the prior cells.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 490
+ ],
+ "score": 1.0,
+ "content": "Since each cell can look on all the prior reasoning steps and their corresponding memories retrieved",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 488,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "from the Knowledge Base, it can virtually capture any directed acyclic graph, while still having",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 223,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 223,
+ 513
+ ],
+ "score": 1.0,
+ "content": "physically sequential layout.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 515,
+ 505,
+ 572
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 506,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 283,
+ 531
+ ],
+ "score": 1.0,
+ "content": "More formally, the current MAC cell, of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 283,
+ 516,
+ 296,
+ 527
+ ],
+ "score": 0.89,
+ "content": "i ^ { t h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 296,
+ 514,
+ 434,
+ 531
+ ],
+ "score": 1.0,
+ "content": "iteration, is granted with access to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 434,
+ 519,
+ 479,
+ 528
+ ],
+ "score": 0.9,
+ "content": "c _ { 1 } , . . . , c _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 479,
+ 514,
+ 506,
+ 531
+ ],
+ "score": 1.0,
+ "content": "along",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 200,
+ 541
+ ],
+ "score": 1.0,
+ "content": "with the corresponding",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 200,
+ 529,
+ 254,
+ 540
+ ],
+ "score": 0.89,
+ "content": "m _ { 1 } , . . . , m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 254,
+ 527,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": ", that have been computed by the prior MAC cells. It begins by",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 244,
+ 551
+ ],
+ "score": 1.0,
+ "content": "computing the similarity between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 244,
+ 541,
+ 253,
+ 550
+ ],
+ "score": 0.84,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 538,
+ 271,
+ 551
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 540,
+ 316,
+ 550
+ ],
+ "score": 0.89,
+ "content": "c _ { 1 } , . . . , c _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 538,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": ", and use it to derive an attention map over the",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 546,
+ 506,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 546,
+ 174,
+ 565
+ ],
+ "score": 1.0,
+ "content": "prior MAC cells",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 175,
+ 551,
+ 195,
+ 562
+ ],
+ "score": 0.88,
+ "content": "s a _ { i , j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 196,
+ 546,
+ 212,
+ 565
+ ],
+ "score": 1.0,
+ "content": "for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 212,
+ 550,
+ 275,
+ 561
+ ],
+ "score": 0.91,
+ "content": "j = 1 , . . . , i - 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 546,
+ 425,
+ 565
+ ],
+ "score": 1.0,
+ "content": ". This represents the relevance of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 426,
+ 549,
+ 440,
+ 561
+ ],
+ "score": 0.9,
+ "content": "j ^ { t h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 440,
+ 546,
+ 506,
+ 565
+ ],
+ "score": 1.0,
+ "content": "prior reasoning",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 261,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 197,
+ 573
+ ],
+ "score": 1.0,
+ "content": "step to the current one",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 562,
+ 202,
+ 570
+ ],
+ "score": 0.73,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 560,
+ 261,
+ 573
+ ],
+ "score": 1.0,
+ "content": "(equation 9a).",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 577,
+ 503,
+ 611
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 437,
+ 590
+ ],
+ "score": 1.0,
+ "content": "Then, we average the previous memories according to this resulted attention map",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 438,
+ 579,
+ 456,
+ 590
+ ],
+ "score": 0.91,
+ "content": "s a _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 577,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": ". We obtain",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 588,
+ 506,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 125,
+ 600
+ ],
+ "score": 0.85,
+ "content": "m _ { s a }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 125,
+ 588,
+ 506,
+ 601
+ ],
+ "score": 1.0,
+ "content": ", representing the information from all the other reasoning steps that is relevant to the current",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 600,
+ 182,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 182,
+ 612
+ ],
+ "score": 1.0,
+ "content": "one (equation 9b).",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 616,
+ 505,
+ 661
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 629
+ ],
+ "score": 1.0,
+ "content": "This resembles the approach of Key-Value networks (Miller et al., 2016). The similarity between",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "control states, corresponding to the reasoning operations that are performed in each prior step, allows",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "the model to select which memories should be taken into account, when creating the new memory –",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 464,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 464,
+ 662
+ ],
+ "score": 1.0,
+ "content": "namely, which branches of the reasoning process should be merged together at this point.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 39.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 230,
+ 678,
+ 381,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 230,
+ 678,
+ 381,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 230,
+ 678,
+ 381,
+ 731
+ ],
+ "score": 0.88,
+ "content": "\\begin{array} { c } { { s a _ { i j } = \\displaystyle \\mathrm { s o f t m a x } \\left( W ^ { d , 1 } ( c _ { i } \\circ c _ { j } ) + b ^ { 1 } \\right) } } \\\\ { { ( m _ { s a } ) _ { i } = \\displaystyle \\sum _ { j = 1 } ^ { i - 1 } s a _ { i j } \\cdot m _ { j } } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "b2917add3d0ec9deb18d84cf0919d7c999da2714964aba59d6c8b9129667c984.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 42.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 230,
+ 678,
+ 381,
+ 704.5
+ ],
+ "spans": [],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 230,
+ 704.5,
+ 381,
+ 731.0
+ ],
+ "spans": [],
+ "index": 43
+ }
+ ]
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 303,
+ 751,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 12,
+ "width": 7
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 225,
+ 81,
+ 387,
+ 157
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 225,
+ 81,
+ 387,
+ 157
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 225,
+ 81,
+ 387,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 225,
+ 81,
+ 387,
+ 157
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "da3df6ec5443731e5ef4aaf8424acd7af4458f638b9bc19f84b7d1ed88b9267e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 225,
+ 81,
+ 387,
+ 96.2
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 225,
+ 96.2,
+ 387,
+ 111.4
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 225,
+ 111.4,
+ 387,
+ 126.60000000000001
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 225,
+ 126.60000000000001,
+ 387,
+ 141.8
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 225,
+ 141.8,
+ 387,
+ 157.0
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 168,
+ 504,
+ 191
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 505,
+ 181
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 505,
+ 181
+ ],
+ "score": 1.0,
+ "content": "Figure 5: The Write Unit (WU) diagram. Blue refers to control flow and red to memory flow. See",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 223,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 223,
+ 192
+ ],
+ "score": 1.0,
+ "content": "section 3.2.3 for description.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 209,
+ 505,
+ 243
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 316,
+ 222
+ ],
+ "score": 1.0,
+ "content": "reasoning process. It receives the last memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 212,
+ 340,
+ 222
+ ],
+ "score": 0.89,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 209,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "from the previous MAC cell, along with",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 506,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 403,
+ 234
+ ],
+ "score": 1.0,
+ "content": "the newly retrieved information from the read unit in the current iteration,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 404,
+ 222,
+ 429,
+ 232
+ ],
+ "score": 0.89,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 220,
+ 506,
+ 234
+ ],
+ "score": 1.0,
+ "content": ". See figure 5 for a",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 230,
+ 145,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 230,
+ 145,
+ 245
+ ],
+ "score": 1.0,
+ "content": "diagram.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8,
+ "bbox_fs": [
+ 105,
+ 209,
+ 506,
+ 245
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 248,
+ 504,
+ 271
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 261
+ ],
+ "score": 1.0,
+ "content": "In the main design we have explored, merging the new information with the previous memory state",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 259,
+ 275,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 275,
+ 272
+ ],
+ "score": 1.0,
+ "content": "is done simply by a linear transformation.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 105,
+ 247,
+ 505,
+ 272
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 240,
+ 285,
+ 371,
+ 300
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 240,
+ 285,
+ 371,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 240,
+ 285,
+ 371,
+ 300
+ ],
+ "score": 0.93,
+ "content": "m _ { i } ^ { \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i - 1 } ] + b ^ { d }",
+ "type": "interline_equation",
+ "image_path": "fae487b9caae60fb93edbf054820528c6b18061cf0a3ab7e592c6d4fbd281b4a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 240,
+ 285,
+ 371,
+ 300
+ ],
+ "spans": [],
+ "index": 12
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 306,
+ 505,
+ 372
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 306,
+ 505,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 306,
+ 505,
+ 319
+ ],
+ "score": 1.0,
+ "content": "In addition, we have explored two variations of this design. The first, self-attention, allows consid-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 317,
+ 505,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 343,
+ 330
+ ],
+ "score": 1.0,
+ "content": "ering any previous memories rather than just the last one",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 343,
+ 319,
+ 366,
+ 329
+ ],
+ "score": 0.9,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 317,
+ 505,
+ 330
+ ],
+ "score": 1.0,
+ "content": ", thus providing the network with",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 329,
+ 505,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 329,
+ 505,
+ 341
+ ],
+ "score": 1.0,
+ "content": "the capacity to model non-sequential reasoning processes. The second variation is adding gating",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 340,
+ 505,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 340,
+ 505,
+ 351
+ ],
+ "score": 1.0,
+ "content": "mechanisms to the writing unit. These may allow the model to dynamically adjust the practical",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "length of the computation to the question complexity and stabilize the memory content throughout",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 362,
+ 333,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 333,
+ 373
+ ],
+ "score": 1.0,
+ "content": "the sequential network (similarly to GRUs and LSTMs).",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 105,
+ 306,
+ 505,
+ 373
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 384,
+ 505,
+ 462
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "Self-Attention. The current architecture that we have presented allows the model to perform rea-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "soning steps in a sequence, passing control and memory states from one cell to the following. How-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "ever, we would like to grant the system with more flexibility. Particularly, we would like to allow it",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "to capture more complicated reasoning processes such as trees and graphs - Directed Acyclic Graph",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 428,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 428,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "(DAG) in particular, where several branches of reasoning sub-processes are merged together in later",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 438,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 438,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "stages. Indeed, the CLEVR dataset includes cases where the questions embody tree-like reasoning",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 471,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 471,
+ 462
+ ],
+ "score": 1.0,
+ "content": "process, rather than just sequences, which we would like to address correctly in our model.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 104,
+ 384,
+ 506,
+ 462
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 467,
+ 505,
+ 511
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 505,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 505,
+ 478
+ ],
+ "score": 1.0,
+ "content": "We achieve that by adding self-attention connections between each MAC cell and all the prior cells.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 490
+ ],
+ "score": 1.0,
+ "content": "Since each cell can look on all the prior reasoning steps and their corresponding memories retrieved",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 488,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "from the Knowledge Base, it can virtually capture any directed acyclic graph, while still having",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 223,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 223,
+ 513
+ ],
+ "score": 1.0,
+ "content": "physically sequential layout.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 466,
+ 505,
+ 513
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 515,
+ 505,
+ 572
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 506,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 283,
+ 531
+ ],
+ "score": 1.0,
+ "content": "More formally, the current MAC cell, of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 283,
+ 516,
+ 296,
+ 527
+ ],
+ "score": 0.89,
+ "content": "i ^ { t h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 296,
+ 514,
+ 434,
+ 531
+ ],
+ "score": 1.0,
+ "content": "iteration, is granted with access to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 434,
+ 519,
+ 479,
+ 528
+ ],
+ "score": 0.9,
+ "content": "c _ { 1 } , . . . , c _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 479,
+ 514,
+ 506,
+ 531
+ ],
+ "score": 1.0,
+ "content": "along",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 200,
+ 541
+ ],
+ "score": 1.0,
+ "content": "with the corresponding",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 200,
+ 529,
+ 254,
+ 540
+ ],
+ "score": 0.89,
+ "content": "m _ { 1 } , . . . , m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 254,
+ 527,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": ", that have been computed by the prior MAC cells. It begins by",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 244,
+ 551
+ ],
+ "score": 1.0,
+ "content": "computing the similarity between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 244,
+ 541,
+ 253,
+ 550
+ ],
+ "score": 0.84,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 538,
+ 271,
+ 551
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 540,
+ 316,
+ 550
+ ],
+ "score": 0.89,
+ "content": "c _ { 1 } , . . . , c _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 538,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": ", and use it to derive an attention map over the",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 546,
+ 506,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 546,
+ 174,
+ 565
+ ],
+ "score": 1.0,
+ "content": "prior MAC cells",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 175,
+ 551,
+ 195,
+ 562
+ ],
+ "score": 0.88,
+ "content": "s a _ { i , j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 196,
+ 546,
+ 212,
+ 565
+ ],
+ "score": 1.0,
+ "content": "for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 212,
+ 550,
+ 275,
+ 561
+ ],
+ "score": 0.91,
+ "content": "j = 1 , . . . , i - 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 546,
+ 425,
+ 565
+ ],
+ "score": 1.0,
+ "content": ". This represents the relevance of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 426,
+ 549,
+ 440,
+ 561
+ ],
+ "score": 0.9,
+ "content": "j ^ { t h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 440,
+ 546,
+ 506,
+ 565
+ ],
+ "score": 1.0,
+ "content": "prior reasoning",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 261,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 197,
+ 573
+ ],
+ "score": 1.0,
+ "content": "step to the current one",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 562,
+ 202,
+ 570
+ ],
+ "score": 0.73,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 560,
+ 261,
+ 573
+ ],
+ "score": 1.0,
+ "content": "(equation 9a).",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32,
+ "bbox_fs": [
+ 104,
+ 514,
+ 506,
+ 573
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 577,
+ 503,
+ 611
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 437,
+ 590
+ ],
+ "score": 1.0,
+ "content": "Then, we average the previous memories according to this resulted attention map",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 438,
+ 579,
+ 456,
+ 590
+ ],
+ "score": 0.91,
+ "content": "s a _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 577,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": ". We obtain",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 588,
+ 506,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 125,
+ 600
+ ],
+ "score": 0.85,
+ "content": "m _ { s a }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 125,
+ 588,
+ 506,
+ 601
+ ],
+ "score": 1.0,
+ "content": ", representing the information from all the other reasoning steps that is relevant to the current",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 600,
+ 182,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 182,
+ 612
+ ],
+ "score": 1.0,
+ "content": "one (equation 9b).",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 577,
+ 506,
+ 612
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 616,
+ 505,
+ 661
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 629
+ ],
+ "score": 1.0,
+ "content": "This resembles the approach of Key-Value networks (Miller et al., 2016). The similarity between",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "control states, corresponding to the reasoning operations that are performed in each prior step, allows",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "the model to select which memories should be taken into account, when creating the new memory –",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 464,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 464,
+ 662
+ ],
+ "score": 1.0,
+ "content": "namely, which branches of the reasoning process should be merged together at this point.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 39.5,
+ "bbox_fs": [
+ 105,
+ 616,
+ 506,
+ 662
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 230,
+ 678,
+ 381,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 230,
+ 678,
+ 381,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 230,
+ 678,
+ 381,
+ 731
+ ],
+ "score": 0.88,
+ "content": "\\begin{array} { c } { { s a _ { i j } = \\displaystyle \\mathrm { s o f t m a x } \\left( W ^ { d , 1 } ( c _ { i } \\circ c _ { j } ) + b ^ { 1 } \\right) } } \\\\ { { ( m _ { s a } ) _ { i } = \\displaystyle \\sum _ { j = 1 } ^ { i - 1 } s a _ { i j } \\cdot m _ { j } } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "b2917add3d0ec9deb18d84cf0919d7c999da2714964aba59d6c8b9129667c984.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 42.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 230,
+ 678,
+ 381,
+ 704.5
+ ],
+ "spans": [],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 230,
+ 704.5,
+ 381,
+ 731.0
+ ],
+ "spans": [],
+ "index": 43
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 81,
+ 481,
+ 95
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 483,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 169,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Finally, we use",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 84,
+ 187,
+ 93
+ ],
+ "score": 0.87,
+ "content": "m _ { s a }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 81,
+ 234,
+ 96
+ ],
+ "score": 1.0,
+ "content": "along with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 83,
+ 247,
+ 95
+ ],
+ "score": 0.9,
+ "content": "m _ { i } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 81,
+ 295,
+ 96
+ ],
+ "score": 1.0,
+ "content": "to compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 83,
+ 310,
+ 95
+ ],
+ "score": 0.9,
+ "content": "m _ { i } ^ { \\prime \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 310,
+ 81,
+ 483,
+ 96
+ ],
+ "score": 1.0,
+ "content": ", the new memory content in this variation.",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 244,
+ 112,
+ 367,
+ 127
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 244,
+ 112,
+ 367,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 244,
+ 112,
+ 367,
+ 127
+ ],
+ "score": 0.92,
+ "content": "m _ { i } ^ { \\prime \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i } ^ { \\prime } ] + b ^ { d }",
+ "type": "interline_equation",
+ "image_path": "5e36c1389b375bd38c94672bc787841067c61a64ebb4c06b118a35c4f749cf9f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 244,
+ 112,
+ 367,
+ 127
+ ],
+ "spans": [],
+ "index": 1
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 140,
+ 505,
+ 185
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 141,
+ 505,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 141,
+ 431,
+ 153
+ ],
+ "score": 1.0,
+ "content": "Memory Gate. The currently presented MAC network has some fixed number",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 432,
+ 143,
+ 438,
+ 152
+ ],
+ "score": 0.82,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 439,
+ 141,
+ 505,
+ 153
+ ],
+ "score": 1.0,
+ "content": "of concatenated",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 151,
+ 505,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 505,
+ 164
+ ],
+ "score": 1.0,
+ "content": "MAC cells, representing the length of the overall reasoning process we perform. However, not all",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 163,
+ 505,
+ 176
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 163,
+ 505,
+ 176
+ ],
+ "score": 1.0,
+ "content": "questions require reasoning sequence of the same length. Some questions are simpler while others",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 174,
+ 169,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 174,
+ 169,
+ 186
+ ],
+ "score": 1.0,
+ "content": "more complex.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 190,
+ 504,
+ 235
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 505,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 505,
+ 204
+ ],
+ "score": 1.0,
+ "content": "Motivated by this observation, we add a gate over the new memory computed at each step, that may",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 201,
+ 505,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 301,
+ 214
+ ],
+ "score": 1.0,
+ "content": "selectively keep content of the previous memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 302,
+ 203,
+ 325,
+ 213
+ ],
+ "score": 0.89,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 325,
+ 201,
+ 505,
+ 214
+ ],
+ "score": 1.0,
+ "content": "unchanged. Practically, the gate functions in",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 213,
+ 505,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 213,
+ 505,
+ 225
+ ],
+ "score": 1.0,
+ "content": "a similar way to a highway network (Srivastava et al., 2015), where the gate value is conditioned on",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 223,
+ 249,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 235,
+ 237
+ ],
+ "score": 1.0,
+ "content": "the current reasoning operation,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 226,
+ 244,
+ 235
+ ],
+ "score": 0.84,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 245,
+ 223,
+ 249,
+ 237
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 194,
+ 257,
+ 417,
+ 290
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 194,
+ 257,
+ 417,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 194,
+ 257,
+ 417,
+ 290
+ ],
+ "score": 0.9,
+ "content": "\\begin{array} { c } { { { c _ { i } } ^ { \\prime } = W ^ { d , d } \\cdot c _ { i } + b ^ { d } } } \\\\ { { m _ { i } = \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) \\cdot m _ { i - 1 } + ( 1 - \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) ) \\cdot m _ { i } ^ { \\prime \\prime } } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "06b7c9f2a66fd66b565bd6bf64cfa368a00407f9ea2545d0e729a32d2b2c1b1d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 194,
+ 257,
+ 417,
+ 273.5
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 194,
+ 273.5,
+ 417,
+ 290.0
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 302,
+ 504,
+ 325
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 506,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 286,
+ 316
+ ],
+ "score": 1.0,
+ "content": "The write unit returns the new memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 305,
+ 299,
+ 314
+ ],
+ "score": 0.87,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 302,
+ 425,
+ 316
+ ],
+ "score": 1.0,
+ "content": ", that will be passed along with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 305,
+ 434,
+ 314
+ ],
+ "score": 0.85,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 302,
+ 506,
+ 316
+ ],
+ "score": 1.0,
+ "content": "to the next MAC",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 312,
+ 127,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 127,
+ 326
+ ],
+ "score": 1.0,
+ "content": "cell.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 340,
+ 194,
+ 351
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 195,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 195,
+ 353
+ ],
+ "score": 1.0,
+ "content": "3.2.4 DISCUSSION",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 360,
+ 504,
+ 394
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 505,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 505,
+ 373
+ ],
+ "score": 1.0,
+ "content": "Overall, when designing the MAC cell, we have attempted to formulate the inner workings of an",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 372,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 372,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "elementary, yet generic reasoning skills: the model decomposes the problem into steps, focusing on",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 323,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 323,
+ 394
+ ],
+ "score": 1.0,
+ "content": "one at a time. At each such step, it takes into account:",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 132,
+ 404,
+ 505,
+ 495
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 404,
+ 505,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 404,
+ 195,
+ 417
+ ],
+ "score": 1.0,
+ "content": "• The control",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 406,
+ 204,
+ 416
+ ],
+ "score": 0.82,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 404,
+ 505,
+ 417
+ ],
+ "score": 1.0,
+ "content": ": Some aspect of the task - pointing to the future work that has left to be",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 415,
+ 168,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 415,
+ 168,
+ 428
+ ],
+ "score": 1.0,
+ "content": "done.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 132,
+ 433,
+ 505,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 433,
+ 505,
+ 445
+ ],
+ "score": 1.0,
+ "content": "• The previous memory or memories: The partial solution or evidence the cell has acquired",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 444,
+ 405,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 444,
+ 405,
+ 456
+ ],
+ "score": 1.0,
+ "content": "so far – pointing to the past work that has already been achieved.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 132,
+ 461,
+ 504,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 461,
+ 284,
+ 474
+ ],
+ "score": 1.0,
+ "content": "• The newly retrieved information",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 463,
+ 309,
+ 473
+ ],
+ "score": 0.89,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 310,
+ 461,
+ 485,
+ 474
+ ],
+ "score": 1.0,
+ "content": ": that is retrieved from the knowledge base",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 485,
+ 461,
+ 504,
+ 472
+ ],
+ "score": 0.62,
+ "content": "K B",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 142,
+ 473,
+ 505,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 473,
+ 505,
+ 485
+ ],
+ "score": 1.0,
+ "content": "and may or may not be transitively related to that partial solution or evidence - the present,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 485,
+ 209,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 485,
+ 209,
+ 495
+ ],
+ "score": 1.0,
+ "content": "or current work.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 505,
+ 504,
+ 528
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "Considering these three sources of information together, the cell finally adds the new information",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 517,
+ 447,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 222,
+ 530
+ ],
+ "score": 1.0,
+ "content": "up into its working memory,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 519,
+ 235,
+ 528
+ ],
+ "score": 0.84,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 517,
+ 447,
+ 530
+ ],
+ "score": 1.0,
+ "content": ", progressing one more step towards the final answer.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 544,
+ 213,
+ 555
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 215,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 215,
+ 557
+ ],
+ "score": 1.0,
+ "content": "3.3 THE OUTPUT UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 565,
+ 505,
+ 632
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 321,
+ 579
+ ],
+ "score": 1.0,
+ "content": "The output unit receives the question representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 568,
+ 327,
+ 577
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 328,
+ 566,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": ", along with the memory state passed from",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 185,
+ 590
+ ],
+ "score": 1.0,
+ "content": "the last MAC cell",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 578,
+ 199,
+ 589
+ ],
+ "score": 0.86,
+ "content": "m _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 200,
+ 577,
+ 232,
+ 590
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 233,
+ 579,
+ 240,
+ 588
+ ],
+ "score": 0.8,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 240,
+ 577,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "is the number of MAC cells in the network – representing the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 601
+ ],
+ "score": 1.0,
+ "content": "number of reasoning steps in the whole process. It inspects both and predicts an answer based on",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 599,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "their concatenation. Intuitively, we would like our model to consider both the question as well as",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 104,
+ 608,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 608,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "the relevant information that has been progressively retrieved from the KB, deemed the necessary",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 621,
+ 208,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 208,
+ 633
+ ],
+ "score": 1.0,
+ "content": "information to answer it.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 637,
+ 505,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 218,
+ 650
+ ],
+ "score": 1.0,
+ "content": "Note that considering both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 640,
+ 225,
+ 649
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 225,
+ 638,
+ 244,
+ 650
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 244,
+ 640,
+ 259,
+ 650
+ ],
+ "score": 0.87,
+ "content": "m _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 638,
+ 429,
+ 650
+ ],
+ "score": 1.0,
+ "content": "is critical to answer the question. While",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 640,
+ 444,
+ 650
+ ],
+ "score": 0.86,
+ "content": "m _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 445,
+ 638,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "represents the",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 660
+ ],
+ "score": 1.0,
+ "content": "information collected from KB, we still need to recall what has been asked about it to be able to",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "answer accordingly. This is especially true in our case, when all other interactions between the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 669,
+ 506,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 669,
+ 506,
+ 684
+ ],
+ "score": 1.0,
+ "content": "question and the KB are mediated through attention distributions, rather than being transformed into",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 682,
+ 242,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 682,
+ 242,
+ 694
+ ],
+ "score": 1.0,
+ "content": "a shared continuous vector space.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 698,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "The prediction is built out of a standard 2-layers fully-connected softmax-based classifier with hid-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 168,
+ 722
+ ],
+ "score": 1.0,
+ "content": "den dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 710,
+ 174,
+ 720
+ ],
+ "score": 0.79,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 175,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "and output dimension that matches the number of possible answers in the dataset.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 719,
+ 482,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 719,
+ 197,
+ 734
+ ],
+ "score": 1.0,
+ "content": "The classifier receives",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 721,
+ 226,
+ 733
+ ],
+ "score": 0.93,
+ "content": "[ m _ { p } , q ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 719,
+ 482,
+ 734
+ ],
+ "score": 1.0,
+ "content": "as input and returns a probability distribution over the answers.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 26,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 750,
+ 309,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 750,
+ 309,
+ 761
+ ],
+ "score": 1.0,
+ "content": "8",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 81,
+ 481,
+ 95
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 483,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 169,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Finally, we use",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 84,
+ 187,
+ 93
+ ],
+ "score": 0.87,
+ "content": "m _ { s a }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 81,
+ 234,
+ 96
+ ],
+ "score": 1.0,
+ "content": "along with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 83,
+ 247,
+ 95
+ ],
+ "score": 0.9,
+ "content": "m _ { i } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 81,
+ 295,
+ 96
+ ],
+ "score": 1.0,
+ "content": "to compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 83,
+ 310,
+ 95
+ ],
+ "score": 0.9,
+ "content": "m _ { i } ^ { \\prime \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 310,
+ 81,
+ 483,
+ 96
+ ],
+ "score": 1.0,
+ "content": ", the new memory content in this variation.",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0,
+ "bbox_fs": [
+ 104,
+ 81,
+ 483,
+ 96
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 244,
+ 112,
+ 367,
+ 127
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 244,
+ 112,
+ 367,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 244,
+ 112,
+ 367,
+ 127
+ ],
+ "score": 0.92,
+ "content": "m _ { i } ^ { \\prime \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i } ^ { \\prime } ] + b ^ { d }",
+ "type": "interline_equation",
+ "image_path": "5e36c1389b375bd38c94672bc787841067c61a64ebb4c06b118a35c4f749cf9f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 244,
+ 112,
+ 367,
+ 127
+ ],
+ "spans": [],
+ "index": 1
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 140,
+ 505,
+ 185
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 141,
+ 505,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 141,
+ 431,
+ 153
+ ],
+ "score": 1.0,
+ "content": "Memory Gate. The currently presented MAC network has some fixed number",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 432,
+ 143,
+ 438,
+ 152
+ ],
+ "score": 0.82,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 439,
+ 141,
+ 505,
+ 153
+ ],
+ "score": 1.0,
+ "content": "of concatenated",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 151,
+ 505,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 505,
+ 164
+ ],
+ "score": 1.0,
+ "content": "MAC cells, representing the length of the overall reasoning process we perform. However, not all",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 163,
+ 505,
+ 176
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 163,
+ 505,
+ 176
+ ],
+ "score": 1.0,
+ "content": "questions require reasoning sequence of the same length. Some questions are simpler while others",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 174,
+ 169,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 174,
+ 169,
+ 186
+ ],
+ "score": 1.0,
+ "content": "more complex.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 105,
+ 141,
+ 505,
+ 186
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 190,
+ 504,
+ 235
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 505,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 505,
+ 204
+ ],
+ "score": 1.0,
+ "content": "Motivated by this observation, we add a gate over the new memory computed at each step, that may",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 201,
+ 505,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 301,
+ 214
+ ],
+ "score": 1.0,
+ "content": "selectively keep content of the previous memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 302,
+ 203,
+ 325,
+ 213
+ ],
+ "score": 0.89,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 325,
+ 201,
+ 505,
+ 214
+ ],
+ "score": 1.0,
+ "content": "unchanged. Practically, the gate functions in",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 213,
+ 505,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 213,
+ 505,
+ 225
+ ],
+ "score": 1.0,
+ "content": "a similar way to a highway network (Srivastava et al., 2015), where the gate value is conditioned on",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 223,
+ 249,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 235,
+ 237
+ ],
+ "score": 1.0,
+ "content": "the current reasoning operation,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 226,
+ 244,
+ 235
+ ],
+ "score": 0.84,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 245,
+ 223,
+ 249,
+ 237
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 105,
+ 190,
+ 505,
+ 237
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 194,
+ 257,
+ 417,
+ 290
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 194,
+ 257,
+ 417,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 194,
+ 257,
+ 417,
+ 290
+ ],
+ "score": 0.9,
+ "content": "\\begin{array} { c } { { { c _ { i } } ^ { \\prime } = W ^ { d , d } \\cdot c _ { i } + b ^ { d } } } \\\\ { { m _ { i } = \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) \\cdot m _ { i - 1 } + ( 1 - \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) ) \\cdot m _ { i } ^ { \\prime \\prime } } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "06b7c9f2a66fd66b565bd6bf64cfa368a00407f9ea2545d0e729a32d2b2c1b1d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 194,
+ 257,
+ 417,
+ 273.5
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 194,
+ 273.5,
+ 417,
+ 290.0
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 302,
+ 504,
+ 325
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 506,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 286,
+ 316
+ ],
+ "score": 1.0,
+ "content": "The write unit returns the new memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 305,
+ 299,
+ 314
+ ],
+ "score": 0.87,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 302,
+ 425,
+ 316
+ ],
+ "score": 1.0,
+ "content": ", that will be passed along with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 305,
+ 434,
+ 314
+ ],
+ "score": 0.85,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 302,
+ 506,
+ 316
+ ],
+ "score": 1.0,
+ "content": "to the next MAC",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 312,
+ 127,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 127,
+ 326
+ ],
+ "score": 1.0,
+ "content": "cell.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 105,
+ 302,
+ 506,
+ 326
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 340,
+ 194,
+ 351
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 195,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 195,
+ 353
+ ],
+ "score": 1.0,
+ "content": "3.2.4 DISCUSSION",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 360,
+ 504,
+ 394
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 505,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 505,
+ 373
+ ],
+ "score": 1.0,
+ "content": "Overall, when designing the MAC cell, we have attempted to formulate the inner workings of an",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 372,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 372,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "elementary, yet generic reasoning skills: the model decomposes the problem into steps, focusing on",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 323,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 323,
+ 394
+ ],
+ "score": 1.0,
+ "content": "one at a time. At each such step, it takes into account:",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 106,
+ 360,
+ 505,
+ 394
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 132,
+ 404,
+ 505,
+ 495
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 404,
+ 505,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 404,
+ 195,
+ 417
+ ],
+ "score": 1.0,
+ "content": "• The control",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 406,
+ 204,
+ 416
+ ],
+ "score": 0.82,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 404,
+ 505,
+ 417
+ ],
+ "score": 1.0,
+ "content": ": Some aspect of the task - pointing to the future work that has left to be",
+ "type": "text"
+ }
+ ],
+ "index": 18,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 415,
+ 168,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 415,
+ 168,
+ 428
+ ],
+ "score": 1.0,
+ "content": "done.",
+ "type": "text"
+ }
+ ],
+ "index": 19,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 132,
+ 433,
+ 505,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 433,
+ 505,
+ 445
+ ],
+ "score": 1.0,
+ "content": "• The previous memory or memories: The partial solution or evidence the cell has acquired",
+ "type": "text"
+ }
+ ],
+ "index": 20,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 444,
+ 405,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 444,
+ 405,
+ 456
+ ],
+ "score": 1.0,
+ "content": "so far – pointing to the past work that has already been achieved.",
+ "type": "text"
+ }
+ ],
+ "index": 21,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 132,
+ 461,
+ 504,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 461,
+ 284,
+ 474
+ ],
+ "score": 1.0,
+ "content": "• The newly retrieved information",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 463,
+ 309,
+ 473
+ ],
+ "score": 0.89,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 310,
+ 461,
+ 485,
+ 474
+ ],
+ "score": 1.0,
+ "content": ": that is retrieved from the knowledge base",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 485,
+ 461,
+ 504,
+ 472
+ ],
+ "score": 0.62,
+ "content": "K B",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 22,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 142,
+ 473,
+ 505,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 473,
+ 505,
+ 485
+ ],
+ "score": 1.0,
+ "content": "and may or may not be transitively related to that partial solution or evidence - the present,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 485,
+ 209,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 485,
+ 209,
+ 495
+ ],
+ "score": 1.0,
+ "content": "or current work.",
+ "type": "text"
+ }
+ ],
+ "index": 24,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 131,
+ 404,
+ 505,
+ 495
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 505,
+ 504,
+ 528
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "Considering these three sources of information together, the cell finally adds the new information",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 517,
+ 447,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 222,
+ 530
+ ],
+ "score": 1.0,
+ "content": "up into its working memory,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 519,
+ 235,
+ 528
+ ],
+ "score": 0.84,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 517,
+ 447,
+ 530
+ ],
+ "score": 1.0,
+ "content": ", progressing one more step towards the final answer.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 105,
+ 505,
+ 505,
+ 530
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 544,
+ 213,
+ 555
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 215,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 215,
+ 557
+ ],
+ "score": 1.0,
+ "content": "3.3 THE OUTPUT UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 565,
+ 505,
+ 632
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 321,
+ 579
+ ],
+ "score": 1.0,
+ "content": "The output unit receives the question representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 568,
+ 327,
+ 577
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 328,
+ 566,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": ", along with the memory state passed from",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 185,
+ 590
+ ],
+ "score": 1.0,
+ "content": "the last MAC cell",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 578,
+ 199,
+ 589
+ ],
+ "score": 0.86,
+ "content": "m _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 200,
+ 577,
+ 232,
+ 590
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 233,
+ 579,
+ 240,
+ 588
+ ],
+ "score": 0.8,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 240,
+ 577,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "is the number of MAC cells in the network – representing the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 601
+ ],
+ "score": 1.0,
+ "content": "number of reasoning steps in the whole process. It inspects both and predicts an answer based on",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 599,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "their concatenation. Intuitively, we would like our model to consider both the question as well as",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 104,
+ 608,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 608,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "the relevant information that has been progressively retrieved from the KB, deemed the necessary",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 621,
+ 208,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 208,
+ 633
+ ],
+ "score": 1.0,
+ "content": "information to answer it.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 104,
+ 566,
+ 506,
+ 633
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 637,
+ 505,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 218,
+ 650
+ ],
+ "score": 1.0,
+ "content": "Note that considering both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 640,
+ 225,
+ 649
+ ],
+ "score": 0.78,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 225,
+ 638,
+ 244,
+ 650
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 244,
+ 640,
+ 259,
+ 650
+ ],
+ "score": 0.87,
+ "content": "m _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 638,
+ 429,
+ 650
+ ],
+ "score": 1.0,
+ "content": "is critical to answer the question. While",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 640,
+ 444,
+ 650
+ ],
+ "score": 0.86,
+ "content": "m _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 445,
+ 638,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "represents the",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 660
+ ],
+ "score": 1.0,
+ "content": "information collected from KB, we still need to recall what has been asked about it to be able to",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "answer accordingly. This is especially true in our case, when all other interactions between the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 669,
+ 506,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 669,
+ 506,
+ 684
+ ],
+ "score": 1.0,
+ "content": "question and the KB are mediated through attention distributions, rather than being transformed into",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 682,
+ 242,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 682,
+ 242,
+ 694
+ ],
+ "score": 1.0,
+ "content": "a shared continuous vector space.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 638,
+ 506,
+ 694
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 698,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "The prediction is built out of a standard 2-layers fully-connected softmax-based classifier with hid-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 168,
+ 722
+ ],
+ "score": 1.0,
+ "content": "den dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 710,
+ 174,
+ 720
+ ],
+ "score": 0.79,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 175,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "and output dimension that matches the number of possible answers in the dataset.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 719,
+ 482,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 719,
+ 197,
+ 734
+ ],
+ "score": 1.0,
+ "content": "The classifier receives",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 721,
+ 226,
+ 733
+ ],
+ "score": 0.93,
+ "content": "[ m _ { p } , q ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 719,
+ 482,
+ 734
+ ],
+ "score": 1.0,
+ "content": "as input and returns a probability distribution over the answers.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 105,
+ 698,
+ 505,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 135,
+ 144,
+ 476,
+ 291
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 80,
+ 505,
+ 135
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 505,
+ 92
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 505,
+ 92
+ ],
+ "score": 1.0,
+ "content": "Table 1: CLEVR accuracy by baseline methods, competing methods, and our method (MAC) - for",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 91,
+ 504,
+ 103
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 91,
+ 504,
+ 103
+ ],
+ "score": 1.0,
+ "content": "our method we show test results with GLoVE initialization and validation results with and with-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 101,
+ 506,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 101,
+ 458,
+ 116
+ ],
+ "score": 1.0,
+ "content": "out GLoVE. (*) denotes use of extra supervisory information through program labels.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 459,
+ 101,
+ 469,
+ 114
+ ],
+ "score": 0.28,
+ "content": "( ^ { \\dagger } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 101,
+ 506,
+ 116
+ ],
+ "score": 1.0,
+ "content": "denotes",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 113,
+ 505,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 113,
+ 217,
+ 126
+ ],
+ "score": 1.0,
+ "content": "use of data augmentation.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 217,
+ 113,
+ 228,
+ 124
+ ],
+ "score": 0.52,
+ "content": "( ^ { \\ddagger } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 228,
+ 113,
+ 371,
+ 126
+ ],
+ "score": 1.0,
+ "content": "denotes training from raw pixels.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 113,
+ 383,
+ 124
+ ],
+ "score": 0.42,
+ "content": "( ^ { \\circ } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 113,
+ 505,
+ 126
+ ],
+ "score": 1.0,
+ "content": "denotes GLoVE word vector",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 123,
+ 162,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 123,
+ 162,
+ 136
+ ],
+ "score": 1.0,
+ "content": "initialization.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 135,
+ 144,
+ 476,
+ 291
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 144,
+ 476,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 144,
+ 476,
+ 291
+ ],
+ "score": 0.981,
+ "html": "
| Model | Overall | Count | Exist | Compare Numbers | Query Attribute | Compare Attribute |
| Human (Johnson et al., 2017) | 92.6 | 86.7 | 96.6 | 86.5 | 95.0 | 96.0 |
| Q-type baseline (Johnson et al., 2017) | 41.8 | 34.6 | 50.2 | 51.0 | 36.0 | 51.3 |
| LSTM (Johnson et al., 2017) | 46.8 | 41.7 | 61.1 | 69.8 | 36.8 | 51.8 |
| CNN+LSTM (Johnson et al., 2017) | 52.3 | 43.7 | 65.2 | 67.1 | 49.3 | 53.0 |
| CNN+LSTM+SA (Johnson et al., 2016) | 76.6 | 64.4 | 82.7 | 77.4 | 82.6 | 75.4 |
| N2NMN* (Hu et al. 2017) | 83.7 | 68.5 | 85.7 | 84.9 | 90.0 | 88.7 |
| PG+EE (9K prog.)* (Johnson et al., 2017) | 88.6 | 79.7 | 89.7 | 79.1 | 92.6 | 96.0 |
| PG+EE (70oK prog.)* (Johnson et al., 2017) | 96.9 | 92.7 | 97.1 | 98.7 | 98.1 | 98.9 |
| CNN+LSTM+RN†‡ (Santoro et al.,2017) | 95.5 | 90.1 | 97.8 | 93.6 | 97.9 | 97.1 |
| CNN+GRU+FiLM (Perez et al.,2017) | 97.7 | 94.3 | 99.1 | 96.8 | 99.1 | 99.1 |
| CNN+GRU+FiLM* (Perez et al., 2017) | 97.6 | 94.3 | 99.3 | 93.4 | 99.3 | 99.3 |
| CNN+GRU+FiLM (Perez et al., 2017) | 97.7 | 94.3 | 99.1 | 96.8 | 99.1 | 99.1 |
| MAC (this paper)-val | 98.9 | 97.1 | 99.5 | 99.3 | 99.2 | 99.2 |
| MAC-val | 99.0 | 97.2 | 99.5 | 99.5 | 99.5 | 99.4 |
| MAC-test | 98.9 | 97.2 | 99.5 | 99.4 | 99.3 | 99.5 |
",
+ "type": "table",
+ "image_path": "a8f3f0904dbfbf06b3ae89f9cd0dcd15557eabccb6cf593d1481a4565628f79b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 135,
+ 144,
+ 476,
+ 193.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 135,
+ 193.0,
+ 476,
+ 242.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 135,
+ 242.0,
+ 476,
+ 291.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ }
+ ],
+ "index": 4.0
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 329,
+ 200,
+ 342
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 329,
+ 201,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 329,
+ 201,
+ 344
+ ],
+ "score": 1.0,
+ "content": "4 EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 367,
+ 505,
+ 445
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 505,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 505,
+ 381
+ ],
+ "score": 1.0,
+ "content": "We evaluate our model on the recent CLEVR dataset (Johnson et al., 2016). CLEVR is a syn-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 506,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 506,
+ 392
+ ],
+ "score": 1.0,
+ "content": "thetic dataset consisting of 700K tuples; each consists of a 3D-rendered image featuring objects of",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 390,
+ 505,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 390,
+ 505,
+ 403
+ ],
+ "score": 1.0,
+ "content": "various shapes, colors, materials and sizes, coupled with compositional multi-step questions that",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 414
+ ],
+ "score": 1.0,
+ "content": "measure performance on an array of challenging reasoning skills such as following transitive re-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 424
+ ],
+ "score": 1.0,
+ "content": "lations, counting objects and comparing their properties. In addition, each question is associated",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": "with a formal program, specifying the reasoning operations that should be performed to compute",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 434,
+ 248,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 248,
+ 446
+ ],
+ "score": 1.0,
+ "content": "the answer, among 28 possibilities.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 450,
+ 505,
+ 506
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 505,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 505,
+ 464
+ ],
+ "score": 1.0,
+ "content": "We first perform experiments on the original 700k CLEVR dataset (Johnson et al., 2016), comparing",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 474
+ ],
+ "score": 1.0,
+ "content": "to prior work. As shown in table 1, our model matches or outperforms all existing models both in",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 485
+ ],
+ "score": 1.0,
+ "content": "overall accuracy, as well as in each category, testing different reasoning skills. In particular, for the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 237,
+ 497
+ ],
+ "score": 1.0,
+ "content": "overall performance, we achieve",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 238,
+ 484,
+ 270,
+ 494
+ ],
+ "score": 0.86,
+ "content": "9 8 . 9 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 271,
+ 483,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "accuracy, more than halving the error rate of the prior best",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 242,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 242,
+ 507
+ ],
+ "score": 1.0,
+ "content": "model, FiLM (Perez et al., 2017).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 511,
+ 504,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "Counting and Numerical Comparison. Remarkably, our performance on questions testing count-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 522,
+ 504,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 504,
+ 536
+ ],
+ "score": 1.0,
+ "content": "ing and numerical comparisons is significantly higher than the competing models, which consis-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "tently struggle on this question type. Again, we nearly halve the corresponding error rate. These",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 543,
+ 505,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 505,
+ 559
+ ],
+ "score": 1.0,
+ "content": "results demonstrate the aptitude of attention mechanisms to perform counting, reduction and aggre-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 336,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 336,
+ 567
+ ],
+ "score": 1.0,
+ "content": "gation, in contrast to alternative, CNN-based approaches.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 572,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 504,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 504,
+ 584
+ ],
+ "score": 1.0,
+ "content": "Training Length and Computational-Efficiency. We examine the learning curves of our and com-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "score": 1.0,
+ "content": "peting models. We have trained all models on the same architecture and used the author code for",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "the other models. Aiming at having equal settings for comparison, we ran all models including ours",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "with learned random words vectors. In order to make sure the results are statistically significant we",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "ran each model multiple (10) times, and plotted the averages and confidence intervals (figure 4). The",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "results show that our model learns significantly faster than the other leading methods, FiLM (Perez",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 174,
+ 650
+ ],
+ "score": 1.0,
+ "content": "et al., 2017) and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 175,
+ 638,
+ 208,
+ 649
+ ],
+ "score": 0.75,
+ "content": "\\mathrm { P G + E E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 637,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "(Johnson et al., 2017). While we do not have learning curves for the Re-",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 506,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 506,
+ 662
+ ],
+ "score": 1.0,
+ "content": "lational Network model, Santoro et al. (2017) report approximately 1.4 million iterations to achieve",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 659,
+ 506,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 660,
+ 134,
+ 671
+ ],
+ "score": 0.86,
+ "content": "9 5 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 134,
+ 659,
+ 506,
+ 673
+ ],
+ "score": 1.0,
+ "content": "accuracy, which are equivalent to 125 epochs approximately, whereas our model achieves a",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 506,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 506,
+ 684
+ ],
+ "score": 1.0,
+ "content": "comparable accuracy after 3 epochs only, yielding 40x reduction in the length of the training process.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 701
+ ],
+ "score": 1.0,
+ "content": "Naturally, the smaller number of required training steps also translates to comparably shorter training",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "time. Perez et al. (2017) report training time of 4 days, equivalent to 80 epochs, to reach accuracy",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 117,
+ 722
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 118,
+ 710,
+ 144,
+ 720
+ ],
+ "score": 0.87,
+ "content": "9 7 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 145,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": ". In contrast, we achieve higher accuracy in 6 epochs, taking 9.5 hours overall, leading to",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 107,
+ 721,
+ 230,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 721,
+ 123,
+ 731
+ ],
+ "score": 0.49,
+ "content": "1 0 \\mathrm { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 123,
+ 721,
+ 230,
+ 732
+ ],
+ "score": 1.0,
+ "content": "reduction in training time.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 37.5
+ }
+ ],
+ "page_idx": 8,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "9",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 135,
+ 144,
+ 476,
+ 291
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 80,
+ 505,
+ 135
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 505,
+ 92
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 505,
+ 92
+ ],
+ "score": 1.0,
+ "content": "Table 1: CLEVR accuracy by baseline methods, competing methods, and our method (MAC) - for",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 91,
+ 504,
+ 103
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 91,
+ 504,
+ 103
+ ],
+ "score": 1.0,
+ "content": "our method we show test results with GLoVE initialization and validation results with and with-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 101,
+ 506,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 101,
+ 458,
+ 116
+ ],
+ "score": 1.0,
+ "content": "out GLoVE. (*) denotes use of extra supervisory information through program labels.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 459,
+ 101,
+ 469,
+ 114
+ ],
+ "score": 0.28,
+ "content": "( ^ { \\dagger } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 101,
+ 506,
+ 116
+ ],
+ "score": 1.0,
+ "content": "denotes",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 113,
+ 505,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 113,
+ 217,
+ 126
+ ],
+ "score": 1.0,
+ "content": "use of data augmentation.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 217,
+ 113,
+ 228,
+ 124
+ ],
+ "score": 0.52,
+ "content": "( ^ { \\ddagger } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 228,
+ 113,
+ 371,
+ 126
+ ],
+ "score": 1.0,
+ "content": "denotes training from raw pixels.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 113,
+ 383,
+ 124
+ ],
+ "score": 0.42,
+ "content": "( ^ { \\circ } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 113,
+ 505,
+ 126
+ ],
+ "score": 1.0,
+ "content": "denotes GLoVE word vector",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 123,
+ 162,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 123,
+ 162,
+ 136
+ ],
+ "score": 1.0,
+ "content": "initialization.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 135,
+ 144,
+ 476,
+ 291
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 144,
+ 476,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 144,
+ 476,
+ 291
+ ],
+ "score": 0.981,
+ "html": "| Model | Overall | Count | Exist | Compare Numbers | Query Attribute | Compare Attribute |
| Human (Johnson et al., 2017) | 92.6 | 86.7 | 96.6 | 86.5 | 95.0 | 96.0 |
| Q-type baseline (Johnson et al., 2017) | 41.8 | 34.6 | 50.2 | 51.0 | 36.0 | 51.3 |
| LSTM (Johnson et al., 2017) | 46.8 | 41.7 | 61.1 | 69.8 | 36.8 | 51.8 |
| CNN+LSTM (Johnson et al., 2017) | 52.3 | 43.7 | 65.2 | 67.1 | 49.3 | 53.0 |
| CNN+LSTM+SA (Johnson et al., 2016) | 76.6 | 64.4 | 82.7 | 77.4 | 82.6 | 75.4 |
| N2NMN* (Hu et al. 2017) | 83.7 | 68.5 | 85.7 | 84.9 | 90.0 | 88.7 |
| PG+EE (9K prog.)* (Johnson et al., 2017) | 88.6 | 79.7 | 89.7 | 79.1 | 92.6 | 96.0 |
| PG+EE (70oK prog.)* (Johnson et al., 2017) | 96.9 | 92.7 | 97.1 | 98.7 | 98.1 | 98.9 |
| CNN+LSTM+RN†‡ (Santoro et al.,2017) | 95.5 | 90.1 | 97.8 | 93.6 | 97.9 | 97.1 |
| CNN+GRU+FiLM (Perez et al.,2017) | 97.7 | 94.3 | 99.1 | 96.8 | 99.1 | 99.1 |
| CNN+GRU+FiLM* (Perez et al., 2017) | 97.6 | 94.3 | 99.3 | 93.4 | 99.3 | 99.3 |
| CNN+GRU+FiLM (Perez et al., 2017) | 97.7 | 94.3 | 99.1 | 96.8 | 99.1 | 99.1 |
| MAC (this paper)-val | 98.9 | 97.1 | 99.5 | 99.3 | 99.2 | 99.2 |
| MAC-val | 99.0 | 97.2 | 99.5 | 99.5 | 99.5 | 99.4 |
| MAC-test | 98.9 | 97.2 | 99.5 | 99.4 | 99.3 | 99.5 |
",
+ "type": "table",
+ "image_path": "a8f3f0904dbfbf06b3ae89f9cd0dcd15557eabccb6cf593d1481a4565628f79b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 135,
+ 144,
+ 476,
+ 193.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 135,
+ 193.0,
+ 476,
+ 242.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 135,
+ 242.0,
+ 476,
+ 291.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ }
+ ],
+ "index": 4.0
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 329,
+ 200,
+ 342
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 329,
+ 201,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 329,
+ 201,
+ 344
+ ],
+ "score": 1.0,
+ "content": "4 EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 367,
+ 505,
+ 445
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 505,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 505,
+ 381
+ ],
+ "score": 1.0,
+ "content": "We evaluate our model on the recent CLEVR dataset (Johnson et al., 2016). CLEVR is a syn-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 506,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 506,
+ 392
+ ],
+ "score": 1.0,
+ "content": "thetic dataset consisting of 700K tuples; each consists of a 3D-rendered image featuring objects of",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 390,
+ 505,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 390,
+ 505,
+ 403
+ ],
+ "score": 1.0,
+ "content": "various shapes, colors, materials and sizes, coupled with compositional multi-step questions that",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 414
+ ],
+ "score": 1.0,
+ "content": "measure performance on an array of challenging reasoning skills such as following transitive re-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 424
+ ],
+ "score": 1.0,
+ "content": "lations, counting objects and comparing their properties. In addition, each question is associated",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": "with a formal program, specifying the reasoning operations that should be performed to compute",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 434,
+ 248,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 248,
+ 446
+ ],
+ "score": 1.0,
+ "content": "the answer, among 28 possibilities.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 105,
+ 367,
+ 506,
+ 446
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 450,
+ 505,
+ 506
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 505,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 505,
+ 464
+ ],
+ "score": 1.0,
+ "content": "We first perform experiments on the original 700k CLEVR dataset (Johnson et al., 2016), comparing",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 474
+ ],
+ "score": 1.0,
+ "content": "to prior work. As shown in table 1, our model matches or outperforms all existing models both in",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 485
+ ],
+ "score": 1.0,
+ "content": "overall accuracy, as well as in each category, testing different reasoning skills. In particular, for the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 237,
+ 497
+ ],
+ "score": 1.0,
+ "content": "overall performance, we achieve",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 238,
+ 484,
+ 270,
+ 494
+ ],
+ "score": 0.86,
+ "content": "9 8 . 9 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 271,
+ 483,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "accuracy, more than halving the error rate of the prior best",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 242,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 242,
+ 507
+ ],
+ "score": 1.0,
+ "content": "model, FiLM (Perez et al., 2017).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18,
+ "bbox_fs": [
+ 105,
+ 450,
+ 505,
+ 507
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 511,
+ 504,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "Counting and Numerical Comparison. Remarkably, our performance on questions testing count-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 522,
+ 504,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 522,
+ 504,
+ 536
+ ],
+ "score": 1.0,
+ "content": "ing and numerical comparisons is significantly higher than the competing models, which consis-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "tently struggle on this question type. Again, we nearly halve the corresponding error rate. These",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 543,
+ 505,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 505,
+ 559
+ ],
+ "score": 1.0,
+ "content": "results demonstrate the aptitude of attention mechanisms to perform counting, reduction and aggre-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 336,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 336,
+ 567
+ ],
+ "score": 1.0,
+ "content": "gation, in contrast to alternative, CNN-based approaches.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 105,
+ 511,
+ 505,
+ 567
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 572,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 504,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 504,
+ 584
+ ],
+ "score": 1.0,
+ "content": "Training Length and Computational-Efficiency. We examine the learning curves of our and com-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "score": 1.0,
+ "content": "peting models. We have trained all models on the same architecture and used the author code for",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "the other models. Aiming at having equal settings for comparison, we ran all models including ours",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "with learned random words vectors. In order to make sure the results are statistically significant we",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "ran each model multiple (10) times, and plotted the averages and confidence intervals (figure 4). The",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "results show that our model learns significantly faster than the other leading methods, FiLM (Perez",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 174,
+ 650
+ ],
+ "score": 1.0,
+ "content": "et al., 2017) and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 175,
+ 638,
+ 208,
+ 649
+ ],
+ "score": 0.75,
+ "content": "\\mathrm { P G + E E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 637,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "(Johnson et al., 2017). While we do not have learning curves for the Re-",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 506,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 506,
+ 662
+ ],
+ "score": 1.0,
+ "content": "lational Network model, Santoro et al. (2017) report approximately 1.4 million iterations to achieve",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 659,
+ 506,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 660,
+ 134,
+ 671
+ ],
+ "score": 0.86,
+ "content": "9 5 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 134,
+ 659,
+ 506,
+ 673
+ ],
+ "score": 1.0,
+ "content": "accuracy, which are equivalent to 125 epochs approximately, whereas our model achieves a",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 506,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 506,
+ 684
+ ],
+ "score": 1.0,
+ "content": "comparable accuracy after 3 epochs only, yielding 40x reduction in the length of the training process.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 105,
+ 572,
+ 506,
+ 684
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 701
+ ],
+ "score": 1.0,
+ "content": "Naturally, the smaller number of required training steps also translates to comparably shorter training",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "time. Perez et al. (2017) report training time of 4 days, equivalent to 80 epochs, to reach accuracy",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 117,
+ 722
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 118,
+ 710,
+ 144,
+ 720
+ ],
+ "score": 0.87,
+ "content": "9 7 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 145,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": ". In contrast, we achieve higher accuracy in 6 epochs, taking 9.5 hours overall, leading to",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 107,
+ 721,
+ 230,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 721,
+ 123,
+ 731
+ ],
+ "score": 0.49,
+ "content": "1 0 \\mathrm { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 123,
+ 721,
+ 230,
+ 732
+ ],
+ "score": 1.0,
+ "content": "reduction in training time.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 105,
+ 687,
+ 506,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 180,
+ 86,
+ 428,
+ 207
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 180,
+ 86,
+ 428,
+ 207
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 180,
+ 86,
+ 428,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 180,
+ 86,
+ 428,
+ 207
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "c74201d70aae1d3587e14044064872d252c0e00ec6d3c4fd6b70377bb7088c75.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 180,
+ 86,
+ 428,
+ 126.33333333333334
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 180,
+ 126.33333333333334,
+ 428,
+ 166.66666666666669
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 180,
+ 166.66666666666669,
+ 428,
+ 207.00000000000003
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 222,
+ 505,
+ 278
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 504,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 471,
+ 235
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Training curves and accuracies for MACs (our model), FiLM (Perez et al., 2017),",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 472,
+ 223,
+ 504,
+ 233
+ ],
+ "score": 0.53,
+ "content": "\\mathrm { P G + E E }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 234,
+ 504,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 471,
+ 246
+ ],
+ "score": 1.0,
+ "content": "(Johnson et al., 2017) and stacked-attention (Yang et al., 2016; Johnson et al., 2017). (Note:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 472,
+ 234,
+ 504,
+ 244
+ ],
+ "score": 0.54,
+ "content": "\\mathrm { P G + E E }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 245,
+ 505,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 245,
+ 505,
+ 258
+ ],
+ "score": 1.0,
+ "content": "uses the supported CLEVR programs as strong supervision.) Left: Training curve (accuracy/epoch).",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 506,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 256,
+ 270
+ ],
+ "score": 1.0,
+ "content": "Right: Learning curve: Accuracy for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 256,
+ 256,
+ 271,
+ 266
+ ],
+ "score": 0.82,
+ "content": "1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 271,
+ 255,
+ 273,
+ 270
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 274,
+ 256,
+ 289,
+ 266
+ ],
+ "score": 0.79,
+ "content": "2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 289,
+ 255,
+ 292,
+ 270
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 293,
+ 256,
+ 307,
+ 266
+ ],
+ "score": 0.82,
+ "content": "5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 255,
+ 311,
+ 270
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 312,
+ 256,
+ 331,
+ 266
+ ],
+ "score": 0.84,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 255,
+ 334,
+ 270
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 334,
+ 256,
+ 354,
+ 266
+ ],
+ "score": 0.84,
+ "content": "2 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 354,
+ 255,
+ 357,
+ 270
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 256,
+ 378,
+ 266
+ ],
+ "score": 0.85,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 378,
+ 255,
+ 396,
+ 270
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 256,
+ 420,
+ 266
+ ],
+ "score": 0.87,
+ "content": "100 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 255,
+ 447,
+ 270
+ ],
+ "score": 1.0,
+ "content": "of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 447,
+ 256,
+ 469,
+ 266
+ ],
+ "score": 0.68,
+ "content": "7 0 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 469,
+ 255,
+ 506,
+ 270
+ ],
+ "score": 1.0,
+ "content": "CLEVR",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 266,
+ 144,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 266,
+ 144,
+ 280
+ ],
+ "score": 1.0,
+ "content": "samples.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 181,
+ 366,
+ 429,
+ 448
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 312,
+ 505,
+ 356
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "Table 2: Accuracy on CLEVR-Humans of previous methods and our method (MAC), before (left)",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 323,
+ 506,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 369,
+ 336
+ ],
+ "score": 1.0,
+ "content": "and after (right) fine-tuning on the CLEVR-Humans training data.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 370,
+ 324,
+ 402,
+ 334
+ ],
+ "score": 0.75,
+ "content": "\\mathrm { P G + E E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 402,
+ 323,
+ 506,
+ 336
+ ],
+ "score": 1.0,
+ "content": "uses supervised data. For",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": "our method (MAC) we show test results with GLoVE initialization and validation results with and",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 176,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 176,
+ 356
+ ],
+ "score": 1.0,
+ "content": "without GLoVE.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 181,
+ 366,
+ 429,
+ 448
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 181,
+ 366,
+ 429,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 181,
+ 366,
+ 429,
+ 448
+ ],
+ "score": 0.979,
+ "html": "| Model | Train CLEVR | Train CLEVR + fine-tune HUMANS |
| LSTM (Johnson et al.,2017) | 27.5 | 36.5 |
| CNN+LSTM (Johnson et al.,2017) | 37.7 | 43.2 |
| CNN+LSTM+SA+MLP (Johnson et al.,2016) | 50.4 | 57.6 |
| PG+EE (18K prog.)* (Johnson et al.,2017) | 54.0 | 66.6 |
| CNN+GRU+FiLM (Perez et al., 2017) | 56.6 | 75.9 |
| MAC-val | 56.9 | 81.2 |
| MAC-val | 58.5 | 82.0 |
| MAC-test | 58.6 | 82.5 |
",
+ "type": "table",
+ "image_path": "20d0a9d88276e853b2d4290d832c31f6ed7a6d4ae845cb8a1f4754688e965048.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 181,
+ 366,
+ 429,
+ 393.3333333333333
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 181,
+ 393.3333333333333,
+ 429,
+ 420.66666666666663
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 181,
+ 420.66666666666663,
+ 429,
+ 447.99999999999994
+ ],
+ "spans": [],
+ "index": 14
+ }
+ ]
+ }
+ ],
+ "index": 11.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 491,
+ 209,
+ 502
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 210,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 210,
+ 504
+ ],
+ "score": 1.0,
+ "content": "4.1 DATA EFFICIENCY",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 522,
+ 505,
+ 588
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "We have explored the performance of our and other leading approaches on smaller subsets of the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "CLEVR dataset, in order to study the ability of models to generalize from smaller amount of data.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 545,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 303,
+ 556
+ ],
+ "score": 1.0,
+ "content": "We sampled at random subsets of CLEVR, with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 545,
+ 323,
+ 555
+ ],
+ "score": 0.84,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 324,
+ 545,
+ 326,
+ 556
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 545,
+ 347,
+ 555
+ ],
+ "score": 0.84,
+ "content": "2 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 347,
+ 545,
+ 365,
+ 556
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 366,
+ 545,
+ 385,
+ 555
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 386,
+ 545,
+ 444,
+ 556
+ ],
+ "score": 1.0,
+ "content": "of its original",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 545,
+ 466,
+ 555
+ ],
+ "score": 0.44,
+ "content": "7 0 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 545,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "size, and",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 568
+ ],
+ "score": 1.0,
+ "content": "used them to train our and other 3 proposed models for the CLEVR task: FiLM (Perez et al., 2017),",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 204,
+ 578
+ ],
+ "score": 1.0,
+ "content": "the strongly-supervised",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 566,
+ 237,
+ 577
+ ],
+ "score": 0.67,
+ "content": "\\mathrm { P G + E E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 237,
+ 566,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "(Johnson et al., 2017), and stacked-attention networks (Johnson",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 231,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 231,
+ 590
+ ],
+ "score": 1.0,
+ "content": "et al., 2017; Yang et al., 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 504,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 504,
+ 605
+ ],
+ "score": 1.0,
+ "content": "As shown in figure 4, our model outperforms the other models by a wide margin for all subsets of the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 604,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 604,
+ 193,
+ 617
+ ],
+ "score": 1.0,
+ "content": "CLEVR dataset. For",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 605,
+ 213,
+ 615
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 604,
+ 317,
+ 617
+ ],
+ "score": 1.0,
+ "content": "of the data, equivalent to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 317,
+ 605,
+ 339,
+ 615
+ ],
+ "score": 0.63,
+ "content": "3 5 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 604,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "samples, other models obtain accuracies",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 615,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 615,
+ 176,
+ 629
+ ],
+ "score": 1.0,
+ "content": "ranging between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 616,
+ 196,
+ 627
+ ],
+ "score": 0.88,
+ "content": "70 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 196,
+ 615,
+ 215,
+ 629
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 215,
+ 616,
+ 235,
+ 626
+ ],
+ "score": 0.88,
+ "content": "92 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 615,
+ 345,
+ 629
+ ],
+ "score": 1.0,
+ "content": ", while our model achieves",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 616,
+ 372,
+ 627
+ ],
+ "score": 0.87,
+ "content": "9 7 . 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 373,
+ 615,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": ". The gap becomes larger as the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 205,
+ 640
+ ],
+ "score": 1.0,
+ "content": "dataset size reduces: for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 627,
+ 225,
+ 637
+ ],
+ "score": 0.88,
+ "content": "2 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 225,
+ 626,
+ 327,
+ 640
+ ],
+ "score": 1.0,
+ "content": "of the data, equivalent to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 627,
+ 349,
+ 637
+ ],
+ "score": 0.5,
+ "content": "1 7 5 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 349,
+ 626,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "samples, performance of other models",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 104,
+ 636,
+ 416,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 636,
+ 151,
+ 651
+ ],
+ "score": 1.0,
+ "content": "is between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 151,
+ 638,
+ 171,
+ 649
+ ],
+ "score": 0.86,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 636,
+ 189,
+ 651
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 638,
+ 208,
+ 649
+ ],
+ "score": 0.87,
+ "content": "7 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 636,
+ 348,
+ 651
+ ],
+ "score": 1.0,
+ "content": ", while our model maintains a high",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 348,
+ 638,
+ 375,
+ 649
+ ],
+ "score": 0.83,
+ "content": "9 5 . 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 636,
+ 416,
+ 651
+ ],
+ "score": 1.0,
+ "content": "accuracy.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 654,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 154,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Finally, for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 655,
+ 173,
+ 666
+ ],
+ "score": 0.87,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 174,
+ 655,
+ 219,
+ 667
+ ],
+ "score": 1.0,
+ "content": "of the data",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 655,
+ 244,
+ 666
+ ],
+ "score": 0.29,
+ "content": "- 7 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 244,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "samples, still a sizeable amount – our model is the only one that",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 284,
+ 678
+ ],
+ "score": 1.0,
+ "content": "manages to generalize, with performance of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 284,
+ 666,
+ 311,
+ 676
+ ],
+ "score": 0.87,
+ "content": "8 4 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 311,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "on average, whereas the other three models fail,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 147,
+ 689
+ ],
+ "score": 1.0,
+ "content": "achieving",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 147,
+ 677,
+ 204,
+ 687
+ ],
+ "score": 0.89,
+ "content": "4 7 . 6 \\% - 5 7 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": ". Note that as pointed out by (Johnson et al., 2016) a simple baseline that",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 686,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 686,
+ 429,
+ 702
+ ],
+ "score": 1.0,
+ "content": "predicts the most frequent answer for each of the question types achieves already",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 688,
+ 456,
+ 699
+ ],
+ "score": 0.87,
+ "content": "4 2 . 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 686,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": ", suggesting",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "that answering half of the questions correctly means that the competing models barely learn to",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "generalize from the smaller dataset. These results demonstrate the robustness of our architecture",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 484,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 484,
+ 732
+ ],
+ "score": 1.0,
+ "content": "and its key role as a structural prior guiding our network to learn the intended reasoning skills.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "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 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 301,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 765
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 765
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 15,
+ "width": 14
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 180,
+ 86,
+ 428,
+ 207
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 180,
+ 86,
+ 428,
+ 207
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 180,
+ 86,
+ 428,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 180,
+ 86,
+ 428,
+ 207
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "c74201d70aae1d3587e14044064872d252c0e00ec6d3c4fd6b70377bb7088c75.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 180,
+ 86,
+ 428,
+ 126.33333333333334
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 180,
+ 126.33333333333334,
+ 428,
+ 166.66666666666669
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 180,
+ 166.66666666666669,
+ 428,
+ 207.00000000000003
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 222,
+ 505,
+ 278
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 504,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 471,
+ 235
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Training curves and accuracies for MACs (our model), FiLM (Perez et al., 2017),",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 472,
+ 223,
+ 504,
+ 233
+ ],
+ "score": 0.53,
+ "content": "\\mathrm { P G + E E }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 234,
+ 504,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 471,
+ 246
+ ],
+ "score": 1.0,
+ "content": "(Johnson et al., 2017) and stacked-attention (Yang et al., 2016; Johnson et al., 2017). (Note:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 472,
+ 234,
+ 504,
+ 244
+ ],
+ "score": 0.54,
+ "content": "\\mathrm { P G + E E }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 245,
+ 505,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 245,
+ 505,
+ 258
+ ],
+ "score": 1.0,
+ "content": "uses the supported CLEVR programs as strong supervision.) Left: Training curve (accuracy/epoch).",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 506,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 256,
+ 270
+ ],
+ "score": 1.0,
+ "content": "Right: Learning curve: Accuracy for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 256,
+ 256,
+ 271,
+ 266
+ ],
+ "score": 0.82,
+ "content": "1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 271,
+ 255,
+ 273,
+ 270
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 274,
+ 256,
+ 289,
+ 266
+ ],
+ "score": 0.79,
+ "content": "2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 289,
+ 255,
+ 292,
+ 270
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 293,
+ 256,
+ 307,
+ 266
+ ],
+ "score": 0.82,
+ "content": "5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 255,
+ 311,
+ 270
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 312,
+ 256,
+ 331,
+ 266
+ ],
+ "score": 0.84,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 255,
+ 334,
+ 270
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 334,
+ 256,
+ 354,
+ 266
+ ],
+ "score": 0.84,
+ "content": "2 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 354,
+ 255,
+ 357,
+ 270
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 256,
+ 378,
+ 266
+ ],
+ "score": 0.85,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 378,
+ 255,
+ 396,
+ 270
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 256,
+ 420,
+ 266
+ ],
+ "score": 0.87,
+ "content": "100 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 255,
+ 447,
+ 270
+ ],
+ "score": 1.0,
+ "content": "of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 447,
+ 256,
+ 469,
+ 266
+ ],
+ "score": 0.68,
+ "content": "7 0 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 469,
+ 255,
+ 506,
+ 270
+ ],
+ "score": 1.0,
+ "content": "CLEVR",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 266,
+ 144,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 266,
+ 144,
+ 280
+ ],
+ "score": 1.0,
+ "content": "samples.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 181,
+ 366,
+ 429,
+ 448
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 312,
+ 505,
+ 356
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "Table 2: Accuracy on CLEVR-Humans of previous methods and our method (MAC), before (left)",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 323,
+ 506,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 369,
+ 336
+ ],
+ "score": 1.0,
+ "content": "and after (right) fine-tuning on the CLEVR-Humans training data.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 370,
+ 324,
+ 402,
+ 334
+ ],
+ "score": 0.75,
+ "content": "\\mathrm { P G + E E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 402,
+ 323,
+ 506,
+ 336
+ ],
+ "score": 1.0,
+ "content": "uses supervised data. For",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": "our method (MAC) we show test results with GLoVE initialization and validation results with and",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 176,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 176,
+ 356
+ ],
+ "score": 1.0,
+ "content": "without GLoVE.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 181,
+ 366,
+ 429,
+ 448
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 181,
+ 366,
+ 429,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 181,
+ 366,
+ 429,
+ 448
+ ],
+ "score": 0.979,
+ "html": "| Model | Train CLEVR | Train CLEVR + fine-tune HUMANS |
| LSTM (Johnson et al.,2017) | 27.5 | 36.5 |
| CNN+LSTM (Johnson et al.,2017) | 37.7 | 43.2 |
| CNN+LSTM+SA+MLP (Johnson et al.,2016) | 50.4 | 57.6 |
| PG+EE (18K prog.)* (Johnson et al.,2017) | 54.0 | 66.6 |
| CNN+GRU+FiLM (Perez et al., 2017) | 56.6 | 75.9 |
| MAC-val | 56.9 | 81.2 |
| MAC-val | 58.5 | 82.0 |
| MAC-test | 58.6 | 82.5 |
",
+ "type": "table",
+ "image_path": "20d0a9d88276e853b2d4290d832c31f6ed7a6d4ae845cb8a1f4754688e965048.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 181,
+ 366,
+ 429,
+ 393.3333333333333
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 181,
+ 393.3333333333333,
+ 429,
+ 420.66666666666663
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 181,
+ 420.66666666666663,
+ 429,
+ 447.99999999999994
+ ],
+ "spans": [],
+ "index": 14
+ }
+ ]
+ }
+ ],
+ "index": 11.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 491,
+ 209,
+ 502
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 210,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 210,
+ 504
+ ],
+ "score": 1.0,
+ "content": "4.1 DATA EFFICIENCY",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 522,
+ 505,
+ 588
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "We have explored the performance of our and other leading approaches on smaller subsets of the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "CLEVR dataset, in order to study the ability of models to generalize from smaller amount of data.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 545,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 303,
+ 556
+ ],
+ "score": 1.0,
+ "content": "We sampled at random subsets of CLEVR, with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 545,
+ 323,
+ 555
+ ],
+ "score": 0.84,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 324,
+ 545,
+ 326,
+ 556
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 545,
+ 347,
+ 555
+ ],
+ "score": 0.84,
+ "content": "2 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 347,
+ 545,
+ 365,
+ 556
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 366,
+ 545,
+ 385,
+ 555
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 386,
+ 545,
+ 444,
+ 556
+ ],
+ "score": 1.0,
+ "content": "of its original",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 545,
+ 466,
+ 555
+ ],
+ "score": 0.44,
+ "content": "7 0 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 545,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "size, and",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 568
+ ],
+ "score": 1.0,
+ "content": "used them to train our and other 3 proposed models for the CLEVR task: FiLM (Perez et al., 2017),",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 204,
+ 578
+ ],
+ "score": 1.0,
+ "content": "the strongly-supervised",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 566,
+ 237,
+ 577
+ ],
+ "score": 0.67,
+ "content": "\\mathrm { P G + E E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 237,
+ 566,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "(Johnson et al., 2017), and stacked-attention networks (Johnson",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 231,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 231,
+ 590
+ ],
+ "score": 1.0,
+ "content": "et al., 2017; Yang et al., 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 105,
+ 523,
+ 506,
+ 590
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 504,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 504,
+ 605
+ ],
+ "score": 1.0,
+ "content": "As shown in figure 4, our model outperforms the other models by a wide margin for all subsets of the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 604,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 604,
+ 193,
+ 617
+ ],
+ "score": 1.0,
+ "content": "CLEVR dataset. For",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 605,
+ 213,
+ 615
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 604,
+ 317,
+ 617
+ ],
+ "score": 1.0,
+ "content": "of the data, equivalent to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 317,
+ 605,
+ 339,
+ 615
+ ],
+ "score": 0.63,
+ "content": "3 5 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 604,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "samples, other models obtain accuracies",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 615,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 615,
+ 176,
+ 629
+ ],
+ "score": 1.0,
+ "content": "ranging between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 616,
+ 196,
+ 627
+ ],
+ "score": 0.88,
+ "content": "70 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 196,
+ 615,
+ 215,
+ 629
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 215,
+ 616,
+ 235,
+ 626
+ ],
+ "score": 0.88,
+ "content": "92 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 615,
+ 345,
+ 629
+ ],
+ "score": 1.0,
+ "content": ", while our model achieves",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 616,
+ 372,
+ 627
+ ],
+ "score": 0.87,
+ "content": "9 7 . 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 373,
+ 615,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": ". The gap becomes larger as the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 205,
+ 640
+ ],
+ "score": 1.0,
+ "content": "dataset size reduces: for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 627,
+ 225,
+ 637
+ ],
+ "score": 0.88,
+ "content": "2 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 225,
+ 626,
+ 327,
+ 640
+ ],
+ "score": 1.0,
+ "content": "of the data, equivalent to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 627,
+ 349,
+ 637
+ ],
+ "score": 0.5,
+ "content": "1 7 5 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 349,
+ 626,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "samples, performance of other models",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 104,
+ 636,
+ 416,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 636,
+ 151,
+ 651
+ ],
+ "score": 1.0,
+ "content": "is between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 151,
+ 638,
+ 171,
+ 649
+ ],
+ "score": 0.86,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 636,
+ 189,
+ 651
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 638,
+ 208,
+ 649
+ ],
+ "score": 0.87,
+ "content": "7 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 636,
+ 348,
+ 651
+ ],
+ "score": 1.0,
+ "content": ", while our model maintains a high",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 348,
+ 638,
+ 375,
+ 649
+ ],
+ "score": 0.83,
+ "content": "9 5 . 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 636,
+ 416,
+ 651
+ ],
+ "score": 1.0,
+ "content": "accuracy.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 104,
+ 594,
+ 506,
+ 651
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 654,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 154,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Finally, for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 655,
+ 173,
+ 666
+ ],
+ "score": 0.87,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 174,
+ 655,
+ 219,
+ 667
+ ],
+ "score": 1.0,
+ "content": "of the data",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 655,
+ 244,
+ 666
+ ],
+ "score": 0.29,
+ "content": "- 7 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 244,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "samples, still a sizeable amount – our model is the only one that",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 284,
+ 678
+ ],
+ "score": 1.0,
+ "content": "manages to generalize, with performance of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 284,
+ 666,
+ 311,
+ 676
+ ],
+ "score": 0.87,
+ "content": "8 4 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 311,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "on average, whereas the other three models fail,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 147,
+ 689
+ ],
+ "score": 1.0,
+ "content": "achieving",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 147,
+ 677,
+ 204,
+ 687
+ ],
+ "score": 0.89,
+ "content": "4 7 . 6 \\% - 5 7 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": ". Note that as pointed out by (Johnson et al., 2016) a simple baseline that",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 686,
+ 506,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 686,
+ 429,
+ 702
+ ],
+ "score": 1.0,
+ "content": "predicts the most frequent answer for each of the question types achieves already",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 688,
+ 456,
+ 699
+ ],
+ "score": 0.87,
+ "content": "4 2 . 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 686,
+ 506,
+ 702
+ ],
+ "score": 1.0,
+ "content": ", suggesting",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "that answering half of the questions correctly means that the competing models barely learn to",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "generalize from the smaller dataset. These results demonstrate the robustness of our architecture",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 484,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 484,
+ 732
+ ],
+ "score": 1.0,
+ "content": "and its key role as a structural prior guiding our network to learn the intended reasoning skills.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 104,
+ 655,
+ 506,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 83,
+ 363,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 364,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 364,
+ 96
+ ],
+ "score": 1.0,
+ "content": "4.2 CLEVR HUMANS - NATURAL LANGUAGE QUESTIONS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 104,
+ 505,
+ 138
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "We analyze our model performance on the CLEVR-Humans dataset (Johnson et al., 2017), consist-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 116,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 116,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "ing of natural language questions collected through crowdsourcing. As such, the dataset has diverse",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 452,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 452,
+ 139
+ ],
+ "score": 1.0,
+ "content": "vocabulary and linguistic variations, and it also demands more varied reasoning skills.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 143,
+ 505,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 328,
+ 156
+ ],
+ "score": 1.0,
+ "content": "Since the training set is relatively small, consisting of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 144,
+ 345,
+ 154
+ ],
+ "score": 0.33,
+ "content": "1 8 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 143,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "samples, we use it to finetune a model",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 168
+ ],
+ "score": 1.0,
+ "content": "pretrained on the standard CLEVR dataset. However, since most of the vocabulary in CLEVR-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 505,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 505,
+ 180
+ ],
+ "score": 1.0,
+ "content": "Humans is not covered by CLEVR, we do not train the word vectors during the pre-training stage,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 505,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 505,
+ 190
+ ],
+ "score": 1.0,
+ "content": "so to prevent drift in their meaning compared to other uncovered words in CLEVR-Humans that",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 188,
+ 222,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 222,
+ 200
+ ],
+ "score": 1.0,
+ "content": "may be semantically related.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 204,
+ 505,
+ 237
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 216
+ ],
+ "score": 1.0,
+ "content": "As shown in table 2, our model achieves state-of-the-art performance on CLEVR-Humans both",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 215,
+ 504,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 481,
+ 227
+ ],
+ "score": 1.0,
+ "content": "before and after fine-tuning. It surpasses the next-best FiLM model, (Perez et al., 2017) by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 215,
+ 504,
+ 226
+ ],
+ "score": 0.79,
+ "content": "6 . 6 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 227,
+ 213,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 182,
+ 239
+ ],
+ "score": 1.0,
+ "content": "percent, achieving",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 227,
+ 209,
+ 237
+ ],
+ "score": 0.85,
+ "content": "8 2 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 227,
+ 213,
+ 239
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 243,
+ 505,
+ 299
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "The results substantiate the model’s robustness against linguistic variations and noise, as well as",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "its ability to adapt to diverse vocabulary and varied reasoning skills. Arguably, the soft attention",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 265,
+ 506,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 265,
+ 506,
+ 279
+ ],
+ "score": 1.0,
+ "content": "performed over the question words allows the model to focus on the words that are most critical to",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "answer the question and translate them to corresponding reasoning operations, giving less attention",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 287,
+ 240,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 240,
+ 300
+ ],
+ "score": 1.0,
+ "content": "to irrelevant linguistic variations.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 317,
+ 182,
+ 327
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 183,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 183,
+ 330
+ ],
+ "score": 1.0,
+ "content": "4.3 ABLATIONS",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 348,
+ 501,
+ 609
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 108,
+ 348,
+ 501,
+ 609
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 348,
+ 501,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 348,
+ 501,
+ 609
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "d542dee22ea6b6971ea171ea3647286d90bc0976dc1e0a42885a1582ff9227cc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 348,
+ 501,
+ 435.0
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 108,
+ 435.0,
+ 501,
+ 522.0
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 108,
+ 522.0,
+ 501,
+ 609.0
+ ],
+ "spans": [],
+ "index": 20
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 145,
+ 624,
+ 464,
+ 636
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 623,
+ 465,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 623,
+ 465,
+ 637
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Ablations and variations of the MAC network. See text for full detail.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 654,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Based on the validation set, we have conducted an ablation study on our model to understand better",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "the contribution of each of its component to the overall performance. We tested each setting on the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 289,
+ 690
+ ],
+ "score": 1.0,
+ "content": "standard 700K CLEVR dataset as well as on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 677,
+ 308,
+ 687
+ ],
+ "score": 0.87,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 676,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "subset of the dataset. See table 3 for the numer-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "ical results. In addition, figure 4.3 presents the training curves for the different settings trained on",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "the standard dataset. Overall, the results demonstrate the robustness of the model to hyperparam-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 721
+ ],
+ "score": 1.0,
+ "content": "eter variations such as network dimension and length, and also the impact of different aspect and",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 272,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 272,
+ 734
+ ],
+ "score": 1.0,
+ "content": "components of MAC on its performance.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "page_idx": 10,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 310,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 83,
+ 363,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 364,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 364,
+ 96
+ ],
+ "score": 1.0,
+ "content": "4.2 CLEVR HUMANS - NATURAL LANGUAGE QUESTIONS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0,
+ "bbox_fs": [
+ 105,
+ 81,
+ 364,
+ 96
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 104,
+ 505,
+ 138
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "We analyze our model performance on the CLEVR-Humans dataset (Johnson et al., 2017), consist-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 116,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 116,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "ing of natural language questions collected through crowdsourcing. As such, the dataset has diverse",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 452,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 452,
+ 139
+ ],
+ "score": 1.0,
+ "content": "vocabulary and linguistic variations, and it also demands more varied reasoning skills.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2,
+ "bbox_fs": [
+ 105,
+ 104,
+ 505,
+ 139
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 143,
+ 505,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 328,
+ 156
+ ],
+ "score": 1.0,
+ "content": "Since the training set is relatively small, consisting of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 144,
+ 345,
+ 154
+ ],
+ "score": 0.33,
+ "content": "1 8 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 143,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "samples, we use it to finetune a model",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 168
+ ],
+ "score": 1.0,
+ "content": "pretrained on the standard CLEVR dataset. However, since most of the vocabulary in CLEVR-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 505,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 505,
+ 180
+ ],
+ "score": 1.0,
+ "content": "Humans is not covered by CLEVR, we do not train the word vectors during the pre-training stage,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 505,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 505,
+ 190
+ ],
+ "score": 1.0,
+ "content": "so to prevent drift in their meaning compared to other uncovered words in CLEVR-Humans that",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 188,
+ 222,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 222,
+ 200
+ ],
+ "score": 1.0,
+ "content": "may be semantically related.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 143,
+ 505,
+ 200
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 204,
+ 505,
+ 237
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 216
+ ],
+ "score": 1.0,
+ "content": "As shown in table 2, our model achieves state-of-the-art performance on CLEVR-Humans both",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 215,
+ 504,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 481,
+ 227
+ ],
+ "score": 1.0,
+ "content": "before and after fine-tuning. It surpasses the next-best FiLM model, (Perez et al., 2017) by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 215,
+ 504,
+ 226
+ ],
+ "score": 0.79,
+ "content": "6 . 6 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 227,
+ 213,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 182,
+ 239
+ ],
+ "score": 1.0,
+ "content": "percent, achieving",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 227,
+ 209,
+ 237
+ ],
+ "score": 0.85,
+ "content": "8 2 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 227,
+ 213,
+ 239
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 105,
+ 204,
+ 505,
+ 239
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 243,
+ 505,
+ 299
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "The results substantiate the model’s robustness against linguistic variations and noise, as well as",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "its ability to adapt to diverse vocabulary and varied reasoning skills. Arguably, the soft attention",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 265,
+ 506,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 265,
+ 506,
+ 279
+ ],
+ "score": 1.0,
+ "content": "performed over the question words allows the model to focus on the words that are most critical to",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "answer the question and translate them to corresponding reasoning operations, giving less attention",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 287,
+ 240,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 240,
+ 300
+ ],
+ "score": 1.0,
+ "content": "to irrelevant linguistic variations.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14,
+ "bbox_fs": [
+ 104,
+ 243,
+ 506,
+ 300
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 317,
+ 182,
+ 327
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 183,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 183,
+ 330
+ ],
+ "score": 1.0,
+ "content": "4.3 ABLATIONS",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 348,
+ 501,
+ 609
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 108,
+ 348,
+ 501,
+ 609
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 348,
+ 501,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 348,
+ 501,
+ 609
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "d542dee22ea6b6971ea171ea3647286d90bc0976dc1e0a42885a1582ff9227cc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 348,
+ 501,
+ 435.0
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 108,
+ 435.0,
+ 501,
+ 522.0
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 108,
+ 522.0,
+ 501,
+ 609.0
+ ],
+ "spans": [],
+ "index": 20
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 145,
+ 624,
+ 464,
+ 636
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 623,
+ 465,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 623,
+ 465,
+ 637
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Ablations and variations of the MAC network. See text for full detail.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 654,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Based on the validation set, we have conducted an ablation study on our model to understand better",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "the contribution of each of its component to the overall performance. We tested each setting on the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 289,
+ 690
+ ],
+ "score": 1.0,
+ "content": "standard 700K CLEVR dataset as well as on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 677,
+ 308,
+ 687
+ ],
+ "score": 0.87,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 676,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "subset of the dataset. See table 3 for the numer-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "ical results. In addition, figure 4.3 presents the training curves for the different settings trained on",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "the standard dataset. Overall, the results demonstrate the robustness of the model to hyperparam-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 721
+ ],
+ "score": 1.0,
+ "content": "eter variations such as network dimension and length, and also the impact of different aspect and",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 272,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 272,
+ 734
+ ],
+ "score": 1.0,
+ "content": "components of MAC on its performance.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 105,
+ 655,
+ 505,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 187,
+ 111,
+ 423,
+ 265
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 80,
+ 504,
+ 102
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 79,
+ 505,
+ 92
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 79,
+ 383,
+ 92
+ ],
+ "score": 1.0,
+ "content": "Table 3: Ablations Results for the full standard CLEVR dataset and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 384,
+ 81,
+ 403,
+ 91
+ ],
+ "score": 0.85,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 403,
+ 79,
+ 505,
+ 92
+ ],
+ "score": 1.0,
+ "content": "subset of it. See text for",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 89,
+ 155,
+ 104
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 89,
+ 155,
+ 104
+ ],
+ "score": 1.0,
+ "content": "full details.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 187,
+ 111,
+ 423,
+ 265
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 187,
+ 111,
+ 423,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 187,
+ 111,
+ 423,
+ 265
+ ],
+ "score": 0.976,
+ "html": "| Model | Standard CLEVR | 10% CLEVR |
| MAC standard model (dimension 512) | 98.9 | 84.5 |
| dimension 256 | 98.4 | 76.3 |
| dimension 128 | 97.6 | 77.0 |
| w/o control | 55.6 | 51.5 |
| w/o word-attention | 95.3 | 63.2 |
| full-question-as-control | 80.7 | 65.0 |
| w/o positional encoding | 99.1 | 80.0 |
| linear position encoding | 99.0 | 83.2 |
| identity writing unit | 98.5 | 84.5 |
| only-gating writing unit | 99.3 | 83.1 |
| + self-attention | 83.2 | 99.2 |
| + gating | 99.4 | 83.1 |
| + self-attention and gating | 99.5 | 85.5 |
| gate bias 0 | 98.7 | 84.9 |
| gate bias 1 | 99.4 | 68.5 |
| gate bias -1 | 99.0 | 77.1 |
| unshared weights | 97.8 | 67.5 |
| prediction without question information | 97.8 | 64.7 |
",
+ "type": "table",
+ "image_path": "cbfa912b869921d24b8997ef84de01c88a55b56dcf5407ed4090070e65e15576.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 187,
+ 111,
+ 423,
+ 125.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 187,
+ 125.0,
+ 423,
+ 139.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 187,
+ 139.0,
+ 423,
+ 153.0
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 187,
+ 153.0,
+ 423,
+ 167.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 187,
+ 167.0,
+ 423,
+ 181.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 187,
+ 181.0,
+ 423,
+ 195.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 187,
+ 195.0,
+ 423,
+ 209.0
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 187,
+ 209.0,
+ 423,
+ 223.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 187,
+ 223.0,
+ 423,
+ 237.0
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 187,
+ 237.0,
+ 423,
+ 251.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 187,
+ 251.0,
+ 423,
+ 265.0
+ ],
+ "spans": [],
+ "index": 12
+ }
+ ]
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 302,
+ 505,
+ 380
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 302,
+ 505,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 302,
+ 505,
+ 314
+ ],
+ "score": 1.0,
+ "content": "Network Length. We have tested the model performance as a function of the network’s length –",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "the number of MAC cells that were sequenced together. The results show the positive correlation",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 323,
+ 505,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 505,
+ 336
+ ],
+ "score": 1.0,
+ "content": "between the network length and its performance. We can see that for 1 cell the scores are relatively",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 335,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 335,
+ 152,
+ 345
+ ],
+ "score": 0.4,
+ "content": "1 0 \\mathrm { w } - 7 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 335,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": ", but adding at least one more cell leads to a significant increase in performance above",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 505,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 346,
+ 126,
+ 357
+ ],
+ "score": 0.85,
+ "content": "9 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 126,
+ 345,
+ 403,
+ 359
+ ],
+ "score": 1.0,
+ "content": ". The performance keeps improving up to lengths 8-16 that achieve",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 404,
+ 346,
+ 451,
+ 357
+ ],
+ "score": 0.87,
+ "content": "9 8 . 9 – 9 9 . 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 345,
+ 505,
+ 359
+ ],
+ "score": 1.0,
+ "content": ". The results",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 506,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 506,
+ 370
+ ],
+ "score": 1.0,
+ "content": "also teach us about the complexity of the dataset, by showing the relatively significant benefits of",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 367,
+ 327,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 327,
+ 381
+ ],
+ "score": 1.0,
+ "content": "having at least 4 cells, each modeling a reasoning step.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 385,
+ 505,
+ 440
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 385,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 385,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "Network Dimension. We have varied the state dimension to check the robustness of the model to",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 395,
+ 506,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 395,
+ 506,
+ 408
+ ],
+ "score": 1.0,
+ "content": "hyperparameters. The results on the standard CLEVR dataset show the model is able to maintain",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 474,
+ 420
+ ],
+ "score": 1.0,
+ "content": "high performance with dimension of 128, albeit after a longer training process, achieving",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 407,
+ 501,
+ 418
+ ],
+ "score": 0.84,
+ "content": "9 7 . 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 406,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 160,
+ 432
+ ],
+ "score": 1.0,
+ "content": "compared to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 418,
+ 193,
+ 428
+ ],
+ "score": 0.88,
+ "content": "9 8 . 9 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 193,
+ 416,
+ 392,
+ 432
+ ],
+ "score": 1.0,
+ "content": "achieved with dimension of 512. However, for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 418,
+ 411,
+ 428
+ ],
+ "score": 0.87,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 411,
+ 416,
+ 506,
+ 432
+ ],
+ "score": 1.0,
+ "content": "of CLEVR, the larger",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 400,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 282,
+ 442
+ ],
+ "score": 1.0,
+ "content": "512-dimension allows accuracy increase by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 429,
+ 304,
+ 439
+ ],
+ "score": 0.86,
+ "content": "7 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 305,
+ 428,
+ 400,
+ 442
+ ],
+ "score": 1.0,
+ "content": "over dimension of 128.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 445,
+ 505,
+ 511
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 457
+ ],
+ "score": 1.0,
+ "content": "Weight Sharing. We have tested the impact of sharing weights between cell has on the model",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 249,
+ 469
+ ],
+ "score": 1.0,
+ "content": "performance for network of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 249,
+ 457,
+ 281,
+ 468
+ ],
+ "score": 0.91,
+ "content": "p = 1 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 456,
+ 506,
+ 469
+ ],
+ "score": 1.0,
+ "content": ". The results show that for the standard dataset there is",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 308,
+ 480
+ ],
+ "score": 1.0,
+ "content": "only a small difference between these settings of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 468,
+ 323,
+ 478
+ ],
+ "score": 0.86,
+ "content": "1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 323,
+ 467,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": ". However, for less data, we see much more",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 172,
+ 491
+ ],
+ "score": 1.0,
+ "content": "significant drop",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 478,
+ 199,
+ 489
+ ],
+ "score": 0.86,
+ "content": "1 6 . 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 199,
+ 479,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "in the unshared-parameters setting compared to the shared one. Indeed, we",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "observe that a model with less parameter is more data-efficient and has a lower tendency to overfit",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 499,
+ 143,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 499,
+ 143,
+ 513
+ ],
+ "score": 1.0,
+ "content": "the data.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 639
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "Control Unit. We have performed several ablations in the control unit to understand its contribution",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "to the overall model performance. Based on the results, first, we can see the the question information",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "is crucial for the model to handle the questions, as can be noted by the low performance of the model",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "when there is no use of control signal whatsoever. Second, we have tested the model performance",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 575
+ ],
+ "score": 1.0,
+ "content": "when using the continuous control state computed by question (2) in section 3.2.1, without having",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 573,
+ 505,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 505,
+ 584
+ ],
+ "score": 1.0,
+ "content": "word-attention, in order to understand its relative contribution. Based on the results, we can indeed",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 596
+ ],
+ "score": 1.0,
+ "content": "see that using word-attention is useful for accelerating the training process and achieving higher",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "accuracies both for the standard dataset as well as for the small subset, where using word-attention",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 504,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 183,
+ 617
+ ],
+ "score": 1.0,
+ "content": "increases results in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 605,
+ 211,
+ 615
+ ],
+ "score": 0.87,
+ "content": "2 1 . 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 211,
+ 604,
+ 504,
+ 617
+ ],
+ "score": 1.0,
+ "content": ". We also see that using the “contextual words” produced by the question-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 629
+ ],
+ "score": 1.0,
+ "content": "unit LSTM is useful in accelerating the model performance, when compared to using the word-",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 172,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 172,
+ 639
+ ],
+ "score": 1.0,
+ "content": "vectors directly.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 643,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 643,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 643,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "Reading Unit. We have conducted several ablations for the reading unit to better understand its",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "behavior and contribution to the performance of the model. The standard MAC reading unit uses the",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "control state – which averages the question words based on attention distributions computed per each",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "reasoning step. In this ablation experiment, we have tested using the full question representation",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 690,
+ 113,
+ 699
+ ],
+ "score": 0.74,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 113,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "instead across all reasoning steps to gain better understanding of the the contribution of word-",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 104,
+ 698,
+ 504,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 698,
+ 374,
+ 712
+ ],
+ "score": 1.0,
+ "content": "attention to the model performance. Indeed, we can see that using",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 701,
+ 381,
+ 710
+ ],
+ "score": 0.77,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 698,
+ 495,
+ 712
+ ],
+ "score": 1.0,
+ "content": "rather then the control state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 701,
+ 504,
+ 710
+ ],
+ "score": 0.8,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 504,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 282,
+ 721
+ ],
+ "score": 1.0,
+ "content": "results in a significant drops in performance",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 710,
+ 316,
+ 720
+ ],
+ "score": 0.8,
+ "content": "- 1 9 . 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 710,
+ 443,
+ 721
+ ],
+ "score": 1.0,
+ "content": "for the full CLEVR dataset and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 710,
+ 470,
+ 720
+ ],
+ "score": 0.86,
+ "content": "1 9 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 710,
+ 484,
+ 721
+ ],
+ "score": 1.0,
+ "content": "for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 485,
+ 710,
+ 504,
+ 720
+ ],
+ "score": 0.85,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 154,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 154,
+ 732
+ ],
+ "score": 1.0,
+ "content": "of the data.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 45.5
+ }
+ ],
+ "page_idx": 11,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "12",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 187,
+ 111,
+ 423,
+ 265
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 80,
+ 504,
+ 102
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 79,
+ 505,
+ 92
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 79,
+ 383,
+ 92
+ ],
+ "score": 1.0,
+ "content": "Table 3: Ablations Results for the full standard CLEVR dataset and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 384,
+ 81,
+ 403,
+ 91
+ ],
+ "score": 0.85,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 403,
+ 79,
+ 505,
+ 92
+ ],
+ "score": 1.0,
+ "content": "subset of it. See text for",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 89,
+ 155,
+ 104
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 89,
+ 155,
+ 104
+ ],
+ "score": 1.0,
+ "content": "full details.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 187,
+ 111,
+ 423,
+ 265
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 187,
+ 111,
+ 423,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 187,
+ 111,
+ 423,
+ 265
+ ],
+ "score": 0.976,
+ "html": "| Model | Standard CLEVR | 10% CLEVR |
| MAC standard model (dimension 512) | 98.9 | 84.5 |
| dimension 256 | 98.4 | 76.3 |
| dimension 128 | 97.6 | 77.0 |
| w/o control | 55.6 | 51.5 |
| w/o word-attention | 95.3 | 63.2 |
| full-question-as-control | 80.7 | 65.0 |
| w/o positional encoding | 99.1 | 80.0 |
| linear position encoding | 99.0 | 83.2 |
| identity writing unit | 98.5 | 84.5 |
| only-gating writing unit | 99.3 | 83.1 |
| + self-attention | 83.2 | 99.2 |
| + gating | 99.4 | 83.1 |
| + self-attention and gating | 99.5 | 85.5 |
| gate bias 0 | 98.7 | 84.9 |
| gate bias 1 | 99.4 | 68.5 |
| gate bias -1 | 99.0 | 77.1 |
| unshared weights | 97.8 | 67.5 |
| prediction without question information | 97.8 | 64.7 |
",
+ "type": "table",
+ "image_path": "cbfa912b869921d24b8997ef84de01c88a55b56dcf5407ed4090070e65e15576.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 187,
+ 111,
+ 423,
+ 125.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 187,
+ 125.0,
+ 423,
+ 139.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 187,
+ 139.0,
+ 423,
+ 153.0
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 187,
+ 153.0,
+ 423,
+ 167.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 187,
+ 167.0,
+ 423,
+ 181.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 187,
+ 181.0,
+ 423,
+ 195.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 187,
+ 195.0,
+ 423,
+ 209.0
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 187,
+ 209.0,
+ 423,
+ 223.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 187,
+ 223.0,
+ 423,
+ 237.0
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 187,
+ 237.0,
+ 423,
+ 251.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 187,
+ 251.0,
+ 423,
+ 265.0
+ ],
+ "spans": [],
+ "index": 12
+ }
+ ]
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 302,
+ 505,
+ 380
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 302,
+ 505,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 302,
+ 505,
+ 314
+ ],
+ "score": 1.0,
+ "content": "Network Length. We have tested the model performance as a function of the network’s length –",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 312,
+ 505,
+ 325
+ ],
+ "score": 1.0,
+ "content": "the number of MAC cells that were sequenced together. The results show the positive correlation",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 323,
+ 505,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 505,
+ 336
+ ],
+ "score": 1.0,
+ "content": "between the network length and its performance. We can see that for 1 cell the scores are relatively",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 335,
+ 506,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 335,
+ 152,
+ 345
+ ],
+ "score": 0.4,
+ "content": "1 0 \\mathrm { w } - 7 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 335,
+ 506,
+ 347
+ ],
+ "score": 1.0,
+ "content": ", but adding at least one more cell leads to a significant increase in performance above",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 505,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 346,
+ 126,
+ 357
+ ],
+ "score": 0.85,
+ "content": "9 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 126,
+ 345,
+ 403,
+ 359
+ ],
+ "score": 1.0,
+ "content": ". The performance keeps improving up to lengths 8-16 that achieve",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 404,
+ 346,
+ 451,
+ 357
+ ],
+ "score": 0.87,
+ "content": "9 8 . 9 – 9 9 . 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 345,
+ 505,
+ 359
+ ],
+ "score": 1.0,
+ "content": ". The results",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 506,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 506,
+ 370
+ ],
+ "score": 1.0,
+ "content": "also teach us about the complexity of the dataset, by showing the relatively significant benefits of",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 367,
+ 327,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 327,
+ 381
+ ],
+ "score": 1.0,
+ "content": "having at least 4 cells, each modeling a reasoning step.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 105,
+ 302,
+ 506,
+ 381
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 385,
+ 505,
+ 440
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 385,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 385,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "Network Dimension. We have varied the state dimension to check the robustness of the model to",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 395,
+ 506,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 395,
+ 506,
+ 408
+ ],
+ "score": 1.0,
+ "content": "hyperparameters. The results on the standard CLEVR dataset show the model is able to maintain",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 474,
+ 420
+ ],
+ "score": 1.0,
+ "content": "high performance with dimension of 128, albeit after a longer training process, achieving",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 407,
+ 501,
+ 418
+ ],
+ "score": 0.84,
+ "content": "9 7 . 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 406,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 160,
+ 432
+ ],
+ "score": 1.0,
+ "content": "compared to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 418,
+ 193,
+ 428
+ ],
+ "score": 0.88,
+ "content": "9 8 . 9 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 193,
+ 416,
+ 392,
+ 432
+ ],
+ "score": 1.0,
+ "content": "achieved with dimension of 512. However, for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 418,
+ 411,
+ 428
+ ],
+ "score": 0.87,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 411,
+ 416,
+ 506,
+ 432
+ ],
+ "score": 1.0,
+ "content": "of CLEVR, the larger",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 400,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 282,
+ 442
+ ],
+ "score": 1.0,
+ "content": "512-dimension allows accuracy increase by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 429,
+ 304,
+ 439
+ ],
+ "score": 0.86,
+ "content": "7 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 305,
+ 428,
+ 400,
+ 442
+ ],
+ "score": 1.0,
+ "content": "over dimension of 128.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 385,
+ 506,
+ 442
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 445,
+ 505,
+ 511
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 457
+ ],
+ "score": 1.0,
+ "content": "Weight Sharing. We have tested the impact of sharing weights between cell has on the model",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 249,
+ 469
+ ],
+ "score": 1.0,
+ "content": "performance for network of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 249,
+ 457,
+ 281,
+ 468
+ ],
+ "score": 0.91,
+ "content": "p = 1 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 456,
+ 506,
+ 469
+ ],
+ "score": 1.0,
+ "content": ". The results show that for the standard dataset there is",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 308,
+ 480
+ ],
+ "score": 1.0,
+ "content": "only a small difference between these settings of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 468,
+ 323,
+ 478
+ ],
+ "score": 0.86,
+ "content": "1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 323,
+ 467,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": ". However, for less data, we see much more",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 172,
+ 491
+ ],
+ "score": 1.0,
+ "content": "significant drop",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 478,
+ 199,
+ 489
+ ],
+ "score": 0.86,
+ "content": "1 6 . 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 199,
+ 479,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "in the unshared-parameters setting compared to the shared one. Indeed, we",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "observe that a model with less parameter is more data-efficient and has a lower tendency to overfit",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 499,
+ 143,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 499,
+ 143,
+ 513
+ ],
+ "score": 1.0,
+ "content": "the data.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 446,
+ 506,
+ 513
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 639
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "Control Unit. We have performed several ablations in the control unit to understand its contribution",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "to the overall model performance. Based on the results, first, we can see the the question information",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "is crucial for the model to handle the questions, as can be noted by the low performance of the model",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "when there is no use of control signal whatsoever. Second, we have tested the model performance",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 575
+ ],
+ "score": 1.0,
+ "content": "when using the continuous control state computed by question (2) in section 3.2.1, without having",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 573,
+ 505,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 505,
+ 584
+ ],
+ "score": 1.0,
+ "content": "word-attention, in order to understand its relative contribution. Based on the results, we can indeed",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 596
+ ],
+ "score": 1.0,
+ "content": "see that using word-attention is useful for accelerating the training process and achieving higher",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "accuracies both for the standard dataset as well as for the small subset, where using word-attention",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 504,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 183,
+ 617
+ ],
+ "score": 1.0,
+ "content": "increases results in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 605,
+ 211,
+ 615
+ ],
+ "score": 0.87,
+ "content": "2 1 . 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 211,
+ 604,
+ 504,
+ 617
+ ],
+ "score": 1.0,
+ "content": ". We also see that using the “contextual words” produced by the question-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 629
+ ],
+ "score": 1.0,
+ "content": "unit LSTM is useful in accelerating the model performance, when compared to using the word-",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 172,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 172,
+ 639
+ ],
+ "score": 1.0,
+ "content": "vectors directly.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 517,
+ 506,
+ 639
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 643,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 643,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 643,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "Reading Unit. We have conducted several ablations for the reading unit to better understand its",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "behavior and contribution to the performance of the model. The standard MAC reading unit uses the",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "control state – which averages the question words based on attention distributions computed per each",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "reasoning step. In this ablation experiment, we have tested using the full question representation",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 690,
+ 113,
+ 699
+ ],
+ "score": 0.74,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 113,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "instead across all reasoning steps to gain better understanding of the the contribution of word-",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 104,
+ 698,
+ 504,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 698,
+ 374,
+ 712
+ ],
+ "score": 1.0,
+ "content": "attention to the model performance. Indeed, we can see that using",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 701,
+ 381,
+ 710
+ ],
+ "score": 0.77,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 698,
+ 495,
+ 712
+ ],
+ "score": 1.0,
+ "content": "rather then the control state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 701,
+ 504,
+ 710
+ ],
+ "score": 0.8,
+ "content": "c _ { i }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 504,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 282,
+ 721
+ ],
+ "score": 1.0,
+ "content": "results in a significant drops in performance",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 710,
+ 316,
+ 720
+ ],
+ "score": 0.8,
+ "content": "- 1 9 . 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 710,
+ 443,
+ 721
+ ],
+ "score": 1.0,
+ "content": "for the full CLEVR dataset and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 710,
+ 470,
+ 720
+ ],
+ "score": 0.86,
+ "content": "1 9 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 710,
+ 484,
+ 721
+ ],
+ "score": 1.0,
+ "content": "for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 485,
+ 710,
+ 504,
+ 720
+ ],
+ "score": 0.85,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 154,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 154,
+ 732
+ ],
+ "score": 1.0,
+ "content": "of the data.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 45.5,
+ "bbox_fs": [
+ 104,
+ 643,
+ 505,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 159
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "We have conducted additional ablation experiment to better understand the contribution of using",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "the KB features directly in the first-stage information retrieval process described in section 3.2.2,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 102,
+ 506,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 102,
+ 477,
+ 119
+ ],
+ "score": 1.0,
+ "content": "compared to using only the dot-products of the KB elements with the previous memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 478,
+ 106,
+ 501,
+ 116
+ ],
+ "score": 0.88,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 102,
+ 506,
+ 119
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "For the full CLEVR dataset, we can see that this component has only a small impact in the final",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 127,
+ 258,
+ 138
+ ],
+ "score": 1.0,
+ "content": "performance - ultimately resulting in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 126,
+ 285,
+ 137
+ ],
+ "score": 0.87,
+ "content": "0 . 0 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 127,
+ 458,
+ 138
+ ],
+ "score": 1.0,
+ "content": "performance difference. However, for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 458,
+ 127,
+ 478,
+ 137
+ ],
+ "score": 0.86,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 127,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "of the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "score": 1.0,
+ "content": "data, we can see that the difference in performance when ablating this component is much larger -",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 148,
+ 138,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 148,
+ 133,
+ 159
+ ],
+ "score": 0.83,
+ "content": "1 1 . 2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 134,
+ 148,
+ 138,
+ 160
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 165,
+ 504,
+ 253
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 165,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 165,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Writing Unit Ablations. In our main MAC model variant, the memory unit merges the new in-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 149,
+ 190
+ ],
+ "score": 1.0,
+ "content": "formation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 149,
+ 178,
+ 174,
+ 187
+ ],
+ "score": 0.88,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 174,
+ 176,
+ 306,
+ 190
+ ],
+ "score": 1.0,
+ "content": "with the previous memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 306,
+ 178,
+ 329,
+ 188
+ ],
+ "score": 0.89,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 330,
+ 176,
+ 506,
+ 190
+ ],
+ "score": 1.0,
+ "content": "by combining them through a linear trans-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 504,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 452,
+ 201
+ ],
+ "score": 1.0,
+ "content": "formation. In this experiment, we have explored other variations, such as assigning",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 452,
+ 189,
+ 477,
+ 199
+ ],
+ "score": 0.88,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 187,
+ 490,
+ 201
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 189,
+ 504,
+ 198
+ ],
+ "score": 0.83,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 199,
+ 506,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 436,
+ 210
+ ],
+ "score": 1.0,
+ "content": "directly – ignoring previous memories, or doing a linear transformation based on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 436,
+ 200,
+ 461,
+ 209
+ ],
+ "score": 0.88,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 199,
+ 506,
+ 210
+ ],
+ "score": 1.0,
+ "content": "only. The",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 443,
+ 221
+ ],
+ "score": 1.0,
+ "content": "results show that in fact such variant is only slightly worse than our main variant –",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 210,
+ 465,
+ 220
+ ],
+ "score": 0.73,
+ "content": "0 . 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 465,
+ 208,
+ 506,
+ 221
+ ],
+ "score": 1.0,
+ "content": ". We also",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 505,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 505,
+ 233
+ ],
+ "score": 1.0,
+ "content": "conducted an experiment in which we merge the new information with the previous memory just by",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 243
+ ],
+ "score": 1.0,
+ "content": "a having a gate that does a weighted average of them. The results show that this variant performs",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 243,
+ 339,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 339,
+ 253
+ ],
+ "score": 1.0,
+ "content": "equivalently to our standard linear-transformation variant.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 258,
+ 504,
+ 336
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 271
+ ],
+ "score": 1.0,
+ "content": "Writing Unit Additions. We have explored the impact of the writing unit variants described in sec-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 269,
+ 506,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 269,
+ 506,
+ 283
+ ],
+ "score": 1.0,
+ "content": "tion 3.2.3 – adding self-attention, gating mechanisms, or both, compared to our standard main model",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 406,
+ 294
+ ],
+ "score": 1.0,
+ "content": "that uses a linear transformation to merge the newly retrieved information",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 282,
+ 431,
+ 292
+ ],
+ "score": 0.89,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 432,
+ 280,
+ 506,
+ 294
+ ],
+ "score": 1.0,
+ "content": "with the previous",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 104,
+ 290,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 290,
+ 142,
+ 306
+ ],
+ "score": 1.0,
+ "content": "memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 294,
+ 155,
+ 303
+ ],
+ "score": 0.86,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 290,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": ". For the complete CLEVR dataset we can see that indeed both these variants are very",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 315
+ ],
+ "score": 1.0,
+ "content": "helpful in increasing the model performance. Compared to our standard MAC model that achieves",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 313,
+ 506,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 138,
+ 324
+ ],
+ "score": 0.87,
+ "content": "9 8 . 9 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 139,
+ 313,
+ 362,
+ 326
+ ],
+ "score": 1.0,
+ "content": "on the validation set, self-attention yields accuracy of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 362,
+ 314,
+ 394,
+ 325
+ ],
+ "score": 0.87,
+ "content": "9 9 . 2 3 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 394,
+ 313,
+ 453,
+ 326
+ ],
+ "score": 1.0,
+ "content": ", gating yields",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 454,
+ 314,
+ 486,
+ 324
+ ],
+ "score": 0.87,
+ "content": "9 9 . 3 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 313,
+ 506,
+ 326
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 228,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 192,
+ 337
+ ],
+ "score": 1.0,
+ "content": "adding both achieves",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 325,
+ 225,
+ 335
+ ],
+ "score": 0.87,
+ "content": "9 9 . 4 8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 225,
+ 325,
+ 228,
+ 337
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 342,
+ 505,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 505,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 505,
+ 354
+ ],
+ "score": 1.0,
+ "content": "Output Unit. In our standard model, the final predictions made in the output unit are based on the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 353,
+ 183,
+ 365
+ ],
+ "score": 1.0,
+ "content": "final memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 354,
+ 198,
+ 365
+ ],
+ "score": 0.87,
+ "content": "m _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 199,
+ 353,
+ 336,
+ 365
+ ],
+ "score": 1.0,
+ "content": "as well as question representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 337,
+ 354,
+ 343,
+ 364
+ ],
+ "score": 0.77,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 344,
+ 353,
+ 505,
+ 365
+ ],
+ "score": 1.0,
+ "content": "(stands for the final hidden states of the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 376
+ ],
+ "score": 1.0,
+ "content": "backward and forwards passes of the LSTM). We have explored the contribution of basing the model",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 388
+ ],
+ "score": 1.0,
+ "content": "prediction on the latter, by testing the model performance when prediction is based on memory",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 225,
+ 398
+ ],
+ "score": 1.0,
+ "content": "alone, for the complete and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 385,
+ 245,
+ 396
+ ],
+ "score": 0.87,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 245,
+ 385,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": "datasets. We can see that in both settings basing the model’s",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 409
+ ],
+ "score": 1.0,
+ "content": "predictions on the question representation allows faster training and higher accuracies. Notable is",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 231,
+ 420
+ ],
+ "score": 1.0,
+ "content": "the gap in performance for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 407,
+ 251,
+ 418
+ ],
+ "score": 0.85,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 408,
+ 294,
+ 420
+ ],
+ "score": 1.0,
+ "content": "CLEVR -",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 294,
+ 407,
+ 321,
+ 418
+ ],
+ "score": 0.82,
+ "content": "1 9 . 8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 408,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "increase by using the question representation",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 431
+ ],
+ "score": 1.0,
+ "content": "to make predictions. These results are very reasonable intuitively, since the model is structured such",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "that the memory holds only information that was retrieved from the image. Thus, questions that may",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 454
+ ],
+ "score": 1.0,
+ "content": "ask for instance on different aspects (such as color or shape) of the same object in the image may",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "result in the same memory content, which is thus does not directly contain enough information to",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 462,
+ 204,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 204,
+ 474
+ ],
+ "score": 1.0,
+ "content": "respond such questions.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 479,
+ 505,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "Position. In our standard model, similarly to the practice of competing models (Santoro et al., 2017;",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "Perez et al., 2017; Hu et al., 2017), we have concatenated positional information to each region of the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "image, in order to increase the model capability to perform spatial reasoning. We have explored both",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 237,
+ 525
+ ],
+ "score": 1.0,
+ "content": "simple linear maps at a constant",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 238,
+ 512,
+ 266,
+ 524
+ ],
+ "score": 0.9,
+ "content": "[ - 1 , 1 ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 267,
+ 511,
+ 505,
+ 525
+ ],
+ "score": 1.0,
+ "content": "as well as more complex positional encoding suggested by",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 522,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 522,
+ 429,
+ 534
+ ],
+ "score": 1.0,
+ "content": "(Vaswani et al., 2017). However, the results for both the standard dataset and the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 523,
+ 449,
+ 533
+ ],
+ "score": 0.87,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 449,
+ 522,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "version show",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 104,
+ 533,
+ 506,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 533,
+ 506,
+ 547
+ ],
+ "score": 1.0,
+ "content": "a very negligible improvement at best when adding positional encoding information, demonstrating",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 546,
+ 424,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 546,
+ 424,
+ 557
+ ],
+ "score": 1.0,
+ "content": "the capability of MAC to perform spatial reasoning without data augmentation.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 561,
+ 505,
+ 650
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "Gate Bias Initialization. For our model variant with gating mechanism (described in section 3.2.3)",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 382,
+ 585
+ ],
+ "score": 1.0,
+ "content": "we have tested the effect of setting different values for the gate bias -",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 382,
+ 573,
+ 406,
+ 584
+ ],
+ "score": 0.85,
+ "content": "- 1 , 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 406,
+ 573,
+ 448,
+ 585
+ ],
+ "score": 1.0,
+ "content": "and 1. for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 448,
+ 573,
+ 462,
+ 583
+ ],
+ "score": 0.66,
+ "content": "- 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 573,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "the model",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 596
+ ],
+ "score": 1.0,
+ "content": "is initialized to biased for keeping the previous memory value whereas for 1 it will be biased for",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 595,
+ 505,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 505,
+ 607
+ ],
+ "score": 1.0,
+ "content": "using the new memory instead. We can see that for the complete dataset setting the bias to 1 is",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "optimal – apparently since the model has enough data to learn to apply each cell effectively. In",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 195,
+ 629
+ ],
+ "score": 1.0,
+ "content": "contrast, for the small",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 617,
+ 215,
+ 627
+ ],
+ "score": 0.85,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 616,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "CLEVR data, setting the bias to 0 shows better performance, biasing the",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "model to using less cells overall which results ultimately in a theoretically-simpler model that can",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 223,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 223,
+ 651
+ ],
+ "score": 1.0,
+ "content": "fit less data more effectively.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 44.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 674,
+ 213,
+ 685
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 674,
+ 215,
+ 686
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 674,
+ 215,
+ 686
+ ],
+ "score": 1.0,
+ "content": "4.4 INTERPRETABILITY",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 49
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "We have looked into attention maps over the image and question that the model produces during its",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "computation and provide a few examples in figure 4.4. The first example shows us how the model",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "parses the question in steps, first focusing on the main entity that the question is about, then on",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "page_idx": 12,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 301,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 298,
+ 750,
+ 312,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 298,
+ 750,
+ 312,
+ 763
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 13,
+ "width": 14
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 159
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "We have conducted additional ablation experiment to better understand the contribution of using",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "the KB features directly in the first-stage information retrieval process described in section 3.2.2,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 102,
+ 506,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 102,
+ 477,
+ 119
+ ],
+ "score": 1.0,
+ "content": "compared to using only the dot-products of the KB elements with the previous memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 478,
+ 106,
+ 501,
+ 116
+ ],
+ "score": 0.88,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 102,
+ 506,
+ 119
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "For the full CLEVR dataset, we can see that this component has only a small impact in the final",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 127,
+ 258,
+ 138
+ ],
+ "score": 1.0,
+ "content": "performance - ultimately resulting in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 126,
+ 285,
+ 137
+ ],
+ "score": 0.87,
+ "content": "0 . 0 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 127,
+ 458,
+ 138
+ ],
+ "score": 1.0,
+ "content": "performance difference. However, for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 458,
+ 127,
+ 478,
+ 137
+ ],
+ "score": 0.86,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 127,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "of the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "score": 1.0,
+ "content": "data, we can see that the difference in performance when ablating this component is much larger -",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 148,
+ 138,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 148,
+ 133,
+ 159
+ ],
+ "score": 0.83,
+ "content": "1 1 . 2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 134,
+ 148,
+ 138,
+ 160
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 104,
+ 81,
+ 506,
+ 160
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 165,
+ 504,
+ 253
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 165,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 165,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Writing Unit Ablations. In our main MAC model variant, the memory unit merges the new in-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 149,
+ 190
+ ],
+ "score": 1.0,
+ "content": "formation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 149,
+ 178,
+ 174,
+ 187
+ ],
+ "score": 0.88,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 174,
+ 176,
+ 306,
+ 190
+ ],
+ "score": 1.0,
+ "content": "with the previous memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 306,
+ 178,
+ 329,
+ 188
+ ],
+ "score": 0.89,
+ "content": "m _ { i - 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 330,
+ 176,
+ 506,
+ 190
+ ],
+ "score": 1.0,
+ "content": "by combining them through a linear trans-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 504,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 452,
+ 201
+ ],
+ "score": 1.0,
+ "content": "formation. In this experiment, we have explored other variations, such as assigning",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 452,
+ 189,
+ 477,
+ 199
+ ],
+ "score": 0.88,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 187,
+ 490,
+ 201
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 189,
+ 504,
+ 198
+ ],
+ "score": 0.83,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 199,
+ 506,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 436,
+ 210
+ ],
+ "score": 1.0,
+ "content": "directly – ignoring previous memories, or doing a linear transformation based on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 436,
+ 200,
+ 461,
+ 209
+ ],
+ "score": 0.88,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 199,
+ 506,
+ 210
+ ],
+ "score": 1.0,
+ "content": "only. The",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 443,
+ 221
+ ],
+ "score": 1.0,
+ "content": "results show that in fact such variant is only slightly worse than our main variant –",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 210,
+ 465,
+ 220
+ ],
+ "score": 0.73,
+ "content": "0 . 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 465,
+ 208,
+ 506,
+ 221
+ ],
+ "score": 1.0,
+ "content": ". We also",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 505,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 505,
+ 233
+ ],
+ "score": 1.0,
+ "content": "conducted an experiment in which we merge the new information with the previous memory just by",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 506,
+ 243
+ ],
+ "score": 1.0,
+ "content": "a having a gate that does a weighted average of them. The results show that this variant performs",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 243,
+ 339,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 339,
+ 253
+ ],
+ "score": 1.0,
+ "content": "equivalently to our standard linear-transformation variant.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 105,
+ 165,
+ 506,
+ 253
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 258,
+ 504,
+ 336
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 271
+ ],
+ "score": 1.0,
+ "content": "Writing Unit Additions. We have explored the impact of the writing unit variants described in sec-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 269,
+ 506,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 269,
+ 506,
+ 283
+ ],
+ "score": 1.0,
+ "content": "tion 3.2.3 – adding self-attention, gating mechanisms, or both, compared to our standard main model",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 506,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 406,
+ 294
+ ],
+ "score": 1.0,
+ "content": "that uses a linear transformation to merge the newly retrieved information",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 282,
+ 431,
+ 292
+ ],
+ "score": 0.89,
+ "content": "m _ { n e w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 432,
+ 280,
+ 506,
+ 294
+ ],
+ "score": 1.0,
+ "content": "with the previous",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 104,
+ 290,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 290,
+ 142,
+ 306
+ ],
+ "score": 1.0,
+ "content": "memory",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 294,
+ 155,
+ 303
+ ],
+ "score": 0.86,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 290,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": ". For the complete CLEVR dataset we can see that indeed both these variants are very",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 315
+ ],
+ "score": 1.0,
+ "content": "helpful in increasing the model performance. Compared to our standard MAC model that achieves",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 313,
+ 506,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 138,
+ 324
+ ],
+ "score": 0.87,
+ "content": "9 8 . 9 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 139,
+ 313,
+ 362,
+ 326
+ ],
+ "score": 1.0,
+ "content": "on the validation set, self-attention yields accuracy of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 362,
+ 314,
+ 394,
+ 325
+ ],
+ "score": 0.87,
+ "content": "9 9 . 2 3 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 394,
+ 313,
+ 453,
+ 326
+ ],
+ "score": 1.0,
+ "content": ", gating yields",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 454,
+ 314,
+ 486,
+ 324
+ ],
+ "score": 0.87,
+ "content": "9 9 . 3 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 313,
+ 506,
+ 326
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 228,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 192,
+ 337
+ ],
+ "score": 1.0,
+ "content": "adding both achieves",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 325,
+ 225,
+ 335
+ ],
+ "score": 0.87,
+ "content": "9 9 . 4 8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 225,
+ 325,
+ 228,
+ 337
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18,
+ "bbox_fs": [
+ 104,
+ 259,
+ 506,
+ 337
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 342,
+ 505,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 505,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 505,
+ 354
+ ],
+ "score": 1.0,
+ "content": "Output Unit. In our standard model, the final predictions made in the output unit are based on the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 353,
+ 183,
+ 365
+ ],
+ "score": 1.0,
+ "content": "final memory state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 354,
+ 198,
+ 365
+ ],
+ "score": 0.87,
+ "content": "m _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 199,
+ 353,
+ 336,
+ 365
+ ],
+ "score": 1.0,
+ "content": "as well as question representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 337,
+ 354,
+ 343,
+ 364
+ ],
+ "score": 0.77,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 344,
+ 353,
+ 505,
+ 365
+ ],
+ "score": 1.0,
+ "content": "(stands for the final hidden states of the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 376
+ ],
+ "score": 1.0,
+ "content": "backward and forwards passes of the LSTM). We have explored the contribution of basing the model",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 388
+ ],
+ "score": 1.0,
+ "content": "prediction on the latter, by testing the model performance when prediction is based on memory",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 225,
+ 398
+ ],
+ "score": 1.0,
+ "content": "alone, for the complete and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 385,
+ 245,
+ 396
+ ],
+ "score": 0.87,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 245,
+ 385,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": "datasets. We can see that in both settings basing the model’s",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 409
+ ],
+ "score": 1.0,
+ "content": "predictions on the question representation allows faster training and higher accuracies. Notable is",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 231,
+ 420
+ ],
+ "score": 1.0,
+ "content": "the gap in performance for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 407,
+ 251,
+ 418
+ ],
+ "score": 0.85,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 408,
+ 294,
+ 420
+ ],
+ "score": 1.0,
+ "content": "CLEVR -",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 294,
+ 407,
+ 321,
+ 418
+ ],
+ "score": 0.82,
+ "content": "1 9 . 8 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 408,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "increase by using the question representation",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 431
+ ],
+ "score": 1.0,
+ "content": "to make predictions. These results are very reasonable intuitively, since the model is structured such",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "that the memory holds only information that was retrieved from the image. Thus, questions that may",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 505,
+ 454
+ ],
+ "score": 1.0,
+ "content": "ask for instance on different aspects (such as color or shape) of the same object in the image may",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "result in the same memory content, which is thus does not directly contain enough information to",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 462,
+ 204,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 204,
+ 474
+ ],
+ "score": 1.0,
+ "content": "respond such questions.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 342,
+ 506,
+ 474
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 479,
+ 505,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "Position. In our standard model, similarly to the practice of competing models (Santoro et al., 2017;",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "Perez et al., 2017; Hu et al., 2017), we have concatenated positional information to each region of the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "image, in order to increase the model capability to perform spatial reasoning. We have explored both",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 237,
+ 525
+ ],
+ "score": 1.0,
+ "content": "simple linear maps at a constant",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 238,
+ 512,
+ 266,
+ 524
+ ],
+ "score": 0.9,
+ "content": "[ - 1 , 1 ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 267,
+ 511,
+ 505,
+ 525
+ ],
+ "score": 1.0,
+ "content": "as well as more complex positional encoding suggested by",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 522,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 522,
+ 429,
+ 534
+ ],
+ "score": 1.0,
+ "content": "(Vaswani et al., 2017). However, the results for both the standard dataset and the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 523,
+ 449,
+ 533
+ ],
+ "score": 0.87,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 449,
+ 522,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "version show",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 104,
+ 533,
+ 506,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 533,
+ 506,
+ 547
+ ],
+ "score": 1.0,
+ "content": "a very negligible improvement at best when adding positional encoding information, demonstrating",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 546,
+ 424,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 546,
+ 424,
+ 557
+ ],
+ "score": 1.0,
+ "content": "the capability of MAC to perform spatial reasoning without data augmentation.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 104,
+ 478,
+ 506,
+ 557
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 561,
+ 505,
+ 650
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "Gate Bias Initialization. For our model variant with gating mechanism (described in section 3.2.3)",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 382,
+ 585
+ ],
+ "score": 1.0,
+ "content": "we have tested the effect of setting different values for the gate bias -",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 382,
+ 573,
+ 406,
+ 584
+ ],
+ "score": 0.85,
+ "content": "- 1 , 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 406,
+ 573,
+ 448,
+ 585
+ ],
+ "score": 1.0,
+ "content": "and 1. for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 448,
+ 573,
+ 462,
+ 583
+ ],
+ "score": 0.66,
+ "content": "- 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 573,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "the model",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 596
+ ],
+ "score": 1.0,
+ "content": "is initialized to biased for keeping the previous memory value whereas for 1 it will be biased for",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 595,
+ 505,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 505,
+ 607
+ ],
+ "score": 1.0,
+ "content": "using the new memory instead. We can see that for the complete dataset setting the bias to 1 is",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "optimal – apparently since the model has enough data to learn to apply each cell effectively. In",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 195,
+ 629
+ ],
+ "score": 1.0,
+ "content": "contrast, for the small",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 617,
+ 215,
+ 627
+ ],
+ "score": 0.85,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 616,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "CLEVR data, setting the bias to 0 shows better performance, biasing the",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "model to using less cells overall which results ultimately in a theoretically-simpler model that can",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 223,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 223,
+ 651
+ ],
+ "score": 1.0,
+ "content": "fit less data more effectively.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 44.5,
+ "bbox_fs": [
+ 105,
+ 561,
+ 506,
+ 651
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 674,
+ 213,
+ 685
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 674,
+ 215,
+ 686
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 674,
+ 215,
+ 686
+ ],
+ "score": 1.0,
+ "content": "4.4 INTERPRETABILITY",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 49
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "We have looked into attention maps over the image and question that the model produces during its",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "computation and provide a few examples in figure 4.4. The first example shows us how the model",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "parses the question in steps, first focusing on the main entity that the question is about, then on",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 51,
+ "bbox_fs": [
+ 105,
+ 698,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "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": "relation of this entity to the “brown matte thing” which is then located in the image. Finally, the",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 460,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 460,
+ 106
+ ],
+ "score": 1.0,
+ "content": "model correctly focuses on the small brown cube and predicts the right answer – brown.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 110,
+ 504,
+ 143
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 109,
+ 506,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 109,
+ 506,
+ 123
+ ],
+ "score": 1.0,
+ "content": "The second example shows a model with 4 cells instead of 6, that similarly parse the question in",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 120,
+ 506,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 120,
+ 506,
+ 135
+ ],
+ "score": 1.0,
+ "content": "iterations and focuses on the relevant objects at each step, though we can see that the reasoning",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 132,
+ 405,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 132,
+ 405,
+ 144
+ ],
+ "score": 1.0,
+ "content": "process looks somewhat different when the MAC network has fewer cells.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 149,
+ 505,
+ 182
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "The last example shows how how the model handles counting and OR operations. It starts from",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 175
+ ],
+ "score": 1.0,
+ "content": "identifying the task - computing a number, and then red objects as well as the cylinder, one at a time,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 171,
+ 354,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 354,
+ 183
+ ],
+ "score": 1.0,
+ "content": "allowing it ultimately to respond correctly, with the answer 2.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 200,
+ 217,
+ 413,
+ 583
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 200,
+ 217,
+ 413,
+ 583
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 200,
+ 217,
+ 413,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 200,
+ 217,
+ 413,
+ 583
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "e280158a489b95ddeabb4461dda86460910408860df91a493fe23a53648a08b4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 200,
+ 217,
+ 413,
+ 230.55555555555554
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 200,
+ 230.55555555555554,
+ 413,
+ 244.1111111111111
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 200,
+ 244.1111111111111,
+ 413,
+ 257.66666666666663
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 200,
+ 257.66666666666663,
+ 413,
+ 271.2222222222222
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 200,
+ 271.2222222222222,
+ 413,
+ 284.7777777777777
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 200,
+ 284.7777777777777,
+ 413,
+ 298.33333333333326
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 200,
+ 298.33333333333326,
+ 413,
+ 311.8888888888888
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 200,
+ 311.8888888888888,
+ 413,
+ 325.44444444444434
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 200,
+ 325.44444444444434,
+ 413,
+ 338.9999999999999
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 200,
+ 338.9999999999999,
+ 413,
+ 352.55555555555543
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 200,
+ 352.55555555555543,
+ 413,
+ 366.111111111111
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 200,
+ 366.111111111111,
+ 413,
+ 379.6666666666665
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 200,
+ 379.6666666666665,
+ 413,
+ 393.22222222222206
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 200,
+ 393.22222222222206,
+ 413,
+ 406.7777777777776
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 200,
+ 406.7777777777776,
+ 413,
+ 420.33333333333314
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 200,
+ 420.33333333333314,
+ 413,
+ 433.8888888888887
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 200,
+ 433.8888888888887,
+ 413,
+ 447.44444444444423
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 200,
+ 447.44444444444423,
+ 413,
+ 460.9999999999998
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 200,
+ 460.9999999999998,
+ 413,
+ 474.5555555555553
+ ],
+ "spans": [],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 200,
+ 474.5555555555553,
+ 413,
+ 488.11111111111086
+ ],
+ "spans": [],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 200,
+ 488.11111111111086,
+ 413,
+ 501.6666666666664
+ ],
+ "spans": [],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 200,
+ 501.6666666666664,
+ 413,
+ 515.222222222222
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 200,
+ 515.222222222222,
+ 413,
+ 528.7777777777775
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 200,
+ 528.7777777777775,
+ 413,
+ 542.333333333333
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 200,
+ 542.333333333333,
+ 413,
+ 555.8888888888886
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 200,
+ 555.8888888888886,
+ 413,
+ 569.4444444444441
+ ],
+ "spans": [],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 200,
+ 569.4444444444441,
+ 413,
+ 582.9999999999997
+ ],
+ "spans": [],
+ "index": 34
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 617,
+ 504,
+ 630
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 631
+ ],
+ "score": 1.0,
+ "content": "Figure 8: Three examples of attention maps for the image and the question. See text for full details.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 652,
+ 195,
+ 665
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 651,
+ 197,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 651,
+ 197,
+ 668
+ ],
+ "score": 1.0,
+ "content": "5 CONCLUSION",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 677,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "We have given a first demonstration of how a sequence of Memory, Attention and Control (MAC)",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "cells combined into a Compositional Attention Network provides a very effective tool for neural rea-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "soning. In future work, we wish to explore this promising architecture for other tasks and domains,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 445,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 445,
+ 724
+ ],
+ "score": 1.0,
+ "content": "including real-world VQA, machine comprehension and textual question answering.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 38.5
+ }
+ ],
+ "page_idx": 13,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "14",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "relation of this entity to the “brown matte thing” which is then located in the image. Finally, the",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 460,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 460,
+ 106
+ ],
+ "score": 1.0,
+ "content": "model correctly focuses on the small brown cube and predicts the right answer – brown.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 505,
+ 106
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 110,
+ 504,
+ 143
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 109,
+ 506,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 109,
+ 506,
+ 123
+ ],
+ "score": 1.0,
+ "content": "The second example shows a model with 4 cells instead of 6, that similarly parse the question in",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 120,
+ 506,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 120,
+ 506,
+ 135
+ ],
+ "score": 1.0,
+ "content": "iterations and focuses on the relevant objects at each step, though we can see that the reasoning",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 132,
+ 405,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 132,
+ 405,
+ 144
+ ],
+ "score": 1.0,
+ "content": "process looks somewhat different when the MAC network has fewer cells.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 105,
+ 109,
+ 506,
+ 144
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 149,
+ 505,
+ 182
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "The last example shows how how the model handles counting and OR operations. It starts from",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 175
+ ],
+ "score": 1.0,
+ "content": "identifying the task - computing a number, and then red objects as well as the cylinder, one at a time,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 171,
+ 354,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 354,
+ 183
+ ],
+ "score": 1.0,
+ "content": "allowing it ultimately to respond correctly, with the answer 2.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 148,
+ 506,
+ 183
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 200,
+ 217,
+ 413,
+ 583
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 200,
+ 217,
+ 413,
+ 583
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 200,
+ 217,
+ 413,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 200,
+ 217,
+ 413,
+ 583
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "e280158a489b95ddeabb4461dda86460910408860df91a493fe23a53648a08b4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 200,
+ 217,
+ 413,
+ 230.55555555555554
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 200,
+ 230.55555555555554,
+ 413,
+ 244.1111111111111
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 200,
+ 244.1111111111111,
+ 413,
+ 257.66666666666663
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 200,
+ 257.66666666666663,
+ 413,
+ 271.2222222222222
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 200,
+ 271.2222222222222,
+ 413,
+ 284.7777777777777
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 200,
+ 284.7777777777777,
+ 413,
+ 298.33333333333326
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 200,
+ 298.33333333333326,
+ 413,
+ 311.8888888888888
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 200,
+ 311.8888888888888,
+ 413,
+ 325.44444444444434
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 200,
+ 325.44444444444434,
+ 413,
+ 338.9999999999999
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 200,
+ 338.9999999999999,
+ 413,
+ 352.55555555555543
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 200,
+ 352.55555555555543,
+ 413,
+ 366.111111111111
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 200,
+ 366.111111111111,
+ 413,
+ 379.6666666666665
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 200,
+ 379.6666666666665,
+ 413,
+ 393.22222222222206
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 200,
+ 393.22222222222206,
+ 413,
+ 406.7777777777776
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 200,
+ 406.7777777777776,
+ 413,
+ 420.33333333333314
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 200,
+ 420.33333333333314,
+ 413,
+ 433.8888888888887
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 200,
+ 433.8888888888887,
+ 413,
+ 447.44444444444423
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 200,
+ 447.44444444444423,
+ 413,
+ 460.9999999999998
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 200,
+ 460.9999999999998,
+ 413,
+ 474.5555555555553
+ ],
+ "spans": [],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 200,
+ 474.5555555555553,
+ 413,
+ 488.11111111111086
+ ],
+ "spans": [],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 200,
+ 488.11111111111086,
+ 413,
+ 501.6666666666664
+ ],
+ "spans": [],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 200,
+ 501.6666666666664,
+ 413,
+ 515.222222222222
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 200,
+ 515.222222222222,
+ 413,
+ 528.7777777777775
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 200,
+ 528.7777777777775,
+ 413,
+ 542.333333333333
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 200,
+ 542.333333333333,
+ 413,
+ 555.8888888888886
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 200,
+ 555.8888888888886,
+ 413,
+ 569.4444444444441
+ ],
+ "spans": [],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 200,
+ 569.4444444444441,
+ 413,
+ 582.9999999999997
+ ],
+ "spans": [],
+ "index": 34
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 617,
+ 504,
+ 630
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 631
+ ],
+ "score": 1.0,
+ "content": "Figure 8: Three examples of attention maps for the image and the question. See text for full details.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 616,
+ 505,
+ 631
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 652,
+ 195,
+ 665
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 651,
+ 197,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 651,
+ 197,
+ 668
+ ],
+ "score": 1.0,
+ "content": "5 CONCLUSION",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 677,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "We have given a first demonstration of how a sequence of Memory, Attention and Control (MAC)",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "cells combined into a Compositional Attention Network provides a very effective tool for neural rea-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "soning. In future work, we wish to explore this promising architecture for other tasks and domains,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 445,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 445,
+ 724
+ ],
+ "score": 1.0,
+ "content": "including real-world VQA, machine comprehension and textual question answering.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 38.5,
+ "bbox_fs": [
+ 105,
+ 677,
+ 505,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 81,
+ 176,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 176,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 176,
+ 94
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 100,
+ 504,
+ 123
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 113
+ ],
+ "score": 1.0,
+ "content": "Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 116,
+ 111,
+ 356,
+ 122
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 111,
+ 356,
+ 122
+ ],
+ "score": 1.0,
+ "content": "answering models. arXiv preprint arXiv:1606.07356, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 128,
+ 504,
+ 163
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 505,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 505,
+ 141
+ ],
+ "score": 1.0,
+ "content": "Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 139,
+ 506,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 139,
+ 506,
+ 154
+ ],
+ "score": 1.0,
+ "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 39–48,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 149,
+ 147,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 149,
+ 147,
+ 164
+ ],
+ "score": 1.0,
+ "content": "2016a.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 169,
+ 504,
+ 193
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 168,
+ 505,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 168,
+ 505,
+ 183
+ ],
+ "score": 1.0,
+ "content": "Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 115,
+ 181,
+ 420,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 181,
+ 420,
+ 193
+ ],
+ "score": 1.0,
+ "content": "networks for question answering. arXiv preprint arXiv:1601.01705, 2016b.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 199,
+ 504,
+ 233
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 504,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 504,
+ 212
+ ],
+ "score": 1.0,
+ "content": "Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zit-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 116,
+ 210,
+ 505,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 210,
+ 505,
+ 223
+ ],
+ "score": 1.0,
+ "content": "nick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 117,
+ 221,
+ 339,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 221,
+ 339,
+ 234
+ ],
+ "score": 1.0,
+ "content": "Conference on Computer Vision, pp. 2425–2433, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 239,
+ 504,
+ 262
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 253
+ ],
+ "score": 1.0,
+ "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 250,
+ 219,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 250,
+ 219,
+ 262
+ ],
+ "score": 1.0,
+ "content": "arXiv:1607.06450, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 268,
+ 504,
+ 291
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 504,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 504,
+ 282
+ ],
+ "score": 1.0,
+ "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 280,
+ 398,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 280,
+ 398,
+ 292
+ ],
+ "score": 1.0,
+ "content": "learning to align and translate. arXiv preprint arXiv:1409.0473, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 298,
+ 504,
+ 321
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 505,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 505,
+ 311
+ ],
+ "score": 1.0,
+ "content": "Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network ´",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 308,
+ 451,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 308,
+ 451,
+ 322
+ ],
+ "score": 1.0,
+ "content": "learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 327,
+ 504,
+ 361
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 505,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 505,
+ 341
+ ],
+ "score": 1.0,
+ "content": "Alex Graves, Navdeep Jaitly, and Abdel-rahman Mohamed. Hybrid speech recognition with deep",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 116,
+ 339,
+ 505,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 339,
+ 505,
+ 351
+ ],
+ "score": 1.0,
+ "content": "bidirectional lstm. In Automatic Speech Recognition and Understanding (ASRU), 2013 IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 117,
+ 349,
+ 281,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 349,
+ 281,
+ 363
+ ],
+ "score": 1.0,
+ "content": "Workshop on, pp. 273–278. IEEE, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 367,
+ 505,
+ 391
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 381
+ ],
+ "score": 1.0,
+ "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 379,
+ 214,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 379,
+ 214,
+ 390
+ ],
+ "score": 1.0,
+ "content": "arXiv:1410.5401, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 397,
+ 505,
+ 442
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 397,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 397,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 115,
+ 408,
+ 505,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 408,
+ 505,
+ 421
+ ],
+ "score": 1.0,
+ "content": "Barwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 419,
+ 506,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 419,
+ 506,
+ 432
+ ],
+ "score": 1.0,
+ "content": "et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 431,
+ 210,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 431,
+ 210,
+ 443
+ ],
+ "score": 1.0,
+ "content": "(7626):471–476, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 448,
+ 504,
+ 471
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "Akshay Kumar Gupta. Survey of visual question answering: Datasets and techniques. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 460,
+ 220,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 460,
+ 220,
+ 470
+ ],
+ "score": 1.0,
+ "content": "arXiv:1705.03865, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 478,
+ 503,
+ 512
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 505,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 505,
+ 493
+ ],
+ "score": 1.0,
+ "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 488,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 488,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 116,
+ 500,
+ 181,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 500,
+ 181,
+ 511
+ ],
+ "score": 1.0,
+ "content": "770–778, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 103,
+ 518,
+ 504,
+ 541
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 532
+ ],
+ "score": 1.0,
+ "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 116,
+ 529,
+ 192,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 529,
+ 192,
+ 541
+ ],
+ "score": 1.0,
+ "content": "1735–1780, 1997.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 504,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 504,
+ 560
+ ],
+ "score": 1.0,
+ "content": "Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learn-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 115,
+ 559,
+ 505,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 559,
+ 505,
+ 572
+ ],
+ "score": 1.0,
+ "content": "ing to reason: End-to-end module networks for visual question answering. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 569,
+ 220,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 569,
+ 220,
+ 581
+ ],
+ "score": 1.0,
+ "content": "arXiv:1704.05526, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 588,
+ 504,
+ 622
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 602
+ ],
+ "score": 1.0,
+ "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 599,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 599,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 114,
+ 609,
+ 142,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 609,
+ 142,
+ 622
+ ],
+ "score": 1.0,
+ "content": "2015.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 628,
+ 504,
+ 663
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 639,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 639,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 651,
+ 324,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 651,
+ 324,
+ 663
+ ],
+ "score": 1.0,
+ "content": "reasoning. arXiv preprint arXiv:1612.06890, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 669,
+ 503,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 669,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 669,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Judy Hoffman, Li Fei-Fei, C Lawrence",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 679,
+ 505,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 679,
+ 505,
+ 693
+ ],
+ "score": 1.0,
+ "content": "Zitnick, and Ross Girshick. Inferring and executing programs for visual reasoning. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 116,
+ 691,
+ 220,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 691,
+ 220,
+ 702
+ ],
+ "score": 1.0,
+ "content": "arXiv:1705.03633, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 709,
+ 503,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 214,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 214,
+ 731
+ ],
+ "score": 1.0,
+ "content": "arXiv:1412.6980, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45.5
+ }
+ ],
+ "page_idx": 14,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "15",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 81,
+ 176,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 176,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 176,
+ 94
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 100,
+ 504,
+ 123
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 113
+ ],
+ "score": 1.0,
+ "content": "Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 116,
+ 111,
+ 356,
+ 122
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 111,
+ 356,
+ 122
+ ],
+ "score": 1.0,
+ "content": "answering models. arXiv preprint arXiv:1606.07356, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1.5,
+ "bbox_fs": [
+ 106,
+ 100,
+ 505,
+ 122
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 128,
+ 504,
+ 163
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 505,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 505,
+ 141
+ ],
+ "score": 1.0,
+ "content": "Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 139,
+ 506,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 139,
+ 506,
+ 154
+ ],
+ "score": 1.0,
+ "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 39–48,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 149,
+ 147,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 149,
+ 147,
+ 164
+ ],
+ "score": 1.0,
+ "content": "2016a.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 106,
+ 129,
+ 506,
+ 164
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 169,
+ 504,
+ 193
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 168,
+ 505,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 168,
+ 505,
+ 183
+ ],
+ "score": 1.0,
+ "content": "Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 115,
+ 181,
+ 420,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 181,
+ 420,
+ 193
+ ],
+ "score": 1.0,
+ "content": "networks for question answering. arXiv preprint arXiv:1601.01705, 2016b.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 105,
+ 168,
+ 505,
+ 193
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 199,
+ 504,
+ 233
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 504,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 504,
+ 212
+ ],
+ "score": 1.0,
+ "content": "Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zit-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 116,
+ 210,
+ 505,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 210,
+ 505,
+ 223
+ ],
+ "score": 1.0,
+ "content": "nick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 117,
+ 221,
+ 339,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 221,
+ 339,
+ 234
+ ],
+ "score": 1.0,
+ "content": "Conference on Computer Vision, pp. 2425–2433, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9,
+ "bbox_fs": [
+ 106,
+ 199,
+ 505,
+ 234
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 239,
+ 504,
+ 262
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 253
+ ],
+ "score": 1.0,
+ "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 250,
+ 219,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 250,
+ 219,
+ 262
+ ],
+ "score": 1.0,
+ "content": "arXiv:1607.06450, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 238,
+ 505,
+ 262
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 268,
+ 504,
+ 291
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 504,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 504,
+ 282
+ ],
+ "score": 1.0,
+ "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 280,
+ 398,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 280,
+ 398,
+ 292
+ ],
+ "score": 1.0,
+ "content": "learning to align and translate. arXiv preprint arXiv:1409.0473, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 106,
+ 269,
+ 504,
+ 292
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 298,
+ 504,
+ 321
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 505,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 505,
+ 311
+ ],
+ "score": 1.0,
+ "content": "Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network ´",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 308,
+ 451,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 308,
+ 451,
+ 322
+ ],
+ "score": 1.0,
+ "content": "learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 106,
+ 298,
+ 505,
+ 322
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 327,
+ 504,
+ 361
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 505,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 505,
+ 341
+ ],
+ "score": 1.0,
+ "content": "Alex Graves, Navdeep Jaitly, and Abdel-rahman Mohamed. Hybrid speech recognition with deep",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 116,
+ 339,
+ 505,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 339,
+ 505,
+ 351
+ ],
+ "score": 1.0,
+ "content": "bidirectional lstm. In Automatic Speech Recognition and Understanding (ASRU), 2013 IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 117,
+ 349,
+ 281,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 349,
+ 281,
+ 363
+ ],
+ "score": 1.0,
+ "content": "Workshop on, pp. 273–278. IEEE, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18,
+ "bbox_fs": [
+ 105,
+ 326,
+ 505,
+ 363
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 367,
+ 505,
+ 391
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 381
+ ],
+ "score": 1.0,
+ "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 379,
+ 214,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 379,
+ 214,
+ 390
+ ],
+ "score": 1.0,
+ "content": "arXiv:1410.5401, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 105,
+ 366,
+ 506,
+ 390
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 397,
+ 505,
+ 442
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 397,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 397,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 115,
+ 408,
+ 505,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 408,
+ 505,
+ 421
+ ],
+ "score": 1.0,
+ "content": "Barwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 419,
+ 506,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 419,
+ 506,
+ 432
+ ],
+ "score": 1.0,
+ "content": "et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 431,
+ 210,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 431,
+ 210,
+ 443
+ ],
+ "score": 1.0,
+ "content": "(7626):471–476, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 106,
+ 397,
+ 506,
+ 443
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 448,
+ 504,
+ 471
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "Akshay Kumar Gupta. Survey of visual question answering: Datasets and techniques. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 460,
+ 220,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 460,
+ 220,
+ 470
+ ],
+ "score": 1.0,
+ "content": "arXiv:1705.03865, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 448,
+ 505,
+ 470
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 478,
+ 503,
+ 512
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 505,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 505,
+ 493
+ ],
+ "score": 1.0,
+ "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 488,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 488,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 116,
+ 500,
+ 181,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 500,
+ 181,
+ 511
+ ],
+ "score": 1.0,
+ "content": "770–778, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 29,
+ "bbox_fs": [
+ 105,
+ 476,
+ 505,
+ 511
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 103,
+ 518,
+ 504,
+ 541
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 532
+ ],
+ "score": 1.0,
+ "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 116,
+ 529,
+ 192,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 529,
+ 192,
+ 541
+ ],
+ "score": 1.0,
+ "content": "1735–1780, 1997.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31.5,
+ "bbox_fs": [
+ 105,
+ 518,
+ 505,
+ 541
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 504,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 504,
+ 560
+ ],
+ "score": 1.0,
+ "content": "Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learn-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 115,
+ 559,
+ 505,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 559,
+ 505,
+ 572
+ ],
+ "score": 1.0,
+ "content": "ing to reason: End-to-end module networks for visual question answering. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 569,
+ 220,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 569,
+ 220,
+ 581
+ ],
+ "score": 1.0,
+ "content": "arXiv:1704.05526, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 106,
+ 547,
+ 505,
+ 581
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 588,
+ 504,
+ 622
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 602
+ ],
+ "score": 1.0,
+ "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 599,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 599,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 114,
+ 609,
+ 142,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 609,
+ 142,
+ 622
+ ],
+ "score": 1.0,
+ "content": "2015.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 105,
+ 587,
+ 505,
+ 622
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 628,
+ 504,
+ 663
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 639,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 639,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 651,
+ 324,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 651,
+ 324,
+ 663
+ ],
+ "score": 1.0,
+ "content": "reasoning. arXiv preprint arXiv:1612.06890, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 106,
+ 628,
+ 505,
+ 663
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 669,
+ 503,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 669,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 669,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Judy Hoffman, Li Fei-Fei, C Lawrence",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 679,
+ 505,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 679,
+ 505,
+ 693
+ ],
+ "score": 1.0,
+ "content": "Zitnick, and Ross Girshick. Inferring and executing programs for visual reasoning. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 116,
+ 691,
+ 220,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 691,
+ 220,
+ 702
+ ],
+ "score": 1.0,
+ "content": "arXiv:1705.03633, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43,
+ "bbox_fs": [
+ 106,
+ 669,
+ 505,
+ 702
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 709,
+ 503,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 214,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 214,
+ 731
+ ],
+ "score": 1.0,
+ "content": "arXiv:1412.6980, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45.5,
+ "bbox_fs": [
+ 105,
+ 708,
+ 505,
+ 731
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 504,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 114,
+ 103,
+ 143,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 103,
+ 143,
+ 116
+ ],
+ "score": 1.0,
+ "content": "2012.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 123,
+ 505,
+ 167
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 123,
+ 506,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 123,
+ 506,
+ 136
+ ],
+ "score": 1.0,
+ "content": "Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 134,
+ 506,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 134,
+ 506,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 114,
+ 145,
+ 506,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 145,
+ 506,
+ 158
+ ],
+ "score": 1.0,
+ "content": "natural language processing. In International Conference on Machine Learning, pp. 1378–1387,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 114,
+ 155,
+ 143,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 155,
+ 143,
+ 169
+ ],
+ "score": 1.0,
+ "content": "2016.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 174,
+ 503,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 174,
+ 505,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 174,
+ 505,
+ 187
+ ],
+ "score": 1.0,
+ "content": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 116,
+ 186,
+ 417,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 186,
+ 417,
+ 198
+ ],
+ "score": 1.0,
+ "content": "document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 204,
+ 506,
+ 238
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 217
+ ],
+ "score": 1.0,
+ "content": "Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 216,
+ 506,
+ 229
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 216,
+ 506,
+ 229
+ ],
+ "score": 1.0,
+ "content": "for visual question answering. In Advances In Neural Information Processing Systems, pp. 289–",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 226,
+ 162,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 226,
+ 162,
+ 239
+ ],
+ "score": 1.0,
+ "content": "297, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 245,
+ 506,
+ 279
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 258
+ ],
+ "score": 1.0,
+ "content": "Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Ja-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 256,
+ 505,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 256,
+ 505,
+ 269
+ ],
+ "score": 1.0,
+ "content": "son Weston. Key-value memory networks for directly reading documents. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 116,
+ 268,
+ 219,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 268,
+ 219,
+ 279
+ ],
+ "score": 1.0,
+ "content": "arXiv:1606.03126, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 286,
+ 506,
+ 320
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 299
+ ],
+ "score": 1.0,
+ "content": "Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global vectors for word",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 297,
+ 506,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 297,
+ 506,
+ 311
+ ],
+ "score": 1.0,
+ "content": "representation. In Proceedings of the 2014 conference on empirical methods in natural language",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 309,
+ 296,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 309,
+ 296,
+ 321
+ ],
+ "score": 1.0,
+ "content": "processing (EMNLP), pp. 1532–1543, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 327,
+ 505,
+ 351
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 327,
+ 505,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 327,
+ 505,
+ 340
+ ],
+ "score": 1.0,
+ "content": "Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 339,
+ 457,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 339,
+ 457,
+ 351
+ ],
+ "score": 1.0,
+ "content": "reasoning with a general conditioning layer. arXiv preprint arXiv:1709.07871, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 357,
+ 504,
+ 391
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 357,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 357,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "Adam Santoro, David Raposo, David GT Barrett, Mateusz Malinowski, Razvan Pascanu, Peter",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 368,
+ 505,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 368,
+ 505,
+ 381
+ ],
+ "score": 1.0,
+ "content": "Battaglia, and Timothy Lillicrap. A simple neural network module for relational reasoning. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 380,
+ 254,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 380,
+ 254,
+ 391
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1706.01427, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 397,
+ 505,
+ 421
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 411
+ ],
+ "score": 1.0,
+ "content": "Rupesh Kumar Srivastava, Klaus Greff, and Jurgen Schmidhuber. Highway networks. ¨ arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 409,
+ 219,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 409,
+ 219,
+ 420
+ ],
+ "score": 1.0,
+ "content": "arXiv:1505.00387, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 428,
+ 504,
+ 451
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "Bob L Sturm. A simple method to determine if a music information retrieval system is a horse.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 117,
+ 439,
+ 356,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 439,
+ 356,
+ 450
+ ],
+ "score": 1.0,
+ "content": "IEEE Transactions on Multimedia, 16(6):1636–1644, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 457,
+ 505,
+ 492
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 115,
+ 469,
+ 505,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 469,
+ 505,
+ 482
+ ],
+ "score": 1.0,
+ "content": "Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 479,
+ 144,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 479,
+ 144,
+ 492
+ ],
+ "score": 1.0,
+ "content": "2017.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 498,
+ 504,
+ 532
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 498,
+ 506,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 498,
+ 506,
+ 512
+ ],
+ "score": 1.0,
+ "content": "Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 510,
+ 505,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 510,
+ 505,
+ 522
+ ],
+ "score": 1.0,
+ "content": "textual question answering. In International Conference on Machine Learning, pp. 2397–2406,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 114,
+ 519,
+ 143,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 519,
+ 143,
+ 533
+ ],
+ "score": 1.0,
+ "content": "2016.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 539,
+ 505,
+ 573
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 540,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 540,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 116,
+ 551,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 551,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "for image question answering. In Proceedings of the IEEE Conference on Computer Vision and",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 561,
+ 271,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 561,
+ 271,
+ 574
+ ],
+ "score": 1.0,
+ "content": "Pattern Recognition, pp. 21–29, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "page_idx": 15,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "16",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 504,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 114,
+ 103,
+ 143,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 103,
+ 143,
+ 116
+ ],
+ "score": 1.0,
+ "content": "2012.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 105,
+ 81,
+ 505,
+ 116
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 123,
+ 505,
+ 167
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 123,
+ 506,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 123,
+ 506,
+ 136
+ ],
+ "score": 1.0,
+ "content": "Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 134,
+ 506,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 134,
+ 506,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 114,
+ 145,
+ 506,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 145,
+ 506,
+ 158
+ ],
+ "score": 1.0,
+ "content": "natural language processing. In International Conference on Machine Learning, pp. 1378–1387,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 114,
+ 155,
+ 143,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 155,
+ 143,
+ 169
+ ],
+ "score": 1.0,
+ "content": "2016.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 105,
+ 123,
+ 506,
+ 169
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 174,
+ 503,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 174,
+ 505,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 174,
+ 505,
+ 187
+ ],
+ "score": 1.0,
+ "content": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 116,
+ 186,
+ 417,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 186,
+ 417,
+ 198
+ ],
+ "score": 1.0,
+ "content": "document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 106,
+ 174,
+ 505,
+ 198
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 204,
+ 506,
+ 238
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 217
+ ],
+ "score": 1.0,
+ "content": "Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 216,
+ 506,
+ 229
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 216,
+ 506,
+ 229
+ ],
+ "score": 1.0,
+ "content": "for visual question answering. In Advances In Neural Information Processing Systems, pp. 289–",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 226,
+ 162,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 226,
+ 162,
+ 239
+ ],
+ "score": 1.0,
+ "content": "297, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 105,
+ 204,
+ 506,
+ 239
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 245,
+ 506,
+ 279
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 258
+ ],
+ "score": 1.0,
+ "content": "Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Ja-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 256,
+ 505,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 256,
+ 505,
+ 269
+ ],
+ "score": 1.0,
+ "content": "son Weston. Key-value memory networks for directly reading documents. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 116,
+ 268,
+ 219,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 268,
+ 219,
+ 279
+ ],
+ "score": 1.0,
+ "content": "arXiv:1606.03126, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 106,
+ 245,
+ 505,
+ 279
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 286,
+ 506,
+ 320
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 299
+ ],
+ "score": 1.0,
+ "content": "Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global vectors for word",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 297,
+ 506,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 297,
+ 506,
+ 311
+ ],
+ "score": 1.0,
+ "content": "representation. In Proceedings of the 2014 conference on empirical methods in natural language",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 309,
+ 296,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 309,
+ 296,
+ 321
+ ],
+ "score": 1.0,
+ "content": "processing (EMNLP), pp. 1532–1543, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 105,
+ 286,
+ 506,
+ 321
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 327,
+ 505,
+ 351
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 327,
+ 505,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 327,
+ 505,
+ 340
+ ],
+ "score": 1.0,
+ "content": "Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 339,
+ 457,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 339,
+ 457,
+ 351
+ ],
+ "score": 1.0,
+ "content": "reasoning with a general conditioning layer. arXiv preprint arXiv:1709.07871, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 106,
+ 327,
+ 505,
+ 351
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 357,
+ 504,
+ 391
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 357,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 357,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "Adam Santoro, David Raposo, David GT Barrett, Mateusz Malinowski, Razvan Pascanu, Peter",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 368,
+ 505,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 368,
+ 505,
+ 381
+ ],
+ "score": 1.0,
+ "content": "Battaglia, and Timothy Lillicrap. A simple neural network module for relational reasoning. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 380,
+ 254,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 380,
+ 254,
+ 391
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1706.01427, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 106,
+ 357,
+ 505,
+ 391
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 397,
+ 505,
+ 421
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 505,
+ 411
+ ],
+ "score": 1.0,
+ "content": "Rupesh Kumar Srivastava, Klaus Greff, and Jurgen Schmidhuber. Highway networks. ¨ arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 409,
+ 219,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 409,
+ 219,
+ 420
+ ],
+ "score": 1.0,
+ "content": "arXiv:1505.00387, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 105,
+ 397,
+ 505,
+ 420
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 428,
+ 504,
+ 451
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "Bob L Sturm. A simple method to determine if a music information retrieval system is a horse.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 117,
+ 439,
+ 356,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 439,
+ 356,
+ 450
+ ],
+ "score": 1.0,
+ "content": "IEEE Transactions on Multimedia, 16(6):1636–1644, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 105,
+ 426,
+ 505,
+ 450
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 457,
+ 505,
+ 492
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 115,
+ 469,
+ 505,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 469,
+ 505,
+ 482
+ ],
+ "score": 1.0,
+ "content": "Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 479,
+ 144,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 479,
+ 144,
+ 492
+ ],
+ "score": 1.0,
+ "content": "2017.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28,
+ "bbox_fs": [
+ 105,
+ 456,
+ 506,
+ 492
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 498,
+ 504,
+ 532
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 498,
+ 506,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 498,
+ 506,
+ 512
+ ],
+ "score": 1.0,
+ "content": "Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 510,
+ 505,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 510,
+ 505,
+ 522
+ ],
+ "score": 1.0,
+ "content": "textual question answering. In International Conference on Machine Learning, pp. 2397–2406,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 114,
+ 519,
+ 143,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 519,
+ 143,
+ 533
+ ],
+ "score": 1.0,
+ "content": "2016.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 106,
+ 498,
+ 506,
+ 533
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 539,
+ 505,
+ 573
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 540,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 540,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 116,
+ 551,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 551,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "for image question answering. In Proceedings of the IEEE Conference on Computer Vision and",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 561,
+ 271,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 561,
+ 271,
+ 574
+ ],
+ "score": 1.0,
+ "content": "Pattern Recognition, pp. 21–29, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 106,
+ 540,
+ 505,
+ 574
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 81,
+ 255,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 257,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 257,
+ 95
+ ],
+ "score": 1.0,
+ "content": "SUPPLEMENTARY MATERIAL",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 110,
+ 255,
+ 124
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 110,
+ 257,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 110,
+ 257,
+ 126
+ ],
+ "score": 1.0,
+ "content": "A DETAILS OF INPUT UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 138,
+ 505,
+ 183
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 505,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 505,
+ 151
+ ],
+ "score": 1.0,
+ "content": "The input unit processes the raw inputs given to the system into distributed vector representations.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 149,
+ 506,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 149,
+ 506,
+ 163
+ ],
+ "score": 1.0,
+ "content": "It receives a text question (or in general, a query), and an image (or in general, a Knowledge Base",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 174
+ ],
+ "score": 1.0,
+ "content": "(KB)) and processes each of them with a matching sub-unit. Here we provide details of the Query",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 172,
+ 279,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 172,
+ 279,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Unit and the Image Unit used in this work.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 200,
+ 219,
+ 212
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 220,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 220,
+ 214
+ ],
+ "score": 1.0,
+ "content": "A.0.1 THE QUERY UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 221,
+ 505,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 193,
+ 234
+ ],
+ "score": 1.0,
+ "content": "We encode a query of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 222,
+ 201,
+ 232
+ ],
+ "score": 0.83,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 221,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "words into a continuous representation using a bidirectional LSTM (Hochre-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "iter & Schmidhuber, 1997; Graves et al., 2013). Each word is associated with a word embedding",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 107,
+ 244,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 245,
+ 119,
+ 255
+ ],
+ "score": 0.85,
+ "content": "w _ { s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 244,
+ 151,
+ 257
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 151,
+ 244,
+ 201,
+ 255
+ ],
+ "score": 0.92,
+ "content": "s = 1 , . . . , S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 244,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": ". In our case, we use GloVE words embeddings (Pennington et al., 2014).",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 255,
+ 482,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 255,
+ 423,
+ 267
+ ],
+ "score": 1.0,
+ "content": "Then, these embeddings are processed by a bidirectional LST M of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 255,
+ 430,
+ 264
+ ],
+ "score": 0.77,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 430,
+ 255,
+ 482,
+ 267
+ ],
+ "score": 1.0,
+ "content": "that outputs:",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 133,
+ 279,
+ 505,
+ 333
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 279,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 279,
+ 237,
+ 293
+ ],
+ "score": 1.0,
+ "content": "• a matching sequence of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 238,
+ 281,
+ 244,
+ 290
+ ],
+ "score": 0.83,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 244,
+ 279,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "-dimensional output states, which we refer to as contextual words,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 143,
+ 291,
+ 200,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 143,
+ 291,
+ 200,
+ 303
+ ],
+ "score": 0.79,
+ "content": "[ c w _ { 1 } , . . . , c w _ { S } ]",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 132,
+ 309,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 309,
+ 142,
+ 322
+ ],
+ "score": 1.0,
+ "content": "•",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 311,
+ 149,
+ 320
+ ],
+ "score": 0.78,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 149,
+ 309,
+ 252,
+ 322
+ ],
+ "score": 1.0,
+ "content": "-dimensional hidden state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 309,
+ 314,
+ 322
+ ],
+ "score": 0.93,
+ "content": "q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 315,
+ 309,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": ", the concatenation of the hidden states from the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 321,
+ 444,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 321,
+ 312,
+ 334
+ ],
+ "score": 1.0,
+ "content": "backward and forward passes. We refer to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 312,
+ 324,
+ 318,
+ 333
+ ],
+ "score": 0.8,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 319,
+ 321,
+ 444,
+ 334
+ ],
+ "score": 1.0,
+ "content": "as the question representation.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 12.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 346,
+ 504,
+ 370
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 506,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 245,
+ 359
+ ],
+ "score": 1.0,
+ "content": "Intuitively, each contextual word",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 245,
+ 349,
+ 262,
+ 358
+ ],
+ "score": 0.87,
+ "content": "c w _ { s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 345,
+ 374,
+ 359
+ ],
+ "score": 1.0,
+ "content": "represents the meaning of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 346,
+ 389,
+ 357
+ ],
+ "score": 0.89,
+ "content": "s ^ { t h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 389,
+ 345,
+ 506,
+ 359
+ ],
+ "score": 1.0,
+ "content": "word, in the context of the",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 357,
+ 496,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 357,
+ 235,
+ 371
+ ],
+ "score": 1.0,
+ "content": "question, while the hidden state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 360,
+ 241,
+ 370
+ ],
+ "score": 0.8,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 241,
+ 357,
+ 496,
+ 371
+ ],
+ "score": 1.0,
+ "content": "represents the overall (compositional) meaning of the question.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 386,
+ 217,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 385,
+ 218,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 385,
+ 218,
+ 399
+ ],
+ "score": 1.0,
+ "content": "A.0.2 THE IMAGE UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 407,
+ 505,
+ 463
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 408,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 408,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "Given an image, and following prior work on CLEVR (?Santoro et al., 2017; Perez et al., 2017), we",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "extract conv4 features from ResNet101 (He et al., 2016) pretrained on ImageNet (Krizhevsky et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 377,
+ 442
+ ],
+ "score": 1.0,
+ "content": "2012) which we treat as a fixed initial representation of the image,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 377,
+ 432,
+ 384,
+ 439
+ ],
+ "score": 0.69,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 384,
+ 428,
+ 441,
+ 442
+ ],
+ "score": 1.0,
+ "content": "of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 441,
+ 430,
+ 477,
+ 441
+ ],
+ "score": 0.88,
+ "content": "H , W , C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 428,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 107,
+ 441,
+ 506,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 441,
+ 167,
+ 451
+ ],
+ "score": 0.9,
+ "content": "H = W = 1 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 441,
+ 390,
+ 452
+ ],
+ "score": 1.0,
+ "content": "are the height and width of the transformed image and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 391,
+ 441,
+ 435,
+ 451
+ ],
+ "score": 0.89,
+ "content": "C = 1 0 2 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 441,
+ 506,
+ 452
+ ],
+ "score": 1.0,
+ "content": "is the number of",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 396,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 199,
+ 466
+ ],
+ "score": 1.0,
+ "content": "channels. Each feature",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 453,
+ 220,
+ 464
+ ],
+ "score": 0.9,
+ "content": "x _ { h , w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 220,
+ 450,
+ 396,
+ 466
+ ],
+ "score": 1.0,
+ "content": "represents one region in the original image.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 468,
+ 505,
+ 535
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 506,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 506,
+ 481
+ ],
+ "score": 1.0,
+ "content": "Similar to prior work (Hu et al., 2017; Santoro et al., 2017; Perez et al., 2017), we would like to",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 480,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "allow our model to reason explicitly about spatial locations, as required by many of the questions in",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 490,
+ 506,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 490,
+ 506,
+ 503
+ ],
+ "score": 1.0,
+ "content": "CLEVR, and therefore we concatenate to this representation a spatial map that represents each of the",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 104,
+ 500,
+ 506,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 500,
+ 506,
+ 515
+ ],
+ "score": 1.0,
+ "content": "positions in the image. However, in contrast to prior work that uses a linear meshgrid feature map",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 506,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 169,
+ 525
+ ],
+ "score": 1.0,
+ "content": "with 2 features",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 170,
+ 513,
+ 177,
+ 522
+ ],
+ "score": 0.8,
+ "content": "h",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 511,
+ 196,
+ 525
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 514,
+ 205,
+ 523
+ ],
+ "score": 0.73,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 511,
+ 262,
+ 525
+ ],
+ "score": 1.0,
+ "content": "ranging from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 262,
+ 513,
+ 276,
+ 523
+ ],
+ "score": 0.79,
+ "content": "- 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 511,
+ 506,
+ 525
+ ],
+ "score": 1.0,
+ "content": "to 1, and to allow better representation of the positions,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 406,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 406,
+ 536
+ ],
+ "score": 1.0,
+ "content": "we use the positional encoding scheme proposed by Vaswani et al. (2017):",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 238,
+ 560,
+ 372,
+ 607
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 238,
+ 560,
+ 372,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 238,
+ 560,
+ 372,
+ 607
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { l } { p _ { ( h , 2 i ) } = \\sin { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\\\ { p _ { ( h , 2 i + 1 ) } = \\cos { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "23f67e68bd8fe769889cb65ef6b26ad32ea8c028817ac3af21e7f4e098381e73.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 29.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 238,
+ 560,
+ 372,
+ 583.5
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 238,
+ 583.5,
+ 372,
+ 607.0
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 621,
+ 504,
+ 644
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 620,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 620,
+ 178,
+ 633
+ ],
+ "score": 1.0,
+ "content": "And similarly for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 179,
+ 623,
+ 187,
+ 631
+ ],
+ "score": 0.77,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 620,
+ 218,
+ 633
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 623,
+ 230,
+ 632
+ ],
+ "score": 0.86,
+ "content": "p _ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 230,
+ 620,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "is a hyperparameter. Overall, the positional encoding of a feature at",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 449,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 141,
+ 645
+ ],
+ "score": 1.0,
+ "content": "position",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 632,
+ 167,
+ 644
+ ],
+ "score": 0.92,
+ "content": "( h , w )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 632,
+ 178,
+ 645
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 632,
+ 210,
+ 644
+ ],
+ "score": 0.92,
+ "content": "[ p _ { h } , p _ { w } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 211,
+ 632,
+ 411,
+ 645
+ ],
+ "score": 1.0,
+ "content": ", the concatenation of the positional encodings for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 411,
+ 632,
+ 419,
+ 642
+ ],
+ "score": 0.84,
+ "content": "h",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 632,
+ 436,
+ 645
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 437,
+ 633,
+ 445,
+ 642
+ ],
+ "score": 0.8,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 445,
+ 632,
+ 449,
+ 645
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 648,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "This positional encoding scheme allows better correspondence between the distance of 2 positions",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 107,
+ 659,
+ 505,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 660,
+ 131,
+ 672
+ ],
+ "score": 0.91,
+ "content": "( x , y )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 131,
+ 659,
+ 149,
+ 673
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 660,
+ 173,
+ 672
+ ],
+ "score": 0.92,
+ "content": "( x , y )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 174,
+ 659,
+ 483,
+ 673
+ ],
+ "score": 1.0,
+ "content": "in the image and a vector similarity of their positional encodings, even when",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 483,
+ 661,
+ 495,
+ 672
+ ],
+ "score": 0.86,
+ "content": "p _ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 659,
+ 505,
+ 673
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 670,
+ 172,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 172,
+ 684
+ ],
+ "score": 1.0,
+ "content": "larger than two.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 316,
+ 701
+ ],
+ "score": 1.0,
+ "content": "We then concatenate the obtained spatial map with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 691,
+ 322,
+ 698
+ ],
+ "score": 0.67,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 687,
+ 506,
+ 701
+ ],
+ "score": 1.0,
+ "content": ", receiving a spatially-aware image represen-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 132,
+ 711
+ ],
+ "score": 1.0,
+ "content": "tation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 700,
+ 144,
+ 711
+ ],
+ "score": 0.86,
+ "content": "x _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 144,
+ 699,
+ 413,
+ 711
+ ],
+ "score": 1.0,
+ "content": ". Then, we pass this representation through two CNN layers with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 699,
+ 420,
+ 709
+ ],
+ "score": 0.76,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 420,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "output channels and",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 504,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 477,
+ 722
+ ],
+ "score": 1.0,
+ "content": "obtain a final representation of the image, which we refer to as our Visual Knowledge Base",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 477,
+ 710,
+ 504,
+ 721
+ ],
+ "score": 0.85,
+ "content": "( K B _ { V }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 299,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 299,
+ 732
+ ],
+ "score": 1.0,
+ "content": "that is used in further components of the model.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 37.5
+ }
+ ],
+ "page_idx": 16,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "17",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 81,
+ 255,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 257,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 80,
+ 257,
+ 95
+ ],
+ "score": 1.0,
+ "content": "SUPPLEMENTARY MATERIAL",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0,
+ "bbox_fs": [
+ 106,
+ 80,
+ 257,
+ 95
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 110,
+ 255,
+ 124
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 110,
+ 257,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 110,
+ 257,
+ 126
+ ],
+ "score": 1.0,
+ "content": "A DETAILS OF INPUT UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 138,
+ 505,
+ 183
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 505,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 505,
+ 151
+ ],
+ "score": 1.0,
+ "content": "The input unit processes the raw inputs given to the system into distributed vector representations.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 149,
+ 506,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 149,
+ 506,
+ 163
+ ],
+ "score": 1.0,
+ "content": "It receives a text question (or in general, a query), and an image (or in general, a Knowledge Base",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 174
+ ],
+ "score": 1.0,
+ "content": "(KB)) and processes each of them with a matching sub-unit. Here we provide details of the Query",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 172,
+ 279,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 172,
+ 279,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Unit and the Image Unit used in this work.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 104,
+ 138,
+ 506,
+ 184
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 200,
+ 219,
+ 212
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 220,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 220,
+ 214
+ ],
+ "score": 1.0,
+ "content": "A.0.1 THE QUERY UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 221,
+ 505,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 193,
+ 234
+ ],
+ "score": 1.0,
+ "content": "We encode a query of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 222,
+ 201,
+ 232
+ ],
+ "score": 0.83,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 221,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "words into a continuous representation using a bidirectional LSTM (Hochre-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "iter & Schmidhuber, 1997; Graves et al., 2013). Each word is associated with a word embedding",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 107,
+ 244,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 245,
+ 119,
+ 255
+ ],
+ "score": 0.85,
+ "content": "w _ { s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 244,
+ 151,
+ 257
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 151,
+ 244,
+ 201,
+ 255
+ ],
+ "score": 0.92,
+ "content": "s = 1 , . . . , S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 244,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": ". In our case, we use GloVE words embeddings (Pennington et al., 2014).",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 255,
+ 482,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 255,
+ 423,
+ 267
+ ],
+ "score": 1.0,
+ "content": "Then, these embeddings are processed by a bidirectional LST M of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 255,
+ 430,
+ 264
+ ],
+ "score": 0.77,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 430,
+ 255,
+ 482,
+ 267
+ ],
+ "score": 1.0,
+ "content": "that outputs:",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 105,
+ 221,
+ 505,
+ 267
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 133,
+ 279,
+ 505,
+ 333
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 279,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 279,
+ 237,
+ 293
+ ],
+ "score": 1.0,
+ "content": "• a matching sequence of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 238,
+ 281,
+ 244,
+ 290
+ ],
+ "score": 0.83,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 244,
+ 279,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "-dimensional output states, which we refer to as contextual words,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 143,
+ 291,
+ 200,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 143,
+ 291,
+ 200,
+ 303
+ ],
+ "score": 0.79,
+ "content": "[ c w _ { 1 } , . . . , c w _ { S } ]",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 132,
+ 309,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 309,
+ 142,
+ 322
+ ],
+ "score": 1.0,
+ "content": "•",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 311,
+ 149,
+ 320
+ ],
+ "score": 0.78,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 149,
+ 309,
+ 252,
+ 322
+ ],
+ "score": 1.0,
+ "content": "-dimensional hidden state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 309,
+ 314,
+ 322
+ ],
+ "score": 0.93,
+ "content": "q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 315,
+ 309,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": ", the concatenation of the hidden states from the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 321,
+ 444,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 321,
+ 312,
+ 334
+ ],
+ "score": 1.0,
+ "content": "backward and forward passes. We refer to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 312,
+ 324,
+ 318,
+ 333
+ ],
+ "score": 0.8,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 319,
+ 321,
+ 444,
+ 334
+ ],
+ "score": 1.0,
+ "content": "as the question representation.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 131,
+ 279,
+ 506,
+ 334
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 346,
+ 504,
+ 370
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 506,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 245,
+ 359
+ ],
+ "score": 1.0,
+ "content": "Intuitively, each contextual word",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 245,
+ 349,
+ 262,
+ 358
+ ],
+ "score": 0.87,
+ "content": "c w _ { s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 345,
+ 374,
+ 359
+ ],
+ "score": 1.0,
+ "content": "represents the meaning of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 346,
+ 389,
+ 357
+ ],
+ "score": 0.89,
+ "content": "s ^ { t h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 389,
+ 345,
+ 506,
+ 359
+ ],
+ "score": 1.0,
+ "content": "word, in the context of the",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 357,
+ 496,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 357,
+ 235,
+ 371
+ ],
+ "score": 1.0,
+ "content": "question, while the hidden state",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 360,
+ 241,
+ 370
+ ],
+ "score": 0.8,
+ "content": "q",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 241,
+ 357,
+ 496,
+ 371
+ ],
+ "score": 1.0,
+ "content": "represents the overall (compositional) meaning of the question.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 105,
+ 345,
+ 506,
+ 371
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 386,
+ 217,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 385,
+ 218,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 385,
+ 218,
+ 399
+ ],
+ "score": 1.0,
+ "content": "A.0.2 THE IMAGE UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 407,
+ 505,
+ 463
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 408,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 408,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "Given an image, and following prior work on CLEVR (?Santoro et al., 2017; Perez et al., 2017), we",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "extract conv4 features from ResNet101 (He et al., 2016) pretrained on ImageNet (Krizhevsky et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 377,
+ 442
+ ],
+ "score": 1.0,
+ "content": "2012) which we treat as a fixed initial representation of the image,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 377,
+ 432,
+ 384,
+ 439
+ ],
+ "score": 0.69,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 384,
+ 428,
+ 441,
+ 442
+ ],
+ "score": 1.0,
+ "content": "of dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 441,
+ 430,
+ 477,
+ 441
+ ],
+ "score": 0.88,
+ "content": "H , W , C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 428,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 107,
+ 441,
+ 506,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 441,
+ 167,
+ 451
+ ],
+ "score": 0.9,
+ "content": "H = W = 1 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 441,
+ 390,
+ 452
+ ],
+ "score": 1.0,
+ "content": "are the height and width of the transformed image and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 391,
+ 441,
+ 435,
+ 451
+ ],
+ "score": 0.89,
+ "content": "C = 1 0 2 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 441,
+ 506,
+ 452
+ ],
+ "score": 1.0,
+ "content": "is the number of",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 396,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 199,
+ 466
+ ],
+ "score": 1.0,
+ "content": "channels. Each feature",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 453,
+ 220,
+ 464
+ ],
+ "score": 0.9,
+ "content": "x _ { h , w }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 220,
+ 450,
+ 396,
+ 466
+ ],
+ "score": 1.0,
+ "content": "represents one region in the original image.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 20,
+ "bbox_fs": [
+ 105,
+ 408,
+ 506,
+ 466
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 468,
+ 505,
+ 535
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 506,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 506,
+ 481
+ ],
+ "score": 1.0,
+ "content": "Similar to prior work (Hu et al., 2017; Santoro et al., 2017; Perez et al., 2017), we would like to",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 480,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "allow our model to reason explicitly about spatial locations, as required by many of the questions in",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 490,
+ 506,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 490,
+ 506,
+ 503
+ ],
+ "score": 1.0,
+ "content": "CLEVR, and therefore we concatenate to this representation a spatial map that represents each of the",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 104,
+ 500,
+ 506,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 500,
+ 506,
+ 515
+ ],
+ "score": 1.0,
+ "content": "positions in the image. However, in contrast to prior work that uses a linear meshgrid feature map",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 506,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 169,
+ 525
+ ],
+ "score": 1.0,
+ "content": "with 2 features",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 170,
+ 513,
+ 177,
+ 522
+ ],
+ "score": 0.8,
+ "content": "h",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 511,
+ 196,
+ 525
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 514,
+ 205,
+ 523
+ ],
+ "score": 0.73,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 511,
+ 262,
+ 525
+ ],
+ "score": 1.0,
+ "content": "ranging from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 262,
+ 513,
+ 276,
+ 523
+ ],
+ "score": 0.79,
+ "content": "- 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 511,
+ 506,
+ 525
+ ],
+ "score": 1.0,
+ "content": "to 1, and to allow better representation of the positions,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 406,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 406,
+ 536
+ ],
+ "score": 1.0,
+ "content": "we use the positional encoding scheme proposed by Vaswani et al. (2017):",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 104,
+ 468,
+ 506,
+ 536
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 238,
+ 560,
+ 372,
+ 607
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 238,
+ 560,
+ 372,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 238,
+ 560,
+ 372,
+ 607
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { l } { p _ { ( h , 2 i ) } = \\sin { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\\\ { p _ { ( h , 2 i + 1 ) } = \\cos { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "23f67e68bd8fe769889cb65ef6b26ad32ea8c028817ac3af21e7f4e098381e73.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 29.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 238,
+ 560,
+ 372,
+ 583.5
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 238,
+ 583.5,
+ 372,
+ 607.0
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 621,
+ 504,
+ 644
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 620,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 620,
+ 178,
+ 633
+ ],
+ "score": 1.0,
+ "content": "And similarly for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 179,
+ 623,
+ 187,
+ 631
+ ],
+ "score": 0.77,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 620,
+ 218,
+ 633
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 623,
+ 230,
+ 632
+ ],
+ "score": 0.86,
+ "content": "p _ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 230,
+ 620,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "is a hyperparameter. Overall, the positional encoding of a feature at",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 449,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 141,
+ 645
+ ],
+ "score": 1.0,
+ "content": "position",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 632,
+ 167,
+ 644
+ ],
+ "score": 0.92,
+ "content": "( h , w )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 632,
+ 178,
+ 645
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 632,
+ 210,
+ 644
+ ],
+ "score": 0.92,
+ "content": "[ p _ { h } , p _ { w } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 211,
+ 632,
+ 411,
+ 645
+ ],
+ "score": 1.0,
+ "content": ", the concatenation of the positional encodings for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 411,
+ 632,
+ 419,
+ 642
+ ],
+ "score": 0.84,
+ "content": "h",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 632,
+ 436,
+ 645
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 437,
+ 633,
+ 445,
+ 642
+ ],
+ "score": 0.8,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 445,
+ 632,
+ 449,
+ 645
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31.5,
+ "bbox_fs": [
+ 105,
+ 620,
+ 505,
+ 645
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 648,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "This positional encoding scheme allows better correspondence between the distance of 2 positions",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 107,
+ 659,
+ 505,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 660,
+ 131,
+ 672
+ ],
+ "score": 0.91,
+ "content": "( x , y )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 131,
+ 659,
+ 149,
+ 673
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 660,
+ 173,
+ 672
+ ],
+ "score": 0.92,
+ "content": "( x , y )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 174,
+ 659,
+ 483,
+ 673
+ ],
+ "score": 1.0,
+ "content": "in the image and a vector similarity of their positional encodings, even when",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 483,
+ 661,
+ 495,
+ 672
+ ],
+ "score": 0.86,
+ "content": "p _ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 659,
+ 505,
+ 673
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 670,
+ 172,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 172,
+ 684
+ ],
+ "score": 1.0,
+ "content": "larger than two.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 105,
+ 648,
+ 505,
+ 684
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 316,
+ 701
+ ],
+ "score": 1.0,
+ "content": "We then concatenate the obtained spatial map with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 691,
+ 322,
+ 698
+ ],
+ "score": 0.67,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 687,
+ 506,
+ 701
+ ],
+ "score": 1.0,
+ "content": ", receiving a spatially-aware image represen-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 132,
+ 711
+ ],
+ "score": 1.0,
+ "content": "tation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 700,
+ 144,
+ 711
+ ],
+ "score": 0.86,
+ "content": "x _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 144,
+ 699,
+ 413,
+ 711
+ ],
+ "score": 1.0,
+ "content": ". Then, we pass this representation through two CNN layers with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 699,
+ 420,
+ 709
+ ],
+ "score": 0.76,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 420,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "output channels and",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 504,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 477,
+ 722
+ ],
+ "score": 1.0,
+ "content": "obtain a final representation of the image, which we refer to as our Visual Knowledge Base",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 477,
+ 710,
+ 504,
+ 721
+ ],
+ "score": 0.85,
+ "content": "( K B _ { V }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 299,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 299,
+ 732
+ ],
+ "score": 1.0,
+ "content": "that is used in further components of the model.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 105,
+ 687,
+ 506,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 81,
+ 348,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 349,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 349,
+ 95
+ ],
+ "score": 1.0,
+ "content": "B IMPLEMENTATION AND TRAINING DETAILS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 107,
+ 505,
+ 217
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 107,
+ 505,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 107,
+ 505,
+ 120
+ ],
+ "score": 1.0,
+ "content": "For the question processing, we use GloVE (Pennington et al., 2014) word-vectors with dimension",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 131
+ ],
+ "score": 1.0,
+ "content": "300. For the image processing, we extract conv4 features from ResNet101 (He et al., 2016) pre-",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 129,
+ 505,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 129,
+ 362,
+ 141
+ ],
+ "score": 1.0,
+ "content": "trained on ImageNet (Krizhevsky et al., 2012), with dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 362,
+ 129,
+ 398,
+ 141
+ ],
+ "score": 0.78,
+ "content": "H , W , C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 129,
+ 426,
+ 141
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 426,
+ 129,
+ 487,
+ 140
+ ],
+ "score": 0.9,
+ "content": "H = W = 1 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 129,
+ 505,
+ 141
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 139,
+ 506,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 140,
+ 149,
+ 151
+ ],
+ "score": 0.9,
+ "content": "C = 1 0 2 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 149,
+ 139,
+ 450,
+ 153
+ ],
+ "score": 1.0,
+ "content": ", followed by 2 CNN layers with kernel size 2. We use MAC network with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 450,
+ 141,
+ 481,
+ 151
+ ],
+ "score": 0.89,
+ "content": "p = 1 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 481,
+ 139,
+ 506,
+ 153
+ ],
+ "score": 1.0,
+ "content": "cells,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 104,
+ 150,
+ 506,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 151,
+ 379,
+ 163
+ ],
+ "score": 1.0,
+ "content": "and train it using Adam (Kingma & Ba, 2014), with learning rate",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 150,
+ 401,
+ 162
+ ],
+ "score": 0.9,
+ "content": "1 0 ^ { - 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 401,
+ 151,
+ 506,
+ 163
+ ],
+ "score": 1.0,
+ "content": ". We train our model for",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 161,
+ 506,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 162,
+ 141,
+ 173
+ ],
+ "score": 0.76,
+ "content": "1 0 - 2 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 141,
+ 161,
+ 506,
+ 175
+ ],
+ "score": 1.0,
+ "content": "epochs, with batch size 64, and use early stopping based on validation accuracies. During",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 186
+ ],
+ "score": 1.0,
+ "content": "training, the moving averages of all weights of the model are maintained with the exponential de-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 184,
+ 505,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 505,
+ 197
+ ],
+ "score": 1.0,
+ "content": "cay rate of 0.999. At test time, the moving averages instead of the raw weights are used. We use",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "dropout 0.85, and ELU (Clevert et al., 2015) which in our experience has reduce the training process",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 206,
+ 481,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 206,
+ 481,
+ 218
+ ],
+ "score": 1.0,
+ "content": "compared to RELU.The training process takes roughly 10-20 hours on a single Titan X GPU.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 236,
+ 342,
+ 248
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 344,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 344,
+ 250
+ ],
+ "score": 1.0,
+ "content": "C FURTHER DISCUSSION OF RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 261,
+ 505,
+ 328
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 261,
+ 505,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 261,
+ 505,
+ 273
+ ],
+ "score": 1.0,
+ "content": "In this section we provide detailed discussion of related work. Several models have been applied",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 271,
+ 505,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 271,
+ 505,
+ 286
+ ],
+ "score": 1.0,
+ "content": "to the CLEVR task. These can be partitioned into two groups, module networks that use the strong",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 296
+ ],
+ "score": 1.0,
+ "content": "supervision provided as a tree-structured functional program associated with each instance, and",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 306
+ ],
+ "score": 1.0,
+ "content": "end-to-end, fully differentiable networks that combine a fairly standard stack of CNNs with com-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 506,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 506,
+ 317
+ ],
+ "score": 1.0,
+ "content": "ponents that aid them in performing reasoning tasks. We also discuss the relation of MAC to other",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 317,
+ 353,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 317,
+ 353,
+ 330
+ ],
+ "score": 1.0,
+ "content": "approaches, such as memory networks and neural computers.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 343,
+ 224,
+ 354
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 226,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 226,
+ 356
+ ],
+ "score": 1.0,
+ "content": "C.1 MODULE NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 364,
+ 505,
+ 452
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 377
+ ],
+ "score": 1.0,
+ "content": "The modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017) first translates",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 376,
+ 505,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 505,
+ 388
+ ],
+ "score": 1.0,
+ "content": "the given question into a tree-structured action plan, aiming to imitate the ground-truth programs",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 387,
+ 506,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 506,
+ 398
+ ],
+ "score": 1.0,
+ "content": "provided as a form of strong-supervision. Then, it constructs a tailor-made network that executes",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "the plan on the image in multiple steps. This network is composed of discrete units selected out of",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 407,
+ 506,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 407,
+ 506,
+ 421
+ ],
+ "score": 1.0,
+ "content": "a collection of predefined modules, each responsible for an elementary reasoning operation, such as",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 433
+ ],
+ "score": 1.0,
+ "content": "identifying an objects color, filtering them for their shape, or comparing two amounts. Each module",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 431,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "has its own set of learned parameters (Johnson et al., 2017), or even hand-crafted design (Andreas",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 325,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 325,
+ 453
+ ],
+ "score": 1.0,
+ "content": "et al., 2016a) to guide it towards its intended behavior.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 458,
+ 505,
+ 502
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 459,
+ 505,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 459,
+ 505,
+ 470
+ ],
+ "score": 1.0,
+ "content": "Overall, this approach makes discrete choices at two levels: the identity of each module – the be-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 481
+ ],
+ "score": 1.0,
+ "content": "havior it should learn among a fixed set of possible types of behaviors, and the network layout – the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 505,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 505,
+ 493
+ ],
+ "score": 1.0,
+ "content": "way in which these modules are wired together to compute the answer progressively. Hence, their",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 490,
+ 498,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 498,
+ 505
+ ],
+ "score": 1.0,
+ "content": "differentiability is confined to the boundaries of a single module, disallowing end-to-end training.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 508,
+ 505,
+ 650
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "Several key differences exist between our approaches. First, our model replaces the fixed modules",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 518,
+ 506,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 506,
+ 532
+ ],
+ "score": 1.0,
+ "content": "collection with one versatile and universal cell that shares both its architecture and parameters across",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 530,
+ 505,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 530,
+ 505,
+ 542
+ ],
+ "score": 1.0,
+ "content": "all of its instantiations, and is applied across all the reasoning steps. Second, it replaces the dynamic",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "recursive tree structures with a sequential topology, augmented by soft attention mechanisms, as",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 552,
+ 505,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 505,
+ 564
+ ],
+ "score": 1.0,
+ "content": "done in Bahdanau et al. (2014). This confers our network with a virtual capacity to represent arbi-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 575
+ ],
+ "score": 1.0,
+ "content": "trarily complex Directed Acyclic Graphs (DAGs) while still having efficient and readily deployed",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 587
+ ],
+ "score": 1.0,
+ "content": "physical sequential structure. Together, both of these relaxations allow us to effectively train our",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "model end-to-end by backpropagation alone, whereas module networks demand a more involved",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 609
+ ],
+ "score": 1.0,
+ "content": "training scheme that relies on the strongly-supervised programs at the first stage, and on various",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 607,
+ 505,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 607,
+ 505,
+ 619
+ ],
+ "score": 1.0,
+ "content": "Reinforcement Learning (RL) techniques at the second. Furthermore, while our model can be train",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "without the strong supervisory programs, developing adaptive reasoning skills to address the task",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "is it trained for, the modular approach reliance on questions structured and formal representation",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 277,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 277,
+ 652
+ ],
+ "score": 1.0,
+ "content": "hinder its applicability to real-world tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 666,
+ 359,
+ 677
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 665,
+ 360,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 665,
+ 360,
+ 679
+ ],
+ "score": 1.0,
+ "content": "C.2 AUGMENTED CONVOLUTIONAL NEURAL NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 44
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 687,
+ 504,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 686,
+ 506,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 686,
+ 506,
+ 701
+ ],
+ "score": 1.0,
+ "content": "Alternative approaches for the CLEVR task that do not rely on the provided programs as a strong",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "supervision signal are Santoro et al. (2017) and Perez et al. (2017). Both complement standard",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "multi-layer Convolutional Neural Networks (CNNs) with components that aid them in handling",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 265,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 265,
+ 733
+ ],
+ "score": 1.0,
+ "content": "compositional and relational questions.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 46.5
+ }
+ ],
+ "page_idx": 17,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 301,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 763
+ ],
+ "score": 1.0,
+ "content": "18",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 81,
+ 348,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 349,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 349,
+ 95
+ ],
+ "score": 1.0,
+ "content": "B IMPLEMENTATION AND TRAINING DETAILS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 107,
+ 505,
+ 217
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 107,
+ 505,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 107,
+ 505,
+ 120
+ ],
+ "score": 1.0,
+ "content": "For the question processing, we use GloVE (Pennington et al., 2014) word-vectors with dimension",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 131
+ ],
+ "score": 1.0,
+ "content": "300. For the image processing, we extract conv4 features from ResNet101 (He et al., 2016) pre-",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 129,
+ 505,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 129,
+ 362,
+ 141
+ ],
+ "score": 1.0,
+ "content": "trained on ImageNet (Krizhevsky et al., 2012), with dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 362,
+ 129,
+ 398,
+ 141
+ ],
+ "score": 0.78,
+ "content": "H , W , C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 129,
+ 426,
+ 141
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 426,
+ 129,
+ 487,
+ 140
+ ],
+ "score": 0.9,
+ "content": "H = W = 1 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 129,
+ 505,
+ 141
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 139,
+ 506,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 140,
+ 149,
+ 151
+ ],
+ "score": 0.9,
+ "content": "C = 1 0 2 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 149,
+ 139,
+ 450,
+ 153
+ ],
+ "score": 1.0,
+ "content": ", followed by 2 CNN layers with kernel size 2. We use MAC network with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 450,
+ 141,
+ 481,
+ 151
+ ],
+ "score": 0.89,
+ "content": "p = 1 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 481,
+ 139,
+ 506,
+ 153
+ ],
+ "score": 1.0,
+ "content": "cells,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 104,
+ 150,
+ 506,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 151,
+ 379,
+ 163
+ ],
+ "score": 1.0,
+ "content": "and train it using Adam (Kingma & Ba, 2014), with learning rate",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 150,
+ 401,
+ 162
+ ],
+ "score": 0.9,
+ "content": "1 0 ^ { - 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 401,
+ 151,
+ 506,
+ 163
+ ],
+ "score": 1.0,
+ "content": ". We train our model for",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 161,
+ 506,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 162,
+ 141,
+ 173
+ ],
+ "score": 0.76,
+ "content": "1 0 - 2 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 141,
+ 161,
+ 506,
+ 175
+ ],
+ "score": 1.0,
+ "content": "epochs, with batch size 64, and use early stopping based on validation accuracies. During",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 186
+ ],
+ "score": 1.0,
+ "content": "training, the moving averages of all weights of the model are maintained with the exponential de-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 184,
+ 505,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 505,
+ 197
+ ],
+ "score": 1.0,
+ "content": "cay rate of 0.999. At test time, the moving averages instead of the raw weights are used. We use",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 195,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "dropout 0.85, and ELU (Clevert et al., 2015) which in our experience has reduce the training process",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 206,
+ 481,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 206,
+ 481,
+ 218
+ ],
+ "score": 1.0,
+ "content": "compared to RELU.The training process takes roughly 10-20 hours on a single Titan X GPU.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 104,
+ 107,
+ 506,
+ 218
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 236,
+ 342,
+ 248
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 344,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 344,
+ 250
+ ],
+ "score": 1.0,
+ "content": "C FURTHER DISCUSSION OF RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 261,
+ 505,
+ 328
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 261,
+ 505,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 261,
+ 505,
+ 273
+ ],
+ "score": 1.0,
+ "content": "In this section we provide detailed discussion of related work. Several models have been applied",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 271,
+ 505,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 271,
+ 505,
+ 286
+ ],
+ "score": 1.0,
+ "content": "to the CLEVR task. These can be partitioned into two groups, module networks that use the strong",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 296
+ ],
+ "score": 1.0,
+ "content": "supervision provided as a tree-structured functional program associated with each instance, and",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 306
+ ],
+ "score": 1.0,
+ "content": "end-to-end, fully differentiable networks that combine a fairly standard stack of CNNs with com-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 506,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 506,
+ 317
+ ],
+ "score": 1.0,
+ "content": "ponents that aid them in performing reasoning tasks. We also discuss the relation of MAC to other",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 317,
+ 353,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 317,
+ 353,
+ 330
+ ],
+ "score": 1.0,
+ "content": "approaches, such as memory networks and neural computers.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 105,
+ 261,
+ 506,
+ 330
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 343,
+ 224,
+ 354
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 226,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 226,
+ 356
+ ],
+ "score": 1.0,
+ "content": "C.1 MODULE NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 364,
+ 505,
+ 452
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 377
+ ],
+ "score": 1.0,
+ "content": "The modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017) first translates",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 376,
+ 505,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 505,
+ 388
+ ],
+ "score": 1.0,
+ "content": "the given question into a tree-structured action plan, aiming to imitate the ground-truth programs",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 387,
+ 506,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 506,
+ 398
+ ],
+ "score": 1.0,
+ "content": "provided as a form of strong-supervision. Then, it constructs a tailor-made network that executes",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": "the plan on the image in multiple steps. This network is composed of discrete units selected out of",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 407,
+ 506,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 407,
+ 506,
+ 421
+ ],
+ "score": 1.0,
+ "content": "a collection of predefined modules, each responsible for an elementary reasoning operation, such as",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 433
+ ],
+ "score": 1.0,
+ "content": "identifying an objects color, filtering them for their shape, or comparing two amounts. Each module",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 431,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "has its own set of learned parameters (Johnson et al., 2017), or even hand-crafted design (Andreas",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 325,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 325,
+ 453
+ ],
+ "score": 1.0,
+ "content": "et al., 2016a) to guide it towards its intended behavior.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 104,
+ 364,
+ 506,
+ 453
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 458,
+ 505,
+ 502
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 459,
+ 505,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 459,
+ 505,
+ 470
+ ],
+ "score": 1.0,
+ "content": "Overall, this approach makes discrete choices at two levels: the identity of each module – the be-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 469,
+ 505,
+ 481
+ ],
+ "score": 1.0,
+ "content": "havior it should learn among a fixed set of possible types of behaviors, and the network layout – the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 505,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 505,
+ 493
+ ],
+ "score": 1.0,
+ "content": "way in which these modules are wired together to compute the answer progressively. Hence, their",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 490,
+ 498,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 498,
+ 505
+ ],
+ "score": 1.0,
+ "content": "differentiability is confined to the boundaries of a single module, disallowing end-to-end training.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 105,
+ 459,
+ 505,
+ 505
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 508,
+ 505,
+ 650
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "Several key differences exist between our approaches. First, our model replaces the fixed modules",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 518,
+ 506,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 506,
+ 532
+ ],
+ "score": 1.0,
+ "content": "collection with one versatile and universal cell that shares both its architecture and parameters across",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 530,
+ 505,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 530,
+ 505,
+ 542
+ ],
+ "score": 1.0,
+ "content": "all of its instantiations, and is applied across all the reasoning steps. Second, it replaces the dynamic",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "recursive tree structures with a sequential topology, augmented by soft attention mechanisms, as",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 552,
+ 505,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 505,
+ 564
+ ],
+ "score": 1.0,
+ "content": "done in Bahdanau et al. (2014). This confers our network with a virtual capacity to represent arbi-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 575
+ ],
+ "score": 1.0,
+ "content": "trarily complex Directed Acyclic Graphs (DAGs) while still having efficient and readily deployed",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 587
+ ],
+ "score": 1.0,
+ "content": "physical sequential structure. Together, both of these relaxations allow us to effectively train our",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "model end-to-end by backpropagation alone, whereas module networks demand a more involved",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 609
+ ],
+ "score": 1.0,
+ "content": "training scheme that relies on the strongly-supervised programs at the first stage, and on various",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 607,
+ 505,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 607,
+ 505,
+ 619
+ ],
+ "score": 1.0,
+ "content": "Reinforcement Learning (RL) techniques at the second. Furthermore, while our model can be train",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "without the strong supervisory programs, developing adaptive reasoning skills to address the task",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "is it trained for, the modular approach reliance on questions structured and formal representation",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 277,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 277,
+ 652
+ ],
+ "score": 1.0,
+ "content": "hinder its applicability to real-world tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 105,
+ 509,
+ 506,
+ 652
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 666,
+ 359,
+ 677
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 665,
+ 360,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 665,
+ 360,
+ 679
+ ],
+ "score": 1.0,
+ "content": "C.2 AUGMENTED CONVOLUTIONAL NEURAL NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 44
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 687,
+ 504,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 686,
+ 506,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 686,
+ 506,
+ 701
+ ],
+ "score": 1.0,
+ "content": "Alternative approaches for the CLEVR task that do not rely on the provided programs as a strong",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "supervision signal are Santoro et al. (2017) and Perez et al. (2017). Both complement standard",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "multi-layer Convolutional Neural Networks (CNNs) with components that aid them in handling",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 265,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 265,
+ 733
+ ],
+ "score": 1.0,
+ "content": "compositional and relational questions.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 46.5,
+ "bbox_fs": [
+ 105,
+ 686,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 214
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Relational Networks. Santoro et al. (2017) appends a Relation Network (RN) layer to the CNN.",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": "This layer inspects all pairs of pixels in the image, thereby enhancing the network capacity to reason",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "over binary relations between objects. While this approach is very simple and elegant conceptually,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 114,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 114,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "it suffers from quadratic computational complexity, in contrast to our and other leading approaches.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "But beyond that, closer inspection reveals that this direct pairwise comparison might be unnecessary.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 506,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 506,
+ 151
+ ],
+ "score": 1.0,
+ "content": "Based on the analogy suggested by Santoro et al. (2017), according to which pixels are equivalent to",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 161
+ ],
+ "score": 1.0,
+ "content": "objects and their pairwise interactions to relations, a RN layer attempts to grasp the induced graph",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 171
+ ],
+ "score": 1.0,
+ "content": "between objects all at once in one shallow and broad layer. Conversely, our attention-based model",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 185
+ ],
+ "score": 1.0,
+ "content": "proceeds in steps. It basically compares the image to its current memory and control for this step,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 181,
+ 505,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 181,
+ 505,
+ 195
+ ],
+ "score": 1.0,
+ "content": "aggregates the attended regions into the new memory, and repeats the process. By the same analogy,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 192,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 192,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "it traverses a narrow and deep path, progressively following transitive relations. Consequently, our",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 202,
+ 460,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 460,
+ 217
+ ],
+ "score": 1.0,
+ "content": "model exhibits a relational capacity while circumventing the computational inefficiency.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 219,
+ 505,
+ 286
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 505,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 505,
+ 233
+ ],
+ "score": 1.0,
+ "content": "FiLM. FiLM (Perez et al., 2017) is a recently proposed method that interleaves standard CNN",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 244
+ ],
+ "score": 1.0,
+ "content": "layers that process the given image with linear layers, reminiscent of layer normalization techniques",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 242,
+ 505,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 242,
+ 505,
+ 254
+ ],
+ "score": 1.0,
+ "content": "(Ba et al., 2016; Ioffe & Szegedy, 2015). Each of these layers, called FiLM, is conditioned on",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "the question: the question words are processed by a GRU, and its output is linearly transformed",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 263,
+ 505,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 505,
+ 277
+ ],
+ "score": 1.0,
+ "content": "into matching biases and variances for each of the CNN layers, tilting its activations to reflect the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 274,
+ 424,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 424,
+ 289
+ ],
+ "score": 1.0,
+ "content": "specifics of the given question and affect the computation done over the image.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 291,
+ 505,
+ 468
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "Similarly to our model, this approach features distant modulation between the question and the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 505,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 505,
+ 316
+ ],
+ "score": 1.0,
+ "content": "image, where rather than being fused together into the same vector space, the question can affect the",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 314,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 314,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "image processing only through constrained means – for the case of FiLM – linear transformations.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 337
+ ],
+ "score": 1.0,
+ "content": "However, since the same transformation is applied to all the activations homogeneously, agnostic to",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "both their spatial location as well as the features values, this approach does not allow the question",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 346,
+ 506,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 346,
+ 506,
+ 360
+ ],
+ "score": 1.0,
+ "content": "to differentiate between regions in the image based on the objects or concepts they represent – on",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 357,
+ 505,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 505,
+ 370
+ ],
+ "score": 1.0,
+ "content": "the content of the image. This stands in stark contrast to our attention-based model, which readily",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 368,
+ 505,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 505,
+ 381
+ ],
+ "score": 1.0,
+ "content": "allows and actually encourages the question to inform the model about relevant regions to focus",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 379,
+ 505,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 379,
+ 505,
+ 392
+ ],
+ "score": 1.0,
+ "content": "on. We speculate that this still distant, yet more direct interaction between the question and the",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 390,
+ 505,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 390,
+ 505,
+ 403
+ ],
+ "score": 1.0,
+ "content": "data, or image, for the case of VQA, facilitates learning and increases generalizability. It may be",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 414
+ ],
+ "score": 1.0,
+ "content": "more suitable to VQA tasks, and CLEVR in particular, where the questions demand the responder",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 425
+ ],
+ "score": 1.0,
+ "content": "to focus on specific objects, and reason about their properties or relations, rather than respond based",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 436
+ ],
+ "score": 1.0,
+ "content": "only on a holistic view of the image that may lead to sub-optimal results (Yang et al., 2016), as is",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 434,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "the case of FiLM. Indeed, as demonstrated in 4, there is significant evidence showing our models",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 445,
+ 506,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 445,
+ 506,
+ 458
+ ],
+ "score": 1.0,
+ "content": "better generalization capacity, allowing it to achieve high accuracies much faster, and from less data",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 275,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 275,
+ 469
+ ],
+ "score": 1.0,
+ "content": "than FiLM and other competing methods.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 500,
+ 246,
+ 511
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 249,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 249,
+ 512
+ ],
+ "score": 1.0,
+ "content": "C.3 MEMORY AND ATTENTION",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 528,
+ 505,
+ 627
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 528,
+ 504,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 528,
+ 504,
+ 540
+ ],
+ "score": 1.0,
+ "content": "Our architecture draws inspiration from recent research on memory and attention (Kumar et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "2016; Xiong et al., 2016; Graves et al., 2014; 2016). Kumar et al. (2016); Xiong et al. (2016)",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 550,
+ 506,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 550,
+ 506,
+ 563
+ ],
+ "score": 1.0,
+ "content": "propose the Dynamic Memory Network model that proceeds in an iterative process, applying soft",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 573
+ ],
+ "score": 1.0,
+ "content": "attention to retrieve relevant information from a visual or textual KB, which is in turn accumulated",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "into memory passed from one iteration to the next. However, in contrast to our model, it views the",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "question as an atomic unit, whereas our model decomposes it into a multi-step action plan informing",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 595,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "each cell in our sequential network about its current objective. Another key difference is the distant",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "interaction between the question and the KB that characterizes our model. Conversely, DMN fuses",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 394,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 394,
+ 630
+ ],
+ "score": 1.0,
+ "content": "their corresponding representations together into the same vector space.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 39
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 633,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 504,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 504,
+ 645
+ ],
+ "score": 1.0,
+ "content": "Graves et al. (2016; 2014) complements a neural network with a memory array it can interact with,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 656
+ ],
+ "score": 1.0,
+ "content": "through the means of soft attention. Analogously to our model, it partitions the model into a core",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "neural network, called controller, as well as reading and writing heads that interact with external",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "memory array. However, a main point distinguishing our model from this approach, is the use of",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "dynamic memory, as in Kumar et al. (2016), instead of a fixed-array memory. Each MAC cell is",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "associated with a memory state, our reading unit inspects only the latest memory passed from the",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "previous state, and our writing unit creates a new memory state rather than writing to multiple slots",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "in a fixed shared external memory. Notably, our approach is much more reminiscent of the widely",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 380,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 380,
+ 732
+ ],
+ "score": 1.0,
+ "content": "successful RNN structure, rather than to Graves et al. (2016; 2014) .",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "page_idx": 18,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 301,
+ 752,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 765
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 765
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 15,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 214
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Relational Networks. Santoro et al. (2017) appends a Relation Network (RN) layer to the CNN.",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": "This layer inspects all pairs of pixels in the image, thereby enhancing the network capacity to reason",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "over binary relations between objects. While this approach is very simple and elegant conceptually,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 114,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 114,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "it suffers from quadratic computational complexity, in contrast to our and other leading approaches.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "But beyond that, closer inspection reveals that this direct pairwise comparison might be unnecessary.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 506,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 506,
+ 151
+ ],
+ "score": 1.0,
+ "content": "Based on the analogy suggested by Santoro et al. (2017), according to which pixels are equivalent to",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 505,
+ 161
+ ],
+ "score": 1.0,
+ "content": "objects and their pairwise interactions to relations, a RN layer attempts to grasp the induced graph",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 171
+ ],
+ "score": 1.0,
+ "content": "between objects all at once in one shallow and broad layer. Conversely, our attention-based model",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 185
+ ],
+ "score": 1.0,
+ "content": "proceeds in steps. It basically compares the image to its current memory and control for this step,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 181,
+ 505,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 181,
+ 505,
+ 195
+ ],
+ "score": 1.0,
+ "content": "aggregates the attended regions into the new memory, and repeats the process. By the same analogy,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 192,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 192,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "it traverses a narrow and deep path, progressively following transitive relations. Consequently, our",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 202,
+ 460,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 460,
+ 217
+ ],
+ "score": 1.0,
+ "content": "model exhibits a relational capacity while circumventing the computational inefficiency.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 104,
+ 83,
+ 506,
+ 217
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 219,
+ 505,
+ 286
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 505,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 505,
+ 233
+ ],
+ "score": 1.0,
+ "content": "FiLM. FiLM (Perez et al., 2017) is a recently proposed method that interleaves standard CNN",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 231,
+ 505,
+ 244
+ ],
+ "score": 1.0,
+ "content": "layers that process the given image with linear layers, reminiscent of layer normalization techniques",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 242,
+ 505,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 242,
+ 505,
+ 254
+ ],
+ "score": 1.0,
+ "content": "(Ba et al., 2016; Ioffe & Szegedy, 2015). Each of these layers, called FiLM, is conditioned on",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "the question: the question words are processed by a GRU, and its output is linearly transformed",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 263,
+ 505,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 505,
+ 277
+ ],
+ "score": 1.0,
+ "content": "into matching biases and variances for each of the CNN layers, tilting its activations to reflect the",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 274,
+ 424,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 424,
+ 289
+ ],
+ "score": 1.0,
+ "content": "specifics of the given question and affect the computation done over the image.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 105,
+ 220,
+ 505,
+ 289
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 291,
+ 505,
+ 468
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "Similarly to our model, this approach features distant modulation between the question and the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 505,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 505,
+ 316
+ ],
+ "score": 1.0,
+ "content": "image, where rather than being fused together into the same vector space, the question can affect the",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 314,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 314,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "image processing only through constrained means – for the case of FiLM – linear transformations.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 337
+ ],
+ "score": 1.0,
+ "content": "However, since the same transformation is applied to all the activations homogeneously, agnostic to",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "both their spatial location as well as the features values, this approach does not allow the question",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 346,
+ 506,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 346,
+ 506,
+ 360
+ ],
+ "score": 1.0,
+ "content": "to differentiate between regions in the image based on the objects or concepts they represent – on",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 357,
+ 505,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 505,
+ 370
+ ],
+ "score": 1.0,
+ "content": "the content of the image. This stands in stark contrast to our attention-based model, which readily",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 368,
+ 505,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 505,
+ 381
+ ],
+ "score": 1.0,
+ "content": "allows and actually encourages the question to inform the model about relevant regions to focus",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 379,
+ 505,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 379,
+ 505,
+ 392
+ ],
+ "score": 1.0,
+ "content": "on. We speculate that this still distant, yet more direct interaction between the question and the",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 390,
+ 505,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 390,
+ 505,
+ 403
+ ],
+ "score": 1.0,
+ "content": "data, or image, for the case of VQA, facilitates learning and increases generalizability. It may be",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 414
+ ],
+ "score": 1.0,
+ "content": "more suitable to VQA tasks, and CLEVR in particular, where the questions demand the responder",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 425
+ ],
+ "score": 1.0,
+ "content": "to focus on specific objects, and reason about their properties or relations, rather than respond based",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 436
+ ],
+ "score": 1.0,
+ "content": "only on a holistic view of the image that may lead to sub-optimal results (Yang et al., 2016), as is",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 434,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "the case of FiLM. Indeed, as demonstrated in 4, there is significant evidence showing our models",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 445,
+ 506,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 445,
+ 506,
+ 458
+ ],
+ "score": 1.0,
+ "content": "better generalization capacity, allowing it to achieve high accuracies much faster, and from less data",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 275,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 275,
+ 469
+ ],
+ "score": 1.0,
+ "content": "than FiLM and other competing methods.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 105,
+ 291,
+ 506,
+ 469
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 500,
+ 246,
+ 511
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 249,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 249,
+ 512
+ ],
+ "score": 1.0,
+ "content": "C.3 MEMORY AND ATTENTION",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 528,
+ 505,
+ 627
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 528,
+ 504,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 528,
+ 504,
+ 540
+ ],
+ "score": 1.0,
+ "content": "Our architecture draws inspiration from recent research on memory and attention (Kumar et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "2016; Xiong et al., 2016; Graves et al., 2014; 2016). Kumar et al. (2016); Xiong et al. (2016)",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 550,
+ 506,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 550,
+ 506,
+ 563
+ ],
+ "score": 1.0,
+ "content": "propose the Dynamic Memory Network model that proceeds in an iterative process, applying soft",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 573
+ ],
+ "score": 1.0,
+ "content": "attention to retrieve relevant information from a visual or textual KB, which is in turn accumulated",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "into memory passed from one iteration to the next. However, in contrast to our model, it views the",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "question as an atomic unit, whereas our model decomposes it into a multi-step action plan informing",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 595,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "each cell in our sequential network about its current objective. Another key difference is the distant",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "interaction between the question and the KB that characterizes our model. Conversely, DMN fuses",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 394,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 394,
+ 630
+ ],
+ "score": 1.0,
+ "content": "their corresponding representations together into the same vector space.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 39,
+ "bbox_fs": [
+ 104,
+ 528,
+ 506,
+ 630
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 633,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 504,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 504,
+ 645
+ ],
+ "score": 1.0,
+ "content": "Graves et al. (2016; 2014) complements a neural network with a memory array it can interact with,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 656
+ ],
+ "score": 1.0,
+ "content": "through the means of soft attention. Analogously to our model, it partitions the model into a core",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "neural network, called controller, as well as reading and writing heads that interact with external",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "memory array. However, a main point distinguishing our model from this approach, is the use of",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "dynamic memory, as in Kumar et al. (2016), instead of a fixed-array memory. Each MAC cell is",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "associated with a memory state, our reading unit inspects only the latest memory passed from the",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "previous state, and our writing unit creates a new memory state rather than writing to multiple slots",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "in a fixed shared external memory. Notably, our approach is much more reminiscent of the widely",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 380,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 380,
+ 732
+ ],
+ "score": 1.0,
+ "content": "successful RNN structure, rather than to Graves et al. (2016; 2014) .",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 48,
+ "bbox_fs": [
+ 105,
+ 633,
+ 506,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 149
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 505,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 505,
+ 94
+ ],
+ "score": 1.0,
+ "content": "Finally, our approach has potential ties to the VQA models Hu et al. (2017); Lu et al. (2016) which",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 94,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 94,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "also attend both the to question words and the image while progressively addressing the given ques-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "tion. However, both of these models have distinct specialized designs for each of their attention",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 116,
+ 505,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 116,
+ 505,
+ 127
+ ],
+ "score": 1.0,
+ "content": "layers or modules, and have a discrete or fixed layout in which they are composed together. In con-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "trast, our approach relax both of these limitations, having one universal cell design and one universal",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 271,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 271,
+ 150
+ ],
+ "score": 1.0,
+ "content": "self-attending sequential network layout.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 162,
+ 266,
+ 173
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 162,
+ 268,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 162,
+ 268,
+ 175
+ ],
+ "score": 1.0,
+ "content": "C.4 ATTENTION VS. CONVOLUTION",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 183,
+ 505,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 505,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 505,
+ 195
+ ],
+ "score": 1.0,
+ "content": "Compared to other leading methods, our model stands out by being heavily based on soft attention,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "score": 1.0,
+ "content": "whereas most competing approaches are CNN-based, surprisingly lack any attention mechanism.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 218
+ ],
+ "score": 1.0,
+ "content": "Since attention is commonly used in models designed for standard VQA (Antol et al., 2015; Gupta,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 228
+ ],
+ "score": 1.0,
+ "content": "2017; Lu et al., 2016; Yang et al., 2016), it is reasonable to assume that it would be beneficial",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 239
+ ],
+ "score": 1.0,
+ "content": "to incorporate such methods into visual reasoning systems for the CLEVR task as well. In fact,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 238,
+ 506,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 506,
+ 250
+ ],
+ "score": 1.0,
+ "content": "attention mechanisms should be especially useful for multi-step reasoning questions such as those",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 248,
+ 506,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 248,
+ 506,
+ 262
+ ],
+ "score": 1.0,
+ "content": "present in CLEVR. Such questions refer to several relations between different objects in the image",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 272
+ ],
+ "score": 1.0,
+ "content": "and feature compositional structure that may be approached one step at a time. Thus, it should be",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 283
+ ],
+ "score": 1.0,
+ "content": "beneficial for a cogent responder to have the capacity to selectively focus on on one or some objects",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 294
+ ],
+ "score": 1.0,
+ "content": "at each step, traversing the relevant relational links one after the other, both at the image level, and",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 293,
+ 191,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 293,
+ 191,
+ 304
+ ],
+ "score": 1.0,
+ "content": "at the question level.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 309,
+ 505,
+ 408
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 321
+ ],
+ "score": 1.0,
+ "content": "Moreover, attention mechanisms enhance our model’s ability to perform reasoning skills that pertain",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "to aggregation of information across different regions, such as counting, finding maximum value,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "or performing other reduction operations over information that is spread across the image. Indeed,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 355
+ ],
+ "score": 1.0,
+ "content": "as discussed in 4, all existing models for visual reasoning, most of which lacking any attention",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 366
+ ],
+ "score": 1.0,
+ "content": "mechanism, struggle with the counting and numerical comparisons questions present in CLEVR.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 377
+ ],
+ "score": 1.0,
+ "content": "Conversely, our model proves much more capable of performing these reasoning skills, outperform-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "score": 1.0,
+ "content": "ing the other approaches by a wide margin. Noticeably, incorporating soft attention into our model",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 400
+ ],
+ "score": 1.0,
+ "content": "makes it much more adept at performing such aggregation reasoning skills, successfully addressing",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 397,
+ 211,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 211,
+ 410
+ ],
+ "score": 1.0,
+ "content": "the this type of questions.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 414,
+ 504,
+ 469
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 414,
+ 505,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 414,
+ 505,
+ 427
+ ],
+ "score": 1.0,
+ "content": "Finally, as pointed out by Lu et al. (2016); Yang et al. (2016), soft attention confers the model with",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 504,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 504,
+ 438
+ ],
+ "score": 1.0,
+ "content": "robustness to noise introduced from irrelevant information presents in the image, and higher capacity",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 436,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 436,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "for handling larger and more diverse vocabulary, the latter being demonstrated in 4. It allows the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 460
+ ],
+ "score": 1.0,
+ "content": "model to separate the wheat from the chaff, selectively attending to the relevant information only,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 458,
+ 402,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 458,
+ 402,
+ 471
+ ],
+ "score": 1.0,
+ "content": "and arguably, being more resilient to both visual and linguistic variations.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "page_idx": 19,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 298,
+ 750,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 298,
+ 750,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 15
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 149
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 505,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 83,
+ 505,
+ 94
+ ],
+ "score": 1.0,
+ "content": "Finally, our approach has potential ties to the VQA models Hu et al. (2017); Lu et al. (2016) which",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 94,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 94,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "also attend both the to question words and the image while progressively addressing the given ques-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "tion. However, both of these models have distinct specialized designs for each of their attention",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 116,
+ 505,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 116,
+ 505,
+ 127
+ ],
+ "score": 1.0,
+ "content": "layers or modules, and have a discrete or fixed layout in which they are composed together. In con-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "trast, our approach relax both of these limitations, having one universal cell design and one universal",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 271,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 271,
+ 150
+ ],
+ "score": 1.0,
+ "content": "self-attending sequential network layout.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 105,
+ 83,
+ 505,
+ 150
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 162,
+ 266,
+ 173
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 162,
+ 268,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 162,
+ 268,
+ 175
+ ],
+ "score": 1.0,
+ "content": "C.4 ATTENTION VS. CONVOLUTION",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 183,
+ 505,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 505,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 505,
+ 195
+ ],
+ "score": 1.0,
+ "content": "Compared to other leading methods, our model stands out by being heavily based on soft attention,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "score": 1.0,
+ "content": "whereas most competing approaches are CNN-based, surprisingly lack any attention mechanism.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 218
+ ],
+ "score": 1.0,
+ "content": "Since attention is commonly used in models designed for standard VQA (Antol et al., 2015; Gupta,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 228
+ ],
+ "score": 1.0,
+ "content": "2017; Lu et al., 2016; Yang et al., 2016), it is reasonable to assume that it would be beneficial",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 239
+ ],
+ "score": 1.0,
+ "content": "to incorporate such methods into visual reasoning systems for the CLEVR task as well. In fact,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 238,
+ 506,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 506,
+ 250
+ ],
+ "score": 1.0,
+ "content": "attention mechanisms should be especially useful for multi-step reasoning questions such as those",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 248,
+ 506,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 248,
+ 506,
+ 262
+ ],
+ "score": 1.0,
+ "content": "present in CLEVR. Such questions refer to several relations between different objects in the image",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 272
+ ],
+ "score": 1.0,
+ "content": "and feature compositional structure that may be approached one step at a time. Thus, it should be",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 283
+ ],
+ "score": 1.0,
+ "content": "beneficial for a cogent responder to have the capacity to selectively focus on on one or some objects",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 294
+ ],
+ "score": 1.0,
+ "content": "at each step, traversing the relevant relational links one after the other, both at the image level, and",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 293,
+ 191,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 293,
+ 191,
+ 304
+ ],
+ "score": 1.0,
+ "content": "at the question level.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 105,
+ 182,
+ 506,
+ 304
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 309,
+ 505,
+ 408
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 321
+ ],
+ "score": 1.0,
+ "content": "Moreover, attention mechanisms enhance our model’s ability to perform reasoning skills that pertain",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "to aggregation of information across different regions, such as counting, finding maximum value,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "or performing other reduction operations over information that is spread across the image. Indeed,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 355
+ ],
+ "score": 1.0,
+ "content": "as discussed in 4, all existing models for visual reasoning, most of which lacking any attention",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 353,
+ 505,
+ 366
+ ],
+ "score": 1.0,
+ "content": "mechanism, struggle with the counting and numerical comparisons questions present in CLEVR.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 505,
+ 377
+ ],
+ "score": 1.0,
+ "content": "Conversely, our model proves much more capable of performing these reasoning skills, outperform-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "score": 1.0,
+ "content": "ing the other approaches by a wide margin. Noticeably, incorporating soft attention into our model",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 505,
+ 400
+ ],
+ "score": 1.0,
+ "content": "makes it much more adept at performing such aggregation reasoning skills, successfully addressing",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 397,
+ 211,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 397,
+ 211,
+ 410
+ ],
+ "score": 1.0,
+ "content": "the this type of questions.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 309,
+ 505,
+ 410
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 414,
+ 504,
+ 469
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 414,
+ 505,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 414,
+ 505,
+ 427
+ ],
+ "score": 1.0,
+ "content": "Finally, as pointed out by Lu et al. (2016); Yang et al. (2016), soft attention confers the model with",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 504,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 504,
+ 438
+ ],
+ "score": 1.0,
+ "content": "robustness to noise introduced from irrelevant information presents in the image, and higher capacity",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 436,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 436,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "for handling larger and more diverse vocabulary, the latter being demonstrated in 4. It allows the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 506,
+ 460
+ ],
+ "score": 1.0,
+ "content": "model to separate the wheat from the chaff, selectively attending to the relevant information only,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 458,
+ 402,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 458,
+ 402,
+ 471
+ ],
+ "score": 1.0,
+ "content": "and arguably, being more resilient to both visual and linguistic variations.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29,
+ "bbox_fs": [
+ 105,
+ 414,
+ 506,
+ 471
+ ]
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file