diff --git "a/parse/train/S19eAF9ee/S19eAF9ee_middle.json" "b/parse/train/S19eAF9ee/S19eAF9ee_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/train/S19eAF9ee/S19eAF9ee_middle.json"
@@ -0,0 +1,29541 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 79,
+ 483,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 426,
+ 99
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 426,
+ 99
+ ],
+ "score": 1.0,
+ "content": "STRUCTURED SEQUENCE MODELING WITH",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 99,
+ 483,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 99,
+ 483,
+ 117
+ ],
+ "score": 1.0,
+ "content": "GRAPH CONVOLUTIONAL RECURRENT NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 136,
+ 234,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 135,
+ 176,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 135,
+ 176,
+ 148
+ ],
+ "score": 1.0,
+ "content": "Youngjoo Seo",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 146,
+ 191,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 146,
+ 191,
+ 157
+ ],
+ "score": 1.0,
+ "content": "EPFL, Switzerland",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 111,
+ 158,
+ 234,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 158,
+ 234,
+ 169
+ ],
+ "score": 1.0,
+ "content": "youngjoo.seo@epfl.ch",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 341,
+ 135,
+ 498,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 341,
+ 135,
+ 426,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 341,
+ 135,
+ 426,
+ 146
+ ],
+ "score": 1.0,
+ "content": "Michael Defferrard¨",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 341,
+ 147,
+ 419,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 341,
+ 147,
+ 419,
+ 157
+ ],
+ "score": 1.0,
+ "content": "EPFL, Switzerland",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 340,
+ 158,
+ 498,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 340,
+ 158,
+ 498,
+ 169
+ ],
+ "score": 1.0,
+ "content": "michael.defferrard@epfl.ch",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 186,
+ 281,
+ 219
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 185,
+ 208,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 185,
+ 208,
+ 197
+ ],
+ "score": 1.0,
+ "content": "Pierre Vandergheynst",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 112,
+ 197,
+ 191,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 197,
+ 191,
+ 207
+ ],
+ "score": 1.0,
+ "content": "EPFL, Switzerland",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 111,
+ 208,
+ 281,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 208,
+ 281,
+ 219
+ ],
+ "score": 1.0,
+ "content": "pierre.vandergheynst@epfl.ch",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 328,
+ 186,
+ 462,
+ 218
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 327,
+ 185,
+ 396,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 327,
+ 185,
+ 396,
+ 197
+ ],
+ "score": 1.0,
+ "content": "Xavier Bresson",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 328,
+ 197,
+ 406,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 328,
+ 197,
+ 406,
+ 207
+ ],
+ "score": 1.0,
+ "content": "EPFL, Switzerland",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 328,
+ 208,
+ 462,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 328,
+ 208,
+ 462,
+ 219
+ ],
+ "score": 1.0,
+ "content": "xavier.bresson@epfl.ch",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 247,
+ 333,
+ 259
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 247,
+ 335,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 247,
+ 335,
+ 261
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 274,
+ 468,
+ 406
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 140,
+ 272,
+ 470,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 272,
+ 470,
+ 289
+ ],
+ "score": 1.0,
+ "content": "This paper introduces Graph Convolutional Recurrent Network (GCRN), a deep",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 285,
+ 470,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 285,
+ 470,
+ 298
+ ],
+ "score": 1.0,
+ "content": "learning model able to predict structured sequences of data. Precisely, GCRN is",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 296,
+ 470,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 296,
+ 470,
+ 308
+ ],
+ "score": 1.0,
+ "content": "a generalization of classical recurrent neural networks (RNN) to data structured",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 142,
+ 308,
+ 469,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 308,
+ 469,
+ 319
+ ],
+ "score": 1.0,
+ "content": "by an arbitrary graph. Such structured sequences can represent series of frames in",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 318,
+ 470,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 318,
+ 470,
+ 330
+ ],
+ "score": 1.0,
+ "content": "videos, spatio-temporal measurements on a network of sensors, or random walks",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 329,
+ 470,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 329,
+ 470,
+ 342
+ ],
+ "score": 1.0,
+ "content": "on a vocabulary graph for natural language modeling. The proposed model com-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 340,
+ 469,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 340,
+ 469,
+ 352
+ ],
+ "score": 1.0,
+ "content": "bines convolutional neural networks (CNN) on graphs to identify spatial structures",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 351,
+ 470,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 351,
+ 470,
+ 363
+ ],
+ "score": 1.0,
+ "content": "and RNN to find dynamic patterns. We study two possible architectures of GCRN,",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 362,
+ 470,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 362,
+ 470,
+ 375
+ ],
+ "score": 1.0,
+ "content": "and apply the models to two practical problems: predicting moving MNIST data,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 142,
+ 373,
+ 469,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 373,
+ 469,
+ 385
+ ],
+ "score": 1.0,
+ "content": "and modeling natural language with the Penn Treebank dataset. Experiments show",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 384,
+ 470,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 384,
+ 470,
+ 396
+ ],
+ "score": 1.0,
+ "content": "that exploiting simultaneously graph spatial and dynamic information about data",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 395,
+ 333,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 395,
+ 333,
+ 407
+ ],
+ "score": 1.0,
+ "content": "can improve both precision and learning speed.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 430,
+ 206,
+ 443
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 208,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 208,
+ 446
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 456,
+ 505,
+ 545
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 470
+ ],
+ "score": 1.0,
+ "content": "Many real-world data can be cast as structured sequences, with spatio-temporal sequences being a",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "special case. A well-studied example of spatio-temporal data are videos, where succeeding frames",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "share temporal and spatial structures. Many works, such as Donahue et al. (2015); Karpathy & Fei-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "Fei (2015); Vinyals et al. (2015), leveraged a combination of CNN and RNN to exploit such spatial",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "and temporal regularities. Their models are able to process possibly time-varying visual inputs for",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 512,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 512,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "variable-length prediction. These neural network architectures consist of combining a CNN for",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "visual feature extraction followed by a RNN for sequence learning. Such architectures have been",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 534,
+ 461,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 534,
+ 461,
+ 546
+ ],
+ "score": 1.0,
+ "content": "successfully used for video activity recognition, image captioning and video description.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 31.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 550,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "More recently, interest has grown in properly fusing the CNN and RNN models for spatio-temporal",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "sequence modeling. Inspired by language modeling, Ranzato et al. (2014) proposed a model to",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "represent complex deformations and motion patterns by discovering both spatial and temporal cor-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "relations. They showed that prediction of the next video frame and interpolation of intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 607
+ ],
+ "score": 1.0,
+ "content": "frames can be achieved by building a RNN-based language model on the visual words obtained by",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "quantizing the image patches. Their highest-performing model, recursive CNN (rCNN), uses convo-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 617,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 617,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "lutions for both inputs and states. Shi et al. (2015) then proposed the convolutional LSTM network",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "(convLSTM), a recurrent model for spatio-temporal sequence modeling which uses 2D-grid convo-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "lution to leverage the spatial correlations in input data. They successfully applied their model to the",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 410,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 410,
+ 663
+ ],
+ "score": 1.0,
+ "content": "prediction of the evolution of radar echo maps for precipitation nowcasting.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 505,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 505,
+ 679
+ ],
+ "score": 1.0,
+ "content": "The spatial structure of many important problems may however not be as simple as regular grids.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 689
+ ],
+ "score": 1.0,
+ "content": "For instance, the data measured from meteorological stations lie on a irregular grid, i.e. a network of",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "heterogeneous spatial distribution of stations. More challenging, the spatial structure of data may not",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "even be spatial, as it is the case for social or biological networks. Eventually, the interpretation that",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "sentences can be regarded as random walks on vocabulary graphs, a view popularized by Mikolov",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 492,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 492,
+ 733
+ ],
+ "score": 1.0,
+ "content": "et al. (2013), allows us to cast language analysis problems as graph-structured sequence models.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 48.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 303,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 761
+ ],
+ "score": 1.0,
+ "content": "1",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 79,
+ 483,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 426,
+ 99
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 426,
+ 99
+ ],
+ "score": 1.0,
+ "content": "STRUCTURED SEQUENCE MODELING WITH",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 99,
+ 483,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 99,
+ 483,
+ 117
+ ],
+ "score": 1.0,
+ "content": "GRAPH CONVOLUTIONAL RECURRENT NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 136,
+ 234,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 135,
+ 176,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 135,
+ 176,
+ 148
+ ],
+ "score": 1.0,
+ "content": "Youngjoo Seo",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 146,
+ 191,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 146,
+ 191,
+ 157
+ ],
+ "score": 1.0,
+ "content": "EPFL, Switzerland",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 111,
+ 158,
+ 234,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 158,
+ 234,
+ 169
+ ],
+ "score": 1.0,
+ "content": "youngjoo.seo@epfl.ch",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 111,
+ 135,
+ 234,
+ 169
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 341,
+ 135,
+ 498,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 341,
+ 135,
+ 426,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 341,
+ 135,
+ 426,
+ 146
+ ],
+ "score": 1.0,
+ "content": "Michael Defferrard¨",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 341,
+ 147,
+ 419,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 341,
+ 147,
+ 419,
+ 157
+ ],
+ "score": 1.0,
+ "content": "EPFL, Switzerland",
+ "type": "text"
+ }
+ ],
+ "index": 5,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 340,
+ 158,
+ 498,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 340,
+ 158,
+ 498,
+ 169
+ ],
+ "score": 1.0,
+ "content": "michael.defferrard@epfl.ch",
+ "type": "text"
+ }
+ ],
+ "index": 7,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 112,
+ 185,
+ 208,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 185,
+ 208,
+ 197
+ ],
+ "score": 1.0,
+ "content": "Pierre Vandergheynst",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 112,
+ 197,
+ 191,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 197,
+ 191,
+ 207
+ ],
+ "score": 1.0,
+ "content": "EPFL, Switzerland",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 208,
+ 281,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 208,
+ 281,
+ 219
+ ],
+ "score": 1.0,
+ "content": "pierre.vandergheynst@epfl.ch",
+ "type": "text"
+ }
+ ],
+ "index": 12,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 327,
+ 185,
+ 396,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 327,
+ 185,
+ 396,
+ 197
+ ],
+ "score": 1.0,
+ "content": "Xavier Bresson",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 328,
+ 197,
+ 406,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 328,
+ 197,
+ 406,
+ 207
+ ],
+ "score": 1.0,
+ "content": "EPFL, Switzerland",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 328,
+ 208,
+ 462,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 328,
+ 208,
+ 462,
+ 219
+ ],
+ "score": 1.0,
+ "content": "xavier.bresson@epfl.ch",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_start_line": true
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 340,
+ 135,
+ 498,
+ 169
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 112,
+ 186,
+ 281,
+ 219
+ ],
+ "lines": [],
+ "index": 10,
+ "bbox_fs": [
+ 111,
+ 185,
+ 281,
+ 219
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 328,
+ 186,
+ 462,
+ 218
+ ],
+ "lines": [],
+ "index": 11,
+ "bbox_fs": [
+ 327,
+ 185,
+ 462,
+ 219
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 247,
+ 333,
+ 259
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 247,
+ 335,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 247,
+ 335,
+ 261
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 274,
+ 468,
+ 406
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 140,
+ 272,
+ 470,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 272,
+ 470,
+ 289
+ ],
+ "score": 1.0,
+ "content": "This paper introduces Graph Convolutional Recurrent Network (GCRN), a deep",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 285,
+ 470,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 285,
+ 470,
+ 298
+ ],
+ "score": 1.0,
+ "content": "learning model able to predict structured sequences of data. Precisely, GCRN is",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 296,
+ 470,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 296,
+ 470,
+ 308
+ ],
+ "score": 1.0,
+ "content": "a generalization of classical recurrent neural networks (RNN) to data structured",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 142,
+ 308,
+ 469,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 308,
+ 469,
+ 319
+ ],
+ "score": 1.0,
+ "content": "by an arbitrary graph. Such structured sequences can represent series of frames in",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 318,
+ 470,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 318,
+ 470,
+ 330
+ ],
+ "score": 1.0,
+ "content": "videos, spatio-temporal measurements on a network of sensors, or random walks",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 329,
+ 470,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 329,
+ 470,
+ 342
+ ],
+ "score": 1.0,
+ "content": "on a vocabulary graph for natural language modeling. The proposed model com-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 340,
+ 469,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 340,
+ 469,
+ 352
+ ],
+ "score": 1.0,
+ "content": "bines convolutional neural networks (CNN) on graphs to identify spatial structures",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 351,
+ 470,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 351,
+ 470,
+ 363
+ ],
+ "score": 1.0,
+ "content": "and RNN to find dynamic patterns. We study two possible architectures of GCRN,",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 362,
+ 470,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 362,
+ 470,
+ 375
+ ],
+ "score": 1.0,
+ "content": "and apply the models to two practical problems: predicting moving MNIST data,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 142,
+ 373,
+ 469,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 373,
+ 469,
+ 385
+ ],
+ "score": 1.0,
+ "content": "and modeling natural language with the Penn Treebank dataset. Experiments show",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 384,
+ 470,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 384,
+ 470,
+ 396
+ ],
+ "score": 1.0,
+ "content": "that exploiting simultaneously graph spatial and dynamic information about data",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 395,
+ 333,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 395,
+ 333,
+ 407
+ ],
+ "score": 1.0,
+ "content": "can improve both precision and learning speed.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 140,
+ 272,
+ 470,
+ 407
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 430,
+ 206,
+ 443
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 208,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 208,
+ 446
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 456,
+ 505,
+ 545
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 470
+ ],
+ "score": 1.0,
+ "content": "Many real-world data can be cast as structured sequences, with spatio-temporal sequences being a",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "special case. A well-studied example of spatio-temporal data are videos, where succeeding frames",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "share temporal and spatial structures. Many works, such as Donahue et al. (2015); Karpathy & Fei-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "Fei (2015); Vinyals et al. (2015), leveraged a combination of CNN and RNN to exploit such spatial",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "and temporal regularities. Their models are able to process possibly time-varying visual inputs for",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 512,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 512,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "variable-length prediction. These neural network architectures consist of combining a CNN for",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "visual feature extraction followed by a RNN for sequence learning. Such architectures have been",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 534,
+ 461,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 534,
+ 461,
+ 546
+ ],
+ "score": 1.0,
+ "content": "successfully used for video activity recognition, image captioning and video description.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 31.5,
+ "bbox_fs": [
+ 105,
+ 456,
+ 506,
+ 546
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 550,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "More recently, interest has grown in properly fusing the CNN and RNN models for spatio-temporal",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "sequence modeling. Inspired by language modeling, Ranzato et al. (2014) proposed a model to",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "represent complex deformations and motion patterns by discovering both spatial and temporal cor-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "relations. They showed that prediction of the next video frame and interpolation of intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 607
+ ],
+ "score": 1.0,
+ "content": "frames can be achieved by building a RNN-based language model on the visual words obtained by",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "quantizing the image patches. Their highest-performing model, recursive CNN (rCNN), uses convo-",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 617,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 617,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "lutions for both inputs and states. Shi et al. (2015) then proposed the convolutional LSTM network",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "(convLSTM), a recurrent model for spatio-temporal sequence modeling which uses 2D-grid convo-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "lution to leverage the spatial correlations in input data. They successfully applied their model to the",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 410,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 410,
+ 663
+ ],
+ "score": 1.0,
+ "content": "prediction of the evolution of radar echo maps for precipitation nowcasting.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 550,
+ 505,
+ 663
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 505,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 505,
+ 679
+ ],
+ "score": 1.0,
+ "content": "The spatial structure of many important problems may however not be as simple as regular grids.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 689
+ ],
+ "score": 1.0,
+ "content": "For instance, the data measured from meteorological stations lie on a irregular grid, i.e. a network of",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "heterogeneous spatial distribution of stations. More challenging, the spatial structure of data may not",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "even be spatial, as it is the case for social or biological networks. Eventually, the interpretation that",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "sentences can be regarded as random walks on vocabulary graphs, a view popularized by Mikolov",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 492,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 492,
+ 733
+ ],
+ "score": 1.0,
+ "content": "et al. (2013), allows us to cast language analysis problems as graph-structured sequence models.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 48.5,
+ "bbox_fs": [
+ 105,
+ 664,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 109,
+ 83,
+ 339,
+ 299
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 109,
+ 83,
+ 339,
+ 299
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 83,
+ 339,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 83,
+ 339,
+ 299
+ ],
+ "score": 0.955,
+ "type": "image",
+ "image_path": "bc5c5ccd950b0931976650108dec932da5c09eea0c33f2811817daf66f4eb37c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 109,
+ 83,
+ 339,
+ 96.5
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 109,
+ 96.5,
+ 339,
+ 110.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 109,
+ 110.0,
+ 339,
+ 123.5
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 109,
+ 123.5,
+ 339,
+ 137.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 109,
+ 137.0,
+ 339,
+ 150.5
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 109,
+ 150.5,
+ 339,
+ 164.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 109,
+ 164.0,
+ 339,
+ 177.5
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 109,
+ 177.5,
+ 339,
+ 191.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 109,
+ 191.0,
+ 339,
+ 204.5
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 109,
+ 204.5,
+ 339,
+ 218.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 109,
+ 218.0,
+ 339,
+ 231.5
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 109,
+ 231.5,
+ 339,
+ 245.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 109,
+ 245.0,
+ 339,
+ 258.5
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 109,
+ 258.5,
+ 339,
+ 272.0
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 109,
+ 272.0,
+ 339,
+ 285.5
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 109,
+ 285.5,
+ 339,
+ 299.0
+ ],
+ "spans": [],
+ "index": 15
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 314,
+ 342,
+ 370
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 343,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 343,
+ 326
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Illustration of the proposed GCRN model for",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 343,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 343,
+ 337
+ ],
+ "score": 1.0,
+ "content": "spatio-temporal prediction of graph-structured data. The",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 336,
+ 343,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 343,
+ 349
+ ],
+ "score": 1.0,
+ "content": "technique combines at the same time CNN on graphs and",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 343,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 343,
+ 360
+ ],
+ "score": 1.0,
+ "content": "RNN. RNN can be easily exchanged with LSTM or GRU",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 359,
+ 148,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 359,
+ 148,
+ 370
+ ],
+ "score": 1.0,
+ "content": "networks.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 13.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 356,
+ 81,
+ 506,
+ 303
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 356,
+ 81,
+ 506,
+ 303
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 356,
+ 81,
+ 506,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 356,
+ 81,
+ 506,
+ 303
+ ],
+ "score": 0.894,
+ "type": "image",
+ "image_path": "a8c0f2e3882ab26eeb26dbb84bd4e7b15d885d5b9212c832e5fae885167d1d8d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 356,
+ 81,
+ 506,
+ 303
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 356,
+ 315,
+ 505,
+ 370
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 355,
+ 313,
+ 504,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 355,
+ 313,
+ 504,
+ 327
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Illustration of the neigh-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 355,
+ 325,
+ 505,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 355,
+ 325,
+ 505,
+ 338
+ ],
+ "score": 1.0,
+ "content": "borhood on an 8-nearest-neighbor",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 355,
+ 336,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 355,
+ 336,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "grid graph. Isotropic spectral filters",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 355,
+ 348,
+ 506,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 355,
+ 348,
+ 399,
+ 359
+ ],
+ "score": 1.0,
+ "content": "of support",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 399,
+ 348,
+ 410,
+ 358
+ ],
+ "score": 0.76,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 410,
+ 348,
+ 506,
+ 359
+ ],
+ "score": 1.0,
+ "content": "have access to nodes at",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 354,
+ 357,
+ 441,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 354,
+ 357,
+ 387,
+ 372
+ ],
+ "score": 1.0,
+ "content": "most at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 359,
+ 415,
+ 369
+ ],
+ "score": 0.87,
+ "content": "K - 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 415,
+ 357,
+ 441,
+ 372
+ ],
+ "score": 1.0,
+ "content": "hops.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 20.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 388,
+ 505,
+ 444
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 504,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 504,
+ 401
+ ],
+ "score": 1.0,
+ "content": "This work leverages on the recent models of Defferrard et al. (2016); Ranzato et al. (2014); Shi et al.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 399,
+ 504,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 399,
+ 504,
+ 412
+ ],
+ "score": 1.0,
+ "content": "(2015) to design the GCRN model for modeling and predicting time-varying graph-based data. The",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 410,
+ 505,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 505,
+ 424
+ ],
+ "score": 1.0,
+ "content": "core idea is to merge CNN for graph-structured data and RNN to identify simultaneously meaningful",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 421,
+ 506,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 421,
+ 506,
+ 435
+ ],
+ "score": 1.0,
+ "content": "spatial structures and dynamic patterns. A generic illustration of the proposed GCRN architecture",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 433,
+ 191,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 191,
+ 446
+ ],
+ "score": 1.0,
+ "content": "is given by Figure 1.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 460,
+ 208,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 458,
+ 209,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 458,
+ 209,
+ 475
+ ],
+ "score": 1.0,
+ "content": "2 PRELIMINARIES",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 484,
+ 290,
+ 496
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 290,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 290,
+ 497
+ ],
+ "score": 1.0,
+ "content": "2.1 STRUCTURED SEQUENCE MODELING",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 505,
+ 504,
+ 527
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 505,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 416,
+ 519
+ ],
+ "score": 1.0,
+ "content": "Sequence modeling is the problem of predicting the most likely future length-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 506,
+ 426,
+ 515
+ ],
+ "score": 0.82,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 504,
+ 505,
+ 519
+ ],
+ "score": 1.0,
+ "content": "sequence given the",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 208,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 143,
+ 528
+ ],
+ "score": 1.0,
+ "content": "previous",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 144,
+ 516,
+ 151,
+ 526
+ ],
+ "score": 0.82,
+ "content": "J",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 516,
+ 208,
+ 528
+ ],
+ "score": 1.0,
+ "content": "observations:",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 167,
+ 531,
+ 442,
+ 552
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 167,
+ 531,
+ 442,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 167,
+ 531,
+ 442,
+ 552
+ ],
+ "score": 0.91,
+ "content": "\\hat { x } _ { t + 1 } , \\ldots , \\hat { x } _ { t + K } = \\operatorname * { a r g m a x } _ { x _ { t + 1 } , \\ldots , x _ { t + K } } P ( x _ { t + 1 } , \\ldots , x _ { t + K } | x _ { t - J + 1 } , \\ldots , x _ { t } ) ,",
+ "type": "interline_equation",
+ "image_path": "aaa755830e784ee249ddca4a3a4f7d53451ec3cc4a7dba72cabdd70b500cfc9a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 36,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 167,
+ 531,
+ 442,
+ 552
+ ],
+ "spans": [],
+ "index": 36
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 555,
+ 505,
+ 600
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 135,
+ 567
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 555,
+ 175,
+ 566
+ ],
+ "score": 0.91,
+ "content": "x _ { t } ~ \\in ~ \\mathbf { D }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 554,
+ 289,
+ 567
+ ],
+ "score": 1.0,
+ "content": "is an observation at time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 556,
+ 295,
+ 565
+ ],
+ "score": 0.8,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 295,
+ 554,
+ 316,
+ 567
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 555,
+ 327,
+ 565
+ ],
+ "score": 0.72,
+ "content": "\\mathbf { D }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 327,
+ 554,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "denotes the domain of the observed fea-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 290,
+ 578
+ ],
+ "score": 1.0,
+ "content": "tures. The archetypal application being the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 568,
+ 298,
+ 576
+ ],
+ "score": 0.77,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 298,
+ 566,
+ 418,
+ 578
+ ],
+ "score": 1.0,
+ "content": "-gram language model (with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 418,
+ 566,
+ 470,
+ 577
+ ],
+ "score": 0.9,
+ "content": "n = J + 1 ,",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 471,
+ 566,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "), where",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 107,
+ 576,
+ 504,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 577,
+ 207,
+ 589
+ ],
+ "score": 0.9,
+ "content": "P ( x _ { t + 1 } | x _ { t - J + 1 } , . . . , x _ { t } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 207,
+ 576,
+ 336,
+ 591
+ ],
+ "score": 1.0,
+ "content": "models the probability of word",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 337,
+ 578,
+ 357,
+ 589
+ ],
+ "score": 0.9,
+ "content": "x _ { t + 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 576,
+ 496,
+ 591
+ ],
+ "score": 1.0,
+ "content": "to appear conditioned on the past",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 578,
+ 504,
+ 587
+ ],
+ "score": 0.79,
+ "content": "J",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 588,
+ 260,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 260,
+ 600
+ ],
+ "score": 1.0,
+ "content": "words in the sentence (Graves, 2013).",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 38.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 605,
+ 505,
+ 639
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "In this paper, we are interested in special structured sequences, i.e. sequences where features of",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 173,
+ 628
+ ],
+ "score": 1.0,
+ "content": "the observations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 618,
+ 184,
+ 627
+ ],
+ "score": 0.83,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "are not independent but linked by pairwise relationships. Such relationships are",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 272,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 272,
+ 640
+ ],
+ "score": 1.0,
+ "content": "universally modeled by weighted graphs.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 643,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 642,
+ 505,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 642,
+ 128,
+ 657
+ ],
+ "score": 1.0,
+ "content": "Data",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 646,
+ 138,
+ 654
+ ],
+ "score": 0.85,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 139,
+ 642,
+ 505,
+ 657
+ ],
+ "score": 1.0,
+ "content": "can be viewed as a graph signal, i.e. a signal defined on an undirected and weighted graph",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 504,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 168,
+ 667
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { G } = ( \\nu , \\mathcal { E } , A )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 653,
+ 201,
+ 667
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 201,
+ 655,
+ 210,
+ 665
+ ],
+ "score": 0.78,
+ "content": "\\nu",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 210,
+ 653,
+ 281,
+ 667
+ ],
+ "score": 1.0,
+ "content": "is a finite set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 655,
+ 318,
+ 667
+ ],
+ "score": 0.93,
+ "content": "| \\nu | = n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 318,
+ 653,
+ 356,
+ 667
+ ],
+ "score": 1.0,
+ "content": "vertices,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 357,
+ 655,
+ 364,
+ 665
+ ],
+ "score": 0.79,
+ "content": "\\mathcal { E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 653,
+ 455,
+ 667
+ ],
+ "score": 1.0,
+ "content": "is a set of edges and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 655,
+ 504,
+ 665
+ ],
+ "score": 0.91,
+ "content": "A \\in \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "is a weighted adjacency matrix encoding the connection weight between two vertices. A signal",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 107,
+ 674,
+ 507,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 676,
+ 166,
+ 688
+ ],
+ "score": 0.92,
+ "content": "x _ { t } : \\mathcal { V } \\overset { \\cdot } { } \\mathbb { R } ^ { d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 674,
+ 424,
+ 690
+ ],
+ "score": 1.0,
+ "content": "defined on the nodes of the graph may be regarded as a matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 676,
+ 475,
+ 688
+ ],
+ "score": 0.93,
+ "content": "\\boldsymbol { x } _ { t } \\in \\mathbb { R } ^ { n \\times d _ { \\boldsymbol { x } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 475,
+ 674,
+ 507,
+ 690
+ ],
+ "score": 1.0,
+ "content": "whose",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 139,
+ 700
+ ],
+ "score": 1.0,
+ "content": "column",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 139,
+ 689,
+ 144,
+ 698
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 145,
+ 687,
+ 170,
+ 700
+ ],
+ "score": 1.0,
+ "content": "is the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 688,
+ 182,
+ 699
+ ],
+ "score": 0.88,
+ "content": "d _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 687,
+ 271,
+ 700
+ ],
+ "score": 1.0,
+ "content": "-dimensional value of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 689,
+ 282,
+ 699
+ ],
+ "score": 0.84,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 282,
+ 687,
+ 309,
+ 700
+ ],
+ "score": 1.0,
+ "content": "at the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 687,
+ 322,
+ 697
+ ],
+ "score": 0.87,
+ "content": "i ^ { t h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "node. While the number of free variables in",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 103,
+ 698,
+ 507,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 698,
+ 237,
+ 712
+ ],
+ "score": 1.0,
+ "content": "a structured sequence of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 699,
+ 248,
+ 709
+ ],
+ "score": 0.84,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 248,
+ 698,
+ 309,
+ 712
+ ],
+ "score": 1.0,
+ "content": "is in principle",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 698,
+ 357,
+ 711
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { O } ( n ^ { K } d _ { x } { } ^ { K } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 358,
+ 698,
+ 507,
+ 712
+ ],
+ "score": 1.0,
+ "content": ", we seek to exploit the structure of",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "the space of possible predictions to reduce the dimensionality and hence make those problems more",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 146,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 146,
+ 731
+ ],
+ "score": 1.0,
+ "content": "tractable.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 47.5
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "score": 1.0,
+ "content": "2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 109,
+ 83,
+ 339,
+ 299
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 109,
+ 83,
+ 339,
+ 299
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 83,
+ 339,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 83,
+ 339,
+ 299
+ ],
+ "score": 0.955,
+ "type": "image",
+ "image_path": "bc5c5ccd950b0931976650108dec932da5c09eea0c33f2811817daf66f4eb37c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 109,
+ 83,
+ 339,
+ 96.5
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 109,
+ 96.5,
+ 339,
+ 110.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 109,
+ 110.0,
+ 339,
+ 123.5
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 109,
+ 123.5,
+ 339,
+ 137.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 109,
+ 137.0,
+ 339,
+ 150.5
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 109,
+ 150.5,
+ 339,
+ 164.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 109,
+ 164.0,
+ 339,
+ 177.5
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 109,
+ 177.5,
+ 339,
+ 191.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 109,
+ 191.0,
+ 339,
+ 204.5
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 109,
+ 204.5,
+ 339,
+ 218.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 109,
+ 218.0,
+ 339,
+ 231.5
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 109,
+ 231.5,
+ 339,
+ 245.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 109,
+ 245.0,
+ 339,
+ 258.5
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 109,
+ 258.5,
+ 339,
+ 272.0
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 109,
+ 272.0,
+ 339,
+ 285.5
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 109,
+ 285.5,
+ 339,
+ 299.0
+ ],
+ "spans": [],
+ "index": 15
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 314,
+ 342,
+ 370
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 343,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 343,
+ 326
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Illustration of the proposed GCRN model for",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 343,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 343,
+ 337
+ ],
+ "score": 1.0,
+ "content": "spatio-temporal prediction of graph-structured data. The",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 336,
+ 343,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 343,
+ 349
+ ],
+ "score": 1.0,
+ "content": "technique combines at the same time CNN on graphs and",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 343,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 343,
+ 360
+ ],
+ "score": 1.0,
+ "content": "RNN. RNN can be easily exchanged with LSTM or GRU",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 359,
+ 148,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 359,
+ 148,
+ 370
+ ],
+ "score": 1.0,
+ "content": "networks.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 13.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 356,
+ 81,
+ 506,
+ 303
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 356,
+ 81,
+ 506,
+ 303
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 356,
+ 81,
+ 506,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 356,
+ 81,
+ 506,
+ 303
+ ],
+ "score": 0.894,
+ "type": "image",
+ "image_path": "a8c0f2e3882ab26eeb26dbb84bd4e7b15d885d5b9212c832e5fae885167d1d8d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 356,
+ 81,
+ 506,
+ 303
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 356,
+ 315,
+ 505,
+ 370
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 355,
+ 313,
+ 504,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 355,
+ 313,
+ 504,
+ 327
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Illustration of the neigh-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 355,
+ 325,
+ 505,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 355,
+ 325,
+ 505,
+ 338
+ ],
+ "score": 1.0,
+ "content": "borhood on an 8-nearest-neighbor",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 355,
+ 336,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 355,
+ 336,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "grid graph. Isotropic spectral filters",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 355,
+ 348,
+ 506,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 355,
+ 348,
+ 399,
+ 359
+ ],
+ "score": 1.0,
+ "content": "of support",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 399,
+ 348,
+ 410,
+ 358
+ ],
+ "score": 0.76,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 410,
+ 348,
+ 506,
+ 359
+ ],
+ "score": 1.0,
+ "content": "have access to nodes at",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 354,
+ 357,
+ 441,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 354,
+ 357,
+ 387,
+ 372
+ ],
+ "score": 1.0,
+ "content": "most at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 359,
+ 415,
+ 369
+ ],
+ "score": 0.87,
+ "content": "K - 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 415,
+ 357,
+ 441,
+ 372
+ ],
+ "score": 1.0,
+ "content": "hops.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 20.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 388,
+ 505,
+ 444
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 504,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 504,
+ 401
+ ],
+ "score": 1.0,
+ "content": "This work leverages on the recent models of Defferrard et al. (2016); Ranzato et al. (2014); Shi et al.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 399,
+ 504,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 399,
+ 504,
+ 412
+ ],
+ "score": 1.0,
+ "content": "(2015) to design the GCRN model for modeling and predicting time-varying graph-based data. The",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 410,
+ 505,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 505,
+ 424
+ ],
+ "score": 1.0,
+ "content": "core idea is to merge CNN for graph-structured data and RNN to identify simultaneously meaningful",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 421,
+ 506,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 421,
+ 506,
+ 435
+ ],
+ "score": 1.0,
+ "content": "spatial structures and dynamic patterns. A generic illustration of the proposed GCRN architecture",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 433,
+ 191,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 191,
+ 446
+ ],
+ "score": 1.0,
+ "content": "is given by Figure 1.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29,
+ "bbox_fs": [
+ 105,
+ 389,
+ 506,
+ 446
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 460,
+ 208,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 458,
+ 209,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 458,
+ 209,
+ 475
+ ],
+ "score": 1.0,
+ "content": "2 PRELIMINARIES",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 484,
+ 290,
+ 496
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 290,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 290,
+ 497
+ ],
+ "score": 1.0,
+ "content": "2.1 STRUCTURED SEQUENCE MODELING",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 505,
+ 504,
+ 527
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 505,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 416,
+ 519
+ ],
+ "score": 1.0,
+ "content": "Sequence modeling is the problem of predicting the most likely future length-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 506,
+ 426,
+ 515
+ ],
+ "score": 0.82,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 504,
+ 505,
+ 519
+ ],
+ "score": 1.0,
+ "content": "sequence given the",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 208,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 143,
+ 528
+ ],
+ "score": 1.0,
+ "content": "previous",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 144,
+ 516,
+ 151,
+ 526
+ ],
+ "score": 0.82,
+ "content": "J",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 516,
+ 208,
+ 528
+ ],
+ "score": 1.0,
+ "content": "observations:",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34.5,
+ "bbox_fs": [
+ 105,
+ 504,
+ 505,
+ 528
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 167,
+ 531,
+ 442,
+ 552
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 167,
+ 531,
+ 442,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 167,
+ 531,
+ 442,
+ 552
+ ],
+ "score": 0.91,
+ "content": "\\hat { x } _ { t + 1 } , \\ldots , \\hat { x } _ { t + K } = \\operatorname * { a r g m a x } _ { x _ { t + 1 } , \\ldots , x _ { t + K } } P ( x _ { t + 1 } , \\ldots , x _ { t + K } | x _ { t - J + 1 } , \\ldots , x _ { t } ) ,",
+ "type": "interline_equation",
+ "image_path": "aaa755830e784ee249ddca4a3a4f7d53451ec3cc4a7dba72cabdd70b500cfc9a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 36,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 167,
+ 531,
+ 442,
+ 552
+ ],
+ "spans": [],
+ "index": 36
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 555,
+ 505,
+ 600
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 135,
+ 567
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 555,
+ 175,
+ 566
+ ],
+ "score": 0.91,
+ "content": "x _ { t } ~ \\in ~ \\mathbf { D }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 554,
+ 289,
+ 567
+ ],
+ "score": 1.0,
+ "content": "is an observation at time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 556,
+ 295,
+ 565
+ ],
+ "score": 0.8,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 295,
+ 554,
+ 316,
+ 567
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 555,
+ 327,
+ 565
+ ],
+ "score": 0.72,
+ "content": "\\mathbf { D }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 327,
+ 554,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "denotes the domain of the observed fea-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 290,
+ 578
+ ],
+ "score": 1.0,
+ "content": "tures. The archetypal application being the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 291,
+ 568,
+ 298,
+ 576
+ ],
+ "score": 0.77,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 298,
+ 566,
+ 418,
+ 578
+ ],
+ "score": 1.0,
+ "content": "-gram language model (with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 418,
+ 566,
+ 470,
+ 577
+ ],
+ "score": 0.9,
+ "content": "n = J + 1 ,",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 471,
+ 566,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "), where",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 107,
+ 576,
+ 504,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 577,
+ 207,
+ 589
+ ],
+ "score": 0.9,
+ "content": "P ( x _ { t + 1 } | x _ { t - J + 1 } , . . . , x _ { t } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 207,
+ 576,
+ 336,
+ 591
+ ],
+ "score": 1.0,
+ "content": "models the probability of word",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 337,
+ 578,
+ 357,
+ 589
+ ],
+ "score": 0.9,
+ "content": "x _ { t + 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 576,
+ 496,
+ 591
+ ],
+ "score": 1.0,
+ "content": "to appear conditioned on the past",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 578,
+ 504,
+ 587
+ ],
+ "score": 0.79,
+ "content": "J",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 588,
+ 260,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 260,
+ 600
+ ],
+ "score": 1.0,
+ "content": "words in the sentence (Graves, 2013).",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 38.5,
+ "bbox_fs": [
+ 105,
+ 554,
+ 505,
+ 600
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 605,
+ 505,
+ 639
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "In this paper, we are interested in special structured sequences, i.e. sequences where features of",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 173,
+ 628
+ ],
+ "score": 1.0,
+ "content": "the observations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 618,
+ 184,
+ 627
+ ],
+ "score": 0.83,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "are not independent but linked by pairwise relationships. Such relationships are",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 272,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 272,
+ 640
+ ],
+ "score": 1.0,
+ "content": "universally modeled by weighted graphs.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 105,
+ 605,
+ 505,
+ 640
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 643,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 642,
+ 505,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 642,
+ 128,
+ 657
+ ],
+ "score": 1.0,
+ "content": "Data",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 646,
+ 138,
+ 654
+ ],
+ "score": 0.85,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 139,
+ 642,
+ 505,
+ 657
+ ],
+ "score": 1.0,
+ "content": "can be viewed as a graph signal, i.e. a signal defined on an undirected and weighted graph",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 504,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 168,
+ 667
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { G } = ( \\nu , \\mathcal { E } , A )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 653,
+ 201,
+ 667
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 201,
+ 655,
+ 210,
+ 665
+ ],
+ "score": 0.78,
+ "content": "\\nu",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 210,
+ 653,
+ 281,
+ 667
+ ],
+ "score": 1.0,
+ "content": "is a finite set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 655,
+ 318,
+ 667
+ ],
+ "score": 0.93,
+ "content": "| \\nu | = n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 318,
+ 653,
+ 356,
+ 667
+ ],
+ "score": 1.0,
+ "content": "vertices,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 357,
+ 655,
+ 364,
+ 665
+ ],
+ "score": 0.79,
+ "content": "\\mathcal { E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 653,
+ 455,
+ 667
+ ],
+ "score": 1.0,
+ "content": "is a set of edges and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 655,
+ 504,
+ 665
+ ],
+ "score": 0.91,
+ "content": "A \\in \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "is a weighted adjacency matrix encoding the connection weight between two vertices. A signal",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 107,
+ 674,
+ 507,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 676,
+ 166,
+ 688
+ ],
+ "score": 0.92,
+ "content": "x _ { t } : \\mathcal { V } \\overset { \\cdot } { } \\mathbb { R } ^ { d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 674,
+ 424,
+ 690
+ ],
+ "score": 1.0,
+ "content": "defined on the nodes of the graph may be regarded as a matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 676,
+ 475,
+ 688
+ ],
+ "score": 0.93,
+ "content": "\\boldsymbol { x } _ { t } \\in \\mathbb { R } ^ { n \\times d _ { \\boldsymbol { x } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 475,
+ 674,
+ 507,
+ 690
+ ],
+ "score": 1.0,
+ "content": "whose",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 139,
+ 700
+ ],
+ "score": 1.0,
+ "content": "column",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 139,
+ 689,
+ 144,
+ 698
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 145,
+ 687,
+ 170,
+ 700
+ ],
+ "score": 1.0,
+ "content": "is the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 688,
+ 182,
+ 699
+ ],
+ "score": 0.88,
+ "content": "d _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 687,
+ 271,
+ 700
+ ],
+ "score": 1.0,
+ "content": "-dimensional value of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 689,
+ 282,
+ 699
+ ],
+ "score": 0.84,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 282,
+ 687,
+ 309,
+ 700
+ ],
+ "score": 1.0,
+ "content": "at the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 687,
+ 322,
+ 697
+ ],
+ "score": 0.87,
+ "content": "i ^ { t h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "node. While the number of free variables in",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 103,
+ 698,
+ 507,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 698,
+ 237,
+ 712
+ ],
+ "score": 1.0,
+ "content": "a structured sequence of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 699,
+ 248,
+ 709
+ ],
+ "score": 0.84,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 248,
+ 698,
+ 309,
+ 712
+ ],
+ "score": 1.0,
+ "content": "is in principle",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 698,
+ 357,
+ 711
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { O } ( n ^ { K } d _ { x } { } ^ { K } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 358,
+ 698,
+ 507,
+ 712
+ ],
+ "score": 1.0,
+ "content": ", we seek to exploit the structure of",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "the space of possible predictions to reduce the dimensionality and hence make those problems more",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 146,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 146,
+ 731
+ ],
+ "score": 1.0,
+ "content": "tractable.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 47.5,
+ "bbox_fs": [
+ 103,
+ 642,
+ 507,
+ 731
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 259,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 261,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 261,
+ 95
+ ],
+ "score": 1.0,
+ "content": "2.2 LONG SHORT-TERM MEMORY",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 102,
+ 506,
+ 180
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 506,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 506,
+ 116
+ ],
+ "score": 1.0,
+ "content": "A special class of recurrent neural networks (RNN) that prevents the gradient from vanishing too",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 114,
+ 506,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 114,
+ 506,
+ 127
+ ],
+ "score": 1.0,
+ "content": "quickly is the popular long short-term memory (LSTM) introduced by Hochreiter & Schmidhuber",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 124,
+ 506,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 124,
+ 506,
+ 138
+ ],
+ "score": 1.0,
+ "content": "(1997). This architecture has proven stable and powerful for modeling long-range dependencies in",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 506,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 506,
+ 148
+ ],
+ "score": 1.0,
+ "content": "various general-purpose sequence modeling tasks (Graves, 2013; Srivastava et al., 2015; Sutskever",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 146,
+ 506,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 146,
+ 506,
+ 159
+ ],
+ "score": 1.0,
+ "content": "et al., 2014). A fully-connected LSTM (FC-LSTM) may be seen as a multivariate version of LSTM",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 157,
+ 506,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 157,
+ 172,
+ 171
+ ],
+ "score": 1.0,
+ "content": "where the input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 173,
+ 157,
+ 213,
+ 169
+ ],
+ "score": 0.93,
+ "content": "x _ { t } \\in \\mathbb { R } ^ { d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 214,
+ 157,
+ 264,
+ 171
+ ],
+ "score": 1.0,
+ "content": ", cell output",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 157,
+ 325,
+ 169
+ ],
+ "score": 0.93,
+ "content": "h _ { t } \\in [ - 1 , 1 ] ^ { d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 326,
+ 157,
+ 370,
+ 171
+ ],
+ "score": 1.0,
+ "content": "and states",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 157,
+ 410,
+ 169
+ ],
+ "score": 0.93,
+ "content": "c _ { t } \\in \\mathbb { R } ^ { d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 410,
+ 157,
+ 506,
+ 171
+ ],
+ "score": 1.0,
+ "content": "are all vectors. In this",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 168,
+ 388,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 168,
+ 388,
+ 182
+ ],
+ "score": 1.0,
+ "content": "paper, we follow the FC-LSTM formulation of Graves (2013), that is:",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 194,
+ 184,
+ 416,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 194,
+ 184,
+ 416,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 194,
+ 184,
+ 416,
+ 255
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r l } & { i = \\sigma ( W _ { x i } x _ { t } + W _ { h i } h _ { t - 1 } + w _ { c i } \\odot c _ { t - 1 } + b _ { i } ) , } \\\\ & { f = \\sigma ( W _ { x f } x _ { t } + W _ { h f } h _ { t - 1 } + w _ { c f } \\odot c _ { t - 1 } + b _ { f } ) , } \\\\ & { c _ { t } = f _ { t } \\odot c _ { t - 1 } + i _ { t } \\odot \\operatorname { t a n h } ( W _ { x c } x _ { t } + W _ { h c } h _ { t - 1 } + b _ { c } ) , } \\\\ & { o = \\sigma ( W _ { x o } x _ { t } + W _ { h o } h _ { t - 1 } + w _ { c o } \\odot c _ { t } + b _ { o } ) , } \\\\ & { h _ { t } = o \\odot \\operatorname { t a n h } ( c _ { t } ) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "5283c64643a0bc2bd491f8bc46a4643d9f206f0c452a1e0b30d3362102151064.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 194,
+ 184,
+ 416,
+ 198.2
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 194,
+ 198.2,
+ 416,
+ 212.39999999999998
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 194,
+ 212.39999999999998,
+ 416,
+ 226.59999999999997
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 194,
+ 226.59999999999997,
+ 416,
+ 240.79999999999995
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 194,
+ 240.79999999999995,
+ 416,
+ 254.99999999999994
+ ],
+ "spans": [],
+ "index": 12
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 260,
+ 505,
+ 327
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 506,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 135,
+ 273
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 262,
+ 144,
+ 271
+ ],
+ "score": 0.82,
+ "content": "\\odot",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 145,
+ 259,
+ 281,
+ 273
+ ],
+ "score": 1.0,
+ "content": "denotes the Hadamard product,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 261,
+ 299,
+ 273
+ ],
+ "score": 0.91,
+ "content": "\\sigma ( \\cdot )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 299,
+ 259,
+ 392,
+ 273
+ ],
+ "score": 1.0,
+ "content": "the sigmoid function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 260,
+ 485,
+ 272
+ ],
+ "score": 0.91,
+ "content": "\\sigma ( x ) ~ = ~ 1 / ( 1 + e ^ { - x } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 259,
+ 506,
+ 273
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 107,
+ 270,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 271,
+ 171,
+ 283
+ ],
+ "score": 0.91,
+ "content": "i , f , o \\in [ 0 , 1 ] ^ { d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 270,
+ 374,
+ 284
+ ],
+ "score": 1.0,
+ "content": "are the input, forget and output gates. The weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 272,
+ 389,
+ 282
+ ],
+ "score": 0.81,
+ "content": "W _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 388,
+ 272,
+ 435,
+ 282
+ ],
+ "score": 0.69,
+ "content": "\\ l _ { \\cdot } \\in \\mathbb { R } ^ { d _ { h } \\times d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 270,
+ 439,
+ 284
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 442,
+ 271,
+ 501,
+ 282
+ ],
+ "score": 0.67,
+ "content": "\\dot { W } _ { h \\cdot } \\in \\mathbb { R } ^ { d _ { h } \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 270,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 281,
+ 506,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 283,
+ 150,
+ 294
+ ],
+ "score": 0.6,
+ "content": "w _ { c } . \\in \\mathbb { R } ^ { d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 281,
+ 196,
+ 296
+ ],
+ "score": 1.0,
+ "content": "and biases",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 197,
+ 282,
+ 274,
+ 295
+ ],
+ "score": 0.91,
+ "content": "b _ { i } , \\bar { b } _ { f } , b _ { c } , \\bar { b _ { o } } \\in \\mathbb { R } ^ { d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 274,
+ 281,
+ 506,
+ 296
+ ],
+ "score": 1.0,
+ "content": "are the model parameters.1 Such a model is called fully-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 294,
+ 505,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 294,
+ 258,
+ 305
+ ],
+ "score": 1.0,
+ "content": "connected because the dense matrices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 294,
+ 273,
+ 304
+ ],
+ "score": 0.86,
+ "content": "W _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 294,
+ 294,
+ 305
+ ],
+ "score": 1.0,
+ "content": "· and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 294,
+ 294,
+ 309,
+ 304
+ ],
+ "score": 0.89,
+ "content": "W _ { h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 294,
+ 469,
+ 305
+ ],
+ "score": 1.0,
+ "content": "· linearly combine all the components of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 470,
+ 295,
+ 477,
+ 303
+ ],
+ "score": 0.78,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 294,
+ 494,
+ 305
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 294,
+ 501,
+ 303
+ ],
+ "score": 0.8,
+ "content": "h",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 294,
+ 505,
+ 305
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 304,
+ 505,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 304,
+ 248,
+ 317
+ ],
+ "score": 1.0,
+ "content": "The optional peephole connections",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 307,
+ 260,
+ 316
+ ],
+ "score": 0.63,
+ "content": "w _ { c }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 304,
+ 264,
+ 317
+ ],
+ "score": 1.0,
+ "content": "·",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 306,
+ 282,
+ 316
+ ],
+ "score": 0.56,
+ "content": "\\odot c _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 283,
+ 304,
+ 505,
+ 317
+ ],
+ "score": 1.0,
+ "content": ", introduced by Gers & Schmidhuber (2000), have been",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 316,
+ 297,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 316,
+ 297,
+ 327
+ ],
+ "score": 1.0,
+ "content": "found to improve performance on certain tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 340,
+ 350,
+ 351
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 339,
+ 352,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 339,
+ 352,
+ 353
+ ],
+ "score": 1.0,
+ "content": "2.3 CONVOLUTIONAL NEURAL NETWORKS ON GRAPHS",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 360,
+ 505,
+ 569
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 361,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 361,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "Generalizing convolutional neural networks (CNNs) to arbitrary graphs is a recent area of interest.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "Two approaches have been explored in the literature: (i) a generalization of the spatial definition",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 395
+ ],
+ "score": 1.0,
+ "content": "of a convolution (Masci et al., 2015; Niepert et al., 2016) and (ii), a multiplication in the graph",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 504,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 504,
+ 406
+ ],
+ "score": 1.0,
+ "content": "Fourier domain by the way of the convolution theorem (Bruna et al., 2014; Defferrard et al., 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 505,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 404,
+ 505,
+ 417
+ ],
+ "score": 1.0,
+ "content": "Masci et al. (2015) introduced a spatial generalization of CNNs to 3D meshes. The authors used",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "geodesic polar coordinates to define convolution operations on mesh patches, and formulated a",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "deep learning architecture which allows comparison across different meshes. Hence, this method",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 437,
+ 506,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 506,
+ 451
+ ],
+ "score": 1.0,
+ "content": "is tailored to manifolds and is not directly generalizable to arbitrary graphs. Niepert et al. (2016)",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 448,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "proposed a spatial approach which may be decomposed in three steps: (i) select a node, (ii) construct",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 460,
+ 506,
+ 472
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 506,
+ 472
+ ],
+ "score": 1.0,
+ "content": "its neighborhood and (iii) normalize the selected sub-graph, i.e. order the neighboring nodes. The",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 483
+ ],
+ "score": 1.0,
+ "content": "extracted patches are then fed into a conventional 1D Euclidean CNN. As graphs generally do not",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 481,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": "possess a natural ordering (temporal, spatial or otherwise), a labeling procedure should be used to",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "impose it. Bruna et al. (2014) were the first to introduce the spectral framework described below in",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 502,
+ 506,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 383,
+ 516
+ ],
+ "score": 1.0,
+ "content": "the context of graph CNNs. The major drawback of this method is its",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 503,
+ 410,
+ 515
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { O } ( n ^ { 2 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 411,
+ 502,
+ 506,
+ 516
+ ],
+ "score": 1.0,
+ "content": "complexity, which was",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 513,
+ 504,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 513,
+ 475,
+ 528
+ ],
+ "score": 1.0,
+ "content": "overcome with the technique of Defferrard et al. (2016), which offers a linear complexity",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 475,
+ 514,
+ 504,
+ 526
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { O } ( | \\mathcal { E } | )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 506,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 506,
+ 538
+ ],
+ "score": 1.0,
+ "content": "and provides strictly localized filters. Kipf & Welling (2016) took a first-order approximation of",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 548
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 548
+ ],
+ "score": 1.0,
+ "content": "the spectral filters proposed by Defferrard et al. (2016) and successfully used it for semi-supervised",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 560
+ ],
+ "score": 1.0,
+ "content": "classification of nodes. While we focus on the framework introduced by Defferrard et al. (2016), the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 423,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 407,
+ 573
+ ],
+ "score": 1.0,
+ "content": "proposed model is agnostic to the choice of the graph convolution operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 407,
+ 560,
+ 419,
+ 570
+ ],
+ "score": 0.86,
+ "content": "^ { \\ast _ { \\mathcal { G } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 558,
+ 423,
+ 573
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 574,
+ 505,
+ 620
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 575,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 575,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "As it is difficult to express a meaningful translation operator in the vertex domain (Bruna et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 586,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 586,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "2014; Niepert et al., 2016), Defferrard et al. (2016) chose a spectral formulation for the convolution",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 180,
+ 610
+ ],
+ "score": 1.0,
+ "content": "operator on graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 181,
+ 599,
+ 192,
+ 608
+ ],
+ "score": 0.84,
+ "content": "^ { \\ast _ { \\mathcal { G } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 192,
+ 596,
+ 332,
+ 610
+ ],
+ "score": 1.0,
+ "content": ". By this definition, a graph signal",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 332,
+ 596,
+ 379,
+ 607
+ ],
+ "score": 0.93,
+ "content": "x \\in \\overline { { \\mathbb { R } ^ { n \\times d _ { x } } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 380,
+ 596,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "is filtered by a non-parametric",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 608,
+ 415,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 608,
+ 134,
+ 621
+ ],
+ "score": 1.0,
+ "content": "kernel",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 608,
+ 203,
+ 620
+ ],
+ "score": 0.93,
+ "content": "g _ { \\theta } ( \\Lambda ) = \\mathrm { d i a g } ( \\theta )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 608,
+ 234,
+ 621
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 608,
+ 265,
+ 618
+ ],
+ "score": 0.91,
+ "content": "\\boldsymbol { \\theta } \\in \\mathbb { R } ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 265,
+ 608,
+ 415,
+ 621
+ ],
+ "score": 1.0,
+ "content": "is a vector of Fourier coefficients, as",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 172,
+ 624,
+ 438,
+ 640
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 624,
+ 438,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 624,
+ 438,
+ 640
+ ],
+ "score": 0.9,
+ "content": "y = g _ { \\theta } * _ { \\mathcal { G } } x = g _ { \\theta } ( L ) x = g _ { \\theta } ( U \\Lambda U ^ { T } ) x = U g _ { \\theta } ( \\Lambda ) U ^ { T } x \\in \\mathbb { R } ^ { n \\times d _ { x } } ,",
+ "type": "interline_equation",
+ "image_path": "3822f4f70840b8d281bfa285deec5d7c6ba15f6433ea6dc5ddbc6d5c67ac1227.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 43,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 172,
+ 624,
+ 438,
+ 640
+ ],
+ "spans": [],
+ "index": 43
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 713
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 133,
+ 657
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 644,
+ 180,
+ 655
+ ],
+ "score": 0.92,
+ "content": "U \\in \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 642,
+ 316,
+ 657
+ ],
+ "score": 1.0,
+ "content": "is the matrix of eigenvectors and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 317,
+ 644,
+ 362,
+ 655
+ ],
+ "score": 0.91,
+ "content": "\\boldsymbol { \\Lambda } \\in \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 362,
+ 642,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "the diagonal matrix of eigenvalues",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 104,
+ 652,
+ 507,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 652,
+ 251,
+ 671
+ ],
+ "score": 1.0,
+ "content": "of the normalized graph Laplacian",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 655,
+ 450,
+ 667
+ ],
+ "score": 0.88,
+ "content": "\\bar { L } \\ = \\ I _ { n } - D ^ { - 1 / 2 } A D ^ { - 1 / 2 } \\ = U \\bar { \\Lambda } U ^ { T } \\ \\in \\ \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 652,
+ 482,
+ 671
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 657,
+ 493,
+ 667
+ ],
+ "score": 0.87,
+ "content": "I _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 494,
+ 652,
+ 507,
+ 671
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 104,
+ 664,
+ 507,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 664,
+ 204,
+ 682
+ ],
+ "score": 1.0,
+ "content": "the identity matrix and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 667,
+ 255,
+ 678
+ ],
+ "score": 0.92,
+ "content": "D \\in \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 664,
+ 403,
+ 682
+ ],
+ "score": 1.0,
+ "content": "is the diagonal degree matrix with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 403,
+ 667,
+ 468,
+ 681
+ ],
+ "score": 0.94,
+ "content": "D _ { i i } = \\textstyle \\sum _ { j } A _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 468,
+ 664,
+ 507,
+ 682
+ ],
+ "score": 1.0,
+ "content": "(Chung,",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 678,
+ 505,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 678,
+ 224,
+ 693
+ ],
+ "score": 1.0,
+ "content": "1997). Note that the signal",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 681,
+ 232,
+ 689
+ ],
+ "score": 0.79,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 678,
+ 290,
+ 693
+ ],
+ "score": 1.0,
+ "content": "is filtered by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 681,
+ 300,
+ 691
+ ],
+ "score": 0.84,
+ "content": "g _ { \\boldsymbol { \\theta } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 301,
+ 678,
+ 505,
+ 693
+ ],
+ "score": 1.0,
+ "content": "with an element-wise multiplication of its graph",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 689,
+ 506,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 181,
+ 704
+ ],
+ "score": 1.0,
+ "content": "Fourier transform",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 690,
+ 203,
+ 701
+ ],
+ "score": 0.9,
+ "content": "U ^ { T } x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 689,
+ 227,
+ 704
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 693,
+ 237,
+ 702
+ ],
+ "score": 0.82,
+ "content": "g _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 237,
+ 689,
+ 506,
+ 704
+ ],
+ "score": 1.0,
+ "content": "(Shuman et al., 2013). Evaluating (3) is however expensive, as",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 202,
+ 714
+ ],
+ "score": 1.0,
+ "content": "the multiplication with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 702,
+ 211,
+ 711
+ ],
+ "score": 0.82,
+ "content": "U",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 212,
+ 700,
+ 223,
+ 714
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 223,
+ 702,
+ 250,
+ 713
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { O } ( n ^ { 2 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 700,
+ 469,
+ 714
+ ],
+ "score": 1.0,
+ "content": ". Furthermore, computing the eigendecomposition of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 702,
+ 477,
+ 711
+ ],
+ "score": 0.8,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 700,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "might",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 46.5
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 117,
+ 721,
+ 477,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 719,
+ 478,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 719,
+ 271,
+ 734
+ ],
+ "score": 1.0,
+ "content": "1A practical trick is to initialize the biases",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 721,
+ 293,
+ 732
+ ],
+ "score": 0.34,
+ "content": "b _ { i } , b _ { f }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 293,
+ 719,
+ 309,
+ 734
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 722,
+ 318,
+ 731
+ ],
+ "score": 0.87,
+ "content": "b _ { o }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 319,
+ 719,
+ 478,
+ 734
+ ],
+ "score": 1.0,
+ "content": "to one such that the gates are initially open.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 259,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 261,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 261,
+ 95
+ ],
+ "score": 1.0,
+ "content": "2.2 LONG SHORT-TERM MEMORY",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 102,
+ 506,
+ 180
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 506,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 506,
+ 116
+ ],
+ "score": 1.0,
+ "content": "A special class of recurrent neural networks (RNN) that prevents the gradient from vanishing too",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 114,
+ 506,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 114,
+ 506,
+ 127
+ ],
+ "score": 1.0,
+ "content": "quickly is the popular long short-term memory (LSTM) introduced by Hochreiter & Schmidhuber",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 124,
+ 506,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 124,
+ 506,
+ 138
+ ],
+ "score": 1.0,
+ "content": "(1997). This architecture has proven stable and powerful for modeling long-range dependencies in",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 506,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 506,
+ 148
+ ],
+ "score": 1.0,
+ "content": "various general-purpose sequence modeling tasks (Graves, 2013; Srivastava et al., 2015; Sutskever",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 146,
+ 506,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 146,
+ 506,
+ 159
+ ],
+ "score": 1.0,
+ "content": "et al., 2014). A fully-connected LSTM (FC-LSTM) may be seen as a multivariate version of LSTM",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 157,
+ 506,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 157,
+ 172,
+ 171
+ ],
+ "score": 1.0,
+ "content": "where the input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 173,
+ 157,
+ 213,
+ 169
+ ],
+ "score": 0.93,
+ "content": "x _ { t } \\in \\mathbb { R } ^ { d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 214,
+ 157,
+ 264,
+ 171
+ ],
+ "score": 1.0,
+ "content": ", cell output",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 157,
+ 325,
+ 169
+ ],
+ "score": 0.93,
+ "content": "h _ { t } \\in [ - 1 , 1 ] ^ { d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 326,
+ 157,
+ 370,
+ 171
+ ],
+ "score": 1.0,
+ "content": "and states",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 157,
+ 410,
+ 169
+ ],
+ "score": 0.93,
+ "content": "c _ { t } \\in \\mathbb { R } ^ { d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 410,
+ 157,
+ 506,
+ 171
+ ],
+ "score": 1.0,
+ "content": "are all vectors. In this",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 168,
+ 388,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 168,
+ 388,
+ 182
+ ],
+ "score": 1.0,
+ "content": "paper, we follow the FC-LSTM formulation of Graves (2013), that is:",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 104,
+ 102,
+ 506,
+ 182
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 194,
+ 184,
+ 416,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 194,
+ 184,
+ 416,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 194,
+ 184,
+ 416,
+ 255
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r l } & { i = \\sigma ( W _ { x i } x _ { t } + W _ { h i } h _ { t - 1 } + w _ { c i } \\odot c _ { t - 1 } + b _ { i } ) , } \\\\ & { f = \\sigma ( W _ { x f } x _ { t } + W _ { h f } h _ { t - 1 } + w _ { c f } \\odot c _ { t - 1 } + b _ { f } ) , } \\\\ & { c _ { t } = f _ { t } \\odot c _ { t - 1 } + i _ { t } \\odot \\operatorname { t a n h } ( W _ { x c } x _ { t } + W _ { h c } h _ { t - 1 } + b _ { c } ) , } \\\\ & { o = \\sigma ( W _ { x o } x _ { t } + W _ { h o } h _ { t - 1 } + w _ { c o } \\odot c _ { t } + b _ { o } ) , } \\\\ & { h _ { t } = o \\odot \\operatorname { t a n h } ( c _ { t } ) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "5283c64643a0bc2bd491f8bc46a4643d9f206f0c452a1e0b30d3362102151064.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 194,
+ 184,
+ 416,
+ 198.2
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 194,
+ 198.2,
+ 416,
+ 212.39999999999998
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 194,
+ 212.39999999999998,
+ 416,
+ 226.59999999999997
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 194,
+ 226.59999999999997,
+ 416,
+ 240.79999999999995
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 194,
+ 240.79999999999995,
+ 416,
+ 254.99999999999994
+ ],
+ "spans": [],
+ "index": 12
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 260,
+ 505,
+ 327
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 506,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 135,
+ 273
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 262,
+ 144,
+ 271
+ ],
+ "score": 0.82,
+ "content": "\\odot",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 145,
+ 259,
+ 281,
+ 273
+ ],
+ "score": 1.0,
+ "content": "denotes the Hadamard product,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 261,
+ 299,
+ 273
+ ],
+ "score": 0.91,
+ "content": "\\sigma ( \\cdot )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 299,
+ 259,
+ 392,
+ 273
+ ],
+ "score": 1.0,
+ "content": "the sigmoid function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 260,
+ 485,
+ 272
+ ],
+ "score": 0.91,
+ "content": "\\sigma ( x ) ~ = ~ 1 / ( 1 + e ^ { - x } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 259,
+ 506,
+ 273
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 107,
+ 270,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 271,
+ 171,
+ 283
+ ],
+ "score": 0.91,
+ "content": "i , f , o \\in [ 0 , 1 ] ^ { d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 270,
+ 374,
+ 284
+ ],
+ "score": 1.0,
+ "content": "are the input, forget and output gates. The weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 272,
+ 389,
+ 282
+ ],
+ "score": 0.81,
+ "content": "W _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 388,
+ 272,
+ 435,
+ 282
+ ],
+ "score": 0.69,
+ "content": "\\ l _ { \\cdot } \\in \\mathbb { R } ^ { d _ { h } \\times d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 270,
+ 439,
+ 284
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 442,
+ 271,
+ 501,
+ 282
+ ],
+ "score": 0.67,
+ "content": "\\dot { W } _ { h \\cdot } \\in \\mathbb { R } ^ { d _ { h } \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 270,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 281,
+ 506,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 283,
+ 150,
+ 294
+ ],
+ "score": 0.6,
+ "content": "w _ { c } . \\in \\mathbb { R } ^ { d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 281,
+ 196,
+ 296
+ ],
+ "score": 1.0,
+ "content": "and biases",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 197,
+ 282,
+ 274,
+ 295
+ ],
+ "score": 0.91,
+ "content": "b _ { i } , \\bar { b } _ { f } , b _ { c } , \\bar { b _ { o } } \\in \\mathbb { R } ^ { d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 274,
+ 281,
+ 506,
+ 296
+ ],
+ "score": 1.0,
+ "content": "are the model parameters.1 Such a model is called fully-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 294,
+ 505,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 294,
+ 258,
+ 305
+ ],
+ "score": 1.0,
+ "content": "connected because the dense matrices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 294,
+ 273,
+ 304
+ ],
+ "score": 0.86,
+ "content": "W _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 294,
+ 294,
+ 305
+ ],
+ "score": 1.0,
+ "content": "· and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 294,
+ 294,
+ 309,
+ 304
+ ],
+ "score": 0.89,
+ "content": "W _ { h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 294,
+ 469,
+ 305
+ ],
+ "score": 1.0,
+ "content": "· linearly combine all the components of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 470,
+ 295,
+ 477,
+ 303
+ ],
+ "score": 0.78,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 294,
+ 494,
+ 305
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 294,
+ 501,
+ 303
+ ],
+ "score": 0.8,
+ "content": "h",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 294,
+ 505,
+ 305
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 304,
+ 505,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 304,
+ 248,
+ 317
+ ],
+ "score": 1.0,
+ "content": "The optional peephole connections",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 307,
+ 260,
+ 316
+ ],
+ "score": 0.63,
+ "content": "w _ { c }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 304,
+ 264,
+ 317
+ ],
+ "score": 1.0,
+ "content": "·",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 306,
+ 282,
+ 316
+ ],
+ "score": 0.56,
+ "content": "\\odot c _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 283,
+ 304,
+ 505,
+ 317
+ ],
+ "score": 1.0,
+ "content": ", introduced by Gers & Schmidhuber (2000), have been",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 316,
+ 297,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 316,
+ 297,
+ 327
+ ],
+ "score": 1.0,
+ "content": "found to improve performance on certain tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 105,
+ 259,
+ 506,
+ 327
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 340,
+ 350,
+ 351
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 339,
+ 352,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 339,
+ 352,
+ 353
+ ],
+ "score": 1.0,
+ "content": "2.3 CONVOLUTIONAL NEURAL NETWORKS ON GRAPHS",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 360,
+ 505,
+ 569
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 361,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 361,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "Generalizing convolutional neural networks (CNNs) to arbitrary graphs is a recent area of interest.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "Two approaches have been explored in the literature: (i) a generalization of the spatial definition",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 395
+ ],
+ "score": 1.0,
+ "content": "of a convolution (Masci et al., 2015; Niepert et al., 2016) and (ii), a multiplication in the graph",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 504,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 504,
+ 406
+ ],
+ "score": 1.0,
+ "content": "Fourier domain by the way of the convolution theorem (Bruna et al., 2014; Defferrard et al., 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 505,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 404,
+ 505,
+ 417
+ ],
+ "score": 1.0,
+ "content": "Masci et al. (2015) introduced a spatial generalization of CNNs to 3D meshes. The authors used",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "geodesic polar coordinates to define convolution operations on mesh patches, and formulated a",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 505,
+ 439
+ ],
+ "score": 1.0,
+ "content": "deep learning architecture which allows comparison across different meshes. Hence, this method",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 437,
+ 506,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 506,
+ 451
+ ],
+ "score": 1.0,
+ "content": "is tailored to manifolds and is not directly generalizable to arbitrary graphs. Niepert et al. (2016)",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 448,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "proposed a spatial approach which may be decomposed in three steps: (i) select a node, (ii) construct",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 460,
+ 506,
+ 472
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 506,
+ 472
+ ],
+ "score": 1.0,
+ "content": "its neighborhood and (iii) normalize the selected sub-graph, i.e. order the neighboring nodes. The",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 506,
+ 483
+ ],
+ "score": 1.0,
+ "content": "extracted patches are then fed into a conventional 1D Euclidean CNN. As graphs generally do not",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 481,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 481,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": "possess a natural ordering (temporal, spatial or otherwise), a labeling procedure should be used to",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 492,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "impose it. Bruna et al. (2014) were the first to introduce the spectral framework described below in",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 502,
+ 506,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 383,
+ 516
+ ],
+ "score": 1.0,
+ "content": "the context of graph CNNs. The major drawback of this method is its",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 503,
+ 410,
+ 515
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { O } ( n ^ { 2 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 411,
+ 502,
+ 506,
+ 516
+ ],
+ "score": 1.0,
+ "content": "complexity, which was",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 513,
+ 504,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 513,
+ 475,
+ 528
+ ],
+ "score": 1.0,
+ "content": "overcome with the technique of Defferrard et al. (2016), which offers a linear complexity",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 475,
+ 514,
+ 504,
+ 526
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { O } ( | \\mathcal { E } | )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 506,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 506,
+ 538
+ ],
+ "score": 1.0,
+ "content": "and provides strictly localized filters. Kipf & Welling (2016) took a first-order approximation of",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 548
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 548
+ ],
+ "score": 1.0,
+ "content": "the spectral filters proposed by Defferrard et al. (2016) and successfully used it for semi-supervised",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 547,
+ 505,
+ 560
+ ],
+ "score": 1.0,
+ "content": "classification of nodes. While we focus on the framework introduced by Defferrard et al. (2016), the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 423,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 407,
+ 573
+ ],
+ "score": 1.0,
+ "content": "proposed model is agnostic to the choice of the graph convolution operator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 407,
+ 560,
+ 419,
+ 570
+ ],
+ "score": 0.86,
+ "content": "^ { \\ast _ { \\mathcal { G } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 558,
+ 423,
+ 573
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 29,
+ "bbox_fs": [
+ 105,
+ 361,
+ 506,
+ 573
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 574,
+ 505,
+ 620
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 575,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 575,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "As it is difficult to express a meaningful translation operator in the vertex domain (Bruna et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 586,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 586,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "2014; Niepert et al., 2016), Defferrard et al. (2016) chose a spectral formulation for the convolution",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 180,
+ 610
+ ],
+ "score": 1.0,
+ "content": "operator on graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 181,
+ 599,
+ 192,
+ 608
+ ],
+ "score": 0.84,
+ "content": "^ { \\ast _ { \\mathcal { G } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 192,
+ 596,
+ 332,
+ 610
+ ],
+ "score": 1.0,
+ "content": ". By this definition, a graph signal",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 332,
+ 596,
+ 379,
+ 607
+ ],
+ "score": 0.93,
+ "content": "x \\in \\overline { { \\mathbb { R } ^ { n \\times d _ { x } } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 380,
+ 596,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "is filtered by a non-parametric",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 608,
+ 415,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 608,
+ 134,
+ 621
+ ],
+ "score": 1.0,
+ "content": "kernel",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 608,
+ 203,
+ 620
+ ],
+ "score": 0.93,
+ "content": "g _ { \\theta } ( \\Lambda ) = \\mathrm { d i a g } ( \\theta )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 608,
+ 234,
+ 621
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 608,
+ 265,
+ 618
+ ],
+ "score": 0.91,
+ "content": "\\boldsymbol { \\theta } \\in \\mathbb { R } ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 265,
+ 608,
+ 415,
+ 621
+ ],
+ "score": 1.0,
+ "content": "is a vector of Fourier coefficients, as",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 575,
+ 506,
+ 621
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 172,
+ 624,
+ 438,
+ 640
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 624,
+ 438,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 624,
+ 438,
+ 640
+ ],
+ "score": 0.9,
+ "content": "y = g _ { \\theta } * _ { \\mathcal { G } } x = g _ { \\theta } ( L ) x = g _ { \\theta } ( U \\Lambda U ^ { T } ) x = U g _ { \\theta } ( \\Lambda ) U ^ { T } x \\in \\mathbb { R } ^ { n \\times d _ { x } } ,",
+ "type": "interline_equation",
+ "image_path": "3822f4f70840b8d281bfa285deec5d7c6ba15f6433ea6dc5ddbc6d5c67ac1227.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 43,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 172,
+ 624,
+ 438,
+ 640
+ ],
+ "spans": [],
+ "index": 43
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 713
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 642,
+ 133,
+ 657
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 644,
+ 180,
+ 655
+ ],
+ "score": 0.92,
+ "content": "U \\in \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 642,
+ 316,
+ 657
+ ],
+ "score": 1.0,
+ "content": "is the matrix of eigenvectors and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 317,
+ 644,
+ 362,
+ 655
+ ],
+ "score": 0.91,
+ "content": "\\boldsymbol { \\Lambda } \\in \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 362,
+ 642,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "the diagonal matrix of eigenvalues",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 104,
+ 652,
+ 507,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 652,
+ 251,
+ 671
+ ],
+ "score": 1.0,
+ "content": "of the normalized graph Laplacian",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 655,
+ 450,
+ 667
+ ],
+ "score": 0.88,
+ "content": "\\bar { L } \\ = \\ I _ { n } - D ^ { - 1 / 2 } A D ^ { - 1 / 2 } \\ = U \\bar { \\Lambda } U ^ { T } \\ \\in \\ \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 652,
+ 482,
+ 671
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 657,
+ 493,
+ 667
+ ],
+ "score": 0.87,
+ "content": "I _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 494,
+ 652,
+ 507,
+ 671
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 104,
+ 664,
+ 507,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 664,
+ 204,
+ 682
+ ],
+ "score": 1.0,
+ "content": "the identity matrix and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 667,
+ 255,
+ 678
+ ],
+ "score": 0.92,
+ "content": "D \\in \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 664,
+ 403,
+ 682
+ ],
+ "score": 1.0,
+ "content": "is the diagonal degree matrix with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 403,
+ 667,
+ 468,
+ 681
+ ],
+ "score": 0.94,
+ "content": "D _ { i i } = \\textstyle \\sum _ { j } A _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 468,
+ 664,
+ 507,
+ 682
+ ],
+ "score": 1.0,
+ "content": "(Chung,",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 678,
+ 505,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 678,
+ 224,
+ 693
+ ],
+ "score": 1.0,
+ "content": "1997). Note that the signal",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 681,
+ 232,
+ 689
+ ],
+ "score": 0.79,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 678,
+ 290,
+ 693
+ ],
+ "score": 1.0,
+ "content": "is filtered by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 681,
+ 300,
+ 691
+ ],
+ "score": 0.84,
+ "content": "g _ { \\boldsymbol { \\theta } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 301,
+ 678,
+ 505,
+ 693
+ ],
+ "score": 1.0,
+ "content": "with an element-wise multiplication of its graph",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 689,
+ 506,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 181,
+ 704
+ ],
+ "score": 1.0,
+ "content": "Fourier transform",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 690,
+ 203,
+ 701
+ ],
+ "score": 0.9,
+ "content": "U ^ { T } x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 689,
+ 227,
+ 704
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 693,
+ 237,
+ 702
+ ],
+ "score": 0.82,
+ "content": "g _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 237,
+ 689,
+ 506,
+ 704
+ ],
+ "score": 1.0,
+ "content": "(Shuman et al., 2013). Evaluating (3) is however expensive, as",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 202,
+ 714
+ ],
+ "score": 1.0,
+ "content": "the multiplication with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 702,
+ 211,
+ 711
+ ],
+ "score": 0.82,
+ "content": "U",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 212,
+ 700,
+ 223,
+ 714
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 223,
+ 702,
+ 250,
+ 713
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { O } ( n ^ { 2 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 700,
+ 469,
+ 714
+ ],
+ "score": 1.0,
+ "content": ". Furthermore, computing the eigendecomposition of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 702,
+ 477,
+ 711
+ ],
+ "score": 0.8,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 700,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "might",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "be prohibitively expensive for large graphs. To circumvent this problem, Defferrard et al. (2016)",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 94,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 94,
+ 160,
+ 106
+ ],
+ "score": 1.0,
+ "content": "parametrizes",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 160,
+ 96,
+ 171,
+ 105
+ ],
+ "score": 0.8,
+ "content": "g _ { \\theta }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 171,
+ 94,
+ 326,
+ 106
+ ],
+ "score": 1.0,
+ "content": "as a truncated expansion, up to order",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 326,
+ 94,
+ 354,
+ 104
+ ],
+ "score": 0.9,
+ "content": "K - 1",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 355,
+ 94,
+ 470,
+ 106
+ ],
+ "score": 1.0,
+ "content": ", of Chebyshev polynomials",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 470,
+ 94,
+ 482,
+ 105
+ ],
+ "score": 0.89,
+ "content": "T _ { k }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 483,
+ 94,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "such",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 103,
+ 126,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 103,
+ 126,
+ 117
+ ],
+ "score": 1.0,
+ "content": "that",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 46.5,
+ "bbox_fs": [
+ 104,
+ 642,
+ 507,
+ 714
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 114
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "be prohibitively expensive for large graphs. To circumvent this problem, Defferrard et al. (2016)",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 94,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 94,
+ 160,
+ 106
+ ],
+ "score": 1.0,
+ "content": "parametrizes",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 96,
+ 171,
+ 105
+ ],
+ "score": 0.8,
+ "content": "g _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 94,
+ 326,
+ 106
+ ],
+ "score": 1.0,
+ "content": "as a truncated expansion, up to order",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 326,
+ 94,
+ 354,
+ 104
+ ],
+ "score": 0.9,
+ "content": "K - 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 355,
+ 94,
+ 470,
+ 106
+ ],
+ "score": 1.0,
+ "content": ", of Chebyshev polynomials",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 470,
+ 94,
+ 482,
+ 105
+ ],
+ "score": 0.89,
+ "content": "T _ { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 483,
+ 94,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "such",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 103,
+ 126,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 103,
+ 126,
+ 117
+ ],
+ "score": 1.0,
+ "content": "that",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 257,
+ 111,
+ 353,
+ 147
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 257,
+ 111,
+ 353,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 257,
+ 111,
+ 353,
+ 147
+ ],
+ "score": 0.94,
+ "content": "g _ { \\theta } ( \\Lambda ) = \\sum _ { k = 0 } ^ { K - 1 } \\theta _ { k } T _ { k } ( \\tilde { \\Lambda } ) ,",
+ "type": "interline_equation",
+ "image_path": "ca1fa376dcb83ba222b220186098e686f6ca5b2a4db0060aa601fe75f132bc17.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 257,
+ 111,
+ 353,
+ 129.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 257,
+ 129.0,
+ 353,
+ 147.0
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 148,
+ 505,
+ 183
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 194,
+ 162
+ ],
+ "score": 1.0,
+ "content": "where the parameter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 149,
+ 232,
+ 160
+ ],
+ "score": 0.92,
+ "content": "\\theta \\in \\mathbb { R } ^ { K }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 147,
+ 410,
+ 162
+ ],
+ "score": 1.0,
+ "content": "is a vector of Chebyshev coefficients and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 410,
+ 147,
+ 477,
+ 162
+ ],
+ "score": 0.92,
+ "content": "T _ { k } ( \\tilde { \\Lambda } ) \\ \\in \\ \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 147,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "is the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 176
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 237,
+ 176
+ ],
+ "score": 1.0,
+ "content": "Chebyshev polynomial of order",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 163,
+ 244,
+ 172
+ ],
+ "score": 0.82,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 244,
+ 160,
+ 296,
+ 176
+ ],
+ "score": 1.0,
+ "content": "evaluated at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 161,
+ 380,
+ 174
+ ],
+ "score": 0.92,
+ "content": "\\tilde { \\Lambda } = 2 \\Lambda / \\lambda _ { m a x } - I _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 160,
+ 505,
+ 176
+ ],
+ "score": 1.0,
+ "content": ". The graph filtering operation",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 173,
+ 197,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 173,
+ 197,
+ 185
+ ],
+ "score": 1.0,
+ "content": "can then be written as",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 220,
+ 180,
+ 390,
+ 216
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 220,
+ 180,
+ 390,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 220,
+ 180,
+ 390,
+ 216
+ ],
+ "score": 0.94,
+ "content": "y = g _ { \\theta } * _ { \\mathcal { G } } x = g _ { \\theta } ( L ) x = \\sum _ { k = 0 } ^ { K - 1 } \\theta _ { k } T _ { k } ( \\tilde { L } ) x ,",
+ "type": "interline_equation",
+ "image_path": "21ebcfd5a78b0aa995a82680ed098591dda94481e43dcd7d572a6e44709702de.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 220,
+ 180,
+ 390,
+ 198.0
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 220,
+ 198.0,
+ 390,
+ 216.0
+ ],
+ "spans": [],
+ "index": 9
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 217,
+ 505,
+ 297
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 217,
+ 506,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 217,
+ 134,
+ 232
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 217,
+ 200,
+ 231
+ ],
+ "score": 0.92,
+ "content": "T _ { k } ( \\tilde { L } ) \\in \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 200,
+ 217,
+ 359,
+ 232
+ ],
+ "score": 1.0,
+ "content": "is the Chebyshev polynomial of order",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 359,
+ 219,
+ 366,
+ 228
+ ],
+ "score": 0.82,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 366,
+ 217,
+ 506,
+ 232
+ ],
+ "score": 1.0,
+ "content": "evaluated at the scaled Laplacian",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 107,
+ 227,
+ 507,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 230,
+ 192,
+ 243
+ ],
+ "score": 0.88,
+ "content": "\\tilde { L } = 2 L / \\lambda _ { m a x } - I _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 192,
+ 227,
+ 348,
+ 247
+ ],
+ "score": 1.0,
+ "content": ". Using the stable recurrence relation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 348,
+ 230,
+ 483,
+ 243
+ ],
+ "score": 0.93,
+ "content": "T _ { k } ( x ) = 2 x T _ { k - 1 } ( x ) - T _ { k - 2 } ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 483,
+ 227,
+ 507,
+ 247
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 243,
+ 140,
+ 253
+ ],
+ "score": 0.9,
+ "content": "T _ { 0 } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 240,
+ 159,
+ 255
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 243,
+ 194,
+ 253
+ ],
+ "score": 0.89,
+ "content": "T _ { 1 } = x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 194,
+ 240,
+ 295,
+ 255
+ ],
+ "score": 1.0,
+ "content": ", one can evaluate (5) in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 242,
+ 334,
+ 254
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { O } ( K \\vert \\mathcal { E } \\vert )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 240,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "operations, i.e. linearly with the number",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 353,
+ 266
+ ],
+ "score": 1.0,
+ "content": "of edges. Note that as the filtering operation (5) is an order",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 354,
+ 254,
+ 365,
+ 263
+ ],
+ "score": 0.84,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 253,
+ 506,
+ 266
+ ],
+ "score": 1.0,
+ "content": "polynomial of the Laplacian, it is",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 107,
+ 264,
+ 504,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 264,
+ 117,
+ 274
+ ],
+ "score": 0.84,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 117,
+ 264,
+ 356,
+ 276
+ ],
+ "score": 1.0,
+ "content": "-localized and depends only on nodes that are at maximum",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 356,
+ 264,
+ 367,
+ 274
+ ],
+ "score": 0.79,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 264,
+ 504,
+ 276
+ ],
+ "score": 1.0,
+ "content": "hops away from the central node,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 287
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 122,
+ 287
+ ],
+ "score": 1.0,
+ "content": "the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 122,
+ 275,
+ 132,
+ 285
+ ],
+ "score": 0.84,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 275,
+ 505,
+ 287
+ ],
+ "score": 1.0,
+ "content": "-neighborhood. The reader is referred to Defferrard et al. (2016) for details and an in-depth",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 285,
+ 153,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 285,
+ 153,
+ 298
+ ],
+ "score": 1.0,
+ "content": "discussion.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 312,
+ 217,
+ 325
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 311,
+ 218,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 311,
+ 218,
+ 328
+ ],
+ "score": 1.0,
+ "content": "3 RELATED WORKS",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 337,
+ 504,
+ 382
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 337,
+ 505,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 337,
+ 505,
+ 350
+ ],
+ "score": 1.0,
+ "content": "Shi et al. (2015) introduced a model for regular grid-structured sequences, which can be seen as",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 361
+ ],
+ "score": 1.0,
+ "content": "a special case of the proposed model where the graph is an image grid where the nodes are well",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 360,
+ 505,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 360,
+ 505,
+ 372
+ ],
+ "score": 1.0,
+ "content": "ordered. Their model is essentially the classical FC-LSTM (2) where the multiplications by dense",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 368,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 142,
+ 382
+ ],
+ "score": 1.0,
+ "content": "matrices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 371,
+ 155,
+ 380
+ ],
+ "score": 0.65,
+ "content": "W",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 370,
+ 352,
+ 382
+ ],
+ "score": 1.0,
+ "content": "have been replaced by convolutions with kernels",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 371,
+ 363,
+ 380
+ ],
+ "score": 0.73,
+ "content": "W",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 370,
+ 368,
+ 382
+ ],
+ "score": 1.0,
+ "content": ":",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 185,
+ 384,
+ 426,
+ 456
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 185,
+ 384,
+ 426,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 185,
+ 384,
+ 426,
+ 456
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r l } & { i = \\sigma ( W _ { x i } * x _ { t } + W _ { h i } * h _ { t - 1 } + w _ { c i } \\odot c _ { t - 1 } + b _ { i } ) , } \\\\ & { f = \\sigma ( W _ { x f } * x _ { t } + W _ { h f } * h _ { t - 1 } + w _ { c f } \\odot c _ { t - 1 } + b _ { f } ) , } \\\\ & { c _ { t } = f _ { t } \\odot c _ { t - 1 } + i _ { t } \\odot \\operatorname { t a n h } ( W _ { x c } * x _ { t } + W _ { h c } * h _ { t - 1 } + b _ { c } ) , } \\\\ & { o = \\sigma ( W _ { x o } * x _ { t } + W _ { h o } * h _ { t - 1 } + w _ { c o } \\odot c _ { t } + b _ { o } ) , } \\\\ & { h _ { t } = o \\odot \\operatorname { t a n h } ( c _ { t } ) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "d551ef75a9d0ee8af4623ee5acaffef2323d8292b37a1ef738751acab248e26d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 24,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 185,
+ 384,
+ 426,
+ 398.4
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 185,
+ 398.4,
+ 426,
+ 412.79999999999995
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 185,
+ 412.79999999999995,
+ 426,
+ 427.19999999999993
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 185,
+ 427.19999999999993,
+ 426,
+ 441.5999999999999
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 185,
+ 441.5999999999999,
+ 426,
+ 455.9999999999999
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 457,
+ 506,
+ 546
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 134,
+ 471
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 460,
+ 141,
+ 468
+ ],
+ "score": 0.8,
+ "content": "^ *",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 456,
+ 481,
+ 471
+ ],
+ "score": 1.0,
+ "content": "denotes the 2D convolution by a set of kernels. In their setting, the input tensor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 459,
+ 505,
+ 469
+ ],
+ "score": 0.86,
+ "content": "\\boldsymbol { x } _ { t } ~ \\in",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 107,
+ 466,
+ 506,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 468,
+ 154,
+ 479
+ ],
+ "score": 0.91,
+ "content": "\\mathbb { R } ^ { n _ { r } \\times n _ { c } \\times d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 154,
+ 466,
+ 242,
+ 483
+ ],
+ "score": 1.0,
+ "content": "is the observation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 469,
+ 253,
+ 480
+ ],
+ "score": 0.89,
+ "content": "d _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 254,
+ 466,
+ 347,
+ 483
+ ],
+ "score": 1.0,
+ "content": "measurements at time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 470,
+ 352,
+ 479
+ ],
+ "score": 0.74,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 353,
+ 466,
+ 506,
+ 483
+ ],
+ "score": 1.0,
+ "content": "of a dynamical system over a spatial",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 480,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 238,
+ 492
+ ],
+ "score": 1.0,
+ "content": "region represented by a grid of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 238,
+ 482,
+ 250,
+ 491
+ ],
+ "score": 0.87,
+ "content": "n _ { r }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 480,
+ 293,
+ 492
+ ],
+ "score": 1.0,
+ "content": "rows and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 293,
+ 482,
+ 305,
+ 491
+ ],
+ "score": 0.85,
+ "content": "n _ { c }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 305,
+ 480,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": "columns. The model holds spatially distributed",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 490,
+ 506,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 490,
+ 231,
+ 503
+ ],
+ "score": 1.0,
+ "content": "hidden and cell states of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 491,
+ 243,
+ 502
+ ],
+ "score": 0.87,
+ "content": "d _ { h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 490,
+ 333,
+ 503
+ ],
+ "score": 1.0,
+ "content": "given by the tensors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 333,
+ 490,
+ 418,
+ 501
+ ],
+ "score": 0.87,
+ "content": "c _ { t } , h _ { t } \\ \\in \\ \\mathbb { R } ^ { n _ { r } \\times n _ { c } \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 490,
+ 465,
+ 503
+ ],
+ "score": 1.0,
+ "content": ". The size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 492,
+ 476,
+ 501
+ ],
+ "score": 0.73,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 476,
+ 490,
+ 506,
+ 503
+ ],
+ "score": 1.0,
+ "content": "of the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 104,
+ 499,
+ 507,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 499,
+ 195,
+ 514
+ ],
+ "score": 1.0,
+ "content": "convolutional kernels",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 501,
+ 286,
+ 513
+ ],
+ "score": 0.84,
+ "content": "W _ { h . \\cdot } \\in \\mathbb { R } ^ { m \\times \\bar { m } \\times d _ { h } \\times \\bar { d } _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 287,
+ 499,
+ 306,
+ 514
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 306,
+ 501,
+ 321,
+ 513
+ ],
+ "score": 0.81,
+ "content": "W _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 499,
+ 327,
+ 514
+ ],
+ "score": 1.0,
+ "content": "·",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 501,
+ 396,
+ 512
+ ],
+ "score": 0.6,
+ "content": "\\in \\mathbb { R } ^ { m \\times m \\times d _ { h } \\times d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 396,
+ 499,
+ 507,
+ 514
+ ],
+ "score": 1.0,
+ "content": "determines the number of",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 512,
+ 506,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 512,
+ 303,
+ 525
+ ],
+ "score": 1.0,
+ "content": "parameters, which is independent of the grid size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 514,
+ 337,
+ 524
+ ],
+ "score": 0.91,
+ "content": "n _ { r } \\times n _ { c }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 512,
+ 506,
+ 525
+ ],
+ "score": 1.0,
+ "content": ". Earlier, Ranzato et al. (2014) proposed a",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 524,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 524,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "similar RNN variation which uses convolutional layers instead of fully connected layers. The hidden",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 535,
+ 209,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 535,
+ 157,
+ 547
+ ],
+ "score": 1.0,
+ "content": "state at time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 157,
+ 536,
+ 162,
+ 545
+ ],
+ "score": 0.8,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 163,
+ 535,
+ 209,
+ 547
+ ],
+ "score": 1.0,
+ "content": "is given by",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 196,
+ 549,
+ 413,
+ 563
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 196,
+ 549,
+ 413,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 196,
+ 549,
+ 413,
+ 563
+ ],
+ "score": 0.88,
+ "content": "h _ { t } = \\operatorname { t a n h } ( \\sigma ( W _ { x 2 } \\ast \\sigma ( W _ { x 1 } \\ast x _ { t } ) ) + \\sigma ( W _ { h } \\ast h _ { t - 1 } ) ) ,",
+ "type": "interline_equation",
+ "image_path": "c0e6bba61d667cad5d57e19a01b68c8f4baea77f00a566816d97f20d548841cf.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 35,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 196,
+ 549,
+ 413,
+ 563
+ ],
+ "spans": [],
+ "index": 35
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 565,
+ 502,
+ 589
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 235,
+ 579
+ ],
+ "score": 1.0,
+ "content": "where the convolutional kernels",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 565,
+ 294,
+ 577
+ ],
+ "score": 0.93,
+ "content": "W _ { h } \\in \\mathbb { R } ^ { d _ { h } \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 294,
+ 563,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "are restricted to filters of size 1x1 (effectively a fully",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 313,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 313,
+ 590
+ ],
+ "score": 1.0,
+ "content": "connected layer shared across all spatial locations).",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 593,
+ 505,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "Observing that natural language exhibits syntactic properties that naturally combine words into",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "phrases, Tai et al. (2015) proposed a model for tree-structured topologies, where each LSTM has",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 628
+ ],
+ "score": 1.0,
+ "content": "access to the states of its children. They obtained state-of-the-art results on semantic relatedness and",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "sentiment classification. Liang et al. (2016) followed up and proposed a variant on graphs. Their",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "sophisticated network architecture obtained state-of-the-art results for semantic object parsing on",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 506,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 506,
+ 662
+ ],
+ "score": 1.0,
+ "content": "four datasets. In those models, the states are gathered from the neighborhood by way of a weighted",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 673
+ ],
+ "score": 1.0,
+ "content": "sum with trainable weight matrices. Those weights are however not shared across the graph, which",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 684
+ ],
+ "score": 1.0,
+ "content": "would otherwise have required some ordering of the nodes, alike any other spatial definition of graph",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 694
+ ],
+ "score": 1.0,
+ "content": "convolution. Moreover, their formulations are limited to the one-neighborhood of the current node,",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 693,
+ 275,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 693,
+ 275,
+ 705
+ ],
+ "score": 1.0,
+ "content": "with equal weight given to each neighbor.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 42.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "Motivated by spatio-temporal problems like modeling human motion and object interactions, Jain",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 732
+ ],
+ "score": 1.0,
+ "content": "et al. (2016) developed a method to cast a spatio-temporal graph as a rich RNN mixture which",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 48.5
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 12,
+ "width": 8
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 114
+ ],
+ "lines": [],
+ "index": 1,
+ "bbox_fs": [
+ 104,
+ 82,
+ 505,
+ 117
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 257,
+ 111,
+ 353,
+ 147
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 257,
+ 111,
+ 353,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 257,
+ 111,
+ 353,
+ 147
+ ],
+ "score": 0.94,
+ "content": "g _ { \\theta } ( \\Lambda ) = \\sum _ { k = 0 } ^ { K - 1 } \\theta _ { k } T _ { k } ( \\tilde { \\Lambda } ) ,",
+ "type": "interline_equation",
+ "image_path": "ca1fa376dcb83ba222b220186098e686f6ca5b2a4db0060aa601fe75f132bc17.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 257,
+ 111,
+ 353,
+ 129.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 257,
+ 129.0,
+ 353,
+ 147.0
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 148,
+ 505,
+ 183
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 194,
+ 162
+ ],
+ "score": 1.0,
+ "content": "where the parameter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 149,
+ 232,
+ 160
+ ],
+ "score": 0.92,
+ "content": "\\theta \\in \\mathbb { R } ^ { K }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 147,
+ 410,
+ 162
+ ],
+ "score": 1.0,
+ "content": "is a vector of Chebyshev coefficients and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 410,
+ 147,
+ 477,
+ 162
+ ],
+ "score": 0.92,
+ "content": "T _ { k } ( \\tilde { \\Lambda } ) \\ \\in \\ \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 147,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "is the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 505,
+ 176
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 237,
+ 176
+ ],
+ "score": 1.0,
+ "content": "Chebyshev polynomial of order",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 163,
+ 244,
+ 172
+ ],
+ "score": 0.82,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 244,
+ 160,
+ 296,
+ 176
+ ],
+ "score": 1.0,
+ "content": "evaluated at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 161,
+ 380,
+ 174
+ ],
+ "score": 0.92,
+ "content": "\\tilde { \\Lambda } = 2 \\Lambda / \\lambda _ { m a x } - I _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 160,
+ 505,
+ 176
+ ],
+ "score": 1.0,
+ "content": ". The graph filtering operation",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 173,
+ 197,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 173,
+ 197,
+ 185
+ ],
+ "score": 1.0,
+ "content": "can then be written as",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 147,
+ 505,
+ 185
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 220,
+ 180,
+ 390,
+ 216
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 220,
+ 180,
+ 390,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 220,
+ 180,
+ 390,
+ 216
+ ],
+ "score": 0.94,
+ "content": "y = g _ { \\theta } * _ { \\mathcal { G } } x = g _ { \\theta } ( L ) x = \\sum _ { k = 0 } ^ { K - 1 } \\theta _ { k } T _ { k } ( \\tilde { L } ) x ,",
+ "type": "interline_equation",
+ "image_path": "21ebcfd5a78b0aa995a82680ed098591dda94481e43dcd7d572a6e44709702de.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 220,
+ 180,
+ 390,
+ 198.0
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 220,
+ 198.0,
+ 390,
+ 216.0
+ ],
+ "spans": [],
+ "index": 9
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 217,
+ 505,
+ 297
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 217,
+ 506,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 217,
+ 134,
+ 232
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 217,
+ 200,
+ 231
+ ],
+ "score": 0.92,
+ "content": "T _ { k } ( \\tilde { L } ) \\in \\mathbb { R } ^ { n \\times n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 200,
+ 217,
+ 359,
+ 232
+ ],
+ "score": 1.0,
+ "content": "is the Chebyshev polynomial of order",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 359,
+ 219,
+ 366,
+ 228
+ ],
+ "score": 0.82,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 366,
+ 217,
+ 506,
+ 232
+ ],
+ "score": 1.0,
+ "content": "evaluated at the scaled Laplacian",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 107,
+ 227,
+ 507,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 230,
+ 192,
+ 243
+ ],
+ "score": 0.88,
+ "content": "\\tilde { L } = 2 L / \\lambda _ { m a x } - I _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 192,
+ 227,
+ 348,
+ 247
+ ],
+ "score": 1.0,
+ "content": ". Using the stable recurrence relation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 348,
+ 230,
+ 483,
+ 243
+ ],
+ "score": 0.93,
+ "content": "T _ { k } ( x ) = 2 x T _ { k - 1 } ( x ) - T _ { k - 2 } ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 483,
+ 227,
+ 507,
+ 247
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 243,
+ 140,
+ 253
+ ],
+ "score": 0.9,
+ "content": "T _ { 0 } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 240,
+ 159,
+ 255
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 243,
+ 194,
+ 253
+ ],
+ "score": 0.89,
+ "content": "T _ { 1 } = x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 194,
+ 240,
+ 295,
+ 255
+ ],
+ "score": 1.0,
+ "content": ", one can evaluate (5) in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 242,
+ 334,
+ 254
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { O } ( K \\vert \\mathcal { E } \\vert )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 240,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "operations, i.e. linearly with the number",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 353,
+ 266
+ ],
+ "score": 1.0,
+ "content": "of edges. Note that as the filtering operation (5) is an order",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 354,
+ 254,
+ 365,
+ 263
+ ],
+ "score": 0.84,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 253,
+ 506,
+ 266
+ ],
+ "score": 1.0,
+ "content": "polynomial of the Laplacian, it is",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 107,
+ 264,
+ 504,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 264,
+ 117,
+ 274
+ ],
+ "score": 0.84,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 117,
+ 264,
+ 356,
+ 276
+ ],
+ "score": 1.0,
+ "content": "-localized and depends only on nodes that are at maximum",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 356,
+ 264,
+ 367,
+ 274
+ ],
+ "score": 0.79,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 264,
+ 504,
+ 276
+ ],
+ "score": 1.0,
+ "content": "hops away from the central node,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 287
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 122,
+ 287
+ ],
+ "score": 1.0,
+ "content": "the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 122,
+ 275,
+ 132,
+ 285
+ ],
+ "score": 0.84,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 275,
+ 505,
+ 287
+ ],
+ "score": 1.0,
+ "content": "-neighborhood. The reader is referred to Defferrard et al. (2016) for details and an in-depth",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 285,
+ 153,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 285,
+ 153,
+ 298
+ ],
+ "score": 1.0,
+ "content": "discussion.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 105,
+ 217,
+ 507,
+ 298
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 312,
+ 217,
+ 325
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 311,
+ 218,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 311,
+ 218,
+ 328
+ ],
+ "score": 1.0,
+ "content": "3 RELATED WORKS",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 337,
+ 504,
+ 382
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 337,
+ 505,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 337,
+ 505,
+ 350
+ ],
+ "score": 1.0,
+ "content": "Shi et al. (2015) introduced a model for regular grid-structured sequences, which can be seen as",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 361
+ ],
+ "score": 1.0,
+ "content": "a special case of the proposed model where the graph is an image grid where the nodes are well",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 360,
+ 505,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 360,
+ 505,
+ 372
+ ],
+ "score": 1.0,
+ "content": "ordered. Their model is essentially the classical FC-LSTM (2) where the multiplications by dense",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 368,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 142,
+ 382
+ ],
+ "score": 1.0,
+ "content": "matrices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 371,
+ 155,
+ 380
+ ],
+ "score": 0.65,
+ "content": "W",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 370,
+ 352,
+ 382
+ ],
+ "score": 1.0,
+ "content": "have been replaced by convolutions with kernels",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 371,
+ 363,
+ 380
+ ],
+ "score": 0.73,
+ "content": "W",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 370,
+ 368,
+ 382
+ ],
+ "score": 1.0,
+ "content": ":",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19.5,
+ "bbox_fs": [
+ 105,
+ 337,
+ 505,
+ 382
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 185,
+ 384,
+ 426,
+ 456
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 185,
+ 384,
+ 426,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 185,
+ 384,
+ 426,
+ 456
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r l } & { i = \\sigma ( W _ { x i } * x _ { t } + W _ { h i } * h _ { t - 1 } + w _ { c i } \\odot c _ { t - 1 } + b _ { i } ) , } \\\\ & { f = \\sigma ( W _ { x f } * x _ { t } + W _ { h f } * h _ { t - 1 } + w _ { c f } \\odot c _ { t - 1 } + b _ { f } ) , } \\\\ & { c _ { t } = f _ { t } \\odot c _ { t - 1 } + i _ { t } \\odot \\operatorname { t a n h } ( W _ { x c } * x _ { t } + W _ { h c } * h _ { t - 1 } + b _ { c } ) , } \\\\ & { o = \\sigma ( W _ { x o } * x _ { t } + W _ { h o } * h _ { t - 1 } + w _ { c o } \\odot c _ { t } + b _ { o } ) , } \\\\ & { h _ { t } = o \\odot \\operatorname { t a n h } ( c _ { t } ) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "d551ef75a9d0ee8af4623ee5acaffef2323d8292b37a1ef738751acab248e26d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 24,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 185,
+ 384,
+ 426,
+ 398.4
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 185,
+ 398.4,
+ 426,
+ 412.79999999999995
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 185,
+ 412.79999999999995,
+ 426,
+ 427.19999999999993
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 185,
+ 427.19999999999993,
+ 426,
+ 441.5999999999999
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 185,
+ 441.5999999999999,
+ 426,
+ 455.9999999999999
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 457,
+ 506,
+ 546
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 134,
+ 471
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 460,
+ 141,
+ 468
+ ],
+ "score": 0.8,
+ "content": "^ *",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 456,
+ 481,
+ 471
+ ],
+ "score": 1.0,
+ "content": "denotes the 2D convolution by a set of kernels. In their setting, the input tensor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 459,
+ 505,
+ 469
+ ],
+ "score": 0.86,
+ "content": "\\boldsymbol { x } _ { t } ~ \\in",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 107,
+ 466,
+ 506,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 468,
+ 154,
+ 479
+ ],
+ "score": 0.91,
+ "content": "\\mathbb { R } ^ { n _ { r } \\times n _ { c } \\times d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 154,
+ 466,
+ 242,
+ 483
+ ],
+ "score": 1.0,
+ "content": "is the observation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 469,
+ 253,
+ 480
+ ],
+ "score": 0.89,
+ "content": "d _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 254,
+ 466,
+ 347,
+ 483
+ ],
+ "score": 1.0,
+ "content": "measurements at time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 470,
+ 352,
+ 479
+ ],
+ "score": 0.74,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 353,
+ 466,
+ 506,
+ 483
+ ],
+ "score": 1.0,
+ "content": "of a dynamical system over a spatial",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 480,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 238,
+ 492
+ ],
+ "score": 1.0,
+ "content": "region represented by a grid of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 238,
+ 482,
+ 250,
+ 491
+ ],
+ "score": 0.87,
+ "content": "n _ { r }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 480,
+ 293,
+ 492
+ ],
+ "score": 1.0,
+ "content": "rows and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 293,
+ 482,
+ 305,
+ 491
+ ],
+ "score": 0.85,
+ "content": "n _ { c }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 305,
+ 480,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": "columns. The model holds spatially distributed",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 490,
+ 506,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 490,
+ 231,
+ 503
+ ],
+ "score": 1.0,
+ "content": "hidden and cell states of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 491,
+ 243,
+ 502
+ ],
+ "score": 0.87,
+ "content": "d _ { h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 490,
+ 333,
+ 503
+ ],
+ "score": 1.0,
+ "content": "given by the tensors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 333,
+ 490,
+ 418,
+ 501
+ ],
+ "score": 0.87,
+ "content": "c _ { t } , h _ { t } \\ \\in \\ \\mathbb { R } ^ { n _ { r } \\times n _ { c } \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 490,
+ 465,
+ 503
+ ],
+ "score": 1.0,
+ "content": ". The size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 492,
+ 476,
+ 501
+ ],
+ "score": 0.73,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 476,
+ 490,
+ 506,
+ 503
+ ],
+ "score": 1.0,
+ "content": "of the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 104,
+ 499,
+ 507,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 499,
+ 195,
+ 514
+ ],
+ "score": 1.0,
+ "content": "convolutional kernels",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 501,
+ 286,
+ 513
+ ],
+ "score": 0.84,
+ "content": "W _ { h . \\cdot } \\in \\mathbb { R } ^ { m \\times \\bar { m } \\times d _ { h } \\times \\bar { d } _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 287,
+ 499,
+ 306,
+ 514
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 306,
+ 501,
+ 321,
+ 513
+ ],
+ "score": 0.81,
+ "content": "W _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 321,
+ 499,
+ 327,
+ 514
+ ],
+ "score": 1.0,
+ "content": "·",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 501,
+ 396,
+ 512
+ ],
+ "score": 0.6,
+ "content": "\\in \\mathbb { R } ^ { m \\times m \\times d _ { h } \\times d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 396,
+ 499,
+ 507,
+ 514
+ ],
+ "score": 1.0,
+ "content": "determines the number of",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 512,
+ 506,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 512,
+ 303,
+ 525
+ ],
+ "score": 1.0,
+ "content": "parameters, which is independent of the grid size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 514,
+ 337,
+ 524
+ ],
+ "score": 0.91,
+ "content": "n _ { r } \\times n _ { c }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 512,
+ 506,
+ 525
+ ],
+ "score": 1.0,
+ "content": ". Earlier, Ranzato et al. (2014) proposed a",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 524,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 524,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "similar RNN variation which uses convolutional layers instead of fully connected layers. The hidden",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 535,
+ 209,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 535,
+ 157,
+ 547
+ ],
+ "score": 1.0,
+ "content": "state at time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 157,
+ 536,
+ 162,
+ 545
+ ],
+ "score": 0.8,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 163,
+ 535,
+ 209,
+ 547
+ ],
+ "score": 1.0,
+ "content": "is given by",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 104,
+ 456,
+ 507,
+ 547
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 196,
+ 549,
+ 413,
+ 563
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 196,
+ 549,
+ 413,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 196,
+ 549,
+ 413,
+ 563
+ ],
+ "score": 0.88,
+ "content": "h _ { t } = \\operatorname { t a n h } ( \\sigma ( W _ { x 2 } \\ast \\sigma ( W _ { x 1 } \\ast x _ { t } ) ) + \\sigma ( W _ { h } \\ast h _ { t - 1 } ) ) ,",
+ "type": "interline_equation",
+ "image_path": "c0e6bba61d667cad5d57e19a01b68c8f4baea77f00a566816d97f20d548841cf.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 35,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 196,
+ 549,
+ 413,
+ 563
+ ],
+ "spans": [],
+ "index": 35
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 565,
+ 502,
+ 589
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 563,
+ 235,
+ 579
+ ],
+ "score": 1.0,
+ "content": "where the convolutional kernels",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 565,
+ 294,
+ 577
+ ],
+ "score": 0.93,
+ "content": "W _ { h } \\in \\mathbb { R } ^ { d _ { h } \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 294,
+ 563,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "are restricted to filters of size 1x1 (effectively a fully",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 313,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 313,
+ 590
+ ],
+ "score": 1.0,
+ "content": "connected layer shared across all spatial locations).",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 105,
+ 563,
+ 505,
+ 590
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 593,
+ 505,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "Observing that natural language exhibits syntactic properties that naturally combine words into",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "phrases, Tai et al. (2015) proposed a model for tree-structured topologies, where each LSTM has",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 628
+ ],
+ "score": 1.0,
+ "content": "access to the states of its children. They obtained state-of-the-art results on semantic relatedness and",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "sentiment classification. Liang et al. (2016) followed up and proposed a variant on graphs. Their",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "sophisticated network architecture obtained state-of-the-art results for semantic object parsing on",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 506,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 506,
+ 662
+ ],
+ "score": 1.0,
+ "content": "four datasets. In those models, the states are gathered from the neighborhood by way of a weighted",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 673
+ ],
+ "score": 1.0,
+ "content": "sum with trainable weight matrices. Those weights are however not shared across the graph, which",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 684
+ ],
+ "score": 1.0,
+ "content": "would otherwise have required some ordering of the nodes, alike any other spatial definition of graph",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 681,
+ 505,
+ 694
+ ],
+ "score": 1.0,
+ "content": "convolution. Moreover, their formulations are limited to the one-neighborhood of the current node,",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 693,
+ 275,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 693,
+ 275,
+ 705
+ ],
+ "score": 1.0,
+ "content": "with equal weight given to each neighbor.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 105,
+ 593,
+ 506,
+ 705
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "Motivated by spatio-temporal problems like modeling human motion and object interactions, Jain",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 732
+ ],
+ "score": 1.0,
+ "content": "et al. (2016) developed a method to cast a spatio-temporal graph as a rich RNN mixture which",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 82,
+ 504,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 504,
+ 95
+ ],
+ "score": 1.0,
+ "content": "essentially associates a RNN to each node and edge. Again, the communication is limited to directly",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 220,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 220,
+ 107
+ ],
+ "score": 1.0,
+ "content": "connected nodes and edges.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 48.5,
+ "bbox_fs": [
+ 105,
+ 709,
+ 505,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 504,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 504,
+ 95
+ ],
+ "score": 1.0,
+ "content": "essentially associates a RNN to each node and edge. Again, the communication is limited to directly",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 220,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 220,
+ 107
+ ],
+ "score": 1.0,
+ "content": "connected nodes and edges.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 505,
+ 187
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 111,
+ 505,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 111,
+ 505,
+ 123
+ ],
+ "score": 1.0,
+ "content": "The closest model to our work is probably the one proposed by Li et al. (2015), which showed",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 133
+ ],
+ "score": 1.0,
+ "content": "stat-of-the-art performance on a problem from program verification. Whereas they use the iterative",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "procedure of the Graph Neural Networks (GNNs) model introduced by Scarselli et al. (2009) to",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "propagate node representations until convergence, we instead use the graph CNN introduced by",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 167
+ ],
+ "score": 1.0,
+ "content": "Defferrard et al. (2016) to diffuse information across the nodes. While their motivations are quite",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 178
+ ],
+ "score": 1.0,
+ "content": "different, those models are related by the fact that a spectral filter defined as a polynomial of order",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 107,
+ 175,
+ 285,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 177,
+ 117,
+ 186
+ ],
+ "score": 0.76,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 118,
+ 175,
+ 218,
+ 188
+ ],
+ "score": 1.0,
+ "content": "can be implemented as a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 177,
+ 229,
+ 186
+ ],
+ "score": 0.79,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 229,
+ 175,
+ 285,
+ 188
+ ],
+ "score": 1.0,
+ "content": "-layer GNN.2",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 203,
+ 267,
+ 216
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 268,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 268,
+ 219
+ ],
+ "score": 1.0,
+ "content": "4 PROPOSED GCRN MODELS",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 228,
+ 504,
+ 250
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 506,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 506,
+ 241
+ ],
+ "score": 1.0,
+ "content": "We propose two GCRN architectures that are quite natural, and investigate their performances in",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 253,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 253,
+ 250
+ ],
+ "score": 1.0,
+ "content": "real-world applications in Section 5.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 103,
+ 262,
+ 504,
+ 285
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 262,
+ 505,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 262,
+ 505,
+ 275
+ ],
+ "score": 1.0,
+ "content": "Model 1. The most straightforward definition is to stack a graph CNN, defined as (5), for feature",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 272,
+ 359,
+ 287
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 359,
+ 287
+ ],
+ "score": 1.0,
+ "content": "extraction and an LSTM, defined as (2), for sequence learning:",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 181,
+ 289,
+ 428,
+ 383
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 181,
+ 289,
+ 428,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 181,
+ 289,
+ 428,
+ 383
+ ],
+ "score": 0.95,
+ "content": "\\begin{array} { r l } & { x _ { t } ^ { \\mathrm { { C N N } } } = { \\mathrm { C N N } } _ { \\mathscr { G } } ( x _ { t } ) } \\\\ & { \\qquad i = \\sigma \\big ( { W _ { x i } } { x _ { t } ^ { \\mathrm { { C N N } } } } + { W _ { h i } } { h _ { t - 1 } } + { w _ { c i } } \\odot { c _ { t - 1 } } + { b _ { i } } \\big ) , } \\\\ & { f = \\sigma \\big ( { W _ { x f } } { x _ { t } ^ { { \\mathrm { C N N } } } } + { W _ { h f } } { h _ { t - 1 } } + { w _ { c f } } \\odot { c _ { t - 1 } } + { b _ { f } } \\big ) , } \\\\ & { c _ { t } = f _ { t } \\odot c _ { t - 1 } + i _ { t } \\odot { \\mathrm { t a n h } } ( { W _ { x c } } { x _ { t } ^ { \\mathrm { { C N N } } } } + { W _ { h c } } { h _ { t - 1 } } + { b _ { c } } \\big ) , } \\\\ & { o = \\sigma \\big ( { W _ { x o } } { x _ { t } ^ { { \\mathrm { C N N } } } } + { W _ { h o } } { h _ { t - 1 } } + { w _ { c o } } \\odot { c _ { t } } + { b _ { o } } \\big ) , } \\\\ & { h _ { t } = o \\odot { \\mathrm { t a n h } } \\big ( { c _ { t } } \\big ) . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "48c6ec5b9a58b44fa5b53070dc92ffee5ebe8c077532e457a3ac55ac4539eea6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 181,
+ 289,
+ 428,
+ 320.3333333333333
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 181,
+ 320.3333333333333,
+ 428,
+ 351.66666666666663
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 181,
+ 351.66666666666663,
+ 428,
+ 382.99999999999994
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 387,
+ 505,
+ 487
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 384,
+ 507,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 384,
+ 235,
+ 402
+ ],
+ "score": 1.0,
+ "content": "In that setting, the input matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 387,
+ 285,
+ 399
+ ],
+ "score": 0.92,
+ "content": "\\boldsymbol { x } _ { t } \\in \\mathbb { R } ^ { n \\times d _ { \\boldsymbol { x } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 286,
+ 384,
+ 422,
+ 402
+ ],
+ "score": 1.0,
+ "content": "may represent the observation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 388,
+ 434,
+ 399
+ ],
+ "score": 0.88,
+ "content": "d _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 384,
+ 507,
+ 402
+ ],
+ "score": 1.0,
+ "content": "measurements at",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 104,
+ 395,
+ 507,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 395,
+ 126,
+ 413
+ ],
+ "score": 1.0,
+ "content": "time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 127,
+ 399,
+ 132,
+ 408
+ ],
+ "score": 0.74,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 395,
+ 445,
+ 413
+ ],
+ "score": 1.0,
+ "content": "of a dynamical system over a network whose organization is given by a graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 446,
+ 399,
+ 453,
+ 409
+ ],
+ "score": 0.59,
+ "content": "\\mathcal { G }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 454,
+ 395,
+ 457,
+ 413
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 458,
+ 398,
+ 480,
+ 409
+ ],
+ "score": 0.68,
+ "content": "\\boldsymbol { x } _ { t } ^ { \\mathrm { C N N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 480,
+ 395,
+ 507,
+ 413
+ ],
+ "score": 1.0,
+ "content": "is the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 102,
+ 407,
+ 506,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 407,
+ 416,
+ 426
+ ],
+ "score": 1.0,
+ "content": "output of the graph CNN gate. For a proof of concept, we simply choose here",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 417,
+ 409,
+ 502,
+ 421
+ ],
+ "score": 0.89,
+ "content": "\\begin{array} { r } { x _ { t } ^ { \\mathrm { { C N N } } } = W ^ { \\mathrm { { C N N } } } * _ { \\mathcal { G } } x _ { t } . } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 407,
+ 506,
+ 426
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 104,
+ 417,
+ 507,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 417,
+ 134,
+ 434
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 420,
+ 220,
+ 431
+ ],
+ "score": 0.92,
+ "content": "W ^ { \\mathrm { C N N } } \\in \\mathbf { \\bar { \\Gamma } } \\mathbb { R } ^ { K \\times d _ { x } \\times \\mathbf { \\bar { d } } _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 220,
+ 417,
+ 507,
+ 434
+ ],
+ "score": 1.0,
+ "content": "are the Chebyshev coefficients for the graph convolutional kernels of",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 431,
+ 506,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 139,
+ 444
+ ],
+ "score": 1.0,
+ "content": "support",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 139,
+ 432,
+ 149,
+ 441
+ ],
+ "score": 0.78,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 431,
+ 440,
+ 444
+ ],
+ "score": 1.0,
+ "content": ". The model also holds spatially distributed hidden and cell states of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 441,
+ 432,
+ 452,
+ 442
+ ],
+ "score": 0.87,
+ "content": "d _ { h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 431,
+ 506,
+ 444
+ ],
+ "score": 1.0,
+ "content": "given by the",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 440,
+ 503,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 440,
+ 143,
+ 455
+ ],
+ "score": 1.0,
+ "content": "matrices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 144,
+ 442,
+ 206,
+ 453
+ ],
+ "score": 0.9,
+ "content": "c _ { t } , h _ { t } \\in \\mathbb { R } ^ { n \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 207,
+ 440,
+ 328,
+ 455
+ ],
+ "score": 1.0,
+ "content": ". Peepholes are controlled by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 442,
+ 382,
+ 453
+ ],
+ "score": 0.89,
+ "content": "w _ { c } . \\in \\mathbb { R } ^ { n \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 440,
+ 440,
+ 455
+ ],
+ "score": 1.0,
+ "content": ". The weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 446,
+ 442,
+ 503,
+ 453
+ ],
+ "score": 0.82,
+ "content": "V _ { h \\cdot } \\in \\mathbb { R } ^ { d _ { h } \\times d _ { h } }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 506,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 124,
+ 467
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 124,
+ 453,
+ 139,
+ 464
+ ],
+ "score": 0.81,
+ "content": "W _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 451,
+ 144,
+ 467
+ ],
+ "score": 1.0,
+ "content": "·",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 144,
+ 453,
+ 187,
+ 464
+ ],
+ "score": 0.74,
+ "content": "\\mathbf { \\Sigma } \\in \\mathbb { R } ^ { d _ { h } \\times d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 451,
+ 506,
+ 467
+ ],
+ "score": 1.0,
+ "content": "are the parameters of the fully connected layers. An architecture such as (8)",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 478
+ ],
+ "score": 1.0,
+ "content": "may be enough to capture the data distribution by exploiting local stationarity and compositionality",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 286,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 286,
+ 488
+ ],
+ "score": 1.0,
+ "content": "properties as well as the dynamic properties.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 498,
+ 503,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 504,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 504,
+ 511
+ ],
+ "score": 1.0,
+ "content": "Model 2. To generalize the convLSTM model (6) to graphs we replace the Euclidean 2D convo-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 508,
+ 257,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 132,
+ 523
+ ],
+ "score": 1.0,
+ "content": "lution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 511,
+ 139,
+ 519
+ ],
+ "score": 0.73,
+ "content": "^ *",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 139,
+ 508,
+ 241,
+ 523
+ ],
+ "score": 1.0,
+ "content": "by the graph convolution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 511,
+ 253,
+ 522
+ ],
+ "score": 0.84,
+ "content": "^ { \\ast _ { \\mathcal { G } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 508,
+ 257,
+ 523
+ ],
+ "score": 1.0,
+ "content": ":",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 179,
+ 524,
+ 432,
+ 596
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 179,
+ 524,
+ 432,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 179,
+ 524,
+ 432,
+ 596
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r l } & { i = \\sigma ( W _ { x i } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h i } \\ast _ { \\mathcal { G } } h _ { t - 1 } + w _ { c i } \\odot c _ { t - 1 } + b _ { i } ) , } \\\\ & { f = \\sigma ( W _ { x f } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h f } \\ast _ { \\mathcal { G } } h _ { t - 1 } + w _ { c f } \\odot c _ { t - 1 } + b _ { f } ) , } \\\\ & { c _ { t } = f _ { t } \\odot c _ { t - 1 } + i _ { t } \\odot \\operatorname { t a n h } ( W _ { x c } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h c } \\ast _ { \\mathcal { G } } h _ { t - 1 } + b _ { c } ) , } \\\\ & { o = \\sigma ( W _ { x o } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h o } \\ast _ { \\mathcal { G } } h _ { t - 1 } + w _ { c o } \\odot c _ { t } + b _ { o } ) , } \\\\ & { h _ { t } = o \\odot \\operatorname { t a n h } ( c _ { t } ) . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "7096a1d24cce9a1021dc9d187cb4106d4d0a0beb1b8000df88a46edc086de9a8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 29,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 179,
+ 524,
+ 432,
+ 548.0
+ ],
+ "spans": [],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 179,
+ 548.0,
+ 432,
+ 572.0
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 179,
+ 572.0,
+ 432,
+ 596.0
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 598,
+ 505,
+ 675
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 505,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 216,
+ 610
+ ],
+ "score": 1.0,
+ "content": "In that setting, the support",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 216,
+ 599,
+ 227,
+ 608
+ ],
+ "score": 0.85,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 598,
+ 505,
+ 610
+ ],
+ "score": 1.0,
+ "content": "of the graph convolutional kernels defined by the Chebyshev coeffi-",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 104,
+ 605,
+ 507,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 605,
+ 133,
+ 623
+ ],
+ "score": 1.0,
+ "content": "cients",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 609,
+ 148,
+ 620
+ ],
+ "score": 0.81,
+ "content": "W _ { h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 149,
+ 605,
+ 154,
+ 623
+ ],
+ "score": 1.0,
+ "content": "·",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 608,
+ 210,
+ 620
+ ],
+ "score": 0.83,
+ "content": "\\doteq \\mathbb { R } ^ { K \\times d _ { h } ^ { \\bullet } \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 211,
+ 605,
+ 231,
+ 623
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 609,
+ 246,
+ 620
+ ],
+ "score": 0.84,
+ "content": "W _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 605,
+ 252,
+ 623
+ ],
+ "score": 1.0,
+ "content": "·",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 608,
+ 308,
+ 619
+ ],
+ "score": 0.84,
+ "content": "\\breve { \\in \\mathbb { R } ^ { K \\times d _ { h } \\times d _ { x } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 605,
+ 507,
+ 623
+ ],
+ "score": 1.0,
+ "content": "determines the number of parameters, which is",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 253,
+ 633
+ ],
+ "score": 1.0,
+ "content": "independent of the number of nodes",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 622,
+ 261,
+ 630
+ ],
+ "score": 0.76,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 618,
+ 419,
+ 633
+ ],
+ "score": 1.0,
+ "content": ". To keep the notation simple, we write",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 620,
+ 462,
+ 632
+ ],
+ "score": 0.92,
+ "content": "W _ { x i } * _ { \\mathcal { G } } x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 618,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "to mean a",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 631,
+ 506,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 631,
+ 191,
+ 644
+ ],
+ "score": 1.0,
+ "content": "graph convolution of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 191,
+ 632,
+ 201,
+ 642
+ ],
+ "score": 0.84,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 202,
+ 631,
+ 222,
+ 644
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 631,
+ 244,
+ 642
+ ],
+ "score": 0.91,
+ "content": "d _ { h } d _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 244,
+ 631,
+ 441,
+ 644
+ ],
+ "score": 1.0,
+ "content": "filters which are functions of the graph Laplacian",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 442,
+ 632,
+ 450,
+ 641
+ ],
+ "score": 0.79,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 631,
+ 506,
+ 644
+ ],
+ "score": 1.0,
+ "content": "parametrized",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 642,
+ 505,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 642,
+ 119,
+ 654
+ ],
+ "score": 1.0,
+ "content": "by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 119,
+ 642,
+ 129,
+ 652
+ ],
+ "score": 0.8,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 130,
+ 642,
+ 459,
+ 654
+ ],
+ "score": 1.0,
+ "content": "Chebyshev coefficients, as noted in (4) and (5). In a distributed computing setting,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 459,
+ 642,
+ 469,
+ 652
+ ],
+ "score": 0.8,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 642,
+ 505,
+ 654
+ ],
+ "score": 1.0,
+ "content": "controls",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 397,
+ 666
+ ],
+ "score": 1.0,
+ "content": "the communication overhead, i.e. the number of nodes any given node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 654,
+ 402,
+ 663
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 403,
+ 653,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "should exchange with in",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 664,
+ 238,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 664,
+ 238,
+ 676
+ ],
+ "score": 1.0,
+ "content": "order to compute its local states.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 680,
+ 503,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 505,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 505,
+ 693
+ ],
+ "score": 1.0,
+ "content": "The proposed blend of RNNs and graph CNNs is not limited to LSTMs and is straightforward to",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 691,
+ 504,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 691,
+ 383,
+ 706
+ ],
+ "score": 1.0,
+ "content": "apply to any kind of recursive networks. For example, a vanilla RNN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 384,
+ 692,
+ 504,
+ 704
+ ],
+ "score": 0.88,
+ "content": "h _ { t } = \\operatorname { t a n h } ( W _ { x } x _ { t } + W _ { h } h _ { t - 1 } )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38.5
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 711,
+ 506,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 709,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 709,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "2The basic idea is to set the transition function as a diffusion and the output function such as to realize the",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 417,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 230,
+ 732
+ ],
+ "score": 1.0,
+ "content": "polynomial recurrence, then stack",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 722,
+ 240,
+ 730
+ ],
+ "score": 0.8,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 240,
+ 721,
+ 417,
+ 732
+ ],
+ "score": 1.0,
+ "content": "of those. See Defferrard et al. (2016) for details.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [],
+ "index": 0.5,
+ "bbox_fs": [
+ 106,
+ 82,
+ 504,
+ 107
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 505,
+ 187
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 111,
+ 505,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 111,
+ 505,
+ 123
+ ],
+ "score": 1.0,
+ "content": "The closest model to our work is probably the one proposed by Li et al. (2015), which showed",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 133
+ ],
+ "score": 1.0,
+ "content": "stat-of-the-art performance on a problem from program verification. Whereas they use the iterative",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "procedure of the Graph Neural Networks (GNNs) model introduced by Scarselli et al. (2009) to",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 156
+ ],
+ "score": 1.0,
+ "content": "propagate node representations until convergence, we instead use the graph CNN introduced by",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 167
+ ],
+ "score": 1.0,
+ "content": "Defferrard et al. (2016) to diffuse information across the nodes. While their motivations are quite",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 178
+ ],
+ "score": 1.0,
+ "content": "different, those models are related by the fact that a spectral filter defined as a polynomial of order",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 107,
+ 175,
+ 285,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 177,
+ 117,
+ 186
+ ],
+ "score": 0.76,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 118,
+ 175,
+ 218,
+ 188
+ ],
+ "score": 1.0,
+ "content": "can be implemented as a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 177,
+ 229,
+ 186
+ ],
+ "score": 0.79,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 229,
+ 175,
+ 285,
+ 188
+ ],
+ "score": 1.0,
+ "content": "-layer GNN.2",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 105,
+ 111,
+ 506,
+ 188
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 203,
+ 267,
+ 216
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 268,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 268,
+ 219
+ ],
+ "score": 1.0,
+ "content": "4 PROPOSED GCRN MODELS",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 228,
+ 504,
+ 250
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 506,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 506,
+ 241
+ ],
+ "score": 1.0,
+ "content": "We propose two GCRN architectures that are quite natural, and investigate their performances in",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 253,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 253,
+ 250
+ ],
+ "score": 1.0,
+ "content": "real-world applications in Section 5.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 105,
+ 227,
+ 506,
+ 250
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 103,
+ 262,
+ 504,
+ 285
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 262,
+ 505,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 262,
+ 505,
+ 275
+ ],
+ "score": 1.0,
+ "content": "Model 1. The most straightforward definition is to stack a graph CNN, defined as (5), for feature",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 272,
+ 359,
+ 287
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 359,
+ 287
+ ],
+ "score": 1.0,
+ "content": "extraction and an LSTM, defined as (2), for sequence learning:",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 105,
+ 262,
+ 505,
+ 287
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 181,
+ 289,
+ 428,
+ 383
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 181,
+ 289,
+ 428,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 181,
+ 289,
+ 428,
+ 383
+ ],
+ "score": 0.95,
+ "content": "\\begin{array} { r l } & { x _ { t } ^ { \\mathrm { { C N N } } } = { \\mathrm { C N N } } _ { \\mathscr { G } } ( x _ { t } ) } \\\\ & { \\qquad i = \\sigma \\big ( { W _ { x i } } { x _ { t } ^ { \\mathrm { { C N N } } } } + { W _ { h i } } { h _ { t - 1 } } + { w _ { c i } } \\odot { c _ { t - 1 } } + { b _ { i } } \\big ) , } \\\\ & { f = \\sigma \\big ( { W _ { x f } } { x _ { t } ^ { { \\mathrm { C N N } } } } + { W _ { h f } } { h _ { t - 1 } } + { w _ { c f } } \\odot { c _ { t - 1 } } + { b _ { f } } \\big ) , } \\\\ & { c _ { t } = f _ { t } \\odot c _ { t - 1 } + i _ { t } \\odot { \\mathrm { t a n h } } ( { W _ { x c } } { x _ { t } ^ { \\mathrm { { C N N } } } } + { W _ { h c } } { h _ { t - 1 } } + { b _ { c } } \\big ) , } \\\\ & { o = \\sigma \\big ( { W _ { x o } } { x _ { t } ^ { { \\mathrm { C N N } } } } + { W _ { h o } } { h _ { t - 1 } } + { w _ { c o } } \\odot { c _ { t } } + { b _ { o } } \\big ) , } \\\\ & { h _ { t } = o \\odot { \\mathrm { t a n h } } \\big ( { c _ { t } } \\big ) . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "48c6ec5b9a58b44fa5b53070dc92ffee5ebe8c077532e457a3ac55ac4539eea6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 181,
+ 289,
+ 428,
+ 320.3333333333333
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 181,
+ 320.3333333333333,
+ 428,
+ 351.66666666666663
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 181,
+ 351.66666666666663,
+ 428,
+ 382.99999999999994
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 387,
+ 505,
+ 487
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 384,
+ 507,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 384,
+ 235,
+ 402
+ ],
+ "score": 1.0,
+ "content": "In that setting, the input matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 387,
+ 285,
+ 399
+ ],
+ "score": 0.92,
+ "content": "\\boldsymbol { x } _ { t } \\in \\mathbb { R } ^ { n \\times d _ { \\boldsymbol { x } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 286,
+ 384,
+ 422,
+ 402
+ ],
+ "score": 1.0,
+ "content": "may represent the observation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 388,
+ 434,
+ 399
+ ],
+ "score": 0.88,
+ "content": "d _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 384,
+ 507,
+ 402
+ ],
+ "score": 1.0,
+ "content": "measurements at",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 104,
+ 395,
+ 507,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 395,
+ 126,
+ 413
+ ],
+ "score": 1.0,
+ "content": "time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 127,
+ 399,
+ 132,
+ 408
+ ],
+ "score": 0.74,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 395,
+ 445,
+ 413
+ ],
+ "score": 1.0,
+ "content": "of a dynamical system over a network whose organization is given by a graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 446,
+ 399,
+ 453,
+ 409
+ ],
+ "score": 0.59,
+ "content": "\\mathcal { G }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 454,
+ 395,
+ 457,
+ 413
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 458,
+ 398,
+ 480,
+ 409
+ ],
+ "score": 0.68,
+ "content": "\\boldsymbol { x } _ { t } ^ { \\mathrm { C N N } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 480,
+ 395,
+ 507,
+ 413
+ ],
+ "score": 1.0,
+ "content": "is the",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 102,
+ 407,
+ 506,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 407,
+ 416,
+ 426
+ ],
+ "score": 1.0,
+ "content": "output of the graph CNN gate. For a proof of concept, we simply choose here",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 417,
+ 409,
+ 502,
+ 421
+ ],
+ "score": 0.89,
+ "content": "\\begin{array} { r } { x _ { t } ^ { \\mathrm { { C N N } } } = W ^ { \\mathrm { { C N N } } } * _ { \\mathcal { G } } x _ { t } . } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 407,
+ 506,
+ 426
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 104,
+ 417,
+ 507,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 417,
+ 134,
+ 434
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 420,
+ 220,
+ 431
+ ],
+ "score": 0.92,
+ "content": "W ^ { \\mathrm { C N N } } \\in \\mathbf { \\bar { \\Gamma } } \\mathbb { R } ^ { K \\times d _ { x } \\times \\mathbf { \\bar { d } } _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 220,
+ 417,
+ 507,
+ 434
+ ],
+ "score": 1.0,
+ "content": "are the Chebyshev coefficients for the graph convolutional kernels of",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 431,
+ 506,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 139,
+ 444
+ ],
+ "score": 1.0,
+ "content": "support",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 139,
+ 432,
+ 149,
+ 441
+ ],
+ "score": 0.78,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 431,
+ 440,
+ 444
+ ],
+ "score": 1.0,
+ "content": ". The model also holds spatially distributed hidden and cell states of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 441,
+ 432,
+ 452,
+ 442
+ ],
+ "score": 0.87,
+ "content": "d _ { h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 431,
+ 506,
+ 444
+ ],
+ "score": 1.0,
+ "content": "given by the",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 440,
+ 503,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 440,
+ 143,
+ 455
+ ],
+ "score": 1.0,
+ "content": "matrices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 144,
+ 442,
+ 206,
+ 453
+ ],
+ "score": 0.9,
+ "content": "c _ { t } , h _ { t } \\in \\mathbb { R } ^ { n \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 207,
+ 440,
+ 328,
+ 455
+ ],
+ "score": 1.0,
+ "content": ". Peepholes are controlled by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 442,
+ 382,
+ 453
+ ],
+ "score": 0.89,
+ "content": "w _ { c } . \\in \\mathbb { R } ^ { n \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 440,
+ 440,
+ 455
+ ],
+ "score": 1.0,
+ "content": ". The weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 446,
+ 442,
+ 503,
+ 453
+ ],
+ "score": 0.82,
+ "content": "V _ { h \\cdot } \\in \\mathbb { R } ^ { d _ { h } \\times d _ { h } }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 506,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 124,
+ 467
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 124,
+ 453,
+ 139,
+ 464
+ ],
+ "score": 0.81,
+ "content": "W _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 451,
+ 144,
+ 467
+ ],
+ "score": 1.0,
+ "content": "·",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 144,
+ 453,
+ 187,
+ 464
+ ],
+ "score": 0.74,
+ "content": "\\mathbf { \\Sigma } \\in \\mathbb { R } ^ { d _ { h } \\times d _ { x } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 451,
+ 506,
+ 467
+ ],
+ "score": 1.0,
+ "content": "are the parameters of the fully connected layers. An architecture such as (8)",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 478
+ ],
+ "score": 1.0,
+ "content": "may be enough to capture the data distribution by exploiting local stationarity and compositionality",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 286,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 286,
+ 488
+ ],
+ "score": 1.0,
+ "content": "properties as well as the dynamic properties.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 102,
+ 384,
+ 507,
+ 488
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 498,
+ 503,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 504,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 504,
+ 511
+ ],
+ "score": 1.0,
+ "content": "Model 2. To generalize the convLSTM model (6) to graphs we replace the Euclidean 2D convo-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 508,
+ 257,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 132,
+ 523
+ ],
+ "score": 1.0,
+ "content": "lution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 511,
+ 139,
+ 519
+ ],
+ "score": 0.73,
+ "content": "^ *",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 139,
+ 508,
+ 241,
+ 523
+ ],
+ "score": 1.0,
+ "content": "by the graph convolution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 511,
+ 253,
+ 522
+ ],
+ "score": 0.84,
+ "content": "^ { \\ast _ { \\mathcal { G } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 508,
+ 257,
+ 523
+ ],
+ "score": 1.0,
+ "content": ":",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 497,
+ 504,
+ 523
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 179,
+ 524,
+ 432,
+ 596
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 179,
+ 524,
+ 432,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 179,
+ 524,
+ 432,
+ 596
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r l } & { i = \\sigma ( W _ { x i } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h i } \\ast _ { \\mathcal { G } } h _ { t - 1 } + w _ { c i } \\odot c _ { t - 1 } + b _ { i } ) , } \\\\ & { f = \\sigma ( W _ { x f } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h f } \\ast _ { \\mathcal { G } } h _ { t - 1 } + w _ { c f } \\odot c _ { t - 1 } + b _ { f } ) , } \\\\ & { c _ { t } = f _ { t } \\odot c _ { t - 1 } + i _ { t } \\odot \\operatorname { t a n h } ( W _ { x c } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h c } \\ast _ { \\mathcal { G } } h _ { t - 1 } + b _ { c } ) , } \\\\ & { o = \\sigma ( W _ { x o } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h o } \\ast _ { \\mathcal { G } } h _ { t - 1 } + w _ { c o } \\odot c _ { t } + b _ { o } ) , } \\\\ & { h _ { t } = o \\odot \\operatorname { t a n h } ( c _ { t } ) . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "7096a1d24cce9a1021dc9d187cb4106d4d0a0beb1b8000df88a46edc086de9a8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 29,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 179,
+ 524,
+ 432,
+ 548.0
+ ],
+ "spans": [],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 179,
+ 548.0,
+ 432,
+ 572.0
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 179,
+ 572.0,
+ 432,
+ 596.0
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 598,
+ 505,
+ 675
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 505,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 216,
+ 610
+ ],
+ "score": 1.0,
+ "content": "In that setting, the support",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 216,
+ 599,
+ 227,
+ 608
+ ],
+ "score": 0.85,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 598,
+ 505,
+ 610
+ ],
+ "score": 1.0,
+ "content": "of the graph convolutional kernels defined by the Chebyshev coeffi-",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 104,
+ 605,
+ 507,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 605,
+ 133,
+ 623
+ ],
+ "score": 1.0,
+ "content": "cients",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 609,
+ 148,
+ 620
+ ],
+ "score": 0.81,
+ "content": "W _ { h }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 149,
+ 605,
+ 154,
+ 623
+ ],
+ "score": 1.0,
+ "content": "·",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 608,
+ 210,
+ 620
+ ],
+ "score": 0.83,
+ "content": "\\doteq \\mathbb { R } ^ { K \\times d _ { h } ^ { \\bullet } \\times d _ { h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 211,
+ 605,
+ 231,
+ 623
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 609,
+ 246,
+ 620
+ ],
+ "score": 0.84,
+ "content": "W _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 605,
+ 252,
+ 623
+ ],
+ "score": 1.0,
+ "content": "·",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 608,
+ 308,
+ 619
+ ],
+ "score": 0.84,
+ "content": "\\breve { \\in \\mathbb { R } ^ { K \\times d _ { h } \\times d _ { x } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 605,
+ 507,
+ 623
+ ],
+ "score": 1.0,
+ "content": "determines the number of parameters, which is",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 253,
+ 633
+ ],
+ "score": 1.0,
+ "content": "independent of the number of nodes",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 622,
+ 261,
+ 630
+ ],
+ "score": 0.76,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 618,
+ 419,
+ 633
+ ],
+ "score": 1.0,
+ "content": ". To keep the notation simple, we write",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 620,
+ 462,
+ 632
+ ],
+ "score": 0.92,
+ "content": "W _ { x i } * _ { \\mathcal { G } } x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 618,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "to mean a",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 631,
+ 506,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 631,
+ 191,
+ 644
+ ],
+ "score": 1.0,
+ "content": "graph convolution of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 191,
+ 632,
+ 201,
+ 642
+ ],
+ "score": 0.84,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 202,
+ 631,
+ 222,
+ 644
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 631,
+ 244,
+ 642
+ ],
+ "score": 0.91,
+ "content": "d _ { h } d _ { x }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 244,
+ 631,
+ 441,
+ 644
+ ],
+ "score": 1.0,
+ "content": "filters which are functions of the graph Laplacian",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 442,
+ 632,
+ 450,
+ 641
+ ],
+ "score": 0.79,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 631,
+ 506,
+ 644
+ ],
+ "score": 1.0,
+ "content": "parametrized",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 642,
+ 505,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 642,
+ 119,
+ 654
+ ],
+ "score": 1.0,
+ "content": "by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 119,
+ 642,
+ 129,
+ 652
+ ],
+ "score": 0.8,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 130,
+ 642,
+ 459,
+ 654
+ ],
+ "score": 1.0,
+ "content": "Chebyshev coefficients, as noted in (4) and (5). In a distributed computing setting,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 459,
+ 642,
+ 469,
+ 652
+ ],
+ "score": 0.8,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 642,
+ 505,
+ 654
+ ],
+ "score": 1.0,
+ "content": "controls",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 397,
+ 666
+ ],
+ "score": 1.0,
+ "content": "the communication overhead, i.e. the number of nodes any given node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 654,
+ 402,
+ 663
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 403,
+ 653,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "should exchange with in",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 664,
+ 238,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 664,
+ 238,
+ 676
+ ],
+ "score": 1.0,
+ "content": "order to compute its local states.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 104,
+ 598,
+ 507,
+ 676
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 680,
+ 503,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 505,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 505,
+ 693
+ ],
+ "score": 1.0,
+ "content": "The proposed blend of RNNs and graph CNNs is not limited to LSTMs and is straightforward to",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 691,
+ 504,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 691,
+ 383,
+ 706
+ ],
+ "score": 1.0,
+ "content": "apply to any kind of recursive networks. For example, a vanilla RNN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 384,
+ 692,
+ 504,
+ 704
+ ],
+ "score": 0.88,
+ "content": "h _ { t } = \\operatorname { t a n h } ( W _ { x } x _ { t } + W _ { h } h _ { t - 1 } )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38.5,
+ "bbox_fs": [
+ 105,
+ 680,
+ 505,
+ 706
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 118,
+ 80,
+ 491,
+ 193
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 118,
+ 80,
+ 491,
+ 193
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 80,
+ 491,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 80,
+ 491,
+ 193
+ ],
+ "score": 0.983,
+ "html": "
| Architecture | Structure | Filter size | Parameters | Runtime | Test(w/o Rot) | Test(Rot) |
| FC-LSTM | N/A | N/A | 142,667,776 | N/A | 4832 | = |
| LSTM+CNN | N/A | 5×5 | 13,524,496 | 2.10 | 3851 | 4339 |
| LSTM+CNN | N/A | 9×9 | 43,802,128 | 6.10 | 3903 | 4208 |
| LSTM+GCNN | knn=8 | K=3 | 1,629,712 | 0.82 | 3866 | 4367 |
| LSTM+GCNN | knn=8 | K=5 | 2,711,056 | 1.24 | 3495 | 3932 |
| LSTM+GCNN | knn=8 | K=7 | 3,792,400 | 1.61 | 3400 | 3803 |
| LSTM+GCNN | knn=8 | K=9 | 4,873,744 | 2.15 | 3395 | 3814 |
| LSTM+GCNN | knn=4 | K=7 | 3,792,400 | 1.61 | 3446 | 3844 |
| LSTM+GCNN | knn=16 | K=7 | 3,792,400 | 1.61 | 3578 | 3963 |
",
+ "type": "table",
+ "image_path": "1f72a796efa62aa678281afc28c1ae1f8c2af391c3fb550841999e3356c792db.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 118,
+ 80,
+ 491,
+ 117.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 118,
+ 117.66666666666666,
+ 491,
+ 155.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 118,
+ 155.33333333333331,
+ 491,
+ 192.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 201,
+ 504,
+ 235
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 504,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 504,
+ 213
+ ],
+ "score": 1.0,
+ "content": "Table 1: Comparison between models. Runtime is the time spent per each mini-batch in seconds.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 212,
+ 506,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 469,
+ 225
+ ],
+ "score": 1.0,
+ "content": "Test cross-entropies correspond to moving MNIST, and rotating and moving MNIST. LSTM",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 213,
+ 475,
+ 222
+ ],
+ "score": 0.27,
+ "content": "\\cdot +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 476,
+ 212,
+ 506,
+ 225
+ ],
+ "score": 1.0,
+ "content": "GCNN",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 222,
+ 449,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 222,
+ 449,
+ 236
+ ],
+ "score": 1.0,
+ "content": "is Model 2 defined in (9). Cross-entropy of FC-LSTM is taken from Shi et al. (2015).",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 255,
+ 194,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 195,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 195,
+ 267
+ ],
+ "score": 1.0,
+ "content": "would be modified as",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 227,
+ 265,
+ 383,
+ 279
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 227,
+ 265,
+ 383,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 227,
+ 265,
+ 383,
+ 279
+ ],
+ "score": 0.92,
+ "content": "h _ { t } = \\operatorname { t a n h } ( W _ { x } * _ { \\mathcal { G } } x _ { t } + W _ { h } * _ { \\mathcal { G } } h _ { t - 1 } ) ,",
+ "type": "interline_equation",
+ "image_path": "d79616516db4ad1d5910e404e4644b58bff9f4005b164de6b7b572773622014c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 227,
+ 265,
+ 383,
+ 279
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 281,
+ 332,
+ 293
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 332,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 332,
+ 295
+ ],
+ "score": 1.0,
+ "content": "and a Gated Recurrent Unit (GRU) (Cho et al., 2014) as",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 210,
+ 297,
+ 399,
+ 361
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 210,
+ 297,
+ 399,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 210,
+ 297,
+ 399,
+ 361
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r l } & { z = \\sigma ( W _ { x z } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h z } \\ast _ { \\mathcal { G } } h _ { t - 1 } ) , } \\\\ & { r = \\sigma ( W _ { x r } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h r } \\ast _ { \\mathcal { G } } h _ { t - 1 } ) , } \\\\ & { \\tilde { h } = \\operatorname { t a n h } ( W _ { x h } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h h } \\ast _ { \\mathcal { G } } ( r \\odot h _ { t - 1 } ) ) , } \\\\ & { h _ { t } = z \\odot h _ { t - 1 } + ( 1 - z ) \\odot \\tilde { h } . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "422b7026d5a39385da9c5717dc7ba20bfdf03c246eb59113c9e75d0d6450e48c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 210,
+ 297,
+ 399,
+ 313.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 210,
+ 313.0,
+ 399,
+ 329.0
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 210,
+ 329.0,
+ 399,
+ 345.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 210,
+ 345.0,
+ 399,
+ 361.0
+ ],
+ "spans": [],
+ "index": 12
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 370,
+ 505,
+ 414
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 506,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 506,
+ 383
+ ],
+ "score": 1.0,
+ "content": "As demonstrated by Shi et al. (2015), structure-aware LSTM cells can be stacked and used as",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "sequence-to-sequence models using an architecture composed of an encoder, which processes the",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "input sequence, and a decoder, which generates an output sequence. A standard practice for machine",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 403,
+ 369,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 369,
+ 415
+ ],
+ "score": 1.0,
+ "content": "translation using RNNs (Cho et al., 2014; Sutskever et al., 2014).",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 430,
+ 200,
+ 443
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 201,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 201,
+ 445
+ ],
+ "score": 1.0,
+ "content": "5 EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 455,
+ 405,
+ 466
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 454,
+ 407,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 454,
+ 407,
+ 468
+ ],
+ "score": 1.0,
+ "content": "5.1 SPATIO-TEMPORAL SEQUENCE MODELING ON MOVING-MNIST",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 476,
+ 505,
+ 608
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 489
+ ],
+ "score": 1.0,
+ "content": "For this synthetic experiment, we use the moving-MNIST dataset generated by Shi et al. (2015).",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 499
+ ],
+ "score": 1.0,
+ "content": "All sequences are 20 frames long (10 frames as input and 10 frames for prediction) and contain",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 506,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 277,
+ 511
+ ],
+ "score": 1.0,
+ "content": "two handwritten digits bouncing inside a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 498,
+ 312,
+ 509
+ ],
+ "score": 0.9,
+ "content": "6 4 \\times 6 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 497,
+ 506,
+ 511
+ ],
+ "score": 1.0,
+ "content": "patch. Following their experimental setup, all",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 509,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 509,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "models are trained by minimizing the binary cross-entropy loss using back-propagation through",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 519,
+ 506,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 519,
+ 317,
+ 533
+ ],
+ "score": 1.0,
+ "content": "time (BPTT) and RMSProp with a learning rate of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 317,
+ 519,
+ 339,
+ 530
+ ],
+ "score": 0.9,
+ "content": "1 0 ^ { - 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 519,
+ 506,
+ 533
+ ],
+ "score": 1.0,
+ "content": "and a decay rate of 0.9. We choose the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 531,
+ 506,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 506,
+ 544
+ ],
+ "score": 1.0,
+ "content": "best model with early-stopping on validation set. All implementations are based on their Theano",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 271,
+ 555
+ ],
+ "score": 1.0,
+ "content": "code and dataset.3 The adjacency matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 542,
+ 280,
+ 552
+ ],
+ "score": 0.67,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 541,
+ 357,
+ 555
+ ],
+ "score": 1.0,
+ "content": "is constructed as a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 357,
+ 542,
+ 364,
+ 552
+ ],
+ "score": 0.28,
+ "content": "\\mathbf { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 541,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": "-nearest-neighbor (knn) graph with",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Euclidean distance and Gaussian kernel between pixel locations. For a fair comparison with Shi",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 504,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 504,
+ 576
+ ],
+ "score": 1.0,
+ "content": "et al. (2015) defined in (6), all GCRN experiments are conducted with Model 2 defined in (9), which",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 104,
+ 574,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 574,
+ 304,
+ 587
+ ],
+ "score": 1.0,
+ "content": "is the same architecture with the 2D convolution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 577,
+ 312,
+ 585
+ ],
+ "score": 0.76,
+ "content": "^ *",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 574,
+ 444,
+ 587
+ ],
+ "score": 1.0,
+ "content": "replaced by a graph convolution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 576,
+ 457,
+ 587
+ ],
+ "score": 0.85,
+ "content": "^ { \\ast _ { \\mathcal { G } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 457,
+ 574,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": ". To further",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 585,
+ 506,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 585,
+ 506,
+ 599
+ ],
+ "score": 1.0,
+ "content": "explore the impact of the isotropic property of our filters, we generated a variant of the moving",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 349,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 349,
+ 609
+ ],
+ "score": 1.0,
+ "content": "MNIST dataset where digits are also rotating (see Figure 4).",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 613,
+ 505,
+ 713
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 626
+ ],
+ "score": 1.0,
+ "content": "Table 1 shows the performance of various models: (i) the baseline FC-LSTM from Shi et al. (2015),",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 624,
+ 505,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 624,
+ 194,
+ 637
+ ],
+ "score": 1.0,
+ "content": "(ii) the 1-layer LSTM",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 626,
+ 201,
+ 634
+ ],
+ "score": 0.35,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 624,
+ 505,
+ 637
+ ],
+ "score": 1.0,
+ "content": "CNN from Shi et al. (2015) with different filter sizes, and (iii) the proposed",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 635,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 370,
+ 649
+ ],
+ "score": 1.0,
+ "content": "LSTM+graph CNN(GCNN) defined in (9) with different supports",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 370,
+ 636,
+ 380,
+ 645
+ ],
+ "score": 0.77,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 635,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": ". These results show the ability",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 646,
+ 506,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 506,
+ 659
+ ],
+ "score": 1.0,
+ "content": "of the proposed method to capture spatio-temporal structures. Perhaps surprisingly, GCNNs can",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 657,
+ 506,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 657,
+ 506,
+ 671
+ ],
+ "score": 1.0,
+ "content": "offer better performance than regular CNNs, even when the domain is a 2D grid and the data is",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 668,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 668,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "images, the problem CNNs were initially developed for. The explanation is to be found in the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 679,
+ 504,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 679,
+ 473,
+ 691
+ ],
+ "score": 1.0,
+ "content": "differences between 2D filters and spectral graph filters. While a spectral filter of support",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 680,
+ 504,
+ 690
+ ],
+ "score": 0.89,
+ "content": "K = 3",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 690,
+ 505,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 690,
+ 274,
+ 703
+ ],
+ "score": 1.0,
+ "content": "corresponds to the reach of a patch of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 275,
+ 691,
+ 296,
+ 701
+ ],
+ "score": 0.87,
+ "content": "5 \\times 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 296,
+ 690,
+ 505,
+ 703
+ ],
+ "score": 1.0,
+ "content": "(see Figure 2), the difference resides in the isotropic",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 314,
+ 714
+ ],
+ "score": 1.0,
+ "content": "nature of the former and the number of parameters:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 314,
+ 702,
+ 342,
+ 711
+ ],
+ "score": 0.88,
+ "content": "K = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 700,
+ 419,
+ 714
+ ],
+ "score": 1.0,
+ "content": "for the former and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 701,
+ 453,
+ 712
+ ],
+ "score": 0.92,
+ "content": "5 ^ { 2 } = 2 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 453,
+ 700,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "for the later.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 118,
+ 722,
+ 367,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 719,
+ 368,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 719,
+ 368,
+ 734
+ ],
+ "score": 1.0,
+ "content": "3http://www.wanghao.in/code/SPARNN-release.zip",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 118,
+ 80,
+ 491,
+ 193
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 118,
+ 80,
+ 491,
+ 193
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 80,
+ 491,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 80,
+ 491,
+ 193
+ ],
+ "score": 0.983,
+ "html": "| Architecture | Structure | Filter size | Parameters | Runtime | Test(w/o Rot) | Test(Rot) |
| FC-LSTM | N/A | N/A | 142,667,776 | N/A | 4832 | = |
| LSTM+CNN | N/A | 5×5 | 13,524,496 | 2.10 | 3851 | 4339 |
| LSTM+CNN | N/A | 9×9 | 43,802,128 | 6.10 | 3903 | 4208 |
| LSTM+GCNN | knn=8 | K=3 | 1,629,712 | 0.82 | 3866 | 4367 |
| LSTM+GCNN | knn=8 | K=5 | 2,711,056 | 1.24 | 3495 | 3932 |
| LSTM+GCNN | knn=8 | K=7 | 3,792,400 | 1.61 | 3400 | 3803 |
| LSTM+GCNN | knn=8 | K=9 | 4,873,744 | 2.15 | 3395 | 3814 |
| LSTM+GCNN | knn=4 | K=7 | 3,792,400 | 1.61 | 3446 | 3844 |
| LSTM+GCNN | knn=16 | K=7 | 3,792,400 | 1.61 | 3578 | 3963 |
",
+ "type": "table",
+ "image_path": "1f72a796efa62aa678281afc28c1ae1f8c2af391c3fb550841999e3356c792db.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 118,
+ 80,
+ 491,
+ 117.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 118,
+ 117.66666666666666,
+ 491,
+ 155.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 118,
+ 155.33333333333331,
+ 491,
+ 192.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 201,
+ 504,
+ 235
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 504,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 504,
+ 213
+ ],
+ "score": 1.0,
+ "content": "Table 1: Comparison between models. Runtime is the time spent per each mini-batch in seconds.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 212,
+ 506,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 469,
+ 225
+ ],
+ "score": 1.0,
+ "content": "Test cross-entropies correspond to moving MNIST, and rotating and moving MNIST. LSTM",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 213,
+ 475,
+ 222
+ ],
+ "score": 0.27,
+ "content": "\\cdot +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 476,
+ 212,
+ 506,
+ 225
+ ],
+ "score": 1.0,
+ "content": "GCNN",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 222,
+ 449,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 222,
+ 449,
+ 236
+ ],
+ "score": 1.0,
+ "content": "is Model 2 defined in (9). Cross-entropy of FC-LSTM is taken from Shi et al. (2015).",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 105,
+ 201,
+ 506,
+ 236
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 255,
+ 194,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 195,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 195,
+ 267
+ ],
+ "score": 1.0,
+ "content": "would be modified as",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 106,
+ 254,
+ 195,
+ 267
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 227,
+ 265,
+ 383,
+ 279
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 227,
+ 265,
+ 383,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 227,
+ 265,
+ 383,
+ 279
+ ],
+ "score": 0.92,
+ "content": "h _ { t } = \\operatorname { t a n h } ( W _ { x } * _ { \\mathcal { G } } x _ { t } + W _ { h } * _ { \\mathcal { G } } h _ { t - 1 } ) ,",
+ "type": "interline_equation",
+ "image_path": "d79616516db4ad1d5910e404e4644b58bff9f4005b164de6b7b572773622014c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 227,
+ 265,
+ 383,
+ 279
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 281,
+ 332,
+ 293
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 332,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 332,
+ 295
+ ],
+ "score": 1.0,
+ "content": "and a Gated Recurrent Unit (GRU) (Cho et al., 2014) as",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8,
+ "bbox_fs": [
+ 105,
+ 280,
+ 332,
+ 295
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 210,
+ 297,
+ 399,
+ 361
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 210,
+ 297,
+ 399,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 210,
+ 297,
+ 399,
+ 361
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r l } & { z = \\sigma ( W _ { x z } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h z } \\ast _ { \\mathcal { G } } h _ { t - 1 } ) , } \\\\ & { r = \\sigma ( W _ { x r } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h r } \\ast _ { \\mathcal { G } } h _ { t - 1 } ) , } \\\\ & { \\tilde { h } = \\operatorname { t a n h } ( W _ { x h } \\ast _ { \\mathcal { G } } x _ { t } + W _ { h h } \\ast _ { \\mathcal { G } } ( r \\odot h _ { t - 1 } ) ) , } \\\\ & { h _ { t } = z \\odot h _ { t - 1 } + ( 1 - z ) \\odot \\tilde { h } . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "422b7026d5a39385da9c5717dc7ba20bfdf03c246eb59113c9e75d0d6450e48c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 210,
+ 297,
+ 399,
+ 313.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 210,
+ 313.0,
+ 399,
+ 329.0
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 210,
+ 329.0,
+ 399,
+ 345.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 210,
+ 345.0,
+ 399,
+ 361.0
+ ],
+ "spans": [],
+ "index": 12
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 370,
+ 505,
+ 414
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 506,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 506,
+ 383
+ ],
+ "score": 1.0,
+ "content": "As demonstrated by Shi et al. (2015), structure-aware LSTM cells can be stacked and used as",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "sequence-to-sequence models using an architecture composed of an encoder, which processes the",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "input sequence, and a decoder, which generates an output sequence. A standard practice for machine",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 403,
+ 369,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 369,
+ 415
+ ],
+ "score": 1.0,
+ "content": "translation using RNNs (Cho et al., 2014; Sutskever et al., 2014).",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 105,
+ 370,
+ 506,
+ 415
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 430,
+ 200,
+ 443
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 201,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 201,
+ 445
+ ],
+ "score": 1.0,
+ "content": "5 EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 455,
+ 405,
+ 466
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 454,
+ 407,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 454,
+ 407,
+ 468
+ ],
+ "score": 1.0,
+ "content": "5.1 SPATIO-TEMPORAL SEQUENCE MODELING ON MOVING-MNIST",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 476,
+ 505,
+ 608
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 489
+ ],
+ "score": 1.0,
+ "content": "For this synthetic experiment, we use the moving-MNIST dataset generated by Shi et al. (2015).",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 499
+ ],
+ "score": 1.0,
+ "content": "All sequences are 20 frames long (10 frames as input and 10 frames for prediction) and contain",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 506,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 277,
+ 511
+ ],
+ "score": 1.0,
+ "content": "two handwritten digits bouncing inside a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 498,
+ 312,
+ 509
+ ],
+ "score": 0.9,
+ "content": "6 4 \\times 6 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 497,
+ 506,
+ 511
+ ],
+ "score": 1.0,
+ "content": "patch. Following their experimental setup, all",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 509,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 509,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "models are trained by minimizing the binary cross-entropy loss using back-propagation through",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 519,
+ 506,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 519,
+ 317,
+ 533
+ ],
+ "score": 1.0,
+ "content": "time (BPTT) and RMSProp with a learning rate of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 317,
+ 519,
+ 339,
+ 530
+ ],
+ "score": 0.9,
+ "content": "1 0 ^ { - 3 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 519,
+ 506,
+ 533
+ ],
+ "score": 1.0,
+ "content": "and a decay rate of 0.9. We choose the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 531,
+ 506,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 506,
+ 544
+ ],
+ "score": 1.0,
+ "content": "best model with early-stopping on validation set. All implementations are based on their Theano",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 271,
+ 555
+ ],
+ "score": 1.0,
+ "content": "code and dataset.3 The adjacency matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 542,
+ 280,
+ 552
+ ],
+ "score": 0.67,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 541,
+ 357,
+ 555
+ ],
+ "score": 1.0,
+ "content": "is constructed as a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 357,
+ 542,
+ 364,
+ 552
+ ],
+ "score": 0.28,
+ "content": "\\mathbf { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 541,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": "-nearest-neighbor (knn) graph with",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Euclidean distance and Gaussian kernel between pixel locations. For a fair comparison with Shi",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 504,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 504,
+ 576
+ ],
+ "score": 1.0,
+ "content": "et al. (2015) defined in (6), all GCRN experiments are conducted with Model 2 defined in (9), which",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 104,
+ 574,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 574,
+ 304,
+ 587
+ ],
+ "score": 1.0,
+ "content": "is the same architecture with the 2D convolution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 577,
+ 312,
+ 585
+ ],
+ "score": 0.76,
+ "content": "^ *",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 312,
+ 574,
+ 444,
+ 587
+ ],
+ "score": 1.0,
+ "content": "replaced by a graph convolution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 576,
+ 457,
+ 587
+ ],
+ "score": 0.85,
+ "content": "^ { \\ast _ { \\mathcal { G } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 457,
+ 574,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": ". To further",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 585,
+ 506,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 585,
+ 506,
+ 599
+ ],
+ "score": 1.0,
+ "content": "explore the impact of the isotropic property of our filters, we generated a variant of the moving",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 349,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 349,
+ 609
+ ],
+ "score": 1.0,
+ "content": "MNIST dataset where digits are also rotating (see Figure 4).",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 104,
+ 475,
+ 506,
+ 609
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 613,
+ 505,
+ 713
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 612,
+ 505,
+ 626
+ ],
+ "score": 1.0,
+ "content": "Table 1 shows the performance of various models: (i) the baseline FC-LSTM from Shi et al. (2015),",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 624,
+ 505,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 624,
+ 194,
+ 637
+ ],
+ "score": 1.0,
+ "content": "(ii) the 1-layer LSTM",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 626,
+ 201,
+ 634
+ ],
+ "score": 0.35,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 624,
+ 505,
+ 637
+ ],
+ "score": 1.0,
+ "content": "CNN from Shi et al. (2015) with different filter sizes, and (iii) the proposed",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 635,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 370,
+ 649
+ ],
+ "score": 1.0,
+ "content": "LSTM+graph CNN(GCNN) defined in (9) with different supports",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 370,
+ 636,
+ 380,
+ 645
+ ],
+ "score": 0.77,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 635,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": ". These results show the ability",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 646,
+ 506,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 506,
+ 659
+ ],
+ "score": 1.0,
+ "content": "of the proposed method to capture spatio-temporal structures. Perhaps surprisingly, GCNNs can",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 657,
+ 506,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 657,
+ 506,
+ 671
+ ],
+ "score": 1.0,
+ "content": "offer better performance than regular CNNs, even when the domain is a 2D grid and the data is",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 668,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 668,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "images, the problem CNNs were initially developed for. The explanation is to be found in the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 679,
+ 504,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 679,
+ 473,
+ 691
+ ],
+ "score": 1.0,
+ "content": "differences between 2D filters and spectral graph filters. While a spectral filter of support",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 680,
+ 504,
+ 690
+ ],
+ "score": 0.89,
+ "content": "K = 3",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 690,
+ 505,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 690,
+ 274,
+ 703
+ ],
+ "score": 1.0,
+ "content": "corresponds to the reach of a patch of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 275,
+ 691,
+ 296,
+ 701
+ ],
+ "score": 0.87,
+ "content": "5 \\times 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 296,
+ 690,
+ 505,
+ 703
+ ],
+ "score": 1.0,
+ "content": "(see Figure 2), the difference resides in the isotropic",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 700,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 314,
+ 714
+ ],
+ "score": 1.0,
+ "content": "nature of the former and the number of parameters:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 314,
+ 702,
+ 342,
+ 711
+ ],
+ "score": 0.88,
+ "content": "K = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 700,
+ 419,
+ 714
+ ],
+ "score": 1.0,
+ "content": "for the former and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 701,
+ 453,
+ 712
+ ],
+ "score": 0.92,
+ "content": "5 ^ { 2 } = 2 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 453,
+ 700,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "for the later.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 612,
+ 506,
+ 714
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 84,
+ 504,
+ 233
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 84,
+ 504,
+ 233
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 504,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 504,
+ 233
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "c0e894075958a2fe6ae45958014950a3aaf3f9b843faf87a98d2580afacc1d3b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 504,
+ 133.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 133.66666666666666,
+ 504,
+ 183.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 183.33333333333331,
+ 504,
+ 232.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 245,
+ 504,
+ 268
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 258
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Cross-entropy on validation set: Left: performance of graph CNN with various filter",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 256,
+ 337,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 138,
+ 270
+ ],
+ "score": 1.0,
+ "content": "support",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 139,
+ 257,
+ 149,
+ 266
+ ],
+ "score": 0.76,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 256,
+ 337,
+ 270
+ ],
+ "score": 1.0,
+ "content": ". Right: performance w.r.t. graph construction.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 284,
+ 503,
+ 371
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 108,
+ 284,
+ 503,
+ 371
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 284,
+ 503,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 284,
+ 503,
+ 371
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "1cc92b78c2a49f24902eeb6f096c1459249ef396fac9c72850f84cda785f4e32.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 284,
+ 503,
+ 313.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 108,
+ 313.0,
+ 503,
+ 342.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 108,
+ 342.0,
+ 503,
+ 371.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 387,
+ 505,
+ 421
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Qualitative results for moving MNIST, and rotating and moving MNIST. First row",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 104,
+ 398,
+ 505,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 398,
+ 505,
+ 412
+ ],
+ "score": 1.0,
+ "content": "is the input sequence, second the ground truth, and third and fourth are the predictions of the",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 410,
+ 347,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 190,
+ 421
+ ],
+ "score": 0.68,
+ "content": "\\mathrm { L S T M + C N N } ( 5 \\times 5 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 410,
+ 208,
+ 422
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 410,
+ 343,
+ 421
+ ],
+ "score": 0.62,
+ "content": "\\mathrm { L S T M + G C N N } ( k n n = 8 , K = 7 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 410,
+ 347,
+ 422
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 442,
+ 505,
+ 530
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 505,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 224,
+ 454
+ ],
+ "score": 1.0,
+ "content": "Table 1 indeed shows that LS",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 443,
+ 296,
+ 453
+ ],
+ "score": 0.66,
+ "content": "\\mathrm { T M } { + } \\mathrm { C N N } ( 5 \\times 5 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 296,
+ 442,
+ 321,
+ 454
+ ],
+ "score": 1.0,
+ "content": "rivals",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 443,
+ 385,
+ 453
+ ],
+ "score": 0.26,
+ "content": "\\mathrm { L S T M + G C N N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 442,
+ 405,
+ 454
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 443,
+ 435,
+ 453
+ ],
+ "score": 0.88,
+ "content": "K = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 442,
+ 505,
+ 454
+ ],
+ "score": 1.0,
+ "content": ". However, when",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 213,
+ 465
+ ],
+ "score": 1.0,
+ "content": "increasing the filter size to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 213,
+ 454,
+ 236,
+ 464
+ ],
+ "score": 0.9,
+ "content": "9 \\times 9",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 236,
+ 453,
+ 247,
+ 465
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 247,
+ 454,
+ 276,
+ 464
+ ],
+ "score": 0.9,
+ "content": "K = 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 276,
+ 453,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": ", the GCNN variant clearly outperforms the CNN variant.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 477
+ ],
+ "score": 1.0,
+ "content": "This experiment demonstrates that graph spectral filters can obtain superior performance on regular",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 488
+ ],
+ "score": 1.0,
+ "content": "domains with much less parameters thanks to their isotropic nature, a controversial property. Indeed,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 499
+ ],
+ "score": 1.0,
+ "content": "as the nodes are not ordered, there is no notion of an edge going up, down, on the right or on the",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 509
+ ],
+ "score": 1.0,
+ "content": "left. All edges are treated equally, inducing some sort of rotation invariance. Additionally, Table 1",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 509,
+ 506,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 506,
+ 520
+ ],
+ "score": 1.0,
+ "content": "shows that the computational complexity of each model is linear with the filter size, and Figure 3",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 519,
+ 314,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 519,
+ 314,
+ 531
+ ],
+ "score": 1.0,
+ "content": "shows the learning dynamic of some of the models.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 545,
+ 367,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 368,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 368,
+ 558
+ ],
+ "score": 1.0,
+ "content": "5.2 NATURAL LANGUAGE MODELING ON PENN TREEBANK",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 567,
+ 505,
+ 700
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "The Penn Treebank dataset has 1,036,580 words. It was pre-processed in Zaremba et al. (2014) and",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 210,
+ 590
+ ],
+ "score": 1.0,
+ "content": "split4 into a training set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 210,
+ 578,
+ 232,
+ 588
+ ],
+ "score": 0.47,
+ "content": "9 2 9 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 576,
+ 440,
+ 590
+ ],
+ "score": 1.0,
+ "content": "words, a validation set of 73k words, and a test set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 441,
+ 578,
+ 457,
+ 588
+ ],
+ "score": 0.45,
+ "content": "{ 8 2 } \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 457,
+ 576,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "words. The",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 588,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "size of the vocabulary of this corpus is 10,000. We use the gensim library5 to compute a word2vec",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "model (Mikolov et al., 2013) for embedding the words of the dictionary in a 200-dimensional space.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": "Then we build the adjacency matrix of the word embedding using a 4-nearest neighbor graph with",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "cosine distance. Figure 6 presents the computed adjacency matrix, and its 3D visualization. We used",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 632,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 632,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "the hyperparameters of the small configuration given by the code6 based on Zaremba et al. (2014):",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 643,
+ 506,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 643,
+ 506,
+ 655
+ ],
+ "score": 1.0,
+ "content": "the size of the data mini-batch is 20, the number of temporal steps to unroll is 20, the dimension of",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "the hidden state is 200. The global learning rate is 1.0 and the norm of the gradient is bounded by",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 663,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 663,
+ 297,
+ 680
+ ],
+ "score": 1.0,
+ "content": "5. The learning decay function is selected to be",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 665,
+ 368,
+ 677
+ ],
+ "score": 0.78,
+ "content": "\\mathrm { 0 . 5 ^ { m a x ( 0 , \\# e p o c h - 4 ) } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 663,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": ". All experiments have 13 epochs,",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 676,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 676,
+ 423,
+ 690
+ ],
+ "score": 1.0,
+ "content": "and dropout value is 0.75. For Zaremba et al. (2014), the input representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 679,
+ 434,
+ 688
+ ],
+ "score": 0.85,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 676,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "can be either the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "200-dim embedding vector of the word, or the 10,000-dim one-hot representation of the word. For",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 25.5
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 118,
+ 711,
+ 404,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 708,
+ 309,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 708,
+ 309,
+ 723
+ ],
+ "score": 1.0,
+ "content": "4https://github.com/wojzaremba/lstm",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 118,
+ 719,
+ 404,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 719,
+ 404,
+ 733
+ ],
+ "score": 1.0,
+ "content": "5https://radimrehurek.com/gensim/models/word2vec.html",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "7",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 84,
+ 504,
+ 233
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 84,
+ 504,
+ 233
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 504,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 504,
+ 233
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "c0e894075958a2fe6ae45958014950a3aaf3f9b843faf87a98d2580afacc1d3b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 504,
+ 133.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 133.66666666666666,
+ 504,
+ 183.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 183.33333333333331,
+ 504,
+ 232.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 245,
+ 504,
+ 268
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 505,
+ 258
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Cross-entropy on validation set: Left: performance of graph CNN with various filter",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 256,
+ 337,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 138,
+ 270
+ ],
+ "score": 1.0,
+ "content": "support",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 139,
+ 257,
+ 149,
+ 266
+ ],
+ "score": 0.76,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 256,
+ 337,
+ 270
+ ],
+ "score": 1.0,
+ "content": ". Right: performance w.r.t. graph construction.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 284,
+ 503,
+ 371
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 108,
+ 284,
+ 503,
+ 371
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 284,
+ 503,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 284,
+ 503,
+ 371
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "1cc92b78c2a49f24902eeb6f096c1459249ef396fac9c72850f84cda785f4e32.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 284,
+ 503,
+ 313.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 108,
+ 313.0,
+ 503,
+ 342.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 108,
+ 342.0,
+ 503,
+ 371.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 387,
+ 505,
+ 421
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Qualitative results for moving MNIST, and rotating and moving MNIST. First row",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 104,
+ 398,
+ 505,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 398,
+ 505,
+ 412
+ ],
+ "score": 1.0,
+ "content": "is the input sequence, second the ground truth, and third and fourth are the predictions of the",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 410,
+ 347,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 190,
+ 421
+ ],
+ "score": 0.68,
+ "content": "\\mathrm { L S T M + C N N } ( 5 \\times 5 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 410,
+ 208,
+ 422
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 410,
+ 343,
+ 421
+ ],
+ "score": 0.62,
+ "content": "\\mathrm { L S T M + G C N N } ( k n n = 8 , K = 7 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 410,
+ 347,
+ 422
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 442,
+ 505,
+ 530
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 505,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 224,
+ 454
+ ],
+ "score": 1.0,
+ "content": "Table 1 indeed shows that LS",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 443,
+ 296,
+ 453
+ ],
+ "score": 0.66,
+ "content": "\\mathrm { T M } { + } \\mathrm { C N N } ( 5 \\times 5 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 296,
+ 442,
+ 321,
+ 454
+ ],
+ "score": 1.0,
+ "content": "rivals",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 443,
+ 385,
+ 453
+ ],
+ "score": 0.26,
+ "content": "\\mathrm { L S T M + G C N N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 442,
+ 405,
+ 454
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 443,
+ 435,
+ 453
+ ],
+ "score": 0.88,
+ "content": "K = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 442,
+ 505,
+ 454
+ ],
+ "score": 1.0,
+ "content": ". However, when",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 213,
+ 465
+ ],
+ "score": 1.0,
+ "content": "increasing the filter size to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 213,
+ 454,
+ 236,
+ 464
+ ],
+ "score": 0.9,
+ "content": "9 \\times 9",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 236,
+ 453,
+ 247,
+ 465
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 247,
+ 454,
+ 276,
+ 464
+ ],
+ "score": 0.9,
+ "content": "K = 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 276,
+ 453,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": ", the GCNN variant clearly outperforms the CNN variant.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 505,
+ 477
+ ],
+ "score": 1.0,
+ "content": "This experiment demonstrates that graph spectral filters can obtain superior performance on regular",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 488
+ ],
+ "score": 1.0,
+ "content": "domains with much less parameters thanks to their isotropic nature, a controversial property. Indeed,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 499
+ ],
+ "score": 1.0,
+ "content": "as the nodes are not ordered, there is no notion of an edge going up, down, on the right or on the",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 509
+ ],
+ "score": 1.0,
+ "content": "left. All edges are treated equally, inducing some sort of rotation invariance. Additionally, Table 1",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 509,
+ 506,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 506,
+ 520
+ ],
+ "score": 1.0,
+ "content": "shows that the computational complexity of each model is linear with the filter size, and Figure 3",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 519,
+ 314,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 519,
+ 314,
+ 531
+ ],
+ "score": 1.0,
+ "content": "shows the learning dynamic of some of the models.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 105,
+ 442,
+ 506,
+ 531
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 545,
+ 367,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 368,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 368,
+ 558
+ ],
+ "score": 1.0,
+ "content": "5.2 NATURAL LANGUAGE MODELING ON PENN TREEBANK",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 567,
+ 505,
+ 700
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "The Penn Treebank dataset has 1,036,580 words. It was pre-processed in Zaremba et al. (2014) and",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 210,
+ 590
+ ],
+ "score": 1.0,
+ "content": "split4 into a training set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 210,
+ 578,
+ 232,
+ 588
+ ],
+ "score": 0.47,
+ "content": "9 2 9 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 576,
+ 440,
+ 590
+ ],
+ "score": 1.0,
+ "content": "words, a validation set of 73k words, and a test set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 441,
+ 578,
+ 457,
+ 588
+ ],
+ "score": 0.45,
+ "content": "{ 8 2 } \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 457,
+ 576,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "words. The",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 588,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "size of the vocabulary of this corpus is 10,000. We use the gensim library5 to compute a word2vec",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "model (Mikolov et al., 2013) for embedding the words of the dictionary in a 200-dimensional space.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": "Then we build the adjacency matrix of the word embedding using a 4-nearest neighbor graph with",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "cosine distance. Figure 6 presents the computed adjacency matrix, and its 3D visualization. We used",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 632,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 632,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "the hyperparameters of the small configuration given by the code6 based on Zaremba et al. (2014):",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 643,
+ 506,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 643,
+ 506,
+ 655
+ ],
+ "score": 1.0,
+ "content": "the size of the data mini-batch is 20, the number of temporal steps to unroll is 20, the dimension of",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "the hidden state is 200. The global learning rate is 1.0 and the norm of the gradient is bounded by",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 663,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 663,
+ 297,
+ 680
+ ],
+ "score": 1.0,
+ "content": "5. The learning decay function is selected to be",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 665,
+ 368,
+ 677
+ ],
+ "score": 0.78,
+ "content": "\\mathrm { 0 . 5 ^ { m a x ( 0 , \\# e p o c h - 4 ) } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 663,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": ". All experiments have 13 epochs,",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 676,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 676,
+ 423,
+ 690
+ ],
+ "score": 1.0,
+ "content": "and dropout value is 0.75. For Zaremba et al. (2014), the input representation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 679,
+ 434,
+ 688
+ ],
+ "score": 0.85,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 676,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "can be either the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "200-dim embedding vector of the word, or the 10,000-dim one-hot representation of the word. For",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 104,
+ 566,
+ 506,
+ 700
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 83,
+ 504,
+ 221
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 105,
+ 83,
+ 504,
+ 221
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 504,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 504,
+ 221
+ ],
+ "score": 0.967,
+ "type": "image",
+ "image_path": "5aea0fe6f1f13974198278cfe12fec785421338be7b8b8b7b0603d98e18d7d64.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 504,
+ 129.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 129.0,
+ 504,
+ 175.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 175.0,
+ 504,
+ 221.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 119,
+ 254,
+ 489,
+ 349
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 234,
+ 503,
+ 246
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 504,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 504,
+ 248
+ ],
+ "score": 1.0,
+ "content": "Figure 5: Learning dynamic of LSTM with and without graph structure and dropout regularization.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 119,
+ 254,
+ 489,
+ 349
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 254,
+ 489,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 254,
+ 489,
+ 349
+ ],
+ "score": 0.974,
+ "html": "| Architecture | Representation | Parameters | Train Perplexity | Test Perplexity |
| Zaremba et al. (2014) code6 | embedding | 681,800 | 36.96 | 117.29 |
| Zaremba et al. (2014) code6 | one-hot | 34,011,600 | 53.89 | 118.82 |
| LSTM | embedding | 681,800 | 48.38 | 120.90 |
| LSTM | one-hot | 34,011,600 | 54.41 | 120.16 |
| LSTM, dropout | one-hot | 34,011,600 | 145.59 | 112.98 |
| GCRN-M1 | one-hot | 42,011,602 | 18.49 | 177.14 |
| GCRN-M1, dropout | one-hot | 42,011,602 | 114.29 | 98.67 |
",
+ "type": "table",
+ "image_path": "2fd5db6c10cb39ec3cac44638afebe278f3ba5cff9ed142b470c25c771268388.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 119,
+ 254,
+ 489,
+ 285.6666666666667
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 119,
+ 285.6666666666667,
+ 489,
+ 317.33333333333337
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 119,
+ 317.33333333333337,
+ 489,
+ 349.00000000000006
+ ],
+ "spans": [],
+ "index": 6
+ }
+ ]
+ }
+ ],
+ "index": 4.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 358,
+ 505,
+ 425
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 370
+ ],
+ "score": 1.0,
+ "content": "Table 2: Comparison of models in terms of perplexity. Zaremba et al. (2014) code6 is ran as bench-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 370,
+ 505,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 370,
+ 477,
+ 382
+ ],
+ "score": 1.0,
+ "content": "mark algorithm. The original Zaremba et al. (2014) code used as input representation for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 478,
+ 371,
+ 488,
+ 381
+ ],
+ "score": 0.84,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 488,
+ 370,
+ 505,
+ 382
+ ],
+ "score": 1.0,
+ "content": "the",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 380,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 380,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "200-dim embedding representation of words, computed here by the gensim library5. As our model",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 391,
+ 505,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 505,
+ 404
+ ],
+ "score": 1.0,
+ "content": "runs on the 10,000-dim one-hot representation of words, we also ran Zaremba et al. (2014) code on",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 403,
+ 505,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 505,
+ 414
+ ],
+ "score": 1.0,
+ "content": "this representation. We re-implemented Zaremba et al. (2014) code with the same architecture and",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 413,
+ 449,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 449,
+ 426
+ ],
+ "score": 1.0,
+ "content": "hyperparameters. We remind that GCRN-M1 refers to GCRN Model 1 defined in (8).",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 444,
+ 503,
+ 466
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 443,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 443,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "our models, the input representation is a one-hot representation of the word. This choice allows us",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 455,
+ 264,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 264,
+ 467
+ ],
+ "score": 1.0,
+ "content": "to use the graph structure of the words.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 472,
+ 505,
+ 505
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 485
+ ],
+ "score": 1.0,
+ "content": "Table 2 reports the final train and test perplexity values for each investigated model and Figure 5",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "plots the perplexity value vs. the number of epochs for the train and test sets with and without",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 494,
+ 322,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 494,
+ 322,
+ 506
+ ],
+ "score": 1.0,
+ "content": "dropout regularization. Numerical experiments show:",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 128,
+ 514,
+ 505,
+ 701
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 514,
+ 504,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 514,
+ 504,
+ 527
+ ],
+ "score": 1.0,
+ "content": "1. Given the same experimental conditions in terms of architecture and no dropout regulariza-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 525,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 525,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "tion, the standalone model of LSTM is more accurate than LSTM using the spatial graph",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 534,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 534,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "information (120.16 vs. 177.14), extracted by graph CNN with the GCRN architecture of",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 142,
+ 547,
+ 214,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 547,
+ 214,
+ 559
+ ],
+ "score": 1.0,
+ "content": "Model 1, Eq. (8).",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 128,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "2. However, using dropout regularization, the graph LSTM model overcomes the standalone",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 573,
+ 334,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 573,
+ 334,
+ 585
+ ],
+ "score": 1.0,
+ "content": "LSTM with perplexity values 98.67 vs. 112.98.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 128,
+ 586,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 586,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "3. The use of spatial graph information found by graph CNN speeds up the learning process,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 597,
+ 506,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 597,
+ 506,
+ 612
+ ],
+ "score": 1.0,
+ "content": "and overfits the training dataset in the absence of dropout regularization. The graph struc-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 608,
+ 506,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 608,
+ 506,
+ 622
+ ],
+ "score": 1.0,
+ "content": "ture likely acts a constraint on the learning system that is forced to move in the space of",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 141,
+ 619,
+ 209,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 619,
+ 209,
+ 633
+ ],
+ "score": 1.0,
+ "content": "language topics.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 129,
+ 635,
+ 505,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 635,
+ 505,
+ 647
+ ],
+ "score": 1.0,
+ "content": "4. We performed the same experiments with LSTM and Model 2 defined in (9). Model 1",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 141,
+ 645,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 645,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": "significantly outperformed Model 2, and Model 2 did worse than standalone LSTM. This",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 141,
+ 656,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 656,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "bad performance may be the result of the large increase of dimensionality in Model 2, as",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 141,
+ 667,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 667,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "the dimension of the hidden and cell states changes from 200 to 10,000, the size of the",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 142,
+ 679,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 679,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "vocabulary. A solution would be to downsize the data dimensionality, as done in Shi et al.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 141,
+ 689,
+ 275,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 689,
+ 275,
+ 702
+ ],
+ "score": 1.0,
+ "content": "(2015) in the case of image data.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 25.5
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 712,
+ 469,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 709,
+ 469,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 709,
+ 469,
+ 723
+ ],
+ "score": 1.0,
+ "content": "6https://github.com/tensorflow/tensorflow/blob/master/tensorflow/",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 265,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 265,
+ 734
+ ],
+ "score": 1.0,
+ "content": "models/rnn/ptb/ptb_word_lm.py",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 750,
+ 309,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 750,
+ 309,
+ 761
+ ],
+ "score": 1.0,
+ "content": "8",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 83,
+ 504,
+ 221
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 105,
+ 83,
+ 504,
+ 221
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 504,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 504,
+ 221
+ ],
+ "score": 0.967,
+ "type": "image",
+ "image_path": "5aea0fe6f1f13974198278cfe12fec785421338be7b8b8b7b0603d98e18d7d64.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 504,
+ 129.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 129.0,
+ 504,
+ 175.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 175.0,
+ 504,
+ 221.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 119,
+ 254,
+ 489,
+ 349
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 234,
+ 503,
+ 246
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 504,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 504,
+ 248
+ ],
+ "score": 1.0,
+ "content": "Figure 5: Learning dynamic of LSTM with and without graph structure and dropout regularization.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 119,
+ 254,
+ 489,
+ 349
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 254,
+ 489,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 254,
+ 489,
+ 349
+ ],
+ "score": 0.974,
+ "html": "| Architecture | Representation | Parameters | Train Perplexity | Test Perplexity |
| Zaremba et al. (2014) code6 | embedding | 681,800 | 36.96 | 117.29 |
| Zaremba et al. (2014) code6 | one-hot | 34,011,600 | 53.89 | 118.82 |
| LSTM | embedding | 681,800 | 48.38 | 120.90 |
| LSTM | one-hot | 34,011,600 | 54.41 | 120.16 |
| LSTM, dropout | one-hot | 34,011,600 | 145.59 | 112.98 |
| GCRN-M1 | one-hot | 42,011,602 | 18.49 | 177.14 |
| GCRN-M1, dropout | one-hot | 42,011,602 | 114.29 | 98.67 |
",
+ "type": "table",
+ "image_path": "2fd5db6c10cb39ec3cac44638afebe278f3ba5cff9ed142b470c25c771268388.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 119,
+ 254,
+ 489,
+ 285.6666666666667
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 119,
+ 285.6666666666667,
+ 489,
+ 317.33333333333337
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 119,
+ 317.33333333333337,
+ 489,
+ 349.00000000000006
+ ],
+ "spans": [],
+ "index": 6
+ }
+ ]
+ }
+ ],
+ "index": 4.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 358,
+ 505,
+ 425
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 505,
+ 370
+ ],
+ "score": 1.0,
+ "content": "Table 2: Comparison of models in terms of perplexity. Zaremba et al. (2014) code6 is ran as bench-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 370,
+ 505,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 370,
+ 477,
+ 382
+ ],
+ "score": 1.0,
+ "content": "mark algorithm. The original Zaremba et al. (2014) code used as input representation for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 478,
+ 371,
+ 488,
+ 381
+ ],
+ "score": 0.84,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 488,
+ 370,
+ 505,
+ 382
+ ],
+ "score": 1.0,
+ "content": "the",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 380,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 380,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "200-dim embedding representation of words, computed here by the gensim library5. As our model",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 391,
+ 505,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 505,
+ 404
+ ],
+ "score": 1.0,
+ "content": "runs on the 10,000-dim one-hot representation of words, we also ran Zaremba et al. (2014) code on",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 403,
+ 505,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 505,
+ 414
+ ],
+ "score": 1.0,
+ "content": "this representation. We re-implemented Zaremba et al. (2014) code with the same architecture and",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 413,
+ 449,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 449,
+ 426
+ ],
+ "score": 1.0,
+ "content": "hyperparameters. We remind that GCRN-M1 refers to GCRN Model 1 defined in (8).",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 105,
+ 358,
+ 505,
+ 426
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 444,
+ 503,
+ 466
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 443,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 443,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "our models, the input representation is a one-hot representation of the word. This choice allows us",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 455,
+ 264,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 264,
+ 467
+ ],
+ "score": 1.0,
+ "content": "to use the graph structure of the words.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 106,
+ 443,
+ 505,
+ 467
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 472,
+ 505,
+ 505
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 485
+ ],
+ "score": 1.0,
+ "content": "Table 2 reports the final train and test perplexity values for each investigated model and Figure 5",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "plots the perplexity value vs. the number of epochs for the train and test sets with and without",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 494,
+ 322,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 494,
+ 322,
+ 506
+ ],
+ "score": 1.0,
+ "content": "dropout regularization. Numerical experiments show:",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 105,
+ 471,
+ 505,
+ 506
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 128,
+ 514,
+ 505,
+ 701
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 514,
+ 504,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 514,
+ 504,
+ 527
+ ],
+ "score": 1.0,
+ "content": "1. Given the same experimental conditions in terms of architecture and no dropout regulariza-",
+ "type": "text"
+ }
+ ],
+ "index": 18,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 525,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 525,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "tion, the standalone model of LSTM is more accurate than LSTM using the spatial graph",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 534,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 534,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "information (120.16 vs. 177.14), extracted by graph CNN with the GCRN architecture of",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 142,
+ 547,
+ 214,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 547,
+ 214,
+ 559
+ ],
+ "score": 1.0,
+ "content": "Model 1, Eq. (8).",
+ "type": "text"
+ }
+ ],
+ "index": 21,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "2. However, using dropout regularization, the graph LSTM model overcomes the standalone",
+ "type": "text"
+ }
+ ],
+ "index": 22,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 573,
+ 334,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 573,
+ 334,
+ 585
+ ],
+ "score": 1.0,
+ "content": "LSTM with perplexity values 98.67 vs. 112.98.",
+ "type": "text"
+ }
+ ],
+ "index": 23,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 586,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 586,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "3. The use of spatial graph information found by graph CNN speeds up the learning process,",
+ "type": "text"
+ }
+ ],
+ "index": 24,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 597,
+ 506,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 597,
+ 506,
+ 612
+ ],
+ "score": 1.0,
+ "content": "and overfits the training dataset in the absence of dropout regularization. The graph struc-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 608,
+ 506,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 608,
+ 506,
+ 622
+ ],
+ "score": 1.0,
+ "content": "ture likely acts a constraint on the learning system that is forced to move in the space of",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 141,
+ 619,
+ 209,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 619,
+ 209,
+ 633
+ ],
+ "score": 1.0,
+ "content": "language topics.",
+ "type": "text"
+ }
+ ],
+ "index": 27,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 129,
+ 635,
+ 505,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 635,
+ 505,
+ 647
+ ],
+ "score": 1.0,
+ "content": "4. We performed the same experiments with LSTM and Model 2 defined in (9). Model 1",
+ "type": "text"
+ }
+ ],
+ "index": 28,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 645,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 645,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": "significantly outperformed Model 2, and Model 2 did worse than standalone LSTM. This",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 141,
+ 656,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 656,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "bad performance may be the result of the large increase of dimensionality in Model 2, as",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 141,
+ 667,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 667,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "the dimension of the hidden and cell states changes from 200 to 10,000, the size of the",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 142,
+ 679,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 679,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "vocabulary. A solution would be to downsize the data dimensionality, as done in Shi et al.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 141,
+ 689,
+ 275,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 689,
+ 275,
+ 702
+ ],
+ "score": 1.0,
+ "content": "(2015) in the case of image data.",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 128,
+ 514,
+ 506,
+ 702
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 359,
+ 88,
+ 484,
+ 217
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 359,
+ 88,
+ 484,
+ 217
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 359,
+ 88,
+ 484,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 359,
+ 88,
+ 484,
+ 217
+ ],
+ "score": 0.58,
+ "type": "image",
+ "image_path": "db4ee1673e3b3918a68b4668574358c8b6f41e4249c011cc1ed5e715b01017b4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2.0,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 359,
+ 88,
+ 484,
+ 152.5
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 359,
+ 152.5,
+ 484,
+ 217.0
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ }
+ ],
+ "index": 2.0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 156,
+ 91,
+ 285,
+ 212
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 156,
+ 91,
+ 285,
+ 212
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 156,
+ 91,
+ 285,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 156,
+ 91,
+ 285,
+ 212
+ ],
+ "score": 0.885,
+ "type": "image",
+ "image_path": "1a0958b69c9ef0cec7f225791ed9902ac5872a3df2683f257db5466fc71df530.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1.0,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 156,
+ 91,
+ 285,
+ 151.5
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 156,
+ 151.5,
+ 285,
+ 212.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 225,
+ 502,
+ 236
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 224,
+ 504,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 224,
+ 504,
+ 238
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Left: adjacency matrix of word embeddings. Right: 3D visualization of words’ structure.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 257,
+ 303,
+ 270
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 304,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 304,
+ 272
+ ],
+ "score": 1.0,
+ "content": "6 CONCLUSION AND FUTURE WORK",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 284,
+ 505,
+ 427
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 505,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 505,
+ 297
+ ],
+ "score": 1.0,
+ "content": "This work aims at learning spatio-temporal structures from graph-structured and time-varying data.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "In this context, the main challenge is to identify the best possible architecture that combines simul-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 307,
+ 505,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 505,
+ 318
+ ],
+ "score": 1.0,
+ "content": "taneously recurrent neural networks like vanilla RNN, LSTM or GRU with convolutional neural",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 317,
+ 506,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 506,
+ 330
+ ],
+ "score": 1.0,
+ "content": "networks for graph-structured data. We have investigated here two architectures, one using a stack",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 327,
+ 505,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 327,
+ 505,
+ 341
+ ],
+ "score": 1.0,
+ "content": "of CNN and RNN (Model 1), and one using convLSTM that considers convolutions instead of fully",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 338,
+ 505,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 505,
+ 352
+ ],
+ "score": 1.0,
+ "content": "connected operations in the RNN definition (Model 2). We have then considered two applications:",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 350,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 350,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "video prediction and natural language modeling. Model 2 has shown good performances in the",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 360,
+ 506,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 360,
+ 506,
+ 374
+ ],
+ "score": 1.0,
+ "content": "case of video prediction, by improving the results of Shi et al. (2015). Model 1 has also provided",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 372,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "promising performances in the case of language modeling, particularly in terms of learning speed.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "It has been shown that (i) isotropic filters, maybe surprisingly, can outperform classical 2D filters",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "on images while requiring much less parameters, and (ii) that graphs coupled with graph CNN and",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 403,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 403,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "RNN are a versatile way of introducing and exploiting side-information, e.g. the semantic of words,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 221,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 221,
+ 428
+ ],
+ "score": 1.0,
+ "content": "by structuring a data matrix.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 433,
+ 505,
+ 510
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 445
+ ],
+ "score": 1.0,
+ "content": "Future work will investigate applications to data naturally structured as dynamic graph signals, for",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 445,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 445,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "instance fMRI and sensor networks. The graph CNN model we have used is rotationally-invariant",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 505,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 505,
+ 467
+ ],
+ "score": 1.0,
+ "content": "and such spatial property seems quite attractive in real situations where motion is beyond translation.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 465,
+ 505,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 465,
+ 505,
+ 479
+ ],
+ "score": 1.0,
+ "content": "We will also investigate how to benefit of the fast learning property of our system to speed up",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 490
+ ],
+ "score": 1.0,
+ "content": "language modeling models. Eventually, it will be interesting to analyze the underlying dynamical",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 487,
+ 505,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 487,
+ 505,
+ 501
+ ],
+ "score": 1.0,
+ "content": "property of generic RNN architectures in the case of graphs. Graph structures may introduce stability",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 499,
+ 407,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 499,
+ 407,
+ 511
+ ],
+ "score": 1.0,
+ "content": "to RNN systems, and prevent them to express unstable dynamic behaviors.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 529,
+ 212,
+ 541
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 529,
+ 214,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 529,
+ 214,
+ 543
+ ],
+ "score": 1.0,
+ "content": "ACKNOWLEDGMENT",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 555,
+ 504,
+ 577
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 555,
+ 505,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 555,
+ 505,
+ 568
+ ],
+ "score": 1.0,
+ "content": "This research was supported in part by the European Union’s H2020 Framework Programme",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 479,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 479,
+ 579
+ ],
+ "score": 1.0,
+ "content": "(H2020-MSCA-ITN-2014) under grant No. 642685 MacSeNet, and Nvidia equipment grant.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 595,
+ 175,
+ 607
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 176,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 176,
+ 609
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 614,
+ 505,
+ 635
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 614,
+ 505,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 614,
+ 505,
+ 625
+ ],
+ "score": 1.0,
+ "content": "Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral Networks and Locally Connected",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 116,
+ 624,
+ 460,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 624,
+ 460,
+ 636
+ ],
+ "score": 1.0,
+ "content": "Networks on Graphs. In International Conference on Learning Representations (ICML), 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 646,
+ 504,
+ 677
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 505,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 505,
+ 658
+ ],
+ "score": 1.0,
+ "content": "Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger ¨",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 657,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 657,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 116,
+ 667,
+ 280,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 667,
+ 280,
+ 677
+ ],
+ "score": 1.0,
+ "content": "machine translation. arXiv:1406.1078, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 689,
+ 396,
+ 699
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 397,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 397,
+ 701
+ ],
+ "score": 1.0,
+ "content": "F. R. K. Chung. Spectral Graph Theory. American Mathematical Society, 1997.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 711,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with ¨",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 483,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 483,
+ 732
+ ],
+ "score": 1.0,
+ "content": "fast localized spectral filtering. In Advances in Neural Information Processing Systems (NIPS), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5
+ }
+ ],
+ "page_idx": 8,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2017",
+ "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": "image",
+ "bbox": [
+ 359,
+ 88,
+ 484,
+ 217
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 359,
+ 88,
+ 484,
+ 217
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 359,
+ 88,
+ 484,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 359,
+ 88,
+ 484,
+ 217
+ ],
+ "score": 0.58,
+ "type": "image",
+ "image_path": "db4ee1673e3b3918a68b4668574358c8b6f41e4249c011cc1ed5e715b01017b4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2.0,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 359,
+ 88,
+ 484,
+ 152.5
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 359,
+ 152.5,
+ 484,
+ 217.0
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ }
+ ],
+ "index": 2.0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 156,
+ 91,
+ 285,
+ 212
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 156,
+ 91,
+ 285,
+ 212
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 156,
+ 91,
+ 285,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 156,
+ 91,
+ 285,
+ 212
+ ],
+ "score": 0.885,
+ "type": "image",
+ "image_path": "1a0958b69c9ef0cec7f225791ed9902ac5872a3df2683f257db5466fc71df530.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1.0,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 156,
+ 91,
+ 285,
+ 151.5
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 156,
+ 151.5,
+ 285,
+ 212.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 225,
+ 502,
+ 236
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 224,
+ 504,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 224,
+ 504,
+ 238
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Left: adjacency matrix of word embeddings. Right: 3D visualization of words’ structure.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 257,
+ 303,
+ 270
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 304,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 304,
+ 272
+ ],
+ "score": 1.0,
+ "content": "6 CONCLUSION AND FUTURE WORK",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 284,
+ 505,
+ 427
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 505,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 505,
+ 297
+ ],
+ "score": 1.0,
+ "content": "This work aims at learning spatio-temporal structures from graph-structured and time-varying data.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "In this context, the main challenge is to identify the best possible architecture that combines simul-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 307,
+ 505,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 505,
+ 318
+ ],
+ "score": 1.0,
+ "content": "taneously recurrent neural networks like vanilla RNN, LSTM or GRU with convolutional neural",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 317,
+ 506,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 506,
+ 330
+ ],
+ "score": 1.0,
+ "content": "networks for graph-structured data. We have investigated here two architectures, one using a stack",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 327,
+ 505,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 327,
+ 505,
+ 341
+ ],
+ "score": 1.0,
+ "content": "of CNN and RNN (Model 1), and one using convLSTM that considers convolutions instead of fully",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 338,
+ 505,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 505,
+ 352
+ ],
+ "score": 1.0,
+ "content": "connected operations in the RNN definition (Model 2). We have then considered two applications:",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 350,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 350,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "video prediction and natural language modeling. Model 2 has shown good performances in the",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 360,
+ 506,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 360,
+ 506,
+ 374
+ ],
+ "score": 1.0,
+ "content": "case of video prediction, by improving the results of Shi et al. (2015). Model 1 has also provided",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 372,
+ 505,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 505,
+ 384
+ ],
+ "score": 1.0,
+ "content": "promising performances in the case of language modeling, particularly in terms of learning speed.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "It has been shown that (i) isotropic filters, maybe surprisingly, can outperform classical 2D filters",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "on images while requiring much less parameters, and (ii) that graphs coupled with graph CNN and",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 403,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 403,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "RNN are a versatile way of introducing and exploiting side-information, e.g. the semantic of words,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 221,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 221,
+ 428
+ ],
+ "score": 1.0,
+ "content": "by structuring a data matrix.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 104,
+ 284,
+ 506,
+ 428
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 433,
+ 505,
+ 510
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 445
+ ],
+ "score": 1.0,
+ "content": "Future work will investigate applications to data naturally structured as dynamic graph signals, for",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 445,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 445,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "instance fMRI and sensor networks. The graph CNN model we have used is rotationally-invariant",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 505,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 505,
+ 467
+ ],
+ "score": 1.0,
+ "content": "and such spatial property seems quite attractive in real situations where motion is beyond translation.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 465,
+ 505,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 465,
+ 505,
+ 479
+ ],
+ "score": 1.0,
+ "content": "We will also investigate how to benefit of the fast learning property of our system to speed up",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 490
+ ],
+ "score": 1.0,
+ "content": "language modeling models. Eventually, it will be interesting to analyze the underlying dynamical",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 487,
+ 505,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 487,
+ 505,
+ 501
+ ],
+ "score": 1.0,
+ "content": "property of generic RNN architectures in the case of graphs. Graph structures may introduce stability",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 499,
+ 407,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 499,
+ 407,
+ 511
+ ],
+ "score": 1.0,
+ "content": "to RNN systems, and prevent them to express unstable dynamic behaviors.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 433,
+ 506,
+ 511
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 529,
+ 212,
+ 541
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 529,
+ 214,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 529,
+ 214,
+ 543
+ ],
+ "score": 1.0,
+ "content": "ACKNOWLEDGMENT",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 555,
+ 504,
+ 577
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 555,
+ 505,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 555,
+ 505,
+ 568
+ ],
+ "score": 1.0,
+ "content": "This research was supported in part by the European Union’s H2020 Framework Programme",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 479,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 479,
+ 579
+ ],
+ "score": 1.0,
+ "content": "(H2020-MSCA-ITN-2014) under grant No. 642685 MacSeNet, and Nvidia equipment grant.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 555,
+ 505,
+ 579
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 595,
+ 175,
+ 607
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 176,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 176,
+ 609
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 614,
+ 505,
+ 635
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 614,
+ 505,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 614,
+ 505,
+ 625
+ ],
+ "score": 1.0,
+ "content": "Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral Networks and Locally Connected",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 116,
+ 624,
+ 460,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 624,
+ 460,
+ 636
+ ],
+ "score": 1.0,
+ "content": "Networks on Graphs. In International Conference on Learning Representations (ICML), 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 106,
+ 614,
+ 505,
+ 636
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 646,
+ 504,
+ 677
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 505,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 505,
+ 658
+ ],
+ "score": 1.0,
+ "content": "Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger ¨",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 657,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 657,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 116,
+ 667,
+ 280,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 667,
+ 280,
+ 677
+ ],
+ "score": 1.0,
+ "content": "machine translation. arXiv:1406.1078, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 106,
+ 646,
+ 505,
+ 677
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 689,
+ 396,
+ 699
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 397,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 397,
+ 701
+ ],
+ "score": 1.0,
+ "content": "F. R. K. Chung. Spectral Graph Theory. American Mathematical Society, 1997.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 689,
+ 397,
+ 701
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 711,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with ¨",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 483,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 483,
+ 732
+ ],
+ "score": 1.0,
+ "content": "fast localized spectral filtering. In Advances in Neural Information Processing Systems (NIPS), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 106,
+ 711,
+ 505,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 83,
+ 503,
+ 114
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 92,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 92,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and descrip-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 116,
+ 103,
+ 429,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 103,
+ 429,
+ 115
+ ],
+ "score": 1.0,
+ "content": "tion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 120,
+ 503,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 505,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 505,
+ 132
+ ],
+ "score": 1.0,
+ "content": "Felix A Gers and Jurgen Schmidhuber. Recurrent nets that time and count. In ¨ IEEE-INNS-ENNS International",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 131,
+ 278,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 131,
+ 278,
+ 142
+ ],
+ "score": 1.0,
+ "content": "Joint Conference on Neural Networks, 2000.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 149,
+ 441,
+ 160
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 149,
+ 441,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 149,
+ 441,
+ 161
+ ],
+ "score": 1.0,
+ "content": "Alex Graves. Generating sequences with recurrent neural networks. arXiv:1308.0850, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 167,
+ 455,
+ 178
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 166,
+ 456,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 166,
+ 456,
+ 180
+ ],
+ "score": 1.0,
+ "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 1997.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 184,
+ 502,
+ 205
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 504,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 504,
+ 196
+ ],
+ "score": 1.0,
+ "content": "Ashesh Jain, Amir R. Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-RNN: Deep Learning on Spatio-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 116,
+ 194,
+ 478,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 194,
+ 478,
+ 207
+ ],
+ "score": 1.0,
+ "content": "Temporal Graphs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 212,
+ 504,
+ 234
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 213,
+ 505,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 213,
+ 505,
+ 224
+ ],
+ "score": 1.0,
+ "content": "Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 116,
+ 222,
+ 379,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 222,
+ 379,
+ 234
+ ],
+ "score": 1.0,
+ "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 240,
+ 503,
+ 261
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "Thomas N. Kipf and Max Welling. Semi-Supervised Classification with Graph Convolutional Networks.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 250,
+ 209,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 250,
+ 209,
+ 261
+ ],
+ "score": 1.0,
+ "content": "arXiv:1609.02907, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 268,
+ 503,
+ 289
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 268,
+ 504,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 268,
+ 504,
+ 281
+ ],
+ "score": 1.0,
+ "content": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 279,
+ 262,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 279,
+ 262,
+ 289
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:1511.05493, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 295,
+ 504,
+ 317
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 504,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 504,
+ 308
+ ],
+ "score": 1.0,
+ "content": "Xiaodan Liang, Xiaohui Shen, Jiashi Feng, Liang Lin, and Shuicheng Yan. Semantic object parsing with graph",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 306,
+ 229,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 306,
+ 229,
+ 317
+ ],
+ "score": 1.0,
+ "content": "lstm. arXiv:1603.07063, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 324,
+ 505,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 505,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 505,
+ 335
+ ],
+ "score": 1.0,
+ "content": "Jonathan Masci, Davide Boscaini, Michael M. Bronstein, and Pierre Vandergheynst. Geodesic convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 333,
+ 505,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 333,
+ 505,
+ 346
+ ],
+ "score": 1.0,
+ "content": "neural networks on riemannian manifolds. In IEEE International Conference on Computer Vision (ICCV)",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 116,
+ 344,
+ 182,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 344,
+ 182,
+ 355
+ ],
+ "score": 1.0,
+ "content": "Workshops, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 361,
+ 504,
+ 383
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 374
+ ],
+ "score": 1.0,
+ "content": "T. Mikolov, K. Chen, G. Corrado, and J. Dean. Estimation of Word Representations in Vector Space. In",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 116,
+ 372,
+ 368,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 372,
+ 368,
+ 383
+ ],
+ "score": 1.0,
+ "content": "International Conference on Learning Representations (ICLR), 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 389,
+ 504,
+ 411
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 506,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 506,
+ 402
+ ],
+ "score": 1.0,
+ "content": "Mathias Niepert, Mohamed Ahmed, and Konstantin Kutzkov. Learning Convolutional Neural Networks for",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 117,
+ 400,
+ 385,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 400,
+ 385,
+ 411
+ ],
+ "score": 1.0,
+ "content": "Graphs. In International Conference on Machine Learning (ICML), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 417,
+ 504,
+ 439
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 428,
+ 469,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 428,
+ 469,
+ 439
+ ],
+ "score": 1.0,
+ "content": "(language) modeling: a baseline for generative models of natural videos. arXiv:1412.6604, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 445,
+ 504,
+ 467
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 505,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 505,
+ 458
+ ],
+ "score": 1.0,
+ "content": "Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 116,
+ 456,
+ 370,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 456,
+ 370,
+ 466
+ ],
+ "score": 1.0,
+ "content": "neural network model. IEEE Transactions on Neural Networks, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 473,
+ 504,
+ 505
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 483,
+ 506,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 483,
+ 506,
+ 496
+ ],
+ "score": 1.0,
+ "content": "LSTM Network: A Machine Learning Approach for Precipitation Nowcasting. In Advances in Neural",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 116,
+ 493,
+ 286,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 493,
+ 286,
+ 505
+ ],
+ "score": 1.0,
+ "content": "Information Processing Systems (NIPS), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 511,
+ 505,
+ 542
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "D. Shuman, S. Narang, P. Frossard, A. Ortega, and P. Vandergheynst. The Emerging Field of Signal Processing",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 522,
+ 505,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 522,
+ 505,
+ 533
+ ],
+ "score": 1.0,
+ "content": "on Graphs: Extending High-Dimensional Data Analysis to Networks and other Irregular Domains. IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 531,
+ 246,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 531,
+ 246,
+ 543
+ ],
+ "score": 1.0,
+ "content": "Signal Processing Magazine, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 549,
+ 504,
+ 570
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 548,
+ 505,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 548,
+ 505,
+ 561
+ ],
+ "score": 1.0,
+ "content": "Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 559,
+ 399,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 559,
+ 399,
+ 571
+ ],
+ "score": 1.0,
+ "content": "using lstms. In International Conference on Machine Learning (ICML), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 577,
+ 503,
+ 598
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 506,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 506,
+ 590
+ ],
+ "score": 1.0,
+ "content": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 587,
+ 355,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 587,
+ 355,
+ 599
+ ],
+ "score": 1.0,
+ "content": "Advances in neural information processing systems (NIPS), 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 605,
+ 504,
+ 626
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "Kai Sheng Tai, Richard Socher, and Christopher D. Manning. Improved semantic representations from tree-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 615,
+ 497,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 615,
+ 497,
+ 627
+ ],
+ "score": 1.0,
+ "content": "structured long short-term memory networks. In Association for Computational Linguistics (ACL), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 633,
+ 504,
+ 654
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 646
+ ],
+ "score": 1.0,
+ "content": "Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 643,
+ 449,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 643,
+ 449,
+ 655
+ ],
+ "score": 1.0,
+ "content": "generator. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 660,
+ 503,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. Recurrent neural network regularization.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 670,
+ 205,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 670,
+ 205,
+ 681
+ ],
+ "score": 1.0,
+ "content": "arXiv:1409.2329, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42.5
+ }
+ ],
+ "page_idx": 9,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "10",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2017",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 83,
+ 503,
+ 114
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 92,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 92,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and descrip-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 116,
+ 103,
+ 429,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 103,
+ 429,
+ 115
+ ],
+ "score": 1.0,
+ "content": "tion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 105,
+ 83,
+ 505,
+ 115
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 120,
+ 503,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 505,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 505,
+ 132
+ ],
+ "score": 1.0,
+ "content": "Felix A Gers and Jurgen Schmidhuber. Recurrent nets that time and count. In ¨ IEEE-INNS-ENNS International",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 131,
+ 278,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 131,
+ 278,
+ 142
+ ],
+ "score": 1.0,
+ "content": "Joint Conference on Neural Networks, 2000.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 106,
+ 121,
+ 505,
+ 142
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 149,
+ 441,
+ 160
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 149,
+ 441,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 149,
+ 441,
+ 161
+ ],
+ "score": 1.0,
+ "content": "Alex Graves. Generating sequences with recurrent neural networks. arXiv:1308.0850, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 106,
+ 149,
+ 441,
+ 161
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 167,
+ 455,
+ 178
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 166,
+ 456,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 166,
+ 456,
+ 180
+ ],
+ "score": 1.0,
+ "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 1997.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 166,
+ 456,
+ 180
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 184,
+ 502,
+ 205
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 504,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 504,
+ 196
+ ],
+ "score": 1.0,
+ "content": "Ashesh Jain, Amir R. Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-RNN: Deep Learning on Spatio-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 116,
+ 194,
+ 478,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 194,
+ 478,
+ 207
+ ],
+ "score": 1.0,
+ "content": "Temporal Graphs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 106,
+ 184,
+ 504,
+ 207
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 212,
+ 504,
+ 234
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 213,
+ 505,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 213,
+ 505,
+ 224
+ ],
+ "score": 1.0,
+ "content": "Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 116,
+ 222,
+ 379,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 222,
+ 379,
+ 234
+ ],
+ "score": 1.0,
+ "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 106,
+ 213,
+ 505,
+ 234
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 240,
+ 503,
+ 261
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "Thomas N. Kipf and Max Welling. Semi-Supervised Classification with Graph Convolutional Networks.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 250,
+ 209,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 250,
+ 209,
+ 261
+ ],
+ "score": 1.0,
+ "content": "arXiv:1609.02907, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 106,
+ 240,
+ 505,
+ 261
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 268,
+ 503,
+ 289
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 268,
+ 504,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 268,
+ 504,
+ 281
+ ],
+ "score": 1.0,
+ "content": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 279,
+ 262,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 279,
+ 262,
+ 289
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:1511.05493, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 106,
+ 268,
+ 504,
+ 289
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 295,
+ 504,
+ 317
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 504,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 504,
+ 308
+ ],
+ "score": 1.0,
+ "content": "Xiaodan Liang, Xiaohui Shen, Jiashi Feng, Liang Lin, and Shuicheng Yan. Semantic object parsing with graph",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 306,
+ 229,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 306,
+ 229,
+ 317
+ ],
+ "score": 1.0,
+ "content": "lstm. arXiv:1603.07063, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 105,
+ 294,
+ 504,
+ 317
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 324,
+ 505,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 505,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 505,
+ 335
+ ],
+ "score": 1.0,
+ "content": "Jonathan Masci, Davide Boscaini, Michael M. Bronstein, and Pierre Vandergheynst. Geodesic convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 333,
+ 505,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 333,
+ 505,
+ 346
+ ],
+ "score": 1.0,
+ "content": "neural networks on riemannian manifolds. In IEEE International Conference on Computer Vision (ICCV)",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 116,
+ 344,
+ 182,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 344,
+ 182,
+ 355
+ ],
+ "score": 1.0,
+ "content": "Workshops, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18,
+ "bbox_fs": [
+ 106,
+ 325,
+ 505,
+ 355
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 361,
+ 504,
+ 383
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 374
+ ],
+ "score": 1.0,
+ "content": "T. Mikolov, K. Chen, G. Corrado, and J. Dean. Estimation of Word Representations in Vector Space. In",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 116,
+ 372,
+ 368,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 372,
+ 368,
+ 383
+ ],
+ "score": 1.0,
+ "content": "International Conference on Learning Representations (ICLR), 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 105,
+ 361,
+ 506,
+ 383
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 389,
+ 504,
+ 411
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 506,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 506,
+ 402
+ ],
+ "score": 1.0,
+ "content": "Mathias Niepert, Mohamed Ahmed, and Konstantin Kutzkov. Learning Convolutional Neural Networks for",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 117,
+ 400,
+ 385,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 400,
+ 385,
+ 411
+ ],
+ "score": 1.0,
+ "content": "Graphs. In International Conference on Machine Learning (ICML), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 105,
+ 389,
+ 506,
+ 411
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 417,
+ 504,
+ 439
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 428,
+ 469,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 428,
+ 469,
+ 439
+ ],
+ "score": 1.0,
+ "content": "(language) modeling: a baseline for generative models of natural videos. arXiv:1412.6604, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 105,
+ 417,
+ 505,
+ 439
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 445,
+ 504,
+ 467
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 505,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 505,
+ 458
+ ],
+ "score": 1.0,
+ "content": "Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 116,
+ 456,
+ 370,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 456,
+ 370,
+ 466
+ ],
+ "score": 1.0,
+ "content": "neural network model. IEEE Transactions on Neural Networks, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 444,
+ 505,
+ 466
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 473,
+ 504,
+ 505
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 483,
+ 506,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 483,
+ 506,
+ 496
+ ],
+ "score": 1.0,
+ "content": "LSTM Network: A Machine Learning Approach for Precipitation Nowcasting. In Advances in Neural",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 116,
+ 493,
+ 286,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 493,
+ 286,
+ 505
+ ],
+ "score": 1.0,
+ "content": "Information Processing Systems (NIPS), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 29,
+ "bbox_fs": [
+ 106,
+ 474,
+ 506,
+ 505
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 511,
+ 505,
+ 542
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "D. Shuman, S. Narang, P. Frossard, A. Ortega, and P. Vandergheynst. The Emerging Field of Signal Processing",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 522,
+ 505,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 522,
+ 505,
+ 533
+ ],
+ "score": 1.0,
+ "content": "on Graphs: Extending High-Dimensional Data Analysis to Networks and other Irregular Domains. IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 531,
+ 246,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 531,
+ 246,
+ 543
+ ],
+ "score": 1.0,
+ "content": "Signal Processing Magazine, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32,
+ "bbox_fs": [
+ 105,
+ 509,
+ 505,
+ 543
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 549,
+ 504,
+ 570
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 548,
+ 505,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 548,
+ 505,
+ 561
+ ],
+ "score": 1.0,
+ "content": "Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 559,
+ 399,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 559,
+ 399,
+ 571
+ ],
+ "score": 1.0,
+ "content": "using lstms. In International Conference on Machine Learning (ICML), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34.5,
+ "bbox_fs": [
+ 105,
+ 548,
+ 505,
+ 571
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 577,
+ 503,
+ 598
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 506,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 506,
+ 590
+ ],
+ "score": 1.0,
+ "content": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 587,
+ 355,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 587,
+ 355,
+ 599
+ ],
+ "score": 1.0,
+ "content": "Advances in neural information processing systems (NIPS), 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 105,
+ 576,
+ 506,
+ 599
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 605,
+ 504,
+ 626
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "Kai Sheng Tai, Richard Socher, and Christopher D. Manning. Improved semantic representations from tree-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 615,
+ 497,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 615,
+ 497,
+ 627
+ ],
+ "score": 1.0,
+ "content": "structured long short-term memory networks. In Association for Computational Linguistics (ACL), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38.5,
+ "bbox_fs": [
+ 105,
+ 604,
+ 505,
+ 627
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 633,
+ 504,
+ 654
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 646
+ ],
+ "score": 1.0,
+ "content": "Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 643,
+ 449,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 643,
+ 449,
+ 655
+ ],
+ "score": 1.0,
+ "content": "generator. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 106,
+ 633,
+ 505,
+ 655
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 660,
+ 503,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. Recurrent neural network regularization.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 670,
+ 205,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 670,
+ 205,
+ 681
+ ],
+ "score": 1.0,
+ "content": "arXiv:1409.2329, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 106,
+ 660,
+ 505,
+ 681
+ ]
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file