diff --git "a/parse/train/5FRJWsiLRmA/5FRJWsiLRmA_middle.json" "b/parse/train/5FRJWsiLRmA/5FRJWsiLRmA_middle.json" new file mode 100644--- /dev/null +++ "b/parse/train/5FRJWsiLRmA/5FRJWsiLRmA_middle.json" @@ -0,0 +1,40933 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 318, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 320, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 320, + 98 + ], + "score": 1.0, + "content": "RESERVOIR TRANSFORMERS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 115, + 244, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 116, + 201, + 127 + ], + "spans": [ + { + "bbox": [ + 113, + 116, + 201, + 127 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 112, + 126, + 245, + 138 + ], + "spans": [ + { + "bbox": [ + 112, + 126, + 245, + 138 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 278, + 167, + 333, + 178 + ], + "lines": [ + { + "bbox": [ + 276, + 165, + 336, + 181 + ], + "spans": [ + { + "bbox": [ + 276, + 165, + 336, + 181 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 194, + 468, + 259 + ], + "lines": [ + { + "bbox": [ + 142, + 194, + 469, + 206 + ], + "spans": [ + { + "bbox": [ + 142, + 194, + 469, + 206 + ], + "score": 1.0, + "content": "We demonstrate that transformers obtain impressive performance even when some", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 205, + 469, + 217 + ], + "spans": [ + { + "bbox": [ + 142, + 205, + 469, + 217 + ], + "score": 1.0, + "content": "of the layers are randomly initialized and never updated. Inspired by old and well-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 215, + 470, + 228 + ], + "spans": [ + { + "bbox": [ + 141, + 215, + 470, + 228 + ], + "score": 1.0, + "content": "established ideas in machine learning, we explore a variety of non-linear “reser-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 227, + 469, + 239 + ], + "spans": [ + { + "bbox": [ + 142, + 227, + 469, + 239 + ], + "score": 1.0, + "content": "voir” layers interspersed with regular transformer layers, and show improvements", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 237, + 469, + 250 + ], + "spans": [ + { + "bbox": [ + 141, + 237, + 469, + 250 + ], + "score": 1.0, + "content": "in wall-clock compute time until convergence, as well as overall performance, on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 248, + 417, + 261 + ], + "spans": [ + { + "bbox": [ + 141, + 248, + 417, + 261 + ], + "score": 1.0, + "content": "various machine translation and (masked) language modelling tasks.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 108, + 286, + 205, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 208, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 208, + 302 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 505, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "score": 1.0, + "content": "Transformers (Vaswani et al., 2017) have dominated natural language processing (NLP) in recent", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "score": 1.0, + "content": "years, from large scale machine translation (Ott et al., 2018) to pre-trained (masked) language mod-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "score": 1.0, + "content": "eling (Devlin et al., 2018; Radford et al., 2018), and are becoming more popular in other fields as", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "score": 1.0, + "content": "well, from reinforcement learning (Vinyals et al., 2019) to speech recognition (Baevski et al., 2019)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "and computer vision (Carion et al., 2020). Their success is enabled in part by ever increasing com-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "putational demands, which has naturally led to an increased interest in improving their efficiency.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 379, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 390 + ], + "score": 1.0, + "content": "Scalability gains in transformers could facilitate bigger, deeper networks with longer contexts (Ki-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "taev et al., 2020; Wang et al., 2020; Beltagy et al., 2020; Kaplan et al., 2020; Tay et al., 2020b).", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "Conversely, improved efficiency could reduce environmental costs (Strubell et al., 2019) and hope-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 411, + 262, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 262, + 425 + ], + "score": 1.0, + "content": "fully help democratize the technology.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 429, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "In this work, we explore a simple question: if some layers of the transformer are kept frozen—i.e.,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 440, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 451 + ], + "score": 1.0, + "content": "never updated after random initialization—can we match the performance of fully learned trans-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "score": 1.0, + "content": "formers, while being more efficient? Surprisingly, the answer is resoundingly yes; and what is", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 461, + 387, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 387, + 475 + ], + "score": 1.0, + "content": "more, we find that freezing layers may actually improve performance.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 492 + ], + "score": 1.0, + "content": "Beyond desirable efficiency gains, random layers are interesting for several additional reasons.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 489, + 504, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 504, + 502 + ], + "score": 1.0, + "content": "Fixed randomly initialized networks (Gallicchio & Scardapane, 2020) converge to Gaussian pro-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "cesses in the limit of infinite width (Daniely et al., 2016), have intriguing interpretations in metric", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 512, + 504, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 504, + 523 + ], + "score": 1.0, + "content": "learning (Rosenfeld & Tsotsos, 2019; Giryes et al., 2016), and have been shown to provide excel-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "lent “priors” either for subsequent learning (Ulyanov et al., 2018) or pruning (Frankle & Carbin,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "2018). Fixed layers allow for efficient low-cost hardware implementations (Schrauwen et al., 2007)", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 543, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 558 + ], + "score": 1.0, + "content": "and can be characterized using only a random number generator and its seed, which might have", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 555, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 567 + ], + "score": 1.0, + "content": "repercussions in distributed training and enables highly efficient deployment to edge devices. The", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "score": 1.0, + "content": "strong performance of networks with fixed layers also sheds new light on the inner workings of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 577, + 504, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 504, + 589 + ], + "score": 1.0, + "content": "BERT (Devlin et al., 2018), and layer-wise interpretations of such models (Rogers et al., 2020; Ten-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "ney et al., 2019). It appears that “not all layers are created equal” (Zhang et al., 2019) is true to such", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 599, + 363, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 363, + 611 + ], + "score": 1.0, + "content": "an extent that some layers can simply remain random and fixed.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 504, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "These ideas have a long history in machine learning. By Cover’s theorem (Cover, 1965), any high-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "dimensional non-linear transformation is more likely to be linearly separable than its lower-or-equal-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "dimensional input space. By Johnson-Lindenstrauss (Johnson & Lindenstrauss, 1984), random pro-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "jections distort Euclidean distances very little under mild assumptions, which is useful e.g. for", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "dimensionality reduction and random indexing (Sahlgren, 2005). Fixed random layers in neural", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 671, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 682 + ], + "score": 1.0, + "content": "networks pre-date deep learning by far (Gamba et al., 1961; Baum, 1988). Indeed, random kernel", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 682, + 465, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 465, + 694 + ], + "score": 1.0, + "content": "methods have been an impactful idea in machine learning (Rahimi & Recht, 2008; 2009).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 697, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 697, + 505, + 712 + ], + "score": 1.0, + "content": "One way to think of such layers is as “reservoirs” (Lukosevi ˇ cius & Jaeger, 2009), where a highly ˇ", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "non-linear high-dimensional black box representation is provided to a lightweight “readout” net-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "work, as in echo state networks (Jaeger, 2003) and liquid state machines (Maass et al., 2002). The", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 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 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 318, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 320, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 320, + 98 + ], + "score": 1.0, + "content": "RESERVOIR TRANSFORMERS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 115, + 244, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 116, + 201, + 127 + ], + "spans": [ + { + "bbox": [ + 113, + 116, + 201, + 127 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 112, + 126, + 245, + 138 + ], + "spans": [ + { + "bbox": [ + 112, + 126, + 245, + 138 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 112, + 116, + 245, + 138 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 167, + 333, + 178 + ], + "lines": [ + { + "bbox": [ + 276, + 165, + 336, + 181 + ], + "spans": [ + { + "bbox": [ + 276, + 165, + 336, + 181 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 194, + 468, + 259 + ], + "lines": [ + { + "bbox": [ + 142, + 194, + 469, + 206 + ], + "spans": [ + { + "bbox": [ + 142, + 194, + 469, + 206 + ], + "score": 1.0, + "content": "We demonstrate that transformers obtain impressive performance even when some", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 205, + 469, + 217 + ], + "spans": [ + { + "bbox": [ + 142, + 205, + 469, + 217 + ], + "score": 1.0, + "content": "of the layers are randomly initialized and never updated. Inspired by old and well-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 215, + 470, + 228 + ], + "spans": [ + { + "bbox": [ + 141, + 215, + 470, + 228 + ], + "score": 1.0, + "content": "established ideas in machine learning, we explore a variety of non-linear “reser-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 227, + 469, + 239 + ], + "spans": [ + { + "bbox": [ + 142, + 227, + 469, + 239 + ], + "score": 1.0, + "content": "voir” layers interspersed with regular transformer layers, and show improvements", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 237, + 469, + 250 + ], + "spans": [ + { + "bbox": [ + 141, + 237, + 469, + 250 + ], + "score": 1.0, + "content": "in wall-clock compute time until convergence, as well as overall performance, on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 248, + 417, + 261 + ], + "spans": [ + { + "bbox": [ + 141, + 248, + 417, + 261 + ], + "score": 1.0, + "content": "various machine translation and (masked) language modelling tasks.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5, + "bbox_fs": [ + 141, + 194, + 470, + 261 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 286, + 205, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 208, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 208, + 302 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 505, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "score": 1.0, + "content": "Transformers (Vaswani et al., 2017) have dominated natural language processing (NLP) in recent", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "score": 1.0, + "content": "years, from large scale machine translation (Ott et al., 2018) to pre-trained (masked) language mod-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "score": 1.0, + "content": "eling (Devlin et al., 2018; Radford et al., 2018), and are becoming more popular in other fields as", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "score": 1.0, + "content": "well, from reinforcement learning (Vinyals et al., 2019) to speech recognition (Baevski et al., 2019)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "and computer vision (Carion et al., 2020). Their success is enabled in part by ever increasing com-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "putational demands, which has naturally led to an increased interest in improving their efficiency.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 379, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 390 + ], + "score": 1.0, + "content": "Scalability gains in transformers could facilitate bigger, deeper networks with longer contexts (Ki-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "taev et al., 2020; Wang et al., 2020; Beltagy et al., 2020; Kaplan et al., 2020; Tay et al., 2020b).", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "Conversely, improved efficiency could reduce environmental costs (Strubell et al., 2019) and hope-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 411, + 262, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 262, + 425 + ], + "score": 1.0, + "content": "fully help democratize the technology.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 312, + 505, + 425 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 429, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "In this work, we explore a simple question: if some layers of the transformer are kept frozen—i.e.,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 440, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 451 + ], + "score": 1.0, + "content": "never updated after random initialization—can we match the performance of fully learned trans-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "score": 1.0, + "content": "formers, while being more efficient? Surprisingly, the answer is resoundingly yes; and what is", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 461, + 387, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 387, + 475 + ], + "score": 1.0, + "content": "more, we find that freezing layers may actually improve performance.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 428, + 506, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 492 + ], + "score": 1.0, + "content": "Beyond desirable efficiency gains, random layers are interesting for several additional reasons.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 489, + 504, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 504, + 502 + ], + "score": 1.0, + "content": "Fixed randomly initialized networks (Gallicchio & Scardapane, 2020) converge to Gaussian pro-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "cesses in the limit of infinite width (Daniely et al., 2016), have intriguing interpretations in metric", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 512, + 504, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 504, + 523 + ], + "score": 1.0, + "content": "learning (Rosenfeld & Tsotsos, 2019; Giryes et al., 2016), and have been shown to provide excel-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "lent “priors” either for subsequent learning (Ulyanov et al., 2018) or pruning (Frankle & Carbin,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "2018). Fixed layers allow for efficient low-cost hardware implementations (Schrauwen et al., 2007)", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 543, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 558 + ], + "score": 1.0, + "content": "and can be characterized using only a random number generator and its seed, which might have", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 555, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 567 + ], + "score": 1.0, + "content": "repercussions in distributed training and enables highly efficient deployment to edge devices. The", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "score": 1.0, + "content": "strong performance of networks with fixed layers also sheds new light on the inner workings of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 577, + 504, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 504, + 589 + ], + "score": 1.0, + "content": "BERT (Devlin et al., 2018), and layer-wise interpretations of such models (Rogers et al., 2020; Ten-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "ney et al., 2019). It appears that “not all layers are created equal” (Zhang et al., 2019) is true to such", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 599, + 363, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 363, + 611 + ], + "score": 1.0, + "content": "an extent that some layers can simply remain random and fixed.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 477, + 506, + 611 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 504, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "These ideas have a long history in machine learning. By Cover’s theorem (Cover, 1965), any high-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "dimensional non-linear transformation is more likely to be linearly separable than its lower-or-equal-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "dimensional input space. By Johnson-Lindenstrauss (Johnson & Lindenstrauss, 1984), random pro-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "jections distort Euclidean distances very little under mild assumptions, which is useful e.g. for", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "dimensionality reduction and random indexing (Sahlgren, 2005). Fixed random layers in neural", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 671, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 682 + ], + "score": 1.0, + "content": "networks pre-date deep learning by far (Gamba et al., 1961; Baum, 1988). Indeed, random kernel", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 682, + 465, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 465, + 694 + ], + "score": 1.0, + "content": "methods have been an impactful idea in machine learning (Rahimi & Recht, 2008; 2009).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 615, + 506, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 697, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 697, + 505, + 712 + ], + "score": 1.0, + "content": "One way to think of such layers is as “reservoirs” (Lukosevi ˇ cius & Jaeger, 2009), where a highly ˇ", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "non-linear high-dimensional black box representation is provided to a lightweight “readout” net-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "work, as in echo state networks (Jaeger, 2003) and liquid state machines (Maass et al., 2002). The", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "benefit of such an approach is that the reservoir has fixed parameters and is computationally efficient,", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 419, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 419, + 106 + ], + "score": 1.0, + "content": "as it can be pre-computed and does not (necessarily) require backpropagation.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 45, + "bbox_fs": [ + 106, + 697, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "benefit of such an approach is that the reservoir has fixed parameters and is computationally efficient,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 419, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 419, + 106 + ], + "score": 1.0, + "content": "as it can be pre-computed and does not (necessarily) require backpropagation.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "score": 1.0, + "content": "In NLP, Wieting & Kiela (2019) showed that random sentence encoders present a strong baseline for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 135 + ], + "score": 1.0, + "content": "text classification, with subsequent work showing applications in a variety of NLP tasks (Enguehard", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "et al., 2019; Garg et al., 2020; Pilault et al., 2020). To our knowledge, this work is the first to examine", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "this phenomenon in transformers, and the first to recursively alternate reservoirs with subsequent", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 166 + ], + "score": 1.0, + "content": "transformer layers acting as readout functions. We introduce “reservoir transformers”, wherein fixed", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "random reservoir layers are interspersed with regular updateable transformer layers. The goal of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "score": 1.0, + "content": "this work is not necessarily to set a new state of the art, but to put our understanding of transformer", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "models on a more solid footing by providing empirical evidence of their capabilities even when", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 381, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 381, + 210 + ], + "score": 1.0, + "content": "some of their parameters are fixed. Our contributions are as follows:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 132, + 220, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 135, + 219, + 504, + 232 + ], + "spans": [ + { + "bbox": [ + 135, + 219, + 504, + 232 + ], + "score": 1.0, + "content": "• We introduce a new area under the convergence curve metric for measuring performance-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 231, + 504, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 231, + 504, + 244 + ], + "score": 1.0, + "content": "efficiency trade-offs, and show that replacing regular transformer layers with reservoir lay-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 242, + 303, + 254 + ], + "spans": [ + { + "bbox": [ + 142, + 242, + 303, + 254 + ], + "score": 1.0, + "content": "ers leads to better results on that metric.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 134, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "• We show that the addition of reservoir layers in fact leads to improved test set generalization", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 269, + 318, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 269, + 318, + 281 + ], + "score": 1.0, + "content": "on a variety of tasks in a variety of settings.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 137, + 283, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 137, + 283, + 505, + 295 + ], + "score": 1.0, + "content": "We show that pre-trained masked language modelling architectures like BERT and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 142, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "RoBERTa (Liu et al., 2019) can benefit from having some of their layers frozen, both", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 305, + 420, + 317 + ], + "spans": [ + { + "bbox": [ + 141, + 305, + 420, + 317 + ], + "score": 1.0, + "content": "during pre-training as well as when fine-tuning on downstream tasks.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "• In addition, we experiment with different types of reservoir layers, including convolutional", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 331, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 142, + 331, + 505, + 343 + ], + "score": 1.0, + "content": "and recurrent neural network-based ones. We also show empirical evidence that the back-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 142, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "ward pass can be entirely skipped by approximating top-layer gradients using an approach", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 353, + 423, + 366 + ], + "spans": [ + { + "bbox": [ + 142, + 353, + 423, + 366 + ], + "score": 1.0, + "content": "we call backskipping, with a relatively small sacrifice in performance.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 107, + 380, + 183, + 393 + ], + "lines": [ + { + "bbox": [ + 104, + 379, + 185, + 396 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 185, + 396 + ], + "score": 1.0, + "content": "2 APPROACH", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 406, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "This paper is based on a very simple idea. Neural networks are trained via backpropagation, which", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 416, + 383, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 383, + 430 + ], + "score": 1.0, + "content": "involves consecutive steps of matrix addition and multiplication, i.e.,", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 196, + 441, + 415, + 468 + ], + "lines": [ + { + "bbox": [ + 196, + 441, + 415, + 468 + ], + "spans": [ + { + "bbox": [ + 196, + 441, + 415, + 468 + ], + "score": 0.92, + "content": "\\theta _ { t + 1 } \\theta _ { t } - \\eta \\frac { \\partial J } { \\partial \\theta _ { t } } ; \\frac { \\partial J } { \\partial \\theta _ { t } } = \\frac { \\partial J } { \\partial L _ { n } } \\frac { \\partial L _ { n } } { \\partial L _ { n - 1 } } \\cdot \\cdot \\cdot \\frac { \\partial L _ { 1 } } { \\partial L _ { 0 } } \\frac { \\partial L _ { 0 } } { \\partial x }", + "type": "interline_equation", + "image_path": "23b84a2a7e961774c2016abd8eaae52ef87ba9ba1fd35110f827aa7126795c23.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 441, + 415, + 454.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 196, + 454.5, + 415, + 468.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 506, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 188, + 489 + ], + "score": 1.0, + "content": "for some objective", + "type": "text" + }, + { + "bbox": [ + 189, + 477, + 196, + 486 + ], + "score": 0.79, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 475, + 273, + 489 + ], + "score": 1.0, + "content": ", parameterization", + "type": "text" + }, + { + "bbox": [ + 273, + 476, + 280, + 486 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 475, + 358, + 489 + ], + "score": 1.0, + "content": "and learning rate", + "type": "text" + }, + { + "bbox": [ + 358, + 478, + 365, + 488 + ], + "score": 0.81, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 475, + 506, + 489 + ], + "score": 1.0, + "content": ", with the gradient computed via", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 199, + 499 + ], + "score": 1.0, + "content": "the chain rule, where", + "type": "text" + }, + { + "bbox": [ + 199, + 487, + 211, + 498 + ], + "score": 0.88, + "content": "L _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 486, + 240, + 499 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 241, + 488, + 246, + 497 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 486, + 396, + 499 + ], + "score": 1.0, + "content": "-th layer of the neural network and", + "type": "text" + }, + { + "bbox": [ + 397, + 489, + 404, + 497 + ], + "score": 0.77, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 486, + 482, + 499 + ], + "score": 1.0, + "content": "is the input. Let", + "type": "text" + }, + { + "bbox": [ + 482, + 487, + 505, + 498 + ], + "score": 0.83, + "content": "L \\ =", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 497, + 456, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 157, + 511 + ], + "score": 1.0, + "content": "Transformer", + "type": "text" + }, + { + "bbox": [ + 157, + 498, + 174, + 510 + ], + "score": 0.84, + "content": "( X )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 497, + 456, + 511 + ], + "score": 1.0, + "content": "be a single layer in a Transformer network (Vaswani et al., 2017), i.e.,", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 528, + 403, + 559 + ], + "lines": [ + { + "bbox": [ + 208, + 528, + 403, + 559 + ], + "spans": [ + { + "bbox": [ + 208, + 528, + 403, + 559 + ], + "score": 0.88, + "content": "\\begin{array} { r } { H = \\mathbf { M } \\mathbf { u } ] \\mathrm { t i } \\mathbf { H e a d } \\mathbf { S e l f A t t n } ( \\mathbf { L a y e r N o r m } ( X ) ) + X } \\\\ { L = \\mathbf { F F N } ( \\mathbf { L a y e r N o r m } ( H ) ) + H } \\end{array}", + "type": "interline_equation", + "image_path": "06ee18caa5e9791e3351b33dd81d6b538c4f644c50d2d27558db3d31d1aa49a7.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 528, + 403, + 543.5 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 208, + 543.5, + 403, + 559.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 570, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 420, + 582 + ], + "score": 1.0, + "content": "Now, during every “backward pass”, we compute the Jacobian for parameters", + "type": "text" + }, + { + "bbox": [ + 420, + 569, + 432, + 581 + ], + "score": 0.89, + "content": "\\theta ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 569, + 466, + 582 + ], + "score": 1.0, + "content": "at layer", + "type": "text" + }, + { + "bbox": [ + 466, + 571, + 474, + 580 + ], + "score": 0.79, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 569, + 505, + 582 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 579, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 258, + 596 + ], + "score": 1.0, + "content": "are used to update the parameters of", + "type": "text" + }, + { + "bbox": [ + 258, + 581, + 266, + 591 + ], + "score": 0.57, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 579, + 270, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 271, + 581, + 283, + 593 + ], + "score": 0.79, + "content": "\\theta _ { t } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 579, + 506, + 596 + ], + "score": 1.0, + "content": ", as well as to compute the next layer’s Jacobian, thus", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 591, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 506, + 606 + ], + "score": 1.0, + "content": "back-propagating the gradients. In this work however, for some of the layers, we still backpropagate", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 603, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 506, + 615 + ], + "score": 1.0, + "content": "through them to compute gradients for earlier layers, but we never update their parameters. As a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 614, + 473, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 473, + 626 + ], + "score": 1.0, + "content": "result, these layers stay fixed at their random initialization, saving computational resources.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 107, + 639, + 194, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 194, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 194, + 651 + ], + "score": 1.0, + "content": "2.1 BACKGROUND", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "Naturally, never updating some of the parameters is computationally more efficient, as some matrix", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "addition operations can be skipped in the backward pass, but why is this not detrimental to the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 682, + 224, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 224, + 692 + ], + "score": 1.0, + "content": "performance of the network?", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "In the early days of neural networks, the bottom layers were often kept fixed as “associators” (Block,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "1962), or what Minsky & Papert (2017) called the Gamba perceptron (Gamba et al., 1961; Borsellino", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "& Gamba, 1961). Fixed random networks (Baum, 1988; Schmidt et al., 1992; Pao et al., 1994) have", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 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 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 503, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "score": 1.0, + "content": "In NLP, Wieting & Kiela (2019) showed that random sentence encoders present a strong baseline for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 135 + ], + "score": 1.0, + "content": "text classification, with subsequent work showing applications in a variety of NLP tasks (Enguehard", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "et al., 2019; Garg et al., 2020; Pilault et al., 2020). To our knowledge, this work is the first to examine", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "this phenomenon in transformers, and the first to recursively alternate reservoirs with subsequent", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 166 + ], + "score": 1.0, + "content": "transformer layers acting as readout functions. We introduce “reservoir transformers”, wherein fixed", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "random reservoir layers are interspersed with regular updateable transformer layers. The goal of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "score": 1.0, + "content": "this work is not necessarily to set a new state of the art, but to put our understanding of transformer", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "models on a more solid footing by providing empirical evidence of their capabilities even when", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 381, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 381, + 210 + ], + "score": 1.0, + "content": "some of their parameters are fixed. Our contributions are as follows:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 110, + 506, + 210 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 220, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 135, + 219, + 504, + 232 + ], + "spans": [ + { + "bbox": [ + 135, + 219, + 504, + 232 + ], + "score": 1.0, + "content": "• We introduce a new area under the convergence curve metric for measuring performance-", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 231, + 504, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 231, + 504, + 244 + ], + "score": 1.0, + "content": "efficiency trade-offs, and show that replacing regular transformer layers with reservoir lay-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 242, + 303, + 254 + ], + "spans": [ + { + "bbox": [ + 142, + 242, + 303, + 254 + ], + "score": 1.0, + "content": "ers leads to better results on that metric.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 134, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 134, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "• We show that the addition of reservoir layers in fact leads to improved test set generalization", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 269, + 318, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 269, + 318, + 281 + ], + "score": 1.0, + "content": "on a variety of tasks in a variety of settings.", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 137, + 283, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 137, + 283, + 505, + 295 + ], + "score": 1.0, + "content": "We show that pre-trained masked language modelling architectures like BERT and", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 142, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "RoBERTa (Liu et al., 2019) can benefit from having some of their layers frozen, both", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 305, + 420, + 317 + ], + "spans": [ + { + "bbox": [ + 141, + 305, + 420, + 317 + ], + "score": 1.0, + "content": "during pre-training as well as when fine-tuning on downstream tasks.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "• In addition, we experiment with different types of reservoir layers, including convolutional", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 331, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 142, + 331, + 505, + 343 + ], + "score": 1.0, + "content": "and recurrent neural network-based ones. We also show empirical evidence that the back-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 142, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "ward pass can be entirely skipped by approximating top-layer gradients using an approach", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 353, + 423, + 366 + ], + "spans": [ + { + "bbox": [ + 142, + 353, + 423, + 366 + ], + "score": 1.0, + "content": "we call backskipping, with a relatively small sacrifice in performance.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + } + ], + "index": 16.5, + "bbox_fs": [ + 132, + 219, + 505, + 366 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 380, + 183, + 393 + ], + "lines": [ + { + "bbox": [ + 104, + 379, + 185, + 396 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 185, + 396 + ], + "score": 1.0, + "content": "2 APPROACH", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 406, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "This paper is based on a very simple idea. Neural networks are trained via backpropagation, which", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 416, + 383, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 383, + 430 + ], + "score": 1.0, + "content": "involves consecutive steps of matrix addition and multiplication, i.e.,", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 406, + 505, + 430 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 196, + 441, + 415, + 468 + ], + "lines": [ + { + "bbox": [ + 196, + 441, + 415, + 468 + ], + "spans": [ + { + "bbox": [ + 196, + 441, + 415, + 468 + ], + "score": 0.92, + "content": "\\theta _ { t + 1 } \\theta _ { t } - \\eta \\frac { \\partial J } { \\partial \\theta _ { t } } ; \\frac { \\partial J } { \\partial \\theta _ { t } } = \\frac { \\partial J } { \\partial L _ { n } } \\frac { \\partial L _ { n } } { \\partial L _ { n - 1 } } \\cdot \\cdot \\cdot \\frac { \\partial L _ { 1 } } { \\partial L _ { 0 } } \\frac { \\partial L _ { 0 } } { \\partial x }", + "type": "interline_equation", + "image_path": "23b84a2a7e961774c2016abd8eaae52ef87ba9ba1fd35110f827aa7126795c23.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 441, + 415, + 454.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 196, + 454.5, + 415, + 468.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 506, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 188, + 489 + ], + "score": 1.0, + "content": "for some objective", + "type": "text" + }, + { + "bbox": [ + 189, + 477, + 196, + 486 + ], + "score": 0.79, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 475, + 273, + 489 + ], + "score": 1.0, + "content": ", parameterization", + "type": "text" + }, + { + "bbox": [ + 273, + 476, + 280, + 486 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 475, + 358, + 489 + ], + "score": 1.0, + "content": "and learning rate", + "type": "text" + }, + { + "bbox": [ + 358, + 478, + 365, + 488 + ], + "score": 0.81, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 475, + 506, + 489 + ], + "score": 1.0, + "content": ", with the gradient computed via", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 199, + 499 + ], + "score": 1.0, + "content": "the chain rule, where", + "type": "text" + }, + { + "bbox": [ + 199, + 487, + 211, + 498 + ], + "score": 0.88, + "content": "L _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 486, + 240, + 499 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 241, + 488, + 246, + 497 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 486, + 396, + 499 + ], + "score": 1.0, + "content": "-th layer of the neural network and", + "type": "text" + }, + { + "bbox": [ + 397, + 489, + 404, + 497 + ], + "score": 0.77, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 486, + 482, + 499 + ], + "score": 1.0, + "content": "is the input. Let", + "type": "text" + }, + { + "bbox": [ + 482, + 487, + 505, + 498 + ], + "score": 0.83, + "content": "L \\ =", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 497, + 456, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 157, + 511 + ], + "score": 1.0, + "content": "Transformer", + "type": "text" + }, + { + "bbox": [ + 157, + 498, + 174, + 510 + ], + "score": 0.84, + "content": "( X )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 497, + 456, + 511 + ], + "score": 1.0, + "content": "be a single layer in a Transformer network (Vaswani et al., 2017), i.e.,", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 475, + 506, + 511 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 528, + 403, + 559 + ], + "lines": [ + { + "bbox": [ + 208, + 528, + 403, + 559 + ], + "spans": [ + { + "bbox": [ + 208, + 528, + 403, + 559 + ], + "score": 0.88, + "content": "\\begin{array} { r } { H = \\mathbf { M } \\mathbf { u } ] \\mathrm { t i } \\mathbf { H e a d } \\mathbf { S e l f A t t n } ( \\mathbf { L a y e r N o r m } ( X ) ) + X } \\\\ { L = \\mathbf { F F N } ( \\mathbf { L a y e r N o r m } ( H ) ) + H } \\end{array}", + "type": "interline_equation", + "image_path": "06ee18caa5e9791e3351b33dd81d6b538c4f644c50d2d27558db3d31d1aa49a7.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 528, + 403, + 543.5 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 208, + 543.5, + 403, + 559.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 570, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 420, + 582 + ], + "score": 1.0, + "content": "Now, during every “backward pass”, we compute the Jacobian for parameters", + "type": "text" + }, + { + "bbox": [ + 420, + 569, + 432, + 581 + ], + "score": 0.89, + "content": "\\theta ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 569, + 466, + 582 + ], + "score": 1.0, + "content": "at layer", + "type": "text" + }, + { + "bbox": [ + 466, + 571, + 474, + 580 + ], + "score": 0.79, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 569, + 505, + 582 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 579, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 258, + 596 + ], + "score": 1.0, + "content": "are used to update the parameters of", + "type": "text" + }, + { + "bbox": [ + 258, + 581, + 266, + 591 + ], + "score": 0.57, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 579, + 270, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 271, + 581, + 283, + 593 + ], + "score": 0.79, + "content": "\\theta _ { t } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 579, + 506, + 596 + ], + "score": 1.0, + "content": ", as well as to compute the next layer’s Jacobian, thus", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 591, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 506, + 606 + ], + "score": 1.0, + "content": "back-propagating the gradients. In this work however, for some of the layers, we still backpropagate", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 603, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 506, + 615 + ], + "score": 1.0, + "content": "through them to compute gradients for earlier layers, but we never update their parameters. As a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 614, + 473, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 473, + 626 + ], + "score": 1.0, + "content": "result, these layers stay fixed at their random initialization, saving computational resources.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 569, + 506, + 626 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 639, + 194, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 194, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 194, + 651 + ], + "score": 1.0, + "content": "2.1 BACKGROUND", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "Naturally, never updating some of the parameters is computationally more efficient, as some matrix", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "addition operations can be skipped in the backward pass, but why is this not detrimental to the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 682, + 224, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 224, + 692 + ], + "score": 1.0, + "content": "performance of the network?", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 660, + 505, + 692 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "In the early days of neural networks, the bottom layers were often kept fixed as “associators” (Block,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "1962), or what Minsky & Papert (2017) called the Gamba perceptron (Gamba et al., 1961; Borsellino", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "& Gamba, 1961). Fixed random networks (Baum, 1988; Schmidt et al., 1992; Pao et al., 1994) have", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "been explored from many angles, including as “random kitchen sink” kernel machines (Rahimi", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "& Recht, 2008; 2009), “extreme learning machines” (Huang et al., 2006) and reservoir comput-", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "ing (Jaeger, 2003; Maass et al., 2002; Lukosevi ˇ cius & Jaeger, 2009). In reservoir computing, in- ˇ", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "put data are represented through fixed random high-dimensional non-linear representations, called", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "“reservoirs”, which are followed by a regular (often but not necessarily linear) “readout” network to", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 258, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 258, + 149 + ], + "score": 1.0, + "content": "make the final classification decision.", + "type": "text", + "cross_page": true + } + ], + "index": 5 + } + ], + "index": 43, + "bbox_fs": [ + 106, + 699, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "been explored from many angles, including as “random kitchen sink” kernel machines (Rahimi", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "& Recht, 2008; 2009), “extreme learning machines” (Huang et al., 2006) and reservoir comput-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "ing (Jaeger, 2003; Maass et al., 2002; Lukosevi ˇ cius & Jaeger, 2009). In reservoir computing, in- ˇ", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "put data are represented through fixed random high-dimensional non-linear representations, called", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "“reservoirs”, which are followed by a regular (often but not necessarily linear) “readout” network to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 258, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 258, + 149 + ], + "score": 1.0, + "content": "make the final classification decision.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 504, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "The theoretical justification for these approaches lies in two well-known results in machine learn-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "ing: Cover’s theorem (Cover, 1965) on the separability of patterns states that high-dimensional", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "non-linear transformations are more likely to be linearly separable; and the Johnson-Lindenstrauss", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "score": 1.0, + "content": "lemma (Johnson & Lindenstrauss, 1984) shows that random projections distort Euclidean distances", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 248, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 248, + 212 + ], + "score": 1.0, + "content": "very little under mild assumptions.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "Practically, random layers can be seen as a cheap way to increase network depth. There are inter-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "esting advantages to this approach. Fixed layers are known to have particularly low-cost hardware", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "requirements and can be easily implemented on high-bandwidth FPGAs with low power consump-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "tion (Hadaeghi et al., 2017; Tanaka et al., 2019), or on optical devices (Hicke et al., 2013). This", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "score": 1.0, + "content": "might yield interesting possibilities for training in a distributed fashion across multiple devices, as", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 269, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 283 + ], + "score": 1.0, + "content": "well as for neurmorphic hardware (Neftci et al., 2017). This approach also facilitates lower-latency", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "deployment of neural networks to edge devices, since weights can be shared simply by sending the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 292, + 417, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 417, + 304 + ], + "score": 1.0, + "content": "seed number, assuming the random number generator is known on both ends.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 317, + 253, + 328 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 255, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 255, + 329 + ], + "score": 1.0, + "content": "2.2 RESERVOIR TRANSFORMERS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 108, + 337, + 504, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 336, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 506, + 350 + ], + "score": 1.0, + "content": "This work explores inserting random non-linear transformations, or what we call reservoir layers,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 348, + 462, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 462, + 361 + ], + "score": 1.0, + "content": "into transformer networks. Specifically, we experiment with a variety of reservoir layers:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 132, + 370, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 132, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "• Transformer Reservoir: The standard transformer layer as described above, but with all", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 381, + 432, + 393 + ], + "spans": [ + { + "bbox": [ + 141, + 381, + 432, + 393 + ], + "score": 1.0, + "content": "parameters fixed after initialization, including the self-attention module.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 395, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 505, + 409 + ], + "score": 1.0, + "content": "• FFN Reservoir: A transformer-style fixed feed-forward layer without any self-attention,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 407, + 372, + 419 + ], + "spans": [ + { + "bbox": [ + 141, + 407, + 299, + 419 + ], + "score": 1.0, + "content": "i.e., FFN(LayerNorm(Previous layer))", + "type": "text" + }, + { + "bbox": [ + 299, + 407, + 309, + 417 + ], + "score": 0.62, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 407, + 372, + 419 + ], + "score": 1.0, + "content": "Previous layer.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "• BiGRU Reservoir: A fixed bidirectional Gated Recurrent Unit (Cho et al., 2014) layer,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 433, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 142, + 433, + 505, + 445 + ], + "score": 1.0, + "content": "which is closer in spirit to previous work on reservoir computing, most of which builds on", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 142, + 445, + 299, + 455 + ], + "spans": [ + { + "bbox": [ + 142, + 445, + 299, + 455 + ], + "score": 1.0, + "content": "recurrent neural network architectures.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 140, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 140, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "CNN Reservoir: A fixed Convolutional Neural Network (LeCun et al., 1998) layer, specifi-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 142, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 142, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "cally light dynamical convolution layers (Wu et al., 2019), which are known to be compet-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 481, + 360, + 493 + ], + "spans": [ + { + "bbox": [ + 141, + 481, + 360, + 493 + ], + "score": 1.0, + "content": "itive with transformers in sequence-to-sequence tasks.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 108, + 502, + 503, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 515 + ], + "score": 1.0, + "content": "We find that all these approaches work well, to a certain extent. For clarity, we focus primarily on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 513, + 423, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 423, + 526 + ], + "score": 1.0, + "content": "the first two reservoir layers, but include a broader comparison in Appendix A.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 531, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "In each case, contrary to traditional reservoir computing, our reservoir layers are interspersed", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "score": 1.0, + "content": "throughout a regular transformer network, or what we call a reservoir transformer. A good jus-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 550, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 566 + ], + "score": 1.0, + "content": "tification for this approach is that while random projections are not learned and might introduce", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "noise, subsequent normal transformer “readout” layers might allow us to recover from any adverse", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "effects of randomness. For example, previous work has shown that ResNets, with all of their param-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "score": 1.0, + "content": "eters fixed except for the scale and shift parameters of batch normalization, can still achieve high", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "performance, simply by scaling and shifting random features (Frankle et al., 2020). Adding noise to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "the parameters of neural networks is also known to help convergence and generalization (Jim et al.,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 618, + 317, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 317, + 630 + ], + "score": 1.0, + "content": "1995; 1996; Gulcehre et al., 2016; Noh et al., 2017).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 108, + 646, + 193, + 658 + ], + "lines": [ + { + "bbox": [ + 104, + 644, + 195, + 662 + ], + "spans": [ + { + "bbox": [ + 104, + 644, + 195, + 662 + ], + "score": 1.0, + "content": "3 EVALUATION", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 671, + 503, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 685 + ], + "score": 1.0, + "content": "We evaluate the proposed approach on a variety of well-known tasks in natural language processing,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 680, + 474, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 474, + 695 + ], + "score": 1.0, + "content": "namely: machine translation, language modelling and masked language model pre-training.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "In this work, we are not necessarily interested in obtaining the state of the art on any task or even", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "in improving overall task performance via this method. The main objective is to examine effi-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "ciency, i.e. the relationship between compute time and task performance. This is closely related", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 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 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 148 + ], + "lines": [], + "index": 2.5, + "bbox_fs": [ + 105, + 83, + 506, + 149 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 504, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "The theoretical justification for these approaches lies in two well-known results in machine learn-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "ing: Cover’s theorem (Cover, 1965) on the separability of patterns states that high-dimensional", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "non-linear transformations are more likely to be linearly separable; and the Johnson-Lindenstrauss", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "score": 1.0, + "content": "lemma (Johnson & Lindenstrauss, 1984) shows that random projections distort Euclidean distances", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 248, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 248, + 212 + ], + "score": 1.0, + "content": "very little under mild assumptions.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 153, + 506, + 212 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "Practically, random layers can be seen as a cheap way to increase network depth. There are inter-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "esting advantages to this approach. Fixed layers are known to have particularly low-cost hardware", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "requirements and can be easily implemented on high-bandwidth FPGAs with low power consump-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "tion (Hadaeghi et al., 2017; Tanaka et al., 2019), or on optical devices (Hicke et al., 2013). This", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "score": 1.0, + "content": "might yield interesting possibilities for training in a distributed fashion across multiple devices, as", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 269, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 283 + ], + "score": 1.0, + "content": "well as for neurmorphic hardware (Neftci et al., 2017). This approach also facilitates lower-latency", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "deployment of neural networks to edge devices, since weights can be shared simply by sending the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 292, + 417, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 417, + 304 + ], + "score": 1.0, + "content": "seed number, assuming the random number generator is known on both ends.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 214, + 505, + 304 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 317, + 253, + 328 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 255, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 255, + 329 + ], + "score": 1.0, + "content": "2.2 RESERVOIR TRANSFORMERS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 108, + 337, + 504, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 336, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 506, + 350 + ], + "score": 1.0, + "content": "This work explores inserting random non-linear transformations, or what we call reservoir layers,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 348, + 462, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 462, + 361 + ], + "score": 1.0, + "content": "into transformer networks. Specifically, we experiment with a variety of reservoir layers:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 336, + 506, + 361 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 370, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 132, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "• Transformer Reservoir: The standard transformer layer as described above, but with all", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 381, + 432, + 393 + ], + "spans": [ + { + "bbox": [ + 141, + 381, + 432, + 393 + ], + "score": 1.0, + "content": "parameters fixed after initialization, including the self-attention module.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 395, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 505, + 409 + ], + "score": 1.0, + "content": "• FFN Reservoir: A transformer-style fixed feed-forward layer without any self-attention,", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 407, + 372, + 419 + ], + "spans": [ + { + "bbox": [ + 141, + 407, + 299, + 419 + ], + "score": 1.0, + "content": "i.e., FFN(LayerNorm(Previous layer))", + "type": "text" + }, + { + "bbox": [ + 299, + 407, + 309, + 417 + ], + "score": 0.62, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 407, + 372, + 419 + ], + "score": 1.0, + "content": "Previous layer.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "• BiGRU Reservoir: A fixed bidirectional Gated Recurrent Unit (Cho et al., 2014) layer,", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 433, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 142, + 433, + 505, + 445 + ], + "score": 1.0, + "content": "which is closer in spirit to previous work on reservoir computing, most of which builds on", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 142, + 445, + 299, + 455 + ], + "spans": [ + { + "bbox": [ + 142, + 445, + 299, + 455 + ], + "score": 1.0, + "content": "recurrent neural network architectures.", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 140, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 140, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "CNN Reservoir: A fixed Convolutional Neural Network (LeCun et al., 1998) layer, specifi-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 142, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 142, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "cally light dynamical convolution layers (Wu et al., 2019), which are known to be compet-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 481, + 360, + 493 + ], + "spans": [ + { + "bbox": [ + 141, + 481, + 360, + 493 + ], + "score": 1.0, + "content": "itive with transformers in sequence-to-sequence tasks.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + } + ], + "index": 26.5, + "bbox_fs": [ + 132, + 370, + 505, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 502, + 503, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 515 + ], + "score": 1.0, + "content": "We find that all these approaches work well, to a certain extent. For clarity, we focus primarily on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 513, + 423, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 423, + 526 + ], + "score": 1.0, + "content": "the first two reservoir layers, but include a broader comparison in Appendix A.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 501, + 505, + 526 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 531, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "In each case, contrary to traditional reservoir computing, our reservoir layers are interspersed", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "score": 1.0, + "content": "throughout a regular transformer network, or what we call a reservoir transformer. A good jus-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 550, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 566 + ], + "score": 1.0, + "content": "tification for this approach is that while random projections are not learned and might introduce", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "noise, subsequent normal transformer “readout” layers might allow us to recover from any adverse", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "effects of randomness. For example, previous work has shown that ResNets, with all of their param-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "score": 1.0, + "content": "eters fixed except for the scale and shift parameters of batch normalization, can still achieve high", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "performance, simply by scaling and shifting random features (Frankle et al., 2020). Adding noise to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "the parameters of neural networks is also known to help convergence and generalization (Jim et al.,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 618, + 317, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 317, + 630 + ], + "score": 1.0, + "content": "1995; 1996; Gulcehre et al., 2016; Noh et al., 2017).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 531, + 506, + 630 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 646, + 193, + 658 + ], + "lines": [ + { + "bbox": [ + 104, + 644, + 195, + 662 + ], + "spans": [ + { + "bbox": [ + 104, + 644, + 195, + 662 + ], + "score": 1.0, + "content": "3 EVALUATION", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 671, + 503, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 685 + ], + "score": 1.0, + "content": "We evaluate the proposed approach on a variety of well-known tasks in natural language processing,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 680, + 474, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 474, + 695 + ], + "score": 1.0, + "content": "namely: machine translation, language modelling and masked language model pre-training.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 668, + 505, + 695 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "In this work, we are not necessarily interested in obtaining the state of the art on any task or even", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "in improving overall task performance via this method. The main objective is to examine effi-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "ciency, i.e. the relationship between compute time and task performance. This is closely related", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 286, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 300 + ], + "score": 1.0, + "content": "to efforts in Green AI, which are concerned with the trade-offs between compute, data, and per-", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 297, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 312 + ], + "score": 1.0, + "content": "formance (Schwartz et al., 2019). We propose a new metric for our purposes, the area under the", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "convergence curve (AUCC): similarly to how the area under the receiver operating characteristic", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "(Bradley, 1997, AUC-ROC) measures a classifier’s performance independent of the classification", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "score": 1.0, + "content": "threshold, AUCC measures a model’s performance independent of the specific compute budget.", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 342, + 284, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 284, + 354 + ], + "score": 1.0, + "content": "Specifically, AUCC is computed as follows:", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "spans": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "score": 0.967, + "type": "image", + "image_path": "f12829b70f6d779da8d13c4ab6c635efbc557e9a0d0480b52615b7795f025ea3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 97, + 486, + 139.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 139.66666666666666, + 486, + 182.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 182.33333333333331, + 486, + 224.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 238, + 505, + 261 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "Figure 1: Validation BLEU AUCC and test BLEU for IWSLT (high is good). Comparison of regular", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 249, + 455, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 455, + 262 + ], + "score": 1.0, + "content": "transformer and reservoir transformer with FFN or Transformer reservoir layers added.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 300 + ], + "score": 1.0, + "content": "to efforts in Green AI, which are concerned with the trade-offs between compute, data, and per-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 297, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 312 + ], + "score": 1.0, + "content": "formance (Schwartz et al., 2019). We propose a new metric for our purposes, the area under the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "convergence curve (AUCC): similarly to how the area under the receiver operating characteristic", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "(Bradley, 1997, AUC-ROC) measures a classifier’s performance independent of the classification", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "score": 1.0, + "content": "threshold, AUCC measures a model’s performance independent of the specific compute budget.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 342, + 284, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 284, + 354 + ], + "score": 1.0, + "content": "Specifically, AUCC is computed as follows:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 259, + 372, + 353, + 408 + ], + "lines": [ + { + "bbox": [ + 259, + 372, + 353, + 408 + ], + "spans": [ + { + "bbox": [ + 259, + 372, + 353, + 408 + ], + "score": 0.93, + "content": "\\int _ { t = 0 } ^ { \\hat { T } } \\sum _ { x , y \\in \\mathcal { D } } g _ { t } ( f ( x ) , y )", + "type": "interline_equation", + "image_path": "fd35fc393f71279405ce5c08762c208f32673d365da86fffd205a0b917f48f39.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 259, + 372, + 353, + 390.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 259, + 390.0, + 353, + 408.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 133, + 437 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 425, + 141, + 437 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 423, + 221, + 437 + ], + "score": 1.0, + "content": "is the network and", + "type": "text" + }, + { + "bbox": [ + 221, + 427, + 227, + 437 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 423, + 465, + 437 + ], + "score": 1.0, + "content": "is the evaluation metric, measured until convergence time", + "type": "text" + }, + { + "bbox": [ + 465, + 423, + 474, + 435 + ], + "score": 0.84, + "content": "\\hat { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 423, + 506, + 437 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "is the maximum convergence time of all models included in the comparison. Note that time here is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "wall-clock time, not iterations. By convergence, we mean that validation performance has stopped", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "improving, and hence the convergence curve whose area we measure plots the desired metric over", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "time. Runs are averaged over multiple seeds and reported with standard deviation. We normalize", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 479, + 485, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 428, + 492 + ], + "score": 1.0, + "content": "raw AUCC scores by their maximum score to ensure a more easily interpretable", + "type": "text" + }, + { + "bbox": [ + 428, + 479, + 456, + 492 + ], + "score": 0.93, + "content": "[ 0 - 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 479, + 485, + 492 + ], + "score": 1.0, + "content": "range.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 496, + 505, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 506, + 509 + ], + "score": 1.0, + "content": "One potential downside of this approach is that the AUCC metric could lead to higher scores for a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "model that converges quickly but to ultimately worse performance, if measured in a small window.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 273, + 532 + ], + "score": 1.0, + "content": "We account for this by making sure that", + "type": "text" + }, + { + "bbox": [ + 273, + 518, + 282, + 530 + ], + "score": 0.85, + "content": "\\hat { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "is set sufficiently high. We include the raw validation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 531, + 430, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 430, + 544 + ], + "score": 1.0, + "content": "curves in the appendix and also report test set generalization in each experiment.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 107, + 560, + 385, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 387, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 387, + 573 + ], + "score": 1.0, + "content": "3.1 EXPERIMENTAL SETTINGS AND IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 505, + 661 + ], + "lines": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "We evaluate on IWSLT de-en (Cettolo et al., 2015) and WMT en-de (Bojar et al., 2014) for ma-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "chine translation; enwiki8 (LLC, 2009) for language modelling; and experiment with RoBERTa", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "(Liu et al., 2019) in our pretraining experiments. For IWSLT, we follow the pre-processing steps in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 296, + 627 + ], + "score": 1.0, + "content": "Edunov et al. (2018). The train/val/test split is", + "type": "text" + }, + { + "bbox": [ + 297, + 616, + 356, + 627 + ], + "score": 0.6, + "content": "\\_", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "sentences. For WMT, we follow the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 368, + 640 + ], + "score": 1.0, + "content": "pre-processing steps in Ott et al. (2018). The train/val/test split is", + "type": "text" + }, + { + "bbox": [ + 368, + 627, + 429, + 638 + ], + "score": 0.36, + "content": "4 . 5 \\mathrm { M } / 1 6 . 5 \\mathrm { k } / 3 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "sentences. For en-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "wiki8, we follow the pre-processing steps in Dai et al. (2019). The train/val/test split is 1M/54k/56k", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 649, + 477, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 477, + 661 + ], + "score": 1.0, + "content": "sentences. For RoBERTa pretraining, we follow the pre-processing steps in Liu et al. (2019).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 504, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 504, + 677 + ], + "score": 1.0, + "content": "We use 8 Volta V100 GPUs for WMT and enwik8, 32 V100 GPUs for RoBERTa and a single V100", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "for IWSLT. The hyperparameters for IWSLT14 and WMT16 were set to the best-performing values", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "from Ott et al. (2018) and Kasai et al. (2020) respectively. The enwik8 experiment settings followed", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Bachlechner et al. (2020) and the RoBERTa experiments followed Liu et al. (2019). All experiments", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 711, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 504, + 722 + ], + "score": 1.0, + "content": "were conducted using fairseq (Ott et al., 2019). Our code and experimental settings will be made", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 721, + 310, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 310, + 733 + ], + "score": 1.0, + "content": "open source at [ANONYMIZED-GITHUB-URL].", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 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 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "spans": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "score": 0.967, + "type": "image", + "image_path": "f12829b70f6d779da8d13c4ab6c635efbc557e9a0d0480b52615b7795f025ea3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 97, + 486, + 139.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 139.66666666666666, + 486, + 182.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 182.33333333333331, + 486, + 224.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 238, + 505, + 261 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "Figure 1: Validation BLEU AUCC and test BLEU for IWSLT (high is good). Comparison of regular", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 249, + 455, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 455, + 262 + ], + "score": 1.0, + "content": "transformer and reservoir transformer with FFN or Transformer reservoir layers added.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 505, + 353 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 286, + 505, + 354 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 259, + 372, + 353, + 408 + ], + "lines": [ + { + "bbox": [ + 259, + 372, + 353, + 408 + ], + "spans": [ + { + "bbox": [ + 259, + 372, + 353, + 408 + ], + "score": 0.93, + "content": "\\int _ { t = 0 } ^ { \\hat { T } } \\sum _ { x , y \\in \\mathcal { D } } g _ { t } ( f ( x ) , y )", + "type": "interline_equation", + "image_path": "fd35fc393f71279405ce5c08762c208f32673d365da86fffd205a0b917f48f39.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 259, + 372, + 353, + 390.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 259, + 390.0, + 353, + 408.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 133, + 437 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 425, + 141, + 437 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 423, + 221, + 437 + ], + "score": 1.0, + "content": "is the network and", + "type": "text" + }, + { + "bbox": [ + 221, + 427, + 227, + 437 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 423, + 465, + 437 + ], + "score": 1.0, + "content": "is the evaluation metric, measured until convergence time", + "type": "text" + }, + { + "bbox": [ + 465, + 423, + 474, + 435 + ], + "score": 0.84, + "content": "\\hat { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 423, + 506, + 437 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "is the maximum convergence time of all models included in the comparison. Note that time here is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "wall-clock time, not iterations. By convergence, we mean that validation performance has stopped", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "improving, and hence the convergence curve whose area we measure plots the desired metric over", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "time. Runs are averaged over multiple seeds and reported with standard deviation. We normalize", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 479, + 485, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 428, + 492 + ], + "score": 1.0, + "content": "raw AUCC scores by their maximum score to ensure a more easily interpretable", + "type": "text" + }, + { + "bbox": [ + 428, + 479, + 456, + 492 + ], + "score": 0.93, + "content": "[ 0 - 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 479, + 485, + 492 + ], + "score": 1.0, + "content": "range.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 423, + 506, + 492 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 496, + 505, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 506, + 509 + ], + "score": 1.0, + "content": "One potential downside of this approach is that the AUCC metric could lead to higher scores for a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "model that converges quickly but to ultimately worse performance, if measured in a small window.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 273, + 532 + ], + "score": 1.0, + "content": "We account for this by making sure that", + "type": "text" + }, + { + "bbox": [ + 273, + 518, + 282, + 530 + ], + "score": 0.85, + "content": "\\hat { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "is set sufficiently high. We include the raw validation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 531, + 430, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 430, + 544 + ], + "score": 1.0, + "content": "curves in the appendix and also report test set generalization in each experiment.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 497, + 506, + 544 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 560, + 385, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 387, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 387, + 573 + ], + "score": 1.0, + "content": "3.1 EXPERIMENTAL SETTINGS AND IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 505, + 661 + ], + "lines": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "We evaluate on IWSLT de-en (Cettolo et al., 2015) and WMT en-de (Bojar et al., 2014) for ma-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "chine translation; enwiki8 (LLC, 2009) for language modelling; and experiment with RoBERTa", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "(Liu et al., 2019) in our pretraining experiments. For IWSLT, we follow the pre-processing steps in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 296, + 627 + ], + "score": 1.0, + "content": "Edunov et al. (2018). The train/val/test split is", + "type": "text" + }, + { + "bbox": [ + 297, + 616, + 356, + 627 + ], + "score": 0.6, + "content": "\\_", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "sentences. For WMT, we follow the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 368, + 640 + ], + "score": 1.0, + "content": "pre-processing steps in Ott et al. (2018). The train/val/test split is", + "type": "text" + }, + { + "bbox": [ + 368, + 627, + 429, + 638 + ], + "score": 0.36, + "content": "4 . 5 \\mathrm { M } / 1 6 . 5 \\mathrm { k } / 3 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "sentences. For en-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "wiki8, we follow the pre-processing steps in Dai et al. (2019). The train/val/test split is 1M/54k/56k", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 649, + 477, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 477, + 661 + ], + "score": 1.0, + "content": "sentences. For RoBERTa pretraining, we follow the pre-processing steps in Liu et al. (2019).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 582, + 505, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 504, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 504, + 677 + ], + "score": 1.0, + "content": "We use 8 Volta V100 GPUs for WMT and enwik8, 32 V100 GPUs for RoBERTa and a single V100", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "for IWSLT. The hyperparameters for IWSLT14 and WMT16 were set to the best-performing values", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "from Ott et al. (2018) and Kasai et al. (2020) respectively. The enwik8 experiment settings followed", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Bachlechner et al. (2020) and the RoBERTa experiments followed Liu et al. (2019). All experiments", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 711, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 504, + 722 + ], + "score": 1.0, + "content": "were conducted using fairseq (Ott et al., 2019). Our code and experimental settings will be made", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 721, + 310, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 310, + 733 + ], + "score": 1.0, + "content": "open source at [ANONYMIZED-GITHUB-URL].", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 666, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 80, + 505, + 249 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 80, + 505, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 505, + 249 + ], + "score": 0.985, + "html": "
Model#LayersFrozenMax BLEUTrain time until max (in hours)Ratio# Params Trainable (Total)Train Time each epoch (in seconds)
Transformer6034.52 ± 0.072.548 ± 0.06126.8M122.73 ± 1.16
8034.59 ± 0.112.557 ± 0.05131.1M142.28 ± 1.87
10034.56 ± 0.053.173 ± 0.04135.3M161.66 ± 1.54
12034.29 ± 0.123.521 ± 0.09139.5M172.45 ± 1.98
TReservoir6234.37 ± 0.122.422 ± 0.030.9522.6M (26.8M)120.59 ± 1.32
8234.80 ± 0.072.450 ± 0.060.9626.8M (31.1M)134.49 ± 1.76
10234.70 ± 0.032.831 ± 0.050.8931.1M (35.3M)144.42 ± 1.98
12234.78 ± 0.043.476 ± 0.040.9835.3M (39.5M)159.43 ± 1.67
FFN Reservoir6234.43 ± 0.152.120 ± 0.040.8322.6M (25.8M)107.71 ± 1.73
8234.56 ± 0.162.203 ± 0.060.8626.8M (29.1M)120.07 ± 1.65
10234.66 ± 0.022.493 ± 0.050.7931.1M (33.3M)130.11 ± 1.43
12234.76 ± 0.033.241 ± 0.040.9235.3M (37.5M)156.32 ± 1.87
LayerDrop6234.59 ± 0.152.364 ± 0.080.9222.6M (26.8M)119.30 ± 1.36
8234.58 ± 0.162.554 ± 0.050.9926.8M (31.1M)138.62 ± 1.44
10234.57 ± 0.073.404 ± 0.061.0731.1M (35.3M)140.88 ± 1.62
12233.65 ± 0.243.251 ± 0.040.9235.3M (39.5M)160.85 ± 1.49
", + "type": "table", + "image_path": "4dd693f4fd0bc00d5e71df0ea059dae5cfc447650b30fe1d8ce5446d40f80a14.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 505, + 136.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 136.33333333333334, + 505, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 192.66666666666669, + 505, + 249.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 258, + 506, + 291 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 271 + ], + "score": 1.0, + "content": "Table 1: Wall-clock time (averaged over multiple runs) saved for IWSLT for different model types", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "and encoder depths. Max BLEU is for validation. Number of layers is for encoder, decoder depth is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 280, + 492, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 492, + 293 + ], + "score": 1.0, + "content": "kept fixed at 2. Ratio is computed compared to comparable number of layers in the normal case.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "All the experiments in this paper were run with 3 random seeds and the mean and standard deviation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 307, + 338 + ], + "score": 1.0, + "content": "are reported. For the relatively small IWSLT, the", + "type": "text" + }, + { + "bbox": [ + 308, + 325, + 317, + 336 + ], + "score": 0.84, + "content": "\\hat { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 326, + 505, + 338 + ], + "score": 1.0, + "content": "value in the AUCC metric was set to 4 hours.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "score": 1.0, + "content": "For WMT, which is larger, we set it to 20 hours. For enwiki8, it was 30 hours; and for the RoBERTa", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 349, + 300, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 300, + 361 + ], + "score": 1.0, + "content": "pre-training experiments, it was set to 60 hours.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "The projection weights in random layers were initialized using orthogonal initialization (Saxe", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "score": 1.0, + "content": "et al., 2013), which makes sense since random orthogonal projections should be most information-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 387, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 401 + ], + "score": 1.0, + "content": "preserving, and which was found to work well empirically for initializing fixed random representa-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "tions in previous work (Wieting & Kiela, 2019). Biases and layer norm parameters were initialized", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 409, + 452, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 452, + 423 + ], + "score": 1.0, + "content": "using their respective PyTorch defaults (based on Xavier init; Glorot & Bengio, 2010).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 439 + ], + "score": 1.0, + "content": "We intersperse reservoir layers in alternating fashion starting from the middle. Specifically, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "score": 1.0, + "content": "alternate one reservoir layer with one transformer layer, and place the alternating block in the middle.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 447, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 461 + ], + "score": 1.0, + "content": "For example: a 7-layer encoder LLLLLLL in which we replace three layers with reservoirs becomes", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 456, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 474 + ], + "score": 1.0, + "content": "LRLRLRL, and with two becomes LLRLRLL. See Appendix C for a study comparing this strategy", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 470, + 386, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 386, + 483 + ], + "score": 1.0, + "content": "to alternative approaches (e.g., freezing in the bottom, middle or top).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 107, + 499, + 200, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 201, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 201, + 514 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "In what follows, we first show our main result: reservoir transformers often have better AUCC met-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "rics, less training time per epoch, less convergence time until the best validation performance is", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 546, + 504, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 504, + 559 + ], + "score": 1.0, + "content": "achieved, and even improved test set generalization metrics, on a variety of tasks. As a strong base-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "line method, we compare to LayerDrop (Fan et al., 2019). LayerDrop can also be seen as a method", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "that dynamically bypasses parts of the computation during Transformer training in an attempt to im-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 579, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 591 + ], + "score": 1.0, + "content": "prove efficiency, and is a suitable comparison to examine our methods.. We also examine whether", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "we can minimize the expectation over the gradients of upper layers in the transformer network such", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 600, + 461, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 461, + 615 + ], + "score": 1.0, + "content": "that we do not have to pass the true gradients through the reservoir for further efficiency.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 107, + 628, + 238, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 627, + 239, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 239, + 640 + ], + "score": 1.0, + "content": "4.1 MACHINE TRANSLATION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "Machine translation (MT) is one of the core tasks of NLP. We demonstrate on two well-known MT", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "datasets, IWSLT’14 German-English and WMT’16 English-German, that reservoir transformers", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "obtain a better AUCC. For the raw validation plots over time that were used to calculate the AUCC,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 681, + 218, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 218, + 694 + ], + "score": 1.0, + "content": "please refer to Appendix F.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Following Kasai et al. (2020), the architecture of the network is an N-layer reservoir transformer", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "encoder, followed by a regular shallow one- or two-layer decoder. This design choice has been", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "shown to lead to very good speed and efficiency trade-offs, and serves as a good baseline for our", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 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 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 80, + 505, + 249 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 80, + 505, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 505, + 249 + ], + "score": 0.985, + "html": "
Model#LayersFrozenMax BLEUTrain time until max (in hours)Ratio# Params Trainable (Total)Train Time each epoch (in seconds)
Transformer6034.52 ± 0.072.548 ± 0.06126.8M122.73 ± 1.16
8034.59 ± 0.112.557 ± 0.05131.1M142.28 ± 1.87
10034.56 ± 0.053.173 ± 0.04135.3M161.66 ± 1.54
12034.29 ± 0.123.521 ± 0.09139.5M172.45 ± 1.98
TReservoir6234.37 ± 0.122.422 ± 0.030.9522.6M (26.8M)120.59 ± 1.32
8234.80 ± 0.072.450 ± 0.060.9626.8M (31.1M)134.49 ± 1.76
10234.70 ± 0.032.831 ± 0.050.8931.1M (35.3M)144.42 ± 1.98
12234.78 ± 0.043.476 ± 0.040.9835.3M (39.5M)159.43 ± 1.67
FFN Reservoir6234.43 ± 0.152.120 ± 0.040.8322.6M (25.8M)107.71 ± 1.73
8234.56 ± 0.162.203 ± 0.060.8626.8M (29.1M)120.07 ± 1.65
10234.66 ± 0.022.493 ± 0.050.7931.1M (33.3M)130.11 ± 1.43
12234.76 ± 0.033.241 ± 0.040.9235.3M (37.5M)156.32 ± 1.87
LayerDrop6234.59 ± 0.152.364 ± 0.080.9222.6M (26.8M)119.30 ± 1.36
8234.58 ± 0.162.554 ± 0.050.9926.8M (31.1M)138.62 ± 1.44
10234.57 ± 0.073.404 ± 0.061.0731.1M (35.3M)140.88 ± 1.62
12233.65 ± 0.243.251 ± 0.040.9235.3M (39.5M)160.85 ± 1.49
", + "type": "table", + "image_path": "4dd693f4fd0bc00d5e71df0ea059dae5cfc447650b30fe1d8ce5446d40f80a14.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 505, + 136.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 136.33333333333334, + 505, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 192.66666666666669, + 505, + 249.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 258, + 506, + 291 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 271 + ], + "score": 1.0, + "content": "Table 1: Wall-clock time (averaged over multiple runs) saved for IWSLT for different model types", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "and encoder depths. Max BLEU is for validation. Number of layers is for encoder, decoder depth is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 280, + 492, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 492, + 293 + ], + "score": 1.0, + "content": "kept fixed at 2. Ratio is computed compared to comparable number of layers in the normal case.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "All the experiments in this paper were run with 3 random seeds and the mean and standard deviation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 307, + 338 + ], + "score": 1.0, + "content": "are reported. For the relatively small IWSLT, the", + "type": "text" + }, + { + "bbox": [ + 308, + 325, + 317, + 336 + ], + "score": 0.84, + "content": "\\hat { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 326, + 505, + 338 + ], + "score": 1.0, + "content": "value in the AUCC metric was set to 4 hours.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "score": 1.0, + "content": "For WMT, which is larger, we set it to 20 hours. For enwiki8, it was 30 hours; and for the RoBERTa", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 349, + 300, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 300, + 361 + ], + "score": 1.0, + "content": "pre-training experiments, it was set to 60 hours.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 314, + 505, + 361 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "The projection weights in random layers were initialized using orthogonal initialization (Saxe", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "score": 1.0, + "content": "et al., 2013), which makes sense since random orthogonal projections should be most information-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 387, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 401 + ], + "score": 1.0, + "content": "preserving, and which was found to work well empirically for initializing fixed random representa-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "tions in previous work (Wieting & Kiela, 2019). Biases and layer norm parameters were initialized", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 409, + 452, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 452, + 423 + ], + "score": 1.0, + "content": "using their respective PyTorch defaults (based on Xavier init; Glorot & Bengio, 2010).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 365, + 506, + 423 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 439 + ], + "score": 1.0, + "content": "We intersperse reservoir layers in alternating fashion starting from the middle. Specifically, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "score": 1.0, + "content": "alternate one reservoir layer with one transformer layer, and place the alternating block in the middle.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 447, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 461 + ], + "score": 1.0, + "content": "For example: a 7-layer encoder LLLLLLL in which we replace three layers with reservoirs becomes", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 456, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 474 + ], + "score": 1.0, + "content": "LRLRLRL, and with two becomes LLRLRLL. See Appendix C for a study comparing this strategy", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 470, + 386, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 386, + 483 + ], + "score": 1.0, + "content": "to alternative approaches (e.g., freezing in the bottom, middle or top).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 425, + 505, + 483 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 499, + 200, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 201, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 201, + 514 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "In what follows, we first show our main result: reservoir transformers often have better AUCC met-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "rics, less training time per epoch, less convergence time until the best validation performance is", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 546, + 504, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 504, + 559 + ], + "score": 1.0, + "content": "achieved, and even improved test set generalization metrics, on a variety of tasks. As a strong base-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "line method, we compare to LayerDrop (Fan et al., 2019). LayerDrop can also be seen as a method", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "that dynamically bypasses parts of the computation during Transformer training in an attempt to im-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 579, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 591 + ], + "score": 1.0, + "content": "prove efficiency, and is a suitable comparison to examine our methods.. We also examine whether", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "we can minimize the expectation over the gradients of upper layers in the transformer network such", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 600, + 461, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 461, + 615 + ], + "score": 1.0, + "content": "that we do not have to pass the true gradients through the reservoir for further efficiency.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 524, + 506, + 615 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 628, + 238, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 627, + 239, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 239, + 640 + ], + "score": 1.0, + "content": "4.1 MACHINE TRANSLATION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "Machine translation (MT) is one of the core tasks of NLP. We demonstrate on two well-known MT", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "datasets, IWSLT’14 German-English and WMT’16 English-German, that reservoir transformers", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "obtain a better AUCC. For the raw validation plots over time that were used to calculate the AUCC,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 681, + 218, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 218, + 694 + ], + "score": 1.0, + "content": "please refer to Appendix F.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 649, + 505, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Following Kasai et al. (2020), the architecture of the network is an N-layer reservoir transformer", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "encoder, followed by a regular shallow one- or two-layer decoder. This design choice has been", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "shown to lead to very good speed and efficiency trade-offs, and serves as a good baseline for our", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "experiments. Moreover, shallow decoders make it easier to decide where to place reservoir layers (in", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 511, + 493, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 493, + 523 + ], + "score": 1.0, + "content": "the encoder) and makes it more straightforward to identify where performance gains come from.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 698, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "spans": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "score": 0.924, + "type": "image", + "image_path": "90e2f338200568e105b601ef0f6d470de812574f5e1411facb2387a68f64cf2e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 97, + 486, + 139.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 139.66666666666666, + 486, + 182.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 182.33333333333331, + 486, + 224.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 238, + 505, + 261 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "Figure 2: Validation BLEU AUCC and test BLEU for WMT (high is good). Comparison of regular", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 249, + 455, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 455, + 262 + ], + "score": 1.0, + "content": "transformer and reservoir transformer with FFN or Transformer reservoir layers added.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 114, + 300, + 486, + 429 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 300, + 486, + 429 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 300, + 486, + 429 + ], + "spans": [ + { + "bbox": [ + 114, + 300, + 486, + 429 + ], + "score": 0.965, + "type": "image", + "image_path": "fbf8f8b964d88c1ee463d7c528fdd11742523e63086217892f8a7d6bfafcc93f.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 114, + 300, + 486, + 343.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 343.0, + 486, + 386.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 114, + 386.0, + 486, + 429.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 443, + 503, + 466 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 442, + 504, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 504, + 457 + ], + "score": 1.0, + "content": "Figure 3: Validation BPC AUCC and test BPC on the enwik8 language modelling task (low is good).", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 455, + 384, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 384, + 468 + ], + "score": 1.0, + "content": "Comparison of regular and reservoir transformers for varying depths.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 504, + 523 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "experiments. Moreover, shallow decoders make it easier to decide where to place reservoir layers (in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 511, + 493, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 493, + 523 + ], + "score": 1.0, + "content": "the encoder) and makes it more straightforward to identify where performance gains come from.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "Figure 1 shows the results for IWSLT. On the y-axis we show validation AUCC for the BLEU", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 167, + 551 + ], + "score": 1.0, + "content": "metric; on the", + "type": "text" + }, + { + "bbox": [ + 168, + 541, + 174, + 550 + ], + "score": 0.26, + "content": "\\mathbf { X } ^ { } -", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "-axis we show the number of updatable layers in the encoder. The performance", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 551, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 505, + 562 + ], + "score": 1.0, + "content": "of a regular transformer encoder with 6 layers and a reservoir transformer encoder with 6 layers", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 561, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 340, + 573 + ], + "score": 1.0, + "content": "plus N additional reservoir layers are plotted for the same", + "type": "text" + }, + { + "bbox": [ + 340, + 563, + 347, + 572 + ], + "score": 0.4, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 561, + 506, + 573 + ], + "score": 1.0, + "content": "-axis value to show the total number of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "updated layers. Plots for the total number of layers (updatable plus not-updatable, so essentially", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "shifted versions) are shown in Appendix E. Table 1 shows the time it took to achieve the maximum", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "validation BLEU score and how that relates to the regular transformer, demonstrating that reservoir", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 412, + 617 + ], + "score": 1.0, + "content": "transformers consistently converge faster in terms of wall-clock time, up to", + "type": "text" + }, + { + "bbox": [ + 412, + 605, + 432, + 615 + ], + "score": 0.87, + "content": "22 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "as much with the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 325, + 628 + ], + "score": 1.0, + "content": "same number of updateable layers. We save as much as", + "type": "text" + }, + { + "bbox": [ + 325, + 616, + 345, + 627 + ], + "score": 0.87, + "content": "27 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "time until convergence a 24 layer model", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 627, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 638 + ], + "score": 1.0, + "content": "on WMT, as shown in Table 3. One other noticeable point is that we can see that the T Reservoir", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 637, + 504, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 504, + 650 + ], + "score": 1.0, + "content": "achieves similar performance to LayerDrop on IWSLT and WMT in terms of wall-clock per epoch", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "and wall-clock time to the best performance. However, on both tasks, FFN Reservoir performs much", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "better than LayerDrop in terms of efficiency per epoch and achieves better/similar performance in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 671, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 506, + 683 + ], + "score": 1.0, + "content": "less time in each case. As a point of reference, a half hour gain on IWSLT translates to a gain of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 681, + 466, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 466, + 693 + ], + "score": 1.0, + "content": "several days in the training of bigger transformer models like GPT-3 (Brown et al., 2020).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 506, + 712 + ], + "score": 1.0, + "content": "We observe that reservoir transformers consistently perform better than, or are competitive to, reg-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 709, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 721 + ], + "score": 1.0, + "content": "ular transformers, both in terms of validation BLEU AUCC as well as test time BLEU, for all", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 721, + 213, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 213, + 733 + ], + "score": 1.0, + "content": "examined encoder depths.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 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 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "spans": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "score": 0.924, + "type": "image", + "image_path": "90e2f338200568e105b601ef0f6d470de812574f5e1411facb2387a68f64cf2e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 97, + 486, + 139.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 139.66666666666666, + 486, + 182.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 182.33333333333331, + 486, + 224.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 238, + 505, + 261 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "Figure 2: Validation BLEU AUCC and test BLEU for WMT (high is good). Comparison of regular", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 249, + 455, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 455, + 262 + ], + "score": 1.0, + "content": "transformer and reservoir transformer with FFN or Transformer reservoir layers added.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 114, + 300, + 486, + 429 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 300, + 486, + 429 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 300, + 486, + 429 + ], + "spans": [ + { + "bbox": [ + 114, + 300, + 486, + 429 + ], + "score": 0.965, + "type": "image", + "image_path": "fbf8f8b964d88c1ee463d7c528fdd11742523e63086217892f8a7d6bfafcc93f.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 114, + 300, + 486, + 343.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 343.0, + 486, + 386.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 114, + 386.0, + 486, + 429.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 443, + 503, + 466 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 442, + 504, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 504, + 457 + ], + "score": 1.0, + "content": "Figure 3: Validation BPC AUCC and test BPC on the enwik8 language modelling task (low is good).", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 455, + 384, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 384, + 468 + ], + "score": 1.0, + "content": "Comparison of regular and reservoir transformers for varying depths.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 504, + 523 + ], + "lines": [], + "index": 10.5, + "bbox_fs": [ + 105, + 500, + 505, + 523 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "Figure 1 shows the results for IWSLT. On the y-axis we show validation AUCC for the BLEU", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 167, + 551 + ], + "score": 1.0, + "content": "metric; on the", + "type": "text" + }, + { + "bbox": [ + 168, + 541, + 174, + 550 + ], + "score": 0.26, + "content": "\\mathbf { X } ^ { } -", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "-axis we show the number of updatable layers in the encoder. The performance", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 551, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 505, + 562 + ], + "score": 1.0, + "content": "of a regular transformer encoder with 6 layers and a reservoir transformer encoder with 6 layers", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 561, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 340, + 573 + ], + "score": 1.0, + "content": "plus N additional reservoir layers are plotted for the same", + "type": "text" + }, + { + "bbox": [ + 340, + 563, + 347, + 572 + ], + "score": 0.4, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 561, + 506, + 573 + ], + "score": 1.0, + "content": "-axis value to show the total number of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "updated layers. Plots for the total number of layers (updatable plus not-updatable, so essentially", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "shifted versions) are shown in Appendix E. Table 1 shows the time it took to achieve the maximum", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "validation BLEU score and how that relates to the regular transformer, demonstrating that reservoir", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 412, + 617 + ], + "score": 1.0, + "content": "transformers consistently converge faster in terms of wall-clock time, up to", + "type": "text" + }, + { + "bbox": [ + 412, + 605, + 432, + 615 + ], + "score": 0.87, + "content": "22 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "as much with the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 325, + 628 + ], + "score": 1.0, + "content": "same number of updateable layers. We save as much as", + "type": "text" + }, + { + "bbox": [ + 325, + 616, + 345, + 627 + ], + "score": 0.87, + "content": "27 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "time until convergence a 24 layer model", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 627, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 638 + ], + "score": 1.0, + "content": "on WMT, as shown in Table 3. One other noticeable point is that we can see that the T Reservoir", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 637, + 504, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 504, + 650 + ], + "score": 1.0, + "content": "achieves similar performance to LayerDrop on IWSLT and WMT in terms of wall-clock per epoch", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "and wall-clock time to the best performance. However, on both tasks, FFN Reservoir performs much", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "better than LayerDrop in terms of efficiency per epoch and achieves better/similar performance in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 671, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 506, + 683 + ], + "score": 1.0, + "content": "less time in each case. As a point of reference, a half hour gain on IWSLT translates to a gain of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 681, + 466, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 466, + 693 + ], + "score": 1.0, + "content": "several days in the training of bigger transformer models like GPT-3 (Brown et al., 2020).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 528, + 506, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 506, + 712 + ], + "score": 1.0, + "content": "We observe that reservoir transformers consistently perform better than, or are competitive to, reg-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 709, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 721 + ], + "score": 1.0, + "content": "ular transformers, both in terms of validation BLEU AUCC as well as test time BLEU, for all", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 721, + 213, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 213, + 733 + ], + "score": 1.0, + "content": "examined encoder depths.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 697, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 97, + 486, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 97, + 486, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 97, + 486, + 225 + ], + "spans": [ + { + "bbox": [ + 117, + 97, + 486, + 225 + ], + "score": 0.972, + "type": "image", + "image_path": "d4ae168b422fd5b5699ef8859159ebcc355c0e8a55b9feb4da847e6f8b55cace.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 97, + 486, + 139.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 139.66666666666666, + 486, + 182.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 182.33333333333331, + 486, + 224.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 115, + 238, + 492, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 237, + 493, + 252 + ], + "spans": [ + { + "bbox": [ + 116, + 237, + 493, + 252 + ], + "score": 1.0, + "content": "Figure 4: Downstream RoBERTa performance on SST-2 (left) and MultiNLI-matched (right).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 106, + 278, + 505, + 366 + ], + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "Figure 2 shows a similar trend for WMT. WMT is much larger and requires a much deeper encoder,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 290, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 302 + ], + "score": 1.0, + "content": "as illustrated by the fact that a certain minimum depth is required for reservoir transformers to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "score": 1.0, + "content": "achieve a comparable validation AUCC. At test time, reservoir transformers outperform regular", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "transformers for almost all encoder depths. The FFN reservoir transformer seems to work best in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "both cases, which is surprising because it does not have any self-attention component at all. This", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "score": 1.0, + "content": "finding shows that self-attention, or the mechanism to summarize context information, should be", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "learned if present. Once the context features have been gathered, a random projection via a fixed", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 356, + 322, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 322, + 367 + ], + "score": 1.0, + "content": "FFN module appears to be beneficial, at least for MT.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 108, + 388, + 237, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 237, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 237, + 400 + ], + "score": 1.0, + "content": "4.2 LANGUAGE MODELLING", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 505, + 467 + ], + "lines": [ + { + "bbox": [ + 106, + 411, + 504, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 504, + 423 + ], + "score": 1.0, + "content": "To examine whether the same findings hold for other tasks, we evaluate on the enwiki8 (LLC, 2009)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "language modelling task. We examine the BPC (bits per character) rate for a variety of network", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "depths (since the task is language modelling, these layers are in the decoder). The results show", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 442, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 506, + 459 + ], + "score": 1.0, + "content": "that we obtain consistently better BPC for lower depths, except for the 64-layer regular transformer,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 455, + 495, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 495, + 469 + ], + "score": 1.0, + "content": "which appears to be particularly optimal for this task. We observe similar trends during test time.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 109, + 488, + 317, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 320, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 320, + 500 + ], + "score": 1.0, + "content": "4.3 MASKED LANGUAGE MODEL PRETRAINING", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "We train RoBERTa (Liu et al., 2019) models from scratch at a variety of depths, both in the normal", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "and reservoir setting. We find that these networks show minor differences in their best perplexity", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "and similar AUCC perplexity (see Appendix D). We then examine the performance of these mod-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "els when fine-tuned on downstream tasks, specifically the well known SST-2 (Socher et al., 2013)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "and MultiNLI1 (Williams et al., 2017) tasks. When fine-tuning the reservoir models, we keep the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 567, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 482, + 578 + ], + "score": 1.0, + "content": "reservoir layers fixed (including them in fine-tuning did not work very well, see Appendix D).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 583, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "Figure 4 shows the results of fine-tuning. We observe that the reservoir transformer outperforms", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "normal RoBERTa at all depths in both tasks. At lower depth, the improvements are substantial. As", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 604, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 104, + 604, + 506, + 619 + ], + "score": 1.0, + "content": "a sanity check, we also experiment with freezing some of the layers in normal RoBERTa during", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "fine-tuning (Transformer frozen finetuned) and show that this helps a little but is still outperformed", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 628, + 222, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 222, + 640 + ], + "score": 1.0, + "content": "by the reservoir transformer.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 504, + 699 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "These findings suggest that you can train a RoBERTa model without updating all of the layers,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "achieve similar perplexity at a similar computational cost, but with better downstream performance.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "score": 1.0, + "content": "The fact that some layers can be kept random and entirely fixed during training, without sacrificing", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 504, + 690 + ], + "score": 1.0, + "content": "any performance, raises intriguing questions for “BERTology” (Rogers et al., 2020) and for the study", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 687, + 289, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 289, + 701 + ], + "score": 1.0, + "content": "of what different layers in transformers learn.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 721, + 271, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 720, + 272, + 733 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 272, + 733 + ], + "score": 1.0, + "content": "1We report results for MultiNLI-Matched.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 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 2021", + "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": [ + 117, + 97, + 486, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 97, + 486, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 97, + 486, + 225 + ], + "spans": [ + { + "bbox": [ + 117, + 97, + 486, + 225 + ], + "score": 0.972, + "type": "image", + "image_path": "d4ae168b422fd5b5699ef8859159ebcc355c0e8a55b9feb4da847e6f8b55cace.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 97, + 486, + 139.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 139.66666666666666, + 486, + 182.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 182.33333333333331, + 486, + 224.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 115, + 238, + 492, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 237, + 493, + 252 + ], + "spans": [ + { + "bbox": [ + 116, + 237, + 493, + 252 + ], + "score": 1.0, + "content": "Figure 4: Downstream RoBERTa performance on SST-2 (left) and MultiNLI-matched (right).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 106, + 278, + 505, + 366 + ], + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "Figure 2 shows a similar trend for WMT. WMT is much larger and requires a much deeper encoder,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 290, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 302 + ], + "score": 1.0, + "content": "as illustrated by the fact that a certain minimum depth is required for reservoir transformers to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "score": 1.0, + "content": "achieve a comparable validation AUCC. At test time, reservoir transformers outperform regular", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "transformers for almost all encoder depths. The FFN reservoir transformer seems to work best in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "both cases, which is surprising because it does not have any self-attention component at all. This", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "score": 1.0, + "content": "finding shows that self-attention, or the mechanism to summarize context information, should be", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "learned if present. Once the context features have been gathered, a random projection via a fixed", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 356, + 322, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 322, + 367 + ], + "score": 1.0, + "content": "FFN module appears to be beneficial, at least for MT.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 278, + 506, + 367 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 388, + 237, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 237, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 237, + 400 + ], + "score": 1.0, + "content": "4.2 LANGUAGE MODELLING", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 505, + 467 + ], + "lines": [ + { + "bbox": [ + 106, + 411, + 504, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 504, + 423 + ], + "score": 1.0, + "content": "To examine whether the same findings hold for other tasks, we evaluate on the enwiki8 (LLC, 2009)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "language modelling task. We examine the BPC (bits per character) rate for a variety of network", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "depths (since the task is language modelling, these layers are in the decoder). The results show", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 442, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 506, + 459 + ], + "score": 1.0, + "content": "that we obtain consistently better BPC for lower depths, except for the 64-layer regular transformer,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 455, + 495, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 495, + 469 + ], + "score": 1.0, + "content": "which appears to be particularly optimal for this task. We observe similar trends during test time.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 411, + 506, + 469 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 488, + 317, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 320, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 320, + 500 + ], + "score": 1.0, + "content": "4.3 MASKED LANGUAGE MODEL PRETRAINING", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "We train RoBERTa (Liu et al., 2019) models from scratch at a variety of depths, both in the normal", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "and reservoir setting. We find that these networks show minor differences in their best perplexity", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "and similar AUCC perplexity (see Appendix D). We then examine the performance of these mod-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "els when fine-tuned on downstream tasks, specifically the well known SST-2 (Socher et al., 2013)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "and MultiNLI1 (Williams et al., 2017) tasks. When fine-tuning the reservoir models, we keep the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 567, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 482, + 578 + ], + "score": 1.0, + "content": "reservoir layers fixed (including them in fine-tuning did not work very well, see Appendix D).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 510, + 505, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 583, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "Figure 4 shows the results of fine-tuning. We observe that the reservoir transformer outperforms", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "normal RoBERTa at all depths in both tasks. At lower depth, the improvements are substantial. As", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 604, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 104, + 604, + 506, + 619 + ], + "score": 1.0, + "content": "a sanity check, we also experiment with freezing some of the layers in normal RoBERTa during", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "fine-tuning (Transformer frozen finetuned) and show that this helps a little but is still outperformed", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 628, + 222, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 222, + 640 + ], + "score": 1.0, + "content": "by the reservoir transformer.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 104, + 583, + 506, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 504, + 699 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "These findings suggest that you can train a RoBERTa model without updating all of the layers,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "achieve similar perplexity at a similar computational cost, but with better downstream performance.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "score": 1.0, + "content": "The fact that some layers can be kept random and entirely fixed during training, without sacrificing", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 504, + 690 + ], + "score": 1.0, + "content": "any performance, raises intriguing questions for “BERTology” (Rogers et al., 2020) and for the study", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 687, + 289, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 289, + 701 + ], + "score": 1.0, + "content": "of what different layers in transformers learn.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 644, + 506, + 701 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 204, + 89, + 403, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 204, + 89, + 403, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 204, + 89, + 403, + 198 + ], + "spans": [ + { + "bbox": [ + 204, + 89, + 403, + 198 + ], + "score": 0.968, + "type": "image", + "image_path": "8dc191b8c7100f587adeeeb66439469b208ef40391084c7b050f44a7cb3a4cb5.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 89, + 403, + 102.625 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 204, + 102.625, + 403, + 116.25 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 204, + 116.25, + 403, + 129.875 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 204, + 129.875, + 403, + 143.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 204, + 143.5, + 403, + 157.125 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 204, + 157.125, + 403, + 170.75 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 204, + 170.75, + 403, + 184.375 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 204, + 184.375, + 403, + 198.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 177, + 209, + 434, + 221 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 177, + 208, + 434, + 222 + ], + "spans": [ + { + "bbox": [ + 177, + 208, + 434, + 222 + ], + "score": 1.0, + "content": "Figure 5: IWSLT comparison of normal v frozen v backskipped", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 5.75 + }, + { + "type": "title", + "bbox": [ + 107, + 245, + 196, + 256 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 197, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 197, + 257 + ], + "score": 1.0, + "content": "4.4 BACKSKIPPING", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 265, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 280 + ], + "score": 1.0, + "content": "With the reservoir transformers as described above, we obtain better efficiency by skipping the “gra-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 277, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 505, + 291 + ], + "score": 1.0, + "content": "dient application” matrix addition step in some of the layers (i.e., updating the weights). One step", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 289, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 505, + 301 + ], + "score": 1.0, + "content": "further would be to investigate skipping the entire backward pass for reservoirs altogether, which", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "would save us from having to do the much more expensive matrix multiplication for these layers", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 309, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 324 + ], + "score": 1.0, + "content": "that is required for the propagation of gradients. We report on preliminary experiments where in the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 320, + 335 + ], + "score": 1.0, + "content": "backward pass we replace the gradients for the layer", + "type": "text" + }, + { + "bbox": [ + 320, + 322, + 332, + 333 + ], + "score": 0.87, + "content": "L _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 321, + 430, + 335 + ], + "score": 1.0, + "content": "going into the reservoir", + "type": "text" + }, + { + "bbox": [ + 430, + 321, + 451, + 333 + ], + "score": 0.91, + "content": "L _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "with a noisy", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 331, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 346 + ], + "score": 1.0, + "content": "estimate (Jaderberg et al., 2017; Czarnecki et al., 2017). Promisingly, Oktay et al. (2020) recently", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "asked “why spend resources on exact gradients when we’re going to use stochastic optimization?”", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 354, + 414, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 414, + 367 + ], + "score": 1.0, + "content": "and show that you can do randomized auto-differentiation quite successfully.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 371, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 504, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 313, + 386 + ], + "score": 1.0, + "content": "Here, rather than minimizing the actual gradients", + "type": "text" + }, + { + "bbox": [ + 313, + 371, + 333, + 385 + ], + "score": 0.91, + "content": "\\frac { \\partial L _ { i } } { \\partial \\theta ^ { L _ { i } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 369, + 504, + 386 + ], + "score": 1.0, + "content": ", we minimize their expectation and train", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 381, + 504, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 366, + 397 + ], + "score": 1.0, + "content": "via continuous-action REINFORCE (Williams, 1992). That is,", + "type": "text" + }, + { + "bbox": [ + 367, + 384, + 378, + 394 + ], + "score": 0.89, + "content": "L _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 381, + 455, + 397 + ], + "score": 1.0, + "content": "becomes a policy", + "type": "text" + }, + { + "bbox": [ + 455, + 385, + 466, + 394 + ], + "score": 0.78, + "content": "\\pi _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 381, + 473, + 397 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 473, + 384, + 504, + 395 + ], + "score": 0.84, + "content": "s \\mu", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 510, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 125, + 422 + ], + "score": 1.0, + "content": "whei.e.,", + "type": "text" + }, + { + "bbox": [ + 126, + 404, + 220, + 418 + ], + "score": 0.91, + "content": "\\textstyle { \\frac { 1 } { n } } \\sum _ { i = 0 } ^ { n } ( { \\dot { R } } ^ { i } - V ^ { i } ( a ) ) ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 394, + 262, + 406 + ], + "score": 0.92, + "content": "a \\sim \\mathcal { N } ( \\mu , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 394, + 336, + 422 + ], + "score": 1.0, + "content": ". We train to miniREINFORCE loss", + "type": "text" + }, + { + "bbox": [ + 336, + 405, + 433, + 417 + ], + "score": 0.92, + "content": "\\mathbb { E } _ { a } \\left[ \\log ( { \\bar { a } } ) \\left( R - { \\bar { V } } ( a ) \\right) \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 394, + 510, + 422 + ], + "score": 1.0, + "content": "ion loss via MSE,, where the value", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 415, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 104, + 415, + 141, + 429 + ], + "score": 1.0, + "content": "network", + "type": "text" + }, + { + "bbox": [ + 142, + 417, + 150, + 426 + ], + "score": 0.66, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 415, + 233, + 429 + ], + "score": 1.0, + "content": "acts as the baseline.", + "type": "text" + }, + { + "bbox": [ + 233, + 416, + 242, + 426 + ], + "score": 0.79, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 415, + 460, + 429 + ], + "score": 1.0, + "content": "is defined as the mean of the gradients of the top layer", + "type": "text" + }, + { + "bbox": [ + 460, + 416, + 481, + 428 + ], + "score": 0.93, + "content": "L _ { i + 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 415, + 506, + 429 + ], + "score": 1.0, + "content": ", with", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 427, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 440 + ], + "score": 1.0, + "content": "the sign flipped. Thus, simply put, we train to minimize the expectation of the true gradients at the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 439, + 504, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 504, + 450 + ], + "score": 1.0, + "content": "layer directly following the reservoir. We employ an annealing scheme where we first train the value", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "network and propagate the true gradients during warmup. Afterwards, we anneal the probability", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "of backskipping rather than performing a true backward pass (multiplying the probability by 0.99", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 387, + 484 + ], + "score": 1.0, + "content": "every iteration until we only backskip). We experimented with setting", + "type": "text" + }, + { + "bbox": [ + 388, + 471, + 397, + 481 + ], + "score": 0.82, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 471, + 506, + 484 + ], + "score": 1.0, + "content": "to the negation of the total", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 480, + 478, + 496 + ], + "spans": [ + { + "bbox": [ + 104, + 480, + 478, + 496 + ], + "score": 1.0, + "content": "loss as well but found the current reward to work better. We call this approach backskipping.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 498, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "Figure 5 shows the results as validation BLEU over time. We observe that this approach helps", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 510, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 522 + ], + "score": 1.0, + "content": "especially during the earlier stages of training. Although it does not match the performance of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 520, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 533 + ], + "score": 1.0, + "content": "the approach with true gradients quite yet, it actually performs competitively. Backskipping looks", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "promising as an approach to further reduce computational costs, and would be even more efficient", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "score": 1.0, + "content": "from a hardware perspective since the circuitry for such layers (which do not need to propagate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 554, + 299, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 299, + 567 + ], + "score": 1.0, + "content": "gradients) can effectively be hardwired entirely.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 108, + 583, + 210, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 213, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 213, + 598 + ], + "score": 1.0, + "content": "5 RELATED WORK", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "score": 1.0, + "content": "Recent work has shown that modern NLP models are able to function with different numbers of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "layers for different examples (Elbayad et al., 2019; Fan et al., 2019); that different layers special-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "ize for different purposes (Zhang et al., 2019); that layers can be compressed (Li et al., 2020);", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "and, that layers can be reordered (Press et al., 2019). There is a growing body of work in efficient", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "self-attention networks (Tay et al., 2020b), such as linear attention (Wang et al., 2020), on how", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "score": 1.0, + "content": "to process long context information (Beltagy et al., 2020) and on approximations to make trans-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "formers more scalable (Kitaev et al., 2020; Katharopoulos et al., 2020). BigBIRD (Zaheer et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "score": 1.0, + "content": "2020) provides random keys as additional inputs to its attention mechanism. Locality sensitive", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "hashing (LSH) as employed e.g. in Reformer (Kitaev et al., 2020) utilizes a fixed random projec-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "tion. Performer (Choromanski et al., 2020) computes the transformer’s multi-head attention weights", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "as a fixed orthogonal random projection. Closely related to this work, Tay et al. (2020a) showed", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 41 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 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 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 204, + 89, + 403, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 204, + 89, + 403, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 204, + 89, + 403, + 198 + ], + "spans": [ + { + "bbox": [ + 204, + 89, + 403, + 198 + ], + "score": 0.968, + "type": "image", + "image_path": "8dc191b8c7100f587adeeeb66439469b208ef40391084c7b050f44a7cb3a4cb5.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 89, + 403, + 102.625 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 204, + 102.625, + 403, + 116.25 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 204, + 116.25, + 403, + 129.875 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 204, + 129.875, + 403, + 143.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 204, + 143.5, + 403, + 157.125 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 204, + 157.125, + 403, + 170.75 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 204, + 170.75, + 403, + 184.375 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 204, + 184.375, + 403, + 198.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 177, + 209, + 434, + 221 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 177, + 208, + 434, + 222 + ], + "spans": [ + { + "bbox": [ + 177, + 208, + 434, + 222 + ], + "score": 1.0, + "content": "Figure 5: IWSLT comparison of normal v frozen v backskipped", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 5.75 + }, + { + "type": "title", + "bbox": [ + 107, + 245, + 196, + 256 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 197, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 197, + 257 + ], + "score": 1.0, + "content": "4.4 BACKSKIPPING", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 265, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 280 + ], + "score": 1.0, + "content": "With the reservoir transformers as described above, we obtain better efficiency by skipping the “gra-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 277, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 505, + 291 + ], + "score": 1.0, + "content": "dient application” matrix addition step in some of the layers (i.e., updating the weights). One step", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 289, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 505, + 301 + ], + "score": 1.0, + "content": "further would be to investigate skipping the entire backward pass for reservoirs altogether, which", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "would save us from having to do the much more expensive matrix multiplication for these layers", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 309, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 324 + ], + "score": 1.0, + "content": "that is required for the propagation of gradients. We report on preliminary experiments where in the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 320, + 335 + ], + "score": 1.0, + "content": "backward pass we replace the gradients for the layer", + "type": "text" + }, + { + "bbox": [ + 320, + 322, + 332, + 333 + ], + "score": 0.87, + "content": "L _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 321, + 430, + 335 + ], + "score": 1.0, + "content": "going into the reservoir", + "type": "text" + }, + { + "bbox": [ + 430, + 321, + 451, + 333 + ], + "score": 0.91, + "content": "L _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "with a noisy", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 331, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 346 + ], + "score": 1.0, + "content": "estimate (Jaderberg et al., 2017; Czarnecki et al., 2017). Promisingly, Oktay et al. (2020) recently", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "asked “why spend resources on exact gradients when we’re going to use stochastic optimization?”", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 354, + 414, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 414, + 367 + ], + "score": 1.0, + "content": "and show that you can do randomized auto-differentiation quite successfully.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 265, + 505, + 367 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 371, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 504, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 313, + 386 + ], + "score": 1.0, + "content": "Here, rather than minimizing the actual gradients", + "type": "text" + }, + { + "bbox": [ + 313, + 371, + 333, + 385 + ], + "score": 0.91, + "content": "\\frac { \\partial L _ { i } } { \\partial \\theta ^ { L _ { i } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 369, + 504, + 386 + ], + "score": 1.0, + "content": ", we minimize their expectation and train", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 381, + 504, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 366, + 397 + ], + "score": 1.0, + "content": "via continuous-action REINFORCE (Williams, 1992). That is,", + "type": "text" + }, + { + "bbox": [ + 367, + 384, + 378, + 394 + ], + "score": 0.89, + "content": "L _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 381, + 455, + 397 + ], + "score": 1.0, + "content": "becomes a policy", + "type": "text" + }, + { + "bbox": [ + 455, + 385, + 466, + 394 + ], + "score": 0.78, + "content": "\\pi _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 381, + 473, + 397 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 473, + 384, + 504, + 395 + ], + "score": 0.84, + "content": "s \\mu", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 510, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 125, + 422 + ], + "score": 1.0, + "content": "whei.e.,", + "type": "text" + }, + { + "bbox": [ + 126, + 404, + 220, + 418 + ], + "score": 0.91, + "content": "\\textstyle { \\frac { 1 } { n } } \\sum _ { i = 0 } ^ { n } ( { \\dot { R } } ^ { i } - V ^ { i } ( a ) ) ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 394, + 262, + 406 + ], + "score": 0.92, + "content": "a \\sim \\mathcal { N } ( \\mu , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 394, + 336, + 422 + ], + "score": 1.0, + "content": ". We train to miniREINFORCE loss", + "type": "text" + }, + { + "bbox": [ + 336, + 405, + 433, + 417 + ], + "score": 0.92, + "content": "\\mathbb { E } _ { a } \\left[ \\log ( { \\bar { a } } ) \\left( R - { \\bar { V } } ( a ) \\right) \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 394, + 510, + 422 + ], + "score": 1.0, + "content": "ion loss via MSE,, where the value", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 415, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 104, + 415, + 141, + 429 + ], + "score": 1.0, + "content": "network", + "type": "text" + }, + { + "bbox": [ + 142, + 417, + 150, + 426 + ], + "score": 0.66, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 415, + 233, + 429 + ], + "score": 1.0, + "content": "acts as the baseline.", + "type": "text" + }, + { + "bbox": [ + 233, + 416, + 242, + 426 + ], + "score": 0.79, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 415, + 460, + 429 + ], + "score": 1.0, + "content": "is defined as the mean of the gradients of the top layer", + "type": "text" + }, + { + "bbox": [ + 460, + 416, + 481, + 428 + ], + "score": 0.93, + "content": "L _ { i + 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 415, + 506, + 429 + ], + "score": 1.0, + "content": ", with", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 427, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 440 + ], + "score": 1.0, + "content": "the sign flipped. Thus, simply put, we train to minimize the expectation of the true gradients at the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 439, + 504, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 504, + 450 + ], + "score": 1.0, + "content": "layer directly following the reservoir. We employ an annealing scheme where we first train the value", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "network and propagate the true gradients during warmup. Afterwards, we anneal the probability", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "of backskipping rather than performing a true backward pass (multiplying the probability by 0.99", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 387, + 484 + ], + "score": 1.0, + "content": "every iteration until we only backskip). We experimented with setting", + "type": "text" + }, + { + "bbox": [ + 388, + 471, + 397, + 481 + ], + "score": 0.82, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 471, + 506, + 484 + ], + "score": 1.0, + "content": "to the negation of the total", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 480, + 478, + 496 + ], + "spans": [ + { + "bbox": [ + 104, + 480, + 478, + 496 + ], + "score": 1.0, + "content": "loss as well but found the current reward to work better. We call this approach backskipping.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 369, + 510, + 496 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 498, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "Figure 5 shows the results as validation BLEU over time. We observe that this approach helps", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 510, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 522 + ], + "score": 1.0, + "content": "especially during the earlier stages of training. Although it does not match the performance of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 520, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 533 + ], + "score": 1.0, + "content": "the approach with true gradients quite yet, it actually performs competitively. Backskipping looks", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "promising as an approach to further reduce computational costs, and would be even more efficient", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "score": 1.0, + "content": "from a hardware perspective since the circuitry for such layers (which do not need to propagate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 554, + 299, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 299, + 567 + ], + "score": 1.0, + "content": "gradients) can effectively be hardwired entirely.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 498, + 506, + 567 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 583, + 210, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 213, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 213, + 598 + ], + "score": 1.0, + "content": "5 RELATED WORK", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "score": 1.0, + "content": "Recent work has shown that modern NLP models are able to function with different numbers of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "layers for different examples (Elbayad et al., 2019; Fan et al., 2019); that different layers special-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "ize for different purposes (Zhang et al., 2019); that layers can be compressed (Li et al., 2020);", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "and, that layers can be reordered (Press et al., 2019). There is a growing body of work in efficient", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "self-attention networks (Tay et al., 2020b), such as linear attention (Wang et al., 2020), on how", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "score": 1.0, + "content": "to process long context information (Beltagy et al., 2020) and on approximations to make trans-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "formers more scalable (Kitaev et al., 2020; Katharopoulos et al., 2020). BigBIRD (Zaheer et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "score": 1.0, + "content": "2020) provides random keys as additional inputs to its attention mechanism. Locality sensitive", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "hashing (LSH) as employed e.g. in Reformer (Kitaev et al., 2020) utilizes a fixed random projec-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "tion. Performer (Choromanski et al., 2020) computes the transformer’s multi-head attention weights", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "as a fixed orthogonal random projection. Closely related to this work, Tay et al. (2020a) showed", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "that randomized alignment matrices in their “Synthesizer” architecture are sufficient for many NLP", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "tasks. While these works focus on random attention, we show that entire layers can be random and", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "fixed. We also show that entire layers can be replaced by fixed random projections that do not have", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 210, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 210, + 128 + ], + "score": 1.0, + "content": "any attention whatsoever.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 611, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "that randomized alignment matrices in their “Synthesizer” architecture are sufficient for many NLP", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "tasks. While these works focus on random attention, we show that entire layers can be random and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "fixed. We also show that entire layers can be replaced by fixed random projections that do not have", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 210, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 210, + 128 + ], + "score": 1.0, + "content": "any attention whatsoever.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 241 + ], + "lines": [ + { + "bbox": [ + 106, + 133, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 505, + 144 + ], + "score": 1.0, + "content": "Beyond transformers, random features have been extensively explored. Examples of this include", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "FreezeOut (Brock et al., 2017), deep reservoir computing networks (Scardapane & Wang, 2017;", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "Gallicchio & Micheli, 2017), as well as applications in domains as varied as text classification (Con-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 166, + 504, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 504, + 177 + ], + "score": 1.0, + "content": "neau et al., 2017; Zhang & Bowman, 2018; Wieting & Kiela, 2019) or music classification (Pons", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "& Serra, 2019). It is well known that randomly initialized networks can display impressive perfor-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "mance on their own (Ulyanov et al., 2018; Rosenfeld & Tsotsos, 2019; Ramanujan et al., 2020),", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "which underlies, for example, the recently popularized lottery ticket hypothesis (Frankle & Carbin,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "score": 1.0, + "content": "2018; Zhou et al., 2019). We know that learning deep overparameterized networks appears to help", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "in general (Li & Liang, 2018; Du et al., 2019). Our method represents an easy and cheap way to add", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 315, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 315, + 244 + ], + "score": 1.0, + "content": "both depth and parameters to transformer networks.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 258, + 195, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 197, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 197, + 274 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 282, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "This work demonstrated that state-of-the-art transformer architectures can be trained without updat-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "ing all of the layers. This complements a long history in machine learning of harnessing the power", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 303, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 318 + ], + "score": 1.0, + "content": "of random features. In most cases, “reservoir transformers” achieve better performance-efficiency", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "trade-offs as measured by our newly introduced AUCC metric, and better test set generalization,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "on a variety of tasks and in a variety of settings. Future work includes further investigating hybrid", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "score": 1.0, + "content": "networks and backskipping architectures, as well as utilizing pruning strategies at inference time, in", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 349, + 375, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 375, + 361 + ], + "score": 1.0, + "content": "order to try to obtain even better performance/efficiency trade-offs.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 376, + 175, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 377, + 176, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 176, + 389 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 394, + 506, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "Thomas Bachlechner, Bodhisattwa Prasad Majumder, Huanru Henry Mao, Garrison W Cottrell,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 116, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "and Julian McAuley. Rezero is all you need: Fast convergence at large depth. arXiv preprint", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 417, + 218, + 428 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 218, + 428 + ], + "score": 1.0, + "content": "arXiv:2003.04887, 2020.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "Alexei Baevski, Steffen Schneider, and Michael Auli. vq-wav2vec: Self-supervised learning of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 446, + 408, + 458 + ], + "spans": [ + { + "bbox": [ + 116, + 446, + 408, + 458 + ], + "score": 1.0, + "content": "discrete speech representations. arXiv preprint arXiv:1910.05453, 2019.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "Eric B Baum. On the capabilities of multilayer perceptrons. Journal of complexity, 4(3):193–215,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 474, + 142, + 487 + ], + "spans": [ + { + "bbox": [ + 116, + 474, + 142, + 487 + ], + "score": 1.0, + "content": "1988.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 491, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 506 + ], + "score": 1.0, + "content": "Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 117, + 504, + 278, + 515 + ], + "spans": [ + { + "bbox": [ + 117, + 504, + 278, + 515 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2004.05150, 2020.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "score": 1.0, + "content": "Hans-Dieter Block. The perceptron: A model for brain functioning. i. Reviews of Modern Physics,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 532, + 185, + 543 + ], + "spans": [ + { + "bbox": [ + 117, + 532, + 185, + 543 + ], + "score": 1.0, + "content": "34(1):123, 1962.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Ondˇrej Bojar, Christian Buck, Christian Federmann, Barry Haddow, Philipp Koehn, Johannes Lev-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 115, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "eling, Christof Monz, Pavel Pecina, Matt Post, Herve Saint-Amand, Radu Soricut, Lucia Specia,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 116, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "and Ales Tamchyna. Findings of the 2014 workshop on statistical machine translation. In ˇ Pro-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 116, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "ceedings of the Ninth Workshop on Statistical Machine Translation, Baltimore, Maryland, USA,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 594, + 337, + 607 + ], + "spans": [ + { + "bbox": [ + 115, + 594, + 337, + 607 + ], + "score": 1.0, + "content": "June 2014. Association for Computational Linguistics.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "A Borsellino and A Gamba. An outline of a mathematical theory of papa. Il Nuovo Cimento (1955-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 117, + 623, + 234, + 634 + ], + "spans": [ + { + "bbox": [ + 117, + 623, + 234, + 634 + ], + "score": 1.0, + "content": "1965), 20(2):221–231, 1961.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "score": 1.0, + "content": "Andrew P Bradley. The use of the area under the roc curve in the evaluation of machine learning", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 652, + 347, + 664 + ], + "spans": [ + { + "bbox": [ + 116, + 652, + 347, + 664 + ], + "score": 1.0, + "content": "algorithms. Pattern recognition, 30(7):1145–1159, 1997.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 683 + ], + "score": 1.0, + "content": "Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Freezeout: Accelerate training", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 681, + 412, + 694 + ], + "spans": [ + { + "bbox": [ + 116, + 681, + 412, + 694 + ], + "score": 1.0, + "content": "by progressively freezing layers. arXiv preprint arXiv:1706.04983, 2017.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 721, + 353, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 353, + 733 + ], + "score": 1.0, + "content": "few-shot learners. arXiv preprint arXiv:2005.14165, 2020.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 35 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 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 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 126 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 83, + 506, + 128 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 241 + ], + "lines": [ + { + "bbox": [ + 106, + 133, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 505, + 144 + ], + "score": 1.0, + "content": "Beyond transformers, random features have been extensively explored. Examples of this include", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "FreezeOut (Brock et al., 2017), deep reservoir computing networks (Scardapane & Wang, 2017;", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "Gallicchio & Micheli, 2017), as well as applications in domains as varied as text classification (Con-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 166, + 504, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 504, + 177 + ], + "score": 1.0, + "content": "neau et al., 2017; Zhang & Bowman, 2018; Wieting & Kiela, 2019) or music classification (Pons", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "& Serra, 2019). It is well known that randomly initialized networks can display impressive perfor-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "mance on their own (Ulyanov et al., 2018; Rosenfeld & Tsotsos, 2019; Ramanujan et al., 2020),", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "which underlies, for example, the recently popularized lottery ticket hypothesis (Frankle & Carbin,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "score": 1.0, + "content": "2018; Zhou et al., 2019). We know that learning deep overparameterized networks appears to help", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "in general (Li & Liang, 2018; Du et al., 2019). Our method represents an easy and cheap way to add", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 315, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 315, + 244 + ], + "score": 1.0, + "content": "both depth and parameters to transformer networks.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 133, + 505, + 244 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 258, + 195, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 197, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 197, + 274 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 282, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "This work demonstrated that state-of-the-art transformer architectures can be trained without updat-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "ing all of the layers. This complements a long history in machine learning of harnessing the power", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 303, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 318 + ], + "score": 1.0, + "content": "of random features. In most cases, “reservoir transformers” achieve better performance-efficiency", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "trade-offs as measured by our newly introduced AUCC metric, and better test set generalization,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "on a variety of tasks and in a variety of settings. Future work includes further investigating hybrid", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "score": 1.0, + "content": "networks and backskipping architectures, as well as utilizing pruning strategies at inference time, in", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 349, + 375, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 375, + 361 + ], + "score": 1.0, + "content": "order to try to obtain even better performance/efficiency trade-offs.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 282, + 505, + 361 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 376, + 175, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 377, + 176, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 176, + 389 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "list", + "bbox": [ + 105, + 394, + 506, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "Thomas Bachlechner, Bodhisattwa Prasad Majumder, Huanru Henry Mao, Garrison W Cottrell,", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 116, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "and Julian McAuley. Rezero is all you need: Fast convergence at large depth. arXiv preprint", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 417, + 218, + 428 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 218, + 428 + ], + "score": 1.0, + "content": "arXiv:2003.04887, 2020.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "Alexei Baevski, Steffen Schneider, and Michael Auli. vq-wav2vec: Self-supervised learning of", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 446, + 408, + 458 + ], + "spans": [ + { + "bbox": [ + 116, + 446, + 408, + 458 + ], + "score": 1.0, + "content": "discrete speech representations. arXiv preprint arXiv:1910.05453, 2019.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "Eric B Baum. On the capabilities of multilayer perceptrons. Journal of complexity, 4(3):193–215,", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 474, + 142, + 487 + ], + "spans": [ + { + "bbox": [ + 116, + 474, + 142, + 487 + ], + "score": 1.0, + "content": "1988.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 491, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 506 + ], + "score": 1.0, + "content": "Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer.", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 117, + 504, + 278, + 515 + ], + "spans": [ + { + "bbox": [ + 117, + 504, + 278, + 515 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2004.05150, 2020.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "score": 1.0, + "content": "Hans-Dieter Block. The perceptron: A model for brain functioning. i. Reviews of Modern Physics,", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 117, + 532, + 185, + 543 + ], + "spans": [ + { + "bbox": [ + 117, + 532, + 185, + 543 + ], + "score": 1.0, + "content": "34(1):123, 1962.", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Ondˇrej Bojar, Christian Buck, Christian Federmann, Barry Haddow, Philipp Koehn, Johannes Lev-", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 115, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "eling, Christof Monz, Pavel Pecina, Matt Post, Herve Saint-Amand, Radu Soricut, Lucia Specia,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 116, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "and Ales Tamchyna. Findings of the 2014 workshop on statistical machine translation. In ˇ Pro-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 116, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "ceedings of the Ninth Workshop on Statistical Machine Translation, Baltimore, Maryland, USA,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 594, + 337, + 607 + ], + "spans": [ + { + "bbox": [ + 115, + 594, + 337, + 607 + ], + "score": 1.0, + "content": "June 2014. Association for Computational Linguistics.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "A Borsellino and A Gamba. An outline of a mathematical theory of papa. Il Nuovo Cimento (1955-", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 117, + 623, + 234, + 634 + ], + "spans": [ + { + "bbox": [ + 117, + 623, + 234, + 634 + ], + "score": 1.0, + "content": "1965), 20(2):221–231, 1961.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "score": 1.0, + "content": "Andrew P Bradley. The use of the area under the roc curve in the evaluation of machine learning", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 652, + 347, + 664 + ], + "spans": [ + { + "bbox": [ + 116, + 652, + 347, + 664 + ], + "score": 1.0, + "content": "algorithms. Pattern recognition, 30(7):1145–1159, 1997.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 683 + ], + "score": 1.0, + "content": "Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Freezeout: Accelerate training", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 681, + 412, + 694 + ], + "spans": [ + { + "bbox": [ + 116, + 681, + 412, + 694 + ], + "score": 1.0, + "content": "by progressively freezing layers. arXiv preprint arXiv:1706.04983, 2017.", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal,", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 721, + 353, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 353, + 733 + ], + "score": 1.0, + "content": "few-shot learners. arXiv preprint arXiv:2005.14165, 2020.", + "type": "text" + } + ], + "index": 47, + "is_list_end_line": true + } + ], + "index": 35, + "bbox_fs": [ + 105, + 394, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "and Sergey Zagoruyko. End-to-end object detection with transformers. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 219, + 115 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 219, + 115 + ], + "score": 1.0, + "content": "arXiv:2005.12872, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 104, + 124, + 504, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "score": 1.0, + "content": "M. Cettolo, J. Niehues, S. Stuker, L. Bentivogli, and Marcello Federico. Report on the 11 th iwslt ¨", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 385, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 385, + 147 + ], + "score": 1.0, + "content": "evaluation campaign , iwslt 2014. In Proceedings of IWSLT, 2015.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol- ¨", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 166, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 115, + 166, + 505, + 179 + ], + "score": 1.0, + "content": "ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 177, + 412, + 189 + ], + "spans": [ + { + "bbox": [ + 116, + 177, + 412, + 189 + ], + "score": 1.0, + "content": "for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Jared Davis, Tamas", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 116, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "Sarlos, David Belanger, Lucy Colwell, and Adrian Weller. Masked language modeling for pro-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 218, + 491, + 232 + ], + "spans": [ + { + "bbox": [ + 115, + 218, + 491, + 232 + ], + "score": 1.0, + "content": "teins via linearly scalable long-context transformers. arXiv preprint arXiv:2006.03555, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 503, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 252 + ], + "score": 1.0, + "content": "Alexis Conneau, Douwe Kiela, Holger Schwenk, Loic Barrault, and Antoine Bordes. Super-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 249, + 504, + 262 + ], + "spans": [ + { + "bbox": [ + 117, + 249, + 504, + 262 + ], + "score": 1.0, + "content": "vised learning of universal sentence representations from natural language inference data. arXiv", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 262, + 254, + 272 + ], + "spans": [ + { + "bbox": [ + 115, + 262, + 254, + 272 + ], + "score": 1.0, + "content": "preprint arXiv:1705.02364, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 105, + 280, + 504, + 304 + ], + "lines": [ + { + "bbox": [ + 105, + 279, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 295 + ], + "score": 1.0, + "content": "Thomas M Cover. Geometrical and statistical properties of systems of linear inequalities with ap-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 291, + 504, + 304 + ], + "spans": [ + { + "bbox": [ + 116, + 291, + 504, + 304 + ], + "score": 1.0, + "content": "plications in pattern recognition. IEEE transactions on electronic computers, (3):326–334, 1965.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 311, + 505, + 346 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 326 + ], + "score": 1.0, + "content": "Wojciech Marian Czarnecki, Grzegorz Swirszcz, Max Jaderberg, Simon Osindero, Oriol Vinyals, ´", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 323, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 115, + 323, + 505, + 337 + ], + "score": 1.0, + "content": "and Koray Kavukcuoglu. Understanding synthetic gradients and decoupled neural interfaces.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 335, + 279, + 347 + ], + "spans": [ + { + "bbox": [ + 116, + 335, + 279, + 347 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1703.00522, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "Transformer-XL: Attentive language models beyond a fixed-length context. In Proceedings of the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 376, + 505, + 389 + ], + "score": 1.0, + "content": "57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, July 2019.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 388, + 290, + 400 + ], + "spans": [ + { + "bbox": [ + 116, + 388, + 290, + 400 + ], + "score": 1.0, + "content": "Association for Computational Linguistics.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 407, + 504, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 407, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 421 + ], + "score": 1.0, + "content": "Amit Daniely, Roy Frostig, and Yoram Singer. Toward deeper understanding of neural networks:", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 116, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "The power of initialization and a dual view on expressivity. In Advances In Neural Information", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 429, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 116, + 429, + 289, + 443 + ], + "score": 1.0, + "content": "Processing Systems, pp. 2253–2261, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 104, + 449, + 504, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 464 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 460, + 500, + 473 + ], + "spans": [ + { + "bbox": [ + 116, + 460, + 500, + 473 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 504, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "Simon Du, Jason Lee, Haochuan Li, Liwei Wang, and Xiyu Zhai. Gradient descent finds global", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 115, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "minima of deep neural networks. In International Conference on Machine Learning, pp. 1675–", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 501, + 167, + 514 + ], + "spans": [ + { + "bbox": [ + 116, + 501, + 167, + 514 + ], + "score": 1.0, + "content": "1685, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 522, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "Sergey Edunov, Myle Ott, Michael Auli, David Grangier, and Marc’Aurelio Ranzato. Classical", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "structured prediction losses for sequence to sequence learning. In Proceedings of the 2018 Con-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 543, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 114, + 543, + 505, + 557 + ], + "score": 1.0, + "content": "ference of the North American Chapter of the Association for Computational Linguistics: Human", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 115, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "Language Technologies, Volume 1 (Long Papers), New Orleans, Louisiana, June 2018. Associa-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 567, + 259, + 578 + ], + "spans": [ + { + "bbox": [ + 116, + 567, + 259, + 578 + ], + "score": 1.0, + "content": "tion for Computational Linguistics.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 104, + 585, + 504, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive transformer. arXiv", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 597, + 254, + 608 + ], + "spans": [ + { + "bbox": [ + 115, + 597, + 254, + 608 + ], + "score": 1.0, + "content": "preprint arXiv:1910.10073, 2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 104, + 616, + 504, + 640 + ], + "lines": [ + { + "bbox": [ + 104, + 615, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "Joseph Enguehard, Dan Busbridge, Vitalii Zhelezniak, and Nils Hammerla. Neural language priors.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 628, + 278, + 639 + ], + "spans": [ + { + "bbox": [ + 116, + 628, + 278, + 639 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1910.03492, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 104, + 647, + 504, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 505, + 660 + ], + "score": 1.0, + "content": "Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 658, + 358, + 671 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 358, + 671 + ], + "score": 1.0, + "content": "structured dropout. arXiv preprint arXiv:1909.11556, 2019.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 690, + 321, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 321, + 702 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1803.03635, 2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "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": "Jonathan Frankle, David J Schwab, and Ari S Morcos. Training batchnorm and only batchnorm: On", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 721, + 476, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 476, + 732 + ], + "score": 1.0, + "content": "the expressive power of random features in cnns. arXiv preprint arXiv:2003.00152, 2020.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 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 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "and Sergey Zagoruyko. End-to-end object detection with transformers. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 219, + 115 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 219, + 115 + ], + "score": 1.0, + "content": "arXiv:2005.12872, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 115 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 124, + 504, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "score": 1.0, + "content": "M. Cettolo, J. Niehues, S. Stuker, L. Bentivogli, and Marcello Federico. Report on the 11 th iwslt ¨", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 385, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 385, + 147 + ], + "score": 1.0, + "content": "evaluation campaign , iwslt 2014. In Proceedings of IWSLT, 2015.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 123, + 505, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol- ¨", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 166, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 115, + 166, + 505, + 179 + ], + "score": 1.0, + "content": "ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 177, + 412, + 189 + ], + "spans": [ + { + "bbox": [ + 116, + 177, + 412, + 189 + ], + "score": 1.0, + "content": "for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 154, + 505, + 189 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Jared Davis, Tamas", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 116, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "Sarlos, David Belanger, Lucy Colwell, and Adrian Weller. Masked language modeling for pro-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 218, + 491, + 232 + ], + "spans": [ + { + "bbox": [ + 115, + 218, + 491, + 232 + ], + "score": 1.0, + "content": "teins via linearly scalable long-context transformers. arXiv preprint arXiv:2006.03555, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 196, + 505, + 232 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 503, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 252 + ], + "score": 1.0, + "content": "Alexis Conneau, Douwe Kiela, Holger Schwenk, Loic Barrault, and Antoine Bordes. Super-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 249, + 504, + 262 + ], + "spans": [ + { + "bbox": [ + 117, + 249, + 504, + 262 + ], + "score": 1.0, + "content": "vised learning of universal sentence representations from natural language inference data. arXiv", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 262, + 254, + 272 + ], + "spans": [ + { + "bbox": [ + 115, + 262, + 254, + 272 + ], + "score": 1.0, + "content": "preprint arXiv:1705.02364, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 237, + 505, + 272 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 280, + 504, + 304 + ], + "lines": [ + { + "bbox": [ + 105, + 279, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 295 + ], + "score": 1.0, + "content": "Thomas M Cover. Geometrical and statistical properties of systems of linear inequalities with ap-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 291, + 504, + 304 + ], + "spans": [ + { + "bbox": [ + 116, + 291, + 504, + 304 + ], + "score": 1.0, + "content": "plications in pattern recognition. IEEE transactions on electronic computers, (3):326–334, 1965.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 279, + 505, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 311, + 505, + 346 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 326 + ], + "score": 1.0, + "content": "Wojciech Marian Czarnecki, Grzegorz Swirszcz, Max Jaderberg, Simon Osindero, Oriol Vinyals, ´", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 323, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 115, + 323, + 505, + 337 + ], + "score": 1.0, + "content": "and Koray Kavukcuoglu. Understanding synthetic gradients and decoupled neural interfaces.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 335, + 279, + 347 + ], + "spans": [ + { + "bbox": [ + 116, + 335, + 279, + 347 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1703.00522, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 312, + 505, + 347 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "Transformer-XL: Attentive language models beyond a fixed-length context. In Proceedings of the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 376, + 505, + 389 + ], + "score": 1.0, + "content": "57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, July 2019.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 388, + 290, + 400 + ], + "spans": [ + { + "bbox": [ + 116, + 388, + 290, + 400 + ], + "score": 1.0, + "content": "Association for Computational Linguistics.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 354, + 506, + 400 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 407, + 504, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 407, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 421 + ], + "score": 1.0, + "content": "Amit Daniely, Roy Frostig, and Yoram Singer. Toward deeper understanding of neural networks:", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 116, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "The power of initialization and a dual view on expressivity. In Advances In Neural Information", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 429, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 116, + 429, + 289, + 443 + ], + "score": 1.0, + "content": "Processing Systems, pp. 2253–2261, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 407, + 505, + 443 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 449, + 504, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 464 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 460, + 500, + 473 + ], + "spans": [ + { + "bbox": [ + 116, + 460, + 500, + 473 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 448, + 506, + 473 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 504, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "Simon Du, Jason Lee, Haochuan Li, Liwei Wang, and Xiyu Zhai. Gradient descent finds global", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 115, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "minima of deep neural networks. In International Conference on Machine Learning, pp. 1675–", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 501, + 167, + 514 + ], + "spans": [ + { + "bbox": [ + 116, + 501, + 167, + 514 + ], + "score": 1.0, + "content": "1685, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 480, + 506, + 514 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 522, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "Sergey Edunov, Myle Ott, Michael Auli, David Grangier, and Marc’Aurelio Ranzato. Classical", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "structured prediction losses for sequence to sequence learning. In Proceedings of the 2018 Con-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 543, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 114, + 543, + 505, + 557 + ], + "score": 1.0, + "content": "ference of the North American Chapter of the Association for Computational Linguistics: Human", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 115, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "Language Technologies, Volume 1 (Long Papers), New Orleans, Louisiana, June 2018. Associa-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 567, + 259, + 578 + ], + "spans": [ + { + "bbox": [ + 116, + 567, + 259, + 578 + ], + "score": 1.0, + "content": "tion for Computational Linguistics.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 522, + 505, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 585, + 504, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive transformer. arXiv", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 597, + 254, + 608 + ], + "spans": [ + { + "bbox": [ + 115, + 597, + 254, + 608 + ], + "score": 1.0, + "content": "preprint arXiv:1910.10073, 2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 586, + 505, + 608 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 616, + 504, + 640 + ], + "lines": [ + { + "bbox": [ + 104, + 615, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "Joseph Enguehard, Dan Busbridge, Vitalii Zhelezniak, and Nils Hammerla. Neural language priors.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 628, + 278, + 639 + ], + "spans": [ + { + "bbox": [ + 116, + 628, + 278, + 639 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1910.03492, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 104, + 615, + 505, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 647, + 504, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 505, + 660 + ], + "score": 1.0, + "content": "Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 658, + 358, + 671 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 358, + 671 + ], + "score": 1.0, + "content": "structured dropout. arXiv preprint arXiv:1909.11556, 2019.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 647, + 505, + 671 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 690, + 321, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 321, + 702 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1803.03635, 2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 678, + 505, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Jonathan Frankle, David J Schwab, and Ari S Morcos. Training batchnorm and only batchnorm: On", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 721, + 476, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 476, + 732 + ], + "score": 1.0, + "content": "the expressive power of random features in cnns. arXiv preprint arXiv:2003.00152, 2020.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 504, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 504, + 96 + ], + "score": 1.0, + "content": "Claudio Gallicchio and Alessio Micheli. Echo state property of deep reservoir computing networks.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 299, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 299, + 106 + ], + "score": 1.0, + "content": "Cognitive Computation, 9(3):337–350, 2017.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "Claudio Gallicchio and Simone Scardapane. Deep randomized neural networks. In Recent Trends", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 322, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 322, + 135 + ], + "score": 1.0, + "content": "in Learning From Data, pp. 43–68. Springer, 2020.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 105, + 140, + 504, + 164 + ], + "lines": [ + { + "bbox": [ + 105, + 139, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 505, + 155 + ], + "score": 1.0, + "content": "A. Gamba, L. Gamberini, G. Palmieri, and R. Sanna. Further experiments with papa. Il Nuovo", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 152, + 297, + 164 + ], + "spans": [ + { + "bbox": [ + 116, + 152, + 297, + 164 + ], + "score": 1.0, + "content": "Cimento (1955-1965), 20(2):112–115, 1961.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 105, + 170, + 505, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 168, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 506, + 184 + ], + "score": 1.0, + "content": "Ankush Garg, Yuan Cao, and Qi Ge. Echo state neural machine translation. arXiv preprint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 181, + 219, + 193 + ], + "spans": [ + { + "bbox": [ + 115, + 181, + 219, + 193 + ], + "score": 1.0, + "content": "arXiv:2002.11847, 2020.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 199, + 504, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "Raja Giryes, Guillermo Sapiro, and Alex M Bronstein. Deep neural networks with random gaussian", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 210, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 115, + 210, + 505, + 224 + ], + "score": 1.0, + "content": "weights: A universal classification strategy? IEEE Transactions on Signal Processing, 64(13):", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 222, + 191, + 233 + ], + "spans": [ + { + "bbox": [ + 116, + 222, + 191, + 233 + ], + "score": 1.0, + "content": "3444–3457, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 108, + 239, + 504, + 274 + ], + "lines": [ + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 251, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 115, + 251, + 505, + 264 + ], + "score": 1.0, + "content": "networks. In Proceedings of the thirteenth international conference on artificial intelligence and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 262, + 236, + 274 + ], + "spans": [ + { + "bbox": [ + 115, + 262, + 236, + 274 + ], + "score": 1.0, + "content": "statistics, pp. 249–256, 2010.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 105, + 280, + 503, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 504, + 293 + ], + "score": 1.0, + "content": "Caglar Gulcehre, Marcin Moczulski, Misha Denil, and Yoshua Bengio. Noisy activation functions.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 291, + 405, + 304 + ], + "spans": [ + { + "bbox": [ + 115, + 291, + 405, + 304 + ], + "score": 1.0, + "content": "In International conference on machine learning, pp. 3059–3068, 2016.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 105, + 309, + 504, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 323 + ], + "score": 1.0, + "content": "Fatemeh Hadaeghi, Xu He, and Herbert Jaeger. Unconventional Information Processing Systems,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 320, + 293, + 332 + ], + "spans": [ + { + "bbox": [ + 116, + 320, + 293, + 332 + ], + "score": 1.0, + "content": "Novel Hardware: A Tour D’Horizon. 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 338, + 504, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 339, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 506, + 351 + ], + "score": 1.0, + "content": "Konstantin Hicke, Miguel Escalona-Moran, Daniel Brunner, Miguel Soriano, Ingo Fischer, and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 115, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "Claudio Mirasso. Information processing using transient dynamics of semiconductor lasers sub-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 361, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 114, + 361, + 506, + 375 + ], + "score": 1.0, + "content": "ject to delayed feedback. Selected Topics in Quantum Electronics, IEEE Journal of, 19:1501610–", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 372, + 342, + 384 + ], + "spans": [ + { + "bbox": [ + 116, + 372, + 342, + 384 + ], + "score": 1.0, + "content": "1501610, 07 2013. doi: 10.1109/JSTQE.2013.2241738.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 105, + 390, + 505, + 414 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 403 + ], + "score": 1.0, + "content": "Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. Extreme learning machine: theory and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 402, + 340, + 414 + ], + "spans": [ + { + "bbox": [ + 116, + 402, + 340, + 414 + ], + "score": 1.0, + "content": "applications. Neurocomputing, 70(1-3):489–501, 2006.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 419, + 504, + 454 + ], + "lines": [ + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "Max Jaderberg, Wojciech Marian Czarnecki, Simon Osindero, Oriol Vinyals, Alex Graves, David", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 117, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "Silver, and Koray Kavukcuoglu. Decoupled neural interfaces using synthetic gradients. In Inter-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 441, + 414, + 455 + ], + "spans": [ + { + "bbox": [ + 115, + 441, + 414, + 455 + ], + "score": 1.0, + "content": "national Conference on Machine Learning, pp. 1627–1635. PMLR, 2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 105, + 460, + 504, + 483 + ], + "lines": [ + { + "bbox": [ + 106, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "Herbert Jaeger. Adaptive nonlinear system identification with echo state networks. In Advances in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 471, + 300, + 483 + ], + "spans": [ + { + "bbox": [ + 115, + 471, + 300, + 483 + ], + "score": 1.0, + "content": "neural information processing systems, 2003.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 505, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "Ganesh Jawahar, Benoˆıt Sagot, and Djame Seddah. What does BERT learn about the structure of ´", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 116, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "language? In Proceedings of the 57th Annual Meeting of the Association for Computational", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 511, + 190, + 524 + ], + "spans": [ + { + "bbox": [ + 116, + 511, + 190, + 524 + ], + "score": 1.0, + "content": "Linguistics, 2019.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 104, + 529, + 504, + 553 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "Kam Jim, Bill G Horne, and C Lee Giles. Effects of noise on convergence and generalization in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 541, + 497, + 554 + ], + "spans": [ + { + "bbox": [ + 115, + 541, + 497, + 554 + ], + "score": 1.0, + "content": "recurrent networks. In Advances in neural information processing systems, pp. 649–656, 1995.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 105, + 558, + 503, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 504, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 504, + 570 + ], + "score": 1.0, + "content": "Kam-Chuen Jim, C Lee Giles, and Bill G Horne. An analysis of noise in recurrent neural networks:", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 570, + 502, + 582 + ], + "spans": [ + { + "bbox": [ + 116, + 570, + 502, + 582 + ], + "score": 1.0, + "content": "convergence and generalization. IEEE Transactions on neural networks, 7(6):1424–1438, 1996.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 105, + 588, + 503, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 505, + 602 + ], + "score": 1.0, + "content": "William B Johnson and Joram Lindenstrauss. Extensions of lipschitz mappings into a hilbert space.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 117, + 600, + 320, + 611 + ], + "spans": [ + { + "bbox": [ + 117, + 600, + 320, + 611 + ], + "score": 1.0, + "content": "Contemporary mathematics, 26(189-206):1, 1984.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 116, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 640, + 313, + 651 + ], + "spans": [ + { + "bbox": [ + 116, + 640, + 313, + 651 + ], + "score": 1.0, + "content": "models. arXiv preprint arXiv:2001.08361, 2020.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 658, + 504, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 658, + 504, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 504, + 671 + ], + "score": 1.0, + "content": "Jungo Kasai, Nikolaos Pappas, Hao Peng, James Cross, and Noah A Smith. Deep encoder, shal-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 115, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "low decoder: Reevaluating the speed-quality tradeoff in machine translation. arXiv preprint", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 680, + 218, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 218, + 692 + ], + "score": 1.0, + "content": "arXiv:2006.10369, 2020.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Franc¸ois Fleuret. Transformers are", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "rnns: Fast autoregressive transformers with linear attention. arXiv preprint arXiv:2006.16236,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 27, + 307, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 504, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 504, + 96 + ], + "score": 1.0, + "content": "Claudio Gallicchio and Alessio Micheli. Echo state property of deep reservoir computing networks.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 299, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 299, + 106 + ], + "score": 1.0, + "content": "Cognitive Computation, 9(3):337–350, 2017.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 504, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "Claudio Gallicchio and Simone Scardapane. Deep randomized neural networks. In Recent Trends", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 322, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 322, + 135 + ], + "score": 1.0, + "content": "in Learning From Data, pp. 43–68. Springer, 2020.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 111, + 505, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 140, + 504, + 164 + ], + "lines": [ + { + "bbox": [ + 105, + 139, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 505, + 155 + ], + "score": 1.0, + "content": "A. Gamba, L. Gamberini, G. Palmieri, and R. Sanna. Further experiments with papa. Il Nuovo", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 152, + 297, + 164 + ], + "spans": [ + { + "bbox": [ + 116, + 152, + 297, + 164 + ], + "score": 1.0, + "content": "Cimento (1955-1965), 20(2):112–115, 1961.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 139, + 505, + 164 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 170, + 505, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 168, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 506, + 184 + ], + "score": 1.0, + "content": "Ankush Garg, Yuan Cao, and Qi Ge. Echo state neural machine translation. arXiv preprint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 181, + 219, + 193 + ], + "spans": [ + { + "bbox": [ + 115, + 181, + 219, + 193 + ], + "score": 1.0, + "content": "arXiv:2002.11847, 2020.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 168, + 506, + 193 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 199, + 504, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "Raja Giryes, Guillermo Sapiro, and Alex M Bronstein. Deep neural networks with random gaussian", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 210, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 115, + 210, + 505, + 224 + ], + "score": 1.0, + "content": "weights: A universal classification strategy? IEEE Transactions on Signal Processing, 64(13):", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 222, + 191, + 233 + ], + "spans": [ + { + "bbox": [ + 116, + 222, + 191, + 233 + ], + "score": 1.0, + "content": "3444–3457, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 199, + 505, + 233 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 239, + 504, + 274 + ], + "lines": [ + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 251, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 115, + 251, + 505, + 264 + ], + "score": 1.0, + "content": "networks. In Proceedings of the thirteenth international conference on artificial intelligence and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 262, + 236, + 274 + ], + "spans": [ + { + "bbox": [ + 115, + 262, + 236, + 274 + ], + "score": 1.0, + "content": "statistics, pp. 249–256, 2010.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 239, + 505, + 274 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 280, + 503, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 504, + 293 + ], + "score": 1.0, + "content": "Caglar Gulcehre, Marcin Moczulski, Misha Denil, and Yoshua Bengio. Noisy activation functions.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 291, + 405, + 304 + ], + "spans": [ + { + "bbox": [ + 115, + 291, + 405, + 304 + ], + "score": 1.0, + "content": "In International conference on machine learning, pp. 3059–3068, 2016.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 280, + 504, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 309, + 504, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 323 + ], + "score": 1.0, + "content": "Fatemeh Hadaeghi, Xu He, and Herbert Jaeger. Unconventional Information Processing Systems,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 320, + 293, + 332 + ], + "spans": [ + { + "bbox": [ + 116, + 320, + 293, + 332 + ], + "score": 1.0, + "content": "Novel Hardware: A Tour D’Horizon. 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 308, + 505, + 332 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 338, + 504, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 339, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 506, + 351 + ], + "score": 1.0, + "content": "Konstantin Hicke, Miguel Escalona-Moran, Daniel Brunner, Miguel Soriano, Ingo Fischer, and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 115, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "Claudio Mirasso. Information processing using transient dynamics of semiconductor lasers sub-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 361, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 114, + 361, + 506, + 375 + ], + "score": 1.0, + "content": "ject to delayed feedback. Selected Topics in Quantum Electronics, IEEE Journal of, 19:1501610–", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 372, + 342, + 384 + ], + "spans": [ + { + "bbox": [ + 116, + 372, + 342, + 384 + ], + "score": 1.0, + "content": "1501610, 07 2013. doi: 10.1109/JSTQE.2013.2241738.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 339, + 506, + 384 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 390, + 505, + 414 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 403 + ], + "score": 1.0, + "content": "Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. Extreme learning machine: theory and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 402, + 340, + 414 + ], + "spans": [ + { + "bbox": [ + 116, + 402, + 340, + 414 + ], + "score": 1.0, + "content": "applications. Neurocomputing, 70(1-3):489–501, 2006.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 389, + 505, + 414 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 419, + 504, + 454 + ], + "lines": [ + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "Max Jaderberg, Wojciech Marian Czarnecki, Simon Osindero, Oriol Vinyals, Alex Graves, David", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 117, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "Silver, and Koray Kavukcuoglu. Decoupled neural interfaces using synthetic gradients. In Inter-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 441, + 414, + 455 + ], + "spans": [ + { + "bbox": [ + 115, + 441, + 414, + 455 + ], + "score": 1.0, + "content": "national Conference on Machine Learning, pp. 1627–1635. PMLR, 2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 420, + 505, + 455 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 460, + 504, + 483 + ], + "lines": [ + { + "bbox": [ + 106, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "Herbert Jaeger. Adaptive nonlinear system identification with echo state networks. In Advances in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 471, + 300, + 483 + ], + "spans": [ + { + "bbox": [ + 115, + 471, + 300, + 483 + ], + "score": 1.0, + "content": "neural information processing systems, 2003.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 460, + 505, + 483 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 505, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "Ganesh Jawahar, Benoˆıt Sagot, and Djame Seddah. What does BERT learn about the structure of ´", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 116, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "language? In Proceedings of the 57th Annual Meeting of the Association for Computational", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 511, + 190, + 524 + ], + "spans": [ + { + "bbox": [ + 116, + 511, + 190, + 524 + ], + "score": 1.0, + "content": "Linguistics, 2019.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 489, + 505, + 524 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 529, + 504, + 553 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "Kam Jim, Bill G Horne, and C Lee Giles. Effects of noise on convergence and generalization in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 541, + 497, + 554 + ], + "spans": [ + { + "bbox": [ + 115, + 541, + 497, + 554 + ], + "score": 1.0, + "content": "recurrent networks. In Advances in neural information processing systems, pp. 649–656, 1995.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 529, + 505, + 554 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 558, + 503, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 504, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 504, + 570 + ], + "score": 1.0, + "content": "Kam-Chuen Jim, C Lee Giles, and Bill G Horne. An analysis of noise in recurrent neural networks:", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 570, + 502, + 582 + ], + "spans": [ + { + "bbox": [ + 116, + 570, + 502, + 582 + ], + "score": 1.0, + "content": "convergence and generalization. IEEE Transactions on neural networks, 7(6):1424–1438, 1996.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 559, + 504, + 582 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 588, + 503, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 505, + 602 + ], + "score": 1.0, + "content": "William B Johnson and Joram Lindenstrauss. Extensions of lipschitz mappings into a hilbert space.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 117, + 600, + 320, + 611 + ], + "spans": [ + { + "bbox": [ + 117, + 600, + 320, + 611 + ], + "score": 1.0, + "content": "Contemporary mathematics, 26(189-206):1, 1984.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 587, + 505, + 611 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 116, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 640, + 313, + 651 + ], + "spans": [ + { + "bbox": [ + 116, + 640, + 313, + 651 + ], + "score": 1.0, + "content": "models. arXiv preprint arXiv:2001.08361, 2020.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 618, + 505, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 658, + 504, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 658, + 504, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 504, + 671 + ], + "score": 1.0, + "content": "Jungo Kasai, Nikolaos Pappas, Hao Peng, James Cross, and Noah A Smith. Deep encoder, shal-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 115, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "low decoder: Reevaluating the speed-quality tradeoff in machine translation. arXiv preprint", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 680, + 218, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 218, + 692 + ], + "score": 1.0, + "content": "arXiv:2006.10369, 2020.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 658, + 505, + 692 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Franc¸ois Fleuret. Transformers are", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "rnns: Fast autoregressive transformers with linear attention. arXiv preprint arXiv:2006.16236,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 698, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 158, + 95 + ], + "score": 1.0, + "content": "Yoon Kim.", + "type": "text" + }, + { + "bbox": [ + 169, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "Convolutional neural networks for sentence classification. arXiv preprint", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 214, + 104 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 214, + 104 + ], + "score": 1.0, + "content": "arXiv:1408.5882, 2014.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 112, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "score": 1.0, + "content": "Nikita Kitaev, Łukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. arXiv", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 124, + 254, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 124, + 254, + 135 + ], + "score": 1.0, + "content": "preprint arXiv:2001.04451, 2020.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 504, + 166 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 155, + 417, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 417, + 167 + ], + "score": 1.0, + "content": "document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 108, + 174, + 504, + 207 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 506, + 187 + ], + "score": 1.0, + "content": "Yuanzhi Li and Yingyu Liang. Learning overparameterized neural networks via stochastic gradient", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 185, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 506, + 198 + ], + "score": 1.0, + "content": "descent on structured data. In Advances in Neural Information Processing Systems, pp. 8157–", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 196, + 166, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 196, + 166, + 208 + ], + "score": 1.0, + "content": "8166, 2018.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 215, + 504, + 249 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joseph E Gon-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 227, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 116, + 227, + 506, + 239 + ], + "score": 1.0, + "content": "zalez. Train large, then compress: Rethinking model size for efficient training and inference of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 238, + 335, + 250 + ], + "spans": [ + { + "bbox": [ + 116, + 238, + 335, + 250 + ], + "score": 1.0, + "content": "transformers. arXiv preprint arXiv:2002.11794, 2020.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 257, + 504, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 267, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 115, + 267, + 505, + 282 + ], + "score": 1.0, + "content": "Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 280, + 321, + 291 + ], + "spans": [ + { + "bbox": [ + 116, + 280, + 321, + 291 + ], + "score": 1.0, + "content": "approach. arXiv preprint arXiv:1907.11692, 2019.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 352, + 310 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 352, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 352, + 311 + ], + "score": 1.0, + "content": "MultiMedia LLC. Large text compression benchmark. 2009.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 105, + 318, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "Mantas Lukosevi ˇ cius and Herbert Jaeger. Reservoir computing approaches to recurrent neural net- ˇ", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 330, + 332, + 342 + ], + "spans": [ + { + "bbox": [ + 115, + 330, + 332, + 342 + ], + "score": 1.0, + "content": "work training. Computer Science Review, 3(3), 2009.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 348, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 362 + ], + "score": 1.0, + "content": "Wolfgang Maass, Thomas Natschlager, and Henry Markram. Real-time computing without stable ¨", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 115, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "states: A new framework for neural computation based on perturbations. Neural computation, 14", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 371, + 210, + 382 + ], + "spans": [ + { + "bbox": [ + 115, + 371, + 210, + 382 + ], + "score": 1.0, + "content": "(11):2531–2560, 2002.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 105, + 390, + 504, + 413 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 505, + 405 + ], + "score": 1.0, + "content": "Marvin Minsky and Seymour A Papert. Perceptrons: An introduction to computational geometry.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 401, + 188, + 414 + ], + "spans": [ + { + "bbox": [ + 115, + 401, + 188, + 414 + ], + "score": 1.0, + "content": "MIT press, 2017.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 420, + 504, + 455 + ], + "lines": [ + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "Emre O Neftci, Charles Augustine, Somnath Paul, and Georgios Detorakis. Event-driven random", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 432, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 116, + 432, + 506, + 444 + ], + "score": 1.0, + "content": "back-propagation: Enabling neuromorphic deep learning machines. Frontiers in neuroscience,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 443, + 174, + 455 + ], + "spans": [ + { + "bbox": [ + 116, + 443, + 174, + 455 + ], + "score": 1.0, + "content": "11:324, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "Hyeonwoo Noh, Tackgeun You, Jonghwan Mun, and Bohyung Han. Regularizing deep neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 116, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "networks by noise: Its interpretation and optimization. In Advances in Neural Information Pro-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 485, + 275, + 497 + ], + "spans": [ + { + "bbox": [ + 116, + 485, + 275, + 497 + ], + "score": 1.0, + "content": "cessing Systems, pp. 5109–5118, 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 504, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "Deniz Oktay, Nick McGreivy, Joshua Aduol, Alex Beatson, and Ryan P Adams. Randomized", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 515, + 384, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 384, + 528 + ], + "score": 1.0, + "content": "automatic differentiation. arXiv preprint arXiv:2007.10412, 2020.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 504, + 558 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "Myle Ott, Sergey Edunov, David Grangier, and Michael Auli. Scaling neural machine translation.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 546, + 279, + 557 + ], + "spans": [ + { + "bbox": [ + 115, + 546, + 279, + 557 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1806.00187, 2018.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 419, + 578 + ], + "score": 1.0, + "content": "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan", + "type": "text" + }, + { + "bbox": [ + 419, + 565, + 433, + 577 + ], + "score": 0.25, + "content": "\\mathrm { N g }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 564, + 505, + 578 + ], + "score": 1.0, + "content": ", David Grangier,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 575, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 575, + 506, + 590 + ], + "score": 1.0, + "content": "and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. In Proceedings of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 587, + 290, + 599 + ], + "spans": [ + { + "bbox": [ + 116, + 587, + 290, + 599 + ], + "score": 1.0, + "content": "NAACL-HLT 2019: Demonstrations, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 105, + 606, + 504, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "Yoh-Han Pao, Gwang-Hoon Park, and Dejan J Sobajic. Learning and generalization characteristics", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 117, + 618, + 437, + 630 + ], + "spans": [ + { + "bbox": [ + 117, + 618, + 437, + 630 + ], + "score": 1.0, + "content": "of the random vector functional-link net. Neurocomputing, 6(2):163–180, 1994.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 105, + 637, + 503, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "score": 1.0, + "content": "Jonathan Pilault, Jaehong Park, and Christopher Pal. On the impressive performance of randomly", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 648, + 453, + 660 + ], + "spans": [ + { + "bbox": [ + 115, + 648, + 453, + 660 + ], + "score": 1.0, + "content": "weighted encoders in summarization tasks. arXiv preprint arXiv:2002.09084, 2020.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "Jordi Pons and Xavier Serra. Randomly weighted cnns for (music) audio classification. In ICASSP", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "2019-2019 IEEE international conference on acoustics, speech and signal processing (ICASSP),", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 690, + 223, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 223, + 702 + ], + "score": 1.0, + "content": "pp. 336–340. IEEE, 2019.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 504, + 722 + ], + "score": 1.0, + "content": "Ofir Press, Noah A Smith, and Omer Levy. Improving transformer models by reordering their", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 322, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 322, + 732 + ], + "score": 1.0, + "content": "sublayers. arXiv preprint arXiv:1911.03864, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 11, + "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": "12", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 27, + 307, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 158, + 95 + ], + "score": 1.0, + "content": "Yoon Kim.", + "type": "text" + }, + { + "bbox": [ + 169, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "Convolutional neural networks for sentence classification. arXiv preprint", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 214, + 104 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 214, + 104 + ], + "score": 1.0, + "content": "arXiv:1408.5882, 2014.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 80, + 506, + 104 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 112, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "score": 1.0, + "content": "Nikita Kitaev, Łukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. arXiv", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 124, + 254, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 124, + 254, + 135 + ], + "score": 1.0, + "content": "preprint arXiv:2001.04451, 2020.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 112, + 505, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 504, + 166 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 155, + 417, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 417, + 167 + ], + "score": 1.0, + "content": "document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 142, + 505, + 167 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 174, + 504, + 207 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 506, + 187 + ], + "score": 1.0, + "content": "Yuanzhi Li and Yingyu Liang. Learning overparameterized neural networks via stochastic gradient", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 185, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 506, + 198 + ], + "score": 1.0, + "content": "descent on structured data. In Advances in Neural Information Processing Systems, pp. 8157–", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 196, + 166, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 196, + 166, + 208 + ], + "score": 1.0, + "content": "8166, 2018.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 173, + 506, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 215, + 504, + 249 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joseph E Gon-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 227, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 116, + 227, + 506, + 239 + ], + "score": 1.0, + "content": "zalez. Train large, then compress: Rethinking model size for efficient training and inference of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 238, + 335, + 250 + ], + "spans": [ + { + "bbox": [ + 116, + 238, + 335, + 250 + ], + "score": 1.0, + "content": "transformers. arXiv preprint arXiv:2002.11794, 2020.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 215, + 506, + 250 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 257, + 504, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 267, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 115, + 267, + 505, + 282 + ], + "score": 1.0, + "content": "Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 280, + 321, + 291 + ], + "spans": [ + { + "bbox": [ + 116, + 280, + 321, + 291 + ], + "score": 1.0, + "content": "approach. arXiv preprint arXiv:1907.11692, 2019.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 257, + 505, + 291 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 352, + 310 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 352, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 352, + 311 + ], + "score": 1.0, + "content": "MultiMedia LLC. Large text compression benchmark. 2009.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 106, + 298, + 352, + 311 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 318, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "Mantas Lukosevi ˇ cius and Herbert Jaeger. Reservoir computing approaches to recurrent neural net- ˇ", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 330, + 332, + 342 + ], + "spans": [ + { + "bbox": [ + 115, + 330, + 332, + 342 + ], + "score": 1.0, + "content": "work training. Computer Science Review, 3(3), 2009.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 318, + 505, + 342 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 348, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 362 + ], + "score": 1.0, + "content": "Wolfgang Maass, Thomas Natschlager, and Henry Markram. Real-time computing without stable ¨", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 115, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "states: A new framework for neural computation based on perturbations. Neural computation, 14", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 371, + 210, + 382 + ], + "spans": [ + { + "bbox": [ + 115, + 371, + 210, + 382 + ], + "score": 1.0, + "content": "(11):2531–2560, 2002.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 348, + 506, + 382 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 390, + 504, + 413 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 505, + 405 + ], + "score": 1.0, + "content": "Marvin Minsky and Seymour A Papert. Perceptrons: An introduction to computational geometry.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 401, + 188, + 414 + ], + "spans": [ + { + "bbox": [ + 115, + 401, + 188, + 414 + ], + "score": 1.0, + "content": "MIT press, 2017.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 388, + 505, + 414 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 420, + 504, + 455 + ], + "lines": [ + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "Emre O Neftci, Charles Augustine, Somnath Paul, and Georgios Detorakis. Event-driven random", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 432, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 116, + 432, + 506, + 444 + ], + "score": 1.0, + "content": "back-propagation: Enabling neuromorphic deep learning machines. Frontiers in neuroscience,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 443, + 174, + 455 + ], + "spans": [ + { + "bbox": [ + 116, + 443, + 174, + 455 + ], + "score": 1.0, + "content": "11:324, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 421, + 506, + 455 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "Hyeonwoo Noh, Tackgeun You, Jonghwan Mun, and Bohyung Han. Regularizing deep neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 116, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "networks by noise: Its interpretation and optimization. In Advances in Neural Information Pro-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 485, + 275, + 497 + ], + "spans": [ + { + "bbox": [ + 116, + 485, + 275, + 497 + ], + "score": 1.0, + "content": "cessing Systems, pp. 5109–5118, 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 462, + 505, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 504, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "Deniz Oktay, Nick McGreivy, Joshua Aduol, Alex Beatson, and Ryan P Adams. Randomized", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 515, + 384, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 384, + 528 + ], + "score": 1.0, + "content": "automatic differentiation. arXiv preprint arXiv:2007.10412, 2020.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 504, + 505, + 528 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 504, + 558 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "Myle Ott, Sergey Edunov, David Grangier, and Michael Auli. Scaling neural machine translation.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 546, + 279, + 557 + ], + "spans": [ + { + "bbox": [ + 115, + 546, + 279, + 557 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1806.00187, 2018.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 534, + 505, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 419, + 578 + ], + "score": 1.0, + "content": "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan", + "type": "text" + }, + { + "bbox": [ + 419, + 565, + 433, + 577 + ], + "score": 0.25, + "content": "\\mathrm { N g }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 564, + 505, + 578 + ], + "score": 1.0, + "content": ", David Grangier,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 575, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 575, + 506, + 590 + ], + "score": 1.0, + "content": "and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. In Proceedings of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 587, + 290, + 599 + ], + "spans": [ + { + "bbox": [ + 116, + 587, + 290, + 599 + ], + "score": 1.0, + "content": "NAACL-HLT 2019: Demonstrations, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 564, + 506, + 599 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 606, + 504, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "Yoh-Han Pao, Gwang-Hoon Park, and Dejan J Sobajic. Learning and generalization characteristics", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 117, + 618, + 437, + 630 + ], + "spans": [ + { + "bbox": [ + 117, + 618, + 437, + 630 + ], + "score": 1.0, + "content": "of the random vector functional-link net. Neurocomputing, 6(2):163–180, 1994.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 606, + 505, + 630 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 637, + 503, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "score": 1.0, + "content": "Jonathan Pilault, Jaehong Park, and Christopher Pal. On the impressive performance of randomly", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 648, + 453, + 660 + ], + "spans": [ + { + "bbox": [ + 115, + 648, + 453, + 660 + ], + "score": 1.0, + "content": "weighted encoders in summarization tasks. arXiv preprint arXiv:2002.09084, 2020.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 636, + 505, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "Jordi Pons and Xavier Serra. Randomly weighted cnns for (music) audio classification. In ICASSP", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "2019-2019 IEEE international conference on acoustics, speech and signal processing (ICASSP),", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 690, + 223, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 223, + 702 + ], + "score": 1.0, + "content": "pp. 336–340. IEEE, 2019.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 668, + 505, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 504, + 722 + ], + "score": 1.0, + "content": "Ofir Press, Noah A Smith, and Omer Levy. Improving transformer models by reordering their", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 322, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 322, + 732 + ], + "score": 1.0, + "content": "sublayers. arXiv preprint arXiv:1911.03864, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 708, + 504, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 506, + 97 + ], + "score": 1.0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 320, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 320, + 106 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 506, + 124 + ], + "score": 1.0, + "content": "Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. In Advances in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 365, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 365, + 135 + ], + "score": 1.0, + "content": "neural information processing systems, pp. 1177–1184, 2008.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 140, + 504, + 174 + ], + "lines": [ + { + "bbox": [ + 106, + 141, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 505, + 154 + ], + "score": 1.0, + "content": "Ali Rahimi and Benjamin Recht. Weighted sums of random kitchen sinks: Replacing minimization", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 152, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 116, + 152, + 505, + 164 + ], + "score": 1.0, + "content": "with randomization in learning. In Advances in neural information processing systems, pp. 1313–", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 163, + 166, + 174 + ], + "spans": [ + { + "bbox": [ + 117, + 163, + 166, + 174 + ], + "score": 1.0, + "content": "1320, 2009.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 181, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "Vivek Ramanujan, Mitchell Wortsman, Aniruddha Kembhavi, Ali Farhadi, and Mohammad Raste-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 115, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "gari. What’s hidden in a randomly weighted neural network? In Proceedings of the IEEE/CVF", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 203, + 447, + 216 + ], + "spans": [ + { + "bbox": [ + 116, + 203, + 447, + 216 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, pp. 11893–11902, 2020.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 105, + 221, + 504, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 505, + 234 + ], + "score": 1.0, + "content": "Anna Rogers, Olga Kovaleva, and Anna Rumshisky. A primer in bertology: What we know about", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 233, + 345, + 245 + ], + "spans": [ + { + "bbox": [ + 115, + 233, + 345, + 245 + ], + "score": 1.0, + "content": "how bert works. arXiv preprint arXiv:2002.12327, 2020.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 250, + 504, + 285 + ], + "lines": [ + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "score": 1.0, + "content": "Amir Rosenfeld and John K Tsotsos. Intriguing properties of randomly weighted networks: Gener-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 262, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 116, + 262, + 505, + 274 + ], + "score": 1.0, + "content": "alizing while learning next to nothing. In 2019 16th Conference on Computer and Robot Vision", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 273, + 238, + 285 + ], + "spans": [ + { + "bbox": [ + 116, + 273, + 238, + 285 + ], + "score": 1.0, + "content": "(CRV), pp. 9–16. IEEE, 2019.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 291, + 504, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "Magnus Sahlgren. An introduction to random indexing. In Methods and applications of semantic in-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 114, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "dexing workshop at the 7th international conference on terminology and knowledge engineering,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 312, + 144, + 326 + ], + "spans": [ + { + "bbox": [ + 115, + 312, + 144, + 326 + ], + "score": 1.0, + "content": "2005.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 105, + 331, + 504, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 504, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 504, + 344 + ], + "score": 1.0, + "content": "Andrew M Saxe, James L McClelland, and Surya Ganguli. Exact solutions to the nonlinear dynam-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 343, + 460, + 355 + ], + "spans": [ + { + "bbox": [ + 115, + 343, + 460, + 355 + ], + "score": 1.0, + "content": "ics of learning in deep linear neural networks. arXiv preprint arXiv:1312.6120, 2013.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 105, + 360, + 503, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "score": 1.0, + "content": "Simone Scardapane and Dianhui Wang. Randomness in neural networks: an overview. Wiley Inter-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 372, + 443, + 384 + ], + "spans": [ + { + "bbox": [ + 115, + 372, + 443, + 384 + ], + "score": 1.0, + "content": "disciplinary Reviews: Data Mining and Knowledge Discovery, 7(2):e1200, 2017.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 504, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "Wouter F Schmidt, Martin A Kraaijveld, and Robert PW Duin. Feedforward neural networks with", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 116, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "random weights. In Proceedings of the 11th International Conference on Pattern Recognition,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 412, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 116, + 412, + 482, + 426 + ], + "score": 1.0, + "content": "1992. Vol. II. Conference B: Pattern Recognition Methodology and Systems, pp. 1–4, 1992.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 105, + 430, + 504, + 464 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 504, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 504, + 443 + ], + "score": 1.0, + "content": "Benjamin Schrauwen, Michiel D’Haene, David Verstraeten, and Jan Campenhout. Compact hard-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 441, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 115, + 441, + 506, + 454 + ], + "score": 1.0, + "content": "ware for real-time speech recognition using a liquid state machine. pp. 1097 – 1102, 09 2007.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 453, + 263, + 464 + ], + "spans": [ + { + "bbox": [ + 116, + 453, + 263, + 464 + ], + "score": 1.0, + "content": "doi: 10.1109/IJCNN.2007.4371111.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 105, + 470, + 503, + 494 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 484 + ], + "score": 1.0, + "content": "Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. Green ai. arXiv preprint", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 481, + 220, + 494 + ], + "spans": [ + { + "bbox": [ + 115, + 481, + 220, + 494 + ], + "score": 1.0, + "content": "arXiv:1907.10597, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 500, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 514 + ], + "score": 1.0, + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 116, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 116, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "treebank. In Proceedings of the 2013 conference on empirical methods in natural language pro-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 534, + 241, + 545 + ], + "spans": [ + { + "bbox": [ + 115, + 534, + 241, + 545 + ], + "score": 1.0, + "content": "cessing, pp. 1631–1642, 2013.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 104, + 551, + 504, + 575 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 566 + ], + "score": 1.0, + "content": "Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 563, + 343, + 575 + ], + "spans": [ + { + "bbox": [ + 115, + 563, + 343, + 575 + ], + "score": 1.0, + "content": "learning in nlp. arXiv preprint arXiv:1906.02243, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 595 + ], + "score": 1.0, + "content": "Gouhei Tanaka, Toshiyuki Yamane, Jean Benoit Heroux, Ryosho Nakane, Naoki Kanazawa, Seiji ´", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 116, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "Takeda, Hidetoshi Numata, Daiju Nakano, and Akira Hirose. Recent advances in physical reser-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 603, + 385, + 615 + ], + "spans": [ + { + "bbox": [ + 116, + 603, + 385, + 615 + ], + "score": 1.0, + "content": "voir computing: A review. Neural Networks, 115:100 – 123, 2019.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 503, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "Yi Tay, Dara Bahri, Donald Metzler, Da-Cheng Juan, Zhe Zhao, and Che Zheng. Synthesizer:", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 632, + 479, + 645 + ], + "spans": [ + { + "bbox": [ + 116, + 632, + 479, + 645 + ], + "score": 1.0, + "content": "Rethinking self-attention in transformer models. arXiv preprint arXiv:2005.00743, 2020a.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 504, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. Efficient transformers: A survey. arXiv", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 662, + 260, + 673 + ], + "spans": [ + { + "bbox": [ + 114, + 662, + 260, + 673 + ], + "score": 1.0, + "content": "preprint arXiv:2009.06732, 2020b.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 693 + ], + "score": 1.0, + "content": "Ian Tenney, Dipanjan Das, and Ellie Pavlick. Bert rediscovers the classical nlp pipeline. arXiv", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 691, + 254, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 254, + 703 + ], + "score": 1.0, + "content": "preprint arXiv:1905.05950, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 720, + 462, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 462, + 733 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition, pp. 9446–9454, 2018.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 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 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 506, + 97 + ], + "score": 1.0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 320, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 320, + 106 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 79, + 506, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 506, + 124 + ], + "score": 1.0, + "content": "Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. In Advances in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 365, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 365, + 135 + ], + "score": 1.0, + "content": "neural information processing systems, pp. 1177–1184, 2008.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 111, + 506, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 140, + 504, + 174 + ], + "lines": [ + { + "bbox": [ + 106, + 141, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 505, + 154 + ], + "score": 1.0, + "content": "Ali Rahimi and Benjamin Recht. Weighted sums of random kitchen sinks: Replacing minimization", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 152, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 116, + 152, + 505, + 164 + ], + "score": 1.0, + "content": "with randomization in learning. In Advances in neural information processing systems, pp. 1313–", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 163, + 166, + 174 + ], + "spans": [ + { + "bbox": [ + 117, + 163, + 166, + 174 + ], + "score": 1.0, + "content": "1320, 2009.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 141, + 505, + 174 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 181, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "Vivek Ramanujan, Mitchell Wortsman, Aniruddha Kembhavi, Ali Farhadi, and Mohammad Raste-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 115, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "gari. What’s hidden in a randomly weighted neural network? In Proceedings of the IEEE/CVF", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 203, + 447, + 216 + ], + "spans": [ + { + "bbox": [ + 116, + 203, + 447, + 216 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, pp. 11893–11902, 2020.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 106, + 181, + 506, + 216 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 221, + 504, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 505, + 234 + ], + "score": 1.0, + "content": "Anna Rogers, Olga Kovaleva, and Anna Rumshisky. A primer in bertology: What we know about", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 233, + 345, + 245 + ], + "spans": [ + { + "bbox": [ + 115, + 233, + 345, + 245 + ], + "score": 1.0, + "content": "how bert works. arXiv preprint arXiv:2002.12327, 2020.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 221, + 505, + 245 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 250, + 504, + 285 + ], + "lines": [ + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "score": 1.0, + "content": "Amir Rosenfeld and John K Tsotsos. Intriguing properties of randomly weighted networks: Gener-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 262, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 116, + 262, + 505, + 274 + ], + "score": 1.0, + "content": "alizing while learning next to nothing. In 2019 16th Conference on Computer and Robot Vision", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 273, + 238, + 285 + ], + "spans": [ + { + "bbox": [ + 116, + 273, + 238, + 285 + ], + "score": 1.0, + "content": "(CRV), pp. 9–16. IEEE, 2019.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 250, + 505, + 285 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 291, + 504, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "Magnus Sahlgren. An introduction to random indexing. In Methods and applications of semantic in-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 114, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "dexing workshop at the 7th international conference on terminology and knowledge engineering,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 312, + 144, + 326 + ], + "spans": [ + { + "bbox": [ + 115, + 312, + 144, + 326 + ], + "score": 1.0, + "content": "2005.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 291, + 505, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 331, + 504, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 504, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 504, + 344 + ], + "score": 1.0, + "content": "Andrew M Saxe, James L McClelland, and Surya Ganguli. Exact solutions to the nonlinear dynam-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 343, + 460, + 355 + ], + "spans": [ + { + "bbox": [ + 115, + 343, + 460, + 355 + ], + "score": 1.0, + "content": "ics of learning in deep linear neural networks. arXiv preprint arXiv:1312.6120, 2013.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 331, + 504, + 355 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 360, + 503, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "score": 1.0, + "content": "Simone Scardapane and Dianhui Wang. Randomness in neural networks: an overview. Wiley Inter-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 372, + 443, + 384 + ], + "spans": [ + { + "bbox": [ + 115, + 372, + 443, + 384 + ], + "score": 1.0, + "content": "disciplinary Reviews: Data Mining and Knowledge Discovery, 7(2):e1200, 2017.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 360, + 505, + 384 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 504, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "Wouter F Schmidt, Martin A Kraaijveld, and Robert PW Duin. Feedforward neural networks with", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 116, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "random weights. In Proceedings of the 11th International Conference on Pattern Recognition,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 412, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 116, + 412, + 482, + 426 + ], + "score": 1.0, + "content": "1992. Vol. II. Conference B: Pattern Recognition Methodology and Systems, pp. 1–4, 1992.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 390, + 505, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 430, + 504, + 464 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 504, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 504, + 443 + ], + "score": 1.0, + "content": "Benjamin Schrauwen, Michiel D’Haene, David Verstraeten, and Jan Campenhout. Compact hard-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 441, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 115, + 441, + 506, + 454 + ], + "score": 1.0, + "content": "ware for real-time speech recognition using a liquid state machine. pp. 1097 – 1102, 09 2007.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 453, + 263, + 464 + ], + "spans": [ + { + "bbox": [ + 116, + 453, + 263, + 464 + ], + "score": 1.0, + "content": "doi: 10.1109/IJCNN.2007.4371111.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 430, + 506, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 470, + 503, + 494 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 484 + ], + "score": 1.0, + "content": "Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. Green ai. arXiv preprint", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 481, + 220, + 494 + ], + "spans": [ + { + "bbox": [ + 115, + 481, + 220, + 494 + ], + "score": 1.0, + "content": "arXiv:1907.10597, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 469, + 505, + 494 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 500, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 514 + ], + "score": 1.0, + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 116, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 116, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "treebank. In Proceedings of the 2013 conference on empirical methods in natural language pro-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 534, + 241, + 545 + ], + "spans": [ + { + "bbox": [ + 115, + 534, + 241, + 545 + ], + "score": 1.0, + "content": "cessing, pp. 1631–1642, 2013.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 498, + 505, + 545 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 551, + 504, + 575 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 566 + ], + "score": 1.0, + "content": "Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 563, + 343, + 575 + ], + "spans": [ + { + "bbox": [ + 115, + 563, + 343, + 575 + ], + "score": 1.0, + "content": "learning in nlp. arXiv preprint arXiv:1906.02243, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 550, + 506, + 575 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 595 + ], + "score": 1.0, + "content": "Gouhei Tanaka, Toshiyuki Yamane, Jean Benoit Heroux, Ryosho Nakane, Naoki Kanazawa, Seiji ´", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 116, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "Takeda, Hidetoshi Numata, Daiju Nakano, and Akira Hirose. Recent advances in physical reser-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 603, + 385, + 615 + ], + "spans": [ + { + "bbox": [ + 116, + 603, + 385, + 615 + ], + "score": 1.0, + "content": "voir computing: A review. Neural Networks, 115:100 – 123, 2019.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 579, + 505, + 615 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 503, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "Yi Tay, Dara Bahri, Donald Metzler, Da-Cheng Juan, Zhe Zhao, and Che Zheng. Synthesizer:", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 632, + 479, + 645 + ], + "spans": [ + { + "bbox": [ + 116, + 632, + 479, + 645 + ], + "score": 1.0, + "content": "Rethinking self-attention in transformer models. arXiv preprint arXiv:2005.00743, 2020a.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 621, + 505, + 645 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 504, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. Efficient transformers: A survey. arXiv", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 662, + 260, + 673 + ], + "spans": [ + { + "bbox": [ + 114, + 662, + 260, + 673 + ], + "score": 1.0, + "content": "preprint arXiv:2009.06732, 2020b.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 650, + 505, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 693 + ], + "score": 1.0, + "content": "Ian Tenney, Dipanjan Das, and Ellie Pavlick. Bert rediscovers the classical nlp pipeline. arXiv", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 691, + 254, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 254, + 703 + ], + "score": 1.0, + "content": "preprint arXiv:1905.05950, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 678, + 505, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 720, + 462, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 462, + 733 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition, pp. 9446–9454, 2018.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 106, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 287, + 116 + ], + "score": 1.0, + "content": "processing systems, pp. 5998–6008, 2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "score": 1.0, + "content": "Oriol Vinyals, Igor Babuschkin, Wojciech M. Czarnecki, Michael Mathieu, Andrew Dudzik, Juny- ¨", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "oung Chung, David H. Choi, Richard Powell, Timo Ewalds, Petko Georgiev, Junhyuk Oh, Dan", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 505, + 158 + ], + "score": 1.0, + "content": "Horgan, Manuel Kroiss, Ivo Danihelka, Aja Huang, Laurent Sifre, Trevor Cai, John P. Agapiou,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "Max Jaderberg, Alexander S. Vezhnevets, Remi Leblond, Tobias Pohlen, Valentin Dalibard, David ´", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 167, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 115, + 167, + 506, + 181 + ], + "score": 1.0, + "content": "Budden, Yury Sulsky, James Molloy, Tom L. Paine, Caglar Gulcehre, Ziyu Wang, Tobias Pfaff,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 177, + 505, + 191 + ], + "score": 1.0, + "content": "Yuhuai Wu, Roman Ring, Dani Yogatama, Dario Wunsch, Katrina McKinney, Oliver Smith, Tom ¨", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 189, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 115, + 189, + 505, + 202 + ], + "score": 1.0, + "content": "Schaul, Timothy Lillicrap, Koray Kavukcuoglu, Demis Hassabis, Chris Apps, and David Silver.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 199, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 115, + 199, + 505, + 214 + ], + "score": 1.0, + "content": "Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature, 575(7782):", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 211, + 225, + 223 + ], + "spans": [ + { + "bbox": [ + 116, + 211, + 225, + 223 + ], + "score": 1.0, + "content": "350–354, November 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 105, + 230, + 504, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "score": 1.0, + "content": "Sinong Wang, Belinda Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 240, + 353, + 253 + ], + "spans": [ + { + "bbox": [ + 115, + 240, + 353, + 253 + ], + "score": 1.0, + "content": "linear complexity. arXiv preprint arXiv:2006.04768, 2020.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 259, + 504, + 282 + ], + "lines": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "John Wieting and Douwe Kiela. No training required: Exploring random encoders for sentence", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 270, + 336, + 283 + ], + "spans": [ + { + "bbox": [ + 115, + 270, + 336, + 283 + ], + "score": 1.0, + "content": "classification. arXiv preprint arXiv:1901.10444, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 503, + 312 + ], + "lines": [ + { + "bbox": [ + 106, + 288, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 303 + ], + "score": 1.0, + "content": "Adina Williams, Nikita Nangia, and Samuel R Bowman. A broad-coverage challenge corpus for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 301, + 452, + 313 + ], + "spans": [ + { + "bbox": [ + 115, + 301, + 452, + 313 + ], + "score": 1.0, + "content": "sentence understanding through inference. arXiv preprint arXiv:1704.05426, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 105, + 319, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 333 + ], + "score": 1.0, + "content": "Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 330, + 324, + 343 + ], + "spans": [ + { + "bbox": [ + 116, + 330, + 324, + 343 + ], + "score": 1.0, + "content": "learning. Machine learning, 8(3-4):229–256, 1992.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 105, + 349, + 504, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "Felix Wu, Angela Fan, Alexei Baevski, Yann N Dauphin, and Michael Auli. Pay less attention with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 360, + 438, + 372 + ], + "spans": [ + { + "bbox": [ + 116, + 360, + 438, + 372 + ], + "score": 1.0, + "content": "lightweight and dynamic convolutions. arXiv preprint arXiv:1901.10430, 2019.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 379, + 504, + 413 + ], + "lines": [ + { + "bbox": [ + 105, + 378, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 393 + ], + "score": 1.0, + "content": "Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 116, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 402, + 325, + 413 + ], + "spans": [ + { + "bbox": [ + 116, + 402, + 325, + 413 + ], + "score": 1.0, + "content": "sequences. arXiv preprint arXiv:2007.14062, 2020.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 105, + 420, + 504, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "score": 1.0, + "content": "Chiyuan Zhang, Samy Bengio, and Yoram Singer. Are all layers created equal? arXiv preprint", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 430, + 219, + 443 + ], + "spans": [ + { + "bbox": [ + 115, + 430, + 219, + 443 + ], + "score": 1.0, + "content": "arXiv:1902.01996, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "score": 1.0, + "content": "Kelly Zhang and Samuel Bowman. Language modeling teaches you more than translation does:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 115, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "Lessons learned through auxiliary syntactic task analysis. In Proceedings of the 2018 EMNLP", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 471, + 462, + 485 + ], + "spans": [ + { + "bbox": [ + 116, + 471, + 462, + 485 + ], + "score": 1.0, + "content": "Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 506, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "score": 1.0, + "content": "Hattie Zhou, Janice Lan, Rosanne Liu, and Jason Yosinski. Deconstructing lottery tickets: Zeros,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 116, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "signs, and the supermask. In Advances in Neural Information Processing Systems, pp. 3597–3607,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 511, + 143, + 526 + ], + "spans": [ + { + "bbox": [ + 115, + 511, + 143, + 526 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 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 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 287, + 116 + ], + "score": 1.0, + "content": "processing systems, pp. 5998–6008, 2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 81, + 505, + 116 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "score": 1.0, + "content": "Oriol Vinyals, Igor Babuschkin, Wojciech M. Czarnecki, Michael Mathieu, Andrew Dudzik, Juny- ¨", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "oung Chung, David H. Choi, Richard Powell, Timo Ewalds, Petko Georgiev, Junhyuk Oh, Dan", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 505, + 158 + ], + "score": 1.0, + "content": "Horgan, Manuel Kroiss, Ivo Danihelka, Aja Huang, Laurent Sifre, Trevor Cai, John P. Agapiou,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "Max Jaderberg, Alexander S. Vezhnevets, Remi Leblond, Tobias Pohlen, Valentin Dalibard, David ´", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 167, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 115, + 167, + 506, + 181 + ], + "score": 1.0, + "content": "Budden, Yury Sulsky, James Molloy, Tom L. Paine, Caglar Gulcehre, Ziyu Wang, Tobias Pfaff,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 177, + 505, + 191 + ], + "score": 1.0, + "content": "Yuhuai Wu, Roman Ring, Dani Yogatama, Dario Wunsch, Katrina McKinney, Oliver Smith, Tom ¨", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 189, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 115, + 189, + 505, + 202 + ], + "score": 1.0, + "content": "Schaul, Timothy Lillicrap, Koray Kavukcuoglu, Demis Hassabis, Chris Apps, and David Silver.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 199, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 115, + 199, + 505, + 214 + ], + "score": 1.0, + "content": "Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature, 575(7782):", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 211, + 225, + 223 + ], + "spans": [ + { + "bbox": [ + 116, + 211, + 225, + 223 + ], + "score": 1.0, + "content": "350–354, November 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 122, + 506, + 223 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 230, + 504, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "score": 1.0, + "content": "Sinong Wang, Belinda Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 240, + 353, + 253 + ], + "spans": [ + { + "bbox": [ + 115, + 240, + 353, + 253 + ], + "score": 1.0, + "content": "linear complexity. arXiv preprint arXiv:2006.04768, 2020.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 230, + 505, + 253 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 259, + 504, + 282 + ], + "lines": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "John Wieting and Douwe Kiela. No training required: Exploring random encoders for sentence", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 270, + 336, + 283 + ], + "spans": [ + { + "bbox": [ + 115, + 270, + 336, + 283 + ], + "score": 1.0, + "content": "classification. arXiv preprint arXiv:1901.10444, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 260, + 505, + 283 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 503, + 312 + ], + "lines": [ + { + "bbox": [ + 106, + 288, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 303 + ], + "score": 1.0, + "content": "Adina Williams, Nikita Nangia, and Samuel R Bowman. A broad-coverage challenge corpus for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 301, + 452, + 313 + ], + "spans": [ + { + "bbox": [ + 115, + 301, + 452, + 313 + ], + "score": 1.0, + "content": "sentence understanding through inference. arXiv preprint arXiv:1704.05426, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 288, + 505, + 313 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 319, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 333 + ], + "score": 1.0, + "content": "Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 330, + 324, + 343 + ], + "spans": [ + { + "bbox": [ + 116, + 330, + 324, + 343 + ], + "score": 1.0, + "content": "learning. Machine learning, 8(3-4):229–256, 1992.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 318, + 505, + 343 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 349, + 504, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "Felix Wu, Angela Fan, Alexei Baevski, Yann N Dauphin, and Michael Auli. Pay less attention with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 360, + 438, + 372 + ], + "spans": [ + { + "bbox": [ + 116, + 360, + 438, + 372 + ], + "score": 1.0, + "content": "lightweight and dynamic convolutions. arXiv preprint arXiv:1901.10430, 2019.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 348, + 505, + 372 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 379, + 504, + 413 + ], + "lines": [ + { + "bbox": [ + 105, + 378, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 393 + ], + "score": 1.0, + "content": "Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 116, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 402, + 325, + 413 + ], + "spans": [ + { + "bbox": [ + 116, + 402, + 325, + 413 + ], + "score": 1.0, + "content": "sequences. arXiv preprint arXiv:2007.14062, 2020.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 378, + 506, + 413 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 420, + 504, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "score": 1.0, + "content": "Chiyuan Zhang, Samy Bengio, and Yoram Singer. Are all layers created equal? arXiv preprint", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 430, + 219, + 443 + ], + "spans": [ + { + "bbox": [ + 115, + 430, + 219, + 443 + ], + "score": 1.0, + "content": "arXiv:1902.01996, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 419, + 505, + 443 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "score": 1.0, + "content": "Kelly Zhang and Samuel Bowman. Language modeling teaches you more than translation does:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 115, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "Lessons learned through auxiliary syntactic task analysis. In Proceedings of the 2018 EMNLP", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 471, + 462, + 485 + ], + "spans": [ + { + "bbox": [ + 116, + 471, + 462, + 485 + ], + "score": 1.0, + "content": "Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 450, + 506, + 485 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 506, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "score": 1.0, + "content": "Hattie Zhou, Janice Lan, Rosanne Liu, and Jason Yosinski. Deconstructing lottery tickets: Zeros,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 116, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "signs, and the supermask. In Advances in Neural Information Processing Systems, pp. 3597–3607,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 511, + 143, + 526 + ], + "spans": [ + { + "bbox": [ + 115, + 511, + 143, + 526 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 489, + 505, + 526 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "spans": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "score": 0.956, + "type": "image", + "image_path": "b02380450760b4c0b3afd0aed260a881b2948fde2eb7140d57dab6c452cac15a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 97, + 486, + 139.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 139.66666666666666, + 486, + 182.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 182.33333333333331, + 486, + 224.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 115, + 238, + 491, + 251 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 238, + 492, + 252 + ], + "spans": [ + { + "bbox": [ + 118, + 238, + 492, + 252 + ], + "score": 1.0, + "content": "Figure 6: IWSLT comparison of different hybrid architectures with different reservoir layers.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 113, + 281, + 486, + 411 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 281, + 486, + 411 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 113, + 281, + 486, + 411 + ], + "spans": [ + { + "bbox": [ + 113, + 281, + 486, + 411 + ], + "score": 0.96, + "type": "image", + "image_path": "91e3cdb50285f6e0d0f5645a8f28b599dc264c710e34a17c857f20ebe2df2ca1.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 113, + 281, + 486, + 324.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 113, + 324.3333333333333, + 486, + 367.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 113, + 367.66666666666663, + 486, + 410.99999999999994 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 160, + 425, + 451, + 437 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 159, + 423, + 451, + 439 + ], + "spans": [ + { + "bbox": [ + 159, + 423, + 451, + 439 + ], + "score": 1.0, + "content": "Figure 7: IWSLT validation AUCC and test BLEU with 6-layer decoder.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 107, + 460, + 437, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 460, + 438, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 438, + 475 + ], + "score": 1.0, + "content": "A HYBRID NETWORKS AND NON-TRANSFORMER RESERVOIRS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 504, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 504, + 501 + ], + "score": 1.0, + "content": "We investigate whether reservoir layers need to be transformer-based (or transformers-without-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "attention, i.e., FFN). We examine two different alternatives: bidirectional Gated Recurrent Units", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 512, + 504, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 504, + 523 + ], + "score": 1.0, + "content": "(Cho et al., 2014) and Convolutional Neural Networks (LeCun et al., 1998; Kim, 2014), specifi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 522, + 504, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 504, + 534 + ], + "score": 1.0, + "content": "cally light dynamical convolutions (Wu et al., 2019). Figure 6 shows the results for these hybrids:", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "depending on the setting, they may obtain a better AUCC than the regular transformer, but this is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 544, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 557 + ], + "score": 1.0, + "content": "less consistent than with the other reservoir layers, most likely because these layers have different", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "computational properties. It’s possible that these hybrids simply require further tuning, as we found", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 104, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "e.g. up-projecting to help for BiGRUs, but studying this is outside of the scope of the current work.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 108, + 599, + 214, + 611 + ], + "lines": [ + { + "bbox": [ + 104, + 597, + 216, + 614 + ], + "spans": [ + { + "bbox": [ + 104, + 597, + 216, + 614 + ], + "score": 1.0, + "content": "B DEEP DECODERS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 504, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "We show that the same results hold for a 6-layer decoder on IWSLT (although less pronounced for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 638, + 478, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 478, + 650 + ], + "score": 1.0, + "content": "AUCC, probably because the decoder is computationally heavier). See Figure 7 and Table 2.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 108, + 670, + 238, + 683 + ], + "lines": [ + { + "bbox": [ + 105, + 669, + 240, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 240, + 685 + ], + "score": 1.0, + "content": "C FREEZING STRATEGY", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 504, + 712 + ], + "score": 1.0, + "content": "We explored different strategies for the placement of reservoir layers and found the “alternating”", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "strategy reported in the main body of the paper to work best. Generally, we found repetitive appli-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 721, + 446, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 446, + 733 + ], + "score": 1.0, + "content": "cation of reservoirs to yield diminishing returns, as might be expected. See Figure 8.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 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 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 97, + 486, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "spans": [ + { + "bbox": [ + 112, + 97, + 486, + 225 + ], + "score": 0.956, + "type": "image", + "image_path": "b02380450760b4c0b3afd0aed260a881b2948fde2eb7140d57dab6c452cac15a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 97, + 486, + 139.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 139.66666666666666, + 486, + 182.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 182.33333333333331, + 486, + 224.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 115, + 238, + 491, + 251 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 238, + 492, + 252 + ], + "spans": [ + { + "bbox": [ + 118, + 238, + 492, + 252 + ], + "score": 1.0, + "content": "Figure 6: IWSLT comparison of different hybrid architectures with different reservoir layers.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 113, + 281, + 486, + 411 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 281, + 486, + 411 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 113, + 281, + 486, + 411 + ], + "spans": [ + { + "bbox": [ + 113, + 281, + 486, + 411 + ], + "score": 0.96, + "type": "image", + "image_path": "91e3cdb50285f6e0d0f5645a8f28b599dc264c710e34a17c857f20ebe2df2ca1.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 113, + 281, + 486, + 324.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 113, + 324.3333333333333, + 486, + 367.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 113, + 367.66666666666663, + 486, + 410.99999999999994 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 160, + 425, + 451, + 437 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 159, + 423, + 451, + 439 + ], + "spans": [ + { + "bbox": [ + 159, + 423, + 451, + 439 + ], + "score": 1.0, + "content": "Figure 7: IWSLT validation AUCC and test BLEU with 6-layer decoder.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 107, + 460, + 437, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 460, + 438, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 438, + 475 + ], + "score": 1.0, + "content": "A HYBRID NETWORKS AND NON-TRANSFORMER RESERVOIRS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 460, + 438, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 504, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 504, + 501 + ], + "score": 1.0, + "content": "We investigate whether reservoir layers need to be transformer-based (or transformers-without-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "attention, i.e., FFN). We examine two different alternatives: bidirectional Gated Recurrent Units", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 512, + 504, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 504, + 523 + ], + "score": 1.0, + "content": "(Cho et al., 2014) and Convolutional Neural Networks (LeCun et al., 1998; Kim, 2014), specifi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 522, + 504, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 504, + 534 + ], + "score": 1.0, + "content": "cally light dynamical convolutions (Wu et al., 2019). Figure 6 shows the results for these hybrids:", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "depending on the setting, they may obtain a better AUCC than the regular transformer, but this is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 544, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 557 + ], + "score": 1.0, + "content": "less consistent than with the other reservoir layers, most likely because these layers have different", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "computational properties. It’s possible that these hybrids simply require further tuning, as we found", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 104, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "e.g. up-projecting to help for BiGRUs, but studying this is outside of the scope of the current work.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5, + "bbox_fs": [ + 104, + 490, + 506, + 578 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 599, + 214, + 611 + ], + "lines": [ + { + "bbox": [ + 104, + 597, + 216, + 614 + ], + "spans": [ + { + "bbox": [ + 104, + 597, + 216, + 614 + ], + "score": 1.0, + "content": "B DEEP DECODERS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 504, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "We show that the same results hold for a 6-layer decoder on IWSLT (although less pronounced for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 638, + 478, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 478, + 650 + ], + "score": 1.0, + "content": "AUCC, probably because the decoder is computationally heavier). See Figure 7 and Table 2.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 627, + 505, + 650 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 670, + 238, + 683 + ], + "lines": [ + { + "bbox": [ + 105, + 669, + 240, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 240, + 685 + ], + "score": 1.0, + "content": "C FREEZING STRATEGY", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 504, + 712 + ], + "score": 1.0, + "content": "We explored different strategies for the placement of reservoir layers and found the “alternating”", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "strategy reported in the main body of the paper to work best. Generally, we found repetitive appli-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 721, + 446, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 446, + 733 + ], + "score": 1.0, + "content": "cation of reservoirs to yield diminishing returns, as might be expected. See Figure 8.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 698, + 504, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 80, + 505, + 249 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 80, + 505, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 505, + 249 + ], + "score": 0.985, + "html": "
Model#LayersFrozenMax BLEUTrain time until max (in hours)Ratio# Params Trainable (Total)Train Time each epoch (in seconds)
Transformer6034.97 ± 0.051.984 ± 0.02139.5M177.84 ± 2.98
8034.99 ± 0.082.161 ± 0.03143.7M206.59 ± 3.47
10034.98 ± 0.042.345± 0.02147.9M236.72 ± 3.52
12034.78 ± 0.112.535 ± 0.05152.0M265.90 ± 4.97
TReservoir6234.73 ± 0.111.838 ± 0.010.9235.3M (39.5M)166.11 ± 2.21
8235.07 ± 0.051.912 ± 0.030.8839.5M (43.7M)190.08 ± 3.73
10235.02 ± 0.011.970 ± 0.040.8443.7M (47.9M)204.42 ± 2.89
12235.06 ± 0.022.429 ± 0.020.9547.8M (52.0M)236.41 ± 4.35
FFN Reservoir6234.85 ± 0.101.729 ± 0.030.8735.3M (37.4M)161.72 ± 2.32
8234.99 ± 0.111.751 ± 0.020.8139.5M (41.6M)180.21 ± 2.68
10234.92 ± 0.031.907 ± 0.020.8143.7M (45.8M)191.40 ± 2.49
12235.16 ± 0.042.395 ± 0.010.9447.8M (49.9M)216.08 ± 2.57
LayerDrop6２234.51 ± 0.121.908 ± 0.040.9635.3M (39.5M)169.62 ± 3.16
834.77 ± 0.112.023 ± 0.020.9439.5M (43.7M)186.71 ± 2.17
10234.06 ± 0.051.912 ± 0.020.9743.7M (47.9M)205.52 ± 3.31
12234.08 ± 0.132.524 ± 0.010.9947.8M (52.0M)222.45 ± 2.21
", + "type": "table", + "image_path": "096f963793969ec1e760d733f7cd6004b153a4f7197889afc9ad33f53fb344ac.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 505, + 136.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 136.33333333333334, + 505, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 192.66666666666669, + 505, + 249.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 258, + 506, + 292 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "score": 1.0, + "content": "Table 2: Wall-clock time (averaged over multiple runs) saved for IWSLT for different model types", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "and encoder depths. Max BLEU is for validation. Number of layers is for encoder, decoder depth is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 280, + 492, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 492, + 293 + ], + "score": 1.0, + "content": "kept fixed at 6. Ratio is computed compared to comparable number of layers in the normal case.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 108, + 307, + 505, + 477 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 307, + 505, + 477 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 307, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 108, + 307, + 505, + 477 + ], + "score": 0.985, + "html": "
Model # LayersFrozenMax BLEUTrain time until max (in hours)Ratio# Params Trainable (Total)Train Time each epoch (in hours)
Transformer12024.46 ± 0.0415.15 ± 0.15175.6M0.505 ± 0.005
16024.52 ± 0.0316.05 ± 0.1888.2M0.643 ± 0.006
24024.69 ± 0.0517.61 ± 0.851113.4M0.877 ± 0.029
32024.83 ± 0.0418.42 ± 0.281138.6M1.036 ± 0.010
TReservoir12424.26 ± 0.0814.11 ± 0.210.9372.4M (75.6M)0.472 ± 0.007
16424.50 ± 0.0515.25 ± 0.280.9575.6M (88.2M)0.596 ± 0.009
24425.11 ± 0.0715.89 ± 0.740.90100.8M (113.4M)0.776 ± 0.024
32424.66 ± 0.0416.38 ± 0.240.88126.0M (138.6M)0.998 ± 0.009
FFN Reservoir12424.42 ± 0.0514.01 ± 0.090.9272.4M (71.4M)0.441 ± 0.003
16424.65 ± 0.0714.53 ± 0.170.9175.6M (83.9M)0.524 ± 0.006
24424.93 ± 0.0412.62 ± 1.530.71100.8M (109.2M)0.743 ± 0.018
32424.98 ± 0.0313.96 ± 0.190.73126.0M (134.4M)0.964 ± 0.007
LayerDrop12424.27 ± 0.0314.61 ± 0.140.9672.4M (75.6M)0.489 ± 0.006
16424.15 ± 0.0615.55 ± 0.540.9775.6M (88.2M)0.597 ± 0.017
24424.37 ± 0.0516.25 ± 0.360.92100.8M (113.4M)0.823 ± 0.013
32423.84 ± 0.0315.27 ± 0.380.83126.0M (138.6M)1.028 ± 0.012
", + "type": "table", + "image_path": "d1cac1c7bd6e33cd4703a3ee16df8151c3579d2fb6cc9b58629b215468cd37aa.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 108, + 307, + 505, + 363.6666666666667 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 363.6666666666667, + 505, + 420.33333333333337 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 108, + 420.33333333333337, + 505, + 477.00000000000006 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 486, + 505, + 520 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "score": 1.0, + "content": "Table 3: Wall-clock time (averaged over multiple runs) saved for WMT for different model types", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "and encoder depths. Max BLEU is for validation. Number of layers is for encoder, decoder depth is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 507, + 492, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 492, + 520 + ], + "score": 1.0, + "content": "kept fixed at 1. Ratio is computed compared to comparable number of layers in the normal case.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 544, + 232, + 557 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 234, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 234, + 559 + ], + "score": 1.0, + "content": "D ROBERTA RESULTS", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "Here we present the additional RoBERTa results for convergence plot and AUCC in various decoder", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "depth setting in Figure 10. As stated in the main paper, the difference of AUCC / Convergence", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "Plot between RoBERTa model with or without Reservoir layers are limited. Moreover, we plot the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "downstream task performance for SST-2 and MNLI compared to the pretraining wall-clock time", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 397, + 630 + ], + "score": 1.0, + "content": "in Figure 9. It can be seen that the FFN Reservoir can achieve up to", + "type": "text" + }, + { + "bbox": [ + 397, + 617, + 417, + 627 + ], + "score": 0.86, + "content": "\\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 615, + 437, + 630 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 617, + 457, + 627 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "pretraining", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "time savings while matching the best performance of vanilla transformers for MNLI-m and SST2,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 638, + 159, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 159, + 651 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 107, + 671, + 338, + 684 + ], + "lines": [ + { + "bbox": [ + 104, + 669, + 339, + 686 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 339, + 686 + ], + "score": 1.0, + "content": "E RESERVOIR RESULTS FOR TOTAL LAYERS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "Here we present the shifted Reservoir Results for IWSLT14, WMT16, Enwik8 and RoBERTa fine-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "tuning in Figure 11, 12, 13, 14, respectively. We show the same results also hold when it comes to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 721, + 399, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 399, + 732 + ], + "score": 1.0, + "content": "replace normal transformer blocks with Reservoir blocks at least for MT.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 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 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 80, + 505, + 249 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 80, + 505, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 505, + 249 + ], + "score": 0.985, + "html": "
Model#LayersFrozenMax BLEUTrain time until max (in hours)Ratio# Params Trainable (Total)Train Time each epoch (in seconds)
Transformer6034.97 ± 0.051.984 ± 0.02139.5M177.84 ± 2.98
8034.99 ± 0.082.161 ± 0.03143.7M206.59 ± 3.47
10034.98 ± 0.042.345± 0.02147.9M236.72 ± 3.52
12034.78 ± 0.112.535 ± 0.05152.0M265.90 ± 4.97
TReservoir6234.73 ± 0.111.838 ± 0.010.9235.3M (39.5M)166.11 ± 2.21
8235.07 ± 0.051.912 ± 0.030.8839.5M (43.7M)190.08 ± 3.73
10235.02 ± 0.011.970 ± 0.040.8443.7M (47.9M)204.42 ± 2.89
12235.06 ± 0.022.429 ± 0.020.9547.8M (52.0M)236.41 ± 4.35
FFN Reservoir6234.85 ± 0.101.729 ± 0.030.8735.3M (37.4M)161.72 ± 2.32
8234.99 ± 0.111.751 ± 0.020.8139.5M (41.6M)180.21 ± 2.68
10234.92 ± 0.031.907 ± 0.020.8143.7M (45.8M)191.40 ± 2.49
12235.16 ± 0.042.395 ± 0.010.9447.8M (49.9M)216.08 ± 2.57
LayerDrop6２234.51 ± 0.121.908 ± 0.040.9635.3M (39.5M)169.62 ± 3.16
834.77 ± 0.112.023 ± 0.020.9439.5M (43.7M)186.71 ± 2.17
10234.06 ± 0.051.912 ± 0.020.9743.7M (47.9M)205.52 ± 3.31
12234.08 ± 0.132.524 ± 0.010.9947.8M (52.0M)222.45 ± 2.21
", + "type": "table", + "image_path": "096f963793969ec1e760d733f7cd6004b153a4f7197889afc9ad33f53fb344ac.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 505, + 136.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 136.33333333333334, + 505, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 192.66666666666669, + 505, + 249.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 258, + 506, + 292 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "score": 1.0, + "content": "Table 2: Wall-clock time (averaged over multiple runs) saved for IWSLT for different model types", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "and encoder depths. Max BLEU is for validation. Number of layers is for encoder, decoder depth is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 280, + 492, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 492, + 293 + ], + "score": 1.0, + "content": "kept fixed at 6. Ratio is computed compared to comparable number of layers in the normal case.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 108, + 307, + 505, + 477 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 307, + 505, + 477 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 307, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 108, + 307, + 505, + 477 + ], + "score": 0.985, + "html": "
Model # LayersFrozenMax BLEUTrain time until max (in hours)Ratio# Params Trainable (Total)Train Time each epoch (in hours)
Transformer12024.46 ± 0.0415.15 ± 0.15175.6M0.505 ± 0.005
16024.52 ± 0.0316.05 ± 0.1888.2M0.643 ± 0.006
24024.69 ± 0.0517.61 ± 0.851113.4M0.877 ± 0.029
32024.83 ± 0.0418.42 ± 0.281138.6M1.036 ± 0.010
TReservoir12424.26 ± 0.0814.11 ± 0.210.9372.4M (75.6M)0.472 ± 0.007
16424.50 ± 0.0515.25 ± 0.280.9575.6M (88.2M)0.596 ± 0.009
24425.11 ± 0.0715.89 ± 0.740.90100.8M (113.4M)0.776 ± 0.024
32424.66 ± 0.0416.38 ± 0.240.88126.0M (138.6M)0.998 ± 0.009
FFN Reservoir12424.42 ± 0.0514.01 ± 0.090.9272.4M (71.4M)0.441 ± 0.003
16424.65 ± 0.0714.53 ± 0.170.9175.6M (83.9M)0.524 ± 0.006
24424.93 ± 0.0412.62 ± 1.530.71100.8M (109.2M)0.743 ± 0.018
32424.98 ± 0.0313.96 ± 0.190.73126.0M (134.4M)0.964 ± 0.007
LayerDrop12424.27 ± 0.0314.61 ± 0.140.9672.4M (75.6M)0.489 ± 0.006
16424.15 ± 0.0615.55 ± 0.540.9775.6M (88.2M)0.597 ± 0.017
24424.37 ± 0.0516.25 ± 0.360.92100.8M (113.4M)0.823 ± 0.013
32423.84 ± 0.0315.27 ± 0.380.83126.0M (138.6M)1.028 ± 0.012
", + "type": "table", + "image_path": "d1cac1c7bd6e33cd4703a3ee16df8151c3579d2fb6cc9b58629b215468cd37aa.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 108, + 307, + 505, + 363.6666666666667 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 363.6666666666667, + 505, + 420.33333333333337 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 108, + 420.33333333333337, + 505, + 477.00000000000006 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 486, + 505, + 520 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "score": 1.0, + "content": "Table 3: Wall-clock time (averaged over multiple runs) saved for WMT for different model types", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "and encoder depths. Max BLEU is for validation. Number of layers is for encoder, decoder depth is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 507, + 492, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 492, + 520 + ], + "score": 1.0, + "content": "kept fixed at 1. Ratio is computed compared to comparable number of layers in the normal case.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 544, + 232, + 557 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 234, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 234, + 559 + ], + "score": 1.0, + "content": "D ROBERTA RESULTS", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "Here we present the additional RoBERTa results for convergence plot and AUCC in various decoder", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "depth setting in Figure 10. As stated in the main paper, the difference of AUCC / Convergence", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "Plot between RoBERTa model with or without Reservoir layers are limited. Moreover, we plot the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "downstream task performance for SST-2 and MNLI compared to the pretraining wall-clock time", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 397, + 630 + ], + "score": 1.0, + "content": "in Figure 9. It can be seen that the FFN Reservoir can achieve up to", + "type": "text" + }, + { + "bbox": [ + 397, + 617, + 417, + 627 + ], + "score": 0.86, + "content": "\\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 615, + 437, + 630 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 617, + 457, + 627 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "pretraining", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "time savings while matching the best performance of vanilla transformers for MNLI-m and SST2,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 638, + 159, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 159, + 651 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 572, + 506, + 651 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 671, + 338, + 684 + ], + "lines": [ + { + "bbox": [ + 104, + 669, + 339, + 686 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 339, + 686 + ], + "score": 1.0, + "content": "E RESERVOIR RESULTS FOR TOTAL LAYERS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "Here we present the shifted Reservoir Results for IWSLT14, WMT16, Enwik8 and RoBERTa fine-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "tuning in Figure 11, 12, 13, 14, respectively. We show the same results also hold when it comes to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 721, + 399, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 399, + 732 + ], + "score": 1.0, + "content": "replace normal transformer blocks with Reservoir blocks at least for MT.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 80, + 505, + 202 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 80, + 505, + 202 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 505, + 202 + ], + "score": 0.982, + "html": "
Mode1#LayersIWSLT-Dec2 Train time until 95% max (in hours)Max BLEU (95%)#LayersIWSLT-Dec6 Train time until 95% max (in hours)Max BLEU (95%)#LayersWMT-Dec1 Train time until 95% max (in hours)Max BLEU (95%)
Transformer60.647 ± 0.0332.89 ± 0.0460.642 ± 0.0233.36 ± 0.03163.788 ± 0.05323.36 ± 0.06
0.711 ± 0.0533.04 ± 0.0380.765 ± 0.0333.41 ± 0.083.820 ± 0.07223.41 ± 0.05
0.808 ± 0.0233.96 ± 0.08100.898 ± 0.0433.32 ± 0.075.262 ± 0.60723.50 ± 0.03
121.037 ± 0.0333.07 ± 0.09121.037 ± 0.0333.07 ± 0.11五6.212 ± 0.23223.81 ±0.04
0.569 ± 0.0232.78 ±0.0360.599 ± 0.0133.09 ± 0.053.563 ± 0.06123.21 ± 0.04
80.619 ± 0.0433.12 ± 0.0580.726 ± 0.0233.38 ± 0.0913.603 ± 0.05623.80 ± 0.06
T Reservoir0.729 ± 0.0433.13 ± 0.07100.738 ± 0.0333.37 ± 0.04244.923 ± 0.77123.75 ± 0.02
120.982 ± 0.0233.03 ± 0.11120.958 ± 0.0133.46± 0.09325.780 ± 0.21423.71 ±0.03
60.521 ± 0.0532.85 ± 0.0260.594 ± 0.0333.13 ± 0.04123.417 ± 0.04623.22 ± 0.07
80.533 ± 0.0333.84 ± 0.0480.651 ± 0.0433.36 ± 0.06163.527 ± 0.06323.54 ± 0.05
100.614 ± 0.0133.05 ± 0.08100.627 ± 0.0533.26 ± 0.03244.197 ± 0.69723.74 ± 0.06
LayerDrop120.811 ± 0.0233.26 ± 0.10120.780 ± 0.0233.46 ± 0.08324.984 ± 0.32123.82 ± 0.02
0.837 ±0.0832.87 ± 0.0560.706 ±0.0133.08 ± 0.03123.912 ± 0.06823.33 ± 0.08
60.934 ± 0.0733.12 ± 0.0380.753 ± 0.0433.14 ± 0.05163.581 ± 0.07623.17 ± 0.04
100.901 ± 0.0633.18 ±0.02100.691 ± 0.0332.39 ± 0.0534.875 ± 0.72823.43 ± 0.07
120.914 ± 0.0132.33 ± 0.06120.803 ± 0.0232.94 ± 0.105.980 ± 0.21922.97 ± 0.08
", + "type": "table", + "image_path": "4d3fb4527e6f03bb61941ed5effff97c8cabc76bd21ddb711c93f44c7bf76a44.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 505, + 120.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 120.66666666666666, + 505, + 161.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 161.33333333333331, + 505, + 201.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 210, + 504, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 210, + 504, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 504, + 221 + ], + "score": 1.0, + "content": "Table 4: Wall-clock time (averaged over multiple runs) saved for IWSLT/WMT for different model", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 220, + 351, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 212, + 234 + ], + "score": 1.0, + "content": "types and encoder depths.", + "type": "text" + }, + { + "bbox": [ + 212, + 221, + 232, + 232 + ], + "score": 0.81, + "content": "\\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 220, + 351, + 234 + ], + "score": 1.0, + "content": "Max BLEU is for validation.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 107, + 244, + 505, + 366 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 107, + 244, + 505, + 366 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 244, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 107, + 244, + 505, + 366 + ], + "score": 0.982, + "html": "
Model#LayersIWSLT-Dec2 Train timeIWSLT-Dec6WMT-Dec1
until 99 % max (in hours)Max BLEU (99%)#LayersTrain time until 99 % max (in hours)Max BLEU (99%)#LayersTrain time until 99 % max (in hours)Max BLEU (99%)
Transformer1.454 ± 0.0634.24 ± 0.0561.297 ± 0.0334.69 ± 0.0519.961 ± 0.05324.27 ± 0.04
1.475 ± 0.0934.32 ± 0.0981.390 ± 0.0212.623 ± 0.07224.35 ± 0.06
101.526 ± 0.0434.25 ± 0.041.622 ± 0.0534.75 ± 0.09 34.64 ± 0.03313.412 ± 0.83724.49 ± 0.07
122.259 ± 0.0734.24 ± 0.1110 121.748 ± 0.0134.66 ± 0.0815.117 ± 0.23224.56 ± 0.02
TReservoir1.257 ± 0.0434.05 ± 0.0961.291 ± 0.0334.51 ± 0.1018.314 ± 0.06224.15 ± 0.06
101.472 ± 0.0634.47 ± 0.051.339 ± 0.0334.80 ± 0.049.221 ± 0.07324.41 ± 0.05
121.530 ± 0.0334.36 ± 0.02101.419 ± 0.0434.72 ± 0.0310.413 ± 0.58024.56 ± 0.03
2.043 ± 0.0534.53 ± 0.07121.642 ± 0.0234.87 ± 0.02311.465 ± 0.22724.49 ±0.01
6801.138 ± 0.0334.10 ± 0.1361.169 ± 0.0234.71 ± 0.097.407 ± 0.08724.33 ± 0.08
FFN Reservoir1.101 ± 0.0734.32 ± 0.11 34.36 ± 0.038 101.201 ± 0.03 1.276 ± 0.0334.79 ±0.08 34.63 ± 0.039.336 ± 0.03624.42 ± 0.05
121.281 ± 0.019.978 ± 0.54624.91 ± 0.07
1.785 ± 0.0334.42 ± 0.06121.440 ± 0.0134.87 ± 0.0210.524 ± 0.34124.96 ± 0.01
81.363 ± 0.0534.58 ± 0.1461.253 ± 0.0134.42 ± 0.108.372 ± 0.05924.17 ± 0.04
1.468 ± 0.0334.50 ± 0.1281.244 ± 0.0434.44 ± 0.099.741 ± 0.04323.93 ± 0.08
LayerDrop101.678 ± 0.0434.52 ± 0.071033.83 ±0.0616 310.145 ± 0.62824.07 ± 0.09
122.071 ± 0.0233.45 ± 0.23121.343 ± 0.04 1.423 ± 0.0233.97 ± 0.1210.168 ± 0.32923.81 ± 0.03
", + "type": "table", + "image_path": "cc2eb78eb0712912677703c3e20189a9030afe6a087941113b4d0f72dac4402a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 107, + 244, + 505, + 284.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 284.6666666666667, + 505, + 325.33333333333337 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 107, + 325.33333333333337, + 505, + 366.00000000000006 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 374, + 504, + 396 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "Table 5: Wall-clock time (averaged over multiple runs) saved for IWSLT/WMT for different model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 384, + 351, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 212, + 397 + ], + "score": 1.0, + "content": "types and encoder depths.", + "type": "text" + }, + { + "bbox": [ + 212, + 385, + 232, + 396 + ], + "score": 0.84, + "content": "9 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 384, + 351, + 397 + ], + "score": 1.0, + "content": "Max BLEU is for validation.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "title", + "bbox": [ + 108, + 416, + 227, + 429 + ], + "lines": [ + { + "bbox": [ + 104, + 415, + 228, + 432 + ], + "spans": [ + { + "bbox": [ + 104, + 415, + 228, + 432 + ], + "score": 1.0, + "content": "F VALIDATION PLOTS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 441, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 506, + 454 + ], + "score": 1.0, + "content": "Here we present the validation plots for training a 8-layer encoder, 2-layer decoder model for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "IWSLT14, a 24-layer encoder, 1-layer decoder model for WMT14, a 48-layer decoder model for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 463, + 504, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 504, + 474 + ], + "score": 1.0, + "content": "enwik8 and a 12-layer decoder model for RoBERTa for detailed steps to calculate the AUCC. It can", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 474, + 504, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 504, + 487 + ], + "score": 1.0, + "content": "be clearly observed that given the configurations from Section 3.1, all the models have converged.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "So when we compute the area under the convergence curve, this depicts the training efficiency of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "model (basically time x performance) until convergence. Specifically, we set T sufficiently high for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "computing the AUCC, which is 4h for IWSLT, 20h for WMT, 30h for enwik8 and 60h for RoBERTa", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "pretraning. From the training plot in the appendix, we can see that each model has converged at", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "that point. The Reservoir model in Figure 15 has 2 layers frozen for IWSLT14, 8 layers frozen for", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 540, + 331, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 331, + 551 + ], + "score": 1.0, + "content": "enwik8, and 4 layers frozen for WMT14 and RoBERTa.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 107, + 567, + 234, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 235, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 235, + 583 + ], + "score": 1.0, + "content": "G ROBERTA PROBING", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 506, + 605 + ], + "score": 1.0, + "content": "We follow Jawahar et al. (2019) and investigate what the frozen layers in the Reservoir Transformer", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "score": 1.0, + "content": "have actually “learned” (while being forzen) as measured by probing tasks, reported in Table 6. The", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "results are gathered over 3 random seeds for reporting the mean and standard deviation. From the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 624, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 638 + ], + "score": 1.0, + "content": "table, we can see that generally probing performance is quite similar between Transformer and the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 635, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 649 + ], + "score": 1.0, + "content": "T Reservoir model. We also noticed that the representations collected after the frozen layer (3, 5,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 647, + 504, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 504, + 659 + ], + "score": 1.0, + "content": "7, 9) in the T Reservoir actually have significantly better performance over the regular Transformer", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "representations across all the probing tasks. This has interesting repercussions for the study of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 668, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 104, + 668, + 506, + 683 + ], + "score": 1.0, + "content": "“BERTology”, as it clearly shows, somewhat confusingly, that even completely random and frozen", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 682, + 261, + 692 + ], + "spans": [ + { + "bbox": [ + 107, + 682, + 261, + 692 + ], + "score": 1.0, + "content": "layers represent linguistic phenomena.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 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 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 80, + 505, + 202 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 80, + 505, + 202 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 505, + 202 + ], + "score": 0.982, + "html": "
Mode1#LayersIWSLT-Dec2 Train time until 95% max (in hours)Max BLEU (95%)#LayersIWSLT-Dec6 Train time until 95% max (in hours)Max BLEU (95%)#LayersWMT-Dec1 Train time until 95% max (in hours)Max BLEU (95%)
Transformer60.647 ± 0.0332.89 ± 0.0460.642 ± 0.0233.36 ± 0.03163.788 ± 0.05323.36 ± 0.06
0.711 ± 0.0533.04 ± 0.0380.765 ± 0.0333.41 ± 0.083.820 ± 0.07223.41 ± 0.05
0.808 ± 0.0233.96 ± 0.08100.898 ± 0.0433.32 ± 0.075.262 ± 0.60723.50 ± 0.03
121.037 ± 0.0333.07 ± 0.09121.037 ± 0.0333.07 ± 0.11五6.212 ± 0.23223.81 ±0.04
0.569 ± 0.0232.78 ±0.0360.599 ± 0.0133.09 ± 0.053.563 ± 0.06123.21 ± 0.04
80.619 ± 0.0433.12 ± 0.0580.726 ± 0.0233.38 ± 0.0913.603 ± 0.05623.80 ± 0.06
T Reservoir0.729 ± 0.0433.13 ± 0.07100.738 ± 0.0333.37 ± 0.04244.923 ± 0.77123.75 ± 0.02
120.982 ± 0.0233.03 ± 0.11120.958 ± 0.0133.46± 0.09325.780 ± 0.21423.71 ±0.03
60.521 ± 0.0532.85 ± 0.0260.594 ± 0.0333.13 ± 0.04123.417 ± 0.04623.22 ± 0.07
80.533 ± 0.0333.84 ± 0.0480.651 ± 0.0433.36 ± 0.06163.527 ± 0.06323.54 ± 0.05
100.614 ± 0.0133.05 ± 0.08100.627 ± 0.0533.26 ± 0.03244.197 ± 0.69723.74 ± 0.06
LayerDrop120.811 ± 0.0233.26 ± 0.10120.780 ± 0.0233.46 ± 0.08324.984 ± 0.32123.82 ± 0.02
0.837 ±0.0832.87 ± 0.0560.706 ±0.0133.08 ± 0.03123.912 ± 0.06823.33 ± 0.08
60.934 ± 0.0733.12 ± 0.0380.753 ± 0.0433.14 ± 0.05163.581 ± 0.07623.17 ± 0.04
100.901 ± 0.0633.18 ±0.02100.691 ± 0.0332.39 ± 0.0534.875 ± 0.72823.43 ± 0.07
120.914 ± 0.0132.33 ± 0.06120.803 ± 0.0232.94 ± 0.105.980 ± 0.21922.97 ± 0.08
", + "type": "table", + "image_path": "4d3fb4527e6f03bb61941ed5effff97c8cabc76bd21ddb711c93f44c7bf76a44.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 505, + 120.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 120.66666666666666, + 505, + 161.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 161.33333333333331, + 505, + 201.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 210, + 504, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 210, + 504, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 504, + 221 + ], + "score": 1.0, + "content": "Table 4: Wall-clock time (averaged over multiple runs) saved for IWSLT/WMT for different model", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 220, + 351, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 212, + 234 + ], + "score": 1.0, + "content": "types and encoder depths.", + "type": "text" + }, + { + "bbox": [ + 212, + 221, + 232, + 232 + ], + "score": 0.81, + "content": "\\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 220, + 351, + 234 + ], + "score": 1.0, + "content": "Max BLEU is for validation.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 107, + 244, + 505, + 366 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 107, + 244, + 505, + 366 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 244, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 107, + 244, + 505, + 366 + ], + "score": 0.982, + "html": "
Model#LayersIWSLT-Dec2 Train timeIWSLT-Dec6WMT-Dec1
until 99 % max (in hours)Max BLEU (99%)#LayersTrain time until 99 % max (in hours)Max BLEU (99%)#LayersTrain time until 99 % max (in hours)Max BLEU (99%)
Transformer1.454 ± 0.0634.24 ± 0.0561.297 ± 0.0334.69 ± 0.0519.961 ± 0.05324.27 ± 0.04
1.475 ± 0.0934.32 ± 0.0981.390 ± 0.0212.623 ± 0.07224.35 ± 0.06
101.526 ± 0.0434.25 ± 0.041.622 ± 0.0534.75 ± 0.09 34.64 ± 0.03313.412 ± 0.83724.49 ± 0.07
122.259 ± 0.0734.24 ± 0.1110 121.748 ± 0.0134.66 ± 0.0815.117 ± 0.23224.56 ± 0.02
TReservoir1.257 ± 0.0434.05 ± 0.0961.291 ± 0.0334.51 ± 0.1018.314 ± 0.06224.15 ± 0.06
101.472 ± 0.0634.47 ± 0.051.339 ± 0.0334.80 ± 0.049.221 ± 0.07324.41 ± 0.05
121.530 ± 0.0334.36 ± 0.02101.419 ± 0.0434.72 ± 0.0310.413 ± 0.58024.56 ± 0.03
2.043 ± 0.0534.53 ± 0.07121.642 ± 0.0234.87 ± 0.02311.465 ± 0.22724.49 ±0.01
6801.138 ± 0.0334.10 ± 0.1361.169 ± 0.0234.71 ± 0.097.407 ± 0.08724.33 ± 0.08
FFN Reservoir1.101 ± 0.0734.32 ± 0.11 34.36 ± 0.038 101.201 ± 0.03 1.276 ± 0.0334.79 ±0.08 34.63 ± 0.039.336 ± 0.03624.42 ± 0.05
121.281 ± 0.019.978 ± 0.54624.91 ± 0.07
1.785 ± 0.0334.42 ± 0.06121.440 ± 0.0134.87 ± 0.0210.524 ± 0.34124.96 ± 0.01
81.363 ± 0.0534.58 ± 0.1461.253 ± 0.0134.42 ± 0.108.372 ± 0.05924.17 ± 0.04
1.468 ± 0.0334.50 ± 0.1281.244 ± 0.0434.44 ± 0.099.741 ± 0.04323.93 ± 0.08
LayerDrop101.678 ± 0.0434.52 ± 0.071033.83 ±0.0616 310.145 ± 0.62824.07 ± 0.09
122.071 ± 0.0233.45 ± 0.23121.343 ± 0.04 1.423 ± 0.0233.97 ± 0.1210.168 ± 0.32923.81 ± 0.03
", + "type": "table", + "image_path": "cc2eb78eb0712912677703c3e20189a9030afe6a087941113b4d0f72dac4402a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 107, + 244, + 505, + 284.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 284.6666666666667, + 505, + 325.33333333333337 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 107, + 325.33333333333337, + 505, + 366.00000000000006 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 374, + 504, + 396 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "Table 5: Wall-clock time (averaged over multiple runs) saved for IWSLT/WMT for different model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 384, + 351, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 212, + 397 + ], + "score": 1.0, + "content": "types and encoder depths.", + "type": "text" + }, + { + "bbox": [ + 212, + 385, + 232, + 396 + ], + "score": 0.84, + "content": "9 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 384, + 351, + 397 + ], + "score": 1.0, + "content": "Max BLEU is for validation.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "title", + "bbox": [ + 108, + 416, + 227, + 429 + ], + "lines": [ + { + "bbox": [ + 104, + 415, + 228, + 432 + ], + "spans": [ + { + "bbox": [ + 104, + 415, + 228, + 432 + ], + "score": 1.0, + "content": "F VALIDATION PLOTS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 441, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 506, + 454 + ], + "score": 1.0, + "content": "Here we present the validation plots for training a 8-layer encoder, 2-layer decoder model for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "IWSLT14, a 24-layer encoder, 1-layer decoder model for WMT14, a 48-layer decoder model for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 463, + 504, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 504, + 474 + ], + "score": 1.0, + "content": "enwik8 and a 12-layer decoder model for RoBERTa for detailed steps to calculate the AUCC. It can", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 474, + 504, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 504, + 487 + ], + "score": 1.0, + "content": "be clearly observed that given the configurations from Section 3.1, all the models have converged.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "So when we compute the area under the convergence curve, this depicts the training efficiency of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "model (basically time x performance) until convergence. Specifically, we set T sufficiently high for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "computing the AUCC, which is 4h for IWSLT, 20h for WMT, 30h for enwik8 and 60h for RoBERTa", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "pretraning. From the training plot in the appendix, we can see that each model has converged at", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "that point. The Reservoir model in Figure 15 has 2 layers frozen for IWSLT14, 8 layers frozen for", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 540, + 331, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 331, + 551 + ], + "score": 1.0, + "content": "enwik8, and 4 layers frozen for WMT14 and RoBERTa.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 441, + 506, + 551 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 567, + 234, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 235, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 235, + 583 + ], + "score": 1.0, + "content": "G ROBERTA PROBING", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 506, + 605 + ], + "score": 1.0, + "content": "We follow Jawahar et al. (2019) and investigate what the frozen layers in the Reservoir Transformer", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "score": 1.0, + "content": "have actually “learned” (while being forzen) as measured by probing tasks, reported in Table 6. The", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "results are gathered over 3 random seeds for reporting the mean and standard deviation. From the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 624, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 638 + ], + "score": 1.0, + "content": "table, we can see that generally probing performance is quite similar between Transformer and the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 635, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 649 + ], + "score": 1.0, + "content": "T Reservoir model. We also noticed that the representations collected after the frozen layer (3, 5,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 647, + 504, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 504, + 659 + ], + "score": 1.0, + "content": "7, 9) in the T Reservoir actually have significantly better performance over the regular Transformer", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "representations across all the probing tasks. This has interesting repercussions for the study of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 668, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 104, + 668, + 506, + 683 + ], + "score": 1.0, + "content": "“BERTology”, as it clearly shows, somewhat confusingly, that even completely random and frozen", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 682, + 261, + 692 + ], + "spans": [ + { + "bbox": [ + 107, + 682, + 261, + 692 + ], + "score": 1.0, + "content": "layers represent linguistic phenomena.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 591, + 506, + 692 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 192, + 118, + 402, + 270 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 192, + 118, + 402, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 192, + 118, + 402, + 270 + ], + "spans": [ + { + "bbox": [ + 192, + 118, + 402, + 270 + ], + "score": 0.962, + "type": "image", + "image_path": "bdd1e0f5437432a0076b1c35076cc1e6057826414f43599a327bf57236d29faf.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 192, + 118, + 402, + 133.2 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 192, + 133.2, + 402, + 148.39999999999998 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 192, + 148.39999999999998, + 402, + 163.59999999999997 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 192, + 163.59999999999997, + 402, + 178.79999999999995 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 192, + 178.79999999999995, + 402, + 193.99999999999994 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 192, + 193.99999999999994, + 402, + 209.19999999999993 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 192, + 209.19999999999993, + 402, + 224.39999999999992 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 192, + 224.39999999999992, + 402, + 239.5999999999999 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 192, + 239.5999999999999, + 402, + 254.7999999999999 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 192, + 254.7999999999999, + 402, + 269.9999999999999 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 161, + 285, + 449, + 298 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 160, + 283, + 450, + 301 + ], + "spans": [ + { + "bbox": [ + 160, + 283, + 450, + 301 + ], + "score": 1.0, + "content": "Figure 8: IWSLT with 2-layer decoder using different freezing strategy.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 7.25 + }, + { + "type": "image", + "bbox": [ + 117, + 354, + 486, + 483 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 354, + 486, + 483 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 117, + 354, + 486, + 483 + ], + "spans": [ + { + "bbox": [ + 117, + 354, + 486, + 483 + ], + "score": 0.968, + "type": "image", + "image_path": "8efb8273099ff74135796331759c0532b0ccf7448a7b83e8c437cb3b5dc2a6fb.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 117, + 354, + 486, + 397.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 117, + 397.0, + 486, + 440.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 117, + 440.0, + 486, + 483.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 497, + 505, + 520 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 495, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 510 + ], + "score": 1.0, + "content": "Figure 9: RoBERTa Reservoir Results, Pre-training versus downstream task plot for 12 layer", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 505, + 274, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 274, + 521 + ], + "score": 1.0, + "content": "RoBERTa. MNLI-m (left). SST-2 (right).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + } + ], + "index": 13.25 + }, + { + "type": "image", + "bbox": [ + 124, + 573, + 490, + 678 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 573, + 490, + 678 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 124, + 573, + 490, + 678 + ], + "spans": [ + { + "bbox": [ + 124, + 573, + 490, + 678 + ], + "score": 0.942, + "type": "image", + "image_path": "16c67d74e5743eb685c5b29836cc09d4540ceae3fe880d614d534330288bb088.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 124, + 573, + 490, + 608.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 124, + 608.0, + 490, + 643.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 124, + 643.0, + 490, + 678.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 689, + 507, + 712 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "Figure 10: RoBERTa Reservoir Results, Training plot for 12 layer RoBERTa (left). AUCC result", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 699, + 138, + 714 + ], + "spans": [ + { + "bbox": [ + 104, + 699, + 138, + 714 + ], + "score": 1.0, + "content": "(right).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + } + ], + "index": 18.25 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 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 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 192, + 118, + 402, + 270 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 192, + 118, + 402, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 192, + 118, + 402, + 270 + ], + "spans": [ + { + "bbox": [ + 192, + 118, + 402, + 270 + ], + "score": 0.962, + "type": "image", + "image_path": "bdd1e0f5437432a0076b1c35076cc1e6057826414f43599a327bf57236d29faf.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 192, + 118, + 402, + 133.2 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 192, + 133.2, + 402, + 148.39999999999998 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 192, + 148.39999999999998, + 402, + 163.59999999999997 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 192, + 163.59999999999997, + 402, + 178.79999999999995 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 192, + 178.79999999999995, + 402, + 193.99999999999994 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 192, + 193.99999999999994, + 402, + 209.19999999999993 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 192, + 209.19999999999993, + 402, + 224.39999999999992 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 192, + 224.39999999999992, + 402, + 239.5999999999999 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 192, + 239.5999999999999, + 402, + 254.7999999999999 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 192, + 254.7999999999999, + 402, + 269.9999999999999 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 161, + 285, + 449, + 298 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 160, + 283, + 450, + 301 + ], + "spans": [ + { + "bbox": [ + 160, + 283, + 450, + 301 + ], + "score": 1.0, + "content": "Figure 8: IWSLT with 2-layer decoder using different freezing strategy.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 7.25 + }, + { + "type": "image", + "bbox": [ + 117, + 354, + 486, + 483 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 354, + 486, + 483 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 117, + 354, + 486, + 483 + ], + "spans": [ + { + "bbox": [ + 117, + 354, + 486, + 483 + ], + "score": 0.968, + "type": "image", + "image_path": "8efb8273099ff74135796331759c0532b0ccf7448a7b83e8c437cb3b5dc2a6fb.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 117, + 354, + 486, + 397.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 117, + 397.0, + 486, + 440.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 117, + 440.0, + 486, + 483.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 497, + 505, + 520 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 495, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 510 + ], + "score": 1.0, + "content": "Figure 9: RoBERTa Reservoir Results, Pre-training versus downstream task plot for 12 layer", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 505, + 274, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 274, + 521 + ], + "score": 1.0, + "content": "RoBERTa. MNLI-m (left). SST-2 (right).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + } + ], + "index": 13.25 + }, + { + "type": "image", + "bbox": [ + 124, + 573, + 490, + 678 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 573, + 490, + 678 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 124, + 573, + 490, + 678 + ], + "spans": [ + { + "bbox": [ + 124, + 573, + 490, + 678 + ], + "score": 0.942, + "type": "image", + "image_path": "16c67d74e5743eb685c5b29836cc09d4540ceae3fe880d614d534330288bb088.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 124, + 573, + 490, + 608.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 124, + 608.0, + 490, + 643.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 124, + 643.0, + 490, + 678.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 689, + 507, + 712 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "Figure 10: RoBERTa Reservoir Results, Training plot for 12 layer RoBERTa (left). AUCC result", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 699, + 138, + 714 + ], + "spans": [ + { + "bbox": [ + 104, + 699, + 138, + 714 + ], + "score": 1.0, + "content": "(right).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + } + ], + "index": 18.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 111, + 486, + 240 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 111, + 486, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 111, + 486, + 240 + ], + "spans": [ + { + "bbox": [ + 112, + 111, + 486, + 240 + ], + "score": 0.962, + "type": "image", + "image_path": "e38bef769826f1e91e25614fc341e46647bcb6ea7bc3f2a9aafc23ea160c51d5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 111, + 486, + 154.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 154.0, + 486, + 197.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 197.0, + 486, + 240.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 253, + 506, + 277 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "Figure 11: Validation BLEU AUCC and test BLEU for IWSLT (high is good). Comparison of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 264, + 486, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 486, + 277 + ], + "score": 1.0, + "content": "regular transformer and reservoir transformer with FFN or Transformer reservoir layers added.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 111, + 330, + 487, + 459 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 330, + 487, + 459 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 111, + 330, + 487, + 459 + ], + "spans": [ + { + "bbox": [ + 111, + 330, + 487, + 459 + ], + "score": 0.721, + "type": "image", + "image_path": "a79c4059de5d7c1e244e50c3037c5c3ba9d32882a35b29d942cc4721a4bbab54.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 111, + 330, + 487, + 373.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 111, + 373.0, + 487, + 416.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 416.0, + 487, + 459.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 472, + 505, + 496 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 472, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 505, + 485 + ], + "score": 1.0, + "content": "Figure 12: Validation BLEU AUCC and test BLEU for WMT (high is good). Comparison of regular", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 483, + 454, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 454, + 495 + ], + "score": 1.0, + "content": "transformer and reservoir transformer with FFN or Transformer reservoir layers added.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "image", + "bbox": [ + 114, + 547, + 486, + 677 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 547, + 486, + 677 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 114, + 547, + 486, + 677 + ], + "spans": [ + { + "bbox": [ + 114, + 547, + 486, + 677 + ], + "score": 0.895, + "type": "image", + "image_path": "341111a25bde86caccbfa885e06e66acf9b016afdcaeecdf8630ea4dc0291102.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 114, + 547, + 486, + 590.3333333333334 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 114, + 590.3333333333334, + 486, + 633.6666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 114, + 633.6666666666667, + 486, + 677.0000000000001 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 691, + 506, + 714 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 690, + 506, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 506, + 705 + ], + "score": 1.0, + "content": "Figure 13: Validation BPC AUCC and test BPC on the enwik8 language modelling task (low is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 702, + 413, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 413, + 715 + ], + "score": 1.0, + "content": "good). Comparison of regular and reservoir transformers for varying depths.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + } + ], + "index": 12.25 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 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 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 111, + 486, + 240 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 111, + 486, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 111, + 486, + 240 + ], + "spans": [ + { + "bbox": [ + 112, + 111, + 486, + 240 + ], + "score": 0.962, + "type": "image", + "image_path": "e38bef769826f1e91e25614fc341e46647bcb6ea7bc3f2a9aafc23ea160c51d5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 111, + 486, + 154.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 154.0, + 486, + 197.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 197.0, + 486, + 240.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 253, + 506, + 277 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "Figure 11: Validation BLEU AUCC and test BLEU for IWSLT (high is good). Comparison of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 264, + 486, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 486, + 277 + ], + "score": 1.0, + "content": "regular transformer and reservoir transformer with FFN or Transformer reservoir layers added.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 111, + 330, + 487, + 459 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 330, + 487, + 459 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 111, + 330, + 487, + 459 + ], + "spans": [ + { + "bbox": [ + 111, + 330, + 487, + 459 + ], + "score": 0.721, + "type": "image", + "image_path": "a79c4059de5d7c1e244e50c3037c5c3ba9d32882a35b29d942cc4721a4bbab54.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 111, + 330, + 487, + 373.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 111, + 373.0, + 487, + 416.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 416.0, + 487, + 459.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 472, + 505, + 496 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 472, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 505, + 485 + ], + "score": 1.0, + "content": "Figure 12: Validation BLEU AUCC and test BLEU for WMT (high is good). Comparison of regular", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 483, + 454, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 454, + 495 + ], + "score": 1.0, + "content": "transformer and reservoir transformer with FFN or Transformer reservoir layers added.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "image", + "bbox": [ + 114, + 547, + 486, + 677 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 547, + 486, + 677 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 114, + 547, + 486, + 677 + ], + "spans": [ + { + "bbox": [ + 114, + 547, + 486, + 677 + ], + "score": 0.895, + "type": "image", + "image_path": "341111a25bde86caccbfa885e06e66acf9b016afdcaeecdf8630ea4dc0291102.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 114, + 547, + 486, + 590.3333333333334 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 114, + 590.3333333333334, + 486, + 633.6666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 114, + 633.6666666666667, + 486, + 677.0000000000001 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 691, + 506, + 714 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 690, + 506, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 506, + 705 + ], + "score": 1.0, + "content": "Figure 13: Validation BPC AUCC and test BPC on the enwik8 language modelling task (low is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 702, + 413, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 413, + 715 + ], + "score": 1.0, + "content": "good). Comparison of regular and reservoir transformers for varying depths.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + } + ], + "index": 12.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 101, + 486, + 228 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 101, + 486, + 228 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 101, + 486, + 228 + ], + "spans": [ + { + "bbox": [ + 117, + 101, + 486, + 228 + ], + "score": 0.969, + "type": "image", + "image_path": "dbab4e8809ef2a132617f00c1dc1acd380a0973768e0196c29742da904ec4002.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 101, + 486, + 143.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 143.33333333333334, + 486, + 185.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 185.66666666666669, + 486, + 228.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 112, + 242, + 495, + 254 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 241, + 496, + 256 + ], + "spans": [ + { + "bbox": [ + 114, + 241, + 496, + 256 + ], + "score": 1.0, + "content": "Figure 14: Downstream RoBERTa performance on SST-2 (left) and MultiNLI-matched (right).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 122, + 280, + 487, + 469 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 280, + 487, + 469 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 122, + 280, + 487, + 469 + ], + "spans": [ + { + "bbox": [ + 122, + 280, + 487, + 469 + ], + "score": 0.973, + "type": "image", + "image_path": "41ebdc6849103e9713e0791384d38890f3c671a95f174869f62c2c94346167cf.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 122, + 280, + 487, + 343.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 122, + 343.0, + 487, + 406.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 122, + 406.0, + 487, + 469.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 479, + 504, + 513 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Figure 15: IWSLT with 2-layer decoder validation plot (upper left). WMT with 24-layer decoder", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 490, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 502 + ], + "score": 1.0, + "content": "validation plot (upper right). Enwik8 with 48-layer decoder validation plot (lower left). RoBERTa", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 501, + 311, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 311, + 514 + ], + "score": 1.0, + "content": "with 12-layer decoder validation plot (lower right).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "table", + "bbox": [ + 108, + 528, + 505, + 704 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 528, + 505, + 704 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 528, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 108, + 528, + 505, + 704 + ], + "score": 0.982, + "html": "
ModelLayerSentLen (Surface)TreeDepth (Syntactic)TopConst (Syntactic)BShift (Syntactic)Tense (Semantic)SubjNum (Semantic)ObjNum (Semantic)SOMO (Semantic)CoordInv (Semantic)
Transformer84.56 ± 0.5432.30 ± 0.4154.40 ± 0.3349.99 ± 0.0180.98 ± 0.3276.26 ± 0.0950.01 ± 0.1976.38 ± 0.6154.33 ± 0.47
87.22 ± 0.0733.63 ± 0.5758.38 ± 0.2050.12 ± 0.1782.84 ± 0.6878.65 ± 0.1951.47 ± 0.5378.00 ± 1.1254.66 ± 0.55
84.25 ± 0.1632.60 ± 0.1754.41 ± 0.1050.02 ± 0.0181.72 ± 0.5977.00 ± 0.1351.32 ± 0.6476.57 ± 1.1354.13 ± 0.51
87.37 ± 0.2032.59 ± 0.2950.06 ± 0.2169.76 ± 0.2681.63 ± 1.1776.47 ± 0.0952.41 ± 1.4976.15 ± 0.8452.62 ± 1.34
584.61 ± 0.2431.14 ± 0.4844.76 ± 0.3874.82 ± 0.1180.16 ± 0.1973.66 ± 0.1652.95 ± 1.7772.90 ± 0.2151.26 ± 1.14
682.56 ± 0.2530.31 ± 0.4039.30 ± 0.4078.80 ±0.3881.88 ± 0.4775.30 ± 0.0756.21 ± 1.2674.37 ± 0.1651.44 ± 1.04
70.85 ± 0.1326.65 ± 0.7240.70 ± 0.1378.98 ± 0.3285.11 ± 0.3172.03 ± 0.4658.15 ± 0.4668.71 ± 0.9155.39 ± 0.27
866.23 ± 1.3323.46 ± 0.4425.19 ± 1.0277.42 ± 0.2780.35 ± 0.4567.55 ± 0.9954.94 ± 2.0463.69 ± 2.3250.58 ± 0.83
971.17 ± 0.2931.21 ± 0.3158.42 ± 0.2985.55 ± 0.4486.77 ± 0.1980.30 ± 0.0864.36 ± 1.2081.68 ± 0.4566.90 ± 0.49
1073.19 ± 0.5027.74 ± 0.5341.01 ± 0.2283.56 ± 0.9686.13 ± 0.3583.04 ± 0.0462.01 ± 0.5979.73 ± 0.2162.60 ± 1.04
1171.37 ± 0.4230.22 ± 0.2848.58 ± 0.3584.40 ± 0.4487.28 ± 0.5982.34 ± 0.1561.10 ± 0.1480.00 ± 0.4064.44 ± 0.38
1271.66 ± 0.1233.43 ± 0.1864.38 ± 0.2087.38 ± 0.0288.41 ± 0.0984.46 ± 0.2563.01 ± 0.0581.80 ± 0.2765.72 ± 0.16
87.75 ± 0.1031.60 ± 0.2150.38 ± 0.2350.00 ± 0.0080.40 ± 0.1876.47 ± 0.2050.53 ± 0.1473.48 ± 0.15
TReservoir281.28 ± 0.2334.20 ± 0.4160.64 ± 0.6553.55 ± 0.70
61.41 ± 0.4281.50 ± 0.7776.33 ± 0.0850.73 ± 0.3474.28 ± 0.6756.82 ± 0.10
389.28 ± 0.0936.42 ± 0.1167.36 ± 0.4575.64 ± 0.5285.42 ± 0.1880.53 ± 0.0252.50 ± 1.8078.47 ± 1.8157.16 ± 0.27
74.31 ± 0.3232.42 ± 0.8355.19 ± 0.3373.41 ± 0.0079.56 ± 0.0075.15 ± 0.0853.68 ± 0.6675.02 ± 0.1956.89 ± 0.08
88.03 ± 0.22 74.55 ± 0.3738.34 ± 0.64 33.13 ± 0.2968.65 ± 0.2982.25 ± 0.1286.80 ± 0.0282.27 ± 0.3357.95 ± 0.2480.82 ± 0.9158.05 ± 0.10
52.70 ± 0.8179.21 ± 0.1385.70 ± 0.3677.43 ± 0.0357.26 ± 0.1975.38 ± 0.6651.95 ± 1.30
85.82 ± 0.37 71.69 ± 0.7137.63 ± 0.13 30.32 ± 0.0170.43 ± 0.05 48.44 ± 0.3084.12 ± 0.35 79.12 ± 0.1286.88 ± 0.0782.86 ± 0.3061.17 ± 0.2180.79 ± 0.1761.83 ± 0.95
8 985.86 ± 0.1237.89 ± 0.0369.53 ± 0.3785.55 ± 0.1284.75 ± 0.09 87.98 ± 0.2279.23 ± 0.11 84.13 ± 0.0159.53 ± 0.1676.80 ± 0.4157.34 ± 0.14
69.22 ± 0.2325.58 ± 0.3529.20 ± 0.5878.57 ± 0.0963.06± 0.0182.55 ± 0.3166.07 ± 0.05
1047.56 ± 0.0285.02 ± 0.0375.68 ± 0.1657.55 ± 1.5774.70 ± 0.0255.02 ± 0.64
1165.70 ± 0.0530.57 ± 0.0381.20 ± 0.0086.78 ± 0.0283.73 ± 0.0560.38 ± 0.1780.59 ± 0.1562.50 ± 0.11
1270.61 ± 0.1834.45± 0.2064.19 ± 0.1084.53 ± 0.0387.48 ± 0.1684.86 ± 0.1462.75 ± 0.1482.08 ± 0.0364.73 ± 0.06
", + "type": "table", + "image_path": "1a0ada09fc6e820280ef99ea702bea624120a34c1aa4e724a8efcb050a9688cd.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 528, + 505, + 586.6666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 586.6666666666666, + 505, + 645.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 645.3333333333333, + 505, + 703.9999999999999 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 713, + 503, + 725 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 712, + 505, + 726 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 505, + 726 + ], + "score": 1.0, + "content": "Table 6: RoBERTa Probing Results. The line in bold text are the the frozen layers in the T Reservoir.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + } + ], + "index": 12.0 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 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 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 101, + 486, + 228 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 101, + 486, + 228 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 101, + 486, + 228 + ], + "spans": [ + { + "bbox": [ + 117, + 101, + 486, + 228 + ], + "score": 0.969, + "type": "image", + "image_path": "dbab4e8809ef2a132617f00c1dc1acd380a0973768e0196c29742da904ec4002.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 101, + 486, + 143.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 143.33333333333334, + 486, + 185.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 185.66666666666669, + 486, + 228.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 112, + 242, + 495, + 254 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 241, + 496, + 256 + ], + "spans": [ + { + "bbox": [ + 114, + 241, + 496, + 256 + ], + "score": 1.0, + "content": "Figure 14: Downstream RoBERTa performance on SST-2 (left) and MultiNLI-matched (right).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 122, + 280, + 487, + 469 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 280, + 487, + 469 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 122, + 280, + 487, + 469 + ], + "spans": [ + { + "bbox": [ + 122, + 280, + 487, + 469 + ], + "score": 0.973, + "type": "image", + "image_path": "41ebdc6849103e9713e0791384d38890f3c671a95f174869f62c2c94346167cf.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 122, + 280, + 487, + 343.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 122, + 343.0, + 487, + 406.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 122, + 406.0, + 487, + 469.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 479, + 504, + 513 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Figure 15: IWSLT with 2-layer decoder validation plot (upper left). WMT with 24-layer decoder", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 490, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 502 + ], + "score": 1.0, + "content": "validation plot (upper right). Enwik8 with 48-layer decoder validation plot (lower left). RoBERTa", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 501, + 311, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 311, + 514 + ], + "score": 1.0, + "content": "with 12-layer decoder validation plot (lower right).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "table", + "bbox": [ + 108, + 528, + 505, + 704 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 528, + 505, + 704 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 528, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 108, + 528, + 505, + 704 + ], + "score": 0.982, + "html": "
ModelLayerSentLen (Surface)TreeDepth (Syntactic)TopConst (Syntactic)BShift (Syntactic)Tense (Semantic)SubjNum (Semantic)ObjNum (Semantic)SOMO (Semantic)CoordInv (Semantic)
Transformer84.56 ± 0.5432.30 ± 0.4154.40 ± 0.3349.99 ± 0.0180.98 ± 0.3276.26 ± 0.0950.01 ± 0.1976.38 ± 0.6154.33 ± 0.47
87.22 ± 0.0733.63 ± 0.5758.38 ± 0.2050.12 ± 0.1782.84 ± 0.6878.65 ± 0.1951.47 ± 0.5378.00 ± 1.1254.66 ± 0.55
84.25 ± 0.1632.60 ± 0.1754.41 ± 0.1050.02 ± 0.0181.72 ± 0.5977.00 ± 0.1351.32 ± 0.6476.57 ± 1.1354.13 ± 0.51
87.37 ± 0.2032.59 ± 0.2950.06 ± 0.2169.76 ± 0.2681.63 ± 1.1776.47 ± 0.0952.41 ± 1.4976.15 ± 0.8452.62 ± 1.34
584.61 ± 0.2431.14 ± 0.4844.76 ± 0.3874.82 ± 0.1180.16 ± 0.1973.66 ± 0.1652.95 ± 1.7772.90 ± 0.2151.26 ± 1.14
682.56 ± 0.2530.31 ± 0.4039.30 ± 0.4078.80 ±0.3881.88 ± 0.4775.30 ± 0.0756.21 ± 1.2674.37 ± 0.1651.44 ± 1.04
70.85 ± 0.1326.65 ± 0.7240.70 ± 0.1378.98 ± 0.3285.11 ± 0.3172.03 ± 0.4658.15 ± 0.4668.71 ± 0.9155.39 ± 0.27
866.23 ± 1.3323.46 ± 0.4425.19 ± 1.0277.42 ± 0.2780.35 ± 0.4567.55 ± 0.9954.94 ± 2.0463.69 ± 2.3250.58 ± 0.83
971.17 ± 0.2931.21 ± 0.3158.42 ± 0.2985.55 ± 0.4486.77 ± 0.1980.30 ± 0.0864.36 ± 1.2081.68 ± 0.4566.90 ± 0.49
1073.19 ± 0.5027.74 ± 0.5341.01 ± 0.2283.56 ± 0.9686.13 ± 0.3583.04 ± 0.0462.01 ± 0.5979.73 ± 0.2162.60 ± 1.04
1171.37 ± 0.4230.22 ± 0.2848.58 ± 0.3584.40 ± 0.4487.28 ± 0.5982.34 ± 0.1561.10 ± 0.1480.00 ± 0.4064.44 ± 0.38
1271.66 ± 0.1233.43 ± 0.1864.38 ± 0.2087.38 ± 0.0288.41 ± 0.0984.46 ± 0.2563.01 ± 0.0581.80 ± 0.2765.72 ± 0.16
87.75 ± 0.1031.60 ± 0.2150.38 ± 0.2350.00 ± 0.0080.40 ± 0.1876.47 ± 0.2050.53 ± 0.1473.48 ± 0.15
TReservoir281.28 ± 0.2334.20 ± 0.4160.64 ± 0.6553.55 ± 0.70
61.41 ± 0.4281.50 ± 0.7776.33 ± 0.0850.73 ± 0.3474.28 ± 0.6756.82 ± 0.10
389.28 ± 0.0936.42 ± 0.1167.36 ± 0.4575.64 ± 0.5285.42 ± 0.1880.53 ± 0.0252.50 ± 1.8078.47 ± 1.8157.16 ± 0.27
74.31 ± 0.3232.42 ± 0.8355.19 ± 0.3373.41 ± 0.0079.56 ± 0.0075.15 ± 0.0853.68 ± 0.6675.02 ± 0.1956.89 ± 0.08
88.03 ± 0.22 74.55 ± 0.3738.34 ± 0.64 33.13 ± 0.2968.65 ± 0.2982.25 ± 0.1286.80 ± 0.0282.27 ± 0.3357.95 ± 0.2480.82 ± 0.9158.05 ± 0.10
52.70 ± 0.8179.21 ± 0.1385.70 ± 0.3677.43 ± 0.0357.26 ± 0.1975.38 ± 0.6651.95 ± 1.30
85.82 ± 0.37 71.69 ± 0.7137.63 ± 0.13 30.32 ± 0.0170.43 ± 0.05 48.44 ± 0.3084.12 ± 0.35 79.12 ± 0.1286.88 ± 0.0782.86 ± 0.3061.17 ± 0.2180.79 ± 0.1761.83 ± 0.95
8 985.86 ± 0.1237.89 ± 0.0369.53 ± 0.3785.55 ± 0.1284.75 ± 0.09 87.98 ± 0.2279.23 ± 0.11 84.13 ± 0.0159.53 ± 0.1676.80 ± 0.4157.34 ± 0.14
69.22 ± 0.2325.58 ± 0.3529.20 ± 0.5878.57 ± 0.0963.06± 0.0182.55 ± 0.3166.07 ± 0.05
1047.56 ± 0.0285.02 ± 0.0375.68 ± 0.1657.55 ± 1.5774.70 ± 0.0255.02 ± 0.64
1165.70 ± 0.0530.57 ± 0.0381.20 ± 0.0086.78 ± 0.0283.73 ± 0.0560.38 ± 0.1780.59 ± 0.1562.50 ± 0.11
1270.61 ± 0.1834.45± 0.2064.19 ± 0.1084.53 ± 0.0387.48 ± 0.1684.86 ± 0.1462.75 ± 0.1482.08 ± 0.0364.73 ± 0.06
", + "type": "table", + "image_path": "1a0ada09fc6e820280ef99ea702bea624120a34c1aa4e724a8efcb050a9688cd.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 528, + 505, + 586.6666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 586.6666666666666, + 505, + 645.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 645.3333333333333, + 505, + 703.9999999999999 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 713, + 503, + 725 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 712, + 505, + 726 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 505, + 726 + ], + "score": 1.0, + "content": "Table 6: RoBERTa Probing Results. The line in bold text are the the frozen layers in the T Reservoir.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + } + ], + "index": 12.0 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file