{ "pdf_info": [ { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 78, 450, 116 ], "lines": [ { "bbox": [ 105, 78, 335, 96 ], "spans": [ { "bbox": [ 105, 78, 335, 96 ], "score": 1.0, "content": "TOWARDS A UNIFIED VIEW OF", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 97, 451, 117 ], "spans": [ { "bbox": [ 105, 97, 451, 117 ], "score": 1.0, "content": "PARAMETER-EFFICIENT TRANSFER LEARNING", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 113, 135, 227, 168 ], "lines": [ { "bbox": [ 111, 135, 167, 146 ], "spans": [ { "bbox": [ 111, 135, 149, 146 ], "score": 1.0, "content": "Junxian", "type": "text" }, { "bbox": [ 149, 135, 167, 145 ], "score": 0.36, "content": "\\mathbf { H e } ^ { * }", "type": "inline_equation" } ], "index": 2 }, { "bbox": [ 112, 145, 226, 158 ], "spans": [ { "bbox": [ 112, 145, 226, 158 ], "score": 1.0, "content": "Carnegie Mellon University", "type": "text" } ], "index": 4 }, { "bbox": [ 113, 158, 228, 169 ], "spans": [ { "bbox": [ 113, 158, 228, 169 ], "score": 1.0, "content": "junxianh@cs.cmu.edu", "type": "text" } ], "index": 6 } ], "index": 4 }, { "type": "text", "bbox": [ 383, 135, 498, 168 ], "lines": [ { "bbox": [ 382, 135, 455, 147 ], "spans": [ { "bbox": [ 382, 135, 455, 147 ], "score": 1.0, "content": "Chunting Zhou∗", "type": "text" } ], "index": 3 }, { "bbox": [ 382, 144, 497, 159 ], "spans": [ { "bbox": [ 382, 144, 497, 159 ], "score": 1.0, "content": "Carnegie Mellon University", "type": "text" } ], "index": 5 }, { "bbox": [ 383, 156, 499, 169 ], "spans": [ { "bbox": [ 383, 156, 499, 169 ], "score": 1.0, "content": "chuntinz@cs.cmu.edu", "type": "text" } ], "index": 7 } ], "index": 5 }, { "type": "text", "bbox": [ 113, 185, 249, 218 ], "lines": [ { "bbox": [ 112, 185, 164, 197 ], "spans": [ { "bbox": [ 112, 185, 164, 197 ], "score": 1.0, "content": "Xuezhe Ma", "type": "text" } ], "index": 8 }, { "bbox": [ 112, 196, 249, 207 ], "spans": [ { "bbox": [ 112, 196, 249, 207 ], "score": 1.0, "content": "University of Southern California", "type": "text" } ], "index": 9 }, { "bbox": [ 112, 208, 210, 218 ], "spans": [ { "bbox": [ 112, 208, 210, 218 ], "score": 1.0, "content": "xuezhema@isi.edu", "type": "text" } ], "index": 14 } ], "index": 9 }, { "type": "text", "bbox": [ 261, 185, 369, 218 ], "lines": [ { "bbox": [ 260, 184, 368, 198 ], "spans": [ { "bbox": [ 260, 184, 368, 198 ], "score": 1.0, "content": "Taylor Berg-Kirkpatrick", "type": "text" } ], "index": 11 }, { "bbox": [ 260, 194, 322, 209 ], "spans": [ { "bbox": [ 260, 194, 322, 209 ], "score": 1.0, "content": "UC San Diego", "type": "text" } ], "index": 12 }, { "bbox": [ 260, 207, 371, 220 ], "spans": [ { "bbox": [ 260, 207, 371, 220 ], "score": 1.0, "content": "tberg@eng.ucsd.edu", "type": "text" } ], "index": 15 } ], "index": 12 }, { "type": "text", "bbox": [ 382, 185, 495, 218 ], "lines": [ { "bbox": [ 382, 183, 455, 199 ], "spans": [ { "bbox": [ 382, 183, 455, 199 ], "score": 1.0, "content": "Graham Neubig", "type": "text" } ], "index": 10 }, { "bbox": [ 382, 195, 496, 209 ], "spans": [ { "bbox": [ 382, 195, 496, 209 ], "score": 1.0, "content": "Carnegie Mellon University", "type": "text" } ], "index": 13 }, { "bbox": [ 383, 208, 492, 220 ], "spans": [ { "bbox": [ 383, 208, 492, 220 ], "score": 1.0, "content": "gneubig@cs.cmu.edu", "type": "text" } ], "index": 16 } ], "index": 13 }, { "type": "title", "bbox": [ 278, 248, 333, 259 ], "lines": [ { "bbox": [ 276, 246, 336, 262 ], "spans": [ { "bbox": [ 276, 246, 336, 262 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 143, 273, 468, 493 ], "lines": [ { "bbox": [ 141, 273, 470, 286 ], "spans": [ { "bbox": [ 141, 273, 470, 286 ], "score": 1.0, "content": "Fine-tuning large pretrained language models on downstream tasks has become", "type": "text" } ], "index": 18 }, { "bbox": [ 141, 284, 469, 297 ], "spans": [ { "bbox": [ 141, 284, 469, 297 ], "score": 1.0, "content": "the de-facto learning paradigm in NLP. However, conventional approaches fine-", "type": "text" } ], "index": 19 }, { "bbox": [ 141, 295, 469, 308 ], "spans": [ { "bbox": [ 141, 295, 469, 308 ], "score": 1.0, "content": "tune all the parameters of the pretrained model, which becomes prohibitive as", "type": "text" } ], "index": 20 }, { "bbox": [ 142, 307, 469, 318 ], "spans": [ { "bbox": [ 142, 307, 469, 318 ], "score": 1.0, "content": "the model size and the number of tasks grow. Recent work has proposed a va-", "type": "text" } ], "index": 21 }, { "bbox": [ 141, 317, 470, 329 ], "spans": [ { "bbox": [ 141, 317, 470, 329 ], "score": 1.0, "content": "riety of parameter-efficient transfer learning methods that only fine-tune a small", "type": "text" } ], "index": 22 }, { "bbox": [ 141, 328, 470, 340 ], "spans": [ { "bbox": [ 141, 328, 470, 340 ], "score": 1.0, "content": "number of (extra) parameters to attain strong performance. While effective, the", "type": "text" } ], "index": 23 }, { "bbox": [ 142, 339, 469, 351 ], "spans": [ { "bbox": [ 142, 339, 469, 351 ], "score": 1.0, "content": "critical ingredients for success and the connections among the various methods", "type": "text" } ], "index": 24 }, { "bbox": [ 141, 351, 469, 362 ], "spans": [ { "bbox": [ 141, 351, 469, 362 ], "score": 1.0, "content": "are poorly understood. In this paper, we break down the design of state-of-the-art", "type": "text" } ], "index": 25 }, { "bbox": [ 141, 361, 470, 372 ], "spans": [ { "bbox": [ 141, 361, 470, 372 ], "score": 1.0, "content": "parameter-efficient transfer learning methods and present a unified framework that", "type": "text" } ], "index": 26 }, { "bbox": [ 141, 372, 469, 384 ], "spans": [ { "bbox": [ 141, 372, 469, 384 ], "score": 1.0, "content": "establishes connections between them. Specifically, we re-frame them as modifi-", "type": "text" } ], "index": 27 }, { "bbox": [ 141, 382, 470, 396 ], "spans": [ { "bbox": [ 141, 382, 470, 396 ], "score": 1.0, "content": "cations to specific hidden states in pretrained models, and define a set of design", "type": "text" } ], "index": 28 }, { "bbox": [ 141, 393, 470, 407 ], "spans": [ { "bbox": [ 141, 393, 470, 407 ], "score": 1.0, "content": "dimensions along which different methods vary, such as the function to compute", "type": "text" } ], "index": 29 }, { "bbox": [ 141, 405, 469, 417 ], "spans": [ { "bbox": [ 141, 405, 469, 417 ], "score": 1.0, "content": "the modification and the position to apply the modification. Through comprehen-", "type": "text" } ], "index": 30 }, { "bbox": [ 141, 415, 469, 428 ], "spans": [ { "bbox": [ 141, 415, 469, 428 ], "score": 1.0, "content": "sive empirical studies across machine translation, text summarization, language", "type": "text" } ], "index": 31 }, { "bbox": [ 142, 427, 469, 438 ], "spans": [ { "bbox": [ 142, 427, 469, 438 ], "score": 1.0, "content": "understanding, and text classification benchmarks, we utilize the unified view to", "type": "text" } ], "index": 32 }, { "bbox": [ 141, 438, 470, 449 ], "spans": [ { "bbox": [ 141, 438, 470, 449 ], "score": 1.0, "content": "identify important design choices in previous methods. Furthermore, our unified", "type": "text" } ], "index": 33 }, { "bbox": [ 141, 448, 469, 461 ], "spans": [ { "bbox": [ 141, 448, 469, 461 ], "score": 1.0, "content": "framework enables the transfer of design elements across different approaches,", "type": "text" } ], "index": 34 }, { "bbox": [ 141, 459, 470, 472 ], "spans": [ { "bbox": [ 141, 459, 470, 472 ], "score": 1.0, "content": "and as a result we are able to instantiate new parameter-efficient fine-tuning meth-", "type": "text" } ], "index": 35 }, { "bbox": [ 141, 470, 470, 484 ], "spans": [ { "bbox": [ 141, 470, 470, 484 ], "score": 1.0, "content": "ods that tune less parameters than previous methods while being more effective,", "type": "text" } ], "index": 36 }, { "bbox": [ 141, 482, 448, 493 ], "spans": [ { "bbox": [ 141, 482, 448, 493 ], "score": 1.0, "content": "achieving comparable results to fine-tuning all parameters on all four tasks.1", "type": "text" } ], "index": 37 } ], "index": 27.5 }, { "type": "title", "bbox": [ 108, 515, 206, 528 ], "lines": [ { "bbox": [ 105, 514, 208, 531 ], "spans": [ { "bbox": [ 105, 514, 208, 531 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 38 } ], "index": 38 }, { "type": "text", "bbox": [ 107, 541, 505, 630 ], "lines": [ { "bbox": [ 105, 541, 505, 555 ], "spans": [ { "bbox": [ 105, 541, 505, 555 ], "score": 1.0, "content": "Transfer learning from pre-trained language models (PLMs) is now the prevalent paradigm in natural", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 553, 505, 564 ], "spans": [ { "bbox": [ 105, 553, 505, 564 ], "score": 1.0, "content": "language processing, yielding strong performance on many tasks (Peters et al., 2018; Devlin et al.,", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 563, 505, 576 ], "spans": [ { "bbox": [ 105, 563, 505, 576 ], "score": 1.0, "content": "2019; Qiu et al., 2020). The most common way to adapt general-purpose PLMs to downstream tasks", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 574, 505, 587 ], "spans": [ { "bbox": [ 105, 574, 505, 587 ], "score": 1.0, "content": "is to fine-tune all the model parameters (full fine-tuning). However, this results in a separate copy of", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 586, 505, 598 ], "spans": [ { "bbox": [ 105, 586, 505, 598 ], "score": 1.0, "content": "fine-tuned model parameters for each task, which is prohibitively expensive when serving models", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 597, 505, 609 ], "spans": [ { "bbox": [ 105, 597, 505, 609 ], "score": 1.0, "content": "that perform a large number of tasks. This issue is particularly salient with the ever-increasing size", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 607, 505, 619 ], "spans": [ { "bbox": [ 105, 607, 505, 619 ], "score": 1.0, "content": "of PLMs, which now range from hundreds of millions (Radford et al., 2019; Lewis et al., 2020) to", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 618, 476, 630 ], "spans": [ { "bbox": [ 105, 618, 476, 630 ], "score": 1.0, "content": "hundreds of billions (Brown et al., 2020) or even trillions of parameters (Fedus et al., 2021).", "type": "text" } ], "index": 46 } ], "index": 42.5 }, { "type": "text", "bbox": [ 107, 635, 504, 702 ], "lines": [ { "bbox": [ 106, 635, 505, 648 ], "spans": [ { "bbox": [ 106, 635, 505, 648 ], "score": 1.0, "content": "To mitigate this issue, a few lightweight alternatives have been proposed to update only a small num-", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 645, 505, 659 ], "spans": [ { "bbox": [ 105, 645, 505, 659 ], "score": 1.0, "content": "ber of extra parameters while keeping most pretrained parameters frozen. For example, adapter tun-", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 657, 506, 670 ], "spans": [ { "bbox": [ 105, 657, 506, 670 ], "score": 1.0, "content": "ing (Houlsby et al., 2019) inserts small neural modules called adapters to each layer of the pretrained", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 667, 506, 681 ], "spans": [ { "bbox": [ 105, 667, 506, 681 ], "score": 1.0, "content": "network and only the adapters are trained at fine-tuning time. Inspired by the success of prompting", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 678, 506, 692 ], "spans": [ { "bbox": [ 105, 678, 506, 692 ], "score": 1.0, "content": "methods that control PLMs through textual prompts (Brown et al., 2020; Liu et al., 2021a), prefix", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 690, 505, 702 ], "spans": [ { "bbox": [ 105, 690, 467, 702 ], "score": 1.0, "content": "tuning (Li & Liang, 2021) and prompt tuning (Lester et al., 2021) prepend an additional", "type": "text" }, { "bbox": [ 467, 691, 471, 700 ], "score": 0.37, "content": "l", "type": "inline_equation" }, { "bbox": [ 472, 690, 505, 702 ], "score": 1.0, "content": "tunable", "type": "text" } ], "index": 52 } ], "index": 49.5 } ], "page_idx": 0, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 118, 711, 401, 732 ], "lines": [ { "bbox": [ 118, 708, 349, 723 ], "spans": [ { "bbox": [ 118, 708, 349, 723 ], "score": 1.0, "content": "∗Equal Contribution. Order determined by random dice rolling.", "type": "text" } ] }, { "bbox": [ 118, 720, 401, 733 ], "spans": [ { "bbox": [ 118, 720, 401, 733 ], "score": 1.0, "content": "1Code is available at https://github.com/jxhe/unify-parameter-efficient-tuning.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "1", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 78, 450, 116 ], "lines": [ { "bbox": [ 105, 78, 335, 96 ], "spans": [ { "bbox": [ 105, 78, 335, 96 ], "score": 1.0, "content": "TOWARDS A UNIFIED VIEW OF", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 97, 451, 117 ], "spans": [ { "bbox": [ 105, 97, 451, 117 ], "score": 1.0, "content": "PARAMETER-EFFICIENT TRANSFER LEARNING", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 113, 135, 227, 168 ], "lines": [ { "bbox": [ 111, 135, 167, 146 ], "spans": [ { "bbox": [ 111, 135, 149, 146 ], "score": 1.0, "content": "Junxian", "type": "text" }, { "bbox": [ 149, 135, 167, 145 ], "score": 0.36, "content": "\\mathbf { H e } ^ { * }", "type": "inline_equation" } ], "index": 2 }, { "bbox": [ 112, 145, 226, 158 ], "spans": [ { "bbox": [ 112, 145, 226, 158 ], "score": 1.0, "content": "Carnegie Mellon University", "type": "text" } ], "index": 4 }, { "bbox": [ 113, 158, 228, 169 ], "spans": [ { "bbox": [ 113, 158, 228, 169 ], "score": 1.0, "content": "junxianh@cs.cmu.edu", "type": "text" } ], "index": 6 } ], "index": 4, "bbox_fs": [ 111, 135, 228, 169 ] }, { "type": "text", "bbox": [ 383, 135, 498, 168 ], "lines": [ { "bbox": [ 382, 135, 455, 147 ], "spans": [ { "bbox": [ 382, 135, 455, 147 ], "score": 1.0, "content": "Chunting Zhou∗", "type": "text" } ], "index": 3 }, { "bbox": [ 382, 144, 497, 159 ], "spans": [ { "bbox": [ 382, 144, 497, 159 ], "score": 1.0, "content": "Carnegie Mellon University", "type": "text" } ], "index": 5 }, { "bbox": [ 383, 156, 499, 169 ], "spans": [ { "bbox": [ 383, 156, 499, 169 ], "score": 1.0, "content": "chuntinz@cs.cmu.edu", "type": "text" } ], "index": 7 } ], "index": 5, "bbox_fs": [ 382, 135, 499, 169 ] }, { "type": "text", "bbox": [ 113, 185, 249, 218 ], "lines": [ { "bbox": [ 112, 185, 164, 197 ], "spans": [ { "bbox": [ 112, 185, 164, 197 ], "score": 1.0, "content": "Xuezhe Ma", "type": "text" } ], "index": 8 }, { "bbox": [ 112, 196, 249, 207 ], "spans": [ { "bbox": [ 112, 196, 249, 207 ], "score": 1.0, "content": "University of Southern California", "type": "text" } ], "index": 9 }, { "bbox": [ 112, 208, 210, 218 ], "spans": [ { "bbox": [ 112, 208, 210, 218 ], "score": 1.0, "content": "xuezhema@isi.edu", "type": "text" } ], "index": 14 } ], "index": 9, "bbox_fs": [ 112, 185, 249, 218 ] }, { "type": "text", "bbox": [ 261, 185, 369, 218 ], "lines": [ { "bbox": [ 260, 184, 368, 198 ], "spans": [ { "bbox": [ 260, 184, 368, 198 ], "score": 1.0, "content": "Taylor Berg-Kirkpatrick", "type": "text" } ], "index": 11 }, { "bbox": [ 260, 194, 322, 209 ], "spans": [ { "bbox": [ 260, 194, 322, 209 ], "score": 1.0, "content": "UC San Diego", "type": "text" } ], "index": 12 }, { "bbox": [ 260, 207, 371, 220 ], "spans": [ { "bbox": [ 260, 207, 371, 220 ], "score": 1.0, "content": "tberg@eng.ucsd.edu", "type": "text" } ], "index": 15 } ], "index": 12, "bbox_fs": [ 260, 184, 371, 220 ] }, { "type": "text", "bbox": [ 382, 185, 495, 218 ], "lines": [ { "bbox": [ 382, 183, 455, 199 ], "spans": [ { "bbox": [ 382, 183, 455, 199 ], "score": 1.0, "content": "Graham Neubig", "type": "text" } ], "index": 10 }, { "bbox": [ 382, 195, 496, 209 ], "spans": [ { "bbox": [ 382, 195, 496, 209 ], "score": 1.0, "content": "Carnegie Mellon University", "type": "text" } ], "index": 13 }, { "bbox": [ 383, 208, 492, 220 ], "spans": [ { "bbox": [ 383, 208, 492, 220 ], "score": 1.0, "content": "gneubig@cs.cmu.edu", "type": "text" } ], "index": 16 } ], "index": 13, "bbox_fs": [ 382, 183, 496, 220 ] }, { "type": "title", "bbox": [ 278, 248, 333, 259 ], "lines": [ { "bbox": [ 276, 246, 336, 262 ], "spans": [ { "bbox": [ 276, 246, 336, 262 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 143, 273, 468, 493 ], "lines": [ { "bbox": [ 141, 273, 470, 286 ], "spans": [ { "bbox": [ 141, 273, 470, 286 ], "score": 1.0, "content": "Fine-tuning large pretrained language models on downstream tasks has become", "type": "text" } ], "index": 18 }, { "bbox": [ 141, 284, 469, 297 ], "spans": [ { "bbox": [ 141, 284, 469, 297 ], "score": 1.0, "content": "the de-facto learning paradigm in NLP. However, conventional approaches fine-", "type": "text" } ], "index": 19 }, { "bbox": [ 141, 295, 469, 308 ], "spans": [ { "bbox": [ 141, 295, 469, 308 ], "score": 1.0, "content": "tune all the parameters of the pretrained model, which becomes prohibitive as", "type": "text" } ], "index": 20 }, { "bbox": [ 142, 307, 469, 318 ], "spans": [ { "bbox": [ 142, 307, 469, 318 ], "score": 1.0, "content": "the model size and the number of tasks grow. Recent work has proposed a va-", "type": "text" } ], "index": 21 }, { "bbox": [ 141, 317, 470, 329 ], "spans": [ { "bbox": [ 141, 317, 470, 329 ], "score": 1.0, "content": "riety of parameter-efficient transfer learning methods that only fine-tune a small", "type": "text" } ], "index": 22 }, { "bbox": [ 141, 328, 470, 340 ], "spans": [ { "bbox": [ 141, 328, 470, 340 ], "score": 1.0, "content": "number of (extra) parameters to attain strong performance. While effective, the", "type": "text" } ], "index": 23 }, { "bbox": [ 142, 339, 469, 351 ], "spans": [ { "bbox": [ 142, 339, 469, 351 ], "score": 1.0, "content": "critical ingredients for success and the connections among the various methods", "type": "text" } ], "index": 24 }, { "bbox": [ 141, 351, 469, 362 ], "spans": [ { "bbox": [ 141, 351, 469, 362 ], "score": 1.0, "content": "are poorly understood. In this paper, we break down the design of state-of-the-art", "type": "text" } ], "index": 25 }, { "bbox": [ 141, 361, 470, 372 ], "spans": [ { "bbox": [ 141, 361, 470, 372 ], "score": 1.0, "content": "parameter-efficient transfer learning methods and present a unified framework that", "type": "text" } ], "index": 26 }, { "bbox": [ 141, 372, 469, 384 ], "spans": [ { "bbox": [ 141, 372, 469, 384 ], "score": 1.0, "content": "establishes connections between them. Specifically, we re-frame them as modifi-", "type": "text" } ], "index": 27 }, { "bbox": [ 141, 382, 470, 396 ], "spans": [ { "bbox": [ 141, 382, 470, 396 ], "score": 1.0, "content": "cations to specific hidden states in pretrained models, and define a set of design", "type": "text" } ], "index": 28 }, { "bbox": [ 141, 393, 470, 407 ], "spans": [ { "bbox": [ 141, 393, 470, 407 ], "score": 1.0, "content": "dimensions along which different methods vary, such as the function to compute", "type": "text" } ], "index": 29 }, { "bbox": [ 141, 405, 469, 417 ], "spans": [ { "bbox": [ 141, 405, 469, 417 ], "score": 1.0, "content": "the modification and the position to apply the modification. Through comprehen-", "type": "text" } ], "index": 30 }, { "bbox": [ 141, 415, 469, 428 ], "spans": [ { "bbox": [ 141, 415, 469, 428 ], "score": 1.0, "content": "sive empirical studies across machine translation, text summarization, language", "type": "text" } ], "index": 31 }, { "bbox": [ 142, 427, 469, 438 ], "spans": [ { "bbox": [ 142, 427, 469, 438 ], "score": 1.0, "content": "understanding, and text classification benchmarks, we utilize the unified view to", "type": "text" } ], "index": 32 }, { "bbox": [ 141, 438, 470, 449 ], "spans": [ { "bbox": [ 141, 438, 470, 449 ], "score": 1.0, "content": "identify important design choices in previous methods. Furthermore, our unified", "type": "text" } ], "index": 33 }, { "bbox": [ 141, 448, 469, 461 ], "spans": [ { "bbox": [ 141, 448, 469, 461 ], "score": 1.0, "content": "framework enables the transfer of design elements across different approaches,", "type": "text" } ], "index": 34 }, { "bbox": [ 141, 459, 470, 472 ], "spans": [ { "bbox": [ 141, 459, 470, 472 ], "score": 1.0, "content": "and as a result we are able to instantiate new parameter-efficient fine-tuning meth-", "type": "text" } ], "index": 35 }, { "bbox": [ 141, 470, 470, 484 ], "spans": [ { "bbox": [ 141, 470, 470, 484 ], "score": 1.0, "content": "ods that tune less parameters than previous methods while being more effective,", "type": "text" } ], "index": 36 }, { "bbox": [ 141, 482, 448, 493 ], "spans": [ { "bbox": [ 141, 482, 448, 493 ], "score": 1.0, "content": "achieving comparable results to fine-tuning all parameters on all four tasks.1", "type": "text" } ], "index": 37 } ], "index": 27.5, "bbox_fs": [ 141, 273, 470, 493 ] }, { "type": "title", "bbox": [ 108, 515, 206, 528 ], "lines": [ { "bbox": [ 105, 514, 208, 531 ], "spans": [ { "bbox": [ 105, 514, 208, 531 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 38 } ], "index": 38 }, { "type": "text", "bbox": [ 107, 541, 505, 630 ], "lines": [ { "bbox": [ 105, 541, 505, 555 ], "spans": [ { "bbox": [ 105, 541, 505, 555 ], "score": 1.0, "content": "Transfer learning from pre-trained language models (PLMs) is now the prevalent paradigm in natural", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 553, 505, 564 ], "spans": [ { "bbox": [ 105, 553, 505, 564 ], "score": 1.0, "content": "language processing, yielding strong performance on many tasks (Peters et al., 2018; Devlin et al.,", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 563, 505, 576 ], "spans": [ { "bbox": [ 105, 563, 505, 576 ], "score": 1.0, "content": "2019; Qiu et al., 2020). The most common way to adapt general-purpose PLMs to downstream tasks", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 574, 505, 587 ], "spans": [ { "bbox": [ 105, 574, 505, 587 ], "score": 1.0, "content": "is to fine-tune all the model parameters (full fine-tuning). However, this results in a separate copy of", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 586, 505, 598 ], "spans": [ { "bbox": [ 105, 586, 505, 598 ], "score": 1.0, "content": "fine-tuned model parameters for each task, which is prohibitively expensive when serving models", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 597, 505, 609 ], "spans": [ { "bbox": [ 105, 597, 505, 609 ], "score": 1.0, "content": "that perform a large number of tasks. This issue is particularly salient with the ever-increasing size", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 607, 505, 619 ], "spans": [ { "bbox": [ 105, 607, 505, 619 ], "score": 1.0, "content": "of PLMs, which now range from hundreds of millions (Radford et al., 2019; Lewis et al., 2020) to", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 618, 476, 630 ], "spans": [ { "bbox": [ 105, 618, 476, 630 ], "score": 1.0, "content": "hundreds of billions (Brown et al., 2020) or even trillions of parameters (Fedus et al., 2021).", "type": "text" } ], "index": 46 } ], "index": 42.5, "bbox_fs": [ 105, 541, 505, 630 ] }, { "type": "text", "bbox": [ 107, 635, 504, 702 ], "lines": [ { "bbox": [ 106, 635, 505, 648 ], "spans": [ { "bbox": [ 106, 635, 505, 648 ], "score": 1.0, "content": "To mitigate this issue, a few lightweight alternatives have been proposed to update only a small num-", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 645, 505, 659 ], "spans": [ { "bbox": [ 105, 645, 505, 659 ], "score": 1.0, "content": "ber of extra parameters while keeping most pretrained parameters frozen. For example, adapter tun-", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 657, 506, 670 ], "spans": [ { "bbox": [ 105, 657, 506, 670 ], "score": 1.0, "content": "ing (Houlsby et al., 2019) inserts small neural modules called adapters to each layer of the pretrained", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 667, 506, 681 ], "spans": [ { "bbox": [ 105, 667, 506, 681 ], "score": 1.0, "content": "network and only the adapters are trained at fine-tuning time. Inspired by the success of prompting", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 678, 506, 692 ], "spans": [ { "bbox": [ 105, 678, 506, 692 ], "score": 1.0, "content": "methods that control PLMs through textual prompts (Brown et al., 2020; Liu et al., 2021a), prefix", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 690, 505, 702 ], "spans": [ { "bbox": [ 105, 690, 467, 702 ], "score": 1.0, "content": "tuning (Li & Liang, 2021) and prompt tuning (Lester et al., 2021) prepend an additional", "type": "text" }, { "bbox": [ 467, 691, 471, 700 ], "score": 0.37, "content": "l", "type": "inline_equation" }, { "bbox": [ 472, 690, 505, 702 ], "score": 1.0, "content": "tunable", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 324, 505, 338 ], "spans": [ { "bbox": [ 105, 324, 505, 338 ], "score": 1.0, "content": "prefix tokens to the input or hidden layers and only train these soft prompts when fine-tuning on", "type": "text", "cross_page": true } ], "index": 32 }, { "bbox": [ 105, 335, 505, 349 ], "spans": [ { "bbox": [ 105, 335, 505, 349 ], "score": 1.0, "content": "downstream tasks. More recently, Hu et al. (2021) learn low-rank matrices to approximate param-", "type": "text", "cross_page": true } ], "index": 33 }, { "bbox": [ 105, 347, 505, 360 ], "spans": [ { "bbox": [ 105, 347, 505, 360 ], "score": 1.0, "content": "eter updates. We illustrate these methods in Figure 1. These approaches have all been reported to", "type": "text", "cross_page": true } ], "index": 34 }, { "bbox": [ 105, 357, 505, 371 ], "spans": [ { "bbox": [ 105, 357, 505, 371 ], "score": 1.0, "content": "demonstrate comparable performance to full fine-tuning on different sets of tasks, often through up-", "type": "text", "cross_page": true } ], "index": 35 }, { "bbox": [ 105, 368, 506, 382 ], "spans": [ { "bbox": [ 105, 368, 169, 382 ], "score": 1.0, "content": "dating less than", "type": "text", "cross_page": true }, { "bbox": [ 170, 369, 184, 379 ], "score": 0.84, "content": "1 \\%", "type": "inline_equation", "cross_page": true }, { "bbox": [ 185, 368, 506, 382 ], "score": 1.0, "content": "of the original model parameters. Besides parameter savings, parameter-efficient", "type": "text", "cross_page": true } ], "index": 36 }, { "bbox": [ 105, 379, 506, 393 ], "spans": [ { "bbox": [ 105, 379, 506, 393 ], "score": 1.0, "content": "tuning makes it possible to quickly adapt to new tasks without catastrophic forgetting (Pfeiffer et al.,", "type": "text", "cross_page": true } ], "index": 37 }, { "bbox": [ 105, 390, 497, 404 ], "spans": [ { "bbox": [ 105, 390, 497, 404 ], "score": 1.0, "content": "2021) and often exhibits superior robustness in out-of-distribution evaluation (Li & Liang, 2021).", "type": "text", "cross_page": true } ], "index": 38 } ], "index": 49.5, "bbox_fs": [ 105, 635, 506, 702 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 107, 92, 300, 259 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 92, 300, 259 ], "group_id": 0, "lines": [ { "bbox": [ 107, 92, 300, 259 ], "spans": [ { "bbox": [ 107, 92, 300, 259 ], "score": 0.972, "type": "image", "image_path": "817d3636b1fd3352037dfcc37df9a02d25048bdc48931d091e4c6fddbc6d0118.jpg" } ] } ], "index": 5.5, "virtual_lines": [ { "bbox": [ 107, 92, 300, 105.91666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 107, 105.91666666666667, 300, 119.83333333333334 ], "spans": [], "index": 1 }, { "bbox": [ 107, 119.83333333333334, 300, 133.75 ], "spans": [], "index": 2 }, { "bbox": [ 107, 133.75, 300, 147.66666666666666 ], "spans": [], "index": 3 }, { "bbox": [ 107, 147.66666666666666, 300, 161.58333333333331 ], "spans": [], "index": 4 }, { "bbox": [ 107, 161.58333333333331, 300, 175.49999999999997 ], "spans": [], "index": 5 }, { "bbox": [ 107, 175.49999999999997, 300, 189.41666666666663 ], "spans": [], "index": 6 }, { "bbox": [ 107, 189.41666666666663, 300, 203.3333333333333 ], "spans": [], "index": 7 }, { "bbox": [ 107, 203.3333333333333, 300, 217.24999999999994 ], "spans": [], "index": 8 }, { "bbox": [ 107, 217.24999999999994, 300, 231.1666666666666 ], "spans": [], "index": 9 }, { "bbox": [ 107, 231.1666666666666, 300, 245.08333333333326 ], "spans": [], "index": 10 }, { "bbox": [ 107, 245.08333333333326, 300, 258.99999999999994 ], "spans": [], "index": 11 } ] }, { "type": "image_caption", "bbox": [ 106, 269, 302, 309 ], "group_id": 0, "lines": [ { "bbox": [ 106, 269, 303, 279 ], "spans": [ { "bbox": [ 106, 269, 303, 279 ], "score": 1.0, "content": "Figure 1: Illustration of the transformer architecture", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 277, 303, 290 ], "spans": [ { "bbox": [ 105, 277, 303, 290 ], "score": 1.0, "content": "and several state-of-the-art parameter-efficient tuning", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 289, 303, 299 ], "spans": [ { "bbox": [ 106, 289, 303, 299 ], "score": 1.0, "content": "methods. We use blocks with dashed borderlines to", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 299, 277, 309 ], "spans": [ { "bbox": [ 105, 299, 277, 309 ], "score": 1.0, "content": "represent the added modules by those methods.", "type": "text" } ], "index": 15 } ], "index": 13.5 } ], "index": 9.5 }, { "type": "image", "bbox": [ 326, 95, 491, 257 ], "blocks": [ { "type": "image_body", "bbox": [ 326, 95, 491, 257 ], "group_id": 1, "lines": [ { "bbox": [ 326, 95, 491, 257 ], "spans": [ { "bbox": [ 326, 95, 491, 257 ], "score": 0.967, "type": "image", "image_path": "ed2872e8987c3789fd45b53e042a8d249dbe4d22df5d13958ff5aee7ae6ba30f.jpg" } ] } ], "index": 21.5, "virtual_lines": [ { "bbox": [ 326, 95, 491, 108.5 ], "spans": [], "index": 16 }, { "bbox": [ 326, 108.5, 491, 122.0 ], "spans": [], "index": 17 }, { "bbox": [ 326, 122.0, 491, 135.5 ], "spans": [], "index": 18 }, { "bbox": [ 326, 135.5, 491, 149.0 ], "spans": [], "index": 19 }, { "bbox": [ 326, 149.0, 491, 162.5 ], "spans": [], "index": 20 }, { "bbox": [ 326, 162.5, 491, 176.0 ], "spans": [], "index": 21 }, { "bbox": [ 326, 176.0, 491, 189.5 ], "spans": [], "index": 22 }, { "bbox": [ 326, 189.5, 491, 203.0 ], "spans": [], "index": 23 }, { "bbox": [ 326, 203.0, 491, 216.5 ], "spans": [], "index": 24 }, { "bbox": [ 326, 216.5, 491, 230.0 ], "spans": [], "index": 25 }, { "bbox": [ 326, 230.0, 491, 243.5 ], "spans": [], "index": 26 }, { "bbox": [ 326, 243.5, 491, 257.0 ], "spans": [], "index": 27 } ] }, { "type": "image_caption", "bbox": [ 317, 265, 502, 306 ], "group_id": 1, "lines": [ { "bbox": [ 317, 265, 502, 276 ], "spans": [ { "bbox": [ 317, 265, 502, 276 ], "score": 1.0, "content": "Figure 2: Performance of different methods on the", "type": "text" } ], "index": 28 }, { "bbox": [ 317, 275, 502, 286 ], "spans": [ { "bbox": [ 317, 275, 502, 286 ], "score": 1.0, "content": "XSum (Narayan et al., 2018) summarization task.", "type": "text" } ], "index": 29 }, { "bbox": [ 317, 285, 503, 297 ], "spans": [ { "bbox": [ 317, 285, 503, 297 ], "score": 1.0, "content": "The number of fine-tuned parameters is relative to", "type": "text" } ], "index": 30 }, { "bbox": [ 317, 294, 463, 307 ], "spans": [ { "bbox": [ 317, 294, 463, 307 ], "score": 1.0, "content": "the tuned parameters in full fine-tuning.", "type": "text" } ], "index": 31 } ], "index": 29.5 } ], "index": 25.5 }, { "type": "text", "bbox": [ 107, 325, 505, 402 ], "lines": [ { "bbox": [ 105, 324, 505, 338 ], "spans": [ { "bbox": [ 105, 324, 505, 338 ], "score": 1.0, "content": "prefix tokens to the input or hidden layers and only train these soft prompts when fine-tuning on", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 335, 505, 349 ], "spans": [ { "bbox": [ 105, 335, 505, 349 ], "score": 1.0, "content": "downstream tasks. More recently, Hu et al. (2021) learn low-rank matrices to approximate param-", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 347, 505, 360 ], "spans": [ { "bbox": [ 105, 347, 505, 360 ], "score": 1.0, "content": "eter updates. We illustrate these methods in Figure 1. These approaches have all been reported to", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 357, 505, 371 ], "spans": [ { "bbox": [ 105, 357, 505, 371 ], "score": 1.0, "content": "demonstrate comparable performance to full fine-tuning on different sets of tasks, often through up-", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 368, 506, 382 ], "spans": [ { "bbox": [ 105, 368, 169, 382 ], "score": 1.0, "content": "dating less than", "type": "text" }, { "bbox": [ 170, 369, 184, 379 ], "score": 0.84, "content": "1 \\%", "type": "inline_equation" }, { "bbox": [ 185, 368, 506, 382 ], "score": 1.0, "content": "of the original model parameters. Besides parameter savings, parameter-efficient", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 379, 506, 393 ], "spans": [ { "bbox": [ 105, 379, 506, 393 ], "score": 1.0, "content": "tuning makes it possible to quickly adapt to new tasks without catastrophic forgetting (Pfeiffer et al.,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 390, 497, 404 ], "spans": [ { "bbox": [ 105, 390, 497, 404 ], "score": 1.0, "content": "2021) and often exhibits superior robustness in out-of-distribution evaluation (Li & Liang, 2021).", "type": "text" } ], "index": 38 } ], "index": 35 }, { "type": "text", "bbox": [ 107, 408, 504, 462 ], "lines": [ { "bbox": [ 106, 408, 504, 419 ], "spans": [ { "bbox": [ 106, 408, 504, 419 ], "score": 1.0, "content": "However, we contend that the important ingredients that contribute to the success of these parameter-", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 417, 505, 432 ], "spans": [ { "bbox": [ 105, 417, 505, 432 ], "score": 1.0, "content": "efficient tuning methods are poorly understood, and the connections between them are still unclear.", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 429, 506, 443 ], "spans": [ { "bbox": [ 105, 429, 506, 443 ], "score": 1.0, "content": "In this paper, we aim to answer three questions: (1) How are these methods connected? (2) Do these", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 440, 506, 453 ], "spans": [ { "bbox": [ 105, 440, 506, 453 ], "score": 1.0, "content": "methods share design elements that are essential for their effectiveness, and what are they? (3) Can", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 452, 494, 464 ], "spans": [ { "bbox": [ 105, 452, 494, 464 ], "score": 1.0, "content": "the effective ingredients of each method be transferred to others to yield more effective variants?", "type": "text" } ], "index": 43 } ], "index": 41 }, { "type": "text", "bbox": [ 106, 468, 505, 622 ], "lines": [ { "bbox": [ 106, 469, 505, 480 ], "spans": [ { "bbox": [ 106, 469, 505, 480 ], "score": 1.0, "content": "In order to answer these questions, we first derive an alternative form of prefix tuning that reveals", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 479, 505, 492 ], "spans": [ { "bbox": [ 105, 479, 505, 492 ], "score": 1.0, "content": "prefix tuning’s close connections with adapters (§3.1). Based on this we then devise a unified frame-", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 490, 505, 503 ], "spans": [ { "bbox": [ 105, 490, 505, 503 ], "score": 1.0, "content": "work that frames the aforementioned methods as different ways to modify the hidden representations", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 501, 506, 514 ], "spans": [ { "bbox": [ 105, 501, 506, 514 ], "score": 1.0, "content": "of frozen PLMs (§3.2). Our unified framework decomposes previous methods along a shared set", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 512, 505, 524 ], "spans": [ { "bbox": [ 105, 512, 505, 524 ], "score": 1.0, "content": "of design dimensions, such as the function used to perform the modification, the position in which", "type": "text" } ], "index": 48 }, { "bbox": [ 106, 524, 506, 536 ], "spans": [ { "bbox": [ 106, 524, 506, 536 ], "score": 1.0, "content": "to impose this modification, and how to integrate the modification. This framework allows us to", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 534, 505, 547 ], "spans": [ { "bbox": [ 105, 534, 505, 547 ], "score": 1.0, "content": "transfer design choices across approaches to propose new variants such as adapters with multiple", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 546, 505, 558 ], "spans": [ { "bbox": [ 105, 546, 505, 558 ], "score": 1.0, "content": "heads (§3.3). In experiments, we first show that existing parameter-efficient tuning methods still", "type": "text" } ], "index": 51 }, { "bbox": [ 104, 555, 505, 570 ], "spans": [ { "bbox": [ 104, 555, 505, 570 ], "score": 1.0, "content": "lag behind full fine-tuning on higher-resource and challenging tasks (§4.2), as exemplified in Fig-", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 567, 506, 580 ], "spans": [ { "bbox": [ 105, 567, 506, 580 ], "score": 1.0, "content": "ure 2. Then we utilize the unified framework to identify critical design choices and validate the", "type": "text" } ], "index": 53 }, { "bbox": [ 104, 578, 506, 591 ], "spans": [ { "bbox": [ 104, 578, 506, 591 ], "score": 1.0, "content": "proposed variants empirically (§4.3-4.6). Our experiments on four NLP benchmarks covering text", "type": "text" } ], "index": 54 }, { "bbox": [ 104, 587, 505, 603 ], "spans": [ { "bbox": [ 104, 587, 505, 603 ], "score": 1.0, "content": "summarization, machine translation (MT), text classification, and general language understanding,", "type": "text" } ], "index": 55 }, { "bbox": [ 106, 600, 506, 613 ], "spans": [ { "bbox": [ 106, 600, 506, 613 ], "score": 1.0, "content": "demonstrate that the proposed variant uses less parameters than existing methods while being more", "type": "text" } ], "index": 56 }, { "bbox": [ 106, 612, 345, 623 ], "spans": [ { "bbox": [ 106, 612, 345, 623 ], "score": 1.0, "content": "effective, matching full fine-tuning results on all four tasks.", "type": "text" } ], "index": 57 } ], "index": 50.5 }, { "type": "title", "bbox": [ 108, 645, 207, 658 ], "lines": [ { "bbox": [ 104, 643, 209, 660 ], "spans": [ { "bbox": [ 104, 643, 209, 660 ], "score": 1.0, "content": "2 PRELIMINARIES", "type": "text" } ], "index": 58 } ], "index": 58 }, { "type": "title", "bbox": [ 107, 676, 332, 686 ], "lines": [ { "bbox": [ 105, 675, 333, 687 ], "spans": [ { "bbox": [ 105, 675, 333, 687 ], "score": 1.0, "content": "2.1 RECAP OF THE TRANSFORMER ARCHITECTURE", "type": "text" } ], "index": 59 } ], "index": 59 }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 505, 711 ], "score": 1.0, "content": "The transformer model (Vaswani et al., 2017) is now the workhorse architecture behind most state-", "type": "text" } ], "index": 60 }, { "bbox": [ 106, 709, 505, 722 ], "spans": [ { "bbox": [ 106, 709, 505, 722 ], "score": 1.0, "content": "of-the-art PLMs. In this section we recap the equations of this model for completeness. Transformer", "type": "text" } ], "index": 61 }, { "bbox": [ 106, 720, 505, 733 ], "spans": [ { "bbox": [ 106, 720, 207, 733 ], "score": 1.0, "content": "models are composed of", "type": "text" }, { "bbox": [ 207, 721, 216, 730 ], "score": 0.81, "content": "L", "type": "inline_equation" }, { "bbox": [ 216, 720, 505, 733 ], "score": 1.0, "content": "stacked blocks, where each block (Figure 1) contains two types of sub-", "type": "text" } ], "index": 62 } ], "index": 61 } ], "page_idx": 1, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 294, 39 ], "spans": [ { "bbox": [ 106, 25, 294, 39 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 760 ], "lines": [ { "bbox": [ 301, 750, 310, 763 ], "spans": [ { "bbox": [ 301, 750, 310, 763 ], "score": 1.0, "content": "2", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 107, 92, 300, 259 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 92, 300, 259 ], "group_id": 0, "lines": [ { "bbox": [ 107, 92, 300, 259 ], "spans": [ { "bbox": [ 107, 92, 300, 259 ], "score": 0.972, "type": "image", "image_path": "817d3636b1fd3352037dfcc37df9a02d25048bdc48931d091e4c6fddbc6d0118.jpg" } ] } ], "index": 5.5, "virtual_lines": [ { "bbox": [ 107, 92, 300, 105.91666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 107, 105.91666666666667, 300, 119.83333333333334 ], "spans": [], "index": 1 }, { "bbox": [ 107, 119.83333333333334, 300, 133.75 ], "spans": [], "index": 2 }, { "bbox": [ 107, 133.75, 300, 147.66666666666666 ], "spans": [], "index": 3 }, { "bbox": [ 107, 147.66666666666666, 300, 161.58333333333331 ], "spans": [], "index": 4 }, { "bbox": [ 107, 161.58333333333331, 300, 175.49999999999997 ], "spans": [], "index": 5 }, { "bbox": [ 107, 175.49999999999997, 300, 189.41666666666663 ], "spans": [], "index": 6 }, { "bbox": [ 107, 189.41666666666663, 300, 203.3333333333333 ], "spans": [], "index": 7 }, { "bbox": [ 107, 203.3333333333333, 300, 217.24999999999994 ], "spans": [], "index": 8 }, { "bbox": [ 107, 217.24999999999994, 300, 231.1666666666666 ], "spans": [], "index": 9 }, { "bbox": [ 107, 231.1666666666666, 300, 245.08333333333326 ], "spans": [], "index": 10 }, { "bbox": [ 107, 245.08333333333326, 300, 258.99999999999994 ], "spans": [], "index": 11 } ] }, { "type": "image_caption", "bbox": [ 106, 269, 302, 309 ], "group_id": 0, "lines": [ { "bbox": [ 106, 269, 303, 279 ], "spans": [ { "bbox": [ 106, 269, 303, 279 ], "score": 1.0, "content": "Figure 1: Illustration of the transformer architecture", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 277, 303, 290 ], "spans": [ { "bbox": [ 105, 277, 303, 290 ], "score": 1.0, "content": "and several state-of-the-art parameter-efficient tuning", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 289, 303, 299 ], "spans": [ { "bbox": [ 106, 289, 303, 299 ], "score": 1.0, "content": "methods. We use blocks with dashed borderlines to", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 299, 277, 309 ], "spans": [ { "bbox": [ 105, 299, 277, 309 ], "score": 1.0, "content": "represent the added modules by those methods.", "type": "text" } ], "index": 15 } ], "index": 13.5 } ], "index": 9.5 }, { "type": "image", "bbox": [ 326, 95, 491, 257 ], "blocks": [ { "type": "image_body", "bbox": [ 326, 95, 491, 257 ], "group_id": 1, "lines": [ { "bbox": [ 326, 95, 491, 257 ], "spans": [ { "bbox": [ 326, 95, 491, 257 ], "score": 0.967, "type": "image", "image_path": "ed2872e8987c3789fd45b53e042a8d249dbe4d22df5d13958ff5aee7ae6ba30f.jpg" } ] } ], "index": 21.5, "virtual_lines": [ { "bbox": [ 326, 95, 491, 108.5 ], "spans": [], "index": 16 }, { "bbox": [ 326, 108.5, 491, 122.0 ], "spans": [], "index": 17 }, { "bbox": [ 326, 122.0, 491, 135.5 ], "spans": [], "index": 18 }, { "bbox": [ 326, 135.5, 491, 149.0 ], "spans": [], "index": 19 }, { "bbox": [ 326, 149.0, 491, 162.5 ], "spans": [], "index": 20 }, { "bbox": [ 326, 162.5, 491, 176.0 ], "spans": [], "index": 21 }, { "bbox": [ 326, 176.0, 491, 189.5 ], "spans": [], "index": 22 }, { "bbox": [ 326, 189.5, 491, 203.0 ], "spans": [], "index": 23 }, { "bbox": [ 326, 203.0, 491, 216.5 ], "spans": [], "index": 24 }, { "bbox": [ 326, 216.5, 491, 230.0 ], "spans": [], "index": 25 }, { "bbox": [ 326, 230.0, 491, 243.5 ], "spans": [], "index": 26 }, { "bbox": [ 326, 243.5, 491, 257.0 ], "spans": [], "index": 27 } ] }, { "type": "image_caption", "bbox": [ 317, 265, 502, 306 ], "group_id": 1, "lines": [ { "bbox": [ 317, 265, 502, 276 ], "spans": [ { "bbox": [ 317, 265, 502, 276 ], "score": 1.0, "content": "Figure 2: Performance of different methods on the", "type": "text" } ], "index": 28 }, { "bbox": [ 317, 275, 502, 286 ], "spans": [ { "bbox": [ 317, 275, 502, 286 ], "score": 1.0, "content": "XSum (Narayan et al., 2018) summarization task.", "type": "text" } ], "index": 29 }, { "bbox": [ 317, 285, 503, 297 ], "spans": [ { "bbox": [ 317, 285, 503, 297 ], "score": 1.0, "content": "The number of fine-tuned parameters is relative to", "type": "text" } ], "index": 30 }, { "bbox": [ 317, 294, 463, 307 ], "spans": [ { "bbox": [ 317, 294, 463, 307 ], "score": 1.0, "content": "the tuned parameters in full fine-tuning.", "type": "text" } ], "index": 31 } ], "index": 29.5 } ], "index": 25.5 }, { "type": "text", "bbox": [ 107, 325, 505, 402 ], "lines": [], "index": 35, "bbox_fs": [ 105, 324, 506, 404 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 408, 504, 462 ], "lines": [ { "bbox": [ 106, 408, 504, 419 ], "spans": [ { "bbox": [ 106, 408, 504, 419 ], "score": 1.0, "content": "However, we contend that the important ingredients that contribute to the success of these parameter-", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 417, 505, 432 ], "spans": [ { "bbox": [ 105, 417, 505, 432 ], "score": 1.0, "content": "efficient tuning methods are poorly understood, and the connections between them are still unclear.", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 429, 506, 443 ], "spans": [ { "bbox": [ 105, 429, 506, 443 ], "score": 1.0, "content": "In this paper, we aim to answer three questions: (1) How are these methods connected? (2) Do these", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 440, 506, 453 ], "spans": [ { "bbox": [ 105, 440, 506, 453 ], "score": 1.0, "content": "methods share design elements that are essential for their effectiveness, and what are they? (3) Can", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 452, 494, 464 ], "spans": [ { "bbox": [ 105, 452, 494, 464 ], "score": 1.0, "content": "the effective ingredients of each method be transferred to others to yield more effective variants?", "type": "text" } ], "index": 43 } ], "index": 41, "bbox_fs": [ 105, 408, 506, 464 ] }, { "type": "text", "bbox": [ 106, 468, 505, 622 ], "lines": [ { "bbox": [ 106, 469, 505, 480 ], "spans": [ { "bbox": [ 106, 469, 505, 480 ], "score": 1.0, "content": "In order to answer these questions, we first derive an alternative form of prefix tuning that reveals", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 479, 505, 492 ], "spans": [ { "bbox": [ 105, 479, 505, 492 ], "score": 1.0, "content": "prefix tuning’s close connections with adapters (§3.1). Based on this we then devise a unified frame-", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 490, 505, 503 ], "spans": [ { "bbox": [ 105, 490, 505, 503 ], "score": 1.0, "content": "work that frames the aforementioned methods as different ways to modify the hidden representations", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 501, 506, 514 ], "spans": [ { "bbox": [ 105, 501, 506, 514 ], "score": 1.0, "content": "of frozen PLMs (§3.2). Our unified framework decomposes previous methods along a shared set", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 512, 505, 524 ], "spans": [ { "bbox": [ 105, 512, 505, 524 ], "score": 1.0, "content": "of design dimensions, such as the function used to perform the modification, the position in which", "type": "text" } ], "index": 48 }, { "bbox": [ 106, 524, 506, 536 ], "spans": [ { "bbox": [ 106, 524, 506, 536 ], "score": 1.0, "content": "to impose this modification, and how to integrate the modification. This framework allows us to", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 534, 505, 547 ], "spans": [ { "bbox": [ 105, 534, 505, 547 ], "score": 1.0, "content": "transfer design choices across approaches to propose new variants such as adapters with multiple", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 546, 505, 558 ], "spans": [ { "bbox": [ 105, 546, 505, 558 ], "score": 1.0, "content": "heads (§3.3). In experiments, we first show that existing parameter-efficient tuning methods still", "type": "text" } ], "index": 51 }, { "bbox": [ 104, 555, 505, 570 ], "spans": [ { "bbox": [ 104, 555, 505, 570 ], "score": 1.0, "content": "lag behind full fine-tuning on higher-resource and challenging tasks (§4.2), as exemplified in Fig-", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 567, 506, 580 ], "spans": [ { "bbox": [ 105, 567, 506, 580 ], "score": 1.0, "content": "ure 2. Then we utilize the unified framework to identify critical design choices and validate the", "type": "text" } ], "index": 53 }, { "bbox": [ 104, 578, 506, 591 ], "spans": [ { "bbox": [ 104, 578, 506, 591 ], "score": 1.0, "content": "proposed variants empirically (§4.3-4.6). Our experiments on four NLP benchmarks covering text", "type": "text" } ], "index": 54 }, { "bbox": [ 104, 587, 505, 603 ], "spans": [ { "bbox": [ 104, 587, 505, 603 ], "score": 1.0, "content": "summarization, machine translation (MT), text classification, and general language understanding,", "type": "text" } ], "index": 55 }, { "bbox": [ 106, 600, 506, 613 ], "spans": [ { "bbox": [ 106, 600, 506, 613 ], "score": 1.0, "content": "demonstrate that the proposed variant uses less parameters than existing methods while being more", "type": "text" } ], "index": 56 }, { "bbox": [ 106, 612, 345, 623 ], "spans": [ { "bbox": [ 106, 612, 345, 623 ], "score": 1.0, "content": "effective, matching full fine-tuning results on all four tasks.", "type": "text" } ], "index": 57 } ], "index": 50.5, "bbox_fs": [ 104, 469, 506, 623 ] }, { "type": "title", "bbox": [ 108, 645, 207, 658 ], "lines": [ { "bbox": [ 104, 643, 209, 660 ], "spans": [ { "bbox": [ 104, 643, 209, 660 ], "score": 1.0, "content": "2 PRELIMINARIES", "type": "text" } ], "index": 58 } ], "index": 58 }, { "type": "title", "bbox": [ 107, 676, 332, 686 ], "lines": [ { "bbox": [ 105, 675, 333, 687 ], "spans": [ { "bbox": [ 105, 675, 333, 687 ], "score": 1.0, "content": "2.1 RECAP OF THE TRANSFORMER ARCHITECTURE", "type": "text" } ], "index": 59 } ], "index": 59 }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 505, 711 ], "score": 1.0, "content": "The transformer model (Vaswani et al., 2017) is now the workhorse architecture behind most state-", "type": "text" } ], "index": 60 }, { "bbox": [ 106, 709, 505, 722 ], "spans": [ { "bbox": [ 106, 709, 505, 722 ], "score": 1.0, "content": "of-the-art PLMs. In this section we recap the equations of this model for completeness. Transformer", "type": "text" } ], "index": 61 }, { "bbox": [ 106, 720, 505, 733 ], "spans": [ { "bbox": [ 106, 720, 207, 733 ], "score": 1.0, "content": "models are composed of", "type": "text" }, { "bbox": [ 207, 721, 216, 730 ], "score": 0.81, "content": "L", "type": "inline_equation" }, { "bbox": [ 216, 720, 505, 733 ], "score": 1.0, "content": "stacked blocks, where each block (Figure 1) contains two types of sub-", "type": "text" } ], "index": 62 } ], "index": 61, "bbox_fs": [ 106, 699, 505, 733 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 106, 82, 502, 105 ], "lines": [ { "bbox": [ 106, 82, 504, 95 ], "spans": [ { "bbox": [ 106, 82, 504, 95 ], "score": 1.0, "content": "layers: multi-head self-attention and a fully connected feed-forward network (FFN).2 The conven-", "type": "text" } ], "index": 0 }, { "bbox": [ 104, 91, 499, 106 ], "spans": [ { "bbox": [ 104, 91, 259, 106 ], "score": 1.0, "content": "tional attention function maps queries", "type": "text" }, { "bbox": [ 260, 93, 308, 105 ], "score": 0.93, "content": "\\boldsymbol { Q } \\in \\mathbb { R } ^ { n \\times d _ { k } }", "type": "inline_equation" }, { "bbox": [ 309, 91, 390, 106 ], "score": 1.0, "content": "and key-value pairs", "type": "text" }, { "bbox": [ 391, 93, 499, 105 ], "score": 0.39, "content": "\\pmb { K } \\in \\mathbb { R } ^ { m \\times d _ { k } } , \\pmb { V } \\in \\mathbb { R } ^ { m \\times d _ { v } }", "type": "inline_equation" } ], "index": 1 } ], "index": 0.5 }, { "type": "interline_equation", "bbox": [ 225, 108, 385, 136 ], "lines": [ { "bbox": [ 225, 108, 385, 136 ], "spans": [ { "bbox": [ 225, 108, 385, 136 ], "score": 0.94, "content": "\\mathrm { A t t n } ( Q , K , V ) = \\mathrm { s o f t m a x } \\big ( \\frac { Q K ^ { T } } { \\sqrt { d _ { k } } } \\big ) V ,", "type": "interline_equation", "image_path": "15299331db8ef09694fe081540d4f93bbab6da1345a0b6935d59910e0aa8dfe3.jpg" } ] } ], "index": 2.5, "virtual_lines": [ { "bbox": [ 225, 108, 385, 122.0 ], "spans": [], "index": 2 }, { "bbox": [ 225, 122.0, 385, 136.0 ], "spans": [], "index": 3 } ] }, { "type": "text", "bbox": [ 106, 138, 505, 198 ], "lines": [ { "bbox": [ 106, 138, 504, 151 ], "spans": [ { "bbox": [ 106, 138, 133, 151 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 133, 141, 140, 149 ], "score": 0.78, "content": "n", "type": "inline_equation" }, { "bbox": [ 141, 138, 158, 151 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 158, 141, 168, 149 ], "score": 0.76, "content": "m", "type": "inline_equation" }, { "bbox": [ 168, 138, 504, 151 ], "score": 1.0, "content": "are the number of queries and key-value pairs respectively. Multi-head attention per-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 149, 505, 162 ], "spans": [ { "bbox": [ 105, 149, 282, 162 ], "score": 1.0, "content": "forms the attention function in parallel over", "type": "text" }, { "bbox": [ 282, 150, 297, 161 ], "score": 0.89, "content": "N _ { h }", "type": "inline_equation" }, { "bbox": [ 297, 149, 505, 162 ], "score": 1.0, "content": "heads, where each head is separately parameterized", "type": "text" } ], "index": 5 }, { "bbox": [ 101, 156, 510, 179 ], "spans": [ { "bbox": [ 101, 156, 119, 179 ], "score": 1.0, "content": "by", "type": "text" }, { "bbox": [ 120, 160, 142, 176 ], "score": 0.4, "content": "W _ { q } ^ { ( i ) }", "type": "inline_equation" }, { "bbox": [ 142, 156, 146, 179 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 146, 160, 168, 176 ], "score": 0.27, "content": "\\boldsymbol { W } _ { k } ^ { ( i ) }", "type": "inline_equation" }, { "bbox": [ 169, 156, 172, 179 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 172, 160, 235, 174 ], "score": 0.65, "content": "W _ { v } ^ { ( i ) } \\in \\mathbb { R } ^ { d \\times d _ { h } }", "type": "inline_equation" }, { "bbox": [ 235, 156, 510, 179 ], "score": 1.0, "content": "to project inputs to queries, keys, and values. Given a sequence of", "type": "text" } ], "index": 6 }, { "bbox": [ 107, 173, 506, 189 ], "spans": [ { "bbox": [ 107, 178, 117, 186 ], "score": 0.72, "content": "m", "type": "inline_equation" }, { "bbox": [ 117, 173, 150, 189 ], "score": 1.0, "content": "vectors", "type": "text" }, { "bbox": [ 150, 175, 198, 186 ], "score": 0.9, "content": "C \\in \\mathbb { R } ^ { m \\times d }", "type": "inline_equation" }, { "bbox": [ 199, 173, 468, 189 ], "score": 1.0, "content": "over which we would like to perform attention and a query vector", "type": "text" }, { "bbox": [ 469, 175, 501, 186 ], "score": 0.9, "content": "\\pmb { x } \\in \\mathbb { R } ^ { d }", "type": "inline_equation" }, { "bbox": [ 501, 173, 506, 189 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 7 }, { "bbox": [ 104, 186, 460, 199 ], "spans": [ { "bbox": [ 104, 186, 460, 199 ], "score": 1.0, "content": "multi-head attention (MHA) computes the output on each head and concatenates them:3", "type": "text" } ], "index": 8 } ], "index": 6 }, { "type": "interline_equation", "bbox": [ 114, 201, 483, 219 ], "lines": [ { "bbox": [ 114, 201, 483, 219 ], "spans": [ { "bbox": [ 114, 201, 483, 219 ], "score": 0.9, "content": "\\mathrm { M H A } ( C , { \\pmb x } ) = \\mathrm { C o n c a t } ( \\mathrm { h e a d } _ { 1 } , \\cdots , \\mathrm { h e a d } _ { \\mathrm { h } } ) { \\pmb W } _ { o } , \\ \\mathrm { h e a d } _ { \\mathrm { i } } = \\mathrm { A t t n } ( { \\pmb x } { \\pmb W } _ { q } ^ { ( i ) } , C { \\pmb W } _ { k } ^ { ( i ) } , C { \\pmb W } _ { v } ^ { ( i ) } ) ,", "type": "interline_equation", "image_path": "b4207abc5dbe958fa42175659003fde26b9c44e9d3358993a78afc32d25b4733.jpg" } ] } ], "index": 9, "virtual_lines": [ { "bbox": [ 114, 201, 483, 219 ], "spans": [], "index": 9 } ] }, { "type": "text", "bbox": [ 107, 222, 505, 267 ], "lines": [ { "bbox": [ 105, 221, 506, 237 ], "spans": [ { "bbox": [ 105, 221, 134, 237 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 135, 222, 188, 235 ], "score": 0.93, "content": "W _ { o } \\in \\mathbb { R } ^ { d \\times d }", "type": "inline_equation" }, { "bbox": [ 189, 221, 195, 237 ], "score": 1.0, "content": ".", "type": "text" }, { "bbox": [ 195, 223, 202, 234 ], "score": 0.73, "content": "d", "type": "inline_equation" }, { "bbox": [ 203, 221, 360, 237 ], "score": 1.0, "content": "is the model dimension, and in MHA", "type": "text" }, { "bbox": [ 360, 224, 372, 235 ], "score": 0.88, "content": "d _ { h }", "type": "inline_equation" }, { "bbox": [ 372, 221, 448, 237 ], "score": 1.0, "content": "is typically set to", "type": "text" }, { "bbox": [ 448, 223, 472, 235 ], "score": 0.92, "content": "d / N _ { h }", "type": "inline_equation" }, { "bbox": [ 472, 221, 506, 237 ], "score": 1.0, "content": "to save", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 234, 505, 247 ], "spans": [ { "bbox": [ 105, 234, 505, 247 ], "score": 1.0, "content": "parameters, which indicates that each attention head is operating on a lower-dimensional space. The", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 246, 505, 258 ], "spans": [ { "bbox": [ 106, 246, 505, 258 ], "score": 1.0, "content": "other important sublayer is the fully connected feed-forward network (FFN) which consists of two", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 256, 376, 268 ], "spans": [ { "bbox": [ 105, 256, 376, 268 ], "score": 1.0, "content": "linear transformations with a ReLU activation function in between:", "type": "text" } ], "index": 13 } ], "index": 11.5 }, { "type": "interline_equation", "bbox": [ 221, 271, 389, 285 ], "lines": [ { "bbox": [ 221, 271, 389, 285 ], "spans": [ { "bbox": [ 221, 271, 389, 285 ], "score": 0.93, "content": "\\mathrm { F F N } ( \\pmb { x } ) = \\mathrm { R e L U } ( \\pmb { x } \\pmb { W } _ { 1 } + \\pmb { b } _ { 1 } ) \\pmb { W } _ { 2 } + \\pmb { b } _ { 2 } ,", "type": "interline_equation", "image_path": "91465fc942a612d822db153689232aaafecdb1971dc32ba7922da75d199c8dca.jpg" } ] } ], "index": 14, "virtual_lines": [ { "bbox": [ 221, 271, 389, 285 ], "spans": [], "index": 14 } ] }, { "type": "text", "bbox": [ 103, 288, 504, 311 ], "lines": [ { "bbox": [ 105, 285, 506, 303 ], "spans": [ { "bbox": [ 105, 285, 133, 303 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 287, 190, 299 ], "score": 0.9, "content": "W _ { 1 } \\in \\mathbb { R } ^ { d \\times d _ { m } }", "type": "inline_equation" }, { "bbox": [ 191, 285, 195, 303 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 195, 287, 253, 299 ], "score": 0.92, "content": "W _ { 2 } \\in \\mathbb { R } ^ { d _ { m } \\times d }", "type": "inline_equation" }, { "bbox": [ 253, 285, 395, 303 ], "score": 1.0, "content": ". Transformers typically use a large", "type": "text" }, { "bbox": [ 395, 289, 409, 299 ], "score": 0.88, "content": "d _ { m }", "type": "inline_equation" }, { "bbox": [ 409, 285, 430, 303 ], "score": 1.0, "content": ", e.g.", "type": "text" }, { "bbox": [ 430, 289, 468, 299 ], "score": 0.91, "content": "d _ { m } = 4 d", "type": "inline_equation" }, { "bbox": [ 469, 285, 506, 303 ], "score": 1.0, "content": ". Finally,", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 298, 424, 311 ], "spans": [ { "bbox": [ 105, 298, 424, 311 ], "score": 1.0, "content": "a residual connection is used followed by layer normalization (Ba et al., 2016).", "type": "text" } ], "index": 16 } ], "index": 15.5 }, { "type": "title", "bbox": [ 108, 323, 425, 335 ], "lines": [ { "bbox": [ 105, 323, 426, 336 ], "spans": [ { "bbox": [ 105, 323, 426, 336 ], "score": 1.0, "content": "2.2 OVERVIEW OF PREVIOUS PARAMETER-EFFICIENT TUNING METHODS", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 107, 344, 504, 367 ], "lines": [ { "bbox": [ 106, 344, 505, 357 ], "spans": [ { "bbox": [ 106, 344, 505, 357 ], "score": 1.0, "content": "Below and in Figure 1, we introduce several state-of-the-art parameter-efficient tuning methods.", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 355, 478, 368 ], "spans": [ { "bbox": [ 106, 355, 478, 368 ], "score": 1.0, "content": "Unless otherwise specified, they only tune the added parameters while the PLM’s are frozen.", "type": "text" } ], "index": 19 } ], "index": 18.5 }, { "type": "text", "bbox": [ 106, 371, 505, 427 ], "lines": [ { "bbox": [ 106, 372, 506, 385 ], "spans": [ { "bbox": [ 106, 372, 506, 385 ], "score": 1.0, "content": "Adapters (Houlsby et al., 2019): The adapter approach inserts small modules (adapters) between", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 381, 507, 397 ], "spans": [ { "bbox": [ 104, 381, 426, 397 ], "score": 1.0, "content": "transformer layers. The adapter layer generally uses a down-projection with", "type": "text" }, { "bbox": [ 426, 383, 492, 394 ], "score": 0.92, "content": "W _ { \\mathrm { d o w n } } \\ \\in \\ \\mathbb { R } ^ { d \\times r }", "type": "inline_equation" }, { "bbox": [ 493, 381, 507, 397 ], "score": 1.0, "content": "to", "type": "text" } ], "index": 21 }, { "bbox": [ 104, 393, 505, 408 ], "spans": [ { "bbox": [ 104, 393, 175, 408 ], "score": 1.0, "content": "project the input", "type": "text" }, { "bbox": [ 175, 394, 183, 404 ], "score": 0.79, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 184, 393, 444, 408 ], "score": 1.0, "content": "to a lower-dimensional space specified by bottleneck dimension", "type": "text" }, { "bbox": [ 444, 396, 450, 404 ], "score": 0.75, "content": "r", "type": "inline_equation" }, { "bbox": [ 451, 393, 505, 408 ], "score": 1.0, "content": ", followed by", "type": "text" } ], "index": 22 }, { "bbox": [ 103, 402, 507, 419 ], "spans": [ { "bbox": [ 103, 402, 234, 419 ], "score": 1.0, "content": "a nonlinear activation function", "type": "text" }, { "bbox": [ 235, 405, 252, 417 ], "score": 0.91, "content": "f ( \\cdot )", "type": "inline_equation" }, { "bbox": [ 252, 402, 362, 419 ], "score": 1.0, "content": ", and a up-projection with", "type": "text" }, { "bbox": [ 362, 404, 419, 417 ], "score": 0.93, "content": "W _ { \\mathsf { u p } } \\in \\mathbb { R } ^ { r \\times d }", "type": "inline_equation" }, { "bbox": [ 419, 402, 507, 419 ], "score": 1.0, "content": ". These adapters are", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 416, 349, 428 ], "spans": [ { "bbox": [ 106, 416, 349, 428 ], "score": 1.0, "content": "surrounded by a residual connection, leading to a final form:", "type": "text" } ], "index": 24 } ], "index": 22 }, { "type": "interline_equation", "bbox": [ 249, 430, 361, 444 ], "lines": [ { "bbox": [ 249, 430, 361, 444 ], "spans": [ { "bbox": [ 249, 430, 361, 444 ], "score": 0.92, "content": "h h + f ( h W _ { \\mathrm { d o w n } } ) W _ { \\mathrm { u p } } .", "type": "interline_equation", "image_path": "d051c5840ac628806d97ea98a2dc823b86c10625a3c67375d4f75bab612e1628.jpg" } ] } ], "index": 25, "virtual_lines": [ { "bbox": [ 249, 430, 361, 444 ], "spans": [], "index": 25 } ] }, { "type": "text", "bbox": [ 106, 448, 505, 482 ], "lines": [ { "bbox": [ 105, 447, 506, 461 ], "spans": [ { "bbox": [ 105, 447, 506, 461 ], "score": 1.0, "content": "Houlsby et al. (2019) places two adapters sequentially within one layer of the transformer, one after", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 458, 505, 471 ], "spans": [ { "bbox": [ 105, 458, 505, 471 ], "score": 1.0, "content": "the multi-head attention and one after the FFN sub-layer. Pfeiffer et al. (2021) have proposed a more", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 469, 463, 483 ], "spans": [ { "bbox": [ 105, 469, 463, 483 ], "score": 1.0, "content": "efficient adapter variant that is inserted only after the FFN “add & layer norm” sub-layer.", "type": "text" } ], "index": 28 } ], "index": 27 }, { "type": "text", "bbox": [ 106, 486, 505, 542 ], "lines": [ { "bbox": [ 106, 487, 505, 499 ], "spans": [ { "bbox": [ 106, 487, 505, 499 ], "score": 1.0, "content": "Prefix Tuning (Li & Liang, 2021): Inspired by the success of textual prompting methods (Liu", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 497, 504, 510 ], "spans": [ { "bbox": [ 105, 497, 260, 510 ], "score": 1.0, "content": "et al., 2021a), prefix tuning prepends", "type": "text" }, { "bbox": [ 260, 498, 265, 508 ], "score": 0.64, "content": "l", "type": "inline_equation" }, { "bbox": [ 265, 497, 504, 510 ], "score": 1.0, "content": "tunable prefix vectors to the keys and values of the multi-", "type": "text" } ], "index": 30 }, { "bbox": [ 104, 508, 506, 522 ], "spans": [ { "bbox": [ 104, 508, 373, 522 ], "score": 1.0, "content": "head attention at every layer. Specifically, two sets of prefix vectors", "type": "text" }, { "bbox": [ 373, 508, 435, 520 ], "score": 0.92, "content": "P _ { k } , \\dot { P } _ { v } \\in \\mathbb R ^ { l \\times d }", "type": "inline_equation" }, { "bbox": [ 436, 508, 506, 522 ], "score": 1.0, "content": "are concatenated", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 519, 506, 532 ], "spans": [ { "bbox": [ 106, 519, 192, 532 ], "score": 1.0, "content": "with the original key", "type": "text" }, { "bbox": [ 193, 520, 204, 530 ], "score": 0.75, "content": "\\kappa", "type": "inline_equation" }, { "bbox": [ 205, 519, 247, 532 ], "score": 1.0, "content": "and value", "type": "text" }, { "bbox": [ 247, 520, 257, 530 ], "score": 0.76, "content": "V", "type": "inline_equation" }, { "bbox": [ 258, 519, 506, 532 ], "score": 1.0, "content": ". Then multi-head attention is performed on the new prefixed", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 529, 356, 544 ], "spans": [ { "bbox": [ 105, 529, 255, 544 ], "score": 1.0, "content": "keys and values. The computation of", "type": "text" }, { "bbox": [ 256, 531, 281, 542 ], "score": 0.84, "content": "{ \\mathrm { h e a d } } _ { i }", "type": "inline_equation" }, { "bbox": [ 281, 529, 356, 544 ], "score": 1.0, "content": "in Eq. 2 becomes:", "type": "text" } ], "index": 33 } ], "index": 31 }, { "type": "interline_equation", "bbox": [ 163, 545, 446, 563 ], "lines": [ { "bbox": [ 163, 545, 446, 563 ], "spans": [ { "bbox": [ 163, 545, 446, 563 ], "score": 0.9, "content": "\\mathrm { h e a d } _ { i } = \\mathrm { A t t n } ( \\pmb { x } \\pmb { W } _ { q } ^ { ( i ) } , \\mathrm { c o n c a t } ( \\pmb { P } _ { k } ^ { ( i ) } , \\pmb { C } \\pmb { W } _ { k } ^ { ( i ) } ) , \\mathrm { c o n c a t } ( \\pmb { P } _ { v } ^ { ( i ) } , \\pmb { C } \\pmb { W } _ { v } ^ { ( i ) } ) ) ,", "type": "interline_equation", "image_path": "3b583e4542c85d5a77c153f9671be5620121202770b5f0460f1bd76e61784b91.jpg" } ] } ], "index": 34, "virtual_lines": [ { "bbox": [ 163, 545, 446, 563 ], "spans": [], "index": 34 } ] }, { "type": "text", "bbox": [ 107, 567, 505, 603 ], "lines": [ { "bbox": [ 107, 559, 506, 588 ], "spans": [ { "bbox": [ 107, 570, 120, 581 ], "score": 0.87, "content": "P _ { k }", "type": "inline_equation" }, { "bbox": [ 120, 559, 139, 588 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 140, 570, 153, 581 ], "score": 0.88, "content": "P _ { v }", "type": "inline_equation" }, { "bbox": [ 153, 559, 209, 588 ], "score": 1.0, "content": "are split into", "type": "text" }, { "bbox": [ 209, 570, 223, 581 ], "score": 0.9, "content": "N _ { h }", "type": "inline_equation" }, { "bbox": [ 224, 559, 348, 588 ], "score": 1.0, "content": "head vectors respectively and", "type": "text" }, { "bbox": [ 348, 566, 440, 582 ], "score": 0.89, "content": "P _ { k } ^ { ( i ) } , P _ { v } ^ { ( i ) } \\in \\mathbb R ^ { l \\times d / N _ { h } }", "type": "inline_equation" }, { "bbox": [ 441, 559, 488, 588 ], "score": 1.0, "content": "denote the", "type": "text" }, { "bbox": [ 488, 570, 493, 579 ], "score": 0.74, "content": "i", "type": "inline_equation" }, { "bbox": [ 493, 559, 506, 588 ], "score": 1.0, "content": "-th", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 580, 505, 593 ], "spans": [ { "bbox": [ 105, 580, 505, 593 ], "score": 1.0, "content": "head vector. Prompt-tuning (Lester et al., 2021) simplifies prefix-tuning by only prepending to the", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 591, 492, 604 ], "spans": [ { "bbox": [ 106, 591, 378, 604 ], "score": 1.0, "content": "input word embeddings in the first layer; similar work also includes", "type": "text" }, { "bbox": [ 378, 592, 385, 601 ], "score": 0.37, "content": "\\mathrm { \\bf P }", "type": "inline_equation" }, { "bbox": [ 386, 591, 492, 604 ], "score": 1.0, "content": "-tuning (Liu et al., 2021b).", "type": "text" } ], "index": 37 } ], "index": 36 }, { "type": "text", "bbox": [ 106, 608, 505, 676 ], "lines": [ { "bbox": [ 105, 607, 506, 621 ], "spans": [ { "bbox": [ 105, 607, 506, 621 ], "score": 1.0, "content": "LoRA (Hu et al., 2021): LoRA injects trainable low-rank matrices into transformer layers to", "type": "text" } ], "index": 38 }, { "bbox": [ 104, 617, 506, 633 ], "spans": [ { "bbox": [ 104, 617, 369, 633 ], "score": 1.0, "content": "approximate the weight updates. For a pre-trained weight matrix", "type": "text" }, { "bbox": [ 370, 618, 418, 630 ], "score": 0.92, "content": "W \\in \\mathbb { R } ^ { d \\times k }", "type": "inline_equation" }, { "bbox": [ 419, 617, 506, 633 ], "score": 1.0, "content": ", LoRA represents its", "type": "text" } ], "index": 39 }, { "bbox": [ 104, 626, 505, 646 ], "spans": [ { "bbox": [ 104, 626, 260, 646 ], "score": 1.0, "content": "update with a low-rank decomposition", "type": "text" }, { "bbox": [ 260, 630, 380, 642 ], "score": 0.9, "content": "W + \\Delta W = W + W _ { \\mathrm { d o w n } } W _ { \\mathrm { u p } }", "type": "inline_equation" }, { "bbox": [ 380, 626, 410, 646 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 410, 630, 471, 642 ], "score": 0.88, "content": "W _ { \\mathrm { d o w n } } \\in \\mathbb { R } ^ { \\hat { d } \\times r }", "type": "inline_equation" }, { "bbox": [ 471, 626, 475, 646 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 475, 630, 505, 643 ], "score": 0.84, "content": "W _ { \\mathrm { u p } } \\in", "type": "inline_equation" } ], "index": 40 }, { "bbox": [ 106, 639, 506, 658 ], "spans": [ { "bbox": [ 106, 641, 130, 653 ], "score": 0.9, "content": "\\mathbb { R } ^ { r \\times k }", "type": "inline_equation" }, { "bbox": [ 130, 639, 506, 658 ], "score": 1.0, "content": "are tunable parameters. LoRA applies this update to the query and value projection matrices", "type": "text" } ], "index": 41 }, { "bbox": [ 107, 653, 506, 667 ], "spans": [ { "bbox": [ 107, 654, 150, 666 ], "score": 0.9, "content": "\\left( W _ { q } , W _ { v } \\right)", "type": "inline_equation" }, { "bbox": [ 151, 653, 470, 667 ], "score": 1.0, "content": "in the multi-head attention sub-layer, as shown in Figure 1. For a specific input", "type": "text" }, { "bbox": [ 471, 656, 479, 664 ], "score": 0.79, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 479, 653, 506, 667 ], "score": 1.0, "content": "to the", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 665, 442, 677 ], "spans": [ { "bbox": [ 106, 665, 420, 677 ], "score": 1.0, "content": "linear projection in multi-head attention, LoRA modifies the projection output", "type": "text" }, { "bbox": [ 420, 666, 428, 675 ], "score": 0.82, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 428, 665, 442, 677 ], "score": 1.0, "content": "as:", "type": "text" } ], "index": 43 } ], "index": 40.5 }, { "type": "interline_equation", "bbox": [ 250, 680, 360, 694 ], "lines": [ { "bbox": [ 250, 680, 360, 694 ], "spans": [ { "bbox": [ 250, 680, 360, 694 ], "score": 0.92, "content": "h h + s \\cdot x W _ { \\mathrm { d o w n } } W _ { \\mathrm { u p } } ,", "type": "interline_equation", "image_path": "089f93b64428c027bf2a950afbf1f0c71a98197bfdbfdd26d5231d9173a4f135.jpg" } ] } ], "index": 44, "virtual_lines": [ { "bbox": [ 250, 680, 360, 694 ], "spans": [], "index": 44 } ] } ], "page_idx": 2, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 700, 505, 732 ], "lines": [ { "bbox": [ 118, 698, 505, 713 ], "spans": [ { "bbox": [ 118, 698, 505, 713 ], "score": 1.0, "content": "2In an encoder-decoder architecture, the transformer decoder usually has another multi-head cross-attention", "type": "text" } ] }, { "bbox": [ 105, 709, 392, 722 ], "spans": [ { "bbox": [ 105, 709, 392, 722 ], "score": 1.0, "content": "module between the self-attention and FFN, which we omit here for simplicity.", "type": "text" } ] }, { "bbox": [ 118, 719, 462, 734 ], "spans": [ { "bbox": [ 118, 719, 282, 734 ], "score": 1.0, "content": "3Below, we sometimes ignore the head index", "type": "text" }, { "bbox": [ 283, 722, 287, 730 ], "score": 0.8, "content": "_ { i }", "type": "inline_equation" }, { "bbox": [ 287, 719, 462, 734 ], "score": 1.0, "content": "to simplify notation when there is no confusion.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 26, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 294, 38 ], "spans": [ { "bbox": [ 106, 25, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 760 ], "lines": [ { "bbox": [ 301, 750, 310, 762 ], "spans": [ { "bbox": [ 301, 750, 310, 762 ], "score": 1.0, "content": "3", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 106, 82, 502, 105 ], "lines": [ { "bbox": [ 106, 82, 504, 95 ], "spans": [ { "bbox": [ 106, 82, 504, 95 ], "score": 1.0, "content": "layers: multi-head self-attention and a fully connected feed-forward network (FFN).2 The conven-", "type": "text" } ], "index": 0 }, { "bbox": [ 104, 91, 499, 106 ], "spans": [ { "bbox": [ 104, 91, 259, 106 ], "score": 1.0, "content": "tional attention function maps queries", "type": "text" }, { "bbox": [ 260, 93, 308, 105 ], "score": 0.93, "content": "\\boldsymbol { Q } \\in \\mathbb { R } ^ { n \\times d _ { k } }", "type": "inline_equation" }, { "bbox": [ 309, 91, 390, 106 ], "score": 1.0, "content": "and key-value pairs", "type": "text" }, { "bbox": [ 391, 93, 499, 105 ], "score": 0.39, "content": "\\pmb { K } \\in \\mathbb { R } ^ { m \\times d _ { k } } , \\pmb { V } \\in \\mathbb { R } ^ { m \\times d _ { v } }", "type": "inline_equation" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 104, 82, 504, 106 ] }, { "type": "interline_equation", "bbox": [ 225, 108, 385, 136 ], "lines": [ { "bbox": [ 225, 108, 385, 136 ], "spans": [ { "bbox": [ 225, 108, 385, 136 ], "score": 0.94, "content": "\\mathrm { A t t n } ( Q , K , V ) = \\mathrm { s o f t m a x } \\big ( \\frac { Q K ^ { T } } { \\sqrt { d _ { k } } } \\big ) V ,", "type": "interline_equation", "image_path": "15299331db8ef09694fe081540d4f93bbab6da1345a0b6935d59910e0aa8dfe3.jpg" } ] } ], "index": 2.5, "virtual_lines": [ { "bbox": [ 225, 108, 385, 122.0 ], "spans": [], "index": 2 }, { "bbox": [ 225, 122.0, 385, 136.0 ], "spans": [], "index": 3 } ] }, { "type": "text", "bbox": [ 106, 138, 505, 198 ], "lines": [ { "bbox": [ 106, 138, 504, 151 ], "spans": [ { "bbox": [ 106, 138, 133, 151 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 133, 141, 140, 149 ], "score": 0.78, "content": "n", "type": "inline_equation" }, { "bbox": [ 141, 138, 158, 151 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 158, 141, 168, 149 ], "score": 0.76, "content": "m", "type": "inline_equation" }, { "bbox": [ 168, 138, 504, 151 ], "score": 1.0, "content": "are the number of queries and key-value pairs respectively. Multi-head attention per-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 149, 505, 162 ], "spans": [ { "bbox": [ 105, 149, 282, 162 ], "score": 1.0, "content": "forms the attention function in parallel over", "type": "text" }, { "bbox": [ 282, 150, 297, 161 ], "score": 0.89, "content": "N _ { h }", "type": "inline_equation" }, { "bbox": [ 297, 149, 505, 162 ], "score": 1.0, "content": "heads, where each head is separately parameterized", "type": "text" } ], "index": 5 }, { "bbox": [ 101, 156, 510, 179 ], "spans": [ { "bbox": [ 101, 156, 119, 179 ], "score": 1.0, "content": "by", "type": "text" }, { "bbox": [ 120, 160, 142, 176 ], "score": 0.4, "content": "W _ { q } ^ { ( i ) }", "type": "inline_equation" }, { "bbox": [ 142, 156, 146, 179 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 146, 160, 168, 176 ], "score": 0.27, "content": "\\boldsymbol { W } _ { k } ^ { ( i ) }", "type": "inline_equation" }, { "bbox": [ 169, 156, 172, 179 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 172, 160, 235, 174 ], "score": 0.65, "content": "W _ { v } ^ { ( i ) } \\in \\mathbb { R } ^ { d \\times d _ { h } }", "type": "inline_equation" }, { "bbox": [ 235, 156, 510, 179 ], "score": 1.0, "content": "to project inputs to queries, keys, and values. Given a sequence of", "type": "text" } ], "index": 6 }, { "bbox": [ 107, 173, 506, 189 ], "spans": [ { "bbox": [ 107, 178, 117, 186 ], "score": 0.72, "content": "m", "type": "inline_equation" }, { "bbox": [ 117, 173, 150, 189 ], "score": 1.0, "content": "vectors", "type": "text" }, { "bbox": [ 150, 175, 198, 186 ], "score": 0.9, "content": "C \\in \\mathbb { R } ^ { m \\times d }", "type": "inline_equation" }, { "bbox": [ 199, 173, 468, 189 ], "score": 1.0, "content": "over which we would like to perform attention and a query vector", "type": "text" }, { "bbox": [ 469, 175, 501, 186 ], "score": 0.9, "content": "\\pmb { x } \\in \\mathbb { R } ^ { d }", "type": "inline_equation" }, { "bbox": [ 501, 173, 506, 189 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 7 }, { "bbox": [ 104, 186, 460, 199 ], "spans": [ { "bbox": [ 104, 186, 460, 199 ], "score": 1.0, "content": "multi-head attention (MHA) computes the output on each head and concatenates them:3", "type": "text" } ], "index": 8 } ], "index": 6, "bbox_fs": [ 101, 138, 510, 199 ] }, { "type": "interline_equation", "bbox": [ 114, 201, 483, 219 ], "lines": [ { "bbox": [ 114, 201, 483, 219 ], "spans": [ { "bbox": [ 114, 201, 483, 219 ], "score": 0.9, "content": "\\mathrm { M H A } ( C , { \\pmb x } ) = \\mathrm { C o n c a t } ( \\mathrm { h e a d } _ { 1 } , \\cdots , \\mathrm { h e a d } _ { \\mathrm { h } } ) { \\pmb W } _ { o } , \\ \\mathrm { h e a d } _ { \\mathrm { i } } = \\mathrm { A t t n } ( { \\pmb x } { \\pmb W } _ { q } ^ { ( i ) } , C { \\pmb W } _ { k } ^ { ( i ) } , C { \\pmb W } _ { v } ^ { ( i ) } ) ,", "type": "interline_equation", "image_path": "b4207abc5dbe958fa42175659003fde26b9c44e9d3358993a78afc32d25b4733.jpg" } ] } ], "index": 9, "virtual_lines": [ { "bbox": [ 114, 201, 483, 219 ], "spans": [], "index": 9 } ] }, { "type": "text", "bbox": [ 107, 222, 505, 267 ], "lines": [ { "bbox": [ 105, 221, 506, 237 ], "spans": [ { "bbox": [ 105, 221, 134, 237 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 135, 222, 188, 235 ], "score": 0.93, "content": "W _ { o } \\in \\mathbb { R } ^ { d \\times d }", "type": "inline_equation" }, { "bbox": [ 189, 221, 195, 237 ], "score": 1.0, "content": ".", "type": "text" }, { "bbox": [ 195, 223, 202, 234 ], "score": 0.73, "content": "d", "type": "inline_equation" }, { "bbox": [ 203, 221, 360, 237 ], "score": 1.0, "content": "is the model dimension, and in MHA", "type": "text" }, { "bbox": [ 360, 224, 372, 235 ], "score": 0.88, "content": "d _ { h }", "type": "inline_equation" }, { "bbox": [ 372, 221, 448, 237 ], "score": 1.0, "content": "is typically set to", "type": "text" }, { "bbox": [ 448, 223, 472, 235 ], "score": 0.92, "content": "d / N _ { h }", "type": "inline_equation" }, { "bbox": [ 472, 221, 506, 237 ], "score": 1.0, "content": "to save", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 234, 505, 247 ], "spans": [ { "bbox": [ 105, 234, 505, 247 ], "score": 1.0, "content": "parameters, which indicates that each attention head is operating on a lower-dimensional space. The", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 246, 505, 258 ], "spans": [ { "bbox": [ 106, 246, 505, 258 ], "score": 1.0, "content": "other important sublayer is the fully connected feed-forward network (FFN) which consists of two", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 256, 376, 268 ], "spans": [ { "bbox": [ 105, 256, 376, 268 ], "score": 1.0, "content": "linear transformations with a ReLU activation function in between:", "type": "text" } ], "index": 13 } ], "index": 11.5, "bbox_fs": [ 105, 221, 506, 268 ] }, { "type": "interline_equation", "bbox": [ 221, 271, 389, 285 ], "lines": [ { "bbox": [ 221, 271, 389, 285 ], "spans": [ { "bbox": [ 221, 271, 389, 285 ], "score": 0.93, "content": "\\mathrm { F F N } ( \\pmb { x } ) = \\mathrm { R e L U } ( \\pmb { x } \\pmb { W } _ { 1 } + \\pmb { b } _ { 1 } ) \\pmb { W } _ { 2 } + \\pmb { b } _ { 2 } ,", "type": "interline_equation", "image_path": "91465fc942a612d822db153689232aaafecdb1971dc32ba7922da75d199c8dca.jpg" } ] } ], "index": 14, "virtual_lines": [ { "bbox": [ 221, 271, 389, 285 ], "spans": [], "index": 14 } ] }, { "type": "text", "bbox": [ 103, 288, 504, 311 ], "lines": [ { "bbox": [ 105, 285, 506, 303 ], "spans": [ { "bbox": [ 105, 285, 133, 303 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 287, 190, 299 ], "score": 0.9, "content": "W _ { 1 } \\in \\mathbb { R } ^ { d \\times d _ { m } }", "type": "inline_equation" }, { "bbox": [ 191, 285, 195, 303 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 195, 287, 253, 299 ], "score": 0.92, "content": "W _ { 2 } \\in \\mathbb { R } ^ { d _ { m } \\times d }", "type": "inline_equation" }, { "bbox": [ 253, 285, 395, 303 ], "score": 1.0, "content": ". Transformers typically use a large", "type": "text" }, { "bbox": [ 395, 289, 409, 299 ], "score": 0.88, "content": "d _ { m }", "type": "inline_equation" }, { "bbox": [ 409, 285, 430, 303 ], "score": 1.0, "content": ", e.g.", "type": "text" }, { "bbox": [ 430, 289, 468, 299 ], "score": 0.91, "content": "d _ { m } = 4 d", "type": "inline_equation" }, { "bbox": [ 469, 285, 506, 303 ], "score": 1.0, "content": ". Finally,", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 298, 424, 311 ], "spans": [ { "bbox": [ 105, 298, 424, 311 ], "score": 1.0, "content": "a residual connection is used followed by layer normalization (Ba et al., 2016).", "type": "text" } ], "index": 16 } ], "index": 15.5, "bbox_fs": [ 105, 285, 506, 311 ] }, { "type": "title", "bbox": [ 108, 323, 425, 335 ], "lines": [ { "bbox": [ 105, 323, 426, 336 ], "spans": [ { "bbox": [ 105, 323, 426, 336 ], "score": 1.0, "content": "2.2 OVERVIEW OF PREVIOUS PARAMETER-EFFICIENT TUNING METHODS", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "list", "bbox": [ 107, 344, 504, 367 ], "lines": [ { "bbox": [ 106, 344, 505, 357 ], "spans": [ { "bbox": [ 106, 344, 505, 357 ], "score": 1.0, "content": "Below and in Figure 1, we introduce several state-of-the-art parameter-efficient tuning methods.", "type": "text" } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 106, 355, 478, 368 ], "spans": [ { "bbox": [ 106, 355, 478, 368 ], "score": 1.0, "content": "Unless otherwise specified, they only tune the added parameters while the PLM’s are frozen.", "type": "text" } ], "index": 19, "is_list_start_line": true, "is_list_end_line": true } ], "index": 18.5, "bbox_fs": [ 106, 344, 505, 368 ] }, { "type": "text", "bbox": [ 106, 371, 505, 427 ], "lines": [ { "bbox": [ 106, 372, 506, 385 ], "spans": [ { "bbox": [ 106, 372, 506, 385 ], "score": 1.0, "content": "Adapters (Houlsby et al., 2019): The adapter approach inserts small modules (adapters) between", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 381, 507, 397 ], "spans": [ { "bbox": [ 104, 381, 426, 397 ], "score": 1.0, "content": "transformer layers. The adapter layer generally uses a down-projection with", "type": "text" }, { "bbox": [ 426, 383, 492, 394 ], "score": 0.92, "content": "W _ { \\mathrm { d o w n } } \\ \\in \\ \\mathbb { R } ^ { d \\times r }", "type": "inline_equation" }, { "bbox": [ 493, 381, 507, 397 ], "score": 1.0, "content": "to", "type": "text" } ], "index": 21 }, { "bbox": [ 104, 393, 505, 408 ], "spans": [ { "bbox": [ 104, 393, 175, 408 ], "score": 1.0, "content": "project the input", "type": "text" }, { "bbox": [ 175, 394, 183, 404 ], "score": 0.79, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 184, 393, 444, 408 ], "score": 1.0, "content": "to a lower-dimensional space specified by bottleneck dimension", "type": "text" }, { "bbox": [ 444, 396, 450, 404 ], "score": 0.75, "content": "r", "type": "inline_equation" }, { "bbox": [ 451, 393, 505, 408 ], "score": 1.0, "content": ", followed by", "type": "text" } ], "index": 22 }, { "bbox": [ 103, 402, 507, 419 ], "spans": [ { "bbox": [ 103, 402, 234, 419 ], "score": 1.0, "content": "a nonlinear activation function", "type": "text" }, { "bbox": [ 235, 405, 252, 417 ], "score": 0.91, "content": "f ( \\cdot )", "type": "inline_equation" }, { "bbox": [ 252, 402, 362, 419 ], "score": 1.0, "content": ", and a up-projection with", "type": "text" }, { "bbox": [ 362, 404, 419, 417 ], "score": 0.93, "content": "W _ { \\mathsf { u p } } \\in \\mathbb { R } ^ { r \\times d }", "type": "inline_equation" }, { "bbox": [ 419, 402, 507, 419 ], "score": 1.0, "content": ". These adapters are", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 416, 349, 428 ], "spans": [ { "bbox": [ 106, 416, 349, 428 ], "score": 1.0, "content": "surrounded by a residual connection, leading to a final form:", "type": "text" } ], "index": 24 } ], "index": 22, "bbox_fs": [ 103, 372, 507, 428 ] }, { "type": "interline_equation", "bbox": [ 249, 430, 361, 444 ], "lines": [ { "bbox": [ 249, 430, 361, 444 ], "spans": [ { "bbox": [ 249, 430, 361, 444 ], "score": 0.92, "content": "h h + f ( h W _ { \\mathrm { d o w n } } ) W _ { \\mathrm { u p } } .", "type": "interline_equation", "image_path": "d051c5840ac628806d97ea98a2dc823b86c10625a3c67375d4f75bab612e1628.jpg" } ] } ], "index": 25, "virtual_lines": [ { "bbox": [ 249, 430, 361, 444 ], "spans": [], "index": 25 } ] }, { "type": "text", "bbox": [ 106, 448, 505, 482 ], "lines": [ { "bbox": [ 105, 447, 506, 461 ], "spans": [ { "bbox": [ 105, 447, 506, 461 ], "score": 1.0, "content": "Houlsby et al. (2019) places two adapters sequentially within one layer of the transformer, one after", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 458, 505, 471 ], "spans": [ { "bbox": [ 105, 458, 505, 471 ], "score": 1.0, "content": "the multi-head attention and one after the FFN sub-layer. Pfeiffer et al. (2021) have proposed a more", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 469, 463, 483 ], "spans": [ { "bbox": [ 105, 469, 463, 483 ], "score": 1.0, "content": "efficient adapter variant that is inserted only after the FFN “add & layer norm” sub-layer.", "type": "text" } ], "index": 28 } ], "index": 27, "bbox_fs": [ 105, 447, 506, 483 ] }, { "type": "text", "bbox": [ 106, 486, 505, 542 ], "lines": [ { "bbox": [ 106, 487, 505, 499 ], "spans": [ { "bbox": [ 106, 487, 505, 499 ], "score": 1.0, "content": "Prefix Tuning (Li & Liang, 2021): Inspired by the success of textual prompting methods (Liu", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 497, 504, 510 ], "spans": [ { "bbox": [ 105, 497, 260, 510 ], "score": 1.0, "content": "et al., 2021a), prefix tuning prepends", "type": "text" }, { "bbox": [ 260, 498, 265, 508 ], "score": 0.64, "content": "l", "type": "inline_equation" }, { "bbox": [ 265, 497, 504, 510 ], "score": 1.0, "content": "tunable prefix vectors to the keys and values of the multi-", "type": "text" } ], "index": 30 }, { "bbox": [ 104, 508, 506, 522 ], "spans": [ { "bbox": [ 104, 508, 373, 522 ], "score": 1.0, "content": "head attention at every layer. Specifically, two sets of prefix vectors", "type": "text" }, { "bbox": [ 373, 508, 435, 520 ], "score": 0.92, "content": "P _ { k } , \\dot { P } _ { v } \\in \\mathbb R ^ { l \\times d }", "type": "inline_equation" }, { "bbox": [ 436, 508, 506, 522 ], "score": 1.0, "content": "are concatenated", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 519, 506, 532 ], "spans": [ { "bbox": [ 106, 519, 192, 532 ], "score": 1.0, "content": "with the original key", "type": "text" }, { "bbox": [ 193, 520, 204, 530 ], "score": 0.75, "content": "\\kappa", "type": "inline_equation" }, { "bbox": [ 205, 519, 247, 532 ], "score": 1.0, "content": "and value", "type": "text" }, { "bbox": [ 247, 520, 257, 530 ], "score": 0.76, "content": "V", "type": "inline_equation" }, { "bbox": [ 258, 519, 506, 532 ], "score": 1.0, "content": ". Then multi-head attention is performed on the new prefixed", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 529, 356, 544 ], "spans": [ { "bbox": [ 105, 529, 255, 544 ], "score": 1.0, "content": "keys and values. The computation of", "type": "text" }, { "bbox": [ 256, 531, 281, 542 ], "score": 0.84, "content": "{ \\mathrm { h e a d } } _ { i }", "type": "inline_equation" }, { "bbox": [ 281, 529, 356, 544 ], "score": 1.0, "content": "in Eq. 2 becomes:", "type": "text" } ], "index": 33 } ], "index": 31, "bbox_fs": [ 104, 487, 506, 544 ] }, { "type": "interline_equation", "bbox": [ 163, 545, 446, 563 ], "lines": [ { "bbox": [ 163, 545, 446, 563 ], "spans": [ { "bbox": [ 163, 545, 446, 563 ], "score": 0.9, "content": "\\mathrm { h e a d } _ { i } = \\mathrm { A t t n } ( \\pmb { x } \\pmb { W } _ { q } ^ { ( i ) } , \\mathrm { c o n c a t } ( \\pmb { P } _ { k } ^ { ( i ) } , \\pmb { C } \\pmb { W } _ { k } ^ { ( i ) } ) , \\mathrm { c o n c a t } ( \\pmb { P } _ { v } ^ { ( i ) } , \\pmb { C } \\pmb { W } _ { v } ^ { ( i ) } ) ) ,", "type": "interline_equation", "image_path": "3b583e4542c85d5a77c153f9671be5620121202770b5f0460f1bd76e61784b91.jpg" } ] } ], "index": 34, "virtual_lines": [ { "bbox": [ 163, 545, 446, 563 ], "spans": [], "index": 34 } ] }, { "type": "text", "bbox": [ 107, 567, 505, 603 ], "lines": [ { "bbox": [ 107, 559, 506, 588 ], "spans": [ { "bbox": [ 107, 570, 120, 581 ], "score": 0.87, "content": "P _ { k }", "type": "inline_equation" }, { "bbox": [ 120, 559, 139, 588 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 140, 570, 153, 581 ], "score": 0.88, "content": "P _ { v }", "type": "inline_equation" }, { "bbox": [ 153, 559, 209, 588 ], "score": 1.0, "content": "are split into", "type": "text" }, { "bbox": [ 209, 570, 223, 581 ], "score": 0.9, "content": "N _ { h }", "type": "inline_equation" }, { "bbox": [ 224, 559, 348, 588 ], "score": 1.0, "content": "head vectors respectively and", "type": "text" }, { "bbox": [ 348, 566, 440, 582 ], "score": 0.89, "content": "P _ { k } ^ { ( i ) } , P _ { v } ^ { ( i ) } \\in \\mathbb R ^ { l \\times d / N _ { h } }", "type": "inline_equation" }, { "bbox": [ 441, 559, 488, 588 ], "score": 1.0, "content": "denote the", "type": "text" }, { "bbox": [ 488, 570, 493, 579 ], "score": 0.74, "content": "i", "type": "inline_equation" }, { "bbox": [ 493, 559, 506, 588 ], "score": 1.0, "content": "-th", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 580, 505, 593 ], "spans": [ { "bbox": [ 105, 580, 505, 593 ], "score": 1.0, "content": "head vector. Prompt-tuning (Lester et al., 2021) simplifies prefix-tuning by only prepending to the", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 591, 492, 604 ], "spans": [ { "bbox": [ 106, 591, 378, 604 ], "score": 1.0, "content": "input word embeddings in the first layer; similar work also includes", "type": "text" }, { "bbox": [ 378, 592, 385, 601 ], "score": 0.37, "content": "\\mathrm { \\bf P }", "type": "inline_equation" }, { "bbox": [ 386, 591, 492, 604 ], "score": 1.0, "content": "-tuning (Liu et al., 2021b).", "type": "text" } ], "index": 37 } ], "index": 36, "bbox_fs": [ 105, 559, 506, 604 ] }, { "type": "text", "bbox": [ 106, 608, 505, 676 ], "lines": [ { "bbox": [ 105, 607, 506, 621 ], "spans": [ { "bbox": [ 105, 607, 506, 621 ], "score": 1.0, "content": "LoRA (Hu et al., 2021): LoRA injects trainable low-rank matrices into transformer layers to", "type": "text" } ], "index": 38 }, { "bbox": [ 104, 617, 506, 633 ], "spans": [ { "bbox": [ 104, 617, 369, 633 ], "score": 1.0, "content": "approximate the weight updates. For a pre-trained weight matrix", "type": "text" }, { "bbox": [ 370, 618, 418, 630 ], "score": 0.92, "content": "W \\in \\mathbb { R } ^ { d \\times k }", "type": "inline_equation" }, { "bbox": [ 419, 617, 506, 633 ], "score": 1.0, "content": ", LoRA represents its", "type": "text" } ], "index": 39 }, { "bbox": [ 104, 626, 505, 646 ], "spans": [ { "bbox": [ 104, 626, 260, 646 ], "score": 1.0, "content": "update with a low-rank decomposition", "type": "text" }, { "bbox": [ 260, 630, 380, 642 ], "score": 0.9, "content": "W + \\Delta W = W + W _ { \\mathrm { d o w n } } W _ { \\mathrm { u p } }", "type": "inline_equation" }, { "bbox": [ 380, 626, 410, 646 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 410, 630, 471, 642 ], "score": 0.88, "content": "W _ { \\mathrm { d o w n } } \\in \\mathbb { R } ^ { \\hat { d } \\times r }", "type": "inline_equation" }, { "bbox": [ 471, 626, 475, 646 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 475, 630, 505, 643 ], "score": 0.84, "content": "W _ { \\mathrm { u p } } \\in", "type": "inline_equation" } ], "index": 40 }, { "bbox": [ 106, 639, 506, 658 ], "spans": [ { "bbox": [ 106, 641, 130, 653 ], "score": 0.9, "content": "\\mathbb { R } ^ { r \\times k }", "type": "inline_equation" }, { "bbox": [ 130, 639, 506, 658 ], "score": 1.0, "content": "are tunable parameters. LoRA applies this update to the query and value projection matrices", "type": "text" } ], "index": 41 }, { "bbox": [ 107, 653, 506, 667 ], "spans": [ { "bbox": [ 107, 654, 150, 666 ], "score": 0.9, "content": "\\left( W _ { q } , W _ { v } \\right)", "type": "inline_equation" }, { "bbox": [ 151, 653, 470, 667 ], "score": 1.0, "content": "in the multi-head attention sub-layer, as shown in Figure 1. For a specific input", "type": "text" }, { "bbox": [ 471, 656, 479, 664 ], "score": 0.79, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 479, 653, 506, 667 ], "score": 1.0, "content": "to the", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 665, 442, 677 ], "spans": [ { "bbox": [ 106, 665, 420, 677 ], "score": 1.0, "content": "linear projection in multi-head attention, LoRA modifies the projection output", "type": "text" }, { "bbox": [ 420, 666, 428, 675 ], "score": 0.82, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 428, 665, 442, 677 ], "score": 1.0, "content": "as:", "type": "text" } ], "index": 43 } ], "index": 40.5, "bbox_fs": [ 104, 607, 506, 677 ] }, { "type": "interline_equation", "bbox": [ 250, 680, 360, 694 ], "lines": [ { "bbox": [ 250, 680, 360, 694 ], "spans": [ { "bbox": [ 250, 680, 360, 694 ], "score": 0.92, "content": "h h + s \\cdot x W _ { \\mathrm { d o w n } } W _ { \\mathrm { u p } } ,", "type": "interline_equation", "image_path": "089f93b64428c027bf2a950afbf1f0c71a98197bfdbfdd26d5231d9173a4f135.jpg" } ] } ], "index": 44, "virtual_lines": [ { "bbox": [ 250, 680, 360, 694 ], "spans": [], "index": 44 } ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 106, 78, 502, 158 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 78, 502, 158 ], "group_id": 0, "lines": [ { "bbox": [ 106, 78, 502, 158 ], "spans": [ { "bbox": [ 106, 78, 502, 158 ], "score": 0.968, "type": "image", "image_path": "0cae0143a371d3d20d802c7754661581469a23b2c5a99fc1b9f5bba922f4f87f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 78, 502, 104.66666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 106, 104.66666666666667, 502, 131.33333333333334 ], "spans": [], "index": 1 }, { "bbox": [ 106, 131.33333333333334, 502, 158.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 161, 502, 192 ], "group_id": 0, "lines": [ { "bbox": [ 106, 161, 505, 173 ], "spans": [ { "bbox": [ 106, 161, 505, 173 ], "score": 1.0, "content": "Figure 3: Graphical illustration of existing methods and the proposed variants. “PLM module” represents a", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 171, 504, 183 ], "spans": [ { "bbox": [ 105, 171, 504, 183 ], "score": 1.0, "content": "certain sublayer of the PLM (e.g. attention or FFN) that is frozen. “Scaled PA” denotes scaled parallel adapter.", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 181, 343, 193 ], "spans": [ { "bbox": [ 106, 181, 343, 193 ], "score": 1.0, "content": "We do not include multi-head parallel adapter here to save space.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "text", "bbox": [ 106, 200, 301, 212 ], "lines": [ { "bbox": [ 106, 199, 303, 214 ], "spans": [ { "bbox": [ 106, 199, 133, 214 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 133, 201, 158, 212 ], "score": 0.89, "content": "s \\geq 1", "type": "inline_equation" }, { "bbox": [ 158, 199, 303, 214 ], "score": 1.0, "content": "is a tunable scalar hyperparameter.4", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 106, 217, 506, 251 ], "lines": [ { "bbox": [ 105, 217, 506, 230 ], "spans": [ { "bbox": [ 105, 217, 506, 230 ], "score": 1.0, "content": "Others: Other parameter-efficient tuning methods include BitFit (Ben Zaken et al., 2021), which", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 228, 505, 241 ], "spans": [ { "bbox": [ 106, 228, 505, 241 ], "score": 1.0, "content": "only fine-tunes bias vectors in the pre-trained model, and diff-pruning (Guo et al., 2021), which", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 239, 267, 253 ], "spans": [ { "bbox": [ 105, 239, 267, 253 ], "score": 1.0, "content": "learns a sparse parameter update vector.", "type": "text" } ], "index": 9 } ], "index": 8 }, { "type": "title", "bbox": [ 107, 266, 329, 279 ], "lines": [ { "bbox": [ 104, 264, 330, 281 ], "spans": [ { "bbox": [ 104, 264, 330, 281 ], "score": 1.0, "content": "3 BRIDGING THE GAP – A UNIFIED VIEW", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 106, 292, 505, 325 ], "lines": [ { "bbox": [ 106, 292, 505, 304 ], "spans": [ { "bbox": [ 106, 292, 505, 304 ], "score": 1.0, "content": "We first derive an equivalent form of prefix tuning to establish its connection with adapters. We", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 303, 505, 316 ], "spans": [ { "bbox": [ 106, 303, 505, 316 ], "score": 1.0, "content": "then propose a unified framework for parameter-efficient tuning that includes several state-of-the-art", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 314, 212, 326 ], "spans": [ { "bbox": [ 106, 314, 212, 326 ], "score": 1.0, "content": "methods as instantiations.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "title", "bbox": [ 107, 338, 289, 350 ], "lines": [ { "bbox": [ 106, 338, 289, 351 ], "spans": [ { "bbox": [ 106, 338, 289, 351 ], "score": 1.0, "content": "3.1 A CLOSER LOOK AT PREFIX TUNING", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 358, 504, 393 ], "lines": [ { "bbox": [ 106, 359, 505, 371 ], "spans": [ { "bbox": [ 106, 359, 505, 371 ], "score": 1.0, "content": "Eq. 5 describes the mechanism of prefix tuning which changes the attention module through", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 370, 506, 383 ], "spans": [ { "bbox": [ 105, 370, 154, 383 ], "score": 1.0, "content": "prepending", "type": "text" }, { "bbox": [ 154, 371, 159, 380 ], "score": 0.62, "content": "l", "type": "inline_equation" }, { "bbox": [ 159, 370, 506, 383 ], "score": 1.0, "content": "learnable vectors to the original attention keys and values. Here, we derive an equiva-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 380, 378, 395 ], "spans": [ { "bbox": [ 105, 380, 378, 395 ], "score": 1.0, "content": "lent form of Eq. 5 and provide an alternative view of prefix tuning:5", "type": "text" } ], "index": 17 } ], "index": 16 }, { "type": "interline_equation", "bbox": [ 153, 396, 457, 484 ], "lines": [ { "bbox": [ 153, 396, 457, 484 ], "spans": [ { "bbox": [ 153, 396, 457, 484 ], "score": 0.94, "content": "\\begin{array} { r l } & { \\mathrm { h e a d } = \\mathrm { A t } \\mathrm { t n } ( x W _ { q } , \\mathrm { c o n c a t } ( P _ { k } , C W _ { k } ) , \\mathrm { c o n c a t } ( P _ { v } , C W _ { v } ) ) } \\\\ & { \\ = \\mathrm { s o f t m a x } \\big ( x W _ { q } \\mathrm { c o n c a t } ( P _ { k } , C W _ { k } ) ^ { \\top } \\big ) \\Big [ \\begin{array} { l } { P _ { v } } \\\\ { C W _ { v } } \\end{array} \\Big ] } \\\\ & { \\ = ( 1 - \\lambda ( \\pmb { x } ) ) \\mathrm { s o f t m a x } ( { \\pmb x } W _ { q } W _ { k } ^ { \\top } C ^ { \\top } ) C W _ { v } + \\lambda ( { \\pmb x } ) \\mathrm { s o f t m a x } ( { \\pmb x } W _ { q } P _ { k } ^ { \\top } ) P _ { v } } \\\\ & { \\ = ( 1 - \\lambda ( \\pmb { x } ) ) \\underbrace { \\mathrm { A t t n } ( { \\pmb x } W _ { q } , C W _ { k } , C W _ { v } ) } _ { \\mathrm { s t a n d a r d a t e n t i o n } } + \\lambda ( \\pmb { x } ) \\underbrace { \\mathrm { A t t n } ( { \\pmb x } W _ { q } , P _ { k } , P _ { v } ) } _ { \\mathrm { i n d e p e n d e n t o f } C } , } \\end{array}", "type": "interline_equation", "image_path": "8ef42e61f3f24f6dff976b460e696011d571e0cf7e273a9d7d42598b0f57144f.jpg" } ] } ], "index": 19, "virtual_lines": [ { "bbox": [ 153, 396, 457, 425.3333333333333 ], "spans": [], "index": 18 }, { "bbox": [ 153, 425.3333333333333, 457, 454.66666666666663 ], "spans": [], "index": 19 }, { "bbox": [ 153, 454.66666666666663, 457, 483.99999999999994 ], "spans": [], "index": 20 } ] }, { "type": "text", "bbox": [ 104, 488, 485, 500 ], "lines": [ { "bbox": [ 106, 488, 484, 502 ], "spans": [ { "bbox": [ 106, 488, 133, 502 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 133, 489, 154, 501 ], "score": 0.92, "content": "\\lambda ( { \\pmb x } )", "type": "inline_equation" }, { "bbox": [ 155, 488, 484, 502 ], "score": 1.0, "content": "is a scalar that represents the sum of normalized attention weights on the prefixes:", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "interline_equation", "bbox": [ 189, 504, 422, 535 ], "lines": [ { "bbox": [ 189, 504, 422, 535 ], "spans": [ { "bbox": [ 189, 504, 422, 535 ], "score": 0.94, "content": "\\lambda ( \\pmb { x } ) = \\frac { \\sum _ { i } \\exp ( \\pmb { x } \\pmb { W _ { q } } \\pmb { P } _ { k } ^ { \\top } ) _ { i } } { \\sum _ { i } \\exp ( \\pmb { x } \\pmb { W _ { q } } \\pmb { P } _ { k } ^ { \\top } ) _ { i } + \\sum _ { j } \\exp ( \\pmb { x } \\pmb { W _ { q } } \\pmb { W } _ { k } ^ { \\top } \\pmb { C } ^ { \\top } ) _ { j } } .", "type": "interline_equation", "image_path": "64b10e3182022eabd7ae3d8dd60a2f0cae8790691bb51604f897bb0d6bf3728b.jpg" } ] } ], "index": 22.5, "virtual_lines": [ { "bbox": [ 189, 504, 422, 519.5 ], "spans": [], "index": 22 }, { "bbox": [ 189, 519.5, 422, 535.0 ], "spans": [], "index": 23 } ] }, { "type": "text", "bbox": [ 107, 538, 505, 584 ], "lines": [ { "bbox": [ 105, 538, 506, 552 ], "spans": [ { "bbox": [ 105, 538, 232, 552 ], "score": 1.0, "content": "Note that the first term in Eq. 7,", "type": "text" }, { "bbox": [ 233, 539, 339, 552 ], "score": 0.88, "content": "\\mathrm { A t t n } ( x W _ { q } , C W _ { k } , C W _ { v } )", "type": "inline_equation" }, { "bbox": [ 339, 538, 506, 552 ], "score": 1.0, "content": ", is the original attention without prefixes,", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 551, 505, 563 ], "spans": [ { "bbox": [ 106, 551, 397, 563 ], "score": 1.0, "content": "whereas the second term is a position-wise modification independent of", "type": "text" }, { "bbox": [ 397, 551, 407, 560 ], "score": 0.81, "content": "C", "type": "inline_equation" }, { "bbox": [ 407, 551, 505, 563 ], "score": 1.0, "content": ". Eq. 7 gives an alterna-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 561, 505, 574 ], "spans": [ { "bbox": [ 105, 561, 505, 574 ], "score": 1.0, "content": "tive view of prefix tuning that essentially applies a position-wise modification to the original head", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 573, 296, 585 ], "spans": [ { "bbox": [ 106, 573, 172, 585 ], "score": 1.0, "content": "attention output", "type": "text" }, { "bbox": [ 172, 573, 180, 582 ], "score": 0.79, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 180, 573, 296, 585 ], "score": 1.0, "content": "through linear interpolation:", "type": "text" } ], "index": 27 } ], "index": 25.5 }, { "type": "interline_equation", "bbox": [ 174, 588, 437, 603 ], "lines": [ { "bbox": [ 174, 588, 437, 603 ], "spans": [ { "bbox": [ 174, 588, 437, 603 ], "score": 0.86, "content": "\\begin{array} { r } { \\pmb { h } ( 1 - \\lambda ( \\pmb { x } ) ) \\pmb { h } + \\lambda ( \\pmb { x } ) \\Delta \\pmb { h } , \\quad \\Delta \\pmb { h } : = \\mathrm { s o f t m a x } ( \\pmb { x } \\pmb { W } _ { q } \\pmb { P } _ { k } ^ { \\top } ) \\pmb { P } _ { v } . } \\end{array}", "type": "interline_equation", "image_path": "649caf7eb49868967eb23e03d6682dac9585b1e77a503e63210bf71549b3ac19.jpg" } ] } ], "index": 28, "virtual_lines": [ { "bbox": [ 174, 588, 437, 603 ], "spans": [], "index": 28 } ] }, { "type": "text", "bbox": [ 106, 613, 504, 627 ], "lines": [ { "bbox": [ 106, 612, 505, 628 ], "spans": [ { "bbox": [ 106, 612, 291, 628 ], "score": 1.0, "content": "The Connection with Adapters: We define", "type": "text" }, { "bbox": [ 291, 614, 343, 627 ], "score": 0.91, "content": "W _ { 1 } { = } W _ { q } P _ { k } ^ { \\top }", "type": "inline_equation" }, { "bbox": [ 343, 612, 347, 628 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 347, 614, 380, 626 ], "score": 0.86, "content": "W _ { 2 } { = } P _ { v }", "type": "inline_equation" }, { "bbox": [ 381, 612, 384, 628 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 385, 615, 393, 627 ], "score": 0.52, "content": "f \\colon", "type": "inline_equation" }, { "bbox": [ 393, 612, 505, 628 ], "score": 1.0, "content": "=softmax, and rewrite Eq. 9:", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "interline_equation", "bbox": [ 224, 631, 386, 645 ], "lines": [ { "bbox": [ 224, 631, 386, 645 ], "spans": [ { "bbox": [ 224, 631, 386, 645 ], "score": 0.9, "content": "\\begin{array} { r } { \\pmb { h } ( 1 - \\lambda ( \\pmb { x } ) ) \\pmb { h } + \\lambda ( \\pmb { x } ) f ( \\pmb { x } \\pmb { W } _ { 1 } ) \\pmb { W } _ { 2 } , } \\end{array}", "type": "interline_equation", "image_path": "e48ece8aaeae6c4520b396dbfbfa85c3cac2a8997db2274fc247664c93254df9.jpg" } ] } ], "index": 30, "virtual_lines": [ { "bbox": [ 224, 631, 386, 645 ], "spans": [], "index": 30 } ] }, { "type": "text", "bbox": [ 107, 649, 505, 671 ], "lines": [ { "bbox": [ 105, 648, 505, 662 ], "spans": [ { "bbox": [ 105, 648, 505, 662 ], "score": 1.0, "content": "which reaches a very similar form to the adapter function in Eq. 4, except that prefix tuning is", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 659, 506, 672 ], "spans": [ { "bbox": [ 105, 659, 506, 672 ], "score": 1.0, "content": "performing weighted addition while the adapter one is unweighted.6 Figure 3b demonstrates the", "type": "text" } ], "index": 32 } ], "index": 31.5 } ], "page_idx": 3, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 105, 25, 294, 39 ], "spans": [ { "bbox": [ 105, 25, 294, 39 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 106, 678, 505, 732 ], "lines": [ { "bbox": [ 118, 677, 506, 691 ], "spans": [ { "bbox": [ 118, 677, 407, 691 ], "score": 1.0, "content": "4The public code of LoRA at https://github.com/microsoft/LoRA uses different", "type": "text" }, { "bbox": [ 407, 681, 413, 688 ], "score": 0.71, "content": "s", "type": "inline_equation" }, { "bbox": [ 413, 677, 506, 691 ], "score": 1.0, "content": "in different datasets, and", "type": "text" } ] }, { "bbox": [ 106, 689, 379, 700 ], "spans": [ { "bbox": [ 106, 689, 212, 700 ], "score": 1.0, "content": "we have verified the value of", "type": "text" }, { "bbox": [ 212, 691, 218, 698 ], "score": 0.74, "content": "s", "type": "inline_equation" }, { "bbox": [ 218, 689, 379, 700 ], "score": 1.0, "content": "could have a significant effect on the results.", "type": "text" } ] }, { "bbox": [ 117, 698, 457, 713 ], "spans": [ { "bbox": [ 117, 698, 368, 713 ], "score": 1.0, "content": "5Without loss of generalization, we ignore the softmax scaling factor", "type": "text" }, { "bbox": [ 369, 700, 382, 710 ], "score": 0.9, "content": "\\sqrt { d }", "type": "inline_equation" }, { "bbox": [ 382, 698, 457, 713 ], "score": 1.0, "content": "for ease of notation.", "type": "text" } ] }, { "bbox": [ 119, 708, 506, 725 ], "spans": [ { "bbox": [ 119, 711, 130, 721 ], "score": 0.83, "content": "^ 6 h", "type": "inline_equation" }, { "bbox": [ 131, 708, 506, 725 ], "score": 1.0, "content": "in adapters and prefix tuning are usually different, as described more below. However, here we mainly", "type": "text" } ] }, { "bbox": [ 105, 721, 408, 733 ], "spans": [ { "bbox": [ 105, 721, 408, 733 ], "score": 1.0, "content": "discuss the functional form as adapters can, in principle, be inserted at any position.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 759 ], "lines": [ { "bbox": [ 301, 750, 310, 762 ], "spans": [ { "bbox": [ 301, 750, 310, 762 ], "score": 1.0, "content": "4", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 106, 78, 502, 158 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 78, 502, 158 ], "group_id": 0, "lines": [ { "bbox": [ 106, 78, 502, 158 ], "spans": [ { "bbox": [ 106, 78, 502, 158 ], "score": 0.968, "type": "image", "image_path": "0cae0143a371d3d20d802c7754661581469a23b2c5a99fc1b9f5bba922f4f87f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 78, 502, 104.66666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 106, 104.66666666666667, 502, 131.33333333333334 ], "spans": [], "index": 1 }, { "bbox": [ 106, 131.33333333333334, 502, 158.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 161, 502, 192 ], "group_id": 0, "lines": [ { "bbox": [ 106, 161, 505, 173 ], "spans": [ { "bbox": [ 106, 161, 505, 173 ], "score": 1.0, "content": "Figure 3: Graphical illustration of existing methods and the proposed variants. “PLM module” represents a", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 171, 504, 183 ], "spans": [ { "bbox": [ 105, 171, 504, 183 ], "score": 1.0, "content": "certain sublayer of the PLM (e.g. attention or FFN) that is frozen. “Scaled PA” denotes scaled parallel adapter.", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 181, 343, 193 ], "spans": [ { "bbox": [ 106, 181, 343, 193 ], "score": 1.0, "content": "We do not include multi-head parallel adapter here to save space.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "text", "bbox": [ 106, 200, 301, 212 ], "lines": [ { "bbox": [ 106, 199, 303, 214 ], "spans": [ { "bbox": [ 106, 199, 133, 214 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 133, 201, 158, 212 ], "score": 0.89, "content": "s \\geq 1", "type": "inline_equation" }, { "bbox": [ 158, 199, 303, 214 ], "score": 1.0, "content": "is a tunable scalar hyperparameter.4", "type": "text" } ], "index": 6 } ], "index": 6, "bbox_fs": [ 106, 199, 303, 214 ] }, { "type": "text", "bbox": [ 106, 217, 506, 251 ], "lines": [ { "bbox": [ 105, 217, 506, 230 ], "spans": [ { "bbox": [ 105, 217, 506, 230 ], "score": 1.0, "content": "Others: Other parameter-efficient tuning methods include BitFit (Ben Zaken et al., 2021), which", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 228, 505, 241 ], "spans": [ { "bbox": [ 106, 228, 505, 241 ], "score": 1.0, "content": "only fine-tunes bias vectors in the pre-trained model, and diff-pruning (Guo et al., 2021), which", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 239, 267, 253 ], "spans": [ { "bbox": [ 105, 239, 267, 253 ], "score": 1.0, "content": "learns a sparse parameter update vector.", "type": "text" } ], "index": 9 } ], "index": 8, "bbox_fs": [ 105, 217, 506, 253 ] }, { "type": "title", "bbox": [ 107, 266, 329, 279 ], "lines": [ { "bbox": [ 104, 264, 330, 281 ], "spans": [ { "bbox": [ 104, 264, 330, 281 ], "score": 1.0, "content": "3 BRIDGING THE GAP – A UNIFIED VIEW", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 106, 292, 505, 325 ], "lines": [ { "bbox": [ 106, 292, 505, 304 ], "spans": [ { "bbox": [ 106, 292, 505, 304 ], "score": 1.0, "content": "We first derive an equivalent form of prefix tuning to establish its connection with adapters. We", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 303, 505, 316 ], "spans": [ { "bbox": [ 106, 303, 505, 316 ], "score": 1.0, "content": "then propose a unified framework for parameter-efficient tuning that includes several state-of-the-art", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 314, 212, 326 ], "spans": [ { "bbox": [ 106, 314, 212, 326 ], "score": 1.0, "content": "methods as instantiations.", "type": "text" } ], "index": 13 } ], "index": 12, "bbox_fs": [ 106, 292, 505, 326 ] }, { "type": "title", "bbox": [ 107, 338, 289, 350 ], "lines": [ { "bbox": [ 106, 338, 289, 351 ], "spans": [ { "bbox": [ 106, 338, 289, 351 ], "score": 1.0, "content": "3.1 A CLOSER LOOK AT PREFIX TUNING", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 358, 504, 393 ], "lines": [ { "bbox": [ 106, 359, 505, 371 ], "spans": [ { "bbox": [ 106, 359, 505, 371 ], "score": 1.0, "content": "Eq. 5 describes the mechanism of prefix tuning which changes the attention module through", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 370, 506, 383 ], "spans": [ { "bbox": [ 105, 370, 154, 383 ], "score": 1.0, "content": "prepending", "type": "text" }, { "bbox": [ 154, 371, 159, 380 ], "score": 0.62, "content": "l", "type": "inline_equation" }, { "bbox": [ 159, 370, 506, 383 ], "score": 1.0, "content": "learnable vectors to the original attention keys and values. Here, we derive an equiva-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 380, 378, 395 ], "spans": [ { "bbox": [ 105, 380, 378, 395 ], "score": 1.0, "content": "lent form of Eq. 5 and provide an alternative view of prefix tuning:5", "type": "text" } ], "index": 17 } ], "index": 16, "bbox_fs": [ 105, 359, 506, 395 ] }, { "type": "interline_equation", "bbox": [ 153, 396, 457, 484 ], "lines": [ { "bbox": [ 153, 396, 457, 484 ], "spans": [ { "bbox": [ 153, 396, 457, 484 ], "score": 0.94, "content": "\\begin{array} { r l } & { \\mathrm { h e a d } = \\mathrm { A t } \\mathrm { t n } ( x W _ { q } , \\mathrm { c o n c a t } ( P _ { k } , C W _ { k } ) , \\mathrm { c o n c a t } ( P _ { v } , C W _ { v } ) ) } \\\\ & { \\ = \\mathrm { s o f t m a x } \\big ( x W _ { q } \\mathrm { c o n c a t } ( P _ { k } , C W _ { k } ) ^ { \\top } \\big ) \\Big [ \\begin{array} { l } { P _ { v } } \\\\ { C W _ { v } } \\end{array} \\Big ] } \\\\ & { \\ = ( 1 - \\lambda ( \\pmb { x } ) ) \\mathrm { s o f t m a x } ( { \\pmb x } W _ { q } W _ { k } ^ { \\top } C ^ { \\top } ) C W _ { v } + \\lambda ( { \\pmb x } ) \\mathrm { s o f t m a x } ( { \\pmb x } W _ { q } P _ { k } ^ { \\top } ) P _ { v } } \\\\ & { \\ = ( 1 - \\lambda ( \\pmb { x } ) ) \\underbrace { \\mathrm { A t t n } ( { \\pmb x } W _ { q } , C W _ { k } , C W _ { v } ) } _ { \\mathrm { s t a n d a r d a t e n t i o n } } + \\lambda ( \\pmb { x } ) \\underbrace { \\mathrm { A t t n } ( { \\pmb x } W _ { q } , P _ { k } , P _ { v } ) } _ { \\mathrm { i n d e p e n d e n t o f } C } , } \\end{array}", "type": "interline_equation", "image_path": "8ef42e61f3f24f6dff976b460e696011d571e0cf7e273a9d7d42598b0f57144f.jpg" } ] } ], "index": 19, "virtual_lines": [ { "bbox": [ 153, 396, 457, 425.3333333333333 ], "spans": [], "index": 18 }, { "bbox": [ 153, 425.3333333333333, 457, 454.66666666666663 ], "spans": [], "index": 19 }, { "bbox": [ 153, 454.66666666666663, 457, 483.99999999999994 ], "spans": [], "index": 20 } ] }, { "type": "text", "bbox": [ 104, 488, 485, 500 ], "lines": [ { "bbox": [ 106, 488, 484, 502 ], "spans": [ { "bbox": [ 106, 488, 133, 502 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 133, 489, 154, 501 ], "score": 0.92, "content": "\\lambda ( { \\pmb x } )", "type": "inline_equation" }, { "bbox": [ 155, 488, 484, 502 ], "score": 1.0, "content": "is a scalar that represents the sum of normalized attention weights on the prefixes:", "type": "text" } ], "index": 21 } ], "index": 21, "bbox_fs": [ 106, 488, 484, 502 ] }, { "type": "interline_equation", "bbox": [ 189, 504, 422, 535 ], "lines": [ { "bbox": [ 189, 504, 422, 535 ], "spans": [ { "bbox": [ 189, 504, 422, 535 ], "score": 0.94, "content": "\\lambda ( \\pmb { x } ) = \\frac { \\sum _ { i } \\exp ( \\pmb { x } \\pmb { W _ { q } } \\pmb { P } _ { k } ^ { \\top } ) _ { i } } { \\sum _ { i } \\exp ( \\pmb { x } \\pmb { W _ { q } } \\pmb { P } _ { k } ^ { \\top } ) _ { i } + \\sum _ { j } \\exp ( \\pmb { x } \\pmb { W _ { q } } \\pmb { W } _ { k } ^ { \\top } \\pmb { C } ^ { \\top } ) _ { j } } .", "type": "interline_equation", "image_path": "64b10e3182022eabd7ae3d8dd60a2f0cae8790691bb51604f897bb0d6bf3728b.jpg" } ] } ], "index": 22.5, "virtual_lines": [ { "bbox": [ 189, 504, 422, 519.5 ], "spans": [], "index": 22 }, { "bbox": [ 189, 519.5, 422, 535.0 ], "spans": [], "index": 23 } ] }, { "type": "text", "bbox": [ 107, 538, 505, 584 ], "lines": [ { "bbox": [ 105, 538, 506, 552 ], "spans": [ { "bbox": [ 105, 538, 232, 552 ], "score": 1.0, "content": "Note that the first term in Eq. 7,", "type": "text" }, { "bbox": [ 233, 539, 339, 552 ], "score": 0.88, "content": "\\mathrm { A t t n } ( x W _ { q } , C W _ { k } , C W _ { v } )", "type": "inline_equation" }, { "bbox": [ 339, 538, 506, 552 ], "score": 1.0, "content": ", is the original attention without prefixes,", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 551, 505, 563 ], "spans": [ { "bbox": [ 106, 551, 397, 563 ], "score": 1.0, "content": "whereas the second term is a position-wise modification independent of", "type": "text" }, { "bbox": [ 397, 551, 407, 560 ], "score": 0.81, "content": "C", "type": "inline_equation" }, { "bbox": [ 407, 551, 505, 563 ], "score": 1.0, "content": ". Eq. 7 gives an alterna-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 561, 505, 574 ], "spans": [ { "bbox": [ 105, 561, 505, 574 ], "score": 1.0, "content": "tive view of prefix tuning that essentially applies a position-wise modification to the original head", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 573, 296, 585 ], "spans": [ { "bbox": [ 106, 573, 172, 585 ], "score": 1.0, "content": "attention output", "type": "text" }, { "bbox": [ 172, 573, 180, 582 ], "score": 0.79, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 180, 573, 296, 585 ], "score": 1.0, "content": "through linear interpolation:", "type": "text" } ], "index": 27 } ], "index": 25.5, "bbox_fs": [ 105, 538, 506, 585 ] }, { "type": "interline_equation", "bbox": [ 174, 588, 437, 603 ], "lines": [ { "bbox": [ 174, 588, 437, 603 ], "spans": [ { "bbox": [ 174, 588, 437, 603 ], "score": 0.86, "content": "\\begin{array} { r } { \\pmb { h } ( 1 - \\lambda ( \\pmb { x } ) ) \\pmb { h } + \\lambda ( \\pmb { x } ) \\Delta \\pmb { h } , \\quad \\Delta \\pmb { h } : = \\mathrm { s o f t m a x } ( \\pmb { x } \\pmb { W } _ { q } \\pmb { P } _ { k } ^ { \\top } ) \\pmb { P } _ { v } . } \\end{array}", "type": "interline_equation", "image_path": "649caf7eb49868967eb23e03d6682dac9585b1e77a503e63210bf71549b3ac19.jpg" } ] } ], "index": 28, "virtual_lines": [ { "bbox": [ 174, 588, 437, 603 ], "spans": [], "index": 28 } ] }, { "type": "text", "bbox": [ 106, 613, 504, 627 ], "lines": [ { "bbox": [ 106, 612, 505, 628 ], "spans": [ { "bbox": [ 106, 612, 291, 628 ], "score": 1.0, "content": "The Connection with Adapters: We define", "type": "text" }, { "bbox": [ 291, 614, 343, 627 ], "score": 0.91, "content": "W _ { 1 } { = } W _ { q } P _ { k } ^ { \\top }", "type": "inline_equation" }, { "bbox": [ 343, 612, 347, 628 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 347, 614, 380, 626 ], "score": 0.86, "content": "W _ { 2 } { = } P _ { v }", "type": "inline_equation" }, { "bbox": [ 381, 612, 384, 628 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 385, 615, 393, 627 ], "score": 0.52, "content": "f \\colon", "type": "inline_equation" }, { "bbox": [ 393, 612, 505, 628 ], "score": 1.0, "content": "=softmax, and rewrite Eq. 9:", "type": "text" } ], "index": 29 } ], "index": 29, "bbox_fs": [ 106, 612, 505, 628 ] }, { "type": "interline_equation", "bbox": [ 224, 631, 386, 645 ], "lines": [ { "bbox": [ 224, 631, 386, 645 ], "spans": [ { "bbox": [ 224, 631, 386, 645 ], "score": 0.9, "content": "\\begin{array} { r } { \\pmb { h } ( 1 - \\lambda ( \\pmb { x } ) ) \\pmb { h } + \\lambda ( \\pmb { x } ) f ( \\pmb { x } \\pmb { W } _ { 1 } ) \\pmb { W } _ { 2 } , } \\end{array}", "type": "interline_equation", "image_path": "e48ece8aaeae6c4520b396dbfbfa85c3cac2a8997db2274fc247664c93254df9.jpg" } ] } ], "index": 30, "virtual_lines": [ { "bbox": [ 224, 631, 386, 645 ], "spans": [], "index": 30 } ] }, { "type": "text", "bbox": [ 107, 649, 505, 671 ], "lines": [ { "bbox": [ 105, 648, 505, 662 ], "spans": [ { "bbox": [ 105, 648, 505, 662 ], "score": 1.0, "content": "which reaches a very similar form to the adapter function in Eq. 4, except that prefix tuning is", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 659, 506, 672 ], "spans": [ { "bbox": [ 105, 659, 506, 672 ], "score": 1.0, "content": "performing weighted addition while the adapter one is unweighted.6 Figure 3b demonstrates the", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 223, 505, 236 ], "spans": [ { "bbox": [ 106, 223, 505, 236 ], "score": 1.0, "content": "computation graph of prefix tuning from this view, which allows for abstraction of prefix tuning", "type": "text", "cross_page": true } ], "index": 6 }, { "bbox": [ 105, 231, 507, 247 ], "spans": [ { "bbox": [ 105, 231, 336, 247 ], "score": 1.0, "content": "as a plug-in module like adapters. Further, we note that", "type": "text", "cross_page": true }, { "bbox": [ 336, 234, 392, 245 ], "score": 0.93, "content": "W _ { 1 } \\in \\mathbb { R } ^ { d _ { h } \\times l }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 393, 231, 411, 247 ], "score": 1.0, "content": "and", "type": "text", "cross_page": true }, { "bbox": [ 412, 233, 467, 245 ], "score": 0.94, "content": "W _ { 2 } \\in \\mathbb { R } ^ { l ^ { \\cdot } \\times d _ { h } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 467, 231, 507, 247 ], "score": 1.0, "content": "are low-", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 104, 244, 506, 259 ], "spans": [ { "bbox": [ 104, 244, 190, 259 ], "score": 1.0, "content": "rank matrices when", "type": "text", "cross_page": true }, { "bbox": [ 190, 245, 195, 255 ], "score": 0.53, "content": "l", "type": "inline_equation", "cross_page": true }, { "bbox": [ 196, 244, 398, 259 ], "score": 1.0, "content": "is small, and thus they function similarly to the", "type": "text", "cross_page": true }, { "bbox": [ 398, 245, 426, 256 ], "score": 0.9, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 427, 244, 447, 259 ], "score": 1.0, "content": "and", "type": "text", "cross_page": true }, { "bbox": [ 447, 245, 466, 257 ], "score": 0.89, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 467, 244, 506, 259 ], "score": 1.0, "content": "matrices", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 105, 255, 506, 269 ], "spans": [ { "bbox": [ 105, 255, 402, 269 ], "score": 1.0, "content": "in adapters. This view also suggests that the number of prefix vectors,", "type": "text", "cross_page": true }, { "bbox": [ 402, 257, 407, 266 ], "score": 0.57, "content": "l", "type": "inline_equation", "cross_page": true }, { "bbox": [ 407, 255, 506, 269 ], "score": 1.0, "content": ", plays a similar role to", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 106, 267, 504, 279 ], "spans": [ { "bbox": [ 106, 267, 212, 279 ], "score": 1.0, "content": "the bottleneck dimension", "type": "text", "cross_page": true }, { "bbox": [ 212, 269, 218, 277 ], "score": 0.67, "content": "r", "type": "inline_equation", "cross_page": true }, { "bbox": [ 218, 267, 504, 279 ], "score": 1.0, "content": "in adapters: they both represent the rank limitation of computing the", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 106, 278, 505, 289 ], "spans": [ { "bbox": [ 106, 279, 189, 289 ], "score": 1.0, "content": "modification vector", "type": "text", "cross_page": true }, { "bbox": [ 189, 278, 205, 288 ], "score": 0.73, "content": "\\Delta h", "type": "inline_equation", "cross_page": true }, { "bbox": [ 205, 279, 292, 289 ], "score": 1.0, "content": ". Thus we also refer", "type": "text", "cross_page": true }, { "bbox": [ 293, 279, 298, 288 ], "score": 0.47, "content": "l", "type": "inline_equation", "cross_page": true }, { "bbox": [ 298, 279, 505, 289 ], "score": 1.0, "content": "as the bottleneck dimension. Intuitively, the rank", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 105, 288, 492, 302 ], "spans": [ { "bbox": [ 105, 288, 197, 302 ], "score": 1.0, "content": "limitation implies that", "type": "text", "cross_page": true }, { "bbox": [ 198, 289, 213, 299 ], "score": 0.83, "content": "\\Delta h", "type": "inline_equation", "cross_page": true }, { "bbox": [ 214, 288, 356, 302 ], "score": 1.0, "content": "is a linear combination of the same", "type": "text", "cross_page": true }, { "bbox": [ 356, 289, 361, 299 ], "score": 0.31, "content": "l", "type": "inline_equation", "cross_page": true }, { "bbox": [ 361, 288, 376, 302 ], "score": 1.0, "content": "(or", "type": "text", "cross_page": true }, { "bbox": [ 376, 289, 392, 300 ], "score": 0.78, "content": "\\leq l", "type": "inline_equation", "cross_page": true }, { "bbox": [ 392, 288, 480, 302 ], "score": 1.0, "content": ") basis vectors for any", "type": "text", "cross_page": true }, { "bbox": [ 480, 291, 488, 299 ], "score": 0.68, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 488, 288, 492, 302 ], "score": 1.0, "content": ".", "type": "text", "cross_page": true } ], "index": 12 } ], "index": 31.5, "bbox_fs": [ 105, 648, 506, 672 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 107, 114, 505, 208 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 81, 505, 111 ], "group_id": 0, "lines": [ { "bbox": [ 105, 79, 505, 92 ], "spans": [ { "bbox": [ 105, 79, 505, 92 ], "score": 1.0, "content": "Table 1: Parameter-efficient tuning methods decomposed along the defined design dimensions. Here, for clarity,", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 91, 505, 101 ], "spans": [ { "bbox": [ 106, 91, 505, 101 ], "score": 1.0, "content": "we directly write the adapter nonlinear function as ReLU which is commonly used. The bottom part of the table", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 101, 392, 112 ], "spans": [ { "bbox": [ 105, 101, 392, 112 ], "score": 1.0, "content": "exemplifies new variants by transferring design choices of existing approaches.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 107, 114, 505, 208 ], "group_id": 0, "lines": [ { "bbox": [ 107, 114, 505, 208 ], "spans": [ { "bbox": [ 107, 114, 505, 208 ], "score": 0.981, "html": "
Method△h functional forminsertion formmodified representationcomposition function
Existing Methods
Prefix Tuning softmax(xWqPT)Puparallelhead attnh←(1-λ)h+λ△h
AdapterReLU(hWdown)Wupsequentialffn/attnh←h+△h
LoRAxWdownWupparallelattn key/valh←h+s·△h
Proposed Variants
Parallel adapterReLU(hWdown)Wupparallelffn/attnh←h+△h
Muti-head parallel adapterReLU(hWdown)Wupparallelhead attnh←h+△h
Scaled parallel adapterReLU(hWdown)Wupparallelffn/attnh←h+s·△h
", "type": "table", "image_path": "b28a62ed2483e98fd1f71cf0f6114593aa0a48ba96f113edaec0cb39d5a64ea2.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 107, 114, 505, 145.33333333333334 ], "spans": [], "index": 3 }, { "bbox": [ 107, 145.33333333333334, 505, 176.66666666666669 ], "spans": [], "index": 4 }, { "bbox": [ 107, 176.66666666666669, 505, 208.00000000000003 ], "spans": [], "index": 5 } ] } ], "index": 2.5 }, { "type": "text", "bbox": [ 106, 223, 505, 300 ], "lines": [ { "bbox": [ 106, 223, 505, 236 ], "spans": [ { "bbox": [ 106, 223, 505, 236 ], "score": 1.0, "content": "computation graph of prefix tuning from this view, which allows for abstraction of prefix tuning", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 231, 507, 247 ], "spans": [ { "bbox": [ 105, 231, 336, 247 ], "score": 1.0, "content": "as a plug-in module like adapters. Further, we note that", "type": "text" }, { "bbox": [ 336, 234, 392, 245 ], "score": 0.93, "content": "W _ { 1 } \\in \\mathbb { R } ^ { d _ { h } \\times l }", "type": "inline_equation" }, { "bbox": [ 393, 231, 411, 247 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 412, 233, 467, 245 ], "score": 0.94, "content": "W _ { 2 } \\in \\mathbb { R } ^ { l ^ { \\cdot } \\times d _ { h } }", "type": "inline_equation" }, { "bbox": [ 467, 231, 507, 247 ], "score": 1.0, "content": "are low-", "type": "text" } ], "index": 7 }, { "bbox": [ 104, 244, 506, 259 ], "spans": [ { "bbox": [ 104, 244, 190, 259 ], "score": 1.0, "content": "rank matrices when", "type": "text" }, { "bbox": [ 190, 245, 195, 255 ], "score": 0.53, "content": "l", "type": "inline_equation" }, { "bbox": [ 196, 244, 398, 259 ], "score": 1.0, "content": "is small, and thus they function similarly to the", "type": "text" }, { "bbox": [ 398, 245, 426, 256 ], "score": 0.9, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation" }, { "bbox": [ 427, 244, 447, 259 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 447, 245, 466, 257 ], "score": 0.89, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation" }, { "bbox": [ 467, 244, 506, 259 ], "score": 1.0, "content": "matrices", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 255, 506, 269 ], "spans": [ { "bbox": [ 105, 255, 402, 269 ], "score": 1.0, "content": "in adapters. This view also suggests that the number of prefix vectors,", "type": "text" }, { "bbox": [ 402, 257, 407, 266 ], "score": 0.57, "content": "l", "type": "inline_equation" }, { "bbox": [ 407, 255, 506, 269 ], "score": 1.0, "content": ", plays a similar role to", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 267, 504, 279 ], "spans": [ { "bbox": [ 106, 267, 212, 279 ], "score": 1.0, "content": "the bottleneck dimension", "type": "text" }, { "bbox": [ 212, 269, 218, 277 ], "score": 0.67, "content": "r", "type": "inline_equation" }, { "bbox": [ 218, 267, 504, 279 ], "score": 1.0, "content": "in adapters: they both represent the rank limitation of computing the", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 278, 505, 289 ], "spans": [ { "bbox": [ 106, 279, 189, 289 ], "score": 1.0, "content": "modification vector", "type": "text" }, { "bbox": [ 189, 278, 205, 288 ], "score": 0.73, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 205, 279, 292, 289 ], "score": 1.0, "content": ". Thus we also refer", "type": "text" }, { "bbox": [ 293, 279, 298, 288 ], "score": 0.47, "content": "l", "type": "inline_equation" }, { "bbox": [ 298, 279, 505, 289 ], "score": 1.0, "content": "as the bottleneck dimension. Intuitively, the rank", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 288, 492, 302 ], "spans": [ { "bbox": [ 105, 288, 197, 302 ], "score": 1.0, "content": "limitation implies that", "type": "text" }, { "bbox": [ 198, 289, 213, 299 ], "score": 0.83, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 214, 288, 356, 302 ], "score": 1.0, "content": "is a linear combination of the same", "type": "text" }, { "bbox": [ 356, 289, 361, 299 ], "score": 0.31, "content": "l", "type": "inline_equation" }, { "bbox": [ 361, 288, 376, 302 ], "score": 1.0, "content": "(or", "type": "text" }, { "bbox": [ 376, 289, 392, 300 ], "score": 0.78, "content": "\\leq l", "type": "inline_equation" }, { "bbox": [ 392, 288, 480, 302 ], "score": 1.0, "content": ") basis vectors for any", "type": "text" }, { "bbox": [ 480, 291, 488, 299 ], "score": 0.68, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 488, 288, 492, 302 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 12 } ], "index": 9 }, { "type": "text", "bbox": [ 107, 306, 505, 437 ], "lines": [ { "bbox": [ 106, 306, 505, 318 ], "spans": [ { "bbox": [ 106, 306, 384, 318 ], "score": 1.0, "content": "The Difference from Adapters: In addition to the gating variable", "type": "text" }, { "bbox": [ 384, 307, 391, 316 ], "score": 0.71, "content": "\\lambda", "type": "inline_equation" }, { "bbox": [ 391, 306, 505, 318 ], "score": 1.0, "content": ", we emphasize three differ-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 317, 504, 329 ], "spans": [ { "bbox": [ 105, 317, 479, 329 ], "score": 1.0, "content": "ences between prefix tuning and adapters. (1) As demonstrated in Figure 3, prefix tuning uses", "type": "text" }, { "bbox": [ 479, 319, 487, 327 ], "score": 0.68, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 487, 317, 504, 329 ], "score": 1.0, "content": ", the", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 327, 505, 341 ], "spans": [ { "bbox": [ 105, 327, 252, 341 ], "score": 1.0, "content": "input of the PLM layer, to compute", "type": "text" }, { "bbox": [ 252, 328, 268, 338 ], "score": 0.83, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 268, 327, 349, 341 ], "score": 1.0, "content": ", while adapters use", "type": "text" }, { "bbox": [ 349, 328, 357, 338 ], "score": 0.73, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 357, 327, 505, 341 ], "score": 1.0, "content": ", the output of the PLM layer. Thus,", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 338, 505, 352 ], "spans": [ { "bbox": [ 105, 338, 505, 352 ], "score": 1.0, "content": "prefix tuning can be thought of as a “parallel” computation to the PLM layer, whereas the typical", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 349, 505, 363 ], "spans": [ { "bbox": [ 105, 349, 505, 363 ], "score": 1.0, "content": "adapter is “sequential” computation. (2) Adapters are more flexible with respect to where they are", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 360, 505, 374 ], "spans": [ { "bbox": [ 105, 360, 505, 374 ], "score": 1.0, "content": "inserted than prefix tuning: adapters typically modify attention or FFN outputs, while prefix tuning", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 371, 505, 385 ], "spans": [ { "bbox": [ 105, 371, 505, 385 ], "score": 1.0, "content": "only modifies the attention output of each head. Empirically, this makes a large difference as we will", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 383, 505, 395 ], "spans": [ { "bbox": [ 105, 383, 141, 395 ], "score": 1.0, "content": "show in", "type": "text" }, { "bbox": [ 141, 383, 160, 394 ], "score": 0.86, "content": "\\ S 4 . 4", "type": "inline_equation" }, { "bbox": [ 160, 383, 505, 395 ], "score": 1.0, "content": ". (3) Eq. 10 applies to each attention head, while adapters are always single-headed,", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 393, 506, 407 ], "spans": [ { "bbox": [ 105, 393, 417, 407 ], "score": 1.0, "content": "which makes prefix tuning more expressive: head attention is of dimension", "type": "text" }, { "bbox": [ 418, 393, 442, 406 ], "score": 0.9, "content": "d / \\dot { N _ { h } }", "type": "inline_equation" }, { "bbox": [ 442, 393, 506, 407 ], "score": 1.0, "content": "– basically we", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 404, 505, 418 ], "spans": [ { "bbox": [ 105, 404, 298, 418 ], "score": 1.0, "content": "have full rank updates to each attention head if", "type": "text" }, { "bbox": [ 298, 405, 339, 417 ], "score": 0.93, "content": "l \\geq d / N _ { h }", "type": "inline_equation" }, { "bbox": [ 340, 404, 505, 418 ], "score": 1.0, "content": ", but we only get full-rank updates to the", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 415, 505, 428 ], "spans": [ { "bbox": [ 105, 415, 263, 428 ], "score": 1.0, "content": "whole attention output with adapters if", "type": "text" }, { "bbox": [ 263, 416, 288, 426 ], "score": 0.91, "content": "r \\geq d", "type": "inline_equation" }, { "bbox": [ 288, 415, 505, 428 ], "score": 1.0, "content": ". Notably, prefix tuning is not adding more parameters", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 426, 453, 439 ], "spans": [ { "bbox": [ 105, 426, 186, 439 ], "score": 1.0, "content": "than adapters when", "type": "text" }, { "bbox": [ 186, 426, 208, 436 ], "score": 0.89, "content": "{ \\dot { l } } = r", "type": "inline_equation" }, { "bbox": [ 208, 426, 429, 439 ], "score": 1.0, "content": ".7 We empirically validate such multi-head influence in", "type": "text" }, { "bbox": [ 429, 426, 448, 437 ], "score": 0.86, "content": "\\ S 4 . 4", "type": "inline_equation" }, { "bbox": [ 449, 426, 453, 439 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 24 } ], "index": 18.5 }, { "type": "title", "bbox": [ 108, 454, 248, 465 ], "lines": [ { "bbox": [ 105, 452, 250, 467 ], "spans": [ { "bbox": [ 105, 452, 250, 467 ], "score": 1.0, "content": "3.2 THE UNIFIED FRAMEWORK", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 476, 505, 565 ], "lines": [ { "bbox": [ 105, 476, 505, 490 ], "spans": [ { "bbox": [ 105, 476, 505, 490 ], "score": 1.0, "content": "Inspired by the connections between prefix tuning and adapters, we propose a general framework", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 488, 505, 500 ], "spans": [ { "bbox": [ 105, 488, 505, 500 ], "score": 1.0, "content": "that aims to unify several state-of-the-art parameter-efficient tuning methods. Specifically, we cast", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 498, 505, 511 ], "spans": [ { "bbox": [ 105, 498, 270, 511 ], "score": 1.0, "content": "them as learning a modification vector", "type": "text" }, { "bbox": [ 271, 499, 286, 509 ], "score": 0.82, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 287, 498, 505, 511 ], "score": 1.0, "content": ", which is applied to various hidden representations.", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 510, 505, 522 ], "spans": [ { "bbox": [ 105, 510, 409, 522 ], "score": 1.0, "content": "Formally, we denote the hidden representation to be directly modified as", "type": "text" }, { "bbox": [ 409, 510, 417, 520 ], "score": 0.72, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 417, 510, 505, 522 ], "score": 1.0, "content": ", and the direct input", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 521, 506, 533 ], "spans": [ { "bbox": [ 105, 521, 265, 533 ], "score": 1.0, "content": "to the PLM sub-module that computes", "type": "text" }, { "bbox": [ 265, 521, 273, 531 ], "score": 0.73, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 274, 521, 286, 533 ], "score": 1.0, "content": "as", "type": "text" }, { "bbox": [ 286, 522, 294, 531 ], "score": 0.72, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 295, 521, 316, 533 ], "score": 1.0, "content": "(e.g.", "type": "text" }, { "bbox": [ 317, 521, 325, 531 ], "score": 0.65, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 325, 521, 344, 533 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 344, 522, 352, 531 ], "score": 0.65, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 352, 521, 506, 533 ], "score": 1.0, "content": "can be the attention output and input", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 532, 506, 544 ], "spans": [ { "bbox": [ 105, 532, 506, 544 ], "score": 1.0, "content": "respectively). To characterize this modification process, we define a set of design dimensions, and", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 542, 505, 555 ], "spans": [ { "bbox": [ 105, 542, 505, 555 ], "score": 1.0, "content": "different methods can be instantiated by varying values along these dimensions. We detail the design", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 552, 503, 566 ], "spans": [ { "bbox": [ 106, 552, 503, 566 ], "score": 1.0, "content": "dimensions below, and illustrate how adapters, prefix tuning, and LoRA fall along them in Table 1:", "type": "text" } ], "index": 33 } ], "index": 29.5 }, { "type": "text", "bbox": [ 107, 570, 505, 615 ], "lines": [ { "bbox": [ 105, 570, 505, 582 ], "spans": [ { "bbox": [ 105, 570, 333, 582 ], "score": 1.0, "content": "Functional Form is the specific function that computes", "type": "text" }, { "bbox": [ 333, 570, 349, 581 ], "score": 0.73, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 349, 570, 505, 582 ], "score": 1.0, "content": ". We have detailed the functional form", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 582, 505, 593 ], "spans": [ { "bbox": [ 105, 582, 505, 593 ], "score": 1.0, "content": "for adapters, prefix tuning, and LoRA in Eq. 4, 6, and 10 respectively. The functional forms of all", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 592, 505, 605 ], "spans": [ { "bbox": [ 105, 592, 293, 605 ], "score": 1.0, "content": "these methods are similar with a proj down", "type": "text" }, { "bbox": [ 293, 594, 308, 603 ], "score": 0.81, "content": "", "type": "inline_equation" }, { "bbox": [ 308, 592, 368, 605 ], "score": 1.0, "content": "nonlinear", "type": "text" }, { "bbox": [ 368, 593, 411, 604 ], "score": 0.37, "content": "\\to \\mathsf { p r o j }", "type": "inline_equation" }, { "bbox": [ 412, 592, 505, 605 ], "score": 1.0, "content": "up architecture, while", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 603, 338, 615 ], "spans": [ { "bbox": [ 106, 603, 338, 615 ], "score": 1.0, "content": "“nonlinear” degenerates to the identity function in LoRA.", "type": "text" } ], "index": 37 } ], "index": 35.5 }, { "type": "text", "bbox": [ 107, 620, 452, 632 ], "lines": [ { "bbox": [ 105, 619, 454, 633 ], "spans": [ { "bbox": [ 105, 619, 454, 633 ], "score": 1.0, "content": "Modified Representation indicates which hidden representation is directly modified.8", "type": "text" } ], "index": 38 } ], "index": 38 }, { "type": "text", "bbox": [ 107, 637, 505, 681 ], "lines": [ { "bbox": [ 106, 637, 505, 649 ], "spans": [ { "bbox": [ 106, 637, 505, 649 ], "score": 1.0, "content": "Insertion Form is how the added module is inserted into the network. As mentioned in the previous", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 648, 505, 660 ], "spans": [ { "bbox": [ 106, 648, 505, 660 ], "score": 1.0, "content": "section and shown in Figure 3, traditionally adapters are inserted at a position in a sequential manner,", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 660, 504, 671 ], "spans": [ { "bbox": [ 106, 660, 251, 671 ], "score": 1.0, "content": "where both the input and output are", "type": "text" }, { "bbox": [ 251, 660, 259, 669 ], "score": 0.6, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 259, 660, 504, 671 ], "score": 1.0, "content": ". Prefix tuning and LoRA – although not originally described", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 671, 431, 682 ], "spans": [ { "bbox": [ 105, 671, 373, 682 ], "score": 1.0, "content": "in this way – turn out to be equivalent to a parallel insertion where", "type": "text" }, { "bbox": [ 374, 672, 381, 680 ], "score": 0.77, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 381, 671, 431, 682 ], "score": 1.0, "content": "is the input.", "type": "text" } ], "index": 42 } ], "index": 40.5 } ], "page_idx": 4, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 700, 504, 732 ], "lines": [ { "bbox": [ 118, 698, 398, 713 ], "spans": [ { "bbox": [ 118, 698, 184, 713 ], "score": 1.0, "content": "7We will detail in", "type": "text" }, { "bbox": [ 184, 701, 201, 711 ], "score": 0.88, "content": "\\ S 4 . 1", "type": "inline_equation" }, { "bbox": [ 201, 698, 398, 713 ], "score": 1.0, "content": "the number of parameters added of different methods.", "type": "text" } ] }, { "bbox": [ 118, 709, 506, 724 ], "spans": [ { "bbox": [ 118, 709, 506, 724 ], "score": 1.0, "content": "8Strictly speaking, all the hidden representations would be indirectly influenced by modifying the ones", "type": "text" } ] }, { "bbox": [ 105, 720, 426, 732 ], "spans": [ { "bbox": [ 105, 720, 426, 732 ], "score": 1.0, "content": "before them. Here we refer to the position being directly modified by the added module.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "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": [ 107, 114, 505, 208 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 81, 505, 111 ], "group_id": 0, "lines": [ { "bbox": [ 105, 79, 505, 92 ], "spans": [ { "bbox": [ 105, 79, 505, 92 ], "score": 1.0, "content": "Table 1: Parameter-efficient tuning methods decomposed along the defined design dimensions. Here, for clarity,", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 91, 505, 101 ], "spans": [ { "bbox": [ 106, 91, 505, 101 ], "score": 1.0, "content": "we directly write the adapter nonlinear function as ReLU which is commonly used. The bottom part of the table", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 101, 392, 112 ], "spans": [ { "bbox": [ 105, 101, 392, 112 ], "score": 1.0, "content": "exemplifies new variants by transferring design choices of existing approaches.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 107, 114, 505, 208 ], "group_id": 0, "lines": [ { "bbox": [ 107, 114, 505, 208 ], "spans": [ { "bbox": [ 107, 114, 505, 208 ], "score": 0.981, "html": "
Method△h functional forminsertion formmodified representationcomposition function
Existing Methods
Prefix Tuning softmax(xWqPT)Puparallelhead attnh←(1-λ)h+λ△h
AdapterReLU(hWdown)Wupsequentialffn/attnh←h+△h
LoRAxWdownWupparallelattn key/valh←h+s·△h
Proposed Variants
Parallel adapterReLU(hWdown)Wupparallelffn/attnh←h+△h
Muti-head parallel adapterReLU(hWdown)Wupparallelhead attnh←h+△h
Scaled parallel adapterReLU(hWdown)Wupparallelffn/attnh←h+s·△h
", "type": "table", "image_path": "b28a62ed2483e98fd1f71cf0f6114593aa0a48ba96f113edaec0cb39d5a64ea2.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 107, 114, 505, 145.33333333333334 ], "spans": [], "index": 3 }, { "bbox": [ 107, 145.33333333333334, 505, 176.66666666666669 ], "spans": [], "index": 4 }, { "bbox": [ 107, 176.66666666666669, 505, 208.00000000000003 ], "spans": [], "index": 5 } ] } ], "index": 2.5 }, { "type": "text", "bbox": [ 106, 223, 505, 300 ], "lines": [], "index": 9, "bbox_fs": [ 104, 223, 507, 302 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 306, 505, 437 ], "lines": [ { "bbox": [ 106, 306, 505, 318 ], "spans": [ { "bbox": [ 106, 306, 384, 318 ], "score": 1.0, "content": "The Difference from Adapters: In addition to the gating variable", "type": "text" }, { "bbox": [ 384, 307, 391, 316 ], "score": 0.71, "content": "\\lambda", "type": "inline_equation" }, { "bbox": [ 391, 306, 505, 318 ], "score": 1.0, "content": ", we emphasize three differ-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 317, 504, 329 ], "spans": [ { "bbox": [ 105, 317, 479, 329 ], "score": 1.0, "content": "ences between prefix tuning and adapters. (1) As demonstrated in Figure 3, prefix tuning uses", "type": "text" }, { "bbox": [ 479, 319, 487, 327 ], "score": 0.68, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 487, 317, 504, 329 ], "score": 1.0, "content": ", the", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 327, 505, 341 ], "spans": [ { "bbox": [ 105, 327, 252, 341 ], "score": 1.0, "content": "input of the PLM layer, to compute", "type": "text" }, { "bbox": [ 252, 328, 268, 338 ], "score": 0.83, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 268, 327, 349, 341 ], "score": 1.0, "content": ", while adapters use", "type": "text" }, { "bbox": [ 349, 328, 357, 338 ], "score": 0.73, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 357, 327, 505, 341 ], "score": 1.0, "content": ", the output of the PLM layer. Thus,", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 338, 505, 352 ], "spans": [ { "bbox": [ 105, 338, 505, 352 ], "score": 1.0, "content": "prefix tuning can be thought of as a “parallel” computation to the PLM layer, whereas the typical", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 349, 505, 363 ], "spans": [ { "bbox": [ 105, 349, 505, 363 ], "score": 1.0, "content": "adapter is “sequential” computation. (2) Adapters are more flexible with respect to where they are", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 360, 505, 374 ], "spans": [ { "bbox": [ 105, 360, 505, 374 ], "score": 1.0, "content": "inserted than prefix tuning: adapters typically modify attention or FFN outputs, while prefix tuning", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 371, 505, 385 ], "spans": [ { "bbox": [ 105, 371, 505, 385 ], "score": 1.0, "content": "only modifies the attention output of each head. Empirically, this makes a large difference as we will", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 383, 505, 395 ], "spans": [ { "bbox": [ 105, 383, 141, 395 ], "score": 1.0, "content": "show in", "type": "text" }, { "bbox": [ 141, 383, 160, 394 ], "score": 0.86, "content": "\\ S 4 . 4", "type": "inline_equation" }, { "bbox": [ 160, 383, 505, 395 ], "score": 1.0, "content": ". (3) Eq. 10 applies to each attention head, while adapters are always single-headed,", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 393, 506, 407 ], "spans": [ { "bbox": [ 105, 393, 417, 407 ], "score": 1.0, "content": "which makes prefix tuning more expressive: head attention is of dimension", "type": "text" }, { "bbox": [ 418, 393, 442, 406 ], "score": 0.9, "content": "d / \\dot { N _ { h } }", "type": "inline_equation" }, { "bbox": [ 442, 393, 506, 407 ], "score": 1.0, "content": "– basically we", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 404, 505, 418 ], "spans": [ { "bbox": [ 105, 404, 298, 418 ], "score": 1.0, "content": "have full rank updates to each attention head if", "type": "text" }, { "bbox": [ 298, 405, 339, 417 ], "score": 0.93, "content": "l \\geq d / N _ { h }", "type": "inline_equation" }, { "bbox": [ 340, 404, 505, 418 ], "score": 1.0, "content": ", but we only get full-rank updates to the", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 415, 505, 428 ], "spans": [ { "bbox": [ 105, 415, 263, 428 ], "score": 1.0, "content": "whole attention output with adapters if", "type": "text" }, { "bbox": [ 263, 416, 288, 426 ], "score": 0.91, "content": "r \\geq d", "type": "inline_equation" }, { "bbox": [ 288, 415, 505, 428 ], "score": 1.0, "content": ". Notably, prefix tuning is not adding more parameters", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 426, 453, 439 ], "spans": [ { "bbox": [ 105, 426, 186, 439 ], "score": 1.0, "content": "than adapters when", "type": "text" }, { "bbox": [ 186, 426, 208, 436 ], "score": 0.89, "content": "{ \\dot { l } } = r", "type": "inline_equation" }, { "bbox": [ 208, 426, 429, 439 ], "score": 1.0, "content": ".7 We empirically validate such multi-head influence in", "type": "text" }, { "bbox": [ 429, 426, 448, 437 ], "score": 0.86, "content": "\\ S 4 . 4", "type": "inline_equation" }, { "bbox": [ 449, 426, 453, 439 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 24 } ], "index": 18.5, "bbox_fs": [ 105, 306, 506, 439 ] }, { "type": "title", "bbox": [ 108, 454, 248, 465 ], "lines": [ { "bbox": [ 105, 452, 250, 467 ], "spans": [ { "bbox": [ 105, 452, 250, 467 ], "score": 1.0, "content": "3.2 THE UNIFIED FRAMEWORK", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 476, 505, 565 ], "lines": [ { "bbox": [ 105, 476, 505, 490 ], "spans": [ { "bbox": [ 105, 476, 505, 490 ], "score": 1.0, "content": "Inspired by the connections between prefix tuning and adapters, we propose a general framework", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 488, 505, 500 ], "spans": [ { "bbox": [ 105, 488, 505, 500 ], "score": 1.0, "content": "that aims to unify several state-of-the-art parameter-efficient tuning methods. Specifically, we cast", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 498, 505, 511 ], "spans": [ { "bbox": [ 105, 498, 270, 511 ], "score": 1.0, "content": "them as learning a modification vector", "type": "text" }, { "bbox": [ 271, 499, 286, 509 ], "score": 0.82, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 287, 498, 505, 511 ], "score": 1.0, "content": ", which is applied to various hidden representations.", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 510, 505, 522 ], "spans": [ { "bbox": [ 105, 510, 409, 522 ], "score": 1.0, "content": "Formally, we denote the hidden representation to be directly modified as", "type": "text" }, { "bbox": [ 409, 510, 417, 520 ], "score": 0.72, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 417, 510, 505, 522 ], "score": 1.0, "content": ", and the direct input", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 521, 506, 533 ], "spans": [ { "bbox": [ 105, 521, 265, 533 ], "score": 1.0, "content": "to the PLM sub-module that computes", "type": "text" }, { "bbox": [ 265, 521, 273, 531 ], "score": 0.73, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 274, 521, 286, 533 ], "score": 1.0, "content": "as", "type": "text" }, { "bbox": [ 286, 522, 294, 531 ], "score": 0.72, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 295, 521, 316, 533 ], "score": 1.0, "content": "(e.g.", "type": "text" }, { "bbox": [ 317, 521, 325, 531 ], "score": 0.65, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 325, 521, 344, 533 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 344, 522, 352, 531 ], "score": 0.65, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 352, 521, 506, 533 ], "score": 1.0, "content": "can be the attention output and input", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 532, 506, 544 ], "spans": [ { "bbox": [ 105, 532, 506, 544 ], "score": 1.0, "content": "respectively). To characterize this modification process, we define a set of design dimensions, and", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 542, 505, 555 ], "spans": [ { "bbox": [ 105, 542, 505, 555 ], "score": 1.0, "content": "different methods can be instantiated by varying values along these dimensions. We detail the design", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 552, 503, 566 ], "spans": [ { "bbox": [ 106, 552, 503, 566 ], "score": 1.0, "content": "dimensions below, and illustrate how adapters, prefix tuning, and LoRA fall along them in Table 1:", "type": "text" } ], "index": 33 } ], "index": 29.5, "bbox_fs": [ 105, 476, 506, 566 ] }, { "type": "text", "bbox": [ 107, 570, 505, 615 ], "lines": [ { "bbox": [ 105, 570, 505, 582 ], "spans": [ { "bbox": [ 105, 570, 333, 582 ], "score": 1.0, "content": "Functional Form is the specific function that computes", "type": "text" }, { "bbox": [ 333, 570, 349, 581 ], "score": 0.73, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 349, 570, 505, 582 ], "score": 1.0, "content": ". We have detailed the functional form", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 582, 505, 593 ], "spans": [ { "bbox": [ 105, 582, 505, 593 ], "score": 1.0, "content": "for adapters, prefix tuning, and LoRA in Eq. 4, 6, and 10 respectively. The functional forms of all", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 592, 505, 605 ], "spans": [ { "bbox": [ 105, 592, 293, 605 ], "score": 1.0, "content": "these methods are similar with a proj down", "type": "text" }, { "bbox": [ 293, 594, 308, 603 ], "score": 0.81, "content": "", "type": "inline_equation" }, { "bbox": [ 308, 592, 368, 605 ], "score": 1.0, "content": "nonlinear", "type": "text" }, { "bbox": [ 368, 593, 411, 604 ], "score": 0.37, "content": "\\to \\mathsf { p r o j }", "type": "inline_equation" }, { "bbox": [ 412, 592, 505, 605 ], "score": 1.0, "content": "up architecture, while", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 603, 338, 615 ], "spans": [ { "bbox": [ 106, 603, 338, 615 ], "score": 1.0, "content": "“nonlinear” degenerates to the identity function in LoRA.", "type": "text" } ], "index": 37 } ], "index": 35.5, "bbox_fs": [ 105, 570, 505, 615 ] }, { "type": "text", "bbox": [ 107, 620, 452, 632 ], "lines": [ { "bbox": [ 105, 619, 454, 633 ], "spans": [ { "bbox": [ 105, 619, 454, 633 ], "score": 1.0, "content": "Modified Representation indicates which hidden representation is directly modified.8", "type": "text" } ], "index": 38 } ], "index": 38, "bbox_fs": [ 105, 619, 454, 633 ] }, { "type": "text", "bbox": [ 107, 637, 505, 681 ], "lines": [ { "bbox": [ 106, 637, 505, 649 ], "spans": [ { "bbox": [ 106, 637, 505, 649 ], "score": 1.0, "content": "Insertion Form is how the added module is inserted into the network. As mentioned in the previous", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 648, 505, 660 ], "spans": [ { "bbox": [ 106, 648, 505, 660 ], "score": 1.0, "content": "section and shown in Figure 3, traditionally adapters are inserted at a position in a sequential manner,", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 660, 504, 671 ], "spans": [ { "bbox": [ 106, 660, 251, 671 ], "score": 1.0, "content": "where both the input and output are", "type": "text" }, { "bbox": [ 251, 660, 259, 669 ], "score": 0.6, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 259, 660, 504, 671 ], "score": 1.0, "content": ". Prefix tuning and LoRA – although not originally described", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 671, 431, 682 ], "spans": [ { "bbox": [ 105, 671, 373, 682 ], "score": 1.0, "content": "in this way – turn out to be equivalent to a parallel insertion where", "type": "text" }, { "bbox": [ 374, 672, 381, 680 ], "score": 0.77, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 381, 671, 431, 682 ], "score": 1.0, "content": "is the input.", "type": "text" } ], "index": 42 } ], "index": 40.5, "bbox_fs": [ 105, 637, 505, 682 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 504, 126 ], "lines": [ { "bbox": [ 106, 82, 505, 95 ], "spans": [ { "bbox": [ 106, 82, 312, 95 ], "score": 1.0, "content": "Composition Function is how the modified vector", "type": "text" }, { "bbox": [ 312, 83, 328, 93 ], "score": 0.72, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 329, 82, 505, 95 ], "score": 1.0, "content": "is composed with the original hidden repre-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 94, 505, 106 ], "spans": [ { "bbox": [ 106, 94, 145, 106 ], "score": 1.0, "content": "sentation", "type": "text" }, { "bbox": [ 145, 94, 153, 104 ], "score": 0.75, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 154, 94, 505, 106 ], "score": 1.0, "content": "to form the new hidden representation. For example, adapters perform simple additive", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 105, 506, 117 ], "spans": [ { "bbox": [ 105, 105, 506, 117 ], "score": 1.0, "content": "composition, prefix tuning uses a gated additive composition as shown in Eq. 10, and LoRA scales", "type": "text" } ], "index": 2 }, { "bbox": [ 107, 115, 444, 128 ], "spans": [ { "bbox": [ 107, 115, 123, 126 ], "score": 0.74, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 123, 115, 444, 128 ], "score": 1.0, "content": "by a constant factor and adds it to the original hidden representation as in Eq. 6.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "text", "bbox": [ 107, 132, 505, 198 ], "lines": [ { "bbox": [ 106, 132, 505, 145 ], "spans": [ { "bbox": [ 106, 132, 505, 145 ], "score": 1.0, "content": "We note that many other methods not present in Table 1 fit into this framework as well. For example,", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 143, 506, 156 ], "spans": [ { "bbox": [ 105, 143, 506, 156 ], "score": 1.0, "content": "prompt tuning modifies the head attention in the first layer in a way similar to prefix tuning, and", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 154, 505, 167 ], "spans": [ { "bbox": [ 105, 154, 505, 167 ], "score": 1.0, "content": "various adapter variants (Pfeiffer et al., 2021; Mahabadi et al., 2021) can be represented in a similar", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 165, 505, 179 ], "spans": [ { "bbox": [ 105, 165, 505, 179 ], "score": 1.0, "content": "way as adapters. Critically, the unified framework allows us to study parameter-efficient tuning", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 177, 505, 188 ], "spans": [ { "bbox": [ 106, 177, 505, 188 ], "score": 1.0, "content": "methods along these design dimensions, identify the critical design choices, and potentially transfer", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 188, 359, 199 ], "spans": [ { "bbox": [ 106, 188, 359, 199 ], "score": 1.0, "content": "design elements across approaches, as in the following section.", "type": "text" } ], "index": 9 } ], "index": 6.5 }, { "type": "title", "bbox": [ 109, 210, 284, 221 ], "lines": [ { "bbox": [ 106, 209, 285, 222 ], "spans": [ { "bbox": [ 106, 209, 285, 222 ], "score": 1.0, "content": "3.3 TRANSFERRING DESIGN ELEMENTS", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 107, 230, 505, 329 ], "lines": [ { "bbox": [ 105, 231, 504, 243 ], "spans": [ { "bbox": [ 105, 231, 504, 243 ], "score": 1.0, "content": "Here, and in Figure 3, we describe just a few novel methods that can be derived through our uni-", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 241, 506, 254 ], "spans": [ { "bbox": [ 105, 241, 506, 254 ], "score": 1.0, "content": "fied view above by transferring design elements across methods: (1) Parallel Adapter is the variant", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 253, 505, 265 ], "spans": [ { "bbox": [ 105, 253, 505, 265 ], "score": 1.0, "content": "by transferring the parallel insertion of prefix tuning into adapters. Interestingly, while we moti-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 263, 505, 276 ], "spans": [ { "bbox": [ 105, 263, 505, 276 ], "score": 1.0, "content": "vate the parallel adapter due to its similarity to prefix tuning, concurrent work (Zhu et al., 2021)", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 273, 506, 288 ], "spans": [ { "bbox": [ 105, 273, 506, 288 ], "score": 1.0, "content": "independently proposed this variant and studied it empirically; (2) Multi-head Parallel Adapter is", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 285, 505, 299 ], "spans": [ { "bbox": [ 105, 285, 505, 299 ], "score": 1.0, "content": "a further step to make adapters more similar to prefix tuning: we apply parallel adapters to modify", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 296, 505, 309 ], "spans": [ { "bbox": [ 105, 296, 505, 309 ], "score": 1.0, "content": "head attention outputs as prefix tuning. This way the variant improves the capacity for free by uti-", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 306, 505, 321 ], "spans": [ { "bbox": [ 105, 306, 304, 321 ], "score": 1.0, "content": "lizing the multi-head projections as we discuss in", "type": "text" }, { "bbox": [ 304, 308, 323, 319 ], "score": 0.86, "content": "\\ S 3 . 1", "type": "inline_equation" }, { "bbox": [ 323, 306, 505, 321 ], "score": 1.0, "content": ". (3) Scaled Parallel Adapter is the variant by", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 318, 484, 331 ], "spans": [ { "bbox": [ 105, 318, 484, 331 ], "score": 1.0, "content": "transferring the composition and insertion form of LoRA into adapters, as shown in Figure 3e.", "type": "text" } ], "index": 19 } ], "index": 15 }, { "type": "text", "bbox": [ 108, 335, 505, 368 ], "lines": [ { "bbox": [ 105, 334, 505, 348 ], "spans": [ { "bbox": [ 105, 334, 505, 348 ], "score": 1.0, "content": "Our discussion and formulation so far raise a few questions: Do methods varying the design elements", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 347, 505, 359 ], "spans": [ { "bbox": [ 106, 347, 505, 359 ], "score": 1.0, "content": "above exhibit distinct properties? Which design dimensions are particularly important? Do the novel", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 358, 446, 370 ], "spans": [ { "bbox": [ 106, 358, 446, 370 ], "score": 1.0, "content": "methods described above yield better performance? We answer these questions next.", "type": "text" } ], "index": 22 } ], "index": 21 }, { "type": "title", "bbox": [ 108, 385, 200, 398 ], "lines": [ { "bbox": [ 105, 385, 201, 399 ], "spans": [ { "bbox": [ 105, 385, 201, 399 ], "score": 1.0, "content": "4 EXPERIMENTS", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "title", "bbox": [ 108, 409, 204, 421 ], "lines": [ { "bbox": [ 105, 408, 205, 423 ], "spans": [ { "bbox": [ 105, 408, 205, 423 ], "score": 1.0, "content": "4.1 GENERAL SETUP", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 430, 505, 496 ], "lines": [ { "bbox": [ 105, 430, 505, 442 ], "spans": [ { "bbox": [ 105, 430, 505, 442 ], "score": 1.0, "content": "Datasets: We study four downstream tasks: (1) XSum (Narayan et al., 2018) is an English sum-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 442, 506, 454 ], "spans": [ { "bbox": [ 105, 442, 506, 454 ], "score": 1.0, "content": "marization dataset where models predict a summary given a news article; (2) English to Romanian", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 452, 505, 465 ], "spans": [ { "bbox": [ 105, 452, 505, 465 ], "score": 1.0, "content": "translation using the WMT 2016 en-ro dataset (Bojar et al., 2016); (3) MNLI (Williams et al., 2018)", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 464, 505, 475 ], "spans": [ { "bbox": [ 106, 464, 505, 475 ], "score": 1.0, "content": "is an English natural language inference dataset where models predict whether one sentence entails,", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 474, 505, 486 ], "spans": [ { "bbox": [ 105, 474, 505, 486 ], "score": 1.0, "content": "contradicts, or is neutral to another. (4) SST2 (Socher et al., 2013) is an English sentiment classifi-", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 485, 486, 498 ], "spans": [ { "bbox": [ 105, 485, 486, 498 ], "score": 1.0, "content": "cation benchmark where models predict whether a sentence’s sentiment is positive or negative.", "type": "text" } ], "index": 30 } ], "index": 27.5 }, { "type": "text", "bbox": [ 107, 502, 505, 601 ], "lines": [ { "bbox": [ 105, 501, 506, 516 ], "spans": [ { "bbox": [ 105, 501, 175, 516 ], "score": 1.0, "content": "Setup: We use", "type": "text" }, { "bbox": [ 176, 503, 222, 514 ], "score": 0.73, "content": "{ \\tt B A R T } _ { \\tt L A R G E }", "type": "inline_equation" }, { "bbox": [ 222, 501, 506, 516 ], "score": 1.0, "content": "(Lewis et al., 2020) and a multilingual version of it, mBARTLARGE (Liu", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 514, 505, 526 ], "spans": [ { "bbox": [ 105, 514, 505, 526 ], "score": 1.0, "content": "et al., 2020a), as the underlying pretrained models for XSum and en-ro translation respectively, and", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 523, 506, 537 ], "spans": [ { "bbox": [ 105, 523, 506, 537 ], "score": 1.0, "content": "we use RoBERTaBASE (Liu et al., 2019) for MNLI and SST2. We vary the bottleneck dimension", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 533, 506, 549 ], "spans": [ { "bbox": [ 105, 533, 136, 549 ], "score": 1.0, "content": "within", "type": "text" }, { "bbox": [ 136, 535, 230, 547 ], "score": 0.88, "content": "\\{ 1 , 3 0 , 2 0 0 , 5 1 2 , 1 0 2 4 \\}", "type": "inline_equation" }, { "bbox": [ 230, 533, 506, 549 ], "score": 1.0, "content": "if needed.9 We mainly study adapters, prefix tuning (prefix), and", "type": "text" } ], "index": 34 }, { "bbox": [ 104, 545, 506, 560 ], "spans": [ { "bbox": [ 104, 545, 506, 560 ], "score": 1.0, "content": "LoRA which greatly outperform bitfit and prompt tuning in our experiments. In the analysis sections", "type": "text" } ], "index": 35 }, { "bbox": [ 107, 557, 505, 570 ], "spans": [ { "bbox": [ 107, 557, 147, 568 ], "score": 0.45, "content": "( \\ S 4 . 3 – 4 . 5 )", "type": "inline_equation" }, { "bbox": [ 147, 557, 505, 570 ], "score": 1.0, "content": "we insert adapters either at the attention or FFN layers for easier analysis, but include the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 568, 505, 580 ], "spans": [ { "bbox": [ 105, 568, 505, 580 ], "score": 1.0, "content": "results of inserting at both places in the final comparison (§4.6). We re-implement these methods", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 577, 506, 592 ], "spans": [ { "bbox": [ 105, 577, 506, 592 ], "score": 1.0, "content": "based on their respective public code.10 We use the huggingface transformers library (Wolf et al.,", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 589, 442, 603 ], "spans": [ { "bbox": [ 105, 589, 442, 603 ], "score": 1.0, "content": "2020) for our implementation. Complete setup details can be found in Appendix A.", "type": "text" } ], "index": 39 } ], "index": 35 }, { "type": "text", "bbox": [ 107, 606, 505, 651 ], "lines": [ { "bbox": [ 105, 606, 505, 619 ], "spans": [ { "bbox": [ 105, 606, 245, 619 ], "score": 1.0, "content": "Evaluation: We report ROUGE", "type": "text" }, { "bbox": [ 245, 607, 268, 618 ], "score": 0.32, "content": "1 / 2 / \\mathrm { L }", "type": "inline_equation" }, { "bbox": [ 268, 606, 505, 619 ], "score": 1.0, "content": "scores (R-1/2/L, Lin (2004)) on the XSum test set, BLEU", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 618, 505, 631 ], "spans": [ { "bbox": [ 105, 618, 505, 631 ], "score": 1.0, "content": "scores (Papineni et al., 2002) on the en-ro test set, and accuracy on the MNLI and SST2 dev set.", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 629, 506, 642 ], "spans": [ { "bbox": [ 105, 629, 506, 642 ], "score": 1.0, "content": "For MNLI and SST2, we take the median of five random runs. We also report the number of tuned", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 640, 330, 652 ], "spans": [ { "bbox": [ 105, 640, 330, 652 ], "score": 1.0, "content": "parameters relative to that in full fine-tuning (#params).", "type": "text" } ], "index": 43 } ], "index": 41.5 }, { "type": "text", "bbox": [ 107, 657, 503, 690 ], "lines": [ { "bbox": [ 106, 657, 505, 669 ], "spans": [ { "bbox": [ 106, 657, 505, 669 ], "score": 1.0, "content": "Number of Tunable Parameters: BART and mBART have an encoder-decoder structure that has", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 668, 504, 679 ], "spans": [ { "bbox": [ 106, 668, 504, 679 ], "score": 1.0, "content": "three types of attention: encoder self-attention, decoder self-attention, and decoder cross-attention.", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 678, 505, 692 ], "spans": [ { "bbox": [ 105, 678, 505, 692 ], "score": 1.0, "content": "RoBERTa only has encoder self-attention. For each attention sub-layer, the number of parameters", "type": "text" } ], "index": 46 } ], "index": 45 } ], "page_idx": 5, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 700, 505, 732 ], "lines": [ { "bbox": [ 118, 698, 484, 713 ], "spans": [ { "bbox": [ 118, 698, 484, 713 ], "score": 1.0, "content": "9In some settings we use other values to match the number of added parameters of different methods.", "type": "text" } ] }, { "bbox": [ 115, 708, 506, 725 ], "spans": [ { "bbox": [ 115, 708, 506, 725 ], "score": 1.0, "content": "10We verify that our re-implementation can reproduce adapter and prefix tuning on XSum, and LoRA on", "type": "text" } ] }, { "bbox": [ 106, 721, 379, 733 ], "spans": [ { "bbox": [ 106, 721, 379, 733 ], "score": 1.0, "content": "MNLI, by comparing with the results of running the original released code.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "6", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 504, 126 ], "lines": [ { "bbox": [ 106, 82, 505, 95 ], "spans": [ { "bbox": [ 106, 82, 312, 95 ], "score": 1.0, "content": "Composition Function is how the modified vector", "type": "text" }, { "bbox": [ 312, 83, 328, 93 ], "score": 0.72, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 329, 82, 505, 95 ], "score": 1.0, "content": "is composed with the original hidden repre-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 94, 505, 106 ], "spans": [ { "bbox": [ 106, 94, 145, 106 ], "score": 1.0, "content": "sentation", "type": "text" }, { "bbox": [ 145, 94, 153, 104 ], "score": 0.75, "content": "^ { h }", "type": "inline_equation" }, { "bbox": [ 154, 94, 505, 106 ], "score": 1.0, "content": "to form the new hidden representation. For example, adapters perform simple additive", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 105, 506, 117 ], "spans": [ { "bbox": [ 105, 105, 506, 117 ], "score": 1.0, "content": "composition, prefix tuning uses a gated additive composition as shown in Eq. 10, and LoRA scales", "type": "text" } ], "index": 2 }, { "bbox": [ 107, 115, 444, 128 ], "spans": [ { "bbox": [ 107, 115, 123, 126 ], "score": 0.74, "content": "\\Delta h", "type": "inline_equation" }, { "bbox": [ 123, 115, 444, 128 ], "score": 1.0, "content": "by a constant factor and adds it to the original hidden representation as in Eq. 6.", "type": "text" } ], "index": 3 } ], "index": 1.5, "bbox_fs": [ 105, 82, 506, 128 ] }, { "type": "text", "bbox": [ 107, 132, 505, 198 ], "lines": [ { "bbox": [ 106, 132, 505, 145 ], "spans": [ { "bbox": [ 106, 132, 505, 145 ], "score": 1.0, "content": "We note that many other methods not present in Table 1 fit into this framework as well. For example,", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 143, 506, 156 ], "spans": [ { "bbox": [ 105, 143, 506, 156 ], "score": 1.0, "content": "prompt tuning modifies the head attention in the first layer in a way similar to prefix tuning, and", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 154, 505, 167 ], "spans": [ { "bbox": [ 105, 154, 505, 167 ], "score": 1.0, "content": "various adapter variants (Pfeiffer et al., 2021; Mahabadi et al., 2021) can be represented in a similar", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 165, 505, 179 ], "spans": [ { "bbox": [ 105, 165, 505, 179 ], "score": 1.0, "content": "way as adapters. Critically, the unified framework allows us to study parameter-efficient tuning", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 177, 505, 188 ], "spans": [ { "bbox": [ 106, 177, 505, 188 ], "score": 1.0, "content": "methods along these design dimensions, identify the critical design choices, and potentially transfer", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 188, 359, 199 ], "spans": [ { "bbox": [ 106, 188, 359, 199 ], "score": 1.0, "content": "design elements across approaches, as in the following section.", "type": "text" } ], "index": 9 } ], "index": 6.5, "bbox_fs": [ 105, 132, 506, 199 ] }, { "type": "title", "bbox": [ 109, 210, 284, 221 ], "lines": [ { "bbox": [ 106, 209, 285, 222 ], "spans": [ { "bbox": [ 106, 209, 285, 222 ], "score": 1.0, "content": "3.3 TRANSFERRING DESIGN ELEMENTS", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 107, 230, 505, 329 ], "lines": [ { "bbox": [ 105, 231, 504, 243 ], "spans": [ { "bbox": [ 105, 231, 504, 243 ], "score": 1.0, "content": "Here, and in Figure 3, we describe just a few novel methods that can be derived through our uni-", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 241, 506, 254 ], "spans": [ { "bbox": [ 105, 241, 506, 254 ], "score": 1.0, "content": "fied view above by transferring design elements across methods: (1) Parallel Adapter is the variant", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 253, 505, 265 ], "spans": [ { "bbox": [ 105, 253, 505, 265 ], "score": 1.0, "content": "by transferring the parallel insertion of prefix tuning into adapters. Interestingly, while we moti-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 263, 505, 276 ], "spans": [ { "bbox": [ 105, 263, 505, 276 ], "score": 1.0, "content": "vate the parallel adapter due to its similarity to prefix tuning, concurrent work (Zhu et al., 2021)", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 273, 506, 288 ], "spans": [ { "bbox": [ 105, 273, 506, 288 ], "score": 1.0, "content": "independently proposed this variant and studied it empirically; (2) Multi-head Parallel Adapter is", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 285, 505, 299 ], "spans": [ { "bbox": [ 105, 285, 505, 299 ], "score": 1.0, "content": "a further step to make adapters more similar to prefix tuning: we apply parallel adapters to modify", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 296, 505, 309 ], "spans": [ { "bbox": [ 105, 296, 505, 309 ], "score": 1.0, "content": "head attention outputs as prefix tuning. This way the variant improves the capacity for free by uti-", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 306, 505, 321 ], "spans": [ { "bbox": [ 105, 306, 304, 321 ], "score": 1.0, "content": "lizing the multi-head projections as we discuss in", "type": "text" }, { "bbox": [ 304, 308, 323, 319 ], "score": 0.86, "content": "\\ S 3 . 1", "type": "inline_equation" }, { "bbox": [ 323, 306, 505, 321 ], "score": 1.0, "content": ". (3) Scaled Parallel Adapter is the variant by", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 318, 484, 331 ], "spans": [ { "bbox": [ 105, 318, 484, 331 ], "score": 1.0, "content": "transferring the composition and insertion form of LoRA into adapters, as shown in Figure 3e.", "type": "text" } ], "index": 19 } ], "index": 15, "bbox_fs": [ 105, 231, 506, 331 ] }, { "type": "text", "bbox": [ 108, 335, 505, 368 ], "lines": [ { "bbox": [ 105, 334, 505, 348 ], "spans": [ { "bbox": [ 105, 334, 505, 348 ], "score": 1.0, "content": "Our discussion and formulation so far raise a few questions: Do methods varying the design elements", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 347, 505, 359 ], "spans": [ { "bbox": [ 106, 347, 505, 359 ], "score": 1.0, "content": "above exhibit distinct properties? Which design dimensions are particularly important? Do the novel", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 358, 446, 370 ], "spans": [ { "bbox": [ 106, 358, 446, 370 ], "score": 1.0, "content": "methods described above yield better performance? We answer these questions next.", "type": "text" } ], "index": 22 } ], "index": 21, "bbox_fs": [ 105, 334, 505, 370 ] }, { "type": "title", "bbox": [ 108, 385, 200, 398 ], "lines": [ { "bbox": [ 105, 385, 201, 399 ], "spans": [ { "bbox": [ 105, 385, 201, 399 ], "score": 1.0, "content": "4 EXPERIMENTS", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "title", "bbox": [ 108, 409, 204, 421 ], "lines": [ { "bbox": [ 105, 408, 205, 423 ], "spans": [ { "bbox": [ 105, 408, 205, 423 ], "score": 1.0, "content": "4.1 GENERAL SETUP", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 430, 505, 496 ], "lines": [ { "bbox": [ 105, 430, 505, 442 ], "spans": [ { "bbox": [ 105, 430, 505, 442 ], "score": 1.0, "content": "Datasets: We study four downstream tasks: (1) XSum (Narayan et al., 2018) is an English sum-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 442, 506, 454 ], "spans": [ { "bbox": [ 105, 442, 506, 454 ], "score": 1.0, "content": "marization dataset where models predict a summary given a news article; (2) English to Romanian", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 452, 505, 465 ], "spans": [ { "bbox": [ 105, 452, 505, 465 ], "score": 1.0, "content": "translation using the WMT 2016 en-ro dataset (Bojar et al., 2016); (3) MNLI (Williams et al., 2018)", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 464, 505, 475 ], "spans": [ { "bbox": [ 106, 464, 505, 475 ], "score": 1.0, "content": "is an English natural language inference dataset where models predict whether one sentence entails,", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 474, 505, 486 ], "spans": [ { "bbox": [ 105, 474, 505, 486 ], "score": 1.0, "content": "contradicts, or is neutral to another. (4) SST2 (Socher et al., 2013) is an English sentiment classifi-", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 485, 486, 498 ], "spans": [ { "bbox": [ 105, 485, 486, 498 ], "score": 1.0, "content": "cation benchmark where models predict whether a sentence’s sentiment is positive or negative.", "type": "text" } ], "index": 30 } ], "index": 27.5, "bbox_fs": [ 105, 430, 506, 498 ] }, { "type": "text", "bbox": [ 107, 502, 505, 601 ], "lines": [ { "bbox": [ 105, 501, 506, 516 ], "spans": [ { "bbox": [ 105, 501, 175, 516 ], "score": 1.0, "content": "Setup: We use", "type": "text" }, { "bbox": [ 176, 503, 222, 514 ], "score": 0.73, "content": "{ \\tt B A R T } _ { \\tt L A R G E }", "type": "inline_equation" }, { "bbox": [ 222, 501, 506, 516 ], "score": 1.0, "content": "(Lewis et al., 2020) and a multilingual version of it, mBARTLARGE (Liu", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 514, 505, 526 ], "spans": [ { "bbox": [ 105, 514, 505, 526 ], "score": 1.0, "content": "et al., 2020a), as the underlying pretrained models for XSum and en-ro translation respectively, and", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 523, 506, 537 ], "spans": [ { "bbox": [ 105, 523, 506, 537 ], "score": 1.0, "content": "we use RoBERTaBASE (Liu et al., 2019) for MNLI and SST2. We vary the bottleneck dimension", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 533, 506, 549 ], "spans": [ { "bbox": [ 105, 533, 136, 549 ], "score": 1.0, "content": "within", "type": "text" }, { "bbox": [ 136, 535, 230, 547 ], "score": 0.88, "content": "\\{ 1 , 3 0 , 2 0 0 , 5 1 2 , 1 0 2 4 \\}", "type": "inline_equation" }, { "bbox": [ 230, 533, 506, 549 ], "score": 1.0, "content": "if needed.9 We mainly study adapters, prefix tuning (prefix), and", "type": "text" } ], "index": 34 }, { "bbox": [ 104, 545, 506, 560 ], "spans": [ { "bbox": [ 104, 545, 506, 560 ], "score": 1.0, "content": "LoRA which greatly outperform bitfit and prompt tuning in our experiments. In the analysis sections", "type": "text" } ], "index": 35 }, { "bbox": [ 107, 557, 505, 570 ], "spans": [ { "bbox": [ 107, 557, 147, 568 ], "score": 0.45, "content": "( \\ S 4 . 3 – 4 . 5 )", "type": "inline_equation" }, { "bbox": [ 147, 557, 505, 570 ], "score": 1.0, "content": "we insert adapters either at the attention or FFN layers for easier analysis, but include the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 568, 505, 580 ], "spans": [ { "bbox": [ 105, 568, 505, 580 ], "score": 1.0, "content": "results of inserting at both places in the final comparison (§4.6). We re-implement these methods", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 577, 506, 592 ], "spans": [ { "bbox": [ 105, 577, 506, 592 ], "score": 1.0, "content": "based on their respective public code.10 We use the huggingface transformers library (Wolf et al.,", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 589, 442, 603 ], "spans": [ { "bbox": [ 105, 589, 442, 603 ], "score": 1.0, "content": "2020) for our implementation. Complete setup details can be found in Appendix A.", "type": "text" } ], "index": 39 } ], "index": 35, "bbox_fs": [ 104, 501, 506, 603 ] }, { "type": "text", "bbox": [ 107, 606, 505, 651 ], "lines": [ { "bbox": [ 105, 606, 505, 619 ], "spans": [ { "bbox": [ 105, 606, 245, 619 ], "score": 1.0, "content": "Evaluation: We report ROUGE", "type": "text" }, { "bbox": [ 245, 607, 268, 618 ], "score": 0.32, "content": "1 / 2 / \\mathrm { L }", "type": "inline_equation" }, { "bbox": [ 268, 606, 505, 619 ], "score": 1.0, "content": "scores (R-1/2/L, Lin (2004)) on the XSum test set, BLEU", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 618, 505, 631 ], "spans": [ { "bbox": [ 105, 618, 505, 631 ], "score": 1.0, "content": "scores (Papineni et al., 2002) on the en-ro test set, and accuracy on the MNLI and SST2 dev set.", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 629, 506, 642 ], "spans": [ { "bbox": [ 105, 629, 506, 642 ], "score": 1.0, "content": "For MNLI and SST2, we take the median of five random runs. We also report the number of tuned", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 640, 330, 652 ], "spans": [ { "bbox": [ 105, 640, 330, 652 ], "score": 1.0, "content": "parameters relative to that in full fine-tuning (#params).", "type": "text" } ], "index": 43 } ], "index": 41.5, "bbox_fs": [ 105, 606, 506, 652 ] }, { "type": "text", "bbox": [ 107, 657, 503, 690 ], "lines": [ { "bbox": [ 106, 657, 505, 669 ], "spans": [ { "bbox": [ 106, 657, 505, 669 ], "score": 1.0, "content": "Number of Tunable Parameters: BART and mBART have an encoder-decoder structure that has", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 668, 504, 679 ], "spans": [ { "bbox": [ 106, 668, 504, 679 ], "score": 1.0, "content": "three types of attention: encoder self-attention, decoder self-attention, and decoder cross-attention.", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 678, 505, 692 ], "spans": [ { "bbox": [ 105, 678, 505, 692 ], "score": 1.0, "content": "RoBERTa only has encoder self-attention. For each attention sub-layer, the number of parameters", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 360, 504, 372 ], "spans": [ { "bbox": [ 105, 360, 307, 372 ], "score": 1.0, "content": "used of each method is: (1) prefix tuning prepends", "type": "text", "cross_page": true }, { "bbox": [ 308, 361, 312, 370 ], "score": 0.51, "content": "l", "type": "inline_equation", "cross_page": true }, { "bbox": [ 313, 360, 468, 372 ], "score": 1.0, "content": "vectors to the keys and values and uses", "type": "text", "cross_page": true }, { "bbox": [ 469, 361, 504, 371 ], "score": 0.93, "content": "2 \\times l \\times d", "type": "inline_equation", "cross_page": true } ], "index": 27 }, { "bbox": [ 104, 371, 506, 384 ], "spans": [ { "bbox": [ 104, 371, 216, 384 ], "score": 1.0, "content": "parameters; (2) adapter has", "type": "text", "cross_page": true }, { "bbox": [ 216, 372, 244, 383 ], "score": 0.91, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 244, 371, 261, 384 ], "score": 1.0, "content": "and", "type": "text", "cross_page": true }, { "bbox": [ 262, 371, 281, 383 ], "score": 0.89, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 281, 371, 319, 384 ], "score": 1.0, "content": "thus uses", "type": "text", "cross_page": true }, { "bbox": [ 320, 372, 355, 382 ], "score": 0.92, "content": "2 \\times r \\times d", "type": "inline_equation", "cross_page": true }, { "bbox": [ 355, 371, 506, 384 ], "score": 1.0, "content": "parameters; (3) LoRA employs a pair", "type": "text", "cross_page": true } ], "index": 28 }, { "bbox": [ 105, 382, 506, 395 ], "spans": [ { "bbox": [ 105, 382, 117, 395 ], "score": 1.0, "content": "of", "type": "text", "cross_page": true }, { "bbox": [ 117, 382, 145, 393 ], "score": 0.91, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 145, 382, 163, 395 ], "score": 1.0, "content": "and", "type": "text", "cross_page": true }, { "bbox": [ 163, 382, 182, 394 ], "score": 0.88, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 182, 382, 356, 395 ], "score": 1.0, "content": "for query and value projections, hence uses", "type": "text", "cross_page": true }, { "bbox": [ 356, 383, 393, 393 ], "score": 0.92, "content": "4 \\times r \\times d", "type": "inline_equation", "cross_page": true }, { "bbox": [ 393, 382, 506, 395 ], "score": 1.0, "content": "parameters. For the adapter", "type": "text", "cross_page": true } ], "index": 29 }, { "bbox": [ 105, 392, 505, 406 ], "spans": [ { "bbox": [ 105, 392, 212, 406 ], "score": 1.0, "content": "modification at ffn, it uses", "type": "text", "cross_page": true }, { "bbox": [ 212, 393, 249, 404 ], "score": 0.92, "content": "2 \\times r \\times d", "type": "inline_equation", "cross_page": true }, { "bbox": [ 250, 392, 505, 406 ], "score": 1.0, "content": "parameters which is the same as adapter at attention. Therefore,", "type": "text", "cross_page": true } ], "index": 30 }, { "bbox": [ 105, 403, 505, 416 ], "spans": [ { "bbox": [ 105, 403, 197, 416 ], "score": 1.0, "content": "for a specific value of", "type": "text", "cross_page": true }, { "bbox": [ 198, 406, 204, 414 ], "score": 0.76, "content": "r", "type": "inline_equation", "cross_page": true }, { "bbox": [ 204, 403, 216, 416 ], "score": 1.0, "content": "or", "type": "text", "cross_page": true }, { "bbox": [ 217, 405, 221, 414 ], "score": 0.6, "content": "l", "type": "inline_equation", "cross_page": true }, { "bbox": [ 221, 403, 505, 416 ], "score": 1.0, "content": ", prefix tuning uses the same number of parameters as adapters, while", "type": "text", "cross_page": true } ], "index": 31 }, { "bbox": [ 105, 415, 396, 427 ], "spans": [ { "bbox": [ 105, 415, 396, 427 ], "score": 1.0, "content": "LoRA uses more parameters. More details can be found in Appendix B.", "type": "text", "cross_page": true } ], "index": 32 } ], "index": 45, "bbox_fs": [ 105, 657, 505, 692 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 106, 91, 342, 207 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 91, 342, 207 ], "group_id": 0, "lines": [ { "bbox": [ 106, 91, 342, 207 ], "spans": [ { "bbox": [ 106, 91, 342, 207 ], "score": 0.971, "type": "image", "image_path": "bfd47b7ac5dcc585f8293ecd8c4a66c357d735ee263aa3622fc820936609f5be.jpg" } ] } ], "index": 3.5, "virtual_lines": [ { "bbox": [ 106, 91, 342, 105.5 ], "spans": [], "index": 0 }, { "bbox": [ 106, 105.5, 342, 120.0 ], "spans": [], "index": 1 }, { "bbox": [ 106, 120.0, 342, 134.5 ], "spans": [], "index": 2 }, { "bbox": [ 106, 134.5, 342, 149.0 ], "spans": [], "index": 3 }, { "bbox": [ 106, 149.0, 342, 163.5 ], "spans": [], "index": 4 }, { "bbox": [ 106, 163.5, 342, 178.0 ], "spans": [], "index": 11 }, { "bbox": [ 106, 178.0, 342, 192.5 ], "spans": [], "index": 6 }, { "bbox": [ 106, 192.5, 342, 207.0 ], "spans": [], "index": 7 } ] }, { "type": "image_caption", "bbox": [ 105, 213, 345, 233 ], "group_id": 0, "lines": [ { "bbox": [ 106, 212, 346, 224 ], "spans": [ { "bbox": [ 106, 212, 346, 224 ], "score": 1.0, "content": "Figure 4: Performance of previous state-of-the-art parameter-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 222, 316, 234 ], "spans": [ { "bbox": [ 105, 222, 207, 234 ], "score": 1.0, "content": "efficient tuning methods on", "type": "text" }, { "bbox": [ 207, 223, 231, 232 ], "score": 0.26, "content": "\\bar { \\mathrm { X S u m } }", "type": "inline_equation" }, { "bbox": [ 232, 222, 316, 234 ], "score": 1.0, "content": "(left) and en-ro (right).", "type": "text" } ], "index": 9 } ], "index": 8.5 } ], "index": 6.0 }, { "type": "table", "bbox": [ 360, 147, 488, 224 ], "blocks": [ { "type": "table_caption", "bbox": [ 357, 102, 502, 142 ], "group_id": 1, "lines": [ { "bbox": [ 357, 101, 503, 113 ], "spans": [ { "bbox": [ 357, 101, 503, 113 ], "score": 1.0, "content": "Table 2: Accuracy on the dev set of", "type": "text" } ], "index": 10 }, { "bbox": [ 356, 110, 503, 123 ], "spans": [ { "bbox": [ 356, 110, 503, 123 ], "score": 1.0, "content": "MNLI and SST2. MAM Adapter is", "type": "text" } ], "index": 5 }, { "bbox": [ 356, 122, 503, 132 ], "spans": [ { "bbox": [ 356, 122, 405, 132 ], "score": 1.0, "content": "proposed in", "type": "text" }, { "bbox": [ 405, 122, 423, 132 ], "score": 0.86, "content": "\\ S 4 . 6", "type": "inline_equation" }, { "bbox": [ 423, 122, 503, 132 ], "score": 1.0, "content": ". Bitfit numbers are", "type": "text" } ], "index": 12 }, { "bbox": [ 356, 131, 467, 143 ], "spans": [ { "bbox": [ 356, 131, 467, 143 ], "score": 1.0, "content": "from Ben Zaken et al. (2021).", "type": "text" } ], "index": 13 } ], "index": 11.0 }, { "type": "table_body", "bbox": [ 360, 147, 488, 224 ], "group_id": 1, "lines": [ { "bbox": [ 360, 147, 488, 224 ], "spans": [ { "bbox": [ 360, 147, 488, 224 ], "score": 0.966, "html": "
Method (# params)MNLISST2
Full-FT (100%)87.6±.494.6±.4 93.7
Bitfit (0.1 %) Prefix (0.5%) LoRA (0.5%) Adapter (0.5%)84.7 86.3±.4 87.2±.4 87.2±.294.0±.1 94.2±.2 94.2±.1
MAM Adapter (0.5%) 87.4±.3 94.2±.3
", "type": "table", "image_path": "2be889d74237c1555787a7c44f92253980d2236397ebd647588409b5e5f62eb7.jpg" } ] } ], "index": 14.5, "virtual_lines": [ { "bbox": [ 360, 147, 488, 185.5 ], "spans": [], "index": 14 }, { "bbox": [ 360, 185.5, 488, 224.0 ], "spans": [], "index": 15 } ] } ], "index": 12.75 }, { "type": "table", "bbox": [ 116, 279, 329, 352 ], "blocks": [ { "type": "table_caption", "bbox": [ 105, 245, 334, 275 ], "group_id": 0, "lines": [ { "bbox": [ 105, 243, 335, 256 ], "spans": [ { "bbox": [ 105, 243, 335, 256 ], "score": 1.0, "content": "Table 3: Comparison of different insertion forms for adapters,", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 255, 334, 265 ], "spans": [ { "bbox": [ 105, 255, 334, 265 ], "score": 1.0, "content": "i.e. sequential adapter (SA) and parallel adapter (PA). We in-", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 264, 299, 276 ], "spans": [ { "bbox": [ 106, 264, 299, 276 ], "score": 1.0, "content": "clude the results of prefix tuning as a reference point.", "type": "text" } ], "index": 18 } ], "index": 17 }, { "type": "table_body", "bbox": [ 116, 279, 329, 352 ], "group_id": 0, "lines": [ { "bbox": [ 116, 279, 329, 352 ], "spans": [ { "bbox": [ 116, 279, 329, 352 ], "score": 0.977, "html": "
Method# paramsXSum (R-1/2/L)MT (BLEU)
Prefix,l=2003.6%43.40/20.46/35.5135.6
SA (attn), r=2003.6%42.01/19.30/34.4035.3
SA (ffn),r=2002.4%43.21/19.98/35.0835.6
PA (attn), r=2003.6%43.58/20.31/35.3435.6
PA (ffn),r=2002.4%43.93/20.66/35.6336.4
", "type": "table", "image_path": "dca06a014081f54d894a2ba4c0aed826ca5d7d5c7e868147bf469cfcded10223.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 116, 279, 329, 293.6 ], "spans": [], "index": 21 }, { "bbox": [ 116, 293.6, 329, 308.20000000000005 ], "spans": [], "index": 22 }, { "bbox": [ 116, 308.20000000000005, 329, 322.80000000000007 ], "spans": [], "index": 23 }, { "bbox": [ 116, 322.80000000000007, 329, 337.4000000000001 ], "spans": [], "index": 24 }, { "bbox": [ 116, 337.4000000000001, 329, 352.0000000000001 ], "spans": [], "index": 25 } ] } ], "index": 20.0 }, { "type": "table", "bbox": [ 350, 258, 498, 352 ], "blocks": [ { "type": "table_caption", "bbox": [ 360, 244, 480, 254 ], "group_id": 2, "lines": [ { "bbox": [ 359, 244, 481, 255 ], "spans": [ { "bbox": [ 359, 244, 481, 255 ], "score": 1.0, "content": "Table 4: Results on en-ro dataset.", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "table_body", "bbox": [ 350, 258, 498, 352 ], "group_id": 2, "lines": [ { "bbox": [ 350, 258, 498, 352 ], "spans": [ { "bbox": [ 350, 258, 498, 352 ], "score": 0.97, "html": "
Method# params MT (BLEU)
PA (attn),r=200 Prefix,l=2003.6% 35.6 3.6% 35.6
MH PA (attn),r=2003.6% 35.8
Prefix,l=300.1% 35.2
-gating,l=300.1% 34.9
PA (ffn),r=300.1% 33.0
PA (attn),r=30 MH PA (attn),r=300.1% 33.7 0.1% 35.3
", "type": "table", "image_path": "dc702ee3077619be3b01a1dc0d94ca4c45755440c068f489489ac840181210d7.jpg" } ] } ], "index": 23.0, "virtual_lines": [ { "bbox": [ 350, 258, 498, 305.0 ], "spans": [], "index": 20 }, { "bbox": [ 350, 305.0, 498, 352.0 ], "spans": [], "index": 26 } ] } ], "index": 21.0 }, { "type": "text", "bbox": [ 107, 360, 505, 426 ], "lines": [ { "bbox": [ 105, 360, 504, 372 ], "spans": [ { "bbox": [ 105, 360, 307, 372 ], "score": 1.0, "content": "used of each method is: (1) prefix tuning prepends", "type": "text" }, { "bbox": [ 308, 361, 312, 370 ], "score": 0.51, "content": "l", "type": "inline_equation" }, { "bbox": [ 313, 360, 468, 372 ], "score": 1.0, "content": "vectors to the keys and values and uses", "type": "text" }, { "bbox": [ 469, 361, 504, 371 ], "score": 0.93, "content": "2 \\times l \\times d", "type": "inline_equation" } ], "index": 27 }, { "bbox": [ 104, 371, 506, 384 ], "spans": [ { "bbox": [ 104, 371, 216, 384 ], "score": 1.0, "content": "parameters; (2) adapter has", "type": "text" }, { "bbox": [ 216, 372, 244, 383 ], "score": 0.91, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation" }, { "bbox": [ 244, 371, 261, 384 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 262, 371, 281, 383 ], "score": 0.89, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation" }, { "bbox": [ 281, 371, 319, 384 ], "score": 1.0, "content": "thus uses", "type": "text" }, { "bbox": [ 320, 372, 355, 382 ], "score": 0.92, "content": "2 \\times r \\times d", "type": "inline_equation" }, { "bbox": [ 355, 371, 506, 384 ], "score": 1.0, "content": "parameters; (3) LoRA employs a pair", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 382, 506, 395 ], "spans": [ { "bbox": [ 105, 382, 117, 395 ], "score": 1.0, "content": "of", "type": "text" }, { "bbox": [ 117, 382, 145, 393 ], "score": 0.91, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation" }, { "bbox": [ 145, 382, 163, 395 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 163, 382, 182, 394 ], "score": 0.88, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation" }, { "bbox": [ 182, 382, 356, 395 ], "score": 1.0, "content": "for query and value projections, hence uses", "type": "text" }, { "bbox": [ 356, 383, 393, 393 ], "score": 0.92, "content": "4 \\times r \\times d", "type": "inline_equation" }, { "bbox": [ 393, 382, 506, 395 ], "score": 1.0, "content": "parameters. For the adapter", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 392, 505, 406 ], "spans": [ { "bbox": [ 105, 392, 212, 406 ], "score": 1.0, "content": "modification at ffn, it uses", "type": "text" }, { "bbox": [ 212, 393, 249, 404 ], "score": 0.92, "content": "2 \\times r \\times d", "type": "inline_equation" }, { "bbox": [ 250, 392, 505, 406 ], "score": 1.0, "content": "parameters which is the same as adapter at attention. Therefore,", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 403, 505, 416 ], "spans": [ { "bbox": [ 105, 403, 197, 416 ], "score": 1.0, "content": "for a specific value of", "type": "text" }, { "bbox": [ 198, 406, 204, 414 ], "score": 0.76, "content": "r", "type": "inline_equation" }, { "bbox": [ 204, 403, 216, 416 ], "score": 1.0, "content": "or", "type": "text" }, { "bbox": [ 217, 405, 221, 414 ], "score": 0.6, "content": "l", "type": "inline_equation" }, { "bbox": [ 221, 403, 505, 416 ], "score": 1.0, "content": ", prefix tuning uses the same number of parameters as adapters, while", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 415, 396, 427 ], "spans": [ { "bbox": [ 105, 415, 396, 427 ], "score": 1.0, "content": "LoRA uses more parameters. More details can be found in Appendix B.", "type": "text" } ], "index": 32 } ], "index": 29.5 }, { "type": "title", "bbox": [ 108, 439, 295, 451 ], "lines": [ { "bbox": [ 105, 438, 296, 452 ], "spans": [ { "bbox": [ 105, 438, 296, 452 ], "score": 1.0, "content": "4.2 THE RESULTS OF EXISTING METHODS", "type": "text" } ], "index": 33 } ], "index": 33 }, { "type": "text", "bbox": [ 107, 460, 505, 624 ], "lines": [ { "bbox": [ 106, 460, 505, 472 ], "spans": [ { "bbox": [ 106, 460, 505, 472 ], "score": 1.0, "content": "We first overview the results of existing methods on the four tasks. As shown in Figure 4 and Table 2,", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 471, 506, 484 ], "spans": [ { "bbox": [ 106, 471, 506, 484 ], "score": 1.0, "content": "while existing methods can achieve competitive performance on MNLI and SST2 by tuning fewer", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 480, 506, 496 ], "spans": [ { "bbox": [ 105, 480, 126, 496 ], "score": 1.0, "content": "than", "type": "text" }, { "bbox": [ 126, 482, 140, 492 ], "score": 0.84, "content": "1 \\%", "type": "inline_equation" }, { "bbox": [ 141, 480, 329, 496 ], "score": 1.0, "content": "parameters, a large gap is still present if we add", "type": "text" }, { "bbox": [ 329, 482, 344, 492 ], "score": 0.86, "content": "5 \\%", "type": "inline_equation" }, { "bbox": [ 344, 480, 506, 496 ], "score": 1.0, "content": "parameters in XSum and en-ro. The gap", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 492, 506, 507 ], "spans": [ { "bbox": [ 105, 492, 403, 507 ], "score": 1.0, "content": "remains significant even though we increase the relative parameter size to", "type": "text" }, { "bbox": [ 403, 493, 431, 504 ], "score": 0.9, "content": "> 1 0 \\%", "type": "inline_equation" }, { "bbox": [ 431, 492, 506, 507 ], "score": 1.0, "content": ". Even larger gaps", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 504, 505, 515 ], "spans": [ { "bbox": [ 106, 504, 505, 515 ], "score": 1.0, "content": "have been observed in Raffel et al. (2020) on high-resource MT tasks. This shows that many methods", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 515, 505, 527 ], "spans": [ { "bbox": [ 105, 515, 505, 527 ], "score": 1.0, "content": "that claimed comparable results to full fine-tuning on the GLUE benchmark with an encoder-only", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 525, 505, 538 ], "spans": [ { "bbox": [ 105, 525, 505, 538 ], "score": 1.0, "content": "model (Guo et al., 2021; Ben Zaken et al., 2021; Mahabadi et al., 2021), or on relatively simple", "type": "text" } ], "index": 40 }, { "bbox": [ 104, 536, 506, 549 ], "spans": [ { "bbox": [ 104, 536, 506, 549 ], "score": 1.0, "content": "generation benchmarks such as E2E (Novikova et al., 2017) with an encoder-decoder model (Li &", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 547, 505, 560 ], "spans": [ { "bbox": [ 105, 547, 505, 560 ], "score": 1.0, "content": "Liang, 2021), may not generalize well to other standard benchmarks. The influencing factors could", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 558, 506, 572 ], "spans": [ { "bbox": [ 105, 558, 506, 572 ], "score": 1.0, "content": "be complicated including the number of training samples, task complexity, or model architecture.", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 569, 506, 582 ], "spans": [ { "bbox": [ 106, 569, 506, 582 ], "score": 1.0, "content": "We thus advocate for future research on this line to report results on more diverse benchmarks to", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 580, 506, 593 ], "spans": [ { "bbox": [ 105, 580, 506, 593 ], "score": 1.0, "content": "exhibit a more complete picture of their performance profile. Below, our analysis will mainly focus", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 591, 506, 604 ], "spans": [ { "bbox": [ 105, 591, 506, 604 ], "score": 1.0, "content": "on the XSum and en-ro datasets to better distinguish different design choices. We note that these two", "type": "text" } ], "index": 46 }, { "bbox": [ 106, 603, 505, 614 ], "spans": [ { "bbox": [ 106, 603, 505, 614 ], "score": 1.0, "content": "benchmarks are relatively high-resource performed with an encoder-decoder model (BART), while", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 614, 491, 625 ], "spans": [ { "bbox": [ 106, 614, 469, 625 ], "score": 1.0, "content": "we will discuss the results on MNLI and SST2 with an encoder-only model (RoBERTa) in", "type": "text" }, { "bbox": [ 469, 614, 487, 624 ], "score": 0.87, "content": "\\ S 4 . 6", "type": "inline_equation" }, { "bbox": [ 488, 614, 491, 625 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 48 } ], "index": 41 }, { "type": "title", "bbox": [ 113, 637, 371, 649 ], "lines": [ { "bbox": [ 111, 637, 374, 650 ], "spans": [ { "bbox": [ 111, 637, 374, 650 ], "score": 1.0, "content": ".3 WHICH INSERTION FORM – SEQUENTIAL OR PARALLEL?", "type": "text" } ], "index": 49 } ], "index": 49 }, { "type": "text", "bbox": [ 107, 658, 504, 713 ], "lines": [ { "bbox": [ 107, 659, 505, 671 ], "spans": [ { "bbox": [ 107, 659, 505, 671 ], "score": 1.0, "content": "We first study the insertion form design dimension, comparing the proposed parallel adapter (PA)", "type": "text" } ], "index": 50 }, { "bbox": [ 106, 669, 505, 681 ], "spans": [ { "bbox": [ 106, 669, 505, 681 ], "score": 1.0, "content": "variant to the conventional sequential adapter (SA) over both the attention (att) and FFN modifica-", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 680, 506, 693 ], "spans": [ { "bbox": [ 105, 680, 506, 693 ], "score": 1.0, "content": "tion. We also include prefix tuning as a reference point. As shown in Table 3, prefix tuning, which", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 691, 506, 704 ], "spans": [ { "bbox": [ 105, 691, 506, 704 ], "score": 1.0, "content": "uses parallel insertion, outperforms attention sequential adapters. Further, the parallel adapter is able", "type": "text" } ], "index": 53 }, { "bbox": [ 105, 701, 506, 715 ], "spans": [ { "bbox": [ 105, 701, 433, 715 ], "score": 1.0, "content": "to beat sequential adapters in all cases,11 with PA (ffn) outperforming SA (ffn) by", "type": "text" }, { "bbox": [ 434, 702, 465, 713 ], "score": 0.43, "content": "1 . 7 \\mathrm { R } \\mathrm { - } 2", "type": "inline_equation" }, { "bbox": [ 465, 701, 506, 715 ], "score": 1.0, "content": "points on", "type": "text" } ], "index": 54 } ], "index": 52 } ], "page_idx": 6, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 115, 721, 463, 732 ], "lines": [ { "bbox": [ 115, 719, 466, 734 ], "spans": [ { "bbox": [ 115, 719, 221, 734 ], "score": 1.0, "content": "11More results with different", "type": "text" }, { "bbox": [ 221, 724, 227, 730 ], "score": 0.61, "content": "r", "type": "inline_equation" }, { "bbox": [ 227, 719, 466, 734 ], "score": 1.0, "content": "can be found in Appendix C, which exhibits similar observations.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 294, 39 ], "spans": [ { "bbox": [ 106, 25, 294, 39 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 759 ], "lines": [ { "bbox": [ 302, 750, 309, 762 ], "spans": [ { "bbox": [ 302, 750, 309, 762 ], "score": 1.0, "content": "7", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 106, 91, 342, 207 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 91, 342, 207 ], "group_id": 0, "lines": [ { "bbox": [ 106, 91, 342, 207 ], "spans": [ { "bbox": [ 106, 91, 342, 207 ], "score": 0.971, "type": "image", "image_path": "bfd47b7ac5dcc585f8293ecd8c4a66c357d735ee263aa3622fc820936609f5be.jpg" } ] } ], "index": 3.5, "virtual_lines": [ { "bbox": [ 106, 91, 342, 105.5 ], "spans": [], "index": 0 }, { "bbox": [ 106, 105.5, 342, 120.0 ], "spans": [], "index": 1 }, { "bbox": [ 106, 120.0, 342, 134.5 ], "spans": [], "index": 2 }, { "bbox": [ 106, 134.5, 342, 149.0 ], "spans": [], "index": 3 }, { "bbox": [ 106, 149.0, 342, 163.5 ], "spans": [], "index": 4 }, { "bbox": [ 106, 163.5, 342, 178.0 ], "spans": [], "index": 11 }, { "bbox": [ 106, 178.0, 342, 192.5 ], "spans": [], "index": 6 }, { "bbox": [ 106, 192.5, 342, 207.0 ], "spans": [], "index": 7 } ] }, { "type": "image_caption", "bbox": [ 105, 213, 345, 233 ], "group_id": 0, "lines": [ { "bbox": [ 106, 212, 346, 224 ], "spans": [ { "bbox": [ 106, 212, 346, 224 ], "score": 1.0, "content": "Figure 4: Performance of previous state-of-the-art parameter-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 222, 316, 234 ], "spans": [ { "bbox": [ 105, 222, 207, 234 ], "score": 1.0, "content": "efficient tuning methods on", "type": "text" }, { "bbox": [ 207, 223, 231, 232 ], "score": 0.26, "content": "\\bar { \\mathrm { X S u m } }", "type": "inline_equation" }, { "bbox": [ 232, 222, 316, 234 ], "score": 1.0, "content": "(left) and en-ro (right).", "type": "text" } ], "index": 9 } ], "index": 8.5 } ], "index": 6.0 }, { "type": "table", "bbox": [ 360, 147, 488, 224 ], "blocks": [ { "type": "table_caption", "bbox": [ 357, 102, 502, 142 ], "group_id": 1, "lines": [ { "bbox": [ 357, 101, 503, 113 ], "spans": [ { "bbox": [ 357, 101, 503, 113 ], "score": 1.0, "content": "Table 2: Accuracy on the dev set of", "type": "text" } ], "index": 10 }, { "bbox": [ 356, 110, 503, 123 ], "spans": [ { "bbox": [ 356, 110, 503, 123 ], "score": 1.0, "content": "MNLI and SST2. MAM Adapter is", "type": "text" } ], "index": 5 }, { "bbox": [ 356, 122, 503, 132 ], "spans": [ { "bbox": [ 356, 122, 405, 132 ], "score": 1.0, "content": "proposed in", "type": "text" }, { "bbox": [ 405, 122, 423, 132 ], "score": 0.86, "content": "\\ S 4 . 6", "type": "inline_equation" }, { "bbox": [ 423, 122, 503, 132 ], "score": 1.0, "content": ". Bitfit numbers are", "type": "text" } ], "index": 12 }, { "bbox": [ 356, 131, 467, 143 ], "spans": [ { "bbox": [ 356, 131, 467, 143 ], "score": 1.0, "content": "from Ben Zaken et al. (2021).", "type": "text" } ], "index": 13 } ], "index": 11.0 }, { "type": "table_body", "bbox": [ 360, 147, 488, 224 ], "group_id": 1, "lines": [ { "bbox": [ 360, 147, 488, 224 ], "spans": [ { "bbox": [ 360, 147, 488, 224 ], "score": 0.966, "html": "
Method (# params)MNLISST2
Full-FT (100%)87.6±.494.6±.4 93.7
Bitfit (0.1 %) Prefix (0.5%) LoRA (0.5%) Adapter (0.5%)84.7 86.3±.4 87.2±.4 87.2±.294.0±.1 94.2±.2 94.2±.1
MAM Adapter (0.5%) 87.4±.3 94.2±.3
", "type": "table", "image_path": "2be889d74237c1555787a7c44f92253980d2236397ebd647588409b5e5f62eb7.jpg" } ] } ], "index": 14.5, "virtual_lines": [ { "bbox": [ 360, 147, 488, 185.5 ], "spans": [], "index": 14 }, { "bbox": [ 360, 185.5, 488, 224.0 ], "spans": [], "index": 15 } ] } ], "index": 12.75 }, { "type": "table", "bbox": [ 116, 279, 329, 352 ], "blocks": [ { "type": "table_caption", "bbox": [ 105, 245, 334, 275 ], "group_id": 0, "lines": [ { "bbox": [ 105, 243, 335, 256 ], "spans": [ { "bbox": [ 105, 243, 335, 256 ], "score": 1.0, "content": "Table 3: Comparison of different insertion forms for adapters,", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 255, 334, 265 ], "spans": [ { "bbox": [ 105, 255, 334, 265 ], "score": 1.0, "content": "i.e. sequential adapter (SA) and parallel adapter (PA). We in-", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 264, 299, 276 ], "spans": [ { "bbox": [ 106, 264, 299, 276 ], "score": 1.0, "content": "clude the results of prefix tuning as a reference point.", "type": "text" } ], "index": 18 } ], "index": 17 }, { "type": "table_body", "bbox": [ 116, 279, 329, 352 ], "group_id": 0, "lines": [ { "bbox": [ 116, 279, 329, 352 ], "spans": [ { "bbox": [ 116, 279, 329, 352 ], "score": 0.977, "html": "
Method# paramsXSum (R-1/2/L)MT (BLEU)
Prefix,l=2003.6%43.40/20.46/35.5135.6
SA (attn), r=2003.6%42.01/19.30/34.4035.3
SA (ffn),r=2002.4%43.21/19.98/35.0835.6
PA (attn), r=2003.6%43.58/20.31/35.3435.6
PA (ffn),r=2002.4%43.93/20.66/35.6336.4
", "type": "table", "image_path": "dca06a014081f54d894a2ba4c0aed826ca5d7d5c7e868147bf469cfcded10223.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 116, 279, 329, 293.6 ], "spans": [], "index": 21 }, { "bbox": [ 116, 293.6, 329, 308.20000000000005 ], "spans": [], "index": 22 }, { "bbox": [ 116, 308.20000000000005, 329, 322.80000000000007 ], "spans": [], "index": 23 }, { "bbox": [ 116, 322.80000000000007, 329, 337.4000000000001 ], "spans": [], "index": 24 }, { "bbox": [ 116, 337.4000000000001, 329, 352.0000000000001 ], "spans": [], "index": 25 } ] } ], "index": 20.0 }, { "type": "table", "bbox": [ 350, 258, 498, 352 ], "blocks": [ { "type": "table_caption", "bbox": [ 360, 244, 480, 254 ], "group_id": 2, "lines": [ { "bbox": [ 359, 244, 481, 255 ], "spans": [ { "bbox": [ 359, 244, 481, 255 ], "score": 1.0, "content": "Table 4: Results on en-ro dataset.", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "table_body", "bbox": [ 350, 258, 498, 352 ], "group_id": 2, "lines": [ { "bbox": [ 350, 258, 498, 352 ], "spans": [ { "bbox": [ 350, 258, 498, 352 ], "score": 0.97, "html": "
Method# params MT (BLEU)
PA (attn),r=200 Prefix,l=2003.6% 35.6 3.6% 35.6
MH PA (attn),r=2003.6% 35.8
Prefix,l=300.1% 35.2
-gating,l=300.1% 34.9
PA (ffn),r=300.1% 33.0
PA (attn),r=30 MH PA (attn),r=300.1% 33.7 0.1% 35.3
", "type": "table", "image_path": "dc702ee3077619be3b01a1dc0d94ca4c45755440c068f489489ac840181210d7.jpg" } ] } ], "index": 23.0, "virtual_lines": [ { "bbox": [ 350, 258, 498, 305.0 ], "spans": [], "index": 20 }, { "bbox": [ 350, 305.0, 498, 352.0 ], "spans": [], "index": 26 } ] } ], "index": 21.0 }, { "type": "text", "bbox": [ 107, 360, 505, 426 ], "lines": [], "index": 29.5, "bbox_fs": [ 104, 360, 506, 427 ], "lines_deleted": true }, { "type": "title", "bbox": [ 108, 439, 295, 451 ], "lines": [ { "bbox": [ 105, 438, 296, 452 ], "spans": [ { "bbox": [ 105, 438, 296, 452 ], "score": 1.0, "content": "4.2 THE RESULTS OF EXISTING METHODS", "type": "text" } ], "index": 33 } ], "index": 33 }, { "type": "text", "bbox": [ 107, 460, 505, 624 ], "lines": [ { "bbox": [ 106, 460, 505, 472 ], "spans": [ { "bbox": [ 106, 460, 505, 472 ], "score": 1.0, "content": "We first overview the results of existing methods on the four tasks. As shown in Figure 4 and Table 2,", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 471, 506, 484 ], "spans": [ { "bbox": [ 106, 471, 506, 484 ], "score": 1.0, "content": "while existing methods can achieve competitive performance on MNLI and SST2 by tuning fewer", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 480, 506, 496 ], "spans": [ { "bbox": [ 105, 480, 126, 496 ], "score": 1.0, "content": "than", "type": "text" }, { "bbox": [ 126, 482, 140, 492 ], "score": 0.84, "content": "1 \\%", "type": "inline_equation" }, { "bbox": [ 141, 480, 329, 496 ], "score": 1.0, "content": "parameters, a large gap is still present if we add", "type": "text" }, { "bbox": [ 329, 482, 344, 492 ], "score": 0.86, "content": "5 \\%", "type": "inline_equation" }, { "bbox": [ 344, 480, 506, 496 ], "score": 1.0, "content": "parameters in XSum and en-ro. The gap", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 492, 506, 507 ], "spans": [ { "bbox": [ 105, 492, 403, 507 ], "score": 1.0, "content": "remains significant even though we increase the relative parameter size to", "type": "text" }, { "bbox": [ 403, 493, 431, 504 ], "score": 0.9, "content": "> 1 0 \\%", "type": "inline_equation" }, { "bbox": [ 431, 492, 506, 507 ], "score": 1.0, "content": ". Even larger gaps", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 504, 505, 515 ], "spans": [ { "bbox": [ 106, 504, 505, 515 ], "score": 1.0, "content": "have been observed in Raffel et al. (2020) on high-resource MT tasks. This shows that many methods", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 515, 505, 527 ], "spans": [ { "bbox": [ 105, 515, 505, 527 ], "score": 1.0, "content": "that claimed comparable results to full fine-tuning on the GLUE benchmark with an encoder-only", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 525, 505, 538 ], "spans": [ { "bbox": [ 105, 525, 505, 538 ], "score": 1.0, "content": "model (Guo et al., 2021; Ben Zaken et al., 2021; Mahabadi et al., 2021), or on relatively simple", "type": "text" } ], "index": 40 }, { "bbox": [ 104, 536, 506, 549 ], "spans": [ { "bbox": [ 104, 536, 506, 549 ], "score": 1.0, "content": "generation benchmarks such as E2E (Novikova et al., 2017) with an encoder-decoder model (Li &", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 547, 505, 560 ], "spans": [ { "bbox": [ 105, 547, 505, 560 ], "score": 1.0, "content": "Liang, 2021), may not generalize well to other standard benchmarks. The influencing factors could", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 558, 506, 572 ], "spans": [ { "bbox": [ 105, 558, 506, 572 ], "score": 1.0, "content": "be complicated including the number of training samples, task complexity, or model architecture.", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 569, 506, 582 ], "spans": [ { "bbox": [ 106, 569, 506, 582 ], "score": 1.0, "content": "We thus advocate for future research on this line to report results on more diverse benchmarks to", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 580, 506, 593 ], "spans": [ { "bbox": [ 105, 580, 506, 593 ], "score": 1.0, "content": "exhibit a more complete picture of their performance profile. Below, our analysis will mainly focus", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 591, 506, 604 ], "spans": [ { "bbox": [ 105, 591, 506, 604 ], "score": 1.0, "content": "on the XSum and en-ro datasets to better distinguish different design choices. We note that these two", "type": "text" } ], "index": 46 }, { "bbox": [ 106, 603, 505, 614 ], "spans": [ { "bbox": [ 106, 603, 505, 614 ], "score": 1.0, "content": "benchmarks are relatively high-resource performed with an encoder-decoder model (BART), while", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 614, 491, 625 ], "spans": [ { "bbox": [ 106, 614, 469, 625 ], "score": 1.0, "content": "we will discuss the results on MNLI and SST2 with an encoder-only model (RoBERTa) in", "type": "text" }, { "bbox": [ 469, 614, 487, 624 ], "score": 0.87, "content": "\\ S 4 . 6", "type": "inline_equation" }, { "bbox": [ 488, 614, 491, 625 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 48 } ], "index": 41, "bbox_fs": [ 104, 460, 506, 625 ] }, { "type": "title", "bbox": [ 113, 637, 371, 649 ], "lines": [ { "bbox": [ 111, 637, 374, 650 ], "spans": [ { "bbox": [ 111, 637, 374, 650 ], "score": 1.0, "content": ".3 WHICH INSERTION FORM – SEQUENTIAL OR PARALLEL?", "type": "text" } ], "index": 49 } ], "index": 49 }, { "type": "text", "bbox": [ 107, 658, 504, 713 ], "lines": [ { "bbox": [ 107, 659, 505, 671 ], "spans": [ { "bbox": [ 107, 659, 505, 671 ], "score": 1.0, "content": "We first study the insertion form design dimension, comparing the proposed parallel adapter (PA)", "type": "text" } ], "index": 50 }, { "bbox": [ 106, 669, 505, 681 ], "spans": [ { "bbox": [ 106, 669, 505, 681 ], "score": 1.0, "content": "variant to the conventional sequential adapter (SA) over both the attention (att) and FFN modifica-", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 680, 506, 693 ], "spans": [ { "bbox": [ 105, 680, 506, 693 ], "score": 1.0, "content": "tion. We also include prefix tuning as a reference point. As shown in Table 3, prefix tuning, which", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 691, 506, 704 ], "spans": [ { "bbox": [ 105, 691, 506, 704 ], "score": 1.0, "content": "uses parallel insertion, outperforms attention sequential adapters. Further, the parallel adapter is able", "type": "text" } ], "index": 53 }, { "bbox": [ 105, 701, 506, 715 ], "spans": [ { "bbox": [ 105, 701, 433, 715 ], "score": 1.0, "content": "to beat sequential adapters in all cases,11 with PA (ffn) outperforming SA (ffn) by", "type": "text" }, { "bbox": [ 434, 702, 465, 713 ], "score": 0.43, "content": "1 . 7 \\mathrm { R } \\mathrm { - } 2", "type": "inline_equation" }, { "bbox": [ 465, 701, 506, 715 ], "score": 1.0, "content": "points on", "type": "text" } ], "index": 54 } ], "index": 52, "bbox_fs": [ 105, 659, 506, 715 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 116, 82, 473, 177 ], "blocks": [ { "type": "image_body", "bbox": [ 116, 82, 473, 177 ], "group_id": 0, "lines": [ { "bbox": [ 116, 82, 473, 177 ], "spans": [ { "bbox": [ 116, 82, 473, 177 ], "score": 0.614, "type": "image", "image_path": "08efaea948e4b8e31896a063d3d9ff883d744e25c54932f026083a5374a43c9f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 116, 82, 473, 113.66666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 116, 113.66666666666667, 473, 145.33333333333334 ], "spans": [], "index": 1 }, { "bbox": [ 116, 145.33333333333334, 473, 177.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 179, 504, 200 ], "group_id": 0, "lines": [ { "bbox": [ 105, 177, 505, 192 ], "spans": [ { "bbox": [ 105, 177, 505, 192 ], "score": 1.0, "content": "Figure 5: Results on XSum (left) and en-ro (right). PA represents parallel adapter. Blue and red markers apply", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 189, 401, 200 ], "spans": [ { "bbox": [ 106, 189, 401, 200 ], "score": 1.0, "content": "modifications at attention and FFN sub-layers respectively (best viewed in color).", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "text", "bbox": [ 107, 206, 503, 228 ], "lines": [ { "bbox": [ 105, 205, 505, 219 ], "spans": [ { "bbox": [ 105, 205, 505, 219 ], "score": 1.0, "content": "XSum and 0.8 BLEU points on en-ro respectively. Given the superior results of parallel adapters", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 217, 439, 230 ], "spans": [ { "bbox": [ 105, 217, 439, 230 ], "score": 1.0, "content": "over sequential adapters, we focus on parallel adapter results in following sections.", "type": "text" } ], "index": 6 } ], "index": 5.5 }, { "type": "title", "bbox": [ 107, 242, 391, 253 ], "lines": [ { "bbox": [ 106, 241, 392, 254 ], "spans": [ { "bbox": [ 106, 241, 392, 254 ], "score": 1.0, "content": "4.4 WHICH MODIFIED REPRESENTATION – ATTENTION OR FFN?", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 106, 262, 505, 362 ], "lines": [ { "bbox": [ 106, 261, 505, 275 ], "spans": [ { "bbox": [ 106, 261, 505, 275 ], "score": 1.0, "content": "Setup: We now study the effect of modifying different representations. We mainly compare at-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 273, 505, 286 ], "spans": [ { "bbox": [ 105, 273, 505, 286 ], "score": 1.0, "content": "tention and FFN modification. For easier analysis we categorize methods that modifies any hidden", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 285, 505, 298 ], "spans": [ { "bbox": [ 105, 285, 505, 298 ], "score": 1.0, "content": "representations in the attention sub-layer (e.g. the head output, query, etc) as modifying the atten-", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 296, 505, 308 ], "spans": [ { "bbox": [ 106, 296, 505, 308 ], "score": 1.0, "content": "tion module. We compare parallel adapters at attention and FFN and prefix tuning. We also transfer", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 306, 505, 319 ], "spans": [ { "bbox": [ 106, 306, 505, 319 ], "score": 1.0, "content": "the FFN modification to LoRA to have a LoRA (ffn) variant for a complete comparison. Specifi-", "type": "text" } ], "index": 12 }, { "bbox": [ 104, 316, 507, 330 ], "spans": [ { "bbox": [ 104, 316, 428, 330 ], "score": 1.0, "content": "cally, we use LoRA to approximate the parameter updates for the FFN weights", "type": "text" }, { "bbox": [ 428, 316, 486, 329 ], "score": 0.93, "content": "\\dot { W _ { 1 } } \\in \\mathbb { R } ^ { d \\times \\dot { d _ { m } } }", "type": "inline_equation" }, { "bbox": [ 486, 316, 507, 330 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 13 }, { "bbox": [ 107, 326, 506, 343 ], "spans": [ { "bbox": [ 107, 327, 164, 339 ], "score": 0.92, "content": "\\pmb { W } _ { 2 } \\in \\mathbb { R } ^ { d _ { m } \\times d }", "type": "inline_equation" }, { "bbox": [ 164, 326, 214, 343 ], "score": 1.0, "content": ". In this case", "type": "text" }, { "bbox": [ 215, 329, 234, 340 ], "score": 0.89, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation" }, { "bbox": [ 235, 326, 285, 343 ], "score": 1.0, "content": "in LoRA for", "type": "text" }, { "bbox": [ 286, 329, 302, 339 ], "score": 0.87, "content": "W _ { 1 }", "type": "inline_equation" }, { "bbox": [ 302, 326, 350, 343 ], "score": 1.0, "content": "(similar for", "type": "text" }, { "bbox": [ 350, 329, 378, 340 ], "score": 0.91, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation" }, { "bbox": [ 378, 326, 389, 343 ], "score": 1.0, "content": "of", "type": "text" }, { "bbox": [ 389, 329, 406, 340 ], "score": 0.86, "content": "W _ { 2 }", "type": "inline_equation" }, { "bbox": [ 406, 326, 506, 343 ], "score": 1.0, "content": ") would have dimensions", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 339, 505, 352 ], "spans": [ { "bbox": [ 106, 339, 118, 352 ], "score": 1.0, "content": "of", "type": "text" }, { "bbox": [ 118, 340, 149, 350 ], "score": 0.91, "content": "r \\times d _ { m }", "type": "inline_equation" }, { "bbox": [ 149, 339, 181, 352 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 181, 339, 221, 350 ], "score": 0.92, "content": "d _ { m } = 4 d", "type": "inline_equation" }, { "bbox": [ 221, 339, 285, 352 ], "score": 1.0, "content": "as described in", "type": "text" }, { "bbox": [ 286, 340, 304, 350 ], "score": 0.83, "content": "\\ S 2 . 1", "type": "inline_equation" }, { "bbox": [ 304, 339, 434, 352 ], "score": 1.0, "content": ". Thus we typically use smaller", "type": "text" }, { "bbox": [ 434, 341, 440, 349 ], "score": 0.72, "content": "r", "type": "inline_equation" }, { "bbox": [ 441, 339, 505, 352 ], "score": 1.0, "content": "for LoRA (ffn)", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 350, 417, 363 ], "spans": [ { "bbox": [ 106, 350, 417, 363 ], "score": 1.0, "content": "than other methods to match their overall parameter size in later experiments.", "type": "text" } ], "index": 16 } ], "index": 12 }, { "type": "text", "bbox": [ 106, 367, 505, 477 ], "lines": [ { "bbox": [ 106, 367, 504, 379 ], "spans": [ { "bbox": [ 106, 367, 504, 379 ], "score": 1.0, "content": "Results: As shown in Figure 5, any method with FFN modification outperforms all the methods", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 379, 504, 390 ], "spans": [ { "bbox": [ 106, 379, 504, 390 ], "score": 1.0, "content": "with attention modification in all cases (the red markers are generally above all the blue ones, the", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 389, 505, 402 ], "spans": [ { "bbox": [ 106, 389, 225, 402 ], "score": 1.0, "content": "only exception is ffn-PA with", "type": "text" }, { "bbox": [ 225, 389, 247, 400 ], "score": 0.86, "content": "2 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 247, 389, 505, 402 ], "score": 1.0, "content": "params), often with fewer parameters. Second, the same method", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 399, 505, 414 ], "spans": [ { "bbox": [ 105, 399, 505, 414 ], "score": 1.0, "content": "applied at FFN always improves over its attention counterpart. For example, LoRA (ffn) improves", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 410, 505, 425 ], "spans": [ { "bbox": [ 105, 410, 505, 425 ], "score": 1.0, "content": "LoRA (attn) by 1 R-2 points on XSum. We also highlight that prefix tuning does not keep improving", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 422, 505, 435 ], "spans": [ { "bbox": [ 106, 422, 505, 435 ], "score": 1.0, "content": "when we further increase the capacity, which is also observed in Li & Liang (2021). These results", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 433, 506, 447 ], "spans": [ { "bbox": [ 105, 433, 506, 447 ], "score": 1.0, "content": "suggest that FFN modification can utilize the added parameters more effectively than attention, no", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 444, 505, 457 ], "spans": [ { "bbox": [ 105, 444, 505, 457 ], "score": 1.0, "content": "matter what the functional form or composition function is. We hypothesize that this is because", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 455, 505, 468 ], "spans": [ { "bbox": [ 106, 455, 505, 468 ], "score": 1.0, "content": "the FFN learns task-specific textual patterns (Geva et al., 2021), while attention learns pairwise", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 466, 445, 479 ], "spans": [ { "bbox": [ 105, 466, 445, 479 ], "score": 1.0, "content": "positional interactions which do not require large capacity for adapting to new tasks.", "type": "text" } ], "index": 26 } ], "index": 21.5 }, { "type": "text", "bbox": [ 107, 483, 505, 614 ], "lines": [ { "bbox": [ 104, 482, 506, 496 ], "spans": [ { "bbox": [ 104, 482, 257, 496 ], "score": 1.0, "content": "Is the story different when we use", "type": "text" }, { "bbox": [ 258, 482, 280, 494 ], "score": 0.87, "content": "0 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 280, 482, 357, 496 ], "score": 1.0, "content": "parameters? In", "type": "text" }, { "bbox": [ 357, 483, 376, 494 ], "score": 0.82, "content": "\\ S 3 . 1", "type": "inline_equation" }, { "bbox": [ 376, 482, 506, 496 ], "score": 1.0, "content": "we reason that prefix tuning is", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 495, 505, 506 ], "spans": [ { "bbox": [ 105, 495, 505, 506 ], "score": 1.0, "content": "more expressive than adapters (attn), which, however, is not reflected in Figure 5. We conjecture", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 505, 505, 517 ], "spans": [ { "bbox": [ 105, 505, 505, 517 ], "score": 1.0, "content": "that this is because multi-head attention is only superior when the parameter budget is small. To", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 515, 506, 529 ], "spans": [ { "bbox": [ 105, 515, 455, 529 ], "score": 1.0, "content": "validate this hypothesis, we compare prefix tuning to parallel adapters when they add", "type": "text" }, { "bbox": [ 455, 516, 478, 527 ], "score": 0.88, "content": "0 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 478, 515, 506, 529 ], "score": 1.0, "content": "of the", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 527, 505, 539 ], "spans": [ { "bbox": [ 105, 527, 505, 539 ], "score": 1.0, "content": "pretrained parameters. To ablate the impact of the composition function, we also report the results", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 538, 505, 550 ], "spans": [ { "bbox": [ 105, 538, 273, 550 ], "score": 1.0, "content": "of removing the gating in prefix tuning as", "type": "text" }, { "bbox": [ 273, 538, 308, 549 ], "score": 0.91, "content": "h + \\Delta h", "type": "inline_equation" }, { "bbox": [ 308, 538, 505, 550 ], "score": 1.0, "content": ". We include the results of the multi-head parallel", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 549, 505, 561 ], "spans": [ { "bbox": [ 105, 549, 261, 561 ], "score": 1.0, "content": "adapter variant (MH PA) described in", "type": "text" }, { "bbox": [ 261, 549, 280, 560 ], "score": 0.87, "content": "\\ S 3 . 3", "type": "inline_equation" }, { "bbox": [ 280, 549, 505, 561 ], "score": 1.0, "content": ". As shown in Table 4, the multi-head methods – prefix", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 560, 506, 572 ], "spans": [ { "bbox": [ 105, 560, 470, 572 ], "score": 1.0, "content": "tuning and MH PA (attn) – outperform all others by at least 1.6 BLEU points when using", "type": "text" }, { "bbox": [ 470, 560, 493, 571 ], "score": 0.87, "content": "0 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 493, 560, 506, 572 ], "score": 1.0, "content": "of", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 569, 506, 585 ], "spans": [ { "bbox": [ 105, 569, 261, 585 ], "score": 1.0, "content": "the parameters. Surprisingly, reducing", "type": "text" }, { "bbox": [ 261, 571, 266, 581 ], "score": 0.32, "content": "l", "type": "inline_equation" }, { "bbox": [ 266, 569, 506, 585 ], "score": 1.0, "content": "from 200 to 30 only causes 0.4 BLEU loss for prefix tuning", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 582, 505, 594 ], "spans": [ { "bbox": [ 106, 582, 505, 594 ], "score": 1.0, "content": "while PA (attn) loses 1.9 points. The gating composition function in prefix tuning slightly helps the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 593, 505, 605 ], "spans": [ { "bbox": [ 105, 593, 505, 605 ], "score": 1.0, "content": "results by 0.3 points. We highlight that the MH parallel adapter improves the single-headed version", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 604, 438, 615 ], "spans": [ { "bbox": [ 105, 604, 438, 615 ], "score": 1.0, "content": "by 1.6 points, which again verifies the effectiveness of the multi-head formulation.", "type": "text" } ], "index": 38 } ], "index": 32.5 }, { "type": "text", "bbox": [ 107, 620, 505, 665 ], "lines": [ { "bbox": [ 105, 620, 505, 632 ], "spans": [ { "bbox": [ 105, 620, 505, 632 ], "score": 1.0, "content": "Combining the results in Figure 5 and Table 4, we conclude that modifying head attention shows the", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 632, 505, 644 ], "spans": [ { "bbox": [ 105, 632, 505, 644 ], "score": 1.0, "content": "best results when the parameter budget is very small, while the FFN can better utilize modifications", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 643, 505, 655 ], "spans": [ { "bbox": [ 106, 643, 505, 655 ], "score": 1.0, "content": "at larger capacities. This suggests that it may be effective to allocate a larger parameter budget to", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 653, 471, 667 ], "spans": [ { "bbox": [ 105, 653, 471, 667 ], "score": 1.0, "content": "FFN modification instead of treating attention and FFN equally as in Houlsby et al. (2019).", "type": "text" } ], "index": 42 } ], "index": 40.5 }, { "type": "title", "bbox": [ 108, 678, 279, 689 ], "lines": [ { "bbox": [ 106, 678, 280, 691 ], "spans": [ { "bbox": [ 106, 678, 280, 691 ], "score": 1.0, "content": "4.5 WHICH COMPOSITION FUNCTION?", "type": "text" } ], "index": 43 } ], "index": 43 }, { "type": "text", "bbox": [ 108, 699, 504, 731 ], "lines": [ { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 312, 711 ], "score": 1.0, "content": "We have presented three composition functions in", "type": "text" }, { "bbox": [ 313, 699, 331, 710 ], "score": 0.86, "content": "\\ S 3 . 2", "type": "inline_equation" }, { "bbox": [ 332, 699, 505, 711 ], "score": 1.0, "content": ": simple addition (adapter), gated addition", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 710, 505, 721 ], "spans": [ { "bbox": [ 106, 710, 505, 721 ], "score": 1.0, "content": "(prefix tuning) and scaled addition (LoRA). As it is unnatural to incorporate the exact gated ad-", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 720, 504, 732 ], "spans": [ { "bbox": [ 106, 720, 504, 732 ], "score": 1.0, "content": "dition into methods whose functional form does not use softmax, we examine the other two by", "type": "text" } ], "index": 46 } ], "index": 45 } ], "page_idx": 7, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 294, 39 ], "spans": [ { "bbox": [ 106, 25, 294, 39 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "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" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 116, 82, 473, 177 ], "blocks": [ { "type": "image_body", "bbox": [ 116, 82, 473, 177 ], "group_id": 0, "lines": [ { "bbox": [ 116, 82, 473, 177 ], "spans": [ { "bbox": [ 116, 82, 473, 177 ], "score": 0.614, "type": "image", "image_path": "08efaea948e4b8e31896a063d3d9ff883d744e25c54932f026083a5374a43c9f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 116, 82, 473, 113.66666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 116, 113.66666666666667, 473, 145.33333333333334 ], "spans": [], "index": 1 }, { "bbox": [ 116, 145.33333333333334, 473, 177.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 179, 504, 200 ], "group_id": 0, "lines": [ { "bbox": [ 105, 177, 505, 192 ], "spans": [ { "bbox": [ 105, 177, 505, 192 ], "score": 1.0, "content": "Figure 5: Results on XSum (left) and en-ro (right). PA represents parallel adapter. Blue and red markers apply", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 189, 401, 200 ], "spans": [ { "bbox": [ 106, 189, 401, 200 ], "score": 1.0, "content": "modifications at attention and FFN sub-layers respectively (best viewed in color).", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "text", "bbox": [ 107, 206, 503, 228 ], "lines": [ { "bbox": [ 105, 205, 505, 219 ], "spans": [ { "bbox": [ 105, 205, 505, 219 ], "score": 1.0, "content": "XSum and 0.8 BLEU points on en-ro respectively. Given the superior results of parallel adapters", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 217, 439, 230 ], "spans": [ { "bbox": [ 105, 217, 439, 230 ], "score": 1.0, "content": "over sequential adapters, we focus on parallel adapter results in following sections.", "type": "text" } ], "index": 6 } ], "index": 5.5, "bbox_fs": [ 105, 205, 505, 230 ] }, { "type": "title", "bbox": [ 107, 242, 391, 253 ], "lines": [ { "bbox": [ 106, 241, 392, 254 ], "spans": [ { "bbox": [ 106, 241, 392, 254 ], "score": 1.0, "content": "4.4 WHICH MODIFIED REPRESENTATION – ATTENTION OR FFN?", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 106, 262, 505, 362 ], "lines": [ { "bbox": [ 106, 261, 505, 275 ], "spans": [ { "bbox": [ 106, 261, 505, 275 ], "score": 1.0, "content": "Setup: We now study the effect of modifying different representations. We mainly compare at-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 273, 505, 286 ], "spans": [ { "bbox": [ 105, 273, 505, 286 ], "score": 1.0, "content": "tention and FFN modification. For easier analysis we categorize methods that modifies any hidden", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 285, 505, 298 ], "spans": [ { "bbox": [ 105, 285, 505, 298 ], "score": 1.0, "content": "representations in the attention sub-layer (e.g. the head output, query, etc) as modifying the atten-", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 296, 505, 308 ], "spans": [ { "bbox": [ 106, 296, 505, 308 ], "score": 1.0, "content": "tion module. We compare parallel adapters at attention and FFN and prefix tuning. We also transfer", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 306, 505, 319 ], "spans": [ { "bbox": [ 106, 306, 505, 319 ], "score": 1.0, "content": "the FFN modification to LoRA to have a LoRA (ffn) variant for a complete comparison. Specifi-", "type": "text" } ], "index": 12 }, { "bbox": [ 104, 316, 507, 330 ], "spans": [ { "bbox": [ 104, 316, 428, 330 ], "score": 1.0, "content": "cally, we use LoRA to approximate the parameter updates for the FFN weights", "type": "text" }, { "bbox": [ 428, 316, 486, 329 ], "score": 0.93, "content": "\\dot { W _ { 1 } } \\in \\mathbb { R } ^ { d \\times \\dot { d _ { m } } }", "type": "inline_equation" }, { "bbox": [ 486, 316, 507, 330 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 13 }, { "bbox": [ 107, 326, 506, 343 ], "spans": [ { "bbox": [ 107, 327, 164, 339 ], "score": 0.92, "content": "\\pmb { W } _ { 2 } \\in \\mathbb { R } ^ { d _ { m } \\times d }", "type": "inline_equation" }, { "bbox": [ 164, 326, 214, 343 ], "score": 1.0, "content": ". In this case", "type": "text" }, { "bbox": [ 215, 329, 234, 340 ], "score": 0.89, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation" }, { "bbox": [ 235, 326, 285, 343 ], "score": 1.0, "content": "in LoRA for", "type": "text" }, { "bbox": [ 286, 329, 302, 339 ], "score": 0.87, "content": "W _ { 1 }", "type": "inline_equation" }, { "bbox": [ 302, 326, 350, 343 ], "score": 1.0, "content": "(similar for", "type": "text" }, { "bbox": [ 350, 329, 378, 340 ], "score": 0.91, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation" }, { "bbox": [ 378, 326, 389, 343 ], "score": 1.0, "content": "of", "type": "text" }, { "bbox": [ 389, 329, 406, 340 ], "score": 0.86, "content": "W _ { 2 }", "type": "inline_equation" }, { "bbox": [ 406, 326, 506, 343 ], "score": 1.0, "content": ") would have dimensions", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 339, 505, 352 ], "spans": [ { "bbox": [ 106, 339, 118, 352 ], "score": 1.0, "content": "of", "type": "text" }, { "bbox": [ 118, 340, 149, 350 ], "score": 0.91, "content": "r \\times d _ { m }", "type": "inline_equation" }, { "bbox": [ 149, 339, 181, 352 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 181, 339, 221, 350 ], "score": 0.92, "content": "d _ { m } = 4 d", "type": "inline_equation" }, { "bbox": [ 221, 339, 285, 352 ], "score": 1.0, "content": "as described in", "type": "text" }, { "bbox": [ 286, 340, 304, 350 ], "score": 0.83, "content": "\\ S 2 . 1", "type": "inline_equation" }, { "bbox": [ 304, 339, 434, 352 ], "score": 1.0, "content": ". Thus we typically use smaller", "type": "text" }, { "bbox": [ 434, 341, 440, 349 ], "score": 0.72, "content": "r", "type": "inline_equation" }, { "bbox": [ 441, 339, 505, 352 ], "score": 1.0, "content": "for LoRA (ffn)", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 350, 417, 363 ], "spans": [ { "bbox": [ 106, 350, 417, 363 ], "score": 1.0, "content": "than other methods to match their overall parameter size in later experiments.", "type": "text" } ], "index": 16 } ], "index": 12, "bbox_fs": [ 104, 261, 507, 363 ] }, { "type": "text", "bbox": [ 106, 367, 505, 477 ], "lines": [ { "bbox": [ 106, 367, 504, 379 ], "spans": [ { "bbox": [ 106, 367, 504, 379 ], "score": 1.0, "content": "Results: As shown in Figure 5, any method with FFN modification outperforms all the methods", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 379, 504, 390 ], "spans": [ { "bbox": [ 106, 379, 504, 390 ], "score": 1.0, "content": "with attention modification in all cases (the red markers are generally above all the blue ones, the", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 389, 505, 402 ], "spans": [ { "bbox": [ 106, 389, 225, 402 ], "score": 1.0, "content": "only exception is ffn-PA with", "type": "text" }, { "bbox": [ 225, 389, 247, 400 ], "score": 0.86, "content": "2 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 247, 389, 505, 402 ], "score": 1.0, "content": "params), often with fewer parameters. Second, the same method", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 399, 505, 414 ], "spans": [ { "bbox": [ 105, 399, 505, 414 ], "score": 1.0, "content": "applied at FFN always improves over its attention counterpart. For example, LoRA (ffn) improves", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 410, 505, 425 ], "spans": [ { "bbox": [ 105, 410, 505, 425 ], "score": 1.0, "content": "LoRA (attn) by 1 R-2 points on XSum. We also highlight that prefix tuning does not keep improving", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 422, 505, 435 ], "spans": [ { "bbox": [ 106, 422, 505, 435 ], "score": 1.0, "content": "when we further increase the capacity, which is also observed in Li & Liang (2021). These results", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 433, 506, 447 ], "spans": [ { "bbox": [ 105, 433, 506, 447 ], "score": 1.0, "content": "suggest that FFN modification can utilize the added parameters more effectively than attention, no", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 444, 505, 457 ], "spans": [ { "bbox": [ 105, 444, 505, 457 ], "score": 1.0, "content": "matter what the functional form or composition function is. We hypothesize that this is because", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 455, 505, 468 ], "spans": [ { "bbox": [ 106, 455, 505, 468 ], "score": 1.0, "content": "the FFN learns task-specific textual patterns (Geva et al., 2021), while attention learns pairwise", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 466, 445, 479 ], "spans": [ { "bbox": [ 105, 466, 445, 479 ], "score": 1.0, "content": "positional interactions which do not require large capacity for adapting to new tasks.", "type": "text" } ], "index": 26 } ], "index": 21.5, "bbox_fs": [ 105, 367, 506, 479 ] }, { "type": "text", "bbox": [ 107, 483, 505, 614 ], "lines": [ { "bbox": [ 104, 482, 506, 496 ], "spans": [ { "bbox": [ 104, 482, 257, 496 ], "score": 1.0, "content": "Is the story different when we use", "type": "text" }, { "bbox": [ 258, 482, 280, 494 ], "score": 0.87, "content": "0 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 280, 482, 357, 496 ], "score": 1.0, "content": "parameters? In", "type": "text" }, { "bbox": [ 357, 483, 376, 494 ], "score": 0.82, "content": "\\ S 3 . 1", "type": "inline_equation" }, { "bbox": [ 376, 482, 506, 496 ], "score": 1.0, "content": "we reason that prefix tuning is", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 495, 505, 506 ], "spans": [ { "bbox": [ 105, 495, 505, 506 ], "score": 1.0, "content": "more expressive than adapters (attn), which, however, is not reflected in Figure 5. We conjecture", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 505, 505, 517 ], "spans": [ { "bbox": [ 105, 505, 505, 517 ], "score": 1.0, "content": "that this is because multi-head attention is only superior when the parameter budget is small. To", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 515, 506, 529 ], "spans": [ { "bbox": [ 105, 515, 455, 529 ], "score": 1.0, "content": "validate this hypothesis, we compare prefix tuning to parallel adapters when they add", "type": "text" }, { "bbox": [ 455, 516, 478, 527 ], "score": 0.88, "content": "0 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 478, 515, 506, 529 ], "score": 1.0, "content": "of the", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 527, 505, 539 ], "spans": [ { "bbox": [ 105, 527, 505, 539 ], "score": 1.0, "content": "pretrained parameters. To ablate the impact of the composition function, we also report the results", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 538, 505, 550 ], "spans": [ { "bbox": [ 105, 538, 273, 550 ], "score": 1.0, "content": "of removing the gating in prefix tuning as", "type": "text" }, { "bbox": [ 273, 538, 308, 549 ], "score": 0.91, "content": "h + \\Delta h", "type": "inline_equation" }, { "bbox": [ 308, 538, 505, 550 ], "score": 1.0, "content": ". We include the results of the multi-head parallel", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 549, 505, 561 ], "spans": [ { "bbox": [ 105, 549, 261, 561 ], "score": 1.0, "content": "adapter variant (MH PA) described in", "type": "text" }, { "bbox": [ 261, 549, 280, 560 ], "score": 0.87, "content": "\\ S 3 . 3", "type": "inline_equation" }, { "bbox": [ 280, 549, 505, 561 ], "score": 1.0, "content": ". As shown in Table 4, the multi-head methods – prefix", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 560, 506, 572 ], "spans": [ { "bbox": [ 105, 560, 470, 572 ], "score": 1.0, "content": "tuning and MH PA (attn) – outperform all others by at least 1.6 BLEU points when using", "type": "text" }, { "bbox": [ 470, 560, 493, 571 ], "score": 0.87, "content": "0 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 493, 560, 506, 572 ], "score": 1.0, "content": "of", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 569, 506, 585 ], "spans": [ { "bbox": [ 105, 569, 261, 585 ], "score": 1.0, "content": "the parameters. Surprisingly, reducing", "type": "text" }, { "bbox": [ 261, 571, 266, 581 ], "score": 0.32, "content": "l", "type": "inline_equation" }, { "bbox": [ 266, 569, 506, 585 ], "score": 1.0, "content": "from 200 to 30 only causes 0.4 BLEU loss for prefix tuning", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 582, 505, 594 ], "spans": [ { "bbox": [ 106, 582, 505, 594 ], "score": 1.0, "content": "while PA (attn) loses 1.9 points. The gating composition function in prefix tuning slightly helps the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 593, 505, 605 ], "spans": [ { "bbox": [ 105, 593, 505, 605 ], "score": 1.0, "content": "results by 0.3 points. We highlight that the MH parallel adapter improves the single-headed version", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 604, 438, 615 ], "spans": [ { "bbox": [ 105, 604, 438, 615 ], "score": 1.0, "content": "by 1.6 points, which again verifies the effectiveness of the multi-head formulation.", "type": "text" } ], "index": 38 } ], "index": 32.5, "bbox_fs": [ 104, 482, 506, 615 ] }, { "type": "text", "bbox": [ 107, 620, 505, 665 ], "lines": [ { "bbox": [ 105, 620, 505, 632 ], "spans": [ { "bbox": [ 105, 620, 505, 632 ], "score": 1.0, "content": "Combining the results in Figure 5 and Table 4, we conclude that modifying head attention shows the", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 632, 505, 644 ], "spans": [ { "bbox": [ 105, 632, 505, 644 ], "score": 1.0, "content": "best results when the parameter budget is very small, while the FFN can better utilize modifications", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 643, 505, 655 ], "spans": [ { "bbox": [ 106, 643, 505, 655 ], "score": 1.0, "content": "at larger capacities. This suggests that it may be effective to allocate a larger parameter budget to", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 653, 471, 667 ], "spans": [ { "bbox": [ 105, 653, 471, 667 ], "score": 1.0, "content": "FFN modification instead of treating attention and FFN equally as in Houlsby et al. (2019).", "type": "text" } ], "index": 42 } ], "index": 40.5, "bbox_fs": [ 105, 620, 505, 667 ] }, { "type": "title", "bbox": [ 108, 678, 279, 689 ], "lines": [ { "bbox": [ 106, 678, 280, 691 ], "spans": [ { "bbox": [ 106, 678, 280, 691 ], "score": 1.0, "content": "4.5 WHICH COMPOSITION FUNCTION?", "type": "text" } ], "index": 43 } ], "index": 43 }, { "type": "text", "bbox": [ 108, 699, 504, 731 ], "lines": [ { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 312, 711 ], "score": 1.0, "content": "We have presented three composition functions in", "type": "text" }, { "bbox": [ 313, 699, 331, 710 ], "score": 0.86, "content": "\\ S 3 . 2", "type": "inline_equation" }, { "bbox": [ 332, 699, 505, 711 ], "score": 1.0, "content": ": simple addition (adapter), gated addition", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 710, 505, 721 ], "spans": [ { "bbox": [ 106, 710, 505, 721 ], "score": 1.0, "content": "(prefix tuning) and scaled addition (LoRA). As it is unnatural to incorporate the exact gated ad-", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 720, 504, 732 ], "spans": [ { "bbox": [ 106, 720, 504, 732 ], "score": 1.0, "content": "dition into methods whose functional form does not use softmax, we examine the other two by", "type": "text" } ], "index": 46 }, { "bbox": [ 106, 284, 503, 296 ], "spans": [ { "bbox": [ 106, 284, 503, 296 ], "score": 1.0, "content": "ablating on LoRA and comparing with the proposed scaled parallel adapter (Scaled PA), we con-", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 106, 294, 501, 307 ], "spans": [ { "bbox": [ 106, 294, 482, 307 ], "score": 1.0, "content": "strain modified representation to be FFN since it is generally more effective as shown in", "type": "text", "cross_page": true }, { "bbox": [ 482, 295, 501, 306 ], "score": 0.86, "content": "\\ S 4 . 4", "type": "inline_equation", "cross_page": true } ], "index": 9 } ], "index": 45, "bbox_fs": [ 106, 699, 505, 732 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 116, 135, 495, 277 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 80, 505, 131 ], "group_id": 0, "lines": [ { "bbox": [ 106, 79, 505, 92 ], "spans": [ { "bbox": [ 106, 79, 505, 92 ], "score": 1.0, "content": "Table 6: Comparison of various parameter-efficient tuning methods and the proposed variants. “†” are results", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 90, 505, 102 ], "spans": [ { "bbox": [ 105, 90, 505, 102 ], "score": 1.0, "content": "copied from Lewis et al. (2020) and Liu et al. (2020b). We could not reproduce exactly the same full fine-", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 100, 505, 112 ], "spans": [ { "bbox": [ 106, 100, 505, 112 ], "score": 1.0, "content": "tuning numbers with the same hyperparameters or even searching them. The reason may be the different", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 109, 505, 122 ], "spans": [ { "bbox": [ 106, 109, 505, 122 ], "score": 1.0, "content": "libraries which the training code is based on – full fine-tuning is very sensitive to training hyperparameters. For", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 120, 463, 131 ], "spans": [ { "bbox": [ 106, 120, 463, 131 ], "score": 1.0, "content": "the most performant methods we run with 3 random seeds and report mean and standard deviation.", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "table_body", "bbox": [ 116, 135, 495, 277 ], "group_id": 0, "lines": [ { "bbox": [ 116, 135, 495, 277 ], "spans": [ { "bbox": [ 116, 135, 495, 277 ], "score": 0.984, "html": "
Method# paramsXSum (R-1/2/L)MT (BLEU)
Full fine-tuning+100%45.14/22.27/37.2537.7
Full fine-tuning (our run)100%44.81/21.94/36.8337.3
Bitfit (Ben Zaken et al., 2021)0.1%40.64/17.32/32.1926.4
Prompt tuning (Lester et al., 2021)0.1%38.91/15.98/30.8321.0
Prefix tuning (Li& Liang,2021),l=2003.6%43.40/20.46/35.5135.6
Pfeiffer adapter (Pfeiffer et al.,2021),r=6007.2%44.03/20.89/35.89±.13/.10/.0836.9±.1
LoRA (ffn),r=1027.2%44.53/21.29/36.28±.14/.07/.1036.8±.3
Parallel adapter (PA,ffn),r=102412.3%44.71/21.41/36.41±.16/.17/.1637.2±.1
PA (attn,r=30) + PA (ffn,r=512)6.7%44.29/21.06/36.12±.31/.19/.1837.2±.1
Prefix tuning (attn,l=3O) + LoRA (ffn,r=102)6.7%44.84/21.71/36.77±.07/.05/.0337.0±.1
MAM Adapter (our variant, l=30,r=512)6.7%45.06/21.90/36.87±.08/01/.0437.5±.1
", "type": "table", "image_path": "f586b789e71f4c2f99ed5106b7d47d4f035cf46bc1368a7ee339eb982a339215.jpg" } ] } ], "index": 6, "virtual_lines": [ { "bbox": [ 116, 135, 495, 182.33333333333334 ], "spans": [], "index": 5 }, { "bbox": [ 116, 182.33333333333334, 495, 229.66666666666669 ], "spans": [], "index": 6 }, { "bbox": [ 116, 229.66666666666669, 495, 277.0 ], "spans": [], "index": 7 } ] } ], "index": 4.0 }, { "type": "text", "bbox": [ 108, 284, 501, 306 ], "lines": [ { "bbox": [ 106, 284, 503, 296 ], "spans": [ { "bbox": [ 106, 284, 503, 296 ], "score": 1.0, "content": "ablating on LoRA and comparing with the proposed scaled parallel adapter (Scaled PA), we con-", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 294, 501, 307 ], "spans": [ { "bbox": [ 106, 294, 482, 307 ], "score": 1.0, "content": "strain modified representation to be FFN since it is generally more effective as shown in", "type": "text" }, { "bbox": [ 482, 295, 501, 306 ], "score": 0.86, "content": "\\ S 4 . 4", "type": "inline_equation" } ], "index": 9 } ], "index": 8.5 }, { "type": "text", "bbox": [ 107, 306, 316, 438 ], "lines": [ { "bbox": [ 106, 305, 317, 317 ], "spans": [ { "bbox": [ 106, 305, 281, 317 ], "score": 1.0, "content": "Table 5 reports the results on XSum. We set", "type": "text" }, { "bbox": [ 282, 308, 288, 316 ], "score": 0.64, "content": "r", "type": "inline_equation" }, { "bbox": [ 288, 305, 317, 317 ], "score": 1.0, "content": "as 512", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 316, 317, 328 ], "spans": [ { "bbox": [ 106, 316, 317, 328 ], "score": 1.0, "content": "for adapters and 102 for LoRA so that their tuned", "type": "text" } ], "index": 11 }, { "bbox": [ 104, 328, 317, 339 ], "spans": [ { "bbox": [ 104, 328, 282, 339 ], "score": 1.0, "content": "parameter sizes are the same. We select", "type": "text" }, { "bbox": [ 282, 330, 289, 338 ], "score": 0.57, "content": "s", "type": "inline_equation" }, { "bbox": [ 289, 328, 317, 339 ], "score": 1.0, "content": "based", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 338, 317, 350 ], "spans": [ { "bbox": [ 105, 338, 317, 350 ], "score": 1.0, "content": "on the R-2 score on the dev set. We observe that", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 349, 317, 361 ], "spans": [ { "bbox": [ 105, 349, 137, 361 ], "score": 1.0, "content": "LoRA", "type": "text" }, { "bbox": [ 137, 350, 163, 360 ], "score": 0.84, "content": "s = 4", "type": "inline_equation" }, { "bbox": [ 163, 349, 317, 361 ], "score": 1.0, "content": ") performs better than parallel adapter.", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 359, 317, 373 ], "spans": [ { "bbox": [ 105, 359, 317, 373 ], "score": 1.0, "content": "However, the advantage disappears if we remove the", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 370, 317, 384 ], "spans": [ { "bbox": [ 105, 370, 183, 384 ], "score": 1.0, "content": "scaling by setting", "type": "text" }, { "bbox": [ 184, 372, 214, 381 ], "score": 0.9, "content": "s ~ = ~ 1", "type": "inline_equation" }, { "bbox": [ 215, 370, 317, 384 ], "score": 1.0, "content": ". Through plugging the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 381, 317, 395 ], "spans": [ { "bbox": [ 105, 381, 317, 395 ], "score": 1.0, "content": "composition function of LoRA into parallel adapter,", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 393, 317, 405 ], "spans": [ { "bbox": [ 106, 393, 317, 405 ], "score": 1.0, "content": "the resulted Scaled PA improves the vanilla parallel", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 404, 317, 417 ], "spans": [ { "bbox": [ 106, 404, 317, 417 ], "score": 1.0, "content": "adapter by 0.56 ROUGE-2 points. We also experi-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 415, 317, 427 ], "spans": [ { "bbox": [ 105, 415, 317, 427 ], "score": 1.0, "content": "ment with a learned scalar which does not give bet-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 425, 317, 439 ], "spans": [ { "bbox": [ 105, 425, 317, 439 ], "score": 1.0, "content": "ter results. Therefore, we conclude that the scaling", "type": "text" } ], "index": 21 } ], "index": 15.5 }, { "type": "table", "bbox": [ 327, 358, 505, 429 ], "blocks": [ { "type": "table_caption", "bbox": [ 324, 314, 504, 354 ], "group_id": 1, "lines": [ { "bbox": [ 324, 313, 505, 325 ], "spans": [ { "bbox": [ 324, 313, 505, 325 ], "score": 1.0, "content": "Table 5: Results on XSum when using different", "type": "text" } ], "index": 22 }, { "bbox": [ 324, 324, 505, 335 ], "spans": [ { "bbox": [ 324, 324, 505, 335 ], "score": 1.0, "content": "composition functions. The modified representa-", "type": "text" } ], "index": 23 }, { "bbox": [ 324, 334, 504, 344 ], "spans": [ { "bbox": [ 324, 334, 469, 344 ], "score": 1.0, "content": "tion is FFN. The bottleneck dimension", "type": "text" }, { "bbox": [ 469, 334, 504, 343 ], "score": 0.88, "content": "\\bar { r } = 5 1 2", "type": "inline_equation" } ], "index": 24 }, { "bbox": [ 323, 344, 470, 354 ], "spans": [ { "bbox": [ 323, 344, 397, 354 ], "score": 1.0, "content": "for (Scaled) PA and", "type": "text" }, { "bbox": [ 397, 344, 429, 353 ], "score": 0.88, "content": "r = 1 0 2", "type": "inline_equation" }, { "bbox": [ 429, 344, 470, 354 ], "score": 1.0, "content": "for LoRA.", "type": "text" } ], "index": 25 } ], "index": 23.5 }, { "type": "table_body", "bbox": [ 327, 358, 505, 429 ], "group_id": 1, "lines": [ { "bbox": [ 327, 358, 505, 429 ], "spans": [ { "bbox": [ 327, 358, 505, 429 ], "score": 0.971, "html": "
Method (# params)XSum (R-1/2/LSum)
LoRA (6.1%), s=444.59/21.31/36.25
LoRA (6.1%), s=144.17/20.83/35.74
PA (6.1%)44.35/20.98/35.98
Scaled PA (6.1%), s=444.85/21.54/36.58
Scaled PA(6.1%),trainable s44.56/21.31/36.29
", "type": "table", "image_path": "5e6da241c881465525b4a45c36e16ee3be4d24fb4a3a6f5ec4c1fc5c60f42e1c.jpg" } ] } ], "index": 28, "virtual_lines": [ { "bbox": [ 327, 358, 505, 372.2 ], "spans": [], "index": 26 }, { "bbox": [ 327, 372.2, 505, 386.4 ], "spans": [], "index": 27 }, { "bbox": [ 327, 386.4, 505, 400.59999999999997 ], "spans": [], "index": 28 }, { "bbox": [ 327, 400.59999999999997, 505, 414.79999999999995 ], "spans": [], "index": 29 }, { "bbox": [ 327, 414.79999999999995, 505, 428.99999999999994 ], "spans": [], "index": 30 } ] } ], "index": 25.75 }, { "type": "text", "bbox": [ 107, 438, 465, 448 ], "lines": [ { "bbox": [ 105, 436, 465, 451 ], "spans": [ { "bbox": [ 105, 436, 465, 451 ], "score": 1.0, "content": "composition function is better than the vanilla additive one while being easily applicable.", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "title", "bbox": [ 106, 461, 480, 473 ], "lines": [ { "bbox": [ 106, 461, 482, 473 ], "spans": [ { "bbox": [ 106, 461, 482, 473 ], "score": 1.0, "content": "4.6 AN EFFECTIVE INTEGRATION BY TRANSFERRING FAVORABLE DESIGN ELEMENTS", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 107, 482, 505, 647 ], "lines": [ { "bbox": [ 106, 481, 506, 495 ], "spans": [ { "bbox": [ 106, 481, 506, 495 ], "score": 1.0, "content": "We first highlight three findings in previous sections: (1) Scaled parallel adapter is the best variant", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 493, 505, 506 ], "spans": [ { "bbox": [ 105, 493, 505, 506 ], "score": 1.0, "content": "to modify FFN; (2) FFN can better utilize modification at larger capacities; and (3) modifying head", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 504, 505, 516 ], "spans": [ { "bbox": [ 105, 504, 394, 516 ], "score": 1.0, "content": "attentions like prefix tuning can achieve strong performance with only", "type": "text" }, { "bbox": [ 394, 504, 417, 515 ], "score": 0.85, "content": "0 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 417, 504, 505, 516 ], "score": 1.0, "content": "parameters. Inspired", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 515, 505, 528 ], "spans": [ { "bbox": [ 106, 515, 505, 528 ], "score": 1.0, "content": "by them, we mix and match the favorable designs behind these findings: specifically, we use prefix", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 525, 506, 539 ], "spans": [ { "bbox": [ 105, 525, 282, 539 ], "score": 1.0, "content": "tuning with a small bottleneck dimension", "type": "text" }, { "bbox": [ 282, 526, 315, 537 ], "score": 0.87, "content": "\\mathit { l } \\ : = \\ : 3 0 ", "type": "inline_equation" }, { "bbox": [ 316, 525, 506, 539 ], "score": 1.0, "content": ") at the attention sub-layers and allocate more", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 537, 505, 550 ], "spans": [ { "bbox": [ 105, 537, 437, 550 ], "score": 1.0, "content": "parameter budgets to modify FFN representation using the scaled parallel adapter", "type": "text" }, { "bbox": [ 437, 537, 473, 548 ], "score": 0.84, "content": "( r = 5 1 2", "type": "inline_equation" }, { "bbox": [ 473, 537, 505, 550 ], "score": 1.0, "content": "). Since", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 548, 505, 560 ], "spans": [ { "bbox": [ 105, 548, 505, 560 ], "score": 1.0, "content": "prefix tuning can be viewed as a form of adapter in our unified framework, we name this variant", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 559, 505, 571 ], "spans": [ { "bbox": [ 105, 559, 505, 571 ], "score": 1.0, "content": "as Mix-And-Match adapter (MAM Adapter). In Table 6, we compare MAM adapter with various", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 569, 505, 583 ], "spans": [ { "bbox": [ 105, 569, 505, 583 ], "score": 1.0, "content": "parameter-efficient tuning methods. For completeness, we also present results of other combination", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 579, 505, 594 ], "spans": [ { "bbox": [ 105, 579, 505, 594 ], "score": 1.0, "content": "versions in Table 6: using parallel adapters at both attention and FFN layers and combining prefix", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 591, 505, 604 ], "spans": [ { "bbox": [ 105, 591, 505, 604 ], "score": 1.0, "content": "tuning (attn) with LoRA (ffn) – both of these combined versions can improve over their respective", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 602, 506, 615 ], "spans": [ { "bbox": [ 105, 602, 506, 615 ], "score": 1.0, "content": "prototypes. However, MAM Adapter achieves the best performance on both tasks and is able to", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 613, 505, 626 ], "spans": [ { "bbox": [ 105, 613, 344, 626 ], "score": 1.0, "content": "match the results of our full fine-tuning by only updating", "type": "text" }, { "bbox": [ 344, 613, 367, 624 ], "score": 0.87, "content": "6 . 7 \\%", "type": "inline_equation" }, { "bbox": [ 367, 613, 505, 626 ], "score": 1.0, "content": "of the pre-trained parameters. In", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 624, 506, 637 ], "spans": [ { "bbox": [ 106, 624, 506, 637 ], "score": 1.0, "content": "Table 2, we present the results of MAM Adapter on MNLI and SST2 as well, where MAM Adapter", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 636, 481, 648 ], "spans": [ { "bbox": [ 105, 636, 355, 648 ], "score": 1.0, "content": "achieves comparable results to full fine-tuning by adding only", "type": "text" }, { "bbox": [ 355, 636, 378, 646 ], "score": 0.85, "content": "0 . 5 \\%", "type": "inline_equation" }, { "bbox": [ 378, 636, 481, 648 ], "score": 1.0, "content": "of pretrained parameters.", "type": "text" } ], "index": 47 } ], "index": 40 }, { "type": "title", "bbox": [ 108, 662, 190, 675 ], "lines": [ { "bbox": [ 105, 662, 192, 677 ], "spans": [ { "bbox": [ 105, 662, 192, 677 ], "score": 1.0, "content": "5 DISCUSSION", "type": "text" } ], "index": 48 } ], "index": 48 }, { "type": "text", "bbox": [ 108, 688, 504, 732 ], "lines": [ { "bbox": [ 106, 688, 505, 699 ], "spans": [ { "bbox": [ 106, 688, 505, 699 ], "score": 1.0, "content": "We provide a unified framework for several performant parameter-tuning methods, which enables", "type": "text" } ], "index": 49 }, { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 505, 711 ], "score": 1.0, "content": "us to instantiate a more effective model that matches the performance of full fine-tuning method", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 709, 506, 722 ], "spans": [ { "bbox": [ 105, 709, 506, 722 ], "score": 1.0, "content": "through transferring techniques across approaches. We hope our work can provide insights and", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 720, 342, 734 ], "spans": [ { "bbox": [ 105, 720, 342, 734 ], "score": 1.0, "content": "guidance for future research on parameter-efficient tuning.", "type": "text" } ], "index": 52 } ], "index": 50.5 } ], "page_idx": 8, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 759 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "9", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 116, 135, 495, 277 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 80, 505, 131 ], "group_id": 0, "lines": [ { "bbox": [ 106, 79, 505, 92 ], "spans": [ { "bbox": [ 106, 79, 505, 92 ], "score": 1.0, "content": "Table 6: Comparison of various parameter-efficient tuning methods and the proposed variants. “†” are results", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 90, 505, 102 ], "spans": [ { "bbox": [ 105, 90, 505, 102 ], "score": 1.0, "content": "copied from Lewis et al. (2020) and Liu et al. (2020b). We could not reproduce exactly the same full fine-", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 100, 505, 112 ], "spans": [ { "bbox": [ 106, 100, 505, 112 ], "score": 1.0, "content": "tuning numbers with the same hyperparameters or even searching them. The reason may be the different", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 109, 505, 122 ], "spans": [ { "bbox": [ 106, 109, 505, 122 ], "score": 1.0, "content": "libraries which the training code is based on – full fine-tuning is very sensitive to training hyperparameters. For", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 120, 463, 131 ], "spans": [ { "bbox": [ 106, 120, 463, 131 ], "score": 1.0, "content": "the most performant methods we run with 3 random seeds and report mean and standard deviation.", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "table_body", "bbox": [ 116, 135, 495, 277 ], "group_id": 0, "lines": [ { "bbox": [ 116, 135, 495, 277 ], "spans": [ { "bbox": [ 116, 135, 495, 277 ], "score": 0.984, "html": "
Method# paramsXSum (R-1/2/L)MT (BLEU)
Full fine-tuning+100%45.14/22.27/37.2537.7
Full fine-tuning (our run)100%44.81/21.94/36.8337.3
Bitfit (Ben Zaken et al., 2021)0.1%40.64/17.32/32.1926.4
Prompt tuning (Lester et al., 2021)0.1%38.91/15.98/30.8321.0
Prefix tuning (Li& Liang,2021),l=2003.6%43.40/20.46/35.5135.6
Pfeiffer adapter (Pfeiffer et al.,2021),r=6007.2%44.03/20.89/35.89±.13/.10/.0836.9±.1
LoRA (ffn),r=1027.2%44.53/21.29/36.28±.14/.07/.1036.8±.3
Parallel adapter (PA,ffn),r=102412.3%44.71/21.41/36.41±.16/.17/.1637.2±.1
PA (attn,r=30) + PA (ffn,r=512)6.7%44.29/21.06/36.12±.31/.19/.1837.2±.1
Prefix tuning (attn,l=3O) + LoRA (ffn,r=102)6.7%44.84/21.71/36.77±.07/.05/.0337.0±.1
MAM Adapter (our variant, l=30,r=512)6.7%45.06/21.90/36.87±.08/01/.0437.5±.1
", "type": "table", "image_path": "f586b789e71f4c2f99ed5106b7d47d4f035cf46bc1368a7ee339eb982a339215.jpg" } ] } ], "index": 6, "virtual_lines": [ { "bbox": [ 116, 135, 495, 182.33333333333334 ], "spans": [], "index": 5 }, { "bbox": [ 116, 182.33333333333334, 495, 229.66666666666669 ], "spans": [], "index": 6 }, { "bbox": [ 116, 229.66666666666669, 495, 277.0 ], "spans": [], "index": 7 } ] } ], "index": 4.0 }, { "type": "text", "bbox": [ 108, 284, 501, 306 ], "lines": [], "index": 8.5, "bbox_fs": [ 106, 284, 503, 307 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 306, 316, 438 ], "lines": [ { "bbox": [ 106, 305, 317, 317 ], "spans": [ { "bbox": [ 106, 305, 281, 317 ], "score": 1.0, "content": "Table 5 reports the results on XSum. We set", "type": "text" }, { "bbox": [ 282, 308, 288, 316 ], "score": 0.64, "content": "r", "type": "inline_equation" }, { "bbox": [ 288, 305, 317, 317 ], "score": 1.0, "content": "as 512", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 316, 317, 328 ], "spans": [ { "bbox": [ 106, 316, 317, 328 ], "score": 1.0, "content": "for adapters and 102 for LoRA so that their tuned", "type": "text" } ], "index": 11 }, { "bbox": [ 104, 328, 317, 339 ], "spans": [ { "bbox": [ 104, 328, 282, 339 ], "score": 1.0, "content": "parameter sizes are the same. We select", "type": "text" }, { "bbox": [ 282, 330, 289, 338 ], "score": 0.57, "content": "s", "type": "inline_equation" }, { "bbox": [ 289, 328, 317, 339 ], "score": 1.0, "content": "based", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 338, 317, 350 ], "spans": [ { "bbox": [ 105, 338, 317, 350 ], "score": 1.0, "content": "on the R-2 score on the dev set. We observe that", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 349, 317, 361 ], "spans": [ { "bbox": [ 105, 349, 137, 361 ], "score": 1.0, "content": "LoRA", "type": "text" }, { "bbox": [ 137, 350, 163, 360 ], "score": 0.84, "content": "s = 4", "type": "inline_equation" }, { "bbox": [ 163, 349, 317, 361 ], "score": 1.0, "content": ") performs better than parallel adapter.", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 359, 317, 373 ], "spans": [ { "bbox": [ 105, 359, 317, 373 ], "score": 1.0, "content": "However, the advantage disappears if we remove the", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 370, 317, 384 ], "spans": [ { "bbox": [ 105, 370, 183, 384 ], "score": 1.0, "content": "scaling by setting", "type": "text" }, { "bbox": [ 184, 372, 214, 381 ], "score": 0.9, "content": "s ~ = ~ 1", "type": "inline_equation" }, { "bbox": [ 215, 370, 317, 384 ], "score": 1.0, "content": ". Through plugging the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 381, 317, 395 ], "spans": [ { "bbox": [ 105, 381, 317, 395 ], "score": 1.0, "content": "composition function of LoRA into parallel adapter,", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 393, 317, 405 ], "spans": [ { "bbox": [ 106, 393, 317, 405 ], "score": 1.0, "content": "the resulted Scaled PA improves the vanilla parallel", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 404, 317, 417 ], "spans": [ { "bbox": [ 106, 404, 317, 417 ], "score": 1.0, "content": "adapter by 0.56 ROUGE-2 points. We also experi-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 415, 317, 427 ], "spans": [ { "bbox": [ 105, 415, 317, 427 ], "score": 1.0, "content": "ment with a learned scalar which does not give bet-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 425, 317, 439 ], "spans": [ { "bbox": [ 105, 425, 317, 439 ], "score": 1.0, "content": "ter results. Therefore, we conclude that the scaling", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 436, 465, 451 ], "spans": [ { "bbox": [ 105, 436, 465, 451 ], "score": 1.0, "content": "composition function is better than the vanilla additive one while being easily applicable.", "type": "text" } ], "index": 31 } ], "index": 15.5, "bbox_fs": [ 104, 305, 317, 439 ] }, { "type": "table", "bbox": [ 327, 358, 505, 429 ], "blocks": [ { "type": "table_caption", "bbox": [ 324, 314, 504, 354 ], "group_id": 1, "lines": [ { "bbox": [ 324, 313, 505, 325 ], "spans": [ { "bbox": [ 324, 313, 505, 325 ], "score": 1.0, "content": "Table 5: Results on XSum when using different", "type": "text" } ], "index": 22 }, { "bbox": [ 324, 324, 505, 335 ], "spans": [ { "bbox": [ 324, 324, 505, 335 ], "score": 1.0, "content": "composition functions. The modified representa-", "type": "text" } ], "index": 23 }, { "bbox": [ 324, 334, 504, 344 ], "spans": [ { "bbox": [ 324, 334, 469, 344 ], "score": 1.0, "content": "tion is FFN. The bottleneck dimension", "type": "text" }, { "bbox": [ 469, 334, 504, 343 ], "score": 0.88, "content": "\\bar { r } = 5 1 2", "type": "inline_equation" } ], "index": 24 }, { "bbox": [ 323, 344, 470, 354 ], "spans": [ { "bbox": [ 323, 344, 397, 354 ], "score": 1.0, "content": "for (Scaled) PA and", "type": "text" }, { "bbox": [ 397, 344, 429, 353 ], "score": 0.88, "content": "r = 1 0 2", "type": "inline_equation" }, { "bbox": [ 429, 344, 470, 354 ], "score": 1.0, "content": "for LoRA.", "type": "text" } ], "index": 25 } ], "index": 23.5 }, { "type": "table_body", "bbox": [ 327, 358, 505, 429 ], "group_id": 1, "lines": [ { "bbox": [ 327, 358, 505, 429 ], "spans": [ { "bbox": [ 327, 358, 505, 429 ], "score": 0.971, "html": "
Method (# params)XSum (R-1/2/LSum)
LoRA (6.1%), s=444.59/21.31/36.25
LoRA (6.1%), s=144.17/20.83/35.74
PA (6.1%)44.35/20.98/35.98
Scaled PA (6.1%), s=444.85/21.54/36.58
Scaled PA(6.1%),trainable s44.56/21.31/36.29
", "type": "table", "image_path": "5e6da241c881465525b4a45c36e16ee3be4d24fb4a3a6f5ec4c1fc5c60f42e1c.jpg" } ] } ], "index": 28, "virtual_lines": [ { "bbox": [ 327, 358, 505, 372.2 ], "spans": [], "index": 26 }, { "bbox": [ 327, 372.2, 505, 386.4 ], "spans": [], "index": 27 }, { "bbox": [ 327, 386.4, 505, 400.59999999999997 ], "spans": [], "index": 28 }, { "bbox": [ 327, 400.59999999999997, 505, 414.79999999999995 ], "spans": [], "index": 29 }, { "bbox": [ 327, 414.79999999999995, 505, 428.99999999999994 ], "spans": [], "index": 30 } ] } ], "index": 25.75 }, { "type": "text", "bbox": [ 107, 438, 465, 448 ], "lines": [], "index": 31, "bbox_fs": [ 105, 436, 465, 451 ], "lines_deleted": true }, { "type": "title", "bbox": [ 106, 461, 480, 473 ], "lines": [ { "bbox": [ 106, 461, 482, 473 ], "spans": [ { "bbox": [ 106, 461, 482, 473 ], "score": 1.0, "content": "4.6 AN EFFECTIVE INTEGRATION BY TRANSFERRING FAVORABLE DESIGN ELEMENTS", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 107, 482, 505, 647 ], "lines": [ { "bbox": [ 106, 481, 506, 495 ], "spans": [ { "bbox": [ 106, 481, 506, 495 ], "score": 1.0, "content": "We first highlight three findings in previous sections: (1) Scaled parallel adapter is the best variant", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 493, 505, 506 ], "spans": [ { "bbox": [ 105, 493, 505, 506 ], "score": 1.0, "content": "to modify FFN; (2) FFN can better utilize modification at larger capacities; and (3) modifying head", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 504, 505, 516 ], "spans": [ { "bbox": [ 105, 504, 394, 516 ], "score": 1.0, "content": "attentions like prefix tuning can achieve strong performance with only", "type": "text" }, { "bbox": [ 394, 504, 417, 515 ], "score": 0.85, "content": "0 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 417, 504, 505, 516 ], "score": 1.0, "content": "parameters. Inspired", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 515, 505, 528 ], "spans": [ { "bbox": [ 106, 515, 505, 528 ], "score": 1.0, "content": "by them, we mix and match the favorable designs behind these findings: specifically, we use prefix", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 525, 506, 539 ], "spans": [ { "bbox": [ 105, 525, 282, 539 ], "score": 1.0, "content": "tuning with a small bottleneck dimension", "type": "text" }, { "bbox": [ 282, 526, 315, 537 ], "score": 0.87, "content": "\\mathit { l } \\ : = \\ : 3 0 ", "type": "inline_equation" }, { "bbox": [ 316, 525, 506, 539 ], "score": 1.0, "content": ") at the attention sub-layers and allocate more", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 537, 505, 550 ], "spans": [ { "bbox": [ 105, 537, 437, 550 ], "score": 1.0, "content": "parameter budgets to modify FFN representation using the scaled parallel adapter", "type": "text" }, { "bbox": [ 437, 537, 473, 548 ], "score": 0.84, "content": "( r = 5 1 2", "type": "inline_equation" }, { "bbox": [ 473, 537, 505, 550 ], "score": 1.0, "content": "). Since", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 548, 505, 560 ], "spans": [ { "bbox": [ 105, 548, 505, 560 ], "score": 1.0, "content": "prefix tuning can be viewed as a form of adapter in our unified framework, we name this variant", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 559, 505, 571 ], "spans": [ { "bbox": [ 105, 559, 505, 571 ], "score": 1.0, "content": "as Mix-And-Match adapter (MAM Adapter). In Table 6, we compare MAM adapter with various", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 569, 505, 583 ], "spans": [ { "bbox": [ 105, 569, 505, 583 ], "score": 1.0, "content": "parameter-efficient tuning methods. For completeness, we also present results of other combination", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 579, 505, 594 ], "spans": [ { "bbox": [ 105, 579, 505, 594 ], "score": 1.0, "content": "versions in Table 6: using parallel adapters at both attention and FFN layers and combining prefix", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 591, 505, 604 ], "spans": [ { "bbox": [ 105, 591, 505, 604 ], "score": 1.0, "content": "tuning (attn) with LoRA (ffn) – both of these combined versions can improve over their respective", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 602, 506, 615 ], "spans": [ { "bbox": [ 105, 602, 506, 615 ], "score": 1.0, "content": "prototypes. However, MAM Adapter achieves the best performance on both tasks and is able to", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 613, 505, 626 ], "spans": [ { "bbox": [ 105, 613, 344, 626 ], "score": 1.0, "content": "match the results of our full fine-tuning by only updating", "type": "text" }, { "bbox": [ 344, 613, 367, 624 ], "score": 0.87, "content": "6 . 7 \\%", "type": "inline_equation" }, { "bbox": [ 367, 613, 505, 626 ], "score": 1.0, "content": "of the pre-trained parameters. In", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 624, 506, 637 ], "spans": [ { "bbox": [ 106, 624, 506, 637 ], "score": 1.0, "content": "Table 2, we present the results of MAM Adapter on MNLI and SST2 as well, where MAM Adapter", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 636, 481, 648 ], "spans": [ { "bbox": [ 105, 636, 355, 648 ], "score": 1.0, "content": "achieves comparable results to full fine-tuning by adding only", "type": "text" }, { "bbox": [ 355, 636, 378, 646 ], "score": 0.85, "content": "0 . 5 \\%", "type": "inline_equation" }, { "bbox": [ 378, 636, 481, 648 ], "score": 1.0, "content": "of pretrained parameters.", "type": "text" } ], "index": 47 } ], "index": 40, "bbox_fs": [ 105, 481, 506, 648 ] }, { "type": "title", "bbox": [ 108, 662, 190, 675 ], "lines": [ { "bbox": [ 105, 662, 192, 677 ], "spans": [ { "bbox": [ 105, 662, 192, 677 ], "score": 1.0, "content": "5 DISCUSSION", "type": "text" } ], "index": 48 } ], "index": 48 }, { "type": "text", "bbox": [ 108, 688, 504, 732 ], "lines": [ { "bbox": [ 106, 688, 505, 699 ], "spans": [ { "bbox": [ 106, 688, 505, 699 ], "score": 1.0, "content": "We provide a unified framework for several performant parameter-tuning methods, which enables", "type": "text" } ], "index": 49 }, { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 505, 711 ], "score": 1.0, "content": "us to instantiate a more effective model that matches the performance of full fine-tuning method", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 709, 506, 722 ], "spans": [ { "bbox": [ 105, 709, 506, 722 ], "score": 1.0, "content": "through transferring techniques across approaches. We hope our work can provide insights and", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 720, 342, 734 ], "spans": [ { "bbox": [ 105, 720, 342, 734 ], "score": 1.0, "content": "guidance for future research on parameter-efficient tuning.", "type": "text" } ], "index": 52 } ], "index": 50.5, "bbox_fs": [ 105, 688, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 81, 210, 94 ], "lines": [ { "bbox": [ 106, 81, 213, 95 ], "spans": [ { "bbox": [ 106, 81, 213, 95 ], "score": 1.0, "content": "ETHICS STATEMENT", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 106, 505, 183 ], "lines": [ { "bbox": [ 105, 105, 505, 120 ], "spans": [ { "bbox": [ 105, 105, 505, 120 ], "score": 1.0, "content": "Our work proposes a method for efficient fine-tuning of pre-trained models, in particular language", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 117, 505, 130 ], "spans": [ { "bbox": [ 105, 117, 505, 130 ], "score": 1.0, "content": "models. Pre-trained language models have a wide variety of positive applications, such as the appli-", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 129, 504, 140 ], "spans": [ { "bbox": [ 106, 129, 504, 140 ], "score": 1.0, "content": "cations to summarization, translation, or language understanding described in our paper. At the same", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 140, 505, 151 ], "spans": [ { "bbox": [ 106, 140, 505, 151 ], "score": 1.0, "content": "time, there are a number of ethical concerns with language models in general, including concerns", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 150, 506, 163 ], "spans": [ { "bbox": [ 105, 150, 506, 163 ], "score": 1.0, "content": "regarding the generation of biased or discriminative text (Bordia & Bowman, 2019), the leakage of", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 161, 506, 174 ], "spans": [ { "bbox": [ 105, 161, 506, 174 ], "score": 1.0, "content": "private information from training data (Carlini et al., 2020), and environmental impact of training or", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 172, 247, 185 ], "spans": [ { "bbox": [ 105, 172, 247, 185 ], "score": 1.0, "content": "tuning them (Strubell et al., 2019).", "type": "text" } ], "index": 7 } ], "index": 4 }, { "type": "text", "bbox": [ 107, 189, 505, 233 ], "lines": [ { "bbox": [ 105, 189, 505, 201 ], "spans": [ { "bbox": [ 105, 189, 505, 201 ], "score": 1.0, "content": "Our method attempts to train language models making minimal changes to their pre-existing param-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 200, 505, 212 ], "spans": [ { "bbox": [ 105, 200, 505, 212 ], "score": 1.0, "content": "eters. While it is an interesting research question whether parameter-efficient fine-tuning methods", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 211, 505, 223 ], "spans": [ { "bbox": [ 106, 211, 505, 223 ], "score": 1.0, "content": "exacerbate, mitigate, or make little change to issues such as bias or information leakage, to our", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 222, 494, 234 ], "spans": [ { "bbox": [ 106, 222, 494, 234 ], "score": 1.0, "content": "knowledge no previous work has examined this topic. It is an interesting avenue for future work.", "type": "text" } ], "index": 11 } ], "index": 9.5 }, { "type": "text", "bbox": [ 107, 239, 505, 338 ], "lines": [ { "bbox": [ 105, 238, 506, 251 ], "spans": [ { "bbox": [ 105, 238, 506, 251 ], "score": 1.0, "content": "With respect to environmental impact, the methods proposed in this paper add a small number of", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 250, 505, 261 ], "spans": [ { "bbox": [ 105, 250, 505, 261 ], "score": 1.0, "content": "extra parameters and components to existing models, and thus they have a nominal negative impact", "type": "text" } ], "index": 13 }, { "bbox": [ 104, 259, 506, 275 ], "spans": [ { "bbox": [ 104, 259, 414, 275 ], "score": 1.0, "content": "on training and inference time – for example, the final MAM Adapter needs", "type": "text" }, { "bbox": [ 414, 261, 470, 271 ], "score": 0.84, "content": "1 0 0 \\% - 1 5 0 \\%", "type": "inline_equation" }, { "bbox": [ 471, 259, 506, 275 ], "score": 1.0, "content": "training", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 272, 505, 284 ], "spans": [ { "bbox": [ 106, 272, 505, 284 ], "score": 1.0, "content": "time of full fine-tuning in our four benchmarks since parameter-efficient tuning typically needs more", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 282, 505, 297 ], "spans": [ { "bbox": [ 105, 282, 505, 297 ], "score": 1.0, "content": "epochs to converge; the inference time is roughly the same as the model obtained by full fine-tuning.", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 291, 505, 308 ], "spans": [ { "bbox": [ 105, 291, 505, 308 ], "score": 1.0, "content": "On the other hand, as the methods proposed in this paper may obviate the need for full fine-tuning,", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 304, 505, 318 ], "spans": [ { "bbox": [ 105, 304, 505, 318 ], "score": 1.0, "content": "this may also significantly reduce the cost (in terms of memory/deployed servers) of serving models.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 314, 506, 330 ], "spans": [ { "bbox": [ 105, 314, 506, 330 ], "score": 1.0, "content": "Notably, the great majority of the experimentation done for this paper was performed on a data center", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 326, 262, 340 ], "spans": [ { "bbox": [ 105, 326, 262, 340 ], "score": 1.0, "content": "powered entirely by renewable energy.", "type": "text" } ], "index": 20 } ], "index": 16 }, { "type": "title", "bbox": [ 108, 353, 267, 365 ], "lines": [ { "bbox": [ 106, 353, 269, 367 ], "spans": [ { "bbox": [ 106, 353, 269, 367 ], "score": 1.0, "content": "REPRODUCIBILITY STATEMENT", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 108, 378, 505, 411 ], "lines": [ { "bbox": [ 105, 376, 505, 390 ], "spans": [ { "bbox": [ 105, 376, 261, 390 ], "score": 1.0, "content": "In addition to the setup description in", "type": "text" }, { "bbox": [ 261, 378, 279, 389 ], "score": 0.86, "content": "\\ S 4 . 1", "type": "inline_equation" }, { "bbox": [ 280, 376, 505, 390 ], "score": 1.0, "content": ", we have detailed the complete experiments setup such", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 388, 505, 401 ], "spans": [ { "bbox": [ 105, 388, 505, 401 ], "score": 1.0, "content": "as batch size, optimizer, learning rates in Appendix A. Besides, we have publicized our source code.", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 398, 387, 413 ], "spans": [ { "bbox": [ 105, 398, 387, 413 ], "score": 1.0, "content": "These resources should be sufficient to reproduce results of the paper.", "type": "text" } ], "index": 24 } ], "index": 23 }, { "type": "title", "bbox": [ 108, 427, 219, 438 ], "lines": [ { "bbox": [ 106, 427, 220, 441 ], "spans": [ { "bbox": [ 106, 427, 220, 441 ], "score": 1.0, "content": "ACKNOWLEDGEMENT", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 451, 505, 484 ], "lines": [ { "bbox": [ 106, 451, 505, 463 ], "spans": [ { "bbox": [ 106, 451, 505, 463 ], "score": 1.0, "content": "We thank the anonymous reviewers for their comments. This work was supported in part by the", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 462, 505, 474 ], "spans": [ { "bbox": [ 106, 462, 505, 474 ], "score": 1.0, "content": "CMU-Portugal MAIA Project, a Baidu PhD Fellowship for Junxian He, and a CMU Presidential", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 472, 233, 486 ], "spans": [ { "bbox": [ 105, 472, 233, 486 ], "score": 1.0, "content": "Fellowship for Chunting Zhou.", "type": "text" } ], "index": 28 } ], "index": 27 }, { "type": "title", "bbox": [ 108, 500, 175, 512 ], "lines": [ { "bbox": [ 106, 500, 176, 513 ], "spans": [ { "bbox": [ 106, 500, 176, 513 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 106, 524, 504, 547 ], "lines": [ { "bbox": [ 105, 523, 506, 538 ], "spans": [ { "bbox": [ 105, 523, 506, 538 ], "score": 1.0, "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint", "type": "text" } ], "index": 30 }, { "bbox": [ 116, 536, 219, 547 ], "spans": [ { "bbox": [ 116, 536, 219, 547 ], "score": 1.0, "content": "arXiv:1607.06450, 2016.", "type": "text" } ], "index": 31 } ], "index": 30.5 }, { "type": "text", "bbox": [ 106, 552, 505, 576 ], "lines": [ { "bbox": [ 105, 551, 506, 567 ], "spans": [ { "bbox": [ 105, 551, 506, 567 ], "score": 1.0, "content": "Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 563, 465, 576 ], "spans": [ { "bbox": [ 115, 563, 465, 576 ], "score": 1.0, "content": "for transformer-based masked language-models. arXiv e-prints, pp. arXiv–2106, 2021.", "type": "text" } ], "index": 33 } ], "index": 32.5 }, { "type": "text", "bbox": [ 106, 580, 506, 625 ], "lines": [ { "bbox": [ 106, 581, 506, 594 ], "spans": [ { "bbox": [ 106, 581, 506, 594 ], "score": 1.0, "content": "Ondˇrej Bojar, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow, Matthias", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 590, 505, 605 ], "spans": [ { "bbox": [ 115, 590, 505, 605 ], "score": 1.0, "content": "Huck, Antonio Jimeno Yepes, Philipp Koehn, Varvara Logacheva, Christof Monz, et al. Findings", "type": "text" } ], "index": 35 }, { "bbox": [ 115, 602, 506, 616 ], "spans": [ { "bbox": [ 115, 602, 506, 616 ], "score": 1.0, "content": "of the 2016 conference on machine translation. In Proceedings of the First Conference on Machine", "type": "text" } ], "index": 36 }, { "bbox": [ 116, 614, 318, 626 ], "spans": [ { "bbox": [ 116, 614, 318, 626 ], "score": 1.0, "content": "Translation: Volume 2, Shared Task Papers, 2016.", "type": "text" } ], "index": 37 } ], "index": 35.5 }, { "type": "text", "bbox": [ 107, 631, 502, 654 ], "lines": [ { "bbox": [ 105, 629, 504, 645 ], "spans": [ { "bbox": [ 105, 629, 504, 645 ], "score": 1.0, "content": "Shikha Bordia and Samuel R. Bowman. Identifying and reducing gender bias in word-level language", "type": "text" } ], "index": 38 }, { "bbox": [ 116, 642, 439, 655 ], "spans": [ { "bbox": [ 116, 642, 439, 655 ], "score": 1.0, "content": "models. In Proceedings of the 2019 NAACL: Student Research Workshop, 2019.", "type": "text" } ], "index": 39 } ], "index": 38.5 }, { "type": "text", "bbox": [ 107, 659, 503, 693 ], "lines": [ { "bbox": [ 105, 658, 505, 672 ], "spans": [ { "bbox": [ 105, 658, 505, 672 ], "score": 1.0, "content": "Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal,", "type": "text" } ], "index": 40 }, { "bbox": [ 116, 670, 505, 683 ], "spans": [ { "bbox": [ 116, 670, 505, 683 ], "score": 1.0, "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 681, 353, 694 ], "spans": [ { "bbox": [ 115, 681, 353, 694 ], "score": 1.0, "content": "few-shot learners. arXiv preprint arXiv:2005.14165, 2020.", "type": "text" } ], "index": 42 } ], "index": 41 }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 105, 697, 506, 712 ], "spans": [ { "bbox": [ 105, 697, 506, 712 ], "score": 1.0, "content": "Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 709, 505, 722 ], "spans": [ { "bbox": [ 115, 709, 505, 722 ], "score": 1.0, "content": "Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data", "type": "text" } ], "index": 44 }, { "bbox": [ 116, 720, 396, 733 ], "spans": [ { "bbox": [ 116, 720, 396, 733 ], "score": 1.0, "content": "from large language models. arXiv preprint arXiv:2012.07805, 2020.", "type": "text" } ], "index": 45 } ], "index": 44 } ], "page_idx": 9, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 294, 38 ], "spans": [ { "bbox": [ 106, 25, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 765 ], "spans": [ { "bbox": [ 299, 750, 313, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 81, 210, 94 ], "lines": [ { "bbox": [ 106, 81, 213, 95 ], "spans": [ { "bbox": [ 106, 81, 213, 95 ], "score": 1.0, "content": "ETHICS STATEMENT", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 106, 505, 183 ], "lines": [ { "bbox": [ 105, 105, 505, 120 ], "spans": [ { "bbox": [ 105, 105, 505, 120 ], "score": 1.0, "content": "Our work proposes a method for efficient fine-tuning of pre-trained models, in particular language", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 117, 505, 130 ], "spans": [ { "bbox": [ 105, 117, 505, 130 ], "score": 1.0, "content": "models. Pre-trained language models have a wide variety of positive applications, such as the appli-", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 129, 504, 140 ], "spans": [ { "bbox": [ 106, 129, 504, 140 ], "score": 1.0, "content": "cations to summarization, translation, or language understanding described in our paper. At the same", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 140, 505, 151 ], "spans": [ { "bbox": [ 106, 140, 505, 151 ], "score": 1.0, "content": "time, there are a number of ethical concerns with language models in general, including concerns", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 150, 506, 163 ], "spans": [ { "bbox": [ 105, 150, 506, 163 ], "score": 1.0, "content": "regarding the generation of biased or discriminative text (Bordia & Bowman, 2019), the leakage of", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 161, 506, 174 ], "spans": [ { "bbox": [ 105, 161, 506, 174 ], "score": 1.0, "content": "private information from training data (Carlini et al., 2020), and environmental impact of training or", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 172, 247, 185 ], "spans": [ { "bbox": [ 105, 172, 247, 185 ], "score": 1.0, "content": "tuning them (Strubell et al., 2019).", "type": "text" } ], "index": 7 } ], "index": 4, "bbox_fs": [ 105, 105, 506, 185 ] }, { "type": "text", "bbox": [ 107, 189, 505, 233 ], "lines": [ { "bbox": [ 105, 189, 505, 201 ], "spans": [ { "bbox": [ 105, 189, 505, 201 ], "score": 1.0, "content": "Our method attempts to train language models making minimal changes to their pre-existing param-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 200, 505, 212 ], "spans": [ { "bbox": [ 105, 200, 505, 212 ], "score": 1.0, "content": "eters. While it is an interesting research question whether parameter-efficient fine-tuning methods", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 211, 505, 223 ], "spans": [ { "bbox": [ 106, 211, 505, 223 ], "score": 1.0, "content": "exacerbate, mitigate, or make little change to issues such as bias or information leakage, to our", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 222, 494, 234 ], "spans": [ { "bbox": [ 106, 222, 494, 234 ], "score": 1.0, "content": "knowledge no previous work has examined this topic. It is an interesting avenue for future work.", "type": "text" } ], "index": 11 } ], "index": 9.5, "bbox_fs": [ 105, 189, 505, 234 ] }, { "type": "text", "bbox": [ 107, 239, 505, 338 ], "lines": [ { "bbox": [ 105, 238, 506, 251 ], "spans": [ { "bbox": [ 105, 238, 506, 251 ], "score": 1.0, "content": "With respect to environmental impact, the methods proposed in this paper add a small number of", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 250, 505, 261 ], "spans": [ { "bbox": [ 105, 250, 505, 261 ], "score": 1.0, "content": "extra parameters and components to existing models, and thus they have a nominal negative impact", "type": "text" } ], "index": 13 }, { "bbox": [ 104, 259, 506, 275 ], "spans": [ { "bbox": [ 104, 259, 414, 275 ], "score": 1.0, "content": "on training and inference time – for example, the final MAM Adapter needs", "type": "text" }, { "bbox": [ 414, 261, 470, 271 ], "score": 0.84, "content": "1 0 0 \\% - 1 5 0 \\%", "type": "inline_equation" }, { "bbox": [ 471, 259, 506, 275 ], "score": 1.0, "content": "training", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 272, 505, 284 ], "spans": [ { "bbox": [ 106, 272, 505, 284 ], "score": 1.0, "content": "time of full fine-tuning in our four benchmarks since parameter-efficient tuning typically needs more", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 282, 505, 297 ], "spans": [ { "bbox": [ 105, 282, 505, 297 ], "score": 1.0, "content": "epochs to converge; the inference time is roughly the same as the model obtained by full fine-tuning.", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 291, 505, 308 ], "spans": [ { "bbox": [ 105, 291, 505, 308 ], "score": 1.0, "content": "On the other hand, as the methods proposed in this paper may obviate the need for full fine-tuning,", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 304, 505, 318 ], "spans": [ { "bbox": [ 105, 304, 505, 318 ], "score": 1.0, "content": "this may also significantly reduce the cost (in terms of memory/deployed servers) of serving models.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 314, 506, 330 ], "spans": [ { "bbox": [ 105, 314, 506, 330 ], "score": 1.0, "content": "Notably, the great majority of the experimentation done for this paper was performed on a data center", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 326, 262, 340 ], "spans": [ { "bbox": [ 105, 326, 262, 340 ], "score": 1.0, "content": "powered entirely by renewable energy.", "type": "text" } ], "index": 20 } ], "index": 16, "bbox_fs": [ 104, 238, 506, 340 ] }, { "type": "title", "bbox": [ 108, 353, 267, 365 ], "lines": [ { "bbox": [ 106, 353, 269, 367 ], "spans": [ { "bbox": [ 106, 353, 269, 367 ], "score": 1.0, "content": "REPRODUCIBILITY STATEMENT", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 108, 378, 505, 411 ], "lines": [ { "bbox": [ 105, 376, 505, 390 ], "spans": [ { "bbox": [ 105, 376, 261, 390 ], "score": 1.0, "content": "In addition to the setup description in", "type": "text" }, { "bbox": [ 261, 378, 279, 389 ], "score": 0.86, "content": "\\ S 4 . 1", "type": "inline_equation" }, { "bbox": [ 280, 376, 505, 390 ], "score": 1.0, "content": ", we have detailed the complete experiments setup such", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 388, 505, 401 ], "spans": [ { "bbox": [ 105, 388, 505, 401 ], "score": 1.0, "content": "as batch size, optimizer, learning rates in Appendix A. Besides, we have publicized our source code.", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 398, 387, 413 ], "spans": [ { "bbox": [ 105, 398, 387, 413 ], "score": 1.0, "content": "These resources should be sufficient to reproduce results of the paper.", "type": "text" } ], "index": 24 } ], "index": 23, "bbox_fs": [ 105, 376, 505, 413 ] }, { "type": "title", "bbox": [ 108, 427, 219, 438 ], "lines": [ { "bbox": [ 106, 427, 220, 441 ], "spans": [ { "bbox": [ 106, 427, 220, 441 ], "score": 1.0, "content": "ACKNOWLEDGEMENT", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 451, 505, 484 ], "lines": [ { "bbox": [ 106, 451, 505, 463 ], "spans": [ { "bbox": [ 106, 451, 505, 463 ], "score": 1.0, "content": "We thank the anonymous reviewers for their comments. This work was supported in part by the", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 462, 505, 474 ], "spans": [ { "bbox": [ 106, 462, 505, 474 ], "score": 1.0, "content": "CMU-Portugal MAIA Project, a Baidu PhD Fellowship for Junxian He, and a CMU Presidential", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 472, 233, 486 ], "spans": [ { "bbox": [ 105, 472, 233, 486 ], "score": 1.0, "content": "Fellowship for Chunting Zhou.", "type": "text" } ], "index": 28 } ], "index": 27, "bbox_fs": [ 105, 451, 505, 486 ] }, { "type": "title", "bbox": [ 108, 500, 175, 512 ], "lines": [ { "bbox": [ 106, 500, 176, 513 ], "spans": [ { "bbox": [ 106, 500, 176, 513 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 106, 524, 504, 547 ], "lines": [ { "bbox": [ 105, 523, 506, 538 ], "spans": [ { "bbox": [ 105, 523, 506, 538 ], "score": 1.0, "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint", "type": "text" } ], "index": 30 }, { "bbox": [ 116, 536, 219, 547 ], "spans": [ { "bbox": [ 116, 536, 219, 547 ], "score": 1.0, "content": "arXiv:1607.06450, 2016.", "type": "text" } ], "index": 31 } ], "index": 30.5, "bbox_fs": [ 105, 523, 506, 547 ] }, { "type": "text", "bbox": [ 106, 552, 505, 576 ], "lines": [ { "bbox": [ 105, 551, 506, 567 ], "spans": [ { "bbox": [ 105, 551, 506, 567 ], "score": 1.0, "content": "Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 563, 465, 576 ], "spans": [ { "bbox": [ 115, 563, 465, 576 ], "score": 1.0, "content": "for transformer-based masked language-models. arXiv e-prints, pp. arXiv–2106, 2021.", "type": "text" } ], "index": 33 } ], "index": 32.5, "bbox_fs": [ 105, 551, 506, 576 ] }, { "type": "text", "bbox": [ 106, 580, 506, 625 ], "lines": [ { "bbox": [ 106, 581, 506, 594 ], "spans": [ { "bbox": [ 106, 581, 506, 594 ], "score": 1.0, "content": "Ondˇrej Bojar, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow, Matthias", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 590, 505, 605 ], "spans": [ { "bbox": [ 115, 590, 505, 605 ], "score": 1.0, "content": "Huck, Antonio Jimeno Yepes, Philipp Koehn, Varvara Logacheva, Christof Monz, et al. Findings", "type": "text" } ], "index": 35 }, { "bbox": [ 115, 602, 506, 616 ], "spans": [ { "bbox": [ 115, 602, 506, 616 ], "score": 1.0, "content": "of the 2016 conference on machine translation. In Proceedings of the First Conference on Machine", "type": "text" } ], "index": 36 }, { "bbox": [ 116, 614, 318, 626 ], "spans": [ { "bbox": [ 116, 614, 318, 626 ], "score": 1.0, "content": "Translation: Volume 2, Shared Task Papers, 2016.", "type": "text" } ], "index": 37 } ], "index": 35.5, "bbox_fs": [ 106, 581, 506, 626 ] }, { "type": "text", "bbox": [ 107, 631, 502, 654 ], "lines": [ { "bbox": [ 105, 629, 504, 645 ], "spans": [ { "bbox": [ 105, 629, 504, 645 ], "score": 1.0, "content": "Shikha Bordia and Samuel R. Bowman. Identifying and reducing gender bias in word-level language", "type": "text" } ], "index": 38 }, { "bbox": [ 116, 642, 439, 655 ], "spans": [ { "bbox": [ 116, 642, 439, 655 ], "score": 1.0, "content": "models. In Proceedings of the 2019 NAACL: Student Research Workshop, 2019.", "type": "text" } ], "index": 39 } ], "index": 38.5, "bbox_fs": [ 105, 629, 504, 655 ] }, { "type": "text", "bbox": [ 107, 659, 503, 693 ], "lines": [ { "bbox": [ 105, 658, 505, 672 ], "spans": [ { "bbox": [ 105, 658, 505, 672 ], "score": 1.0, "content": "Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal,", "type": "text" } ], "index": 40 }, { "bbox": [ 116, 670, 505, 683 ], "spans": [ { "bbox": [ 116, 670, 505, 683 ], "score": 1.0, "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 681, 353, 694 ], "spans": [ { "bbox": [ 115, 681, 353, 694 ], "score": 1.0, "content": "few-shot learners. arXiv preprint arXiv:2005.14165, 2020.", "type": "text" } ], "index": 42 } ], "index": 41, "bbox_fs": [ 105, 658, 505, 694 ] }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 105, 697, 506, 712 ], "spans": [ { "bbox": [ 105, 697, 506, 712 ], "score": 1.0, "content": "Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 709, 505, 722 ], "spans": [ { "bbox": [ 115, 709, 505, 722 ], "score": 1.0, "content": "Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data", "type": "text" } ], "index": 44 }, { "bbox": [ 116, 720, 396, 733 ], "spans": [ { "bbox": [ 116, 720, 396, 733 ], "score": 1.0, "content": "from large language models. arXiv preprint arXiv:2012.07805, 2020.", "type": "text" } ], "index": 45 } ], "index": 44, "bbox_fs": [ 105, 697, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 82, 504, 106 ], "lines": [ { "bbox": [ 105, 81, 506, 96 ], "spans": [ { "bbox": [ 105, 81, 506, 96 ], "score": 1.0, "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", "type": "text" } ], "index": 0 }, { "bbox": [ 116, 94, 470, 105 ], "spans": [ { "bbox": [ 116, 94, 470, 105 ], "score": 1.0, "content": "bidirectional transformers for language understanding. In Proceedings of NAACL, 2019.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 106, 113, 503, 136 ], "lines": [ { "bbox": [ 106, 112, 505, 126 ], "spans": [ { "bbox": [ 106, 112, 505, 126 ], "score": 1.0, "content": "William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter", "type": "text" } ], "index": 2 }, { "bbox": [ 115, 124, 448, 137 ], "spans": [ { "bbox": [ 115, 124, 448, 137 ], "score": 1.0, "content": "models with simple and efficient sparsity. arXiv preprint arXiv:2101.03961, 2021.", "type": "text" } ], "index": 3 } ], "index": 2.5 }, { "type": "text", "bbox": [ 106, 144, 502, 168 ], "lines": [ { "bbox": [ 105, 143, 504, 158 ], "spans": [ { "bbox": [ 105, 143, 504, 158 ], "score": 1.0, "content": "Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 155, 340, 168 ], "spans": [ { "bbox": [ 116, 155, 340, 168 ], "score": 1.0, "content": "key-value memories. In Proceedings of EMNLP, 2021.", "type": "text" } ], "index": 5 } ], "index": 4.5 }, { "type": "text", "bbox": [ 107, 175, 503, 199 ], "lines": [ { "bbox": [ 105, 174, 504, 189 ], "spans": [ { "bbox": [ 105, 174, 504, 189 ], "score": 1.0, "content": "Demi Guo, Alexander M Rush, and Yoon Kim. Parameter-efficient transfer learning with diff prun-", "type": "text" } ], "index": 6 }, { "bbox": [ 115, 187, 256, 199 ], "spans": [ { "bbox": [ 115, 187, 256, 199 ], "score": 1.0, "content": "ing. In Proceedings of ACL, 2021.", "type": "text" } ], "index": 7 } ], "index": 6.5 }, { "type": "text", "bbox": [ 106, 206, 504, 230 ], "lines": [ { "bbox": [ 105, 205, 506, 221 ], "spans": [ { "bbox": [ 105, 205, 506, 221 ], "score": 1.0, "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing", "type": "text" } ], "index": 8 }, { "bbox": [ 115, 218, 457, 230 ], "spans": [ { "bbox": [ 115, 218, 457, 230 ], "score": 1.0, "content": "human-level performance on imagenet classification. In Proceedings of ICCV, 2015.", "type": "text" } ], "index": 9 } ], "index": 8.5 }, { "type": "text", "bbox": [ 106, 237, 506, 271 ], "lines": [ { "bbox": [ 106, 237, 504, 250 ], "spans": [ { "bbox": [ 106, 237, 504, 250 ], "score": 1.0, "content": "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, An-", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 248, 505, 262 ], "spans": [ { "bbox": [ 115, 248, 505, 262 ], "score": 1.0, "content": "drea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp.", "type": "text" } ], "index": 11 }, { "bbox": [ 116, 260, 243, 272 ], "spans": [ { "bbox": [ 116, 260, 243, 272 ], "score": 1.0, "content": "In Proceedings of ICML, 2019.", "type": "text" } ], "index": 12 } ], "index": 11 }, { "type": "text", "bbox": [ 106, 279, 505, 313 ], "lines": [ { "bbox": [ 105, 279, 505, 293 ], "spans": [ { "bbox": [ 105, 279, 505, 293 ], "score": 1.0, "content": "Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 290, 505, 303 ], "spans": [ { "bbox": [ 115, 290, 505, 303 ], "score": 1.0, "content": "Chen. LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685,", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 300, 144, 315 ], "spans": [ { "bbox": [ 115, 300, 144, 315 ], "score": 1.0, "content": "2021.", "type": "text" } ], "index": 15 } ], "index": 14 }, { "type": "text", "bbox": [ 105, 321, 505, 345 ], "lines": [ { "bbox": [ 105, 320, 506, 335 ], "spans": [ { "bbox": [ 105, 320, 506, 335 ], "score": 1.0, "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 332, 169, 344 ], "spans": [ { "bbox": [ 115, 332, 169, 344 ], "score": 1.0, "content": "ICLR, 2015.", "type": "text" } ], "index": 17 } ], "index": 16.5 }, { "type": "text", "bbox": [ 106, 352, 503, 376 ], "lines": [ { "bbox": [ 105, 351, 505, 367 ], "spans": [ { "bbox": [ 105, 351, 505, 367 ], "score": 1.0, "content": "Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 364, 284, 376 ], "spans": [ { "bbox": [ 115, 364, 284, 376 ], "score": 1.0, "content": "tuning. In Proceedings of EMNLP, 2021.", "type": "text" } ], "index": 19 } ], "index": 18.5 }, { "type": "text", "bbox": [ 106, 383, 505, 428 ], "lines": [ { "bbox": [ 104, 383, 506, 397 ], "spans": [ { "bbox": [ 104, 383, 506, 397 ], "score": 1.0, "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer", "type": "text" } ], "index": 20 }, { "bbox": [ 115, 394, 505, 408 ], "spans": [ { "bbox": [ 115, 394, 505, 408 ], "score": 1.0, "content": "Levy, Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 406, 506, 419 ], "spans": [ { "bbox": [ 115, 406, 506, 419 ], "score": 1.0, "content": "training for natural language generation, translation, and comprehension. In Proceedings of ACL,", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 415, 142, 429 ], "spans": [ { "bbox": [ 115, 415, 142, 429 ], "score": 1.0, "content": "2020.", "type": "text" } ], "index": 23 } ], "index": 21.5 }, { "type": "text", "bbox": [ 107, 436, 504, 460 ], "lines": [ { "bbox": [ 105, 436, 506, 450 ], "spans": [ { "bbox": [ 105, 436, 506, 450 ], "score": 1.0, "content": "Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In", "type": "text" } ], "index": 24 }, { "bbox": [ 116, 448, 227, 461 ], "spans": [ { "bbox": [ 116, 448, 227, 461 ], "score": 1.0, "content": "Proceedings of ACL, 2021.", "type": "text" } ], "index": 25 } ], "index": 24.5 }, { "type": "text", "bbox": [ 105, 468, 504, 491 ], "lines": [ { "bbox": [ 106, 468, 505, 481 ], "spans": [ { "bbox": [ 106, 468, 505, 481 ], "score": 1.0, "content": "Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization", "type": "text" } ], "index": 26 }, { "bbox": [ 116, 479, 201, 490 ], "spans": [ { "bbox": [ 116, 479, 201, 490 ], "score": 1.0, "content": "Branches Out, 2004.", "type": "text" } ], "index": 27 } ], "index": 26.5 }, { "type": "text", "bbox": [ 107, 498, 504, 533 ], "lines": [ { "bbox": [ 105, 498, 504, 512 ], "spans": [ { "bbox": [ 105, 498, 504, 512 ], "score": 1.0, "content": "Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 510, 505, 524 ], "spans": [ { "bbox": [ 115, 510, 505, 524 ], "score": 1.0, "content": "train, prompt, and predict: A systematic survey of prompting methods in natural language pro-", "type": "text" } ], "index": 29 }, { "bbox": [ 116, 522, 319, 533 ], "spans": [ { "bbox": [ 116, 522, 319, 533 ], "score": 1.0, "content": "cessing. arXiv preprint arXiv:2107.13586, 2021a.", "type": "text" } ], "index": 30 } ], "index": 29 }, { "type": "text", "bbox": [ 106, 541, 503, 564 ], "lines": [ { "bbox": [ 105, 540, 505, 554 ], "spans": [ { "bbox": [ 105, 540, 505, 554 ], "score": 1.0, "content": "Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. GPT", "type": "text" } ], "index": 31 }, { "bbox": [ 116, 552, 295, 563 ], "spans": [ { "bbox": [ 116, 552, 295, 563 ], "score": 1.0, "content": "understands, too. arXiv:2103.10385, 2021b.", "type": "text" } ], "index": 32 } ], "index": 31.5 }, { "type": "text", "bbox": [ 106, 572, 506, 606 ], "lines": [ { "bbox": [ 106, 572, 505, 585 ], "spans": [ { "bbox": [ 106, 572, 505, 585 ], "score": 1.0, "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 582, 506, 597 ], "spans": [ { "bbox": [ 115, 582, 506, 597 ], "score": 1.0, "content": "Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized bert pretraining", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 595, 321, 606 ], "spans": [ { "bbox": [ 115, 595, 321, 606 ], "score": 1.0, "content": "approach. arXiv preprint arXiv:1907.11692, 2019.", "type": "text" } ], "index": 35 } ], "index": 34 }, { "type": "text", "bbox": [ 107, 614, 504, 649 ], "lines": [ { "bbox": [ 105, 613, 506, 628 ], "spans": [ { "bbox": [ 105, 613, 506, 628 ], "score": 1.0, "content": "Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis,", "type": "text" } ], "index": 36 }, { "bbox": [ 116, 625, 506, 638 ], "spans": [ { "bbox": [ 116, 625, 506, 638 ], "score": 1.0, "content": "and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Trans-", "type": "text" } ], "index": 37 }, { "bbox": [ 116, 636, 376, 649 ], "spans": [ { "bbox": [ 116, 636, 376, 649 ], "score": 1.0, "content": "actions of the Association for Computational Linguistics, 2020a.", "type": "text" } ], "index": 38 } ], "index": 37 }, { "type": "text", "bbox": [ 107, 656, 505, 701 ], "lines": [ { "bbox": [ 106, 656, 505, 669 ], "spans": [ { "bbox": [ 106, 656, 505, 669 ], "score": 1.0, "content": "Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike", "type": "text" } ], "index": 39 }, { "bbox": [ 114, 667, 505, 680 ], "spans": [ { "bbox": [ 114, 667, 505, 680 ], "score": 1.0, "content": "Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine transla-", "type": "text" } ], "index": 40 }, { "bbox": [ 116, 679, 504, 691 ], "spans": [ { "bbox": [ 116, 679, 504, 691 ], "score": 1.0, "content": "tion. Transactions of the Association for Computational Linguistics, 8:726–742, 2020b. doi:", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 689, 470, 702 ], "spans": [ { "bbox": [ 115, 689, 470, 702 ], "score": 1.0, "content": "10.1162/tacl a 00343. URL https://aclanthology.org/2020.tacl-1.47.", "type": "text" } ], "index": 42 } ], "index": 40.5 }, { "type": "text", "bbox": [ 107, 709, 503, 732 ], "lines": [ { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 721, 375, 732 ], "spans": [ { "bbox": [ 115, 721, 375, 732 ], "score": 1.0, "content": "hypercomplex adapter layers. In Proceedings of NeurIPS, 2021.", "type": "text" } ], "index": 44 } ], "index": 43.5 } ], "page_idx": 10, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "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 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 105, 82, 504, 106 ], "lines": [ { "bbox": [ 105, 81, 506, 96 ], "spans": [ { "bbox": [ 105, 81, 506, 96 ], "score": 1.0, "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", "type": "text" } ], "index": 0 }, { "bbox": [ 116, 94, 470, 105 ], "spans": [ { "bbox": [ 116, 94, 470, 105 ], "score": 1.0, "content": "bidirectional transformers for language understanding. In Proceedings of NAACL, 2019.", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 105, 81, 506, 105 ] }, { "type": "text", "bbox": [ 106, 113, 503, 136 ], "lines": [ { "bbox": [ 106, 112, 505, 126 ], "spans": [ { "bbox": [ 106, 112, 505, 126 ], "score": 1.0, "content": "William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter", "type": "text" } ], "index": 2 }, { "bbox": [ 115, 124, 448, 137 ], "spans": [ { "bbox": [ 115, 124, 448, 137 ], "score": 1.0, "content": "models with simple and efficient sparsity. arXiv preprint arXiv:2101.03961, 2021.", "type": "text" } ], "index": 3 } ], "index": 2.5, "bbox_fs": [ 106, 112, 505, 137 ] }, { "type": "text", "bbox": [ 106, 144, 502, 168 ], "lines": [ { "bbox": [ 105, 143, 504, 158 ], "spans": [ { "bbox": [ 105, 143, 504, 158 ], "score": 1.0, "content": "Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 155, 340, 168 ], "spans": [ { "bbox": [ 116, 155, 340, 168 ], "score": 1.0, "content": "key-value memories. In Proceedings of EMNLP, 2021.", "type": "text" } ], "index": 5 } ], "index": 4.5, "bbox_fs": [ 105, 143, 504, 168 ] }, { "type": "text", "bbox": [ 107, 175, 503, 199 ], "lines": [ { "bbox": [ 105, 174, 504, 189 ], "spans": [ { "bbox": [ 105, 174, 504, 189 ], "score": 1.0, "content": "Demi Guo, Alexander M Rush, and Yoon Kim. Parameter-efficient transfer learning with diff prun-", "type": "text" } ], "index": 6 }, { "bbox": [ 115, 187, 256, 199 ], "spans": [ { "bbox": [ 115, 187, 256, 199 ], "score": 1.0, "content": "ing. In Proceedings of ACL, 2021.", "type": "text" } ], "index": 7 } ], "index": 6.5, "bbox_fs": [ 105, 174, 504, 199 ] }, { "type": "text", "bbox": [ 106, 206, 504, 230 ], "lines": [ { "bbox": [ 105, 205, 506, 221 ], "spans": [ { "bbox": [ 105, 205, 506, 221 ], "score": 1.0, "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing", "type": "text" } ], "index": 8 }, { "bbox": [ 115, 218, 457, 230 ], "spans": [ { "bbox": [ 115, 218, 457, 230 ], "score": 1.0, "content": "human-level performance on imagenet classification. In Proceedings of ICCV, 2015.", "type": "text" } ], "index": 9 } ], "index": 8.5, "bbox_fs": [ 105, 205, 506, 230 ] }, { "type": "text", "bbox": [ 106, 237, 506, 271 ], "lines": [ { "bbox": [ 106, 237, 504, 250 ], "spans": [ { "bbox": [ 106, 237, 504, 250 ], "score": 1.0, "content": "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, An-", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 248, 505, 262 ], "spans": [ { "bbox": [ 115, 248, 505, 262 ], "score": 1.0, "content": "drea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp.", "type": "text" } ], "index": 11 }, { "bbox": [ 116, 260, 243, 272 ], "spans": [ { "bbox": [ 116, 260, 243, 272 ], "score": 1.0, "content": "In Proceedings of ICML, 2019.", "type": "text" } ], "index": 12 } ], "index": 11, "bbox_fs": [ 106, 237, 505, 272 ] }, { "type": "text", "bbox": [ 106, 279, 505, 313 ], "lines": [ { "bbox": [ 105, 279, 505, 293 ], "spans": [ { "bbox": [ 105, 279, 505, 293 ], "score": 1.0, "content": "Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 290, 505, 303 ], "spans": [ { "bbox": [ 115, 290, 505, 303 ], "score": 1.0, "content": "Chen. LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685,", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 300, 144, 315 ], "spans": [ { "bbox": [ 115, 300, 144, 315 ], "score": 1.0, "content": "2021.", "type": "text" } ], "index": 15 } ], "index": 14, "bbox_fs": [ 105, 279, 505, 315 ] }, { "type": "text", "bbox": [ 105, 321, 505, 345 ], "lines": [ { "bbox": [ 105, 320, 506, 335 ], "spans": [ { "bbox": [ 105, 320, 506, 335 ], "score": 1.0, "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 332, 169, 344 ], "spans": [ { "bbox": [ 115, 332, 169, 344 ], "score": 1.0, "content": "ICLR, 2015.", "type": "text" } ], "index": 17 } ], "index": 16.5, "bbox_fs": [ 105, 320, 506, 344 ] }, { "type": "text", "bbox": [ 106, 352, 503, 376 ], "lines": [ { "bbox": [ 105, 351, 505, 367 ], "spans": [ { "bbox": [ 105, 351, 505, 367 ], "score": 1.0, "content": "Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 364, 284, 376 ], "spans": [ { "bbox": [ 115, 364, 284, 376 ], "score": 1.0, "content": "tuning. In Proceedings of EMNLP, 2021.", "type": "text" } ], "index": 19 } ], "index": 18.5, "bbox_fs": [ 105, 351, 505, 376 ] }, { "type": "text", "bbox": [ 106, 383, 505, 428 ], "lines": [ { "bbox": [ 104, 383, 506, 397 ], "spans": [ { "bbox": [ 104, 383, 506, 397 ], "score": 1.0, "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer", "type": "text" } ], "index": 20 }, { "bbox": [ 115, 394, 505, 408 ], "spans": [ { "bbox": [ 115, 394, 505, 408 ], "score": 1.0, "content": "Levy, Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 406, 506, 419 ], "spans": [ { "bbox": [ 115, 406, 506, 419 ], "score": 1.0, "content": "training for natural language generation, translation, and comprehension. In Proceedings of ACL,", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 415, 142, 429 ], "spans": [ { "bbox": [ 115, 415, 142, 429 ], "score": 1.0, "content": "2020.", "type": "text" } ], "index": 23 } ], "index": 21.5, "bbox_fs": [ 104, 383, 506, 429 ] }, { "type": "text", "bbox": [ 107, 436, 504, 460 ], "lines": [ { "bbox": [ 105, 436, 506, 450 ], "spans": [ { "bbox": [ 105, 436, 506, 450 ], "score": 1.0, "content": "Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In", "type": "text" } ], "index": 24 }, { "bbox": [ 116, 448, 227, 461 ], "spans": [ { "bbox": [ 116, 448, 227, 461 ], "score": 1.0, "content": "Proceedings of ACL, 2021.", "type": "text" } ], "index": 25 } ], "index": 24.5, "bbox_fs": [ 105, 436, 506, 461 ] }, { "type": "text", "bbox": [ 105, 468, 504, 491 ], "lines": [ { "bbox": [ 106, 468, 505, 481 ], "spans": [ { "bbox": [ 106, 468, 505, 481 ], "score": 1.0, "content": "Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization", "type": "text" } ], "index": 26 }, { "bbox": [ 116, 479, 201, 490 ], "spans": [ { "bbox": [ 116, 479, 201, 490 ], "score": 1.0, "content": "Branches Out, 2004.", "type": "text" } ], "index": 27 } ], "index": 26.5, "bbox_fs": [ 106, 468, 505, 490 ] }, { "type": "text", "bbox": [ 107, 498, 504, 533 ], "lines": [ { "bbox": [ 105, 498, 504, 512 ], "spans": [ { "bbox": [ 105, 498, 504, 512 ], "score": 1.0, "content": "Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 510, 505, 524 ], "spans": [ { "bbox": [ 115, 510, 505, 524 ], "score": 1.0, "content": "train, prompt, and predict: A systematic survey of prompting methods in natural language pro-", "type": "text" } ], "index": 29 }, { "bbox": [ 116, 522, 319, 533 ], "spans": [ { "bbox": [ 116, 522, 319, 533 ], "score": 1.0, "content": "cessing. arXiv preprint arXiv:2107.13586, 2021a.", "type": "text" } ], "index": 30 } ], "index": 29, "bbox_fs": [ 105, 498, 505, 533 ] }, { "type": "text", "bbox": [ 106, 541, 503, 564 ], "lines": [ { "bbox": [ 105, 540, 505, 554 ], "spans": [ { "bbox": [ 105, 540, 505, 554 ], "score": 1.0, "content": "Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. GPT", "type": "text" } ], "index": 31 }, { "bbox": [ 116, 552, 295, 563 ], "spans": [ { "bbox": [ 116, 552, 295, 563 ], "score": 1.0, "content": "understands, too. arXiv:2103.10385, 2021b.", "type": "text" } ], "index": 32 } ], "index": 31.5, "bbox_fs": [ 105, 540, 505, 563 ] }, { "type": "text", "bbox": [ 106, 572, 506, 606 ], "lines": [ { "bbox": [ 106, 572, 505, 585 ], "spans": [ { "bbox": [ 106, 572, 505, 585 ], "score": 1.0, "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 582, 506, 597 ], "spans": [ { "bbox": [ 115, 582, 506, 597 ], "score": 1.0, "content": "Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized bert pretraining", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 595, 321, 606 ], "spans": [ { "bbox": [ 115, 595, 321, 606 ], "score": 1.0, "content": "approach. arXiv preprint arXiv:1907.11692, 2019.", "type": "text" } ], "index": 35 } ], "index": 34, "bbox_fs": [ 106, 572, 506, 606 ] }, { "type": "text", "bbox": [ 107, 614, 504, 649 ], "lines": [ { "bbox": [ 105, 613, 506, 628 ], "spans": [ { "bbox": [ 105, 613, 506, 628 ], "score": 1.0, "content": "Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis,", "type": "text" } ], "index": 36 }, { "bbox": [ 116, 625, 506, 638 ], "spans": [ { "bbox": [ 116, 625, 506, 638 ], "score": 1.0, "content": "and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Trans-", "type": "text" } ], "index": 37 }, { "bbox": [ 116, 636, 376, 649 ], "spans": [ { "bbox": [ 116, 636, 376, 649 ], "score": 1.0, "content": "actions of the Association for Computational Linguistics, 2020a.", "type": "text" } ], "index": 38 } ], "index": 37, "bbox_fs": [ 105, 613, 506, 649 ] }, { "type": "text", "bbox": [ 107, 656, 505, 701 ], "lines": [ { "bbox": [ 106, 656, 505, 669 ], "spans": [ { "bbox": [ 106, 656, 505, 669 ], "score": 1.0, "content": "Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike", "type": "text" } ], "index": 39 }, { "bbox": [ 114, 667, 505, 680 ], "spans": [ { "bbox": [ 114, 667, 505, 680 ], "score": 1.0, "content": "Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine transla-", "type": "text" } ], "index": 40 }, { "bbox": [ 116, 679, 504, 691 ], "spans": [ { "bbox": [ 116, 679, 504, 691 ], "score": 1.0, "content": "tion. Transactions of the Association for Computational Linguistics, 8:726–742, 2020b. doi:", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 689, 470, 702 ], "spans": [ { "bbox": [ 115, 689, 470, 702 ], "score": 1.0, "content": "10.1162/tacl a 00343. URL https://aclanthology.org/2020.tacl-1.47.", "type": "text" } ], "index": 42 } ], "index": 40.5, "bbox_fs": [ 106, 656, 505, 702 ] }, { "type": "text", "bbox": [ 107, 709, 503, 732 ], "lines": [ { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 721, 375, 732 ], "spans": [ { "bbox": [ 115, 721, 375, 732 ], "score": 1.0, "content": "hypercomplex adapter layers. In Proceedings of NeurIPS, 2021.", "type": "text" } ], "index": 44 } ], "index": 43.5, "bbox_fs": [ 105, 709, 505, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 115 ], "lines": [ { "bbox": [ 106, 82, 504, 94 ], "spans": [ { "bbox": [ 106, 82, 504, 94 ], "score": 1.0, "content": "Shashi Narayan, Shay B. Cohen, and Mirella Lapata. Don’t give me the details, just the sum-", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 506, 107 ], "spans": [ { "bbox": [ 115, 93, 506, 107 ], "score": 1.0, "content": "mary! Topic-aware convolutional neural networks for extreme summarization. In Proceedings of", "type": "text" } ], "index": 1 }, { "bbox": [ 117, 104, 179, 116 ], "spans": [ { "bbox": [ 117, 104, 179, 116 ], "score": 1.0, "content": "EMNLP, 2018.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "text", "bbox": [ 108, 123, 504, 157 ], "lines": [ { "bbox": [ 106, 123, 505, 136 ], "spans": [ { "bbox": [ 106, 123, 505, 136 ], "score": 1.0, "content": "Jekaterina Novikova, Ondˇrej Dusek, and Verena Rieser. The E2E dataset: New challenges for ˇ", "type": "text" } ], "index": 3 }, { "bbox": [ 116, 135, 505, 147 ], "spans": [ { "bbox": [ 116, 135, 505, 147 ], "score": 1.0, "content": "end-to-end generation. In Proceedings of the 18th Annual SIGdial Meeting on Discourse and", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 145, 486, 157 ], "spans": [ { "bbox": [ 116, 145, 486, 157 ], "score": 1.0, "content": "Dialogue, pp. 201–206, Saarbrucken, Germany, August 2017. doi: 10.18653/v1/W17-5525. ¨", "type": "text" } ], "index": 5 } ], "index": 4 }, { "type": "text", "bbox": [ 106, 163, 502, 187 ], "lines": [ { "bbox": [ 106, 163, 504, 176 ], "spans": [ { "bbox": [ 106, 163, 504, 176 ], "score": 1.0, "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic", "type": "text" } ], "index": 6 }, { "bbox": [ 116, 174, 376, 187 ], "spans": [ { "bbox": [ 116, 174, 376, 187 ], "score": 1.0, "content": "evaluation of machine translation. In Proceedings of ACL, 2002.", "type": "text" } ], "index": 7 } ], "index": 6.5 }, { "type": "text", "bbox": [ 106, 193, 504, 217 ], "lines": [ { "bbox": [ 106, 194, 504, 206 ], "spans": [ { "bbox": [ 106, 194, 504, 206 ], "score": 1.0, "content": "Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and", "type": "text" } ], "index": 8 }, { "bbox": [ 116, 205, 500, 218 ], "spans": [ { "bbox": [ 116, 205, 500, 218 ], "score": 1.0, "content": "Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of NAACL, 2018.", "type": "text" } ], "index": 9 } ], "index": 8.5 }, { "type": "text", "bbox": [ 108, 223, 504, 247 ], "lines": [ { "bbox": [ 105, 223, 505, 237 ], "spans": [ { "bbox": [ 105, 223, 505, 237 ], "score": 1.0, "content": "Jonas Pfeiffer, Aishwarya Kamath, Andreas Ruckl ¨ e, Kyunghyun Cho, and Iryna Gurevych. Adapter- ´", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 235, 500, 248 ], "spans": [ { "bbox": [ 115, 235, 500, 248 ], "score": 1.0, "content": "Fusion: Non-destructive task composition for transfer learning. In Proceedings of EACL, 2021.", "type": "text" } ], "index": 11 } ], "index": 10.5 }, { "type": "text", "bbox": [ 107, 253, 504, 277 ], "lines": [ { "bbox": [ 106, 253, 504, 266 ], "spans": [ { "bbox": [ 106, 253, 504, 266 ], "score": 1.0, "content": "Xipeng Qiu, Tianxiang Sun, Yige Xu, Yunfan Shao, Ning Dai, and Xuanjing Huang. Pre-trained", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 264, 502, 278 ], "spans": [ { "bbox": [ 116, 264, 502, 278 ], "score": 1.0, "content": "models for natural language processing: A survey. Science China Technological Sciences, 2020.", "type": "text" } ], "index": 13 } ], "index": 12.5 }, { "type": "text", "bbox": [ 106, 283, 504, 307 ], "lines": [ { "bbox": [ 105, 282, 505, 298 ], "spans": [ { "bbox": [ 105, 282, 505, 298 ], "score": 1.0, "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 295, 376, 307 ], "spans": [ { "bbox": [ 115, 295, 376, 307 ], "score": 1.0, "content": "models are unsupervised multitask learners. OpenAI blog, 2019.", "type": "text" } ], "index": 15 } ], "index": 14.5 }, { "type": "text", "bbox": [ 108, 313, 504, 347 ], "lines": [ { "bbox": [ 105, 312, 505, 327 ], "spans": [ { "bbox": [ 105, 312, 505, 327 ], "score": 1.0, "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 324, 505, 337 ], "spans": [ { "bbox": [ 115, 324, 505, 337 ], "score": 1.0, "content": "Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 336, 354, 348 ], "spans": [ { "bbox": [ 116, 336, 354, 348 ], "score": 1.0, "content": "transformer. Journal of Machine Learning Research, 2020.", "type": "text" } ], "index": 18 } ], "index": 17 }, { "type": "text", "bbox": [ 107, 354, 503, 388 ], "lines": [ { "bbox": [ 105, 352, 502, 369 ], "spans": [ { "bbox": [ 105, 352, 477, 369 ], "score": 1.0, "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew", "type": "text" }, { "bbox": [ 477, 354, 502, 366 ], "score": 0.34, "content": "\\mathrm { ~ Y ~ N ~ g ~ } _ { }", "type": "inline_equation" } ], "index": 19 }, { "bbox": [ 116, 365, 505, 378 ], "spans": [ { "bbox": [ 116, 365, 505, 378 ], "score": 1.0, "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", "type": "text" } ], "index": 20 }, { "bbox": [ 117, 377, 292, 388 ], "spans": [ { "bbox": [ 117, 377, 292, 388 ], "score": 1.0, "content": "treebank. In Proceedings of EMNLP, 2013.", "type": "text" } ], "index": 21 } ], "index": 20 }, { "type": "text", "bbox": [ 106, 395, 504, 418 ], "lines": [ { "bbox": [ 105, 393, 506, 410 ], "spans": [ { "bbox": [ 105, 393, 506, 410 ], "score": 1.0, "content": "Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 406, 308, 419 ], "spans": [ { "bbox": [ 116, 406, 308, 419 ], "score": 1.0, "content": "learning in NLP. In Proceedings of ACL, 2019.", "type": "text" } ], "index": 23 } ], "index": 22.5 }, { "type": "text", "bbox": [ 107, 425, 504, 458 ], "lines": [ { "bbox": [ 105, 424, 506, 438 ], "spans": [ { "bbox": [ 105, 424, 506, 438 ], "score": 1.0, "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 435, 505, 448 ], "spans": [ { "bbox": [ 115, 435, 505, 448 ], "score": 1.0, "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of NeurIPS,", "type": "text" } ], "index": 25 }, { "bbox": [ 114, 445, 144, 459 ], "spans": [ { "bbox": [ 114, 445, 144, 459 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 26 } ], "index": 25 }, { "type": "text", "bbox": [ 106, 466, 504, 489 ], "lines": [ { "bbox": [ 105, 464, 505, 480 ], "spans": [ { "bbox": [ 105, 464, 505, 480 ], "score": 1.0, "content": "Adina Williams, Nikita Nangia, and Samuel Bowman. A broad-coverage challenge corpus for sen-", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 477, 409, 489 ], "spans": [ { "bbox": [ 115, 477, 409, 489 ], "score": 1.0, "content": "tence understanding through inference. In Proceedings of NAACL, 2018.", "type": "text" } ], "index": 28 } ], "index": 27.5 }, { "type": "text", "bbox": [ 106, 496, 504, 551 ], "lines": [ { "bbox": [ 106, 496, 505, 508 ], "spans": [ { "bbox": [ 106, 496, 505, 508 ], "score": 1.0, "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi,", "type": "text" } ], "index": 29 }, { "bbox": [ 115, 506, 506, 519 ], "spans": [ { "bbox": [ 115, 506, 506, 519 ], "score": 1.0, "content": "Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick ´", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 516, 506, 531 ], "spans": [ { "bbox": [ 115, 516, 506, 531 ], "score": 1.0, "content": "von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger,", "type": "text" } ], "index": 31 }, { "bbox": [ 116, 529, 505, 541 ], "spans": [ { "bbox": [ 116, 529, 505, 541 ], "score": 1.0, "content": "Mariama Drame, Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the-art natural", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 540, 438, 551 ], "spans": [ { "bbox": [ 115, 540, 438, 551 ], "score": 1.0, "content": "language processing. In Proceedings of EMNLP: System Demonstrations, 2020.", "type": "text" } ], "index": 33 } ], "index": 31 }, { "type": "text", "bbox": [ 106, 558, 503, 581 ], "lines": [ { "bbox": [ 106, 556, 505, 573 ], "spans": [ { "bbox": [ 106, 556, 505, 573 ], "score": 1.0, "content": "Yaoming Zhu, Jiangtao Feng, Chengqi Zhao, Mingxuan Wang, and Lei Li. Serial or parallel? plug-", "type": "text" } ], "index": 34 }, { "bbox": [ 116, 569, 480, 582 ], "spans": [ { "bbox": [ 116, 569, 480, 582 ], "score": 1.0, "content": "able adapter for multilingual machine translation. arXiv preprint arXiv:2104.08154, 2021.", "type": "text" } ], "index": 35 } ], "index": 34.5 } ], "page_idx": 11, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "12", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 115 ], "lines": [ { "bbox": [ 106, 82, 504, 94 ], "spans": [ { "bbox": [ 106, 82, 504, 94 ], "score": 1.0, "content": "Shashi Narayan, Shay B. Cohen, and Mirella Lapata. Don’t give me the details, just the sum-", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 506, 107 ], "spans": [ { "bbox": [ 115, 93, 506, 107 ], "score": 1.0, "content": "mary! Topic-aware convolutional neural networks for extreme summarization. In Proceedings of", "type": "text" } ], "index": 1 }, { "bbox": [ 117, 104, 179, 116 ], "spans": [ { "bbox": [ 117, 104, 179, 116 ], "score": 1.0, "content": "EMNLP, 2018.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 106, 82, 506, 116 ] }, { "type": "text", "bbox": [ 108, 123, 504, 157 ], "lines": [ { "bbox": [ 106, 123, 505, 136 ], "spans": [ { "bbox": [ 106, 123, 505, 136 ], "score": 1.0, "content": "Jekaterina Novikova, Ondˇrej Dusek, and Verena Rieser. The E2E dataset: New challenges for ˇ", "type": "text" } ], "index": 3 }, { "bbox": [ 116, 135, 505, 147 ], "spans": [ { "bbox": [ 116, 135, 505, 147 ], "score": 1.0, "content": "end-to-end generation. In Proceedings of the 18th Annual SIGdial Meeting on Discourse and", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 145, 486, 157 ], "spans": [ { "bbox": [ 116, 145, 486, 157 ], "score": 1.0, "content": "Dialogue, pp. 201–206, Saarbrucken, Germany, August 2017. doi: 10.18653/v1/W17-5525. ¨", "type": "text" } ], "index": 5 } ], "index": 4, "bbox_fs": [ 106, 123, 505, 157 ] }, { "type": "text", "bbox": [ 106, 163, 502, 187 ], "lines": [ { "bbox": [ 106, 163, 504, 176 ], "spans": [ { "bbox": [ 106, 163, 504, 176 ], "score": 1.0, "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic", "type": "text" } ], "index": 6 }, { "bbox": [ 116, 174, 376, 187 ], "spans": [ { "bbox": [ 116, 174, 376, 187 ], "score": 1.0, "content": "evaluation of machine translation. In Proceedings of ACL, 2002.", "type": "text" } ], "index": 7 } ], "index": 6.5, "bbox_fs": [ 106, 163, 504, 187 ] }, { "type": "text", "bbox": [ 106, 193, 504, 217 ], "lines": [ { "bbox": [ 106, 194, 504, 206 ], "spans": [ { "bbox": [ 106, 194, 504, 206 ], "score": 1.0, "content": "Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and", "type": "text" } ], "index": 8 }, { "bbox": [ 116, 205, 500, 218 ], "spans": [ { "bbox": [ 116, 205, 500, 218 ], "score": 1.0, "content": "Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of NAACL, 2018.", "type": "text" } ], "index": 9 } ], "index": 8.5, "bbox_fs": [ 106, 194, 504, 218 ] }, { "type": "text", "bbox": [ 108, 223, 504, 247 ], "lines": [ { "bbox": [ 105, 223, 505, 237 ], "spans": [ { "bbox": [ 105, 223, 505, 237 ], "score": 1.0, "content": "Jonas Pfeiffer, Aishwarya Kamath, Andreas Ruckl ¨ e, Kyunghyun Cho, and Iryna Gurevych. Adapter- ´", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 235, 500, 248 ], "spans": [ { "bbox": [ 115, 235, 500, 248 ], "score": 1.0, "content": "Fusion: Non-destructive task composition for transfer learning. In Proceedings of EACL, 2021.", "type": "text" } ], "index": 11 } ], "index": 10.5, "bbox_fs": [ 105, 223, 505, 248 ] }, { "type": "text", "bbox": [ 107, 253, 504, 277 ], "lines": [ { "bbox": [ 106, 253, 504, 266 ], "spans": [ { "bbox": [ 106, 253, 504, 266 ], "score": 1.0, "content": "Xipeng Qiu, Tianxiang Sun, Yige Xu, Yunfan Shao, Ning Dai, and Xuanjing Huang. Pre-trained", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 264, 502, 278 ], "spans": [ { "bbox": [ 116, 264, 502, 278 ], "score": 1.0, "content": "models for natural language processing: A survey. Science China Technological Sciences, 2020.", "type": "text" } ], "index": 13 } ], "index": 12.5, "bbox_fs": [ 106, 253, 504, 278 ] }, { "type": "text", "bbox": [ 106, 283, 504, 307 ], "lines": [ { "bbox": [ 105, 282, 505, 298 ], "spans": [ { "bbox": [ 105, 282, 505, 298 ], "score": 1.0, "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 295, 376, 307 ], "spans": [ { "bbox": [ 115, 295, 376, 307 ], "score": 1.0, "content": "models are unsupervised multitask learners. OpenAI blog, 2019.", "type": "text" } ], "index": 15 } ], "index": 14.5, "bbox_fs": [ 105, 282, 505, 307 ] }, { "type": "text", "bbox": [ 108, 313, 504, 347 ], "lines": [ { "bbox": [ 105, 312, 505, 327 ], "spans": [ { "bbox": [ 105, 312, 505, 327 ], "score": 1.0, "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 324, 505, 337 ], "spans": [ { "bbox": [ 115, 324, 505, 337 ], "score": 1.0, "content": "Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 336, 354, 348 ], "spans": [ { "bbox": [ 116, 336, 354, 348 ], "score": 1.0, "content": "transformer. Journal of Machine Learning Research, 2020.", "type": "text" } ], "index": 18 } ], "index": 17, "bbox_fs": [ 105, 312, 505, 348 ] }, { "type": "text", "bbox": [ 107, 354, 503, 388 ], "lines": [ { "bbox": [ 105, 352, 502, 369 ], "spans": [ { "bbox": [ 105, 352, 477, 369 ], "score": 1.0, "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew", "type": "text" }, { "bbox": [ 477, 354, 502, 366 ], "score": 0.34, "content": "\\mathrm { ~ Y ~ N ~ g ~ } _ { }", "type": "inline_equation" } ], "index": 19 }, { "bbox": [ 116, 365, 505, 378 ], "spans": [ { "bbox": [ 116, 365, 505, 378 ], "score": 1.0, "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", "type": "text" } ], "index": 20 }, { "bbox": [ 117, 377, 292, 388 ], "spans": [ { "bbox": [ 117, 377, 292, 388 ], "score": 1.0, "content": "treebank. In Proceedings of EMNLP, 2013.", "type": "text" } ], "index": 21 } ], "index": 20, "bbox_fs": [ 105, 352, 505, 388 ] }, { "type": "text", "bbox": [ 106, 395, 504, 418 ], "lines": [ { "bbox": [ 105, 393, 506, 410 ], "spans": [ { "bbox": [ 105, 393, 506, 410 ], "score": 1.0, "content": "Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 406, 308, 419 ], "spans": [ { "bbox": [ 116, 406, 308, 419 ], "score": 1.0, "content": "learning in NLP. In Proceedings of ACL, 2019.", "type": "text" } ], "index": 23 } ], "index": 22.5, "bbox_fs": [ 105, 393, 506, 419 ] }, { "type": "text", "bbox": [ 107, 425, 504, 458 ], "lines": [ { "bbox": [ 105, 424, 506, 438 ], "spans": [ { "bbox": [ 105, 424, 506, 438 ], "score": 1.0, "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 435, 505, 448 ], "spans": [ { "bbox": [ 115, 435, 505, 448 ], "score": 1.0, "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of NeurIPS,", "type": "text" } ], "index": 25 }, { "bbox": [ 114, 445, 144, 459 ], "spans": [ { "bbox": [ 114, 445, 144, 459 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 26 } ], "index": 25, "bbox_fs": [ 105, 424, 506, 459 ] }, { "type": "text", "bbox": [ 106, 466, 504, 489 ], "lines": [ { "bbox": [ 105, 464, 505, 480 ], "spans": [ { "bbox": [ 105, 464, 505, 480 ], "score": 1.0, "content": "Adina Williams, Nikita Nangia, and Samuel Bowman. A broad-coverage challenge corpus for sen-", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 477, 409, 489 ], "spans": [ { "bbox": [ 115, 477, 409, 489 ], "score": 1.0, "content": "tence understanding through inference. In Proceedings of NAACL, 2018.", "type": "text" } ], "index": 28 } ], "index": 27.5, "bbox_fs": [ 105, 464, 505, 489 ] }, { "type": "text", "bbox": [ 106, 496, 504, 551 ], "lines": [ { "bbox": [ 106, 496, 505, 508 ], "spans": [ { "bbox": [ 106, 496, 505, 508 ], "score": 1.0, "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi,", "type": "text" } ], "index": 29 }, { "bbox": [ 115, 506, 506, 519 ], "spans": [ { "bbox": [ 115, 506, 506, 519 ], "score": 1.0, "content": "Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick ´", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 516, 506, 531 ], "spans": [ { "bbox": [ 115, 516, 506, 531 ], "score": 1.0, "content": "von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger,", "type": "text" } ], "index": 31 }, { "bbox": [ 116, 529, 505, 541 ], "spans": [ { "bbox": [ 116, 529, 505, 541 ], "score": 1.0, "content": "Mariama Drame, Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the-art natural", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 540, 438, 551 ], "spans": [ { "bbox": [ 115, 540, 438, 551 ], "score": 1.0, "content": "language processing. In Proceedings of EMNLP: System Demonstrations, 2020.", "type": "text" } ], "index": 33 } ], "index": 31, "bbox_fs": [ 106, 496, 506, 551 ] }, { "type": "text", "bbox": [ 106, 558, 503, 581 ], "lines": [ { "bbox": [ 106, 556, 505, 573 ], "spans": [ { "bbox": [ 106, 556, 505, 573 ], "score": 1.0, "content": "Yaoming Zhu, Jiangtao Feng, Chengqi Zhao, Mingxuan Wang, and Lei Li. Serial or parallel? plug-", "type": "text" } ], "index": 34 }, { "bbox": [ 116, 569, 480, 582 ], "spans": [ { "bbox": [ 116, 569, 480, 582 ], "score": 1.0, "content": "able adapter for multilingual machine translation. arXiv preprint arXiv:2104.08154, 2021.", "type": "text" } ], "index": 35 } ], "index": 34.5, "bbox_fs": [ 106, 556, 505, 582 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 81, 203, 94 ], "lines": [ { "bbox": [ 106, 81, 204, 95 ], "spans": [ { "bbox": [ 106, 81, 204, 95 ], "score": 1.0, "content": "A EXPERIMENTS", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 110, 167, 122 ], "lines": [ { "bbox": [ 106, 110, 168, 123 ], "spans": [ { "bbox": [ 106, 110, 168, 123 ], "score": 1.0, "content": "A.1 SETUPS", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "table", "bbox": [ 201, 164, 407, 232 ], "blocks": [ { "type": "table_caption", "bbox": [ 227, 144, 383, 155 ], "group_id": 0, "lines": [ { "bbox": [ 226, 144, 384, 155 ], "spans": [ { "bbox": [ 226, 144, 384, 155 ], "score": 1.0, "content": "Table 7: Dataset Statistics of the four tasks.", "type": "text" } ], "index": 2 } ], "index": 2 }, { "type": "table_body", "bbox": [ 201, 164, 407, 232 ], "group_id": 0, "lines": [ { "bbox": [ 201, 164, 407, 232 ], "spans": [ { "bbox": [ 201, 164, 407, 232 ], "score": 0.98, "html": "
Dataset#train#dev#test
XSum204,045113,332113,334
WMT16 en-ro610,3201,9991,999
MNLI392,70298159832
SST-267,3498721,821
", "type": "table", "image_path": "0ba334a0921a08a0c6a1b31b1f0b22359bad748b65df0e005d128bc6735499b0.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 201, 164, 407, 177.6 ], "spans": [], "index": 3 }, { "bbox": [ 201, 177.6, 407, 191.2 ], "spans": [], "index": 4 }, { "bbox": [ 201, 191.2, 407, 204.79999999999998 ], "spans": [], "index": 5 }, { "bbox": [ 201, 204.79999999999998, 407, 218.39999999999998 ], "spans": [], "index": 6 }, { "bbox": [ 201, 218.39999999999998, 407, 231.99999999999997 ], "spans": [], "index": 7 } ] } ], "index": 3.5 }, { "type": "text", "bbox": [ 107, 246, 505, 357 ], "lines": [ { "bbox": [ 106, 247, 505, 259 ], "spans": [ { "bbox": [ 106, 247, 505, 259 ], "score": 1.0, "content": "We implement all the parameter-efficient tuning methods using the huggingface transformers li-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 257, 505, 271 ], "spans": [ { "bbox": [ 105, 257, 505, 271 ], "score": 1.0, "content": "brary (Wolf et al., 2020). We use BARTLARGE(Lewis et al., 2020) and mBARTLARGE (Liu et al.,", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 268, 506, 282 ], "spans": [ { "bbox": [ 105, 268, 506, 282 ], "score": 1.0, "content": "2020b) (mBART-cc25) for the summarization and machine translation tasks respectively, and we", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 279, 506, 293 ], "spans": [ { "bbox": [ 105, 279, 506, 293 ], "score": 1.0, "content": "use RoBERTaBASE (Liu et al., 2019) for MNLI and SST2. BARTLARGE and mBARTLARGE have the", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 291, 505, 303 ], "spans": [ { "bbox": [ 105, 291, 505, 303 ], "score": 1.0, "content": "same encoder-decoder architectures. mBARTLARGE is pre-trained on 25 languages. We use their", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 302, 505, 315 ], "spans": [ { "bbox": [ 105, 302, 505, 315 ], "score": 1.0, "content": "public checkpoints from the transformers library in experiments. For MT and classifications tasks,", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 312, 506, 325 ], "spans": [ { "bbox": [ 105, 312, 506, 325 ], "score": 1.0, "content": "the max token lengths of training data are set to be 150 and 512 respectively. For XSum, we set the", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 324, 505, 336 ], "spans": [ { "bbox": [ 105, 324, 505, 336 ], "score": 1.0, "content": "max length of source articles to be 512 and the max length of the target summary to be 128. The", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 334, 505, 347 ], "spans": [ { "bbox": [ 105, 334, 505, 347 ], "score": 1.0, "content": "detailed dataset statistics is present in Table 7. In our summarization experiments, we only use 1600", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 346, 256, 358 ], "spans": [ { "bbox": [ 106, 346, 256, 358 ], "score": 1.0, "content": "examples for validation to save time.", "type": "text" } ], "index": 17 } ], "index": 12.5 }, { "type": "text", "bbox": [ 107, 362, 505, 428 ], "lines": [ { "bbox": [ 106, 362, 506, 375 ], "spans": [ { "bbox": [ 106, 362, 305, 375 ], "score": 1.0, "content": "While we vary the bottleneck dimension within", "type": "text" }, { "bbox": [ 306, 362, 379, 375 ], "score": 0.9, "content": "\\{ 1 , 3 0 , 5 1 2 , 1 0 2 4 \\}", "type": "inline_equation" }, { "bbox": [ 380, 362, 450, 375 ], "score": 1.0, "content": "as mentioned in", "type": "text" }, { "bbox": [ 450, 363, 468, 374 ], "score": 0.85, "content": "\\ S 4 . 1", "type": "inline_equation" }, { "bbox": [ 469, 362, 506, 375 ], "score": 1.0, "content": ", we test", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 372, 506, 386 ], "spans": [ { "bbox": [ 105, 372, 506, 386 ], "score": 1.0, "content": "bottleneck dimension 1024 only when the modified representation is FFN, because the training of", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 384, 506, 397 ], "spans": [ { "bbox": [ 105, 384, 330, 397 ], "score": 1.0, "content": "prefix tuning does not fit into 48GB GPU memory when", "type": "text" }, { "bbox": [ 330, 385, 368, 395 ], "score": 0.89, "content": "l = 1 0 2 4", "type": "inline_equation" }, { "bbox": [ 368, 384, 506, 397 ], "score": 1.0, "content": ". While other methods do not have", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 395, 506, 408 ], "spans": [ { "bbox": [ 105, 395, 506, 408 ], "score": 1.0, "content": "memory issues, we keep the bottleneck dimension of attention modification at most 512 to have a", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 407, 505, 419 ], "spans": [ { "bbox": [ 105, 407, 505, 419 ], "score": 1.0, "content": "relatively fair comparison with prefix tuning. For LoRA we always tune its scaling hyperparameters", "type": "text" } ], "index": 22 }, { "bbox": [ 107, 418, 174, 429 ], "spans": [ { "bbox": [ 107, 420, 112, 427 ], "score": 0.65, "content": "s", "type": "inline_equation" }, { "bbox": [ 113, 418, 174, 429 ], "score": 1.0, "content": "on the dev set.", "type": "text" } ], "index": 23 } ], "index": 20.5 }, { "type": "title", "bbox": [ 108, 447, 257, 458 ], "lines": [ { "bbox": [ 106, 447, 258, 460 ], "spans": [ { "bbox": [ 106, 447, 258, 460 ], "score": 1.0, "content": "A.2 TRAINING AND EVALUATION", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 469, 505, 569 ], "lines": [ { "bbox": [ 106, 470, 506, 483 ], "spans": [ { "bbox": [ 106, 470, 506, 483 ], "score": 1.0, "content": "We present some training hyperparameters of parameter-efficient tuning methods in Table 8. For all", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 479, 506, 494 ], "spans": [ { "bbox": [ 105, 479, 506, 494 ], "score": 1.0, "content": "the tasks, we train with the Adam optimizer (Kingma & Ba, 2015), and use a polynomial learning", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 491, 506, 505 ], "spans": [ { "bbox": [ 105, 491, 506, 505 ], "score": 1.0, "content": "rate scheduler that linearly decays the learning rate throughout training. We set the warm up steps of", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 501, 506, 517 ], "spans": [ { "bbox": [ 105, 501, 506, 517 ], "score": 1.0, "content": "learning rate to be 0 for both MT and summarization tasks, and for the classification tasks, learning", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 513, 505, 526 ], "spans": [ { "bbox": [ 105, 513, 293, 526 ], "score": 1.0, "content": "rate is linearly warmed up from 0 for the first", "type": "text" }, { "bbox": [ 294, 514, 308, 524 ], "score": 0.85, "content": "6 \\%", "type": "inline_equation" }, { "bbox": [ 309, 513, 505, 526 ], "score": 1.0, "content": "of the total training steps before decay. For full", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 524, 505, 537 ], "spans": [ { "bbox": [ 105, 524, 505, 537 ], "score": 1.0, "content": "fine-tuning we set these training hyperparameters following Lewis et al. (2020) (XSum), Liu et al.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 535, 505, 548 ], "spans": [ { "bbox": [ 105, 535, 505, 548 ], "score": 1.0, "content": "(2020b) (en-ro), and (Liu et al., 2019) (MNLI and SST2). We also did hyperparameter search in the", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 546, 505, 559 ], "spans": [ { "bbox": [ 105, 546, 505, 559 ], "score": 1.0, "content": "full fine-tuning case to try to reproduce their results. We set dropout rate to be 0.1 for all the tasks.", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 556, 504, 571 ], "spans": [ { "bbox": [ 105, 556, 504, 571 ], "score": 1.0, "content": "We use ROUGE-2 and perplexity as the validation metrics for summarization and MT respectively.", "type": "text" } ], "index": 33 } ], "index": 29 }, { "type": "text", "bbox": [ 107, 574, 505, 608 ], "lines": [ { "bbox": [ 106, 574, 505, 586 ], "spans": [ { "bbox": [ 106, 574, 505, 586 ], "score": 1.0, "content": "For MT and text summarization, we use beam search for decoding and set the number of beams to be", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 584, 505, 599 ], "spans": [ { "bbox": [ 105, 584, 505, 599 ], "score": 1.0, "content": "6 and 5 following previous work (Li & Liang, 2021; Liu et al., 2020b). The min and max generation", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 596, 434, 609 ], "spans": [ { "bbox": [ 106, 596, 434, 609 ], "score": 1.0, "content": "lengths for summarization and MT are set to be (10, 60) and (1, 200) respectively.", "type": "text" } ], "index": 36 } ], "index": 35 }, { "type": "title", "bbox": [ 109, 626, 273, 637 ], "lines": [ { "bbox": [ 106, 624, 275, 639 ], "spans": [ { "bbox": [ 106, 624, 275, 639 ], "score": 1.0, "content": "A.3 OTHER EXPERIMENTAL DETAILS", "type": "text" } ], "index": 37 } ], "index": 37 }, { "type": "text", "bbox": [ 107, 648, 504, 682 ], "lines": [ { "bbox": [ 105, 648, 506, 662 ], "spans": [ { "bbox": [ 105, 648, 506, 662 ], "score": 1.0, "content": "Prefix Tuning: Following Li & Liang (2021), we reparameterize the prefix vectors by a MLP", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 659, 504, 672 ], "spans": [ { "bbox": [ 106, 659, 504, 672 ], "score": 1.0, "content": "network which is composed of a small embedding matrix and a large feedforward neural network.", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 670, 414, 684 ], "spans": [ { "bbox": [ 105, 670, 414, 684 ], "score": 1.0, "content": "This is conducive for learning due to the shared parameters across all layers.", "type": "text" } ], "index": 40 } ], "index": 39 }, { "type": "text", "bbox": [ 107, 687, 504, 731 ], "lines": [ { "bbox": [ 105, 686, 505, 700 ], "spans": [ { "bbox": [ 105, 686, 505, 700 ], "score": 1.0, "content": "LoRA: LoRA and adapter employ different parameter initialization methods: LoRA uses a ran-", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 699, 505, 712 ], "spans": [ { "bbox": [ 105, 699, 342, 712 ], "score": 1.0, "content": "dom Kaiming uniform (He et al., 2015) initialization for", "type": "text" }, { "bbox": [ 342, 699, 372, 710 ], "score": 0.9, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation" }, { "bbox": [ 373, 699, 428, 712 ], "score": 1.0, "content": "and zero for", "type": "text" }, { "bbox": [ 429, 699, 450, 711 ], "score": 0.89, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation" }, { "bbox": [ 450, 699, 505, 712 ], "score": 1.0, "content": "(LoRA init),", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "while adapters use the same initialization as BERT (Devlin et al., 2019). We found it beneficial to", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 721, 336, 731 ], "spans": [ { "bbox": [ 106, 721, 336, 731 ], "score": 1.0, "content": "use the same initialization method as LoRA in scaled PA.", "type": "text" } ], "index": 44 } ], "index": 42.5 } ], "page_idx": 12, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 312, 763 ], "spans": [ { "bbox": [ 298, 750, 312, 763 ], "score": 1.0, "content": "", "type": "text", "height": 13, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 81, 203, 94 ], "lines": [ { "bbox": [ 106, 81, 204, 95 ], "spans": [ { "bbox": [ 106, 81, 204, 95 ], "score": 1.0, "content": "A EXPERIMENTS", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 110, 167, 122 ], "lines": [ { "bbox": [ 106, 110, 168, 123 ], "spans": [ { "bbox": [ 106, 110, 168, 123 ], "score": 1.0, "content": "A.1 SETUPS", "type": "text" } ], "index": 1 } ], "index": 1, "bbox_fs": [ 106, 110, 168, 123 ] }, { "type": "table", "bbox": [ 201, 164, 407, 232 ], "blocks": [ { "type": "table_caption", "bbox": [ 227, 144, 383, 155 ], "group_id": 0, "lines": [ { "bbox": [ 226, 144, 384, 155 ], "spans": [ { "bbox": [ 226, 144, 384, 155 ], "score": 1.0, "content": "Table 7: Dataset Statistics of the four tasks.", "type": "text" } ], "index": 2 } ], "index": 2 }, { "type": "table_body", "bbox": [ 201, 164, 407, 232 ], "group_id": 0, "lines": [ { "bbox": [ 201, 164, 407, 232 ], "spans": [ { "bbox": [ 201, 164, 407, 232 ], "score": 0.98, "html": "
Dataset#train#dev#test
XSum204,045113,332113,334
WMT16 en-ro610,3201,9991,999
MNLI392,70298159832
SST-267,3498721,821
", "type": "table", "image_path": "0ba334a0921a08a0c6a1b31b1f0b22359bad748b65df0e005d128bc6735499b0.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 201, 164, 407, 177.6 ], "spans": [], "index": 3 }, { "bbox": [ 201, 177.6, 407, 191.2 ], "spans": [], "index": 4 }, { "bbox": [ 201, 191.2, 407, 204.79999999999998 ], "spans": [], "index": 5 }, { "bbox": [ 201, 204.79999999999998, 407, 218.39999999999998 ], "spans": [], "index": 6 }, { "bbox": [ 201, 218.39999999999998, 407, 231.99999999999997 ], "spans": [], "index": 7 } ] } ], "index": 3.5 }, { "type": "text", "bbox": [ 107, 246, 505, 357 ], "lines": [ { "bbox": [ 106, 247, 505, 259 ], "spans": [ { "bbox": [ 106, 247, 505, 259 ], "score": 1.0, "content": "We implement all the parameter-efficient tuning methods using the huggingface transformers li-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 257, 505, 271 ], "spans": [ { "bbox": [ 105, 257, 505, 271 ], "score": 1.0, "content": "brary (Wolf et al., 2020). We use BARTLARGE(Lewis et al., 2020) and mBARTLARGE (Liu et al.,", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 268, 506, 282 ], "spans": [ { "bbox": [ 105, 268, 506, 282 ], "score": 1.0, "content": "2020b) (mBART-cc25) for the summarization and machine translation tasks respectively, and we", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 279, 506, 293 ], "spans": [ { "bbox": [ 105, 279, 506, 293 ], "score": 1.0, "content": "use RoBERTaBASE (Liu et al., 2019) for MNLI and SST2. BARTLARGE and mBARTLARGE have the", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 291, 505, 303 ], "spans": [ { "bbox": [ 105, 291, 505, 303 ], "score": 1.0, "content": "same encoder-decoder architectures. mBARTLARGE is pre-trained on 25 languages. We use their", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 302, 505, 315 ], "spans": [ { "bbox": [ 105, 302, 505, 315 ], "score": 1.0, "content": "public checkpoints from the transformers library in experiments. For MT and classifications tasks,", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 312, 506, 325 ], "spans": [ { "bbox": [ 105, 312, 506, 325 ], "score": 1.0, "content": "the max token lengths of training data are set to be 150 and 512 respectively. For XSum, we set the", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 324, 505, 336 ], "spans": [ { "bbox": [ 105, 324, 505, 336 ], "score": 1.0, "content": "max length of source articles to be 512 and the max length of the target summary to be 128. The", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 334, 505, 347 ], "spans": [ { "bbox": [ 105, 334, 505, 347 ], "score": 1.0, "content": "detailed dataset statistics is present in Table 7. In our summarization experiments, we only use 1600", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 346, 256, 358 ], "spans": [ { "bbox": [ 106, 346, 256, 358 ], "score": 1.0, "content": "examples for validation to save time.", "type": "text" } ], "index": 17 } ], "index": 12.5, "bbox_fs": [ 105, 247, 506, 358 ] }, { "type": "text", "bbox": [ 107, 362, 505, 428 ], "lines": [ { "bbox": [ 106, 362, 506, 375 ], "spans": [ { "bbox": [ 106, 362, 305, 375 ], "score": 1.0, "content": "While we vary the bottleneck dimension within", "type": "text" }, { "bbox": [ 306, 362, 379, 375 ], "score": 0.9, "content": "\\{ 1 , 3 0 , 5 1 2 , 1 0 2 4 \\}", "type": "inline_equation" }, { "bbox": [ 380, 362, 450, 375 ], "score": 1.0, "content": "as mentioned in", "type": "text" }, { "bbox": [ 450, 363, 468, 374 ], "score": 0.85, "content": "\\ S 4 . 1", "type": "inline_equation" }, { "bbox": [ 469, 362, 506, 375 ], "score": 1.0, "content": ", we test", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 372, 506, 386 ], "spans": [ { "bbox": [ 105, 372, 506, 386 ], "score": 1.0, "content": "bottleneck dimension 1024 only when the modified representation is FFN, because the training of", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 384, 506, 397 ], "spans": [ { "bbox": [ 105, 384, 330, 397 ], "score": 1.0, "content": "prefix tuning does not fit into 48GB GPU memory when", "type": "text" }, { "bbox": [ 330, 385, 368, 395 ], "score": 0.89, "content": "l = 1 0 2 4", "type": "inline_equation" }, { "bbox": [ 368, 384, 506, 397 ], "score": 1.0, "content": ". While other methods do not have", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 395, 506, 408 ], "spans": [ { "bbox": [ 105, 395, 506, 408 ], "score": 1.0, "content": "memory issues, we keep the bottleneck dimension of attention modification at most 512 to have a", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 407, 505, 419 ], "spans": [ { "bbox": [ 105, 407, 505, 419 ], "score": 1.0, "content": "relatively fair comparison with prefix tuning. For LoRA we always tune its scaling hyperparameters", "type": "text" } ], "index": 22 }, { "bbox": [ 107, 418, 174, 429 ], "spans": [ { "bbox": [ 107, 420, 112, 427 ], "score": 0.65, "content": "s", "type": "inline_equation" }, { "bbox": [ 113, 418, 174, 429 ], "score": 1.0, "content": "on the dev set.", "type": "text" } ], "index": 23 } ], "index": 20.5, "bbox_fs": [ 105, 362, 506, 429 ] }, { "type": "title", "bbox": [ 108, 447, 257, 458 ], "lines": [ { "bbox": [ 106, 447, 258, 460 ], "spans": [ { "bbox": [ 106, 447, 258, 460 ], "score": 1.0, "content": "A.2 TRAINING AND EVALUATION", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 469, 505, 569 ], "lines": [ { "bbox": [ 106, 470, 506, 483 ], "spans": [ { "bbox": [ 106, 470, 506, 483 ], "score": 1.0, "content": "We present some training hyperparameters of parameter-efficient tuning methods in Table 8. For all", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 479, 506, 494 ], "spans": [ { "bbox": [ 105, 479, 506, 494 ], "score": 1.0, "content": "the tasks, we train with the Adam optimizer (Kingma & Ba, 2015), and use a polynomial learning", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 491, 506, 505 ], "spans": [ { "bbox": [ 105, 491, 506, 505 ], "score": 1.0, "content": "rate scheduler that linearly decays the learning rate throughout training. We set the warm up steps of", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 501, 506, 517 ], "spans": [ { "bbox": [ 105, 501, 506, 517 ], "score": 1.0, "content": "learning rate to be 0 for both MT and summarization tasks, and for the classification tasks, learning", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 513, 505, 526 ], "spans": [ { "bbox": [ 105, 513, 293, 526 ], "score": 1.0, "content": "rate is linearly warmed up from 0 for the first", "type": "text" }, { "bbox": [ 294, 514, 308, 524 ], "score": 0.85, "content": "6 \\%", "type": "inline_equation" }, { "bbox": [ 309, 513, 505, 526 ], "score": 1.0, "content": "of the total training steps before decay. For full", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 524, 505, 537 ], "spans": [ { "bbox": [ 105, 524, 505, 537 ], "score": 1.0, "content": "fine-tuning we set these training hyperparameters following Lewis et al. (2020) (XSum), Liu et al.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 535, 505, 548 ], "spans": [ { "bbox": [ 105, 535, 505, 548 ], "score": 1.0, "content": "(2020b) (en-ro), and (Liu et al., 2019) (MNLI and SST2). We also did hyperparameter search in the", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 546, 505, 559 ], "spans": [ { "bbox": [ 105, 546, 505, 559 ], "score": 1.0, "content": "full fine-tuning case to try to reproduce their results. We set dropout rate to be 0.1 for all the tasks.", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 556, 504, 571 ], "spans": [ { "bbox": [ 105, 556, 504, 571 ], "score": 1.0, "content": "We use ROUGE-2 and perplexity as the validation metrics for summarization and MT respectively.", "type": "text" } ], "index": 33 } ], "index": 29, "bbox_fs": [ 105, 470, 506, 571 ] }, { "type": "text", "bbox": [ 107, 574, 505, 608 ], "lines": [ { "bbox": [ 106, 574, 505, 586 ], "spans": [ { "bbox": [ 106, 574, 505, 586 ], "score": 1.0, "content": "For MT and text summarization, we use beam search for decoding and set the number of beams to be", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 584, 505, 599 ], "spans": [ { "bbox": [ 105, 584, 505, 599 ], "score": 1.0, "content": "6 and 5 following previous work (Li & Liang, 2021; Liu et al., 2020b). The min and max generation", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 596, 434, 609 ], "spans": [ { "bbox": [ 106, 596, 434, 609 ], "score": 1.0, "content": "lengths for summarization and MT are set to be (10, 60) and (1, 200) respectively.", "type": "text" } ], "index": 36 } ], "index": 35, "bbox_fs": [ 105, 574, 505, 609 ] }, { "type": "title", "bbox": [ 109, 626, 273, 637 ], "lines": [ { "bbox": [ 106, 624, 275, 639 ], "spans": [ { "bbox": [ 106, 624, 275, 639 ], "score": 1.0, "content": "A.3 OTHER EXPERIMENTAL DETAILS", "type": "text" } ], "index": 37 } ], "index": 37 }, { "type": "text", "bbox": [ 107, 648, 504, 682 ], "lines": [ { "bbox": [ 105, 648, 506, 662 ], "spans": [ { "bbox": [ 105, 648, 506, 662 ], "score": 1.0, "content": "Prefix Tuning: Following Li & Liang (2021), we reparameterize the prefix vectors by a MLP", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 659, 504, 672 ], "spans": [ { "bbox": [ 106, 659, 504, 672 ], "score": 1.0, "content": "network which is composed of a small embedding matrix and a large feedforward neural network.", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 670, 414, 684 ], "spans": [ { "bbox": [ 105, 670, 414, 684 ], "score": 1.0, "content": "This is conducive for learning due to the shared parameters across all layers.", "type": "text" } ], "index": 40 } ], "index": 39, "bbox_fs": [ 105, 648, 506, 684 ] }, { "type": "text", "bbox": [ 107, 687, 504, 731 ], "lines": [ { "bbox": [ 105, 686, 505, 700 ], "spans": [ { "bbox": [ 105, 686, 505, 700 ], "score": 1.0, "content": "LoRA: LoRA and adapter employ different parameter initialization methods: LoRA uses a ran-", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 699, 505, 712 ], "spans": [ { "bbox": [ 105, 699, 342, 712 ], "score": 1.0, "content": "dom Kaiming uniform (He et al., 2015) initialization for", "type": "text" }, { "bbox": [ 342, 699, 372, 710 ], "score": 0.9, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation" }, { "bbox": [ 373, 699, 428, 712 ], "score": 1.0, "content": "and zero for", "type": "text" }, { "bbox": [ 429, 699, 450, 711 ], "score": 0.89, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation" }, { "bbox": [ 450, 699, 505, 712 ], "score": 1.0, "content": "(LoRA init),", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "while adapters use the same initialization as BERT (Devlin et al., 2019). We found it beneficial to", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 721, 336, 731 ], "spans": [ { "bbox": [ 106, 721, 336, 731 ], "score": 1.0, "content": "use the same initialization method as LoRA in scaled PA.", "type": "text" } ], "index": 44 } ], "index": 42.5, "bbox_fs": [ 105, 686, 505, 731 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 116, 109, 492, 167 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 80, 503, 101 ], "group_id": 0, "lines": [ { "bbox": [ 106, 79, 505, 92 ], "spans": [ { "bbox": [ 106, 79, 505, 92 ], "score": 1.0, "content": "Table 8: Training hyperparameters of parameter-efficient tuning methods on the four tasks. lr and ls represents", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 89, 276, 102 ], "spans": [ { "bbox": [ 106, 89, 276, 102 ], "score": 1.0, "content": "learning rate and label smoothing respectively.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 116, 109, 492, 167 ], "group_id": 0, "lines": [ { "bbox": [ 116, 109, 492, 167 ], "spans": [ { "bbox": [ 116, 109, 492, 167 ], "score": 0.98, "html": "
Taskslrbatch sizels max grad norm weight decay train steps
XSum5e-564 sents0.10.10.01100K
enro MT5e-516384 tokens0.11.00.0150K
MNLI/SST21e-432 sents01.00.110 epochs
", "type": "table", "image_path": "846d59878d68877f969b500dac028638d0a5b16b4d6e3366752d799d1a1f64ab.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 116, 109, 492, 128.33333333333334 ], "spans": [], "index": 2 }, { "bbox": [ 116, 128.33333333333334, 492, 147.66666666666669 ], "spans": [], "index": 3 }, { "bbox": [ 116, 147.66666666666669, 492, 167.00000000000003 ], "spans": [], "index": 4 } ] } ], "index": 1.75 }, { "type": "title", "bbox": [ 107, 186, 346, 199 ], "lines": [ { "bbox": [ 105, 185, 347, 200 ], "spans": [ { "bbox": [ 105, 185, 347, 200 ], "score": 1.0, "content": "B COMPUTATION OF TUNABLE PARAMETERS", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "table", "bbox": [ 110, 255, 267, 294 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 216, 266, 246 ], "group_id": 1, "lines": [ { "bbox": [ 105, 215, 267, 227 ], "spans": [ { "bbox": [ 105, 215, 267, 227 ], "score": 1.0, "content": "Table 9: Number of attention or FFN sub-", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 226, 267, 237 ], "spans": [ { "bbox": [ 106, 226, 267, 237 ], "score": 1.0, "content": "layers in each layer of the pre-trained mod-", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 236, 122, 247 ], "spans": [ { "bbox": [ 105, 236, 122, 247 ], "score": 1.0, "content": "els.", "type": "text" } ], "index": 8 } ], "index": 7 }, { "type": "table_body", "bbox": [ 110, 255, 267, 294 ], "group_id": 1, "lines": [ { "bbox": [ 110, 255, 267, 294 ], "spans": [ { "bbox": [ 110, 255, 267, 294 ], "score": 0.965, "html": "
BART/mBARTLARGE RoBERTaBASE
Nattn
Nfn1
", "type": "table", "image_path": "70a3e92d41bd5b1ec4a272f6b67b0465e4c5daf6117d24371899744c0722df6e.jpg" } ] } ], "index": 12, "virtual_lines": [ { "bbox": [ 110, 255, 267, 268.0 ], "spans": [], "index": 11 }, { "bbox": [ 110, 268.0, 267, 281.0 ], "spans": [], "index": 12 }, { "bbox": [ 110, 281.0, 267, 294.0 ], "spans": [], "index": 13 } ] } ], "index": 9.5 }, { "type": "table", "bbox": [ 283, 244, 496, 295 ], "blocks": [ { "type": "table_caption", "bbox": [ 275, 215, 501, 236 ], "group_id": 2, "lines": [ { "bbox": [ 273, 214, 502, 226 ], "spans": [ { "bbox": [ 273, 214, 502, 226 ], "score": 1.0, "content": "Table 10: Number of parameters used at each sub-layer for dif-", "type": "text" } ], "index": 9 }, { "bbox": [ 274, 225, 332, 236 ], "spans": [ { "bbox": [ 274, 225, 332, 236 ], "score": 1.0, "content": "ferent methods.", "type": "text" } ], "index": 10 } ], "index": 9.5 }, { "type": "table_body", "bbox": [ 283, 244, 496, 295 ], "group_id": 2, "lines": [ { "bbox": [ 283, 244, 496, 295 ], "spans": [ { "bbox": [ 283, 244, 496, 295 ], "score": 0.969, "html": "
NattnN
Prefix Tuning2ld一
Adapter variants2rd2rd
LoRA2 × 2rd=4rd 2×(rd+4dr)=10rd
", "type": "table", "image_path": "c0d3f7292300c639fc8c751c9d455d68c81d02592a786ed0007a1728a0fcb0ca.jpg" } ] } ], "index": 15, "virtual_lines": [ { "bbox": [ 283, 244, 496, 261.0 ], "spans": [], "index": 14 }, { "bbox": [ 283, 261.0, 496, 278.0 ], "spans": [], "index": 15 }, { "bbox": [ 283, 278.0, 496, 295.0 ], "spans": [], "index": 16 } ] } ], "index": 12.25 }, { "type": "text", "bbox": [ 106, 302, 506, 534 ], "lines": [ { "bbox": [ 106, 302, 505, 314 ], "spans": [ { "bbox": [ 106, 302, 505, 314 ], "score": 1.0, "content": "We compute the number of tunable parameters based on where the tunable module is inserted into", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 312, 504, 324 ], "spans": [ { "bbox": [ 106, 312, 504, 324 ], "score": 1.0, "content": "and how it is parameterized. The pretrained-models for summarization or MT have an encoder-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 323, 506, 337 ], "spans": [ { "bbox": [ 105, 323, 235, 337 ], "score": 1.0, "content": "decoder structure and each has", "type": "text" }, { "bbox": [ 236, 325, 244, 334 ], "score": 0.72, "content": "L", "type": "inline_equation" }, { "bbox": [ 244, 323, 506, 337 ], "score": 1.0, "content": "layers, whereas RoBERTaBASE for classification tasks only has", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 335, 505, 347 ], "spans": [ { "bbox": [ 106, 335, 115, 345 ], "score": 0.65, "content": "L", "type": "inline_equation" }, { "bbox": [ 115, 335, 505, 347 ], "score": 1.0, "content": "encoder layers. To simplify the computation of tunable parameters, we compute the sum of", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 346, 505, 358 ], "spans": [ { "bbox": [ 105, 346, 505, 358 ], "score": 1.0, "content": "parameter used in one encoder layer and one decoder layer as the parameter overhead of one single", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 357, 505, 370 ], "spans": [ { "bbox": [ 106, 357, 374, 370 ], "score": 1.0, "content": "layer of the pre-trained encoder-decoder model. Each layer has", "type": "text" }, { "bbox": [ 374, 357, 399, 368 ], "score": 0.91, "content": "N _ { \\mathrm { a t t n } }", "type": "inline_equation" }, { "bbox": [ 399, 357, 464, 370 ], "score": 1.0, "content": "sub-layers and", "type": "text" }, { "bbox": [ 464, 357, 483, 368 ], "score": 0.89, "content": "N _ { \\mathrm { { f f n } } }", "type": "inline_equation" }, { "bbox": [ 483, 357, 505, 370 ], "score": 1.0, "content": "sub-", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 366, 506, 381 ], "spans": [ { "bbox": [ 105, 366, 274, 381 ], "score": 1.0, "content": "layers. For the encoder-decoder models,", "type": "text" }, { "bbox": [ 275, 368, 320, 379 ], "score": 0.92, "content": "N _ { \\mathrm { a t t n } } = 3", "type": "inline_equation" }, { "bbox": [ 320, 366, 506, 381 ], "score": 1.0, "content": ": the encoder self-attention, the decoder self-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 379, 505, 392 ], "spans": [ { "bbox": [ 105, 379, 390, 392 ], "score": 1.0, "content": "attention and the decoder cross-attention. For the classification tasks,", "type": "text" }, { "bbox": [ 391, 380, 451, 390 ], "score": 0.48, "content": "\\mathtt { R o B E R T a } _ { \\mathtt { B A S E } }", "type": "inline_equation" }, { "bbox": [ 452, 379, 505, 392 ], "score": 1.0, "content": "only has the", "type": "text" } ], "index": 24 }, { "bbox": [ 104, 389, 506, 403 ], "spans": [ { "bbox": [ 104, 389, 221, 403 ], "score": 1.0, "content": "encoder self-attention, thus", "type": "text" }, { "bbox": [ 222, 390, 269, 401 ], "score": 0.92, "content": "N _ { \\mathrm { a t t n } } ~ = ~ 1", "type": "inline_equation" }, { "bbox": [ 270, 389, 506, 403 ], "score": 1.0, "content": ". We present the number of attention and ffn sub-layers", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 400, 506, 414 ], "spans": [ { "bbox": [ 105, 400, 506, 414 ], "score": 1.0, "content": "for different pre-trained models in Table 10. For modifications applied at the attention sub-layers,", "type": "text" } ], "index": 26 }, { "bbox": [ 101, 407, 504, 430 ], "spans": [ { "bbox": [ 101, 407, 316, 430 ], "score": 1.0, "content": "the number of tunable parameters is computed by", "type": "text" }, { "bbox": [ 316, 411, 445, 424 ], "score": 0.93, "content": "| \\Theta | _ { \\mathrm { a t t n } } = \\bar { N } _ { \\mathrm { W } } ^ { \\mathrm { a t t n } } \\times N _ { \\mathrm { a t t n } } \\times L", "type": "inline_equation" }, { "bbox": [ 446, 407, 478, 430 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 478, 412, 504, 424 ], "score": 0.89, "content": "N _ { \\mathrm { W } } ^ { \\mathrm { a t t n } }", "type": "inline_equation" } ], "index": 27 }, { "bbox": [ 104, 421, 506, 437 ], "spans": [ { "bbox": [ 104, 421, 252, 437 ], "score": 1.0, "content": "denotes the number of parameters", "type": "text" }, { "bbox": [ 252, 423, 284, 434 ], "score": 0.87, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation" }, { "bbox": [ 284, 421, 298, 437 ], "score": 1.0, "content": "or", "type": "text" }, { "bbox": [ 298, 423, 320, 435 ], "score": 0.87, "content": "W _ { \\mathrm { u p , } }", "type": "inline_equation" }, { "bbox": [ 320, 421, 506, 437 ], "score": 1.0, "content": ") used for one attention sub-layer. Similarly,", "type": "text" } ], "index": 28 }, { "bbox": [ 102, 430, 505, 452 ], "spans": [ { "bbox": [ 102, 430, 405, 452 ], "score": 1.0, "content": "the number of tunable parameters for the FFN sub-layers is computed by", "type": "text" }, { "bbox": [ 405, 434, 505, 447 ], "score": 0.91, "content": "\\vert \\Theta \\vert _ { \\mathrm { f f n } } = N _ { \\mathrm { W } } ^ { \\mathrm { f f n } } \\times N _ { \\mathrm { f f n } } \\times", "type": "inline_equation" } ], "index": 29 }, { "bbox": [ 106, 446, 506, 459 ], "spans": [ { "bbox": [ 106, 446, 114, 456 ], "score": 0.65, "content": "L", "type": "inline_equation" }, { "bbox": [ 115, 446, 506, 459 ], "score": 1.0, "content": ". In Table 10, we show the number of parameters for one sub-layer. As we have explained in", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 456, 506, 470 ], "spans": [ { "bbox": [ 106, 457, 125, 468 ], "score": 0.86, "content": "\\ S 4 . 4", "type": "inline_equation" }, { "bbox": [ 126, 456, 410, 470 ], "score": 1.0, "content": ", LoRA approximates the update of each weight matrix with a pair of", "type": "text" }, { "bbox": [ 410, 457, 441, 468 ], "score": 0.93, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation" }, { "bbox": [ 441, 456, 460, 470 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 460, 457, 481, 469 ], "score": 0.9, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation" }, { "bbox": [ 482, 456, 506, 470 ], "score": 1.0, "content": ", thus", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 467, 506, 480 ], "spans": [ { "bbox": [ 105, 467, 321, 480 ], "score": 1.0, "content": "LoRA typically uses more parameters with the same", "type": "text" }, { "bbox": [ 321, 470, 327, 478 ], "score": 0.68, "content": "r", "type": "inline_equation" }, { "bbox": [ 327, 467, 506, 480 ], "score": 1.0, "content": "as other methods. Finally, the total number", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 478, 506, 492 ], "spans": [ { "bbox": [ 105, 478, 396, 492 ], "score": 1.0, "content": "of tunable parameters for prefix tuning, adapter variants and LoRA is", "type": "text" }, { "bbox": [ 396, 479, 492, 491 ], "score": 0.91, "content": "| \\Theta | = | \\Theta | _ { \\mathrm { a t t n } } + | \\Theta | _ { \\mathrm { f n } }", "type": "inline_equation" }, { "bbox": [ 493, 478, 506, 492 ], "score": 1.0, "content": "as", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 489, 506, 502 ], "spans": [ { "bbox": [ 105, 489, 255, 502 ], "score": 1.0, "content": "applicable. Prompt tuning prepends", "type": "text" }, { "bbox": [ 255, 490, 260, 500 ], "score": 0.45, "content": "l", "type": "inline_equation" }, { "bbox": [ 260, 489, 436, 502 ], "score": 1.0, "content": "tunable vectors at the input layer and uses", "type": "text" }, { "bbox": [ 436, 491, 459, 500 ], "score": 0.88, "content": "l \\times d", "type": "inline_equation" }, { "bbox": [ 459, 489, 506, 502 ], "score": 1.0, "content": "number of", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 500, 506, 514 ], "spans": [ { "bbox": [ 105, 500, 506, 514 ], "score": 1.0, "content": "parameters. Using MBART/BART as an example, we present the number of parameters used by", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 511, 506, 524 ], "spans": [ { "bbox": [ 105, 511, 506, 524 ], "score": 1.0, "content": "several representative methods throughout our paper in Table 11, where adapter variants include", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 523, 405, 536 ], "spans": [ { "bbox": [ 106, 523, 405, 536 ], "score": 1.0, "content": "sequential adapter, parallel adapter, scaled adapter and multi-head adapter.", "type": "text" } ], "index": 37 } ], "index": 27 }, { "type": "table", "bbox": [ 157, 573, 452, 675 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 545, 504, 565 ], "group_id": 3, "lines": [ { "bbox": [ 105, 543, 505, 556 ], "spans": [ { "bbox": [ 105, 543, 505, 556 ], "score": 1.0, "content": "Table 11: Number of tunable parameters of various parameter-efficient tuning methods with BART/MBART", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 555, 226, 566 ], "spans": [ { "bbox": [ 106, 555, 137, 566 ], "score": 1.0, "content": "models", "type": "text" }, { "bbox": [ 137, 555, 167, 564 ], "score": 0.86, "content": "L = 1 2", "type": "inline_equation" }, { "bbox": [ 167, 555, 226, 566 ], "score": 1.0, "content": ") as an example.", "type": "text" } ], "index": 39 } ], "index": 38.5 }, { "type": "table_body", "bbox": [ 157, 573, 452, 675 ], "group_id": 3, "lines": [ { "bbox": [ 157, 573, 452, 675 ], "spans": [ { "bbox": [ 157, 573, 452, 675 ], "score": 0.981, "html": "
Methodnumber of parameters
Prompt Tuninglxd
Prefix Tuning (attn)2ld×3×12
Adapter variants (attn)2rd×3×12
Adapter variants (ffn)2rd ×2×12
LoRA (attn)4rd×3×12
LoRA (ffn)10rd ×2×12
MAM Adapter (our proposed model)）2ld×3×12+2rd×2×12
", "type": "table", "image_path": "bf1f2c4fbf0d03088ae09a67d8175765c5927430618c1e6c84ee752a7ebb01fd.jpg" } ] } ], "index": 41, "virtual_lines": [ { "bbox": [ 157, 573, 452, 607.0 ], "spans": [], "index": 40 }, { "bbox": [ 157, 607.0, 452, 641.0 ], "spans": [], "index": 41 }, { "bbox": [ 157, 641.0, 452, 675.0 ], "spans": [], "index": 42 } ] } ], "index": 39.75 }, { "type": "title", "bbox": [ 105, 696, 426, 709 ], "lines": [ { "bbox": [ 106, 695, 428, 711 ], "spans": [ { "bbox": [ 106, 695, 428, 711 ], "score": 1.0, "content": "C FULL RESULTS ON DIFFERENT BOTTLENECK DIMENSIONS", "type": "text" } ], "index": 43 } ], "index": 43 } ], "page_idx": 13, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "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" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 116, 109, 492, 167 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 80, 503, 101 ], "group_id": 0, "lines": [ { "bbox": [ 106, 79, 505, 92 ], "spans": [ { "bbox": [ 106, 79, 505, 92 ], "score": 1.0, "content": "Table 8: Training hyperparameters of parameter-efficient tuning methods on the four tasks. lr and ls represents", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 89, 276, 102 ], "spans": [ { "bbox": [ 106, 89, 276, 102 ], "score": 1.0, "content": "learning rate and label smoothing respectively.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 116, 109, 492, 167 ], "group_id": 0, "lines": [ { "bbox": [ 116, 109, 492, 167 ], "spans": [ { "bbox": [ 116, 109, 492, 167 ], "score": 0.98, "html": "
Taskslrbatch sizels max grad norm weight decay train steps
XSum5e-564 sents0.10.10.01100K
enro MT5e-516384 tokens0.11.00.0150K
MNLI/SST21e-432 sents01.00.110 epochs
", "type": "table", "image_path": "846d59878d68877f969b500dac028638d0a5b16b4d6e3366752d799d1a1f64ab.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 116, 109, 492, 128.33333333333334 ], "spans": [], "index": 2 }, { "bbox": [ 116, 128.33333333333334, 492, 147.66666666666669 ], "spans": [], "index": 3 }, { "bbox": [ 116, 147.66666666666669, 492, 167.00000000000003 ], "spans": [], "index": 4 } ] } ], "index": 1.75 }, { "type": "title", "bbox": [ 107, 186, 346, 199 ], "lines": [ { "bbox": [ 105, 185, 347, 200 ], "spans": [ { "bbox": [ 105, 185, 347, 200 ], "score": 1.0, "content": "B COMPUTATION OF TUNABLE PARAMETERS", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "table", "bbox": [ 110, 255, 267, 294 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 216, 266, 246 ], "group_id": 1, "lines": [ { "bbox": [ 105, 215, 267, 227 ], "spans": [ { "bbox": [ 105, 215, 267, 227 ], "score": 1.0, "content": "Table 9: Number of attention or FFN sub-", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 226, 267, 237 ], "spans": [ { "bbox": [ 106, 226, 267, 237 ], "score": 1.0, "content": "layers in each layer of the pre-trained mod-", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 236, 122, 247 ], "spans": [ { "bbox": [ 105, 236, 122, 247 ], "score": 1.0, "content": "els.", "type": "text" } ], "index": 8 } ], "index": 7 }, { "type": "table_body", "bbox": [ 110, 255, 267, 294 ], "group_id": 1, "lines": [ { "bbox": [ 110, 255, 267, 294 ], "spans": [ { "bbox": [ 110, 255, 267, 294 ], "score": 0.965, "html": "
BART/mBARTLARGE RoBERTaBASE
Nattn
Nfn1
", "type": "table", "image_path": "70a3e92d41bd5b1ec4a272f6b67b0465e4c5daf6117d24371899744c0722df6e.jpg" } ] } ], "index": 12, "virtual_lines": [ { "bbox": [ 110, 255, 267, 268.0 ], "spans": [], "index": 11 }, { "bbox": [ 110, 268.0, 267, 281.0 ], "spans": [], "index": 12 }, { "bbox": [ 110, 281.0, 267, 294.0 ], "spans": [], "index": 13 } ] } ], "index": 9.5 }, { "type": "table", "bbox": [ 283, 244, 496, 295 ], "blocks": [ { "type": "table_caption", "bbox": [ 275, 215, 501, 236 ], "group_id": 2, "lines": [ { "bbox": [ 273, 214, 502, 226 ], "spans": [ { "bbox": [ 273, 214, 502, 226 ], "score": 1.0, "content": "Table 10: Number of parameters used at each sub-layer for dif-", "type": "text" } ], "index": 9 }, { "bbox": [ 274, 225, 332, 236 ], "spans": [ { "bbox": [ 274, 225, 332, 236 ], "score": 1.0, "content": "ferent methods.", "type": "text" } ], "index": 10 } ], "index": 9.5 }, { "type": "table_body", "bbox": [ 283, 244, 496, 295 ], "group_id": 2, "lines": [ { "bbox": [ 283, 244, 496, 295 ], "spans": [ { "bbox": [ 283, 244, 496, 295 ], "score": 0.969, "html": "
NattnN
Prefix Tuning2ld一
Adapter variants2rd2rd
LoRA2 × 2rd=4rd 2×(rd+4dr)=10rd
", "type": "table", "image_path": "c0d3f7292300c639fc8c751c9d455d68c81d02592a786ed0007a1728a0fcb0ca.jpg" } ] } ], "index": 15, "virtual_lines": [ { "bbox": [ 283, 244, 496, 261.0 ], "spans": [], "index": 14 }, { "bbox": [ 283, 261.0, 496, 278.0 ], "spans": [], "index": 15 }, { "bbox": [ 283, 278.0, 496, 295.0 ], "spans": [], "index": 16 } ] } ], "index": 12.25 }, { "type": "text", "bbox": [ 106, 302, 506, 534 ], "lines": [ { "bbox": [ 106, 302, 505, 314 ], "spans": [ { "bbox": [ 106, 302, 505, 314 ], "score": 1.0, "content": "We compute the number of tunable parameters based on where the tunable module is inserted into", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 312, 504, 324 ], "spans": [ { "bbox": [ 106, 312, 504, 324 ], "score": 1.0, "content": "and how it is parameterized. The pretrained-models for summarization or MT have an encoder-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 323, 506, 337 ], "spans": [ { "bbox": [ 105, 323, 235, 337 ], "score": 1.0, "content": "decoder structure and each has", "type": "text" }, { "bbox": [ 236, 325, 244, 334 ], "score": 0.72, "content": "L", "type": "inline_equation" }, { "bbox": [ 244, 323, 506, 337 ], "score": 1.0, "content": "layers, whereas RoBERTaBASE for classification tasks only has", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 335, 505, 347 ], "spans": [ { "bbox": [ 106, 335, 115, 345 ], "score": 0.65, "content": "L", "type": "inline_equation" }, { "bbox": [ 115, 335, 505, 347 ], "score": 1.0, "content": "encoder layers. To simplify the computation of tunable parameters, we compute the sum of", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 346, 505, 358 ], "spans": [ { "bbox": [ 105, 346, 505, 358 ], "score": 1.0, "content": "parameter used in one encoder layer and one decoder layer as the parameter overhead of one single", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 357, 505, 370 ], "spans": [ { "bbox": [ 106, 357, 374, 370 ], "score": 1.0, "content": "layer of the pre-trained encoder-decoder model. Each layer has", "type": "text" }, { "bbox": [ 374, 357, 399, 368 ], "score": 0.91, "content": "N _ { \\mathrm { a t t n } }", "type": "inline_equation" }, { "bbox": [ 399, 357, 464, 370 ], "score": 1.0, "content": "sub-layers and", "type": "text" }, { "bbox": [ 464, 357, 483, 368 ], "score": 0.89, "content": "N _ { \\mathrm { { f f n } } }", "type": "inline_equation" }, { "bbox": [ 483, 357, 505, 370 ], "score": 1.0, "content": "sub-", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 366, 506, 381 ], "spans": [ { "bbox": [ 105, 366, 274, 381 ], "score": 1.0, "content": "layers. For the encoder-decoder models,", "type": "text" }, { "bbox": [ 275, 368, 320, 379 ], "score": 0.92, "content": "N _ { \\mathrm { a t t n } } = 3", "type": "inline_equation" }, { "bbox": [ 320, 366, 506, 381 ], "score": 1.0, "content": ": the encoder self-attention, the decoder self-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 379, 505, 392 ], "spans": [ { "bbox": [ 105, 379, 390, 392 ], "score": 1.0, "content": "attention and the decoder cross-attention. For the classification tasks,", "type": "text" }, { "bbox": [ 391, 380, 451, 390 ], "score": 0.48, "content": "\\mathtt { R o B E R T a } _ { \\mathtt { B A S E } }", "type": "inline_equation" }, { "bbox": [ 452, 379, 505, 392 ], "score": 1.0, "content": "only has the", "type": "text" } ], "index": 24 }, { "bbox": [ 104, 389, 506, 403 ], "spans": [ { "bbox": [ 104, 389, 221, 403 ], "score": 1.0, "content": "encoder self-attention, thus", "type": "text" }, { "bbox": [ 222, 390, 269, 401 ], "score": 0.92, "content": "N _ { \\mathrm { a t t n } } ~ = ~ 1", "type": "inline_equation" }, { "bbox": [ 270, 389, 506, 403 ], "score": 1.0, "content": ". We present the number of attention and ffn sub-layers", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 400, 506, 414 ], "spans": [ { "bbox": [ 105, 400, 506, 414 ], "score": 1.0, "content": "for different pre-trained models in Table 10. For modifications applied at the attention sub-layers,", "type": "text" } ], "index": 26 }, { "bbox": [ 101, 407, 504, 430 ], "spans": [ { "bbox": [ 101, 407, 316, 430 ], "score": 1.0, "content": "the number of tunable parameters is computed by", "type": "text" }, { "bbox": [ 316, 411, 445, 424 ], "score": 0.93, "content": "| \\Theta | _ { \\mathrm { a t t n } } = \\bar { N } _ { \\mathrm { W } } ^ { \\mathrm { a t t n } } \\times N _ { \\mathrm { a t t n } } \\times L", "type": "inline_equation" }, { "bbox": [ 446, 407, 478, 430 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 478, 412, 504, 424 ], "score": 0.89, "content": "N _ { \\mathrm { W } } ^ { \\mathrm { a t t n } }", "type": "inline_equation" } ], "index": 27 }, { "bbox": [ 104, 421, 506, 437 ], "spans": [ { "bbox": [ 104, 421, 252, 437 ], "score": 1.0, "content": "denotes the number of parameters", "type": "text" }, { "bbox": [ 252, 423, 284, 434 ], "score": 0.87, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation" }, { "bbox": [ 284, 421, 298, 437 ], "score": 1.0, "content": "or", "type": "text" }, { "bbox": [ 298, 423, 320, 435 ], "score": 0.87, "content": "W _ { \\mathrm { u p , } }", "type": "inline_equation" }, { "bbox": [ 320, 421, 506, 437 ], "score": 1.0, "content": ") used for one attention sub-layer. Similarly,", "type": "text" } ], "index": 28 }, { "bbox": [ 102, 430, 505, 452 ], "spans": [ { "bbox": [ 102, 430, 405, 452 ], "score": 1.0, "content": "the number of tunable parameters for the FFN sub-layers is computed by", "type": "text" }, { "bbox": [ 405, 434, 505, 447 ], "score": 0.91, "content": "\\vert \\Theta \\vert _ { \\mathrm { f f n } } = N _ { \\mathrm { W } } ^ { \\mathrm { f f n } } \\times N _ { \\mathrm { f f n } } \\times", "type": "inline_equation" } ], "index": 29 }, { "bbox": [ 106, 446, 506, 459 ], "spans": [ { "bbox": [ 106, 446, 114, 456 ], "score": 0.65, "content": "L", "type": "inline_equation" }, { "bbox": [ 115, 446, 506, 459 ], "score": 1.0, "content": ". In Table 10, we show the number of parameters for one sub-layer. As we have explained in", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 456, 506, 470 ], "spans": [ { "bbox": [ 106, 457, 125, 468 ], "score": 0.86, "content": "\\ S 4 . 4", "type": "inline_equation" }, { "bbox": [ 126, 456, 410, 470 ], "score": 1.0, "content": ", LoRA approximates the update of each weight matrix with a pair of", "type": "text" }, { "bbox": [ 410, 457, 441, 468 ], "score": 0.93, "content": "W _ { \\mathrm { d o w n } }", "type": "inline_equation" }, { "bbox": [ 441, 456, 460, 470 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 460, 457, 481, 469 ], "score": 0.9, "content": "W _ { \\mathrm { u p } }", "type": "inline_equation" }, { "bbox": [ 482, 456, 506, 470 ], "score": 1.0, "content": ", thus", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 467, 506, 480 ], "spans": [ { "bbox": [ 105, 467, 321, 480 ], "score": 1.0, "content": "LoRA typically uses more parameters with the same", "type": "text" }, { "bbox": [ 321, 470, 327, 478 ], "score": 0.68, "content": "r", "type": "inline_equation" }, { "bbox": [ 327, 467, 506, 480 ], "score": 1.0, "content": "as other methods. Finally, the total number", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 478, 506, 492 ], "spans": [ { "bbox": [ 105, 478, 396, 492 ], "score": 1.0, "content": "of tunable parameters for prefix tuning, adapter variants and LoRA is", "type": "text" }, { "bbox": [ 396, 479, 492, 491 ], "score": 0.91, "content": "| \\Theta | = | \\Theta | _ { \\mathrm { a t t n } } + | \\Theta | _ { \\mathrm { f n } }", "type": "inline_equation" }, { "bbox": [ 493, 478, 506, 492 ], "score": 1.0, "content": "as", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 489, 506, 502 ], "spans": [ { "bbox": [ 105, 489, 255, 502 ], "score": 1.0, "content": "applicable. Prompt tuning prepends", "type": "text" }, { "bbox": [ 255, 490, 260, 500 ], "score": 0.45, "content": "l", "type": "inline_equation" }, { "bbox": [ 260, 489, 436, 502 ], "score": 1.0, "content": "tunable vectors at the input layer and uses", "type": "text" }, { "bbox": [ 436, 491, 459, 500 ], "score": 0.88, "content": "l \\times d", "type": "inline_equation" }, { "bbox": [ 459, 489, 506, 502 ], "score": 1.0, "content": "number of", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 500, 506, 514 ], "spans": [ { "bbox": [ 105, 500, 506, 514 ], "score": 1.0, "content": "parameters. Using MBART/BART as an example, we present the number of parameters used by", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 511, 506, 524 ], "spans": [ { "bbox": [ 105, 511, 506, 524 ], "score": 1.0, "content": "several representative methods throughout our paper in Table 11, where adapter variants include", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 523, 405, 536 ], "spans": [ { "bbox": [ 106, 523, 405, 536 ], "score": 1.0, "content": "sequential adapter, parallel adapter, scaled adapter and multi-head adapter.", "type": "text" } ], "index": 37 } ], "index": 27, "bbox_fs": [ 101, 302, 506, 536 ] }, { "type": "table", "bbox": [ 157, 573, 452, 675 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 545, 504, 565 ], "group_id": 3, "lines": [ { "bbox": [ 105, 543, 505, 556 ], "spans": [ { "bbox": [ 105, 543, 505, 556 ], "score": 1.0, "content": "Table 11: Number of tunable parameters of various parameter-efficient tuning methods with BART/MBART", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 555, 226, 566 ], "spans": [ { "bbox": [ 106, 555, 137, 566 ], "score": 1.0, "content": "models", "type": "text" }, { "bbox": [ 137, 555, 167, 564 ], "score": 0.86, "content": "L = 1 2", "type": "inline_equation" }, { "bbox": [ 167, 555, 226, 566 ], "score": 1.0, "content": ") as an example.", "type": "text" } ], "index": 39 } ], "index": 38.5 }, { "type": "table_body", "bbox": [ 157, 573, 452, 675 ], "group_id": 3, "lines": [ { "bbox": [ 157, 573, 452, 675 ], "spans": [ { "bbox": [ 157, 573, 452, 675 ], "score": 0.981, "html": "
Methodnumber of parameters
Prompt Tuninglxd
Prefix Tuning (attn)2ld×3×12
Adapter variants (attn)2rd×3×12
Adapter variants (ffn)2rd ×2×12
LoRA (attn)4rd×3×12
LoRA (ffn)10rd ×2×12
MAM Adapter (our proposed model)）2ld×3×12+2rd×2×12
", "type": "table", "image_path": "bf1f2c4fbf0d03088ae09a67d8175765c5927430618c1e6c84ee752a7ebb01fd.jpg" } ] } ], "index": 41, "virtual_lines": [ { "bbox": [ 157, 573, 452, 607.0 ], "spans": [], "index": 40 }, { "bbox": [ 157, 607.0, 452, 641.0 ], "spans": [], "index": 41 }, { "bbox": [ 157, 641.0, 452, 675.0 ], "spans": [], "index": 42 } ] } ], "index": 39.75 }, { "type": "title", "bbox": [ 105, 696, 426, 709 ], "lines": [ { "bbox": [ 106, 695, 428, 711 ], "spans": [ { "bbox": [ 106, 695, 428, 711 ], "score": 1.0, "content": "C FULL RESULTS ON DIFFERENT BOTTLENECK DIMENSIONS", "type": "text" } ], "index": 43 } ], "index": 43 } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 149, 321, 462, 508 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 302, 498, 313 ], "group_id": 0, "lines": [ { "bbox": [ 111, 300, 499, 313 ], "spans": [ { "bbox": [ 111, 300, 499, 313 ], "score": 1.0, "content": "Table 12: Performance on the test sets of abstractive summarization (XSum) and WMT EN-RO translation.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "table_body", "bbox": [ 149, 321, 462, 508 ], "group_id": 0, "lines": [ { "bbox": [ 149, 321, 462, 508 ], "spans": [ { "bbox": [ 149, 321, 462, 508 ], "score": 0.982, "html": "
Method# params (%) XSum (R-1/2/L)MTBLEU
Modified Representation: : attention
Prefix Tuning,r = 2003.6 43.40/20.46/35.51 9.235.6
Prefix Tuning,r = 51243.29/20.40/35.3735.1
LoRA,r= 20043.09/20.29/35.3736.2
Sequential Adapter,r = 20042.01/19.30/34.4035.3
Sequential Adapter,r = 51241.05/18.87/33.7134.7
Parallel Adapter,r = 20043.58/20.31/35.3435.6
Parallel Adapter,r = 51243.99/20.83/35.7736.2
Modified Representation: FFN
LoRA,r = 10244.59/21.31/36.2536.5
Sequential Adapter,r = 2002.4 43.21/19.98/35.0835.6
Sequential Adapter,r = 5126.1 43.72/20.75/35.6436.3
Sequential Adapter,r = 102412.3 43.95/21.00/35.9036.7
Parallel Adapter,r = 2002.4 43.93/20.66/35.6336.4
Parallel Adapter,r = 5126.1 44.35/20.98/35.9837.1
Parallel Adapter,r = 102412.3 44.53/21.24/36.2337.3
", "type": "table", "image_path": "48b5db3956ac0e7f5ba107d05714f722771626ac6ba3466e02322f78b7a0070e.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 149, 321, 462, 383.3333333333333 ], "spans": [], "index": 1 }, { "bbox": [ 149, 383.3333333333333, 462, 445.66666666666663 ], "spans": [], "index": 2 }, { "bbox": [ 149, 445.66666666666663, 462, 507.99999999999994 ], "spans": [], "index": 3 } ] } ], "index": 1.0 } ], "page_idx": 14, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "15", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 149, 321, 462, 508 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 302, 498, 313 ], "group_id": 0, "lines": [ { "bbox": [ 111, 300, 499, 313 ], "spans": [ { "bbox": [ 111, 300, 499, 313 ], "score": 1.0, "content": "Table 12: Performance on the test sets of abstractive summarization (XSum) and WMT EN-RO translation.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "table_body", "bbox": [ 149, 321, 462, 508 ], "group_id": 0, "lines": [ { "bbox": [ 149, 321, 462, 508 ], "spans": [ { "bbox": [ 149, 321, 462, 508 ], "score": 0.982, "html": "
Method# params (%) XSum (R-1/2/L)MTBLEU
Modified Representation: : attention
Prefix Tuning,r = 2003.6 43.40/20.46/35.51 9.235.6
Prefix Tuning,r = 51243.29/20.40/35.3735.1
LoRA,r= 20043.09/20.29/35.3736.2
Sequential Adapter,r = 20042.01/19.30/34.4035.3
Sequential Adapter,r = 51241.05/18.87/33.7134.7
Parallel Adapter,r = 20043.58/20.31/35.3435.6
Parallel Adapter,r = 51243.99/20.83/35.7736.2
Modified Representation: FFN
LoRA,r = 10244.59/21.31/36.2536.5
Sequential Adapter,r = 2002.4 43.21/19.98/35.0835.6
Sequential Adapter,r = 5126.1 43.72/20.75/35.6436.3
Sequential Adapter,r = 102412.3 43.95/21.00/35.9036.7
Parallel Adapter,r = 2002.4 43.93/20.66/35.6336.4
Parallel Adapter,r = 5126.1 44.35/20.98/35.9837.1
Parallel Adapter,r = 102412.3 44.53/21.24/36.2337.3
", "type": "table", "image_path": "48b5db3956ac0e7f5ba107d05714f722771626ac6ba3466e02322f78b7a0070e.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 149, 321, 462, 383.3333333333333 ], "spans": [], "index": 1 }, { "bbox": [ 149, 383.3333333333333, 462, 445.66666666666663 ], "spans": [], "index": 2 }, { "bbox": [ 149, 445.66666666666663, 462, 507.99999999999994 ], "spans": [], "index": 3 } ] } ], "index": 1.0 } ] } ], "_backend": "pipeline", "_version_name": "2.2.2" }