diff --git "a/parse/dev/EbMuimAbPbs/EbMuimAbPbs_middle.json" "b/parse/dev/EbMuimAbPbs/EbMuimAbPbs_middle.json" new file mode 100644--- /dev/null +++ "b/parse/dev/EbMuimAbPbs/EbMuimAbPbs_middle.json" @@ -0,0 +1,48662 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 164, + 101, + 450, + 141 + ], + "lines": [ + { + "bbox": [ + 177, + 100, + 450, + 121 + ], + "spans": [ + { + "bbox": [ + 177, + 100, + 450, + 121 + ], + "score": 1.0, + "content": "Flamingo: a Visual Language Model", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 216, + 118, + 390, + 143 + ], + "spans": [ + { + "bbox": [ + 216, + 118, + 390, + 143 + ], + "score": 1.0, + "content": "for Few-Shot Learning", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 115, + 173, + 495, + 186 + ], + "lines": [ + { + "bbox": [ + 115, + 173, + 496, + 188 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 226, + 188 + ], + "score": 1.0, + "content": "Jean-Baptiste Alayrac*,‡", + "type": "text" + }, + { + "bbox": [ + 250, + 173, + 316, + 187 + ], + "score": 1.0, + "content": "Jeff Donahue*", + "type": "text" + }, + { + "bbox": [ + 339, + 173, + 401, + 188 + ], + "score": 1.0, + "content": "Pauline Luc*", + "type": "text" + }, + { + "bbox": [ + 425, + 173, + 496, + 187 + ], + "score": 1.0, + "content": "Antoine Miech*", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 118, + 201, + 492, + 215 + ], + "lines": [ + { + "bbox": [ + 118, + 199, + 491, + 217 + ], + "spans": [ + { + "bbox": [ + 118, + 200, + 169, + 215 + ], + "score": 1.0, + "content": "Iain Barr†", + "type": "text" + }, + { + "bbox": [ + 181, + 199, + 408, + 217 + ], + "score": 1.0, + "content": "Yana Hasson† Karel Lenc† Arthur Mensch†", + "type": "text" + }, + { + "bbox": [ + 420, + 201, + 491, + 215 + ], + "score": 1.0, + "content": "Katie Millican†", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 114, + 229, + 497, + 243 + ], + "lines": [ + { + "bbox": [ + 112, + 228, + 498, + 245 + ], + "spans": [ + { + "bbox": [ + 112, + 228, + 498, + 245 + ], + "score": 1.0, + "content": "Malcolm Reynolds† Roman Ring† Eliza Rutherford† Serkan Cabi Tengda Han", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 119, + 258, + 485, + 271 + ], + "lines": [ + { + "bbox": [ + 127, + 256, + 484, + 273 + ], + "spans": [ + { + "bbox": [ + 127, + 256, + 187, + 273 + ], + "score": 1.0, + "content": "Zhitao Gong", + "type": "text" + }, + { + "bbox": [ + 208, + 258, + 286, + 272 + ], + "score": 1.0, + "content": "Sina Samangooei", + "type": "text" + }, + { + "bbox": [ + 308, + 257, + 397, + 272 + ], + "score": 1.0, + "content": "Marianne Monteiro", + "type": "text" + }, + { + "bbox": [ + 419, + 258, + 484, + 271 + ], + "score": 1.0, + "content": "Jacob Menick", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 120, + 286, + 490, + 299 + ], + "lines": [ + { + "bbox": [ + 118, + 285, + 491, + 301 + ], + "spans": [ + { + "bbox": [ + 118, + 285, + 292, + 301 + ], + "score": 1.0, + "content": "Sebastian Borgeaud Andrew Brock", + "type": "text" + }, + { + "bbox": [ + 304, + 286, + 388, + 299 + ], + "score": 1.0, + "content": "Aida Nematzadeh", + "type": "text" + }, + { + "bbox": [ + 399, + 286, + 491, + 299 + ], + "score": 1.0, + "content": "Sahand Sharifzadeh", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 119, + 315, + 489, + 327 + ], + "lines": [ + { + "bbox": [ + 123, + 313, + 488, + 329 + ], + "spans": [ + { + "bbox": [ + 123, + 315, + 208, + 328 + ], + "score": 1.0, + "content": "Mikolaj Binkowski", + "type": "text" + }, + { + "bbox": [ + 225, + 313, + 304, + 328 + ], + "score": 1.0, + "content": "Ricardo Barreira", + "type": "text" + }, + { + "bbox": [ + 322, + 314, + 385, + 329 + ], + "score": 1.0, + "content": "Oriol Vinyals", + "type": "text" + }, + { + "bbox": [ + 403, + 315, + 488, + 327 + ], + "score": 1.0, + "content": "Andrew Zisserman", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 263, + 343, + 347, + 355 + ], + "lines": [ + { + "bbox": [ + 262, + 342, + 349, + 356 + ], + "spans": [ + { + "bbox": [ + 262, + 342, + 349, + 356 + ], + "score": 1.0, + "content": "Karen Simonyan*,‡", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 125, + 371, + 484, + 394 + ], + "lines": [ + { + "bbox": [ + 127, + 370, + 484, + 385 + ], + "spans": [ + { + "bbox": [ + 127, + 370, + 484, + 385 + ], + "score": 1.0, + "content": "* Equal contributions, ordered alphabetically, † Equal contributions, ordered alphabetically,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 248, + 380, + 362, + 399 + ], + "spans": [ + { + "bbox": [ + 248, + 380, + 362, + 399 + ], + "score": 1.0, + "content": "‡ Equal senior contributions", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 282, + 411, + 328, + 423 + ], + "lines": [ + { + "bbox": [ + 281, + 410, + 330, + 426 + ], + "spans": [ + { + "bbox": [ + 281, + 410, + 330, + 426 + ], + "score": 1.0, + "content": "DeepMind", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 283, + 451, + 328, + 464 + ], + "lines": [ + { + "bbox": [ + 281, + 450, + 331, + 465 + ], + "spans": [ + { + "bbox": [ + 281, + 450, + 331, + 465 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 143, + 475, + 469, + 672 + ], + "lines": [ + { + "bbox": [ + 142, + 476, + 470, + 488 + ], + "spans": [ + { + "bbox": [ + 142, + 476, + 470, + 488 + ], + "score": 1.0, + "content": "Building models that can be rapidly adapted to novel tasks using only a handful of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 488, + 470, + 498 + ], + "spans": [ + { + "bbox": [ + 142, + 488, + 470, + 498 + ], + "score": 1.0, + "content": "annotated examples is an open challenge for multimodal machine learning research.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 497, + 470, + 510 + ], + "spans": [ + { + "bbox": [ + 141, + 497, + 470, + 510 + ], + "score": 1.0, + "content": "We introduce Flamingo, a family of Visual Language Models (VLM) with this", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 509, + 470, + 521 + ], + "spans": [ + { + "bbox": [ + 141, + 509, + 470, + 521 + ], + "score": 1.0, + "content": "ability. We propose key architectural innovations to: (i) bridge powerful pretrained", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 520, + 470, + 532 + ], + "spans": [ + { + "bbox": [ + 141, + 520, + 470, + 532 + ], + "score": 1.0, + "content": "vision-only and language-only models, (ii) handle sequences of arbitrarily inter-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 530, + 469, + 543 + ], + "spans": [ + { + "bbox": [ + 141, + 530, + 469, + 543 + ], + "score": 1.0, + "content": "leaved visual and textual data, and (iii) seamlessly ingest images or videos as", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 541, + 470, + 554 + ], + "spans": [ + { + "bbox": [ + 141, + 541, + 470, + 554 + ], + "score": 1.0, + "content": "inputs. Thanks to their flexibility, Flamingo models can be trained on large-scale", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 552, + 469, + 564 + ], + "spans": [ + { + "bbox": [ + 142, + 552, + 469, + 564 + ], + "score": 1.0, + "content": "multimodal web corpora containing arbitrarily interleaved text and images, which", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 563, + 470, + 576 + ], + "spans": [ + { + "bbox": [ + 141, + 563, + 470, + 576 + ], + "score": 1.0, + "content": "is key to endow them with in-context few-shot learning capabilities. We perform a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 574, + 469, + 587 + ], + "spans": [ + { + "bbox": [ + 141, + 574, + 469, + 587 + ], + "score": 1.0, + "content": "thorough evaluation of our models, exploring and measuring their ability to rapidly", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 585, + 469, + 597 + ], + "spans": [ + { + "bbox": [ + 141, + 585, + 469, + 597 + ], + "score": 1.0, + "content": "adapt to a variety of image and video tasks. These include open-ended tasks such", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 596, + 469, + 608 + ], + "spans": [ + { + "bbox": [ + 141, + 596, + 469, + 608 + ], + "score": 1.0, + "content": "as visual question-answering, where the model is prompted with a question which", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 607, + 469, + 619 + ], + "spans": [ + { + "bbox": [ + 141, + 607, + 469, + 619 + ], + "score": 1.0, + "content": "it has to answer; captioning tasks, which evaluate the ability to describe a scene or", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 617, + 470, + 630 + ], + "spans": [ + { + "bbox": [ + 141, + 617, + 470, + 630 + ], + "score": 1.0, + "content": "an event; and close-ended tasks such as multiple-choice visual question-answering.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 628, + 470, + 641 + ], + "spans": [ + { + "bbox": [ + 141, + 628, + 470, + 641 + ], + "score": 1.0, + "content": "For tasks lying anywhere on this spectrum, a single Flamingo model can achieve a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 639, + 470, + 653 + ], + "spans": [ + { + "bbox": [ + 141, + 639, + 470, + 653 + ], + "score": 1.0, + "content": "new state of the art with few-shot learning, simply by prompting the model with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 142, + 651, + 469, + 663 + ], + "spans": [ + { + "bbox": [ + 142, + 651, + 469, + 663 + ], + "score": 1.0, + "content": "task-specific examples. On numerous benchmarks, Flamingo outperforms models", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 660, + 372, + 673 + ], + "spans": [ + { + "bbox": [ + 141, + 660, + 372, + 673 + ], + "score": 1.0, + "content": "fine-tuned on thousands of times more task-specific data.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 21.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 699, + 404, + 720 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 406, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 406, + 712 + ], + "score": 1.0, + "content": "Corresponding authors: {jalayrac|jeffdonahue|paulineluc|miech}@deepmind.com", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 709, + 386, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 386, + 722 + ], + "score": 1.0, + "content": "36th Conference on Neural Information Processing Systems (NeurIPS 2022).", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 164, + 101, + 450, + 141 + ], + "lines": [ + { + "bbox": [ + 177, + 100, + 450, + 121 + ], + "spans": [ + { + "bbox": [ + 177, + 100, + 450, + 121 + ], + "score": 1.0, + "content": "Flamingo: a Visual Language Model", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 216, + 118, + 390, + 143 + ], + "spans": [ + { + "bbox": [ + 216, + 118, + 390, + 143 + ], + "score": 1.0, + "content": "for Few-Shot Learning", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 115, + 173, + 495, + 186 + ], + "lines": [ + { + "bbox": [ + 115, + 173, + 496, + 188 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 226, + 188 + ], + "score": 1.0, + "content": "Jean-Baptiste Alayrac*,‡", + "type": "text" + }, + { + "bbox": [ + 250, + 173, + 316, + 187 + ], + "score": 1.0, + "content": "Jeff Donahue*", + "type": "text" + }, + { + "bbox": [ + 339, + 173, + 401, + 188 + ], + "score": 1.0, + "content": "Pauline Luc*", + "type": "text" + }, + { + "bbox": [ + 425, + 173, + 496, + 187 + ], + "score": 1.0, + "content": "Antoine Miech*", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 115, + 173, + 496, + 188 + ] + }, + { + "type": "text", + "bbox": [ + 118, + 201, + 492, + 215 + ], + "lines": [ + { + "bbox": [ + 118, + 199, + 491, + 217 + ], + "spans": [ + { + "bbox": [ + 118, + 200, + 169, + 215 + ], + "score": 1.0, + "content": "Iain Barr†", + "type": "text" + }, + { + "bbox": [ + 181, + 199, + 408, + 217 + ], + "score": 1.0, + "content": "Yana Hasson† Karel Lenc† Arthur Mensch†", + "type": "text" + }, + { + "bbox": [ + 420, + 201, + 491, + 215 + ], + "score": 1.0, + "content": "Katie Millican†", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 118, + 199, + 491, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 114, + 229, + 497, + 243 + ], + "lines": [ + { + "bbox": [ + 112, + 228, + 498, + 245 + ], + "spans": [ + { + "bbox": [ + 112, + 228, + 498, + 245 + ], + "score": 1.0, + "content": "Malcolm Reynolds† Roman Ring† Eliza Rutherford† Serkan Cabi Tengda Han", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 112, + 228, + 498, + 245 + ] + }, + { + "type": "text", + "bbox": [ + 119, + 258, + 485, + 271 + ], + "lines": [ + { + "bbox": [ + 127, + 256, + 484, + 273 + ], + "spans": [ + { + "bbox": [ + 127, + 256, + 187, + 273 + ], + "score": 1.0, + "content": "Zhitao Gong", + "type": "text" + }, + { + "bbox": [ + 208, + 258, + 286, + 272 + ], + "score": 1.0, + "content": "Sina Samangooei", + "type": "text" + }, + { + "bbox": [ + 308, + 257, + 397, + 272 + ], + "score": 1.0, + "content": "Marianne Monteiro", + "type": "text" + }, + { + "bbox": [ + 419, + 258, + 484, + 271 + ], + "score": 1.0, + "content": "Jacob Menick", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 127, + 256, + 484, + 273 + ] + }, + { + "type": "text", + "bbox": [ + 120, + 286, + 490, + 299 + ], + "lines": [ + { + "bbox": [ + 118, + 285, + 491, + 301 + ], + "spans": [ + { + "bbox": [ + 118, + 285, + 292, + 301 + ], + "score": 1.0, + "content": "Sebastian Borgeaud Andrew Brock", + "type": "text" + }, + { + "bbox": [ + 304, + 286, + 388, + 299 + ], + "score": 1.0, + "content": "Aida Nematzadeh", + "type": "text" + }, + { + "bbox": [ + 399, + 286, + 491, + 299 + ], + "score": 1.0, + "content": "Sahand Sharifzadeh", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 118, + 285, + 491, + 301 + ] + }, + { + "type": "text", + "bbox": [ + 119, + 315, + 489, + 327 + ], + "lines": [ + { + "bbox": [ + 123, + 313, + 488, + 329 + ], + "spans": [ + { + "bbox": [ + 123, + 315, + 208, + 328 + ], + "score": 1.0, + "content": "Mikolaj Binkowski", + "type": "text" + }, + { + "bbox": [ + 225, + 313, + 304, + 328 + ], + "score": 1.0, + "content": "Ricardo Barreira", + "type": "text" + }, + { + "bbox": [ + 322, + 314, + 385, + 329 + ], + "score": 1.0, + "content": "Oriol Vinyals", + "type": "text" + }, + { + "bbox": [ + 403, + 315, + 488, + 327 + ], + "score": 1.0, + "content": "Andrew Zisserman", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 123, + 313, + 488, + 329 + ] + }, + { + "type": "text", + "bbox": [ + 263, + 343, + 347, + 355 + ], + "lines": [ + { + "bbox": [ + 262, + 342, + 349, + 356 + ], + "spans": [ + { + "bbox": [ + 262, + 342, + 349, + 356 + ], + "score": 1.0, + "content": "Karen Simonyan*,‡", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 262, + 342, + 349, + 356 + ] + }, + { + "type": "text", + "bbox": [ + 125, + 371, + 484, + 394 + ], + "lines": [ + { + "bbox": [ + 127, + 370, + 484, + 385 + ], + "spans": [ + { + "bbox": [ + 127, + 370, + 484, + 385 + ], + "score": 1.0, + "content": "* Equal contributions, ordered alphabetically, † Equal contributions, ordered alphabetically,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 248, + 380, + 362, + 399 + ], + "spans": [ + { + "bbox": [ + 248, + 380, + 362, + 399 + ], + "score": 1.0, + "content": "‡ Equal senior contributions", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 127, + 370, + 484, + 399 + ] + }, + { + "type": "title", + "bbox": [ + 282, + 411, + 328, + 423 + ], + "lines": [ + { + "bbox": [ + 281, + 410, + 330, + 426 + ], + "spans": [ + { + "bbox": [ + 281, + 410, + 330, + 426 + ], + "score": 1.0, + "content": "DeepMind", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 283, + 451, + 328, + 464 + ], + "lines": [ + { + "bbox": [ + 281, + 450, + 331, + 465 + ], + "spans": [ + { + "bbox": [ + 281, + 450, + 331, + 465 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 143, + 475, + 469, + 672 + ], + "lines": [ + { + "bbox": [ + 142, + 476, + 470, + 488 + ], + "spans": [ + { + "bbox": [ + 142, + 476, + 470, + 488 + ], + "score": 1.0, + "content": "Building models that can be rapidly adapted to novel tasks using only a handful of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 488, + 470, + 498 + ], + "spans": [ + { + "bbox": [ + 142, + 488, + 470, + 498 + ], + "score": 1.0, + "content": "annotated examples is an open challenge for multimodal machine learning research.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 497, + 470, + 510 + ], + "spans": [ + { + "bbox": [ + 141, + 497, + 470, + 510 + ], + "score": 1.0, + "content": "We introduce Flamingo, a family of Visual Language Models (VLM) with this", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 509, + 470, + 521 + ], + "spans": [ + { + "bbox": [ + 141, + 509, + 470, + 521 + ], + "score": 1.0, + "content": "ability. We propose key architectural innovations to: (i) bridge powerful pretrained", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 520, + 470, + 532 + ], + "spans": [ + { + "bbox": [ + 141, + 520, + 470, + 532 + ], + "score": 1.0, + "content": "vision-only and language-only models, (ii) handle sequences of arbitrarily inter-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 530, + 469, + 543 + ], + "spans": [ + { + "bbox": [ + 141, + 530, + 469, + 543 + ], + "score": 1.0, + "content": "leaved visual and textual data, and (iii) seamlessly ingest images or videos as", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 541, + 470, + 554 + ], + "spans": [ + { + "bbox": [ + 141, + 541, + 470, + 554 + ], + "score": 1.0, + "content": "inputs. Thanks to their flexibility, Flamingo models can be trained on large-scale", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 552, + 469, + 564 + ], + "spans": [ + { + "bbox": [ + 142, + 552, + 469, + 564 + ], + "score": 1.0, + "content": "multimodal web corpora containing arbitrarily interleaved text and images, which", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 563, + 470, + 576 + ], + "spans": [ + { + "bbox": [ + 141, + 563, + 470, + 576 + ], + "score": 1.0, + "content": "is key to endow them with in-context few-shot learning capabilities. We perform a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 574, + 469, + 587 + ], + "spans": [ + { + "bbox": [ + 141, + 574, + 469, + 587 + ], + "score": 1.0, + "content": "thorough evaluation of our models, exploring and measuring their ability to rapidly", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 585, + 469, + 597 + ], + "spans": [ + { + "bbox": [ + 141, + 585, + 469, + 597 + ], + "score": 1.0, + "content": "adapt to a variety of image and video tasks. These include open-ended tasks such", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 596, + 469, + 608 + ], + "spans": [ + { + "bbox": [ + 141, + 596, + 469, + 608 + ], + "score": 1.0, + "content": "as visual question-answering, where the model is prompted with a question which", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 607, + 469, + 619 + ], + "spans": [ + { + "bbox": [ + 141, + 607, + 469, + 619 + ], + "score": 1.0, + "content": "it has to answer; captioning tasks, which evaluate the ability to describe a scene or", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 617, + 470, + 630 + ], + "spans": [ + { + "bbox": [ + 141, + 617, + 470, + 630 + ], + "score": 1.0, + "content": "an event; and close-ended tasks such as multiple-choice visual question-answering.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 628, + 470, + 641 + ], + "spans": [ + { + "bbox": [ + 141, + 628, + 470, + 641 + ], + "score": 1.0, + "content": "For tasks lying anywhere on this spectrum, a single Flamingo model can achieve a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 639, + 470, + 653 + ], + "spans": [ + { + "bbox": [ + 141, + 639, + 470, + 653 + ], + "score": 1.0, + "content": "new state of the art with few-shot learning, simply by prompting the model with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 142, + 651, + 469, + 663 + ], + "spans": [ + { + "bbox": [ + 142, + 651, + 469, + 663 + ], + "score": 1.0, + "content": "task-specific examples. On numerous benchmarks, Flamingo outperforms models", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 660, + 372, + 673 + ], + "spans": [ + { + "bbox": [ + 141, + 660, + 372, + 673 + ], + "score": 1.0, + "content": "fine-tuned on thousands of times more task-specific data.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 21.5, + "bbox_fs": [ + 141, + 476, + 470, + 673 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 49, + 505, + 688 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 49, + 505, + 688 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 49, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 108, + 49, + 505, + 688 + ], + "score": 0.961, + "type": "image", + "image_path": "500fea4496e438926bc70a717688f49aef7e0ec0799bbdcc53a49ceb3646c261.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 49, + 505, + 262.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 262.0, + 505, + 475.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 475.0, + 505, + 688.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 698, + 505, + 741 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Figure 1: Selected examples of inputs and outputs obtained from Flamingo-80B. Flamingo", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "can rapidly adapt to various image/video understanding tasks with few-shot prompting (top). Out", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "of the box, Flamingo is also capable of multi-image visual dialogue (bottom). More examples in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 730, + 159, + 743 + ], + "spans": [ + { + "bbox": [ + 105, + 730, + 159, + 743 + ], + "score": 1.0, + "content": "Appendix C.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 49, + 505, + 688 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 49, + 505, + 688 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 49, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 108, + 49, + 505, + 688 + ], + "score": 0.961, + "type": "image", + "image_path": "500fea4496e438926bc70a717688f49aef7e0ec0799bbdcc53a49ceb3646c261.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 49, + 505, + 262.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 262.0, + 505, + 475.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 475.0, + 505, + 688.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 698, + 505, + 741 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Figure 1: Selected examples of inputs and outputs obtained from Flamingo-80B. Flamingo", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "can rapidly adapt to various image/video understanding tasks with few-shot prompting (top). Out", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "of the box, Flamingo is also capable of multi-image visual dialogue (bottom). More examples in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 730, + 159, + 743 + ], + "spans": [ + { + "bbox": [ + 105, + 730, + 159, + 743 + ], + "score": 1.0, + "content": "Appendix C.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 74, + 503, + 204 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 74, + 503, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 74, + 503, + 204 + ], + "spans": [ + { + "bbox": [ + 109, + 74, + 503, + 204 + ], + "score": 0.965, + "type": "image", + "image_path": "67d3c4f598fa5f712a5a0d97a635ffbeefb2449fa69c58e60a16d12063b978e9.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 74, + 503, + 117.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 117.33333333333334, + 503, + 160.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 160.66666666666669, + 503, + 204.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 213, + 505, + 268 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 226 + ], + "score": 1.0, + "content": "Figure 2: Flamingo results overview. Left: Our largest model, dubbed Flamingo, outperforms", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "state-of-the-art fine-tuned models on 6 of the 16 tasks we consider with no fine-tuning. For the 9", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 235, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 506, + 247 + ], + "score": 1.0, + "content": "tasks with published few-shot results, Flamingo sets the new few-shot state of the art. Note: We omit", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "RareAct, our 16th benchmark, as it is a zero-shot benchmark with no available fine-tuned results to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 257, + 469, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 469, + 269 + ], + "score": 1.0, + "content": "compare to. Right: Flamingo performance improves with model size and number of shots.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 107, + 276, + 190, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 192, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 192, + 292 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 506, + 444 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 506, + 314 + ], + "score": 1.0, + "content": "One key aspect of intelligence is the ability to quickly learn to perform a new task given a short", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "score": 1.0, + "content": "instruction [33, 70]. While initial progress has been made towards a similar capability in computer", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "score": 1.0, + "content": "vision, the most widely used paradigm still consists of first pretraining on a large amount of supervised", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "data, before fine-tuning the model on the task of interest [66, 118, 143]. However, successful fine-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "tuning often requires many thousands of annotated data points. In addition, it often requires careful", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 355, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 369 + ], + "score": 1.0, + "content": "per-task hyperparameter tuning and is also resource intensive. Recently, multimodal vision-language", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "models trained with a contrastive objective [50, 85] have enabled zero-shot adaptation to novel tasks,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 377, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 506, + 391 + ], + "score": 1.0, + "content": "without the need for fine-tuning. However, because these models simply provide a similarity score", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 389, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 506, + 401 + ], + "score": 1.0, + "content": "between a text and an image, they can only address limited use cases such as classification, where a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 397, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 415 + ], + "score": 1.0, + "content": "finite set of outcomes is provided beforehand. They crucially lack the ability to generate language,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "which makes them less suitable to more open-ended tasks such as captioning or visual question-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "score": 1.0, + "content": "answering. Others have explored visually-conditioned language generation [17, 114, 119, 124, 132]", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 432, + 360, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 360, + 446 + ], + "score": 1.0, + "content": "but have not yet shown good performance in low-data regimes.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 448, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "We introduce Flamingo, a Visual Language Model (VLM) that sets a new state of the art in few-shot", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "learning on a wide range of open-ended vision and language tasks, simply by being prompted with a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 470, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 484 + ], + "score": 1.0, + "content": "few input/output examples, as illustrated in Figure 1. Of the 16 tasks we consider, Flamingo also", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "surpasses the fine-tuned state of the art on 6 tasks, despite using orders of magnitude less task-specific", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 492, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 507 + ], + "score": 1.0, + "content": "training data (see Figure 2). To achieve this, Flamingo takes inspiration from recent work on large", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 502, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 517 + ], + "score": 1.0, + "content": "language models (LMs) which are good few-shot learners [11, 18, 42, 86]. A single large LM can", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "achieve strong performance on many tasks using only its text interface: a few examples of a task are", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "provided to the model as a prompt, along with a query input, and the model generates a continuation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "to produce a predicted output for that query. We show that the same can be done for image and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 506, + 560 + ], + "score": 1.0, + "content": "video understanding tasks such as classification, captioning, or question-answering: these can be", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 558, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 570 + ], + "score": 1.0, + "content": "cast as text prediction problems with visual input conditioning. The difference from a LM is that", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "the model must be able to ingest a multimodal prompt containing images and/or videos interleaved", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "with text. Flamingo models have this capability—they are visually-conditioned autoregressive text", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 590, + 507, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 507, + 605 + ], + "score": 1.0, + "content": "generation models able to ingest a sequence of text tokens interleaved with images and/or videos,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "and produce text as output. Flamingo models leverage two complementary pre-trained and frozen", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "models: a vision model which can “perceive” visual scenes and a large LM which performs a basic", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "form of reasoning. Novel architecture components are added in between these models to connect", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "score": 1.0, + "content": "them in a way that preserves the knowledge they have accumulated during computationally intensive", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "pre-training. Flamingo models are also able to ingest high-resolution images or videos thanks to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 670 + ], + "score": 1.0, + "content": "a Perceiver-based [48] architecture that can produce a small fixed number of visual tokens per", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 667, + 395, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 395, + 680 + ], + "score": 1.0, + "content": "image/video, given a large and variable number of visual input features.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 108, + 682, + 504, + 726 + ], + "lines": [ + { + "bbox": [ + 105, + 682, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 506, + 695 + ], + "score": 1.0, + "content": "A crucial aspect for the performance of large LMs is that they are trained on a large amount of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 694, + 506, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 506, + 706 + ], + "score": 1.0, + "content": "text data. This training provides general-purpose generation capabilities that allows these LMs to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 705, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 705, + 506, + 716 + ], + "score": 1.0, + "content": "perform well when prompted with task examples. Similarly, we demonstrate that the way we train", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 716, + 505, + 727 + ], + "spans": [ + { + "bbox": [ + 106, + 716, + 505, + 727 + ], + "score": 1.0, + "content": "the Flamingo models is crucial for their final performance. They are trained on a carefully chosen", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 74, + 503, + 204 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 74, + 503, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 74, + 503, + 204 + ], + "spans": [ + { + "bbox": [ + 109, + 74, + 503, + 204 + ], + "score": 0.965, + "type": "image", + "image_path": "67d3c4f598fa5f712a5a0d97a635ffbeefb2449fa69c58e60a16d12063b978e9.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 74, + 503, + 117.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 117.33333333333334, + 503, + 160.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 160.66666666666669, + 503, + 204.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 213, + 505, + 268 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 226 + ], + "score": 1.0, + "content": "Figure 2: Flamingo results overview. Left: Our largest model, dubbed Flamingo, outperforms", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "state-of-the-art fine-tuned models on 6 of the 16 tasks we consider with no fine-tuning. For the 9", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 235, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 506, + 247 + ], + "score": 1.0, + "content": "tasks with published few-shot results, Flamingo sets the new few-shot state of the art. Note: We omit", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "RareAct, our 16th benchmark, as it is a zero-shot benchmark with no available fine-tuned results to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 257, + 469, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 469, + 269 + ], + "score": 1.0, + "content": "compare to. Right: Flamingo performance improves with model size and number of shots.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 107, + 276, + 190, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 192, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 192, + 292 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 506, + 444 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 506, + 314 + ], + "score": 1.0, + "content": "One key aspect of intelligence is the ability to quickly learn to perform a new task given a short", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "score": 1.0, + "content": "instruction [33, 70]. While initial progress has been made towards a similar capability in computer", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "score": 1.0, + "content": "vision, the most widely used paradigm still consists of first pretraining on a large amount of supervised", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "data, before fine-tuning the model on the task of interest [66, 118, 143]. However, successful fine-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "tuning often requires many thousands of annotated data points. In addition, it often requires careful", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 355, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 369 + ], + "score": 1.0, + "content": "per-task hyperparameter tuning and is also resource intensive. Recently, multimodal vision-language", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "models trained with a contrastive objective [50, 85] have enabled zero-shot adaptation to novel tasks,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 377, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 506, + 391 + ], + "score": 1.0, + "content": "without the need for fine-tuning. However, because these models simply provide a similarity score", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 389, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 506, + 401 + ], + "score": 1.0, + "content": "between a text and an image, they can only address limited use cases such as classification, where a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 397, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 415 + ], + "score": 1.0, + "content": "finite set of outcomes is provided beforehand. They crucially lack the ability to generate language,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "which makes them less suitable to more open-ended tasks such as captioning or visual question-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "score": 1.0, + "content": "answering. Others have explored visually-conditioned language generation [17, 114, 119, 124, 132]", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 432, + 360, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 360, + 446 + ], + "score": 1.0, + "content": "but have not yet shown good performance in low-data regimes.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 301, + 506, + 446 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 448, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "We introduce Flamingo, a Visual Language Model (VLM) that sets a new state of the art in few-shot", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "learning on a wide range of open-ended vision and language tasks, simply by being prompted with a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 470, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 484 + ], + "score": 1.0, + "content": "few input/output examples, as illustrated in Figure 1. Of the 16 tasks we consider, Flamingo also", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "surpasses the fine-tuned state of the art on 6 tasks, despite using orders of magnitude less task-specific", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 492, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 507 + ], + "score": 1.0, + "content": "training data (see Figure 2). To achieve this, Flamingo takes inspiration from recent work on large", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 502, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 517 + ], + "score": 1.0, + "content": "language models (LMs) which are good few-shot learners [11, 18, 42, 86]. A single large LM can", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "achieve strong performance on many tasks using only its text interface: a few examples of a task are", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "provided to the model as a prompt, along with a query input, and the model generates a continuation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "to produce a predicted output for that query. We show that the same can be done for image and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 506, + 560 + ], + "score": 1.0, + "content": "video understanding tasks such as classification, captioning, or question-answering: these can be", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 558, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 570 + ], + "score": 1.0, + "content": "cast as text prediction problems with visual input conditioning. The difference from a LM is that", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "the model must be able to ingest a multimodal prompt containing images and/or videos interleaved", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "with text. Flamingo models have this capability—they are visually-conditioned autoregressive text", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 590, + 507, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 507, + 605 + ], + "score": 1.0, + "content": "generation models able to ingest a sequence of text tokens interleaved with images and/or videos,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "and produce text as output. Flamingo models leverage two complementary pre-trained and frozen", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "models: a vision model which can “perceive” visual scenes and a large LM which performs a basic", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "form of reasoning. Novel architecture components are added in between these models to connect", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "score": 1.0, + "content": "them in a way that preserves the knowledge they have accumulated during computationally intensive", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "pre-training. Flamingo models are also able to ingest high-resolution images or videos thanks to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 670 + ], + "score": 1.0, + "content": "a Perceiver-based [48] architecture that can produce a small fixed number of visual tokens per", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 667, + 395, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 395, + 680 + ], + "score": 1.0, + "content": "image/video, given a large and variable number of visual input features.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 448, + 507, + 680 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 682, + 504, + 726 + ], + "lines": [ + { + "bbox": [ + 105, + 682, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 506, + 695 + ], + "score": 1.0, + "content": "A crucial aspect for the performance of large LMs is that they are trained on a large amount of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 694, + 506, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 506, + 706 + ], + "score": 1.0, + "content": "text data. This training provides general-purpose generation capabilities that allows these LMs to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 705, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 705, + 506, + 716 + ], + "score": 1.0, + "content": "perform well when prompted with task examples. Similarly, we demonstrate that the way we train", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 716, + 505, + 727 + ], + "spans": [ + { + "bbox": [ + 106, + 716, + 505, + 727 + ], + "score": 1.0, + "content": "the Flamingo models is crucial for their final performance. They are trained on a carefully chosen", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 292, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 505, + 307 + ], + "score": 1.0, + "content": "mixture of complementary large-scale multimodal data coming only from the web, without using any", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "data annotated for machine learning purposes. After this training, a Flamingo model can be directly", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 315, + 448, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 448, + 329 + ], + "score": 1.0, + "content": "adapted to vision tasks via simple few-shot learning without any task-specific tuning.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 682, + 506, + 727 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 70, + 505, + 243 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 70, + 505, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 70, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 112, + 70, + 505, + 243 + ], + "score": 0.973, + "type": "image", + "image_path": "f1607ad89c4ea7b40f8723498b439571ecfb75b7d7f622105db2d86667955992.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 70, + 505, + 127.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 127.66666666666666, + 505, + 185.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 185.33333333333331, + 505, + 242.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 250, + 505, + 273 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "Figure 3: Flamingo architecture overview. Flamingo is a family of visual language models (VLMs)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 260, + 456, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 456, + 274 + ], + "score": 1.0, + "content": "that take as input visual data interleaved with text and produce free-form text as output.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 293, + 505, + 327 + ], + "lines": [ + { + "bbox": [ + 105, + 292, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 505, + 307 + ], + "score": 1.0, + "content": "mixture of complementary large-scale multimodal data coming only from the web, without using any", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "data annotated for machine learning purposes. After this training, a Flamingo model can be directly", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 315, + 448, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 448, + 329 + ], + "score": 1.0, + "content": "adapted to vision tasks via simple few-shot learning without any task-specific tuning.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 331, + 506, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 345 + ], + "score": 1.0, + "content": "Contributions. In summary, our contributions are the following: (i) We introduce the Flamingo", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "family of VLMs which can perform various multimodal tasks (such as captioning, visual dialogue,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 353, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 366 + ], + "score": 1.0, + "content": "or visual question-answering) from only a few input/output examples. Thanks to architectural", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 364, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 506, + 377 + ], + "score": 1.0, + "content": "innovations, the Flamingo models can efficiently accept arbitrarily interleaved visual data and text", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "as input and generate text in an open-ended manner. (ii) We quantitatively evaluate how Flamingo", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 387, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 506, + 398 + ], + "score": 1.0, + "content": "models can be adapted to various tasks via few-shot learning. We notably reserve a large set of held-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "score": 1.0, + "content": "out benchmarks which have not been used for validation of any design decisions or hyperparameters", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 407, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 421 + ], + "score": 1.0, + "content": "of the approach. We use these to estimate unbiased few-shot performance. (iii) Flamingo sets a new", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "state of the art in few-shot learning on a wide array of 16 multimodal language and image/video", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "understanding tasks. On 6 of these 16 tasks, Flamingo also outperforms the fine-tuned state of the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "art despite using only 32 task-specific examples, around 1000 times less task-specific training data", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "than the current state of the art. With a larger annotation budget, Flamingo can also be effectively", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "fine-tuned to set a new state of the art on five additional challenging benchmarks: VQAv2, VATEX,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 473, + 281, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 281, + 486 + ], + "score": 1.0, + "content": "VizWiz, MSRVTTQA, and HatefulMemes.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 107, + 500, + 176, + 514 + ], + "lines": [ + { + "bbox": [ + 103, + 497, + 178, + 517 + ], + "spans": [ + { + "bbox": [ + 103, + 497, + 178, + 517 + ], + "score": 1.0, + "content": "2 Approach", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 524, + 506, + 624 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "This section describes Flamingo: a visual language model that accepts text interleaved with im-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 536, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 506, + 548 + ], + "score": 1.0, + "content": "ages/videos as input and outputs free-form text. The key architectural components shown in Figure 3", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 547, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 506, + 558 + ], + "score": 1.0, + "content": "are chosen to leverage pretrained vision and language models and bridge them effectively. First,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 558, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 506, + 570 + ], + "score": 1.0, + "content": "the Perceiver Resampler (Section 2.1) receives spatio-temporal features from the Vision Encoder", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "score": 1.0, + "content": "(obtained from either an image or a video) and outputs a fixed number of visual tokens. Second,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "score": 1.0, + "content": "these visual tokens are used to condition the frozen LM using freshly initialised cross-attention", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "layers (Section 2.2) that are interleaved between the pretrained LM layers. These new layers offer", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 601, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 613 + ], + "score": 1.0, + "content": "an expressive way for the LM to incorporate visual information for the next-token prediction task.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 613, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 262, + 624 + ], + "score": 1.0, + "content": "Flamingo models the likelihood of text", + "type": "text" + }, + { + "bbox": [ + 263, + 614, + 269, + 624 + ], + "score": 0.81, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 613, + 452, + 624 + ], + "score": 1.0, + "content": "conditioned on interleaved images and videos", + "type": "text" + }, + { + "bbox": [ + 453, + 614, + 460, + 622 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 613, + 506, + 624 + ], + "score": 1.0, + "content": "as follows:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 628, + 365, + 663 + ], + "lines": [ + { + "bbox": [ + 246, + 628, + 365, + 663 + ], + "spans": [ + { + "bbox": [ + 246, + 628, + 365, + 663 + ], + "score": 0.95, + "content": "p ( \\boldsymbol y | \\boldsymbol x ) = \\prod _ { \\ell = 1 } ^ { L } p ( \\boldsymbol y _ { \\ell } | \\boldsymbol y _ { \\angle \\ell } , \\boldsymbol x _ { \\le \\ell } ) ,", + "type": "interline_equation", + "image_path": "baec868931df7df3985395e958c441d5673ad914a1ca464c0cc45ca8b1344dd8.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 246, + 628, + 365, + 645.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 246, + 645.5, + 365, + 663.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 132, + 680 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 669, + 142, + 679 + ], + "score": 0.84, + "content": "y _ { \\ell }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 666, + 167, + 680 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 167, + 668, + 173, + 677 + ], + "score": 0.71, + "content": "\\ell", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 666, + 314, + 680 + ], + "score": 1.0, + "content": "-th language token of the input text,", + "type": "text" + }, + { + "bbox": [ + 314, + 669, + 330, + 679 + ], + "score": 0.83, + "content": "y _ { < \\ell }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 666, + 450, + 680 + ], + "score": 1.0, + "content": "is the set of preceding tokens,", + "type": "text" + }, + { + "bbox": [ + 450, + 669, + 467, + 679 + ], + "score": 0.87, + "content": "x _ { \\le \\ell }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "is the set", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 240, + 691 + ], + "score": 1.0, + "content": "of images/videos preceding token", + "type": "text" + }, + { + "bbox": [ + 241, + 682, + 250, + 690 + ], + "score": 0.71, + "content": "y _ { \\ell }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 678, + 376, + 691 + ], + "score": 1.0, + "content": "in the interleaved sequence and", + "type": "text" + }, + { + "bbox": [ + 376, + 680, + 382, + 690 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "is parametrized by a Flamingo", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "model. The ability to handle interleaved text and visual sequences (Section 2.3) makes it natural", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "to use Flamingo models for in-context few-shot learning, analogously to GPT-3 with few-shot text", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 711, + 478, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 478, + 723 + ], + "score": 1.0, + "content": "prompting. The model is trained on a diverse mixture of datasets as described in Section 2.4.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 11, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 70, + 505, + 243 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 70, + 505, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 70, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 112, + 70, + 505, + 243 + ], + "score": 0.973, + "type": "image", + "image_path": "f1607ad89c4ea7b40f8723498b439571ecfb75b7d7f622105db2d86667955992.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 70, + 505, + 127.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 127.66666666666666, + 505, + 185.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 185.33333333333331, + 505, + 242.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 250, + 505, + 273 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "Figure 3: Flamingo architecture overview. Flamingo is a family of visual language models (VLMs)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 260, + 456, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 456, + 274 + ], + "score": 1.0, + "content": "that take as input visual data interleaved with text and produce free-form text as output.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 293, + 505, + 327 + ], + "lines": [], + "index": 6, + "bbox_fs": [ + 105, + 292, + 505, + 329 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 331, + 506, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 345 + ], + "score": 1.0, + "content": "Contributions. In summary, our contributions are the following: (i) We introduce the Flamingo", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "family of VLMs which can perform various multimodal tasks (such as captioning, visual dialogue,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 353, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 366 + ], + "score": 1.0, + "content": "or visual question-answering) from only a few input/output examples. Thanks to architectural", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 364, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 506, + 377 + ], + "score": 1.0, + "content": "innovations, the Flamingo models can efficiently accept arbitrarily interleaved visual data and text", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "as input and generate text in an open-ended manner. (ii) We quantitatively evaluate how Flamingo", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 387, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 506, + 398 + ], + "score": 1.0, + "content": "models can be adapted to various tasks via few-shot learning. We notably reserve a large set of held-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "score": 1.0, + "content": "out benchmarks which have not been used for validation of any design decisions or hyperparameters", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 407, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 421 + ], + "score": 1.0, + "content": "of the approach. We use these to estimate unbiased few-shot performance. (iii) Flamingo sets a new", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "state of the art in few-shot learning on a wide array of 16 multimodal language and image/video", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "understanding tasks. On 6 of these 16 tasks, Flamingo also outperforms the fine-tuned state of the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "art despite using only 32 task-specific examples, around 1000 times less task-specific training data", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "than the current state of the art. With a larger annotation budget, Flamingo can also be effectively", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "fine-tuned to set a new state of the art on five additional challenging benchmarks: VQAv2, VATEX,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 473, + 281, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 281, + 486 + ], + "score": 1.0, + "content": "VizWiz, MSRVTTQA, and HatefulMemes.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 330, + 506, + 486 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 500, + 176, + 514 + ], + "lines": [ + { + "bbox": [ + 103, + 497, + 178, + 517 + ], + "spans": [ + { + "bbox": [ + 103, + 497, + 178, + 517 + ], + "score": 1.0, + "content": "2 Approach", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 524, + 506, + 624 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "This section describes Flamingo: a visual language model that accepts text interleaved with im-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 536, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 506, + 548 + ], + "score": 1.0, + "content": "ages/videos as input and outputs free-form text. The key architectural components shown in Figure 3", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 547, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 506, + 558 + ], + "score": 1.0, + "content": "are chosen to leverage pretrained vision and language models and bridge them effectively. First,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 558, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 506, + 570 + ], + "score": 1.0, + "content": "the Perceiver Resampler (Section 2.1) receives spatio-temporal features from the Vision Encoder", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "score": 1.0, + "content": "(obtained from either an image or a video) and outputs a fixed number of visual tokens. Second,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "score": 1.0, + "content": "these visual tokens are used to condition the frozen LM using freshly initialised cross-attention", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "layers (Section 2.2) that are interleaved between the pretrained LM layers. These new layers offer", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 601, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 613 + ], + "score": 1.0, + "content": "an expressive way for the LM to incorporate visual information for the next-token prediction task.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 613, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 262, + 624 + ], + "score": 1.0, + "content": "Flamingo models the likelihood of text", + "type": "text" + }, + { + "bbox": [ + 263, + 614, + 269, + 624 + ], + "score": 0.81, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 613, + 452, + 624 + ], + "score": 1.0, + "content": "conditioned on interleaved images and videos", + "type": "text" + }, + { + "bbox": [ + 453, + 614, + 460, + 622 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 613, + 506, + 624 + ], + "score": 1.0, + "content": "as follows:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 524, + 506, + 624 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 628, + 365, + 663 + ], + "lines": [ + { + "bbox": [ + 246, + 628, + 365, + 663 + ], + "spans": [ + { + "bbox": [ + 246, + 628, + 365, + 663 + ], + "score": 0.95, + "content": "p ( \\boldsymbol y | \\boldsymbol x ) = \\prod _ { \\ell = 1 } ^ { L } p ( \\boldsymbol y _ { \\ell } | \\boldsymbol y _ { \\angle \\ell } , \\boldsymbol x _ { \\le \\ell } ) ,", + "type": "interline_equation", + "image_path": "baec868931df7df3985395e958c441d5673ad914a1ca464c0cc45ca8b1344dd8.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 246, + 628, + 365, + 645.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 246, + 645.5, + 365, + 663.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 132, + 680 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 669, + 142, + 679 + ], + "score": 0.84, + "content": "y _ { \\ell }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 666, + 167, + 680 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 167, + 668, + 173, + 677 + ], + "score": 0.71, + "content": "\\ell", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 666, + 314, + 680 + ], + "score": 1.0, + "content": "-th language token of the input text,", + "type": "text" + }, + { + "bbox": [ + 314, + 669, + 330, + 679 + ], + "score": 0.83, + "content": "y _ { < \\ell }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 666, + 450, + 680 + ], + "score": 1.0, + "content": "is the set of preceding tokens,", + "type": "text" + }, + { + "bbox": [ + 450, + 669, + 467, + 679 + ], + "score": 0.87, + "content": "x _ { \\le \\ell }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "is the set", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 240, + 691 + ], + "score": 1.0, + "content": "of images/videos preceding token", + "type": "text" + }, + { + "bbox": [ + 241, + 682, + 250, + 690 + ], + "score": 0.71, + "content": "y _ { \\ell }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 678, + 376, + 691 + ], + "score": 1.0, + "content": "in the interleaved sequence and", + "type": "text" + }, + { + "bbox": [ + 376, + 680, + 382, + 690 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "is parametrized by a Flamingo", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "model. The ability to handle interleaved text and visual sequences (Section 2.3) makes it natural", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "to use Flamingo models for in-context few-shot learning, analogously to GPT-3 with few-shot text", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 711, + 478, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 478, + 723 + ], + "score": 1.0, + "content": "prompting. The model is trained on a diverse mixture of datasets as described in Section 2.4.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 666, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 72, + 504, + 229 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 72, + 504, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 72, + 504, + 229 + ], + "spans": [ + { + "bbox": [ + 107, + 72, + 504, + 229 + ], + "score": 0.97, + "type": "image", + "image_path": "88258190e789163c53cdf24613ff85d84b6075a3d7a575194ef67f24043af676.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 72, + 504, + 124.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 124.33333333333334, + 504, + 176.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 176.66666666666669, + 504, + 229.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 235, + 506, + 290 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 235, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 247 + ], + "score": 1.0, + "content": "Figure 4: GATED XATTN-DENSE layers. To condition the LM on visual inputs, we insert new", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "cross-attention layers between existing pretrained and frozen LM layers. The keys and values in these", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 257, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 506, + 269 + ], + "score": 1.0, + "content": "layers are obtained from the vision features while the queries are derived from the language inputs.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 267, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 506, + 281 + ], + "score": 1.0, + "content": "They are followed by dense feed-forward layers. These layers are gated so that the LM is kept intact", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 277, + 330, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 330, + 291 + ], + "score": 1.0, + "content": "at initialization for improved stability and performance.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 107, + 312, + 329, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 311, + 330, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 330, + 326 + ], + "score": 1.0, + "content": "2.1 Visual processing and the Perceiver Resampler", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "score": 1.0, + "content": "Vision Encoder: from pixels to features. Our vision encoder is a pretrained and frozen Normalizer-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "Free ResNet (NFNet) [10] – we use the F6 model. We pretrain the vision encoder using a contrastive", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "objective on our datasets of image and text pairs, using the two-term contrastive loss from Radford", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "et al. [85]. We use the output of the final stage, a 2D spatial grid of features that is flattened to a 1D", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "sequence. For video inputs, frames are sampled at 1 FPS and encoded independently to obtain a 3D", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "spatio-temporal grid of features to which learned temporal embeddings are added. Features are then", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "flattened to 1D before being fed to the Perceiver Resampler. More details on the contrastive model", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 409, + 461, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 461, + 421 + ], + "score": 1.0, + "content": "training and performance are given in Appendix B.1.3 and Appendix B.3.2, respectively.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 505, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "score": 1.0, + "content": "Perceiver Resampler: from varying-size large feature maps to few visual tokens. This module", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 436, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 506, + 448 + ], + "score": 1.0, + "content": "connects the vision encoder to the frozen language model as shown in Figure 3. It takes as input a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 447, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 459 + ], + "score": 1.0, + "content": "variable number of image or video features from the vision encoder and produces a fixed number of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "visual outputs (64), reducing the computational complexity of the vision-text cross-attention. Similar", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "score": 1.0, + "content": "to Perceiver [48] and DETR [13], we learn a predefined number of latent input queries which are fed", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "score": 1.0, + "content": "to a Transformer and cross-attend to the visual features. We show in our ablation studies (Section 3.3)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 490, + 507, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 507, + 504 + ], + "score": 1.0, + "content": "that using such a vision-language resampler module outperforms a plain Transformer and an MLP.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 501, + 469, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 469, + 514 + ], + "score": 1.0, + "content": "We provide an illustration, more architectural details, and pseudo-code in Appendix A.1.1.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 107, + 526, + 397, + 538 + ], + "lines": [ + { + "bbox": [ + 104, + 525, + 399, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 525, + 399, + 542 + ], + "score": 1.0, + "content": "2.2 Conditioning frozen language models on visual representations", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 109, + 547, + 502, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 560 + ], + "score": 1.0, + "content": "Text generation is performed by a Transformer decoder, conditioned on the visual representations", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 559, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 570 + ], + "score": 1.0, + "content": "produced by the Perceiver Resampler. We interleave pretrained and frozen text-only LM blocks with", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 568, + 488, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 488, + 581 + ], + "score": 1.0, + "content": "blocks trained from scratch that cross-attend to the visual output from the Perceiver Resampler.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 585, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "Interleaving new GATED XATTN-DENSE layers within a frozen pretrained LM. We freeze the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "pretrained LM blocks, and insert gated cross-attention dense blocks (Figure 4) between the original", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "layers, trained from scratch. To ensure that at initialization, the conditioned model yields the same", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 618, + 504, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 504, + 630 + ], + "score": 1.0, + "content": "results as the original language model, we use a tanh-gating mechanism [41]. This multiplies the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 629, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 234, + 642 + ], + "score": 1.0, + "content": "output of a newly added layer by", + "type": "text" + }, + { + "bbox": [ + 235, + 629, + 270, + 641 + ], + "score": 0.58, + "content": "\\operatorname { t a n h } ( \\alpha )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 629, + 505, + 642 + ], + "score": 1.0, + "content": "before adding it to the input representation from the residual", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 640, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 181, + 651 + ], + "score": 1.0, + "content": "connection, where", + "type": "text" + }, + { + "bbox": [ + 181, + 641, + 189, + 650 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 640, + 505, + 651 + ], + "score": 1.0, + "content": "is a layer-specific learnable scalar initialized to 0 [4]. Thus, at initialization, the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 650, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 663 + ], + "score": 1.0, + "content": "model output matches that of the pretrained LM, improving training stability and final performance.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "score": 1.0, + "content": "In our ablation studies (Section 3.3), we compare the proposed GATED XATTN-DENSE layers against", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "recent alternatives [22, 68] and explore the effect of how frequently these additional layers are inserted", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 683, + 451, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 451, + 695 + ], + "score": 1.0, + "content": "to trade off between efficiency and expressivity. See Appendix A.1.2 for more details.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "Varying model sizes. We perform experiments across three models sizes, building on the 1.4B, 7B,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "and 70B parameter Chinchilla models [42]; calling them respectively Flamingo-3B, Flamingo-9B and", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 72, + 504, + 229 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 72, + 504, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 72, + 504, + 229 + ], + "spans": [ + { + "bbox": [ + 107, + 72, + 504, + 229 + ], + "score": 0.97, + "type": "image", + "image_path": "88258190e789163c53cdf24613ff85d84b6075a3d7a575194ef67f24043af676.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 72, + 504, + 124.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 124.33333333333334, + 504, + 176.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 176.66666666666669, + 504, + 229.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 235, + 506, + 290 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 235, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 247 + ], + "score": 1.0, + "content": "Figure 4: GATED XATTN-DENSE layers. To condition the LM on visual inputs, we insert new", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "cross-attention layers between existing pretrained and frozen LM layers. The keys and values in these", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 257, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 506, + 269 + ], + "score": 1.0, + "content": "layers are obtained from the vision features while the queries are derived from the language inputs.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 267, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 506, + 281 + ], + "score": 1.0, + "content": "They are followed by dense feed-forward layers. These layers are gated so that the LM is kept intact", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 277, + 330, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 330, + 291 + ], + "score": 1.0, + "content": "at initialization for improved stability and performance.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 107, + 312, + 329, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 311, + 330, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 330, + 326 + ], + "score": 1.0, + "content": "2.1 Visual processing and the Perceiver Resampler", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "score": 1.0, + "content": "Vision Encoder: from pixels to features. Our vision encoder is a pretrained and frozen Normalizer-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "Free ResNet (NFNet) [10] – we use the F6 model. We pretrain the vision encoder using a contrastive", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "objective on our datasets of image and text pairs, using the two-term contrastive loss from Radford", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "et al. [85]. We use the output of the final stage, a 2D spatial grid of features that is flattened to a 1D", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "sequence. For video inputs, frames are sampled at 1 FPS and encoded independently to obtain a 3D", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "spatio-temporal grid of features to which learned temporal embeddings are added. Features are then", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "flattened to 1D before being fed to the Perceiver Resampler. More details on the contrastive model", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 409, + 461, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 461, + 421 + ], + "score": 1.0, + "content": "training and performance are given in Appendix B.1.3 and Appendix B.3.2, respectively.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 332, + 506, + 421 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 505, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "score": 1.0, + "content": "Perceiver Resampler: from varying-size large feature maps to few visual tokens. This module", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 436, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 506, + 448 + ], + "score": 1.0, + "content": "connects the vision encoder to the frozen language model as shown in Figure 3. It takes as input a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 447, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 459 + ], + "score": 1.0, + "content": "variable number of image or video features from the vision encoder and produces a fixed number of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "visual outputs (64), reducing the computational complexity of the vision-text cross-attention. Similar", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "score": 1.0, + "content": "to Perceiver [48] and DETR [13], we learn a predefined number of latent input queries which are fed", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "score": 1.0, + "content": "to a Transformer and cross-attend to the visual features. We show in our ablation studies (Section 3.3)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 490, + 507, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 507, + 504 + ], + "score": 1.0, + "content": "that using such a vision-language resampler module outperforms a plain Transformer and an MLP.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 501, + 469, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 469, + 514 + ], + "score": 1.0, + "content": "We provide an illustration, more architectural details, and pseudo-code in Appendix A.1.1.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 425, + 507, + 514 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 526, + 397, + 538 + ], + "lines": [ + { + "bbox": [ + 104, + 525, + 399, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 525, + 399, + 542 + ], + "score": 1.0, + "content": "2.2 Conditioning frozen language models on visual representations", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 109, + 547, + 502, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 560 + ], + "score": 1.0, + "content": "Text generation is performed by a Transformer decoder, conditioned on the visual representations", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 559, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 570 + ], + "score": 1.0, + "content": "produced by the Perceiver Resampler. We interleave pretrained and frozen text-only LM blocks with", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 568, + 488, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 488, + 581 + ], + "score": 1.0, + "content": "blocks trained from scratch that cross-attend to the visual output from the Perceiver Resampler.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 546, + 505, + 581 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 585, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "Interleaving new GATED XATTN-DENSE layers within a frozen pretrained LM. We freeze the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "pretrained LM blocks, and insert gated cross-attention dense blocks (Figure 4) between the original", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "layers, trained from scratch. To ensure that at initialization, the conditioned model yields the same", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 618, + 504, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 504, + 630 + ], + "score": 1.0, + "content": "results as the original language model, we use a tanh-gating mechanism [41]. This multiplies the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 629, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 234, + 642 + ], + "score": 1.0, + "content": "output of a newly added layer by", + "type": "text" + }, + { + "bbox": [ + 235, + 629, + 270, + 641 + ], + "score": 0.58, + "content": "\\operatorname { t a n h } ( \\alpha )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 629, + 505, + 642 + ], + "score": 1.0, + "content": "before adding it to the input representation from the residual", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 640, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 181, + 651 + ], + "score": 1.0, + "content": "connection, where", + "type": "text" + }, + { + "bbox": [ + 181, + 641, + 189, + 650 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 640, + 505, + 651 + ], + "score": 1.0, + "content": "is a layer-specific learnable scalar initialized to 0 [4]. Thus, at initialization, the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 650, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 663 + ], + "score": 1.0, + "content": "model output matches that of the pretrained LM, improving training stability and final performance.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "score": 1.0, + "content": "In our ablation studies (Section 3.3), we compare the proposed GATED XATTN-DENSE layers against", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "recent alternatives [22, 68] and explore the effect of how frequently these additional layers are inserted", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 683, + 451, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 451, + 695 + ], + "score": 1.0, + "content": "to trade off between efficiency and expressivity. See Appendix A.1.2 for more details.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 585, + 506, + 695 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "Varying model sizes. We perform experiments across three models sizes, building on the 1.4B, 7B,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "and 70B parameter Chinchilla models [42]; calling them respectively Flamingo-3B, Flamingo-9B and", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 699, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 104, + 71, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 71, + 506, + 87 + ], + "score": 1.0, + "content": "Flamingo-80B. For brevity, we refer to the last as Flamingo throughout the paper. While increasing", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "the parameter count of the frozen LM and the trainable vision-text GATED XATTN-DENSE modules,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 106 + ], + "score": 1.0, + "content": "we maintain a fixed-size frozen vision encoder and trainable Perceiver Resampler across the different", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 106, + 447, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 447, + 118 + ], + "score": 1.0, + "content": "models (small relative to the full model size). See Appendix B.1.1 for further details.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 107, + 131, + 397, + 143 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 399, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 399, + 145 + ], + "score": 1.0, + "content": "2.3 Multi-visual input support: per-image/video attention masking", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 151, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "The image-causal modelling introduced in Equation (1) is obtained by masking the full text-to-image", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 175 + ], + "score": 1.0, + "content": "cross-attention matrix, limiting which visual tokens the model sees at each text token. At a given text", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "token, the model attends to the visual tokens of the image that appeared just before it in the interleaved", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 185, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 197 + ], + "score": 1.0, + "content": "sequence, rather than to all previous images (formalized and illustrated in Appendix A.1.3). Though", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 194, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 209 + ], + "score": 1.0, + "content": "the model only directly attends to a single image at a time, the dependency on all previous images", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 206, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 505, + 219 + ], + "score": 1.0, + "content": "remains via self-attention in the LM. This single-image cross-attention scheme importantly allows", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "score": 1.0, + "content": "the model to seamlessly generalise to any number of visual inputs, regardless of how many are", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 227, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 506, + 241 + ], + "score": 1.0, + "content": "used during training. In particular, we use only up to 5 images per sequence when training on our", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "score": 1.0, + "content": "interleaved datasets, yet our model is able to benefit from sequences of up to 32 pairs (or “shots”) of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 250, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 262 + ], + "score": 1.0, + "content": "images/videos and corresponding texts during evaluation. We show in Section 3.3 that this scheme is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 261, + 450, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 450, + 273 + ], + "score": 1.0, + "content": "more effective than allowing the model to cross-attend to all previous images directly.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10 + }, + { + "type": "title", + "bbox": [ + 107, + 285, + 358, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 284, + 360, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 360, + 300 + ], + "score": 1.0, + "content": "2.4 Training on a mixture of vision and language datasets", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 108, + 306, + 504, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "We train the Flamingo models on a mixture of three kinds of datasets, all scraped from the web: an", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 318, + 495, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 495, + 329 + ], + "score": 1.0, + "content": "interleaved image and text dataset derived from webpages, image-text pairs, and video-text pairs.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 346 + ], + "score": 1.0, + "content": "M3W: Interleaved image and text dataset. The few-shot capabilities of Flamingo models rely on", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "training on interleaved text and image data. For this purpose, we collect the MultiModal MassiveWeb", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 356, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 368 + ], + "score": 1.0, + "content": "(M3W) dataset. We extract both text and images from the HTML of approximately 43 million", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "webpages, determining the positions of images relative to the text based on the relative positions of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 376, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 390 + ], + "score": 1.0, + "content": "the text and image elements in the Document Object Model (DOM). An example is then constructed", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 388, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 506, + 401 + ], + "score": 1.0, + "content": "by inserting tags in plain text at the locations of the images on the page, and inserting a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 398, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 104, + 398, + 136, + 412 + ], + "score": 1.0, + "content": "special", + "type": "text" + }, + { + "bbox": [ + 136, + 400, + 164, + 410 + ], + "score": 0.73, + "content": "\\mathtt { < E O C > }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 398, + 505, + 412 + ], + "score": 1.0, + "content": "(end of chunk) token (added to the vocabulary and learnt) prior to any image and at the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 423, + 423 + ], + "score": 1.0, + "content": "end of the document. From each document, we sample a random subsequence of", + "type": "text" + }, + { + "bbox": [ + 424, + 410, + 460, + 420 + ], + "score": 0.9, + "content": "L = 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "tokens and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 180, + 434 + ], + "score": 1.0, + "content": "take up to the first", + "type": "text" + }, + { + "bbox": [ + 180, + 421, + 209, + 431 + ], + "score": 0.9, + "content": "N = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "images included in the sampled sequence. Further images are discarded in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 432, + 377, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 377, + 444 + ], + "score": 1.0, + "content": "order to save compute. More details are provided in Appendix A.3.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 448, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "Pairs of image/video and text. For our image and text pairs we first leverage the ALIGN [50]", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "score": 1.0, + "content": "dataset, composed of 1.8 billion images paired with alt-text. To complement this dataset, we collect", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "our own dataset of image and text pairs targeting better quality and longer descriptions: LTIP (Long", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 482, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 506, + 493 + ], + "score": 1.0, + "content": "Text & Image Pairs) which consists of 312 million image and text pairs. We also collect a similar", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 492, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 505, + 504 + ], + "score": 1.0, + "content": "dataset but with videos instead of still images: VTP (Video & Text Pairs) consists of 27 million short", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "score": 1.0, + "content": "videos (approximately 22 seconds on average) paired with sentence descriptions. We align the syntax", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 444, + 527 + ], + "score": 1.0, + "content": "of paired datasets with the syntax of M3W by prepending and appending", + "type": "text" + }, + { + "bbox": [ + 445, + 515, + 472, + 524 + ], + "score": 0.81, + "content": "\\mathtt { < E O C > }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "to each", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 525, + 305, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 305, + 537 + ], + "score": 1.0, + "content": "training caption (see Appendix A.3.3 for details).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 502, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "Multi-objective training and optimisation strategy. We train our models by minimizing a weighted", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 552, + 444, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 444, + 564 + ], + "score": 1.0, + "content": "sum of per-dataset expected negative log-likelihoods of text, given the visual inputs:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 569, + 406, + 604 + ], + "lines": [ + { + "bbox": [ + 203, + 569, + 406, + 604 + ], + "spans": [ + { + "bbox": [ + 203, + 569, + 406, + 604 + ], + "score": 0.94, + "content": "\\sum _ { m = 1 } ^ { M } \\lambda _ { m } \\cdot \\mathbb { E } _ { ( x , y ) \\sim \\mathcal { D } _ { m } } \\left[ - \\sum _ { \\ell = 1 } ^ { L } \\log p ( y _ { \\ell } | y _ { < \\ell } , x _ { \\le \\ell } ) \\right] ,", + "type": "interline_equation", + "image_path": "5ccd45cd77c6678250ca4081b93e33cd3118a1887d48169b1637767d9078fb54.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 569, + 406, + 586.5 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 203, + 586.5, + 406, + 604.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 132, + 623 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 611, + 149, + 621 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 609, + 165, + 623 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 166, + 611, + 180, + 621 + ], + "score": 0.9, + "content": "\\lambda _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 609, + 209, + 623 + ], + "score": 1.0, + "content": "are the", + "type": "text" + }, + { + "bbox": [ + 209, + 612, + 219, + 620 + ], + "score": 0.83, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 609, + 505, + 623 + ], + "score": 1.0, + "content": "-th dataset and its weighting, respectively. Tuning the per-dataset weights", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 621, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 107, + 622, + 121, + 632 + ], + "score": 0.88, + "content": "\\lambda _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 621, + 505, + 633 + ], + "score": 1.0, + "content": "is key to performance. We accumulate gradients over all datasets, which we found outperforms a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 631, + 499, + 645 + ], + "spans": [ + { + "bbox": [ + 104, + 631, + 499, + 645 + ], + "score": 1.0, + "content": "“round-robin” approach [17]. We provide further training details and ablations in Appendix B.1.2.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "title", + "bbox": [ + 106, + 657, + 343, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 343, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 343, + 671 + ], + "score": 1.0, + "content": "2.5 Task adaptation with few-shot in-context learning", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 689 + ], + "score": 1.0, + "content": "Once Flamingo is trained, we use it to tackle a visual task by conditioning it on a multimodal", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "interleaved prompt. We evaluate the ability of our models to rapidly adapt to new tasks using in-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "context learning, analogously to GPT-3 [11], by interleaving support example pairs in the form", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 118, + 723 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 711, + 175, + 723 + ], + "score": 0.36, + "content": "( i m a g e , t e x t )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "or (??????????, ????????), followed by the query visual input, to build a prompt (details in", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 104, + 71, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 71, + 506, + 87 + ], + "score": 1.0, + "content": "Flamingo-80B. For brevity, we refer to the last as Flamingo throughout the paper. While increasing", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "the parameter count of the frozen LM and the trainable vision-text GATED XATTN-DENSE modules,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 106 + ], + "score": 1.0, + "content": "we maintain a fixed-size frozen vision encoder and trainable Perceiver Resampler across the different", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 106, + 447, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 447, + 118 + ], + "score": 1.0, + "content": "models (small relative to the full model size). See Appendix B.1.1 for further details.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 104, + 71, + 506, + 118 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 131, + 397, + 143 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 399, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 399, + 145 + ], + "score": 1.0, + "content": "2.3 Multi-visual input support: per-image/video attention masking", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 151, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "The image-causal modelling introduced in Equation (1) is obtained by masking the full text-to-image", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 175 + ], + "score": 1.0, + "content": "cross-attention matrix, limiting which visual tokens the model sees at each text token. At a given text", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "token, the model attends to the visual tokens of the image that appeared just before it in the interleaved", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 185, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 197 + ], + "score": 1.0, + "content": "sequence, rather than to all previous images (formalized and illustrated in Appendix A.1.3). Though", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 194, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 209 + ], + "score": 1.0, + "content": "the model only directly attends to a single image at a time, the dependency on all previous images", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 206, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 505, + 219 + ], + "score": 1.0, + "content": "remains via self-attention in the LM. This single-image cross-attention scheme importantly allows", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "score": 1.0, + "content": "the model to seamlessly generalise to any number of visual inputs, regardless of how many are", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 227, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 506, + 241 + ], + "score": 1.0, + "content": "used during training. In particular, we use only up to 5 images per sequence when training on our", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "score": 1.0, + "content": "interleaved datasets, yet our model is able to benefit from sequences of up to 32 pairs (or “shots”) of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 250, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 262 + ], + "score": 1.0, + "content": "images/videos and corresponding texts during evaluation. We show in Section 3.3 that this scheme is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 261, + 450, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 450, + 273 + ], + "score": 1.0, + "content": "more effective than allowing the model to cross-attend to all previous images directly.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 151, + 506, + 273 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 285, + 358, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 284, + 360, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 360, + 300 + ], + "score": 1.0, + "content": "2.4 Training on a mixture of vision and language datasets", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 108, + 306, + 504, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "We train the Flamingo models on a mixture of three kinds of datasets, all scraped from the web: an", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 318, + 495, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 495, + 329 + ], + "score": 1.0, + "content": "interleaved image and text dataset derived from webpages, image-text pairs, and video-text pairs.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 306, + 505, + 329 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 346 + ], + "score": 1.0, + "content": "M3W: Interleaved image and text dataset. The few-shot capabilities of Flamingo models rely on", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "training on interleaved text and image data. For this purpose, we collect the MultiModal MassiveWeb", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 356, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 368 + ], + "score": 1.0, + "content": "(M3W) dataset. We extract both text and images from the HTML of approximately 43 million", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "webpages, determining the positions of images relative to the text based on the relative positions of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 376, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 390 + ], + "score": 1.0, + "content": "the text and image elements in the Document Object Model (DOM). An example is then constructed", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 388, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 506, + 401 + ], + "score": 1.0, + "content": "by inserting tags in plain text at the locations of the images on the page, and inserting a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 398, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 104, + 398, + 136, + 412 + ], + "score": 1.0, + "content": "special", + "type": "text" + }, + { + "bbox": [ + 136, + 400, + 164, + 410 + ], + "score": 0.73, + "content": "\\mathtt { < E O C > }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 398, + 505, + 412 + ], + "score": 1.0, + "content": "(end of chunk) token (added to the vocabulary and learnt) prior to any image and at the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 423, + 423 + ], + "score": 1.0, + "content": "end of the document. From each document, we sample a random subsequence of", + "type": "text" + }, + { + "bbox": [ + 424, + 410, + 460, + 420 + ], + "score": 0.9, + "content": "L = 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "tokens and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 180, + 434 + ], + "score": 1.0, + "content": "take up to the first", + "type": "text" + }, + { + "bbox": [ + 180, + 421, + 209, + 431 + ], + "score": 0.9, + "content": "N = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "images included in the sampled sequence. Further images are discarded in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 432, + 377, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 377, + 444 + ], + "score": 1.0, + "content": "order to save compute. More details are provided in Appendix A.3.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 333, + 506, + 444 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 448, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "Pairs of image/video and text. For our image and text pairs we first leverage the ALIGN [50]", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "score": 1.0, + "content": "dataset, composed of 1.8 billion images paired with alt-text. To complement this dataset, we collect", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "our own dataset of image and text pairs targeting better quality and longer descriptions: LTIP (Long", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 482, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 506, + 493 + ], + "score": 1.0, + "content": "Text & Image Pairs) which consists of 312 million image and text pairs. We also collect a similar", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 492, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 505, + 504 + ], + "score": 1.0, + "content": "dataset but with videos instead of still images: VTP (Video & Text Pairs) consists of 27 million short", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "score": 1.0, + "content": "videos (approximately 22 seconds on average) paired with sentence descriptions. We align the syntax", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 444, + 527 + ], + "score": 1.0, + "content": "of paired datasets with the syntax of M3W by prepending and appending", + "type": "text" + }, + { + "bbox": [ + 445, + 515, + 472, + 524 + ], + "score": 0.81, + "content": "\\mathtt { < E O C > }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "to each", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 525, + 305, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 305, + 537 + ], + "score": 1.0, + "content": "training caption (see Appendix A.3.3 for details).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 448, + 506, + 537 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 502, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "Multi-objective training and optimisation strategy. We train our models by minimizing a weighted", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 552, + 444, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 444, + 564 + ], + "score": 1.0, + "content": "sum of per-dataset expected negative log-likelihoods of text, given the visual inputs:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 540, + 505, + 564 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 569, + 406, + 604 + ], + "lines": [ + { + "bbox": [ + 203, + 569, + 406, + 604 + ], + "spans": [ + { + "bbox": [ + 203, + 569, + 406, + 604 + ], + "score": 0.94, + "content": "\\sum _ { m = 1 } ^ { M } \\lambda _ { m } \\cdot \\mathbb { E } _ { ( x , y ) \\sim \\mathcal { D } _ { m } } \\left[ - \\sum _ { \\ell = 1 } ^ { L } \\log p ( y _ { \\ell } | y _ { < \\ell } , x _ { \\le \\ell } ) \\right] ,", + "type": "interline_equation", + "image_path": "5ccd45cd77c6678250ca4081b93e33cd3118a1887d48169b1637767d9078fb54.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 569, + 406, + 586.5 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 203, + 586.5, + 406, + 604.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 132, + 623 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 611, + 149, + 621 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 609, + 165, + 623 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 166, + 611, + 180, + 621 + ], + "score": 0.9, + "content": "\\lambda _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 609, + 209, + 623 + ], + "score": 1.0, + "content": "are the", + "type": "text" + }, + { + "bbox": [ + 209, + 612, + 219, + 620 + ], + "score": 0.83, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 609, + 505, + 623 + ], + "score": 1.0, + "content": "-th dataset and its weighting, respectively. Tuning the per-dataset weights", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 621, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 107, + 622, + 121, + 632 + ], + "score": 0.88, + "content": "\\lambda _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 621, + 505, + 633 + ], + "score": 1.0, + "content": "is key to performance. We accumulate gradients over all datasets, which we found outperforms a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 631, + 499, + 645 + ], + "spans": [ + { + "bbox": [ + 104, + 631, + 499, + 645 + ], + "score": 1.0, + "content": "“round-robin” approach [17]. We provide further training details and ablations in Appendix B.1.2.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 104, + 609, + 505, + 645 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 657, + 343, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 343, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 343, + 671 + ], + "score": 1.0, + "content": "2.5 Task adaptation with few-shot in-context learning", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 689 + ], + "score": 1.0, + "content": "Once Flamingo is trained, we use it to tackle a visual task by conditioning it on a multimodal", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "interleaved prompt. We evaluate the ability of our models to rapidly adapt to new tasks using in-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "context learning, analogously to GPT-3 [11], by interleaving support example pairs in the form", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 118, + 723 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 711, + 175, + 723 + ], + "score": 0.36, + "content": "( i m a g e , t e x t )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "or (??????????, ????????), followed by the query visual input, to build a prompt (details in", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 678, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 70, + 503, + 214 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 70, + 503, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 503, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 503, + 214 + ], + "score": 0.977, + "html": "
MethodFTShot OAAYA ZAAO Cco(A) VOAASSVA XIA() ZIMZT3MA ITLTTI(A)VOA!J ooJS TAIY( [PiI[T vY VOAXANR eeeeY
Zero/Few shot SOTAX[124][58][58][135][79] grrreieee
[34][114][143][85]
43.338.232.235.2=-19.2 012.2 0-39.411.6 0-[85] 66.1 (0)40.7
x）(16)(4) 49.2(0)0 27.540.100
Flamingo-3B041.273.028.960.611.032.755.8 64.639.646.130.1 32.721.3 22.453.7 53.658.4
X X4 3243.353.2 57.185.0 99.033.050.0 59.234.072.0 71.214.9 25.635.7 37.776.741.3 41.647.31
X45.9 44.751.879.442.6 30.239.545.5 28.861.513.735.255.041.847.3 48.030.6 31.826.1 23.056.3 57.057.9 -
X0 449.356.393.136.251.734.972.618.237.770.842.850.433.624.762.7-
Flamingo-9B51.060.4106.347.257.444.072.829.440.777.341.250.432.628.463.5-
X32 050.656.384.335.646.731.667.217.440.760.139.752.035.026.746.460.8
美457.463.1103.2 41.756.039.675.123.944.174.542.455.636.530.868.6-
FlamingoX3257.867.6113.852.365.149.875.431.045.386.842.255.637.933.570.0-
54.480.2143.347.976.357.267.4 [150]46.835.4 [135]138.7 [132]36.775.254.725.279.1
√(X)[34] (10K)[140] (444K)[124] [28] (500K) (27K)[153] (500K)[65] (20K)(30K)[51] (130K)(6K)(10K)[128] (46K)[79] (123K)[137] (20K)[129] (38K)[62] (9K)
", + "type": "table", + "image_path": "816c7e30db9daadd54c08b089a4dc8a6ba054108d69bf1833380cb63899414f0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 70, + 503, + 118.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 118.0, + 503, + 166.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 166.0, + 503, + 214.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 218, + 505, + 284 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "score": 1.0, + "content": "Table 1: Comparison to the state of the art. A single Flamingo model reaches the state of the art", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "on a wide array of image (I) and video (V) understanding tasks with few-shot learning, significantly", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "outperforming previous best zero- and few-shot methods with as few as four examples. More", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "score": 1.0, + "content": "importantly, using only 32 examples and without adapting any model weights, Flamingo outperforms", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "the current best methods – fine-tuned on thousands of annotated examples – on seven tasks. Best", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 272, + 375, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 375, + 284 + ], + "score": 1.0, + "content": "few-shot numbers are in bold, best numbers overall are underlined.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 507, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 507, + 311 + ], + "score": 1.0, + "content": "Appendix A.2). We perform open-ended evaluations using beam search for decoding, and close-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "score": 1.0, + "content": "ended evaluations using our model’s log-likelihood to score each possible answer. We explore", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "zero-shot generalization by prompting the model with two text-only examples from the task,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "with no corresponding images. Evaluation hyperparameters and additional details are given in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 342, + 175, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 175, + 354 + ], + "score": 1.0, + "content": "Appendix B.1.5.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 369, + 191, + 382 + ], + "lines": [ + { + "bbox": [ + 103, + 366, + 193, + 386 + ], + "spans": [ + { + "bbox": [ + 103, + 366, + 193, + 386 + ], + "score": 1.0, + "content": "3 Experiments", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 506, + 406 + ], + "score": 1.0, + "content": "Our goal is to develop models that can rapidly adapt to diverse and challenging tasks. For this, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "consider a wide array of 16 popular multimodal image/video and language benchmarks. In order to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 416, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 427 + ], + "score": 1.0, + "content": "validate model design decisions during the course of the project, 5 of these benchmarks were used as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "part of our development (DEV) set: COCO, OKVQA, VQAv2, MSVDQA and VATEX. Performance", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 438, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 449 + ], + "score": 1.0, + "content": "estimates on the DEV benchmarks may be biased, as a result of model selection. We note that this is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 446, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 462 + ], + "score": 1.0, + "content": "also the case for prior work which makes use of similar benchmarks to validate and ablate design", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 456, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 104, + 456, + 506, + 474 + ], + "score": 1.0, + "content": "decisions. To account for this, we report performance on an additional set of 11 benchmarks, spanning", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "captioning, video question-answering, as well as some less commonly explored capabilities such as", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "visual dialogue and multi-choice question-answering tasks. The evaluation benchmarks are described", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 491, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 506 + ], + "score": 1.0, + "content": "in Appendix B.1.4. We keep all evaluation hyperparameters fixed across all benchmarks. Depending", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "on the task, we use four few-shot prompt templates we describe in more detail in Appendix B.1.5.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "We emphasize that we do not validate any design decisions on these 11 benchmarks and use them", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 524, + 402, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 402, + 536 + ], + "score": 1.0, + "content": "solely to estimate unbiased few-shot learning performance of our models.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 552 + ], + "score": 1.0, + "content": "Concretely, estimating few-shot learning performance of a model involves prompting it with a set of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "score": 1.0, + "content": "support samples and evaluating it on a set of query samples. For the DEV benchmarks that are used", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 562, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 505, + 576 + ], + "score": 1.0, + "content": "both to validate design decisions and hyperparameters, as well as to report final performance, we", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 572, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 587 + ], + "score": 1.0, + "content": "therefore use four subsets: validation support, validation query, test support and test query. For other", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 584, + 504, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 504, + 597 + ], + "score": 1.0, + "content": "benchmarks, we need only the latter two. We report in Appendix B.1.4 how we form these subsets.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 601, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 612 + ], + "score": 1.0, + "content": "We report the results of the Flamingo models on few-shot learning in Section 3.1. Section 3.2", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "gives Flamingo fine-tuned results. An ablation study is given in Section 3.3. Appendix B.2 provides", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 623, + 507, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 507, + 635 + ], + "score": 1.0, + "content": "more results including Flamingo’s performance on the ImageNet and Kinetics700 classification tasks,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 634, + 491, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 491, + 646 + ], + "score": 1.0, + "content": "and on our contrastive model’s performance. Appendix C includes additional qualitative results.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 107, + 658, + 313, + 670 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 314, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 314, + 672 + ], + "score": 1.0, + "content": "3.1 Few-shot learning on vision-language tasks", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "Few-shot results. Results are given in Table 1. Flamingo outperforms by a large margin all previous", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "score": 1.0, + "content": "zero-shot or few-shot methods on the 16 benchmarks considered. This is achieved with as few as four", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "examples per task, demonstrating practical and efficient adaptation of vision models to new tasks.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "More importantly, Flamingo is often competitive with state-of-the-art methods additionally fine-tuned", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 753 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 753 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 70, + 503, + 214 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 70, + 503, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 503, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 503, + 214 + ], + "score": 0.977, + "html": "
MethodFTShot OAAYA ZAAO Cco(A) VOAASSVA XIA() ZIMZT3MA ITLTTI(A)VOA!J ooJS TAIY( [PiI[T vY VOAXANR eeeeY
Zero/Few shot SOTAX[124][58][58][135][79] grrreieee
[34][114][143][85]
43.338.232.235.2=-19.2 012.2 0-39.411.6 0-[85] 66.1 (0)40.7
x）(16)(4) 49.2(0)0 27.540.100
Flamingo-3B041.273.028.960.611.032.755.8 64.639.646.130.1 32.721.3 22.453.7 53.658.4
X X4 3243.353.2 57.185.0 99.033.050.0 59.234.072.0 71.214.9 25.635.7 37.776.741.3 41.647.31
X45.9 44.751.879.442.6 30.239.545.5 28.861.513.735.255.041.847.3 48.030.6 31.826.1 23.056.3 57.057.9 -
X0 449.356.393.136.251.734.972.618.237.770.842.850.433.624.762.7-
Flamingo-9B51.060.4106.347.257.444.072.829.440.777.341.250.432.628.463.5-
X32 050.656.384.335.646.731.667.217.440.760.139.752.035.026.746.460.8
美457.463.1103.2 41.756.039.675.123.944.174.542.455.636.530.868.6-
FlamingoX3257.867.6113.852.365.149.875.431.045.386.842.255.637.933.570.0-
54.480.2143.347.976.357.267.4 [150]46.835.4 [135]138.7 [132]36.775.254.725.279.1
√(X)[34] (10K)[140] (444K)[124] [28] (500K) (27K)[153] (500K)[65] (20K)(30K)[51] (130K)(6K)(10K)[128] (46K)[79] (123K)[137] (20K)[129] (38K)[62] (9K)
", + "type": "table", + "image_path": "816c7e30db9daadd54c08b089a4dc8a6ba054108d69bf1833380cb63899414f0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 70, + 503, + 118.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 118.0, + 503, + 166.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 166.0, + 503, + 214.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 218, + 505, + 284 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "score": 1.0, + "content": "Table 1: Comparison to the state of the art. A single Flamingo model reaches the state of the art", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "on a wide array of image (I) and video (V) understanding tasks with few-shot learning, significantly", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "outperforming previous best zero- and few-shot methods with as few as four examples. More", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "score": 1.0, + "content": "importantly, using only 32 examples and without adapting any model weights, Flamingo outperforms", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "the current best methods – fine-tuned on thousands of annotated examples – on seven tasks. Best", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 272, + 375, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 375, + 284 + ], + "score": 1.0, + "content": "few-shot numbers are in bold, best numbers overall are underlined.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 507, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 507, + 311 + ], + "score": 1.0, + "content": "Appendix A.2). We perform open-ended evaluations using beam search for decoding, and close-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "score": 1.0, + "content": "ended evaluations using our model’s log-likelihood to score each possible answer. We explore", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "zero-shot generalization by prompting the model with two text-only examples from the task,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "with no corresponding images. Evaluation hyperparameters and additional details are given in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 342, + 175, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 175, + 354 + ], + "score": 1.0, + "content": "Appendix B.1.5.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 298, + 507, + 354 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 369, + 191, + 382 + ], + "lines": [ + { + "bbox": [ + 103, + 366, + 193, + 386 + ], + "spans": [ + { + "bbox": [ + 103, + 366, + 193, + 386 + ], + "score": 1.0, + "content": "3 Experiments", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 506, + 406 + ], + "score": 1.0, + "content": "Our goal is to develop models that can rapidly adapt to diverse and challenging tasks. For this, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "consider a wide array of 16 popular multimodal image/video and language benchmarks. In order to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 416, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 427 + ], + "score": 1.0, + "content": "validate model design decisions during the course of the project, 5 of these benchmarks were used as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "part of our development (DEV) set: COCO, OKVQA, VQAv2, MSVDQA and VATEX. Performance", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 438, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 449 + ], + "score": 1.0, + "content": "estimates on the DEV benchmarks may be biased, as a result of model selection. We note that this is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 446, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 462 + ], + "score": 1.0, + "content": "also the case for prior work which makes use of similar benchmarks to validate and ablate design", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 456, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 104, + 456, + 506, + 474 + ], + "score": 1.0, + "content": "decisions. To account for this, we report performance on an additional set of 11 benchmarks, spanning", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "captioning, video question-answering, as well as some less commonly explored capabilities such as", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "visual dialogue and multi-choice question-answering tasks. The evaluation benchmarks are described", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 491, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 506 + ], + "score": 1.0, + "content": "in Appendix B.1.4. We keep all evaluation hyperparameters fixed across all benchmarks. Depending", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "on the task, we use four few-shot prompt templates we describe in more detail in Appendix B.1.5.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "We emphasize that we do not validate any design decisions on these 11 benchmarks and use them", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 524, + 402, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 402, + 536 + ], + "score": 1.0, + "content": "solely to estimate unbiased few-shot learning performance of our models.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 21, + "bbox_fs": [ + 104, + 393, + 506, + 536 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 552 + ], + "score": 1.0, + "content": "Concretely, estimating few-shot learning performance of a model involves prompting it with a set of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "score": 1.0, + "content": "support samples and evaluating it on a set of query samples. For the DEV benchmarks that are used", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 562, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 505, + 576 + ], + "score": 1.0, + "content": "both to validate design decisions and hyperparameters, as well as to report final performance, we", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 572, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 587 + ], + "score": 1.0, + "content": "therefore use four subsets: validation support, validation query, test support and test query. For other", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 584, + 504, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 504, + 597 + ], + "score": 1.0, + "content": "benchmarks, we need only the latter two. We report in Appendix B.1.4 how we form these subsets.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 541, + 505, + 597 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 601, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 612 + ], + "score": 1.0, + "content": "We report the results of the Flamingo models on few-shot learning in Section 3.1. Section 3.2", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "gives Flamingo fine-tuned results. An ablation study is given in Section 3.3. Appendix B.2 provides", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 623, + 507, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 507, + 635 + ], + "score": 1.0, + "content": "more results including Flamingo’s performance on the ImageNet and Kinetics700 classification tasks,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 634, + 491, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 491, + 646 + ], + "score": 1.0, + "content": "and on our contrastive model’s performance. Appendix C includes additional qualitative results.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 601, + 507, + 646 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 658, + 313, + 670 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 314, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 314, + 672 + ], + "score": 1.0, + "content": "3.1 Few-shot learning on vision-language tasks", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "Few-shot results. Results are given in Table 1. Flamingo outperforms by a large margin all previous", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "score": 1.0, + "content": "zero-shot or few-shot methods on the 16 benchmarks considered. This is achieved with as few as four", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "examples per task, demonstrating practical and efficient adaptation of vision models to new tasks.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "More importantly, Flamingo is often competitive with state-of-the-art methods additionally fine-tuned", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "score": 1.0, + "content": "on up to hundreds of thousands of annotated examples. On six tasks, Flamingo even outperforms", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 393, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 405 + ], + "score": 1.0, + "content": "the fine-tuned SotA despite using a single set of model weights and only 32 task-specific examples.", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 404, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 416 + ], + "score": 1.0, + "content": "Finally, despite having only used the DEV benchmarks for design decisions, our results generalize", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 415, + 397, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 397, + 427 + ], + "score": 1.0, + "content": "well to the other benchmarks, confirming the generality of our approach.", + "type": "text", + "cross_page": true + } + ], + "index": 15 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 678, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 70, + 504, + 114 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 107, + 70, + 504, + 114 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 70, + 504, + 114 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 504, + 114 + ], + "score": 0.973, + "html": "
MethodVQAV2 test-stdCOCOVATEX testVizWizMSRVTTQA testVisDial valid|test-stdYouCook2TextVQA valid|test-stdHatefulMemes test seen
test-devtesttest-devtest-stdvalid
32 shots67.6：113.865.149.8-31.056.8-86.836.0·70.0
Fine-tuned82.082.1138.184.265.765.447.461.859.7118.657.154.186.6
SotA81.381.3149.681.457.2160.646.875.275.4138.754.773.784.6
[133][133][119][153][65][65][51][79][123][132][137][84][152]
", + "type": "table", + "image_path": "edfb5500208772b66c29347680dcd532d1aa04f52f5f2066608dc920f38abd8f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 70, + 504, + 84.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 84.66666666666667, + 504, + 99.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 99.33333333333334, + 504, + 114.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 119, + 505, + 164 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 118, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 104, + 118, + 505, + 132 + ], + "score": 1.0, + "content": "Table 2: Comparison to SotA when fine-tuning Flamingo. We fine-tune Flamingo on all nine", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "score": 1.0, + "content": "tasks where Flamingo does not achieve SotA with few-shot learning. Flamingo sets a new SotA on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 313, + 154 + ], + "score": 1.0, + "content": "five of them, outperfoming methods (marked with", + "type": "text" + }, + { + "bbox": [ + 313, + 142, + 320, + 153 + ], + "score": 0.39, + "content": "\\dagger .", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 142, + 505, + 154 + ], + "score": 1.0, + "content": ") that use tricks such as model ensembling or", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 153, + 362, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 362, + 164 + ], + "score": 1.0, + "content": "domain-specific metric optimisation (e.g., CIDEr optimisation).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "table", + "bbox": [ + 106, + 174, + 505, + 343 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 174, + 505, + 343 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 505, + 343 + ], + "score": 0.984, + "html": "
settingAblatedFlamingo-3B Changed original value valueParam. Step count↓ time↓CoCo CIDEr↑OKVQA top1个VQAv2 top1个MSVDQA top1个VATEX CIDEr↑Overall score↑
Flamingo-3B model w/o Video-Text pairs3.2B1.74s86.5 84.242.1 43.055.8 53.936.3 34.553.4 46.070.7 67.3
iTraining dataAll dataw/o Image-Text pairs Image-Text pairs→LAION w/oM3W3.2B 3.2B 3.2B 3.2B1.42s 0.95s 1.74s66.3 79.539.2 41.451.6 53.532.0 33.941.6 47.660.9 66.4
(ii)OptimisationAccumulationRound Robin3.2B1.02s 1.68s54.1 76.136.5 39.852.7 52.131.4 33.223.5 40.853.4 62.9
Tanh gating√X3.2B1.74s78.440.552.935.947.566.5
(iv)Cross-attention architectureGATED XATTN-DENSEVANILLA XATTN2.4B1.16s 1.74s80.6 79.241.553.4 50.832.9 32.250.7 47.866.9 63.1
Cross-attentionGRAFTING Single in middle3.3B 2.0B0.87s71.536.1 38.150.229.142.359.8
(v)frequencyEveryEvery 4th Every 2nd2.3B 2.6B1.02s 1.24s82.3 83.742.7 41.055.1 55.834.6 34.550.8 49.768.8 68.2
(vi)ResamplerPerceiverMLP Transformer3.2B 3.2B1.85s 1.81s78.6 83.242.2 41.754.7 55.635.2 31.544.7 48.366.6 66.7
(vii)Vision encoderNFNet-F6CLIP ViT-L/14 NFNet-F03.1B1.58s 1.45s76.5 73.841.653.4 52.833.2 31.144.5 42.964.9 62.7
(vii)X(random init)2.9B 3.2B2.42s74.840.5 31.545.626.950.157.8
Freezing LM√X (pretrained)3.2B2.42s81.233.747.431.053.962.7
", + "type": "table", + "image_path": "02fc77d1d276c1f24da67f57c3821e03b782797a088747233348da55d5370738.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 106, + 174, + 505, + 230.33333333333334 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 230.33333333333334, + 505, + 286.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 286.6666666666667, + 505, + 343.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 344, + 505, + 366 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "score": 1.0, + "content": "Table 3: Ablation studies. Each row should be compared to the baseline Flamingo run (top row).", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 354, + 453, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 453, + 367 + ], + "score": 1.0, + "content": "Step time measures the time spent to perform gradient updates on all training datasets.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + } + ], + "index": 9.25 + }, + { + "type": "text", + "bbox": [ + 107, + 382, + 505, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "score": 1.0, + "content": "on up to hundreds of thousands of annotated examples. On six tasks, Flamingo even outperforms", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 393, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 405 + ], + "score": 1.0, + "content": "the fine-tuned SotA despite using a single set of model weights and only 32 task-specific examples.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 404, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 416 + ], + "score": 1.0, + "content": "Finally, despite having only used the DEV benchmarks for design decisions, our results generalize", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 415, + 397, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 397, + 427 + ], + "score": 1.0, + "content": "well to the other benchmarks, confirming the generality of our approach.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 431, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 506, + 443 + ], + "score": 1.0, + "content": "Scaling with respect to parameters and shots. As shown in Figure 2, the larger the model, the better", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 442, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 506, + 455 + ], + "score": 1.0, + "content": "the few-shot performance, similar to GPT-3 [11]. The performance also improves with the number of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 452, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 466 + ], + "score": 1.0, + "content": "shots. We further find that the largest model better exploits larger numbers of shots. Interestingly,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 464, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 476 + ], + "score": 1.0, + "content": "even though our Flamingo models were trained with sequences limited to only 5 images on M3W,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "they are still able to benefit from up to 32 images or videos during inference. This demonstrates the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 485, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 480, + 498 + ], + "score": 1.0, + "content": "flexibility of the Flamingo architecture for processing a variable number of videos or images.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 108, + 510, + 387, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 508, + 389, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 389, + 525 + ], + "score": 1.0, + "content": "3.2 Fine-tuning Flamingo as a pretrained vision-language model", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 530, + 505, + 607 + ], + "lines": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "While not the main focus of our work, we verify that when given more data, Flamingo models can be", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "adapted to a task by fine-tuning their weights. In Table 2, we explore fine-tuning our largest model,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "Flamingo, for a given task with no limit on the annotation budget. In short, we do so by fine-tuning the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "model on a short schedule with a small learning rate by additionally unfreezing the vision backbone", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 573, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 586 + ], + "score": 1.0, + "content": "to accommodate a higher input resolution (details in Appendix B.2.2). We find that we can improve", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "results over our previously presented in-context few-shot learning results, setting a new state of the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 595, + 465, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 465, + 608 + ], + "score": 1.0, + "content": "art on five additional tasks: VQAv2, VATEX, VizWiz, MSRVTTQA, and HatefulMemes.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 107, + 619, + 199, + 631 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 200, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 200, + 632 + ], + "score": 1.0, + "content": "3.3 Ablation studies", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 504, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 506, + 651 + ], + "score": 1.0, + "content": "In Table 3, we report our ablation results using Flamingo-3B on the validation subsets of the five", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 650, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 662 + ], + "score": 1.0, + "content": "DEV benchmarks with 4 shots. Note that we use smaller batch sizes and a shorter training schedule", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "score": 1.0, + "content": "compared to the final models. The Overall score is obtained by dividing each benchmark score by its", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "score": 1.0, + "content": "state-of-the-art (SotA) performance from Table 1 and averaging the results. More details and results", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 683, + 270, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 270, + 695 + ], + "score": 1.0, + "content": "are given in Appendix B.3 and Table 10.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 502, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "Importance of the training data mixture. As shown in row (i), getting the right training data plays", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 709, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 104, + 709, + 505, + 724 + ], + "score": 1.0, + "content": "a crucial role. In fact, removing the interleaved image-text dataset M3W leads to a decrease of more", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 70, + 504, + 114 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 107, + 70, + 504, + 114 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 70, + 504, + 114 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 504, + 114 + ], + "score": 0.973, + "html": "
MethodVQAV2 test-stdCOCOVATEX testVizWizMSRVTTQA testVisDial valid|test-stdYouCook2TextVQA valid|test-stdHatefulMemes test seen
test-devtesttest-devtest-stdvalid
32 shots67.6：113.865.149.8-31.056.8-86.836.0·70.0
Fine-tuned82.082.1138.184.265.765.447.461.859.7118.657.154.186.6
SotA81.381.3149.681.457.2160.646.875.275.4138.754.773.784.6
[133][133][119][153][65][65][51][79][123][132][137][84][152]
", + "type": "table", + "image_path": "edfb5500208772b66c29347680dcd532d1aa04f52f5f2066608dc920f38abd8f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 70, + 504, + 84.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 84.66666666666667, + 504, + 99.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 99.33333333333334, + 504, + 114.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 119, + 505, + 164 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 118, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 104, + 118, + 505, + 132 + ], + "score": 1.0, + "content": "Table 2: Comparison to SotA when fine-tuning Flamingo. We fine-tune Flamingo on all nine", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "score": 1.0, + "content": "tasks where Flamingo does not achieve SotA with few-shot learning. Flamingo sets a new SotA on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 313, + 154 + ], + "score": 1.0, + "content": "five of them, outperfoming methods (marked with", + "type": "text" + }, + { + "bbox": [ + 313, + 142, + 320, + 153 + ], + "score": 0.39, + "content": "\\dagger .", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 142, + 505, + 154 + ], + "score": 1.0, + "content": ") that use tricks such as model ensembling or", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 153, + 362, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 362, + 164 + ], + "score": 1.0, + "content": "domain-specific metric optimisation (e.g., CIDEr optimisation).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "table", + "bbox": [ + 106, + 174, + 505, + 343 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 174, + 505, + 343 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 505, + 343 + ], + "score": 0.984, + "html": "
settingAblatedFlamingo-3B Changed original value valueParam. Step count↓ time↓CoCo CIDEr↑OKVQA top1个VQAv2 top1个MSVDQA top1个VATEX CIDEr↑Overall score↑
Flamingo-3B model w/o Video-Text pairs3.2B1.74s86.5 84.242.1 43.055.8 53.936.3 34.553.4 46.070.7 67.3
iTraining dataAll dataw/o Image-Text pairs Image-Text pairs→LAION w/oM3W3.2B 3.2B 3.2B 3.2B1.42s 0.95s 1.74s66.3 79.539.2 41.451.6 53.532.0 33.941.6 47.660.9 66.4
(ii)OptimisationAccumulationRound Robin3.2B1.02s 1.68s54.1 76.136.5 39.852.7 52.131.4 33.223.5 40.853.4 62.9
Tanh gating√X3.2B1.74s78.440.552.935.947.566.5
(iv)Cross-attention architectureGATED XATTN-DENSEVANILLA XATTN2.4B1.16s 1.74s80.6 79.241.553.4 50.832.9 32.250.7 47.866.9 63.1
Cross-attentionGRAFTING Single in middle3.3B 2.0B0.87s71.536.1 38.150.229.142.359.8
(v)frequencyEveryEvery 4th Every 2nd2.3B 2.6B1.02s 1.24s82.3 83.742.7 41.055.1 55.834.6 34.550.8 49.768.8 68.2
(vi)ResamplerPerceiverMLP Transformer3.2B 3.2B1.85s 1.81s78.6 83.242.2 41.754.7 55.635.2 31.544.7 48.366.6 66.7
(vii)Vision encoderNFNet-F6CLIP ViT-L/14 NFNet-F03.1B1.58s 1.45s76.5 73.841.653.4 52.833.2 31.144.5 42.964.9 62.7
(vii)X(random init)2.9B 3.2B2.42s74.840.5 31.545.626.950.157.8
Freezing LM√X (pretrained)3.2B2.42s81.233.747.431.053.962.7
", + "type": "table", + "image_path": "02fc77d1d276c1f24da67f57c3821e03b782797a088747233348da55d5370738.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 106, + 174, + 505, + 230.33333333333334 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 230.33333333333334, + 505, + 286.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 286.6666666666667, + 505, + 343.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 344, + 505, + 366 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "score": 1.0, + "content": "Table 3: Ablation studies. Each row should be compared to the baseline Flamingo run (top row).", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 354, + 453, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 453, + 367 + ], + "score": 1.0, + "content": "Step time measures the time spent to perform gradient updates on all training datasets.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + } + ], + "index": 9.25 + }, + { + "type": "text", + "bbox": [ + 107, + 382, + 505, + 426 + ], + "lines": [], + "index": 13.5, + "bbox_fs": [ + 105, + 382, + 506, + 427 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 431, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 506, + 443 + ], + "score": 1.0, + "content": "Scaling with respect to parameters and shots. As shown in Figure 2, the larger the model, the better", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 442, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 506, + 455 + ], + "score": 1.0, + "content": "the few-shot performance, similar to GPT-3 [11]. The performance also improves with the number of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 452, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 466 + ], + "score": 1.0, + "content": "shots. We further find that the largest model better exploits larger numbers of shots. Interestingly,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 464, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 476 + ], + "score": 1.0, + "content": "even though our Flamingo models were trained with sequences limited to only 5 images on M3W,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "they are still able to benefit from up to 32 images or videos during inference. This demonstrates the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 485, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 480, + 498 + ], + "score": 1.0, + "content": "flexibility of the Flamingo architecture for processing a variable number of videos or images.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 431, + 506, + 498 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 510, + 387, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 508, + 389, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 389, + 525 + ], + "score": 1.0, + "content": "3.2 Fine-tuning Flamingo as a pretrained vision-language model", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 530, + 505, + 607 + ], + "lines": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "While not the main focus of our work, we verify that when given more data, Flamingo models can be", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "adapted to a task by fine-tuning their weights. In Table 2, we explore fine-tuning our largest model,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "Flamingo, for a given task with no limit on the annotation budget. In short, we do so by fine-tuning the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "model on a short schedule with a small learning rate by additionally unfreezing the vision backbone", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 573, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 586 + ], + "score": 1.0, + "content": "to accommodate a higher input resolution (details in Appendix B.2.2). We find that we can improve", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "results over our previously presented in-context few-shot learning results, setting a new state of the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 595, + 465, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 465, + 608 + ], + "score": 1.0, + "content": "art on five additional tasks: VQAv2, VATEX, VizWiz, MSRVTTQA, and HatefulMemes.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 530, + 506, + 608 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 619, + 199, + 631 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 200, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 200, + 632 + ], + "score": 1.0, + "content": "3.3 Ablation studies", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 504, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 506, + 651 + ], + "score": 1.0, + "content": "In Table 3, we report our ablation results using Flamingo-3B on the validation subsets of the five", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 650, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 662 + ], + "score": 1.0, + "content": "DEV benchmarks with 4 shots. Note that we use smaller batch sizes and a shorter training schedule", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "score": 1.0, + "content": "compared to the final models. The Overall score is obtained by dividing each benchmark score by its", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "score": 1.0, + "content": "state-of-the-art (SotA) performance from Table 1 and averaging the results. More details and results", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 683, + 270, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 270, + 695 + ], + "score": 1.0, + "content": "are given in Appendix B.3 and Table 10.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 640, + 506, + 695 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 502, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "Importance of the training data mixture. As shown in row (i), getting the right training data plays", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 709, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 104, + 709, + 505, + 724 + ], + "score": 1.0, + "content": "a crucial role. In fact, removing the interleaved image-text dataset M3W leads to a decrease of more", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 127, + 85 + ], + "score": 1.0, + "content": "than", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 127, + 73, + 147, + 83 + ], + "score": 0.87, + "content": "1 7 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 147, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "in performance while removing the conventional paired image-text pairs also decreases", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 174, + 97 + ], + "score": 1.0, + "content": "performance (by", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 175, + 83, + 198, + 95 + ], + "score": 0.87, + "content": "9 . 8 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 198, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "), demonstrating the need for different types of datasets. Moreover, removing", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 108 + ], + "score": 1.0, + "content": "our paired video-text dataset negatively affects performance on all video tasks. We ablate replacing", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 507, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 507, + 119 + ], + "score": 1.0, + "content": "our image-text pairs (ITP) by the publicly available LAION-400M dataset [96], which leads to a", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 116, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 129 + ], + "score": 1.0, + "content": "slight degradation in performance. We show in row (ii) the importance of our gradient accumulation", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 128, + 321, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 321, + 140 + ], + "score": 1.0, + "content": "strategy compared to using round-robin updates [17].", + "type": "text", + "cross_page": true + } + ], + "index": 5 + } + ], + "index": 36.5, + "bbox_fs": [ + 104, + 699, + 505, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 127, + 85 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 127, + 73, + 147, + 83 + ], + "score": 0.87, + "content": "1 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "in performance while removing the conventional paired image-text pairs also decreases", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 174, + 97 + ], + "score": 1.0, + "content": "performance (by", + "type": "text" + }, + { + "bbox": [ + 175, + 83, + 198, + 95 + ], + "score": 0.87, + "content": "9 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "), demonstrating the need for different types of datasets. Moreover, removing", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 108 + ], + "score": 1.0, + "content": "our paired video-text dataset negatively affects performance on all video tasks. We ablate replacing", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 507, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 507, + 119 + ], + "score": 1.0, + "content": "our image-text pairs (ITP) by the publicly available LAION-400M dataset [96], which leads to a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 116, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 129 + ], + "score": 1.0, + "content": "slight degradation in performance. We show in row (ii) the importance of our gradient accumulation", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 128, + 321, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 321, + 140 + ], + "score": 1.0, + "content": "strategy compared to using round-robin updates [17].", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "Visual conditioning of the frozen LM. We ablate the use of the 0-initialized tanh gating when", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 155, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 506, + 167 + ], + "score": 1.0, + "content": "merging the cross-attention output to the frozen LM output in row (iii). Without it, we see a drop of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 129, + 176 + ], + "score": 0.87, + "content": "4 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "in our overall score. Moreover, we have noticed that disabling the 0-initialized tanh gating leads", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "to training instabilities. Next, we ablate different conditioning architectures in row (iv). VANILLA", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 188, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 505, + 199 + ], + "score": 1.0, + "content": "XATTN, refers to the vanilla cross-attention from the original Transformer decoder [115]. In the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "GRAFTING approach from [68], the frozen LM is used as is with no additional layers inserted, and a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "stack of interleaved self-attention and cross-attention layers that take the frozen LM output are learnt", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 220, + 501, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 501, + 232 + ], + "score": 1.0, + "content": "from scratch. Overall, we show that our GATED XATTN-DENSE conditioning approach works best.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 236, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "Compute/Memory vs. performance trade-offs. In row (v), we ablate the frequency at which we", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "add new GATED XATTN-DENSE blocks. Although adding them at every layer is better, it significantly", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 257, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 272 + ], + "score": 1.0, + "content": "increases the number of trainable parameters and time complexity of the model. Notably, inserting", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 269, + 503, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 293, + 281 + ], + "score": 1.0, + "content": "them every fourth block accelerates training by", + "type": "text" + }, + { + "bbox": [ + 293, + 269, + 313, + 280 + ], + "score": 0.89, + "content": "6 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 269, + 480, + 281 + ], + "score": 1.0, + "content": "while only decreasing the overall score by", + "type": "text" + }, + { + "bbox": [ + 481, + 269, + 503, + 280 + ], + "score": 0.89, + "content": "1 . 9 \\%", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 281, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 291 + ], + "score": 1.0, + "content": "In light of this trade-off, we maximize the number of added layers under hardware constraints and add", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 291, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 303 + ], + "score": 1.0, + "content": "a GATED XATTN-DENSE every fourth layer for Flamingo-9B and every seventh for Flamingo-80B.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "score": 1.0, + "content": "We further compare in row (vi) the Perceiver Resampler to a MLP and a vanilla Transformer given a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 313, + 459, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 459, + 326 + ], + "score": 1.0, + "content": "parameter budget. Both underperform the Perceiver Resampler while also being slower.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 329, + 505, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 328, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 506, + 342 + ], + "score": 1.0, + "content": "Vision encoder. In row (vii), we compare our NFNet-F6 vision encoder pretrained with contrastive", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "score": 1.0, + "content": "learning (details in Appendix B.1.3) to the publicly available CLIP ViT-L/14 [85] model trained at", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 255, + 363 + ], + "score": 1.0, + "content": "224 resolution. Our NFNet-F6 has a", + "type": "text" + }, + { + "bbox": [ + 255, + 351, + 286, + 362 + ], + "score": 0.9, + "content": "+ 5 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 352, + 446, + 363 + ], + "score": 1.0, + "content": "advantage over the CLIP ViT-L/14 and", + "type": "text" + }, + { + "bbox": [ + 447, + 351, + 477, + 362 + ], + "score": 0.9, + "content": "+ 8 . 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 352, + 506, + 363 + ], + "score": 1.0, + "content": "over a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 361, + 488, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 488, + 375 + ], + "score": 1.0, + "content": "smaller NFNet-F0 encoder, which highlights the importance of using a strong vision backbone.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 378, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 392 + ], + "score": 1.0, + "content": "Freezing LM components prevents catastrophic forgetting. We verify the importance of freezing", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "the LM layers at training in row (viii). If trained from scratch, we observe a large performance de-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 400, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 144, + 412 + ], + "score": 1.0, + "content": "crease of", + "type": "text" + }, + { + "bbox": [ + 144, + 400, + 178, + 411 + ], + "score": 0.9, + "content": "- 1 2 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 400, + 506, + 412 + ], + "score": 1.0, + "content": ". Interestingly, fine-tuning our pretrained LM also leads to a drop in performance of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 136, + 422 + ], + "score": 0.9, + "content": "- 8 . 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 410, + 506, + 424 + ], + "score": 1.0, + "content": ". This indicates an instance of “catastrophic forgetting” [71], in which the model progressively", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "score": 1.0, + "content": "forgets its pretraining while training on a new objective. In our setting, freezing the language model", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 432, + 470, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 470, + 445 + ], + "score": 1.0, + "content": "is a better alternative to training with the pre-training dataset (MassiveText) in the mixture.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5 + }, + { + "type": "title", + "bbox": [ + 107, + 461, + 194, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 196, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 196, + 476 + ], + "score": 1.0, + "content": "4 Related work", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 487, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "Language modelling and few-shot adaptation. Language modelling has recently made substantial", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 497, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 507, + 511 + ], + "score": 1.0, + "content": "progress following the introduction of Transformers [115]. The paradigm of first pretraining on a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 508, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 506, + 521 + ], + "score": 1.0, + "content": "vast amount of data followed by an adaptation on a downstream task has become standard [11, 23,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "32, 44, 52, 75, 87, 108]. In this work, we build on the 70B Chinchilla language model [42] as the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "base LM for Flamingo. Numerous works have explored techniques to adapt language models to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 540, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 555 + ], + "score": 1.0, + "content": "novel tasks using a few examples. These include adding small adapter modules [43], fine-tuning", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "score": 1.0, + "content": "a small part of the LM [141], showing in-context examples in the prompt [11], or optimizing the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 104, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "prompt [56, 60] through gradient descent. In this paper, we take inspiration from the in-context [11]", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "few-shot learning technique instead of more involved few-shot learning approaches based on metric", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 585, + 380, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 380, + 597 + ], + "score": 1.0, + "content": "learning [24, 103, 112, 117] or meta-learning [6, 7, 27, 31, 91, 155].", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 601, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 615 + ], + "score": 1.0, + "content": "When language meets vision. These LM breakthroughs have been influential for vision-language", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "modelling. In particular, BERT [23] inspired a large body of vision-language work [16, 28, 29,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "38, 59, 61, 66, 101, 106, 107, 109, 118, 121, 142, 143, 151]. We differ from these approaches as", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 505, + 646 + ], + "score": 1.0, + "content": "Flamingo models do not require fine-tuning on new tasks. Another family of vision-language models", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "is based on contrastive learning [2, 5, 49, 50, 57, 74, 82, 85, 138, 140, 146]. Flamingo differs from", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 657, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 506, + 668 + ], + "score": 1.0, + "content": "contrastive models as it can generate text, although we build and rely upon them for our vision encoder.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 506, + 679 + ], + "score": 1.0, + "content": "Similar to our work are VLMs able to generate text in an autoregressive manner [19, 25, 45, 67, 116].", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 676, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 691 + ], + "score": 1.0, + "content": "Concurrent works [17, 58, 119, 124, 154] also propose to formulate numerous vision tasks as text", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 689, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 700 + ], + "score": 1.0, + "content": "generation problems. Building on top of powerful pretrained language models has been explored", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "in several recent works. One recent line of work [26, 68, 78, 114, 136, 144] proposes to freeze the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "pretrained LM weights to prevent catastrophic forgetting [71]. We follow this idea by freezing the", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 48 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 138 + ], + "lines": [], + "index": 2.5, + "bbox_fs": [ + 105, + 72, + 507, + 140 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "Visual conditioning of the frozen LM. We ablate the use of the 0-initialized tanh gating when", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 155, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 506, + 167 + ], + "score": 1.0, + "content": "merging the cross-attention output to the frozen LM output in row (iii). Without it, we see a drop of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 129, + 176 + ], + "score": 0.87, + "content": "4 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "in our overall score. Moreover, we have noticed that disabling the 0-initialized tanh gating leads", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "to training instabilities. Next, we ablate different conditioning architectures in row (iv). VANILLA", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 188, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 505, + 199 + ], + "score": 1.0, + "content": "XATTN, refers to the vanilla cross-attention from the original Transformer decoder [115]. In the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "GRAFTING approach from [68], the frozen LM is used as is with no additional layers inserted, and a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "stack of interleaved self-attention and cross-attention layers that take the frozen LM output are learnt", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 220, + 501, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 501, + 232 + ], + "score": 1.0, + "content": "from scratch. Overall, we show that our GATED XATTN-DENSE conditioning approach works best.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 144, + 506, + 232 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 236, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "Compute/Memory vs. performance trade-offs. In row (v), we ablate the frequency at which we", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "add new GATED XATTN-DENSE blocks. Although adding them at every layer is better, it significantly", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 257, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 272 + ], + "score": 1.0, + "content": "increases the number of trainable parameters and time complexity of the model. Notably, inserting", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 269, + 503, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 293, + 281 + ], + "score": 1.0, + "content": "them every fourth block accelerates training by", + "type": "text" + }, + { + "bbox": [ + 293, + 269, + 313, + 280 + ], + "score": 0.89, + "content": "6 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 269, + 480, + 281 + ], + "score": 1.0, + "content": "while only decreasing the overall score by", + "type": "text" + }, + { + "bbox": [ + 481, + 269, + 503, + 280 + ], + "score": 0.89, + "content": "1 . 9 \\%", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 281, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 291 + ], + "score": 1.0, + "content": "In light of this trade-off, we maximize the number of added layers under hardware constraints and add", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 291, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 303 + ], + "score": 1.0, + "content": "a GATED XATTN-DENSE every fourth layer for Flamingo-9B and every seventh for Flamingo-80B.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "score": 1.0, + "content": "We further compare in row (vi) the Perceiver Resampler to a MLP and a vanilla Transformer given a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 313, + 459, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 459, + 326 + ], + "score": 1.0, + "content": "parameter budget. Both underperform the Perceiver Resampler while also being slower.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 236, + 506, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 329, + 505, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 328, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 506, + 342 + ], + "score": 1.0, + "content": "Vision encoder. In row (vii), we compare our NFNet-F6 vision encoder pretrained with contrastive", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "score": 1.0, + "content": "learning (details in Appendix B.1.3) to the publicly available CLIP ViT-L/14 [85] model trained at", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 255, + 363 + ], + "score": 1.0, + "content": "224 resolution. Our NFNet-F6 has a", + "type": "text" + }, + { + "bbox": [ + 255, + 351, + 286, + 362 + ], + "score": 0.9, + "content": "+ 5 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 352, + 446, + 363 + ], + "score": 1.0, + "content": "advantage over the CLIP ViT-L/14 and", + "type": "text" + }, + { + "bbox": [ + 447, + 351, + 477, + 362 + ], + "score": 0.9, + "content": "+ 8 . 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 352, + 506, + 363 + ], + "score": 1.0, + "content": "over a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 361, + 488, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 488, + 375 + ], + "score": 1.0, + "content": "smaller NFNet-F0 encoder, which highlights the importance of using a strong vision backbone.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 328, + 506, + 375 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 378, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 392 + ], + "score": 1.0, + "content": "Freezing LM components prevents catastrophic forgetting. We verify the importance of freezing", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "the LM layers at training in row (viii). If trained from scratch, we observe a large performance de-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 400, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 144, + 412 + ], + "score": 1.0, + "content": "crease of", + "type": "text" + }, + { + "bbox": [ + 144, + 400, + 178, + 411 + ], + "score": 0.9, + "content": "- 1 2 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 400, + 506, + 412 + ], + "score": 1.0, + "content": ". Interestingly, fine-tuning our pretrained LM also leads to a drop in performance of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 136, + 422 + ], + "score": 0.9, + "content": "- 8 . 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 410, + 506, + 424 + ], + "score": 1.0, + "content": ". This indicates an instance of “catastrophic forgetting” [71], in which the model progressively", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "score": 1.0, + "content": "forgets its pretraining while training on a new objective. In our setting, freezing the language model", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 432, + 470, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 470, + 445 + ], + "score": 1.0, + "content": "is a better alternative to training with the pre-training dataset (MassiveText) in the mixture.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 376, + 506, + 445 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 461, + 194, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 196, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 196, + 476 + ], + "score": 1.0, + "content": "4 Related work", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 487, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "Language modelling and few-shot adaptation. Language modelling has recently made substantial", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 497, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 507, + 511 + ], + "score": 1.0, + "content": "progress following the introduction of Transformers [115]. The paradigm of first pretraining on a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 508, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 506, + 521 + ], + "score": 1.0, + "content": "vast amount of data followed by an adaptation on a downstream task has become standard [11, 23,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "32, 44, 52, 75, 87, 108]. In this work, we build on the 70B Chinchilla language model [42] as the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "base LM for Flamingo. Numerous works have explored techniques to adapt language models to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 540, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 555 + ], + "score": 1.0, + "content": "novel tasks using a few examples. These include adding small adapter modules [43], fine-tuning", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "score": 1.0, + "content": "a small part of the LM [141], showing in-context examples in the prompt [11], or optimizing the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 104, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "prompt [56, 60] through gradient descent. In this paper, we take inspiration from the in-context [11]", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "few-shot learning technique instead of more involved few-shot learning approaches based on metric", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 585, + 380, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 380, + 597 + ], + "score": 1.0, + "content": "learning [24, 103, 112, 117] or meta-learning [6, 7, 27, 31, 91, 155].", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5, + "bbox_fs": [ + 104, + 487, + 507, + 597 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 601, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 615 + ], + "score": 1.0, + "content": "When language meets vision. These LM breakthroughs have been influential for vision-language", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "modelling. In particular, BERT [23] inspired a large body of vision-language work [16, 28, 29,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "38, 59, 61, 66, 101, 106, 107, 109, 118, 121, 142, 143, 151]. We differ from these approaches as", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 505, + 646 + ], + "score": 1.0, + "content": "Flamingo models do not require fine-tuning on new tasks. Another family of vision-language models", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "is based on contrastive learning [2, 5, 49, 50, 57, 74, 82, 85, 138, 140, 146]. Flamingo differs from", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 657, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 506, + 668 + ], + "score": 1.0, + "content": "contrastive models as it can generate text, although we build and rely upon them for our vision encoder.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 506, + 679 + ], + "score": 1.0, + "content": "Similar to our work are VLMs able to generate text in an autoregressive manner [19, 25, 45, 67, 116].", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 676, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 691 + ], + "score": 1.0, + "content": "Concurrent works [17, 58, 119, 124, 154] also propose to formulate numerous vision tasks as text", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 689, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 700 + ], + "score": 1.0, + "content": "generation problems. Building on top of powerful pretrained language models has been explored", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "in several recent works. One recent line of work [26, 68, 78, 114, 136, 144] proposes to freeze the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "pretrained LM weights to prevent catastrophic forgetting [71]. We follow this idea by freezing the", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 48, + "bbox_fs": [ + 105, + 600, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 71, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 71, + 506, + 86 + ], + "score": 1.0, + "content": "Chinchilla LM layers [42] and adding learnable layers within the frozen LM. We differ from prior", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 489, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 489, + 96 + ], + "score": 1.0, + "content": "work by introducing the first LM that can ingest arbitrarily interleaved images, videos, and text.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 504, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 504, + 112 + ], + "score": 1.0, + "content": "Web-scale vision and language training datasets. Manually annotated vision and language datasets", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "are costly to obtain and thus relatively small (10k-100k) in scale [3, 15, 69, 122, 129, 139]. To", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 122, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 133 + ], + "score": 1.0, + "content": "alleviate this lack of data, numerous works [14, 50, 98, 110] automatically scrape readily available", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 145 + ], + "score": 1.0, + "content": "paired vision-text data. In addition to such paired data, we show the importance of also training on", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "entire multimodal webpages containing interleaved images and text as a single sequence. Concurrent", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "work CM3 [1] proposes to generate HTML markup from pages, while we simplify the text prediction", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "task by only generating plain text. We emphasize few-shot learning and vision tasks while CM3 [1]", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 176, + 445, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 445, + 190 + ], + "score": 1.0, + "content": "primarily evaluates on language-only benchmarks in a zero-shot or fine-tuned setup.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 107, + 203, + 179, + 217 + ], + "lines": [ + { + "bbox": [ + 104, + 202, + 181, + 219 + ], + "spans": [ + { + "bbox": [ + 104, + 202, + 181, + 219 + ], + "score": 1.0, + "content": "5 Discussion", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 505, + 283 + ], + "lines": [ + { + "bbox": [ + 106, + 228, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 505, + 240 + ], + "score": 1.0, + "content": "Limitations. First, our models build on pretrained LMs, and as a side effect, directly inherit their", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "weaknesses. For example, LM priors are generally helpful, but may play a role in occasional", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "hallucinations and ungrounded guesses. Furthermore, LMs generalise poorly to sequences longer", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 276 + ], + "score": 1.0, + "content": "than the training ones. They also suffer from poor sample efficiency during training. Addressing", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 496, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 496, + 286 + ], + "score": 1.0, + "content": "these issues can accelerate progress in the field and enhance the abilities of VLMs like Flamingo.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 108, + 289, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 107, + 289, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 107, + 289, + 505, + 300 + ], + "score": 1.0, + "content": "Second, the classification performance of Flamingo lags behind that of state-of-the-art contrastive", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 300, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 505, + 311 + ], + "score": 1.0, + "content": "models [82, 85]. These models directly optimize for text-image retrieval, of which classification is", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 312, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 322 + ], + "score": 1.0, + "content": "a special case. In contrast, our models handle a wider range of tasks, such as open-ended ones. A", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 322, + 450, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 450, + 334 + ], + "score": 1.0, + "content": "unified approach to achieve the best of both worlds is an important research direction.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 337, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "score": 1.0, + "content": "Third, in-context learning has significant advantages over gradient-based few-shot learning methods,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 362 + ], + "score": 1.0, + "content": "but also suffers from drawbacks depending on the characteristics of the application at hand. We", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 359, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 373 + ], + "score": 1.0, + "content": "demonstrate the effectiveness of in-context learning when access is limited to only a few dozen", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "examples. In-context learning also enables simple deployment, requiring only inference, generally", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 382, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 506, + 395 + ], + "score": 1.0, + "content": "with no hyperparameter tuning needed. However, in-context learning is known to be highly sensitive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 393, + 504, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 504, + 405 + ], + "score": 1.0, + "content": "to various aspects of the demonstrations [80, 148], and its inference compute cost and absolute", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "performance scale poorly with the number of shots beyond this low-data regime. There may be", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "score": 1.0, + "content": "opportunities to combine few-shot learning methods to leverage their complementary benefits. We", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 425, + 377, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 377, + 438 + ], + "score": 1.0, + "content": "discuss the limitations of our work in more depth in Appendix D.1.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 455 + ], + "score": 1.0, + "content": "Societal impacts. In terms of societal impacts, Flamingo offers a number of benefits while carrying", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 453, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 464 + ], + "score": 1.0, + "content": "some risks. Its ability to rapidly adapt to a broad range of tasks have the potential to enable non-expert", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 464, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 475 + ], + "score": 1.0, + "content": "users to obtain good performance in data-starved regimes, lowering the barriers to both beneficial", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "and malicious applications. Flamingo is exposed to the same risks as large language models, such", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "as outputting offensive language, propagating social biases and stereotypes, as well as leaking", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "private information [42, 126]. Its ability to additionally handle visual inputs poses specific risks", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "such as gender and racial biases relating to the contents of the input images, similar to a number", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "score": 1.0, + "content": "of visual recognition systems [12, 21, 37, 97, 147]. We refer the reader to Appendix D.2 for a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "more extensive discussion of the societal impacts of our work, both positive and negative; as well", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "as mitigation strategies and early investigations of risks relating to racial or gender bias and toxic", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "outputs. Finally we note that, following prior work focusing on language models [72, 81, 111], the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 562, + 408, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 408, + 574 + ], + "score": 1.0, + "content": "few-shot capabilities of Flamingo could be useful for mitigating such risks.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 504, + 633 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "Conclusion. We proposed Flamingo, a general-purpose family of models that can be applied to image", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "and video tasks with minimal task-specific training data. We also qualitatively explored interactive", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "abilities of Flamingo such as “chatting” with the model, demonstrating flexibility beyond traditional", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "score": 1.0, + "content": "vision benchmarks. Our results suggest that connecting pre-trained large language models with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 621, + 471, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 471, + 635 + ], + "score": 1.0, + "content": "powerful visual models is an important step towards general-purpose visual understanding.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding. This research was funded by DeepMind. We would", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "like to thank many colleagues for useful discussions, suggestions, feedback, and advice, including:", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 667, + 507, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 507, + 680 + ], + "score": 1.0, + "content": "Samuel Albanie, Relja Arandjelovic, Kareem Ayoub, Lorrayne Bennett, Adria Recasens Continente, ´", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "Tom Eccles, Nando de Freitas, Sander Dieleman, Conor Durkan, Aleksa Gordic, Raia Hadsell, ´", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 688, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 507, + 702 + ], + "score": 1.0, + "content": "Will Hawkins, Lisa Anne Hendricks, Felix Hill, Jordan Hoffmann, Geoffrey Irving, Drew Jaegle,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 700, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 507, + 713 + ], + "score": 1.0, + "content": "Koray Kavukcuoglu, Agustin Dal Lago, Mateusz Malinowski, Sona Mokrá, Gaby Pearl, Toby Pohlen, ˇ", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 711, + 472, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 472, + 723 + ], + "score": 1.0, + "content": "Jack Rae, Laurent Sifre, Francis Song, Maria Tsimpoukelli, Gregory Wayne, and Boxi Wu.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 755 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 755 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 71, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 71, + 506, + 86 + ], + "score": 1.0, + "content": "Chinchilla LM layers [42] and adding learnable layers within the frozen LM. We differ from prior", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 489, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 489, + 96 + ], + "score": 1.0, + "content": "work by introducing the first LM that can ingest arbitrarily interleaved images, videos, and text.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 71, + 506, + 96 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 504, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 504, + 112 + ], + "score": 1.0, + "content": "Web-scale vision and language training datasets. Manually annotated vision and language datasets", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "are costly to obtain and thus relatively small (10k-100k) in scale [3, 15, 69, 122, 129, 139]. To", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 122, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 133 + ], + "score": 1.0, + "content": "alleviate this lack of data, numerous works [14, 50, 98, 110] automatically scrape readily available", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 145 + ], + "score": 1.0, + "content": "paired vision-text data. In addition to such paired data, we show the importance of also training on", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "entire multimodal webpages containing interleaved images and text as a single sequence. Concurrent", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "work CM3 [1] proposes to generate HTML markup from pages, while we simplify the text prediction", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "task by only generating plain text. We emphasize few-shot learning and vision tasks while CM3 [1]", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 176, + 445, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 445, + 190 + ], + "score": 1.0, + "content": "primarily evaluates on language-only benchmarks in a zero-shot or fine-tuned setup.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 100, + 505, + 190 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 203, + 179, + 217 + ], + "lines": [ + { + "bbox": [ + 104, + 202, + 181, + 219 + ], + "spans": [ + { + "bbox": [ + 104, + 202, + 181, + 219 + ], + "score": 1.0, + "content": "5 Discussion", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 505, + 283 + ], + "lines": [ + { + "bbox": [ + 106, + 228, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 505, + 240 + ], + "score": 1.0, + "content": "Limitations. First, our models build on pretrained LMs, and as a side effect, directly inherit their", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "weaknesses. For example, LM priors are generally helpful, but may play a role in occasional", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "hallucinations and ungrounded guesses. Furthermore, LMs generalise poorly to sequences longer", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 276 + ], + "score": 1.0, + "content": "than the training ones. They also suffer from poor sample efficiency during training. Addressing", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 496, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 496, + 286 + ], + "score": 1.0, + "content": "these issues can accelerate progress in the field and enhance the abilities of VLMs like Flamingo.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 228, + 506, + 286 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 289, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 107, + 289, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 107, + 289, + 505, + 300 + ], + "score": 1.0, + "content": "Second, the classification performance of Flamingo lags behind that of state-of-the-art contrastive", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 300, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 505, + 311 + ], + "score": 1.0, + "content": "models [82, 85]. These models directly optimize for text-image retrieval, of which classification is", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 312, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 322 + ], + "score": 1.0, + "content": "a special case. In contrast, our models handle a wider range of tasks, such as open-ended ones. A", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 322, + 450, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 450, + 334 + ], + "score": 1.0, + "content": "unified approach to achieve the best of both worlds is an important research direction.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 289, + 505, + 334 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 337, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "score": 1.0, + "content": "Third, in-context learning has significant advantages over gradient-based few-shot learning methods,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 362 + ], + "score": 1.0, + "content": "but also suffers from drawbacks depending on the characteristics of the application at hand. We", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 359, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 373 + ], + "score": 1.0, + "content": "demonstrate the effectiveness of in-context learning when access is limited to only a few dozen", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "examples. In-context learning also enables simple deployment, requiring only inference, generally", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 382, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 506, + 395 + ], + "score": 1.0, + "content": "with no hyperparameter tuning needed. However, in-context learning is known to be highly sensitive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 393, + 504, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 504, + 405 + ], + "score": 1.0, + "content": "to various aspects of the demonstrations [80, 148], and its inference compute cost and absolute", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "performance scale poorly with the number of shots beyond this low-data regime. There may be", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "score": 1.0, + "content": "opportunities to combine few-shot learning methods to leverage their complementary benefits. We", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 425, + 377, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 377, + 438 + ], + "score": 1.0, + "content": "discuss the limitations of our work in more depth in Appendix D.1.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 336, + 506, + 438 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 455 + ], + "score": 1.0, + "content": "Societal impacts. In terms of societal impacts, Flamingo offers a number of benefits while carrying", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 453, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 464 + ], + "score": 1.0, + "content": "some risks. Its ability to rapidly adapt to a broad range of tasks have the potential to enable non-expert", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 464, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 475 + ], + "score": 1.0, + "content": "users to obtain good performance in data-starved regimes, lowering the barriers to both beneficial", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "and malicious applications. Flamingo is exposed to the same risks as large language models, such", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "as outputting offensive language, propagating social biases and stereotypes, as well as leaking", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "private information [42, 126]. Its ability to additionally handle visual inputs poses specific risks", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "such as gender and racial biases relating to the contents of the input images, similar to a number", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "score": 1.0, + "content": "of visual recognition systems [12, 21, 37, 97, 147]. We refer the reader to Appendix D.2 for a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "more extensive discussion of the societal impacts of our work, both positive and negative; as well", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "as mitigation strategies and early investigations of risks relating to racial or gender bias and toxic", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "outputs. Finally we note that, following prior work focusing on language models [72, 81, 111], the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 562, + 408, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 408, + 574 + ], + "score": 1.0, + "content": "few-shot capabilities of Flamingo could be useful for mitigating such risks.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 440, + 506, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 504, + 633 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "Conclusion. We proposed Flamingo, a general-purpose family of models that can be applied to image", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "and video tasks with minimal task-specific training data. We also qualitatively explored interactive", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "abilities of Flamingo such as “chatting” with the model, demonstrating flexibility beyond traditional", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "score": 1.0, + "content": "vision benchmarks. Our results suggest that connecting pre-trained large language models with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 621, + 471, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 471, + 635 + ], + "score": 1.0, + "content": "powerful visual models is an important step towards general-purpose visual understanding.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 577, + 506, + 635 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding. This research was funded by DeepMind. We would", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "like to thank many colleagues for useful discussions, suggestions, feedback, and advice, including:", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 667, + 507, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 507, + 680 + ], + "score": 1.0, + "content": "Samuel Albanie, Relja Arandjelovic, Kareem Ayoub, Lorrayne Bennett, Adria Recasens Continente, ´", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "Tom Eccles, Nando de Freitas, Sander Dieleman, Conor Durkan, Aleksa Gordic, Raia Hadsell, ´", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 688, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 507, + 702 + ], + "score": 1.0, + "content": "Will Hawkins, Lisa Anne Hendricks, Felix Hill, Jordan Hoffmann, Geoffrey Irving, Drew Jaegle,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 700, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 507, + 713 + ], + "score": 1.0, + "content": "Koray Kavukcuoglu, Agustin Dal Lago, Mateusz Malinowski, Sona Mokrá, Gaby Pearl, Toby Pohlen, ˇ", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 711, + 472, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 472, + 723 + ], + "score": 1.0, + "content": "Jack Rae, Laurent Sifre, Francis Song, Maria Tsimpoukelli, Gregory Wayne, and Boxi Wu.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49, + "bbox_fs": [ + 105, + 644, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 57, + 507, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 114, + 89, + 507, + 105 + ], + "spans": [ + { + "bbox": [ + 114, + 89, + 507, + 105 + ], + "score": 1.0, + "content": "[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 100, + 506, + 114 + ], + "spans": [ + { + "bbox": [ + 132, + 100, + 506, + 114 + ], + "score": 1.0, + "content": "Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, and Luke Zettlemoyer. CM3: A", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 113, + 432, + 125 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 432, + 125 + ], + "score": 1.0, + "content": "causal masked multimodal model of the internet. arXiv:2201.07520, 2022.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 131, + 507, + 145 + ], + "spans": [ + { + "bbox": [ + 115, + 131, + 507, + 145 + ], + "score": 1.0, + "content": "[2] Jean-Baptiste Alayrac, Adria Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Rama- ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 142, + 507, + 155 + ], + "spans": [ + { + "bbox": [ + 131, + 142, + 507, + 155 + ], + "score": 1.0, + "content": "puram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 152, + 507, + 169 + ], + "spans": [ + { + "bbox": [ + 129, + 152, + 507, + 169 + ], + "score": 1.0, + "content": "supervised multimodal versatile networks. Conference on Neural Information Processing", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 165, + 194, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 165, + 194, + 177 + ], + "score": 1.0, + "content": "Systems, 2020.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 182, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 182, + 506, + 199 + ], + "score": 1.0, + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 194, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 506, + 209 + ], + "score": 1.0, + "content": "Zitnick, and Devi Parikh. VQA: Visual question answering. In International Conference on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 207, + 230, + 218 + ], + "spans": [ + { + "bbox": [ + 133, + 207, + 230, + 218 + ], + "score": 1.0, + "content": "Computer Vision, 2015.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "[4] Thomas Bachlechner, Bodhisattwa Prasad Majumder, Henry Mao, Gary Cottrell, and Julian", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "McAuley. ReZero is all you need: Fast convergence at large depth. In Uncertainty in Artificial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 248, + 210, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 210, + 259 + ], + "score": 1.0, + "content": "Intelligence, 2021.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 266, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 506, + 280 + ], + "score": 1.0, + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 276, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 506, + 290 + ], + "score": 1.0, + "content": "and image encoder for end-to-end retrieval. In International Conference on Computer Vision,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 287, + 160, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 287, + 160, + 302 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "[6] Luca Bertinetto, João F. Henriques, Jack Valmadre, Philip Torr, and Andrea Vedaldi. Learning", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 318, + 507, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 507, + 332 + ], + "score": 1.0, + "content": "feed-forward one-shot learners. Conference on Neural Information Processing Systems, 2016.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 113, + 336, + 507, + 352 + ], + "spans": [ + { + "bbox": [ + 113, + 336, + 507, + 352 + ], + "score": 1.0, + "content": "[7] Luca Bertinetto, Joao F. Henriques, Philip H. S. Torr, and Andrea Vedaldi. Meta-learning with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 350, + 376, + 361 + ], + "spans": [ + { + "bbox": [ + 133, + 350, + 376, + 361 + ], + "score": 1.0, + "content": "differentiable closed-form solvers. arXiv:1805.08136, 2018.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 368, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 505, + 382 + ], + "score": 1.0, + "content": "[8] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 380, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 505, + 391 + ], + "score": 1.0, + "content": "Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 133, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. URL", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 129, + 399, + 285, + 416 + ], + "spans": [ + { + "bbox": [ + 129, + 399, + 285, + 416 + ], + "score": 1.0, + "content": "http://github.com/google/jax.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 115, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "[9] John S. Bridle. Probabilistic interpretation of feedforward classification network outputs, with", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 432, + 426, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 426, + 444 + ], + "score": 1.0, + "content": "relationships to statistical pattern recognition. In Neurocomputing, 1990.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 111, + 450, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 111, + 450, + 505, + 464 + ], + "score": 1.0, + "content": "[10] Andrew Brock, Soham De, Samuel L. Smith, and Karen Simonyan. High-performance", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 132, + 461, + 451, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 451, + 475 + ], + "score": 1.0, + "content": "large-scale image recognition without normalization. arXiv:2102.06171, 2021.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 110, + 480, + 506, + 494 + ], + "spans": [ + { + "bbox": [ + 110, + 480, + 506, + 494 + ], + "score": 1.0, + "content": "[11] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 132, + 491, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 505, + 504 + ], + "score": 1.0, + "content": "Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agar-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 131, + 501, + 507, + 516 + ], + "spans": [ + { + "bbox": [ + 131, + 501, + 507, + 516 + ], + "score": 1.0, + "content": "wal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 513, + 507, + 527 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 507, + 527 + ], + "score": 1.0, + "content": "Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 132, + 524, + 507, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 507, + 538 + ], + "score": 1.0, + "content": "Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCan-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 535, + 507, + 549 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 507, + 549 + ], + "score": 1.0, + "content": "dlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 131, + 546, + 393, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 393, + 560 + ], + "score": 1.0, + "content": "In Conference on Neural Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 111, + 565, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 111, + 565, + 506, + 579 + ], + "score": 1.0, + "content": "[12] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 576, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 506, + 590 + ], + "score": 1.0, + "content": "commercial gender classification. In ACM Conference on Fairness, Accountability, and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 131, + 588, + 217, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 588, + 217, + 600 + ], + "score": 1.0, + "content": "Transparency, 2018.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 110, + 606, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 110, + 606, + 506, + 621 + ], + "score": 1.0, + "content": "[13] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 133, + 629, + 243, + 641 + ], + "spans": [ + { + "bbox": [ + 133, + 629, + 243, + 641 + ], + "score": 1.0, + "content": "on Computer Vision, 2020.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 109, + 646, + 507, + 663 + ], + "spans": [ + { + "bbox": [ + 109, + 646, + 507, + 663 + ], + "score": 1.0, + "content": "[14] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 132, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 132, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "web-scale image-text pre-training to recognize long-tail visual concepts. In IEEE Computer", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 133, + 671, + 286, + 682 + ], + "spans": [ + { + "bbox": [ + 133, + 671, + 286, + 682 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition, 2021.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 111, + 688, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 111, + 688, + 507, + 702 + ], + "score": 1.0, + "content": "[15] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 131, + 700, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 131, + 700, + 507, + 713 + ], + "score": 1.0, + "content": "and C Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 132, + 710, + 236, + 724 + ], + "spans": [ + { + "bbox": [ + 132, + 710, + 236, + 724 + ], + "score": 1.0, + "content": "arXiv:1504.00325, 2015.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 23.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 310, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 755 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 755 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 107, + 57, + 507, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 114, + 89, + 507, + 105 + ], + "spans": [ + { + "bbox": [ + 114, + 89, + 507, + 105 + ], + "score": 1.0, + "content": "[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 100, + 506, + 114 + ], + "spans": [ + { + "bbox": [ + 132, + 100, + 506, + 114 + ], + "score": 1.0, + "content": "Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, and Luke Zettlemoyer. CM3: A", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 113, + 432, + 125 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 432, + 125 + ], + "score": 1.0, + "content": "causal masked multimodal model of the internet. arXiv:2201.07520, 2022.", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 115, + 131, + 507, + 145 + ], + "spans": [ + { + "bbox": [ + 115, + 131, + 507, + 145 + ], + "score": 1.0, + "content": "[2] Jean-Baptiste Alayrac, Adria Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Rama- ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 142, + 507, + 155 + ], + "spans": [ + { + "bbox": [ + 131, + 142, + 507, + 155 + ], + "score": 1.0, + "content": "puram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 152, + 507, + 169 + ], + "spans": [ + { + "bbox": [ + 129, + 152, + 507, + 169 + ], + "score": 1.0, + "content": "supervised multimodal versatile networks. Conference on Neural Information Processing", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 165, + 194, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 165, + 194, + 177 + ], + "score": 1.0, + "content": "Systems, 2020.", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 114, + 182, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 182, + 506, + 199 + ], + "score": 1.0, + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 194, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 506, + 209 + ], + "score": 1.0, + "content": "Zitnick, and Devi Parikh. VQA: Visual question answering. In International Conference on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 207, + 230, + 218 + ], + "spans": [ + { + "bbox": [ + 133, + 207, + 230, + 218 + ], + "score": 1.0, + "content": "Computer Vision, 2015.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 115, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "[4] Thomas Bachlechner, Bodhisattwa Prasad Majumder, Henry Mao, Gary Cottrell, and Julian", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "McAuley. ReZero is all you need: Fast convergence at large depth. In Uncertainty in Artificial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 248, + 210, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 210, + 259 + ], + "score": 1.0, + "content": "Intelligence, 2021.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 115, + 266, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 506, + 280 + ], + "score": 1.0, + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 276, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 506, + 290 + ], + "score": 1.0, + "content": "and image encoder for end-to-end retrieval. In International Conference on Computer Vision,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 287, + 160, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 287, + 160, + 302 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "[6] Luca Bertinetto, João F. Henriques, Jack Valmadre, Philip Torr, and Andrea Vedaldi. Learning", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 318, + 507, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 507, + 332 + ], + "score": 1.0, + "content": "feed-forward one-shot learners. Conference on Neural Information Processing Systems, 2016.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 113, + 336, + 507, + 352 + ], + "spans": [ + { + "bbox": [ + 113, + 336, + 507, + 352 + ], + "score": 1.0, + "content": "[7] Luca Bertinetto, Joao F. Henriques, Philip H. S. Torr, and Andrea Vedaldi. Meta-learning with", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 133, + 350, + 376, + 361 + ], + "spans": [ + { + "bbox": [ + 133, + 350, + 376, + 361 + ], + "score": 1.0, + "content": "differentiable closed-form solvers. arXiv:1805.08136, 2018.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 115, + 368, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 505, + 382 + ], + "score": 1.0, + "content": "[8] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 380, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 505, + 391 + ], + "score": 1.0, + "content": "Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 133, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. URL", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 129, + 399, + 285, + 416 + ], + "spans": [ + { + "bbox": [ + 129, + 399, + 285, + 416 + ], + "score": 1.0, + "content": "http://github.com/google/jax.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 115, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 115, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "[9] John S. Bridle. Probabilistic interpretation of feedforward classification network outputs, with", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 432, + 426, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 426, + 444 + ], + "score": 1.0, + "content": "relationships to statistical pattern recognition. In Neurocomputing, 1990.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 450, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 111, + 450, + 505, + 464 + ], + "score": 1.0, + "content": "[10] Andrew Brock, Soham De, Samuel L. Smith, and Karen Simonyan. High-performance", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 461, + 451, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 451, + 475 + ], + "score": 1.0, + "content": "large-scale image recognition without normalization. arXiv:2102.06171, 2021.", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 480, + 506, + 494 + ], + "spans": [ + { + "bbox": [ + 110, + 480, + 506, + 494 + ], + "score": 1.0, + "content": "[11] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 491, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 505, + 504 + ], + "score": 1.0, + "content": "Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agar-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 131, + 501, + 507, + 516 + ], + "spans": [ + { + "bbox": [ + 131, + 501, + 507, + 516 + ], + "score": 1.0, + "content": "wal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 513, + 507, + 527 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 507, + 527 + ], + "score": 1.0, + "content": "Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 132, + 524, + 507, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 507, + 538 + ], + "score": 1.0, + "content": "Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCan-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 535, + 507, + 549 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 507, + 549 + ], + "score": 1.0, + "content": "dlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 131, + 546, + 393, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 393, + 560 + ], + "score": 1.0, + "content": "In Conference on Neural Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 565, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 111, + 565, + 506, + 579 + ], + "score": 1.0, + "content": "[12] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 576, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 506, + 590 + ], + "score": 1.0, + "content": "commercial gender classification. In ACM Conference on Fairness, Accountability, and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 131, + 588, + 217, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 588, + 217, + 600 + ], + "score": 1.0, + "content": "Transparency, 2018.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 606, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 110, + 606, + 506, + 621 + ], + "score": 1.0, + "content": "[13] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 133, + 629, + 243, + 641 + ], + "spans": [ + { + "bbox": [ + 133, + 629, + 243, + 641 + ], + "score": 1.0, + "content": "on Computer Vision, 2020.", + "type": "text" + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 646, + 507, + 663 + ], + "spans": [ + { + "bbox": [ + 109, + 646, + 507, + 663 + ], + "score": 1.0, + "content": "[14] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 132, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "web-scale image-text pre-training to recognize long-tail visual concepts. In IEEE Computer", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 133, + 671, + 286, + 682 + ], + "spans": [ + { + "bbox": [ + 133, + 671, + 286, + 682 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition, 2021.", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 688, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 111, + 688, + 507, + 702 + ], + "score": 1.0, + "content": "[15] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár,", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 700, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 131, + 700, + 507, + 713 + ], + "score": 1.0, + "content": "and C Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 132, + 710, + 236, + 724 + ], + "spans": [ + { + "bbox": [ + 132, + 710, + 236, + 724 + ], + "score": 1.0, + "content": "arXiv:1504.00325, 2015.", + "type": "text" + } + ], + "index": 47, + "is_list_end_line": true + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 70, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 111, + 72, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 110, + 71, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 110, + 71, + 506, + 87 + ], + "score": 1.0, + "content": "[16] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 83, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 132, + 83, + 505, + 97 + ], + "score": 1.0, + "content": "and Jingjing Liu. UNITER: Universal image-text representation learning. In European", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 94, + 290, + 107 + ], + "spans": [ + { + "bbox": [ + 132, + 94, + 290, + 107 + ], + "score": 1.0, + "content": "Conference on Computer Vision, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 110, + 112, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 110, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 110, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "[17] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 124, + 412, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 124, + 412, + 136 + ], + "score": 1.0, + "content": "generation. In International Conference on Machine Learning, 2021.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 113, + 141, + 504, + 284 + ], + "lines": [ + { + "bbox": [ + 111, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 111, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 153, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 153, + 506, + 165 + ], + "score": 1.0, + "content": "Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 163, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 506, + 176 + ], + "score": 1.0, + "content": "Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 174, + 506, + 188 + ], + "score": 1.0, + "content": "Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 184, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 131, + 184, + 506, + 199 + ], + "score": 1.0, + "content": "Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 195, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 131, + 195, + 506, + 210 + ], + "score": 1.0, + "content": "Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 207, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 506, + 220 + ], + "score": 1.0, + "content": "Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 217, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 506, + 231 + ], + "score": 1.0, + "content": "Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 230, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 505, + 241 + ], + "score": 1.0, + "content": "Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 240, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 506, + 253 + ], + "score": 1.0, + "content": "Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 262, + 507, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 507, + 275 + ], + "score": 1.0, + "content": "Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM:", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 272, + 411, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 272, + 411, + 285 + ], + "score": 1.0, + "content": "Scaling language modeling with pathways. arXiv:2204.02311, 2022.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 110, + 291, + 503, + 324 + ], + "lines": [ + { + "bbox": [ + 110, + 290, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 110, + 290, + 505, + 305 + ], + "score": 1.0, + "content": "[19] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 301, + 504, + 315 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 504, + 315 + ], + "score": 1.0, + "content": "multimodal generation on clip via vision-language knowledge distillation. In ACL Findings,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 311, + 159, + 325 + ], + "spans": [ + { + "bbox": [ + 131, + 311, + 159, + 325 + ], + "score": 1.0, + "content": "2022.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 110, + 331, + 504, + 364 + ], + "lines": [ + { + "bbox": [ + 111, + 331, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 111, + 331, + 505, + 343 + ], + "score": 1.0, + "content": "[20] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José MF Moura, Devi", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 131, + 342, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 131, + 342, + 506, + 355 + ], + "score": 1.0, + "content": "Parikh, and Dhruv Batra. Visual dialog. In IEEE Computer Vision and Pattern Recognition,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 131, + 352, + 160, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 352, + 160, + 365 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 110, + 371, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 111, + 371, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 111, + 371, + 506, + 384 + ], + "score": 1.0, + "content": "[21] Terrance De Vries, Ishan Misra, Changhan Wang, and Laurens Van der Maaten. Does object", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 382, + 494, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 494, + 395 + ], + "score": 1.0, + "content": "recognition work for everyone? In IEEE Computer Vision and Pattern Recognition, 2019.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 111, + 400, + 505, + 423 + ], + "lines": [ + { + "bbox": [ + 111, + 400, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 111, + 400, + 506, + 413 + ], + "score": 1.0, + "content": "[22] Karan Desai and Justin Johnson. VirTex: Learning visual representations from textual annota-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 411, + 389, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 389, + 424 + ], + "score": 1.0, + "content": "tions. In IEEE Computer Vision and Pattern Recognition, 2021.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 111, + 429, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 110, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 110, + 429, + 506, + 443 + ], + "score": 1.0, + "content": "[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 440, + 477, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 477, + 453 + ], + "score": 1.0, + "content": "deep bidirectional transformers for language understanding. arXiv:1810.04805, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 110, + 458, + 504, + 482 + ], + "lines": [ + { + "bbox": [ + 110, + 457, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 110, + 457, + 505, + 473 + ], + "score": 1.0, + "content": "[24] Carl Doersch, Ankush Gupta, and Andrew Zisserman. CrossTransformers: spatially-aware", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 470, + 455, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 455, + 483 + ], + "score": 1.0, + "content": "few-shot transfer. Conference on Neural Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 113, + 488, + 504, + 532 + ], + "lines": [ + { + "bbox": [ + 110, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 110, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "[25] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 132, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 510, + 507, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 507, + 523 + ], + "score": 1.0, + "content": "for visual recognition and description. In IEEE Computer Vision and Pattern Recognition,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 132, + 519, + 159, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 159, + 534 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 111, + 539, + 504, + 573 + ], + "lines": [ + { + "bbox": [ + 110, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 110, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "[26] Constantin Eichenberg, Sidney Black, Samuel Weinbach, Letitia Parcalabescu, and Anette", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Frank. MAGMA–multimodal augmentation of generative models through adapter-based", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 133, + 561, + 282, + 573 + ], + "spans": [ + { + "bbox": [ + 133, + 561, + 282, + 573 + ], + "score": 1.0, + "content": "finetuning. arXiv:2112.05253, 2021.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 109, + 579, + 504, + 602 + ], + "lines": [ + { + "bbox": [ + 111, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 111, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "[27] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 590, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 482, + 603 + ], + "score": 1.0, + "content": "adaptation of deep networks. In International Conference on Machine Learning, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 111, + 608, + 504, + 642 + ], + "lines": [ + { + "bbox": [ + 110, + 607, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 110, + 607, + 506, + 623 + ], + "score": 1.0, + "content": "[28] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 131, + 618, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 618, + 506, + 633 + ], + "score": 1.0, + "content": "Liu. VIOLET: End-to-end video-language transformers with masked visual-token modeling.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 132, + 631, + 236, + 641 + ], + "spans": [ + { + "bbox": [ + 132, + 631, + 236, + 641 + ], + "score": 1.0, + "content": "arXiv:2111.12681, 2021.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 110, + 648, + 503, + 682 + ], + "lines": [ + { + "bbox": [ + 110, + 647, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 110, + 647, + 506, + 662 + ], + "score": 1.0, + "content": "[29] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 133, + 659, + 504, + 672 + ], + "spans": [ + { + "bbox": [ + 133, + 659, + 504, + 672 + ], + "score": 1.0, + "content": "adversarial training for vision-and-language representation learning. In Conference on Neural", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 132, + 671, + 290, + 683 + ], + "spans": [ + { + "bbox": [ + 132, + 671, + 290, + 683 + ], + "score": 1.0, + "content": "Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 110, + 688, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 110, + 687, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 110, + 687, + 505, + 702 + ], + "score": 1.0, + "content": "[30] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 131, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 131, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "Wallach, Hal Daumé III, and Kate Crawford. Datasheets for datasets. Communications of the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 131, + 711, + 183, + 722 + ], + "spans": [ + { + "bbox": [ + 131, + 711, + 183, + 722 + ], + "score": 1.0, + "content": "ACM, 2021.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 111, + 72, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 110, + 71, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 110, + 71, + 506, + 87 + ], + "score": 1.0, + "content": "[16] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 83, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 132, + 83, + 505, + 97 + ], + "score": 1.0, + "content": "and Jingjing Liu. UNITER: Universal image-text representation learning. In European", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 94, + 290, + 107 + ], + "spans": [ + { + "bbox": [ + 132, + 94, + 290, + 107 + ], + "score": 1.0, + "content": "Conference on Computer Vision, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 110, + 71, + 506, + 107 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 112, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 110, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 110, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "[17] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 124, + 412, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 124, + 412, + 136 + ], + "score": 1.0, + "content": "generation. In International Conference on Machine Learning, 2021.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 110, + 111, + 505, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 113, + 141, + 504, + 284 + ], + "lines": [ + { + "bbox": [ + 111, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 111, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 153, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 153, + 506, + 165 + ], + "score": 1.0, + "content": "Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 163, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 506, + 176 + ], + "score": 1.0, + "content": "Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 174, + 506, + 188 + ], + "score": 1.0, + "content": "Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 184, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 131, + 184, + 506, + 199 + ], + "score": 1.0, + "content": "Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 195, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 131, + 195, + 506, + 210 + ], + "score": 1.0, + "content": "Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 207, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 506, + 220 + ], + "score": 1.0, + "content": "Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 217, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 506, + 231 + ], + "score": 1.0, + "content": "Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 230, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 505, + 241 + ], + "score": 1.0, + "content": "Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 240, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 506, + 253 + ], + "score": 1.0, + "content": "Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 262, + 507, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 507, + 275 + ], + "score": 1.0, + "content": "Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM:", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 272, + 411, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 272, + 411, + 285 + ], + "score": 1.0, + "content": "Scaling language modeling with pathways. arXiv:2204.02311, 2022.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11, + "bbox_fs": [ + 111, + 142, + 507, + 285 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 291, + 503, + 324 + ], + "lines": [ + { + "bbox": [ + 110, + 290, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 110, + 290, + 505, + 305 + ], + "score": 1.0, + "content": "[19] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 301, + 504, + 315 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 504, + 315 + ], + "score": 1.0, + "content": "multimodal generation on clip via vision-language knowledge distillation. In ACL Findings,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 311, + 159, + 325 + ], + "spans": [ + { + "bbox": [ + 131, + 311, + 159, + 325 + ], + "score": 1.0, + "content": "2022.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 110, + 290, + 505, + 325 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 331, + 504, + 364 + ], + "lines": [ + { + "bbox": [ + 111, + 331, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 111, + 331, + 505, + 343 + ], + "score": 1.0, + "content": "[20] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José MF Moura, Devi", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 131, + 342, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 131, + 342, + 506, + 355 + ], + "score": 1.0, + "content": "Parikh, and Dhruv Batra. Visual dialog. In IEEE Computer Vision and Pattern Recognition,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 131, + 352, + 160, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 352, + 160, + 365 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 111, + 331, + 506, + 365 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 371, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 111, + 371, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 111, + 371, + 506, + 384 + ], + "score": 1.0, + "content": "[21] Terrance De Vries, Ishan Misra, Changhan Wang, and Laurens Van der Maaten. Does object", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 382, + 494, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 494, + 395 + ], + "score": 1.0, + "content": "recognition work for everyone? In IEEE Computer Vision and Pattern Recognition, 2019.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 111, + 371, + 506, + 395 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 400, + 505, + 423 + ], + "lines": [ + { + "bbox": [ + 111, + 400, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 111, + 400, + 506, + 413 + ], + "score": 1.0, + "content": "[22] Karan Desai and Justin Johnson. VirTex: Learning visual representations from textual annota-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 411, + 389, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 389, + 424 + ], + "score": 1.0, + "content": "tions. In IEEE Computer Vision and Pattern Recognition, 2021.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 111, + 400, + 506, + 424 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 429, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 110, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 110, + 429, + 506, + 443 + ], + "score": 1.0, + "content": "[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 440, + 477, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 477, + 453 + ], + "score": 1.0, + "content": "deep bidirectional transformers for language understanding. arXiv:1810.04805, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 110, + 429, + 506, + 453 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 458, + 504, + 482 + ], + "lines": [ + { + "bbox": [ + 110, + 457, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 110, + 457, + 505, + 473 + ], + "score": 1.0, + "content": "[24] Carl Doersch, Ankush Gupta, and Andrew Zisserman. CrossTransformers: spatially-aware", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 470, + 455, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 455, + 483 + ], + "score": 1.0, + "content": "few-shot transfer. Conference on Neural Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 110, + 457, + 505, + 483 + ] + }, + { + "type": "text", + "bbox": [ + 113, + 488, + 504, + 532 + ], + "lines": [ + { + "bbox": [ + 110, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 110, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "[25] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 132, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 510, + 507, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 507, + 523 + ], + "score": 1.0, + "content": "for visual recognition and description. In IEEE Computer Vision and Pattern Recognition,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 132, + 519, + 159, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 159, + 534 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 110, + 487, + 507, + 534 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 539, + 504, + 573 + ], + "lines": [ + { + "bbox": [ + 110, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 110, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "[26] Constantin Eichenberg, Sidney Black, Samuel Weinbach, Letitia Parcalabescu, and Anette", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Frank. MAGMA–multimodal augmentation of generative models through adapter-based", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 133, + 561, + 282, + 573 + ], + "spans": [ + { + "bbox": [ + 133, + 561, + 282, + 573 + ], + "score": 1.0, + "content": "finetuning. arXiv:2112.05253, 2021.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 110, + 538, + 505, + 573 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 579, + 504, + 602 + ], + "lines": [ + { + "bbox": [ + 111, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 111, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "[27] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 590, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 482, + 603 + ], + "score": 1.0, + "content": "adaptation of deep networks. In International Conference on Machine Learning, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 111, + 579, + 506, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 608, + 504, + 642 + ], + "lines": [ + { + "bbox": [ + 110, + 607, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 110, + 607, + 506, + 623 + ], + "score": 1.0, + "content": "[28] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 131, + 618, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 618, + 506, + 633 + ], + "score": 1.0, + "content": "Liu. VIOLET: End-to-end video-language transformers with masked visual-token modeling.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 132, + 631, + 236, + 641 + ], + "spans": [ + { + "bbox": [ + 132, + 631, + 236, + 641 + ], + "score": 1.0, + "content": "arXiv:2111.12681, 2021.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 110, + 607, + 506, + 641 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 648, + 503, + 682 + ], + "lines": [ + { + "bbox": [ + 110, + 647, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 110, + 647, + 506, + 662 + ], + "score": 1.0, + "content": "[29] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 133, + 659, + 504, + 672 + ], + "spans": [ + { + "bbox": [ + 133, + 659, + 504, + 672 + ], + "score": 1.0, + "content": "adversarial training for vision-and-language representation learning. In Conference on Neural", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 132, + 671, + 290, + 683 + ], + "spans": [ + { + "bbox": [ + 132, + 671, + 290, + 683 + ], + "score": 1.0, + "content": "Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 110, + 647, + 506, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 688, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 110, + 687, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 110, + 687, + 505, + 702 + ], + "score": 1.0, + "content": "[30] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 131, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 131, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "Wallach, Hal Daumé III, and Kate Crawford. Datasheets for datasets. Communications of the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 131, + 711, + 183, + 722 + ], + "spans": [ + { + "bbox": [ + 131, + 711, + 183, + 722 + ], + "score": 1.0, + "content": "ACM, 2021.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48, + "bbox_fs": [ + 110, + 687, + 505, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 110, + 72, + 505, + 96 + ], + "lines": [ + { + "bbox": [ + 110, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 110, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[31] Jonathan Gordon, John Bronskill, Matthias Bauer, Sebastian Nowozin, and Richard E. Turner.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 83, + 447, + 96 + ], + "spans": [ + { + "bbox": [ + 132, + 83, + 447, + 96 + ], + "score": 1.0, + "content": "Meta-learning probabilistic inference for prediction. arXiv:1805.09921, 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 111, + 103, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 110, + 102, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 110, + 102, + 505, + 117 + ], + "score": 1.0, + "content": "[32] Alex Graves. Generating sequences with recurrent neural networks. arXiv:1308.0850, 2013.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 111, + 123, + 502, + 158 + ], + "lines": [ + { + "bbox": [ + 110, + 123, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 110, + 123, + 505, + 137 + ], + "score": 1.0, + "content": "[33] Thomas L. Griffiths, Frederick Callaway, Michael B. Chang, Erin Grant, Paul M. Krueger,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 135, + 504, + 147 + ], + "spans": [ + { + "bbox": [ + 133, + 135, + 504, + 147 + ], + "score": 1.0, + "content": "and Falk Lieder. Doing more with less: meta-reasoning and meta-learning in humans and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 146, + 367, + 158 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 367, + 158 + ], + "score": 1.0, + "content": "machines. Current Opinion in Behavioral Sciences, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 110, + 165, + 504, + 189 + ], + "lines": [ + { + "bbox": [ + 110, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 110, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "[34] Liangke Gui, Borui Wang, Qiuyuan Huang, Alex Hauptmann, Yonatan Bisk, and Jianfeng Gao.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 177, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 177, + 506, + 189 + ], + "score": 1.0, + "content": "KAT: A knowledge augmented transformer for vision-and-language. arXiv:2112.08614, 2021.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 111, + 196, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 110, + 196, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 110, + 196, + 506, + 210 + ], + "score": 1.0, + "content": "[35] Danna Gurari, Qing Li, Abigale J. Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "Luo, and Jeffrey P. Bigham. VizWiz grand challenge: Answering visual questions from blind", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 219, + 397, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 397, + 232 + ], + "score": 1.0, + "content": "people. In IEEE Computer Vision and Pattern Recognition, 2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 110, + 239, + 506, + 262 + ], + "lines": [ + { + "bbox": [ + 111, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 111, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "[36] Adi Haviv, Ori Ram, Ofir Press, Peter Izsak, and Omer Levy. Transformer language models", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 249, + 487, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 487, + 262 + ], + "score": 1.0, + "content": "without positional encodings still learn positional information. arXiv:2203.16634, 2022.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 111, + 270, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 110, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 110, + 270, + 506, + 282 + ], + "score": 1.0, + "content": "[37] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Women also snowboard: Overcoming bias in captioning models. In European Conference on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 291, + 230, + 305 + ], + "spans": [ + { + "bbox": [ + 133, + 291, + 230, + 305 + ], + "score": 1.0, + "content": "Computer Vision, 2018.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 111, + 311, + 503, + 346 + ], + "lines": [ + { + "bbox": [ + 111, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 111, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "[38] Lisa Anne Hendricks, John Mellor, Rosalia Schneider, Jean-Baptiste Alayrac, and Aida", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 324, + 504, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 324, + 504, + 335 + ], + "score": 1.0, + "content": "Nematzadeh. Decoupling the role of data, attention, and losses in multimodal transformers.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 334, + 423, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 423, + 347 + ], + "score": 1.0, + "content": "Annual Meeting of the Association for Computational Linguistics, 2021.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 110, + 353, + 504, + 377 + ], + "lines": [ + { + "bbox": [ + 111, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 111, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "[39] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (GELUs). arXiv:1606.08415,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 364, + 159, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 364, + 159, + 378 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 111, + 385, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 111, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 111, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "[40] Tom Hennigan, Trevor Cai, Tamara Norman, and Igor Babuschkin. Haiku: Sonnet for JAX,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 395, + 370, + 409 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 370, + 409 + ], + "score": 1.0, + "content": "2020. URL http://github.com/deepmind/dm-haiku.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 110, + 416, + 504, + 439 + ], + "lines": [ + { + "bbox": [ + 110, + 415, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 110, + 415, + 506, + 430 + ], + "score": 1.0, + "content": "[41] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 131, + 425, + 159, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 425, + 159, + 439 + ], + "score": 1.0, + "content": "1997.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 111, + 447, + 506, + 504 + ], + "lines": [ + { + "bbox": [ + 111, + 447, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 111, + 447, + 506, + 459 + ], + "score": 1.0, + "content": "[42] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 469, + 507, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 469, + 507, + 483 + ], + "score": 1.0, + "content": "Eric Noland Tom Hennigan, Katie Millican, George van den Driessche, Bogdan Damoc,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 132, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 491, + 497, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 497, + 504 + ], + "score": 1.0, + "content": "Laurent Sifre. Training compute-optimal large language models. arXiv:2203.15556, 2022.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 110, + 511, + 504, + 545 + ], + "lines": [ + { + "bbox": [ + 110, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 110, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "[43] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 533, + 403, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 403, + 546 + ], + "score": 1.0, + "content": "for NLP. In International Conference on Machine Learning, 2019.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 111, + 553, + 504, + 576 + ], + "lines": [ + { + "bbox": [ + 111, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 111, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "[44] Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classifica-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 133, + 564, + 257, + 576 + ], + "spans": [ + { + "bbox": [ + 133, + 564, + 257, + 576 + ], + "score": 1.0, + "content": "tion. arXiv:1801.06146, 2018.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 111, + 584, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 111, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 111, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "[45] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 595, + 507, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 507, + 609 + ], + "score": 1.0, + "content": "Wang. Scaling up vision-language pre-training for image captioning. arXiv:2111.12233, 2021.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 504, + 639 + ], + "lines": [ + { + "bbox": [ + 110, + 614, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 110, + 614, + 506, + 630 + ], + "score": 1.0, + "content": "[46] Lun Huang, Wenmin Wang, Jie Chen, and Xiao-Yong Wei. Attention on attention for image", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 132, + 627, + 405, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 627, + 405, + 639 + ], + "score": 1.0, + "content": "captioning. In International Conference on Computer Vision, 2019.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 110, + 646, + 505, + 680 + ], + "lines": [ + { + "bbox": [ + 110, + 646, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 110, + 646, + 506, + 660 + ], + "score": 1.0, + "content": "[47] Md Amirul Islam, Matthew Kowal, Sen Jia, Konstantinos G. Derpanis, and Neil D. B. Bruce.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 657, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 132, + 657, + 506, + 671 + ], + "score": 1.0, + "content": "Global pooling, more than meets the eye: Position information is encoded channel-wise in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 133, + 669, + 388, + 681 + ], + "spans": [ + { + "bbox": [ + 133, + 669, + 388, + 681 + ], + "score": 1.0, + "content": "CNNs. In International Conference on Computer Vision, 2021.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 112, + 688, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 110, + 687, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 110, + 687, + 505, + 702 + ], + "score": 1.0, + "content": "[48] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 133, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 133, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "Carreira. Perceiver: General perception with iterative attention. In International Conference", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 132, + 711, + 249, + 722 + ], + "spans": [ + { + "bbox": [ + 132, + 711, + 249, + 722 + ], + "score": 1.0, + "content": "on Machine Learning, 2021.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 110, + 72, + 505, + 96 + ], + "lines": [ + { + "bbox": [ + 110, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 110, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[31] Jonathan Gordon, John Bronskill, Matthias Bauer, Sebastian Nowozin, and Richard E. Turner.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 83, + 447, + 96 + ], + "spans": [ + { + "bbox": [ + 132, + 83, + 447, + 96 + ], + "score": 1.0, + "content": "Meta-learning probabilistic inference for prediction. arXiv:1805.09921, 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 110, + 72, + 507, + 96 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 103, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 110, + 102, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 110, + 102, + 505, + 117 + ], + "score": 1.0, + "content": "[32] Alex Graves. Generating sequences with recurrent neural networks. arXiv:1308.0850, 2013.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 110, + 102, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 123, + 502, + 158 + ], + "lines": [ + { + "bbox": [ + 110, + 123, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 110, + 123, + 505, + 137 + ], + "score": 1.0, + "content": "[33] Thomas L. Griffiths, Frederick Callaway, Michael B. Chang, Erin Grant, Paul M. Krueger,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 135, + 504, + 147 + ], + "spans": [ + { + "bbox": [ + 133, + 135, + 504, + 147 + ], + "score": 1.0, + "content": "and Falk Lieder. Doing more with less: meta-reasoning and meta-learning in humans and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 146, + 367, + 158 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 367, + 158 + ], + "score": 1.0, + "content": "machines. Current Opinion in Behavioral Sciences, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 110, + 123, + 505, + 158 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 165, + 504, + 189 + ], + "lines": [ + { + "bbox": [ + 110, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 110, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "[34] Liangke Gui, Borui Wang, Qiuyuan Huang, Alex Hauptmann, Yonatan Bisk, and Jianfeng Gao.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 177, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 177, + 506, + 189 + ], + "score": 1.0, + "content": "KAT: A knowledge augmented transformer for vision-and-language. arXiv:2112.08614, 2021.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 110, + 165, + 506, + 189 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 196, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 110, + 196, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 110, + 196, + 506, + 210 + ], + "score": 1.0, + "content": "[35] Danna Gurari, Qing Li, Abigale J. Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "Luo, and Jeffrey P. Bigham. VizWiz grand challenge: Answering visual questions from blind", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 219, + 397, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 397, + 232 + ], + "score": 1.0, + "content": "people. In IEEE Computer Vision and Pattern Recognition, 2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 110, + 196, + 506, + 232 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 239, + 506, + 262 + ], + "lines": [ + { + "bbox": [ + 111, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 111, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "[36] Adi Haviv, Ori Ram, Ofir Press, Peter Izsak, and Omer Levy. Transformer language models", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 249, + 487, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 487, + 262 + ], + "score": 1.0, + "content": "without positional encodings still learn positional information. arXiv:2203.16634, 2022.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 111, + 239, + 505, + 262 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 270, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 110, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 110, + 270, + 506, + 282 + ], + "score": 1.0, + "content": "[37] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Women also snowboard: Overcoming bias in captioning models. In European Conference on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 291, + 230, + 305 + ], + "spans": [ + { + "bbox": [ + 133, + 291, + 230, + 305 + ], + "score": 1.0, + "content": "Computer Vision, 2018.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 110, + 270, + 506, + 305 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 311, + 503, + 346 + ], + "lines": [ + { + "bbox": [ + 111, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 111, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "[38] Lisa Anne Hendricks, John Mellor, Rosalia Schneider, Jean-Baptiste Alayrac, and Aida", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 324, + 504, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 324, + 504, + 335 + ], + "score": 1.0, + "content": "Nematzadeh. Decoupling the role of data, attention, and losses in multimodal transformers.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 334, + 423, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 423, + 347 + ], + "score": 1.0, + "content": "Annual Meeting of the Association for Computational Linguistics, 2021.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 111, + 312, + 505, + 347 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 353, + 504, + 377 + ], + "lines": [ + { + "bbox": [ + 111, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 111, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "[39] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (GELUs). arXiv:1606.08415,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 364, + 159, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 364, + 159, + 378 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 111, + 353, + 506, + 378 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 385, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 111, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 111, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "[40] Tom Hennigan, Trevor Cai, Tamara Norman, and Igor Babuschkin. Haiku: Sonnet for JAX,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 395, + 370, + 409 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 370, + 409 + ], + "score": 1.0, + "content": "2020. URL http://github.com/deepmind/dm-haiku.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 111, + 384, + 506, + 409 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 416, + 504, + 439 + ], + "lines": [ + { + "bbox": [ + 110, + 415, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 110, + 415, + 506, + 430 + ], + "score": 1.0, + "content": "[41] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 131, + 425, + 159, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 425, + 159, + 439 + ], + "score": 1.0, + "content": "1997.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 110, + 415, + 506, + 439 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 447, + 506, + 504 + ], + "lines": [ + { + "bbox": [ + 111, + 447, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 111, + 447, + 506, + 459 + ], + "score": 1.0, + "content": "[42] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 469, + 507, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 469, + 507, + 483 + ], + "score": 1.0, + "content": "Eric Noland Tom Hennigan, Katie Millican, George van den Driessche, Bogdan Damoc,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 132, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 491, + 497, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 497, + 504 + ], + "score": 1.0, + "content": "Laurent Sifre. Training compute-optimal large language models. arXiv:2203.15556, 2022.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 111, + 447, + 507, + 504 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 511, + 504, + 545 + ], + "lines": [ + { + "bbox": [ + 110, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 110, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "[43] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 533, + 403, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 403, + 546 + ], + "score": 1.0, + "content": "for NLP. In International Conference on Machine Learning, 2019.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 110, + 511, + 506, + 546 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 553, + 504, + 576 + ], + "lines": [ + { + "bbox": [ + 111, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 111, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "[44] Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classifica-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 133, + 564, + 257, + 576 + ], + "spans": [ + { + "bbox": [ + 133, + 564, + 257, + 576 + ], + "score": 1.0, + "content": "tion. arXiv:1801.06146, 2018.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 111, + 553, + 505, + 576 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 584, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 111, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 111, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "[45] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 595, + 507, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 507, + 609 + ], + "score": 1.0, + "content": "Wang. Scaling up vision-language pre-training for image captioning. arXiv:2111.12233, 2021.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 111, + 584, + 507, + 609 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 504, + 639 + ], + "lines": [ + { + "bbox": [ + 110, + 614, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 110, + 614, + 506, + 630 + ], + "score": 1.0, + "content": "[46] Lun Huang, Wenmin Wang, Jie Chen, and Xiao-Yong Wei. Attention on attention for image", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 132, + 627, + 405, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 627, + 405, + 639 + ], + "score": 1.0, + "content": "captioning. In International Conference on Computer Vision, 2019.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 110, + 614, + 506, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 646, + 505, + 680 + ], + "lines": [ + { + "bbox": [ + 110, + 646, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 110, + 646, + 506, + 660 + ], + "score": 1.0, + "content": "[47] Md Amirul Islam, Matthew Kowal, Sen Jia, Konstantinos G. Derpanis, and Neil D. B. Bruce.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 657, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 132, + 657, + 506, + 671 + ], + "score": 1.0, + "content": "Global pooling, more than meets the eye: Position information is encoded channel-wise in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 133, + 669, + 388, + 681 + ], + "spans": [ + { + "bbox": [ + 133, + 669, + 388, + 681 + ], + "score": 1.0, + "content": "CNNs. In International Conference on Computer Vision, 2021.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 110, + 646, + 506, + 681 + ] + }, + { + "type": "text", + "bbox": [ + 112, + 688, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 110, + 687, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 110, + 687, + 505, + 702 + ], + "score": 1.0, + "content": "[48] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 133, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 133, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "Carreira. Perceiver: General perception with iterative attention. In International Conference", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 132, + 711, + 249, + 722 + ], + "spans": [ + { + "bbox": [ + 132, + 711, + 249, + 722 + ], + "score": 1.0, + "content": "on Machine Learning, 2021.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 110, + 687, + 505, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 41, + 507, + 727 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[49] Aashi Jain, Mandy Guo, Krishna Srinivasan, Ting Chen, Sneha Kudugunta, Chao Jia, Yinfei", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 82, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 132, + 82, + 506, + 97 + ], + "score": 1.0, + "content": "Yang, and Jason Baldridge. MURAL: multimodal, multitask retrieval across languages.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 95, + 236, + 106 + ], + "spans": [ + { + "bbox": [ + 132, + 95, + 236, + 106 + ], + "score": 1.0, + "content": "arXiv:2109.05125, 2021.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 113, + 507, + 127 + ], + "spans": [ + { + "bbox": [ + 111, + 113, + 507, + 127 + ], + "score": 1.0, + "content": "[50] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 124, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 124, + 505, + 138 + ], + "score": 1.0, + "content": "Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 135, + 385, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 135, + 385, + 149 + ], + "score": 1.0, + "content": "learning with noisy text supervision. arXiv:2102.05918, 2021.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 155, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 111, + 155, + 506, + 169 + ], + "score": 1.0, + "content": "[51] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 165, + 506, + 180 + ], + "spans": [ + { + "bbox": [ + 131, + 165, + 506, + 180 + ], + "score": 1.0, + "content": "Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 176, + 290, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 290, + 190 + ], + "score": 1.0, + "content": "pre-training. arXiv:2203.07303, 2022.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 194, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 110, + 194, + 506, + 211 + ], + "score": 1.0, + "content": "[52] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 207, + 370, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 370, + 222 + ], + "score": 1.0, + "content": "the limits of language modeling. arXiv:1602.02410, 2016.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 111, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 111, + 226, + 506, + 240 + ], + "score": 1.0, + "content": "[53] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 248, + 309, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 309, + 261 + ], + "score": 1.0, + "content": "language models. arXiv:2001.08361, 2020.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 111, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 111, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "[54] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 278, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 506, + 292 + ], + "score": 1.0, + "content": "Ringshia, and Davide Testuggine. The Hateful Memes Challenge: Detecting hate speech in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 289, + 465, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 465, + 303 + ], + "score": 1.0, + "content": "multimodal memes. Conference on Neural Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 111, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 111, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "[55] Hugo Larochelle. Few-shot classification by recycling deep learning. Invited Talk at the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 318, + 504, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 504, + 334 + ], + "score": 1.0, + "content": "S2D-OLAD Workshop, ICLR 2021, 2021. URL https://slideslive.com/38955350/", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 329, + 398, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 398, + 345 + ], + "score": 1.0, + "content": "fewshot-classification-by-recycling-deep-learning.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 111, + 349, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 111, + 349, + 506, + 364 + ], + "score": 1.0, + "content": "[56] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 131, + 361, + 300, + 375 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 300, + 375 + ], + "score": 1.0, + "content": "prompt tuning. arXiv:2104.08691, 2021.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 111, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 111, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "[57] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 131, + 391, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 131, + 391, + 506, + 406 + ], + "score": 1.0, + "content": "Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 131, + 402, + 488, + 416 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 488, + 416 + ], + "score": 1.0, + "content": "momentum distillation. In Conference on Neural Information Processing Systems, 2021.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 108, + 419, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 108, + 419, + 506, + 437 + ], + "score": 1.0, + "content": "[58] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 432, + 507, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 507, + 447 + ], + "score": 1.0, + "content": "pre-training for unified vision-language understanding and generation. arXiv:2201.12086,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 443, + 158, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 158, + 457 + ], + "score": 1.0, + "content": "2022.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 111, + 464, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 111, + 464, + 506, + 476 + ], + "score": 1.0, + "content": "[59] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. HERO:", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 131, + 473, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 506, + 488 + ], + "score": 1.0, + "content": "Hierarchical encoder for video+language omni-representation pre-training. arXiv:2005.00200,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 132, + 485, + 160, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 160, + 498 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 111, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 111, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "[60] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 516, + 237, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 237, + 528 + ], + "score": 1.0, + "content": "arXiv:2101.00190, 2021.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 109, + 532, + 507, + 550 + ], + "spans": [ + { + "bbox": [ + 109, + 532, + 507, + 550 + ], + "score": 1.0, + "content": "[61] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 132, + 555, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 506, + 570 + ], + "score": 1.0, + "content": "aligned pre-training for vision-language tasks. In European Conference on Computer Vision,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 131, + 566, + 160, + 581 + ], + "spans": [ + { + "bbox": [ + 131, + 566, + 160, + 581 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 111, + 587, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 111, + 587, + 506, + 599 + ], + "score": 1.0, + "content": "[62] Phillip Lippe, Nithin Holla, Shantanu Chandra, Santhosh Rajamanickam, Georgios Antoniou,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 132, + 599, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 505, + 610 + ], + "score": 1.0, + "content": "Ekaterina Shutova, and Helen Yannakoudakis. A multimodal framework for the detection of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 132, + 609, + 300, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 300, + 621 + ], + "score": 1.0, + "content": "hateful memes. arXiv:2012.12871, 2020.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 110, + 628, + 507, + 641 + ], + "spans": [ + { + "bbox": [ + 110, + 628, + 507, + 641 + ], + "score": 1.0, + "content": "[63] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 132, + 638, + 443, + 653 + ], + "spans": [ + { + "bbox": [ + 132, + 638, + 443, + 653 + ], + "score": 1.0, + "content": "What makes good in-context examples for GPT-3? arXiv:2101.06804, 2021.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 111, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 111, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "[64] Siqi Liu, Zhenhai Zhu, Ning Ye, Sergio Guadarrama, and Kevin Murphy. Optimization of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 132, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 132, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "image description metrics using policy gradient methods. In International Conference on", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 132, + 680, + 231, + 694 + ], + "spans": [ + { + "bbox": [ + 132, + 680, + 231, + 694 + ], + "score": 1.0, + "content": "Computer Vision, 2017.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 108, + 698, + 507, + 715 + ], + "spans": [ + { + "bbox": [ + 108, + 698, + 507, + 715 + ], + "score": 1.0, + "content": "[65] Yu Liu, Lianghua Huang, Liuyihang Song, Bin Wang, Yingya Zhang, and Pan Pan. Enhancing", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 132, + 710, + 462, + 724 + ], + "spans": [ + { + "bbox": [ + 132, + 710, + 462, + 724 + ], + "score": 1.0, + "content": "textual cues in multi-modal transformers for VQA. VizWiz Challenge 2021, 2021.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 23 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 108, + 41, + 507, + 727 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[49] Aashi Jain, Mandy Guo, Krishna Srinivasan, Ting Chen, Sneha Kudugunta, Chao Jia, Yinfei", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 82, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 132, + 82, + 506, + 97 + ], + "score": 1.0, + "content": "Yang, and Jason Baldridge. MURAL: multimodal, multitask retrieval across languages.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 95, + 236, + 106 + ], + "spans": [ + { + "bbox": [ + 132, + 95, + 236, + 106 + ], + "score": 1.0, + "content": "arXiv:2109.05125, 2021.", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 113, + 507, + 127 + ], + "spans": [ + { + "bbox": [ + 111, + 113, + 507, + 127 + ], + "score": 1.0, + "content": "[50] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 124, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 124, + 505, + 138 + ], + "score": 1.0, + "content": "Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 135, + 385, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 135, + 385, + 149 + ], + "score": 1.0, + "content": "learning with noisy text supervision. arXiv:2102.05918, 2021.", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 155, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 111, + 155, + 506, + 169 + ], + "score": 1.0, + "content": "[51] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu,", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 165, + 506, + 180 + ], + "spans": [ + { + "bbox": [ + 131, + 165, + 506, + 180 + ], + "score": 1.0, + "content": "Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 176, + 290, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 290, + 190 + ], + "score": 1.0, + "content": "pre-training. arXiv:2203.07303, 2022.", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 194, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 110, + 194, + 506, + 211 + ], + "score": 1.0, + "content": "[52] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 207, + 370, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 370, + 222 + ], + "score": 1.0, + "content": "the limits of language modeling. arXiv:1602.02410, 2016.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 111, + 226, + 506, + 240 + ], + "score": 1.0, + "content": "[53] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 248, + 309, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 309, + 261 + ], + "score": 1.0, + "content": "language models. arXiv:2001.08361, 2020.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 111, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "[54] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 278, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 506, + 292 + ], + "score": 1.0, + "content": "Ringshia, and Davide Testuggine. The Hateful Memes Challenge: Detecting hate speech in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 289, + 465, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 465, + 303 + ], + "score": 1.0, + "content": "multimodal memes. Conference on Neural Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 111, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "[55] Hugo Larochelle. Few-shot classification by recycling deep learning. Invited Talk at the", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 130, + 318, + 504, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 504, + 334 + ], + "score": 1.0, + "content": "S2D-OLAD Workshop, ICLR 2021, 2021. URL https://slideslive.com/38955350/", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 329, + 398, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 398, + 345 + ], + "score": 1.0, + "content": "fewshot-classification-by-recycling-deep-learning.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 349, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 111, + 349, + 506, + 364 + ], + "score": 1.0, + "content": "[56] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 361, + 300, + 375 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 300, + 375 + ], + "score": 1.0, + "content": "prompt tuning. arXiv:2104.08691, 2021.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 111, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "[57] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 391, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 131, + 391, + 506, + 406 + ], + "score": 1.0, + "content": "Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 131, + 402, + 488, + 416 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 488, + 416 + ], + "score": 1.0, + "content": "momentum distillation. In Conference on Neural Information Processing Systems, 2021.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 108, + 419, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 108, + 419, + 506, + 437 + ], + "score": 1.0, + "content": "[58] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 130, + 432, + 507, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 507, + 447 + ], + "score": 1.0, + "content": "pre-training for unified vision-language understanding and generation. arXiv:2201.12086,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 443, + 158, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 158, + 457 + ], + "score": 1.0, + "content": "2022.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 464, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 111, + 464, + 506, + 476 + ], + "score": 1.0, + "content": "[59] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. HERO:", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 473, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 506, + 488 + ], + "score": 1.0, + "content": "Hierarchical encoder for video+language omni-representation pre-training. arXiv:2005.00200,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 132, + 485, + 160, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 160, + 498 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 111, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "[60] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation.", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 516, + 237, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 237, + 528 + ], + "score": 1.0, + "content": "arXiv:2101.00190, 2021.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 532, + 507, + 550 + ], + "spans": [ + { + "bbox": [ + 109, + 532, + 507, + 550 + ], + "score": 1.0, + "content": "[61] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang,", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 132, + 555, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 506, + 570 + ], + "score": 1.0, + "content": "aligned pre-training for vision-language tasks. In European Conference on Computer Vision,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 131, + 566, + 160, + 581 + ], + "spans": [ + { + "bbox": [ + 131, + 566, + 160, + 581 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 587, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 111, + 587, + 506, + 599 + ], + "score": 1.0, + "content": "[62] Phillip Lippe, Nithin Holla, Shantanu Chandra, Santhosh Rajamanickam, Georgios Antoniou,", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 599, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 505, + 610 + ], + "score": 1.0, + "content": "Ekaterina Shutova, and Helen Yannakoudakis. A multimodal framework for the detection of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 132, + 609, + 300, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 300, + 621 + ], + "score": 1.0, + "content": "hateful memes. arXiv:2012.12871, 2020.", + "type": "text" + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 628, + 507, + 641 + ], + "spans": [ + { + "bbox": [ + 110, + 628, + 507, + 641 + ], + "score": 1.0, + "content": "[63] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen.", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 638, + 443, + 653 + ], + "spans": [ + { + "bbox": [ + 132, + 638, + 443, + 653 + ], + "score": 1.0, + "content": "What makes good in-context examples for GPT-3? arXiv:2101.06804, 2021.", + "type": "text" + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 111, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "[64] Siqi Liu, Zhenhai Zhu, Ning Ye, Sergio Guadarrama, and Kevin Murphy. Optimization of", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 132, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "image description metrics using policy gradient methods. In International Conference on", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 132, + 680, + 231, + 694 + ], + "spans": [ + { + "bbox": [ + 132, + 680, + 231, + 694 + ], + "score": 1.0, + "content": "Computer Vision, 2017.", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 108, + 698, + 507, + 715 + ], + "spans": [ + { + "bbox": [ + 108, + 698, + 507, + 715 + ], + "score": 1.0, + "content": "[65] Yu Liu, Lianghua Huang, Liuyihang Song, Bin Wang, Yingya Zhang, and Pan Pan. Enhancing", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 710, + 462, + 724 + ], + "spans": [ + { + "bbox": [ + 132, + 710, + 462, + 724 + ], + "score": 1.0, + "content": "textual cues in multi-modal transformers for VQA. VizWiz Challenge 2021, 2021.", + "type": "text" + } + ], + "index": 46, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 110, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[66] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViLBERT: Pretraining task-agnostic vi-", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 132, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "siolinguistic representations for vision-and-language tasks. Conference on Neural Information", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 95, + 240, + 106 + ], + "spans": [ + { + "bbox": [ + 132, + 95, + 240, + 106 + ], + "score": 1.0, + "content": "Processing Systems, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 112, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 110, + 112, + 506, + 126 + ], + "score": 1.0, + "content": "[67] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 123, + 507, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 123, + 507, + 137 + ], + "score": 1.0, + "content": "Bharti, and Ming Zhou. UniVL: A unified video and language pre-training model for multi-", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 136, + 387, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 387, + 147 + ], + "score": 1.0, + "content": "modal understanding and generation. arXiv:2002.06353, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 152, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 110, + 152, + 506, + 166 + ], + "score": 1.0, + "content": "[68] Ziyang Luo, Yadong Xi, Rongsheng Zhang, and Jing Ma. VC-GPT: Visual conditioned GPT", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 164, + 476, + 178 + ], + "spans": [ + { + "bbox": [ + 131, + 164, + 476, + 178 + ], + "score": 1.0, + "content": "for end-to-end generative vision-and-language pre-training. arXiv:2201.12723, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 182, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 110, + 182, + 506, + 196 + ], + "score": 1.0, + "content": "[69] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. OK-VQA: A", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 193, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 506, + 207 + ], + "score": 1.0, + "content": "visual question answering benchmark requiring external knowledge. In IEEE Computer Vision", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 203, + 262, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 262, + 217 + ], + "score": 1.0, + "content": "and Pattern Recognition, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 222, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 110, + 222, + 506, + 236 + ], + "score": 1.0, + "content": "[70] Ellen M. Markman. Categorization and naming in children: Problems of induction. MIT", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 234, + 184, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 234, + 184, + 245 + ], + "score": 1.0, + "content": "Press, 1989.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 251, + 507, + 265 + ], + "spans": [ + { + "bbox": [ + 110, + 251, + 507, + 265 + ], + "score": 1.0, + "content": "[71] Michael McCloskey and Neil J. Cohen. Catastrophic interference in connectionist networks:", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 262, + 474, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 474, + 276 + ], + "score": 1.0, + "content": "The sequential learning problem. The Psychology of Learning and Motivation, 1989.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 280, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 110, + 280, + 506, + 294 + ], + "score": 1.0, + "content": "[72] Jacob Menick, Maja Trebacz, Vladimir Mikulik, John Aslanides, Francis Song, Martin", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 130, + 289, + 507, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 507, + 307 + ], + "score": 1.0, + "content": "Chadwick, Mia Glaese, Susannah Young, Lucy Campbell-Gillingham, Geoffrey Irving,", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 303, + 507, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 507, + 317 + ], + "score": 1.0, + "content": "and Nat McAleese. Teaching language models to support answers with verified quotes.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 314, + 236, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 236, + 326 + ], + "score": 1.0, + "content": "arXiv:2203.11147, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 331, + 507, + 345 + ], + "spans": [ + { + "bbox": [ + 110, + 331, + 507, + 345 + ], + "score": 1.0, + "content": "[73] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 343, + 434, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 434, + 355 + ], + "score": 1.0, + "content": "RareAct: A video dataset of unusual interactions. arxiv:2008.01018, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 360, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 110, + 360, + 506, + 374 + ], + "score": 1.0, + "content": "[74] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 371, + 507, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 507, + 385 + ], + "score": 1.0, + "content": "Zisserman. End-to-end learning of visual representations from uncurated instructional videos.", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 382, + 366, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 366, + 396 + ], + "score": 1.0, + "content": "In IEEE Computer Vision and Pattern Recognition, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 400, + 507, + 415 + ], + "spans": [ + { + "bbox": [ + 110, + 400, + 507, + 415 + ], + "score": 1.0, + "content": "[75] Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernocky, and Sanjeev Khudanpur. `", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 412, + 407, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 407, + 424 + ], + "score": 1.0, + "content": "Recurrent neural network based language model. Interspeech, 2010.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 110, + 429, + 506, + 443 + ], + "score": 1.0, + "content": "[76] Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 130, + 439, + 507, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 507, + 457 + ], + "score": 1.0, + "content": "Luke Zettlemoyer. Rethinking the role of demonstrations: What makes in-context learning", + "type": "text", + "cross_page": true + } + ], + "index": 27 + }, + { + "bbox": [ + 132, + 452, + 264, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 264, + 464 + ], + "score": 1.0, + "content": "work? arXiv:2202.12837, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 470, + 507, + 484 + ], + "spans": [ + { + "bbox": [ + 110, + 470, + 507, + 484 + ], + "score": 1.0, + "content": "[77] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 481, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 506, + 495 + ], + "score": 1.0, + "content": "Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model", + "type": "text", + "cross_page": true + } + ], + "index": 30 + }, + { + "bbox": [ + 131, + 491, + 474, + 507 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 474, + 507 + ], + "score": 1.0, + "content": "reporting. In ACM Conference on Fairness, Accountability, and Transparency, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 509, + 507, + 524 + ], + "spans": [ + { + "bbox": [ + 110, + 509, + 507, + 524 + ], + "score": 1.0, + "content": "[78] Ron Mokady, Amir Hertz, and Amit H. Bermano. ClipCap: CLIP prefix for image captioning.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 521, + 237, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 237, + 535 + ], + "score": 1.0, + "content": "arXiv:2111.09734, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 539, + 507, + 552 + ], + "spans": [ + { + "bbox": [ + 110, + 539, + 507, + 552 + ], + "score": 1.0, + "content": "[79] Vishvak Murahari, Dhruv Batra, Devi Parikh, and Abhishek Das. Large-scale pretraining for", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 549, + 507, + 564 + ], + "spans": [ + { + "bbox": [ + 131, + 549, + 507, + 564 + ], + "score": 1.0, + "content": "visual dialog: A simple state-of-the-art baseline. In European Conference on Computer Vision,", + "type": "text", + "cross_page": true + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 561, + 159, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 159, + 574 + ], + "score": 1.0, + "content": "2020.", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 581, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 111, + 581, + 506, + 592 + ], + "score": 1.0, + "content": "[80] Ethan Perez, Douwe Kiela, and Kyunghyun Cho. True few-shot learning with language models.", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 591, + 381, + 603 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 381, + 603 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 609, + 507, + 622 + ], + "spans": [ + { + "bbox": [ + 110, + 609, + 507, + 622 + ], + "score": 1.0, + "content": "[81] Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 130, + 617, + 507, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 507, + 635 + ], + "score": 1.0, + "content": "Glaese, Nat McAleese, and Geoffrey Irving. Red teaming language models with language", + "type": "text", + "cross_page": true + } + ], + "index": 40 + }, + { + "bbox": [ + 131, + 628, + 271, + 645 + ], + "spans": [ + { + "bbox": [ + 131, + 628, + 271, + 645 + ], + "score": 1.0, + "content": "models. arXiv:2202.03286, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 648, + 507, + 662 + ], + "spans": [ + { + "bbox": [ + 110, + 648, + 507, + 662 + ], + "score": 1.0, + "content": "[82] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Lu-", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 130, + 657, + 507, + 675 + ], + "spans": [ + { + "bbox": [ + 130, + 657, + 507, + 675 + ], + "score": 1.0, + "content": "ong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning.", + "type": "text", + "cross_page": true + } + ], + "index": 43 + }, + { + "bbox": [ + 131, + 670, + 236, + 684 + ], + "spans": [ + { + "bbox": [ + 131, + 670, + 236, + 684 + ], + "score": 1.0, + "content": "arXiv:2111.10050, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 110, + 689, + 506, + 703 + ], + "score": 1.0, + "content": "[83] Ofir Press, Noah Smith, and Mike Lewis. Train short, test long: Attention with linear biases", + "type": "text", + "cross_page": true + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 700, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 131, + 700, + 506, + 714 + ], + "score": 1.0, + "content": "enables input length extrapolation. In International Conference on Learning Representations,", + "type": "text", + "cross_page": true + } + ], + "index": 46 + }, + { + "bbox": [ + 132, + 710, + 158, + 723 + ], + "spans": [ + { + "bbox": [ + 132, + 710, + 158, + 723 + ], + "score": 1.0, + "content": "2022.", + "type": "text", + "cross_page": true + } + ], + "index": 47, + "is_list_end_line": true + } + ], + "index": 23, + "bbox_fs": [ + 108, + 72, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 109, + 56, + 507, + 727 + ], + "lines": [ + { + "bbox": [ + 110, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 110, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[66] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViLBERT: Pretraining task-agnostic vi-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 132, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "siolinguistic representations for vision-and-language tasks. Conference on Neural Information", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 95, + 240, + 106 + ], + "spans": [ + { + "bbox": [ + 132, + 95, + 240, + 106 + ], + "score": 1.0, + "content": "Processing Systems, 2019.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 112, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 110, + 112, + 506, + 126 + ], + "score": 1.0, + "content": "[67] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 123, + 507, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 123, + 507, + 137 + ], + "score": 1.0, + "content": "Bharti, and Ming Zhou. UniVL: A unified video and language pre-training model for multi-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 136, + 387, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 387, + 147 + ], + "score": 1.0, + "content": "modal understanding and generation. arXiv:2002.06353, 2020.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 152, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 110, + 152, + 506, + 166 + ], + "score": 1.0, + "content": "[68] Ziyang Luo, Yadong Xi, Rongsheng Zhang, and Jing Ma. VC-GPT: Visual conditioned GPT", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 164, + 476, + 178 + ], + "spans": [ + { + "bbox": [ + 131, + 164, + 476, + 178 + ], + "score": 1.0, + "content": "for end-to-end generative vision-and-language pre-training. arXiv:2201.12723, 2022.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 182, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 110, + 182, + 506, + 196 + ], + "score": 1.0, + "content": "[69] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. OK-VQA: A", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 193, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 506, + 207 + ], + "score": 1.0, + "content": "visual question answering benchmark requiring external knowledge. In IEEE Computer Vision", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 203, + 262, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 262, + 217 + ], + "score": 1.0, + "content": "and Pattern Recognition, 2019.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 110, + 222, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 110, + 222, + 506, + 236 + ], + "score": 1.0, + "content": "[70] Ellen M. Markman. Categorization and naming in children: Problems of induction. MIT", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 234, + 184, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 234, + 184, + 245 + ], + "score": 1.0, + "content": "Press, 1989.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 110, + 251, + 507, + 265 + ], + "spans": [ + { + "bbox": [ + 110, + 251, + 507, + 265 + ], + "score": 1.0, + "content": "[71] Michael McCloskey and Neil J. Cohen. Catastrophic interference in connectionist networks:", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 262, + 474, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 474, + 276 + ], + "score": 1.0, + "content": "The sequential learning problem. The Psychology of Learning and Motivation, 1989.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 110, + 280, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 110, + 280, + 506, + 294 + ], + "score": 1.0, + "content": "[72] Jacob Menick, Maja Trebacz, Vladimir Mikulik, John Aslanides, Francis Song, Martin", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 289, + 507, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 507, + 307 + ], + "score": 1.0, + "content": "Chadwick, Mia Glaese, Susannah Young, Lucy Campbell-Gillingham, Geoffrey Irving,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 303, + 507, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 507, + 317 + ], + "score": 1.0, + "content": "and Nat McAleese. Teaching language models to support answers with verified quotes.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 314, + 236, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 236, + 326 + ], + "score": 1.0, + "content": "arXiv:2203.11147, 2022.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 110, + 331, + 507, + 345 + ], + "spans": [ + { + "bbox": [ + 110, + 331, + 507, + 345 + ], + "score": 1.0, + "content": "[73] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 343, + 434, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 434, + 355 + ], + "score": 1.0, + "content": "RareAct: A video dataset of unusual interactions. arxiv:2008.01018, 2020.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 110, + 360, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 110, + 360, + 506, + 374 + ], + "score": 1.0, + "content": "[74] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 371, + 507, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 507, + 385 + ], + "score": 1.0, + "content": "Zisserman. End-to-end learning of visual representations from uncurated instructional videos.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 382, + 366, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 366, + 396 + ], + "score": 1.0, + "content": "In IEEE Computer Vision and Pattern Recognition, 2020.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 110, + 400, + 507, + 415 + ], + "spans": [ + { + "bbox": [ + 110, + 400, + 507, + 415 + ], + "score": 1.0, + "content": "[75] Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernocky, and Sanjeev Khudanpur. `", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 412, + 407, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 407, + 424 + ], + "score": 1.0, + "content": "Recurrent neural network based language model. Interspeech, 2010.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 110, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 110, + 429, + 506, + 443 + ], + "score": 1.0, + "content": "[76] Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 439, + 507, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 507, + 457 + ], + "score": 1.0, + "content": "Luke Zettlemoyer. Rethinking the role of demonstrations: What makes in-context learning", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 132, + 452, + 264, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 264, + 464 + ], + "score": 1.0, + "content": "work? arXiv:2202.12837, 2022.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 110, + 470, + 507, + 484 + ], + "spans": [ + { + "bbox": [ + 110, + 470, + 507, + 484 + ], + "score": 1.0, + "content": "[77] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 132, + 481, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 506, + 495 + ], + "score": 1.0, + "content": "Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 131, + 491, + 474, + 507 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 474, + 507 + ], + "score": 1.0, + "content": "reporting. In ACM Conference on Fairness, Accountability, and Transparency, 2019.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 110, + 509, + 507, + 524 + ], + "spans": [ + { + "bbox": [ + 110, + 509, + 507, + 524 + ], + "score": 1.0, + "content": "[78] Ron Mokady, Amir Hertz, and Amit H. Bermano. ClipCap: CLIP prefix for image captioning.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 132, + 521, + 237, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 237, + 535 + ], + "score": 1.0, + "content": "arXiv:2111.09734, 2021.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 110, + 539, + 507, + 552 + ], + "spans": [ + { + "bbox": [ + 110, + 539, + 507, + 552 + ], + "score": 1.0, + "content": "[79] Vishvak Murahari, Dhruv Batra, Devi Parikh, and Abhishek Das. Large-scale pretraining for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 131, + 549, + 507, + 564 + ], + "spans": [ + { + "bbox": [ + 131, + 549, + 507, + 564 + ], + "score": 1.0, + "content": "visual dialog: A simple state-of-the-art baseline. In European Conference on Computer Vision,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 561, + 159, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 159, + 574 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 111, + 581, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 111, + 581, + 506, + 592 + ], + "score": 1.0, + "content": "[80] Ethan Perez, Douwe Kiela, and Kyunghyun Cho. True few-shot learning with language models.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 132, + 591, + 381, + 603 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 381, + 603 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems, 2021.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 110, + 609, + 507, + 622 + ], + "spans": [ + { + "bbox": [ + 110, + 609, + 507, + 622 + ], + "score": 1.0, + "content": "[81] Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 130, + 617, + 507, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 507, + 635 + ], + "score": 1.0, + "content": "Glaese, Nat McAleese, and Geoffrey Irving. Red teaming language models with language", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 131, + 628, + 271, + 645 + ], + "spans": [ + { + "bbox": [ + 131, + 628, + 271, + 645 + ], + "score": 1.0, + "content": "models. arXiv:2202.03286, 2022.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 110, + 648, + 507, + 662 + ], + "spans": [ + { + "bbox": [ + 110, + 648, + 507, + 662 + ], + "score": 1.0, + "content": "[82] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Lu-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 130, + 657, + 507, + 675 + ], + "spans": [ + { + "bbox": [ + 130, + 657, + 507, + 675 + ], + "score": 1.0, + "content": "ong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 131, + 670, + 236, + 684 + ], + "spans": [ + { + "bbox": [ + 131, + 670, + 236, + 684 + ], + "score": 1.0, + "content": "arXiv:2111.10050, 2021.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 110, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 110, + 689, + 506, + 703 + ], + "score": 1.0, + "content": "[83] Ofir Press, Noah Smith, and Mike Lewis. Train short, test long: Attention with linear biases", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 131, + 700, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 131, + 700, + 506, + 714 + ], + "score": 1.0, + "content": "enables input length extrapolation. In International Conference on Learning Representations,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 132, + 710, + 158, + 723 + ], + "spans": [ + { + "bbox": [ + 132, + 710, + 158, + 723 + ], + "score": 1.0, + "content": "2022.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 23.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 109, + 56, + 507, + 727 + ], + "lines": [], + "index": 23.5, + "bbox_fs": [ + 110, + 72, + 507, + 723 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 112, + 72, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 110, + 71, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 110, + 71, + 506, + 87 + ], + "score": 1.0, + "content": "[84] Yixuan Qiao, Hao Chen, Jun Wang, Yihao Chen, Xianbin Ye, Ziliang Li, Xianbiao Qi, Peng", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 131, + 82, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 131, + 82, + 506, + 97 + ], + "score": 1.0, + "content": "Gao, and Guotong Xie. Winner team Mia at TextVQA Challenge 2021: Vision-and-language", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 95, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 131, + 95, + 506, + 107 + ], + "score": 1.0, + "content": "representation learning with pre-trained sequence-to-sequence model. arXiv:2106.15332,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 104, + 159, + 117 + ], + "spans": [ + { + "bbox": [ + 131, + 104, + 159, + 117 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 112, + 124, + 505, + 169 + ], + "lines": [ + { + "bbox": [ + 110, + 124, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 110, + 124, + 505, + 136 + ], + "score": 1.0, + "content": "[85] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 135, + 507, + 149 + ], + "spans": [ + { + "bbox": [ + 131, + 135, + 507, + 149 + ], + "score": 1.0, + "content": "Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 146, + 507, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 507, + 159 + ], + "score": 1.0, + "content": "and Ilya Sutskever. Learning transferable visual models from natural language supervision.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 157, + 236, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 236, + 169 + ], + "score": 1.0, + "content": "arXiv:2103.00020, 2021.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 113, + 176, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 111, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 111, + 176, + 506, + 190 + ], + "score": 1.0, + "content": "[86] Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 188, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 188, + 506, + 200 + ], + "score": 1.0, + "content": "Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George van den Driess-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 210, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 210, + 506, + 222 + ], + "score": 1.0, + "content": "che, Lisa Anne Hendricks, Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 220, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 506, + 233 + ], + "score": 1.0, + "content": "Sumanth Dathathri, Saffron Huang, Jonathan Uesato, John Mellor, Irina Higgins, Antonia", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 230, + 507, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 507, + 244 + ], + "score": 1.0, + "content": "Creswell, Nat McAleese, Amy Wu, Erich Elsen, Siddhant Jayakumar, Elena Buchatskaya,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "David Budden, Esme Sutherland, Karen Simonyan, Michela Paganini, Laurent Sifre, Lena", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 252, + 507, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 507, + 267 + ], + "score": 1.0, + "content": "Martens, Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh, Elena Gribovskaya,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "Domenic Donato, Angeliki Lazaridou, Arthur Mensch, Jean-Baptiste Lespiau, Maria Tsim-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "poukelli, Nikolai Grigorev, Doug Fritz, Thibault Sottiaux, Mantas Pajarskas, Toby Pohlen,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "Zhitao Gong, Daniel Toyama, Cyprien de Masson d’Autume, Yujia Li, Tayfun Terzi, Vladimir", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "Mikulik, Igor Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "Bradbury, Matthew Johnson, Blake Hechtman, Laura Weidinger, Iason Gabriel, William Isaac,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "Ed Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem Ayoub, Jeff", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 329, + 507, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 507, + 342 + ], + "score": 1.0, + "content": "Stanway, Lorrayne Bennett, Demis Hassabis, Koray Kavukcuoglu, and Geoffrey Irving. Scal-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "ing language models: Methods, analysis & insights from training Gopher. arXiv:2112.11446,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 351, + 159, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 159, + 363 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 111, + 370, + 503, + 404 + ], + "lines": [ + { + "bbox": [ + 110, + 369, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 110, + 369, + 506, + 384 + ], + "score": 1.0, + "content": "[87] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 393, + 334, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 393, + 334, + 403 + ], + "score": 1.0, + "content": "text-to-text transformer. arXiv:1910.10683, 2019.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 111, + 411, + 502, + 445 + ], + "lines": [ + { + "bbox": [ + 110, + 411, + 504, + 425 + ], + "spans": [ + { + "bbox": [ + 110, + 411, + 504, + 425 + ], + "score": 1.0, + "content": "[88] Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. ZeRO: Memory", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 423, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 504, + 435 + ], + "score": 1.0, + "content": "optimizations toward training trillion parameter models. In International Conference for High", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 132, + 433, + 401, + 447 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 401, + 447 + ], + "score": 1.0, + "content": "Performance Computing, Networking, Storage and Analysis, 2020.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 110, + 452, + 503, + 475 + ], + "lines": [ + { + "bbox": [ + 110, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 110, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "[89] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 464, + 442, + 476 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 442, + 476 + ], + "score": 1.0, + "content": "text-conditional image generation with clip latents. arXiv:2204.06125, 2022.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 111, + 483, + 504, + 516 + ], + "lines": [ + { + "bbox": [ + 111, + 483, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 111, + 483, + 506, + 495 + ], + "score": 1.0, + "content": "[90] Steven J. Rennie, Etienne Marcheret, Youssef Mroueh, Jarret Ross, and Vaibhava Goel.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 131, + 493, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 131, + 493, + 506, + 507 + ], + "score": 1.0, + "content": "Self-critical sequence training for image captioning. In IEEE Computer Vision and Pattern", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 132, + 505, + 211, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 211, + 518 + ], + "score": 1.0, + "content": "Recognition, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 110, + 524, + 504, + 558 + ], + "lines": [ + { + "bbox": [ + 110, + 523, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 110, + 523, + 506, + 537 + ], + "score": 1.0, + "content": "[91] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 534, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 506, + 548 + ], + "score": 1.0, + "content": "Turner. Fast and flexible multi-task classification using conditional neural adaptive processes.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 132, + 546, + 381, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 381, + 559 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems, 2019.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 111, + 565, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 110, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 110, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "[92] Laria Reynolds and Kyle McDonell. Prompt programming for large language models: Beyond", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "the few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 132, + 587, + 252, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 252, + 599 + ], + "score": 1.0, + "content": "in Computing Systems, 2021.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 109, + 606, + 504, + 629 + ], + "lines": [ + { + "bbox": [ + 111, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 111, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "[93] Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. Gender bias", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 132, + 618, + 340, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 340, + 629 + ], + "score": 1.0, + "content": "in coreference resolution. arXiv:1804.09301, 2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 111, + 636, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 110, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 110, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "[94] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 132, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 132, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 132, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 132, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "Fei. ImageNet large scale visual recognition challenge. International Journal of Computer", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 132, + 669, + 187, + 681 + ], + "spans": [ + { + "bbox": [ + 132, + 669, + 187, + 681 + ], + "score": 1.0, + "content": "Vision, 2015.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5 + }, + { + "type": "text", + "bbox": [ + 112, + 689, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 110, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 110, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "[95] Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 131, + 699, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 131, + 699, + 507, + 713 + ], + "score": 1.0, + "content": "Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, Manan Dey, M. Saiful Bari,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 132, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 132, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 112, + 72, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 110, + 71, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 110, + 71, + 506, + 87 + ], + "score": 1.0, + "content": "[84] Yixuan Qiao, Hao Chen, Jun Wang, Yihao Chen, Xianbin Ye, Ziliang Li, Xianbiao Qi, Peng", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 131, + 82, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 131, + 82, + 506, + 97 + ], + "score": 1.0, + "content": "Gao, and Guotong Xie. Winner team Mia at TextVQA Challenge 2021: Vision-and-language", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 95, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 131, + 95, + 506, + 107 + ], + "score": 1.0, + "content": "representation learning with pre-trained sequence-to-sequence model. arXiv:2106.15332,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 104, + 159, + 117 + ], + "spans": [ + { + "bbox": [ + 131, + 104, + 159, + 117 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 110, + 71, + 506, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 112, + 124, + 505, + 169 + ], + "lines": [ + { + "bbox": [ + 110, + 124, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 110, + 124, + 505, + 136 + ], + "score": 1.0, + "content": "[85] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 135, + 507, + 149 + ], + "spans": [ + { + "bbox": [ + 131, + 135, + 507, + 149 + ], + "score": 1.0, + "content": "Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 146, + 507, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 507, + 159 + ], + "score": 1.0, + "content": "and Ilya Sutskever. Learning transferable visual models from natural language supervision.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 157, + 236, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 236, + 169 + ], + "score": 1.0, + "content": "arXiv:2103.00020, 2021.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 110, + 124, + 507, + 169 + ] + }, + { + "type": "text", + "bbox": [ + 113, + 176, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 111, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 111, + 176, + 506, + 190 + ], + "score": 1.0, + "content": "[86] Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 188, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 188, + 506, + 200 + ], + "score": 1.0, + "content": "Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George van den Driess-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 210, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 210, + 506, + 222 + ], + "score": 1.0, + "content": "che, Lisa Anne Hendricks, Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 220, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 506, + 233 + ], + "score": 1.0, + "content": "Sumanth Dathathri, Saffron Huang, Jonathan Uesato, John Mellor, Irina Higgins, Antonia", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 230, + 507, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 507, + 244 + ], + "score": 1.0, + "content": "Creswell, Nat McAleese, Amy Wu, Erich Elsen, Siddhant Jayakumar, Elena Buchatskaya,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "David Budden, Esme Sutherland, Karen Simonyan, Michela Paganini, Laurent Sifre, Lena", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 252, + 507, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 507, + 267 + ], + "score": 1.0, + "content": "Martens, Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh, Elena Gribovskaya,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "Domenic Donato, Angeliki Lazaridou, Arthur Mensch, Jean-Baptiste Lespiau, Maria Tsim-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "poukelli, Nikolai Grigorev, Doug Fritz, Thibault Sottiaux, Mantas Pajarskas, Toby Pohlen,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "Zhitao Gong, Daniel Toyama, Cyprien de Masson d’Autume, Yujia Li, Tayfun Terzi, Vladimir", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "Mikulik, Igor Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "Bradbury, Matthew Johnson, Blake Hechtman, Laura Weidinger, Iason Gabriel, William Isaac,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "Ed Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem Ayoub, Jeff", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 329, + 507, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 507, + 342 + ], + "score": 1.0, + "content": "Stanway, Lorrayne Bennett, Demis Hassabis, Koray Kavukcuoglu, and Geoffrey Irving. Scal-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "ing language models: Methods, analysis & insights from training Gopher. arXiv:2112.11446,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 351, + 159, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 159, + 363 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 16, + "bbox_fs": [ + 111, + 176, + 507, + 363 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 370, + 503, + 404 + ], + "lines": [ + { + "bbox": [ + 110, + 369, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 110, + 369, + 506, + 384 + ], + "score": 1.0, + "content": "[87] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 393, + 334, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 393, + 334, + 403 + ], + "score": 1.0, + "content": "text-to-text transformer. arXiv:1910.10683, 2019.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 110, + 369, + 506, + 403 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 411, + 502, + 445 + ], + "lines": [ + { + "bbox": [ + 110, + 411, + 504, + 425 + ], + "spans": [ + { + "bbox": [ + 110, + 411, + 504, + 425 + ], + "score": 1.0, + "content": "[88] Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. ZeRO: Memory", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 423, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 504, + 435 + ], + "score": 1.0, + "content": "optimizations toward training trillion parameter models. In International Conference for High", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 132, + 433, + 401, + 447 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 401, + 447 + ], + "score": 1.0, + "content": "Performance Computing, Networking, Storage and Analysis, 2020.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 110, + 411, + 504, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 452, + 503, + 475 + ], + "lines": [ + { + "bbox": [ + 110, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 110, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "[89] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 464, + 442, + 476 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 442, + 476 + ], + "score": 1.0, + "content": "text-conditional image generation with clip latents. arXiv:2204.06125, 2022.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 110, + 452, + 505, + 476 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 483, + 504, + 516 + ], + "lines": [ + { + "bbox": [ + 111, + 483, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 111, + 483, + 506, + 495 + ], + "score": 1.0, + "content": "[90] Steven J. Rennie, Etienne Marcheret, Youssef Mroueh, Jarret Ross, and Vaibhava Goel.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 131, + 493, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 131, + 493, + 506, + 507 + ], + "score": 1.0, + "content": "Self-critical sequence training for image captioning. In IEEE Computer Vision and Pattern", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 132, + 505, + 211, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 211, + 518 + ], + "score": 1.0, + "content": "Recognition, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 111, + 483, + 506, + 518 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 524, + 504, + 558 + ], + "lines": [ + { + "bbox": [ + 110, + 523, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 110, + 523, + 506, + 537 + ], + "score": 1.0, + "content": "[91] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 534, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 506, + 548 + ], + "score": 1.0, + "content": "Turner. Fast and flexible multi-task classification using conditional neural adaptive processes.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 132, + 546, + 381, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 381, + 559 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems, 2019.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 110, + 523, + 506, + 559 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 565, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 110, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 110, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "[92] Laria Reynolds and Kyle McDonell. Prompt programming for large language models: Beyond", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "the few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 132, + 587, + 252, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 252, + 599 + ], + "score": 1.0, + "content": "in Computing Systems, 2021.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 110, + 565, + 505, + 599 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 606, + 504, + 629 + ], + "lines": [ + { + "bbox": [ + 111, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 111, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "[93] Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. Gender bias", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 132, + 618, + 340, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 340, + 629 + ], + "score": 1.0, + "content": "in coreference resolution. arXiv:1804.09301, 2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 111, + 606, + 505, + 629 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 636, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 110, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 110, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "[94] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 132, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 132, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 132, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 132, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "Fei. ImageNet large scale visual recognition challenge. International Journal of Computer", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 132, + 669, + 187, + 681 + ], + "spans": [ + { + "bbox": [ + 132, + 669, + 187, + 681 + ], + "score": 1.0, + "content": "Vision, 2015.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5, + "bbox_fs": [ + 110, + 636, + 506, + 681 + ] + }, + { + "type": "text", + "bbox": [ + 112, + 689, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 110, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 110, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "[95] Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 131, + 699, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 131, + 699, + 507, + 713 + ], + "score": 1.0, + "content": "Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, Manan Dey, M. Saiful Bari,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 132, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 132, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49, + "bbox_fs": [ + 110, + 687, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 44, + 507, + 730 + ], + "lines": [ + { + "bbox": [ + 130, + 69, + 509, + 88 + ], + "spans": [ + { + "bbox": [ + 130, + 69, + 509, + 88 + ], + "score": 1.0, + "content": "Chhablani, Nihal Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 131, + 82, + 507, + 98 + ], + "spans": [ + { + "bbox": [ + 131, + 82, + 507, + 98 + ], + "score": 1.0, + "content": "Matteo Manica, Sheng Shen, Zheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 95, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 132, + 95, + 506, + 107 + ], + "score": 1.0, + "content": "Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Fevry, Jason Alan", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 104, + 507, + 119 + ], + "spans": [ + { + "bbox": [ + 130, + 104, + 507, + 119 + ], + "score": 1.0, + "content": "Fries, Ryan Teehan, Stella Biderman, Leo Gao, Tali Bers, Thomas Wolf, and Alexander M.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 115, + 507, + 130 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 507, + 130 + ], + "score": 1.0, + "content": "Rush. Multitask Prompted Training Enables Zero-Shot Task Generalization. In International", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 126, + 327, + 141 + ], + "spans": [ + { + "bbox": [ + 131, + 126, + 327, + 141 + ], + "score": 1.0, + "content": "Conference on Learning Representations, 2022.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 146, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 110, + 146, + 505, + 158 + ], + "score": 1.0, + "content": "[96] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 155, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 131, + 155, + 506, + 171 + ], + "score": 1.0, + "content": "Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 167, + 445, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 445, + 182 + ], + "score": 1.0, + "content": "dataset of clip-filtered 400 million image-text pairs. arXiv:2111.02114, 2021.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 186, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 110, + 186, + 506, + 200 + ], + "score": 1.0, + "content": "[97] Carsten Schwemmer, Carly Knight, Emily D. Bello-Pardo, Stan Oklobdzija, Martijn", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 198, + 507, + 212 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 507, + 212 + ], + "score": 1.0, + "content": "Schoonvelde, and Jeffrey W. Lockhart. Diagnosing gender bias in image recognition systems.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 209, + 190, + 221 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 190, + 221 + ], + "score": 1.0, + "content": "Socius, 2020.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 109, + 227, + 507, + 240 + ], + "spans": [ + { + "bbox": [ + 109, + 227, + 507, + 240 + ], + "score": 1.0, + "content": "[98] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual Captions:", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 238, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 506, + 252 + ], + "score": 1.0, + "content": "A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Annual", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 249, + 392, + 263 + ], + "spans": [ + { + "bbox": [ + 131, + 249, + 392, + 263 + ], + "score": 1.0, + "content": "Meeting of the Association for Computational Linguistics, 2018.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 110, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 110, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "[99] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 133, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "Catanzaro. Megatron-LM: Training multi-billion parameter language models using model", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 289, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 287, + 303 + ], + "score": 1.0, + "content": "parallelism. arXiv:2104.08691, 2019.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "[100] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 318, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 131, + 318, + 506, + 333 + ], + "score": 1.0, + "content": "Parikh, and Marcus Rohrbach. Towards VQA models that can read. In IEEE Computer Vision", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 330, + 261, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 261, + 344 + ], + "score": 1.0, + "content": "and Pattern Recognition, 2019.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 349, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 506, + 361 + ], + "score": 1.0, + "content": "[101] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 131, + 358, + 507, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 507, + 374 + ], + "score": 1.0, + "content": "Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 371, + 267, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 267, + 383 + ], + "score": 1.0, + "content": "model. arXiv:2112.04482, 2021.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 389, + 507, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 507, + 403 + ], + "score": 1.0, + "content": "[102] Lucas Smaira, João Carreira, Eric Noland, Ellen Clancy, Amy Wu, and Andrew Zisserman. A", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 131, + 400, + 473, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 400, + 473, + 414 + ], + "score": 1.0, + "content": "short note on the Kinetics-700-2020 human action dataset. arXiv:2010.10864, 2020.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 417, + 507, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 417, + 507, + 434 + ], + "score": 1.0, + "content": "[103] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 133, + 431, + 381, + 443 + ], + "spans": [ + { + "bbox": [ + 133, + 431, + 381, + 443 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems, 2017.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 448, + 507, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 507, + 462 + ], + "score": 1.0, + "content": "[104] David R So, Wojciech Manke, Hanxiao Liu, Zihang Dai, Noam Shazeer, and Quoc V. Le. ´", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 459, + 507, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 507, + 474 + ], + "score": 1.0, + "content": "Primer: Searching for efficient transformers for language modeling. arXiv:2109.08668, 2021.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "[105] Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 489, + 340, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 340, + 501 + ], + "score": 1.0, + "content": "for deep learning in NLP. arXiv:1906.02243, 2019.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 507, + 507, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 507, + 521 + ], + "score": 1.0, + "content": "[106] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT:", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 131, + 518, + 461, + 532 + ], + "spans": [ + { + "bbox": [ + 131, + 518, + 461, + 532 + ], + "score": 1.0, + "content": "Pre-training of generic visual-linguistic representations. arXiv:1908.08530, 2019.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 538, + 507, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 507, + 550 + ], + "score": 1.0, + "content": "[107] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. VideoBERT:", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 131, + 547, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 505, + 562 + ], + "score": 1.0, + "content": "A joint model for video and language representation learning. In International Conference on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 559, + 231, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 231, + 574 + ], + "score": 1.0, + "content": "Computer Vision, 2019.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 107, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "[108] Ilya Sutskever, James Martens, and Geoffrey E. Hinton. Generating text with recurrent neural", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 132, + 590, + 406, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 406, + 601 + ], + "score": 1.0, + "content": "networks. In International Conference on Machine Learning, 2011.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "score": 1.0, + "content": "[109] Hao Tan and Mohit Bansal. LXMERT: Learning cross-modality encoder representations from", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 620, + 494, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 494, + 632 + ], + "score": 1.0, + "content": "transformer. In Conference on Empirical Methods in Natural Language Processing, 2019.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "[110] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 133, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 133, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "Poland, Damian Borth, and Li-Jia Li. YFCC100M: The new data in multimedia research.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 132, + 659, + 279, + 672 + ], + "spans": [ + { + "bbox": [ + 132, + 659, + 279, + 672 + ], + "score": 1.0, + "content": "Communications of the ACM, 2016.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 677, + 507, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 507, + 692 + ], + "score": 1.0, + "content": "[111] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 130, + 687, + 507, + 704 + ], + "spans": [ + { + "bbox": [ + 130, + 687, + 507, + 704 + ], + "score": 1.0, + "content": "Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 131, + 699, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 131, + 699, + 507, + 714 + ], + "score": 1.0, + "content": "Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 131, + 709, + 507, + 725 + ], + "spans": [ + { + "bbox": [ + 131, + 709, + 507, + 725 + ], + "score": 1.0, + "content": "Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts,", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 23.5 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 44, + 507, + 730 + ], + "lines": [ + { + "bbox": [ + 130, + 69, + 509, + 88 + ], + "spans": [ + { + "bbox": [ + 130, + 69, + 509, + 88 + ], + "score": 1.0, + "content": "Chhablani, Nihal Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 131, + 82, + 507, + 98 + ], + "spans": [ + { + "bbox": [ + 131, + 82, + 507, + 98 + ], + "score": 1.0, + "content": "Matteo Manica, Sheng Shen, Zheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 95, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 132, + 95, + 506, + 107 + ], + "score": 1.0, + "content": "Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Fevry, Jason Alan", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 104, + 507, + 119 + ], + "spans": [ + { + "bbox": [ + 130, + 104, + 507, + 119 + ], + "score": 1.0, + "content": "Fries, Ryan Teehan, Stella Biderman, Leo Gao, Tali Bers, Thomas Wolf, and Alexander M.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 115, + 507, + 130 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 507, + 130 + ], + "score": 1.0, + "content": "Rush. Multitask Prompted Training Enables Zero-Shot Task Generalization. In International", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 126, + 327, + 141 + ], + "spans": [ + { + "bbox": [ + 131, + 126, + 327, + 141 + ], + "score": 1.0, + "content": "Conference on Learning Representations, 2022.", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 146, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 110, + 146, + 505, + 158 + ], + "score": 1.0, + "content": "[96] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 155, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 131, + 155, + 506, + 171 + ], + "score": 1.0, + "content": "Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 167, + 445, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 445, + 182 + ], + "score": 1.0, + "content": "dataset of clip-filtered 400 million image-text pairs. arXiv:2111.02114, 2021.", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 186, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 110, + 186, + 506, + 200 + ], + "score": 1.0, + "content": "[97] Carsten Schwemmer, Carly Knight, Emily D. Bello-Pardo, Stan Oklobdzija, Martijn", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 198, + 507, + 212 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 507, + 212 + ], + "score": 1.0, + "content": "Schoonvelde, and Jeffrey W. Lockhart. Diagnosing gender bias in image recognition systems.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 209, + 190, + 221 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 190, + 221 + ], + "score": 1.0, + "content": "Socius, 2020.", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 227, + 507, + 240 + ], + "spans": [ + { + "bbox": [ + 109, + 227, + 507, + 240 + ], + "score": 1.0, + "content": "[98] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual Captions:", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 238, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 506, + 252 + ], + "score": 1.0, + "content": "A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Annual", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 249, + 392, + 263 + ], + "spans": [ + { + "bbox": [ + 131, + 249, + 392, + 263 + ], + "score": 1.0, + "content": "Meeting of the Association for Computational Linguistics, 2018.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 110, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "[99] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 133, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 133, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "Catanzaro. Megatron-LM: Training multi-billion parameter language models using model", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 289, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 287, + 303 + ], + "score": 1.0, + "content": "parallelism. arXiv:2104.08691, 2019.", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "[100] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 318, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 131, + 318, + 506, + 333 + ], + "score": 1.0, + "content": "Parikh, and Marcus Rohrbach. Towards VQA models that can read. In IEEE Computer Vision", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 330, + 261, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 261, + 344 + ], + "score": 1.0, + "content": "and Pattern Recognition, 2019.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 349, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 506, + 361 + ], + "score": 1.0, + "content": "[101] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba,", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 358, + 507, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 507, + 374 + ], + "score": 1.0, + "content": "Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 371, + 267, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 267, + 383 + ], + "score": 1.0, + "content": "model. arXiv:2112.04482, 2021.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 389, + 507, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 507, + 403 + ], + "score": 1.0, + "content": "[102] Lucas Smaira, João Carreira, Eric Noland, Ellen Clancy, Amy Wu, and Andrew Zisserman. A", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 400, + 473, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 400, + 473, + 414 + ], + "score": 1.0, + "content": "short note on the Kinetics-700-2020 human action dataset. arXiv:2010.10864, 2020.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 417, + 507, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 417, + 507, + 434 + ], + "score": 1.0, + "content": "[103] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning.", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 133, + 431, + 381, + 443 + ], + "spans": [ + { + "bbox": [ + 133, + 431, + 381, + 443 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems, 2017.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 448, + 507, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 507, + 462 + ], + "score": 1.0, + "content": "[104] David R So, Wojciech Manke, Hanxiao Liu, Zihang Dai, Noam Shazeer, and Quoc V. Le. ´", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 459, + 507, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 507, + 474 + ], + "score": 1.0, + "content": "Primer: Searching for efficient transformers for language modeling. arXiv:2109.08668, 2021.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "[105] Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 489, + 340, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 340, + 501 + ], + "score": 1.0, + "content": "for deep learning in NLP. arXiv:1906.02243, 2019.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 507, + 507, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 507, + 521 + ], + "score": 1.0, + "content": "[106] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT:", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 518, + 461, + 532 + ], + "spans": [ + { + "bbox": [ + 131, + 518, + 461, + 532 + ], + "score": 1.0, + "content": "Pre-training of generic visual-linguistic representations. arXiv:1908.08530, 2019.", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 538, + 507, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 507, + 550 + ], + "score": 1.0, + "content": "[107] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. VideoBERT:", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 547, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 505, + 562 + ], + "score": 1.0, + "content": "A joint model for video and language representation learning. In International Conference on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 559, + 231, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 231, + 574 + ], + "score": 1.0, + "content": "Computer Vision, 2019.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 107, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "[108] Ilya Sutskever, James Martens, and Geoffrey E. Hinton. Generating text with recurrent neural", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 590, + 406, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 406, + 601 + ], + "score": 1.0, + "content": "networks. In International Conference on Machine Learning, 2011.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "score": 1.0, + "content": "[109] Hao Tan and Mohit Bansal. LXMERT: Learning cross-modality encoder representations from", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 620, + 494, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 494, + 632 + ], + "score": 1.0, + "content": "transformer. In Conference on Empirical Methods in Natural Language Processing, 2019.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "[110] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 133, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 133, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "Poland, Damian Borth, and Li-Jia Li. YFCC100M: The new data in multimedia research.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 132, + 659, + 279, + 672 + ], + "spans": [ + { + "bbox": [ + 132, + 659, + 279, + 672 + ], + "score": 1.0, + "content": "Communications of the ACM, 2016.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 677, + 507, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 507, + 692 + ], + "score": 1.0, + "content": "[111] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha,", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 130, + 687, + 507, + 704 + ], + "spans": [ + { + "bbox": [ + 130, + 687, + 507, + 704 + ], + "score": 1.0, + "content": "Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 131, + 699, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 131, + 699, + 507, + 714 + ], + "score": 1.0, + "content": "Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 131, + 709, + 507, + 725 + ], + "spans": [ + { + "bbox": [ + 131, + 709, + 507, + 725 + ], + "score": 1.0, + "content": "Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts,", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 69, + 509, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 133, + 72, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 132, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 132, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 131, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 131, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 93, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 131, + 93, + 506, + 108 + ], + "score": 1.0, + "content": "Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 104, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 132, + 104, + 506, + 119 + ], + "score": 1.0, + "content": "Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 126, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 132, + 126, + 506, + 141 + ], + "score": 1.0, + "content": "Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 137, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 505, + 151 + ], + "score": 1.0, + "content": "Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. LaMDA: Language models for dialog", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 149, + 290, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 290, + 162 + ], + "score": 1.0, + "content": "applications. arXiv:2201.08239, 2022.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 170, + 504, + 204 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 505, + 184 + ], + "score": 1.0, + "content": "[112] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B. Tenenbaum, and Phillip Isola. Rethinking", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "few-shot image classification: a good embedding is all you need? In European Conference on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 192, + 229, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 229, + 205 + ], + "score": 1.0, + "content": "Computer Vision, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 505, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 212, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 506, + 227 + ], + "score": 1.0, + "content": "[113] Hugo Touvron, Andrea Vedaldi, Matthijs Douze, and Hervé Jégou. Fixing the train-test", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 225, + 477, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 477, + 237 + ], + "score": 1.0, + "content": "resolution discrepancy. Conference on Neural Information Processing Systems, 2019.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 505, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "[114] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 256, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 505, + 270 + ], + "score": 1.0, + "content": "Multimodal few-shot learning with frozen language models. Conference on Neural Information", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 268, + 241, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 241, + 281 + ], + "score": 1.0, + "content": "Processing Systems, 2021.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 505, + 322 + ], + "lines": [ + { + "bbox": [ + 104, + 288, + 507, + 303 + ], + "spans": [ + { + "bbox": [ + 104, + 288, + 507, + 303 + ], + "score": 1.0, + "content": "[115] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 299, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 505, + 313 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Conference on Neural", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 311, + 291, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 291, + 323 + ], + "score": 1.0, + "content": "Information Processing Systems, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 105, + 332, + 505, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "score": 1.0, + "content": "[116] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 344, + 459, + 356 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 459, + 356 + ], + "score": 1.0, + "content": "image caption generator. In International Conference on Computer Vision, 2015.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 364, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 507, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 507, + 378 + ], + "score": 1.0, + "content": "[117] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 131, + 374, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 374, + 505, + 389 + ], + "score": 1.0, + "content": "Matching networks for one shot learning. Conference on Neural Information Processing", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 386, + 194, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 194, + 399 + ], + "score": 1.0, + "content": "Systems, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 407, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 407, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 421 + ], + "score": 1.0, + "content": "[118] Jianfeng Wang, Xiaowei Hu, Zhe Gan, Zhengyuan Yang, Xiyang Dai, Zicheng Liu, Yumao", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 131, + 418, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 506, + 433 + ], + "score": 1.0, + "content": "Lu, and Lijuan Wang. UFO: A unified transformer for vision-language representation learning.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 430, + 236, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 236, + 441 + ], + "score": 1.0, + "content": "arXiv:2111.10023, 2021.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "score": 1.0, + "content": "[119] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 132, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "Zhou, Jingren Zhou, and Hongxia Yang. Unifying architectures, tasks, and modalities through", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 474, + 446, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 446, + 485 + ], + "score": 1.0, + "content": "a simple sequence-to-sequence learning framework. arXiv:2202.03052, 2022.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 494, + 506, + 528 + ], + "lines": [ + { + "bbox": [ + 104, + 492, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 104, + 492, + 506, + 509 + ], + "score": 1.0, + "content": "[120] Thomas Wang, Adam Roberts, Daniel Hesslow, Teven Le Scao, Hyung Won Chung, Iz Beltagy,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 131, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 131, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "Julien Launay, and Colin Raffel. What language model architecture and pretraining objective", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 516, + 395, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 395, + 528 + ], + "score": 1.0, + "content": "work best for zero-shot generalization? arXiv:2204.05832, 2022.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 104, + 537, + 504, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 535, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 506, + 552 + ], + "score": 1.0, + "content": "[121] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. VLMo: Unified vision-language", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 549, + 425, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 425, + 561 + ], + "score": 1.0, + "content": "pre-training with mixture-of-modality-experts. arXiv:2111.02358, 2021.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 105, + 569, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 507, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 507, + 584 + ], + "score": 1.0, + "content": "[122] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "VATEX: A large-scale, high-quality multilingual dataset for video-and-language research. In", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 133, + 592, + 345, + 604 + ], + "spans": [ + { + "bbox": [ + 133, + 592, + 345, + 604 + ], + "score": 1.0, + "content": "International Conference on Computer Vision, 2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 613, + 506, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 613, + 507, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 507, + 626 + ], + "score": 1.0, + "content": "[123] Yue Wang, Shafiq Joty, Michael Lyu, Irwin King, Caiming Xiong, and Steven Hoi. VD-BERT:", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 131, + 623, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 506, + 637 + ], + "score": 1.0, + "content": "A unified vision and dialog transformer with BERT. In Conference on Empirical Methods in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 635, + 282, + 648 + ], + "spans": [ + { + "bbox": [ + 132, + 635, + 282, + 648 + ], + "score": 1.0, + "content": "Natural Language Processing, 2020.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 103, + 656, + 505, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "[124] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM:", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 132, + 667, + 502, + 680 + ], + "spans": [ + { + "bbox": [ + 132, + 667, + 502, + 680 + ], + "score": 1.0, + "content": "Simple visual language model pretraining with weak supervision. arXiv:2108.10904, 2021.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 701 + ], + "score": 1.0, + "content": "[125] Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 132, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 132, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "Du, Andrew M. Dai, and Quoc V. Le. Finetuned language models are zero-shot learners.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 132, + 711, + 236, + 721 + ], + "spans": [ + { + "bbox": [ + 132, + 711, + 236, + 721 + ], + "score": 1.0, + "content": "arXiv:2109.01652, 2021.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 133, + 72, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 132, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 132, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 131, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 131, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 93, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 131, + 93, + 506, + 108 + ], + "score": 1.0, + "content": "Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 104, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 132, + 104, + 506, + 119 + ], + "score": 1.0, + "content": "Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 126, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 132, + 126, + 506, + 141 + ], + "score": 1.0, + "content": "Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 137, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 505, + 151 + ], + "score": 1.0, + "content": "Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. LaMDA: Language models for dialog", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 149, + 290, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 290, + 162 + ], + "score": 1.0, + "content": "applications. arXiv:2201.08239, 2022.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5, + "bbox_fs": [ + 131, + 72, + 506, + 162 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 170, + 504, + 204 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 505, + 184 + ], + "score": 1.0, + "content": "[112] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B. Tenenbaum, and Phillip Isola. Rethinking", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "few-shot image classification: a good embedding is all you need? In European Conference on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 192, + 229, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 229, + 205 + ], + "score": 1.0, + "content": "Computer Vision, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 169, + 505, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 505, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 212, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 506, + 227 + ], + "score": 1.0, + "content": "[113] Hugo Touvron, Andrea Vedaldi, Matthijs Douze, and Hervé Jégou. Fixing the train-test", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 225, + 477, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 477, + 237 + ], + "score": 1.0, + "content": "resolution discrepancy. Conference on Neural Information Processing Systems, 2019.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 212, + 506, + 237 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 505, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "[114] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 256, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 505, + 270 + ], + "score": 1.0, + "content": "Multimodal few-shot learning with frozen language models. Conference on Neural Information", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 268, + 241, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 241, + 281 + ], + "score": 1.0, + "content": "Processing Systems, 2021.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 245, + 506, + 281 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 505, + 322 + ], + "lines": [ + { + "bbox": [ + 104, + 288, + 507, + 303 + ], + "spans": [ + { + "bbox": [ + 104, + 288, + 507, + 303 + ], + "score": 1.0, + "content": "[115] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 299, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 505, + 313 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Conference on Neural", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 311, + 291, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 291, + 323 + ], + "score": 1.0, + "content": "Information Processing Systems, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 288, + 507, + 323 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 332, + 505, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "score": 1.0, + "content": "[116] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 344, + 459, + 356 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 459, + 356 + ], + "score": 1.0, + "content": "image caption generator. In International Conference on Computer Vision, 2015.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 331, + 506, + 356 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 364, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 507, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 507, + 378 + ], + "score": 1.0, + "content": "[117] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 131, + 374, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 374, + 505, + 389 + ], + "score": 1.0, + "content": "Matching networks for one shot learning. Conference on Neural Information Processing", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 386, + 194, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 194, + 399 + ], + "score": 1.0, + "content": "Systems, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 363, + 507, + 399 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 407, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 407, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 421 + ], + "score": 1.0, + "content": "[118] Jianfeng Wang, Xiaowei Hu, Zhe Gan, Zhengyuan Yang, Xiyang Dai, Zicheng Liu, Yumao", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 131, + 418, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 506, + 433 + ], + "score": 1.0, + "content": "Lu, and Lijuan Wang. UFO: A unified transformer for vision-language representation learning.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 430, + 236, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 236, + 441 + ], + "score": 1.0, + "content": "arXiv:2111.10023, 2021.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 407, + 506, + 441 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "score": 1.0, + "content": "[119] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 132, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "Zhou, Jingren Zhou, and Hongxia Yang. Unifying architectures, tasks, and modalities through", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 474, + 446, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 446, + 485 + ], + "score": 1.0, + "content": "a simple sequence-to-sequence learning framework. arXiv:2202.03052, 2022.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 450, + 506, + 485 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 494, + 506, + 528 + ], + "lines": [ + { + "bbox": [ + 104, + 492, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 104, + 492, + 506, + 509 + ], + "score": 1.0, + "content": "[120] Thomas Wang, Adam Roberts, Daniel Hesslow, Teven Le Scao, Hyung Won Chung, Iz Beltagy,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 131, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 131, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "Julien Launay, and Colin Raffel. What language model architecture and pretraining objective", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 516, + 395, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 395, + 528 + ], + "score": 1.0, + "content": "work best for zero-shot generalization? arXiv:2204.05832, 2022.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 492, + 506, + 528 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 537, + 504, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 535, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 506, + 552 + ], + "score": 1.0, + "content": "[121] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. VLMo: Unified vision-language", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 549, + 425, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 425, + 561 + ], + "score": 1.0, + "content": "pre-training with mixture-of-modality-experts. arXiv:2111.02358, 2021.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 535, + 506, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 569, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 507, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 507, + 584 + ], + "score": 1.0, + "content": "[122] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "VATEX: A large-scale, high-quality multilingual dataset for video-and-language research. In", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 133, + 592, + 345, + 604 + ], + "spans": [ + { + "bbox": [ + 133, + 592, + 345, + 604 + ], + "score": 1.0, + "content": "International Conference on Computer Vision, 2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 568, + 507, + 604 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 613, + 506, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 613, + 507, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 507, + 626 + ], + "score": 1.0, + "content": "[123] Yue Wang, Shafiq Joty, Michael Lyu, Irwin King, Caiming Xiong, and Steven Hoi. VD-BERT:", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 131, + 623, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 506, + 637 + ], + "score": 1.0, + "content": "A unified vision and dialog transformer with BERT. In Conference on Empirical Methods in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 635, + 282, + 648 + ], + "spans": [ + { + "bbox": [ + 132, + 635, + 282, + 648 + ], + "score": 1.0, + "content": "Natural Language Processing, 2020.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 613, + 507, + 648 + ] + }, + { + "type": "text", + "bbox": [ + 103, + 656, + 505, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "[124] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM:", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 132, + 667, + 502, + 680 + ], + "spans": [ + { + "bbox": [ + 132, + 667, + 502, + 680 + ], + "score": 1.0, + "content": "Simple visual language model pretraining with weak supervision. arXiv:2108.10904, 2021.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 656, + 506, + 680 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 701 + ], + "score": 1.0, + "content": "[125] Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 132, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 132, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "Du, Andrew M. Dai, and Quoc V. Le. Finetuned language models are zero-shot learners.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 132, + 711, + 236, + 721 + ], + "spans": [ + { + "bbox": [ + 132, + 711, + 236, + 721 + ], + "score": 1.0, + "content": "arXiv:2109.01652, 2021.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 688, + 506, + 721 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 506, + 128 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 507, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 507, + 87 + ], + "score": 1.0, + "content": "[126] Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 84, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 132, + 84, + 505, + 96 + ], + "score": 1.0, + "content": "Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, Zac Kenton, Sasha Brown, Will", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 132, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "Hawkins, Tom Stepleton, Courtney Biles, Abeba Birhane, Julia Haas, Laura Rimell, Lisa Anne", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 133, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "Hendricks, William Isaac, Sean Legassick, Geoffrey Irving, and Iason Gabriel. Ethical and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 116, + 412, + 128 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 412, + 128 + ], + "score": 1.0, + "content": "social risks of harm from language models. arXiv:2112.04359, 2021.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 134, + 504, + 179 + ], + "lines": [ + { + "bbox": [ + 105, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "[127] Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael Gontijo-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 156, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 506, + 169 + ], + "score": 1.0, + "content": "Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 168, + 414, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 414, + 180 + ], + "score": 1.0, + "content": "accuracy without increasing inference time. arXiv:2203.05482, 2022.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 185, + 504, + 219 + ], + "lines": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "score": 1.0, + "content": "[128] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B. Tenenbaum, and Chuang Gan. STAR: A", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "Benchmark for Situated Reasoning in Real-World Videos. In Conference on Neural Information", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 208, + 241, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 241, + 220 + ], + "score": 1.0, + "content": "Processing Systems, 2021.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 225, + 505, + 259 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 507, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 507, + 239 + ], + "score": 1.0, + "content": "[129] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-QA: Next phase of question-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 236, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 506, + 249 + ], + "score": 1.0, + "content": "answering to explaining temporal actions. In IEEE Computer Vision and Pattern Recognition,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 247, + 159, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 159, + 259 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 265, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 507, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 507, + 280 + ], + "score": 1.0, + "content": "[130] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 276, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 276, + 506, + 290 + ], + "score": 1.0, + "content": "Video question answering via gradually refined attention over appearance and motion. In ACM", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 288, + 209, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 288, + 209, + 299 + ], + "score": 1.0, + "content": "Multimedia, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 504, + 339 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "[131] Hanwei Xu, Yujun Chen, Yulun Du, Nan Shao, Yanggang Wang, Haiyu Li, and Zhilin", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 316, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 505, + 330 + ], + "score": 1.0, + "content": "Yang. Zeroprompt: Scaling prompt-based pretraining to 1,000 tasks improves zero-shot", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 328, + 298, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 328, + 298, + 339 + ], + "score": 1.0, + "content": "generalization. arXiv:2201.06910, 2022.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "[132] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feicht-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "enhofer, Florian Metze, and Luke Zettlemoyer. VLM: Task-agnostic video-language model", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 368, + 387, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 387, + 380 + ], + "score": 1.0, + "content": "pre-training for video understanding. arXiv:2105.09996, 2021.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 506, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 507, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 507, + 399 + ], + "score": 1.0, + "content": "[133] Ming Yan, Haiyang Xu, Chenliang Li, Junfeng Tian, Bin Bi, Wei Wang, Weihua Chen,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 131, + 395, + 507, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 395, + 507, + 411 + ], + "score": 1.0, + "content": "Xianzhe Xu, Fan Wang, Zheng Cao, Zhicheng Zhang, Qiyu Zhang, Ji Zhang, Songfang Huang,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 131, + 407, + 507, + 421 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 507, + 421 + ], + "score": 1.0, + "content": "Fei Huang, Luo Si, and Rong Jin. Achieving human parity on visual question answering.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 419, + 236, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 236, + 431 + ], + "score": 1.0, + "content": "arXiv:2111.08896, 2021.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 105, + 436, + 504, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "[134] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 448, + 459, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 459, + 461 + ], + "score": 1.0, + "content": "Schmid. Multiview transformers for video recognition. arXiv:2201.04288, 2022.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 504, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 507, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 507, + 478 + ], + "score": 1.0, + "content": "[135] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "Learning to answer questions from millions of narrated videos. In International Conference", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 488, + 243, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 488, + 243, + 500 + ], + "score": 1.0, + "content": "on Computer Vision, 2021.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 105, + 506, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "score": 1.0, + "content": "[136] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 517, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 506, + 531 + ], + "score": 1.0, + "content": "Wang. An empirical study of GPT-3 for few-shot knowledge-based VQA. In National", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 133, + 529, + 339, + 541 + ], + "spans": [ + { + "bbox": [ + 133, + 529, + 339, + 541 + ], + "score": 1.0, + "content": "Conference on Artificial Intelligence (AAAI), 2021.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "score": 1.0, + "content": "[137] Zhengyuan Yang, Yijuan Lu, Jianfeng Wang, Xi Yin, Dinei Florencio, Lijuan Wang, Cha", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "Zhang, Lei Zhang, and Jiebo Luo. TAP: Text-aware pre-training for text-VQA and text-caption.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 132, + 568, + 365, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 365, + 582 + ], + "score": 1.0, + "content": "In IEEE Computer Vision and Pattern Recognition, 2021.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 105, + 586, + 504, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "score": 1.0, + "content": "[138] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 131, + 596, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 596, + 505, + 611 + ], + "score": 1.0, + "content": "Zhenguo Li, Xin Jiang, and Chunjing Xu. FILIP: Fine-grained interactive language-image", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 132, + 610, + 289, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 289, + 621 + ], + "score": 1.0, + "content": "pre-training. arXiv:2111.07783, 2021.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 105, + 626, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "[139] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 132, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 132, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "to visual denotations: New similarity metrics for semantic inference over event descriptions.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 132, + 649, + 423, + 661 + ], + "spans": [ + { + "bbox": [ + 132, + 649, + 423, + 661 + ], + "score": 1.0, + "content": "Annual Meeting of the Association for Computational Linguistics, 2014.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 681 + ], + "score": 1.0, + "content": "[140] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 132, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 132, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 131, + 686, + 507, + 704 + ], + "spans": [ + { + "bbox": [ + 131, + 686, + 507, + 704 + ], + "score": 1.0, + "content": "Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 131, + 698, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 131, + 698, + 507, + 714 + ], + "score": 1.0, + "content": "Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 132, + 710, + 235, + 722 + ], + "spans": [ + { + "bbox": [ + 132, + 710, + 235, + 722 + ], + "score": 1.0, + "content": "arXiv:2111.11432, 2021.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "19", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 506, + 128 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 507, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 507, + 87 + ], + "score": 1.0, + "content": "[126] Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 84, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 132, + 84, + 505, + 96 + ], + "score": 1.0, + "content": "Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, Zac Kenton, Sasha Brown, Will", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 132, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "Hawkins, Tom Stepleton, Courtney Biles, Abeba Birhane, Julia Haas, Laura Rimell, Lisa Anne", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 133, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "Hendricks, William Isaac, Sean Legassick, Geoffrey Irving, and Iason Gabriel. Ethical and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 116, + 412, + 128 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 412, + 128 + ], + "score": 1.0, + "content": "social risks of harm from language models. arXiv:2112.04359, 2021.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 72, + 507, + 128 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 134, + 504, + 179 + ], + "lines": [ + { + "bbox": [ + 105, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "[127] Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael Gontijo-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 156, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 506, + 169 + ], + "score": 1.0, + "content": "Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 168, + 414, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 414, + 180 + ], + "score": 1.0, + "content": "accuracy without increasing inference time. arXiv:2203.05482, 2022.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 134, + 506, + 180 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 185, + 504, + 219 + ], + "lines": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "score": 1.0, + "content": "[128] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B. Tenenbaum, and Chuang Gan. STAR: A", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "Benchmark for Situated Reasoning in Real-World Videos. In Conference on Neural Information", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 208, + 241, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 241, + 220 + ], + "score": 1.0, + "content": "Processing Systems, 2021.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 185, + 506, + 220 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 225, + 505, + 259 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 507, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 507, + 239 + ], + "score": 1.0, + "content": "[129] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-QA: Next phase of question-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 236, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 506, + 249 + ], + "score": 1.0, + "content": "answering to explaining temporal actions. In IEEE Computer Vision and Pattern Recognition,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 247, + 159, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 159, + 259 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 225, + 507, + 259 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 265, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 507, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 507, + 280 + ], + "score": 1.0, + "content": "[130] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 276, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 276, + 506, + 290 + ], + "score": 1.0, + "content": "Video question answering via gradually refined attention over appearance and motion. In ACM", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 288, + 209, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 288, + 209, + 299 + ], + "score": 1.0, + "content": "Multimedia, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 264, + 507, + 299 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 504, + 339 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "[131] Hanwei Xu, Yujun Chen, Yulun Du, Nan Shao, Yanggang Wang, Haiyu Li, and Zhilin", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 316, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 505, + 330 + ], + "score": 1.0, + "content": "Yang. Zeroprompt: Scaling prompt-based pretraining to 1,000 tasks improves zero-shot", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 328, + 298, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 328, + 298, + 339 + ], + "score": 1.0, + "content": "generalization. arXiv:2201.06910, 2022.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 106, + 306, + 505, + 339 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "[132] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feicht-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "enhofer, Florian Metze, and Luke Zettlemoyer. VLM: Task-agnostic video-language model", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 368, + 387, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 387, + 380 + ], + "score": 1.0, + "content": "pre-training for video understanding. arXiv:2105.09996, 2021.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 345, + 506, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 506, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 507, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 507, + 399 + ], + "score": 1.0, + "content": "[133] Ming Yan, Haiyang Xu, Chenliang Li, Junfeng Tian, Bin Bi, Wei Wang, Weihua Chen,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 131, + 395, + 507, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 395, + 507, + 411 + ], + "score": 1.0, + "content": "Xianzhe Xu, Fan Wang, Zheng Cao, Zhicheng Zhang, Qiyu Zhang, Ji Zhang, Songfang Huang,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 131, + 407, + 507, + 421 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 507, + 421 + ], + "score": 1.0, + "content": "Fei Huang, Luo Si, and Rong Jin. Achieving human parity on visual question answering.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 419, + 236, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 236, + 431 + ], + "score": 1.0, + "content": "arXiv:2111.08896, 2021.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 385, + 507, + 431 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 436, + 504, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "[134] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 448, + 459, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 459, + 461 + ], + "score": 1.0, + "content": "Schmid. Multiview transformers for video recognition. arXiv:2201.04288, 2022.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 437, + 505, + 461 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 504, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 507, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 507, + 478 + ], + "score": 1.0, + "content": "[135] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "Learning to answer questions from millions of narrated videos. In International Conference", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 488, + 243, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 488, + 243, + 500 + ], + "score": 1.0, + "content": "on Computer Vision, 2021.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 466, + 507, + 500 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 506, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "score": 1.0, + "content": "[136] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 517, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 506, + 531 + ], + "score": 1.0, + "content": "Wang. An empirical study of GPT-3 for few-shot knowledge-based VQA. In National", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 133, + 529, + 339, + 541 + ], + "spans": [ + { + "bbox": [ + 133, + 529, + 339, + 541 + ], + "score": 1.0, + "content": "Conference on Artificial Intelligence (AAAI), 2021.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 505, + 506, + 541 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "score": 1.0, + "content": "[137] Zhengyuan Yang, Yijuan Lu, Jianfeng Wang, Xi Yin, Dinei Florencio, Lijuan Wang, Cha", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "Zhang, Lei Zhang, and Jiebo Luo. TAP: Text-aware pre-training for text-VQA and text-caption.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 132, + 568, + 365, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 365, + 582 + ], + "score": 1.0, + "content": "In IEEE Computer Vision and Pattern Recognition, 2021.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 545, + 506, + 582 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 586, + 504, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "score": 1.0, + "content": "[138] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 131, + 596, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 596, + 505, + 611 + ], + "score": 1.0, + "content": "Zhenguo Li, Xin Jiang, and Chunjing Xu. FILIP: Fine-grained interactive language-image", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 132, + 610, + 289, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 289, + 621 + ], + "score": 1.0, + "content": "pre-training. arXiv:2111.07783, 2021.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 586, + 506, + 621 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 626, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "[139] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 132, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 132, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "to visual denotations: New similarity metrics for semantic inference over event descriptions.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 132, + 649, + 423, + 661 + ], + "spans": [ + { + "bbox": [ + 132, + 649, + 423, + 661 + ], + "score": 1.0, + "content": "Annual Meeting of the Association for Computational Linguistics, 2014.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 626, + 505, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 681 + ], + "score": 1.0, + "content": "[140] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 132, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 132, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 131, + 686, + 507, + 704 + ], + "spans": [ + { + "bbox": [ + 131, + 686, + 507, + 704 + ], + "score": 1.0, + "content": "Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 131, + 698, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 131, + 698, + 507, + 714 + ], + "score": 1.0, + "content": "Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 132, + 710, + 235, + 722 + ], + "spans": [ + { + "bbox": [ + 132, + 710, + 235, + 722 + ], + "score": 1.0, + "content": "arXiv:2111.11432, 2021.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 666, + 507, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 72, + 505, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "[141] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. BitFit: Simple parameter-efficient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 83, + 478, + 96 + ], + "spans": [ + { + "bbox": [ + 132, + 83, + 478, + 96 + ], + "score": 1.0, + "content": "fine-tuning for transformer-based masked language-models. arXiv:2106.10199, 2021.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 505, + 115 + ], + "score": 1.0, + "content": "[142] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 113, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 505, + 126 + ], + "score": 1.0, + "content": "and Yejin Choi. MERLOT: Multimodal neural script knowledge models. Conference on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 124, + 321, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 124, + 321, + 137 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, 2021.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 506, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 155 + ], + "score": 1.0, + "content": "[143] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 153, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 132, + 153, + 506, + 167 + ], + "score": 1.0, + "content": "Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. MERLOT reserve: Neural script", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 164, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 164, + 506, + 178 + ], + "score": 1.0, + "content": "knowledge through vision and language and sound. In IEEE Computer Vision and Pattern", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 176, + 211, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 176, + 211, + 188 + ], + "score": 1.0, + "content": "Recognition, 2022.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 506, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "[144] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 206, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 506, + 217 + ], + "score": 1.0, + "content": "Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 215, + 507, + 230 + ], + "spans": [ + { + "bbox": [ + 131, + 215, + 507, + 230 + ], + "score": 1.0, + "content": "Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 227, + 236, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 236, + 239 + ], + "score": 1.0, + "content": "arXiv:2204.00598, 2022.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 105, + 245, + 504, + 269 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 506, + 259 + ], + "score": 1.0, + "content": "[145] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision trans-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 257, + 273, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 273, + 269 + ], + "score": 1.0, + "content": "formers. arXiv:2106.04560, 2021.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 504, + 309 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 289 + ], + "score": 1.0, + "content": "[146] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 285, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 506, + 301 + ], + "score": 1.0, + "content": "Kolesnikov, and Lucas Beyer. LiT: Zero-shot transfer with locked-image text tuning.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 298, + 236, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 298, + 236, + 309 + ], + "score": 1.0, + "content": "arXiv:2111.07991, 2021.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 105, + 316, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 329 + ], + "score": 1.0, + "content": "[147] Dora Zhao, Angelina Wang, and Olga Russakovsky. Understanding and evaluating racial", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 327, + 477, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 477, + 340 + ], + "score": 1.0, + "content": "biases in image captioning. In IEEE Computer Vision and Pattern Recognition, 2021.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 346, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "score": 1.0, + "content": "[148] Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use:", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "Improving few-shot performance of language models. In International Conference on Machine", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 368, + 199, + 381 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 199, + 381 + ], + "score": 1.0, + "content": "Learning, 2021.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 387, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "[149] Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 397, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 506, + 411 + ], + "score": 1.0, + "content": "from web instructional videos. In National Conference on Artificial Intelligence (AAAI), 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "[150] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "Unified vision-language pre-training for image captioning and VQA. In National Conference", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 133, + 439, + 290, + 451 + ], + "spans": [ + { + "bbox": [ + 133, + 439, + 290, + 451 + ], + "score": 1.0, + "content": "on Artificial Intelligence (AAAI), 2020.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 104, + 457, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "[151] Linchao Zhu and Yi Yang. ActBERT: Learning global-local video-text representations. In", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 468, + 353, + 481 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 353, + 481 + ], + "score": 1.0, + "content": "IEEE Computer Vision and Pattern Recognition, 2020.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 105, + 487, + 505, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "[152] Ron Zhu. Enhance multimodal transformer with external label and in-domain pretrain: Hateful", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 499, + 376, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 376, + 511 + ], + "score": 1.0, + "content": "meme challenge winning solution. arXiv:2012.08290, 2020.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 504, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "[153] Xinxin Zhu, Longteng Guo, Peng Yao, Shichen Lu, Wei Liu, and Jing Liu. Vatex video cap-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 132, + 527, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 506, + 542 + ], + "score": 1.0, + "content": "tioning challenge 2020: Multi-view features and hybrid reward strategies for video captioning.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 539, + 236, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 236, + 550 + ], + "score": 1.0, + "content": "arXiv:1910.11102, 2019.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 558, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 557, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 572 + ], + "score": 1.0, + "content": "[154] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Xiaogang Wang, Hongsheng Li, Xiaohua Wang,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 568, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 505, + 582 + ], + "score": 1.0, + "content": "and Jifeng Dai. Uni-Perceiver: Pre-training unified architecture for generic perception for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 580, + 355, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 355, + 592 + ], + "score": 1.0, + "content": "zero-shot and few-shot tasks. arXiv:2112.01522, 2021.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 105, + 598, + 505, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 598, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 506, + 611 + ], + "score": 1.0, + "content": "[155] Luisa Zintgraf, Kyriacos Shiarli, Vitaly Kurin, Katja Hofmann, and Shimon Whiteson. Fast", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 132, + 609, + 507, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 507, + 622 + ], + "score": 1.0, + "content": "context adaptation via meta-learning. In International Conference on Machine Learning, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 298, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 298, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 72, + 505, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "[141] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. BitFit: Simple parameter-efficient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 83, + 478, + 96 + ], + "spans": [ + { + "bbox": [ + 132, + 83, + 478, + 96 + ], + "score": 1.0, + "content": "fine-tuning for transformer-based masked language-models. arXiv:2106.10199, 2021.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 72, + 505, + 96 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 505, + 115 + ], + "score": 1.0, + "content": "[142] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 113, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 505, + 126 + ], + "score": 1.0, + "content": "and Yejin Choi. MERLOT: Multimodal neural script knowledge models. Conference on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 124, + 321, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 124, + 321, + 137 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, 2021.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 102, + 505, + 137 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 506, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 155 + ], + "score": 1.0, + "content": "[143] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 153, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 132, + 153, + 506, + 167 + ], + "score": 1.0, + "content": "Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. MERLOT reserve: Neural script", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 164, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 164, + 506, + 178 + ], + "score": 1.0, + "content": "knowledge through vision and language and sound. In IEEE Computer Vision and Pattern", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 176, + 211, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 176, + 211, + 188 + ], + "score": 1.0, + "content": "Recognition, 2022.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 142, + 506, + 188 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 506, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "[144] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 206, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 506, + 217 + ], + "score": 1.0, + "content": "Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 215, + 507, + 230 + ], + "spans": [ + { + "bbox": [ + 131, + 215, + 507, + 230 + ], + "score": 1.0, + "content": "Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 227, + 236, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 236, + 239 + ], + "score": 1.0, + "content": "arXiv:2204.00598, 2022.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 194, + 507, + 239 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 245, + 504, + 269 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 506, + 259 + ], + "score": 1.0, + "content": "[145] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision trans-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 257, + 273, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 273, + 269 + ], + "score": 1.0, + "content": "formers. arXiv:2106.04560, 2021.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 245, + 506, + 269 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 504, + 309 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 289 + ], + "score": 1.0, + "content": "[146] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 285, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 506, + 301 + ], + "score": 1.0, + "content": "Kolesnikov, and Lucas Beyer. LiT: Zero-shot transfer with locked-image text tuning.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 298, + 236, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 298, + 236, + 309 + ], + "score": 1.0, + "content": "arXiv:2111.07991, 2021.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 275, + 506, + 309 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 316, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 329 + ], + "score": 1.0, + "content": "[147] Dora Zhao, Angelina Wang, and Olga Russakovsky. Understanding and evaluating racial", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 327, + 477, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 477, + 340 + ], + "score": 1.0, + "content": "biases in image captioning. In IEEE Computer Vision and Pattern Recognition, 2021.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 317, + 505, + 340 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 346, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "score": 1.0, + "content": "[148] Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use:", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "Improving few-shot performance of language models. In International Conference on Machine", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 368, + 199, + 381 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 199, + 381 + ], + "score": 1.0, + "content": "Learning, 2021.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 345, + 506, + 381 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 387, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "[149] Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 397, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 506, + 411 + ], + "score": 1.0, + "content": "from web instructional videos. In National Conference on Artificial Intelligence (AAAI), 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 387, + 506, + 411 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "[150] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "Unified vision-language pre-training for image captioning and VQA. In National Conference", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 133, + 439, + 290, + 451 + ], + "spans": [ + { + "bbox": [ + 133, + 439, + 290, + 451 + ], + "score": 1.0, + "content": "on Artificial Intelligence (AAAI), 2020.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 417, + 505, + 451 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 457, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "[151] Linchao Zhu and Yi Yang. ActBERT: Learning global-local video-text representations. In", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 468, + 353, + 481 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 353, + 481 + ], + "score": 1.0, + "content": "IEEE Computer Vision and Pattern Recognition, 2020.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 457, + 505, + 481 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 487, + 505, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "[152] Ron Zhu. Enhance multimodal transformer with external label and in-domain pretrain: Hateful", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 499, + 376, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 376, + 511 + ], + "score": 1.0, + "content": "meme challenge winning solution. arXiv:2012.08290, 2020.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 487, + 505, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 504, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "[153] Xinxin Zhu, Longteng Guo, Peng Yao, Shichen Lu, Wei Liu, and Jing Liu. Vatex video cap-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 132, + 527, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 506, + 542 + ], + "score": 1.0, + "content": "tioning challenge 2020: Multi-view features and hybrid reward strategies for video captioning.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 539, + 236, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 236, + 550 + ], + "score": 1.0, + "content": "arXiv:1910.11102, 2019.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 516, + 506, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 558, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 557, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 572 + ], + "score": 1.0, + "content": "[154] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Xiaogang Wang, Hongsheng Li, Xiaohua Wang,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 568, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 505, + 582 + ], + "score": 1.0, + "content": "and Jifeng Dai. Uni-Perceiver: Pre-training unified architecture for generic perception for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 580, + 355, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 355, + 592 + ], + "score": 1.0, + "content": "zero-shot and few-shot tasks. arXiv:2112.01522, 2021.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 557, + 506, + 592 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 598, + 505, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 598, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 506, + 611 + ], + "score": 1.0, + "content": "[155] Luisa Zintgraf, Kyriacos Shiarli, Vitaly Kurin, Katja Hofmann, and Shimon Whiteson. Fast", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 132, + 609, + 507, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 507, + 622 + ], + "score": 1.0, + "content": "context adaptation via meta-learning. In International Conference on Machine Learning, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 598, + 507, + 622 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 156, + 84 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "score": 1.0, + "content": "Checklist", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 131, + 92, + 208, + 103 + ], + "lines": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "spans": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 146, + 108, + 505, + 190 + ], + "lines": [ + { + "bbox": [ + 146, + 107, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 146, + 107, + 505, + 120 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 118, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 162, + 118, + 288, + 131 + ], + "score": 1.0, + "content": "contributions and scope? [Yes]", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 130, + 435, + 144 + ], + "spans": [ + { + "bbox": [ + 144, + 130, + 435, + 144 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes] See Section 5.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 146, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes] See", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 161, + 154, + 455, + 167 + ], + "spans": [ + { + "bbox": [ + 161, + 154, + 455, + 167 + ], + "score": 1.0, + "content": "Section 5 for a brief discussion and Appendix D.2 for the full discussion.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 168, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 146, + 168, + 505, + 180 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 161, + 178, + 214, + 191 + ], + "spans": [ + { + "bbox": [ + 161, + 178, + 214, + 191 + ], + "score": 1.0, + "content": "them? [Yes]", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 131, + 193, + 302, + 205 + ], + "lines": [ + { + "bbox": [ + 129, + 192, + 304, + 207 + ], + "spans": [ + { + "bbox": [ + 129, + 192, + 304, + 207 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 146, + 208, + 452, + 233 + ], + "lines": [ + { + "bbox": [ + 145, + 208, + 453, + 221 + ], + "spans": [ + { + "bbox": [ + 145, + 208, + 453, + 221 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 144, + 221, + 424, + 234 + ], + "spans": [ + { + "bbox": [ + 144, + 221, + 424, + 234 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 130, + 237, + 241, + 248 + ], + "lines": [ + { + "bbox": [ + 128, + 235, + 243, + 250 + ], + "spans": [ + { + "bbox": [ + 128, + 235, + 243, + 250 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 146, + 251, + 505, + 388 + ], + "lines": [ + { + "bbox": [ + 145, + 250, + 507, + 263 + ], + "spans": [ + { + "bbox": [ + 145, + 250, + 507, + 263 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 160, + 262, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 160, + 262, + 505, + 274 + ], + "score": 1.0, + "content": "mental results (either in the supplemental material or as a URL)? [No] The code and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 162, + 273, + 259, + 286 + ], + "spans": [ + { + "bbox": [ + 162, + 273, + 259, + 286 + ], + "score": 1.0, + "content": "the data are proprietary.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 146, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 146, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 161, + 296, + 373, + 309 + ], + "spans": [ + { + "bbox": [ + 161, + 296, + 373, + 309 + ], + "score": 1.0, + "content": "were chosen)? [Yes] See Section 3 and Appendix B.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 147, + 310, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 147, + 310, + 506, + 323 + ], + "score": 1.0, + "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 161, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 161, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "ments multiple times)? [No] We do not observe large enough variance in our training", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 161, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 161, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "runs to justify the computation cost incurred by multiple training runs. For the largest", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 161, + 342, + 375, + 356 + ], + "spans": [ + { + "bbox": [ + 161, + 342, + 375, + 356 + ], + "score": 1.0, + "content": "models, it is not feasible within our compute budget.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 146, + 354, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 146, + 354, + 506, + 369 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g.,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 161, + 366, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 161, + 366, + 505, + 379 + ], + "score": 1.0, + "content": "type of GPUs, internal cluster, or cloud provider)? [Yes] Details can be found in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 162, + 376, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 162, + 376, + 505, + 390 + ], + "score": 1.0, + "content": "Appendix B.1.2. In short, our largest run was trained on 1536 TPU chips for 15 days.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 131, + 392, + 504, + 404 + ], + "lines": [ + { + "bbox": [ + 129, + 390, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 129, + 390, + 506, + 406 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 146, + 407, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 146, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 146, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] We properly cited the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 160, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 160, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "prior methods on which our work is based, as well as prior datasets when appropriate", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 161, + 429, + 223, + 441 + ], + "spans": [ + { + "bbox": [ + 161, + 429, + 223, + 441 + ], + "score": 1.0, + "content": "(e.g., ALIGN).", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 146, + 442, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 146, + 442, + 505, + 453 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [N/A] The assets we used are previous work", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 160, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 160, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "for which we cited papers. We do mention the license of all visual assets we use for the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 162, + 464, + 307, + 477 + ], + "spans": [ + { + "bbox": [ + 162, + 464, + 307, + 477 + ], + "score": 1.0, + "content": "figures of the paper in Appendix G.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 147, + 476, + 503, + 489 + ], + "spans": [ + { + "bbox": [ + 147, + 476, + 503, + 489 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [No]", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 146, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 146, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 161, + 499, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 161, + 499, + 506, + 514 + ], + "score": 1.0, + "content": "using/curating? [Yes] Our data was automatically scraped from million of webpages.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 161, + 510, + 307, + 523 + ], + "spans": [ + { + "bbox": [ + 161, + 510, + 307, + 523 + ], + "score": 1.0, + "content": "See Datasheets [30] in Appendix F.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 145, + 523, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 145, + 523, + 505, + 537 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 162, + 534, + 466, + 547 + ], + "spans": [ + { + "bbox": [ + 162, + 534, + 466, + 547 + ], + "score": 1.0, + "content": "information or offensive content? [Yes] See Datasheets [30] in Appendix F.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 131, + 550, + 432, + 561 + ], + "lines": [ + { + "bbox": [ + 128, + 548, + 433, + 564 + ], + "spans": [ + { + "bbox": [ + 128, + 548, + 433, + 564 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 146, + 564, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 145, + 564, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 145, + 564, + 506, + 577 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 162, + 576, + 237, + 588 + ], + "spans": [ + { + "bbox": [ + 162, + 576, + 237, + 588 + ], + "score": 1.0, + "content": "applicable? [N/A]", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 145, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 145, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 161, + 599, + 342, + 612 + ], + "spans": [ + { + "bbox": [ + 161, + 599, + 342, + 612 + ], + "score": 1.0, + "content": "Board (IRB) approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 147, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 147, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 162, + 624, + 333, + 636 + ], + "spans": [ + { + "bbox": [ + 162, + 624, + 333, + 636 + ], + "score": 1.0, + "content": "spent on participant compensation? [N/A]", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 310, + 750 + ], + "lines": [ + { + "bbox": [ + 298, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 298, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 156, + 84 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "score": 1.0, + "content": "Checklist", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 131, + 92, + 208, + 103 + ], + "lines": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "spans": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 129, + 91, + 210, + 105 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 108, + 505, + 190 + ], + "lines": [ + { + "bbox": [ + 146, + 107, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 146, + 107, + 505, + 120 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 118, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 162, + 118, + 288, + 131 + ], + "score": 1.0, + "content": "contributions and scope? [Yes]", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 144, + 130, + 435, + 144 + ], + "spans": [ + { + "bbox": [ + 144, + 130, + 435, + 144 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes] See Section 5.", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 146, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes] See", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 154, + 455, + 167 + ], + "spans": [ + { + "bbox": [ + 161, + 154, + 455, + 167 + ], + "score": 1.0, + "content": "Section 5 for a brief discussion and Appendix D.2 for the full discussion.", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 168, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 146, + 168, + 505, + 180 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 178, + 214, + 191 + ], + "spans": [ + { + "bbox": [ + 161, + 178, + 214, + 191 + ], + "score": 1.0, + "content": "them? [Yes]", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + } + ], + "index": 5, + "bbox_fs": [ + 144, + 107, + 505, + 191 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 193, + 302, + 205 + ], + "lines": [ + { + "bbox": [ + 129, + 192, + 304, + 207 + ], + "spans": [ + { + "bbox": [ + 129, + 192, + 304, + 207 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 129, + 192, + 304, + 207 + ] + }, + { + "type": "text", + "bbox": [ + 146, + 208, + 452, + 233 + ], + "lines": [ + { + "bbox": [ + 145, + 208, + 453, + 221 + ], + "spans": [ + { + "bbox": [ + 145, + 208, + 453, + 221 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 144, + 221, + 424, + 234 + ], + "spans": [ + { + "bbox": [ + 144, + 221, + 424, + 234 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 144, + 208, + 453, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 237, + 241, + 248 + ], + "lines": [ + { + "bbox": [ + 128, + 235, + 243, + 250 + ], + "spans": [ + { + "bbox": [ + 128, + 235, + 243, + 250 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 128, + 235, + 243, + 250 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 251, + 505, + 388 + ], + "lines": [ + { + "bbox": [ + 145, + 250, + 507, + 263 + ], + "spans": [ + { + "bbox": [ + 145, + 250, + 507, + 263 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 160, + 262, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 160, + 262, + 505, + 274 + ], + "score": 1.0, + "content": "mental results (either in the supplemental material or as a URL)? [No] The code and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 162, + 273, + 259, + 286 + ], + "spans": [ + { + "bbox": [ + 162, + 273, + 259, + 286 + ], + "score": 1.0, + "content": "the data are proprietary.", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 146, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 296, + 373, + 309 + ], + "spans": [ + { + "bbox": [ + 161, + 296, + 373, + 309 + ], + "score": 1.0, + "content": "were chosen)? [Yes] See Section 3 and Appendix B.", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 147, + 310, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 147, + 310, + 506, + 323 + ], + "score": 1.0, + "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 161, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "ments multiple times)? [No] We do not observe large enough variance in our training", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 161, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 161, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "runs to justify the computation cost incurred by multiple training runs. For the largest", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 161, + 342, + 375, + 356 + ], + "spans": [ + { + "bbox": [ + 161, + 342, + 375, + 356 + ], + "score": 1.0, + "content": "models, it is not feasible within our compute budget.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 354, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 146, + 354, + 506, + 369 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g.,", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 366, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 161, + 366, + 505, + 379 + ], + "score": 1.0, + "content": "type of GPUs, internal cluster, or cloud provider)? [Yes] Details can be found in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 162, + 376, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 162, + 376, + 505, + 390 + ], + "score": 1.0, + "content": "Appendix B.1.2. In short, our largest run was trained on 1536 TPU chips for 15 days.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 18.5, + "bbox_fs": [ + 145, + 250, + 507, + 390 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 392, + 504, + 404 + ], + "lines": [ + { + "bbox": [ + 129, + 390, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 129, + 390, + 506, + 406 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 129, + 390, + 506, + 406 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 407, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 146, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 146, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] We properly cited the", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 160, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 160, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "prior methods on which our work is based, as well as prior datasets when appropriate", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 161, + 429, + 223, + 441 + ], + "spans": [ + { + "bbox": [ + 161, + 429, + 223, + 441 + ], + "score": 1.0, + "content": "(e.g., ALIGN).", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 442, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 146, + 442, + 505, + 453 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [N/A] The assets we used are previous work", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 160, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 160, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "for which we cited papers. We do mention the license of all visual assets we use for the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 162, + 464, + 307, + 477 + ], + "spans": [ + { + "bbox": [ + 162, + 464, + 307, + 477 + ], + "score": 1.0, + "content": "figures of the paper in Appendix G.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 147, + 476, + 503, + 489 + ], + "spans": [ + { + "bbox": [ + 147, + 476, + 503, + 489 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [No]", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 146, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 146, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 499, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 161, + 499, + 506, + 514 + ], + "score": 1.0, + "content": "using/curating? [Yes] Our data was automatically scraped from million of webpages.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 161, + 510, + 307, + 523 + ], + "spans": [ + { + "bbox": [ + 161, + 510, + 307, + 523 + ], + "score": 1.0, + "content": "See Datasheets [30] in Appendix F.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 523, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 145, + 523, + 505, + 537 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 534, + 466, + 547 + ], + "spans": [ + { + "bbox": [ + 162, + 534, + 466, + 547 + ], + "score": 1.0, + "content": "information or offensive content? [Yes] See Datasheets [30] in Appendix F.", + "type": "text" + } + ], + "index": 37, + "is_list_end_line": true + } + ], + "index": 31.5, + "bbox_fs": [ + 145, + 407, + 506, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 550, + 432, + 561 + ], + "lines": [ + { + "bbox": [ + 128, + 548, + 433, + 564 + ], + "spans": [ + { + "bbox": [ + 128, + 548, + 433, + 564 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 128, + 548, + 433, + 564 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 564, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 145, + 564, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 145, + 564, + 506, + 577 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 576, + 237, + 588 + ], + "spans": [ + { + "bbox": [ + 162, + 576, + 237, + 588 + ], + "score": 1.0, + "content": "applicable? [N/A]", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 145, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 599, + 342, + 612 + ], + "spans": [ + { + "bbox": [ + 161, + 599, + 342, + 612 + ], + "score": 1.0, + "content": "Board (IRB) approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 147, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 147, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 624, + 333, + 636 + ], + "spans": [ + { + "bbox": [ + 162, + 624, + 333, + 636 + ], + "score": 1.0, + "content": "spent on participant compensation? [N/A]", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + } + ], + "index": 41.5, + "bbox_fs": [ + 145, + 564, + 506, + 636 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file