diff --git "a/parse/dev/S5wmbQc1We/S5wmbQc1We_middle.json" "b/parse/dev/S5wmbQc1We/S5wmbQc1We_middle.json" new file mode 100644--- /dev/null +++ "b/parse/dev/S5wmbQc1We/S5wmbQc1We_middle.json" @@ -0,0 +1,63795 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 111, + 97, + 501, + 137 + ], + "lines": [ + { + "bbox": [ + 110, + 97, + 500, + 117 + ], + "spans": [ + { + "bbox": [ + 110, + 97, + 500, + 117 + ], + "score": 1.0, + "content": "The Clock and the Pizza: Two Stories in Mechanistic", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 184, + 118, + 427, + 138 + ], + "spans": [ + { + "bbox": [ + 184, + 118, + 427, + 138 + ], + "score": 1.0, + "content": "Explanation of Neural Networks", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 177, + 178, + 434, + 213 + ], + "lines": [ + { + "bbox": [ + 176, + 178, + 434, + 192 + ], + "spans": [ + { + "bbox": [ + 176, + 178, + 434, + 192 + ], + "score": 1.0, + "content": "Ziqian Zhong*, Ziming Liu*, Max Tegmark, Jacob Andreas", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 226, + 189, + 384, + 204 + ], + "spans": [ + { + "bbox": [ + 226, + 189, + 384, + 204 + ], + "score": 1.0, + "content": "Massachusetts Institute of Technology", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 201, + 406, + 214 + ], + "spans": [ + { + "bbox": [ + 205, + 201, + 406, + 214 + ], + "score": 1.0, + "content": "{ziqianz, zmliu, tegmark, jda}@mit.edu", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 283, + 240, + 328, + 254 + ], + "lines": [ + { + "bbox": [ + 281, + 240, + 330, + 255 + ], + "spans": [ + { + "bbox": [ + 281, + 240, + 330, + 255 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 267, + 469, + 430 + ], + "lines": [ + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "score": 1.0, + "content": "Do neural networks, trained on well-understood algorithmic tasks, reliably re-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 278, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 142, + 278, + 469, + 290 + ], + "score": 1.0, + "content": "discover known algorithms for solving those tasks? Several recent studies, on", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "tasks ranging from group arithmetic to in-context linear regression, have suggested", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 299, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 313 + ], + "score": 1.0, + "content": "that the answer is yes. Using modular addition as a prototypical problem, we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 310, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 470, + 323 + ], + "score": 1.0, + "content": "show that algorithm discovery in neural networks is sometimes more complex.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "score": 1.0, + "content": "Small changes to model hyperparameters and initializations can induce discovery", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "score": 1.0, + "content": "of qualitatively different algorithms from a fixed training set, and even parallel", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 344, + 469, + 355 + ], + "spans": [ + { + "bbox": [ + 142, + 344, + 469, + 355 + ], + "score": 1.0, + "content": "implementations of multiple such algorithms. Some networks trained to perform", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "score": 1.0, + "content": "modular addition implement a familiar Clock algorithm (previously described by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 365, + 470, + 377 + ], + "spans": [ + { + "bbox": [ + 141, + 365, + 470, + 377 + ], + "score": 1.0, + "content": "Nanda et al. [1]); others implement a previously undescribed, less intuitive, but", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 377, + 469, + 388 + ], + "spans": [ + { + "bbox": [ + 142, + 377, + 469, + 388 + ], + "score": 1.0, + "content": "comprehensible procedure we term the Pizza algorithm, or a variety of even more", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 388, + 469, + 399 + ], + "spans": [ + { + "bbox": [ + 142, + 388, + 469, + 399 + ], + "score": 1.0, + "content": "complex procedures. Our results show that even simple learning problems can", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 398, + 469, + 410 + ], + "spans": [ + { + "bbox": [ + 141, + 398, + 469, + 410 + ], + "score": 1.0, + "content": "admit a surprising diversity of solutions, motivating the development of new tools", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 409, + 469, + 421 + ], + "spans": [ + { + "bbox": [ + 141, + 409, + 469, + 421 + ], + "score": 1.0, + "content": "for characterizing the behavior of neural networks across their algorithmic phase", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 140, + 419, + 178, + 432 + ], + "spans": [ + { + "bbox": [ + 140, + 419, + 178, + 432 + ], + "score": 1.0, + "content": "space. 1", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 452, + 190, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 192, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 192, + 468 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 478, + 505, + 544 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 491 + ], + "score": 1.0, + "content": "Mechanistically understanding deep network models—reverse-engineering their learned algorithms", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 489, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 500 + ], + "score": 1.0, + "content": "and representation schemes—remains a major challenge across problem domains. Several recent", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 500, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 512 + ], + "score": 1.0, + "content": "studies [2, 3, 4, 5, 1] have exhibited specific examples of models apparently re-discovering inter-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "pretable (and in some cases familiar) solutions to tasks like curve detection, sequence copying and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "modular arithmetic. Are these models the exception or the rule? Under what conditions do neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 533, + 415, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 415, + 544 + ], + "score": 1.0, + "content": "network models discover familiar algorithmic solutions to algorithmic tasks?", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 549, + 505, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 548, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 506, + 562 + ], + "score": 1.0, + "content": "In this paper, we focus specifically on the problem of learning modular addition, training networks", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 560, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 195, + 572 + ], + "score": 1.0, + "content": "to compute sums like", + "type": "text" + }, + { + "bbox": [ + 195, + 560, + 237, + 570 + ], + "score": 0.9, + "content": "8 + 6 = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 560, + 505, + 572 + ], + "score": 1.0, + "content": "(mod 12). Modular arithmetic can be implemented with a simple", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 570, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 584 + ], + "score": 1.0, + "content": "geometric solution, familiar to anyone who has learned to read a clock: every integer is represented", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 582, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 594 + ], + "score": 1.0, + "content": "as an angle, input angles are added together, and the resulting angle evaluated to obtain a modular", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "sum (Figure 1, left). Nanda et al. [1] show that specific neural network architectures, when trained to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "score": 1.0, + "content": "perform modular addition, implement this Clock algorithm. In this work, we show that the Clock", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "algorithm is only one part of a more complicated picture of algorithm learning in deep networks. In", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 625, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 506, + 639 + ], + "score": 1.0, + "content": "particular, networks structurally similar to the ones trained by Nanda et al. preferentially implement a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "qualitatively different approach to modular arithmetic, which we term the Pizza algorithm (Figure 1,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "right), and sometimes even more complex solutions. Models exhibit sharp algorithmic phase", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "transitions [6] between the Clock and Pizza algorithms as their width and attention strength very, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 669, + 417, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 417, + 681 + ], + "score": 1.0, + "content": "often implement multiple, imperfect copies of the Pizza algorithm in parallel.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 691, + 324, + 712 + ], + "lines": [ + { + "bbox": [ + 119, + 689, + 195, + 702 + ], + "spans": [ + { + "bbox": [ + 119, + 689, + 195, + 702 + ], + "score": 1.0, + "content": "*Equal contribution.", + "type": "text" + } + ] + }, + { + "bbox": [ + 119, + 700, + 324, + 713 + ], + "spans": [ + { + "bbox": [ + 119, + 700, + 324, + 713 + ], + "score": 1.0, + "content": "1Code is available at https://github.com/fjzzq2002/pizza.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 731, + 385, + 742 + ], + "lines": [ + { + "bbox": [ + 105, + 730, + 386, + 744 + ], + "spans": [ + { + "bbox": [ + 105, + 730, + 386, + 744 + ], + "score": 1.0, + "content": "37th Conference on Neural Information Processing Systems (NeurIPS 2023).", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 111, + 97, + 501, + 137 + ], + "lines": [ + { + "bbox": [ + 110, + 97, + 500, + 117 + ], + "spans": [ + { + "bbox": [ + 110, + 97, + 500, + 117 + ], + "score": 1.0, + "content": "The Clock and the Pizza: Two Stories in Mechanistic", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 184, + 118, + 427, + 138 + ], + "spans": [ + { + "bbox": [ + 184, + 118, + 427, + 138 + ], + "score": 1.0, + "content": "Explanation of Neural Networks", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 177, + 178, + 434, + 213 + ], + "lines": [ + { + "bbox": [ + 176, + 178, + 434, + 192 + ], + "spans": [ + { + "bbox": [ + 176, + 178, + 434, + 192 + ], + "score": 1.0, + "content": "Ziqian Zhong*, Ziming Liu*, Max Tegmark, Jacob Andreas", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 226, + 189, + 384, + 204 + ], + "spans": [ + { + "bbox": [ + 226, + 189, + 384, + 204 + ], + "score": 1.0, + "content": "Massachusetts Institute of Technology", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 201, + 406, + 214 + ], + "spans": [ + { + "bbox": [ + 205, + 201, + 406, + 214 + ], + "score": 1.0, + "content": "{ziqianz, zmliu, tegmark, jda}@mit.edu", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 176, + 178, + 434, + 214 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 240, + 328, + 254 + ], + "lines": [ + { + "bbox": [ + 281, + 240, + 330, + 255 + ], + "spans": [ + { + "bbox": [ + 281, + 240, + 330, + 255 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 267, + 469, + 430 + ], + "lines": [ + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "score": 1.0, + "content": "Do neural networks, trained on well-understood algorithmic tasks, reliably re-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 278, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 142, + 278, + 469, + 290 + ], + "score": 1.0, + "content": "discover known algorithms for solving those tasks? Several recent studies, on", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "tasks ranging from group arithmetic to in-context linear regression, have suggested", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 299, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 313 + ], + "score": 1.0, + "content": "that the answer is yes. Using modular addition as a prototypical problem, we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 310, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 470, + 323 + ], + "score": 1.0, + "content": "show that algorithm discovery in neural networks is sometimes more complex.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "score": 1.0, + "content": "Small changes to model hyperparameters and initializations can induce discovery", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "score": 1.0, + "content": "of qualitatively different algorithms from a fixed training set, and even parallel", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 344, + 469, + 355 + ], + "spans": [ + { + "bbox": [ + 142, + 344, + 469, + 355 + ], + "score": 1.0, + "content": "implementations of multiple such algorithms. Some networks trained to perform", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "score": 1.0, + "content": "modular addition implement a familiar Clock algorithm (previously described by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 365, + 470, + 377 + ], + "spans": [ + { + "bbox": [ + 141, + 365, + 470, + 377 + ], + "score": 1.0, + "content": "Nanda et al. [1]); others implement a previously undescribed, less intuitive, but", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 377, + 469, + 388 + ], + "spans": [ + { + "bbox": [ + 142, + 377, + 469, + 388 + ], + "score": 1.0, + "content": "comprehensible procedure we term the Pizza algorithm, or a variety of even more", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 388, + 469, + 399 + ], + "spans": [ + { + "bbox": [ + 142, + 388, + 469, + 399 + ], + "score": 1.0, + "content": "complex procedures. Our results show that even simple learning problems can", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 398, + 469, + 410 + ], + "spans": [ + { + "bbox": [ + 141, + 398, + 469, + 410 + ], + "score": 1.0, + "content": "admit a surprising diversity of solutions, motivating the development of new tools", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 409, + 469, + 421 + ], + "spans": [ + { + "bbox": [ + 141, + 409, + 469, + 421 + ], + "score": 1.0, + "content": "for characterizing the behavior of neural networks across their algorithmic phase", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 140, + 419, + 178, + 432 + ], + "spans": [ + { + "bbox": [ + 140, + 419, + 178, + 432 + ], + "score": 1.0, + "content": "space. 1", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 13, + "bbox_fs": [ + 140, + 267, + 470, + 432 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 452, + 190, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 192, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 192, + 468 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 478, + 505, + 544 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 491 + ], + "score": 1.0, + "content": "Mechanistically understanding deep network models—reverse-engineering their learned algorithms", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 489, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 500 + ], + "score": 1.0, + "content": "and representation schemes—remains a major challenge across problem domains. Several recent", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 500, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 512 + ], + "score": 1.0, + "content": "studies [2, 3, 4, 5, 1] have exhibited specific examples of models apparently re-discovering inter-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "pretable (and in some cases familiar) solutions to tasks like curve detection, sequence copying and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "modular arithmetic. Are these models the exception or the rule? Under what conditions do neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 533, + 415, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 415, + 544 + ], + "score": 1.0, + "content": "network models discover familiar algorithmic solutions to algorithmic tasks?", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 477, + 506, + 544 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 549, + 505, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 548, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 506, + 562 + ], + "score": 1.0, + "content": "In this paper, we focus specifically on the problem of learning modular addition, training networks", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 560, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 195, + 572 + ], + "score": 1.0, + "content": "to compute sums like", + "type": "text" + }, + { + "bbox": [ + 195, + 560, + 237, + 570 + ], + "score": 0.9, + "content": "8 + 6 = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 560, + 505, + 572 + ], + "score": 1.0, + "content": "(mod 12). Modular arithmetic can be implemented with a simple", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 570, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 584 + ], + "score": 1.0, + "content": "geometric solution, familiar to anyone who has learned to read a clock: every integer is represented", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 582, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 594 + ], + "score": 1.0, + "content": "as an angle, input angles are added together, and the resulting angle evaluated to obtain a modular", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "sum (Figure 1, left). Nanda et al. [1] show that specific neural network architectures, when trained to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "score": 1.0, + "content": "perform modular addition, implement this Clock algorithm. In this work, we show that the Clock", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "algorithm is only one part of a more complicated picture of algorithm learning in deep networks. In", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 625, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 506, + 639 + ], + "score": 1.0, + "content": "particular, networks structurally similar to the ones trained by Nanda et al. preferentially implement a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "qualitatively different approach to modular arithmetic, which we term the Pizza algorithm (Figure 1,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "right), and sometimes even more complex solutions. Models exhibit sharp algorithmic phase", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "transitions [6] between the Clock and Pizza algorithms as their width and attention strength very, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 669, + 417, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 417, + 681 + ], + "score": 1.0, + "content": "often implement multiple, imperfect copies of the Pizza algorithm in parallel.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 548, + 506, + 681 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 238, + 77, + 441, + 84 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 237, + 76, + 442, + 84 + ], + "spans": [ + { + "bbox": [ + 237, + 76, + 303, + 84 + ], + "score": 1.0, + "content": "Step 1: Embed token a and", + "type": "text" + }, + { + "bbox": [ + 304, + 77, + 308, + 83 + ], + "score": 0.43, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 76, + 348, + 84 + ], + "score": 1.0, + "content": "to a circle where", + "type": "text" + }, + { + "bbox": [ + 348, + 77, + 376, + 84 + ], + "score": 0.88, + "content": "w _ { k } = 2 \\pi k / p", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 76, + 398, + 84 + ], + "score": 1.0, + "content": "for some", + "type": "text" + }, + { + "bbox": [ + 399, + 77, + 442, + 84 + ], + "score": 0.81, + "content": "k \\in [ 1 , 2 ^ { \\ldots } , p - 1 ]", + "type": "inline_equation" + } + ], + "index": 0 + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 254, + 86, + 430, + 93 + ], + "lines": [ + { + "bbox": [ + 254, + 86, + 430, + 93 + ], + "spans": [ + { + "bbox": [ + 254, + 86, + 430, + 93 + ], + "score": 0.76, + "content": "a \\mathbf { E } _ { a } \\equiv ( \\mathbf { E } _ { a , \\mathrm { x } } , \\mathbf { E } _ { a , \\mathrm { y } } ) = ( \\cos ( w _ { k } a ) , \\sin ( w _ { k } a ) ) , b \\mathbf { E } _ { b } \\equiv ( \\mathbf { E } _ { b , \\mathrm { x } } , \\mathbf { E } _ { b , \\mathrm { y } } ) = ( \\cos ( w _ { k } b ) , \\sin ( w _ { k } b ) )", + "type": "interline_equation", + "image_path": "1d249505fc3f0761cb53e16ff7dd3894f83afd1bfb2d56a6e7abbf7be57a418b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 254, + 86, + 430, + 93 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "image", + "bbox": [ + 110, + 75, + 165, + 168 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 75, + 165, + 168 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 110, + 75, + 165, + 168 + ], + "spans": [ + { + "bbox": [ + 110, + 75, + 165, + 168 + ], + "score": 0.88, + "type": "image", + "image_path": "e855a035fa9f6b20eb042a7751bf3efaaa4efce15ea2db6d154cb1cc3673e02b.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 110, + 75, + 165, + 168 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 227, + 101, + 290, + 111 + ], + "lines": [ + { + "bbox": [ + 225, + 99, + 292, + 112 + ], + "spans": [ + { + "bbox": [ + 225, + 99, + 292, + 112 + ], + "score": 1.0, + "content": "Clock Algorithm", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 389, + 102, + 452, + 111 + ], + "lines": [ + { + "bbox": [ + 388, + 101, + 453, + 112 + ], + "spans": [ + { + "bbox": [ + 388, + 101, + 453, + 112 + ], + "score": 1.0, + "content": "Pizza Algorithm", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 188, + 113, + 311, + 120 + ], + "lines": [ + { + "bbox": [ + 187, + 112, + 312, + 120 + ], + "spans": [ + { + "bbox": [ + 187, + 112, + 312, + 120 + ], + "score": 1.0, + "content": "Step 2: compute the angle sum using multiplication.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 349, + 113, + 434, + 119 + ], + "lines": [ + { + "bbox": [ + 348, + 113, + 435, + 120 + ], + "spans": [ + { + "bbox": [ + 348, + 113, + 435, + 120 + ], + "score": 1.0, + "content": "Step 2.1: compute the vector mean.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 200, + 122, + 316, + 149 + ], + "lines": [ + { + "bbox": [ + 200, + 122, + 316, + 138 + ], + "spans": [ + { + "bbox": [ + 200, + 122, + 316, + 138 + ], + "score": 0.43, + "content": "\\mathbf { E } _ { a b } \\equiv { \\binom { \\mathbf { E } _ { a b , x } } { \\mathbf { E } _ { a b , y } } } = { \\binom { \\mathbf { E } _ { a , x } \\mathbf { E } _ { b , x } - \\mathbf { E } _ { a , y } \\mathbf { E } _ { b , y } } { \\mathbf { E } _ { a , x } \\mathbf { E } _ { b , y } + \\mathbf { E } _ { a , y } \\mathbf { E } _ { b , x } } } = { \\binom { \\cos ( w _ { k } ( a + b ) ) } { \\sin ( w _ { k } ( a + b ) ) } }", + "type": "inline_equation", + "image_path": "04f9a22f324d3223bac22f363b658863254014c0d441d4dea386aa215fe7ae5f.jpg" + } + ], + "index": 7 + }, + { + "bbox": [ + 199, + 139, + 223, + 150 + ], + "spans": [ + { + "bbox": [ + 199, + 139, + 223, + 150 + ], + "score": 1.0, + "content": "Hab = Eab", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.0 + }, + { + "type": "interline_equation", + "bbox": [ + 356, + 122, + 489, + 128 + ], + "lines": [ + { + "bbox": [ + 356, + 122, + 489, + 128 + ], + "spans": [ + { + "bbox": [ + 356, + 122, + 489, + 128 + ], + "score": 0.77, + "content": "\\mathbf { E } _ { a b } = ( \\mathbf { E } _ { a } + \\mathbf { E } _ { b } ) / 2 = ( \\cos ( w _ { k } a ) + \\cos ( w _ { k } b ) , \\sin ( w _ { k } a ) + \\sin ( w _ { k } b ) ) / 2", + "type": "interline_equation", + "image_path": "7c6e6391aac371a512032d8b6f9c884cb9bf068d1e4662987fbc2d0ac59f2372.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 356, + 122, + 489, + 128 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 350, + 131, + 472, + 138 + ], + "lines": [ + { + "bbox": [ + 348, + 130, + 473, + 140 + ], + "spans": [ + { + "bbox": [ + 348, + 130, + 385, + 140 + ], + "score": 1.0, + "content": "Step 2.2: using", + "type": "text" + }, + { + "bbox": [ + 386, + 132, + 394, + 138 + ], + "score": 0.64, + "content": "\\mathbf { E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 130, + 464, + 140 + ], + "score": 1.0, + "content": "and nonlinearities to compute", + "type": "text" + }, + { + "bbox": [ + 464, + 132, + 473, + 138 + ], + "score": 0.3, + "content": "\\mathbf { H } _ { a b }", + "type": "inline_equation" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 357, + 141, + 471, + 148 + ], + "lines": [ + { + "bbox": [ + 357, + 141, + 471, + 148 + ], + "spans": [ + { + "bbox": [ + 357, + 141, + 471, + 148 + ], + "score": 0.68, + "content": "\\mathbf { H } _ { a b } = | \\cos ( w _ { k } ( a - b ) / 2 ) | ( \\cos ( w _ { k } ( a + b ) ) , \\sin ( w _ { k } ( a + b ) ) )", + "type": "interline_equation", + "image_path": "f92e5c6e04a7cfb3d8a4380fbc8aa4fd803e2717c044113dc7e0f8e71a6f72d9.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 357, + 141, + 471, + 148 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 277, + 158, + 402, + 164 + ], + "lines": [ + { + "bbox": [ + 277, + 157, + 402, + 165 + ], + "spans": [ + { + "bbox": [ + 277, + 157, + 350, + 165 + ], + "score": 1.0, + "content": "Step 3: score possible outputs", + "type": "text" + }, + { + "bbox": [ + 350, + 159, + 354, + 164 + ], + "score": 0.31, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 157, + 402, + 165 + ], + "score": 1.0, + "content": "using a dot product.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 281, + 164, + 399, + 172 + ], + "lines": [ + { + "bbox": [ + 280, + 163, + 400, + 172 + ], + "spans": [ + { + "bbox": [ + 280, + 163, + 284, + 172 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 284, + 164, + 316, + 172 + ], + "score": 0.28, + "content": "Q _ { a b c } = \\mathbf { U } _ { c } \\cdot \\mathbf { H } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 163, + 400, + 172 + ], + "score": 1.0, + "content": ", Uc ≡ (Ec,x, Ec,y) = (cos(wkc), sin(wkc))", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 181, + 172, + 500, + 264 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 181, + 172, + 500, + 264 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 181, + 172, + 500, + 264 + ], + "spans": [ + { + "bbox": [ + 181, + 172, + 500, + 264 + ], + "score": 0.823, + "type": "image", + "image_path": "a8e0022db1a64b8295bc2a10d5d9ef3d0796653ed202d0d3d29b73ded004df02.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 181, + 172, + 500, + 202.66666666666666 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 181, + 202.66666666666666, + 500, + 233.33333333333331 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 181, + 233.33333333333331, + 500, + 264.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 185, + 270, + 425, + 282 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 186, + 270, + 425, + 284 + ], + "spans": [ + { + "bbox": [ + 186, + 270, + 425, + 284 + ], + "score": 1.0, + "content": "Figure 1: Illustration of the Clock and the Pizza Algorithm.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + } + ], + "index": 16.0 + }, + { + "type": "text", + "bbox": [ + 107, + 304, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 303, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 317 + ], + "score": 1.0, + "content": "Our results highlight the complexity of mechanistic description in even models trained to perform", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "simple tasks. They point to characterization of algorithmic phase spaces, not just single algorithmic", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 326, + 369, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 369, + 339 + ], + "score": 1.0, + "content": "solutions, as an important goal in algorithm-level interpretability.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "Organization In Section 2, we review the Clock algorithm [1] and show empirical evidence of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "deviation from it in models trained to perform modular addition. In Section 3, we show that these", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 365, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 505, + 376 + ], + "score": 1.0, + "content": "deviations can be explained by an alternative Pizza algorithm. In Section 4, we define additional", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 376, + 504, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 504, + 387 + ], + "score": 1.0, + "content": "metrics to distinguish between these algorithms, and detect phase transitions between these algorithms", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "score": 1.0, + "content": "(and others Non-circular algorithms) when architectures and hyperparameters are varied. We discuss", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "the relationship between these findings and other work on model interpretation in Section 5, and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 408, + 198, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 198, + 418 + ], + "score": 1.0, + "content": "conclude in Section 6.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 107, + 434, + 358, + 448 + ], + "lines": [ + { + "bbox": [ + 104, + 432, + 359, + 452 + ], + "spans": [ + { + "bbox": [ + 104, + 432, + 359, + 452 + ], + "score": 1.0, + "content": "2 Modular Arithmetic and the Clock Algorithm", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 459, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 358, + 472 + ], + "score": 1.0, + "content": "Setup We train neural networks to perform modular addition", + "type": "text" + }, + { + "bbox": [ + 358, + 460, + 436, + 472 + ], + "score": 0.59, + "content": "a + b = c ( \\mathrm { m o d } p )", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 459, + 469, + 472 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 469, + 460, + 505, + 471 + ], + "score": 0.89, + "content": "a , b , c =", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 107, + 471, + 169, + 482 + ], + "score": 0.9, + "content": "0 , 1 , \\cdots , p - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 470, + 208, + 483 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 208, + 471, + 240, + 482 + ], + "score": 0.89, + "content": "p = 5 9", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 470, + 469, + 483 + ], + "score": 1.0, + "content": "throughout the paper. In these networks, every integer", + "type": "text" + }, + { + "bbox": [ + 470, + 472, + 474, + 480 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "has an", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 480, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 480, + 227, + 495 + ], + "score": 1.0, + "content": "associated embedding vector", + "type": "text" + }, + { + "bbox": [ + 228, + 481, + 266, + 492 + ], + "score": 0.91, + "content": "\\mathbf { E } _ { t } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 480, + 421, + 495 + ], + "score": 1.0, + "content": ". Networks take as input embeddings", + "type": "text" + }, + { + "bbox": [ + 422, + 481, + 486, + 493 + ], + "score": 0.92, + "content": "[ \\dot { \\mathbf { E } _ { a } } , \\mathbf { E } _ { b } ] ^ { \\mathbf { ^ { \\prime } } } \\in \\mathbb { R } ^ { 2 d }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 480, + 506, + 495 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 104, + 492, + 217, + 506 + ], + "score": 1.0, + "content": "predict a categorical output", + "type": "text" + }, + { + "bbox": [ + 217, + 495, + 223, + 502 + ], + "score": 0.58, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 492, + 506, + 506 + ], + "score": 1.0, + "content": ". Both embeddings and network parameters are learned. In preliminary", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 504, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 515 + ], + "score": 1.0, + "content": "experiments, we train two different network architectures on the modular arithmetic task, which we", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "score": 1.0, + "content": "refer to as: Model A and Model B. Model A is a one-layer ReLU transformer [7] with constant", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 201, + 537 + ], + "score": 1.0, + "content": "attention, while Model", + "type": "text" + }, + { + "bbox": [ + 201, + 525, + 210, + 535 + ], + "score": 0.46, + "content": "\\mathbf { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 524, + 506, + 537 + ], + "score": 1.0, + "content": "is a standard one-layer ReLU transformer (see Appendix F.1 for details).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 535, + 499, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 499, + 549 + ], + "score": 1.0, + "content": "As attention is not involved in Model A, it can also be understood as a ReLU MLP (Appendix G).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 108, + 560, + 261, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 262, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 262, + 575 + ], + "score": 1.0, + "content": "2.1 Review of the Clock Algorithm", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 505, + 636 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 460, + 593 + ], + "score": 1.0, + "content": "As in past work, we find that after training both Model A and Model B, embeddings", + "type": "text" + }, + { + "bbox": [ + 460, + 581, + 493, + 592 + ], + "score": 0.91, + "content": "( \\mathbf { E } _ { a } , \\mathbf { E } _ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 592, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 505, + 604 + ], + "score": 1.0, + "content": "Figure 1) usually describe a circle [8] in the plane spanned by the first two principal components of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 602, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 239, + 616 + ], + "score": 1.0, + "content": "the embedding matrix. Formally,", + "type": "text" + }, + { + "bbox": [ + 239, + 603, + 352, + 615 + ], + "score": 0.9, + "content": "\\mathbf { E } _ { a } \\approx [ \\cos ( w _ { k } a ) , \\bar { \\sin ( w _ { k } a ) } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 602, + 379, + 616 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 380, + 603, + 432, + 615 + ], + "score": 0.91, + "content": "w _ { k } = 2 \\pi k / p", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 602, + 436, + 616 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 436, + 603, + 443, + 613 + ], + "score": 0.66, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 602, + 505, + 616 + ], + "score": 1.0, + "content": "is an integer in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 613, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 107, + 614, + 145, + 626 + ], + "score": 0.89, + "content": "[ 1 , p - 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 613, + 506, + 628 + ], + "score": 1.0, + "content": ". Nanda et al. [1] discovered a circuit that uses these circular embeddings to implement an", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 624, + 437, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 437, + 638 + ], + "score": 1.0, + "content": "interpretable algorithm for modular arithmetic, which we call the Clock algorithm.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "table", + "bbox": [ + 122, + 654, + 486, + 698 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 122, + 654, + 486, + 698 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 654, + 486, + 698 + ], + "spans": [ + { + "bbox": [ + 122, + 654, + 486, + 698 + ], + "score": 0.975, + "html": "
AlgorithmLearned EmbeddingsGradientSymmetryRequired Non-linearity
ClockCircleNoMultiplication
PizzaCircleYesAbsolute value
Non-circularLine,Lissajous-like curves, etc.N/AN/A
", + "type": "table", + "image_path": "2f25e792d71bb04f39c1bf014fa70aa28c6f7d41afa66156b41e29d261c38c52.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 122, + 654, + 486, + 668.6666666666666 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 122, + 668.6666666666666, + 486, + 683.3333333333333 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 122, + 683.3333333333333, + 486, + 697.9999999999999 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 190, + 701, + 421, + 712 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 189, + 700, + 421, + 713 + ], + "spans": [ + { + "bbox": [ + 189, + 700, + 421, + 713 + ], + "score": 1.0, + "content": "Table 1: Different neural algorithms for modular addition", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + } + ], + "index": 45.0 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 175, + 95, + 192, + 112 + ], + "lines": [] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 340, + 95, + 357, + 111 + ], + "lines": [ + { + "bbox": [ + 339, + 93, + 360, + 114 + ], + "spans": [ + { + "bbox": [ + 339, + 93, + 360, + 114 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 238, + 77, + 441, + 84 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 237, + 76, + 442, + 84 + ], + "spans": [ + { + "bbox": [ + 237, + 76, + 303, + 84 + ], + "score": 1.0, + "content": "Step 1: Embed token a and", + "type": "text" + }, + { + "bbox": [ + 304, + 77, + 308, + 83 + ], + "score": 0.43, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 76, + 348, + 84 + ], + "score": 1.0, + "content": "to a circle where", + "type": "text" + }, + { + "bbox": [ + 348, + 77, + 376, + 84 + ], + "score": 0.88, + "content": "w _ { k } = 2 \\pi k / p", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 76, + 398, + 84 + ], + "score": 1.0, + "content": "for some", + "type": "text" + }, + { + "bbox": [ + 399, + 77, + 442, + 84 + ], + "score": 0.81, + "content": "k \\in [ 1 , 2 ^ { \\ldots } , p - 1 ]", + "type": "inline_equation" + } + ], + "index": 0 + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 254, + 86, + 430, + 93 + ], + "lines": [ + { + "bbox": [ + 254, + 86, + 430, + 93 + ], + "spans": [ + { + "bbox": [ + 254, + 86, + 430, + 93 + ], + "score": 0.76, + "content": "a \\mathbf { E } _ { a } \\equiv ( \\mathbf { E } _ { a , \\mathrm { x } } , \\mathbf { E } _ { a , \\mathrm { y } } ) = ( \\cos ( w _ { k } a ) , \\sin ( w _ { k } a ) ) , b \\mathbf { E } _ { b } \\equiv ( \\mathbf { E } _ { b , \\mathrm { x } } , \\mathbf { E } _ { b , \\mathrm { y } } ) = ( \\cos ( w _ { k } b ) , \\sin ( w _ { k } b ) )", + "type": "interline_equation", + "image_path": "1d249505fc3f0761cb53e16ff7dd3894f83afd1bfb2d56a6e7abbf7be57a418b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 254, + 86, + 430, + 93 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "image", + "bbox": [ + 110, + 75, + 165, + 168 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 75, + 165, + 168 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 110, + 75, + 165, + 168 + ], + "spans": [ + { + "bbox": [ + 110, + 75, + 165, + 168 + ], + "score": 0.88, + "type": "image", + "image_path": "e855a035fa9f6b20eb042a7751bf3efaaa4efce15ea2db6d154cb1cc3673e02b.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 110, + 75, + 165, + 168 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 227, + 101, + 290, + 111 + ], + "lines": [ + { + "bbox": [ + 225, + 99, + 292, + 112 + ], + "spans": [ + { + "bbox": [ + 225, + 99, + 292, + 112 + ], + "score": 1.0, + "content": "Clock Algorithm", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 389, + 102, + 452, + 111 + ], + "lines": [ + { + "bbox": [ + 388, + 101, + 453, + 112 + ], + "spans": [ + { + "bbox": [ + 388, + 101, + 453, + 112 + ], + "score": 1.0, + "content": "Pizza Algorithm", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 188, + 113, + 311, + 120 + ], + "lines": [ + { + "bbox": [ + 187, + 112, + 312, + 120 + ], + "spans": [ + { + "bbox": [ + 187, + 112, + 312, + 120 + ], + "score": 1.0, + "content": "Step 2: compute the angle sum using multiplication.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 187, + 112, + 312, + 120 + ] + }, + { + "type": "text", + "bbox": [ + 349, + 113, + 434, + 119 + ], + "lines": [ + { + "bbox": [ + 348, + 113, + 435, + 120 + ], + "spans": [ + { + "bbox": [ + 348, + 113, + 435, + 120 + ], + "score": 1.0, + "content": "Step 2.1: compute the vector mean.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 348, + 113, + 435, + 120 + ] + }, + { + "type": "text", + "bbox": [ + 200, + 122, + 316, + 149 + ], + "lines": [ + { + "bbox": [ + 200, + 122, + 316, + 138 + ], + "spans": [ + { + "bbox": [ + 200, + 122, + 316, + 138 + ], + "score": 0.43, + "content": "\\mathbf { E } _ { a b } \\equiv { \\binom { \\mathbf { E } _ { a b , x } } { \\mathbf { E } _ { a b , y } } } = { \\binom { \\mathbf { E } _ { a , x } \\mathbf { E } _ { b , x } - \\mathbf { E } _ { a , y } \\mathbf { E } _ { b , y } } { \\mathbf { E } _ { a , x } \\mathbf { E } _ { b , y } + \\mathbf { E } _ { a , y } \\mathbf { E } _ { b , x } } } = { \\binom { \\cos ( w _ { k } ( a + b ) ) } { \\sin ( w _ { k } ( a + b ) ) } }", + "type": "inline_equation", + "image_path": "04f9a22f324d3223bac22f363b658863254014c0d441d4dea386aa215fe7ae5f.jpg" + } + ], + "index": 7 + }, + { + "bbox": [ + 199, + 139, + 223, + 150 + ], + "spans": [ + { + "bbox": [ + 199, + 139, + 223, + 150 + ], + "score": 1.0, + "content": "Hab = Eab", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.0, + "bbox_fs": [ + 199, + 122, + 316, + 150 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 356, + 122, + 489, + 128 + ], + "lines": [ + { + "bbox": [ + 356, + 122, + 489, + 128 + ], + "spans": [ + { + "bbox": [ + 356, + 122, + 489, + 128 + ], + "score": 0.77, + "content": "\\mathbf { E } _ { a b } = ( \\mathbf { E } _ { a } + \\mathbf { E } _ { b } ) / 2 = ( \\cos ( w _ { k } a ) + \\cos ( w _ { k } b ) , \\sin ( w _ { k } a ) + \\sin ( w _ { k } b ) ) / 2", + "type": "interline_equation", + "image_path": "7c6e6391aac371a512032d8b6f9c884cb9bf068d1e4662987fbc2d0ac59f2372.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 356, + 122, + 489, + 128 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 350, + 131, + 472, + 138 + ], + "lines": [ + { + "bbox": [ + 348, + 130, + 473, + 140 + ], + "spans": [ + { + "bbox": [ + 348, + 130, + 385, + 140 + ], + "score": 1.0, + "content": "Step 2.2: using", + "type": "text" + }, + { + "bbox": [ + 386, + 132, + 394, + 138 + ], + "score": 0.64, + "content": "\\mathbf { E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 130, + 464, + 140 + ], + "score": 1.0, + "content": "and nonlinearities to compute", + "type": "text" + }, + { + "bbox": [ + 464, + 132, + 473, + 138 + ], + "score": 0.3, + "content": "\\mathbf { H } _ { a b }", + "type": "inline_equation" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 348, + 130, + 473, + 140 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 357, + 141, + 471, + 148 + ], + "lines": [ + { + "bbox": [ + 357, + 141, + 471, + 148 + ], + "spans": [ + { + "bbox": [ + 357, + 141, + 471, + 148 + ], + "score": 0.68, + "content": "\\mathbf { H } _ { a b } = | \\cos ( w _ { k } ( a - b ) / 2 ) | ( \\cos ( w _ { k } ( a + b ) ) , \\sin ( w _ { k } ( a + b ) ) )", + "type": "interline_equation", + "image_path": "f92e5c6e04a7cfb3d8a4380fbc8aa4fd803e2717c044113dc7e0f8e71a6f72d9.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 357, + 141, + 471, + 148 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 277, + 158, + 402, + 164 + ], + "lines": [ + { + "bbox": [ + 277, + 157, + 402, + 165 + ], + "spans": [ + { + "bbox": [ + 277, + 157, + 350, + 165 + ], + "score": 1.0, + "content": "Step 3: score possible outputs", + "type": "text" + }, + { + "bbox": [ + 350, + 159, + 354, + 164 + ], + "score": 0.31, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 157, + 402, + 165 + ], + "score": 1.0, + "content": "using a dot product.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 277, + 157, + 402, + 165 + ] + }, + { + "type": "text", + "bbox": [ + 281, + 164, + 399, + 172 + ], + "lines": [ + { + "bbox": [ + 280, + 163, + 400, + 172 + ], + "spans": [ + { + "bbox": [ + 280, + 163, + 284, + 172 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 284, + 164, + 316, + 172 + ], + "score": 0.28, + "content": "Q _ { a b c } = \\mathbf { U } _ { c } \\cdot \\mathbf { H } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 163, + 400, + 172 + ], + "score": 1.0, + "content": ", Uc ≡ (Ec,x, Ec,y) = (cos(wkc), sin(wkc))", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 280, + 163, + 400, + 172 + ] + }, + { + "type": "image", + "bbox": [ + 181, + 172, + 500, + 264 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 181, + 172, + 500, + 264 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 181, + 172, + 500, + 264 + ], + "spans": [ + { + "bbox": [ + 181, + 172, + 500, + 264 + ], + "score": 0.823, + "type": "image", + "image_path": "a8e0022db1a64b8295bc2a10d5d9ef3d0796653ed202d0d3d29b73ded004df02.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 181, + 172, + 500, + 202.66666666666666 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 181, + 202.66666666666666, + 500, + 233.33333333333331 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 181, + 233.33333333333331, + 500, + 264.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 185, + 270, + 425, + 282 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 186, + 270, + 425, + 284 + ], + "spans": [ + { + "bbox": [ + 186, + 270, + 425, + 284 + ], + "score": 1.0, + "content": "Figure 1: Illustration of the Clock and the Pizza Algorithm.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + } + ], + "index": 16.0 + }, + { + "type": "text", + "bbox": [ + 107, + 304, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 303, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 317 + ], + "score": 1.0, + "content": "Our results highlight the complexity of mechanistic description in even models trained to perform", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "simple tasks. They point to characterization of algorithmic phase spaces, not just single algorithmic", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 326, + 369, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 369, + 339 + ], + "score": 1.0, + "content": "solutions, as an important goal in algorithm-level interpretability.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 303, + 505, + 339 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "Organization In Section 2, we review the Clock algorithm [1] and show empirical evidence of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "deviation from it in models trained to perform modular addition. In Section 3, we show that these", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 365, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 505, + 376 + ], + "score": 1.0, + "content": "deviations can be explained by an alternative Pizza algorithm. In Section 4, we define additional", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 376, + 504, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 504, + 387 + ], + "score": 1.0, + "content": "metrics to distinguish between these algorithms, and detect phase transitions between these algorithms", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "score": 1.0, + "content": "(and others Non-circular algorithms) when architectures and hyperparameters are varied. We discuss", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "the relationship between these findings and other work on model interpretation in Section 5, and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 408, + 198, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 198, + 418 + ], + "score": 1.0, + "content": "conclude in Section 6.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 342, + 505, + 418 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 434, + 358, + 448 + ], + "lines": [ + { + "bbox": [ + 104, + 432, + 359, + 452 + ], + "spans": [ + { + "bbox": [ + 104, + 432, + 359, + 452 + ], + "score": 1.0, + "content": "2 Modular Arithmetic and the Clock Algorithm", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 459, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 358, + 472 + ], + "score": 1.0, + "content": "Setup We train neural networks to perform modular addition", + "type": "text" + }, + { + "bbox": [ + 358, + 460, + 436, + 472 + ], + "score": 0.59, + "content": "a + b = c ( \\mathrm { m o d } p )", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 459, + 469, + 472 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 469, + 460, + 505, + 471 + ], + "score": 0.89, + "content": "a , b , c =", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 107, + 471, + 169, + 482 + ], + "score": 0.9, + "content": "0 , 1 , \\cdots , p - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 470, + 208, + 483 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 208, + 471, + 240, + 482 + ], + "score": 0.89, + "content": "p = 5 9", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 470, + 469, + 483 + ], + "score": 1.0, + "content": "throughout the paper. In these networks, every integer", + "type": "text" + }, + { + "bbox": [ + 470, + 472, + 474, + 480 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "has an", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 480, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 480, + 227, + 495 + ], + "score": 1.0, + "content": "associated embedding vector", + "type": "text" + }, + { + "bbox": [ + 228, + 481, + 266, + 492 + ], + "score": 0.91, + "content": "\\mathbf { E } _ { t } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 480, + 421, + 495 + ], + "score": 1.0, + "content": ". Networks take as input embeddings", + "type": "text" + }, + { + "bbox": [ + 422, + 481, + 486, + 493 + ], + "score": 0.92, + "content": "[ \\dot { \\mathbf { E } _ { a } } , \\mathbf { E } _ { b } ] ^ { \\mathbf { ^ { \\prime } } } \\in \\mathbb { R } ^ { 2 d }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 480, + 506, + 495 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 104, + 492, + 217, + 506 + ], + "score": 1.0, + "content": "predict a categorical output", + "type": "text" + }, + { + "bbox": [ + 217, + 495, + 223, + 502 + ], + "score": 0.58, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 492, + 506, + 506 + ], + "score": 1.0, + "content": ". Both embeddings and network parameters are learned. In preliminary", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 504, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 515 + ], + "score": 1.0, + "content": "experiments, we train two different network architectures on the modular arithmetic task, which we", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "score": 1.0, + "content": "refer to as: Model A and Model B. Model A is a one-layer ReLU transformer [7] with constant", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 201, + 537 + ], + "score": 1.0, + "content": "attention, while Model", + "type": "text" + }, + { + "bbox": [ + 201, + 525, + 210, + 535 + ], + "score": 0.46, + "content": "\\mathbf { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 524, + 506, + 537 + ], + "score": 1.0, + "content": "is a standard one-layer ReLU transformer (see Appendix F.1 for details).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 535, + 499, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 499, + 549 + ], + "score": 1.0, + "content": "As attention is not involved in Model A, it can also be understood as a ReLU MLP (Appendix G).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5, + "bbox_fs": [ + 104, + 459, + 506, + 549 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 560, + 261, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 262, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 262, + 575 + ], + "score": 1.0, + "content": "2.1 Review of the Clock Algorithm", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 505, + 636 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 460, + 593 + ], + "score": 1.0, + "content": "As in past work, we find that after training both Model A and Model B, embeddings", + "type": "text" + }, + { + "bbox": [ + 460, + 581, + 493, + 592 + ], + "score": 0.91, + "content": "( \\mathbf { E } _ { a } , \\mathbf { E } _ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 592, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 505, + 604 + ], + "score": 1.0, + "content": "Figure 1) usually describe a circle [8] in the plane spanned by the first two principal components of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 602, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 239, + 616 + ], + "score": 1.0, + "content": "the embedding matrix. Formally,", + "type": "text" + }, + { + "bbox": [ + 239, + 603, + 352, + 615 + ], + "score": 0.9, + "content": "\\mathbf { E } _ { a } \\approx [ \\cos ( w _ { k } a ) , \\bar { \\sin ( w _ { k } a ) } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 602, + 379, + 616 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 380, + 603, + 432, + 615 + ], + "score": 0.91, + "content": "w _ { k } = 2 \\pi k / p", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 602, + 436, + 616 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 436, + 603, + 443, + 613 + ], + "score": 0.66, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 602, + 505, + 616 + ], + "score": 1.0, + "content": "is an integer in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 613, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 107, + 614, + 145, + 626 + ], + "score": 0.89, + "content": "[ 1 , p - 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 613, + 506, + 628 + ], + "score": 1.0, + "content": ". Nanda et al. [1] discovered a circuit that uses these circular embeddings to implement an", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 624, + 437, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 437, + 638 + ], + "score": 1.0, + "content": "interpretable algorithm for modular arithmetic, which we call the Clock algorithm.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 581, + 506, + 638 + ] + }, + { + "type": "table", + "bbox": [ + 122, + 654, + 486, + 698 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 122, + 654, + 486, + 698 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 654, + 486, + 698 + ], + "spans": [ + { + "bbox": [ + 122, + 654, + 486, + 698 + ], + "score": 0.975, + "html": "
AlgorithmLearned EmbeddingsGradientSymmetryRequired Non-linearity
ClockCircleNoMultiplication
PizzaCircleYesAbsolute value
Non-circularLine,Lissajous-like curves, etc.N/AN/A
", + "type": "table", + "image_path": "2f25e792d71bb04f39c1bf014fa70aa28c6f7d41afa66156b41e29d261c38c52.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 122, + 654, + 486, + 668.6666666666666 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 122, + 668.6666666666666, + 486, + 683.3333333333333 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 122, + 683.3333333333333, + 486, + 697.9999999999999 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 190, + 701, + 421, + 712 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 189, + 700, + 421, + 713 + ], + "spans": [ + { + "bbox": [ + 189, + 700, + 421, + 713 + ], + "score": 1.0, + "content": "Table 1: Different neural algorithms for modular addition", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + } + ], + "index": 45.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 84 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 84 + ], + "score": 1.0, + "content": "\"If a meeting starts at 10, and lasts for 3 hours, then it will end at 1.\" This familiar fact is a description", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 178, + 96 + ], + "score": 1.0, + "content": "of a modular sum,", + "type": "text" + }, + { + "bbox": [ + 179, + 84, + 222, + 94 + ], + "score": 0.88, + "content": "1 0 + 3 = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "(mod 12), and the movement of a clock describes a simple algorithm for", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 394, + 108 + ], + "score": 1.0, + "content": "modular arithmetic: the numbers 1 through 12 are arranged on a circle in", + "type": "text" + }, + { + "bbox": [ + 394, + 94, + 457, + 106 + ], + "score": 0.92, + "content": "3 6 0 ^ { \\circ } / 1 2 = 3 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 93, + 507, + 108 + ], + "score": 1.0, + "content": "increments,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 105, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 146, + 118 + ], + "score": 1.0, + "content": "angles of", + "type": "text" + }, + { + "bbox": [ + 146, + 106, + 183, + 116 + ], + "score": 0.92, + "content": "1 0 \\times 3 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 105, + 201, + 118 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 202, + 106, + 235, + 116 + ], + "score": 0.92, + "content": "3 \\times 3 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 105, + 506, + 118 + ], + "score": 1.0, + "content": "are added together, then this angle is evaluated to determine that it", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 117, + 205, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 167, + 128 + ], + "score": 1.0, + "content": "corresponds to", + "type": "text" + }, + { + "bbox": [ + 168, + 117, + 200, + 127 + ], + "score": 0.91, + "content": "1 \\times 3 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 117, + 205, + 128 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 132, + 505, + 210 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "Remarkably, Nanda et al. [1] find that neural networks like our Model B implement this Clock", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 349, + 156 + ], + "score": 1.0, + "content": "algorithm, visualized in Figure 1 (left): they represent tokens", + "type": "text" + }, + { + "bbox": [ + 349, + 146, + 356, + 154 + ], + "score": 0.75, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 144, + 373, + 156 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 374, + 144, + 379, + 154 + ], + "score": 0.76, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "as 2D vectors, and adding their", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "polar angles using trigonometric identities. Concretely, the Clock algorithm consists of three steps: In", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 163, + 179 + ], + "score": 1.0, + "content": "step 1, tokens", + "type": "text" + }, + { + "bbox": [ + 163, + 167, + 169, + 175 + ], + "score": 0.75, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 165, + 187, + 179 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 187, + 166, + 193, + 175 + ], + "score": 0.67, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 165, + 262, + 179 + ], + "score": 1.0, + "content": "are embedded as", + "type": "text" + }, + { + "bbox": [ + 262, + 165, + 374, + 177 + ], + "score": 0.88, + "content": "{ \\bf E } _ { a } = [ \\cos ( \\dot { w _ { k } } a ) , \\sin ( w _ { k } a ) \\bar { ] }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 165, + 393, + 179 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 393, + 165, + 502, + 177 + ], + "score": 0.89, + "content": "\\mathbf { E } _ { b } = [ \\cos ( w _ { k } b ) , \\sin ( \\bar { w } _ { k } b ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 165, + 506, + 179 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 176, + 186, + 189 + ], + "score": 1.0, + "content": "respectively, where", + "type": "text" + }, + { + "bbox": [ + 186, + 176, + 240, + 189 + ], + "score": 0.93, + "content": "w _ { k } = 2 \\pi k / p", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 176, + 335, + 189 + ], + "score": 1.0, + "content": "(an everyday clock has", + "type": "text" + }, + { + "bbox": [ + 335, + 178, + 365, + 188 + ], + "score": 0.89, + "content": "p = 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 176, + 383, + 189 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 383, + 177, + 408, + 187 + ], + "score": 0.88, + "content": "k = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "). Then the polar angles", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 117, + 200 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 187, + 131, + 198 + ], + "score": 0.89, + "content": "\\mathbf { E } _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 187, + 150, + 200 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 150, + 188, + 163, + 198 + ], + "score": 0.89, + "content": "\\mathbf { E } _ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "are added (in step 2) and extracted (in step 3) via trigonometric identities. For each", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 196, + 465, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 175, + 212 + ], + "score": 1.0, + "content": "candidate output", + "type": "text" + }, + { + "bbox": [ + 175, + 200, + 181, + 208 + ], + "score": 0.63, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 196, + 263, + 212 + ], + "score": 1.0, + "content": ", we denote the logit", + "type": "text" + }, + { + "bbox": [ + 263, + 198, + 284, + 210 + ], + "score": 0.91, + "content": "Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 196, + 379, + 212 + ], + "score": 1.0, + "content": "; the predicted output is", + "type": "text" + }, + { + "bbox": [ + 379, + 199, + 461, + 210 + ], + "score": 0.89, + "content": "c ^ { * } = \\operatorname { a r g m a x } _ { c } Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 196, + 465, + 212 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 214, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "Crucial to this algorithm is the fact that the attention mechanism can be leveraged to perform", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "multiplication. What happens in model variants when the attention mechanism is absent, as in Model", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 236, + 452, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 452, + 249 + ], + "score": 1.0, + "content": "A? We find two pieces of evidence of deviation from the Clock algorithm in Model A.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 106, + 264, + 380, + 276 + ], + "lines": [ + { + "bbox": [ + 104, + 262, + 381, + 280 + ], + "spans": [ + { + "bbox": [ + 104, + 262, + 381, + 280 + ], + "score": 1.0, + "content": "2.2 First Evidence for Clock Violation: Gradient Symmetricity", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 255, + 298 + ], + "lines": [ + { + "bbox": [ + 106, + 285, + 255, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 255, + 299 + ], + "score": 1.0, + "content": "Since the Clock algorithm has logits:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 306, + 447, + 322 + ], + "lines": [ + { + "bbox": [ + 162, + 306, + 447, + 322 + ], + "spans": [ + { + "bbox": [ + 162, + 306, + 447, + 322 + ], + "score": 0.88, + "content": "Q _ { a b c } ^ { \\mathrm { C l o c k } } = ( { \\bf E } _ { a , x } { \\bf E } _ { b , x } - { \\bf E } _ { a , y } { \\bf E } _ { b , y } ) { \\bf E } _ { c , x } + ( { \\bf E } _ { a , x } { \\bf E } _ { b , y } + { \\bf E } _ { a , y } { \\bf E } _ { b , x } ) { \\bf E } _ { c , y } ,", + "type": "interline_equation", + "image_path": "6fed1a6631949119b31389b7a5afdee812b60f0e29cd7b57f05276bcd07b88e0.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 162, + 306, + 447, + 322 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 330, + 505, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 507, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 237, + 343 + ], + "score": 1.0, + "content": "(see Figure 1) the gradients of", + "type": "text" + }, + { + "bbox": [ + 237, + 331, + 258, + 342 + ], + "score": 0.91, + "content": "Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 330, + 507, + 343 + ], + "score": 1.0, + "content": "generically lack permutation symmetry in argument order:", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 340, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 204, + 353 + ], + "score": 0.9, + "content": "\\nabla _ { \\mathbf { E } _ { a } } Q _ { a b c } \\neq \\nabla _ { \\mathbf { E } _ { b } } Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 340, + 447, + 355 + ], + "score": 1.0, + "content": ". Thus, if learned models exhibit permutation symmetry", + "type": "text" + }, + { + "bbox": [ + 447, + 341, + 505, + 353 + ], + "score": 0.89, + "content": "( \\nabla _ { \\mathbf { E } _ { a } } Q _ { a b c } ~ =", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 351, + 354, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 149, + 364 + ], + "score": 0.91, + "content": "\\nabla _ { \\mathbf { E } _ { b } } Q _ { a b c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 351, + 354, + 365 + ], + "score": 1.0, + "content": ", they must be implementing some other algorithm.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 368, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 381 + ], + "score": 1.0, + "content": "We compute the 6 largest principal components of the input embedding vectors. We then compute the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "gradients of output logits (unnormalized log-probabilities from the model) with respect to the input", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "embeddings. We then project them onto these 6 principal components (since the angles relevant to the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "score": 1.0, + "content": "Clock and Pizza algorithms are encoded in the first few principal components). These projections are", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 411, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 425 + ], + "score": 1.0, + "content": "shown in Figure 2. While Model B demonstrates asymmetry in general, Model A exhibits gradient", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 423, + 151, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 151, + 437 + ], + "score": 1.0, + "content": "symmetry.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + }, + { + "type": "image", + "bbox": [ + 168, + 446, + 443, + 568 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 168, + 446, + 443, + 568 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 168, + 446, + 443, + 568 + ], + "spans": [ + { + "bbox": [ + 168, + 446, + 443, + 568 + ], + "score": 0.969, + "type": "image", + "image_path": "47a0c498b5770d8bfa4c34124b0a6ec340016648a966d6625a5affe4d90c69f0.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 168, + 446, + 443, + 486.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 168, + 486.6666666666667, + 443, + 527.3333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 168, + 527.3333333333334, + 443, + 568.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 576, + 505, + 610 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 403, + 589 + ], + "score": 1.0, + "content": "Figure 2: Gradients on first six principal components of input embeddings.", + "type": "text" + }, + { + "bbox": [ + 403, + 576, + 434, + 588 + ], + "score": 0.93, + "content": "( a , b , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "in the title stands", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 267, + 600 + ], + "score": 1.0, + "content": "for taking gradients on the output logit", + "type": "text" + }, + { + "bbox": [ + 268, + 589, + 274, + 597 + ], + "score": 0.68, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 586, + 313, + 600 + ], + "score": 1.0, + "content": "for input", + "type": "text" + }, + { + "bbox": [ + 314, + 587, + 336, + 599 + ], + "score": 0.92, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 586, + 505, + 600 + ], + "score": 1.0, + "content": ". x and y axes represent the gradients for", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 597, + 501, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 355, + 611 + ], + "score": 1.0, + "content": "embeddings of the first and the second token. The dashed line", + "type": "text" + }, + { + "bbox": [ + 356, + 600, + 381, + 609 + ], + "score": 0.92, + "content": "y = x", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 597, + 501, + 611 + ], + "score": 1.0, + "content": "signals a symmetric gradient.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 107, + 634, + 352, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 353, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 353, + 650 + ], + "score": 1.0, + "content": "2.3 Second Evidence for Clock Violation: Logit Patterns", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "Inspecting models’ outputs, in addition to inputs, reveals further differences. For each input pair", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 666, + 507, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 128, + 679 + ], + "score": 0.91, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 666, + 366, + 681 + ], + "score": 1.0, + "content": ", we compute the output logit assigned to the correct label", + "type": "text" + }, + { + "bbox": [ + 367, + 667, + 390, + 678 + ], + "score": 0.89, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 666, + 507, + 681 + ], + "score": 1.0, + "content": ". We visualize these correct", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 413, + 690 + ], + "score": 1.0, + "content": "logits from Models A and B in Figure 3. Notice that the rows are indexed by", + "type": "text" + }, + { + "bbox": [ + 413, + 678, + 436, + 689 + ], + "score": 0.9, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "and the columns", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 687, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 119, + 702 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 689, + 143, + 699 + ], + "score": 0.9, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 687, + 505, + 702 + ], + "score": 1.0, + "content": ". From Figure 3, we can see that the correct logits of Model A have a clear dependency", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 119, + 712 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 700, + 143, + 711 + ], + "score": 0.87, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "in that within each row, the correct logits are roughly the same, while this pattern is not", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 710, + 494, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 494, + 723 + ], + "score": 1.0, + "content": "observed in Model B. This suggests that Models A and B are implementing different algorithms.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 84 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 84 + ], + "score": 1.0, + "content": "\"If a meeting starts at 10, and lasts for 3 hours, then it will end at 1.\" This familiar fact is a description", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 178, + 96 + ], + "score": 1.0, + "content": "of a modular sum,", + "type": "text" + }, + { + "bbox": [ + 179, + 84, + 222, + 94 + ], + "score": 0.88, + "content": "1 0 + 3 = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "(mod 12), and the movement of a clock describes a simple algorithm for", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 394, + 108 + ], + "score": 1.0, + "content": "modular arithmetic: the numbers 1 through 12 are arranged on a circle in", + "type": "text" + }, + { + "bbox": [ + 394, + 94, + 457, + 106 + ], + "score": 0.92, + "content": "3 6 0 ^ { \\circ } / 1 2 = 3 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 93, + 507, + 108 + ], + "score": 1.0, + "content": "increments,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 105, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 146, + 118 + ], + "score": 1.0, + "content": "angles of", + "type": "text" + }, + { + "bbox": [ + 146, + 106, + 183, + 116 + ], + "score": 0.92, + "content": "1 0 \\times 3 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 105, + 201, + 118 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 202, + 106, + 235, + 116 + ], + "score": 0.92, + "content": "3 \\times 3 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 105, + 506, + 118 + ], + "score": 1.0, + "content": "are added together, then this angle is evaluated to determine that it", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 117, + 205, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 167, + 128 + ], + "score": 1.0, + "content": "corresponds to", + "type": "text" + }, + { + "bbox": [ + 168, + 117, + 200, + 127 + ], + "score": 0.91, + "content": "1 \\times 3 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 117, + 205, + 128 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 72, + 507, + 128 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 132, + 505, + 210 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "Remarkably, Nanda et al. [1] find that neural networks like our Model B implement this Clock", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 349, + 156 + ], + "score": 1.0, + "content": "algorithm, visualized in Figure 1 (left): they represent tokens", + "type": "text" + }, + { + "bbox": [ + 349, + 146, + 356, + 154 + ], + "score": 0.75, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 144, + 373, + 156 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 374, + 144, + 379, + 154 + ], + "score": 0.76, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "as 2D vectors, and adding their", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "polar angles using trigonometric identities. Concretely, the Clock algorithm consists of three steps: In", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 163, + 179 + ], + "score": 1.0, + "content": "step 1, tokens", + "type": "text" + }, + { + "bbox": [ + 163, + 167, + 169, + 175 + ], + "score": 0.75, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 165, + 187, + 179 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 187, + 166, + 193, + 175 + ], + "score": 0.67, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 165, + 262, + 179 + ], + "score": 1.0, + "content": "are embedded as", + "type": "text" + }, + { + "bbox": [ + 262, + 165, + 374, + 177 + ], + "score": 0.88, + "content": "{ \\bf E } _ { a } = [ \\cos ( \\dot { w _ { k } } a ) , \\sin ( w _ { k } a ) \\bar { ] }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 165, + 393, + 179 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 393, + 165, + 502, + 177 + ], + "score": 0.89, + "content": "\\mathbf { E } _ { b } = [ \\cos ( w _ { k } b ) , \\sin ( \\bar { w } _ { k } b ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 165, + 506, + 179 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 176, + 186, + 189 + ], + "score": 1.0, + "content": "respectively, where", + "type": "text" + }, + { + "bbox": [ + 186, + 176, + 240, + 189 + ], + "score": 0.93, + "content": "w _ { k } = 2 \\pi k / p", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 176, + 335, + 189 + ], + "score": 1.0, + "content": "(an everyday clock has", + "type": "text" + }, + { + "bbox": [ + 335, + 178, + 365, + 188 + ], + "score": 0.89, + "content": "p = 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 176, + 383, + 189 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 383, + 177, + 408, + 187 + ], + "score": 0.88, + "content": "k = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "). Then the polar angles", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 117, + 200 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 187, + 131, + 198 + ], + "score": 0.89, + "content": "\\mathbf { E } _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 187, + 150, + 200 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 150, + 188, + 163, + 198 + ], + "score": 0.89, + "content": "\\mathbf { E } _ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "are added (in step 2) and extracted (in step 3) via trigonometric identities. For each", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 196, + 465, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 175, + 212 + ], + "score": 1.0, + "content": "candidate output", + "type": "text" + }, + { + "bbox": [ + 175, + 200, + 181, + 208 + ], + "score": 0.63, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 196, + 263, + 212 + ], + "score": 1.0, + "content": ", we denote the logit", + "type": "text" + }, + { + "bbox": [ + 263, + 198, + 284, + 210 + ], + "score": 0.91, + "content": "Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 196, + 379, + 212 + ], + "score": 1.0, + "content": "; the predicted output is", + "type": "text" + }, + { + "bbox": [ + 379, + 199, + 461, + 210 + ], + "score": 0.89, + "content": "c ^ { * } = \\operatorname { a r g m a x } _ { c } Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 196, + 465, + 212 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8, + "bbox_fs": [ + 104, + 132, + 506, + 212 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 214, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "Crucial to this algorithm is the fact that the attention mechanism can be leveraged to perform", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "multiplication. What happens in model variants when the attention mechanism is absent, as in Model", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 236, + 452, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 452, + 249 + ], + "score": 1.0, + "content": "A? We find two pieces of evidence of deviation from the Clock algorithm in Model A.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 214, + 505, + 249 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 264, + 380, + 276 + ], + "lines": [ + { + "bbox": [ + 104, + 262, + 381, + 280 + ], + "spans": [ + { + "bbox": [ + 104, + 262, + 381, + 280 + ], + "score": 1.0, + "content": "2.2 First Evidence for Clock Violation: Gradient Symmetricity", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 255, + 298 + ], + "lines": [ + { + "bbox": [ + 106, + 285, + 255, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 255, + 299 + ], + "score": 1.0, + "content": "Since the Clock algorithm has logits:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 285, + 255, + 299 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 306, + 447, + 322 + ], + "lines": [ + { + "bbox": [ + 162, + 306, + 447, + 322 + ], + "spans": [ + { + "bbox": [ + 162, + 306, + 447, + 322 + ], + "score": 0.88, + "content": "Q _ { a b c } ^ { \\mathrm { C l o c k } } = ( { \\bf E } _ { a , x } { \\bf E } _ { b , x } - { \\bf E } _ { a , y } { \\bf E } _ { b , y } ) { \\bf E } _ { c , x } + ( { \\bf E } _ { a , x } { \\bf E } _ { b , y } + { \\bf E } _ { a , y } { \\bf E } _ { b , x } ) { \\bf E } _ { c , y } ,", + "type": "interline_equation", + "image_path": "6fed1a6631949119b31389b7a5afdee812b60f0e29cd7b57f05276bcd07b88e0.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 162, + 306, + 447, + 322 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 330, + 505, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 507, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 237, + 343 + ], + "score": 1.0, + "content": "(see Figure 1) the gradients of", + "type": "text" + }, + { + "bbox": [ + 237, + 331, + 258, + 342 + ], + "score": 0.91, + "content": "Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 330, + 507, + 343 + ], + "score": 1.0, + "content": "generically lack permutation symmetry in argument order:", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 340, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 204, + 353 + ], + "score": 0.9, + "content": "\\nabla _ { \\mathbf { E } _ { a } } Q _ { a b c } \\neq \\nabla _ { \\mathbf { E } _ { b } } Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 340, + 447, + 355 + ], + "score": 1.0, + "content": ". Thus, if learned models exhibit permutation symmetry", + "type": "text" + }, + { + "bbox": [ + 447, + 341, + 505, + 353 + ], + "score": 0.89, + "content": "( \\nabla _ { \\mathbf { E } _ { a } } Q _ { a b c } ~ =", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 351, + 354, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 149, + 364 + ], + "score": 0.91, + "content": "\\nabla _ { \\mathbf { E } _ { b } } Q _ { a b c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 351, + 354, + 365 + ], + "score": 1.0, + "content": ", they must be implementing some other algorithm.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 330, + 507, + 365 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 368, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 381 + ], + "score": 1.0, + "content": "We compute the 6 largest principal components of the input embedding vectors. We then compute the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "gradients of output logits (unnormalized log-probabilities from the model) with respect to the input", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "embeddings. We then project them onto these 6 principal components (since the angles relevant to the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "score": 1.0, + "content": "Clock and Pizza algorithms are encoded in the first few principal components). These projections are", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 411, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 425 + ], + "score": 1.0, + "content": "shown in Figure 2. While Model B demonstrates asymmetry in general, Model A exhibits gradient", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 423, + 151, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 151, + 437 + ], + "score": 1.0, + "content": "symmetry.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 367, + 506, + 437 + ] + }, + { + "type": "image", + "bbox": [ + 168, + 446, + 443, + 568 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 168, + 446, + 443, + 568 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 168, + 446, + 443, + 568 + ], + "spans": [ + { + "bbox": [ + 168, + 446, + 443, + 568 + ], + "score": 0.969, + "type": "image", + "image_path": "47a0c498b5770d8bfa4c34124b0a6ec340016648a966d6625a5affe4d90c69f0.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 168, + 446, + 443, + 486.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 168, + 486.6666666666667, + 443, + 527.3333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 168, + 527.3333333333334, + 443, + 568.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 576, + 505, + 610 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 403, + 589 + ], + "score": 1.0, + "content": "Figure 2: Gradients on first six principal components of input embeddings.", + "type": "text" + }, + { + "bbox": [ + 403, + 576, + 434, + 588 + ], + "score": 0.93, + "content": "( a , b , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "in the title stands", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 267, + 600 + ], + "score": 1.0, + "content": "for taking gradients on the output logit", + "type": "text" + }, + { + "bbox": [ + 268, + 589, + 274, + 597 + ], + "score": 0.68, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 586, + 313, + 600 + ], + "score": 1.0, + "content": "for input", + "type": "text" + }, + { + "bbox": [ + 314, + 587, + 336, + 599 + ], + "score": 0.92, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 586, + 505, + 600 + ], + "score": 1.0, + "content": ". x and y axes represent the gradients for", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 597, + 501, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 355, + 611 + ], + "score": 1.0, + "content": "embeddings of the first and the second token. The dashed line", + "type": "text" + }, + { + "bbox": [ + 356, + 600, + 381, + 609 + ], + "score": 0.92, + "content": "y = x", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 597, + 501, + 611 + ], + "score": 1.0, + "content": "signals a symmetric gradient.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 107, + 634, + 352, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 353, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 353, + 650 + ], + "score": 1.0, + "content": "2.3 Second Evidence for Clock Violation: Logit Patterns", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "Inspecting models’ outputs, in addition to inputs, reveals further differences. For each input pair", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 666, + 507, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 128, + 679 + ], + "score": 0.91, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 666, + 366, + 681 + ], + "score": 1.0, + "content": ", we compute the output logit assigned to the correct label", + "type": "text" + }, + { + "bbox": [ + 367, + 667, + 390, + 678 + ], + "score": 0.89, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 666, + 507, + 681 + ], + "score": 1.0, + "content": ". We visualize these correct", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 413, + 690 + ], + "score": 1.0, + "content": "logits from Models A and B in Figure 3. Notice that the rows are indexed by", + "type": "text" + }, + { + "bbox": [ + 413, + 678, + 436, + 689 + ], + "score": 0.9, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "and the columns", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 687, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 119, + 702 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 689, + 143, + 699 + ], + "score": 0.9, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 687, + 505, + 702 + ], + "score": 1.0, + "content": ". From Figure 3, we can see that the correct logits of Model A have a clear dependency", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 119, + 712 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 700, + 143, + 711 + ], + "score": 0.87, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "in that within each row, the correct logits are roughly the same, while this pattern is not", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 710, + 494, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 494, + 723 + ], + "score": 1.0, + "content": "observed in Model B. This suggests that Models A and B are implementing different algorithms.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 656, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 72, + 485, + 193 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 72, + 485, + 193 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 72, + 485, + 193 + ], + "spans": [ + { + "bbox": [ + 128, + 72, + 485, + 193 + ], + "score": 0.964, + "type": "image", + "image_path": "8b77cd016ee8b0bb8159304cbcdbfe4da16e61f7d8d6ec157f44a9e89eb71835.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 128, + 72, + 485, + 112.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 112.33333333333334, + 485, + 152.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 152.66666666666669, + 485, + 193.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 201, + 505, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "Figure 3: Correct Logits of Model A & Model B. The correct logits of Model A (left) have a clear", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 212, + 350, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 168, + 225 + ], + "score": 1.0, + "content": "dependence on", + "type": "text" + }, + { + "bbox": [ + 169, + 213, + 191, + 223 + ], + "score": 0.88, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 212, + 350, + 225 + ], + "score": 1.0, + "content": ", while those of Model B (right) do not.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 106, + 248, + 356, + 262 + ], + "lines": [ + { + "bbox": [ + 104, + 246, + 357, + 265 + ], + "spans": [ + { + "bbox": [ + 104, + 246, + 357, + 265 + ], + "score": 1.0, + "content": "3 An Alternative Solution: the Pizza Algorithm", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 276, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 287 + ], + "score": 1.0, + "content": "How does Model A perform modular arithmetic? Whatever solution it implements must exhibit", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "gradient symmetricity in Figure 2 and the output patterns in Figure 3. In this section, we describe a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 298, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 310 + ], + "score": 1.0, + "content": "new algorithm for modular arithmetic, which we call the Pizza algorithm, and then provide evidence", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 309, + 315, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 315, + 321 + ], + "score": 1.0, + "content": "that this is the procedure implemented by Model A.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 108, + 337, + 217, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 219, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 219, + 352 + ], + "score": 1.0, + "content": "3.1 The Pizza Algorithm", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 360, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 359, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 372 + ], + "score": 1.0, + "content": "Unlike the Clock algorithm, the Pizza algorithm operates inside the circle formed by embeddings", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "(just as pepperoni are spread all over a pizza), instead of operating on the circumference of the circle.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 344, + 395 + ], + "score": 1.0, + "content": "The basic idea is illustrated in Figure 1: given a fixed label", + "type": "text" + }, + { + "bbox": [ + 344, + 384, + 349, + 392 + ], + "score": 0.66, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 380, + 381, + 395 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 381, + 381, + 403, + 393 + ], + "score": 0.91, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 380, + 425, + 395 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 425, + 382, + 466, + 392 + ], + "score": 0.62, + "content": "a + b = c", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 380, + 492, + 395 + ], + "score": 1.0, + "content": "(mod", + "type": "text" + }, + { + "bbox": [ + 492, + 383, + 499, + 393 + ], + "score": 0.3, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 499, + 380, + 506, + 395 + ], + "score": 1.0, + "content": "),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 391, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 147, + 406 + ], + "score": 1.0, + "content": "the points", + "type": "text" + }, + { + "bbox": [ + 148, + 392, + 231, + 405 + ], + "score": 0.92, + "content": "{ \\bf E } _ { a b } = ( { \\bf E } _ { a } + { \\bf E } _ { b } ) / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 391, + 506, + 406 + ], + "score": 1.0, + "content": "lie on a line though the origin of a 2D plane, and the points closer to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 321, + 417 + ], + "score": 1.0, + "content": "this line than to the lines corresponding to any other", + "type": "text" + }, + { + "bbox": [ + 321, + 406, + 327, + 413 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 403, + 394, + 417 + ], + "score": 1.0, + "content": "form two out of", + "type": "text" + }, + { + "bbox": [ + 395, + 404, + 406, + 415 + ], + "score": 0.85, + "content": "2 p", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "mirrored “pizza slices”,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "as shown at the right of the figure. Thus, to perform modular arithmetic, a network can determine", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 438, + 438 + ], + "score": 1.0, + "content": "which slice pair the average of the two embedding vectors lies in. Concretely, the", + "type": "text" + }, + { + "bbox": [ + 438, + 426, + 446, + 435 + ], + "score": 0.28, + "content": "P", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "izza algorithm", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 436, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 457, + 448 + ], + "score": 1.0, + "content": "also consists of three steps. Step 1 is the same as in the Clock algorithm: the tokens", + "type": "text" + }, + { + "bbox": [ + 457, + 438, + 464, + 446 + ], + "score": 0.71, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 436, + 483, + 448 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 483, + 437, + 489, + 446 + ], + "score": 0.66, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 436, + 506, + 448 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 159, + 460 + ], + "score": 1.0, + "content": "embedded at", + "type": "text" + }, + { + "bbox": [ + 160, + 447, + 273, + 459 + ], + "score": 0.9, + "content": "\\mathbf { E } _ { a } = ( \\cos ( \\bar { w } _ { k } a ) , \\bar { \\sin } ( w _ { k } a ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 446, + 293, + 460 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 293, + 447, + 403, + 459 + ], + "score": 0.9, + "content": "\\mathbf { E } _ { b } = ( \\cos ( w _ { k } b ) , \\sin ( w _ { k } b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 446, + 506, + 460 + ], + "score": 1.0, + "content": ", respectively. Step 2 and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 346, + 471 + ], + "score": 1.0, + "content": "Step 3 are different from the Clock algorithm. In Step 2.1,", + "type": "text" + }, + { + "bbox": [ + 347, + 459, + 360, + 469 + ], + "score": 0.85, + "content": "\\mathbf { E } _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 457, + 379, + 471 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 379, + 459, + 392, + 469 + ], + "score": 0.87, + "content": "\\mathbf { E } _ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "are averaged to produce an", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 467, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 152, + 483 + ], + "score": 1.0, + "content": "embedding", + "type": "text" + }, + { + "bbox": [ + 152, + 469, + 169, + 480 + ], + "score": 0.89, + "content": "\\mathbf { E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 467, + 334, + 483 + ], + "score": 1.0, + "content": ". In Step 2.2 and Step 3, the polar angle of", + "type": "text" + }, + { + "bbox": [ + 334, + 469, + 351, + 480 + ], + "score": 0.9, + "content": "{ \\bf E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 467, + 506, + 483 + ], + "score": 1.0, + "content": "is (implicitly) computed by computing", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 478, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 142, + 493 + ], + "score": 1.0, + "content": "the logit", + "type": "text" + }, + { + "bbox": [ + 143, + 480, + 164, + 491 + ], + "score": 0.91, + "content": "Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 478, + 265, + 493 + ], + "score": 1.0, + "content": "for any possible outputs", + "type": "text" + }, + { + "bbox": [ + 266, + 482, + 271, + 489 + ], + "score": 0.69, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 478, + 506, + 493 + ], + "score": 1.0, + "content": ". While one possibility of doing so is to take the absolute", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 490, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 221, + 504 + ], + "score": 1.0, + "content": "value of the dot product of", + "type": "text" + }, + { + "bbox": [ + 221, + 491, + 238, + 502 + ], + "score": 0.9, + "content": "{ \\bf E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 490, + 262, + 504 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 262, + 491, + 367, + 503 + ], + "score": 0.81, + "content": "( \\cos ( w _ { k } c / 2 ) , \\sin ( w _ { k } c / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 490, + 506, + 504 + ], + "score": 1.0, + "content": ", it is not commonly observed in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 461, + 514 + ], + "score": 1.0, + "content": "neural networks (and will result in a different logit pattern). Instead, Step 2.2 transforms", + "type": "text" + }, + { + "bbox": [ + 462, + 502, + 479, + 513 + ], + "score": 0.9, + "content": "{ \\bf E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "into a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 511, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 171, + 525 + ], + "score": 1.0, + "content": "vector encoding", + "type": "text" + }, + { + "bbox": [ + 171, + 512, + 385, + 524 + ], + "score": 0.83, + "content": "| \\cos ( w _ { k } ( a - b ) / 2 ) | ( \\cos ( w _ { k } ( a + b ) \\bar { ) } , \\sin ( w _ { k } ( a + b ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 511, + 506, + 525 + ], + "score": 1.0, + "content": ", which is then dotted with the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 523, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 183, + 537 + ], + "score": 1.0, + "content": "output embedding", + "type": "text" + }, + { + "bbox": [ + 184, + 524, + 295, + 535 + ], + "score": 0.8, + "content": "U _ { c } = ( \\cos ( w _ { k } c ) , \\sin ( w _ { k } c ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 523, + 402, + 537 + ], + "score": 1.0, + "content": ". Finally, the prediction is", + "type": "text" + }, + { + "bbox": [ + 402, + 524, + 483, + 535 + ], + "score": 0.9, + "content": "c ^ { * } = \\operatorname { a r g m a x } _ { c } Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 523, + 506, + 537 + ], + "score": 1.0, + "content": ". See", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 534, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 215, + 547 + ], + "score": 1.0, + "content": "Appendix A and Appendix", + "type": "text" + }, + { + "bbox": [ + 216, + 536, + 223, + 544 + ], + "score": 0.35, + "content": "\\mathrm { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 534, + 468, + 547 + ], + "score": 1.0, + "content": "for a more detailed analysis of a neural circuit that computes", + "type": "text" + }, + { + "bbox": [ + 468, + 535, + 487, + 545 + ], + "score": 0.89, + "content": "{ \\mathbf { H } } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 534, + 506, + 547 + ], + "score": 1.0, + "content": "in a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 545, + 162, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 162, + 557 + ], + "score": 1.0, + "content": "real network.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "The key difference between the two algorithms lies in what non-linear operations are required: Clock", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 573, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 506, + 585 + ], + "score": 1.0, + "content": "requires multiplication of inputs in Step 2, while Pizza requires only absolute value computation,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "which is easily implemented by the ReLU layers. If neural networks lack inductive biases toward", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "implementing multiplication, they may be more likely to implement Pizza rather than Clock, as we", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 606, + 203, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 203, + 618 + ], + "score": 1.0, + "content": "will verify in Section 4.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 107, + 634, + 299, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 300, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 300, + 649 + ], + "score": 1.0, + "content": "3.2 First Evidence for Pizza: Logit Patterns", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 320, + 669 + ], + "score": 1.0, + "content": "Both the Clock and Pizza algorithms compute logits", + "type": "text" + }, + { + "bbox": [ + 321, + 657, + 342, + 668 + ], + "score": 0.91, + "content": "Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "in Step 3, but they have different forms,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 666, + 504, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 237, + 681 + ], + "score": 1.0, + "content": "shown in Figure 1. Specifically,", + "type": "text" + }, + { + "bbox": [ + 237, + 667, + 289, + 679 + ], + "score": 0.92, + "content": "Q _ { a b c } ( P i z z a )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 666, + 425, + 681 + ], + "score": 1.0, + "content": "has an extra multiplicative factor", + "type": "text" + }, + { + "bbox": [ + 425, + 667, + 504, + 679 + ], + "score": 0.95, + "content": "\\vert \\cos ( w _ { k } ( a - b ) / 2 ) \\vert", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 678, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 160, + 691 + ], + "score": 1.0, + "content": "compared to", + "type": "text" + }, + { + "bbox": [ + 160, + 678, + 213, + 690 + ], + "score": 0.91, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 678, + 294, + 691 + ], + "score": 1.0, + "content": ". As a result, given", + "type": "text" + }, + { + "bbox": [ + 294, + 678, + 336, + 689 + ], + "score": 0.78, + "content": "c = a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 678, + 340, + 691 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 341, + 678, + 393, + 690 + ], + "score": 0.89, + "content": "Q _ { a b c } ( P i z z a )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 678, + 462, + 691 + ], + "score": 1.0, + "content": "is dependent on", + "type": "text" + }, + { + "bbox": [ + 462, + 679, + 485, + 689 + ], + "score": 0.87, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 678, + 506, + 691 + ], + "score": 1.0, + "content": ", but", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 107, + 689, + 159, + 700 + ], + "score": 0.9, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "is not. The intuition for the dependence is that a sample is more likely to be classified", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 153, + 712 + ], + "score": 1.0, + "content": "correctly if", + "type": "text" + }, + { + "bbox": [ + 153, + 700, + 170, + 711 + ], + "score": 0.93, + "content": "{ \\bf E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 699, + 354, + 712 + ], + "score": 1.0, + "content": "is longer. The norm of this vector depends on", + "type": "text" + }, + { + "bbox": [ + 354, + 700, + 377, + 711 + ], + "score": 0.89, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 699, + 505, + 712 + ], + "score": 1.0, + "content": ". As we observe in Figure 3, the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 710, + 362, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 335, + 723 + ], + "score": 1.0, + "content": "logits in Model A indeed exhibit a strong dependence on", + "type": "text" + }, + { + "bbox": [ + 335, + 711, + 358, + 721 + ], + "score": 0.89, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 710, + 362, + 723 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + } + ], + "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": [ + 128, + 72, + 485, + 193 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 72, + 485, + 193 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 72, + 485, + 193 + ], + "spans": [ + { + "bbox": [ + 128, + 72, + 485, + 193 + ], + "score": 0.964, + "type": "image", + "image_path": "8b77cd016ee8b0bb8159304cbcdbfe4da16e61f7d8d6ec157f44a9e89eb71835.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 128, + 72, + 485, + 112.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 112.33333333333334, + 485, + 152.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 152.66666666666669, + 485, + 193.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 201, + 505, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "Figure 3: Correct Logits of Model A & Model B. The correct logits of Model A (left) have a clear", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 212, + 350, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 168, + 225 + ], + "score": 1.0, + "content": "dependence on", + "type": "text" + }, + { + "bbox": [ + 169, + 213, + 191, + 223 + ], + "score": 0.88, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 212, + 350, + 225 + ], + "score": 1.0, + "content": ", while those of Model B (right) do not.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 106, + 248, + 356, + 262 + ], + "lines": [ + { + "bbox": [ + 104, + 246, + 357, + 265 + ], + "spans": [ + { + "bbox": [ + 104, + 246, + 357, + 265 + ], + "score": 1.0, + "content": "3 An Alternative Solution: the Pizza Algorithm", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 276, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 287 + ], + "score": 1.0, + "content": "How does Model A perform modular arithmetic? Whatever solution it implements must exhibit", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "gradient symmetricity in Figure 2 and the output patterns in Figure 3. In this section, we describe a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 298, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 310 + ], + "score": 1.0, + "content": "new algorithm for modular arithmetic, which we call the Pizza algorithm, and then provide evidence", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 309, + 315, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 315, + 321 + ], + "score": 1.0, + "content": "that this is the procedure implemented by Model A.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 276, + 505, + 321 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 337, + 217, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 219, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 219, + 352 + ], + "score": 1.0, + "content": "3.1 The Pizza Algorithm", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 360, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 359, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 372 + ], + "score": 1.0, + "content": "Unlike the Clock algorithm, the Pizza algorithm operates inside the circle formed by embeddings", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "(just as pepperoni are spread all over a pizza), instead of operating on the circumference of the circle.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 344, + 395 + ], + "score": 1.0, + "content": "The basic idea is illustrated in Figure 1: given a fixed label", + "type": "text" + }, + { + "bbox": [ + 344, + 384, + 349, + 392 + ], + "score": 0.66, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 380, + 381, + 395 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 381, + 381, + 403, + 393 + ], + "score": 0.91, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 380, + 425, + 395 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 425, + 382, + 466, + 392 + ], + "score": 0.62, + "content": "a + b = c", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 380, + 492, + 395 + ], + "score": 1.0, + "content": "(mod", + "type": "text" + }, + { + "bbox": [ + 492, + 383, + 499, + 393 + ], + "score": 0.3, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 499, + 380, + 506, + 395 + ], + "score": 1.0, + "content": "),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 391, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 147, + 406 + ], + "score": 1.0, + "content": "the points", + "type": "text" + }, + { + "bbox": [ + 148, + 392, + 231, + 405 + ], + "score": 0.92, + "content": "{ \\bf E } _ { a b } = ( { \\bf E } _ { a } + { \\bf E } _ { b } ) / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 391, + 506, + 406 + ], + "score": 1.0, + "content": "lie on a line though the origin of a 2D plane, and the points closer to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 321, + 417 + ], + "score": 1.0, + "content": "this line than to the lines corresponding to any other", + "type": "text" + }, + { + "bbox": [ + 321, + 406, + 327, + 413 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 403, + 394, + 417 + ], + "score": 1.0, + "content": "form two out of", + "type": "text" + }, + { + "bbox": [ + 395, + 404, + 406, + 415 + ], + "score": 0.85, + "content": "2 p", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "mirrored “pizza slices”,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "as shown at the right of the figure. Thus, to perform modular arithmetic, a network can determine", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 438, + 438 + ], + "score": 1.0, + "content": "which slice pair the average of the two embedding vectors lies in. Concretely, the", + "type": "text" + }, + { + "bbox": [ + 438, + 426, + 446, + 435 + ], + "score": 0.28, + "content": "P", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "izza algorithm", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 436, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 457, + 448 + ], + "score": 1.0, + "content": "also consists of three steps. Step 1 is the same as in the Clock algorithm: the tokens", + "type": "text" + }, + { + "bbox": [ + 457, + 438, + 464, + 446 + ], + "score": 0.71, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 436, + 483, + 448 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 483, + 437, + 489, + 446 + ], + "score": 0.66, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 436, + 506, + 448 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 159, + 460 + ], + "score": 1.0, + "content": "embedded at", + "type": "text" + }, + { + "bbox": [ + 160, + 447, + 273, + 459 + ], + "score": 0.9, + "content": "\\mathbf { E } _ { a } = ( \\cos ( \\bar { w } _ { k } a ) , \\bar { \\sin } ( w _ { k } a ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 446, + 293, + 460 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 293, + 447, + 403, + 459 + ], + "score": 0.9, + "content": "\\mathbf { E } _ { b } = ( \\cos ( w _ { k } b ) , \\sin ( w _ { k } b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 446, + 506, + 460 + ], + "score": 1.0, + "content": ", respectively. Step 2 and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 346, + 471 + ], + "score": 1.0, + "content": "Step 3 are different from the Clock algorithm. In Step 2.1,", + "type": "text" + }, + { + "bbox": [ + 347, + 459, + 360, + 469 + ], + "score": 0.85, + "content": "\\mathbf { E } _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 457, + 379, + 471 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 379, + 459, + 392, + 469 + ], + "score": 0.87, + "content": "\\mathbf { E } _ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "are averaged to produce an", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 467, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 152, + 483 + ], + "score": 1.0, + "content": "embedding", + "type": "text" + }, + { + "bbox": [ + 152, + 469, + 169, + 480 + ], + "score": 0.89, + "content": "\\mathbf { E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 467, + 334, + 483 + ], + "score": 1.0, + "content": ". In Step 2.2 and Step 3, the polar angle of", + "type": "text" + }, + { + "bbox": [ + 334, + 469, + 351, + 480 + ], + "score": 0.9, + "content": "{ \\bf E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 467, + 506, + 483 + ], + "score": 1.0, + "content": "is (implicitly) computed by computing", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 478, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 142, + 493 + ], + "score": 1.0, + "content": "the logit", + "type": "text" + }, + { + "bbox": [ + 143, + 480, + 164, + 491 + ], + "score": 0.91, + "content": "Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 478, + 265, + 493 + ], + "score": 1.0, + "content": "for any possible outputs", + "type": "text" + }, + { + "bbox": [ + 266, + 482, + 271, + 489 + ], + "score": 0.69, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 478, + 506, + 493 + ], + "score": 1.0, + "content": ". While one possibility of doing so is to take the absolute", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 490, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 221, + 504 + ], + "score": 1.0, + "content": "value of the dot product of", + "type": "text" + }, + { + "bbox": [ + 221, + 491, + 238, + 502 + ], + "score": 0.9, + "content": "{ \\bf E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 490, + 262, + 504 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 262, + 491, + 367, + 503 + ], + "score": 0.81, + "content": "( \\cos ( w _ { k } c / 2 ) , \\sin ( w _ { k } c / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 490, + 506, + 504 + ], + "score": 1.0, + "content": ", it is not commonly observed in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 461, + 514 + ], + "score": 1.0, + "content": "neural networks (and will result in a different logit pattern). Instead, Step 2.2 transforms", + "type": "text" + }, + { + "bbox": [ + 462, + 502, + 479, + 513 + ], + "score": 0.9, + "content": "{ \\bf E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "into a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 511, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 171, + 525 + ], + "score": 1.0, + "content": "vector encoding", + "type": "text" + }, + { + "bbox": [ + 171, + 512, + 385, + 524 + ], + "score": 0.83, + "content": "| \\cos ( w _ { k } ( a - b ) / 2 ) | ( \\cos ( w _ { k } ( a + b ) \\bar { ) } , \\sin ( w _ { k } ( a + b ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 511, + 506, + 525 + ], + "score": 1.0, + "content": ", which is then dotted with the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 523, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 183, + 537 + ], + "score": 1.0, + "content": "output embedding", + "type": "text" + }, + { + "bbox": [ + 184, + 524, + 295, + 535 + ], + "score": 0.8, + "content": "U _ { c } = ( \\cos ( w _ { k } c ) , \\sin ( w _ { k } c ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 523, + 402, + 537 + ], + "score": 1.0, + "content": ". Finally, the prediction is", + "type": "text" + }, + { + "bbox": [ + 402, + 524, + 483, + 535 + ], + "score": 0.9, + "content": "c ^ { * } = \\operatorname { a r g m a x } _ { c } Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 523, + 506, + 537 + ], + "score": 1.0, + "content": ". See", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 534, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 215, + 547 + ], + "score": 1.0, + "content": "Appendix A and Appendix", + "type": "text" + }, + { + "bbox": [ + 216, + 536, + 223, + 544 + ], + "score": 0.35, + "content": "\\mathrm { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 534, + 468, + 547 + ], + "score": 1.0, + "content": "for a more detailed analysis of a neural circuit that computes", + "type": "text" + }, + { + "bbox": [ + 468, + 535, + 487, + 545 + ], + "score": 0.89, + "content": "{ \\mathbf { H } } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 534, + 506, + 547 + ], + "score": 1.0, + "content": "in a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 545, + 162, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 162, + 557 + ], + "score": 1.0, + "content": "real network.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 359, + 506, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "The key difference between the two algorithms lies in what non-linear operations are required: Clock", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 573, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 506, + 585 + ], + "score": 1.0, + "content": "requires multiplication of inputs in Step 2, while Pizza requires only absolute value computation,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "which is easily implemented by the ReLU layers. If neural networks lack inductive biases toward", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "implementing multiplication, they may be more likely to implement Pizza rather than Clock, as we", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 606, + 203, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 203, + 618 + ], + "score": 1.0, + "content": "will verify in Section 4.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 561, + 506, + 618 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 634, + 299, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 300, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 300, + 649 + ], + "score": 1.0, + "content": "3.2 First Evidence for Pizza: Logit Patterns", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 320, + 669 + ], + "score": 1.0, + "content": "Both the Clock and Pizza algorithms compute logits", + "type": "text" + }, + { + "bbox": [ + 321, + 657, + 342, + 668 + ], + "score": 0.91, + "content": "Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "in Step 3, but they have different forms,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 666, + 504, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 237, + 681 + ], + "score": 1.0, + "content": "shown in Figure 1. Specifically,", + "type": "text" + }, + { + "bbox": [ + 237, + 667, + 289, + 679 + ], + "score": 0.92, + "content": "Q _ { a b c } ( P i z z a )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 666, + 425, + 681 + ], + "score": 1.0, + "content": "has an extra multiplicative factor", + "type": "text" + }, + { + "bbox": [ + 425, + 667, + 504, + 679 + ], + "score": 0.95, + "content": "\\vert \\cos ( w _ { k } ( a - b ) / 2 ) \\vert", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 678, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 160, + 691 + ], + "score": 1.0, + "content": "compared to", + "type": "text" + }, + { + "bbox": [ + 160, + 678, + 213, + 690 + ], + "score": 0.91, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 678, + 294, + 691 + ], + "score": 1.0, + "content": ". As a result, given", + "type": "text" + }, + { + "bbox": [ + 294, + 678, + 336, + 689 + ], + "score": 0.78, + "content": "c = a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 678, + 340, + 691 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 341, + 678, + 393, + 690 + ], + "score": 0.89, + "content": "Q _ { a b c } ( P i z z a )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 678, + 462, + 691 + ], + "score": 1.0, + "content": "is dependent on", + "type": "text" + }, + { + "bbox": [ + 462, + 679, + 485, + 689 + ], + "score": 0.87, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 678, + 506, + 691 + ], + "score": 1.0, + "content": ", but", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 107, + 689, + 159, + 700 + ], + "score": 0.9, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "is not. The intuition for the dependence is that a sample is more likely to be classified", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 153, + 712 + ], + "score": 1.0, + "content": "correctly if", + "type": "text" + }, + { + "bbox": [ + 153, + 700, + 170, + 711 + ], + "score": 0.93, + "content": "{ \\bf E } _ { a b }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 699, + 354, + 712 + ], + "score": 1.0, + "content": "is longer. The norm of this vector depends on", + "type": "text" + }, + { + "bbox": [ + 354, + 700, + 377, + 711 + ], + "score": 0.89, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 699, + 505, + 712 + ], + "score": 1.0, + "content": ". As we observe in Figure 3, the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 710, + 362, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 335, + 723 + ], + "score": 1.0, + "content": "logits in Model A indeed exhibit a strong dependence on", + "type": "text" + }, + { + "bbox": [ + 335, + 711, + 358, + 721 + ], + "score": 0.89, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 710, + 362, + 723 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 656, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 72, + 426, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 428, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 428, + 86 + ], + "score": 1.0, + "content": "3.3 Second Evidence for Pizza: Clearer Logit Patterns via Circle Isolation", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 99, + 505, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 98, + 505, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 443, + 111 + ], + "score": 1.0, + "content": "To better understand the behavior of this algorithm, we replace the embedding matrix", + "type": "text" + }, + { + "bbox": [ + 444, + 99, + 452, + 109 + ], + "score": 0.65, + "content": "\\mathbf { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 98, + 505, + 111 + ], + "score": 1.0, + "content": "with a series", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 109, + 506, + 122 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 506, + 122 + ], + "score": 1.0, + "content": "of rank-2 approximations: using only the first and second principal components, or only the third", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 119, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 505, + 133 + ], + "score": 1.0, + "content": "and fourth, etc. For each such matrix, embeddings lie in a a two-dimensional subspace. For both", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "Model A and Model B, we find that embeddings form a circle in this subspace (Figure 4 and Figure 5,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 142, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 506, + 154 + ], + "score": 1.0, + "content": "bottom). We call this procedure circle isolation. Even after this drastic modification to the trained", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 165 + ], + "score": 1.0, + "content": "models’ parameters, both Model A and Model B continue to behave in interpretable ways: a subset", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 471, + 177 + ], + "score": 1.0, + "content": "of predictions remain highly accurate, with this subset determined by the periodicity of the", + "type": "text" + }, + { + "bbox": [ + 471, + 164, + 478, + 174 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 164, + 506, + 177 + ], + "score": 1.0, + "content": "of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "score": 1.0, + "content": "isolated circle. As predicted by the Pizza and Clock algorithms described in Figure 1, Model A’s", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 284, + 198 + ], + "score": 1.0, + "content": "accuracy drops to zero at specific values of", + "type": "text" + }, + { + "bbox": [ + 285, + 186, + 308, + 196 + ], + "score": 0.9, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 186, + 479, + 198 + ], + "score": 1.0, + "content": ", while Model B’s accuracy is invariant in", + "type": "text" + }, + { + "bbox": [ + 480, + 186, + 503, + 196 + ], + "score": 0.91, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 186, + 506, + 198 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 197, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 210 + ], + "score": 1.0, + "content": "Applying circle isolation to Model A on the two principal components (one circle) yields a model", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 207, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 126, + 222 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 126, + 208, + 154, + 218 + ], + "score": 0.89, + "content": "3 2 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 207, + 505, + 222 + ], + "score": 1.0, + "content": "overall accuracy, while retaining the first six principal components (three circles) yields", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 196, + 231 + ], + "score": 1.0, + "content": "an overall accuracy of", + "type": "text" + }, + { + "bbox": [ + 197, + 218, + 224, + 229 + ], + "score": 0.9, + "content": "9 1 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 218, + 506, + 231 + ], + "score": 1.0, + "content": ". See Appendix D for more discussion. By contrast, Model B achieves", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 228, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 131, + 240 + ], + "score": 0.89, + "content": "1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 228, + 505, + 242 + ], + "score": 1.0, + "content": "when embeddings are truncated to the first six principal components. Circle isolation thus", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "reveals an error correction mechanism achieved via ensembling: when an algorithm (clock or pizza)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "score": 1.0, + "content": "exhibits systematic errors on subset of inputs, models can implement multiple algorithm variants in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 262, + 275, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 275, + 275 + ], + "score": 1.0, + "content": "parallel to obtain more robust predictions.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 8.5 + }, + { + "type": "image", + "bbox": [ + 110, + 298, + 501, + 498 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 298, + 501, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 298, + 501, + 498 + ], + "spans": [ + { + "bbox": [ + 110, + 298, + 501, + 498 + ], + "score": 0.976, + "type": "image", + "image_path": "a4c16929c8d49f65231b30e3beaaf0cbf51a22e43b383cb260880db63d85c7c4.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 110, + 298, + 501, + 364.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 110, + 364.6666666666667, + 501, + 431.33333333333337 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 110, + 431.33333333333337, + 501, + 498.00000000000006 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 505, + 506, + 538 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 504, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 247, + 520 + ], + "score": 1.0, + "content": "Figure 4: Correct logits of Model A", + "type": "text" + }, + { + "bbox": [ + 248, + 506, + 275, + 516 + ], + "score": 0.44, + "content": "( P i z z a )", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 504, + 506, + 520 + ], + "score": 1.0, + "content": "after circle isolation. The rightmost pizza is accompanying", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 480, + 529 + ], + "score": 1.0, + "content": "the third pizza (discussed in Section 3.4 and Appendix D). Top: The logit pattern depends on", + "type": "text" + }, + { + "bbox": [ + 480, + 517, + 503, + 527 + ], + "score": 0.89, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 516, + 506, + 529 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 528, + 255, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 255, + 538 + ], + "score": 1.0, + "content": "Bottom: Embeddings for each circle.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "Using these isolated embeddings, we may additionally calculate the isolated logits directly with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 572, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 586 + ], + "score": 1.0, + "content": "formulas in Figure 1 and compare with the actual logits from Model A. Results are displayed in Table", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 583, + 446, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 168, + 596 + ], + "score": 1.0, + "content": "2. We find that", + "type": "text" + }, + { + "bbox": [ + 168, + 583, + 220, + 595 + ], + "score": 0.92, + "content": "Q _ { a b c } ( P i z z a )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 583, + 389, + 596 + ], + "score": 1.0, + "content": "explains substantially more variance than", + "type": "text" + }, + { + "bbox": [ + 389, + 583, + 442, + 595 + ], + "score": 0.92, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 583, + 446, + 596 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 379, + 613 + ], + "score": 1.0, + "content": "Why do we only analyze correct logits? The logits from the", + "type": "text" + }, + { + "bbox": [ + 379, + 601, + 403, + 610 + ], + "score": 0.27, + "content": "P i z z a", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 598, + 505, + 613 + ], + "score": 1.0, + "content": "algorithm are given by", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 108, + 609, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 108, + 610, + 309, + 622 + ], + "score": 0.38, + "content": "Q _ { a b c } ( P i z z a ) = | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a +", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 609, + 506, + 624 + ], + "score": 1.0, + "content": "b − c)). By contrast, the Clock algorithm has", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 132, + 634 + ], + "score": 1.0, + "content": "logits", + "type": "text" + }, + { + "bbox": [ + 133, + 622, + 288, + 633 + ], + "score": 0.87, + "content": "Q _ { a b c } ( C l o c k ) = \\cos ( w _ { k } ( a + b - c ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 621, + 348, + 634 + ], + "score": 1.0, + "content": ". In a word,", + "type": "text" + }, + { + "bbox": [ + 349, + 622, + 401, + 633 + ], + "score": 0.92, + "content": "Q _ { a b c } ( P i z z a )", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "has an extra multiplica-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 154, + 645 + ], + "score": 1.0, + "content": "tive factor", + "type": "text" + }, + { + "bbox": [ + 154, + 633, + 237, + 644 + ], + "score": 0.87, + "content": "\\vert \\cos ( w _ { k } ( a - b ) / 2 ) \\vert", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 632, + 296, + 645 + ], + "score": 1.0, + "content": "compared to", + "type": "text" + }, + { + "bbox": [ + 296, + 632, + 349, + 644 + ], + "score": 0.9, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 632, + 429, + 645 + ], + "score": 1.0, + "content": ". By constraining", + "type": "text" + }, + { + "bbox": [ + 430, + 633, + 480, + 643 + ], + "score": 0.91, + "content": "c = a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "(thus", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 642, + 404, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 207, + 656 + ], + "score": 0.9, + "content": "\\cos ( w _ { k } ( a + b - c ) ) = 1 \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 642, + 253, + 656 + ], + "score": 1.0, + "content": ", the factor", + "type": "text" + }, + { + "bbox": [ + 253, + 644, + 332, + 656 + ], + "score": 0.91, + "content": "\\vert \\cos ( w _ { k } ( a - b ) / 2 ) \\vert", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 642, + 404, + 656 + ], + "score": 1.0, + "content": "can be identified.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 505, + 726 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 255, + 672 + ], + "score": 1.0, + "content": "(Unexpected) dependence of logits", + "type": "text" + }, + { + "bbox": [ + 255, + 659, + 308, + 672 + ], + "score": 0.78, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 659, + 323, + 672 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 323, + 660, + 347, + 671 + ], + "score": 0.56, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 659, + 505, + 672 + ], + "score": 1.0, + "content": ": Although our analysis above expects", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 131, + 684 + ], + "score": 1.0, + "content": "logits", + "type": "text" + }, + { + "bbox": [ + 131, + 671, + 184, + 682 + ], + "score": 0.92, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 671, + 254, + 684 + ], + "score": 1.0, + "content": "not to depend on", + "type": "text" + }, + { + "bbox": [ + 254, + 672, + 277, + 681 + ], + "score": 0.88, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 671, + 430, + 684 + ], + "score": 1.0, + "content": ", they do not predict its dependence on", + "type": "text" + }, + { + "bbox": [ + 430, + 671, + 453, + 681 + ], + "score": 0.89, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 671, + 506, + 684 + ], + "score": 1.0, + "content": ". In Figure 5,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 205, + 695 + ], + "score": 1.0, + "content": "we surprisingly find that", + "type": "text" + }, + { + "bbox": [ + 205, + 682, + 257, + 693 + ], + "score": 0.91, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 681, + 506, + 695 + ], + "score": 1.0, + "content": "is sensitive to this sum. Our conjecture is that Step 1 and Step", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 693, + 504, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 504, + 704 + ], + "score": 1.0, + "content": "2 of the Clock are implemented (almost) noiselessly, such that same-label samples collapse to the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 702, + 506, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 506, + 717 + ], + "score": 1.0, + "content": "same point after Step 2. However, Step 3 (classification) is imperfect after circle isolation, resulting", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 714, + 436, + 727 + ], + "spans": [ + { + "bbox": [ + 105, + 714, + 312, + 727 + ], + "score": 1.0, + "content": "in fluctuations of logits. Inputs with common sums", + "type": "text" + }, + { + "bbox": [ + 312, + 715, + 335, + 725 + ], + "score": 0.88, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 714, + 436, + 727 + ], + "score": 1.0, + "content": "produce the same logits.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + } + ], + "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": "title", + "bbox": [ + 106, + 72, + 426, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 428, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 428, + 86 + ], + "score": 1.0, + "content": "3.3 Second Evidence for Pizza: Clearer Logit Patterns via Circle Isolation", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 99, + 505, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 98, + 505, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 443, + 111 + ], + "score": 1.0, + "content": "To better understand the behavior of this algorithm, we replace the embedding matrix", + "type": "text" + }, + { + "bbox": [ + 444, + 99, + 452, + 109 + ], + "score": 0.65, + "content": "\\mathbf { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 98, + 505, + 111 + ], + "score": 1.0, + "content": "with a series", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 109, + 506, + 122 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 506, + 122 + ], + "score": 1.0, + "content": "of rank-2 approximations: using only the first and second principal components, or only the third", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 119, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 505, + 133 + ], + "score": 1.0, + "content": "and fourth, etc. For each such matrix, embeddings lie in a a two-dimensional subspace. For both", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "Model A and Model B, we find that embeddings form a circle in this subspace (Figure 4 and Figure 5,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 142, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 506, + 154 + ], + "score": 1.0, + "content": "bottom). We call this procedure circle isolation. Even after this drastic modification to the trained", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 165 + ], + "score": 1.0, + "content": "models’ parameters, both Model A and Model B continue to behave in interpretable ways: a subset", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 471, + 177 + ], + "score": 1.0, + "content": "of predictions remain highly accurate, with this subset determined by the periodicity of the", + "type": "text" + }, + { + "bbox": [ + 471, + 164, + 478, + 174 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 164, + 506, + 177 + ], + "score": 1.0, + "content": "of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "score": 1.0, + "content": "isolated circle. As predicted by the Pizza and Clock algorithms described in Figure 1, Model A’s", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 284, + 198 + ], + "score": 1.0, + "content": "accuracy drops to zero at specific values of", + "type": "text" + }, + { + "bbox": [ + 285, + 186, + 308, + 196 + ], + "score": 0.9, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 186, + 479, + 198 + ], + "score": 1.0, + "content": ", while Model B’s accuracy is invariant in", + "type": "text" + }, + { + "bbox": [ + 480, + 186, + 503, + 196 + ], + "score": 0.91, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 186, + 506, + 198 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 197, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 210 + ], + "score": 1.0, + "content": "Applying circle isolation to Model A on the two principal components (one circle) yields a model", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 207, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 126, + 222 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 126, + 208, + 154, + 218 + ], + "score": 0.89, + "content": "3 2 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 207, + 505, + 222 + ], + "score": 1.0, + "content": "overall accuracy, while retaining the first six principal components (three circles) yields", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 196, + 231 + ], + "score": 1.0, + "content": "an overall accuracy of", + "type": "text" + }, + { + "bbox": [ + 197, + 218, + 224, + 229 + ], + "score": 0.9, + "content": "9 1 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 218, + 506, + 231 + ], + "score": 1.0, + "content": ". See Appendix D for more discussion. By contrast, Model B achieves", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 228, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 131, + 240 + ], + "score": 0.89, + "content": "1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 228, + 505, + 242 + ], + "score": 1.0, + "content": "when embeddings are truncated to the first six principal components. Circle isolation thus", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "reveals an error correction mechanism achieved via ensembling: when an algorithm (clock or pizza)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "score": 1.0, + "content": "exhibits systematic errors on subset of inputs, models can implement multiple algorithm variants in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 262, + 275, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 275, + 275 + ], + "score": 1.0, + "content": "parallel to obtain more robust predictions.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 98, + 506, + 275 + ] + }, + { + "type": "image", + "bbox": [ + 110, + 298, + 501, + 498 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 298, + 501, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 298, + 501, + 498 + ], + "spans": [ + { + "bbox": [ + 110, + 298, + 501, + 498 + ], + "score": 0.976, + "type": "image", + "image_path": "a4c16929c8d49f65231b30e3beaaf0cbf51a22e43b383cb260880db63d85c7c4.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 110, + 298, + 501, + 364.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 110, + 364.6666666666667, + 501, + 431.33333333333337 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 110, + 431.33333333333337, + 501, + 498.00000000000006 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 505, + 506, + 538 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 504, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 247, + 520 + ], + "score": 1.0, + "content": "Figure 4: Correct logits of Model A", + "type": "text" + }, + { + "bbox": [ + 248, + 506, + 275, + 516 + ], + "score": 0.44, + "content": "( P i z z a )", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 504, + 506, + 520 + ], + "score": 1.0, + "content": "after circle isolation. The rightmost pizza is accompanying", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 480, + 529 + ], + "score": 1.0, + "content": "the third pizza (discussed in Section 3.4 and Appendix D). Top: The logit pattern depends on", + "type": "text" + }, + { + "bbox": [ + 480, + 517, + 503, + 527 + ], + "score": 0.89, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 516, + 506, + 529 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 528, + 255, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 255, + 538 + ], + "score": 1.0, + "content": "Bottom: Embeddings for each circle.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "Using these isolated embeddings, we may additionally calculate the isolated logits directly with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 572, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 586 + ], + "score": 1.0, + "content": "formulas in Figure 1 and compare with the actual logits from Model A. Results are displayed in Table", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 583, + 446, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 168, + 596 + ], + "score": 1.0, + "content": "2. We find that", + "type": "text" + }, + { + "bbox": [ + 168, + 583, + 220, + 595 + ], + "score": 0.92, + "content": "Q _ { a b c } ( P i z z a )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 583, + 389, + 596 + ], + "score": 1.0, + "content": "explains substantially more variance than", + "type": "text" + }, + { + "bbox": [ + 389, + 583, + 442, + 595 + ], + "score": 0.92, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 583, + 446, + 596 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 560, + 505, + 596 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 379, + 613 + ], + "score": 1.0, + "content": "Why do we only analyze correct logits? The logits from the", + "type": "text" + }, + { + "bbox": [ + 379, + 601, + 403, + 610 + ], + "score": 0.27, + "content": "P i z z a", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 598, + 505, + 613 + ], + "score": 1.0, + "content": "algorithm are given by", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 108, + 609, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 108, + 610, + 309, + 622 + ], + "score": 0.38, + "content": "Q _ { a b c } ( P i z z a ) = | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a +", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 609, + 506, + 624 + ], + "score": 1.0, + "content": "b − c)). By contrast, the Clock algorithm has", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 132, + 634 + ], + "score": 1.0, + "content": "logits", + "type": "text" + }, + { + "bbox": [ + 133, + 622, + 288, + 633 + ], + "score": 0.87, + "content": "Q _ { a b c } ( C l o c k ) = \\cos ( w _ { k } ( a + b - c ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 621, + 348, + 634 + ], + "score": 1.0, + "content": ". In a word,", + "type": "text" + }, + { + "bbox": [ + 349, + 622, + 401, + 633 + ], + "score": 0.92, + "content": "Q _ { a b c } ( P i z z a )", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "has an extra multiplica-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 154, + 645 + ], + "score": 1.0, + "content": "tive factor", + "type": "text" + }, + { + "bbox": [ + 154, + 633, + 237, + 644 + ], + "score": 0.87, + "content": "\\vert \\cos ( w _ { k } ( a - b ) / 2 ) \\vert", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 632, + 296, + 645 + ], + "score": 1.0, + "content": "compared to", + "type": "text" + }, + { + "bbox": [ + 296, + 632, + 349, + 644 + ], + "score": 0.9, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 632, + 429, + 645 + ], + "score": 1.0, + "content": ". By constraining", + "type": "text" + }, + { + "bbox": [ + 430, + 633, + 480, + 643 + ], + "score": 0.91, + "content": "c = a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "(thus", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 642, + 404, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 207, + 656 + ], + "score": 0.9, + "content": "\\cos ( w _ { k } ( a + b - c ) ) = 1 \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 642, + 253, + 656 + ], + "score": 1.0, + "content": ", the factor", + "type": "text" + }, + { + "bbox": [ + 253, + 644, + 332, + 656 + ], + "score": 0.91, + "content": "\\vert \\cos ( w _ { k } ( a - b ) / 2 ) \\vert", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 642, + 404, + 656 + ], + "score": 1.0, + "content": "can be identified.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 598, + 506, + 656 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 505, + 726 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 255, + 672 + ], + "score": 1.0, + "content": "(Unexpected) dependence of logits", + "type": "text" + }, + { + "bbox": [ + 255, + 659, + 308, + 672 + ], + "score": 0.78, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 659, + 323, + 672 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 323, + 660, + 347, + 671 + ], + "score": 0.56, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 659, + 505, + 672 + ], + "score": 1.0, + "content": ": Although our analysis above expects", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 131, + 684 + ], + "score": 1.0, + "content": "logits", + "type": "text" + }, + { + "bbox": [ + 131, + 671, + 184, + 682 + ], + "score": 0.92, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 671, + 254, + 684 + ], + "score": 1.0, + "content": "not to depend on", + "type": "text" + }, + { + "bbox": [ + 254, + 672, + 277, + 681 + ], + "score": 0.88, + "content": "a - b", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 671, + 430, + 684 + ], + "score": 1.0, + "content": ", they do not predict its dependence on", + "type": "text" + }, + { + "bbox": [ + 430, + 671, + 453, + 681 + ], + "score": 0.89, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 671, + 506, + 684 + ], + "score": 1.0, + "content": ". In Figure 5,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 205, + 695 + ], + "score": 1.0, + "content": "we surprisingly find that", + "type": "text" + }, + { + "bbox": [ + 205, + 682, + 257, + 693 + ], + "score": 0.91, + "content": "Q _ { a b c } ( C l o c k )", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 681, + 506, + 695 + ], + "score": 1.0, + "content": "is sensitive to this sum. Our conjecture is that Step 1 and Step", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 693, + 504, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 504, + 704 + ], + "score": 1.0, + "content": "2 of the Clock are implemented (almost) noiselessly, such that same-label samples collapse to the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 702, + 506, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 506, + 717 + ], + "score": 1.0, + "content": "same point after Step 2. However, Step 3 (classification) is imperfect after circle isolation, resulting", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 714, + 436, + 727 + ], + "spans": [ + { + "bbox": [ + 105, + 714, + 312, + 727 + ], + "score": 1.0, + "content": "in fluctuations of logits. Inputs with common sums", + "type": "text" + }, + { + "bbox": [ + 312, + 715, + 335, + 725 + ], + "score": 0.88, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 714, + 436, + 727 + ], + "score": 1.0, + "content": "produce the same logits.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 659, + 506, + 727 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 83, + 503, + 273 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 83, + 503, + 273 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 83, + 503, + 273 + ], + "spans": [ + { + "bbox": [ + 109, + 83, + 503, + 273 + ], + "score": 0.961, + "type": "image", + "image_path": "262aacc85fad06c0956a5c94d9416eca3a007771e7088076425c9761096eab86.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 83, + 503, + 146.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 146.33333333333334, + 503, + 209.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 209.66666666666669, + 503, + 273.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 103, + 280, + 505, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 279, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 506, + 293 + ], + "score": 1.0, + "content": "Figure 5: Correct logits of Model B (Clock) after circle isolation. Top: The logit pattern depends on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 291, + 283, + 303 + ], + "spans": [ + { + "bbox": [ + 107, + 291, + 129, + 302 + ], + "score": 0.87, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 291, + 283, + 303 + ], + "score": 1.0, + "content": ". Bottom: Embeddings for each circle.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 311, + 438, + 362 + ], + "lines": [ + { + "bbox": [ + 172, + 311, + 438, + 362 + ], + "spans": [ + { + "bbox": [ + 172, + 311, + 438, + 362 + ], + "score": 0.28, + "content": "\\frac { \\mathrm { C i r c l e ~ \\left| ~ \\right.} w _ { k } } { \\# 1 } \\frac { w _ { k } } { 2 \\pi / 5 9 \\cdot 1 7 ~ \\frac { \\ d / { \\operatorname { Q } _ { a b c } ( \\mathrm { c l o c k } ) } \\mathrm { F V E } } { 7 5 . 4 1 \\% } \\frac { \\ d / { Q _ { a b c } ( \\mathrm { p i z z a } ) } \\mathrm { F V E } } { 9 9 . 1 8 \\% } }", + "type": "interline_equation", + "image_path": "31ed86340a77685732e7e9363ad4e39c29c6ea847776269b7ccfee7e7580835a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 172, + 311, + 438, + 328.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 172, + 328.0, + 438, + 345.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 172, + 345.0, + 438, + 362.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "Table 2: After isolating circles in the input embedding, fraction of variance explained (FVE) of all", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 378, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 205, + 389 + ], + "score": 1.0, + "content": "Model A’s output logits", + "type": "text" + }, + { + "bbox": [ + 205, + 378, + 262, + 389 + ], + "score": 0.93, + "content": "( 5 9 \\times 5 9 \\times 5 9", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 378, + 505, + 389 + ], + "score": 1.0, + "content": "of them) by various formulas. Both model output logits and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 388, + 402 + ], + "score": 1.0, + "content": "formula results’ are normalized to mean 0 variance 1 before taking FVE.", + "type": "text" + }, + { + "bbox": [ + 388, + 390, + 401, + 400 + ], + "score": 0.87, + "content": "w _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "’s are calculated according", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 399, + 504, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 430, + 411 + ], + "score": 1.0, + "content": "to the visualization. For example, distance between 0 and 1 in Circle #1 is 17, so", + "type": "text" + }, + { + "bbox": [ + 430, + 399, + 500, + 412 + ], + "score": 0.92, + "content": "w _ { k } = 2 \\pi / 5 9 \\cdot 1 7", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 400, + 504, + 411 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 442, + 402, + 454 + ], + "lines": [ + { + "bbox": [ + 104, + 439, + 404, + 457 + ], + "spans": [ + { + "bbox": [ + 104, + 439, + 404, + 457 + ], + "score": 1.0, + "content": "3.4 Third Evidence for Pizza: Accompanied & Accompanying Pizza", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 423, + 477 + ], + "score": 1.0, + "content": "The Achilles’ heel of the Pizza algorithm is antipodal pairs. If two inputs", + "type": "text" + }, + { + "bbox": [ + 423, + 464, + 445, + 475 + ], + "score": 0.91, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 463, + 506, + 477 + ], + "score": 1.0, + "content": "happen to lie", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "antipodally, then their middle point will lie at the origin, where the correct “pizza slice” is difficult to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "identify. For example in Figure 1 right, antipodal pairs are (1,7), (2,8), (3,9) etc., whose middle points", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "all collapse to the origin, but their class labels are different. Models cannot distinguish between,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 324, + 520 + ], + "score": 1.0, + "content": "and thus correctly classify, these pairs. Even for odd", + "type": "text" + }, + { + "bbox": [ + 324, + 508, + 331, + 519 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "’s where there are no strict antipodal pairs,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "approximately antipodal pairs are also more likely to be classified incorrectly than non-antipodal", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 531, + 131, + 541 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 131, + 541 + ], + "score": 1.0, + "content": "pairs.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 545, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "Intriguingly, neural networks find a clever way to compensate for this failure mode. we find that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 104, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "pizzas usually come with “accompanying pizzas”. An accompanied pizza and its accompanying", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "pizza complement each other in the sense that near-antipodal pairs in the accompanied pizza become", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "adjacent or close (i.e, very non-antipodal) in the accompanying pizza. If we denote the difference", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 278, + 602 + ], + "score": 1.0, + "content": "between adjacent numbers on the circle as", + "type": "text" + }, + { + "bbox": [ + 279, + 590, + 285, + 599 + ], + "score": 0.76, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 589, + 303, + 602 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 303, + 589, + 313, + 600 + ], + "score": 0.82, + "content": "\\delta _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 589, + 316, + 602 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 317, + 589, + 327, + 600 + ], + "score": 0.8, + "content": "\\delta _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "for accompanied and accompanying pizzas,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 179, + 613 + ], + "score": 1.0, + "content": "respectively, then", + "type": "text" + }, + { + "bbox": [ + 179, + 600, + 217, + 611 + ], + "score": 0.89, + "content": "\\delta _ { 1 } = 2 \\delta _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 600, + 243, + 613 + ], + "score": 1.0, + "content": "(mod", + "type": "text" + }, + { + "bbox": [ + 243, + 601, + 250, + 612 + ], + "score": 0.39, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "). In the experiment, we found that pizzas #1/#2/#3 in Figure 4", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 610, + 507, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 507, + 625 + ], + "score": 1.0, + "content": "all have accompanying pizzas, which we call pizzas #4/#5/#6 (see Appendix D for details). However,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 620, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 104, + 620, + 506, + 636 + ], + "score": 1.0, + "content": "these accompanying pizzas do not play a significant role in final model predictions 2. We conjecture", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "that training dynamics are as follows: (1) At initialization, pizzas #1/#2/#3 correspond to three", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "different “lottery tickets” [9]. (2) In early stages of training, to compensate the weaknesses (antipodal", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "pairs) of pizzas #1/#2/#3, pizzas #4/#5/#6 are formed. (3) As training goes on (in the presence of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "weight decay), the neural network gets pruned. As a result, pizzas #4/#5/#6 are not significantly", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 676, + 442, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 442, + 690 + ], + "score": 1.0, + "content": "involved in prediction, although they continue to be visible in the embedding space.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 26 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 701, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 118, + 700, + 280, + 713 + ], + "score": 1.0, + "content": "2Accompanied pizzas #1/#2/#3 can achieve", + "type": "text" + }, + { + "bbox": [ + 280, + 702, + 305, + 711 + ], + "score": 0.78, + "content": "9 9 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "accuracy, but accompanying pizzas #4/#5/#6 can only", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 710, + 197, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 136, + 723 + ], + "score": 1.0, + "content": "achieve", + "type": "text" + }, + { + "bbox": [ + 136, + 712, + 160, + 721 + ], + "score": 0.67, + "content": "1 6 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 710, + 197, + 723 + ], + "score": 1.0, + "content": "accuracy.", + "type": "text" + } + ] + } + ] + }, + { + "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" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 237, + 73, + 371, + 80 + ], + "lines": [ + { + "bbox": [ + 237, + 72, + 372, + 81 + ], + "spans": [ + { + "bbox": [ + 237, + 72, + 372, + 81 + ], + "score": 0.976, + "content": "Correct Logits of Model B after Circle Isolation", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 83, + 503, + 273 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 83, + 503, + 273 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 83, + 503, + 273 + ], + "spans": [ + { + "bbox": [ + 109, + 83, + 503, + 273 + ], + "score": 0.961, + "type": "image", + "image_path": "262aacc85fad06c0956a5c94d9416eca3a007771e7088076425c9761096eab86.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 83, + 503, + 146.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 146.33333333333334, + 503, + 209.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 209.66666666666669, + 503, + 273.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 103, + 280, + 505, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 279, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 506, + 293 + ], + "score": 1.0, + "content": "Figure 5: Correct logits of Model B (Clock) after circle isolation. Top: The logit pattern depends on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 291, + 283, + 303 + ], + "spans": [ + { + "bbox": [ + 107, + 291, + 129, + 302 + ], + "score": 0.87, + "content": "a + b", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 291, + 283, + 303 + ], + "score": 1.0, + "content": ". Bottom: Embeddings for each circle.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 311, + 438, + 362 + ], + "lines": [ + { + "bbox": [ + 172, + 311, + 438, + 362 + ], + "spans": [ + { + "bbox": [ + 172, + 311, + 438, + 362 + ], + "score": 0.28, + "content": "\\frac { \\mathrm { C i r c l e ~ \\left| ~ \\right.} w _ { k } } { \\# 1 } \\frac { w _ { k } } { 2 \\pi / 5 9 \\cdot 1 7 ~ \\frac { \\ d / { \\operatorname { Q } _ { a b c } ( \\mathrm { c l o c k } ) } \\mathrm { F V E } } { 7 5 . 4 1 \\% } \\frac { \\ d / { Q _ { a b c } ( \\mathrm { p i z z a } ) } \\mathrm { F V E } } { 9 9 . 1 8 \\% } }", + "type": "interline_equation", + "image_path": "31ed86340a77685732e7e9363ad4e39c29c6ea847776269b7ccfee7e7580835a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 172, + 311, + 438, + 328.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 172, + 328.0, + 438, + 345.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 172, + 345.0, + 438, + 362.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "Table 2: After isolating circles in the input embedding, fraction of variance explained (FVE) of all", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 378, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 205, + 389 + ], + "score": 1.0, + "content": "Model A’s output logits", + "type": "text" + }, + { + "bbox": [ + 205, + 378, + 262, + 389 + ], + "score": 0.93, + "content": "( 5 9 \\times 5 9 \\times 5 9", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 378, + 505, + 389 + ], + "score": 1.0, + "content": "of them) by various formulas. Both model output logits and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 388, + 402 + ], + "score": 1.0, + "content": "formula results’ are normalized to mean 0 variance 1 before taking FVE.", + "type": "text" + }, + { + "bbox": [ + 388, + 390, + 401, + 400 + ], + "score": 0.87, + "content": "w _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "’s are calculated according", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 399, + 504, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 430, + 411 + ], + "score": 1.0, + "content": "to the visualization. For example, distance between 0 and 1 in Circle #1 is 17, so", + "type": "text" + }, + { + "bbox": [ + 430, + 399, + 500, + 412 + ], + "score": 0.92, + "content": "w _ { k } = 2 \\pi / 5 9 \\cdot 1 7", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 400, + 504, + 411 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 366, + 506, + 412 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 442, + 402, + 454 + ], + "lines": [ + { + "bbox": [ + 104, + 439, + 404, + 457 + ], + "spans": [ + { + "bbox": [ + 104, + 439, + 404, + 457 + ], + "score": 1.0, + "content": "3.4 Third Evidence for Pizza: Accompanied & Accompanying Pizza", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 423, + 477 + ], + "score": 1.0, + "content": "The Achilles’ heel of the Pizza algorithm is antipodal pairs. If two inputs", + "type": "text" + }, + { + "bbox": [ + 423, + 464, + 445, + 475 + ], + "score": 0.91, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 463, + 506, + 477 + ], + "score": 1.0, + "content": "happen to lie", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "antipodally, then their middle point will lie at the origin, where the correct “pizza slice” is difficult to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "identify. For example in Figure 1 right, antipodal pairs are (1,7), (2,8), (3,9) etc., whose middle points", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "all collapse to the origin, but their class labels are different. Models cannot distinguish between,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 324, + 520 + ], + "score": 1.0, + "content": "and thus correctly classify, these pairs. Even for odd", + "type": "text" + }, + { + "bbox": [ + 324, + 508, + 331, + 519 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "’s where there are no strict antipodal pairs,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "approximately antipodal pairs are also more likely to be classified incorrectly than non-antipodal", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 531, + 131, + 541 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 131, + 541 + ], + "score": 1.0, + "content": "pairs.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 104, + 463, + 506, + 541 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 545, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "Intriguingly, neural networks find a clever way to compensate for this failure mode. we find that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 104, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "pizzas usually come with “accompanying pizzas”. An accompanied pizza and its accompanying", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "pizza complement each other in the sense that near-antipodal pairs in the accompanied pizza become", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "adjacent or close (i.e, very non-antipodal) in the accompanying pizza. If we denote the difference", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 278, + 602 + ], + "score": 1.0, + "content": "between adjacent numbers on the circle as", + "type": "text" + }, + { + "bbox": [ + 279, + 590, + 285, + 599 + ], + "score": 0.76, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 589, + 303, + 602 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 303, + 589, + 313, + 600 + ], + "score": 0.82, + "content": "\\delta _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 589, + 316, + 602 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 317, + 589, + 327, + 600 + ], + "score": 0.8, + "content": "\\delta _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "for accompanied and accompanying pizzas,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 179, + 613 + ], + "score": 1.0, + "content": "respectively, then", + "type": "text" + }, + { + "bbox": [ + 179, + 600, + 217, + 611 + ], + "score": 0.89, + "content": "\\delta _ { 1 } = 2 \\delta _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 600, + 243, + 613 + ], + "score": 1.0, + "content": "(mod", + "type": "text" + }, + { + "bbox": [ + 243, + 601, + 250, + 612 + ], + "score": 0.39, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "). In the experiment, we found that pizzas #1/#2/#3 in Figure 4", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 610, + 507, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 507, + 625 + ], + "score": 1.0, + "content": "all have accompanying pizzas, which we call pizzas #4/#5/#6 (see Appendix D for details). However,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 620, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 104, + 620, + 506, + 636 + ], + "score": 1.0, + "content": "these accompanying pizzas do not play a significant role in final model predictions 2. We conjecture", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "that training dynamics are as follows: (1) At initialization, pizzas #1/#2/#3 correspond to three", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "different “lottery tickets” [9]. (2) In early stages of training, to compensate the weaknesses (antipodal", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "pairs) of pizzas #1/#2/#3, pizzas #4/#5/#6 are formed. (3) As training goes on (in the presence of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "weight decay), the neural network gets pruned. As a result, pizzas #4/#5/#6 are not significantly", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 676, + 442, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 442, + 690 + ], + "score": 1.0, + "content": "involved in prediction, although they continue to be visible in the embedding space.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 546, + 507, + 690 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 70, + 276, + 85 + ], + "lines": [ + { + "bbox": [ + 104, + 69, + 278, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 278, + 87 + ], + "score": 1.0, + "content": "4 The Algorithmic Phase Space", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 94, + 506, + 172 + ], + "lines": [ + { + "bbox": [ + 105, + 94, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "In Section 3, we have demonstrated a typical Clock (Model A) and a typical Pizza (Model B). In this", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 106, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 106, + 506, + 118 + ], + "score": 1.0, + "content": "section, we study how architectures and hyperparametes govern the selection of these two algorithmic", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 506, + 129 + ], + "score": 1.0, + "content": "“phases”. In Section 4.1, we propose quantitative metrics that can distinguish between Pizza and Clock.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 507, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 507, + 141 + ], + "score": 1.0, + "content": "In Section 4.2, we observe how these metrics behave with different architectures and hyperparameters,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "score": 1.0, + "content": "demonstrating sharp phase transitions. The results in this section focus Clock and Pizza models, but", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "other algorithmic solutions to modular addition are also discovered, and explored in more detail in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 160, + 160, + 173 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 160, + 173 + ], + "score": 1.0, + "content": "Appendix B.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 107, + 183, + 164, + 195 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 164, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 164, + 196 + ], + "score": 1.0, + "content": "4.1 Metrics", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 217 + ], + "score": 1.0, + "content": "We wish to study the distribution of Pizza and Clock algorithms statistically, which will require us to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "distinguish between two algorithms automatically. In order to do so, we formalize our observations", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 226, + 497, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 497, + 238 + ], + "score": 1.0, + "content": "in Section 2.2 and 2.3, arriving at two metrics: gradient symmetricity and distance irrelevance.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "title", + "bbox": [ + 107, + 248, + 237, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 238, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 238, + 263 + ], + "score": 1.0, + "content": "4.1.1 Gradient Symmetricity", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 266, + 506, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 433, + 280 + ], + "score": 1.0, + "content": "To measure the symmetricity of the gradients, we select some input-output group", + "type": "text" + }, + { + "bbox": [ + 433, + 267, + 464, + 279 + ], + "score": 0.94, + "content": "( a , b , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 266, + 505, + 280 + ], + "score": 1.0, + "content": ", compute", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 277, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 309, + 291 + ], + "score": 1.0, + "content": "the gradient vectors for the output logit at position", + "type": "text" + }, + { + "bbox": [ + 309, + 280, + 316, + 288 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 277, + 505, + 291 + ], + "score": 1.0, + "content": "with respect to the input embeddings, and then", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 288, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 302 + ], + "score": 1.0, + "content": "compute the cosine similarity. Taking the average over many pairs yields the gradient symmetricity.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 299, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 104, + 299, + 342, + 315 + ], + "score": 1.0, + "content": "Definition 4.1 (Gradient Symmetricity). For a fixed set", + "type": "text" + }, + { + "bbox": [ + 342, + 301, + 379, + 315 + ], + "score": 0.91, + "content": "S \\subseteq \\mathbb { Z } _ { p } ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 299, + 506, + 315 + ], + "score": 1.0, + "content": "of input-output pairs3, define", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 312, + 376, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 251, + 325 + ], + "score": 1.0, + "content": "gradient-symmetricity of a network", + "type": "text" + }, + { + "bbox": [ + 251, + 313, + 263, + 322 + ], + "score": 0.72, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 312, + 353, + 325 + ], + "score": 1.0, + "content": "with embedding layer", + "type": "text" + }, + { + "bbox": [ + 353, + 314, + 362, + 322 + ], + "score": 0.68, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 312, + 376, + 325 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 325, + 390, + 357 + ], + "lines": [ + { + "bbox": [ + 219, + 325, + 390, + 357 + ], + "spans": [ + { + "bbox": [ + 219, + 325, + 390, + 357 + ], + "score": 0.95, + "content": "s _ { g } \\equiv \\frac { 1 } { | S | } \\sum _ { ( a , b , c ) \\in S } s i m \\left( \\frac { \\partial Q _ { a b c } } { \\partial { \\bf E } _ { a } } , \\frac { \\partial Q _ { a b c } } { \\partial { \\bf E } _ { b } } \\right) ,", + "type": "interline_equation", + "image_path": "919fd114fce35c7ab580a67983c7d5977104770180d8b0cc953aa581423c6c44.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 325, + 390, + 341.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 219, + 341.0, + 390, + 357.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 360, + 502, + 386 + ], + "lines": [ + { + "bbox": [ + 105, + 357, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 133, + 376 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 359, + 202, + 374 + ], + "score": 0.94, + "content": "\\begin{array} { r } { s i m ( a , b ) = \\frac { a \\cdot b } { | a | | b | } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 357, + 297, + 376 + ], + "score": 1.0, + "content": "is the cosine-similarity,", + "type": "text" + }, + { + "bbox": [ + 297, + 360, + 318, + 372 + ], + "score": 0.91, + "content": "Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 357, + 399, + 376 + ], + "score": 1.0, + "content": "is the logit for class", + "type": "text" + }, + { + "bbox": [ + 399, + 362, + 406, + 370 + ], + "score": 0.28, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 357, + 505, + 376 + ], + "score": 1.0, + "content": "given input a and b. It is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 373, + 201, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 147, + 387 + ], + "score": 1.0, + "content": "clear that", + "type": "text" + }, + { + "bbox": [ + 147, + 374, + 197, + 386 + ], + "score": 0.9, + "content": "s _ { g } \\in [ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 373, + 201, + 387 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 392, + 505, + 426 + ], + "lines": [ + { + "bbox": [ + 106, + 392, + 504, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 504, + 405 + ], + "score": 1.0, + "content": "As we discussed in Section 2.2, the Pizza algorithm has symmetric gradients while the Clock algorithm", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 454, + 415 + ], + "score": 1.0, + "content": "has asymmetric ones. Model A and Model B in Section 3 have gradient symmetricity", + "type": "text" + }, + { + "bbox": [ + 454, + 403, + 487, + 414 + ], + "score": 0.87, + "content": "9 9 . 3 \\hat { 7 } \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 414, + 237, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 138, + 425 + ], + "score": 0.88, + "content": "3 3 . 3 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 414, + 237, + 427 + ], + "score": 1.0, + "content": ", respectively (Figure 2).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 107, + 436, + 225, + 448 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 226, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 226, + 449 + ], + "score": 1.0, + "content": "4.1.2 Distance Irrelevance", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 505, + 499 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "To measure the dependence of correct logits on differences between two inputs, which reflect the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 465, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 478 + ], + "score": 1.0, + "content": "distances of the inputs on the circles, we measure how much of the variance in the correct logit matrix", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "depends on it. We do so by comparing the average standard deviation of correct logits from inputs", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 487, + 381, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 381, + 500 + ], + "score": 1.0, + "content": "with the same differences and the standard deviation from all inputs.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 501, + 503, + 523 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 343, + 515 + ], + "score": 1.0, + "content": "Definition 4.2 (Distance Irrelevance). For some network", + "type": "text" + }, + { + "bbox": [ + 343, + 501, + 355, + 511 + ], + "score": 0.63, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 498, + 461, + 515 + ], + "score": 1.0, + "content": "with correct logit matrix", + "type": "text" + }, + { + "bbox": [ + 461, + 501, + 470, + 511 + ], + "score": 0.39, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 501, + 505, + 514 + ], + "score": 0.75, + "content": "( L _ { i , j } =", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 511, + 280, + 525 + ], + "spans": [ + { + "bbox": [ + 107, + 512, + 139, + 524 + ], + "score": 0.89, + "content": "Q _ { i j , i + j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 511, + 280, + 525 + ], + "score": 1.0, + "content": ", define its distance irrelevance as", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "interline_equation", + "bbox": [ + 231, + 524, + 379, + 556 + ], + "lines": [ + { + "bbox": [ + 231, + 524, + 379, + 556 + ], + "spans": [ + { + "bbox": [ + 231, + 524, + 379, + 556 + ], + "score": 0.94, + "content": "q \\equiv \\frac { \\frac { 1 } { p } \\sum _ { d \\in \\mathbb { Z } _ { p } } \\operatorname { s t d } \\left( L _ { i , i + d } \\mid i \\in \\mathbb { Z } _ { p } \\right) } { \\operatorname { s t d } \\left( L _ { i , j } \\mid i , j \\in \\mathbb { Z } _ { p } ^ { 2 } \\right) } ,", + "type": "interline_equation", + "image_path": "c7b66d3b3104beb6ffda565875d9345bb68581401d1ba011a649067b482cd8d0.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 231, + 524, + 379, + 540.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 231, + 540.0, + 379, + 556.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 558, + 414, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 411, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 374, + 572 + ], + "score": 1.0, + "content": "where std computes the standard deviation of a set. It is clear that", + "type": "text" + }, + { + "bbox": [ + 374, + 558, + 411, + 570 + ], + "score": 0.92, + "content": "q \\in [ 0 , 1 ]", + "type": "inline_equation" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 576, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 589 + ], + "score": 1.0, + "content": "Model A and Model B in Section 3 give distance irrelevance 0.17 and 0.85, respectively (Figure 3).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "A typical distance irrelevance from the Pizza algorithm ranges from 0 to 0.4 while a typical distance", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 597, + 327, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 327, + 610 + ], + "score": 1.0, + "content": "irrelevance from Clock algorithm ranges from 0.4 to 1.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 107, + 619, + 273, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 275, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 275, + 633 + ], + "score": 1.0, + "content": "4.1.3 Which Metric is More Decisive?", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 639, + 505, + 705 + ], + "lines": [ + { + "bbox": [ + 105, + 638, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 652 + ], + "score": 1.0, + "content": "When the two metrics have conflicting results, which one is more decisive? We consider distance", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 663 + ], + "score": 1.0, + "content": "irrelevance as the decisive factor of the Pizza algorithm, as the output logits being dependent on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 661, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 505, + 673 + ], + "score": 1.0, + "content": "the distance is highly suggestive of Pizza. On the other hand, gradient symmetricity can be used", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "score": 1.0, + "content": "to rule out the Clock algorithm, as it requires multiplying (transformed) inputs which will result in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 683, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 505, + 696 + ], + "score": 1.0, + "content": "asymmetric gradients. Figure 6 confirmed that at low distance irrelevance (suggesting pizza) the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 694, + 402, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 402, + 706 + ], + "score": 1.0, + "content": "gradient symmetricity is almost always close to 1 (suggesting non-clock).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 116, + 711, + 473, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 474, + 724 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 299, + 724 + ], + "score": 1.0, + "content": "3To speed-up the calculations, in our experiments", + "type": "text" + }, + { + "bbox": [ + 299, + 713, + 306, + 721 + ], + "score": 0.85, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 709, + 416, + 724 + ], + "score": 1.0, + "content": "is taken as a random subset of", + "type": "text" + }, + { + "bbox": [ + 417, + 711, + 428, + 723 + ], + "score": 0.9, + "content": "\\mathbb { Z } _ { p } ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 709, + 474, + 724 + ], + "score": 1.0, + "content": "of size 100.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 70, + 276, + 85 + ], + "lines": [ + { + "bbox": [ + 104, + 69, + 278, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 278, + 87 + ], + "score": 1.0, + "content": "4 The Algorithmic Phase Space", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 94, + 506, + 172 + ], + "lines": [ + { + "bbox": [ + 105, + 94, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "In Section 3, we have demonstrated a typical Clock (Model A) and a typical Pizza (Model B). In this", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 106, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 106, + 506, + 118 + ], + "score": 1.0, + "content": "section, we study how architectures and hyperparametes govern the selection of these two algorithmic", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 506, + 129 + ], + "score": 1.0, + "content": "“phases”. In Section 4.1, we propose quantitative metrics that can distinguish between Pizza and Clock.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 507, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 507, + 141 + ], + "score": 1.0, + "content": "In Section 4.2, we observe how these metrics behave with different architectures and hyperparameters,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "score": 1.0, + "content": "demonstrating sharp phase transitions. The results in this section focus Clock and Pizza models, but", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "other algorithmic solutions to modular addition are also discovered, and explored in more detail in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 160, + 160, + 173 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 160, + 173 + ], + "score": 1.0, + "content": "Appendix B.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4, + "bbox_fs": [ + 104, + 94, + 507, + 173 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 183, + 164, + 195 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 164, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 164, + 196 + ], + "score": 1.0, + "content": "4.1 Metrics", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 217 + ], + "score": 1.0, + "content": "We wish to study the distribution of Pizza and Clock algorithms statistically, which will require us to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "distinguish between two algorithms automatically. In order to do so, we formalize our observations", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 226, + 497, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 497, + 238 + ], + "score": 1.0, + "content": "in Section 2.2 and 2.3, arriving at two metrics: gradient symmetricity and distance irrelevance.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 203, + 505, + 238 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 248, + 237, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 238, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 238, + 263 + ], + "score": 1.0, + "content": "4.1.1 Gradient Symmetricity", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 266, + 506, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 433, + 280 + ], + "score": 1.0, + "content": "To measure the symmetricity of the gradients, we select some input-output group", + "type": "text" + }, + { + "bbox": [ + 433, + 267, + 464, + 279 + ], + "score": 0.94, + "content": "( a , b , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 266, + 505, + 280 + ], + "score": 1.0, + "content": ", compute", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 277, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 309, + 291 + ], + "score": 1.0, + "content": "the gradient vectors for the output logit at position", + "type": "text" + }, + { + "bbox": [ + 309, + 280, + 316, + 288 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 277, + 505, + 291 + ], + "score": 1.0, + "content": "with respect to the input embeddings, and then", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 288, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 302 + ], + "score": 1.0, + "content": "compute the cosine similarity. Taking the average over many pairs yields the gradient symmetricity.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 299, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 104, + 299, + 342, + 315 + ], + "score": 1.0, + "content": "Definition 4.1 (Gradient Symmetricity). For a fixed set", + "type": "text" + }, + { + "bbox": [ + 342, + 301, + 379, + 315 + ], + "score": 0.91, + "content": "S \\subseteq \\mathbb { Z } _ { p } ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 299, + 506, + 315 + ], + "score": 1.0, + "content": "of input-output pairs3, define", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 312, + 376, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 251, + 325 + ], + "score": 1.0, + "content": "gradient-symmetricity of a network", + "type": "text" + }, + { + "bbox": [ + 251, + 313, + 263, + 322 + ], + "score": 0.72, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 312, + 353, + 325 + ], + "score": 1.0, + "content": "with embedding layer", + "type": "text" + }, + { + "bbox": [ + 353, + 314, + 362, + 322 + ], + "score": 0.68, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 312, + 376, + 325 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 266, + 506, + 325 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 325, + 390, + 357 + ], + "lines": [ + { + "bbox": [ + 219, + 325, + 390, + 357 + ], + "spans": [ + { + "bbox": [ + 219, + 325, + 390, + 357 + ], + "score": 0.95, + "content": "s _ { g } \\equiv \\frac { 1 } { | S | } \\sum _ { ( a , b , c ) \\in S } s i m \\left( \\frac { \\partial Q _ { a b c } } { \\partial { \\bf E } _ { a } } , \\frac { \\partial Q _ { a b c } } { \\partial { \\bf E } _ { b } } \\right) ,", + "type": "interline_equation", + "image_path": "919fd114fce35c7ab580a67983c7d5977104770180d8b0cc953aa581423c6c44.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 325, + 390, + 341.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 219, + 341.0, + 390, + 357.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 360, + 502, + 386 + ], + "lines": [ + { + "bbox": [ + 105, + 357, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 133, + 376 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 359, + 202, + 374 + ], + "score": 0.94, + "content": "\\begin{array} { r } { s i m ( a , b ) = \\frac { a \\cdot b } { | a | | b | } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 357, + 297, + 376 + ], + "score": 1.0, + "content": "is the cosine-similarity,", + "type": "text" + }, + { + "bbox": [ + 297, + 360, + 318, + 372 + ], + "score": 0.91, + "content": "Q _ { a b c }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 357, + 399, + 376 + ], + "score": 1.0, + "content": "is the logit for class", + "type": "text" + }, + { + "bbox": [ + 399, + 362, + 406, + 370 + ], + "score": 0.28, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 357, + 505, + 376 + ], + "score": 1.0, + "content": "given input a and b. It is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 373, + 201, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 147, + 387 + ], + "score": 1.0, + "content": "clear that", + "type": "text" + }, + { + "bbox": [ + 147, + 374, + 197, + 386 + ], + "score": 0.9, + "content": "s _ { g } \\in [ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 373, + 201, + 387 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 357, + 505, + 387 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 392, + 505, + 426 + ], + "lines": [ + { + "bbox": [ + 106, + 392, + 504, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 504, + 405 + ], + "score": 1.0, + "content": "As we discussed in Section 2.2, the Pizza algorithm has symmetric gradients while the Clock algorithm", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 454, + 415 + ], + "score": 1.0, + "content": "has asymmetric ones. Model A and Model B in Section 3 have gradient symmetricity", + "type": "text" + }, + { + "bbox": [ + 454, + 403, + 487, + 414 + ], + "score": 0.87, + "content": "9 9 . 3 \\hat { 7 } \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 414, + 237, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 138, + 425 + ], + "score": 0.88, + "content": "3 3 . 3 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 414, + 237, + 427 + ], + "score": 1.0, + "content": ", respectively (Figure 2).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 392, + 505, + 427 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 436, + 225, + 448 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 226, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 226, + 449 + ], + "score": 1.0, + "content": "4.1.2 Distance Irrelevance", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 505, + 499 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "To measure the dependence of correct logits on differences between two inputs, which reflect the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 465, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 478 + ], + "score": 1.0, + "content": "distances of the inputs on the circles, we measure how much of the variance in the correct logit matrix", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "depends on it. We do so by comparing the average standard deviation of correct logits from inputs", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 487, + 381, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 381, + 500 + ], + "score": 1.0, + "content": "with the same differences and the standard deviation from all inputs.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 455, + 506, + 500 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 501, + 503, + 523 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 343, + 515 + ], + "score": 1.0, + "content": "Definition 4.2 (Distance Irrelevance). For some network", + "type": "text" + }, + { + "bbox": [ + 343, + 501, + 355, + 511 + ], + "score": 0.63, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 498, + 461, + 515 + ], + "score": 1.0, + "content": "with correct logit matrix", + "type": "text" + }, + { + "bbox": [ + 461, + 501, + 470, + 511 + ], + "score": 0.39, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 501, + 505, + 514 + ], + "score": 0.75, + "content": "( L _ { i , j } =", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 511, + 280, + 525 + ], + "spans": [ + { + "bbox": [ + 107, + 512, + 139, + 524 + ], + "score": 0.89, + "content": "Q _ { i j , i + j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 511, + 280, + 525 + ], + "score": 1.0, + "content": ", define its distance irrelevance as", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 498, + 505, + 525 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 231, + 524, + 379, + 556 + ], + "lines": [ + { + "bbox": [ + 231, + 524, + 379, + 556 + ], + "spans": [ + { + "bbox": [ + 231, + 524, + 379, + 556 + ], + "score": 0.94, + "content": "q \\equiv \\frac { \\frac { 1 } { p } \\sum _ { d \\in \\mathbb { Z } _ { p } } \\operatorname { s t d } \\left( L _ { i , i + d } \\mid i \\in \\mathbb { Z } _ { p } \\right) } { \\operatorname { s t d } \\left( L _ { i , j } \\mid i , j \\in \\mathbb { Z } _ { p } ^ { 2 } \\right) } ,", + "type": "interline_equation", + "image_path": "c7b66d3b3104beb6ffda565875d9345bb68581401d1ba011a649067b482cd8d0.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 231, + 524, + 379, + 540.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 231, + 540.0, + 379, + 556.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 558, + 414, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 411, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 374, + 572 + ], + "score": 1.0, + "content": "where std computes the standard deviation of a set. It is clear that", + "type": "text" + }, + { + "bbox": [ + 374, + 558, + 411, + 570 + ], + "score": 0.92, + "content": "q \\in [ 0 , 1 ]", + "type": "inline_equation" + } + ], + "index": 34 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 556, + 411, + 572 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 576, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 589 + ], + "score": 1.0, + "content": "Model A and Model B in Section 3 give distance irrelevance 0.17 and 0.85, respectively (Figure 3).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "A typical distance irrelevance from the Pizza algorithm ranges from 0 to 0.4 while a typical distance", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 597, + 327, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 327, + 610 + ], + "score": 1.0, + "content": "irrelevance from Clock algorithm ranges from 0.4 to 1.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 574, + 505, + 610 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 619, + 273, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 275, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 275, + 633 + ], + "score": 1.0, + "content": "4.1.3 Which Metric is More Decisive?", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 639, + 505, + 705 + ], + "lines": [ + { + "bbox": [ + 105, + 638, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 652 + ], + "score": 1.0, + "content": "When the two metrics have conflicting results, which one is more decisive? We consider distance", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 663 + ], + "score": 1.0, + "content": "irrelevance as the decisive factor of the Pizza algorithm, as the output logits being dependent on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 661, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 505, + 673 + ], + "score": 1.0, + "content": "the distance is highly suggestive of Pizza. On the other hand, gradient symmetricity can be used", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "score": 1.0, + "content": "to rule out the Clock algorithm, as it requires multiplying (transformed) inputs which will result in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 683, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 505, + 696 + ], + "score": 1.0, + "content": "asymmetric gradients. Figure 6 confirmed that at low distance irrelevance (suggesting pizza) the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 694, + 402, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 402, + 706 + ], + "score": 1.0, + "content": "gradient symmetricity is almost always close to 1 (suggesting non-clock).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 638, + 505, + 706 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 75, + 461, + 177 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 150, + 75, + 461, + 177 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 150, + 75, + 461, + 177 + ], + "spans": [ + { + "bbox": [ + 150, + 75, + 461, + 177 + ], + "score": 0.966, + "type": "image", + "image_path": "3b5bcb2407867634c3a3a64e1ba5e3e12dd8262482af1732c6a562c71758e744.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 150, + 75, + 461, + 109.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 150, + 109.0, + 461, + 143.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 150, + 143.0, + 461, + 177.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 131, + 187, + 478, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 131, + 186, + 479, + 201 + ], + "spans": [ + { + "bbox": [ + 131, + 186, + 479, + 201 + ], + "score": 1.0, + "content": "Figure 6: Distance irrelevance vs gradient symmetricity over all the main experiments.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 107, + 220, + 303, + 232 + ], + "lines": [ + { + "bbox": [ + 106, + 220, + 305, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 305, + 234 + ], + "score": 1.0, + "content": "4.2 Identifying algorithmic phase transitions", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 108, + 240, + 504, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "score": 1.0, + "content": "How do models “choose” whether to implement the Clock or Pizza algorithm? We investigate this", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 252, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 506, + 264 + ], + "score": 1.0, + "content": "question by interpolating between Model A (transformer without attention) and Model B (transformer", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 263, + 471, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 354, + 275 + ], + "score": 1.0, + "content": "with attention). To do so, we introduce a new hyperparameter", + "type": "text" + }, + { + "bbox": [ + 354, + 265, + 362, + 273 + ], + "score": 0.74, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 263, + 471, + 275 + ], + "score": 1.0, + "content": "we call the attention rate.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 279, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 234, + 290 + ], + "score": 1.0, + "content": "For a model with attention rate", + "type": "text" + }, + { + "bbox": [ + 234, + 281, + 242, + 289 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 279, + 374, + 290 + ], + "score": 1.0, + "content": ", we modify the attention matrix", + "type": "text" + }, + { + "bbox": [ + 374, + 279, + 386, + 289 + ], + "score": 0.73, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 279, + 505, + 290 + ], + "score": 1.0, + "content": "for each attention head to be", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 107, + 290, + 203, + 302 + ], + "score": 0.91, + "content": "M ^ { \\prime } = M \\alpha + J ( 1 - \\alpha )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 289, + 505, + 302 + ], + "score": 1.0, + "content": ". In other words, we modify this matrix to consist of a linear interpolation", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 431, + 313 + ], + "score": 1.0, + "content": "between the all-one matrix and the original attention (post-softmax), with the rate", + "type": "text" + }, + { + "bbox": [ + 432, + 303, + 439, + 311 + ], + "score": 0.73, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "controlling how", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "score": 1.0, + "content": "much of the attention is kept. The transformer with and without attention corresponds to the case", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 133, + 336 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 323, + 161, + 333 + ], + "score": 0.89, + "content": "\\alpha = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 322, + 245, + 336 + ], + "score": 1.0, + "content": "(attention kept) and", + "type": "text" + }, + { + "bbox": [ + 245, + 323, + 272, + 333 + ], + "score": 0.89, + "content": "\\alpha = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 322, + 506, + 336 + ], + "score": 1.0, + "content": "(constant attention matrix). With this parameter, we can", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 334, + 378, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 378, + 347 + ], + "score": 1.0, + "content": "control the balance of attention versus linear layers in transformers.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 350, + 505, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 350, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 505, + 363 + ], + "score": 1.0, + "content": "We performed the following set of experiments on transformers (see Appendix F.1 for architecture", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 361, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 373 + ], + "score": 1.0, + "content": "and training details). (1) One-layer transformers with width 128 and attention rate uniformly sampled", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 371, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 117, + 385 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 118, + 372, + 138, + 384 + ], + "score": 0.67, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 371, + 506, + 385 + ], + "score": 1.0, + "content": "(Figure 7). (2) One-layer transformers with width log-uniformly sampled in [32, 512] and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 383, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 250, + 395 + ], + "score": 1.0, + "content": "attention rate uniformly sampled in", + "type": "text" + }, + { + "bbox": [ + 250, + 383, + 271, + 394 + ], + "score": 0.58, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 383, + 506, + 395 + ], + "score": 1.0, + "content": "(Figure 7). (3) Transformers with 2 to 4 layers, width 128", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 393, + 339, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 267, + 407 + ], + "score": 1.0, + "content": "and attention rate uniformly sampled in", + "type": "text" + }, + { + "bbox": [ + 267, + 394, + 288, + 406 + ], + "score": 0.82, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 393, + 339, + 407 + ], + "score": 1.0, + "content": "(Figure 11).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 506, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 506, + 431 + ], + "score": 1.0, + "content": "The Pizza and the Clock algorithms are the dominating algorithms with circular embeddings.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "For circular models, most observed models either have low gradient symmetricity (corresponding to", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 439, + 460, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 460, + 453 + ], + "score": 1.0, + "content": "the Clock algorithm) or low distance irrelevance (corresponding to the Pizza algorithm).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "Two-dimensional phase change observed for attention rate and layer width. For the fixed-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "width experiment, we observed a clear phase transition from the Pizza algorithm to the Clock", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "algorithm (characterized by gradient symmetricity and distance irrelevance). We also observe an", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 496, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 507 + ], + "score": 1.0, + "content": "almost linear phase boundary with regards to both attention rate and layer width. In other words, the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 507, + 361, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 361, + 519 + ], + "score": 1.0, + "content": "attention rate transition point increases as the model gets wider.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 530, + 504, + 552 + ], + "lines": [ + { + "bbox": [ + 107, + 530, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 107, + 530, + 506, + 541 + ], + "score": 1.0, + "content": "Dominance of linear layers determines whether the Pizza or the Clock algorithm is preferred.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 541, + 493, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 493, + 553 + ], + "score": 1.0, + "content": "For one-layer transformers, we study the transition point against the attention rate and the width:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 133, + 563, + 505, + 666 + ], + "lines": [ + { + "bbox": [ + 133, + 561, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 133, + 561, + 505, + 576 + ], + "score": 1.0, + "content": "• The Clock algorithm dominates when the attention rate is higher than the phase change", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 574, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 141, + 574, + 506, + 586 + ], + "score": 1.0, + "content": "point, and the Pizza algorithm dominates when the attention rate is lower than the point.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 584, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 141, + 584, + 506, + 597 + ], + "score": 1.0, + "content": "Our explanation is: At a high attention rate, the attention mechanism is more prominent in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 141, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 141, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "the network, giving rise to the clock algorithm. At a low attention rate, the linear layers are", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 141, + 607, + 348, + 619 + ], + "spans": [ + { + "bbox": [ + 141, + 607, + 348, + 619 + ], + "score": 1.0, + "content": "more prominent, giving rise to the pizza algorithm.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 133, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "• The phase change point gets higher when the model width increases. Our explanation is:", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 141, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 141, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "When the model gets wider, the linear layers become more capable while the attention", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 141, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 141, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "mechanism receive less benefit (attentions remain scalars while outputs from linear layers", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 142, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 506, + 666 + ], + "score": 1.0, + "content": "become wider vectors). The linear layer therefore gets more prominence with a wider model.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "Possibly hybrid algorithms between the Clock and the Pizza algorithms. The continuous phase", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 689, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 700 + ], + "score": 1.0, + "content": "change suggests the existence of networks that lie between the Clock and the Pizza algorithms. This is", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "achievable by having some principal components acting as the Clock and some principal components", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 711, + 186, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 186, + 723 + ], + "score": 1.0, + "content": "acting as the Pizza.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 75, + 461, + 177 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 150, + 75, + 461, + 177 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 150, + 75, + 461, + 177 + ], + "spans": [ + { + "bbox": [ + 150, + 75, + 461, + 177 + ], + "score": 0.966, + "type": "image", + "image_path": "3b5bcb2407867634c3a3a64e1ba5e3e12dd8262482af1732c6a562c71758e744.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 150, + 75, + 461, + 109.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 150, + 109.0, + 461, + 143.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 150, + 143.0, + 461, + 177.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 131, + 187, + 478, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 131, + 186, + 479, + 201 + ], + "spans": [ + { + "bbox": [ + 131, + 186, + 479, + 201 + ], + "score": 1.0, + "content": "Figure 6: Distance irrelevance vs gradient symmetricity over all the main experiments.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 107, + 220, + 303, + 232 + ], + "lines": [ + { + "bbox": [ + 106, + 220, + 305, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 305, + 234 + ], + "score": 1.0, + "content": "4.2 Identifying algorithmic phase transitions", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 108, + 240, + 504, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "score": 1.0, + "content": "How do models “choose” whether to implement the Clock or Pizza algorithm? We investigate this", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 252, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 506, + 264 + ], + "score": 1.0, + "content": "question by interpolating between Model A (transformer without attention) and Model B (transformer", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 263, + 471, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 354, + 275 + ], + "score": 1.0, + "content": "with attention). To do so, we introduce a new hyperparameter", + "type": "text" + }, + { + "bbox": [ + 354, + 265, + 362, + 273 + ], + "score": 0.74, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 263, + 471, + 275 + ], + "score": 1.0, + "content": "we call the attention rate.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 106, + 241, + 506, + 275 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 279, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 234, + 290 + ], + "score": 1.0, + "content": "For a model with attention rate", + "type": "text" + }, + { + "bbox": [ + 234, + 281, + 242, + 289 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 279, + 374, + 290 + ], + "score": 1.0, + "content": ", we modify the attention matrix", + "type": "text" + }, + { + "bbox": [ + 374, + 279, + 386, + 289 + ], + "score": 0.73, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 279, + 505, + 290 + ], + "score": 1.0, + "content": "for each attention head to be", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 107, + 290, + 203, + 302 + ], + "score": 0.91, + "content": "M ^ { \\prime } = M \\alpha + J ( 1 - \\alpha )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 289, + 505, + 302 + ], + "score": 1.0, + "content": ". In other words, we modify this matrix to consist of a linear interpolation", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 431, + 313 + ], + "score": 1.0, + "content": "between the all-one matrix and the original attention (post-softmax), with the rate", + "type": "text" + }, + { + "bbox": [ + 432, + 303, + 439, + 311 + ], + "score": 0.73, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "controlling how", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "score": 1.0, + "content": "much of the attention is kept. The transformer with and without attention corresponds to the case", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 133, + 336 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 323, + 161, + 333 + ], + "score": 0.89, + "content": "\\alpha = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 322, + 245, + 336 + ], + "score": 1.0, + "content": "(attention kept) and", + "type": "text" + }, + { + "bbox": [ + 245, + 323, + 272, + 333 + ], + "score": 0.89, + "content": "\\alpha = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 322, + 506, + 336 + ], + "score": 1.0, + "content": "(constant attention matrix). With this parameter, we can", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 334, + 378, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 378, + 347 + ], + "score": 1.0, + "content": "control the balance of attention versus linear layers in transformers.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 279, + 506, + 347 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 350, + 505, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 350, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 505, + 363 + ], + "score": 1.0, + "content": "We performed the following set of experiments on transformers (see Appendix F.1 for architecture", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 361, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 373 + ], + "score": 1.0, + "content": "and training details). (1) One-layer transformers with width 128 and attention rate uniformly sampled", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 371, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 117, + 385 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 118, + 372, + 138, + 384 + ], + "score": 0.67, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 371, + 506, + 385 + ], + "score": 1.0, + "content": "(Figure 7). (2) One-layer transformers with width log-uniformly sampled in [32, 512] and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 383, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 250, + 395 + ], + "score": 1.0, + "content": "attention rate uniformly sampled in", + "type": "text" + }, + { + "bbox": [ + 250, + 383, + 271, + 394 + ], + "score": 0.58, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 383, + 506, + 395 + ], + "score": 1.0, + "content": "(Figure 7). (3) Transformers with 2 to 4 layers, width 128", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 393, + 339, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 267, + 407 + ], + "score": 1.0, + "content": "and attention rate uniformly sampled in", + "type": "text" + }, + { + "bbox": [ + 267, + 394, + 288, + 406 + ], + "score": 0.82, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 393, + 339, + 407 + ], + "score": 1.0, + "content": "(Figure 11).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 350, + 506, + 407 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 506, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 506, + 431 + ], + "score": 1.0, + "content": "The Pizza and the Clock algorithms are the dominating algorithms with circular embeddings.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "For circular models, most observed models either have low gradient symmetricity (corresponding to", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 439, + 460, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 460, + 453 + ], + "score": 1.0, + "content": "the Clock algorithm) or low distance irrelevance (corresponding to the Pizza algorithm).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 416, + 506, + 453 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "Two-dimensional phase change observed for attention rate and layer width. For the fixed-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "width experiment, we observed a clear phase transition from the Pizza algorithm to the Clock", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "algorithm (characterized by gradient symmetricity and distance irrelevance). We also observe an", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 496, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 507 + ], + "score": 1.0, + "content": "almost linear phase boundary with regards to both attention rate and layer width. In other words, the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 507, + 361, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 361, + 519 + ], + "score": 1.0, + "content": "attention rate transition point increases as the model gets wider.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 462, + 506, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 530, + 504, + 552 + ], + "lines": [ + { + "bbox": [ + 107, + 530, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 107, + 530, + 506, + 541 + ], + "score": 1.0, + "content": "Dominance of linear layers determines whether the Pizza or the Clock algorithm is preferred.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 541, + 493, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 493, + 553 + ], + "score": 1.0, + "content": "For one-layer transformers, we study the transition point against the attention rate and the width:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 530, + 506, + 553 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 563, + 505, + 666 + ], + "lines": [ + { + "bbox": [ + 133, + 561, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 133, + 561, + 505, + 576 + ], + "score": 1.0, + "content": "• The Clock algorithm dominates when the attention rate is higher than the phase change", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 574, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 141, + 574, + 506, + 586 + ], + "score": 1.0, + "content": "point, and the Pizza algorithm dominates when the attention rate is lower than the point.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 584, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 141, + 584, + 506, + 597 + ], + "score": 1.0, + "content": "Our explanation is: At a high attention rate, the attention mechanism is more prominent in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 141, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 141, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "the network, giving rise to the clock algorithm. At a low attention rate, the linear layers are", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 141, + 607, + 348, + 619 + ], + "spans": [ + { + "bbox": [ + 141, + 607, + 348, + 619 + ], + "score": 1.0, + "content": "more prominent, giving rise to the pizza algorithm.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 133, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "• The phase change point gets higher when the model width increases. Our explanation is:", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 141, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 141, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "When the model gets wider, the linear layers become more capable while the attention", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 141, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 141, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "mechanism receive less benefit (attentions remain scalars while outputs from linear layers", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 142, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 506, + 666 + ], + "score": 1.0, + "content": "become wider vectors). The linear layer therefore gets more prominence with a wider model.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33, + "bbox_fs": [ + 133, + 561, + 506, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "Possibly hybrid algorithms between the Clock and the Pizza algorithms. The continuous phase", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 689, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 700 + ], + "score": 1.0, + "content": "change suggests the existence of networks that lie between the Clock and the Pizza algorithms. This is", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "achievable by having some principal components acting as the Clock and some principal components", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 711, + 186, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 186, + 723 + ], + "score": 1.0, + "content": "acting as the Pizza.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 677, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 73, + 500, + 384 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 73, + 500, + 384 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 73, + 500, + 384 + ], + "spans": [ + { + "bbox": [ + 108, + 73, + 500, + 384 + ], + "score": 0.978, + "type": "image", + "image_path": "0feb2db95d70a731bc0f9cfba96608247c2b687b80ec80cbd916a74cea8ae715.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 73, + 500, + 176.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 176.66666666666669, + 500, + 280.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 280.33333333333337, + 500, + 384.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 393, + 505, + 448 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 391, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 104, + 391, + 506, + 409 + ], + "score": 1.0, + "content": "Figure 7: Training results from 1-layer transformers. Each point in the plots represents a training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 104, + 404, + 264, + 417 + ], + "score": 1.0, + "content": "run reaching circular embeddings and", + "type": "text" + }, + { + "bbox": [ + 265, + 405, + 290, + 415 + ], + "score": 0.88, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "validation accuracy. See Appendix C for additional", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 415, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 104, + 415, + 506, + 429 + ], + "score": 1.0, + "content": "plots. Top: Model width fixed to be 128. Bottom: Model width varies. The phase transition lines", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 456, + 439 + ], + "score": 1.0, + "content": "are calculated by logistic regression (classify the runs by whether gradient symmetricity", + "type": "text" + }, + { + "bbox": [ + 457, + 426, + 487, + 438 + ], + "score": 0.9, + "content": "> 9 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 437, + 254, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 222, + 450 + ], + "score": 1.0, + "content": "whether distance irrelevance", + "type": "text" + }, + { + "bbox": [ + 222, + 438, + 247, + 448 + ], + "score": 0.81, + "content": "< 0 . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 437, + 254, + 450 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 106, + 471, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "Existence of non-circular algorithms. Although our presentation focuses on circular algorithms", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 483, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 506, + 495 + ], + "score": 1.0, + "content": "(i.e., whose embeddings are circular), we find non-circular algorithms (i.e., whose embeddings do not", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 492, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 507 + ], + "score": 1.0, + "content": "form a circle when projected onto any plane) to be present in neural networks. See Appendix B for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 504, + 507, + 517 + ], + "spans": [ + { + "bbox": [ + 104, + 504, + 507, + 517 + ], + "score": 1.0, + "content": "preliminary findings. We find that deeper networks are more likely to form non-circular algorithms.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 515, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 506, + 528 + ], + "score": 1.0, + "content": "We also observe the appearance of non-circular networks at low attention rates. Nevertheless, the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 525, + 493, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 493, + 539 + ], + "score": 1.0, + "content": "Pizza algorithm continues to be observed (low distance irrelevance, high gradient symmetricity).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 107, + 554, + 197, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 198, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 198, + 569 + ], + "score": 1.0, + "content": "5 Related Work", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 578, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 593 + ], + "score": 1.0, + "content": "Mechanistic interpretability aims to mechanically understand neural networks by reverse engi-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "neering them [2, 3, 5, 4, 10, 11, 1, 12, 13, 14]. One can either look for patterns in weights and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "activations by studying single-neuron behavior (superposition [11], monosemantic neurons [15]), or", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 611, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 626 + ], + "score": 1.0, + "content": "study meaningful modules or circuits grouped by neurons [4, 14]. Mechanistic interpretability is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 623, + 293, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 293, + 636 + ], + "score": 1.0, + "content": "closely related to training dynamics [8, 13, 1].", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 506, + 653 + ], + "score": 1.0, + "content": "Learning mathematical tasks: Mathematical tasks provide useful benchmarks for neural network", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "interpretability, since the tasks themselves are well understood. The setup could be learning from", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "images [16, 17], with trainable embeddings [18], or with number as inputs [19, 5]. Beyond arithmetic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 671, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 687 + ], + "score": 1.0, + "content": "relations, machine learning has been applied to learn other mathematical structures, including", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 684, + 327, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 327, + 696 + ], + "score": 1.0, + "content": "geometry [20], knot theory [21] and group theory [22].", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "Algorithmic phase transitions: Phase transitions are present in classical algorithms [23] and in deep", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 710, + 504, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 504, + 724 + ], + "score": 1.0, + "content": "learning [6, 24, 25]. Usually the phase transition means that the algorithmic performance sharply", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + } + ], + "page_idx": 8, + "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": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 73, + 500, + 384 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 73, + 500, + 384 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 73, + 500, + 384 + ], + "spans": [ + { + "bbox": [ + 108, + 73, + 500, + 384 + ], + "score": 0.978, + "type": "image", + "image_path": "0feb2db95d70a731bc0f9cfba96608247c2b687b80ec80cbd916a74cea8ae715.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 73, + 500, + 176.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 176.66666666666669, + 500, + 280.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 280.33333333333337, + 500, + 384.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 393, + 505, + 448 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 391, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 104, + 391, + 506, + 409 + ], + "score": 1.0, + "content": "Figure 7: Training results from 1-layer transformers. Each point in the plots represents a training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 104, + 404, + 264, + 417 + ], + "score": 1.0, + "content": "run reaching circular embeddings and", + "type": "text" + }, + { + "bbox": [ + 265, + 405, + 290, + 415 + ], + "score": 0.88, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "validation accuracy. See Appendix C for additional", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 415, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 104, + 415, + 506, + 429 + ], + "score": 1.0, + "content": "plots. Top: Model width fixed to be 128. Bottom: Model width varies. The phase transition lines", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 456, + 439 + ], + "score": 1.0, + "content": "are calculated by logistic regression (classify the runs by whether gradient symmetricity", + "type": "text" + }, + { + "bbox": [ + 457, + 426, + 487, + 438 + ], + "score": 0.9, + "content": "> 9 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 437, + 254, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 222, + 450 + ], + "score": 1.0, + "content": "whether distance irrelevance", + "type": "text" + }, + { + "bbox": [ + 222, + 438, + 247, + 448 + ], + "score": 0.81, + "content": "< 0 . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 437, + 254, + 450 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 106, + 471, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "Existence of non-circular algorithms. Although our presentation focuses on circular algorithms", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 483, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 506, + 495 + ], + "score": 1.0, + "content": "(i.e., whose embeddings are circular), we find non-circular algorithms (i.e., whose embeddings do not", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 492, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 507 + ], + "score": 1.0, + "content": "form a circle when projected onto any plane) to be present in neural networks. See Appendix B for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 504, + 507, + 517 + ], + "spans": [ + { + "bbox": [ + 104, + 504, + 507, + 517 + ], + "score": 1.0, + "content": "preliminary findings. We find that deeper networks are more likely to form non-circular algorithms.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 515, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 506, + 528 + ], + "score": 1.0, + "content": "We also observe the appearance of non-circular networks at low attention rates. Nevertheless, the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 525, + 493, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 493, + 539 + ], + "score": 1.0, + "content": "Pizza algorithm continues to be observed (low distance irrelevance, high gradient symmetricity).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5, + "bbox_fs": [ + 104, + 471, + 507, + 539 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 554, + 197, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 198, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 198, + 569 + ], + "score": 1.0, + "content": "5 Related Work", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 578, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 593 + ], + "score": 1.0, + "content": "Mechanistic interpretability aims to mechanically understand neural networks by reverse engi-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "neering them [2, 3, 5, 4, 10, 11, 1, 12, 13, 14]. One can either look for patterns in weights and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "activations by studying single-neuron behavior (superposition [11], monosemantic neurons [15]), or", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 611, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 626 + ], + "score": 1.0, + "content": "study meaningful modules or circuits grouped by neurons [4, 14]. Mechanistic interpretability is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 623, + 293, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 293, + 636 + ], + "score": 1.0, + "content": "closely related to training dynamics [8, 13, 1].", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 578, + 506, + 636 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 506, + 653 + ], + "score": 1.0, + "content": "Learning mathematical tasks: Mathematical tasks provide useful benchmarks for neural network", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "interpretability, since the tasks themselves are well understood. The setup could be learning from", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "images [16, 17], with trainable embeddings [18], or with number as inputs [19, 5]. Beyond arithmetic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 671, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 687 + ], + "score": 1.0, + "content": "relations, machine learning has been applied to learn other mathematical structures, including", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 684, + 327, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 327, + 696 + ], + "score": 1.0, + "content": "geometry [20], knot theory [21] and group theory [22].", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 639, + 506, + 696 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "Algorithmic phase transitions: Phase transitions are present in classical algorithms [23] and in deep", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 710, + 504, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 504, + 724 + ], + "score": 1.0, + "content": "learning [6, 24, 25]. Usually the phase transition means that the algorithmic performance sharply", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "changes when a parameter is varied (e.g., amount of data, network capacity etc). However, the phase", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "transition studied in this paper is representational: both clock and pizza give perfect accuracy, but", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "arrive at answers via different interal computations. These model-internal phase transitions are harder", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 403, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 403, + 119 + ], + "score": 1.0, + "content": "to study, but closer to corresponding phenomena in physical systems [24].", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 699, + 505, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "changes when a parameter is varied (e.g., amount of data, network capacity etc). However, the phase", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "transition studied in this paper is representational: both clock and pizza give perfect accuracy, but", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "arrive at answers via different interal computations. These model-internal phase transitions are harder", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 403, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 403, + 119 + ], + "score": 1.0, + "content": "to study, but closer to corresponding phenomena in physical systems [24].", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 122, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "Algorithm learning in neural networks: Emergent abilities in deep neural networks, especially", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "score": 1.0, + "content": "large language models, have recently attracted significant attention [26]. An ability is “emergent”", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "if the performance on a subtask suddenly increases with growing model sizes, though such claims", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "depend on the choice of metric [27]. It has been hypothesized that the emergence of specific capability", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "in a model corresponds to the emergence of a modular circuit responsible for that capability, and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "score": 1.0, + "content": "that emergence of some model behaviors thus results from a sequence of quantized circuit discovery", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 187, + 147, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 147, + 201 + ], + "score": 1.0, + "content": "steps [5].", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 107, + 214, + 187, + 227 + ], + "lines": [ + { + "bbox": [ + 104, + 211, + 190, + 230 + ], + "spans": [ + { + "bbox": [ + 104, + 211, + 190, + 230 + ], + "score": 1.0, + "content": "6 Conclusions", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 239, + 505, + 403 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 251 + ], + "score": 1.0, + "content": "We have offered a closer look at recent findings that familiar algorithms arise in neural networks", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "trained on specific algorithmic tasks. In modular arithmetic, we have shown that such algorithmic", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "discoveries are not inevitable: in addition to the Clock algorithm reverse-engineered by [1], we find", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "other algorithms (including a Pizza algorithm, and more complicated procedures) to be prevalent in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 282, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 506, + 296 + ], + "score": 1.0, + "content": "trained models. These different algorithmic phases can be distinguished using a variety of new and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 294, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 506, + 306 + ], + "score": 1.0, + "content": "existing interpretability techniques, including logit visualization, isolation of principle components", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 305, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 506, + 317 + ], + "score": 1.0, + "content": "in embedding space, and gradient-based measures of model symmetry. These techniques make it", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 315, + 507, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 507, + 329 + ], + "score": 1.0, + "content": "possible to automatically classify trained networks according to the algorithms they implement,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "and reveal algorithmic phase transitions in the space of model hyperparameters. Here we found", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 337, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 350 + ], + "score": 1.0, + "content": "specifically that the emergence of a Pizza or Clock algorithm depends on the relative strength of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 347, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 506, + 361 + ], + "score": 1.0, + "content": "linear layers and attention outputs. We additionally showed that these algorithms are not implemented", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "in isolation; instead, networks sometimes ensemble multiple copies of an algorithm in parallel.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "These results offer exciting new challenges for mechanistic interpretability: (1) How to find, classify,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 381, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 506, + 394 + ], + "score": 1.0, + "content": "and interpret unfamiliar algorithms in a systematic way? (2) How to disentangle multiple, parallel", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 391, + 343, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 343, + 406 + ], + "score": 1.0, + "content": "algorithm implementations in the presence of ensembling?", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 415, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "Limitations We have focused on a single learning problem: modular addition. Even in this", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "score": 1.0, + "content": "restricted domain, qualitatively different model behaviors emerge across architectures and seeds.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 436, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 505, + 450 + ], + "score": 1.0, + "content": "Significant additional work is needed to scale these techniques to the even more complex models", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 448, + 207, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 207, + 460 + ], + "score": 1.0, + "content": "used in real-world tasks.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 471, + 505, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "score": 1.0, + "content": "Broader Impact We believe interpretability techniques can play a crucial role in creating and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "score": 1.0, + "content": "improving safe AI systems. However, they may also be used to build more accurate systems, with the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "attendant risks inherent in all dual-use technologies. It is therefore necessary to exercise caution and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 505, + 358, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 358, + 517 + ], + "score": 1.0, + "content": "responsible decision-making when deploying such techniques.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 108, + 531, + 202, + 544 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 204, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 204, + 547 + ], + "score": 1.0, + "content": "Acknowledgement", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 556, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "We would like to thank Mingyang Deng and anonymous reviewers for valuable and fruitful discussions", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "and MIT SuperCloud for providing computation resources. ZL and MT are supported by the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "Foundational Questions Institute, the Rothberg Family Fund for Cognitive Science and IAIFI through", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 588, + 474, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 474, + 601 + ], + "score": 1.0, + "content": "NSF grant PHY-2019786. JA is supported by a gift from the OpenPhilanthropy Foundation.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "title", + "bbox": [ + 107, + 615, + 163, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 165, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 165, + 631 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 111, + 637, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 111, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 111, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "[1] Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. Progress", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 127, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "measures for grokking via mechanistic interpretability. In The Eleventh International Conference", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 659, + 273, + 671 + ], + "spans": [ + { + "bbox": [ + 127, + 659, + 273, + 671 + ], + "score": 1.0, + "content": "on Learning Representations, 2023.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 112, + 673, + 507, + 687 + ], + "spans": [ + { + "bbox": [ + 112, + 673, + 507, + 687 + ], + "score": 1.0, + "content": "[2] Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 684, + 498, + 698 + ], + "spans": [ + { + "bbox": [ + 127, + 684, + 498, + 698 + ], + "score": 1.0, + "content": "Zoom in: An introduction to circuits. Distill, 2020. https://distill.pub/2020/circuits/zoom-in.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 112, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 112, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "[3] Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Dawn Drain,", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44 + } + ], + "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": [ + 107, + 73, + 505, + 117 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 73, + 506, + 119 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 122, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "Algorithm learning in neural networks: Emergent abilities in deep neural networks, especially", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "score": 1.0, + "content": "large language models, have recently attracted significant attention [26]. An ability is “emergent”", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "if the performance on a subtask suddenly increases with growing model sizes, though such claims", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "depend on the choice of metric [27]. It has been hypothesized that the emergence of specific capability", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "in a model corresponds to the emergence of a modular circuit responsible for that capability, and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "score": 1.0, + "content": "that emergence of some model behaviors thus results from a sequence of quantized circuit discovery", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 187, + 147, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 147, + 201 + ], + "score": 1.0, + "content": "steps [5].", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 121, + 506, + 201 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 214, + 187, + 227 + ], + "lines": [ + { + "bbox": [ + 104, + 211, + 190, + 230 + ], + "spans": [ + { + "bbox": [ + 104, + 211, + 190, + 230 + ], + "score": 1.0, + "content": "6 Conclusions", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 239, + 505, + 403 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 251 + ], + "score": 1.0, + "content": "We have offered a closer look at recent findings that familiar algorithms arise in neural networks", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "trained on specific algorithmic tasks. In modular arithmetic, we have shown that such algorithmic", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "discoveries are not inevitable: in addition to the Clock algorithm reverse-engineered by [1], we find", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "other algorithms (including a Pizza algorithm, and more complicated procedures) to be prevalent in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 282, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 506, + 296 + ], + "score": 1.0, + "content": "trained models. These different algorithmic phases can be distinguished using a variety of new and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 294, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 506, + 306 + ], + "score": 1.0, + "content": "existing interpretability techniques, including logit visualization, isolation of principle components", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 305, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 506, + 317 + ], + "score": 1.0, + "content": "in embedding space, and gradient-based measures of model symmetry. These techniques make it", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 315, + 507, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 507, + 329 + ], + "score": 1.0, + "content": "possible to automatically classify trained networks according to the algorithms they implement,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "and reveal algorithmic phase transitions in the space of model hyperparameters. Here we found", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 337, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 350 + ], + "score": 1.0, + "content": "specifically that the emergence of a Pizza or Clock algorithm depends on the relative strength of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 347, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 506, + 361 + ], + "score": 1.0, + "content": "linear layers and attention outputs. We additionally showed that these algorithms are not implemented", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "in isolation; instead, networks sometimes ensemble multiple copies of an algorithm in parallel.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "These results offer exciting new challenges for mechanistic interpretability: (1) How to find, classify,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 381, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 506, + 394 + ], + "score": 1.0, + "content": "and interpret unfamiliar algorithms in a systematic way? (2) How to disentangle multiple, parallel", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 391, + 343, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 343, + 406 + ], + "score": 1.0, + "content": "algorithm implementations in the presence of ensembling?", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 239, + 507, + 406 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 415, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "Limitations We have focused on a single learning problem: modular addition. Even in this", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "score": 1.0, + "content": "restricted domain, qualitatively different model behaviors emerge across architectures and seeds.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 436, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 505, + 450 + ], + "score": 1.0, + "content": "Significant additional work is needed to scale these techniques to the even more complex models", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 448, + 207, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 207, + 460 + ], + "score": 1.0, + "content": "used in real-world tasks.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 414, + 506, + 460 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 471, + 505, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "score": 1.0, + "content": "Broader Impact We believe interpretability techniques can play a crucial role in creating and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "score": 1.0, + "content": "improving safe AI systems. However, they may also be used to build more accurate systems, with the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "attendant risks inherent in all dual-use technologies. It is therefore necessary to exercise caution and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 505, + 358, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 358, + 517 + ], + "score": 1.0, + "content": "responsible decision-making when deploying such techniques.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 471, + 506, + 517 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 531, + 202, + 544 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 204, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 204, + 547 + ], + "score": 1.0, + "content": "Acknowledgement", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 556, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "We would like to thank Mingyang Deng and anonymous reviewers for valuable and fruitful discussions", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "and MIT SuperCloud for providing computation resources. ZL and MT are supported by the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "Foundational Questions Institute, the Rothberg Family Fund for Cognitive Science and IAIFI through", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 588, + 474, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 474, + 601 + ], + "score": 1.0, + "content": "NSF grant PHY-2019786. JA is supported by a gift from the OpenPhilanthropy Foundation.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 556, + 505, + 601 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 615, + 163, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 165, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 165, + 631 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "list", + "bbox": [ + 111, + 637, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 111, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 111, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "[1] Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. Progress", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 127, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "measures for grokking via mechanistic interpretability. In The Eleventh International Conference", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 659, + 273, + 671 + ], + "spans": [ + { + "bbox": [ + 127, + 659, + 273, + 671 + ], + "score": 1.0, + "content": "on Learning Representations, 2023.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 112, + 673, + 507, + 687 + ], + "spans": [ + { + "bbox": [ + 112, + 673, + 507, + 687 + ], + "score": 1.0, + "content": "[2] Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter.", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 684, + 498, + 698 + ], + "spans": [ + { + "bbox": [ + 127, + 684, + 498, + 698 + ], + "score": 1.0, + "content": "Zoom in: An introduction to circuits. Distill, 2020. https://distill.pub/2020/circuits/zoom-in.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 112, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 112, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "[3] Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Dawn Drain,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 127, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 127, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Scott Johnston, Andy Jones, Jackson", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 82, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 126, + 82, + 507, + 97 + ], + "score": 1.0, + "content": "Kernion, Liane Lovitt, Kamal Ndousse, Dario Amodei, Tom Brown, Jack Clark, Jared Kaplan,", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 94, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 126, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "Sam McCandlish, and Chris Olah. In-context learning and induction heads. Transformer", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 105, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 126, + 105, + 506, + 119 + ], + "score": 1.0, + "content": "Circuits Thread, 2022. https://transformer-circuits.pub/2022/in-context-learning-and-induction-", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 115, + 201, + 129 + ], + "spans": [ + { + "bbox": [ + 126, + 115, + 201, + 129 + ], + "score": 1.0, + "content": "heads/index.html.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 130, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 110, + 130, + 506, + 145 + ], + "score": 1.0, + "content": "[4] Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 141, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 126, + 141, + 506, + 155 + ], + "score": 1.0, + "content": "Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. In The", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 152, + 414, + 167 + ], + "spans": [ + { + "bbox": [ + 126, + 152, + 414, + 167 + ], + "score": 1.0, + "content": "Eleventh International Conference on Learning Representations, 2023.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 166, + 507, + 182 + ], + "spans": [ + { + "bbox": [ + 109, + 166, + 507, + 182 + ], + "score": 1.0, + "content": "[5] Eric J Michaud, Ziming Liu, Uzay Girit, and Max Tegmark. The quantization model of neural", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 178, + 325, + 191 + ], + "spans": [ + { + "bbox": [ + 125, + 178, + 325, + 191 + ], + "score": 1.0, + "content": "scaling. arXiv preprint arXiv:2303.13506, 2023.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 111, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "[6] Ekin Akyürek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. What learn-", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 124, + 202, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 124, + 202, + 506, + 218 + ], + "score": 1.0, + "content": "ing algorithm is in-context learning? investigations with linear models. In The Eleventh", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 215, + 376, + 227 + ], + "spans": [ + { + "bbox": [ + 127, + 215, + 376, + 227 + ], + "score": 1.0, + "content": "International Conference on Learning Representations, 2023.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 228, + 507, + 243 + ], + "spans": [ + { + "bbox": [ + 110, + 228, + 507, + 243 + ], + "score": 1.0, + "content": "[7] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 239, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 126, + 239, + 506, + 254 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 251, + 252, + 265 + ], + "spans": [ + { + "bbox": [ + 126, + 251, + 252, + 265 + ], + "score": 1.0, + "content": "Processing Systems, 30, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 265, + 507, + 280 + ], + "spans": [ + { + "bbox": [ + 109, + 265, + 507, + 280 + ], + "score": 1.0, + "content": "[8] Ziming Liu, Ouail Kitouni, Niklas S Nolte, Eric Michaud, Max Tegmark, and Mike Williams.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 275, + 507, + 291 + ], + "spans": [ + { + "bbox": [ + 124, + 275, + 507, + 291 + ], + "score": 1.0, + "content": "Towards understanding grokking: An effective theory of representation learning. Advances in", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 286, + 388, + 300 + ], + "spans": [ + { + "bbox": [ + 126, + 286, + 388, + 300 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, 35:34651–34663, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 110, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "[9] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 312, + 458, + 326 + ], + "spans": [ + { + "bbox": [ + 126, + 312, + 458, + 326 + ], + "score": 1.0, + "content": "neural networks. In International Conference on Learning Representations, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 326, + 507, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 507, + 341 + ], + "score": 1.0, + "content": "[10] Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann,", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 337, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 126, + 337, + 506, + 352 + ], + "score": 1.0, + "content": "Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 349, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 127, + 349, + 505, + 363 + ], + "score": 1.0, + "content": "Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt,", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 358, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 126, + 358, + 506, + 374 + ], + "score": 1.0, + "content": "Kamal Ndousse, Dario Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 127, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "Chris Olah. A mathematical framework for transformer circuits. Transformer Circuits Thread,", + "type": "text", + "cross_page": true + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 381, + 398, + 396 + ], + "spans": [ + { + "bbox": [ + 126, + 381, + 398, + 396 + ], + "score": 1.0, + "content": "2021. https://transformer-circuits.pub/2021/framework/index.html.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 396, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 396, + 506, + 410 + ], + "score": 1.0, + "content": "[11] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 128, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam", + "type": "text", + "cross_page": true + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 417, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 126, + 417, + 505, + 432 + ], + "score": 1.0, + "content": "McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy", + "type": "text", + "cross_page": true + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 429, + 374, + 443 + ], + "spans": [ + { + "bbox": [ + 126, + 429, + 374, + 443 + ], + "score": 1.0, + "content": "models of superposition. Transformer Circuits Thread, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 443, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 458 + ], + "score": 1.0, + "content": "[12] Bilal Chughtai, Lawrence Chan, and Neel Nanda. A toy model of universality: Reverse", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 454, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 127, + 454, + 506, + 469 + ], + "score": 1.0, + "content": "engineering how networks learn group operations. In The Fortieth International Conference on", + "type": "text", + "cross_page": true + } + ], + "index": 32 + }, + { + "bbox": [ + 126, + 465, + 232, + 480 + ], + "spans": [ + { + "bbox": [ + 126, + 465, + 232, + 480 + ], + "score": 1.0, + "content": "Machine Learning, 2023.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 480, + 507, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 507, + 494 + ], + "score": 1.0, + "content": "[13] Ziming Liu, Eric J Michaud, and Max Tegmark. Omnigrok: Grokking beyond algorithmic data.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 491, + 442, + 505 + ], + "spans": [ + { + "bbox": [ + 126, + 491, + 442, + 505 + ], + "score": 1.0, + "content": "In The Eleventh International Conference on Learning Representations, 2023.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "[14] Arthur Conmy, Augustine N Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 128, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 37 + }, + { + "bbox": [ + 124, + 527, + 267, + 541 + ], + "spans": [ + { + "bbox": [ + 124, + 527, + 267, + 541 + ], + "score": 1.0, + "content": "preprint arXiv:2304.14997, 2023.", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "score": 1.0, + "content": "[15] Wes Gurnee, Neel Nanda, Matthew Pauly, Katherine Harvey, Dmitrii Troitskii, and Dimitris", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 552, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 126, + 552, + 506, + 567 + ], + "score": 1.0, + "content": "Bertsimas. Finding neurons in a haystack: Case studies with sparse probing. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 40 + }, + { + "bbox": [ + 125, + 563, + 231, + 577 + ], + "spans": [ + { + "bbox": [ + 125, + 563, + 231, + 577 + ], + "score": 1.0, + "content": "arXiv:2305.01610, 2023.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "[16] Yedid Hoshen and Shmuel Peleg. Visual learning of arithmetic operation. In Proceedings of the", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 590, + 327, + 602 + ], + "spans": [ + { + "bbox": [ + 126, + 590, + 327, + 602 + ], + "score": 1.0, + "content": "AAAI Conference on Artificial Intelligence, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 604, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 506, + 619 + ], + "score": 1.0, + "content": "[17] Samuel Kim, Peter Y Lu, Srijon Mukherjee, Michael Gilbert, Li Jing, Vladimir Ceperi ˇ c, and ´", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 127, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "Marin Soljaciˇ c. Integration of neural network-based symbolic regression in deep learning for ´", + "type": "text", + "cross_page": true + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 626, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 126, + 626, + 506, + 641 + ], + "score": 1.0, + "content": "scientific discovery. IEEE Transactions on Neural Networks and Learning Systems, 32(9):4166–", + "type": "text", + "cross_page": true + } + ], + "index": 46 + }, + { + "bbox": [ + 126, + 636, + 178, + 652 + ], + "spans": [ + { + "bbox": [ + 126, + 636, + 178, + 652 + ], + "score": 1.0, + "content": "4177, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 47, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 653, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 506, + 665 + ], + "score": 1.0, + "content": "[18] Alethea Power, Yuri Burda, Harri Edwards, Igor Babuschkin, and Vedant Misra. Grokking: Gen-", + "type": "text", + "cross_page": true + } + ], + "index": 48, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 663, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 126, + 663, + 506, + 677 + ], + "score": 1.0, + "content": "eralization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177,", + "type": "text", + "cross_page": true + } + ], + "index": 49 + }, + { + "bbox": [ + 126, + 673, + 155, + 689 + ], + "spans": [ + { + "bbox": [ + 126, + 673, + 155, + 689 + ], + "score": 1.0, + "content": "2022.", + "type": "text", + "cross_page": true + } + ], + "index": 50, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 687, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 703 + ], + "score": 1.0, + "content": "[19] Boaz Barak, Benjamin Edelman, Surbhi Goel, Sham Kakade, Eran Malach, and Cyril Zhang.", + "type": "text", + "cross_page": true + } + ], + "index": 51, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 127, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "Hidden progress in deep learning: SGD learns parities near the computational limit. Advances", + "type": "text", + "cross_page": true + } + ], + "index": 52 + }, + { + "bbox": [ + 126, + 710, + 398, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 710, + 398, + 724 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, 35:21750–21764, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 53, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "[20] Yang-Hui He. Machine-learning mathematical structures. International Journal of Data Science", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 83, + 330, + 96 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 330, + 96 + ], + "score": 1.0, + "content": "in the Mathematical Sciences, 1(01):23–47, 2023.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 99, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 107, + 99, + 505, + 110 + ], + "score": 1.0, + "content": "[21] Sergei Gukov, James Halverson, Fabian Ruehle, and Piotr Sułkowski. Learning to unknot.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 110, + 387, + 123 + ], + "spans": [ + { + "bbox": [ + 127, + 110, + 387, + 123 + ], + "score": 1.0, + "content": "Machine Learning: Science and Technology, 2(2):025035, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 123, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 506, + 137 + ], + "score": 1.0, + "content": "[22] Alex Davies, Petar Velickovi ˇ c, Lars Buesing, Sam Blackwell, Daniel Zheng, Nenad Tomašev, ´", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 135, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 127, + 135, + 506, + 148 + ], + "score": 1.0, + "content": "Richard Tanburn, Peter Battaglia, Charles Blundell, András Juhász, et al. Advancing mathemat-", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 146, + 418, + 159 + ], + "spans": [ + { + "bbox": [ + 127, + 146, + 418, + 159 + ], + "score": 1.0, + "content": "ics by guiding human intuition with ai. Nature, 600(7887):70–74, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 161, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 505, + 174 + ], + "score": 1.0, + "content": "[23] Alexander K Hartmann and Martin Weigt. Phase transitions in combinatorial optimization", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 172, + 460, + 185 + ], + "spans": [ + { + "bbox": [ + 127, + 172, + 460, + 185 + ], + "score": 1.0, + "content": "problems: basics, algorithms and statistical mechanics. John Wiley & Sons, 2006.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 186, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 506, + 200 + ], + "score": 1.0, + "content": "[24] Lorenza Saitta, Attilio Giordana, and Antoine Cornuejols. Phase transitions in machine learning.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 198, + 270, + 210 + ], + "spans": [ + { + "bbox": [ + 128, + 198, + 270, + 210 + ], + "score": 1.0, + "content": "Cambridge University Press, 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 213, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 107, + 213, + 506, + 225 + ], + "score": 1.0, + "content": "[25] Ekdeep Singh Lubana, Eric J Bigelow, Robert P Dick, David Krueger, and Hidenori Tanaka.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 222, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 126, + 222, + 505, + 237 + ], + "score": 1.0, + "content": "Mechanistic mode connectivity. In International Conference on Machine Learning, pages", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 234, + 245, + 246 + ], + "spans": [ + { + "bbox": [ + 127, + 234, + 245, + 246 + ], + "score": 1.0, + "content": "22965–23004. PMLR, 2023.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "[26] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 260, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 127, + 260, + 506, + 273 + ], + "score": 1.0, + "content": "Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto,", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 270, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 127, + 270, + 505, + 285 + ], + "score": 1.0, + "content": "Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 282, + 459, + 295 + ], + "spans": [ + { + "bbox": [ + 127, + 282, + 459, + 295 + ], + "score": 1.0, + "content": "models. Transactions on Machine Learning Research, 2022. Survey Certification.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 295, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 107, + 295, + 505, + 311 + ], + "score": 1.0, + "content": "[27] Rylan Schaeffer, Brando Miranda, and Sanmi Koyejo. Are emergent abilities of large language", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 308, + 365, + 320 + ], + "spans": [ + { + "bbox": [ + 126, + 308, + 365, + 320 + ], + "score": 1.0, + "content": "models a mirage? arXiv preprint arXiv:2304.15004, 2023.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 107, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "[28] Michael Hassid, Hao Peng, Daniel Rotem, Jungo Kasai, Ivan Montero, Noah A. Smith, and Roy", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 127, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "Schwartz. How much does attention actually attend? questioning the importance of attention in", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 127, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "pretrained transformers. In Findings of the Association for Computational Linguistics: EMNLP", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 128, + 356, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 128, + 356, + 505, + 367 + ], + "score": 1.0, + "content": "2022, pages 1403–1416, Abu Dhabi, United Arab Emirates, December 2022. Association for", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 366, + 239, + 380 + ], + "spans": [ + { + "bbox": [ + 127, + 366, + 239, + 380 + ], + "score": 1.0, + "content": "Computational Linguistics.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 392, + 321, + 405 + ], + "spans": [ + { + "bbox": [ + 128, + 392, + 321, + 405 + ], + "score": 1.0, + "content": "Conference on Learning Representations, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + } + ], + "index": 44, + "bbox_fs": [ + 111, + 637, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 37, + 507, + 730 + ], + "lines": [ + { + "bbox": [ + 127, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 127, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Scott Johnston, Andy Jones, Jackson", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 82, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 126, + 82, + 507, + 97 + ], + "score": 1.0, + "content": "Kernion, Liane Lovitt, Kamal Ndousse, Dario Amodei, Tom Brown, Jack Clark, Jared Kaplan,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 94, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 126, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "Sam McCandlish, and Chris Olah. In-context learning and induction heads. Transformer", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 105, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 126, + 105, + 506, + 119 + ], + "score": 1.0, + "content": "Circuits Thread, 2022. https://transformer-circuits.pub/2022/in-context-learning-and-induction-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 115, + 201, + 129 + ], + "spans": [ + { + "bbox": [ + 126, + 115, + 201, + 129 + ], + "score": 1.0, + "content": "heads/index.html.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 130, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 110, + 130, + 506, + 145 + ], + "score": 1.0, + "content": "[4] Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 141, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 126, + 141, + 506, + 155 + ], + "score": 1.0, + "content": "Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. In The", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 152, + 414, + 167 + ], + "spans": [ + { + "bbox": [ + 126, + 152, + 414, + 167 + ], + "score": 1.0, + "content": "Eleventh International Conference on Learning Representations, 2023.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 109, + 166, + 507, + 182 + ], + "spans": [ + { + "bbox": [ + 109, + 166, + 507, + 182 + ], + "score": 1.0, + "content": "[5] Eric J Michaud, Ziming Liu, Uzay Girit, and Max Tegmark. The quantization model of neural", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 178, + 325, + 191 + ], + "spans": [ + { + "bbox": [ + 125, + 178, + 325, + 191 + ], + "score": 1.0, + "content": "scaling. arXiv preprint arXiv:2303.13506, 2023.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 111, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "[6] Ekin Akyürek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. What learn-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 124, + 202, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 124, + 202, + 506, + 218 + ], + "score": 1.0, + "content": "ing algorithm is in-context learning? investigations with linear models. In The Eleventh", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 215, + 376, + 227 + ], + "spans": [ + { + "bbox": [ + 127, + 215, + 376, + 227 + ], + "score": 1.0, + "content": "International Conference on Learning Representations, 2023.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 110, + 228, + 507, + 243 + ], + "spans": [ + { + "bbox": [ + 110, + 228, + 507, + 243 + ], + "score": 1.0, + "content": "[7] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 239, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 126, + 239, + 506, + 254 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 251, + 252, + 265 + ], + "spans": [ + { + "bbox": [ + 126, + 251, + 252, + 265 + ], + "score": 1.0, + "content": "Processing Systems, 30, 2017.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 109, + 265, + 507, + 280 + ], + "spans": [ + { + "bbox": [ + 109, + 265, + 507, + 280 + ], + "score": 1.0, + "content": "[8] Ziming Liu, Ouail Kitouni, Niklas S Nolte, Eric Michaud, Max Tegmark, and Mike Williams.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 124, + 275, + 507, + 291 + ], + "spans": [ + { + "bbox": [ + 124, + 275, + 507, + 291 + ], + "score": 1.0, + "content": "Towards understanding grokking: An effective theory of representation learning. Advances in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 286, + 388, + 300 + ], + "spans": [ + { + "bbox": [ + 126, + 286, + 388, + 300 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, 35:34651–34663, 2022.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 110, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 110, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "[9] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 312, + 458, + 326 + ], + "spans": [ + { + "bbox": [ + 126, + 312, + 458, + 326 + ], + "score": 1.0, + "content": "neural networks. In International Conference on Learning Representations, 2019.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 326, + 507, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 507, + 341 + ], + "score": 1.0, + "content": "[10] Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 337, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 126, + 337, + 506, + 352 + ], + "score": 1.0, + "content": "Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 349, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 127, + 349, + 505, + 363 + ], + "score": 1.0, + "content": "Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 358, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 126, + 358, + 506, + 374 + ], + "score": 1.0, + "content": "Kamal Ndousse, Dario Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 127, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "Chris Olah. A mathematical framework for transformer circuits. Transformer Circuits Thread,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 381, + 398, + 396 + ], + "spans": [ + { + "bbox": [ + 126, + 381, + 398, + 396 + ], + "score": 1.0, + "content": "2021. https://transformer-circuits.pub/2021/framework/index.html.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 396, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 396, + 506, + 410 + ], + "score": 1.0, + "content": "[11] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 128, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 128, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 417, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 126, + 417, + 505, + 432 + ], + "score": 1.0, + "content": "McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 429, + 374, + 443 + ], + "spans": [ + { + "bbox": [ + 126, + 429, + 374, + 443 + ], + "score": 1.0, + "content": "models of superposition. Transformer Circuits Thread, 2022.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 443, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 458 + ], + "score": 1.0, + "content": "[12] Bilal Chughtai, Lawrence Chan, and Neel Nanda. A toy model of universality: Reverse", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 454, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 127, + 454, + 506, + 469 + ], + "score": 1.0, + "content": "engineering how networks learn group operations. In The Fortieth International Conference on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 126, + 465, + 232, + 480 + ], + "spans": [ + { + "bbox": [ + 126, + 465, + 232, + 480 + ], + "score": 1.0, + "content": "Machine Learning, 2023.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 480, + 507, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 507, + 494 + ], + "score": 1.0, + "content": "[13] Ziming Liu, Eric J Michaud, and Max Tegmark. Omnigrok: Grokking beyond algorithmic data.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 491, + 442, + 505 + ], + "spans": [ + { + "bbox": [ + 126, + 491, + 442, + 505 + ], + "score": 1.0, + "content": "In The Eleventh International Conference on Learning Representations, 2023.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "[14] Arthur Conmy, Augustine N Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 128, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 128, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. arXiv", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 124, + 527, + 267, + 541 + ], + "spans": [ + { + "bbox": [ + 124, + 527, + 267, + 541 + ], + "score": 1.0, + "content": "preprint arXiv:2304.14997, 2023.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "score": 1.0, + "content": "[15] Wes Gurnee, Neel Nanda, Matthew Pauly, Katherine Harvey, Dmitrii Troitskii, and Dimitris", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 126, + 552, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 126, + 552, + 506, + 567 + ], + "score": 1.0, + "content": "Bertsimas. Finding neurons in a haystack: Case studies with sparse probing. arXiv preprint", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 125, + 563, + 231, + 577 + ], + "spans": [ + { + "bbox": [ + 125, + 563, + 231, + 577 + ], + "score": 1.0, + "content": "arXiv:2305.01610, 2023.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "[16] Yedid Hoshen and Shmuel Peleg. Visual learning of arithmetic operation. In Proceedings of the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 126, + 590, + 327, + 602 + ], + "spans": [ + { + "bbox": [ + 126, + 590, + 327, + 602 + ], + "score": 1.0, + "content": "AAAI Conference on Artificial Intelligence, 2016.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 604, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 506, + 619 + ], + "score": 1.0, + "content": "[17] Samuel Kim, Peter Y Lu, Srijon Mukherjee, Michael Gilbert, Li Jing, Vladimir Ceperi ˇ c, and ´", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 127, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "Marin Soljaciˇ c. Integration of neural network-based symbolic regression in deep learning for ´", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 626, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 126, + 626, + 506, + 641 + ], + "score": 1.0, + "content": "scientific discovery. IEEE Transactions on Neural Networks and Learning Systems, 32(9):4166–", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 126, + 636, + 178, + 652 + ], + "spans": [ + { + "bbox": [ + 126, + 636, + 178, + 652 + ], + "score": 1.0, + "content": "4177, 2020.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 653, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 506, + 665 + ], + "score": 1.0, + "content": "[18] Alethea Power, Yuri Burda, Harri Edwards, Igor Babuschkin, and Vedant Misra. Grokking: Gen-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 126, + 663, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 126, + 663, + 506, + 677 + ], + "score": 1.0, + "content": "eralization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 126, + 673, + 155, + 689 + ], + "spans": [ + { + "bbox": [ + 126, + 673, + 155, + 689 + ], + "score": 1.0, + "content": "2022.", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 687, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 703 + ], + "score": 1.0, + "content": "[19] Boaz Barak, Benjamin Edelman, Surbhi Goel, Sham Kakade, Eran Malach, and Cyril Zhang.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 127, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 127, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "Hidden progress in deep learning: SGD learns parities near the computational limit. Advances", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 126, + 710, + 398, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 710, + 398, + 724 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, 35:21750–21764, 2022.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 26.5 + } + ], + "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": [ + 105, + 37, + 507, + 730 + ], + "lines": [], + "index": 26.5, + "bbox_fs": [ + 104, + 72, + 507, + 724 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 71, + 507, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "[20] Yang-Hui He. Machine-learning mathematical structures. International Journal of Data Science", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 83, + 330, + 96 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 330, + 96 + ], + "score": 1.0, + "content": "in the Mathematical Sciences, 1(01):23–47, 2023.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 99, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 107, + 99, + 505, + 110 + ], + "score": 1.0, + "content": "[21] Sergei Gukov, James Halverson, Fabian Ruehle, and Piotr Sułkowski. Learning to unknot.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 110, + 387, + 123 + ], + "spans": [ + { + "bbox": [ + 127, + 110, + 387, + 123 + ], + "score": 1.0, + "content": "Machine Learning: Science and Technology, 2(2):025035, 2021.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 123, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 506, + 137 + ], + "score": 1.0, + "content": "[22] Alex Davies, Petar Velickovi ˇ c, Lars Buesing, Sam Blackwell, Daniel Zheng, Nenad Tomašev, ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 135, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 127, + 135, + 506, + 148 + ], + "score": 1.0, + "content": "Richard Tanburn, Peter Battaglia, Charles Blundell, András Juhász, et al. Advancing mathemat-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 146, + 418, + 159 + ], + "spans": [ + { + "bbox": [ + 127, + 146, + 418, + 159 + ], + "score": 1.0, + "content": "ics by guiding human intuition with ai. Nature, 600(7887):70–74, 2021.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 161, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 505, + 174 + ], + "score": 1.0, + "content": "[23] Alexander K Hartmann and Martin Weigt. Phase transitions in combinatorial optimization", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 172, + 460, + 185 + ], + "spans": [ + { + "bbox": [ + 127, + 172, + 460, + 185 + ], + "score": 1.0, + "content": "problems: basics, algorithms and statistical mechanics. John Wiley & Sons, 2006.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 186, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 506, + 200 + ], + "score": 1.0, + "content": "[24] Lorenza Saitta, Attilio Giordana, and Antoine Cornuejols. Phase transitions in machine learning.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 128, + 198, + 270, + 210 + ], + "spans": [ + { + "bbox": [ + 128, + 198, + 270, + 210 + ], + "score": 1.0, + "content": "Cambridge University Press, 2011.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 213, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 107, + 213, + 506, + 225 + ], + "score": 1.0, + "content": "[25] Ekdeep Singh Lubana, Eric J Bigelow, Robert P Dick, David Krueger, and Hidenori Tanaka.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 222, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 126, + 222, + 505, + 237 + ], + "score": 1.0, + "content": "Mechanistic mode connectivity. In International Conference on Machine Learning, pages", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 234, + 245, + 246 + ], + "spans": [ + { + "bbox": [ + 127, + 234, + 245, + 246 + ], + "score": 1.0, + "content": "22965–23004. PMLR, 2023.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "[26] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 260, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 127, + 260, + 506, + 273 + ], + "score": 1.0, + "content": "Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 270, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 127, + 270, + 505, + 285 + ], + "score": 1.0, + "content": "Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 282, + 459, + 295 + ], + "spans": [ + { + "bbox": [ + 127, + 282, + 459, + 295 + ], + "score": 1.0, + "content": "models. Transactions on Machine Learning Research, 2022. Survey Certification.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 295, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 107, + 295, + 505, + 311 + ], + "score": 1.0, + "content": "[27] Rylan Schaeffer, Brando Miranda, and Sanmi Koyejo. Are emergent abilities of large language", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 308, + 365, + 320 + ], + "spans": [ + { + "bbox": [ + 126, + 308, + 365, + 320 + ], + "score": 1.0, + "content": "models a mirage? arXiv preprint arXiv:2304.15004, 2023.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 107, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "[28] Michael Hassid, Hao Peng, Daniel Rotem, Jungo Kasai, Ivan Montero, Noah A. Smith, and Roy", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 127, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "Schwartz. How much does attention actually attend? questioning the importance of attention in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 127, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "pretrained transformers. In Findings of the Association for Computational Linguistics: EMNLP", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 128, + 356, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 128, + 356, + 505, + 367 + ], + "score": 1.0, + "content": "2022, pages 1403–1416, Abu Dhabi, United Arab Emirates, December 2022. Association for", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 366, + 239, + 380 + ], + "spans": [ + { + "bbox": [ + 127, + 366, + 239, + 380 + ], + "score": 1.0, + "content": "Computational Linguistics.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 128, + 392, + 321, + 405 + ], + "spans": [ + { + "bbox": [ + 128, + 392, + 321, + 405 + ], + "score": 1.0, + "content": "Conference on Learning Representations, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 13 + } + ], + "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": "list", + "bbox": [ + 105, + 71, + 507, + 405 + ], + "lines": [], + "index": 13, + "bbox_fs": [ + 106, + 73, + 506, + 405 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 214, + 69, + 396, + 88 + ], + "lines": [ + { + "bbox": [ + 213, + 68, + 398, + 90 + ], + "spans": [ + { + "bbox": [ + 213, + 68, + 398, + 90 + ], + "score": 1.0, + "content": "Supplementary material", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 105, + 101, + 434, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 434, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 434, + 119 + ], + "score": 1.0, + "content": "A Mathematical Analysis and An Example of Pizza Algorithm", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 506, + 161 + ], + "lines": [ + { + "bbox": [ + 104, + 124, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 124, + 236, + 140 + ], + "score": 1.0, + "content": "In the pizza algorithm, we have", + "type": "text" + }, + { + "bbox": [ + 236, + 126, + 503, + 138 + ], + "score": 0.7, + "content": "E _ { a b } = \\cos ( w _ { k } ( a - b ) / 2 ) \\cdot ( \\cos ( w _ { k } ( a + b ) / 2 ) , \\sin ( w _ { k } ( a + b ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 124, + 507, + 140 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 136, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 104, + 136, + 116, + 151 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 117, + 138, + 317, + 150 + ], + "score": 0.81, + "content": "\\cos x + \\cos y = \\cos ( ( x - y ) / 2 ) ( 2 \\cos ( ( x + y ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 136, + 336, + 151 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 336, + 138, + 506, + 150 + ], + "score": 0.82, + "content": "\\sin x + \\sin y = \\cos ( ( x - y ) / 2 ) ( 2 \\sin ( ( x +", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 146, + 141, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 130, + 161 + ], + "score": 0.84, + "content": "y ) / 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 146, + 141, + 163 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 105, + 164, + 506, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 137, + 178 + ], + "score": 1.0, + "content": "To get", + "type": "text" + }, + { + "bbox": [ + 137, + 164, + 355, + 177 + ], + "score": 0.87, + "content": "| \\mathrm { c o s } ( w _ { k } ( a - b ) / 2 ) | ( \\mathrm { c o s } ( w _ { k } ( a + b ) ) , \\mathrm { s i n } ( w _ { k } ( a + b ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 163, + 451, + 178 + ], + "score": 1.0, + "content": ", we generalize this to", + "type": "text" + }, + { + "bbox": [ + 451, + 165, + 506, + 177 + ], + "score": 0.87, + "content": "| \\cos ( w _ { k } ( a -", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 460, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 213, + 188 + ], + "score": 0.79, + "content": "b ) / 2 ) | \\cos ( w _ { k } ( a + b - u ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 174, + 356, + 189 + ], + "score": 1.0, + "content": "(the two given cases correspond to", + "type": "text" + }, + { + "bbox": [ + 356, + 177, + 381, + 186 + ], + "score": 0.9, + "content": "u = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 174, + 399, + 189 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 399, + 176, + 453, + 188 + ], + "score": 0.92, + "content": "u = \\pi / 2 / w _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 174, + 460, + 189 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "interline_equation", + "bbox": [ + 138, + 192, + 471, + 222 + ], + "lines": [ + { + "bbox": [ + 138, + 192, + 471, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 471, + 222 + ], + "score": 0.88, + "content": "\\begin{array} { r l r } & { } & { | ( \\cos ( w _ { k } u / 2 ) , \\sin ( w _ { k } u / 2 ) ) \\cdot E _ { a b } | = | \\cos ( w _ { k } ( a - b ) / 2 ) \\cos ( w _ { k } ( a + b - u ) / 2 ) | } \\\\ & { } & { | ( - \\sin ( w _ { k } u / 2 ) , \\cos ( w _ { k } u / 2 ) ) \\cdot E _ { a b } | = | \\cos ( w _ { k } ( a - b ) / 2 ) \\sin ( w _ { k } ( a + b - u ) / 2 ) | } \\end{array}", + "type": "interline_equation", + "image_path": "99a9e9d2a7254857d61c6bfa24e92e9119b409ba6cc8e144dc9a6963113d1e31.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 138, + 192, + 471, + 222 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 226, + 253, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 224, + 253, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 253, + 239 + ], + "score": 1.0, + "content": "thus their difference will be equal to", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 243, + 452, + 257 + ], + "lines": [ + { + "bbox": [ + 158, + 243, + 452, + 257 + ], + "spans": [ + { + "bbox": [ + 158, + 243, + 452, + 257 + ], + "score": 0.87, + "content": "| \\cos ( w _ { k } ( a - b ) / 2 ) | ( | \\cos ( w _ { k } ( a + b - u ) / 2 ) | - | \\sin ( w _ { k } ( a + b - u ) / 2 ) | ) .", + "type": "interline_equation", + "image_path": "2fa350e434aee390247637cb1bc350bce65583764bb6e7798a495579481a6360.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 158, + 243, + 452, + 257 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 262, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 260, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 154, + 276 + ], + "score": 1.0, + "content": "Now notice", + "type": "text" + }, + { + "bbox": [ + 155, + 262, + 269, + 274 + ], + "score": 0.84, + "content": "| \\cos ( t ) | - | \\sin ( t ) | \\approx \\cos ( 2 t )", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 260, + 300, + 276 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 300, + 263, + 325, + 273 + ], + "score": 0.89, + "content": "t \\in \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 260, + 505, + 276 + ], + "score": 1.0, + "content": "(Figure 8), so the difference is approximately", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 273, + 272, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 268, + 285 + ], + "score": 0.88, + "content": "| \\mathrm { c o s } ( w _ { k } ( a - b ) / 2 ) | \\mathrm { c o s } ( w _ { k } ( a + b - u ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 273, + 272, + 286 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "image", + "bbox": [ + 187, + 297, + 421, + 413 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 187, + 297, + 421, + 413 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 297, + 421, + 413 + ], + "spans": [ + { + "bbox": [ + 187, + 297, + 421, + 413 + ], + "score": 0.967, + "type": "image", + "image_path": "25459a229478da7afbc938b0a2d2ab789a18c3adb27d1be2413c5ae527e1ea20.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 187, + 297, + 421, + 313.57142857142856 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 187, + 313.57142857142856, + 421, + 330.1428571428571 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 187, + 330.1428571428571, + 421, + 346.71428571428567 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 187, + 346.71428571428567, + 421, + 363.2857142857142 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 187, + 363.2857142857142, + 421, + 379.8571428571428 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 187, + 379.8571428571428, + 421, + 396.42857142857133 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 187, + 396.42857142857133, + 421, + 412.9999999999999 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 169, + 418, + 441, + 431 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 169, + 417, + 440, + 432 + ], + "spans": [ + { + "bbox": [ + 169, + 417, + 209, + 432 + ], + "score": 1.0, + "content": "Figure 8:", + "type": "text" + }, + { + "bbox": [ + 210, + 418, + 282, + 430 + ], + "score": 0.9, + "content": "| \\cos ( t ) | - | \\sin ( t ) |", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 417, + 352, + 432 + ], + "score": 1.0, + "content": "is approximately", + "type": "text" + }, + { + "bbox": [ + 353, + 418, + 384, + 430 + ], + "score": 0.73, + "content": "\\cos ( 2 t )", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 417, + 415, + 432 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 416, + 419, + 440, + 429 + ], + "score": 0.87, + "content": "t \\in \\mathbb { R }", + "type": "inline_equation" + } + ], + "index": 19 + } + ], + "index": 19 + } + ], + "index": 17.0 + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 502, + 465 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 154, + 455 + ], + "score": 1.0, + "content": "Plugging in", + "type": "text" + }, + { + "bbox": [ + 154, + 443, + 180, + 452 + ], + "score": 0.9, + "content": "u = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 441, + 197, + 455 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 197, + 442, + 250, + 454 + ], + "score": 0.93, + "content": "u = \\pi / 2 / w _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 441, + 505, + 455 + ], + "score": 1.0, + "content": "as mentioned, we get the following particular implementation of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 453, + 188, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 188, + 466 + ], + "score": 1.0, + "content": "the pizza algorithm.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 107, + 476, + 224, + 488 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 225, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 225, + 491 + ], + "score": 1.0, + "content": "Algorithm: Pizza, Example", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 110, + 493, + 504, + 516 + ], + "lines": [ + { + "bbox": [ + 111, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 111, + 492, + 210, + 506 + ], + "score": 1.0, + "content": "Step 1 On given input", + "type": "text" + }, + { + "bbox": [ + 210, + 495, + 217, + 503 + ], + "score": 0.75, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 492, + 238, + 506 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 239, + 493, + 244, + 503 + ], + "score": 0.74, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 492, + 505, + 506 + ], + "score": 1.0, + "content": ", circularly embed them to two vectors on the circumference", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 504, + 339, + 517 + ], + "spans": [ + { + "bbox": [ + 142, + 504, + 229, + 516 + ], + "score": 0.9, + "content": "( \\cos ( w _ { k } a ) , \\sin ( w _ { k } a ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 504, + 249, + 517 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 249, + 504, + 335, + 516 + ], + "score": 0.9, + "content": "( \\cos ( w _ { k } b ) , \\sin ( w _ { k } b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 504, + 339, + 517 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 112, + 519, + 183, + 531 + ], + "lines": [ + { + "bbox": [ + 111, + 518, + 185, + 533 + ], + "spans": [ + { + "bbox": [ + 111, + 518, + 185, + 533 + ], + "score": 1.0, + "content": "Step 2 Compute:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 146, + 535, + 499, + 628 + ], + "lines": [ + { + "bbox": [ + 146, + 535, + 499, + 628 + ], + "spans": [ + { + "bbox": [ + 146, + 535, + 499, + 628 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\alpha = | \\cos ( w _ { k } a ) + \\cos ( w _ { k } b ) | / 2 - | \\sin ( w _ { k } a ) + \\sin ( w _ { k } b ) | / 2 } \\\\ & { \\quad \\approx | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a + b ) ) } \\\\ & { \\beta = | \\cos ( w _ { k } a ) + \\cos ( w _ { k } b ) + \\sin ( w _ { k } a ) + \\sin ( w _ { k } b ) | / ( 2 \\sqrt { 2 } ) } \\\\ & { \\quad - | \\cos ( w _ { k } a ) + \\cos ( w _ { k } b ) - \\sin ( w _ { k } a ) - \\sin ( w _ { k } b ) | / ( 2 \\sqrt { 2 } ) } \\\\ & { \\quad = | \\cos ( w _ { k } a - \\pi / 4 ) + \\cos ( w _ { k } b - \\pi / 4 ) | / 2 - | \\sin ( w _ { k } a - \\pi / 4 ) + \\sin ( w _ { k } b - \\pi / 4 ) | / 2 } \\\\ & { \\quad \\approx | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a + b ) - \\pi / 2 ) = | \\cos ( w _ { k } ( a - b ) / 2 ) | \\sin ( w _ { k } ( a + b ) ) } \\end{array}", + "type": "interline_equation", + "image_path": "22396d7ed4ac01a8ba85838d60d8ab7d94398fa581964c76ecc4244c63e47fd5.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 146, + 535, + 499, + 566.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 146, + 566.0, + 499, + 597.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 146, + 597.0, + 499, + 628.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 112, + 632, + 342, + 645 + ], + "lines": [ + { + "bbox": [ + 111, + 632, + 343, + 646 + ], + "spans": [ + { + "bbox": [ + 111, + 632, + 343, + 646 + ], + "score": 1.0, + "content": "Step 3 Output of this pizza is computed as a dot product.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 650, + 478, + 664 + ], + "lines": [ + { + "bbox": [ + 168, + 650, + 478, + 664 + ], + "spans": [ + { + "bbox": [ + 168, + 650, + 478, + 664 + ], + "score": 0.84, + "content": "Q _ { a b c } ^ { \\prime } = \\alpha \\cos ( w _ { k } c ) + \\beta \\sin ( w _ { k } c ) \\approx | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a + b - c ) )", + "type": "interline_equation", + "image_path": "07ca1f822ae49d607b02114515653b44d472ba661eca33f7fce790e554e9baf3.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 168, + 650, + 478, + 664 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 672, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 671, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 686 + ], + "score": 1.0, + "content": "Similar circuits are observed in the wild, but instead of the above two-term approximation, a more", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 683, + 337, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 283, + 696 + ], + "score": 1.0, + "content": "complicated one is observed. See Appendix", + "type": "text" + }, + { + "bbox": [ + 284, + 684, + 291, + 693 + ], + "score": 0.55, + "content": "\\mathrm { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 683, + 337, + 696 + ], + "score": 1.0, + "content": "for details.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 503, + 723 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 148, + 713 + ], + "score": 1.0, + "content": "The extra", + "type": "text" + }, + { + "bbox": [ + 149, + 699, + 228, + 712 + ], + "score": 0.91, + "content": "\\vert \\cos ( w _ { k } ( a - b ) / 2 ) \\vert", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 698, + 505, + 713 + ], + "score": 1.0, + "content": "term is not a coincidence. We can generalize our derivation as the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 708, + 151, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 151, + 725 + ], + "score": 1.0, + "content": "following.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + } + ], + "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": "title", + "bbox": [ + 214, + 69, + 396, + 88 + ], + "lines": [ + { + "bbox": [ + 213, + 68, + 398, + 90 + ], + "spans": [ + { + "bbox": [ + 213, + 68, + 398, + 90 + ], + "score": 1.0, + "content": "Supplementary material", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 105, + 101, + 434, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 434, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 434, + 119 + ], + "score": 1.0, + "content": "A Mathematical Analysis and An Example of Pizza Algorithm", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 100, + 434, + 119 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 506, + 161 + ], + "lines": [ + { + "bbox": [ + 104, + 124, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 124, + 236, + 140 + ], + "score": 1.0, + "content": "In the pizza algorithm, we have", + "type": "text" + }, + { + "bbox": [ + 236, + 126, + 503, + 138 + ], + "score": 0.7, + "content": "E _ { a b } = \\cos ( w _ { k } ( a - b ) / 2 ) \\cdot ( \\cos ( w _ { k } ( a + b ) / 2 ) , \\sin ( w _ { k } ( a + b ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 124, + 507, + 140 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 136, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 104, + 136, + 116, + 151 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 117, + 138, + 317, + 150 + ], + "score": 0.81, + "content": "\\cos x + \\cos y = \\cos ( ( x - y ) / 2 ) ( 2 \\cos ( ( x + y ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 136, + 336, + 151 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 336, + 138, + 506, + 150 + ], + "score": 0.82, + "content": "\\sin x + \\sin y = \\cos ( ( x - y ) / 2 ) ( 2 \\sin ( ( x +", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 146, + 141, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 130, + 161 + ], + "score": 0.84, + "content": "y ) / 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 146, + 141, + 163 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 124, + 507, + 163 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 164, + 506, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 137, + 178 + ], + "score": 1.0, + "content": "To get", + "type": "text" + }, + { + "bbox": [ + 137, + 164, + 355, + 177 + ], + "score": 0.87, + "content": "| \\mathrm { c o s } ( w _ { k } ( a - b ) / 2 ) | ( \\mathrm { c o s } ( w _ { k } ( a + b ) ) , \\mathrm { s i n } ( w _ { k } ( a + b ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 163, + 451, + 178 + ], + "score": 1.0, + "content": ", we generalize this to", + "type": "text" + }, + { + "bbox": [ + 451, + 165, + 506, + 177 + ], + "score": 0.87, + "content": "| \\cos ( w _ { k } ( a -", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 460, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 213, + 188 + ], + "score": 0.79, + "content": "b ) / 2 ) | \\cos ( w _ { k } ( a + b - u ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 174, + 356, + 189 + ], + "score": 1.0, + "content": "(the two given cases correspond to", + "type": "text" + }, + { + "bbox": [ + 356, + 177, + 381, + 186 + ], + "score": 0.9, + "content": "u = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 174, + 399, + 189 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 399, + 176, + 453, + 188 + ], + "score": 0.92, + "content": "u = \\pi / 2 / w _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 174, + 460, + 189 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 163, + 506, + 189 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 138, + 192, + 471, + 222 + ], + "lines": [ + { + "bbox": [ + 138, + 192, + 471, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 471, + 222 + ], + "score": 0.88, + "content": "\\begin{array} { r l r } & { } & { | ( \\cos ( w _ { k } u / 2 ) , \\sin ( w _ { k } u / 2 ) ) \\cdot E _ { a b } | = | \\cos ( w _ { k } ( a - b ) / 2 ) \\cos ( w _ { k } ( a + b - u ) / 2 ) | } \\\\ & { } & { | ( - \\sin ( w _ { k } u / 2 ) , \\cos ( w _ { k } u / 2 ) ) \\cdot E _ { a b } | = | \\cos ( w _ { k } ( a - b ) / 2 ) \\sin ( w _ { k } ( a + b - u ) / 2 ) | } \\end{array}", + "type": "interline_equation", + "image_path": "99a9e9d2a7254857d61c6bfa24e92e9119b409ba6cc8e144dc9a6963113d1e31.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 138, + 192, + 471, + 222 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 226, + 253, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 224, + 253, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 253, + 239 + ], + "score": 1.0, + "content": "thus their difference will be equal to", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 106, + 224, + 253, + 239 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 243, + 452, + 257 + ], + "lines": [ + { + "bbox": [ + 158, + 243, + 452, + 257 + ], + "spans": [ + { + "bbox": [ + 158, + 243, + 452, + 257 + ], + "score": 0.87, + "content": "| \\cos ( w _ { k } ( a - b ) / 2 ) | ( | \\cos ( w _ { k } ( a + b - u ) / 2 ) | - | \\sin ( w _ { k } ( a + b - u ) / 2 ) | ) .", + "type": "interline_equation", + "image_path": "2fa350e434aee390247637cb1bc350bce65583764bb6e7798a495579481a6360.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 158, + 243, + 452, + 257 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 262, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 260, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 154, + 276 + ], + "score": 1.0, + "content": "Now notice", + "type": "text" + }, + { + "bbox": [ + 155, + 262, + 269, + 274 + ], + "score": 0.84, + "content": "| \\cos ( t ) | - | \\sin ( t ) | \\approx \\cos ( 2 t )", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 260, + 300, + 276 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 300, + 263, + 325, + 273 + ], + "score": 0.89, + "content": "t \\in \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 260, + 505, + 276 + ], + "score": 1.0, + "content": "(Figure 8), so the difference is approximately", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 273, + 272, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 268, + 285 + ], + "score": 0.88, + "content": "| \\mathrm { c o s } ( w _ { k } ( a - b ) / 2 ) | \\mathrm { c o s } ( w _ { k } ( a + b - u ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 273, + 272, + 286 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 260, + 505, + 286 + ] + }, + { + "type": "image", + "bbox": [ + 187, + 297, + 421, + 413 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 187, + 297, + 421, + 413 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 297, + 421, + 413 + ], + "spans": [ + { + "bbox": [ + 187, + 297, + 421, + 413 + ], + "score": 0.967, + "type": "image", + "image_path": "25459a229478da7afbc938b0a2d2ab789a18c3adb27d1be2413c5ae527e1ea20.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 187, + 297, + 421, + 313.57142857142856 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 187, + 313.57142857142856, + 421, + 330.1428571428571 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 187, + 330.1428571428571, + 421, + 346.71428571428567 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 187, + 346.71428571428567, + 421, + 363.2857142857142 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 187, + 363.2857142857142, + 421, + 379.8571428571428 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 187, + 379.8571428571428, + 421, + 396.42857142857133 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 187, + 396.42857142857133, + 421, + 412.9999999999999 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 169, + 418, + 441, + 431 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 169, + 417, + 440, + 432 + ], + "spans": [ + { + "bbox": [ + 169, + 417, + 209, + 432 + ], + "score": 1.0, + "content": "Figure 8:", + "type": "text" + }, + { + "bbox": [ + 210, + 418, + 282, + 430 + ], + "score": 0.9, + "content": "| \\cos ( t ) | - | \\sin ( t ) |", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 417, + 352, + 432 + ], + "score": 1.0, + "content": "is approximately", + "type": "text" + }, + { + "bbox": [ + 353, + 418, + 384, + 430 + ], + "score": 0.73, + "content": "\\cos ( 2 t )", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 417, + 415, + 432 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 416, + 419, + 440, + 429 + ], + "score": 0.87, + "content": "t \\in \\mathbb { R }", + "type": "inline_equation" + } + ], + "index": 19 + } + ], + "index": 19 + } + ], + "index": 17.0 + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 502, + 465 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 154, + 455 + ], + "score": 1.0, + "content": "Plugging in", + "type": "text" + }, + { + "bbox": [ + 154, + 443, + 180, + 452 + ], + "score": 0.9, + "content": "u = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 441, + 197, + 455 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 197, + 442, + 250, + 454 + ], + "score": 0.93, + "content": "u = \\pi / 2 / w _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 441, + 505, + 455 + ], + "score": 1.0, + "content": "as mentioned, we get the following particular implementation of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 453, + 188, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 188, + 466 + ], + "score": 1.0, + "content": "the pizza algorithm.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 441, + 505, + 466 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 476, + 224, + 488 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 225, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 225, + 491 + ], + "score": 1.0, + "content": "Algorithm: Pizza, Example", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 110, + 493, + 504, + 516 + ], + "lines": [ + { + "bbox": [ + 111, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 111, + 492, + 210, + 506 + ], + "score": 1.0, + "content": "Step 1 On given input", + "type": "text" + }, + { + "bbox": [ + 210, + 495, + 217, + 503 + ], + "score": 0.75, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 492, + 238, + 506 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 239, + 493, + 244, + 503 + ], + "score": 0.74, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 492, + 505, + 506 + ], + "score": 1.0, + "content": ", circularly embed them to two vectors on the circumference", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 504, + 339, + 517 + ], + "spans": [ + { + "bbox": [ + 142, + 504, + 229, + 516 + ], + "score": 0.9, + "content": "( \\cos ( w _ { k } a ) , \\sin ( w _ { k } a ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 504, + 249, + 517 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 249, + 504, + 335, + 516 + ], + "score": 0.9, + "content": "( \\cos ( w _ { k } b ) , \\sin ( w _ { k } b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 504, + 339, + 517 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 111, + 492, + 505, + 517 + ] + }, + { + "type": "text", + "bbox": [ + 112, + 519, + 183, + 531 + ], + "lines": [ + { + "bbox": [ + 111, + 518, + 185, + 533 + ], + "spans": [ + { + "bbox": [ + 111, + 518, + 185, + 533 + ], + "score": 1.0, + "content": "Step 2 Compute:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 111, + 518, + 185, + 533 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 146, + 535, + 499, + 628 + ], + "lines": [ + { + "bbox": [ + 146, + 535, + 499, + 628 + ], + "spans": [ + { + "bbox": [ + 146, + 535, + 499, + 628 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\alpha = | \\cos ( w _ { k } a ) + \\cos ( w _ { k } b ) | / 2 - | \\sin ( w _ { k } a ) + \\sin ( w _ { k } b ) | / 2 } \\\\ & { \\quad \\approx | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a + b ) ) } \\\\ & { \\beta = | \\cos ( w _ { k } a ) + \\cos ( w _ { k } b ) + \\sin ( w _ { k } a ) + \\sin ( w _ { k } b ) | / ( 2 \\sqrt { 2 } ) } \\\\ & { \\quad - | \\cos ( w _ { k } a ) + \\cos ( w _ { k } b ) - \\sin ( w _ { k } a ) - \\sin ( w _ { k } b ) | / ( 2 \\sqrt { 2 } ) } \\\\ & { \\quad = | \\cos ( w _ { k } a - \\pi / 4 ) + \\cos ( w _ { k } b - \\pi / 4 ) | / 2 - | \\sin ( w _ { k } a - \\pi / 4 ) + \\sin ( w _ { k } b - \\pi / 4 ) | / 2 } \\\\ & { \\quad \\approx | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a + b ) - \\pi / 2 ) = | \\cos ( w _ { k } ( a - b ) / 2 ) | \\sin ( w _ { k } ( a + b ) ) } \\end{array}", + "type": "interline_equation", + "image_path": "22396d7ed4ac01a8ba85838d60d8ab7d94398fa581964c76ecc4244c63e47fd5.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 146, + 535, + 499, + 566.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 146, + 566.0, + 499, + 597.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 146, + 597.0, + 499, + 628.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 112, + 632, + 342, + 645 + ], + "lines": [ + { + "bbox": [ + 111, + 632, + 343, + 646 + ], + "spans": [ + { + "bbox": [ + 111, + 632, + 343, + 646 + ], + "score": 1.0, + "content": "Step 3 Output of this pizza is computed as a dot product.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29, + "bbox_fs": [ + 111, + 632, + 343, + 646 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 650, + 478, + 664 + ], + "lines": [ + { + "bbox": [ + 168, + 650, + 478, + 664 + ], + "spans": [ + { + "bbox": [ + 168, + 650, + 478, + 664 + ], + "score": 0.84, + "content": "Q _ { a b c } ^ { \\prime } = \\alpha \\cos ( w _ { k } c ) + \\beta \\sin ( w _ { k } c ) \\approx | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a + b - c ) )", + "type": "interline_equation", + "image_path": "07ca1f822ae49d607b02114515653b44d472ba661eca33f7fce790e554e9baf3.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 168, + 650, + 478, + 664 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 672, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 671, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 686 + ], + "score": 1.0, + "content": "Similar circuits are observed in the wild, but instead of the above two-term approximation, a more", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 683, + 337, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 283, + 696 + ], + "score": 1.0, + "content": "complicated one is observed. See Appendix", + "type": "text" + }, + { + "bbox": [ + 284, + 684, + 291, + 693 + ], + "score": 0.55, + "content": "\\mathrm { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 683, + 337, + 696 + ], + "score": 1.0, + "content": "for details.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 671, + 505, + 696 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 503, + 723 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 148, + 713 + ], + "score": 1.0, + "content": "The extra", + "type": "text" + }, + { + "bbox": [ + 149, + 699, + 228, + 712 + ], + "score": 0.91, + "content": "\\vert \\cos ( w _ { k } ( a - b ) / 2 ) \\vert", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 698, + 505, + 713 + ], + "score": 1.0, + "content": "term is not a coincidence. We can generalize our derivation as the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 708, + 151, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 151, + 725 + ], + "score": 1.0, + "content": "following.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 698, + 505, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 72, + 504, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 504, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 251, + 86 + ], + "score": 1.0, + "content": "Lemma A.1. A symmetric function", + "type": "text" + }, + { + "bbox": [ + 251, + 73, + 281, + 84 + ], + "score": 0.91, + "content": "f ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 72, + 405, + 86 + ], + "score": 1.0, + "content": "that is a linear combination of", + "type": "text" + }, + { + "bbox": [ + 406, + 73, + 504, + 84 + ], + "score": 0.84, + "content": "\\cos x , \\sin x , \\cos y , \\sin y ^ { 4 }", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 390, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 208, + 97 + ], + "score": 1.0, + "content": "can always be written as", + "type": "text" + }, + { + "bbox": [ + 208, + 84, + 307, + 96 + ], + "score": 0.88, + "content": "\\cos ( ( x - y ) / 2 ) g ( x + y )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 83, + 380, + 97 + ], + "score": 1.0, + "content": "for some function", + "type": "text" + }, + { + "bbox": [ + 380, + 86, + 386, + 95 + ], + "score": 0.68, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 83, + 390, + 97 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 505, + 141 + ], + "lines": [ + { + "bbox": [ + 104, + 105, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 104, + 105, + 165, + 120 + ], + "score": 1.0, + "content": "Proof. Notice", + "type": "text" + }, + { + "bbox": [ + 166, + 107, + 376, + 118 + ], + "score": 0.68, + "content": "\\cos x + \\cos y = \\cos ( ( x - y ) / 2 ) ( 2 \\cos ( ( x + y ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 105, + 398, + 120 + ], + "score": 1.0, + "content": ") and", + "type": "text" + }, + { + "bbox": [ + 398, + 107, + 505, + 118 + ], + "score": 0.73, + "content": "\\sin x + \\sin y = \\cos ( ( x -", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 117, + 507, + 131 + ], + "spans": [ + { + "bbox": [ + 106, + 118, + 206, + 129 + ], + "score": 0.83, + "content": "y ) / 2 ) ( 2 \\sin ( ( x + y ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 117, + 225, + 131 + ], + "score": 1.0, + "content": ", so", + "type": "text" + }, + { + "bbox": [ + 225, + 119, + 507, + 130 + ], + "score": 0.53, + "content": "\\alpha ( \\cos x + \\cos y ) + \\beta ( \\sin x + \\sin y ) = \\cos ( ( x - y ) / 2 ) ( 2 \\alpha \\cos ( ( x + \\cos y ) 2 ) + \\beta ( \\sin x + \\sin y ) = \\cos ( ( x - y ) / 2 )", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 128, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 218, + 141 + ], + "score": 0.83, + "content": "y ) / 2 ) + 2 \\beta \\sin ( ( x + y ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 128, + 224, + 142 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 495, + 129, + 505, + 139 + ], + "score": 0.997, + "content": "□", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 151, + 505, + 175 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 313, + 165 + ], + "score": 1.0, + "content": "This is why we consider the output pattern with the", + "type": "text" + }, + { + "bbox": [ + 314, + 152, + 393, + 164 + ], + "score": 0.91, + "content": "\\vert \\cos ( w _ { k } ( a - b ) / 2 ) \\vert", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "terms rather than the actual", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 163, + 390, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 390, + 176 + ], + "score": 1.0, + "content": "computation circuits as the determinant feature of the pizza algorithm.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 108, + 189, + 255, + 203 + ], + "lines": [ + { + "bbox": [ + 104, + 187, + 256, + 206 + ], + "spans": [ + { + "bbox": [ + 104, + 187, + 256, + 206 + ], + "score": 1.0, + "content": "B Non-Circular algorithms", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 506, + 281 + ], + "lines": [ + { + "bbox": [ + 107, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 107, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "One thing that further complicates our experiment is the existence of non-circular embeddings. While", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "only circular algorithms are reported in the previous works [8, 1], many non-circular embeddings", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "score": 1.0, + "content": "are found in our experiments, e.g., 1D lines or 3D Lissajous-like curves, as shown in Figure 9. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "leave the detailed analysis of these non-circular algorithms for future study. Since circular algorithms", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "are our primary focus of study, we propose the following metric circularity to filter out non-circular", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 269, + 507, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 507, + 282 + ], + "score": 1.0, + "content": "algorithms. The metric reaches maximum 1 when the principal components aligns with cosine waves.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + }, + { + "type": "image", + "bbox": [ + 109, + 294, + 502, + 425 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 294, + 502, + 425 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 294, + 502, + 425 + ], + "spans": [ + { + "bbox": [ + 109, + 294, + 502, + 425 + ], + "score": 0.97, + "type": "image", + "image_path": "60dd0aff3a24dafd4477f18335d37e63462d10075ade3f02f05c85e428094e6e.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 109, + 294, + 502, + 337.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 109, + 337.6666666666667, + 502, + 381.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 109, + 381.33333333333337, + 502, + 425.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 433, + 506, + 478 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 433, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 506, + 446 + ], + "score": 1.0, + "content": "Figure 9: Visualization of the principal components of input embeddings for two trained non-circular", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 445, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 457 + ], + "score": 1.0, + "content": "models. Top: A line-like first principal component. Notice the re-arranged x axis (token id). Bottom:", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "First three principal components forming a three-dimensional non-circular pattern. Each point", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 467, + 255, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 255, + 479 + ], + "score": 1.0, + "content": "represents the embedding of a token.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5 + } + ], + "index": 16.75 + }, + { + "type": "text", + "bbox": [ + 106, + 488, + 504, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "Definition B.1 (Circularity). For some network, suppose the l-th principal component of its input", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 499, + 473, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 165, + 514 + ], + "score": 1.0, + "content": "embeddings is", + "type": "text" + }, + { + "bbox": [ + 166, + 501, + 248, + 512 + ], + "score": 0.89, + "content": "v _ { l , 0 } , v _ { l , 1 } , \\cdots , v _ { l , p - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 499, + 473, + 514 + ], + "score": 1.0, + "content": ", define its circularity based on first four components as", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 175, + 516, + 435, + 563 + ], + "lines": [ + { + "bbox": [ + 175, + 516, + 435, + 563 + ], + "spans": [ + { + "bbox": [ + 175, + 516, + 435, + 563 + ], + "score": 0.96, + "content": "c = \\frac { 1 } { 4 } \\sum _ { l = 1 } ^ { 4 } \\left( \\operatorname* { m a x } _ { k \\in \\left[ 1 , 2 , \\cdots , p - 1 \\right] } \\left( \\frac { 2 } { p { \\sum _ { j = 0 } ^ { p - 1 } { v _ { l , j } ^ { 2 } } } } \\left| \\sum _ { j = 0 } ^ { p - 1 } v _ { l , j } e ^ { 2 \\pi i \\cdot j k / p } \\right| ^ { 2 } \\right) \\right)", + "type": "interline_equation", + "image_path": "b35466e5619ef7d4f835e41cd63134bdb37213dc402449a21de091d2c5fd7fc8.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 175, + 516, + 435, + 531.6666666666666 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 175, + 531.6666666666666, + 435, + 547.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 175, + 547.3333333333333, + 435, + 562.9999999999999 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 567, + 504, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 227, + 581 + ], + "score": 1.0, + "content": "where i is the imaginary unit.", + "type": "text" + }, + { + "bbox": [ + 227, + 567, + 264, + 579 + ], + "score": 0.9, + "content": "c \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 567, + 348, + 581 + ], + "score": 1.0, + "content": "by Fourier analysis.", + "type": "text" + }, + { + "bbox": [ + 349, + 568, + 372, + 578 + ], + "score": 0.88, + "content": "c = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 567, + 505, + 581 + ], + "score": 1.0, + "content": "means first four components are", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 578, + 168, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 168, + 590 + ], + "score": 1.0, + "content": "Fourier waves.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 504, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 376, + 611 + ], + "score": 1.0, + "content": "Both Model A and Model B in Section 3 have a circularity around", + "type": "text" + }, + { + "bbox": [ + 377, + 599, + 404, + 609 + ], + "score": 0.87, + "content": "9 9 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "and we consider models", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 609, + 248, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 169, + 622 + ], + "score": 1.0, + "content": "with circularity", + "type": "text" + }, + { + "bbox": [ + 169, + 609, + 208, + 621 + ], + "score": 0.9, + "content": "\\geq 9 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 609, + 248, + 622 + ], + "score": 1.0, + "content": "circular.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "title", + "bbox": [ + 106, + 635, + 343, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 343, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 343, + 654 + ], + "score": 1.0, + "content": "C More Results from the Main Experiments", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 660, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "Here we provide Figure 7 with non-circular networks unfiltered (Figure 10). We can see more noise", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "score": 1.0, + "content": "emerging in the plot. We also provide the training results from multi-layer transformers (Figure 11).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 691, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 689, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 118, + 689, + 506, + 704 + ], + "score": 1.0, + "content": "4The actual neural networks could be more complicated - even if our neural network is locally linear and", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 702, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 286, + 713 + ], + "score": 1.0, + "content": "symmetric, locally they could be asymmetric (e.g.", + "type": "text" + }, + { + "bbox": [ + 286, + 702, + 318, + 712 + ], + "score": 0.91, + "content": "| x | + | y |", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 702, + 378, + 713 + ], + "score": 1.0, + "content": "could locally be", + "type": "text" + }, + { + "bbox": [ + 378, + 702, + 402, + 712 + ], + "score": 0.86, + "content": "x - y )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 702, + 506, + 713 + ], + "score": 1.0, + "content": ". Nevertheless, the pattern is", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 712, + 228, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 228, + 722 + ], + "score": 1.0, + "content": "observed in our trained networks.", + "type": "text" + } + ] + } + ] + }, + { + "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": "text", + "bbox": [ + 105, + 72, + 504, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 504, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 251, + 86 + ], + "score": 1.0, + "content": "Lemma A.1. A symmetric function", + "type": "text" + }, + { + "bbox": [ + 251, + 73, + 281, + 84 + ], + "score": 0.91, + "content": "f ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 72, + 405, + 86 + ], + "score": 1.0, + "content": "that is a linear combination of", + "type": "text" + }, + { + "bbox": [ + 406, + 73, + 504, + 84 + ], + "score": 0.84, + "content": "\\cos x , \\sin x , \\cos y , \\sin y ^ { 4 }", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 390, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 208, + 97 + ], + "score": 1.0, + "content": "can always be written as", + "type": "text" + }, + { + "bbox": [ + 208, + 84, + 307, + 96 + ], + "score": 0.88, + "content": "\\cos ( ( x - y ) / 2 ) g ( x + y )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 83, + 380, + 97 + ], + "score": 1.0, + "content": "for some function", + "type": "text" + }, + { + "bbox": [ + 380, + 86, + 386, + 95 + ], + "score": 0.68, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 83, + 390, + 97 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 72, + 504, + 97 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 505, + 141 + ], + "lines": [ + { + "bbox": [ + 104, + 105, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 104, + 105, + 165, + 120 + ], + "score": 1.0, + "content": "Proof. Notice", + "type": "text" + }, + { + "bbox": [ + 166, + 107, + 376, + 118 + ], + "score": 0.68, + "content": "\\cos x + \\cos y = \\cos ( ( x - y ) / 2 ) ( 2 \\cos ( ( x + y ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 105, + 398, + 120 + ], + "score": 1.0, + "content": ") and", + "type": "text" + }, + { + "bbox": [ + 398, + 107, + 505, + 118 + ], + "score": 0.73, + "content": "\\sin x + \\sin y = \\cos ( ( x -", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 117, + 507, + 131 + ], + "spans": [ + { + "bbox": [ + 106, + 118, + 206, + 129 + ], + "score": 0.83, + "content": "y ) / 2 ) ( 2 \\sin ( ( x + y ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 117, + 225, + 131 + ], + "score": 1.0, + "content": ", so", + "type": "text" + }, + { + "bbox": [ + 225, + 119, + 507, + 130 + ], + "score": 0.53, + "content": "\\alpha ( \\cos x + \\cos y ) + \\beta ( \\sin x + \\sin y ) = \\cos ( ( x - y ) / 2 ) ( 2 \\alpha \\cos ( ( x + \\cos y ) 2 ) + \\beta ( \\sin x + \\sin y ) = \\cos ( ( x - y ) / 2 )", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 128, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 218, + 141 + ], + "score": 0.83, + "content": "y ) / 2 ) + 2 \\beta \\sin ( ( x + y ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 128, + 224, + 142 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 495, + 129, + 505, + 139 + ], + "score": 0.997, + "content": "□", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 105, + 507, + 142 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 151, + 505, + 175 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 313, + 165 + ], + "score": 1.0, + "content": "This is why we consider the output pattern with the", + "type": "text" + }, + { + "bbox": [ + 314, + 152, + 393, + 164 + ], + "score": 0.91, + "content": "\\vert \\cos ( w _ { k } ( a - b ) / 2 ) \\vert", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "terms rather than the actual", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 163, + 390, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 390, + 176 + ], + "score": 1.0, + "content": "computation circuits as the determinant feature of the pizza algorithm.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 151, + 505, + 176 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 189, + 255, + 203 + ], + "lines": [ + { + "bbox": [ + 104, + 187, + 256, + 206 + ], + "spans": [ + { + "bbox": [ + 104, + 187, + 256, + 206 + ], + "score": 1.0, + "content": "B Non-Circular algorithms", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 506, + 281 + ], + "lines": [ + { + "bbox": [ + 107, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 107, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "One thing that further complicates our experiment is the existence of non-circular embeddings. While", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "only circular algorithms are reported in the previous works [8, 1], many non-circular embeddings", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "score": 1.0, + "content": "are found in our experiments, e.g., 1D lines or 3D Lissajous-like curves, as shown in Figure 9. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "leave the detailed analysis of these non-circular algorithms for future study. Since circular algorithms", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "are our primary focus of study, we propose the following metric circularity to filter out non-circular", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 269, + 507, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 507, + 282 + ], + "score": 1.0, + "content": "algorithms. The metric reaches maximum 1 when the principal components aligns with cosine waves.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 214, + 507, + 282 + ] + }, + { + "type": "image", + "bbox": [ + 109, + 294, + 502, + 425 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 294, + 502, + 425 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 294, + 502, + 425 + ], + "spans": [ + { + "bbox": [ + 109, + 294, + 502, + 425 + ], + "score": 0.97, + "type": "image", + "image_path": "60dd0aff3a24dafd4477f18335d37e63462d10075ade3f02f05c85e428094e6e.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 109, + 294, + 502, + 337.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 109, + 337.6666666666667, + 502, + 381.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 109, + 381.33333333333337, + 502, + 425.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 433, + 506, + 478 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 433, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 506, + 446 + ], + "score": 1.0, + "content": "Figure 9: Visualization of the principal components of input embeddings for two trained non-circular", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 445, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 457 + ], + "score": 1.0, + "content": "models. Top: A line-like first principal component. Notice the re-arranged x axis (token id). Bottom:", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "First three principal components forming a three-dimensional non-circular pattern. Each point", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 467, + 255, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 255, + 479 + ], + "score": 1.0, + "content": "represents the embedding of a token.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5 + } + ], + "index": 16.75 + }, + { + "type": "text", + "bbox": [ + 106, + 488, + 504, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "Definition B.1 (Circularity). For some network, suppose the l-th principal component of its input", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 499, + 473, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 165, + 514 + ], + "score": 1.0, + "content": "embeddings is", + "type": "text" + }, + { + "bbox": [ + 166, + 501, + 248, + 512 + ], + "score": 0.89, + "content": "v _ { l , 0 } , v _ { l , 1 } , \\cdots , v _ { l , p - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 499, + 473, + 514 + ], + "score": 1.0, + "content": ", define its circularity based on first four components as", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 488, + 505, + 514 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 175, + 516, + 435, + 563 + ], + "lines": [ + { + "bbox": [ + 175, + 516, + 435, + 563 + ], + "spans": [ + { + "bbox": [ + 175, + 516, + 435, + 563 + ], + "score": 0.96, + "content": "c = \\frac { 1 } { 4 } \\sum _ { l = 1 } ^ { 4 } \\left( \\operatorname* { m a x } _ { k \\in \\left[ 1 , 2 , \\cdots , p - 1 \\right] } \\left( \\frac { 2 } { p { \\sum _ { j = 0 } ^ { p - 1 } { v _ { l , j } ^ { 2 } } } } \\left| \\sum _ { j = 0 } ^ { p - 1 } v _ { l , j } e ^ { 2 \\pi i \\cdot j k / p } \\right| ^ { 2 } \\right) \\right)", + "type": "interline_equation", + "image_path": "b35466e5619ef7d4f835e41cd63134bdb37213dc402449a21de091d2c5fd7fc8.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 175, + 516, + 435, + 531.6666666666666 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 175, + 531.6666666666666, + 435, + 547.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 175, + 547.3333333333333, + 435, + 562.9999999999999 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 567, + 504, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 227, + 581 + ], + "score": 1.0, + "content": "where i is the imaginary unit.", + "type": "text" + }, + { + "bbox": [ + 227, + 567, + 264, + 579 + ], + "score": 0.9, + "content": "c \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 567, + 348, + 581 + ], + "score": 1.0, + "content": "by Fourier analysis.", + "type": "text" + }, + { + "bbox": [ + 349, + 568, + 372, + 578 + ], + "score": 0.88, + "content": "c = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 567, + 505, + 581 + ], + "score": 1.0, + "content": "means first four components are", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 578, + 168, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 168, + 590 + ], + "score": 1.0, + "content": "Fourier waves.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 567, + 505, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 504, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 376, + 611 + ], + "score": 1.0, + "content": "Both Model A and Model B in Section 3 have a circularity around", + "type": "text" + }, + { + "bbox": [ + 377, + 599, + 404, + 609 + ], + "score": 0.87, + "content": "9 9 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "and we consider models", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 609, + 248, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 169, + 622 + ], + "score": 1.0, + "content": "with circularity", + "type": "text" + }, + { + "bbox": [ + 169, + 609, + 208, + 621 + ], + "score": 0.9, + "content": "\\geq 9 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 609, + 248, + 622 + ], + "score": 1.0, + "content": "circular.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 598, + 505, + 622 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 635, + 343, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 343, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 343, + 654 + ], + "score": 1.0, + "content": "C More Results from the Main Experiments", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 660, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "Here we provide Figure 7 with non-circular networks unfiltered (Figure 10). We can see more noise", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "score": 1.0, + "content": "emerging in the plot. We also provide the training results from multi-layer transformers (Figure 11).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 660, + 506, + 685 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 76, + 499, + 385 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 76, + 499, + 385 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 76, + 499, + 385 + ], + "spans": [ + { + "bbox": [ + 109, + 76, + 499, + 385 + ], + "score": 0.978, + "type": "image", + "image_path": "39705dfeb893c4170e800d2a99e816eb82bb1d03503e16e63da1c1f86bab3291.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 76, + 499, + 179.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 179.0, + 499, + 282.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 282.0, + 499, + 385.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 395, + 506, + 429 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "Figure 10: Training results from 1-layer transformers. Each point in the plots represents a training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 161, + 419 + ], + "score": 1.0, + "content": "run reaching", + "type": "text" + }, + { + "bbox": [ + 162, + 407, + 186, + 417 + ], + "score": 0.85, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 406, + 455, + 419 + ], + "score": 1.0, + "content": "validation accuracy. Among all the trained 1-layer transformers,", + "type": "text" + }, + { + "bbox": [ + 456, + 407, + 488, + 417 + ], + "score": 0.87, + "content": "3 4 . 3 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 418, + 399, + 429 + ], + "spans": [ + { + "bbox": [ + 107, + 418, + 399, + 429 + ], + "score": 1.0, + "content": "circular. Top: Model width fixed to be 128. Bottom: Model width varies.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 450, + 236, + 463 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 237, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 237, + 465 + ], + "score": 1.0, + "content": "D Pizzas Come in Pairs", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "Cautious readers might notice that the pizza algorithm is imperfect - for near antipodal points, the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 487, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 498 + ], + "score": 1.0, + "content": "sum vector will have a very small norm and the result will be noise-sensitive. While the problem", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "is partially elevated by the use of multiple circles instead of one, we also noticed another pattern", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 510, + 241, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 241, + 521 + ], + "score": 1.0, + "content": "emerged: accompanying pizzas.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 423, + 538 + ], + "score": 1.0, + "content": "The idea is the following: suppose the difference between adjacent points is", + "type": "text" + }, + { + "bbox": [ + 424, + 525, + 435, + 535 + ], + "score": 0.79, + "content": "2 k", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 524, + 458, + 538 + ], + "score": 1.0, + "content": "mod", + "type": "text" + }, + { + "bbox": [ + 459, + 527, + 465, + 536 + ], + "score": 0.7, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 524, + 506, + 538 + ], + "score": 1.0, + "content": ", then the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 239, + 547 + ], + "score": 1.0, + "content": "antipodal points have difference", + "type": "text" + }, + { + "bbox": [ + 239, + 536, + 254, + 546 + ], + "score": 0.86, + "content": "\\pm k", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 536, + 482, + 547 + ], + "score": 1.0, + "content": ". Therefore, if we arrange a new circle with a difference", + "type": "text" + }, + { + "bbox": [ + 483, + 536, + 489, + 546 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 547, + 433, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 433, + 560 + ], + "score": 1.0, + "content": "adjacent points, we will get a pizza that works best for formerly antipodal points.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 108, + 571, + 248, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 249, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 249, + 586 + ], + "score": 1.0, + "content": "Algorithm: Accompanying Pizza", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 111, + 587, + 504, + 611 + ], + "lines": [ + { + "bbox": [ + 111, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 111, + 587, + 163, + 601 + ], + "score": 1.0, + "content": "Step 1 Take", + "type": "text" + }, + { + "bbox": [ + 163, + 589, + 177, + 599 + ], + "score": 0.86, + "content": "w _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 587, + 354, + 601 + ], + "score": 1.0, + "content": "as of the accompanied pizza. On given input", + "type": "text" + }, + { + "bbox": [ + 354, + 590, + 361, + 597 + ], + "score": 0.76, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 587, + 378, + 601 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 379, + 588, + 384, + 597 + ], + "score": 0.77, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 587, + 506, + 601 + ], + "score": 1.0, + "content": ", circularly embed them to two", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 598, + 477, + 611 + ], + "spans": [ + { + "bbox": [ + 142, + 598, + 261, + 611 + ], + "score": 1.0, + "content": "vectors on the circumference", + "type": "text" + }, + { + "bbox": [ + 261, + 599, + 358, + 611 + ], + "score": 0.86, + "content": "( \\cos ( 2 w _ { k } a ) , \\sin ( 2 w _ { k } a ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 598, + 378, + 611 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 378, + 599, + 474, + 611 + ], + "score": 0.81, + "content": "( \\cos ( 2 w _ { k } b ) , \\sin ( 2 w _ { k } b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 598, + 477, + 611 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 112, + 614, + 237, + 626 + ], + "lines": [ + { + "bbox": [ + 111, + 613, + 237, + 628 + ], + "spans": [ + { + "bbox": [ + 111, + 613, + 237, + 628 + ], + "score": 1.0, + "content": "Step 2 Compute the midpoint:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 632, + 441, + 657 + ], + "lines": [ + { + "bbox": [ + 207, + 632, + 441, + 657 + ], + "spans": [ + { + "bbox": [ + 207, + 632, + 441, + 657 + ], + "score": 0.91, + "content": "s = { \\frac { 1 } { 2 } } ( \\cos ( 2 w _ { k } a ) + \\cos ( 2 w _ { k } b ) , \\sin ( 2 w _ { k } a ) + \\sin ( 2 w _ { k } b ) )", + "type": "interline_equation", + "image_path": "1e6cf8ed36fef9d92dcdbd371a70bbed922e5e824ad7df98c2b822c61f0b86af.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 207, + 632, + 441, + 657 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 112, + 664, + 342, + 676 + ], + "lines": [ + { + "bbox": [ + 112, + 663, + 343, + 678 + ], + "spans": [ + { + "bbox": [ + 112, + 663, + 343, + 678 + ], + "score": 1.0, + "content": "Step 3 Output of this pizza is computed as a dot product.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 682, + 390, + 696 + ], + "lines": [ + { + "bbox": [ + 257, + 682, + 390, + 696 + ], + "spans": [ + { + "bbox": [ + 257, + 682, + 390, + 696 + ], + "score": 0.9, + "content": "A _ { c } = - ( \\cos ( w _ { k } c ) , \\sin ( w _ { k } c ) ) \\cdot s", + "type": "interline_equation", + "image_path": "fdf88f847c21a28651926eb2c048c93e4eae3af679e27307acd08d82267b3ce1.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 257, + 682, + 390, + 696 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 705, + 505, + 729 + ], + "lines": [ + { + "bbox": [ + 106, + 705, + 505, + 718 + ], + "spans": [ + { + "bbox": [ + 106, + 705, + 505, + 718 + ], + "score": 1.0, + "content": "This is exactly what we observed in Model A (Table 3, Figure 13). With the six circles (pizzas and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 716, + 450, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 716, + 384, + 732 + ], + "score": 1.0, + "content": "accompanying pizzas) included in the embedding, Model A also gets", + "type": "text" + }, + { + "bbox": [ + 384, + 717, + 408, + 727 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 716, + 450, + 732 + ], + "score": 1.0, + "content": "accuracy.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + } + ], + "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": "image", + "bbox": [ + 109, + 76, + 499, + 385 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 76, + 499, + 385 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 76, + 499, + 385 + ], + "spans": [ + { + "bbox": [ + 109, + 76, + 499, + 385 + ], + "score": 0.978, + "type": "image", + "image_path": "39705dfeb893c4170e800d2a99e816eb82bb1d03503e16e63da1c1f86bab3291.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 76, + 499, + 179.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 179.0, + 499, + 282.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 282.0, + 499, + 385.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 395, + 506, + 429 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "Figure 10: Training results from 1-layer transformers. Each point in the plots represents a training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 161, + 419 + ], + "score": 1.0, + "content": "run reaching", + "type": "text" + }, + { + "bbox": [ + 162, + 407, + 186, + 417 + ], + "score": 0.85, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 406, + 455, + 419 + ], + "score": 1.0, + "content": "validation accuracy. Among all the trained 1-layer transformers,", + "type": "text" + }, + { + "bbox": [ + 456, + 407, + 488, + 417 + ], + "score": 0.87, + "content": "3 4 . 3 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 418, + 399, + 429 + ], + "spans": [ + { + "bbox": [ + 107, + 418, + 399, + 429 + ], + "score": 1.0, + "content": "circular. Top: Model width fixed to be 128. Bottom: Model width varies.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 450, + 236, + 463 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 237, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 237, + 465 + ], + "score": 1.0, + "content": "D Pizzas Come in Pairs", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "Cautious readers might notice that the pizza algorithm is imperfect - for near antipodal points, the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 487, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 498 + ], + "score": 1.0, + "content": "sum vector will have a very small norm and the result will be noise-sensitive. While the problem", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "is partially elevated by the use of multiple circles instead of one, we also noticed another pattern", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 510, + 241, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 241, + 521 + ], + "score": 1.0, + "content": "emerged: accompanying pizzas.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 475, + 505, + 521 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 423, + 538 + ], + "score": 1.0, + "content": "The idea is the following: suppose the difference between adjacent points is", + "type": "text" + }, + { + "bbox": [ + 424, + 525, + 435, + 535 + ], + "score": 0.79, + "content": "2 k", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 524, + 458, + 538 + ], + "score": 1.0, + "content": "mod", + "type": "text" + }, + { + "bbox": [ + 459, + 527, + 465, + 536 + ], + "score": 0.7, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 524, + 506, + 538 + ], + "score": 1.0, + "content": ", then the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 239, + 547 + ], + "score": 1.0, + "content": "antipodal points have difference", + "type": "text" + }, + { + "bbox": [ + 239, + 536, + 254, + 546 + ], + "score": 0.86, + "content": "\\pm k", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 536, + 482, + 547 + ], + "score": 1.0, + "content": ". Therefore, if we arrange a new circle with a difference", + "type": "text" + }, + { + "bbox": [ + 483, + 536, + 489, + 546 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 547, + 433, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 433, + 560 + ], + "score": 1.0, + "content": "adjacent points, we will get a pizza that works best for formerly antipodal points.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 524, + 506, + 560 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 571, + 248, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 249, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 249, + 586 + ], + "score": 1.0, + "content": "Algorithm: Accompanying Pizza", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 111, + 587, + 504, + 611 + ], + "lines": [ + { + "bbox": [ + 111, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 111, + 587, + 163, + 601 + ], + "score": 1.0, + "content": "Step 1 Take", + "type": "text" + }, + { + "bbox": [ + 163, + 589, + 177, + 599 + ], + "score": 0.86, + "content": "w _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 587, + 354, + 601 + ], + "score": 1.0, + "content": "as of the accompanied pizza. On given input", + "type": "text" + }, + { + "bbox": [ + 354, + 590, + 361, + 597 + ], + "score": 0.76, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 587, + 378, + 601 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 379, + 588, + 384, + 597 + ], + "score": 0.77, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 587, + 506, + 601 + ], + "score": 1.0, + "content": ", circularly embed them to two", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 598, + 477, + 611 + ], + "spans": [ + { + "bbox": [ + 142, + 598, + 261, + 611 + ], + "score": 1.0, + "content": "vectors on the circumference", + "type": "text" + }, + { + "bbox": [ + 261, + 599, + 358, + 611 + ], + "score": 0.86, + "content": "( \\cos ( 2 w _ { k } a ) , \\sin ( 2 w _ { k } a ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 598, + 378, + 611 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 378, + 599, + 474, + 611 + ], + "score": 0.81, + "content": "( \\cos ( 2 w _ { k } b ) , \\sin ( 2 w _ { k } b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 598, + 477, + 611 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 111, + 587, + 506, + 611 + ] + }, + { + "type": "text", + "bbox": [ + 112, + 614, + 237, + 626 + ], + "lines": [ + { + "bbox": [ + 111, + 613, + 237, + 628 + ], + "spans": [ + { + "bbox": [ + 111, + 613, + 237, + 628 + ], + "score": 1.0, + "content": "Step 2 Compute the midpoint:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 111, + 613, + 237, + 628 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 632, + 441, + 657 + ], + "lines": [ + { + "bbox": [ + 207, + 632, + 441, + 657 + ], + "spans": [ + { + "bbox": [ + 207, + 632, + 441, + 657 + ], + "score": 0.91, + "content": "s = { \\frac { 1 } { 2 } } ( \\cos ( 2 w _ { k } a ) + \\cos ( 2 w _ { k } b ) , \\sin ( 2 w _ { k } a ) + \\sin ( 2 w _ { k } b ) )", + "type": "interline_equation", + "image_path": "1e6cf8ed36fef9d92dcdbd371a70bbed922e5e824ad7df98c2b822c61f0b86af.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 207, + 632, + 441, + 657 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 112, + 664, + 342, + 676 + ], + "lines": [ + { + "bbox": [ + 112, + 663, + 343, + 678 + ], + "spans": [ + { + "bbox": [ + 112, + 663, + 343, + 678 + ], + "score": 1.0, + "content": "Step 3 Output of this pizza is computed as a dot product.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 112, + 663, + 343, + 678 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 682, + 390, + 696 + ], + "lines": [ + { + "bbox": [ + 257, + 682, + 390, + 696 + ], + "spans": [ + { + "bbox": [ + 257, + 682, + 390, + 696 + ], + "score": 0.9, + "content": "A _ { c } = - ( \\cos ( w _ { k } c ) , \\sin ( w _ { k } c ) ) \\cdot s", + "type": "interline_equation", + "image_path": "fdf88f847c21a28651926eb2c048c93e4eae3af679e27307acd08d82267b3ce1.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 257, + 682, + 390, + 696 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 705, + 505, + 729 + ], + "lines": [ + { + "bbox": [ + 106, + 705, + 505, + 718 + ], + "spans": [ + { + "bbox": [ + 106, + 705, + 505, + 718 + ], + "score": 1.0, + "content": "This is exactly what we observed in Model A (Table 3, Figure 13). With the six circles (pizzas and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 716, + 450, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 716, + 384, + 732 + ], + "score": 1.0, + "content": "accompanying pizzas) included in the embedding, Model A also gets", + "type": "text" + }, + { + "bbox": [ + 384, + 717, + 408, + 727 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 716, + 450, + 732 + ], + "score": 1.0, + "content": "accuracy.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 705, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 72, + 502, + 450 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 72, + 502, + 450 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 72, + 502, + 450 + ], + "spans": [ + { + "bbox": [ + 109, + 72, + 502, + 450 + ], + "score": 0.976, + "type": "image", + "image_path": "ab8f6c5657ee7abe49da6fb6d9f1cfe7f621c533c464eb6e9838877566ef324e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 72, + 502, + 198.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 198.0, + 502, + 324.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 324.0, + 502, + 450.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 457, + 504, + 480 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "Figure 11: Training results from transformers with 2, 3 and 4 layers. Among all the trained", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 468, + 460, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 250, + 482 + ], + "score": 1.0, + "content": "transformers with 2, 3 and 4 layers,", + "type": "text" + }, + { + "bbox": [ + 250, + 469, + 277, + 479 + ], + "score": 0.85, + "content": "9 . 9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 468, + 281, + 482 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 281, + 469, + 313, + 479 + ], + "score": 0.86, + "content": "1 1 . 5 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 468, + 331, + 482 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 331, + 469, + 358, + 479 + ], + "score": 0.88, + "content": "6 . 0 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 468, + 460, + 482 + ], + "score": 1.0, + "content": "are circular, respectively.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 147, + 489, + 461, + 672 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 147, + 489, + 461, + 672 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 147, + 489, + 461, + 672 + ], + "spans": [ + { + "bbox": [ + 147, + 489, + 461, + 672 + ], + "score": 0.914, + "type": "image", + "image_path": "a4c73059a68ba2a9da455769c1f5731c615d66815fd2a58be9ffe4ca6d957dce.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 147, + 489, + 461, + 550.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 147, + 550.0, + 461, + 611.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 147, + 611.0, + 461, + 672.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 173, + 677, + 438, + 689 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 173, + 676, + 437, + 691 + ], + "spans": [ + { + "bbox": [ + 173, + 676, + 437, + 691 + ], + "score": 1.0, + "content": "Figure 12: An Illustration on the Accompanying Pizza Algorithm", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 7.0 + } + ], + "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": "image", + "bbox": [ + 109, + 72, + 502, + 450 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 72, + 502, + 450 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 72, + 502, + 450 + ], + "spans": [ + { + "bbox": [ + 109, + 72, + 502, + 450 + ], + "score": 0.976, + "type": "image", + "image_path": "ab8f6c5657ee7abe49da6fb6d9f1cfe7f621c533c464eb6e9838877566ef324e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 72, + 502, + 198.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 198.0, + 502, + 324.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 324.0, + 502, + 450.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 457, + 504, + 480 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "Figure 11: Training results from transformers with 2, 3 and 4 layers. Among all the trained", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 468, + 460, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 250, + 482 + ], + "score": 1.0, + "content": "transformers with 2, 3 and 4 layers,", + "type": "text" + }, + { + "bbox": [ + 250, + 469, + 277, + 479 + ], + "score": 0.85, + "content": "9 . 9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 468, + 281, + 482 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 281, + 469, + 313, + 479 + ], + "score": 0.86, + "content": "1 1 . 5 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 468, + 331, + 482 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 331, + 469, + 358, + 479 + ], + "score": 0.88, + "content": "6 . 0 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 468, + 460, + 482 + ], + "score": 1.0, + "content": "are circular, respectively.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 147, + 489, + 461, + 672 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 147, + 489, + 461, + 672 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 147, + 489, + 461, + 672 + ], + "spans": [ + { + "bbox": [ + 147, + 489, + 461, + 672 + ], + "score": 0.914, + "type": "image", + "image_path": "a4c73059a68ba2a9da455769c1f5731c615d66815fd2a58be9ffe4ca6d957dce.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 147, + 489, + 461, + 550.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 147, + 550.0, + 461, + 611.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 147, + 611.0, + 461, + 672.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 173, + 677, + 438, + 689 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 173, + 676, + 437, + 691 + ], + "spans": [ + { + "bbox": [ + 173, + 676, + 437, + 691 + ], + "score": 1.0, + "content": "Figure 12: An Illustration on the Accompanying Pizza Algorithm", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 7.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 202, + 70, + 406, + 117 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 202, + 70, + 406, + 117 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 202, + 70, + 406, + 117 + ], + "spans": [ + { + "bbox": [ + 202, + 70, + 406, + 117 + ], + "score": 0.972, + "html": "
CircleWkAc FVE
#4 (accompanying #1)2π/59·1797.56%
#5 (accompanying #2)2π/59.397.23%
#6 (accompanying #3)2π/59·4497.69%
", + "type": "table", + "image_path": "5e37af99bd1fbc6dec5e7d6cf988a6e708ed329f036637611f2cc08df2710cb4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 202, + 70, + 406, + 85.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 202, + 85.66666666666667, + 406, + 101.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 202, + 101.33333333333334, + 406, + 117.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 506, + 168 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Table 3: After isolating accompanying circles in the input embedding, fraction of variance explained", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 135, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 106, + 135, + 505, + 146 + ], + "score": 1.0, + "content": "(FVE) of all Model A’s output logits by various formulas. Both model output logits and formula", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "results’ are normalized to mean 0 variance 1 before taking FVE. Accompanying and accompanied", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 156, + 207, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 188, + 169 + ], + "score": 1.0, + "content": "pizza have the same", + "type": "text" + }, + { + "bbox": [ + 189, + 158, + 201, + 167 + ], + "score": 0.85, + "content": "w _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 156, + 207, + 169 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 189, + 189, + 423, + 199 + ], + "lines": [ + { + "bbox": [ + 189, + 189, + 423, + 199 + ], + "spans": [], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 110, + 191, + 502, + 459 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 191, + 502, + 459 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 191, + 502, + 459 + ], + "spans": [ + { + "bbox": [ + 110, + 191, + 502, + 459 + ], + "score": 0.973, + "type": "image", + "image_path": "01bf6b5de01736cd06d52ccf64b92caac2e8f38abe74f3c56b9c06e02d77f1b0.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 110, + 191, + 502, + 280.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 280.3333333333333, + 502, + 369.66666666666663 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 369.66666666666663, + 502, + 458.99999999999994 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 472, + 504, + 495 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "score": 1.0, + "content": "Figure 13: Correct logits of Model A (Pizza) after circle isolation. Only accompanying pizzas are", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 483, + 354, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 354, + 496 + ], + "score": 1.0, + "content": "displayed. Notice the complementing logit pattern (Figure 4).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + } + ], + "index": 10.25 + }, + { + "type": "title", + "bbox": [ + 107, + 514, + 322, + 529 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 322, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 322, + 531 + ], + "score": 1.0, + "content": "E Results in Other Linear Architectures", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 540, + 504, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 506, + 552 + ], + "score": 1.0, + "content": "While this is not the primary focus of our paper, we also ran experiments on the following four", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 550, + 361, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 361, + 564 + ], + "score": 1.0, + "content": "different linear model setups (see Section F.2 for setup details).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 133, + 572, + 505, + 690 + ], + "lines": [ + { + "bbox": [ + 132, + 571, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 330, + 586 + ], + "score": 1.0, + "content": "• For all the models, we first encode input tokens", + "type": "text" + }, + { + "bbox": [ + 330, + 573, + 353, + 585 + ], + "score": 0.91, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 571, + 485, + 586 + ], + "score": 1.0, + "content": "with a trainable embedding layer", + "type": "text" + }, + { + "bbox": [ + 485, + 573, + 502, + 583 + ], + "score": 0.91, + "content": "W _ { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 571, + 506, + 586 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 142, + 583, + 191, + 595 + ], + "score": 0.83, + "content": "x _ { 1 } = W _ { E , a }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 583, + 197, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 198, + 583, + 245, + 595 + ], + "score": 0.84, + "content": "x _ { 2 } = W _ { E , b }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 583, + 444, + 596 + ], + "score": 1.0, + "content": "(positional embedding removed for simplicity).", + "type": "text" + }, + { + "bbox": [ + 444, + 583, + 488, + 595 + ], + "score": 0.92, + "content": "L _ { 1 } , L _ { 2 } , L _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 593, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 141, + 593, + 505, + 608 + ], + "score": 1.0, + "content": "trainable linear layers. The outmost layers (commonly referred as unembed layers) have no", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 604, + 392, + 618 + ], + "spans": [ + { + "bbox": [ + 141, + 604, + 392, + 618 + ], + "score": 1.0, + "content": "biases and the other layers have biases included for generality.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 619, + 390, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 619, + 171, + 632 + ], + "score": 1.0, + "content": "• Model", + "type": "text" + }, + { + "bbox": [ + 171, + 622, + 178, + 630 + ], + "score": 0.67, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 619, + 284, + 632 + ], + "score": 1.0, + "content": ": calculate output logits as", + "type": "text" + }, + { + "bbox": [ + 284, + 619, + 386, + 632 + ], + "score": 0.88, + "content": "L _ { 2 } ( \\mathrm { R e L U } ( L _ { 1 } ( x _ { 1 } + x _ { 2 } ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 619, + 390, + 632 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 634, + 442, + 648 + ], + "spans": [ + { + "bbox": [ + 132, + 634, + 171, + 648 + ], + "score": 1.0, + "content": "• Model", + "type": "text" + }, + { + "bbox": [ + 171, + 635, + 178, + 647 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 634, + 284, + 648 + ], + "score": 1.0, + "content": ": calculate output logits as", + "type": "text" + }, + { + "bbox": [ + 285, + 635, + 437, + 648 + ], + "score": 0.88, + "content": "L _ { 3 } ( \\mathrm { R e L U } ( L _ { 2 } ( \\mathrm { R e L U } ( L _ { 1 } ( x _ { 1 } + x _ { 2 } ) ) ) ) ) ,", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 634, + 442, + 648 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 649, + 461, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 649, + 170, + 664 + ], + "score": 1.0, + "content": "• Model", + "type": "text" + }, + { + "bbox": [ + 171, + 652, + 178, + 662 + ], + "score": 0.78, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 649, + 284, + 664 + ], + "score": 1.0, + "content": ": calculate output logits as", + "type": "text" + }, + { + "bbox": [ + 284, + 650, + 457, + 663 + ], + "score": 0.88, + "content": "L _ { 3 } ( \\mathrm { R e L U } ( L _ { 2 } ( \\mathrm { R e L U } ( L _ { 1 } ( x _ { 1 } ) + L _ { 1 } ( x _ { 2 } ) ) ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 649, + 461, + 664 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 133, + 663, + 385, + 679 + ], + "spans": [ + { + "bbox": [ + 133, + 663, + 171, + 679 + ], + "score": 1.0, + "content": "• Model", + "type": "text" + }, + { + "bbox": [ + 171, + 666, + 177, + 675 + ], + "score": 0.58, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 663, + 283, + 679 + ], + "score": 1.0, + "content": ": calculate output logits as", + "type": "text" + }, + { + "bbox": [ + 283, + 665, + 382, + 677 + ], + "score": 0.88, + "content": "L _ { 2 } \\big ( \\mathrm { R e L U } ( L _ { 1 } ( [ x _ { 1 } ; x _ { 2 } ] ) ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 663, + 385, + 679 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 678, + 349, + 691 + ], + "spans": [ + { + "bbox": [ + 142, + 678, + 145, + 691 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 145, + 678, + 177, + 690 + ], + "score": 0.91, + "content": "[ x _ { 1 } ; x _ { 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 678, + 303, + 691 + ], + "score": 1.0, + "content": "stands for the concatenation of", + "type": "text" + }, + { + "bbox": [ + 303, + 680, + 314, + 689 + ], + "score": 0.85, + "content": "x _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 678, + 332, + 691 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 333, + 681, + 344, + 689 + ], + "score": 0.77, + "content": "x _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 678, + 349, + 691 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 504, + 723 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 351, + 712 + ], + "score": 1.0, + "content": "The results are shown in Figure 14. Rather surprisingly, Model", + "type": "text" + }, + { + "bbox": [ + 351, + 702, + 359, + 710 + ], + "score": 0.74, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 699, + 389, + 712 + ], + "score": 1.0, + "content": ", Model", + "type": "text" + }, + { + "bbox": [ + 390, + 700, + 397, + 711 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 699, + 442, + 712 + ], + "score": 1.0, + "content": "and Model", + "type": "text" + }, + { + "bbox": [ + 442, + 700, + 448, + 710 + ], + "score": 0.78, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 699, + 504, + 712 + ], + "score": 1.0, + "content": "gave radically", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 207, + 723 + ], + "score": 1.0, + "content": "different results. Model", + "type": "text" + }, + { + "bbox": [ + 207, + 711, + 215, + 722 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 710, + 263, + 723 + ], + "score": 1.0, + "content": "and Model", + "type": "text" + }, + { + "bbox": [ + 263, + 712, + 271, + 722 + ], + "score": 0.82, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "are very similar, and in general they are more pizza-like", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + } + ], + "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": "table", + "bbox": [ + 202, + 70, + 406, + 117 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 202, + 70, + 406, + 117 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 202, + 70, + 406, + 117 + ], + "spans": [ + { + "bbox": [ + 202, + 70, + 406, + 117 + ], + "score": 0.972, + "html": "
CircleWkAc FVE
#4 (accompanying #1)2π/59·1797.56%
#5 (accompanying #2)2π/59.397.23%
#6 (accompanying #3)2π/59·4497.69%
", + "type": "table", + "image_path": "5e37af99bd1fbc6dec5e7d6cf988a6e708ed329f036637611f2cc08df2710cb4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 202, + 70, + 406, + 85.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 202, + 85.66666666666667, + 406, + 101.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 202, + 101.33333333333334, + 406, + 117.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 506, + 168 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Table 3: After isolating accompanying circles in the input embedding, fraction of variance explained", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 135, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 106, + 135, + 505, + 146 + ], + "score": 1.0, + "content": "(FVE) of all Model A’s output logits by various formulas. Both model output logits and formula", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "results’ are normalized to mean 0 variance 1 before taking FVE. Accompanying and accompanied", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 156, + 207, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 188, + 169 + ], + "score": 1.0, + "content": "pizza have the same", + "type": "text" + }, + { + "bbox": [ + 189, + 158, + 201, + 167 + ], + "score": 0.85, + "content": "w _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 156, + 207, + 169 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 123, + 505, + 169 + ] + }, + { + "type": "title", + "bbox": [ + 189, + 189, + 423, + 199 + ], + "lines": [ + { + "bbox": [ + 189, + 189, + 423, + 199 + ], + "spans": [], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 110, + 191, + 502, + 459 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 191, + 502, + 459 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 191, + 502, + 459 + ], + "spans": [ + { + "bbox": [ + 110, + 191, + 502, + 459 + ], + "score": 0.973, + "type": "image", + "image_path": "01bf6b5de01736cd06d52ccf64b92caac2e8f38abe74f3c56b9c06e02d77f1b0.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 110, + 191, + 502, + 280.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 280.3333333333333, + 502, + 369.66666666666663 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 369.66666666666663, + 502, + 458.99999999999994 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 472, + 504, + 495 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "score": 1.0, + "content": "Figure 13: Correct logits of Model A (Pizza) after circle isolation. Only accompanying pizzas are", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 483, + 354, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 354, + 496 + ], + "score": 1.0, + "content": "displayed. Notice the complementing logit pattern (Figure 4).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + } + ], + "index": 10.25 + }, + { + "type": "title", + "bbox": [ + 107, + 514, + 322, + 529 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 322, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 322, + 531 + ], + "score": 1.0, + "content": "E Results in Other Linear Architectures", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 540, + 504, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 506, + 552 + ], + "score": 1.0, + "content": "While this is not the primary focus of our paper, we also ran experiments on the following four", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 550, + 361, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 361, + 564 + ], + "score": 1.0, + "content": "different linear model setups (see Section F.2 for setup details).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 538, + 506, + 564 + ] + }, + { + "type": "list", + "bbox": [ + 133, + 572, + 505, + 690 + ], + "lines": [ + { + "bbox": [ + 132, + 571, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 330, + 586 + ], + "score": 1.0, + "content": "• For all the models, we first encode input tokens", + "type": "text" + }, + { + "bbox": [ + 330, + 573, + 353, + 585 + ], + "score": 0.91, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 571, + 485, + 586 + ], + "score": 1.0, + "content": "with a trainable embedding layer", + "type": "text" + }, + { + "bbox": [ + 485, + 573, + 502, + 583 + ], + "score": 0.91, + "content": "W _ { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 571, + 506, + 586 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 142, + 583, + 191, + 595 + ], + "score": 0.83, + "content": "x _ { 1 } = W _ { E , a }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 583, + 197, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 198, + 583, + 245, + 595 + ], + "score": 0.84, + "content": "x _ { 2 } = W _ { E , b }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 583, + 444, + 596 + ], + "score": 1.0, + "content": "(positional embedding removed for simplicity).", + "type": "text" + }, + { + "bbox": [ + 444, + 583, + 488, + 595 + ], + "score": 0.92, + "content": "L _ { 1 } , L _ { 2 } , L _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 593, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 141, + 593, + 505, + 608 + ], + "score": 1.0, + "content": "trainable linear layers. The outmost layers (commonly referred as unembed layers) have no", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 604, + 392, + 618 + ], + "spans": [ + { + "bbox": [ + 141, + 604, + 392, + 618 + ], + "score": 1.0, + "content": "biases and the other layers have biases included for generality.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 134, + 619, + 390, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 619, + 171, + 632 + ], + "score": 1.0, + "content": "• Model", + "type": "text" + }, + { + "bbox": [ + 171, + 622, + 178, + 630 + ], + "score": 0.67, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 619, + 284, + 632 + ], + "score": 1.0, + "content": ": calculate output logits as", + "type": "text" + }, + { + "bbox": [ + 284, + 619, + 386, + 632 + ], + "score": 0.88, + "content": "L _ { 2 } ( \\mathrm { R e L U } ( L _ { 1 } ( x _ { 1 } + x _ { 2 } ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 619, + 390, + 632 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 634, + 442, + 648 + ], + "spans": [ + { + "bbox": [ + 132, + 634, + 171, + 648 + ], + "score": 1.0, + "content": "• Model", + "type": "text" + }, + { + "bbox": [ + 171, + 635, + 178, + 647 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 634, + 284, + 648 + ], + "score": 1.0, + "content": ": calculate output logits as", + "type": "text" + }, + { + "bbox": [ + 285, + 635, + 437, + 648 + ], + "score": 0.88, + "content": "L _ { 3 } ( \\mathrm { R e L U } ( L _ { 2 } ( \\mathrm { R e L U } ( L _ { 1 } ( x _ { 1 } + x _ { 2 } ) ) ) ) ) ,", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 634, + 442, + 648 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 649, + 461, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 649, + 170, + 664 + ], + "score": 1.0, + "content": "• Model", + "type": "text" + }, + { + "bbox": [ + 171, + 652, + 178, + 662 + ], + "score": 0.78, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 649, + 284, + 664 + ], + "score": 1.0, + "content": ": calculate output logits as", + "type": "text" + }, + { + "bbox": [ + 284, + 650, + 457, + 663 + ], + "score": 0.88, + "content": "L _ { 3 } ( \\mathrm { R e L U } ( L _ { 2 } ( \\mathrm { R e L U } ( L _ { 1 } ( x _ { 1 } ) + L _ { 1 } ( x _ { 2 } ) ) ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 649, + 461, + 664 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 133, + 663, + 385, + 679 + ], + "spans": [ + { + "bbox": [ + 133, + 663, + 171, + 679 + ], + "score": 1.0, + "content": "• Model", + "type": "text" + }, + { + "bbox": [ + 171, + 666, + 177, + 675 + ], + "score": 0.58, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 663, + 283, + 679 + ], + "score": 1.0, + "content": ": calculate output logits as", + "type": "text" + }, + { + "bbox": [ + 283, + 665, + 382, + 677 + ], + "score": 0.88, + "content": "L _ { 2 } \\big ( \\mathrm { R e L U } ( L _ { 1 } ( [ x _ { 1 } ; x _ { 2 } ] ) ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 663, + 385, + 679 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 142, + 678, + 349, + 691 + ], + "spans": [ + { + "bbox": [ + 142, + 678, + 145, + 691 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 145, + 678, + 177, + 690 + ], + "score": 0.91, + "content": "[ x _ { 1 } ; x _ { 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 678, + 303, + 691 + ], + "score": 1.0, + "content": "stands for the concatenation of", + "type": "text" + }, + { + "bbox": [ + 303, + 680, + 314, + 689 + ], + "score": 0.85, + "content": "x _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 678, + 332, + 691 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 333, + 681, + 344, + 689 + ], + "score": 0.77, + "content": "x _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 678, + 349, + 691 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + } + ], + "index": 20, + "bbox_fs": [ + 132, + 571, + 506, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 504, + 723 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 351, + 712 + ], + "score": 1.0, + "content": "The results are shown in Figure 14. Rather surprisingly, Model", + "type": "text" + }, + { + "bbox": [ + 351, + 702, + 359, + 710 + ], + "score": 0.74, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 699, + 389, + 712 + ], + "score": 1.0, + "content": ", Model", + "type": "text" + }, + { + "bbox": [ + 390, + 700, + 397, + 711 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 699, + 442, + 712 + ], + "score": 1.0, + "content": "and Model", + "type": "text" + }, + { + "bbox": [ + 442, + 700, + 448, + 710 + ], + "score": 0.78, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 699, + 504, + 712 + ], + "score": 1.0, + "content": "gave radically", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 207, + 723 + ], + "score": 1.0, + "content": "different results. Model", + "type": "text" + }, + { + "bbox": [ + 207, + 711, + 215, + 722 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 710, + 263, + 723 + ], + "score": 1.0, + "content": "and Model", + "type": "text" + }, + { + "bbox": [ + 263, + 712, + 271, + 722 + ], + "score": 0.82, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "are very similar, and in general they are more pizza-like", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 154, + 398 + ], + "score": 1.0, + "content": "than Model", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 154, + 388, + 162, + 396 + ], + "score": 0.75, + "content": "\\alpha", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 162, + 385, + 505, + 398 + ], + "score": 1.0, + "content": ", with lower distance irrelevancy and higher circularity. This could be explained by the", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 397, + 236, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 236, + 409 + ], + "score": 1.0, + "content": "addition of an extra linear layer.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 699, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 73, + 501, + 333 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 73, + 501, + 333 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 73, + 501, + 333 + ], + "spans": [ + { + "bbox": [ + 108, + 73, + 501, + 333 + ], + "score": 0.976, + "type": "image", + "image_path": "e8fd35053c1f30d57d8e0fcbbd9dd4ce144784115afd24206efec7ddd16c126d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 73, + 501, + 159.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 159.66666666666669, + 501, + 246.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 246.33333333333337, + 501, + 333.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 341, + 504, + 364 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 340, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 356 + ], + "score": 1.0, + "content": "Figure 14: Training results from linear models. Each point in the first-row plots represents a training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 351, + 429, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 429, + 365 + ], + "score": 1.0, + "content": "run. The second row are histograms for distance irrelevancy of each model type.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 106, + 386, + 503, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 154, + 398 + ], + "score": 1.0, + "content": "than Model", + "type": "text" + }, + { + "bbox": [ + 154, + 388, + 162, + 396 + ], + "score": 0.75, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 385, + 505, + 398 + ], + "score": 1.0, + "content": ", with lower distance irrelevancy and higher circularity. This could be explained by the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 397, + 236, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 236, + 409 + ], + "score": 1.0, + "content": "addition of an extra linear layer.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 413, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 175, + 426 + ], + "score": 1.0, + "content": "However, Model", + "type": "text" + }, + { + "bbox": [ + 176, + 414, + 182, + 423 + ], + "score": 0.63, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 413, + 339, + 426 + ], + "score": 1.0, + "content": "gave very different results from Model", + "type": "text" + }, + { + "bbox": [ + 339, + 415, + 347, + 424 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "although they are both one-layer linear", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "score": 1.0, + "content": "models. It is more likely to be non-circular and have very high distance irrelevancy in general. In", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "other words, concatenating instead of adding embeddings yields radically different behaviors in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "one-layer linear model. This result, again, alarmed us the significance of induction biases in neural", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 457, + 147, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 147, + 468 + ], + "score": 1.0, + "content": "networks.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 503, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 418, + 486 + ], + "score": 1.0, + "content": "We also want to note that using different embeddings on two tokens of Model", + "type": "text" + }, + { + "bbox": [ + 419, + 475, + 426, + 483 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "doesn’t resolve the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 484, + 245, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 245, + 497 + ], + "score": 1.0, + "content": "discrepancy. The following model", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 134, + 505, + 505, + 532 + ], + "lines": [ + { + "bbox": [ + 129, + 498, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 129, + 498, + 171, + 524 + ], + "score": 1.0, + "content": "• Model", + "type": "text" + }, + { + "bbox": [ + 172, + 506, + 182, + 516 + ], + "score": 0.84, + "content": "\\alpha ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 498, + 292, + 524 + ], + "score": 1.0, + "content": ": calculate output logits as", + "type": "text" + }, + { + "bbox": [ + 293, + 506, + 395, + 518 + ], + "score": 0.89, + "content": "L _ { 2 } ( \\mathrm { R e L U } ( L _ { 1 } ( x _ { 1 } + x _ { 2 } ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 498, + 426, + 524 + ], + "score": 1.0, + "content": ") where", + "type": "text" + }, + { + "bbox": [ + 426, + 505, + 475, + 520 + ], + "score": 0.87, + "content": "x _ { 1 } = W _ { E , a } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 498, + 481, + 524 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 481, + 506, + 505, + 519 + ], + "score": 0.75, + "content": "x _ { 2 } =", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 514, + 415, + 537 + ], + "spans": [ + { + "bbox": [ + 142, + 518, + 165, + 533 + ], + "score": 0.92, + "content": "W _ { E , b } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 514, + 203, + 537 + ], + "score": 1.0, + "content": "on input", + "type": "text" + }, + { + "bbox": [ + 203, + 519, + 225, + 532 + ], + "score": 0.91, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 514, + 243, + 537 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 518, + 284, + 532 + ], + "score": 0.9, + "content": "W _ { E } ^ { A } , W _ { E } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 514, + 415, + 537 + ], + "score": 1.0, + "content": "are different embedding layers.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 410, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 411, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 275, + 557 + ], + "score": 1.0, + "content": "gives roughly the same result as of Model", + "type": "text" + }, + { + "bbox": [ + 275, + 544, + 284, + 553 + ], + "score": 0.75, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 541, + 411, + 557 + ], + "score": 1.0, + "content": "(Figure 14, lower right corner).", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 559, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "Figure 15 shows the correct logits after circle isolation (Section 3.3) of a circular model from Model", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 107, + 571, + 114, + 581 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 569, + 506, + 583 + ], + "score": 1.0, + "content": "implementing the pizza algorithm. Figure 16 shows the correct logits after circle isolation (Section", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 581, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 251, + 593 + ], + "score": 1.0, + "content": "3.3) of a circular model from Model", + "type": "text" + }, + { + "bbox": [ + 251, + 581, + 257, + 591 + ], + "score": 0.56, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 581, + 506, + 593 + ], + "score": 1.0, + "content": ". We can see the pattern is similar but different from the one of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 592, + 419, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 419, + 604 + ], + "score": 1.0, + "content": "clock algorithm (Figure 5). We leave the study of such models to future work.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 106, + 619, + 300, + 633 + ], + "lines": [ + { + "bbox": [ + 104, + 617, + 300, + 635 + ], + "spans": [ + { + "bbox": [ + 104, + 617, + 300, + 635 + ], + "score": 1.0, + "content": "F Architecture and Training Details", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 644, + 189, + 656 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 190, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 190, + 657 + ], + "score": 1.0, + "content": "F.1 Transformers", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 504, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "Here we describe our setup for the main experiments. See Appendix E and Appendix I for experiments", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 675, + 186, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 186, + 689 + ], + "score": 1.0, + "content": "on different setups.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "Architecture We train bidirectional transformers (attention unmasked) to perform modular addition", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 128, + 723 + ], + "score": 1.0, + "content": "mod", + "type": "text" + }, + { + "bbox": [ + 129, + 713, + 135, + 722 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 711, + 162, + 723 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 163, + 711, + 192, + 722 + ], + "score": 0.9, + "content": "p = 5 9", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 711, + 248, + 723 + ], + "score": 1.0, + "content": ". To calculate", + "type": "text" + }, + { + "bbox": [ + 248, + 711, + 278, + 723 + ], + "score": 0.91, + "content": "( a + b )", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 711, + 302, + 723 + ], + "score": 1.0, + "content": "mod", + "type": "text" + }, + { + "bbox": [ + 302, + 715, + 308, + 722 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 711, + 505, + 723 + ], + "score": 1.0, + "content": ", the input is provided to the model as a sequence", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + } + ], + "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": "image", + "bbox": [ + 108, + 73, + 501, + 333 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 73, + 501, + 333 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 73, + 501, + 333 + ], + "spans": [ + { + "bbox": [ + 108, + 73, + 501, + 333 + ], + "score": 0.976, + "type": "image", + "image_path": "e8fd35053c1f30d57d8e0fcbbd9dd4ce144784115afd24206efec7ddd16c126d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 73, + 501, + 159.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 159.66666666666669, + 501, + 246.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 246.33333333333337, + 501, + 333.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 341, + 504, + 364 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 340, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 356 + ], + "score": 1.0, + "content": "Figure 14: Training results from linear models. Each point in the first-row plots represents a training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 351, + 429, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 429, + 365 + ], + "score": 1.0, + "content": "run. The second row are histograms for distance irrelevancy of each model type.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 106, + 386, + 503, + 408 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 106, + 385, + 505, + 409 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 413, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 175, + 426 + ], + "score": 1.0, + "content": "However, Model", + "type": "text" + }, + { + "bbox": [ + 176, + 414, + 182, + 423 + ], + "score": 0.63, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 413, + 339, + 426 + ], + "score": 1.0, + "content": "gave very different results from Model", + "type": "text" + }, + { + "bbox": [ + 339, + 415, + 347, + 424 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "although they are both one-layer linear", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "score": 1.0, + "content": "models. It is more likely to be non-circular and have very high distance irrelevancy in general. In", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "other words, concatenating instead of adding embeddings yields radically different behaviors in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "one-layer linear model. This result, again, alarmed us the significance of induction biases in neural", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 457, + 147, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 147, + 468 + ], + "score": 1.0, + "content": "networks.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 413, + 506, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 503, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 418, + 486 + ], + "score": 1.0, + "content": "We also want to note that using different embeddings on two tokens of Model", + "type": "text" + }, + { + "bbox": [ + 419, + 475, + 426, + 483 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "doesn’t resolve the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 484, + 245, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 245, + 497 + ], + "score": 1.0, + "content": "discrepancy. The following model", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 473, + 505, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 134, + 505, + 505, + 532 + ], + "lines": [ + { + "bbox": [ + 129, + 498, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 129, + 498, + 171, + 524 + ], + "score": 1.0, + "content": "• Model", + "type": "text" + }, + { + "bbox": [ + 172, + 506, + 182, + 516 + ], + "score": 0.84, + "content": "\\alpha ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 498, + 292, + 524 + ], + "score": 1.0, + "content": ": calculate output logits as", + "type": "text" + }, + { + "bbox": [ + 293, + 506, + 395, + 518 + ], + "score": 0.89, + "content": "L _ { 2 } ( \\mathrm { R e L U } ( L _ { 1 } ( x _ { 1 } + x _ { 2 } ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 498, + 426, + 524 + ], + "score": 1.0, + "content": ") where", + "type": "text" + }, + { + "bbox": [ + 426, + 505, + 475, + 520 + ], + "score": 0.87, + "content": "x _ { 1 } = W _ { E , a } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 498, + 481, + 524 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 481, + 506, + 505, + 519 + ], + "score": 0.75, + "content": "x _ { 2 } =", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 514, + 415, + 537 + ], + "spans": [ + { + "bbox": [ + 142, + 518, + 165, + 533 + ], + "score": 0.92, + "content": "W _ { E , b } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 514, + 203, + 537 + ], + "score": 1.0, + "content": "on input", + "type": "text" + }, + { + "bbox": [ + 203, + 519, + 225, + 532 + ], + "score": 0.91, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 514, + 243, + 537 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 518, + 284, + 532 + ], + "score": 0.9, + "content": "W _ { E } ^ { A } , W _ { E } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 514, + 415, + 537 + ], + "score": 1.0, + "content": "are different embedding layers.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 129, + 498, + 505, + 537 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 410, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 411, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 275, + 557 + ], + "score": 1.0, + "content": "gives roughly the same result as of Model", + "type": "text" + }, + { + "bbox": [ + 275, + 544, + 284, + 553 + ], + "score": 0.75, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 541, + 411, + 557 + ], + "score": 1.0, + "content": "(Figure 14, lower right corner).", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 541, + 411, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 559, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "Figure 15 shows the correct logits after circle isolation (Section 3.3) of a circular model from Model", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 107, + 571, + 114, + 581 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 569, + 506, + 583 + ], + "score": 1.0, + "content": "implementing the pizza algorithm. Figure 16 shows the correct logits after circle isolation (Section", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 581, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 251, + 593 + ], + "score": 1.0, + "content": "3.3) of a circular model from Model", + "type": "text" + }, + { + "bbox": [ + 251, + 581, + 257, + 591 + ], + "score": 0.56, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 581, + 506, + 593 + ], + "score": 1.0, + "content": ". We can see the pattern is similar but different from the one of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 592, + 419, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 419, + 604 + ], + "score": 1.0, + "content": "clock algorithm (Figure 5). We leave the study of such models to future work.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 559, + 506, + 604 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 619, + 300, + 633 + ], + "lines": [ + { + "bbox": [ + 104, + 617, + 300, + 635 + ], + "spans": [ + { + "bbox": [ + 104, + 617, + 300, + 635 + ], + "score": 1.0, + "content": "F Architecture and Training Details", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 644, + 189, + 656 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 190, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 190, + 657 + ], + "score": 1.0, + "content": "F.1 Transformers", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 504, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "Here we describe our setup for the main experiments. See Appendix E and Appendix I for experiments", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 675, + 186, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 186, + 689 + ], + "score": 1.0, + "content": "on different setups.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 664, + 505, + 689 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "Architecture We train bidirectional transformers (attention unmasked) to perform modular addition", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 128, + 723 + ], + "score": 1.0, + "content": "mod", + "type": "text" + }, + { + "bbox": [ + 129, + 713, + 135, + 722 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 711, + 162, + 723 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 163, + 711, + 192, + 722 + ], + "score": 0.9, + "content": "p = 5 9", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 711, + 248, + 723 + ], + "score": 1.0, + "content": ". To calculate", + "type": "text" + }, + { + "bbox": [ + 248, + 711, + 278, + 723 + ], + "score": 0.91, + "content": "( a + b )", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 711, + 302, + 723 + ], + "score": 1.0, + "content": "mod", + "type": "text" + }, + { + "bbox": [ + 302, + 715, + 308, + 722 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 711, + 505, + 723 + ], + "score": 1.0, + "content": ", the input is provided to the model as a sequence", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 164, + 679 + ], + "score": 1.0, + "content": "of two tokens", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 164, + 667, + 184, + 679 + ], + "score": 0.91, + "content": "[ a , b ]", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 185, + 667, + 506, + 679 + ], + "score": 1.0, + "content": ". The output logit at the last token is considered as the output of the model. For", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 219, + 690 + ], + "score": 1.0, + "content": "a transformer with “width”", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 220, + 678, + 226, + 688 + ], + "score": 0.7, + "content": "d", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 227, + 677, + 443, + 690 + ], + "score": 1.0, + "content": ", the input embedding and the residue stream will be", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 443, + 678, + 450, + 688 + ], + "score": 0.81, + "content": "d", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 450, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "-dimensional,", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 190, + 701 + ], + "score": 1.0, + "content": "4 attention heads of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 190, + 689, + 214, + 701 + ], + "score": 0.91, + "content": "\\lfloor d / 4 \\rfloor", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 214, + 689, + 438, + 701 + ], + "score": 1.0, + "content": "dimensions will be employed, and the MLP will be of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 439, + 689, + 451, + 699 + ], + "score": 0.75, + "content": "4 d", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 451, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "hidden units.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 150, + 712 + ], + "score": 1.0, + "content": "By default", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 151, + 700, + 186, + 710 + ], + "score": 0.91, + "content": "d = 1 2 8", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 186, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "is chosen. ReLU is used as the activation function and layer normalization isn’t", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "applied. The post-softmax attention matrix is interpolated between an all-one matrix and original as", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "specified by the attention rate (Section 4.2). We want to point out that the setup of constant-attention", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 338, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 338, + 96 + ], + "score": 1.0, + "content": "transformers is also considered in the previous work [28].", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 699, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 89, + 500, + 339 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 182, + 70, + 426, + 82 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 182, + 69, + 428, + 83 + ], + "spans": [ + { + "bbox": [ + 182, + 69, + 335, + 83 + ], + "score": 1.0, + "content": "Figure 15: Correct logits from Model", + "type": "text" + }, + { + "bbox": [ + 335, + 71, + 343, + 82 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 69, + 428, + 83 + ], + "score": 1.0, + "content": "after circle isolation.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image_body", + "bbox": [ + 112, + 89, + 500, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 89, + 500, + 339 + ], + "spans": [ + { + "bbox": [ + 112, + 89, + 500, + 339 + ], + "score": 0.967, + "type": "image", + "image_path": "208a9e146b1796dbfcc6aed685b316be60b30f70153c68d9352806b6d6967502.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 112, + 89, + 500, + 172.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 172.33333333333331, + 500, + 255.66666666666663 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 112, + 255.66666666666663, + 500, + 338.99999999999994 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "image", + "bbox": [ + 110, + 380, + 502, + 631 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 182, + 362, + 426, + 374 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 183, + 361, + 428, + 375 + ], + "spans": [ + { + "bbox": [ + 183, + 361, + 335, + 375 + ], + "score": 1.0, + "content": "Figure 16: Correct logits from Model", + "type": "text" + }, + { + "bbox": [ + 336, + 362, + 342, + 372 + ], + "score": 0.62, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 361, + 428, + 375 + ], + "score": 1.0, + "content": "after circle isolation.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "image_body", + "bbox": [ + 110, + 380, + 502, + 631 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 380, + 502, + 631 + ], + "spans": [ + { + "bbox": [ + 110, + 380, + 502, + 631 + ], + "score": 0.97, + "type": "image", + "image_path": "58674c5336c13104d117005c72a37e55d1073de56a5820d831f5191397e18680.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 110, + 380, + 502, + 463.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 110, + 463.6666666666667, + 502, + 547.3333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 110, + 547.3333333333334, + 502, + 631.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 164, + 679 + ], + "score": 1.0, + "content": "of two tokens", + "type": "text" + }, + { + "bbox": [ + 164, + 667, + 184, + 679 + ], + "score": 0.91, + "content": "[ a , b ]", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 667, + 506, + 679 + ], + "score": 1.0, + "content": ". The output logit at the last token is considered as the output of the model. For", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 219, + 690 + ], + "score": 1.0, + "content": "a transformer with “width”", + "type": "text" + }, + { + "bbox": [ + 220, + 678, + 226, + 688 + ], + "score": 0.7, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 677, + 443, + 690 + ], + "score": 1.0, + "content": ", the input embedding and the residue stream will be", + "type": "text" + }, + { + "bbox": [ + 443, + 678, + 450, + 688 + ], + "score": 0.81, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "-dimensional,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 190, + 701 + ], + "score": 1.0, + "content": "4 attention heads of", + "type": "text" + }, + { + "bbox": [ + 190, + 689, + 214, + 701 + ], + "score": 0.91, + "content": "\\lfloor d / 4 \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 689, + 438, + 701 + ], + "score": 1.0, + "content": "dimensions will be employed, and the MLP will be of", + "type": "text" + }, + { + "bbox": [ + 439, + 689, + 451, + 699 + ], + "score": 0.75, + "content": "4 d", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "hidden units.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 150, + 712 + ], + "score": 1.0, + "content": "By default", + "type": "text" + }, + { + "bbox": [ + 151, + 700, + 186, + 710 + ], + "score": 0.91, + "content": "d = 1 2 8", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "is chosen. ReLU is used as the activation function and layer normalization isn’t", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "applied. The post-softmax attention matrix is interpolated between an all-one matrix and original as", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + } + ], + "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": "image", + "bbox": [ + 112, + 89, + 500, + 339 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 182, + 70, + 426, + 82 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 182, + 69, + 428, + 83 + ], + "spans": [ + { + "bbox": [ + 182, + 69, + 335, + 83 + ], + "score": 1.0, + "content": "Figure 15: Correct logits from Model", + "type": "text" + }, + { + "bbox": [ + 335, + 71, + 343, + 82 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 69, + 428, + 83 + ], + "score": 1.0, + "content": "after circle isolation.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image_body", + "bbox": [ + 112, + 89, + 500, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 89, + 500, + 339 + ], + "spans": [ + { + "bbox": [ + 112, + 89, + 500, + 339 + ], + "score": 0.967, + "type": "image", + "image_path": "208a9e146b1796dbfcc6aed685b316be60b30f70153c68d9352806b6d6967502.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 112, + 89, + 500, + 172.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 172.33333333333331, + 500, + 255.66666666666663 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 112, + 255.66666666666663, + 500, + 338.99999999999994 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "image", + "bbox": [ + 110, + 380, + 502, + 631 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 182, + 362, + 426, + 374 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 183, + 361, + 428, + 375 + ], + "spans": [ + { + "bbox": [ + 183, + 361, + 335, + 375 + ], + "score": 1.0, + "content": "Figure 16: Correct logits from Model", + "type": "text" + }, + { + "bbox": [ + 336, + 362, + 342, + 372 + ], + "score": 0.62, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 361, + 428, + 375 + ], + "score": 1.0, + "content": "after circle isolation.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "image_body", + "bbox": [ + 110, + 380, + 502, + 631 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 380, + 502, + 631 + ], + "spans": [ + { + "bbox": [ + 110, + 380, + 502, + 631 + ], + "score": 0.97, + "type": "image", + "image_path": "58674c5336c13104d117005c72a37e55d1073de56a5820d831f5191397e18680.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 110, + 380, + 502, + 463.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 110, + 463.6666666666667, + 502, + 547.3333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 110, + 547.3333333333334, + 502, + 631.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 506, + 722 + ], + "lines": [], + "index": 10, + "bbox_fs": [ + 104, + 667, + 506, + 723 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "specified by the attention rate (Section 4.2). We want to point out that the setup of constant-attention", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 338, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 338, + 96 + ], + "score": 1.0, + "content": "transformers is also considered in the previous work [28].", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 140 + ], + "lines": [ + { + "bbox": [ + 104, + 104, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 269, + 120 + ], + "score": 1.0, + "content": "Data Among all possible data points", + "type": "text" + }, + { + "bbox": [ + 269, + 106, + 315, + 118 + ], + "score": 0.9, + "content": "\\gamma ^ { 2 } = 3 4 8 1", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 104, + 438, + 120 + ], + "score": 1.0, + "content": "of them), we randomly select", + "type": "text" + }, + { + "bbox": [ + 438, + 106, + 459, + 117 + ], + "score": 0.88, + "content": "8 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 104, + 505, + 120 + ], + "score": 1.0, + "content": "as training", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 505, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 158, + 131 + ], + "score": 1.0, + "content": "samples and", + "type": "text" + }, + { + "bbox": [ + 158, + 118, + 178, + 128 + ], + "score": 0.88, + "content": "2 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 117, + 346, + 131 + ], + "score": 1.0, + "content": "as validation samples. This choice (small", + "type": "text" + }, + { + "bbox": [ + 347, + 120, + 353, + 129 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 117, + 505, + 131 + ], + "score": 1.0, + "content": "and high training data fraction) helps", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 208, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 208, + 142 + ], + "score": 1.0, + "content": "accelerating the training.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 151, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 361, + 164 + ], + "score": 1.0, + "content": "Training We used AdamW optimizer [29] with learning rate", + "type": "text" + }, + { + "bbox": [ + 362, + 153, + 405, + 164 + ], + "score": 0.91, + "content": "\\gamma = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 151, + 506, + 164 + ], + "score": 1.0, + "content": "and weight decay factor", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 107, + 163, + 132, + 174 + ], + "score": 0.89, + "content": "\\beta = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 162, + 505, + 176 + ], + "score": 1.0, + "content": ". We do not use minibatches and the shuffled training data is provided as a whole batch in every", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "epoch. For each run, we start the training from scratch and train for 20, 000 epoches. We removed", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 183, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 213, + 198 + ], + "score": 1.0, + "content": "the runs that did not reach", + "type": "text" + }, + { + "bbox": [ + 213, + 184, + 238, + 195 + ], + "score": 0.89, + "content": "1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 183, + 506, + 198 + ], + "score": 1.0, + "content": "validation accuracy at the end of the training (majority of the runs", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 195, + 171, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 140, + 207 + ], + "score": 1.0, + "content": "reached", + "type": "text" + }, + { + "bbox": [ + 140, + 195, + 165, + 207 + ], + "score": 0.87, + "content": "1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 195, + 171, + 207 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 107, + 219, + 193, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 218, + 194, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 194, + 233 + ], + "score": 1.0, + "content": "F.2 Linear Models", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 240, + 408, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 409, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 409, + 254 + ], + "score": 1.0, + "content": "Here we describe our setup for the linear model experiments (Appendix E).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 470, + 275 + ], + "score": 1.0, + "content": "Architecture We train several types of linear models to perform modular addition mod", + "type": "text" + }, + { + "bbox": [ + 470, + 266, + 477, + 275 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 136, + 285 + ], + "score": 0.9, + "content": "p = 5 9", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 274, + 388, + 286 + ], + "score": 1.0, + "content": ". The input embedding, residue stream and hidden layer are all", + "type": "text" + }, + { + "bbox": [ + 388, + 274, + 423, + 284 + ], + "score": 0.9, + "content": "d = 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "dimensional. ReLU", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "score": 1.0, + "content": "is used as the activation function. The actual structures of network types are specified in Appendix E.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 308, + 370, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 371, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 371, + 321 + ], + "score": 1.0, + "content": "Data & Training Same as in the previous section (Section F.1).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 107, + 332, + 224, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 226, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 226, + 347 + ], + "score": 1.0, + "content": "F.3 Computing Resources", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 504, + 375 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 504, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 504, + 366 + ], + "score": 1.0, + "content": "A total of 226 GPU days of NVidia V100 is spent on this project, although we expect a replication", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 364, + 272, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 272, + 376 + ], + "score": 1.0, + "content": "would take significantly fewer resources.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 107, + 390, + 442, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 444, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 444, + 405 + ], + "score": 1.0, + "content": "G Mathematical Description of Constant-Attention transformer", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 108, + 415, + 504, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "In this section, we examine the structure of constant-attention transformers loosely following the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 426, + 174, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 174, + 438 + ], + "score": 1.0, + "content": "notation of [10].", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 506, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 281, + 457 + ], + "score": 1.0, + "content": "Denote the weight of embedding layer as", + "type": "text" + }, + { + "bbox": [ + 281, + 443, + 298, + 454 + ], + "score": 0.89, + "content": "W _ { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 442, + 464, + 457 + ], + "score": 1.0, + "content": ", the weight of positional embedding as", + "type": "text" + }, + { + "bbox": [ + 465, + 443, + 485, + 455 + ], + "score": 0.9, + "content": "W _ { \\mathrm { p o s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 442, + 506, + 457 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 451, + 508, + 474 + ], + "spans": [ + { + "bbox": [ + 104, + 451, + 280, + 474 + ], + "score": 1.0, + "content": "weight of the value and output matrix of the", + "type": "text" + }, + { + "bbox": [ + 280, + 457, + 286, + 468 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 451, + 343, + 474 + ], + "score": 1.0, + "content": "-th head of the", + "type": "text" + }, + { + "bbox": [ + 344, + 457, + 349, + 466 + ], + "score": 0.76, + "content": "t { \\cdot }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 451, + 393, + 474 + ], + "score": 1.0, + "content": "-th layer as", + "type": "text" + }, + { + "bbox": [ + 393, + 454, + 415, + 468 + ], + "score": 0.91, + "content": "W _ { V } ^ { t , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 451, + 432, + 474 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 433, + 455, + 454, + 469 + ], + "score": 0.91, + "content": "W _ { O } ^ { t , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 451, + 508, + 474 + ], + "score": 1.0, + "content": ", the weights", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 465, + 507, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 299, + 483 + ], + "score": 1.0, + "content": "and biases of the input linear map of MLP in the", + "type": "text" + }, + { + "bbox": [ + 305, + 465, + 349, + 483 + ], + "score": 1.0, + "content": "-th layer as", + "type": "text" + }, + { + "bbox": [ + 349, + 467, + 366, + 479 + ], + "score": 0.91, + "content": "\\it { W _ { \\mathrm { i n } } ^ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 465, + 383, + 483 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 384, + 467, + 395, + 479 + ], + "score": 0.89, + "content": "b _ { \\mathrm { i n } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 465, + 507, + 483 + ], + "score": 1.0, + "content": ", the corresponding weights", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 103, + 474, + 507, + 495 + ], + "spans": [ + { + "bbox": [ + 103, + 474, + 256, + 495 + ], + "score": 1.0, + "content": "and biases of the output linear map as", + "type": "text" + }, + { + "bbox": [ + 256, + 478, + 276, + 490 + ], + "score": 0.92, + "content": "W _ { \\mathrm { o u t } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 474, + 294, + 495 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 294, + 478, + 309, + 490 + ], + "score": 0.91, + "content": "b _ { \\mathrm { o u t } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 474, + 486, + 495 + ], + "score": 1.0, + "content": ", and the weight of the unembedding layer as", + "type": "text" + }, + { + "bbox": [ + 486, + 478, + 503, + 489 + ], + "score": 0.89, + "content": "W _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 474, + 507, + 495 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "score": 1.0, + "content": "Notice that the query and the key matrices are irrelevant as the attention matrix is replaced with an", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 499, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 197, + 512 + ], + "score": 1.0, + "content": "all-one matrix. Denote", + "type": "text" + }, + { + "bbox": [ + 198, + 499, + 208, + 510 + ], + "score": 0.87, + "content": "x ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 499, + 406, + 512 + ], + "score": 1.0, + "content": "as the value of residue stream vector after the first", + "type": "text" + }, + { + "bbox": [ + 407, + 501, + 413, + 511 + ], + "score": 0.84, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 499, + 484, + 512 + ], + "score": 1.0, + "content": "layers and denote", + "type": "text" + }, + { + "bbox": [ + 484, + 502, + 493, + 511 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 499, + 506, + 512 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 511, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 187, + 523 + ], + "score": 1.0, + "content": "the character in the", + "type": "text" + }, + { + "bbox": [ + 187, + 511, + 192, + 520 + ], + "score": 0.77, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 511, + 337, + 523 + ], + "score": 1.0, + "content": "-th position. We use subscripts like", + "type": "text" + }, + { + "bbox": [ + 338, + 512, + 348, + 522 + ], + "score": 0.86, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 511, + 506, + 523 + ], + "score": 1.0, + "content": "to denote taking a specific element of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 523, + 136, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 136, + 534 + ], + "score": 1.0, + "content": "vector.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 538, + 327, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 329, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 329, + 551 + ], + "score": 1.0, + "content": "We can formalize the logit calculation as the following:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 133, + 558, + 402, + 633 + ], + "lines": [ + { + "bbox": [ + 133, + 556, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 133, + 556, + 194, + 574 + ], + "score": 1.0, + "content": "• Embedding:", + "type": "text" + }, + { + "bbox": [ + 194, + 559, + 281, + 572 + ], + "score": 0.91, + "content": "x _ { i } ^ { 0 } = W _ { E , c _ { i } } + W _ { \\mathrm { p o s } , i }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 556, + 286, + 574 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 133, + 573, + 272, + 587 + ], + "spans": [ + { + "bbox": [ + 133, + 573, + 201, + 587 + ], + "score": 1.0, + "content": "• For each layer", + "type": "text" + }, + { + "bbox": [ + 201, + 576, + 207, + 584 + ], + "score": 0.51, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 573, + 247, + 587 + ], + "score": 1.0, + "content": "from 1 to", + "type": "text" + }, + { + "bbox": [ + 248, + 576, + 269, + 587 + ], + "score": 0.83, + "content": "n _ { \\mathrm { l a y e r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 573, + 272, + 587 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 149, + 588, + 401, + 608 + ], + "spans": [ + { + "bbox": [ + 149, + 588, + 247, + 608 + ], + "score": 1.0, + "content": "– Constant Attention:", + "type": "text" + }, + { + "bbox": [ + 247, + 590, + 401, + 606 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\boldsymbol { w } _ { i } ^ { t } = \\boldsymbol { x } _ { i } ^ { t - 1 } + \\sum _ { j } \\boldsymbol { W } _ { O } ^ { t , j } \\boldsymbol { W } _ { V } ^ { t , j } \\sum _ { k } \\boldsymbol { x } _ { k } ^ { t - 1 } . } \\end{array}", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 151, + 605, + 364, + 621 + ], + "spans": [ + { + "bbox": [ + 151, + 605, + 187, + 621 + ], + "score": 1.0, + "content": "– MLP:", + "type": "text" + }, + { + "bbox": [ + 187, + 606, + 360, + 619 + ], + "score": 0.82, + "content": "x ^ { t } = w ^ { t } + b _ { \\mathrm { o u t } } ^ { t } + W _ { \\mathrm { o u t } } ^ { t } \\mathrm { R e L U } ( b _ { \\mathrm { i n } } ^ { t } + W _ { \\mathrm { i n } } ^ { t } w ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 605, + 364, + 621 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 134, + 621, + 240, + 635 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 176, + 635 + ], + "score": 1.0, + "content": "• Output:", + "type": "text" + }, + { + "bbox": [ + 176, + 622, + 236, + 633 + ], + "score": 0.89, + "content": "O = W _ { U } x ^ { n _ { \\mathrm { l a y e r } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 621, + 240, + 635 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 504, + 666 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 392, + 656 + ], + "score": 1.0, + "content": "In the particular case where the input length is 2, the number of layer is", + "type": "text" + }, + { + "bbox": [ + 393, + 644, + 398, + 653 + ], + "score": 0.28, + "content": "1", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 643, + 505, + 656 + ], + "score": 1.0, + "content": ", and we focus on the logit", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 653, + 433, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 346, + 667 + ], + "score": 1.0, + "content": "of the last position, we may restate as the following (denote", + "type": "text" + }, + { + "bbox": [ + 347, + 657, + 353, + 664 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 653, + 365, + 667 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 365, + 654, + 376, + 664 + ], + "score": 0.87, + "content": "\\dot { x } ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 653, + 394, + 667 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 394, + 656, + 401, + 666 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 653, + 413, + 667 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 413, + 654, + 426, + 664 + ], + "score": 0.85, + "content": "w ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 653, + 433, + 667 + ], + "score": 1.0, + "content": "):", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 133, + 675, + 380, + 724 + ], + "lines": [ + { + "bbox": [ + 132, + 672, + 381, + 691 + ], + "spans": [ + { + "bbox": [ + 132, + 672, + 193, + 691 + ], + "score": 1.0, + "content": "• Embedding:", + "type": "text" + }, + { + "bbox": [ + 193, + 676, + 381, + 688 + ], + "score": 0.7, + "content": "x _ { 1 } = W _ { E , c _ { 1 } } + W _ { \\mathrm { p o s } , 1 } , x _ { 2 } = W _ { E , c _ { 2 } } + W _ { \\mathrm { p o s } , 2 } .", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 132, + 690, + 360, + 708 + ], + "spans": [ + { + "bbox": [ + 132, + 690, + 223, + 708 + ], + "score": 1.0, + "content": "• Constant Attention:", + "type": "text" + }, + { + "bbox": [ + 224, + 691, + 355, + 707 + ], + "score": 0.89, + "content": "\\begin{array} { r } { y = x _ { 2 } + \\sum _ { j } W _ { O } ^ { j } W _ { V } ^ { j } ( x _ { 1 } + x _ { 2 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 690, + 360, + 708 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 708, + 329, + 725 + ], + "spans": [ + { + "bbox": [ + 132, + 708, + 168, + 725 + ], + "score": 1.0, + "content": "• MLP:", + "type": "text" + }, + { + "bbox": [ + 168, + 710, + 325, + 723 + ], + "score": 0.88, + "content": "z = y + b _ { \\mathrm { o u t } } ^ { t } + W _ { \\mathrm { o u t } } ^ { t } \\mathrm { R e L U } ( b _ { \\mathrm { i n } } ^ { t } + W _ { \\mathrm { i n } } ^ { t } y )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 708, + 329, + 725 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + } + ], + "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": [ + 106, + 72, + 504, + 95 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 72, + 505, + 96 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 140 + ], + "lines": [ + { + "bbox": [ + 104, + 104, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 269, + 120 + ], + "score": 1.0, + "content": "Data Among all possible data points", + "type": "text" + }, + { + "bbox": [ + 269, + 106, + 315, + 118 + ], + "score": 0.9, + "content": "\\gamma ^ { 2 } = 3 4 8 1", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 104, + 438, + 120 + ], + "score": 1.0, + "content": "of them), we randomly select", + "type": "text" + }, + { + "bbox": [ + 438, + 106, + 459, + 117 + ], + "score": 0.88, + "content": "8 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 104, + 505, + 120 + ], + "score": 1.0, + "content": "as training", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 505, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 158, + 131 + ], + "score": 1.0, + "content": "samples and", + "type": "text" + }, + { + "bbox": [ + 158, + 118, + 178, + 128 + ], + "score": 0.88, + "content": "2 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 117, + 346, + 131 + ], + "score": 1.0, + "content": "as validation samples. This choice (small", + "type": "text" + }, + { + "bbox": [ + 347, + 120, + 353, + 129 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 117, + 505, + 131 + ], + "score": 1.0, + "content": "and high training data fraction) helps", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 208, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 208, + 142 + ], + "score": 1.0, + "content": "accelerating the training.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 104, + 505, + 142 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 151, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 361, + 164 + ], + "score": 1.0, + "content": "Training We used AdamW optimizer [29] with learning rate", + "type": "text" + }, + { + "bbox": [ + 362, + 153, + 405, + 164 + ], + "score": 0.91, + "content": "\\gamma = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 151, + 506, + 164 + ], + "score": 1.0, + "content": "and weight decay factor", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 107, + 163, + 132, + 174 + ], + "score": 0.89, + "content": "\\beta = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 162, + 505, + 176 + ], + "score": 1.0, + "content": ". We do not use minibatches and the shuffled training data is provided as a whole batch in every", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "epoch. For each run, we start the training from scratch and train for 20, 000 epoches. We removed", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 183, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 213, + 198 + ], + "score": 1.0, + "content": "the runs that did not reach", + "type": "text" + }, + { + "bbox": [ + 213, + 184, + 238, + 195 + ], + "score": 0.89, + "content": "1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 183, + 506, + 198 + ], + "score": 1.0, + "content": "validation accuracy at the end of the training (majority of the runs", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 195, + 171, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 140, + 207 + ], + "score": 1.0, + "content": "reached", + "type": "text" + }, + { + "bbox": [ + 140, + 195, + 165, + 207 + ], + "score": 0.87, + "content": "1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 195, + 171, + 207 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 151, + 506, + 207 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 219, + 193, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 218, + 194, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 194, + 233 + ], + "score": 1.0, + "content": "F.2 Linear Models", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 240, + 408, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 409, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 409, + 254 + ], + "score": 1.0, + "content": "Here we describe our setup for the linear model experiments (Appendix E).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 238, + 409, + 254 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 470, + 275 + ], + "score": 1.0, + "content": "Architecture We train several types of linear models to perform modular addition mod", + "type": "text" + }, + { + "bbox": [ + 470, + 266, + 477, + 275 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 136, + 285 + ], + "score": 0.9, + "content": "p = 5 9", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 274, + 388, + 286 + ], + "score": 1.0, + "content": ". The input embedding, residue stream and hidden layer are all", + "type": "text" + }, + { + "bbox": [ + 388, + 274, + 423, + 284 + ], + "score": 0.9, + "content": "d = 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "dimensional. ReLU", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "score": 1.0, + "content": "is used as the activation function. The actual structures of network types are specified in Appendix E.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 263, + 506, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 308, + 370, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 371, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 371, + 321 + ], + "score": 1.0, + "content": "Data & Training Same as in the previous section (Section F.1).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 307, + 371, + 321 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 332, + 224, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 226, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 226, + 347 + ], + "score": 1.0, + "content": "F.3 Computing Resources", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 504, + 375 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 504, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 504, + 366 + ], + "score": 1.0, + "content": "A total of 226 GPU days of NVidia V100 is spent on this project, although we expect a replication", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 364, + 272, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 272, + 376 + ], + "score": 1.0, + "content": "would take significantly fewer resources.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 352, + 504, + 376 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 390, + 442, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 444, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 444, + 405 + ], + "score": 1.0, + "content": "G Mathematical Description of Constant-Attention transformer", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 108, + 415, + 504, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "In this section, we examine the structure of constant-attention transformers loosely following the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 426, + 174, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 174, + 438 + ], + "score": 1.0, + "content": "notation of [10].", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 415, + 505, + 438 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 506, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 281, + 457 + ], + "score": 1.0, + "content": "Denote the weight of embedding layer as", + "type": "text" + }, + { + "bbox": [ + 281, + 443, + 298, + 454 + ], + "score": 0.89, + "content": "W _ { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 442, + 464, + 457 + ], + "score": 1.0, + "content": ", the weight of positional embedding as", + "type": "text" + }, + { + "bbox": [ + 465, + 443, + 485, + 455 + ], + "score": 0.9, + "content": "W _ { \\mathrm { p o s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 442, + 506, + 457 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 451, + 508, + 474 + ], + "spans": [ + { + "bbox": [ + 104, + 451, + 280, + 474 + ], + "score": 1.0, + "content": "weight of the value and output matrix of the", + "type": "text" + }, + { + "bbox": [ + 280, + 457, + 286, + 468 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 451, + 343, + 474 + ], + "score": 1.0, + "content": "-th head of the", + "type": "text" + }, + { + "bbox": [ + 344, + 457, + 349, + 466 + ], + "score": 0.76, + "content": "t { \\cdot }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 451, + 393, + 474 + ], + "score": 1.0, + "content": "-th layer as", + "type": "text" + }, + { + "bbox": [ + 393, + 454, + 415, + 468 + ], + "score": 0.91, + "content": "W _ { V } ^ { t , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 451, + 432, + 474 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 433, + 455, + 454, + 469 + ], + "score": 0.91, + "content": "W _ { O } ^ { t , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 451, + 508, + 474 + ], + "score": 1.0, + "content": ", the weights", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 465, + 507, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 299, + 483 + ], + "score": 1.0, + "content": "and biases of the input linear map of MLP in the", + "type": "text" + }, + { + "bbox": [ + 305, + 465, + 349, + 483 + ], + "score": 1.0, + "content": "-th layer as", + "type": "text" + }, + { + "bbox": [ + 349, + 467, + 366, + 479 + ], + "score": 0.91, + "content": "\\it { W _ { \\mathrm { i n } } ^ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 465, + 383, + 483 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 384, + 467, + 395, + 479 + ], + "score": 0.89, + "content": "b _ { \\mathrm { i n } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 465, + 507, + 483 + ], + "score": 1.0, + "content": ", the corresponding weights", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 103, + 474, + 507, + 495 + ], + "spans": [ + { + "bbox": [ + 103, + 474, + 256, + 495 + ], + "score": 1.0, + "content": "and biases of the output linear map as", + "type": "text" + }, + { + "bbox": [ + 256, + 478, + 276, + 490 + ], + "score": 0.92, + "content": "W _ { \\mathrm { o u t } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 474, + 294, + 495 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 294, + 478, + 309, + 490 + ], + "score": 0.91, + "content": "b _ { \\mathrm { o u t } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 474, + 486, + 495 + ], + "score": 1.0, + "content": ", and the weight of the unembedding layer as", + "type": "text" + }, + { + "bbox": [ + 486, + 478, + 503, + 489 + ], + "score": 0.89, + "content": "W _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 474, + 507, + 495 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "score": 1.0, + "content": "Notice that the query and the key matrices are irrelevant as the attention matrix is replaced with an", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 499, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 197, + 512 + ], + "score": 1.0, + "content": "all-one matrix. Denote", + "type": "text" + }, + { + "bbox": [ + 198, + 499, + 208, + 510 + ], + "score": 0.87, + "content": "x ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 499, + 406, + 512 + ], + "score": 1.0, + "content": "as the value of residue stream vector after the first", + "type": "text" + }, + { + "bbox": [ + 407, + 501, + 413, + 511 + ], + "score": 0.84, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 499, + 484, + 512 + ], + "score": 1.0, + "content": "layers and denote", + "type": "text" + }, + { + "bbox": [ + 484, + 502, + 493, + 511 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 499, + 506, + 512 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 511, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 187, + 523 + ], + "score": 1.0, + "content": "the character in the", + "type": "text" + }, + { + "bbox": [ + 187, + 511, + 192, + 520 + ], + "score": 0.77, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 511, + 337, + 523 + ], + "score": 1.0, + "content": "-th position. We use subscripts like", + "type": "text" + }, + { + "bbox": [ + 338, + 512, + 348, + 522 + ], + "score": 0.86, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 511, + 506, + 523 + ], + "score": 1.0, + "content": "to denote taking a specific element of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 523, + 136, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 136, + 534 + ], + "score": 1.0, + "content": "vector.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5, + "bbox_fs": [ + 103, + 442, + 508, + 534 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 538, + 327, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 329, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 329, + 551 + ], + "score": 1.0, + "content": "We can formalize the logit calculation as the following:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 536, + 329, + 551 + ] + }, + { + "type": "list", + "bbox": [ + 133, + 558, + 402, + 633 + ], + "lines": [ + { + "bbox": [ + 133, + 556, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 133, + 556, + 194, + 574 + ], + "score": 1.0, + "content": "• Embedding:", + "type": "text" + }, + { + "bbox": [ + 194, + 559, + 281, + 572 + ], + "score": 0.91, + "content": "x _ { i } ^ { 0 } = W _ { E , c _ { i } } + W _ { \\mathrm { p o s } , i }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 556, + 286, + 574 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 133, + 573, + 272, + 587 + ], + "spans": [ + { + "bbox": [ + 133, + 573, + 201, + 587 + ], + "score": 1.0, + "content": "• For each layer", + "type": "text" + }, + { + "bbox": [ + 201, + 576, + 207, + 584 + ], + "score": 0.51, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 573, + 247, + 587 + ], + "score": 1.0, + "content": "from 1 to", + "type": "text" + }, + { + "bbox": [ + 248, + 576, + 269, + 587 + ], + "score": 0.83, + "content": "n _ { \\mathrm { l a y e r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 573, + 272, + 587 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 149, + 588, + 401, + 608 + ], + "spans": [ + { + "bbox": [ + 149, + 588, + 247, + 608 + ], + "score": 1.0, + "content": "– Constant Attention:", + "type": "text" + }, + { + "bbox": [ + 247, + 590, + 401, + 606 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\boldsymbol { w } _ { i } ^ { t } = \\boldsymbol { x } _ { i } ^ { t - 1 } + \\sum _ { j } \\boldsymbol { W } _ { O } ^ { t , j } \\boldsymbol { W } _ { V } ^ { t , j } \\sum _ { k } \\boldsymbol { x } _ { k } ^ { t - 1 } . } \\end{array}", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 151, + 605, + 364, + 621 + ], + "spans": [ + { + "bbox": [ + 151, + 605, + 187, + 621 + ], + "score": 1.0, + "content": "– MLP:", + "type": "text" + }, + { + "bbox": [ + 187, + 606, + 360, + 619 + ], + "score": 0.82, + "content": "x ^ { t } = w ^ { t } + b _ { \\mathrm { o u t } } ^ { t } + W _ { \\mathrm { o u t } } ^ { t } \\mathrm { R e L U } ( b _ { \\mathrm { i n } } ^ { t } + W _ { \\mathrm { i n } } ^ { t } w ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 605, + 364, + 621 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 134, + 621, + 240, + 635 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 176, + 635 + ], + "score": 1.0, + "content": "• Output:", + "type": "text" + }, + { + "bbox": [ + 176, + 622, + 236, + 633 + ], + "score": 0.89, + "content": "O = W _ { U } x ^ { n _ { \\mathrm { l a y e r } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 621, + 240, + 635 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 33, + "bbox_fs": [ + 133, + 556, + 401, + 635 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 504, + 666 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 392, + 656 + ], + "score": 1.0, + "content": "In the particular case where the input length is 2, the number of layer is", + "type": "text" + }, + { + "bbox": [ + 393, + 644, + 398, + 653 + ], + "score": 0.28, + "content": "1", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 643, + 505, + 656 + ], + "score": 1.0, + "content": ", and we focus on the logit", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 653, + 433, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 346, + 667 + ], + "score": 1.0, + "content": "of the last position, we may restate as the following (denote", + "type": "text" + }, + { + "bbox": [ + 347, + 657, + 353, + 664 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 653, + 365, + 667 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 365, + 654, + 376, + 664 + ], + "score": 0.87, + "content": "\\dot { x } ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 653, + 394, + 667 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 394, + 656, + 401, + 666 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 653, + 413, + 667 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 413, + 654, + 426, + 664 + ], + "score": 0.85, + "content": "w ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 653, + 433, + 667 + ], + "score": 1.0, + "content": "):", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 643, + 505, + 667 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 675, + 380, + 724 + ], + "lines": [ + { + "bbox": [ + 132, + 672, + 381, + 691 + ], + "spans": [ + { + "bbox": [ + 132, + 672, + 193, + 691 + ], + "score": 1.0, + "content": "• Embedding:", + "type": "text" + }, + { + "bbox": [ + 193, + 676, + 381, + 688 + ], + "score": 0.7, + "content": "x _ { 1 } = W _ { E , c _ { 1 } } + W _ { \\mathrm { p o s } , 1 } , x _ { 2 } = W _ { E , c _ { 2 } } + W _ { \\mathrm { p o s } , 2 } .", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 132, + 690, + 360, + 708 + ], + "spans": [ + { + "bbox": [ + 132, + 690, + 223, + 708 + ], + "score": 1.0, + "content": "• Constant Attention:", + "type": "text" + }, + { + "bbox": [ + 224, + 691, + 355, + 707 + ], + "score": 0.89, + "content": "\\begin{array} { r } { y = x _ { 2 } + \\sum _ { j } W _ { O } ^ { j } W _ { V } ^ { j } ( x _ { 1 } + x _ { 2 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 690, + 360, + 708 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 708, + 329, + 725 + ], + "spans": [ + { + "bbox": [ + 132, + 708, + 168, + 725 + ], + "score": 1.0, + "content": "• MLP:", + "type": "text" + }, + { + "bbox": [ + 168, + 710, + 325, + 723 + ], + "score": 0.88, + "content": "z = y + b _ { \\mathrm { o u t } } ^ { t } + W _ { \\mathrm { o u t } } ^ { t } \\mathrm { R e L U } ( b _ { \\mathrm { i n } } ^ { t } + W _ { \\mathrm { i n } } ^ { t } y )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 708, + 329, + 725 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 132, + 672, + 381, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 133, + 72, + 219, + 84 + ], + "lines": [ + { + "bbox": [ + 133, + 70, + 216, + 87 + ], + "spans": [ + { + "bbox": [ + 133, + 70, + 176, + 87 + ], + "score": 1.0, + "content": "• Output:", + "type": "text" + }, + { + "bbox": [ + 176, + 73, + 216, + 84 + ], + "score": 0.87, + "content": "o = W _ { U } z", + "type": "inline_equation" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 93, + 433, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 92, + 433, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 433, + 107 + ], + "score": 1.0, + "content": "If we remove the skip connections, the network after embedding could be seen as", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 173, + 108, + 437, + 149 + ], + "lines": [ + { + "bbox": [ + 173, + 108, + 437, + 149 + ], + "spans": [ + { + "bbox": [ + 173, + 108, + 437, + 149 + ], + "score": 0.94, + "content": "o = L _ { U } \\left( L _ { \\mathrm { o u t } } \\left( \\operatorname { R e L U } \\left( L _ { \\mathrm { i n } } \\left( \\sum _ { j } L _ { O } ^ { j } \\left( L _ { V } ^ { j } \\left( x _ { 1 } + x _ { 2 } \\right) \\right) \\right) \\right) \\right) \\right)", + "type": "interline_equation", + "image_path": "6ac774d02f9716d0114dd5cafb5d09c68f61d47dc99829d7d5cf0d7f2d73b526.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 173, + 108, + 437, + 121.66666666666667 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 173, + 121.66666666666667, + 437, + 135.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 173, + 135.33333333333334, + 437, + 149.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 154, + 453, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 456, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 133, + 168 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 152, + 221, + 167 + ], + "score": 0.93, + "content": "L _ { V } ^ { j } , L _ { O } ^ { j } , L _ { \\mathrm { i n } } , L _ { \\mathrm { o u t } } , L _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 152, + 456, + 168 + ], + "score": 1.0, + "content": "are a series of linear layers corresponding to the matrices.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 107, + 180, + 234, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 235, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 235, + 196 + ], + "score": 1.0, + "content": "H Pizza with Attention", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 205, + 505, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "Extrapolating from Figure 7, we trained transformers with width 1024 and attention rate 1 (normal", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "attention). After several tries, we are able to observe a trained circular model with distance irrelevance", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 227, + 449, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 449, + 240 + ], + "score": 1.0, + "content": "0.156 and gradient symmetricity 0.995, which fits our definition of Pizza (Figure 17).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 110, + 248, + 502, + 439 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 248, + 502, + 439 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 248, + 502, + 439 + ], + "spans": [ + { + "bbox": [ + 110, + 248, + 502, + 439 + ], + "score": 0.973, + "type": "image", + "image_path": "6a19a13ddcdd757cf6d8f5ef7ee8b2fb27edf3f827fe9c82e37625373f8993c0.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 110, + 248, + 502, + 311.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 110, + 311.6666666666667, + 502, + 375.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 110, + 375.33333333333337, + 502, + 439.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 116, + 445, + 493, + 457 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 444, + 493, + 459 + ], + "spans": [ + { + "bbox": [ + 117, + 444, + 493, + 459 + ], + "score": 1.0, + "content": "Figure 17: Correct logits of the trained model in Section H after circle isolation (Section 3.3).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + } + ], + "index": 12.0 + }, + { + "type": "title", + "bbox": [ + 106, + 478, + 306, + 493 + ], + "lines": [ + { + "bbox": [ + 104, + 477, + 307, + 496 + ], + "spans": [ + { + "bbox": [ + 104, + 477, + 307, + 496 + ], + "score": 1.0, + "content": "I Results on Slightly Different Setups", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 504, + 526 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 515 + ], + "score": 1.0, + "content": "We considered the following variations of our setups (Appendix F.1, Section 4), for which the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 514, + 417, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 417, + 527 + ], + "score": 1.0, + "content": "existence of pizzas and clocks as well as the phase changes are still observed.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 537, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "GeLU instead of ReLU We conducted the same 1-layer transformer experiment with activation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 547, + 423, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 423, + 561 + ], + "score": 1.0, + "content": "function GeLU instead of ReLU. Very similar results are observed (Figure 18).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 571, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 585 + ], + "score": 1.0, + "content": "Encode Two Tokens Differently We conducted the 1-layer transformer experiments but with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "different embedding for the two tokens. Again very similar results are observed (Figure 19). We", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "score": 1.0, + "content": "also discovered that the two tokens’ embeddings are often aligned to implement the Pizza and Clock", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 604, + 198, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 198, + 617 + ], + "score": 1.0, + "content": "algorithm (Figure 20).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 626, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "Adding Equal Sign We conducted the 1-layer transformer experiment with an equal sign added.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 637, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 289, + 650 + ], + "score": 1.0, + "content": "Very similar results are observed (Figure 21).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 107, + 663, + 336, + 678 + ], + "lines": [ + { + "bbox": [ + 104, + 662, + 338, + 681 + ], + "spans": [ + { + "bbox": [ + 104, + 662, + 338, + 681 + ], + "score": 1.0, + "content": "J Pizza Occurs Early in the Clock Training", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "We plotted intermediate states during the training of a model with attention (attention rate 1). Pizza-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "like pattern was observed early in the training, but the pattern gradually disappeared during the run", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 710, + 157, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 157, + 724 + ], + "score": 1.0, + "content": "(Figure 22).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 310, + 750 + ], + "lines": [ + { + "bbox": [ + 298, + 740, + 312, + 755 + ], + "spans": [ + { + "bbox": [ + 298, + 740, + 312, + 755 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 133, + 72, + 219, + 84 + ], + "lines": [ + { + "bbox": [ + 133, + 70, + 216, + 87 + ], + "spans": [ + { + "bbox": [ + 133, + 70, + 176, + 87 + ], + "score": 1.0, + "content": "• Output:", + "type": "text" + }, + { + "bbox": [ + 176, + 73, + 216, + 84 + ], + "score": 0.87, + "content": "o = W _ { U } z", + "type": "inline_equation" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 133, + 70, + 216, + 87 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 93, + 433, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 92, + 433, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 433, + 107 + ], + "score": 1.0, + "content": "If we remove the skip connections, the network after embedding could be seen as", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 92, + 433, + 107 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 173, + 108, + 437, + 149 + ], + "lines": [ + { + "bbox": [ + 173, + 108, + 437, + 149 + ], + "spans": [ + { + "bbox": [ + 173, + 108, + 437, + 149 + ], + "score": 0.94, + "content": "o = L _ { U } \\left( L _ { \\mathrm { o u t } } \\left( \\operatorname { R e L U } \\left( L _ { \\mathrm { i n } } \\left( \\sum _ { j } L _ { O } ^ { j } \\left( L _ { V } ^ { j } \\left( x _ { 1 } + x _ { 2 } \\right) \\right) \\right) \\right) \\right) \\right)", + "type": "interline_equation", + "image_path": "6ac774d02f9716d0114dd5cafb5d09c68f61d47dc99829d7d5cf0d7f2d73b526.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 173, + 108, + 437, + 121.66666666666667 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 173, + 121.66666666666667, + 437, + 135.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 173, + 135.33333333333334, + 437, + 149.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 154, + 453, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 456, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 133, + 168 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 152, + 221, + 167 + ], + "score": 0.93, + "content": "L _ { V } ^ { j } , L _ { O } ^ { j } , L _ { \\mathrm { i n } } , L _ { \\mathrm { o u t } } , L _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 152, + 456, + 168 + ], + "score": 1.0, + "content": "are a series of linear layers corresponding to the matrices.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 152, + 456, + 168 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 180, + 234, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 235, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 235, + 196 + ], + "score": 1.0, + "content": "H Pizza with Attention", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 205, + 505, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "Extrapolating from Figure 7, we trained transformers with width 1024 and attention rate 1 (normal", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "attention). After several tries, we are able to observe a trained circular model with distance irrelevance", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 227, + 449, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 449, + 240 + ], + "score": 1.0, + "content": "0.156 and gradient symmetricity 0.995, which fits our definition of Pizza (Figure 17).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 205, + 505, + 240 + ] + }, + { + "type": "image", + "bbox": [ + 110, + 248, + 502, + 439 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 248, + 502, + 439 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 248, + 502, + 439 + ], + "spans": [ + { + "bbox": [ + 110, + 248, + 502, + 439 + ], + "score": 0.973, + "type": "image", + "image_path": "6a19a13ddcdd757cf6d8f5ef7ee8b2fb27edf3f827fe9c82e37625373f8993c0.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 110, + 248, + 502, + 311.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 110, + 311.6666666666667, + 502, + 375.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 110, + 375.33333333333337, + 502, + 439.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 116, + 445, + 493, + 457 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 444, + 493, + 459 + ], + "spans": [ + { + "bbox": [ + 117, + 444, + 493, + 459 + ], + "score": 1.0, + "content": "Figure 17: Correct logits of the trained model in Section H after circle isolation (Section 3.3).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + } + ], + "index": 12.0 + }, + { + "type": "title", + "bbox": [ + 106, + 478, + 306, + 493 + ], + "lines": [ + { + "bbox": [ + 104, + 477, + 307, + 496 + ], + "spans": [ + { + "bbox": [ + 104, + 477, + 307, + 496 + ], + "score": 1.0, + "content": "I Results on Slightly Different Setups", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 504, + 526 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 515 + ], + "score": 1.0, + "content": "We considered the following variations of our setups (Appendix F.1, Section 4), for which the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 514, + 417, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 417, + 527 + ], + "score": 1.0, + "content": "existence of pizzas and clocks as well as the phase changes are still observed.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 504, + 505, + 527 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 537, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "GeLU instead of ReLU We conducted the same 1-layer transformer experiment with activation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 547, + 423, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 423, + 561 + ], + "score": 1.0, + "content": "function GeLU instead of ReLU. Very similar results are observed (Figure 18).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 536, + 505, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 571, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 585 + ], + "score": 1.0, + "content": "Encode Two Tokens Differently We conducted the 1-layer transformer experiments but with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "different embedding for the two tokens. Again very similar results are observed (Figure 19). We", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "score": 1.0, + "content": "also discovered that the two tokens’ embeddings are often aligned to implement the Pizza and Clock", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 604, + 198, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 198, + 617 + ], + "score": 1.0, + "content": "algorithm (Figure 20).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 570, + 506, + 617 + ] + }, + { + "type": "list", + "bbox": [ + 108, + 626, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "Adding Equal Sign We conducted the 1-layer transformer experiment with an equal sign added.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 637, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 289, + 650 + ], + "score": 1.0, + "content": "Very similar results are observed (Figure 21).", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 626, + 505, + 650 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 663, + 336, + 678 + ], + "lines": [ + { + "bbox": [ + 104, + 662, + 338, + 681 + ], + "spans": [ + { + "bbox": [ + 104, + 662, + 338, + 681 + ], + "score": 1.0, + "content": "J Pizza Occurs Early in the Clock Training", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "We plotted intermediate states during the training of a model with attention (attention rate 1). Pizza-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "like pattern was observed early in the training, but the pattern gradually disappeared during the run", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 710, + 157, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 157, + 724 + ], + "score": 1.0, + "content": "(Figure 22).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 687, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 104, + 500, + 423 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 70, + 507, + 93 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 505, + 82 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 505, + 82 + ], + "score": 1.0, + "content": "Figure 18: Training results from 1-layer transformers with GeLU instead of ReLU as the activation", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 79, + 495, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 388, + 95 + ], + "score": 1.0, + "content": "function. Each point in the plots represents a training run that reached", + "type": "text" + }, + { + "bbox": [ + 388, + 82, + 412, + 92 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 79, + 495, + 95 + ], + "score": 1.0, + "content": "validation accuracy.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "image_body", + "bbox": [ + 109, + 104, + 500, + 423 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 104, + 500, + 423 + ], + "spans": [ + { + "bbox": [ + 109, + 104, + 500, + 423 + ], + "score": 0.976, + "type": "image", + "image_path": "8e3aa8c8d51780d597724dbf1b2353c2c20309c869362e114321b4956eb28135.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 109, + 104, + 500, + 210.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 210.33333333333331, + 500, + 316.66666666666663 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 316.66666666666663, + 500, + 422.99999999999994 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "image", + "bbox": [ + 110, + 514, + 498, + 667 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 459, + 505, + 504 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "Figure 19: Training results from 1-layer transformers where the two tokens use different embeddings", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 129, + 483 + ], + "score": 1.0, + "content": "(feed", + "type": "text" + }, + { + "bbox": [ + 130, + 470, + 168, + 483 + ], + "score": 0.93, + "content": "[ a , b + p ]", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 470, + 260, + 483 + ], + "score": 1.0, + "content": "to the model on input", + "type": "text" + }, + { + "bbox": [ + 260, + 470, + 282, + 483 + ], + "score": 0.89, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 470, + 286, + 483 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 287, + 471, + 299, + 482 + ], + "score": 0.8, + "content": "2 p", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "tokens are handled in the embedding layer). Each", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 331, + 495 + ], + "score": 1.0, + "content": "point in the plots represents a training run that reached", + "type": "text" + }, + { + "bbox": [ + 332, + 482, + 356, + 492 + ], + "score": 0.87, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "validation accuracy. We did not use", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 493, + 366, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 366, + 504 + ], + "score": 1.0, + "content": "circularity to filter the result because it is no longer well-defined.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "image_body", + "bbox": [ + 110, + 514, + 498, + 667 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 514, + 498, + 667 + ], + "spans": [ + { + "bbox": [ + 110, + 514, + 498, + 667 + ], + "score": 0.954, + "type": "image", + "image_path": "ea4290352b0b65ad356165dd30ffc420464d1a19c9fb175bb23275c71e5bbc11.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 110, + 514, + 498, + 565.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 565.0, + 498, + 616.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 110, + 616.0, + 498, + 667.0 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 8.25 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 298, + 740, + 313, + 755 + ], + "spans": [ + { + "bbox": [ + 298, + 740, + 313, + 755 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 104, + 500, + 423 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 70, + 507, + 93 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 505, + 82 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 505, + 82 + ], + "score": 1.0, + "content": "Figure 18: Training results from 1-layer transformers with GeLU instead of ReLU as the activation", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 79, + 495, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 388, + 95 + ], + "score": 1.0, + "content": "function. Each point in the plots represents a training run that reached", + "type": "text" + }, + { + "bbox": [ + 388, + 82, + 412, + 92 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 79, + 495, + 95 + ], + "score": 1.0, + "content": "validation accuracy.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "image_body", + "bbox": [ + 109, + 104, + 500, + 423 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 104, + 500, + 423 + ], + "spans": [ + { + "bbox": [ + 109, + 104, + 500, + 423 + ], + "score": 0.976, + "type": "image", + "image_path": "8e3aa8c8d51780d597724dbf1b2353c2c20309c869362e114321b4956eb28135.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 109, + 104, + 500, + 210.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 210.33333333333331, + 500, + 316.66666666666663 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 316.66666666666663, + 500, + 422.99999999999994 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "image", + "bbox": [ + 110, + 514, + 498, + 667 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 459, + 505, + 504 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "Figure 19: Training results from 1-layer transformers where the two tokens use different embeddings", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 129, + 483 + ], + "score": 1.0, + "content": "(feed", + "type": "text" + }, + { + "bbox": [ + 130, + 470, + 168, + 483 + ], + "score": 0.93, + "content": "[ a , b + p ]", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 470, + 260, + 483 + ], + "score": 1.0, + "content": "to the model on input", + "type": "text" + }, + { + "bbox": [ + 260, + 470, + 282, + 483 + ], + "score": 0.89, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 470, + 286, + 483 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 287, + 471, + 299, + 482 + ], + "score": 0.8, + "content": "2 p", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "tokens are handled in the embedding layer). Each", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 331, + 495 + ], + "score": 1.0, + "content": "point in the plots represents a training run that reached", + "type": "text" + }, + { + "bbox": [ + 332, + 482, + 356, + 492 + ], + "score": 0.87, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "validation accuracy. We did not use", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 493, + 366, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 366, + 504 + ], + "score": 1.0, + "content": "circularity to filter the result because it is no longer well-defined.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "image_body", + "bbox": [ + 110, + 514, + 498, + 667 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 514, + 498, + 667 + ], + "spans": [ + { + "bbox": [ + 110, + 514, + 498, + 667 + ], + "score": 0.954, + "type": "image", + "image_path": "ea4290352b0b65ad356165dd30ffc420464d1a19c9fb175bb23275c71e5bbc11.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 110, + 514, + 498, + 565.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 565.0, + 498, + 616.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 110, + 616.0, + 498, + 667.0 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 8.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 154, + 502, + 403 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 70, + 506, + 147 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 506, + 83 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 506, + 83 + ], + "score": 1.0, + "content": "Figure 20: Correct logits after circle isolation from a trained model where two tokens use different", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 81, + 506, + 94 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 94 + ], + "score": 1.0, + "content": "embeddings. The blue points represent the embeddings for the first token and the green points", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "represent the embeddings for the second token. The model is implementing the Pizza algorithm.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "score": 1.0, + "content": "The correct logit pattern is shifted comparing to the previous patterns because the embeddings of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 115, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 126 + ], + "score": 1.0, + "content": "two tokens do not line up exactly. For example, the third circle has near-maximum correct logit for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 107, + 125, + 159, + 136 + ], + "score": 0.92, + "content": "a = 6 , b = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 125, + 318, + 138 + ], + "score": 1.0, + "content": "(the two points lining up on the top) and", + "type": "text" + }, + { + "bbox": [ + 318, + 125, + 386, + 137 + ], + "score": 0.91, + "content": "( a - b ) / 1 8 \\equiv 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 125, + 506, + 138 + ], + "score": 1.0, + "content": "(mod 59). This is the reason", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 136, + 366, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 366, + 148 + ], + "score": 1.0, + "content": "that the correct logit pattern appears to be shifted 10 units down.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "image_body", + "bbox": [ + 110, + 154, + 502, + 403 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 154, + 502, + 403 + ], + "spans": [ + { + "bbox": [ + 110, + 154, + 502, + 403 + ], + "score": 0.975, + "type": "image", + "image_path": "5ac4b2302c6e22ff00f14df38874e91cbd43407065c9487f049eda28ffe47326.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 110, + 154, + 502, + 237.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 110, + 237.0, + 502, + 320.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 320.0, + 502, + 403.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 5.5 + }, + { + "type": "image", + "bbox": [ + 110, + 487, + 498, + 641 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 422, + 505, + 477 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 421, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 473, + 435 + ], + "score": 1.0, + "content": "Figure 21: Training results from 1-layer transformers where an equal sign is added (feed", + "type": "text" + }, + { + "bbox": [ + 473, + 422, + 504, + 434 + ], + "score": 0.91, + "content": "[ a , b , = ]", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 196, + 446 + ], + "score": 1.0, + "content": "to the model on input", + "type": "text" + }, + { + "bbox": [ + 196, + 433, + 218, + 445 + ], + "score": 0.92, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 432, + 246, + 446 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 246, + 435, + 255, + 443 + ], + "score": 0.79, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 432, + 329, + 446 + ], + "score": 1.0, + "content": "is a special token;", + "type": "text" + }, + { + "bbox": [ + 329, + 434, + 353, + 444 + ], + "score": 0.9, + "content": "p + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "tokens are handled in the embedding", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "layer; context length of the model becomes 3). Each point in the plots represents a training run that", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 453, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 139, + 469 + ], + "score": 1.0, + "content": "reached", + "type": "text" + }, + { + "bbox": [ + 140, + 455, + 163, + 466 + ], + "score": 0.87, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 453, + 506, + 469 + ], + "score": 1.0, + "content": "validation accuracy. We did not use circularity to filter the result because it is no longer", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 466, + 161, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 161, + 477 + ], + "score": 1.0, + "content": "well-defined.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "image_body", + "bbox": [ + 110, + 487, + 498, + 641 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 487, + 498, + 641 + ], + "spans": [ + { + "bbox": [ + 110, + 487, + 498, + 641 + ], + "score": 0.965, + "type": "image", + "image_path": "db427ec9128d569b9efb9358cb5819df97fffd4f84bd382d2986125be94cf974.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 110, + 487, + 498, + 538.3333333333334 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 110, + 538.3333333333334, + 498, + 589.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 110, + 589.6666666666667, + 498, + 641.0000000000001 + ], + "spans": [], + "index": 17 + } + ] + } + ], + "index": 14.0 + }, + { + "type": "title", + "bbox": [ + 106, + 673, + 417, + 688 + ], + "lines": [ + { + "bbox": [ + 104, + 671, + 419, + 691 + ], + "spans": [ + { + "bbox": [ + 104, + 671, + 419, + 691 + ], + "score": 1.0, + "content": "K Accompanying Pizza Occurs Early in the Pizza Training", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "We plotted intermediate states during the training of a model without attention (attention rate 0). We", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "score": 1.0, + "content": "observed the early emergence of a pattern similar to accompanying pizza in training runs (Figure", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + } + ], + "page_idx": 22, + "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": "image", + "bbox": [ + 110, + 154, + 502, + 403 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 70, + 506, + 147 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 506, + 83 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 506, + 83 + ], + "score": 1.0, + "content": "Figure 20: Correct logits after circle isolation from a trained model where two tokens use different", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 81, + 506, + 94 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 94 + ], + "score": 1.0, + "content": "embeddings. The blue points represent the embeddings for the first token and the green points", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "represent the embeddings for the second token. The model is implementing the Pizza algorithm.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "score": 1.0, + "content": "The correct logit pattern is shifted comparing to the previous patterns because the embeddings of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 115, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 126 + ], + "score": 1.0, + "content": "two tokens do not line up exactly. For example, the third circle has near-maximum correct logit for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 107, + 125, + 159, + 136 + ], + "score": 0.92, + "content": "a = 6 , b = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 125, + 318, + 138 + ], + "score": 1.0, + "content": "(the two points lining up on the top) and", + "type": "text" + }, + { + "bbox": [ + 318, + 125, + 386, + 137 + ], + "score": 0.91, + "content": "( a - b ) / 1 8 \\equiv 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 125, + 506, + 138 + ], + "score": 1.0, + "content": "(mod 59). This is the reason", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 136, + 366, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 366, + 148 + ], + "score": 1.0, + "content": "that the correct logit pattern appears to be shifted 10 units down.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "image_body", + "bbox": [ + 110, + 154, + 502, + 403 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 154, + 502, + 403 + ], + "spans": [ + { + "bbox": [ + 110, + 154, + 502, + 403 + ], + "score": 0.975, + "type": "image", + "image_path": "5ac4b2302c6e22ff00f14df38874e91cbd43407065c9487f049eda28ffe47326.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 110, + 154, + 502, + 237.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 110, + 237.0, + 502, + 320.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 320.0, + 502, + 403.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 5.5 + }, + { + "type": "image", + "bbox": [ + 110, + 487, + 498, + 641 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 422, + 505, + 477 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 421, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 473, + 435 + ], + "score": 1.0, + "content": "Figure 21: Training results from 1-layer transformers where an equal sign is added (feed", + "type": "text" + }, + { + "bbox": [ + 473, + 422, + 504, + 434 + ], + "score": 0.91, + "content": "[ a , b , = ]", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 196, + 446 + ], + "score": 1.0, + "content": "to the model on input", + "type": "text" + }, + { + "bbox": [ + 196, + 433, + 218, + 445 + ], + "score": 0.92, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 432, + 246, + 446 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 246, + 435, + 255, + 443 + ], + "score": 0.79, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 432, + 329, + 446 + ], + "score": 1.0, + "content": "is a special token;", + "type": "text" + }, + { + "bbox": [ + 329, + 434, + 353, + 444 + ], + "score": 0.9, + "content": "p + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "tokens are handled in the embedding", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "layer; context length of the model becomes 3). Each point in the plots represents a training run that", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 453, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 139, + 469 + ], + "score": 1.0, + "content": "reached", + "type": "text" + }, + { + "bbox": [ + 140, + 455, + 163, + 466 + ], + "score": 0.87, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 453, + 506, + 469 + ], + "score": 1.0, + "content": "validation accuracy. We did not use circularity to filter the result because it is no longer", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 466, + 161, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 161, + 477 + ], + "score": 1.0, + "content": "well-defined.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "image_body", + "bbox": [ + 110, + 487, + 498, + 641 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 487, + 498, + 641 + ], + "spans": [ + { + "bbox": [ + 110, + 487, + 498, + 641 + ], + "score": 0.965, + "type": "image", + "image_path": "db427ec9128d569b9efb9358cb5819df97fffd4f84bd382d2986125be94cf974.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 110, + 487, + 498, + 538.3333333333334 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 110, + 538.3333333333334, + 498, + 589.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 110, + 589.6666666666667, + 498, + 641.0000000000001 + ], + "spans": [], + "index": 17 + } + ] + } + ], + "index": 14.0 + }, + { + "type": "title", + "bbox": [ + 106, + 673, + 417, + 688 + ], + "lines": [ + { + "bbox": [ + 104, + 671, + 419, + 691 + ], + "spans": [ + { + "bbox": [ + 104, + 671, + 419, + 691 + ], + "score": 1.0, + "content": "K Accompanying Pizza Occurs Early in the Pizza Training", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "We plotted intermediate states during the training of a model without attention (attention rate 0). We", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "score": 1.0, + "content": "observed the early emergence of a pattern similar to accompanying pizza in training runs (Figure", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 700, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 72, + 502, + 625 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 72, + 502, + 625 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 72, + 502, + 625 + ], + "spans": [ + { + "bbox": [ + 112, + 72, + 502, + 625 + ], + "score": 0.968, + "type": "image", + "image_path": "4518f1c3f7b0e97afc8bc880c00613ffc7fb0a21dbf71b429831d34673932c81.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 72, + 502, + 256.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 256.33333333333337, + 502, + 440.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 440.66666666666674, + 502, + 625.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 626, + 504, + 649 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 625, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 505, + 639 + ], + "score": 1.0, + "content": "Figure 22: For a 1-layer transformer with attention, correct logits after principal component (possibly", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 636, + 504, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 504, + 649 + ], + "score": 1.0, + "content": "non-circle) isolations at various states during the training. The pizza-like pattern gradually desolved.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 331, + 702 + ], + "score": 1.0, + "content": "23) and removing that circle brings accuracy down from", + "type": "text" + }, + { + "bbox": [ + 331, + 689, + 358, + 699 + ], + "score": 0.86, + "content": "9 9 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 687, + 369, + 702 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 369, + 689, + 396, + 699 + ], + "score": 0.87, + "content": "9 7 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 687, + 506, + 702 + ], + "score": 1.0, + "content": ". They are less helpful later", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "in the network (removing accompanying pizzas in trained Model A only brings accuracy down to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 709, + 142, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 135, + 722 + ], + "score": 0.82, + "content": "9 9 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 709, + 142, + 724 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + } + ], + "page_idx": 23, + "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": "image", + "bbox": [ + 112, + 72, + 502, + 625 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 72, + 502, + 625 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 72, + 502, + 625 + ], + "spans": [ + { + "bbox": [ + 112, + 72, + 502, + 625 + ], + "score": 0.968, + "type": "image", + "image_path": "4518f1c3f7b0e97afc8bc880c00613ffc7fb0a21dbf71b429831d34673932c81.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 72, + 502, + 256.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 256.33333333333337, + 502, + 440.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 440.66666666666674, + 502, + 625.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 626, + 504, + 649 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 625, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 505, + 639 + ], + "score": 1.0, + "content": "Figure 22: For a 1-layer transformer with attention, correct logits after principal component (possibly", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 636, + 504, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 504, + 649 + ], + "score": 1.0, + "content": "non-circle) isolations at various states during the training. The pizza-like pattern gradually desolved.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 331, + 702 + ], + "score": 1.0, + "content": "23) and removing that circle brings accuracy down from", + "type": "text" + }, + { + "bbox": [ + 331, + 689, + 358, + 699 + ], + "score": 0.86, + "content": "9 9 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 687, + 369, + 702 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 369, + 689, + 396, + 699 + ], + "score": 0.87, + "content": "9 7 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 687, + 506, + 702 + ], + "score": 1.0, + "content": ". They are less helpful later", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "in the network (removing accompanying pizzas in trained Model A only brings accuracy down to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 709, + 142, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 135, + 722 + ], + "score": 0.82, + "content": "9 9 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 709, + 142, + 724 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 687, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 68, + 461, + 362 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 151, + 68, + 461, + 362 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 151, + 68, + 461, + 362 + ], + "spans": [ + { + "bbox": [ + 151, + 68, + 461, + 362 + ], + "score": 0.962, + "type": "image", + "image_path": "abdbd9fca314dc3e8e3bcc0f5d0e0c26da0990664e1e5c45f6144e8ec6ec83bd.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 151, + 68, + 461, + 166.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 151, + 166.0, + 461, + 264.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 151, + 264.0, + 461, + 362.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 370, + 505, + 393 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "score": 1.0, + "content": "Figure 23: Immediate state after 600 epochs of training for a 1-layer transformer with constant", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 379, + 148, + 394 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 148, + 394 + ], + "score": 1.0, + "content": "attention.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 106, + 410, + 328, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 410, + 328, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 328, + 425 + ], + "score": 1.0, + "content": "L A Closer Look at a Linear Pizza Model", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "score": 1.0, + "content": "In this section, we provide a full picture of the linear model shown in Figure 15 by investigating the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 446, + 221, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 221, + 457 + ], + "score": 1.0, + "content": "actual weights in the model.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 107, + 470, + 203, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 204, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 204, + 483 + ], + "score": 1.0, + "content": "L.1 Model Structure", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 111, + 489, + 466, + 502 + ], + "lines": [ + { + "bbox": [ + 109, + 489, + 468, + 504 + ], + "spans": [ + { + "bbox": [ + 109, + 489, + 259, + 504 + ], + "score": 1.0, + "content": "As described in Appendix E, on input", + "type": "text" + }, + { + "bbox": [ + 259, + 490, + 281, + 502 + ], + "score": 0.93, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 489, + 468, + 504 + ], + "score": 1.0, + "content": ", the output logits of the model is computed as", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 505, + 412, + 519 + ], + "lines": [ + { + "bbox": [ + 198, + 505, + 412, + 519 + ], + "spans": [ + { + "bbox": [ + 198, + 505, + 412, + 519 + ], + "score": 0.87, + "content": "L _ { 3 } ( \\mathrm { R e L U } ( L _ { 2 } ( \\mathrm { R e L U } ( L _ { 1 } ( \\mathrm { E m b e d } [ a ] + \\mathrm { E m b e d } [ b ] ) ) ) ) ) .", + "type": "interline_equation", + "image_path": "dd864c968950275c4c817384eb7a4b3cd7f618e282f2c0dec6ec7cc0c8330af7.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 198, + 505, + 412, + 519 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 109, + 527, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 107, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 107, + 527, + 269, + 539 + ], + "score": 1.0, + "content": "Denote the weight of embedding layer as", + "type": "text" + }, + { + "bbox": [ + 270, + 527, + 287, + 538 + ], + "score": 0.9, + "content": "W _ { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 527, + 437, + 539 + ], + "score": 1.0, + "content": ", the weight of the unembedding layer", + "type": "text" + }, + { + "bbox": [ + 437, + 528, + 456, + 538 + ], + "score": 0.82, + "content": "( L _ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 527, + 467, + 539 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 468, + 527, + 484, + 538 + ], + "score": 0.89, + "content": "W _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 527, + 505, + 539 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 107, + 537, + 212, + 550 + ], + "score": 1.0, + "content": "the weights and biases of", + "type": "text" + }, + { + "bbox": [ + 213, + 538, + 225, + 549 + ], + "score": 0.9, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 537, + 244, + 550 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 538, + 256, + 549 + ], + "score": 0.89, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 537, + 269, + 550 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 270, + 538, + 298, + 549 + ], + "score": 0.91, + "content": "W _ { 1 } , b _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 537, + 316, + 550 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 317, + 538, + 345, + 549 + ], + "score": 0.92, + "content": "W _ { 2 } , b _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 537, + 505, + 550 + ], + "score": 1.0, + "content": ", respectively, then the output logits on", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 108, + 549, + 224, + 561 + ], + "spans": [ + { + "bbox": [ + 108, + 549, + 130, + 561 + ], + "score": 1.0, + "content": "input", + "type": "text" + }, + { + "bbox": [ + 130, + 549, + 152, + 561 + ], + "score": 0.92, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 549, + 224, + 561 + ], + "score": 1.0, + "content": "can be written as", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 563, + 420, + 577 + ], + "lines": [ + { + "bbox": [ + 191, + 563, + 420, + 577 + ], + "spans": [ + { + "bbox": [ + 191, + 563, + 420, + 577 + ], + "score": 0.85, + "content": "W _ { U } \\mathrm { R e L U } ( b _ { 2 } + W _ { 2 } \\mathrm { R e L U } ( b _ { 1 } + W _ { 1 } ( W _ { E } [ a ] + W _ { E } [ b ] ) ) ) .", + "type": "interline_equation", + "image_path": "6fc3a1254cdc9bffbae427dacb3b40c15f5ab806783ec6a7a10163f3d911a8d6.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 191, + 563, + 420, + 577 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "title", + "bbox": [ + 106, + 588, + 200, + 600 + ], + "lines": [ + { + "bbox": [ + 104, + 586, + 201, + 602 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 201, + 602 + ], + "score": 1.0, + "content": "L.2 General Picture", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 608, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "We first perform principal component visualizations on the embedding and unembedding matrices.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "From Figure 24, we can see that the embedding and unembedding matrices formed matching circles", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 631, + 326, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 212, + 643 + ], + "score": 1.0, + "content": "(circles with the same gap", + "type": "text" + }, + { + "bbox": [ + 213, + 631, + 219, + 640 + ], + "score": 0.74, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 631, + 326, + 643 + ], + "score": 1.0, + "content": "between adjacent entries).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 646, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "We now give the general overview of the circuit. Each pair of matching circles forms an instance of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 657, + 498, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 498, + 671 + ], + "score": 1.0, + "content": "Pizza and they operate independently (with rather limited interference). Specifically for each pair,", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 133, + 677, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 132, + 675, + 504, + 692 + ], + "spans": [ + { + "bbox": [ + 132, + 675, + 341, + 692 + ], + "score": 1.0, + "content": "• The embedding matrix first places the inputs", + "type": "text" + }, + { + "bbox": [ + 342, + 678, + 357, + 689 + ], + "score": 0.88, + "content": "a , b", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 675, + 461, + 692 + ], + "score": 1.0, + "content": "on the circumference:", + "type": "text" + }, + { + "bbox": [ + 461, + 677, + 504, + 690 + ], + "score": 0.9, + "content": "W _ { E } ^ { \\prime } [ a ] \\ \\approx", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 688, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 142, + 689, + 229, + 700 + ], + "score": 0.88, + "content": "( \\cos ( w _ { k } a ) , \\sin ( w _ { k } a ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 688, + 250, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 250, + 689, + 444, + 701 + ], + "score": 0.77, + "content": "\\bar { W _ { E } ^ { \\prime } [ b ] } \\approx ( \\cos ( \\bar { w _ { k } } b ) , \\sin ( w _ { k } b ) ) ( w _ { k } = 2 \\pi k / p", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 688, + 507, + 702 + ], + "score": 1.0, + "content": "for some inte-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 141, + 700, + 158, + 713 + ], + "score": 1.0, + "content": "ger", + "type": "text" + }, + { + "bbox": [ + 158, + 700, + 214, + 712 + ], + "score": 0.91, + "content": "k \\in [ 1 , p - 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 700, + 287, + 713 + ], + "score": 1.0, + "content": "as in Section 2.1;", + "type": "text" + }, + { + "bbox": [ + 288, + 700, + 305, + 712 + ], + "score": 0.86, + "content": "W _ { E } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "stands for the two currently considered principal", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 711, + 389, + 723 + ], + "spans": [ + { + "bbox": [ + 141, + 711, + 204, + 723 + ], + "score": 1.0, + "content": "components of", + "type": "text" + }, + { + "bbox": [ + 204, + 711, + 221, + 722 + ], + "score": 0.89, + "content": "W _ { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 711, + 389, + 723 + ], + "score": 1.0, + "content": "; rotation and scaling omitted for brevity).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + } + ], + "page_idx": 24, + "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": "image", + "bbox": [ + 151, + 68, + 461, + 362 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 151, + 68, + 461, + 362 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 151, + 68, + 461, + 362 + ], + "spans": [ + { + "bbox": [ + 151, + 68, + 461, + 362 + ], + "score": 0.962, + "type": "image", + "image_path": "abdbd9fca314dc3e8e3bcc0f5d0e0c26da0990664e1e5c45f6144e8ec6ec83bd.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 151, + 68, + 461, + 166.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 151, + 166.0, + 461, + 264.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 151, + 264.0, + 461, + 362.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 370, + 505, + 393 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "score": 1.0, + "content": "Figure 23: Immediate state after 600 epochs of training for a 1-layer transformer with constant", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 379, + 148, + 394 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 148, + 394 + ], + "score": 1.0, + "content": "attention.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 106, + 410, + 328, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 410, + 328, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 328, + 425 + ], + "score": 1.0, + "content": "L A Closer Look at a Linear Pizza Model", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "score": 1.0, + "content": "In this section, we provide a full picture of the linear model shown in Figure 15 by investigating the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 446, + 221, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 221, + 457 + ], + "score": 1.0, + "content": "actual weights in the model.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 434, + 505, + 457 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 470, + 203, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 204, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 204, + 483 + ], + "score": 1.0, + "content": "L.1 Model Structure", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 111, + 489, + 466, + 502 + ], + "lines": [ + { + "bbox": [ + 109, + 489, + 468, + 504 + ], + "spans": [ + { + "bbox": [ + 109, + 489, + 259, + 504 + ], + "score": 1.0, + "content": "As described in Appendix E, on input", + "type": "text" + }, + { + "bbox": [ + 259, + 490, + 281, + 502 + ], + "score": 0.93, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 489, + 468, + 504 + ], + "score": 1.0, + "content": ", the output logits of the model is computed as", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 109, + 489, + 468, + 504 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 505, + 412, + 519 + ], + "lines": [ + { + "bbox": [ + 198, + 505, + 412, + 519 + ], + "spans": [ + { + "bbox": [ + 198, + 505, + 412, + 519 + ], + "score": 0.87, + "content": "L _ { 3 } ( \\mathrm { R e L U } ( L _ { 2 } ( \\mathrm { R e L U } ( L _ { 1 } ( \\mathrm { E m b e d } [ a ] + \\mathrm { E m b e d } [ b ] ) ) ) ) ) .", + "type": "interline_equation", + "image_path": "dd864c968950275c4c817384eb7a4b3cd7f618e282f2c0dec6ec7cc0c8330af7.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 198, + 505, + 412, + 519 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 109, + 527, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 107, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 107, + 527, + 269, + 539 + ], + "score": 1.0, + "content": "Denote the weight of embedding layer as", + "type": "text" + }, + { + "bbox": [ + 270, + 527, + 287, + 538 + ], + "score": 0.9, + "content": "W _ { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 527, + 437, + 539 + ], + "score": 1.0, + "content": ", the weight of the unembedding layer", + "type": "text" + }, + { + "bbox": [ + 437, + 528, + 456, + 538 + ], + "score": 0.82, + "content": "( L _ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 527, + 467, + 539 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 468, + 527, + 484, + 538 + ], + "score": 0.89, + "content": "W _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 527, + 505, + 539 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 107, + 537, + 212, + 550 + ], + "score": 1.0, + "content": "the weights and biases of", + "type": "text" + }, + { + "bbox": [ + 213, + 538, + 225, + 549 + ], + "score": 0.9, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 537, + 244, + 550 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 538, + 256, + 549 + ], + "score": 0.89, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 537, + 269, + 550 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 270, + 538, + 298, + 549 + ], + "score": 0.91, + "content": "W _ { 1 } , b _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 537, + 316, + 550 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 317, + 538, + 345, + 549 + ], + "score": 0.92, + "content": "W _ { 2 } , b _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 537, + 505, + 550 + ], + "score": 1.0, + "content": ", respectively, then the output logits on", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 108, + 549, + 224, + 561 + ], + "spans": [ + { + "bbox": [ + 108, + 549, + 130, + 561 + ], + "score": 1.0, + "content": "input", + "type": "text" + }, + { + "bbox": [ + 130, + 549, + 152, + 561 + ], + "score": 0.92, + "content": "( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 549, + 224, + 561 + ], + "score": 1.0, + "content": "can be written as", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 107, + 527, + 505, + 561 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 563, + 420, + 577 + ], + "lines": [ + { + "bbox": [ + 191, + 563, + 420, + 577 + ], + "spans": [ + { + "bbox": [ + 191, + 563, + 420, + 577 + ], + "score": 0.85, + "content": "W _ { U } \\mathrm { R e L U } ( b _ { 2 } + W _ { 2 } \\mathrm { R e L U } ( b _ { 1 } + W _ { 1 } ( W _ { E } [ a ] + W _ { E } [ b ] ) ) ) .", + "type": "interline_equation", + "image_path": "6fc3a1254cdc9bffbae427dacb3b40c15f5ab806783ec6a7a10163f3d911a8d6.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 191, + 563, + 420, + 577 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "title", + "bbox": [ + 106, + 588, + 200, + 600 + ], + "lines": [ + { + "bbox": [ + 104, + 586, + 201, + 602 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 201, + 602 + ], + "score": 1.0, + "content": "L.2 General Picture", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 608, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "We first perform principal component visualizations on the embedding and unembedding matrices.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "From Figure 24, we can see that the embedding and unembedding matrices formed matching circles", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 631, + 326, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 212, + 643 + ], + "score": 1.0, + "content": "(circles with the same gap", + "type": "text" + }, + { + "bbox": [ + 213, + 631, + 219, + 640 + ], + "score": 0.74, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 631, + 326, + 643 + ], + "score": 1.0, + "content": "between adjacent entries).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 607, + 506, + 643 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 646, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "We now give the general overview of the circuit. Each pair of matching circles forms an instance of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 657, + 498, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 498, + 671 + ], + "score": 1.0, + "content": "Pizza and they operate independently (with rather limited interference). Specifically for each pair,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 675, + 504, + 692 + ], + "spans": [ + { + "bbox": [ + 132, + 675, + 341, + 692 + ], + "score": 1.0, + "content": "• The embedding matrix first places the inputs", + "type": "text" + }, + { + "bbox": [ + 342, + 678, + 357, + 689 + ], + "score": 0.88, + "content": "a , b", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 675, + 461, + 692 + ], + "score": 1.0, + "content": "on the circumference:", + "type": "text" + }, + { + "bbox": [ + 461, + 677, + 504, + 690 + ], + "score": 0.9, + "content": "W _ { E } ^ { \\prime } [ a ] \\ \\approx", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 688, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 142, + 689, + 229, + 700 + ], + "score": 0.88, + "content": "( \\cos ( w _ { k } a ) , \\sin ( w _ { k } a ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 688, + 250, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 250, + 689, + 444, + 701 + ], + "score": 0.77, + "content": "\\bar { W _ { E } ^ { \\prime } [ b ] } \\approx ( \\cos ( \\bar { w _ { k } } b ) , \\sin ( w _ { k } b ) ) ( w _ { k } = 2 \\pi k / p", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 688, + 507, + 702 + ], + "score": 1.0, + "content": "for some inte-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 141, + 700, + 158, + 713 + ], + "score": 1.0, + "content": "ger", + "type": "text" + }, + { + "bbox": [ + 158, + 700, + 214, + 712 + ], + "score": 0.91, + "content": "k \\in [ 1 , p - 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 700, + 287, + 713 + ], + "score": 1.0, + "content": "as in Section 2.1;", + "type": "text" + }, + { + "bbox": [ + 288, + 700, + 305, + 712 + ], + "score": 0.86, + "content": "W _ { E } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "stands for the two currently considered principal", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 711, + 389, + 723 + ], + "spans": [ + { + "bbox": [ + 141, + 711, + 204, + 723 + ], + "score": 1.0, + "content": "components of", + "type": "text" + }, + { + "bbox": [ + 204, + 711, + 221, + 722 + ], + "score": 0.89, + "content": "W _ { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 711, + 389, + 723 + ], + "score": 1.0, + "content": "; rotation and scaling omitted for brevity).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 646, + 505, + 671 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 677, + 505, + 722 + ], + "lines": [], + "index": 22.5, + "bbox_fs": [ + 132, + 675, + 507, + 723 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 134, + 73, + 277, + 84 + ], + "lines": [ + { + "bbox": [ + 132, + 70, + 278, + 87 + ], + "spans": [ + { + "bbox": [ + 132, + 70, + 278, + 87 + ], + "score": 1.0, + "content": "• The embeddings are added to get", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 91, + 448, + 120 + ], + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 120 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 120 + ], + "score": 0.88, + "content": "\\begin{array} { r l } & { \\quad ( \\cos ( w _ { k } a ) + \\cos ( w _ { k } b ) , \\sin ( w _ { k } a ) + \\sin ( w _ { k } b ) ) } \\\\ & { = \\cos ( w _ { k } ( a - b ) / 2 ) \\cdot ( \\cos ( w _ { k } ( a + b ) / 2 ) , \\sin ( w _ { k } ( a + b ) / 2 ) ) } \\end{array}", + "type": "interline_equation", + "image_path": "0ae2d95f43ce3cbd3b3900bde01a24648ad17822897c07fdc74a6cdf00b0bfb9.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 198, + 91, + 448, + 100.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 198, + 100.66666666666667, + 448, + 110.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 198, + 110.33333333333334, + 448, + 120.00000000000001 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 134, + 129, + 504, + 174 + ], + "lines": [ + { + "bbox": [ + 133, + 128, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 133, + 128, + 321, + 142 + ], + "score": 1.0, + "content": "• It is then passed through the first linear layer", + "type": "text" + }, + { + "bbox": [ + 321, + 130, + 333, + 140 + ], + "score": 0.88, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 128, + 505, + 142 + ], + "score": 1.0, + "content": ". Each result entry pre-ReLU will thus be a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 138, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 140, + 138, + 454, + 153 + ], + "score": 1.0, + "content": "linear combination of the two dimensions of the aforementioned vectors, i.e.", + "type": "text" + }, + { + "bbox": [ + 455, + 140, + 505, + 152 + ], + "score": 0.88, + "content": "\\cos ( w _ { k } ( a -", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 150, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 141, + 151, + 356, + 162 + ], + "score": 0.88, + "content": "b ) / 2 ) \\cdot ( \\alpha \\cos ( w _ { k } ( a + b ) / 2 ) + \\beta \\sin ( w _ { k } ( a + b ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 150, + 400, + 164 + ], + "score": 1.0, + "content": ") for some", + "type": "text" + }, + { + "bbox": [ + 400, + 151, + 419, + 162 + ], + "score": 0.9, + "content": "\\alpha , \\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 150, + 505, + 164 + ], + "score": 1.0, + "content": ", which will become", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 162, + 455, + 175 + ], + "spans": [ + { + "bbox": [ + 142, + 162, + 402, + 174 + ], + "score": 0.79, + "content": "\\left| \\cos ( w _ { k } ( a - b ) / 2 ) \\right| | \\alpha \\cos ( w _ { k } ( a + b ) / 2 ) + \\beta \\sin ( w _ { k } ( a + b ) / 2 ) ) |", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 162, + 455, + 175 + ], + "score": 1.0, + "content": "| after ReLU.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 133, + 178, + 504, + 212 + ], + "lines": [ + { + "bbox": [ + 132, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 177, + 387, + 191 + ], + "score": 1.0, + "content": "• These values are then passed through the second linear layer", + "type": "text" + }, + { + "bbox": [ + 387, + 179, + 399, + 190 + ], + "score": 0.87, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 177, + 505, + 191 + ], + "score": 1.0, + "content": ". Empirically the ReLU is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 190, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 141, + 190, + 505, + 202 + ], + "score": 1.0, + "content": "not observed to be effective as the majority of values is positive. The output entries are then", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 200, + 388, + 213 + ], + "spans": [ + { + "bbox": [ + 142, + 200, + 372, + 213 + ], + "score": 1.0, + "content": "simply linear combinations of aforementioned outputs of", + "type": "text" + }, + { + "bbox": [ + 372, + 201, + 384, + 212 + ], + "score": 0.89, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 200, + 388, + 213 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 134, + 217, + 505, + 273 + ], + "lines": [ + { + "bbox": [ + 133, + 216, + 507, + 231 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 507, + 231 + ], + "score": 1.0, + "content": "• The unembedding matrix is finally applied. In the principal components we are considering,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 227, + 507, + 242 + ], + "spans": [ + { + "bbox": [ + 142, + 228, + 272, + 241 + ], + "score": 0.87, + "content": "W _ { U } ^ { \\prime } [ c ] \\approx ( \\cos ( \\bar { w } _ { k } c ) , \\sin ( w _ { k } c ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 227, + 279, + 242 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 279, + 228, + 299, + 241 + ], + "score": 0.87, + "content": "( W _ { U } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 227, + 507, + 242 + ], + "score": 1.0, + "content": "stands for the two currently considered principal", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 239, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 141, + 239, + 207, + 252 + ], + "score": 1.0, + "content": "components of", + "type": "text" + }, + { + "bbox": [ + 208, + 240, + 225, + 250 + ], + "score": 0.87, + "content": "W _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 239, + 506, + 252 + ], + "score": 1.0, + "content": "; rotation and scaling omitted for brevity) and these two principal", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 250, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 141, + 250, + 439, + 262 + ], + "score": 1.0, + "content": "components correspond to a linear combination of the output entries of", + "type": "text" + }, + { + "bbox": [ + 440, + 250, + 452, + 261 + ], + "score": 0.89, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 250, + 506, + 262 + ], + "score": 1.0, + "content": ", which then", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 141, + 261, + 346, + 274 + ], + "score": 1.0, + "content": "correspond to a linear combination of the outputs of", + "type": "text" + }, + { + "bbox": [ + 347, + 262, + 358, + 272 + ], + "score": 0.89, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "(thanks to the non-functional ReLU).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 134, + 277, + 505, + 312 + ], + "lines": [ + { + "bbox": [ + 133, + 276, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 133, + 276, + 233, + 291 + ], + "score": 1.0, + "content": "• Similar to the formula", + "type": "text" + }, + { + "bbox": [ + 233, + 278, + 352, + 290 + ], + "score": 0.92, + "content": "| \\sin ( t ) | - | \\cos ( t ) | \\approx \\cos ( 2 t )", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 276, + 506, + 291 + ], + "score": 1.0, + "content": "discussed in Appendix A, these linear", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 288, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 141, + 288, + 340, + 302 + ], + "score": 1.0, + "content": "combinations provide good approximations for", + "type": "text" + }, + { + "bbox": [ + 340, + 289, + 485, + 301 + ], + "score": 0.89, + "content": "| \\mathrm { { c o s } } ( w _ { k } ( a - b ) / \\bar { 2 } ) \\bar { | } \\cos ( w _ { k } ( a + b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 288, + 506, + 302 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 143, + 299, + 370, + 313 + ], + "spans": [ + { + "bbox": [ + 143, + 299, + 285, + 312 + ], + "score": 0.91, + "content": "| \\mathrm { c o s } ( w _ { k } ( a - b \\bar { ) } / 2 ) | \\mathrm { s i n } \\bar { ( } w _ { k } ( a \\bar { + } b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 299, + 370, + 313 + ], + "score": 1.0, + "content": ". Finally we arrive at", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 317, + 478, + 347 + ], + "lines": [ + { + "bbox": [ + 168, + 317, + 478, + 347 + ], + "spans": [ + { + "bbox": [ + 168, + 317, + 478, + 347 + ], + "score": 0.86, + "content": "\\begin{array} { r l } & { \\quad | \\cos ( w _ { k } ( a - b ) / 2 ) | ( \\cos ( w _ { k } c ) \\cos ( w _ { k } ( a + b ) ) + \\sin ( w _ { k } c ) \\sin ( w _ { k } ( a + b ) ) ) } \\\\ & { = | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a + b - c ) ) } \\end{array}", + "type": "interline_equation", + "image_path": "9da9ab0d9344b20583129c674adb2863f39dad4ec9b66dc3e3a17080209d2c9f.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 168, + 317, + 478, + 327.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 168, + 327.0, + 478, + 337.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 168, + 337.0, + 478, + 347.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "image", + "bbox": [ + 110, + 382, + 502, + 469 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 382, + 502, + 469 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 382, + 502, + 469 + ], + "spans": [ + { + "bbox": [ + 110, + 382, + 502, + 469 + ], + "score": 0.957, + "type": "image", + "image_path": "27e9e673e2d2e485ab1dc14b4b52023a2ca88454d83f6c4353efa0239e77a651.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 110, + 382, + 502, + 411.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 110, + 411.0, + 502, + 440.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 110, + 440.0, + 502, + 469.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 475, + 503, + 487 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "score": 1.0, + "content": "Figure 24: Visualization of the principal components of the embeddings and unembedding matrices.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + } + ], + "index": 24.0 + }, + { + "type": "title", + "bbox": [ + 107, + 513, + 242, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 512, + 243, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 243, + 528 + ], + "score": 1.0, + "content": "L.3 Aligning Weight Matrices", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 568 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 548 + ], + "score": 1.0, + "content": "We first verify that the ReLU from the second layer is not functional. After removing it, the accuracy", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 544, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 198, + 558 + ], + "score": 1.0, + "content": "of the model remains", + "type": "text" + }, + { + "bbox": [ + 199, + 545, + 223, + 556 + ], + "score": 0.88, + "content": "1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 544, + 439, + 558 + ], + "score": 1.0, + "content": "and the cross-entropy loss actually decreased from", + "type": "text" + }, + { + "bbox": [ + 439, + 545, + 493, + 556 + ], + "score": 0.91, + "content": "{ \\mathrm { 6 . 2 0 \\times 1 0 ^ { - 7 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 544, + 506, + 558 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 556, + 162, + 568 + ], + "spans": [ + { + "bbox": [ + 107, + 556, + 158, + 568 + ], + "score": 0.91, + "content": "5 . 8 9 \\times 1 0 ^ { - 7 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 556, + 162, + 567 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 350, + 585 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 351, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 351, + 586 + ], + "score": 1.0, + "content": "Therefore, the model output can be approximately written as", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 591, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 111, + 591, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 111, + 591, + 505, + 605 + ], + "score": 0.86, + "content": "W _ { U } ( b _ { 2 } + W _ { 2 } \\mathrm { R e L U } ( b _ { 1 } + W _ { 1 } ( W _ { E } [ a ] + W _ { E } [ b ] ) ) ) = W _ { U } b _ { 2 } + W _ { U } W _ { 2 } \\mathrm { R e L U } ( b _ { 1 } + W _ { 1 } ( W _ { E } [ a ] + W _ { E } [ b ] ) ) .", + "type": "interline_equation", + "image_path": "8fa5b7823ebbc0a6a9ba4028c30569536cb8d63f33ca30d7152bac0d7d0af27e.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 111, + 591, + 505, + 605 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 252, + 631 + ], + "score": 1.0, + "content": "We now “align” the weight matrices", + "type": "text" + }, + { + "bbox": [ + 253, + 618, + 268, + 629 + ], + "score": 0.9, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 617, + 285, + 631 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 286, + 618, + 301, + 629 + ], + "score": 0.89, + "content": "W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "by mapping through the directions of the principal", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "components of the embeddings and unembeddings. That is, we calculate how these matrices act on", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 639, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 275, + 653 + ], + "score": 1.0, + "content": "and onto the principal directions (consider", + "type": "text" + }, + { + "bbox": [ + 276, + 640, + 296, + 651 + ], + "score": 0.91, + "content": "W _ { 1 } v", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 639, + 410, + 653 + ], + "score": 1.0, + "content": "for every principal direction", + "type": "text" + }, + { + "bbox": [ + 410, + 642, + 417, + 650 + ], + "score": 0.72, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 639, + 428, + 653 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 428, + 640, + 445, + 651 + ], + "score": 0.9, + "content": "W _ { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 639, + 463, + 653 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 464, + 639, + 490, + 651 + ], + "score": 0.92, + "content": "v ^ { T } W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 639, + 506, + 653 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 650, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 208, + 664 + ], + "score": 1.0, + "content": "every principal direction", + "type": "text" + }, + { + "bbox": [ + 208, + 653, + 214, + 660 + ], + "score": 0.68, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 650, + 226, + 664 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 226, + 651, + 243, + 662 + ], + "score": 0.87, + "content": "W _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 650, + 410, + 664 + ], + "score": 1.0, + "content": "). We call the other dimension of aligned", + "type": "text" + }, + { + "bbox": [ + 410, + 651, + 425, + 662 + ], + "score": 0.88, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 650, + 443, + 664 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 443, + 651, + 458, + 662 + ], + "score": 0.89, + "content": "W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 650, + 506, + 664 + ], + "score": 1.0, + "content": "output and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 662, + 285, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 285, + 674 + ], + "score": 1.0, + "content": "source dimensions, respectively (Figure 25).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "In the aligned weight matrices, we can see a clear domino-like pattern: in most output or source", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "dimensions, only two principal components have significant non-zero values, and they correspond to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "a pair of matching circle, or a pizza. In this way, every immediate dimension serves for exactly one", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 712, + 316, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 316, + 723 + ], + "score": 1.0, + "content": "pizza, so the pizzas do not interfere with each other.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + } + ], + "page_idx": 25, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 298, + 740, + 313, + 753 + ], + "spans": [ + { + "bbox": [ + 298, + 740, + 313, + 753 + ], + "score": 1.0, + "content": "26", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 134, + 73, + 277, + 84 + ], + "lines": [ + { + "bbox": [ + 132, + 70, + 278, + 87 + ], + "spans": [ + { + "bbox": [ + 132, + 70, + 278, + 87 + ], + "score": 1.0, + "content": "• The embeddings are added to get", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 132, + 70, + 278, + 87 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 91, + 448, + 120 + ], + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 120 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 120 + ], + "score": 0.88, + "content": "\\begin{array} { r l } & { \\quad ( \\cos ( w _ { k } a ) + \\cos ( w _ { k } b ) , \\sin ( w _ { k } a ) + \\sin ( w _ { k } b ) ) } \\\\ & { = \\cos ( w _ { k } ( a - b ) / 2 ) \\cdot ( \\cos ( w _ { k } ( a + b ) / 2 ) , \\sin ( w _ { k } ( a + b ) / 2 ) ) } \\end{array}", + "type": "interline_equation", + "image_path": "0ae2d95f43ce3cbd3b3900bde01a24648ad17822897c07fdc74a6cdf00b0bfb9.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 198, + 91, + 448, + 100.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 198, + 100.66666666666667, + 448, + 110.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 198, + 110.33333333333334, + 448, + 120.00000000000001 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 134, + 129, + 504, + 174 + ], + "lines": [ + { + "bbox": [ + 133, + 128, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 133, + 128, + 321, + 142 + ], + "score": 1.0, + "content": "• It is then passed through the first linear layer", + "type": "text" + }, + { + "bbox": [ + 321, + 130, + 333, + 140 + ], + "score": 0.88, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 128, + 505, + 142 + ], + "score": 1.0, + "content": ". Each result entry pre-ReLU will thus be a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 138, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 140, + 138, + 454, + 153 + ], + "score": 1.0, + "content": "linear combination of the two dimensions of the aforementioned vectors, i.e.", + "type": "text" + }, + { + "bbox": [ + 455, + 140, + 505, + 152 + ], + "score": 0.88, + "content": "\\cos ( w _ { k } ( a -", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 150, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 141, + 151, + 356, + 162 + ], + "score": 0.88, + "content": "b ) / 2 ) \\cdot ( \\alpha \\cos ( w _ { k } ( a + b ) / 2 ) + \\beta \\sin ( w _ { k } ( a + b ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 150, + 400, + 164 + ], + "score": 1.0, + "content": ") for some", + "type": "text" + }, + { + "bbox": [ + 400, + 151, + 419, + 162 + ], + "score": 0.9, + "content": "\\alpha , \\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 150, + 505, + 164 + ], + "score": 1.0, + "content": ", which will become", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 162, + 455, + 175 + ], + "spans": [ + { + "bbox": [ + 142, + 162, + 402, + 174 + ], + "score": 0.79, + "content": "\\left| \\cos ( w _ { k } ( a - b ) / 2 ) \\right| | \\alpha \\cos ( w _ { k } ( a + b ) / 2 ) + \\beta \\sin ( w _ { k } ( a + b ) / 2 ) ) |", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 162, + 455, + 175 + ], + "score": 1.0, + "content": "| after ReLU.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 133, + 128, + 505, + 175 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 178, + 504, + 212 + ], + "lines": [ + { + "bbox": [ + 132, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 177, + 387, + 191 + ], + "score": 1.0, + "content": "• These values are then passed through the second linear layer", + "type": "text" + }, + { + "bbox": [ + 387, + 179, + 399, + 190 + ], + "score": 0.87, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 177, + 505, + 191 + ], + "score": 1.0, + "content": ". Empirically the ReLU is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 190, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 141, + 190, + 505, + 202 + ], + "score": 1.0, + "content": "not observed to be effective as the majority of values is positive. The output entries are then", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 200, + 388, + 213 + ], + "spans": [ + { + "bbox": [ + 142, + 200, + 372, + 213 + ], + "score": 1.0, + "content": "simply linear combinations of aforementioned outputs of", + "type": "text" + }, + { + "bbox": [ + 372, + 201, + 384, + 212 + ], + "score": 0.89, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 200, + 388, + 213 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 132, + 177, + 505, + 213 + ] + }, + { + "type": "text", + "bbox": [ + 134, + 217, + 505, + 273 + ], + "lines": [ + { + "bbox": [ + 133, + 216, + 507, + 231 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 507, + 231 + ], + "score": 1.0, + "content": "• The unembedding matrix is finally applied. In the principal components we are considering,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 227, + 507, + 242 + ], + "spans": [ + { + "bbox": [ + 142, + 228, + 272, + 241 + ], + "score": 0.87, + "content": "W _ { U } ^ { \\prime } [ c ] \\approx ( \\cos ( \\bar { w } _ { k } c ) , \\sin ( w _ { k } c ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 227, + 279, + 242 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 279, + 228, + 299, + 241 + ], + "score": 0.87, + "content": "( W _ { U } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 227, + 507, + 242 + ], + "score": 1.0, + "content": "stands for the two currently considered principal", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 239, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 141, + 239, + 207, + 252 + ], + "score": 1.0, + "content": "components of", + "type": "text" + }, + { + "bbox": [ + 208, + 240, + 225, + 250 + ], + "score": 0.87, + "content": "W _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 239, + 506, + 252 + ], + "score": 1.0, + "content": "; rotation and scaling omitted for brevity) and these two principal", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 250, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 141, + 250, + 439, + 262 + ], + "score": 1.0, + "content": "components correspond to a linear combination of the output entries of", + "type": "text" + }, + { + "bbox": [ + 440, + 250, + 452, + 261 + ], + "score": 0.89, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 250, + 506, + 262 + ], + "score": 1.0, + "content": ", which then", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 141, + 261, + 346, + 274 + ], + "score": 1.0, + "content": "correspond to a linear combination of the outputs of", + "type": "text" + }, + { + "bbox": [ + 347, + 262, + 358, + 272 + ], + "score": 0.89, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "(thanks to the non-functional ReLU).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 133, + 216, + 507, + 274 + ] + }, + { + "type": "text", + "bbox": [ + 134, + 277, + 505, + 312 + ], + "lines": [ + { + "bbox": [ + 133, + 276, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 133, + 276, + 233, + 291 + ], + "score": 1.0, + "content": "• Similar to the formula", + "type": "text" + }, + { + "bbox": [ + 233, + 278, + 352, + 290 + ], + "score": 0.92, + "content": "| \\sin ( t ) | - | \\cos ( t ) | \\approx \\cos ( 2 t )", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 276, + 506, + 291 + ], + "score": 1.0, + "content": "discussed in Appendix A, these linear", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 288, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 141, + 288, + 340, + 302 + ], + "score": 1.0, + "content": "combinations provide good approximations for", + "type": "text" + }, + { + "bbox": [ + 340, + 289, + 485, + 301 + ], + "score": 0.89, + "content": "| \\mathrm { { c o s } } ( w _ { k } ( a - b ) / \\bar { 2 } ) \\bar { | } \\cos ( w _ { k } ( a + b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 288, + 506, + 302 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 143, + 299, + 370, + 313 + ], + "spans": [ + { + "bbox": [ + 143, + 299, + 285, + 312 + ], + "score": 0.91, + "content": "| \\mathrm { c o s } ( w _ { k } ( a - b \\bar { ) } / 2 ) | \\mathrm { s i n } \\bar { ( } w _ { k } ( a \\bar { + } b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 299, + 370, + 313 + ], + "score": 1.0, + "content": ". Finally we arrive at", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 133, + 276, + 506, + 313 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 317, + 478, + 347 + ], + "lines": [ + { + "bbox": [ + 168, + 317, + 478, + 347 + ], + "spans": [ + { + "bbox": [ + 168, + 317, + 478, + 347 + ], + "score": 0.86, + "content": "\\begin{array} { r l } & { \\quad | \\cos ( w _ { k } ( a - b ) / 2 ) | ( \\cos ( w _ { k } c ) \\cos ( w _ { k } ( a + b ) ) + \\sin ( w _ { k } c ) \\sin ( w _ { k } ( a + b ) ) ) } \\\\ & { = | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a + b - c ) ) } \\end{array}", + "type": "interline_equation", + "image_path": "9da9ab0d9344b20583129c674adb2863f39dad4ec9b66dc3e3a17080209d2c9f.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 168, + 317, + 478, + 327.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 168, + 327.0, + 478, + 337.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 168, + 337.0, + 478, + 347.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "image", + "bbox": [ + 110, + 382, + 502, + 469 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 382, + 502, + 469 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 382, + 502, + 469 + ], + "spans": [ + { + "bbox": [ + 110, + 382, + 502, + 469 + ], + "score": 0.957, + "type": "image", + "image_path": "27e9e673e2d2e485ab1dc14b4b52023a2ca88454d83f6c4353efa0239e77a651.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 110, + 382, + 502, + 411.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 110, + 411.0, + 502, + 440.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 110, + 440.0, + 502, + 469.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 475, + 503, + 487 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "score": 1.0, + "content": "Figure 24: Visualization of the principal components of the embeddings and unembedding matrices.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + } + ], + "index": 24.0 + }, + { + "type": "title", + "bbox": [ + 107, + 513, + 242, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 512, + 243, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 243, + 528 + ], + "score": 1.0, + "content": "L.3 Aligning Weight Matrices", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 568 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 548 + ], + "score": 1.0, + "content": "We first verify that the ReLU from the second layer is not functional. After removing it, the accuracy", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 544, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 198, + 558 + ], + "score": 1.0, + "content": "of the model remains", + "type": "text" + }, + { + "bbox": [ + 199, + 545, + 223, + 556 + ], + "score": 0.88, + "content": "1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 544, + 439, + 558 + ], + "score": 1.0, + "content": "and the cross-entropy loss actually decreased from", + "type": "text" + }, + { + "bbox": [ + 439, + 545, + 493, + 556 + ], + "score": 0.91, + "content": "{ \\mathrm { 6 . 2 0 \\times 1 0 ^ { - 7 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 544, + 506, + 558 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 556, + 162, + 568 + ], + "spans": [ + { + "bbox": [ + 107, + 556, + 158, + 568 + ], + "score": 0.91, + "content": "5 . 8 9 \\times 1 0 ^ { - 7 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 556, + 162, + 567 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 533, + 506, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 350, + 585 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 351, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 351, + 586 + ], + "score": 1.0, + "content": "Therefore, the model output can be approximately written as", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 571, + 351, + 586 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 591, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 111, + 591, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 111, + 591, + 505, + 605 + ], + "score": 0.86, + "content": "W _ { U } ( b _ { 2 } + W _ { 2 } \\mathrm { R e L U } ( b _ { 1 } + W _ { 1 } ( W _ { E } [ a ] + W _ { E } [ b ] ) ) ) = W _ { U } b _ { 2 } + W _ { U } W _ { 2 } \\mathrm { R e L U } ( b _ { 1 } + W _ { 1 } ( W _ { E } [ a ] + W _ { E } [ b ] ) ) .", + "type": "interline_equation", + "image_path": "8fa5b7823ebbc0a6a9ba4028c30569536cb8d63f33ca30d7152bac0d7d0af27e.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 111, + 591, + 505, + 605 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 252, + 631 + ], + "score": 1.0, + "content": "We now “align” the weight matrices", + "type": "text" + }, + { + "bbox": [ + 253, + 618, + 268, + 629 + ], + "score": 0.9, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 617, + 285, + 631 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 286, + 618, + 301, + 629 + ], + "score": 0.89, + "content": "W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "by mapping through the directions of the principal", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "components of the embeddings and unembeddings. That is, we calculate how these matrices act on", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 639, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 275, + 653 + ], + "score": 1.0, + "content": "and onto the principal directions (consider", + "type": "text" + }, + { + "bbox": [ + 276, + 640, + 296, + 651 + ], + "score": 0.91, + "content": "W _ { 1 } v", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 639, + 410, + 653 + ], + "score": 1.0, + "content": "for every principal direction", + "type": "text" + }, + { + "bbox": [ + 410, + 642, + 417, + 650 + ], + "score": 0.72, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 639, + 428, + 653 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 428, + 640, + 445, + 651 + ], + "score": 0.9, + "content": "W _ { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 639, + 463, + 653 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 464, + 639, + 490, + 651 + ], + "score": 0.92, + "content": "v ^ { T } W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 639, + 506, + 653 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 650, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 208, + 664 + ], + "score": 1.0, + "content": "every principal direction", + "type": "text" + }, + { + "bbox": [ + 208, + 653, + 214, + 660 + ], + "score": 0.68, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 650, + 226, + 664 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 226, + 651, + 243, + 662 + ], + "score": 0.87, + "content": "W _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 650, + 410, + 664 + ], + "score": 1.0, + "content": "). We call the other dimension of aligned", + "type": "text" + }, + { + "bbox": [ + 410, + 651, + 425, + 662 + ], + "score": 0.88, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 650, + 443, + 664 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 443, + 651, + 458, + 662 + ], + "score": 0.89, + "content": "W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 650, + 506, + 664 + ], + "score": 1.0, + "content": "output and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 662, + 285, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 285, + 674 + ], + "score": 1.0, + "content": "source dimensions, respectively (Figure 25).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 617, + 506, + 674 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "In the aligned weight matrices, we can see a clear domino-like pattern: in most output or source", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "dimensions, only two principal components have significant non-zero values, and they correspond to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "a pair of matching circle, or a pizza. In this way, every immediate dimension serves for exactly one", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 712, + 316, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 316, + 723 + ], + "score": 1.0, + "content": "pizza, so the pizzas do not interfere with each other.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 678, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 75, + 505, + 245 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 75, + 505, + 245 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 75, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 75, + 505, + 245 + ], + "score": 0.936, + "type": "image", + "image_path": "21d102a425be8f7ecd1521e7f89086bb694334c9adbd6fc9fd7307aea8c45af7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 75, + 505, + 131.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 131.66666666666666, + 505, + 188.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 188.33333333333331, + 505, + 244.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 201, + 252, + 410, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 200, + 249, + 410, + 266 + ], + "spans": [ + { + "bbox": [ + 200, + 249, + 358, + 266 + ], + "score": 1.0, + "content": "Figure 25: Visualization of the aligned", + "type": "text" + }, + { + "bbox": [ + 358, + 252, + 373, + 263 + ], + "score": 0.89, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 249, + 391, + 266 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 392, + 252, + 406, + 263 + ], + "score": 0.89, + "content": "W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 249, + 410, + 266 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 106, + 285, + 197, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 284, + 199, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 199, + 299 + ], + "score": 1.0, + "content": "L.4 Approximation", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 305, + 506, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "score": 1.0, + "content": "Everything becomes much clearer after realigning the matrices. For a pizza and its two corresponding", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 315, + 509, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 299, + 330 + ], + "score": 1.0, + "content": "principal embedding / unembedding dimensions,", + "type": "text" + }, + { + "bbox": [ + 299, + 316, + 509, + 329 + ], + "score": 0.7, + "content": "W _ { E } ^ { \\prime } [ a ] + W _ { E } ^ { \\prime } [ b ] \\approx \\bar { \\cos } ( w _ { k } ( a - b ) / 2 ) \\cdot ( \\cos ( w _ { k } ( a +", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 127, + 339 + ], + "score": 0.71, + "content": "\\bar { b } ) / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 326, + 133, + 340 + ], + "score": 1.0, + "content": "),", + "type": "text" + }, + { + "bbox": [ + 133, + 327, + 207, + 339 + ], + "score": 0.83, + "content": "\\sin ( w _ { k } ( a + \\bar { b } ) / 2 ) ;", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 326, + 329, + 340 + ], + "score": 1.0, + "content": "will be mapped by realigned", + "type": "text" + }, + { + "bbox": [ + 330, + 328, + 344, + 338 + ], + "score": 0.88, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "into its corresponding columns (which", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 270, + 352 + ], + "score": 1.0, + "content": "are different for every pizza), added with", + "type": "text" + }, + { + "bbox": [ + 270, + 339, + 280, + 349 + ], + "score": 0.87, + "content": "b _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 337, + 506, + 352 + ], + "score": 1.0, + "content": "and apply ReLU. The result will then be mapped by the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 348, + 454, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 146, + 362 + ], + "score": 1.0, + "content": "realigned", + "type": "text" + }, + { + "bbox": [ + 146, + 349, + 161, + 360 + ], + "score": 0.89, + "content": "W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 348, + 252, + 362 + ], + "score": 1.0, + "content": ", added with realigned", + "type": "text" + }, + { + "bbox": [ + 252, + 349, + 262, + 360 + ], + "score": 0.87, + "content": "b _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 348, + 364, + 362 + ], + "score": 1.0, + "content": ", and finally multipled by", + "type": "text" + }, + { + "bbox": [ + 365, + 349, + 450, + 361 + ], + "score": 0.87, + "content": "( \\cos ( w _ { k } c ) , \\sin ( w _ { k } c ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 348, + 454, + 362 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 365, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 104, + 364, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 104, + 364, + 296, + 378 + ], + "score": 1.0, + "content": "For the first two principal dimensions, realigned", + "type": "text" + }, + { + "bbox": [ + 297, + 365, + 312, + 376 + ], + "score": 0.89, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 364, + 506, + 378 + ], + "score": 1.0, + "content": "has 44 corresponding columns (with coefficients", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 375, + 504, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 174, + 389 + ], + "score": 1.0, + "content": "of absolute value", + "type": "text" + }, + { + "bbox": [ + 175, + 377, + 201, + 387 + ], + "score": 0.83, + "content": "> 0 . 1 \\AA ,", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 375, + 313, + 389 + ], + "score": 1.0, + "content": "). Let the embedded input be", + "type": "text" + }, + { + "bbox": [ + 313, + 376, + 502, + 388 + ], + "score": 0.91, + "content": "( x , y ) = W _ { E } ^ { \\prime } [ a ] + W _ { E } ^ { \\prime } [ b ] \\approx \\cos ( w _ { k } ( a - b ) / 2 ) \\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 375, + 504, + 389 + ], + "score": 1.0, + "content": "·", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 108, + 387, + 385, + 400 + ], + "spans": [ + { + "bbox": [ + 108, + 387, + 182, + 399 + ], + "score": 0.72, + "content": "( \\cos ( w _ { k } ( a + b ) / 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 387, + 187, + 400 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 187, + 387, + 261, + 399 + ], + "score": 0.79, + "content": "\\sin ( w _ { k } ( a + b ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 387, + 385, + 400 + ], + "score": 1.0, + "content": "), the intermediate columns are", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 105, + 403, + 506, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 507, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 113, + 416 + ], + "score": 1.0, + "content": "R", + "type": "text" + }, + { + "bbox": [ + 113, + 405, + 507, + 415 + ], + "score": 0.53, + "content": "\\operatorname { e L U } ( [ 0 . 5 3 0 x - 1 . 1 3 5 y + 0 . 2 5 3 , - 0 . 1 6 4 x - 1 . 1 0 0 y + 0 . 2 0 5 , 1 . 2 1 0 x - 0 . 3 7 0 y + 0 . 1 9 8 , - 0 . 4 7 8 x - 0 . 0 9 2 x + 0 . 0 9 2 x + 0 . 0 9 2 x )", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 413, + 504, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 134, + 428 + ], + "score": 1.0, + "content": "1.072y", + "type": "text" + }, + { + "bbox": [ + 135, + 415, + 336, + 425 + ], + "score": 0.65, + "content": "+ ~ 0 . 2 1 5 , - 1 . 0 1 7 x ~ + ~ 0 . 7 9 9 y ~ + ~ 0 . 2 4 9 , 0 . 3 4 2 x", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 415, + 416, + 425 + ], + "score": 0.25, + "content": "+ ~ 0 . 7 9 9 y + 0 . 2 4 9 , 0 . 3 4 2 x ~ - ~ 0 . 0 4 8 y ~ + ~ 0 . 0 8 5", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 416, + 504, + 426 + ], + "score": 0.62, + "content": "0 . 3 4 2 x \\mathrm { ~ - ~ } 0 . 0 4 8 y \\mathrm { ~ + ~ } 0 . 0 8 5 , 1 . 1 4 9 x \\mathrm { ~ - ~ } 0 . 5 9 8 y \\mathrm { ~ + ~ }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 429, + 500, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 500, + 437 + ], + "score": 0.4, + "content": "0 . 2 1 2 , - 0 . 4 4 3 x + 1 . 5 5 0 y + 0 . 1 3 9 9 , - 1 . 0 8 0 x - 0 . 0 0 0 y + 0 . 1 5 1 , - 1 . 4 0 5 x + 0 . 4 1 0 y + 0 . 1 ( 8 , 1 . 0 3 8 x + 0 . 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 )", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 434, + 507, + 450 + ], + "spans": [ + { + "bbox": [ + 104, + 434, + 134, + 450 + ], + "score": 1.0, + "content": "0.905y", + "type": "text" + }, + { + "bbox": [ + 135, + 438, + 336, + 447 + ], + "score": 0.51, + "content": "\\mathrm { ~ ~ { ~ \\tau ~ } ~ } + 0 . 1 9 0 , 0 . 5 6 8 x + 1 . 1 8 8 y + 0 . 1 2 8 , 0 . 2 3 5 x -", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 438, + 450, + 447 + ], + "score": 0.59, + "content": "y \\ + \\ 0 . 1 2 8 , 0 . 2 3 5 x \\ - \\ 1 . 3 3 7 y \\ + \\ 0 . 1 6 4 , - 1 . 1 8 0 x", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 434, + 507, + 450 + ], + "score": 1.0, + "content": "+ 1.052y +", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 134, + 460 + ], + "score": 1.0, + "content": "0.139,", + "type": "text" + }, + { + "bbox": [ + 138, + 448, + 506, + 455 + ], + "score": 0.28, + "content": "- 0 . 1 7 3 x + 0 . 9 1 8 u + 0 . 1 4 8 . - 0 . 9 0 0 x + 1 . 0 6 0 u + 0 . 1 7 3 . - 1 . 3 4 9 x + 0 . 3 9 0 u + 0 . 2 5 6 . 0 , 1 0 5 x - 0 . 0 9 8 x + 0 . 2 5 6 . 0 , 0 . 1 2 5 x + 0 . 2 5 6 . 0 , 0 . 1 2 5 x + 0 . 2 5 6 . 0 , 0 . 1 2 5 x + 0 . 2 5 6 . 0 , 0 . 1 2 5 x + 0 . 2 5 6 . 0 , 0 . 2 5 6 . - 0 . 3 2 5 x", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 449, + 450, + 459 + ], + "score": 0.55, + "content": "- 0 . 2 0 0 x + 1 . 0 6 0 y + 0 . 1 7 3 , - 1 . 3 4 2 x + 0 . 3 9 0 y + 0 . 2", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 457, + 507, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 111, + 471 + ], + "score": 1.0, + "content": "1", + "type": "text" + }, + { + "bbox": [ + 111, + 460, + 507, + 468 + ], + "score": 0.57, + "content": "\\ 2 4 6 y \\ + \\ 0 . 2 0 9 , 0 . 1 1 5 x \\ + \\ 1 . 2 9 3 y \\ + \\ 0 . 1 9 7 , 0 . 2 5 2 x \\ + \\ 1 . 2 4 7 y \\ + \\ 0 . 1 4 0 , - 0 . 4 9 3 x \\ + \\ 1 . 2 5 2 y \\ +", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 472, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 505, + 480 + ], + "score": 0.48, + "content": "\\begin{array} { r } . 1 2 0 x + 0 . 2 6 2 y + 0 . 2 3 9 , 0 . 6 6 8 x + 1 . 0 9 6 y + 0 . 2 0 5 , - 0 . 4 8 ( x - 1 . 3 0 2 y + 0 . 1 4 5 , 1 . 1 3 4 x - 0 . 0 0 0 0 ) x + 0 . 0 0 0 0 x + 0 . 0 0 0 0 x = 1 . 0 0 0 0 x + 0 . 0 0 0 0 x + 0 . 0 0 0 0 x + 0 . 0 0 0 0 x = 1 . 0 0 0 0 x - 0 . 0 0 0 0 x + 0 . 0 0 0 0 x + 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 0 x + 0 . 0 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 0 x = 1 . 0 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 0 , 0 0 0 = 1 . 0 0 0 0 x = 1 . 0 0 0 0 0 x = 1 . 0 0 0 0 , 0 0 0 = 1 . 0 0 0 0 x = 1 . 0 0 0 0 0 x = 1 . 0 0 0 0 0 x = 1 . 0 0 0 0 0 , 0 0 = 1 . 0 0 0 0 0 x = 1 . 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 = 1 . 0 0 0 0 0 x = 0 0 . 0 0 0 0 0 = 1 . 0 0 0 0 0 0 x = 0 0 . 0 0 0 0 0 = 1 . 0 0 0 0 0 0 0 x = 0 . 0 0 0 0 0 0 = 0 . 0 0 0 0 0 0 . 0 0 0 0 0 = 0 0 . 0 0 0 0 0 0 . 0 0 0 0 0 0 = 0 0 . 0 0 0 0 0 0 . 0 0 0 0 0 0 . 0 0 0 0 0 0 . 0 0 0 0 0 0 . 0 0 0 0 0 0 0 0 . 0 0 0 0 0 0 0 0 . 0 0 0 0 0 \\end{array}", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 478, + 508, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 111, + 493 + ], + "score": 1.0, + "content": "0", + "type": "text" + }, + { + "bbox": [ + 111, + 482, + 508, + 491 + ], + "score": 0.63, + "content": "1 . 8 6 2 y + 0 . 2 7 3 , 1 . 1 4 3 x + 0 . 4 3 5 y + 0 . 1 7 1 , - 1 . 2 8 5 x - 0 . 6 4 4 y + 0 . 1 4 2 , - 1 . 4 5 4 x - 0 . 2 8 5 y + 0 . 2 2 7 , 0 . 2 8 5 y + 0 . 2 2 7 , 0 . 2 2 7 5 + 0 . 2 2 8 5 y + 0 . 2 2 8 5 y", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 147, + 503 + ], + "score": 1.0, + "content": "0.218, −0", + "type": "text" + }, + { + "bbox": [ + 148, + 493, + 349, + 502 + ], + "score": 0.37, + "content": "9 2 4 x + 1 . 0 6 8 y + 0 . 1 4 5 , - 0 . 4 0 1 x + 0 . 1 6 7 y + 0 . 1 0 6", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 493, + 497, + 501 + ], + "score": 0.48, + "content": ". 1 6 7 y + 0 . 1 0 6 , - 0 . 4 1 1 x - 1 . 3 8 9 y + 0 . 2 4 9 , 1 . 4 2 2 x", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "−", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 500, + 507, + 515 + ], + "spans": [ + { + "bbox": [ + 104, + 500, + 134, + 515 + ], + "score": 1.0, + "content": "0.117y", + "type": "text" + }, + { + "bbox": [ + 135, + 504, + 336, + 513 + ], + "score": 0.53, + "content": "\\mathbf { \\xi } ^ { \\prime } + 0 . 2 2 7 , - 0 . 8 5 9 x - 0 . 7 7 8 y + 0 . 1 2 1 , - 0 . 5 2 8 x", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 503, + 410, + 513 + ], + "score": 0.25, + "content": "x \\ : - \\ : 0 . 7 7 8 y + 0 . 1 2 1 , - 0 . 5 2 8 x \\ : - \\ : 0 . 2 1 6 y + 0 . 0 9 \\ : \\hat ", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 503, + 497, + 513 + ], + "score": 0.53, + "content": "- 0 . 5 2 8 x \\mathrm { ~ - ~ } 0 . 2 1 6 y + 0 . 0 9 7 , - 0 . 8 8 4 x \\mathrm { ~ - ~ } 0 . 7 2 4 y", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 500, + 507, + 515 + ], + "score": 1.0, + "content": "+", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 511, + 510, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 111, + 525 + ], + "score": 1.0, + "content": "0", + "type": "text" + }, + { + "bbox": [ + 111, + 513, + 510, + 523 + ], + "score": 0.54, + "content": "1 . 1 7 1 , 1 . 1 9 3 x + 0 . 7 2 4 y + 0 . 1 3 1 , 1 . 0 8 6 x + 0 . 6 6 7 y + 0 . 2 1 8 , 0 . 4 0 2 x + 1 . 2 4 0 y + 0 . 2 1 3 , 1 . 0 6 9 x - 0 . 9 0 3 y + 0 . 2 2 5 , 0 . 2 4 0 x + 0 . 2 2 5 , 0 . 3 2 5 , 0 . 4 2 5 , x = 0 . 4 2 5 , x = 0 . 4 2 5 , x = 0 . 4 4 , x = 0 . 5 3 0", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 523, + 511, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 111, + 536 + ], + "score": 1.0, + "content": "0", + "type": "text" + }, + { + "bbox": [ + 111, + 525, + 511, + 534 + ], + "score": 0.48, + "content": "1 . 1 2 0 , 0 . 5 0 6 x - 1 . 0 4 2 y + 0 . 1 5 3 , 1 . 4 0 4 x - 0 . 0 6 4 y + 0 . 1 5 2 , 0 . 6 9 6 x - 1 . 2 4 9 y + 0 . 1 9 9 , - 0 . 7 5 2 x - 0 . 0 9 6 x + 0 . 0 9 6 x - 0 . 0 9 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 5 . 0 1 5 , 0 . 0 1 5 , 0 . 0 1 5 , 0 . 1 0 . 1 5 , 0 0 . 1 5 , 0 0 . 1 5 , 0 0 . 1 0 . 1 5 0 , 0 . 1 5 0 , 0 . 1 0 0 .", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 108, + 533, + 297, + 548 + ], + "spans": [ + { + "bbox": [ + 108, + 536, + 293, + 546 + ], + "score": 0.71, + "content": "0 . 8 8 0 y + 0 . 1 0 6 , - 0 . 9 5 6 x - 0 . 5 8 1 y + 0 . 2 2 3 ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 533, + 297, + 548 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 426, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 427, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 427, + 564 + ], + "score": 1.0, + "content": "For the first principal unembedding dimension, it will be taken dot product with", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 566, + 509, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 501, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 501, + 579 + ], + "score": 1.0, + "content": "[1.326, 0.179, 0.142, −0.458, 1.101, −0.083, 0.621, 1.255, −0.709, 0.123, −1.346, −0.571, 1.016,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 578, + 490, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 490, + 590 + ], + "score": 1.0, + "content": "1.337, 0.732, 0.839, 0.129, 0.804, 0.377, 0.078, 1.322, −1.021, −0.799, −0.339, 1.117, −1.162,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 590, + 510, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 231, + 601 + ], + "score": 1.0, + "content": "−1.423, −1.157, 1.363, 0.156,", + "type": "text" + }, + { + "bbox": [ + 231, + 590, + 262, + 600 + ], + "score": 0.4, + "content": "- 0 . 1 6 5", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 590, + 265, + 601 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 266, + 590, + 297, + 600 + ], + "score": 0.26, + "content": "- 0 . 4 5 1", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 590, + 510, + 601 + ], + "score": 1.0, + "content": ", −1.101, −0.572, −1.180, −1.386, −1.346, −0.226,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 600, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 294, + 612 + ], + "score": 0.72, + "content": "1 . 0 9 1 , 1 . 1 5 9 , - 0 . 5 2 4 , 1 . 4 4 1 , - 0 . 9 4 9 , - 1 . 2 4 8 ] .", + "type": "inline_equation" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 506, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 507, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 181, + 630 + ], + "score": 1.0, + "content": "Call this function", + "type": "text" + }, + { + "bbox": [ + 181, + 616, + 211, + 628 + ], + "score": 0.91, + "content": "f ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 615, + 291, + 630 + ], + "score": 1.0, + "content": ". When we plug in", + "type": "text" + }, + { + "bbox": [ + 291, + 616, + 386, + 628 + ], + "score": 0.9, + "content": "x = \\cos ( t ) , y = \\sin ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 615, + 507, + 630 + ], + "score": 1.0, + "content": ", we get a function that well-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 165, + 639 + ], + "score": 1.0, + "content": "approximated", + "type": "text" + }, + { + "bbox": [ + 166, + 628, + 220, + 639 + ], + "score": 0.87, + "content": "8 \\cos ( 2 t + 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 627, + 334, + 639 + ], + "score": 1.0, + "content": "(Figure 26). Therefore, let", + "type": "text" + }, + { + "bbox": [ + 334, + 627, + 405, + 639 + ], + "score": 0.89, + "content": "t = w _ { k } ( a + b ) / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 627, + 506, + 639 + ], + "score": 1.0, + "content": ", the dot product will be", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 638, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 168, + 650 + ], + "score": 1.0, + "content": "approximately", + "type": "text" + }, + { + "bbox": [ + 168, + 638, + 335, + 650 + ], + "score": 0.9, + "content": "8 | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a + b ) + 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 638, + 351, + 650 + ], + "score": 1.0, + "content": ", or", + "type": "text" + }, + { + "bbox": [ + 351, + 638, + 494, + 650 + ], + "score": 0.89, + "content": "| \\mathrm { c o s } ( w _ { k } ( a - b ) / 2 ) | \\mathrm { c o s } ( w _ { k } ( a + b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 638, + 506, + 650 + ], + "score": 1.0, + "content": "if", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 649, + 430, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 430, + 662 + ], + "score": 1.0, + "content": "we ignore the phase and scaling. This completes the picture we described above.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + } + ], + "page_idx": 26, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 298, + 740, + 313, + 755 + ], + "spans": [ + { + "bbox": [ + 298, + 740, + 313, + 755 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 75, + 505, + 245 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 75, + 505, + 245 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 75, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 75, + 505, + 245 + ], + "score": 0.936, + "type": "image", + "image_path": "21d102a425be8f7ecd1521e7f89086bb694334c9adbd6fc9fd7307aea8c45af7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 75, + 505, + 131.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 131.66666666666666, + 505, + 188.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 188.33333333333331, + 505, + 244.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 201, + 252, + 410, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 200, + 249, + 410, + 266 + ], + "spans": [ + { + "bbox": [ + 200, + 249, + 358, + 266 + ], + "score": 1.0, + "content": "Figure 25: Visualization of the aligned", + "type": "text" + }, + { + "bbox": [ + 358, + 252, + 373, + 263 + ], + "score": 0.89, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 249, + 391, + 266 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 392, + 252, + 406, + 263 + ], + "score": 0.89, + "content": "W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 249, + 410, + 266 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 106, + 285, + 197, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 284, + 199, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 199, + 299 + ], + "score": 1.0, + "content": "L.4 Approximation", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 305, + 506, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "score": 1.0, + "content": "Everything becomes much clearer after realigning the matrices. For a pizza and its two corresponding", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 315, + 509, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 299, + 330 + ], + "score": 1.0, + "content": "principal embedding / unembedding dimensions,", + "type": "text" + }, + { + "bbox": [ + 299, + 316, + 509, + 329 + ], + "score": 0.7, + "content": "W _ { E } ^ { \\prime } [ a ] + W _ { E } ^ { \\prime } [ b ] \\approx \\bar { \\cos } ( w _ { k } ( a - b ) / 2 ) \\cdot ( \\cos ( w _ { k } ( a +", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 127, + 339 + ], + "score": 0.71, + "content": "\\bar { b } ) / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 326, + 133, + 340 + ], + "score": 1.0, + "content": "),", + "type": "text" + }, + { + "bbox": [ + 133, + 327, + 207, + 339 + ], + "score": 0.83, + "content": "\\sin ( w _ { k } ( a + \\bar { b } ) / 2 ) ;", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 326, + 329, + 340 + ], + "score": 1.0, + "content": "will be mapped by realigned", + "type": "text" + }, + { + "bbox": [ + 330, + 328, + 344, + 338 + ], + "score": 0.88, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "into its corresponding columns (which", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 270, + 352 + ], + "score": 1.0, + "content": "are different for every pizza), added with", + "type": "text" + }, + { + "bbox": [ + 270, + 339, + 280, + 349 + ], + "score": 0.87, + "content": "b _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 337, + 506, + 352 + ], + "score": 1.0, + "content": "and apply ReLU. The result will then be mapped by the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 348, + 454, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 146, + 362 + ], + "score": 1.0, + "content": "realigned", + "type": "text" + }, + { + "bbox": [ + 146, + 349, + 161, + 360 + ], + "score": 0.89, + "content": "W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 348, + 252, + 362 + ], + "score": 1.0, + "content": ", added with realigned", + "type": "text" + }, + { + "bbox": [ + 252, + 349, + 262, + 360 + ], + "score": 0.87, + "content": "b _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 348, + 364, + 362 + ], + "score": 1.0, + "content": ", and finally multipled by", + "type": "text" + }, + { + "bbox": [ + 365, + 349, + 450, + 361 + ], + "score": 0.87, + "content": "( \\cos ( w _ { k } c ) , \\sin ( w _ { k } c ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 348, + 454, + 362 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 304, + 509, + 362 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 365, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 104, + 364, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 104, + 364, + 296, + 378 + ], + "score": 1.0, + "content": "For the first two principal dimensions, realigned", + "type": "text" + }, + { + "bbox": [ + 297, + 365, + 312, + 376 + ], + "score": 0.89, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 364, + 506, + 378 + ], + "score": 1.0, + "content": "has 44 corresponding columns (with coefficients", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 375, + 504, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 174, + 389 + ], + "score": 1.0, + "content": "of absolute value", + "type": "text" + }, + { + "bbox": [ + 175, + 377, + 201, + 387 + ], + "score": 0.83, + "content": "> 0 . 1 \\AA ,", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 375, + 313, + 389 + ], + "score": 1.0, + "content": "). Let the embedded input be", + "type": "text" + }, + { + "bbox": [ + 313, + 376, + 502, + 388 + ], + "score": 0.91, + "content": "( x , y ) = W _ { E } ^ { \\prime } [ a ] + W _ { E } ^ { \\prime } [ b ] \\approx \\cos ( w _ { k } ( a - b ) / 2 ) \\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 375, + 504, + 389 + ], + "score": 1.0, + "content": "·", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 108, + 387, + 385, + 400 + ], + "spans": [ + { + "bbox": [ + 108, + 387, + 182, + 399 + ], + "score": 0.72, + "content": "( \\cos ( w _ { k } ( a + b ) / 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 387, + 187, + 400 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 187, + 387, + 261, + 399 + ], + "score": 0.79, + "content": "\\sin ( w _ { k } ( a + b ) / 2 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 387, + 385, + 400 + ], + "score": 1.0, + "content": "), the intermediate columns are", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 104, + 364, + 506, + 400 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 403, + 506, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 507, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 113, + 416 + ], + "score": 1.0, + "content": "R", + "type": "text" + }, + { + "bbox": [ + 113, + 405, + 507, + 415 + ], + "score": 0.53, + "content": "\\operatorname { e L U } ( [ 0 . 5 3 0 x - 1 . 1 3 5 y + 0 . 2 5 3 , - 0 . 1 6 4 x - 1 . 1 0 0 y + 0 . 2 0 5 , 1 . 2 1 0 x - 0 . 3 7 0 y + 0 . 1 9 8 , - 0 . 4 7 8 x - 0 . 0 9 2 x + 0 . 0 9 2 x + 0 . 0 9 2 x )", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 413, + 504, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 134, + 428 + ], + "score": 1.0, + "content": "1.072y", + "type": "text" + }, + { + "bbox": [ + 135, + 415, + 336, + 425 + ], + "score": 0.65, + "content": "+ ~ 0 . 2 1 5 , - 1 . 0 1 7 x ~ + ~ 0 . 7 9 9 y ~ + ~ 0 . 2 4 9 , 0 . 3 4 2 x", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 415, + 416, + 425 + ], + "score": 0.25, + "content": "+ ~ 0 . 7 9 9 y + 0 . 2 4 9 , 0 . 3 4 2 x ~ - ~ 0 . 0 4 8 y ~ + ~ 0 . 0 8 5", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 416, + 504, + 426 + ], + "score": 0.62, + "content": "0 . 3 4 2 x \\mathrm { ~ - ~ } 0 . 0 4 8 y \\mathrm { ~ + ~ } 0 . 0 8 5 , 1 . 1 4 9 x \\mathrm { ~ - ~ } 0 . 5 9 8 y \\mathrm { ~ + ~ }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 429, + 500, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 500, + 437 + ], + "score": 0.4, + "content": "0 . 2 1 2 , - 0 . 4 4 3 x + 1 . 5 5 0 y + 0 . 1 3 9 9 , - 1 . 0 8 0 x - 0 . 0 0 0 y + 0 . 1 5 1 , - 1 . 4 0 5 x + 0 . 4 1 0 y + 0 . 1 ( 8 , 1 . 0 3 8 x + 0 . 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 )", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 434, + 507, + 450 + ], + "spans": [ + { + "bbox": [ + 104, + 434, + 134, + 450 + ], + "score": 1.0, + "content": "0.905y", + "type": "text" + }, + { + "bbox": [ + 135, + 438, + 336, + 447 + ], + "score": 0.51, + "content": "\\mathrm { ~ ~ { ~ \\tau ~ } ~ } + 0 . 1 9 0 , 0 . 5 6 8 x + 1 . 1 8 8 y + 0 . 1 2 8 , 0 . 2 3 5 x -", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 438, + 450, + 447 + ], + "score": 0.59, + "content": "y \\ + \\ 0 . 1 2 8 , 0 . 2 3 5 x \\ - \\ 1 . 3 3 7 y \\ + \\ 0 . 1 6 4 , - 1 . 1 8 0 x", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 434, + 507, + 450 + ], + "score": 1.0, + "content": "+ 1.052y +", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 134, + 460 + ], + "score": 1.0, + "content": "0.139,", + "type": "text" + }, + { + "bbox": [ + 138, + 448, + 506, + 455 + ], + "score": 0.28, + "content": "- 0 . 1 7 3 x + 0 . 9 1 8 u + 0 . 1 4 8 . - 0 . 9 0 0 x + 1 . 0 6 0 u + 0 . 1 7 3 . - 1 . 3 4 9 x + 0 . 3 9 0 u + 0 . 2 5 6 . 0 , 1 0 5 x - 0 . 0 9 8 x + 0 . 2 5 6 . 0 , 0 . 1 2 5 x + 0 . 2 5 6 . 0 , 0 . 1 2 5 x + 0 . 2 5 6 . 0 , 0 . 1 2 5 x + 0 . 2 5 6 . 0 , 0 . 1 2 5 x + 0 . 2 5 6 . 0 , 0 . 2 5 6 . - 0 . 3 2 5 x", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 449, + 450, + 459 + ], + "score": 0.55, + "content": "- 0 . 2 0 0 x + 1 . 0 6 0 y + 0 . 1 7 3 , - 1 . 3 4 2 x + 0 . 3 9 0 y + 0 . 2", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 457, + 507, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 111, + 471 + ], + "score": 1.0, + "content": "1", + "type": "text" + }, + { + "bbox": [ + 111, + 460, + 507, + 468 + ], + "score": 0.57, + "content": "\\ 2 4 6 y \\ + \\ 0 . 2 0 9 , 0 . 1 1 5 x \\ + \\ 1 . 2 9 3 y \\ + \\ 0 . 1 9 7 , 0 . 2 5 2 x \\ + \\ 1 . 2 4 7 y \\ + \\ 0 . 1 4 0 , - 0 . 4 9 3 x \\ + \\ 1 . 2 5 2 y \\ +", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 472, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 505, + 480 + ], + "score": 0.48, + "content": "\\begin{array} { r } . 1 2 0 x + 0 . 2 6 2 y + 0 . 2 3 9 , 0 . 6 6 8 x + 1 . 0 9 6 y + 0 . 2 0 5 , - 0 . 4 8 ( x - 1 . 3 0 2 y + 0 . 1 4 5 , 1 . 1 3 4 x - 0 . 0 0 0 0 ) x + 0 . 0 0 0 0 x + 0 . 0 0 0 0 x = 1 . 0 0 0 0 x + 0 . 0 0 0 0 x + 0 . 0 0 0 0 x + 0 . 0 0 0 0 x = 1 . 0 0 0 0 x - 0 . 0 0 0 0 x + 0 . 0 0 0 0 x + 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 0 x + 0 . 0 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 0 x = 1 . 0 0 0 0 , 0 . 0 0 0 0 x = 1 . 0 0 0 0 , 0 0 0 = 1 . 0 0 0 0 x = 1 . 0 0 0 0 0 x = 1 . 0 0 0 0 , 0 0 0 = 1 . 0 0 0 0 x = 1 . 0 0 0 0 0 x = 1 . 0 0 0 0 0 x = 1 . 0 0 0 0 0 , 0 0 = 1 . 0 0 0 0 0 x = 1 . 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 x = 1 . 0 0 0 0 0 0 = 1 . 0 0 0 0 0 x = 0 0 . 0 0 0 0 0 = 1 . 0 0 0 0 0 0 x = 0 0 . 0 0 0 0 0 = 1 . 0 0 0 0 0 0 0 x = 0 . 0 0 0 0 0 0 = 0 . 0 0 0 0 0 0 . 0 0 0 0 0 = 0 0 . 0 0 0 0 0 0 . 0 0 0 0 0 0 = 0 0 . 0 0 0 0 0 0 . 0 0 0 0 0 0 . 0 0 0 0 0 0 . 0 0 0 0 0 0 . 0 0 0 0 0 0 0 0 . 0 0 0 0 0 0 0 0 . 0 0 0 0 0 \\end{array}", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 478, + 508, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 111, + 493 + ], + "score": 1.0, + "content": "0", + "type": "text" + }, + { + "bbox": [ + 111, + 482, + 508, + 491 + ], + "score": 0.63, + "content": "1 . 8 6 2 y + 0 . 2 7 3 , 1 . 1 4 3 x + 0 . 4 3 5 y + 0 . 1 7 1 , - 1 . 2 8 5 x - 0 . 6 4 4 y + 0 . 1 4 2 , - 1 . 4 5 4 x - 0 . 2 8 5 y + 0 . 2 2 7 , 0 . 2 8 5 y + 0 . 2 2 7 , 0 . 2 2 7 5 + 0 . 2 2 8 5 y + 0 . 2 2 8 5 y", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 147, + 503 + ], + "score": 1.0, + "content": "0.218, −0", + "type": "text" + }, + { + "bbox": [ + 148, + 493, + 349, + 502 + ], + "score": 0.37, + "content": "9 2 4 x + 1 . 0 6 8 y + 0 . 1 4 5 , - 0 . 4 0 1 x + 0 . 1 6 7 y + 0 . 1 0 6", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 493, + 497, + 501 + ], + "score": 0.48, + "content": ". 1 6 7 y + 0 . 1 0 6 , - 0 . 4 1 1 x - 1 . 3 8 9 y + 0 . 2 4 9 , 1 . 4 2 2 x", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "−", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 500, + 507, + 515 + ], + "spans": [ + { + "bbox": [ + 104, + 500, + 134, + 515 + ], + "score": 1.0, + "content": "0.117y", + "type": "text" + }, + { + "bbox": [ + 135, + 504, + 336, + 513 + ], + "score": 0.53, + "content": "\\mathbf { \\xi } ^ { \\prime } + 0 . 2 2 7 , - 0 . 8 5 9 x - 0 . 7 7 8 y + 0 . 1 2 1 , - 0 . 5 2 8 x", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 503, + 410, + 513 + ], + "score": 0.25, + "content": "x \\ : - \\ : 0 . 7 7 8 y + 0 . 1 2 1 , - 0 . 5 2 8 x \\ : - \\ : 0 . 2 1 6 y + 0 . 0 9 \\ : \\hat ", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 503, + 497, + 513 + ], + "score": 0.53, + "content": "- 0 . 5 2 8 x \\mathrm { ~ - ~ } 0 . 2 1 6 y + 0 . 0 9 7 , - 0 . 8 8 4 x \\mathrm { ~ - ~ } 0 . 7 2 4 y", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 500, + 507, + 515 + ], + "score": 1.0, + "content": "+", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 511, + 510, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 111, + 525 + ], + "score": 1.0, + "content": "0", + "type": "text" + }, + { + "bbox": [ + 111, + 513, + 510, + 523 + ], + "score": 0.54, + "content": "1 . 1 7 1 , 1 . 1 9 3 x + 0 . 7 2 4 y + 0 . 1 3 1 , 1 . 0 8 6 x + 0 . 6 6 7 y + 0 . 2 1 8 , 0 . 4 0 2 x + 1 . 2 4 0 y + 0 . 2 1 3 , 1 . 0 6 9 x - 0 . 9 0 3 y + 0 . 2 2 5 , 0 . 2 4 0 x + 0 . 2 2 5 , 0 . 3 2 5 , 0 . 4 2 5 , x = 0 . 4 2 5 , x = 0 . 4 2 5 , x = 0 . 4 4 , x = 0 . 5 3 0", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 523, + 511, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 111, + 536 + ], + "score": 1.0, + "content": "0", + "type": "text" + }, + { + "bbox": [ + 111, + 525, + 511, + 534 + ], + "score": 0.48, + "content": "1 . 1 2 0 , 0 . 5 0 6 x - 1 . 0 4 2 y + 0 . 1 5 3 , 1 . 4 0 4 x - 0 . 0 6 4 y + 0 . 1 5 2 , 0 . 6 9 6 x - 1 . 2 4 9 y + 0 . 1 9 9 , - 0 . 7 5 2 x - 0 . 0 9 6 x + 0 . 0 9 6 x - 0 . 0 9 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 7 6 , 0 . 1 5 . 0 1 5 , 0 . 0 1 5 , 0 . 0 1 5 , 0 . 1 0 . 1 5 , 0 0 . 1 5 , 0 0 . 1 5 , 0 0 . 1 0 . 1 5 0 , 0 . 1 5 0 , 0 . 1 0 0 .", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 108, + 533, + 297, + 548 + ], + "spans": [ + { + "bbox": [ + 108, + 536, + 293, + 546 + ], + "score": 0.71, + "content": "0 . 8 8 0 y + 0 . 1 0 6 , - 0 . 9 5 6 x - 0 . 5 8 1 y + 0 . 2 2 3 ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 533, + 297, + 548 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 19, + "bbox_fs": [ + 104, + 403, + 511, + 548 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 426, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 427, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 427, + 564 + ], + "score": 1.0, + "content": "For the first principal unembedding dimension, it will be taken dot product with", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 567, + 501, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 501, + 579 + ], + "score": 1.0, + "content": "[1.326, 0.179, 0.142, −0.458, 1.101, −0.083, 0.621, 1.255, −0.709, 0.123, −1.346, −0.571, 1.016,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 578, + 490, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 490, + 590 + ], + "score": 1.0, + "content": "1.337, 0.732, 0.839, 0.129, 0.804, 0.377, 0.078, 1.322, −1.021, −0.799, −0.339, 1.117, −1.162,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 590, + 510, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 231, + 601 + ], + "score": 1.0, + "content": "−1.423, −1.157, 1.363, 0.156,", + "type": "text" + }, + { + "bbox": [ + 231, + 590, + 262, + 600 + ], + "score": 0.4, + "content": "- 0 . 1 6 5", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 590, + 265, + 601 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 266, + 590, + 297, + 600 + ], + "score": 0.26, + "content": "- 0 . 4 5 1", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 590, + 510, + 601 + ], + "score": 1.0, + "content": ", −1.101, −0.572, −1.180, −1.386, −1.346, −0.226,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 600, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 294, + 612 + ], + "score": 0.72, + "content": "1 . 0 9 1 , 1 . 1 5 9 , - 0 . 5 2 4 , 1 . 4 4 1 , - 0 . 9 4 9 , - 1 . 2 4 8 ] .", + "type": "inline_equation" + } + ], + "index": 30 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 550, + 427, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 566, + 509, + 612 + ], + "lines": [], + "index": 28.5, + "bbox_fs": [ + 105, + 567, + 510, + 612 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 506, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 507, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 181, + 630 + ], + "score": 1.0, + "content": "Call this function", + "type": "text" + }, + { + "bbox": [ + 181, + 616, + 211, + 628 + ], + "score": 0.91, + "content": "f ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 615, + 291, + 630 + ], + "score": 1.0, + "content": ". When we plug in", + "type": "text" + }, + { + "bbox": [ + 291, + 616, + 386, + 628 + ], + "score": 0.9, + "content": "x = \\cos ( t ) , y = \\sin ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 615, + 507, + 630 + ], + "score": 1.0, + "content": ", we get a function that well-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 165, + 639 + ], + "score": 1.0, + "content": "approximated", + "type": "text" + }, + { + "bbox": [ + 166, + 628, + 220, + 639 + ], + "score": 0.87, + "content": "8 \\cos ( 2 t + 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 627, + 334, + 639 + ], + "score": 1.0, + "content": "(Figure 26). Therefore, let", + "type": "text" + }, + { + "bbox": [ + 334, + 627, + 405, + 639 + ], + "score": 0.89, + "content": "t = w _ { k } ( a + b ) / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 627, + 506, + 639 + ], + "score": 1.0, + "content": ", the dot product will be", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 638, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 168, + 650 + ], + "score": 1.0, + "content": "approximately", + "type": "text" + }, + { + "bbox": [ + 168, + 638, + 335, + 650 + ], + "score": 0.9, + "content": "8 | \\cos ( w _ { k } ( a - b ) / 2 ) | \\cos ( w _ { k } ( a + b ) + 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 638, + 351, + 650 + ], + "score": 1.0, + "content": ", or", + "type": "text" + }, + { + "bbox": [ + 351, + 638, + 494, + 650 + ], + "score": 0.89, + "content": "| \\mathrm { c o s } ( w _ { k } ( a - b ) / 2 ) | \\mathrm { c o s } ( w _ { k } ( a + b ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 638, + 506, + 650 + ], + "score": 1.0, + "content": "if", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 649, + 430, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 430, + 662 + ], + "score": 1.0, + "content": "we ignore the phase and scaling. This completes the picture we described above.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 615, + 507, + 662 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 228, + 335, + 383, + 437 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 228, + 335, + 383, + 437 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 228, + 335, + 383, + 437 + ], + "spans": [ + { + "bbox": [ + 228, + 335, + 383, + 437 + ], + "score": 0.962, + "type": "image", + "image_path": "9d0d3575e176f1f2b5dddb06b63e304a3dfff3496d1d4fc543e627666cb32f3d.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 335, + 383, + 352.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 228, + 352.0, + 383, + 369.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 228, + 369.0, + 383, + 386.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 228, + 386.0, + 383, + 403.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 228, + 403.0, + 383, + 420.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 228, + 420.0, + 383, + 437.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 182, + 445, + 429, + 458 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 181, + 444, + 430, + 459 + ], + "spans": [ + { + "bbox": [ + 181, + 444, + 226, + 459 + ], + "score": 1.0, + "content": "Figure 26:", + "type": "text" + }, + { + "bbox": [ + 226, + 445, + 293, + 457 + ], + "score": 0.91, + "content": "f ( \\cos ( t ) , \\sin ( t ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 444, + 371, + 459 + ], + "score": 1.0, + "content": "well-approximates", + "type": "text" + }, + { + "bbox": [ + 372, + 445, + 426, + 457 + ], + "score": 0.91, + "content": "8 \\cos ( 2 t + 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 444, + 430, + 459 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 4.25 + } + ], + "page_idx": 27, + "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": "image", + "bbox": [ + 228, + 335, + 383, + 437 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 228, + 335, + 383, + 437 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 228, + 335, + 383, + 437 + ], + "spans": [ + { + "bbox": [ + 228, + 335, + 383, + 437 + ], + "score": 0.962, + "type": "image", + "image_path": "9d0d3575e176f1f2b5dddb06b63e304a3dfff3496d1d4fc543e627666cb32f3d.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 335, + 383, + 352.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 228, + 352.0, + 383, + 369.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 228, + 369.0, + 383, + 386.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 228, + 386.0, + 383, + 403.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 228, + 403.0, + 383, + 420.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 228, + 420.0, + 383, + 437.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 182, + 445, + 429, + 458 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 181, + 444, + 430, + 459 + ], + "spans": [ + { + "bbox": [ + 181, + 444, + 226, + 459 + ], + "score": 1.0, + "content": "Figure 26:", + "type": "text" + }, + { + "bbox": [ + 226, + 445, + 293, + 457 + ], + "score": 0.91, + "content": "f ( \\cos ( t ) , \\sin ( t ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 444, + 371, + 459 + ], + "score": 1.0, + "content": "well-approximates", + "type": "text" + }, + { + "bbox": [ + 372, + 445, + 426, + 457 + ], + "score": 0.91, + "content": "8 \\cos ( 2 t + 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 444, + 430, + 459 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 4.25 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file