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": "
| Algorithm | Learned Embeddings | GradientSymmetry | Required Non-linearity |
| Clock | Circle | No | Multiplication |
| Pizza | Circle | Yes | Absolute value |
| Non-circular | Line,Lissajous-like curves, etc. | N/A | N/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": "| Algorithm | Learned Embeddings | GradientSymmetry | Required Non-linearity |
| Clock | Circle | No | Multiplication |
| Pizza | Circle | Yes | Absolute value |
| Non-circular | Line,Lissajous-like curves, etc. | N/A | N/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": "| Circle | Wk | Ac FVE |
| #4 (accompanying #1) | 2π/59·17 | 97.56% |
| #5 (accompanying #2) | 2π/59.3 | 97.23% |
| #6 (accompanying #3) | 2π/59·44 | 97.69% |
",
+ "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": "| Circle | Wk | Ac FVE |
| #4 (accompanying #1) | 2π/59·17 | 97.56% |
| #5 (accompanying #2) | 2π/59.3 | 97.23% |
| #6 (accompanying #3) | 2π/59·44 | 97.69% |
",
+ "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